1 /* 2 * Copyright (c) 1997, 2017, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/assembler.hpp" 27 #include "asm/assembler.inline.hpp" 28 #include "compiler/disassembler.hpp" 29 #include "gc/shared/cardTableModRefBS.hpp" 30 #include "gc/shared/collectedHeap.inline.hpp" 31 #include "interpreter/interpreter.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "memory/universe.hpp" 34 #include "oops/klass.inline.hpp" 35 #include "prims/methodHandles.hpp" 36 #include "runtime/biasedLocking.hpp" 37 #include "runtime/interfaceSupport.hpp" 38 #include "runtime/objectMonitor.hpp" 39 #include "runtime/os.hpp" 40 #include "runtime/sharedRuntime.hpp" 41 #include "runtime/stubRoutines.hpp" 42 #include "runtime/thread.hpp" 43 #include "utilities/macros.hpp" 44 #if INCLUDE_ALL_GCS 45 #include "gc/g1/g1CollectedHeap.inline.hpp" 46 #include "gc/g1/g1SATBCardTableModRefBS.hpp" 47 #include "gc/g1/heapRegion.hpp" 48 #endif // INCLUDE_ALL_GCS 49 #include "crc32c.h" 50 #ifdef COMPILER2 51 #include "opto/intrinsicnode.hpp" 52 #endif 53 54 #ifdef PRODUCT 55 #define BLOCK_COMMENT(str) /* nothing */ 56 #define STOP(error) stop(error) 57 #else 58 #define BLOCK_COMMENT(str) block_comment(str) 59 #define STOP(error) block_comment(error); stop(error) 60 #endif 61 62 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":") 63 64 #ifdef ASSERT 65 bool AbstractAssembler::pd_check_instruction_mark() { return true; } 66 #endif 67 68 static Assembler::Condition reverse[] = { 69 Assembler::noOverflow /* overflow = 0x0 */ , 70 Assembler::overflow /* noOverflow = 0x1 */ , 71 Assembler::aboveEqual /* carrySet = 0x2, below = 0x2 */ , 72 Assembler::below /* aboveEqual = 0x3, carryClear = 0x3 */ , 73 Assembler::notZero /* zero = 0x4, equal = 0x4 */ , 74 Assembler::zero /* notZero = 0x5, notEqual = 0x5 */ , 75 Assembler::above /* belowEqual = 0x6 */ , 76 Assembler::belowEqual /* above = 0x7 */ , 77 Assembler::positive /* negative = 0x8 */ , 78 Assembler::negative /* positive = 0x9 */ , 79 Assembler::noParity /* parity = 0xa */ , 80 Assembler::parity /* noParity = 0xb */ , 81 Assembler::greaterEqual /* less = 0xc */ , 82 Assembler::less /* greaterEqual = 0xd */ , 83 Assembler::greater /* lessEqual = 0xe */ , 84 Assembler::lessEqual /* greater = 0xf, */ 85 86 }; 87 88 89 // Implementation of MacroAssembler 90 91 // First all the versions that have distinct versions depending on 32/64 bit 92 // Unless the difference is trivial (1 line or so). 93 94 #ifndef _LP64 95 96 // 32bit versions 97 98 Address MacroAssembler::as_Address(AddressLiteral adr) { 99 return Address(adr.target(), adr.rspec()); 100 } 101 102 Address MacroAssembler::as_Address(ArrayAddress adr) { 103 return Address::make_array(adr); 104 } 105 106 void MacroAssembler::call_VM_leaf_base(address entry_point, 107 int number_of_arguments) { 108 call(RuntimeAddress(entry_point)); 109 increment(rsp, number_of_arguments * wordSize); 110 } 111 112 void MacroAssembler::cmpklass(Address src1, Metadata* obj) { 113 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 114 } 115 116 void MacroAssembler::cmpklass(Register src1, Metadata* obj) { 117 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 118 } 119 120 void MacroAssembler::cmpoop(Address src1, jobject obj) { 121 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate()); 122 } 123 124 void MacroAssembler::cmpoop(Register src1, jobject obj) { 125 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate()); 126 } 127 128 void MacroAssembler::extend_sign(Register hi, Register lo) { 129 // According to Intel Doc. AP-526, "Integer Divide", p.18. 130 if (VM_Version::is_P6() && hi == rdx && lo == rax) { 131 cdql(); 132 } else { 133 movl(hi, lo); 134 sarl(hi, 31); 135 } 136 } 137 138 void MacroAssembler::jC2(Register tmp, Label& L) { 139 // set parity bit if FPU flag C2 is set (via rax) 140 save_rax(tmp); 141 fwait(); fnstsw_ax(); 142 sahf(); 143 restore_rax(tmp); 144 // branch 145 jcc(Assembler::parity, L); 146 } 147 148 void MacroAssembler::jnC2(Register tmp, Label& L) { 149 // set parity bit if FPU flag C2 is set (via rax) 150 save_rax(tmp); 151 fwait(); fnstsw_ax(); 152 sahf(); 153 restore_rax(tmp); 154 // branch 155 jcc(Assembler::noParity, L); 156 } 157 158 // 32bit can do a case table jump in one instruction but we no longer allow the base 159 // to be installed in the Address class 160 void MacroAssembler::jump(ArrayAddress entry) { 161 jmp(as_Address(entry)); 162 } 163 164 // Note: y_lo will be destroyed 165 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) { 166 // Long compare for Java (semantics as described in JVM spec.) 167 Label high, low, done; 168 169 cmpl(x_hi, y_hi); 170 jcc(Assembler::less, low); 171 jcc(Assembler::greater, high); 172 // x_hi is the return register 173 xorl(x_hi, x_hi); 174 cmpl(x_lo, y_lo); 175 jcc(Assembler::below, low); 176 jcc(Assembler::equal, done); 177 178 bind(high); 179 xorl(x_hi, x_hi); 180 increment(x_hi); 181 jmp(done); 182 183 bind(low); 184 xorl(x_hi, x_hi); 185 decrementl(x_hi); 186 187 bind(done); 188 } 189 190 void MacroAssembler::lea(Register dst, AddressLiteral src) { 191 mov_literal32(dst, (int32_t)src.target(), src.rspec()); 192 } 193 194 void MacroAssembler::lea(Address dst, AddressLiteral adr) { 195 // leal(dst, as_Address(adr)); 196 // see note in movl as to why we must use a move 197 mov_literal32(dst, (int32_t) adr.target(), adr.rspec()); 198 } 199 200 void MacroAssembler::leave() { 201 mov(rsp, rbp); 202 pop(rbp); 203 } 204 205 void MacroAssembler::lmul(int x_rsp_offset, int y_rsp_offset) { 206 // Multiplication of two Java long values stored on the stack 207 // as illustrated below. Result is in rdx:rax. 208 // 209 // rsp ---> [ ?? ] \ \ 210 // .... | y_rsp_offset | 211 // [ y_lo ] / (in bytes) | x_rsp_offset 212 // [ y_hi ] | (in bytes) 213 // .... | 214 // [ x_lo ] / 215 // [ x_hi ] 216 // .... 217 // 218 // Basic idea: lo(result) = lo(x_lo * y_lo) 219 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi) 220 Address x_hi(rsp, x_rsp_offset + wordSize); Address x_lo(rsp, x_rsp_offset); 221 Address y_hi(rsp, y_rsp_offset + wordSize); Address y_lo(rsp, y_rsp_offset); 222 Label quick; 223 // load x_hi, y_hi and check if quick 224 // multiplication is possible 225 movl(rbx, x_hi); 226 movl(rcx, y_hi); 227 movl(rax, rbx); 228 orl(rbx, rcx); // rbx, = 0 <=> x_hi = 0 and y_hi = 0 229 jcc(Assembler::zero, quick); // if rbx, = 0 do quick multiply 230 // do full multiplication 231 // 1st step 232 mull(y_lo); // x_hi * y_lo 233 movl(rbx, rax); // save lo(x_hi * y_lo) in rbx, 234 // 2nd step 235 movl(rax, x_lo); 236 mull(rcx); // x_lo * y_hi 237 addl(rbx, rax); // add lo(x_lo * y_hi) to rbx, 238 // 3rd step 239 bind(quick); // note: rbx, = 0 if quick multiply! 240 movl(rax, x_lo); 241 mull(y_lo); // x_lo * y_lo 242 addl(rdx, rbx); // correct hi(x_lo * y_lo) 243 } 244 245 void MacroAssembler::lneg(Register hi, Register lo) { 246 negl(lo); 247 adcl(hi, 0); 248 negl(hi); 249 } 250 251 void MacroAssembler::lshl(Register hi, Register lo) { 252 // Java shift left long support (semantics as described in JVM spec., p.305) 253 // (basic idea for shift counts s >= n: x << s == (x << n) << (s - n)) 254 // shift value is in rcx ! 255 assert(hi != rcx, "must not use rcx"); 256 assert(lo != rcx, "must not use rcx"); 257 const Register s = rcx; // shift count 258 const int n = BitsPerWord; 259 Label L; 260 andl(s, 0x3f); // s := s & 0x3f (s < 0x40) 261 cmpl(s, n); // if (s < n) 262 jcc(Assembler::less, L); // else (s >= n) 263 movl(hi, lo); // x := x << n 264 xorl(lo, lo); 265 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n! 266 bind(L); // s (mod n) < n 267 shldl(hi, lo); // x := x << s 268 shll(lo); 269 } 270 271 272 void MacroAssembler::lshr(Register hi, Register lo, bool sign_extension) { 273 // Java shift right long support (semantics as described in JVM spec., p.306 & p.310) 274 // (basic idea for shift counts s >= n: x >> s == (x >> n) >> (s - n)) 275 assert(hi != rcx, "must not use rcx"); 276 assert(lo != rcx, "must not use rcx"); 277 const Register s = rcx; // shift count 278 const int n = BitsPerWord; 279 Label L; 280 andl(s, 0x3f); // s := s & 0x3f (s < 0x40) 281 cmpl(s, n); // if (s < n) 282 jcc(Assembler::less, L); // else (s >= n) 283 movl(lo, hi); // x := x >> n 284 if (sign_extension) sarl(hi, 31); 285 else xorl(hi, hi); 286 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n! 287 bind(L); // s (mod n) < n 288 shrdl(lo, hi); // x := x >> s 289 if (sign_extension) sarl(hi); 290 else shrl(hi); 291 } 292 293 void MacroAssembler::movoop(Register dst, jobject obj) { 294 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate()); 295 } 296 297 void MacroAssembler::movoop(Address dst, jobject obj) { 298 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate()); 299 } 300 301 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) { 302 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 303 } 304 305 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) { 306 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 307 } 308 309 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) { 310 // scratch register is not used, 311 // it is defined to match parameters of 64-bit version of this method. 312 if (src.is_lval()) { 313 mov_literal32(dst, (intptr_t)src.target(), src.rspec()); 314 } else { 315 movl(dst, as_Address(src)); 316 } 317 } 318 319 void MacroAssembler::movptr(ArrayAddress dst, Register src) { 320 movl(as_Address(dst), src); 321 } 322 323 void MacroAssembler::movptr(Register dst, ArrayAddress src) { 324 movl(dst, as_Address(src)); 325 } 326 327 // src should NEVER be a real pointer. Use AddressLiteral for true pointers 328 void MacroAssembler::movptr(Address dst, intptr_t src) { 329 movl(dst, src); 330 } 331 332 333 void MacroAssembler::pop_callee_saved_registers() { 334 pop(rcx); 335 pop(rdx); 336 pop(rdi); 337 pop(rsi); 338 } 339 340 void MacroAssembler::pop_fTOS() { 341 fld_d(Address(rsp, 0)); 342 addl(rsp, 2 * wordSize); 343 } 344 345 void MacroAssembler::push_callee_saved_registers() { 346 push(rsi); 347 push(rdi); 348 push(rdx); 349 push(rcx); 350 } 351 352 void MacroAssembler::push_fTOS() { 353 subl(rsp, 2 * wordSize); 354 fstp_d(Address(rsp, 0)); 355 } 356 357 358 void MacroAssembler::pushoop(jobject obj) { 359 push_literal32((int32_t)obj, oop_Relocation::spec_for_immediate()); 360 } 361 362 void MacroAssembler::pushklass(Metadata* obj) { 363 push_literal32((int32_t)obj, metadata_Relocation::spec_for_immediate()); 364 } 365 366 void MacroAssembler::pushptr(AddressLiteral src) { 367 if (src.is_lval()) { 368 push_literal32((int32_t)src.target(), src.rspec()); 369 } else { 370 pushl(as_Address(src)); 371 } 372 } 373 374 void MacroAssembler::set_word_if_not_zero(Register dst) { 375 xorl(dst, dst); 376 set_byte_if_not_zero(dst); 377 } 378 379 static void pass_arg0(MacroAssembler* masm, Register arg) { 380 masm->push(arg); 381 } 382 383 static void pass_arg1(MacroAssembler* masm, Register arg) { 384 masm->push(arg); 385 } 386 387 static void pass_arg2(MacroAssembler* masm, Register arg) { 388 masm->push(arg); 389 } 390 391 static void pass_arg3(MacroAssembler* masm, Register arg) { 392 masm->push(arg); 393 } 394 395 #ifndef PRODUCT 396 extern "C" void findpc(intptr_t x); 397 #endif 398 399 void MacroAssembler::debug32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip, char* msg) { 400 // In order to get locks to work, we need to fake a in_VM state 401 JavaThread* thread = JavaThread::current(); 402 JavaThreadState saved_state = thread->thread_state(); 403 thread->set_thread_state(_thread_in_vm); 404 if (ShowMessageBoxOnError) { 405 JavaThread* thread = JavaThread::current(); 406 JavaThreadState saved_state = thread->thread_state(); 407 thread->set_thread_state(_thread_in_vm); 408 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) { 409 ttyLocker ttyl; 410 BytecodeCounter::print(); 411 } 412 // To see where a verify_oop failed, get $ebx+40/X for this frame. 413 // This is the value of eip which points to where verify_oop will return. 414 if (os::message_box(msg, "Execution stopped, print registers?")) { 415 print_state32(rdi, rsi, rbp, rsp, rbx, rdx, rcx, rax, eip); 416 BREAKPOINT; 417 } 418 } else { 419 ttyLocker ttyl; 420 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", msg); 421 } 422 // Don't assert holding the ttyLock 423 assert(false, "DEBUG MESSAGE: %s", msg); 424 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state); 425 } 426 427 void MacroAssembler::print_state32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip) { 428 ttyLocker ttyl; 429 FlagSetting fs(Debugging, true); 430 tty->print_cr("eip = 0x%08x", eip); 431 #ifndef PRODUCT 432 if ((WizardMode || Verbose) && PrintMiscellaneous) { 433 tty->cr(); 434 findpc(eip); 435 tty->cr(); 436 } 437 #endif 438 #define PRINT_REG(rax) \ 439 { tty->print("%s = ", #rax); os::print_location(tty, rax); } 440 PRINT_REG(rax); 441 PRINT_REG(rbx); 442 PRINT_REG(rcx); 443 PRINT_REG(rdx); 444 PRINT_REG(rdi); 445 PRINT_REG(rsi); 446 PRINT_REG(rbp); 447 PRINT_REG(rsp); 448 #undef PRINT_REG 449 // Print some words near top of staack. 450 int* dump_sp = (int*) rsp; 451 for (int col1 = 0; col1 < 8; col1++) { 452 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 453 os::print_location(tty, *dump_sp++); 454 } 455 for (int row = 0; row < 16; row++) { 456 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 457 for (int col = 0; col < 8; col++) { 458 tty->print(" 0x%08x", *dump_sp++); 459 } 460 tty->cr(); 461 } 462 // Print some instructions around pc: 463 Disassembler::decode((address)eip-64, (address)eip); 464 tty->print_cr("--------"); 465 Disassembler::decode((address)eip, (address)eip+32); 466 } 467 468 void MacroAssembler::stop(const char* msg) { 469 ExternalAddress message((address)msg); 470 // push address of message 471 pushptr(message.addr()); 472 { Label L; call(L, relocInfo::none); bind(L); } // push eip 473 pusha(); // push registers 474 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug32))); 475 hlt(); 476 } 477 478 void MacroAssembler::warn(const char* msg) { 479 push_CPU_state(); 480 481 ExternalAddress message((address) msg); 482 // push address of message 483 pushptr(message.addr()); 484 485 call(RuntimeAddress(CAST_FROM_FN_PTR(address, warning))); 486 addl(rsp, wordSize); // discard argument 487 pop_CPU_state(); 488 } 489 490 void MacroAssembler::print_state() { 491 { Label L; call(L, relocInfo::none); bind(L); } // push eip 492 pusha(); // push registers 493 494 push_CPU_state(); 495 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::print_state32))); 496 pop_CPU_state(); 497 498 popa(); 499 addl(rsp, wordSize); 500 } 501 502 #else // _LP64 503 504 // 64 bit versions 505 506 Address MacroAssembler::as_Address(AddressLiteral adr) { 507 // amd64 always does this as a pc-rel 508 // we can be absolute or disp based on the instruction type 509 // jmp/call are displacements others are absolute 510 assert(!adr.is_lval(), "must be rval"); 511 assert(reachable(adr), "must be"); 512 return Address((int32_t)(intptr_t)(adr.target() - pc()), adr.target(), adr.reloc()); 513 514 } 515 516 Address MacroAssembler::as_Address(ArrayAddress adr) { 517 AddressLiteral base = adr.base(); 518 lea(rscratch1, base); 519 Address index = adr.index(); 520 assert(index._disp == 0, "must not have disp"); // maybe it can? 521 Address array(rscratch1, index._index, index._scale, index._disp); 522 return array; 523 } 524 525 void MacroAssembler::call_VM_leaf_base(address entry_point, int num_args) { 526 Label L, E; 527 528 #ifdef _WIN64 529 // Windows always allocates space for it's register args 530 assert(num_args <= 4, "only register arguments supported"); 531 subq(rsp, frame::arg_reg_save_area_bytes); 532 #endif 533 534 // Align stack if necessary 535 testl(rsp, 15); 536 jcc(Assembler::zero, L); 537 538 subq(rsp, 8); 539 { 540 call(RuntimeAddress(entry_point)); 541 } 542 addq(rsp, 8); 543 jmp(E); 544 545 bind(L); 546 { 547 call(RuntimeAddress(entry_point)); 548 } 549 550 bind(E); 551 552 #ifdef _WIN64 553 // restore stack pointer 554 addq(rsp, frame::arg_reg_save_area_bytes); 555 #endif 556 557 } 558 559 void MacroAssembler::cmp64(Register src1, AddressLiteral src2) { 560 assert(!src2.is_lval(), "should use cmpptr"); 561 562 if (reachable(src2)) { 563 cmpq(src1, as_Address(src2)); 564 } else { 565 lea(rscratch1, src2); 566 Assembler::cmpq(src1, Address(rscratch1, 0)); 567 } 568 } 569 570 int MacroAssembler::corrected_idivq(Register reg) { 571 // Full implementation of Java ldiv and lrem; checks for special 572 // case as described in JVM spec., p.243 & p.271. The function 573 // returns the (pc) offset of the idivl instruction - may be needed 574 // for implicit exceptions. 575 // 576 // normal case special case 577 // 578 // input : rax: dividend min_long 579 // reg: divisor (may not be eax/edx) -1 580 // 581 // output: rax: quotient (= rax idiv reg) min_long 582 // rdx: remainder (= rax irem reg) 0 583 assert(reg != rax && reg != rdx, "reg cannot be rax or rdx register"); 584 static const int64_t min_long = 0x8000000000000000; 585 Label normal_case, special_case; 586 587 // check for special case 588 cmp64(rax, ExternalAddress((address) &min_long)); 589 jcc(Assembler::notEqual, normal_case); 590 xorl(rdx, rdx); // prepare rdx for possible special case (where 591 // remainder = 0) 592 cmpq(reg, -1); 593 jcc(Assembler::equal, special_case); 594 595 // handle normal case 596 bind(normal_case); 597 cdqq(); 598 int idivq_offset = offset(); 599 idivq(reg); 600 601 // normal and special case exit 602 bind(special_case); 603 604 return idivq_offset; 605 } 606 607 void MacroAssembler::decrementq(Register reg, int value) { 608 if (value == min_jint) { subq(reg, value); return; } 609 if (value < 0) { incrementq(reg, -value); return; } 610 if (value == 0) { ; return; } 611 if (value == 1 && UseIncDec) { decq(reg) ; return; } 612 /* else */ { subq(reg, value) ; return; } 613 } 614 615 void MacroAssembler::decrementq(Address dst, int value) { 616 if (value == min_jint) { subq(dst, value); return; } 617 if (value < 0) { incrementq(dst, -value); return; } 618 if (value == 0) { ; return; } 619 if (value == 1 && UseIncDec) { decq(dst) ; return; } 620 /* else */ { subq(dst, value) ; return; } 621 } 622 623 void MacroAssembler::incrementq(AddressLiteral dst) { 624 if (reachable(dst)) { 625 incrementq(as_Address(dst)); 626 } else { 627 lea(rscratch1, dst); 628 incrementq(Address(rscratch1, 0)); 629 } 630 } 631 632 void MacroAssembler::incrementq(Register reg, int value) { 633 if (value == min_jint) { addq(reg, value); return; } 634 if (value < 0) { decrementq(reg, -value); return; } 635 if (value == 0) { ; return; } 636 if (value == 1 && UseIncDec) { incq(reg) ; return; } 637 /* else */ { addq(reg, value) ; return; } 638 } 639 640 void MacroAssembler::incrementq(Address dst, int value) { 641 if (value == min_jint) { addq(dst, value); return; } 642 if (value < 0) { decrementq(dst, -value); return; } 643 if (value == 0) { ; return; } 644 if (value == 1 && UseIncDec) { incq(dst) ; return; } 645 /* else */ { addq(dst, value) ; return; } 646 } 647 648 // 32bit can do a case table jump in one instruction but we no longer allow the base 649 // to be installed in the Address class 650 void MacroAssembler::jump(ArrayAddress entry) { 651 lea(rscratch1, entry.base()); 652 Address dispatch = entry.index(); 653 assert(dispatch._base == noreg, "must be"); 654 dispatch._base = rscratch1; 655 jmp(dispatch); 656 } 657 658 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) { 659 ShouldNotReachHere(); // 64bit doesn't use two regs 660 cmpq(x_lo, y_lo); 661 } 662 663 void MacroAssembler::lea(Register dst, AddressLiteral src) { 664 mov_literal64(dst, (intptr_t)src.target(), src.rspec()); 665 } 666 667 void MacroAssembler::lea(Address dst, AddressLiteral adr) { 668 mov_literal64(rscratch1, (intptr_t)adr.target(), adr.rspec()); 669 movptr(dst, rscratch1); 670 } 671 672 void MacroAssembler::leave() { 673 // %%% is this really better? Why not on 32bit too? 674 emit_int8((unsigned char)0xC9); // LEAVE 675 } 676 677 void MacroAssembler::lneg(Register hi, Register lo) { 678 ShouldNotReachHere(); // 64bit doesn't use two regs 679 negq(lo); 680 } 681 682 void MacroAssembler::movoop(Register dst, jobject obj) { 683 mov_literal64(dst, (intptr_t)obj, oop_Relocation::spec_for_immediate()); 684 } 685 686 void MacroAssembler::movoop(Address dst, jobject obj) { 687 mov_literal64(rscratch1, (intptr_t)obj, oop_Relocation::spec_for_immediate()); 688 movq(dst, rscratch1); 689 } 690 691 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) { 692 mov_literal64(dst, (intptr_t)obj, metadata_Relocation::spec_for_immediate()); 693 } 694 695 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) { 696 mov_literal64(rscratch1, (intptr_t)obj, metadata_Relocation::spec_for_immediate()); 697 movq(dst, rscratch1); 698 } 699 700 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) { 701 if (src.is_lval()) { 702 mov_literal64(dst, (intptr_t)src.target(), src.rspec()); 703 } else { 704 if (reachable(src)) { 705 movq(dst, as_Address(src)); 706 } else { 707 lea(scratch, src); 708 movq(dst, Address(scratch, 0)); 709 } 710 } 711 } 712 713 void MacroAssembler::movptr(ArrayAddress dst, Register src) { 714 movq(as_Address(dst), src); 715 } 716 717 void MacroAssembler::movptr(Register dst, ArrayAddress src) { 718 movq(dst, as_Address(src)); 719 } 720 721 // src should NEVER be a real pointer. Use AddressLiteral for true pointers 722 void MacroAssembler::movptr(Address dst, intptr_t src) { 723 mov64(rscratch1, src); 724 movq(dst, rscratch1); 725 } 726 727 // These are mostly for initializing NULL 728 void MacroAssembler::movptr(Address dst, int32_t src) { 729 movslq(dst, src); 730 } 731 732 void MacroAssembler::movptr(Register dst, int32_t src) { 733 mov64(dst, (intptr_t)src); 734 } 735 736 void MacroAssembler::pushoop(jobject obj) { 737 movoop(rscratch1, obj); 738 push(rscratch1); 739 } 740 741 void MacroAssembler::pushklass(Metadata* obj) { 742 mov_metadata(rscratch1, obj); 743 push(rscratch1); 744 } 745 746 void MacroAssembler::pushptr(AddressLiteral src) { 747 lea(rscratch1, src); 748 if (src.is_lval()) { 749 push(rscratch1); 750 } else { 751 pushq(Address(rscratch1, 0)); 752 } 753 } 754 755 void MacroAssembler::reset_last_Java_frame(bool clear_fp) { 756 // we must set sp to zero to clear frame 757 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), NULL_WORD); 758 // must clear fp, so that compiled frames are not confused; it is 759 // possible that we need it only for debugging 760 if (clear_fp) { 761 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD); 762 } 763 764 // Always clear the pc because it could have been set by make_walkable() 765 movptr(Address(r15_thread, JavaThread::last_Java_pc_offset()), NULL_WORD); 766 vzeroupper(); 767 } 768 769 void MacroAssembler::set_last_Java_frame(Register last_java_sp, 770 Register last_java_fp, 771 address last_java_pc) { 772 vzeroupper(); 773 // determine last_java_sp register 774 if (!last_java_sp->is_valid()) { 775 last_java_sp = rsp; 776 } 777 778 // last_java_fp is optional 779 if (last_java_fp->is_valid()) { 780 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), 781 last_java_fp); 782 } 783 784 // last_java_pc is optional 785 if (last_java_pc != NULL) { 786 Address java_pc(r15_thread, 787 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()); 788 lea(rscratch1, InternalAddress(last_java_pc)); 789 movptr(java_pc, rscratch1); 790 } 791 792 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), last_java_sp); 793 } 794 795 static void pass_arg0(MacroAssembler* masm, Register arg) { 796 if (c_rarg0 != arg ) { 797 masm->mov(c_rarg0, arg); 798 } 799 } 800 801 static void pass_arg1(MacroAssembler* masm, Register arg) { 802 if (c_rarg1 != arg ) { 803 masm->mov(c_rarg1, arg); 804 } 805 } 806 807 static void pass_arg2(MacroAssembler* masm, Register arg) { 808 if (c_rarg2 != arg ) { 809 masm->mov(c_rarg2, arg); 810 } 811 } 812 813 static void pass_arg3(MacroAssembler* masm, Register arg) { 814 if (c_rarg3 != arg ) { 815 masm->mov(c_rarg3, arg); 816 } 817 } 818 819 void MacroAssembler::stop(const char* msg) { 820 address rip = pc(); 821 pusha(); // get regs on stack 822 lea(c_rarg0, ExternalAddress((address) msg)); 823 lea(c_rarg1, InternalAddress(rip)); 824 movq(c_rarg2, rsp); // pass pointer to regs array 825 andq(rsp, -16); // align stack as required by ABI 826 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64))); 827 hlt(); 828 } 829 830 void MacroAssembler::warn(const char* msg) { 831 push(rbp); 832 movq(rbp, rsp); 833 andq(rsp, -16); // align stack as required by push_CPU_state and call 834 push_CPU_state(); // keeps alignment at 16 bytes 835 lea(c_rarg0, ExternalAddress((address) msg)); 836 call_VM_leaf(CAST_FROM_FN_PTR(address, warning), c_rarg0); 837 pop_CPU_state(); 838 mov(rsp, rbp); 839 pop(rbp); 840 } 841 842 void MacroAssembler::print_state() { 843 address rip = pc(); 844 pusha(); // get regs on stack 845 push(rbp); 846 movq(rbp, rsp); 847 andq(rsp, -16); // align stack as required by push_CPU_state and call 848 push_CPU_state(); // keeps alignment at 16 bytes 849 850 lea(c_rarg0, InternalAddress(rip)); 851 lea(c_rarg1, Address(rbp, wordSize)); // pass pointer to regs array 852 call_VM_leaf(CAST_FROM_FN_PTR(address, MacroAssembler::print_state64), c_rarg0, c_rarg1); 853 854 pop_CPU_state(); 855 mov(rsp, rbp); 856 pop(rbp); 857 popa(); 858 } 859 860 #ifndef PRODUCT 861 extern "C" void findpc(intptr_t x); 862 #endif 863 864 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[]) { 865 // In order to get locks to work, we need to fake a in_VM state 866 if (ShowMessageBoxOnError) { 867 JavaThread* thread = JavaThread::current(); 868 JavaThreadState saved_state = thread->thread_state(); 869 thread->set_thread_state(_thread_in_vm); 870 #ifndef PRODUCT 871 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) { 872 ttyLocker ttyl; 873 BytecodeCounter::print(); 874 } 875 #endif 876 // To see where a verify_oop failed, get $ebx+40/X for this frame. 877 // XXX correct this offset for amd64 878 // This is the value of eip which points to where verify_oop will return. 879 if (os::message_box(msg, "Execution stopped, print registers?")) { 880 print_state64(pc, regs); 881 BREAKPOINT; 882 assert(false, "start up GDB"); 883 } 884 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state); 885 } else { 886 ttyLocker ttyl; 887 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", 888 msg); 889 assert(false, "DEBUG MESSAGE: %s", msg); 890 } 891 } 892 893 void MacroAssembler::print_state64(int64_t pc, int64_t regs[]) { 894 ttyLocker ttyl; 895 FlagSetting fs(Debugging, true); 896 tty->print_cr("rip = 0x%016lx", pc); 897 #ifndef PRODUCT 898 tty->cr(); 899 findpc(pc); 900 tty->cr(); 901 #endif 902 #define PRINT_REG(rax, value) \ 903 { tty->print("%s = ", #rax); os::print_location(tty, value); } 904 PRINT_REG(rax, regs[15]); 905 PRINT_REG(rbx, regs[12]); 906 PRINT_REG(rcx, regs[14]); 907 PRINT_REG(rdx, regs[13]); 908 PRINT_REG(rdi, regs[8]); 909 PRINT_REG(rsi, regs[9]); 910 PRINT_REG(rbp, regs[10]); 911 PRINT_REG(rsp, regs[11]); 912 PRINT_REG(r8 , regs[7]); 913 PRINT_REG(r9 , regs[6]); 914 PRINT_REG(r10, regs[5]); 915 PRINT_REG(r11, regs[4]); 916 PRINT_REG(r12, regs[3]); 917 PRINT_REG(r13, regs[2]); 918 PRINT_REG(r14, regs[1]); 919 PRINT_REG(r15, regs[0]); 920 #undef PRINT_REG 921 // Print some words near top of staack. 922 int64_t* rsp = (int64_t*) regs[11]; 923 int64_t* dump_sp = rsp; 924 for (int col1 = 0; col1 < 8; col1++) { 925 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (int64_t)dump_sp); 926 os::print_location(tty, *dump_sp++); 927 } 928 for (int row = 0; row < 25; row++) { 929 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (int64_t)dump_sp); 930 for (int col = 0; col < 4; col++) { 931 tty->print(" 0x%016lx", *dump_sp++); 932 } 933 tty->cr(); 934 } 935 // Print some instructions around pc: 936 Disassembler::decode((address)pc-64, (address)pc); 937 tty->print_cr("--------"); 938 Disassembler::decode((address)pc, (address)pc+32); 939 } 940 941 #endif // _LP64 942 943 // Now versions that are common to 32/64 bit 944 945 void MacroAssembler::addptr(Register dst, int32_t imm32) { 946 LP64_ONLY(addq(dst, imm32)) NOT_LP64(addl(dst, imm32)); 947 } 948 949 void MacroAssembler::addptr(Register dst, Register src) { 950 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src)); 951 } 952 953 void MacroAssembler::addptr(Address dst, Register src) { 954 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src)); 955 } 956 957 void MacroAssembler::addsd(XMMRegister dst, AddressLiteral src) { 958 if (reachable(src)) { 959 Assembler::addsd(dst, as_Address(src)); 960 } else { 961 lea(rscratch1, src); 962 Assembler::addsd(dst, Address(rscratch1, 0)); 963 } 964 } 965 966 void MacroAssembler::addss(XMMRegister dst, AddressLiteral src) { 967 if (reachable(src)) { 968 addss(dst, as_Address(src)); 969 } else { 970 lea(rscratch1, src); 971 addss(dst, Address(rscratch1, 0)); 972 } 973 } 974 975 void MacroAssembler::addpd(XMMRegister dst, AddressLiteral src) { 976 if (reachable(src)) { 977 Assembler::addpd(dst, as_Address(src)); 978 } else { 979 lea(rscratch1, src); 980 Assembler::addpd(dst, Address(rscratch1, 0)); 981 } 982 } 983 984 void MacroAssembler::align(int modulus) { 985 align(modulus, offset()); 986 } 987 988 void MacroAssembler::align(int modulus, int target) { 989 if (target % modulus != 0) { 990 nop(modulus - (target % modulus)); 991 } 992 } 993 994 void MacroAssembler::andpd(XMMRegister dst, AddressLiteral src) { 995 // Used in sign-masking with aligned address. 996 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 997 if (reachable(src)) { 998 Assembler::andpd(dst, as_Address(src)); 999 } else { 1000 lea(rscratch1, src); 1001 Assembler::andpd(dst, Address(rscratch1, 0)); 1002 } 1003 } 1004 1005 void MacroAssembler::andps(XMMRegister dst, AddressLiteral src) { 1006 // Used in sign-masking with aligned address. 1007 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 1008 if (reachable(src)) { 1009 Assembler::andps(dst, as_Address(src)); 1010 } else { 1011 lea(rscratch1, src); 1012 Assembler::andps(dst, Address(rscratch1, 0)); 1013 } 1014 } 1015 1016 void MacroAssembler::andptr(Register dst, int32_t imm32) { 1017 LP64_ONLY(andq(dst, imm32)) NOT_LP64(andl(dst, imm32)); 1018 } 1019 1020 void MacroAssembler::atomic_incl(Address counter_addr) { 1021 if (os::is_MP()) 1022 lock(); 1023 incrementl(counter_addr); 1024 } 1025 1026 void MacroAssembler::atomic_incl(AddressLiteral counter_addr, Register scr) { 1027 if (reachable(counter_addr)) { 1028 atomic_incl(as_Address(counter_addr)); 1029 } else { 1030 lea(scr, counter_addr); 1031 atomic_incl(Address(scr, 0)); 1032 } 1033 } 1034 1035 #ifdef _LP64 1036 void MacroAssembler::atomic_incq(Address counter_addr) { 1037 if (os::is_MP()) 1038 lock(); 1039 incrementq(counter_addr); 1040 } 1041 1042 void MacroAssembler::atomic_incq(AddressLiteral counter_addr, Register scr) { 1043 if (reachable(counter_addr)) { 1044 atomic_incq(as_Address(counter_addr)); 1045 } else { 1046 lea(scr, counter_addr); 1047 atomic_incq(Address(scr, 0)); 1048 } 1049 } 1050 #endif 1051 1052 // Writes to stack successive pages until offset reached to check for 1053 // stack overflow + shadow pages. This clobbers tmp. 1054 void MacroAssembler::bang_stack_size(Register size, Register tmp) { 1055 movptr(tmp, rsp); 1056 // Bang stack for total size given plus shadow page size. 1057 // Bang one page at a time because large size can bang beyond yellow and 1058 // red zones. 1059 Label loop; 1060 bind(loop); 1061 movl(Address(tmp, (-os::vm_page_size())), size ); 1062 subptr(tmp, os::vm_page_size()); 1063 subl(size, os::vm_page_size()); 1064 jcc(Assembler::greater, loop); 1065 1066 // Bang down shadow pages too. 1067 // At this point, (tmp-0) is the last address touched, so don't 1068 // touch it again. (It was touched as (tmp-pagesize) but then tmp 1069 // was post-decremented.) Skip this address by starting at i=1, and 1070 // touch a few more pages below. N.B. It is important to touch all 1071 // the way down including all pages in the shadow zone. 1072 for (int i = 1; i < ((int)JavaThread::stack_shadow_zone_size() / os::vm_page_size()); i++) { 1073 // this could be any sized move but this is can be a debugging crumb 1074 // so the bigger the better. 1075 movptr(Address(tmp, (-i*os::vm_page_size())), size ); 1076 } 1077 } 1078 1079 void MacroAssembler::reserved_stack_check() { 1080 // testing if reserved zone needs to be enabled 1081 Label no_reserved_zone_enabling; 1082 Register thread = NOT_LP64(rsi) LP64_ONLY(r15_thread); 1083 NOT_LP64(get_thread(rsi);) 1084 1085 cmpptr(rsp, Address(thread, JavaThread::reserved_stack_activation_offset())); 1086 jcc(Assembler::below, no_reserved_zone_enabling); 1087 1088 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), thread); 1089 jump(RuntimeAddress(StubRoutines::throw_delayed_StackOverflowError_entry())); 1090 should_not_reach_here(); 1091 1092 bind(no_reserved_zone_enabling); 1093 } 1094 1095 int MacroAssembler::biased_locking_enter(Register lock_reg, 1096 Register obj_reg, 1097 Register swap_reg, 1098 Register tmp_reg, 1099 bool swap_reg_contains_mark, 1100 Label& done, 1101 Label* slow_case, 1102 BiasedLockingCounters* counters) { 1103 assert(UseBiasedLocking, "why call this otherwise?"); 1104 assert(swap_reg == rax, "swap_reg must be rax for cmpxchgq"); 1105 assert(tmp_reg != noreg, "tmp_reg must be supplied"); 1106 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg); 1107 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout"); 1108 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes()); 1109 NOT_LP64( Address saved_mark_addr(lock_reg, 0); ) 1110 1111 if (PrintBiasedLockingStatistics && counters == NULL) { 1112 counters = BiasedLocking::counters(); 1113 } 1114 // Biased locking 1115 // See whether the lock is currently biased toward our thread and 1116 // whether the epoch is still valid 1117 // Note that the runtime guarantees sufficient alignment of JavaThread 1118 // pointers to allow age to be placed into low bits 1119 // First check to see whether biasing is even enabled for this object 1120 Label cas_label; 1121 int null_check_offset = -1; 1122 if (!swap_reg_contains_mark) { 1123 null_check_offset = offset(); 1124 movptr(swap_reg, mark_addr); 1125 } 1126 movptr(tmp_reg, swap_reg); 1127 andptr(tmp_reg, markOopDesc::biased_lock_mask_in_place); 1128 cmpptr(tmp_reg, markOopDesc::biased_lock_pattern); 1129 jcc(Assembler::notEqual, cas_label); 1130 // The bias pattern is present in the object's header. Need to check 1131 // whether the bias owner and the epoch are both still current. 1132 #ifndef _LP64 1133 // Note that because there is no current thread register on x86_32 we 1134 // need to store off the mark word we read out of the object to 1135 // avoid reloading it and needing to recheck invariants below. This 1136 // store is unfortunate but it makes the overall code shorter and 1137 // simpler. 1138 movptr(saved_mark_addr, swap_reg); 1139 #endif 1140 if (swap_reg_contains_mark) { 1141 null_check_offset = offset(); 1142 } 1143 load_prototype_header(tmp_reg, obj_reg); 1144 #ifdef _LP64 1145 orptr(tmp_reg, r15_thread); 1146 xorptr(tmp_reg, swap_reg); 1147 Register header_reg = tmp_reg; 1148 #else 1149 xorptr(tmp_reg, swap_reg); 1150 get_thread(swap_reg); 1151 xorptr(swap_reg, tmp_reg); 1152 Register header_reg = swap_reg; 1153 #endif 1154 andptr(header_reg, ~((int) markOopDesc::age_mask_in_place)); 1155 if (counters != NULL) { 1156 cond_inc32(Assembler::zero, 1157 ExternalAddress((address) counters->biased_lock_entry_count_addr())); 1158 } 1159 jcc(Assembler::equal, done); 1160 1161 Label try_revoke_bias; 1162 Label try_rebias; 1163 1164 // At this point we know that the header has the bias pattern and 1165 // that we are not the bias owner in the current epoch. We need to 1166 // figure out more details about the state of the header in order to 1167 // know what operations can be legally performed on the object's 1168 // header. 1169 1170 // If the low three bits in the xor result aren't clear, that means 1171 // the prototype header is no longer biased and we have to revoke 1172 // the bias on this object. 1173 testptr(header_reg, markOopDesc::biased_lock_mask_in_place); 1174 jccb(Assembler::notZero, try_revoke_bias); 1175 1176 // Biasing is still enabled for this data type. See whether the 1177 // epoch of the current bias is still valid, meaning that the epoch 1178 // bits of the mark word are equal to the epoch bits of the 1179 // prototype header. (Note that the prototype header's epoch bits 1180 // only change at a safepoint.) If not, attempt to rebias the object 1181 // toward the current thread. Note that we must be absolutely sure 1182 // that the current epoch is invalid in order to do this because 1183 // otherwise the manipulations it performs on the mark word are 1184 // illegal. 1185 testptr(header_reg, markOopDesc::epoch_mask_in_place); 1186 jccb(Assembler::notZero, try_rebias); 1187 1188 // The epoch of the current bias is still valid but we know nothing 1189 // about the owner; it might be set or it might be clear. Try to 1190 // acquire the bias of the object using an atomic operation. If this 1191 // fails we will go in to the runtime to revoke the object's bias. 1192 // Note that we first construct the presumed unbiased header so we 1193 // don't accidentally blow away another thread's valid bias. 1194 NOT_LP64( movptr(swap_reg, saved_mark_addr); ) 1195 andptr(swap_reg, 1196 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place); 1197 #ifdef _LP64 1198 movptr(tmp_reg, swap_reg); 1199 orptr(tmp_reg, r15_thread); 1200 #else 1201 get_thread(tmp_reg); 1202 orptr(tmp_reg, swap_reg); 1203 #endif 1204 if (os::is_MP()) { 1205 lock(); 1206 } 1207 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1208 // If the biasing toward our thread failed, this means that 1209 // another thread succeeded in biasing it toward itself and we 1210 // need to revoke that bias. The revocation will occur in the 1211 // interpreter runtime in the slow case. 1212 if (counters != NULL) { 1213 cond_inc32(Assembler::zero, 1214 ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr())); 1215 } 1216 if (slow_case != NULL) { 1217 jcc(Assembler::notZero, *slow_case); 1218 } 1219 jmp(done); 1220 1221 bind(try_rebias); 1222 // At this point we know the epoch has expired, meaning that the 1223 // current "bias owner", if any, is actually invalid. Under these 1224 // circumstances _only_, we are allowed to use the current header's 1225 // value as the comparison value when doing the cas to acquire the 1226 // bias in the current epoch. In other words, we allow transfer of 1227 // the bias from one thread to another directly in this situation. 1228 // 1229 // FIXME: due to a lack of registers we currently blow away the age 1230 // bits in this situation. Should attempt to preserve them. 1231 load_prototype_header(tmp_reg, obj_reg); 1232 #ifdef _LP64 1233 orptr(tmp_reg, r15_thread); 1234 #else 1235 get_thread(swap_reg); 1236 orptr(tmp_reg, swap_reg); 1237 movptr(swap_reg, saved_mark_addr); 1238 #endif 1239 if (os::is_MP()) { 1240 lock(); 1241 } 1242 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1243 // If the biasing toward our thread failed, then another thread 1244 // succeeded in biasing it toward itself and we need to revoke that 1245 // bias. The revocation will occur in the runtime in the slow case. 1246 if (counters != NULL) { 1247 cond_inc32(Assembler::zero, 1248 ExternalAddress((address) counters->rebiased_lock_entry_count_addr())); 1249 } 1250 if (slow_case != NULL) { 1251 jcc(Assembler::notZero, *slow_case); 1252 } 1253 jmp(done); 1254 1255 bind(try_revoke_bias); 1256 // The prototype mark in the klass doesn't have the bias bit set any 1257 // more, indicating that objects of this data type are not supposed 1258 // to be biased any more. We are going to try to reset the mark of 1259 // this object to the prototype value and fall through to the 1260 // CAS-based locking scheme. Note that if our CAS fails, it means 1261 // that another thread raced us for the privilege of revoking the 1262 // bias of this particular object, so it's okay to continue in the 1263 // normal locking code. 1264 // 1265 // FIXME: due to a lack of registers we currently blow away the age 1266 // bits in this situation. Should attempt to preserve them. 1267 NOT_LP64( movptr(swap_reg, saved_mark_addr); ) 1268 load_prototype_header(tmp_reg, obj_reg); 1269 if (os::is_MP()) { 1270 lock(); 1271 } 1272 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1273 // Fall through to the normal CAS-based lock, because no matter what 1274 // the result of the above CAS, some thread must have succeeded in 1275 // removing the bias bit from the object's header. 1276 if (counters != NULL) { 1277 cond_inc32(Assembler::zero, 1278 ExternalAddress((address) counters->revoked_lock_entry_count_addr())); 1279 } 1280 1281 bind(cas_label); 1282 1283 return null_check_offset; 1284 } 1285 1286 void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) { 1287 assert(UseBiasedLocking, "why call this otherwise?"); 1288 1289 // Check for biased locking unlock case, which is a no-op 1290 // Note: we do not have to check the thread ID for two reasons. 1291 // First, the interpreter checks for IllegalMonitorStateException at 1292 // a higher level. Second, if the bias was revoked while we held the 1293 // lock, the object could not be rebiased toward another thread, so 1294 // the bias bit would be clear. 1295 movptr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 1296 andptr(temp_reg, markOopDesc::biased_lock_mask_in_place); 1297 cmpptr(temp_reg, markOopDesc::biased_lock_pattern); 1298 jcc(Assembler::equal, done); 1299 } 1300 1301 #ifdef COMPILER2 1302 1303 #if INCLUDE_RTM_OPT 1304 1305 // Update rtm_counters based on abort status 1306 // input: abort_status 1307 // rtm_counters (RTMLockingCounters*) 1308 // flags are killed 1309 void MacroAssembler::rtm_counters_update(Register abort_status, Register rtm_counters) { 1310 1311 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abort_count_offset())); 1312 if (PrintPreciseRTMLockingStatistics) { 1313 for (int i = 0; i < RTMLockingCounters::ABORT_STATUS_LIMIT; i++) { 1314 Label check_abort; 1315 testl(abort_status, (1<<i)); 1316 jccb(Assembler::equal, check_abort); 1317 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abortX_count_offset() + (i * sizeof(uintx)))); 1318 bind(check_abort); 1319 } 1320 } 1321 } 1322 1323 // Branch if (random & (count-1) != 0), count is 2^n 1324 // tmp, scr and flags are killed 1325 void MacroAssembler::branch_on_random_using_rdtsc(Register tmp, Register scr, int count, Label& brLabel) { 1326 assert(tmp == rax, ""); 1327 assert(scr == rdx, ""); 1328 rdtsc(); // modifies EDX:EAX 1329 andptr(tmp, count-1); 1330 jccb(Assembler::notZero, brLabel); 1331 } 1332 1333 // Perform abort ratio calculation, set no_rtm bit if high ratio 1334 // input: rtm_counters_Reg (RTMLockingCounters* address) 1335 // tmpReg, rtm_counters_Reg and flags are killed 1336 void MacroAssembler::rtm_abort_ratio_calculation(Register tmpReg, 1337 Register rtm_counters_Reg, 1338 RTMLockingCounters* rtm_counters, 1339 Metadata* method_data) { 1340 Label L_done, L_check_always_rtm1, L_check_always_rtm2; 1341 1342 if (RTMLockingCalculationDelay > 0) { 1343 // Delay calculation 1344 movptr(tmpReg, ExternalAddress((address) RTMLockingCounters::rtm_calculation_flag_addr()), tmpReg); 1345 testptr(tmpReg, tmpReg); 1346 jccb(Assembler::equal, L_done); 1347 } 1348 // Abort ratio calculation only if abort_count > RTMAbortThreshold 1349 // Aborted transactions = abort_count * 100 1350 // All transactions = total_count * RTMTotalCountIncrRate 1351 // Set no_rtm bit if (Aborted transactions >= All transactions * RTMAbortRatio) 1352 1353 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::abort_count_offset())); 1354 cmpptr(tmpReg, RTMAbortThreshold); 1355 jccb(Assembler::below, L_check_always_rtm2); 1356 imulptr(tmpReg, tmpReg, 100); 1357 1358 Register scrReg = rtm_counters_Reg; 1359 movptr(scrReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset())); 1360 imulptr(scrReg, scrReg, RTMTotalCountIncrRate); 1361 imulptr(scrReg, scrReg, RTMAbortRatio); 1362 cmpptr(tmpReg, scrReg); 1363 jccb(Assembler::below, L_check_always_rtm1); 1364 if (method_data != NULL) { 1365 // set rtm_state to "no rtm" in MDO 1366 mov_metadata(tmpReg, method_data); 1367 if (os::is_MP()) { 1368 lock(); 1369 } 1370 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), NoRTM); 1371 } 1372 jmpb(L_done); 1373 bind(L_check_always_rtm1); 1374 // Reload RTMLockingCounters* address 1375 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters)); 1376 bind(L_check_always_rtm2); 1377 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset())); 1378 cmpptr(tmpReg, RTMLockingThreshold / RTMTotalCountIncrRate); 1379 jccb(Assembler::below, L_done); 1380 if (method_data != NULL) { 1381 // set rtm_state to "always rtm" in MDO 1382 mov_metadata(tmpReg, method_data); 1383 if (os::is_MP()) { 1384 lock(); 1385 } 1386 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), UseRTM); 1387 } 1388 bind(L_done); 1389 } 1390 1391 // Update counters and perform abort ratio calculation 1392 // input: abort_status_Reg 1393 // rtm_counters_Reg, flags are killed 1394 void MacroAssembler::rtm_profiling(Register abort_status_Reg, 1395 Register rtm_counters_Reg, 1396 RTMLockingCounters* rtm_counters, 1397 Metadata* method_data, 1398 bool profile_rtm) { 1399 1400 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1401 // update rtm counters based on rax value at abort 1402 // reads abort_status_Reg, updates flags 1403 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters)); 1404 rtm_counters_update(abort_status_Reg, rtm_counters_Reg); 1405 if (profile_rtm) { 1406 // Save abort status because abort_status_Reg is used by following code. 1407 if (RTMRetryCount > 0) { 1408 push(abort_status_Reg); 1409 } 1410 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1411 rtm_abort_ratio_calculation(abort_status_Reg, rtm_counters_Reg, rtm_counters, method_data); 1412 // restore abort status 1413 if (RTMRetryCount > 0) { 1414 pop(abort_status_Reg); 1415 } 1416 } 1417 } 1418 1419 // Retry on abort if abort's status is 0x6: can retry (0x2) | memory conflict (0x4) 1420 // inputs: retry_count_Reg 1421 // : abort_status_Reg 1422 // output: retry_count_Reg decremented by 1 1423 // flags are killed 1424 void MacroAssembler::rtm_retry_lock_on_abort(Register retry_count_Reg, Register abort_status_Reg, Label& retryLabel) { 1425 Label doneRetry; 1426 assert(abort_status_Reg == rax, ""); 1427 // The abort reason bits are in eax (see all states in rtmLocking.hpp) 1428 // 0x6 = conflict on which we can retry (0x2) | memory conflict (0x4) 1429 // if reason is in 0x6 and retry count != 0 then retry 1430 andptr(abort_status_Reg, 0x6); 1431 jccb(Assembler::zero, doneRetry); 1432 testl(retry_count_Reg, retry_count_Reg); 1433 jccb(Assembler::zero, doneRetry); 1434 pause(); 1435 decrementl(retry_count_Reg); 1436 jmp(retryLabel); 1437 bind(doneRetry); 1438 } 1439 1440 // Spin and retry if lock is busy, 1441 // inputs: box_Reg (monitor address) 1442 // : retry_count_Reg 1443 // output: retry_count_Reg decremented by 1 1444 // : clear z flag if retry count exceeded 1445 // tmp_Reg, scr_Reg, flags are killed 1446 void MacroAssembler::rtm_retry_lock_on_busy(Register retry_count_Reg, Register box_Reg, 1447 Register tmp_Reg, Register scr_Reg, Label& retryLabel) { 1448 Label SpinLoop, SpinExit, doneRetry; 1449 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1450 1451 testl(retry_count_Reg, retry_count_Reg); 1452 jccb(Assembler::zero, doneRetry); 1453 decrementl(retry_count_Reg); 1454 movptr(scr_Reg, RTMSpinLoopCount); 1455 1456 bind(SpinLoop); 1457 pause(); 1458 decrementl(scr_Reg); 1459 jccb(Assembler::lessEqual, SpinExit); 1460 movptr(tmp_Reg, Address(box_Reg, owner_offset)); 1461 testptr(tmp_Reg, tmp_Reg); 1462 jccb(Assembler::notZero, SpinLoop); 1463 1464 bind(SpinExit); 1465 jmp(retryLabel); 1466 bind(doneRetry); 1467 incrementl(retry_count_Reg); // clear z flag 1468 } 1469 1470 // Use RTM for normal stack locks 1471 // Input: objReg (object to lock) 1472 void MacroAssembler::rtm_stack_locking(Register objReg, Register tmpReg, Register scrReg, 1473 Register retry_on_abort_count_Reg, 1474 RTMLockingCounters* stack_rtm_counters, 1475 Metadata* method_data, bool profile_rtm, 1476 Label& DONE_LABEL, Label& IsInflated) { 1477 assert(UseRTMForStackLocks, "why call this otherwise?"); 1478 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking"); 1479 assert(tmpReg == rax, ""); 1480 assert(scrReg == rdx, ""); 1481 Label L_rtm_retry, L_decrement_retry, L_on_abort; 1482 1483 if (RTMRetryCount > 0) { 1484 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort 1485 bind(L_rtm_retry); 1486 } 1487 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); 1488 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased 1489 jcc(Assembler::notZero, IsInflated); 1490 1491 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1492 Label L_noincrement; 1493 if (RTMTotalCountIncrRate > 1) { 1494 // tmpReg, scrReg and flags are killed 1495 branch_on_random_using_rdtsc(tmpReg, scrReg, (int)RTMTotalCountIncrRate, L_noincrement); 1496 } 1497 assert(stack_rtm_counters != NULL, "should not be NULL when profiling RTM"); 1498 atomic_incptr(ExternalAddress((address)stack_rtm_counters->total_count_addr()), scrReg); 1499 bind(L_noincrement); 1500 } 1501 xbegin(L_on_abort); 1502 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // fetch markword 1503 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits 1504 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked 1505 jcc(Assembler::equal, DONE_LABEL); // all done if unlocked 1506 1507 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX 1508 if (UseRTMXendForLockBusy) { 1509 xend(); 1510 movptr(abort_status_Reg, 0x2); // Set the abort status to 2 (so we can retry) 1511 jmp(L_decrement_retry); 1512 } 1513 else { 1514 xabort(0); 1515 } 1516 bind(L_on_abort); 1517 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1518 rtm_profiling(abort_status_Reg, scrReg, stack_rtm_counters, method_data, profile_rtm); 1519 } 1520 bind(L_decrement_retry); 1521 if (RTMRetryCount > 0) { 1522 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4) 1523 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry); 1524 } 1525 } 1526 1527 // Use RTM for inflating locks 1528 // inputs: objReg (object to lock) 1529 // boxReg (on-stack box address (displaced header location) - KILLED) 1530 // tmpReg (ObjectMonitor address + markOopDesc::monitor_value) 1531 void MacroAssembler::rtm_inflated_locking(Register objReg, Register boxReg, Register tmpReg, 1532 Register scrReg, Register retry_on_busy_count_Reg, 1533 Register retry_on_abort_count_Reg, 1534 RTMLockingCounters* rtm_counters, 1535 Metadata* method_data, bool profile_rtm, 1536 Label& DONE_LABEL) { 1537 assert(UseRTMLocking, "why call this otherwise?"); 1538 assert(tmpReg == rax, ""); 1539 assert(scrReg == rdx, ""); 1540 Label L_rtm_retry, L_decrement_retry, L_on_abort; 1541 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1542 1543 // Without cast to int32_t a movptr will destroy r10 which is typically obj 1544 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark())); 1545 movptr(boxReg, tmpReg); // Save ObjectMonitor address 1546 1547 if (RTMRetryCount > 0) { 1548 movl(retry_on_busy_count_Reg, RTMRetryCount); // Retry on lock busy 1549 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort 1550 bind(L_rtm_retry); 1551 } 1552 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1553 Label L_noincrement; 1554 if (RTMTotalCountIncrRate > 1) { 1555 // tmpReg, scrReg and flags are killed 1556 branch_on_random_using_rdtsc(tmpReg, scrReg, (int)RTMTotalCountIncrRate, L_noincrement); 1557 } 1558 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1559 atomic_incptr(ExternalAddress((address)rtm_counters->total_count_addr()), scrReg); 1560 bind(L_noincrement); 1561 } 1562 xbegin(L_on_abort); 1563 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); 1564 movptr(tmpReg, Address(tmpReg, owner_offset)); 1565 testptr(tmpReg, tmpReg); 1566 jcc(Assembler::zero, DONE_LABEL); 1567 if (UseRTMXendForLockBusy) { 1568 xend(); 1569 jmp(L_decrement_retry); 1570 } 1571 else { 1572 xabort(0); 1573 } 1574 bind(L_on_abort); 1575 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX 1576 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1577 rtm_profiling(abort_status_Reg, scrReg, rtm_counters, method_data, profile_rtm); 1578 } 1579 if (RTMRetryCount > 0) { 1580 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4) 1581 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry); 1582 } 1583 1584 movptr(tmpReg, Address(boxReg, owner_offset)) ; 1585 testptr(tmpReg, tmpReg) ; 1586 jccb(Assembler::notZero, L_decrement_retry) ; 1587 1588 // Appears unlocked - try to swing _owner from null to non-null. 1589 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1590 #ifdef _LP64 1591 Register threadReg = r15_thread; 1592 #else 1593 get_thread(scrReg); 1594 Register threadReg = scrReg; 1595 #endif 1596 if (os::is_MP()) { 1597 lock(); 1598 } 1599 cmpxchgptr(threadReg, Address(boxReg, owner_offset)); // Updates tmpReg 1600 1601 if (RTMRetryCount > 0) { 1602 // success done else retry 1603 jccb(Assembler::equal, DONE_LABEL) ; 1604 bind(L_decrement_retry); 1605 // Spin and retry if lock is busy. 1606 rtm_retry_lock_on_busy(retry_on_busy_count_Reg, boxReg, tmpReg, scrReg, L_rtm_retry); 1607 } 1608 else { 1609 bind(L_decrement_retry); 1610 } 1611 } 1612 1613 #endif // INCLUDE_RTM_OPT 1614 1615 // Fast_Lock and Fast_Unlock used by C2 1616 1617 // Because the transitions from emitted code to the runtime 1618 // monitorenter/exit helper stubs are so slow it's critical that 1619 // we inline both the stack-locking fast-path and the inflated fast path. 1620 // 1621 // See also: cmpFastLock and cmpFastUnlock. 1622 // 1623 // What follows is a specialized inline transliteration of the code 1624 // in slow_enter() and slow_exit(). If we're concerned about I$ bloat 1625 // another option would be to emit TrySlowEnter and TrySlowExit methods 1626 // at startup-time. These methods would accept arguments as 1627 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure 1628 // indications in the icc.ZFlag. Fast_Lock and Fast_Unlock would simply 1629 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit. 1630 // In practice, however, the # of lock sites is bounded and is usually small. 1631 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer 1632 // if the processor uses simple bimodal branch predictors keyed by EIP 1633 // Since the helper routines would be called from multiple synchronization 1634 // sites. 1635 // 1636 // An even better approach would be write "MonitorEnter()" and "MonitorExit()" 1637 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites 1638 // to those specialized methods. That'd give us a mostly platform-independent 1639 // implementation that the JITs could optimize and inline at their pleasure. 1640 // Done correctly, the only time we'd need to cross to native could would be 1641 // to park() or unpark() threads. We'd also need a few more unsafe operators 1642 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and 1643 // (b) explicit barriers or fence operations. 1644 // 1645 // TODO: 1646 // 1647 // * Arrange for C2 to pass "Self" into Fast_Lock and Fast_Unlock in one of the registers (scr). 1648 // This avoids manifesting the Self pointer in the Fast_Lock and Fast_Unlock terminals. 1649 // Given TLAB allocation, Self is usually manifested in a register, so passing it into 1650 // the lock operators would typically be faster than reifying Self. 1651 // 1652 // * Ideally I'd define the primitives as: 1653 // fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED. 1654 // fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED 1655 // Unfortunately ADLC bugs prevent us from expressing the ideal form. 1656 // Instead, we're stuck with a rather awkward and brittle register assignments below. 1657 // Furthermore the register assignments are overconstrained, possibly resulting in 1658 // sub-optimal code near the synchronization site. 1659 // 1660 // * Eliminate the sp-proximity tests and just use "== Self" tests instead. 1661 // Alternately, use a better sp-proximity test. 1662 // 1663 // * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value. 1664 // Either one is sufficient to uniquely identify a thread. 1665 // TODO: eliminate use of sp in _owner and use get_thread(tr) instead. 1666 // 1667 // * Intrinsify notify() and notifyAll() for the common cases where the 1668 // object is locked by the calling thread but the waitlist is empty. 1669 // avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll(). 1670 // 1671 // * use jccb and jmpb instead of jcc and jmp to improve code density. 1672 // But beware of excessive branch density on AMD Opterons. 1673 // 1674 // * Both Fast_Lock and Fast_Unlock set the ICC.ZF to indicate success 1675 // or failure of the fast-path. If the fast-path fails then we pass 1676 // control to the slow-path, typically in C. In Fast_Lock and 1677 // Fast_Unlock we often branch to DONE_LABEL, just to find that C2 1678 // will emit a conditional branch immediately after the node. 1679 // So we have branches to branches and lots of ICC.ZF games. 1680 // Instead, it might be better to have C2 pass a "FailureLabel" 1681 // into Fast_Lock and Fast_Unlock. In the case of success, control 1682 // will drop through the node. ICC.ZF is undefined at exit. 1683 // In the case of failure, the node will branch directly to the 1684 // FailureLabel 1685 1686 1687 // obj: object to lock 1688 // box: on-stack box address (displaced header location) - KILLED 1689 // rax,: tmp -- KILLED 1690 // scr: tmp -- KILLED 1691 void MacroAssembler::fast_lock(Register objReg, Register boxReg, Register tmpReg, 1692 Register scrReg, Register cx1Reg, Register cx2Reg, 1693 BiasedLockingCounters* counters, 1694 RTMLockingCounters* rtm_counters, 1695 RTMLockingCounters* stack_rtm_counters, 1696 Metadata* method_data, 1697 bool use_rtm, bool profile_rtm) { 1698 // Ensure the register assignments are disjoint 1699 assert(tmpReg == rax, ""); 1700 1701 if (use_rtm) { 1702 assert_different_registers(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg); 1703 } else { 1704 assert(cx1Reg == noreg, ""); 1705 assert(cx2Reg == noreg, ""); 1706 assert_different_registers(objReg, boxReg, tmpReg, scrReg); 1707 } 1708 1709 if (counters != NULL) { 1710 atomic_incl(ExternalAddress((address)counters->total_entry_count_addr()), scrReg); 1711 } 1712 if (EmitSync & 1) { 1713 // set box->dhw = markOopDesc::unused_mark() 1714 // Force all sync thru slow-path: slow_enter() and slow_exit() 1715 movptr (Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark())); 1716 cmpptr (rsp, (int32_t)NULL_WORD); 1717 } else { 1718 // Possible cases that we'll encounter in fast_lock 1719 // ------------------------------------------------ 1720 // * Inflated 1721 // -- unlocked 1722 // -- Locked 1723 // = by self 1724 // = by other 1725 // * biased 1726 // -- by Self 1727 // -- by other 1728 // * neutral 1729 // * stack-locked 1730 // -- by self 1731 // = sp-proximity test hits 1732 // = sp-proximity test generates false-negative 1733 // -- by other 1734 // 1735 1736 Label IsInflated, DONE_LABEL; 1737 1738 // it's stack-locked, biased or neutral 1739 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage 1740 // order to reduce the number of conditional branches in the most common cases. 1741 // Beware -- there's a subtle invariant that fetch of the markword 1742 // at [FETCH], below, will never observe a biased encoding (*101b). 1743 // If this invariant is not held we risk exclusion (safety) failure. 1744 if (UseBiasedLocking && !UseOptoBiasInlining) { 1745 biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, counters); 1746 } 1747 1748 #if INCLUDE_RTM_OPT 1749 if (UseRTMForStackLocks && use_rtm) { 1750 rtm_stack_locking(objReg, tmpReg, scrReg, cx2Reg, 1751 stack_rtm_counters, method_data, profile_rtm, 1752 DONE_LABEL, IsInflated); 1753 } 1754 #endif // INCLUDE_RTM_OPT 1755 1756 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // [FETCH] 1757 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased 1758 jccb(Assembler::notZero, IsInflated); 1759 1760 // Attempt stack-locking ... 1761 orptr (tmpReg, markOopDesc::unlocked_value); 1762 movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS 1763 if (os::is_MP()) { 1764 lock(); 1765 } 1766 cmpxchgptr(boxReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Updates tmpReg 1767 if (counters != NULL) { 1768 cond_inc32(Assembler::equal, 1769 ExternalAddress((address)counters->fast_path_entry_count_addr())); 1770 } 1771 jcc(Assembler::equal, DONE_LABEL); // Success 1772 1773 // Recursive locking. 1774 // The object is stack-locked: markword contains stack pointer to BasicLock. 1775 // Locked by current thread if difference with current SP is less than one page. 1776 subptr(tmpReg, rsp); 1777 // Next instruction set ZFlag == 1 (Success) if difference is less then one page. 1778 andptr(tmpReg, (int32_t) (NOT_LP64(0xFFFFF003) LP64_ONLY(7 - os::vm_page_size())) ); 1779 movptr(Address(boxReg, 0), tmpReg); 1780 if (counters != NULL) { 1781 cond_inc32(Assembler::equal, 1782 ExternalAddress((address)counters->fast_path_entry_count_addr())); 1783 } 1784 jmp(DONE_LABEL); 1785 1786 bind(IsInflated); 1787 // The object is inflated. tmpReg contains pointer to ObjectMonitor* + markOopDesc::monitor_value 1788 1789 #if INCLUDE_RTM_OPT 1790 // Use the same RTM locking code in 32- and 64-bit VM. 1791 if (use_rtm) { 1792 rtm_inflated_locking(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg, 1793 rtm_counters, method_data, profile_rtm, DONE_LABEL); 1794 } else { 1795 #endif // INCLUDE_RTM_OPT 1796 1797 #ifndef _LP64 1798 // The object is inflated. 1799 1800 // boxReg refers to the on-stack BasicLock in the current frame. 1801 // We'd like to write: 1802 // set box->_displaced_header = markOopDesc::unused_mark(). Any non-0 value suffices. 1803 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers 1804 // additional latency as we have another ST in the store buffer that must drain. 1805 1806 if (EmitSync & 8192) { 1807 movptr(Address(boxReg, 0), 3); // results in ST-before-CAS penalty 1808 get_thread (scrReg); 1809 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2] 1810 movptr(tmpReg, NULL_WORD); // consider: xor vs mov 1811 if (os::is_MP()) { 1812 lock(); 1813 } 1814 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1815 } else 1816 if ((EmitSync & 128) == 0) { // avoid ST-before-CAS 1817 // register juggle because we need tmpReg for cmpxchgptr below 1818 movptr(scrReg, boxReg); 1819 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2] 1820 1821 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes 1822 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) { 1823 // prefetchw [eax + Offset(_owner)-2] 1824 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1825 } 1826 1827 if ((EmitSync & 64) == 0) { 1828 // Optimistic form: consider XORL tmpReg,tmpReg 1829 movptr(tmpReg, NULL_WORD); 1830 } else { 1831 // Can suffer RTS->RTO upgrades on shared or cold $ lines 1832 // Test-And-CAS instead of CAS 1833 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner 1834 testptr(tmpReg, tmpReg); // Locked ? 1835 jccb (Assembler::notZero, DONE_LABEL); 1836 } 1837 1838 // Appears unlocked - try to swing _owner from null to non-null. 1839 // Ideally, I'd manifest "Self" with get_thread and then attempt 1840 // to CAS the register containing Self into m->Owner. 1841 // But we don't have enough registers, so instead we can either try to CAS 1842 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds 1843 // we later store "Self" into m->Owner. Transiently storing a stack address 1844 // (rsp or the address of the box) into m->owner is harmless. 1845 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1846 if (os::is_MP()) { 1847 lock(); 1848 } 1849 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1850 movptr(Address(scrReg, 0), 3); // box->_displaced_header = 3 1851 // If we weren't able to swing _owner from NULL to the BasicLock 1852 // then take the slow path. 1853 jccb (Assembler::notZero, DONE_LABEL); 1854 // update _owner from BasicLock to thread 1855 get_thread (scrReg); // beware: clobbers ICCs 1856 movptr(Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), scrReg); 1857 xorptr(boxReg, boxReg); // set icc.ZFlag = 1 to indicate success 1858 1859 // If the CAS fails we can either retry or pass control to the slow-path. 1860 // We use the latter tactic. 1861 // Pass the CAS result in the icc.ZFlag into DONE_LABEL 1862 // If the CAS was successful ... 1863 // Self has acquired the lock 1864 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it. 1865 // Intentional fall-through into DONE_LABEL ... 1866 } else { 1867 movptr(Address(boxReg, 0), intptr_t(markOopDesc::unused_mark())); // results in ST-before-CAS penalty 1868 movptr(boxReg, tmpReg); 1869 1870 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes 1871 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) { 1872 // prefetchw [eax + Offset(_owner)-2] 1873 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1874 } 1875 1876 if ((EmitSync & 64) == 0) { 1877 // Optimistic form 1878 xorptr (tmpReg, tmpReg); 1879 } else { 1880 // Can suffer RTS->RTO upgrades on shared or cold $ lines 1881 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner 1882 testptr(tmpReg, tmpReg); // Locked ? 1883 jccb (Assembler::notZero, DONE_LABEL); 1884 } 1885 1886 // Appears unlocked - try to swing _owner from null to non-null. 1887 // Use either "Self" (in scr) or rsp as thread identity in _owner. 1888 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1889 get_thread (scrReg); 1890 if (os::is_MP()) { 1891 lock(); 1892 } 1893 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1894 1895 // If the CAS fails we can either retry or pass control to the slow-path. 1896 // We use the latter tactic. 1897 // Pass the CAS result in the icc.ZFlag into DONE_LABEL 1898 // If the CAS was successful ... 1899 // Self has acquired the lock 1900 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it. 1901 // Intentional fall-through into DONE_LABEL ... 1902 } 1903 #else // _LP64 1904 // It's inflated 1905 movq(scrReg, tmpReg); 1906 xorq(tmpReg, tmpReg); 1907 1908 if (os::is_MP()) { 1909 lock(); 1910 } 1911 cmpxchgptr(r15_thread, Address(scrReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1912 // Unconditionally set box->_displaced_header = markOopDesc::unused_mark(). 1913 // Without cast to int32_t movptr will destroy r10 which is typically obj. 1914 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark())); 1915 // Intentional fall-through into DONE_LABEL ... 1916 // Propagate ICC.ZF from CAS above into DONE_LABEL. 1917 #endif // _LP64 1918 #if INCLUDE_RTM_OPT 1919 } // use_rtm() 1920 #endif 1921 // DONE_LABEL is a hot target - we'd really like to place it at the 1922 // start of cache line by padding with NOPs. 1923 // See the AMD and Intel software optimization manuals for the 1924 // most efficient "long" NOP encodings. 1925 // Unfortunately none of our alignment mechanisms suffice. 1926 bind(DONE_LABEL); 1927 1928 // At DONE_LABEL the icc ZFlag is set as follows ... 1929 // Fast_Unlock uses the same protocol. 1930 // ZFlag == 1 -> Success 1931 // ZFlag == 0 -> Failure - force control through the slow-path 1932 } 1933 } 1934 1935 // obj: object to unlock 1936 // box: box address (displaced header location), killed. Must be EAX. 1937 // tmp: killed, cannot be obj nor box. 1938 // 1939 // Some commentary on balanced locking: 1940 // 1941 // Fast_Lock and Fast_Unlock are emitted only for provably balanced lock sites. 1942 // Methods that don't have provably balanced locking are forced to run in the 1943 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock. 1944 // The interpreter provides two properties: 1945 // I1: At return-time the interpreter automatically and quietly unlocks any 1946 // objects acquired the current activation (frame). Recall that the 1947 // interpreter maintains an on-stack list of locks currently held by 1948 // a frame. 1949 // I2: If a method attempts to unlock an object that is not held by the 1950 // the frame the interpreter throws IMSX. 1951 // 1952 // Lets say A(), which has provably balanced locking, acquires O and then calls B(). 1953 // B() doesn't have provably balanced locking so it runs in the interpreter. 1954 // Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O 1955 // is still locked by A(). 1956 // 1957 // The only other source of unbalanced locking would be JNI. The "Java Native Interface: 1958 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter 1959 // should not be unlocked by "normal" java-level locking and vice-versa. The specification 1960 // doesn't specify what will occur if a program engages in such mixed-mode locking, however. 1961 // Arguably given that the spec legislates the JNI case as undefined our implementation 1962 // could reasonably *avoid* checking owner in Fast_Unlock(). 1963 // In the interest of performance we elide m->Owner==Self check in unlock. 1964 // A perfectly viable alternative is to elide the owner check except when 1965 // Xcheck:jni is enabled. 1966 1967 void MacroAssembler::fast_unlock(Register objReg, Register boxReg, Register tmpReg, bool use_rtm) { 1968 assert(boxReg == rax, ""); 1969 assert_different_registers(objReg, boxReg, tmpReg); 1970 1971 if (EmitSync & 4) { 1972 // Disable - inhibit all inlining. Force control through the slow-path 1973 cmpptr (rsp, 0); 1974 } else { 1975 Label DONE_LABEL, Stacked, CheckSucc; 1976 1977 // Critically, the biased locking test must have precedence over 1978 // and appear before the (box->dhw == 0) recursive stack-lock test. 1979 if (UseBiasedLocking && !UseOptoBiasInlining) { 1980 biased_locking_exit(objReg, tmpReg, DONE_LABEL); 1981 } 1982 1983 #if INCLUDE_RTM_OPT 1984 if (UseRTMForStackLocks && use_rtm) { 1985 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking"); 1986 Label L_regular_unlock; 1987 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // fetch markword 1988 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits 1989 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked 1990 jccb(Assembler::notEqual, L_regular_unlock); // if !HLE RegularLock 1991 xend(); // otherwise end... 1992 jmp(DONE_LABEL); // ... and we're done 1993 bind(L_regular_unlock); 1994 } 1995 #endif 1996 1997 cmpptr(Address(boxReg, 0), (int32_t)NULL_WORD); // Examine the displaced header 1998 jcc (Assembler::zero, DONE_LABEL); // 0 indicates recursive stack-lock 1999 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Examine the object's markword 2000 testptr(tmpReg, markOopDesc::monitor_value); // Inflated? 2001 jccb (Assembler::zero, Stacked); 2002 2003 // It's inflated. 2004 #if INCLUDE_RTM_OPT 2005 if (use_rtm) { 2006 Label L_regular_inflated_unlock; 2007 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 2008 movptr(boxReg, Address(tmpReg, owner_offset)); 2009 testptr(boxReg, boxReg); 2010 jccb(Assembler::notZero, L_regular_inflated_unlock); 2011 xend(); 2012 jmpb(DONE_LABEL); 2013 bind(L_regular_inflated_unlock); 2014 } 2015 #endif 2016 2017 // Despite our balanced locking property we still check that m->_owner == Self 2018 // as java routines or native JNI code called by this thread might 2019 // have released the lock. 2020 // Refer to the comments in synchronizer.cpp for how we might encode extra 2021 // state in _succ so we can avoid fetching EntryList|cxq. 2022 // 2023 // I'd like to add more cases in fast_lock() and fast_unlock() -- 2024 // such as recursive enter and exit -- but we have to be wary of 2025 // I$ bloat, T$ effects and BP$ effects. 2026 // 2027 // If there's no contention try a 1-0 exit. That is, exit without 2028 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how 2029 // we detect and recover from the race that the 1-0 exit admits. 2030 // 2031 // Conceptually Fast_Unlock() must execute a STST|LDST "release" barrier 2032 // before it STs null into _owner, releasing the lock. Updates 2033 // to data protected by the critical section must be visible before 2034 // we drop the lock (and thus before any other thread could acquire 2035 // the lock and observe the fields protected by the lock). 2036 // IA32's memory-model is SPO, so STs are ordered with respect to 2037 // each other and there's no need for an explicit barrier (fence). 2038 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 2039 #ifndef _LP64 2040 get_thread (boxReg); 2041 if ((EmitSync & 4096) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) { 2042 // prefetchw [ebx + Offset(_owner)-2] 2043 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2044 } 2045 2046 // Note that we could employ various encoding schemes to reduce 2047 // the number of loads below (currently 4) to just 2 or 3. 2048 // Refer to the comments in synchronizer.cpp. 2049 // In practice the chain of fetches doesn't seem to impact performance, however. 2050 xorptr(boxReg, boxReg); 2051 if ((EmitSync & 65536) == 0 && (EmitSync & 256)) { 2052 // Attempt to reduce branch density - AMD's branch predictor. 2053 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 2054 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2055 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2056 jccb (Assembler::notZero, DONE_LABEL); 2057 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD); 2058 jmpb (DONE_LABEL); 2059 } else { 2060 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 2061 jccb (Assembler::notZero, DONE_LABEL); 2062 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2063 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2064 jccb (Assembler::notZero, CheckSucc); 2065 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD); 2066 jmpb (DONE_LABEL); 2067 } 2068 2069 // The Following code fragment (EmitSync & 65536) improves the performance of 2070 // contended applications and contended synchronization microbenchmarks. 2071 // Unfortunately the emission of the code - even though not executed - causes regressions 2072 // in scimark and jetstream, evidently because of $ effects. Replacing the code 2073 // with an equal number of never-executed NOPs results in the same regression. 2074 // We leave it off by default. 2075 2076 if ((EmitSync & 65536) != 0) { 2077 Label LSuccess, LGoSlowPath ; 2078 2079 bind (CheckSucc); 2080 2081 // Optional pre-test ... it's safe to elide this 2082 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2083 jccb(Assembler::zero, LGoSlowPath); 2084 2085 // We have a classic Dekker-style idiom: 2086 // ST m->_owner = 0 ; MEMBAR; LD m->_succ 2087 // There are a number of ways to implement the barrier: 2088 // (1) lock:andl &m->_owner, 0 2089 // is fast, but mask doesn't currently support the "ANDL M,IMM32" form. 2090 // LOCK: ANDL [ebx+Offset(_Owner)-2], 0 2091 // Encodes as 81 31 OFF32 IMM32 or 83 63 OFF8 IMM8 2092 // (2) If supported, an explicit MFENCE is appealing. 2093 // In older IA32 processors MFENCE is slower than lock:add or xchg 2094 // particularly if the write-buffer is full as might be the case if 2095 // if stores closely precede the fence or fence-equivalent instruction. 2096 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences 2097 // as the situation has changed with Nehalem and Shanghai. 2098 // (3) In lieu of an explicit fence, use lock:addl to the top-of-stack 2099 // The $lines underlying the top-of-stack should be in M-state. 2100 // The locked add instruction is serializing, of course. 2101 // (4) Use xchg, which is serializing 2102 // mov boxReg, 0; xchgl boxReg, [tmpReg + Offset(_owner)-2] also works 2103 // (5) ST m->_owner = 0 and then execute lock:orl &m->_succ, 0. 2104 // The integer condition codes will tell us if succ was 0. 2105 // Since _succ and _owner should reside in the same $line and 2106 // we just stored into _owner, it's likely that the $line 2107 // remains in M-state for the lock:orl. 2108 // 2109 // We currently use (3), although it's likely that switching to (2) 2110 // is correct for the future. 2111 2112 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD); 2113 if (os::is_MP()) { 2114 lock(); addptr(Address(rsp, 0), 0); 2115 } 2116 // Ratify _succ remains non-null 2117 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), 0); 2118 jccb (Assembler::notZero, LSuccess); 2119 2120 xorptr(boxReg, boxReg); // box is really EAX 2121 if (os::is_MP()) { lock(); } 2122 cmpxchgptr(rsp, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2123 // There's no successor so we tried to regrab the lock with the 2124 // placeholder value. If that didn't work, then another thread 2125 // grabbed the lock so we're done (and exit was a success). 2126 jccb (Assembler::notEqual, LSuccess); 2127 // Since we're low on registers we installed rsp as a placeholding in _owner. 2128 // Now install Self over rsp. This is safe as we're transitioning from 2129 // non-null to non=null 2130 get_thread (boxReg); 2131 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), boxReg); 2132 // Intentional fall-through into LGoSlowPath ... 2133 2134 bind (LGoSlowPath); 2135 orptr(boxReg, 1); // set ICC.ZF=0 to indicate failure 2136 jmpb (DONE_LABEL); 2137 2138 bind (LSuccess); 2139 xorptr(boxReg, boxReg); // set ICC.ZF=1 to indicate success 2140 jmpb (DONE_LABEL); 2141 } 2142 2143 bind (Stacked); 2144 // It's not inflated and it's not recursively stack-locked and it's not biased. 2145 // It must be stack-locked. 2146 // Try to reset the header to displaced header. 2147 // The "box" value on the stack is stable, so we can reload 2148 // and be assured we observe the same value as above. 2149 movptr(tmpReg, Address(boxReg, 0)); 2150 if (os::is_MP()) { 2151 lock(); 2152 } 2153 cmpxchgptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Uses RAX which is box 2154 // Intention fall-thru into DONE_LABEL 2155 2156 // DONE_LABEL is a hot target - we'd really like to place it at the 2157 // start of cache line by padding with NOPs. 2158 // See the AMD and Intel software optimization manuals for the 2159 // most efficient "long" NOP encodings. 2160 // Unfortunately none of our alignment mechanisms suffice. 2161 if ((EmitSync & 65536) == 0) { 2162 bind (CheckSucc); 2163 } 2164 #else // _LP64 2165 // It's inflated 2166 if (EmitSync & 1024) { 2167 // Emit code to check that _owner == Self 2168 // We could fold the _owner test into subsequent code more efficiently 2169 // than using a stand-alone check, but since _owner checking is off by 2170 // default we don't bother. We also might consider predicating the 2171 // _owner==Self check on Xcheck:jni or running on a debug build. 2172 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2173 xorptr(boxReg, r15_thread); 2174 } else { 2175 xorptr(boxReg, boxReg); 2176 } 2177 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 2178 jccb (Assembler::notZero, DONE_LABEL); 2179 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2180 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2181 jccb (Assembler::notZero, CheckSucc); 2182 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD); 2183 jmpb (DONE_LABEL); 2184 2185 if ((EmitSync & 65536) == 0) { 2186 // Try to avoid passing control into the slow_path ... 2187 Label LSuccess, LGoSlowPath ; 2188 bind (CheckSucc); 2189 2190 // The following optional optimization can be elided if necessary 2191 // Effectively: if (succ == null) goto SlowPath 2192 // The code reduces the window for a race, however, 2193 // and thus benefits performance. 2194 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2195 jccb (Assembler::zero, LGoSlowPath); 2196 2197 xorptr(boxReg, boxReg); 2198 if ((EmitSync & 16) && os::is_MP()) { 2199 xchgptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2200 } else { 2201 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD); 2202 if (os::is_MP()) { 2203 // Memory barrier/fence 2204 // Dekker pivot point -- fulcrum : ST Owner; MEMBAR; LD Succ 2205 // Instead of MFENCE we use a dummy locked add of 0 to the top-of-stack. 2206 // This is faster on Nehalem and AMD Shanghai/Barcelona. 2207 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences 2208 // We might also restructure (ST Owner=0;barrier;LD _Succ) to 2209 // (mov box,0; xchgq box, &m->Owner; LD _succ) . 2210 lock(); addl(Address(rsp, 0), 0); 2211 } 2212 } 2213 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2214 jccb (Assembler::notZero, LSuccess); 2215 2216 // Rare inopportune interleaving - race. 2217 // The successor vanished in the small window above. 2218 // The lock is contended -- (cxq|EntryList) != null -- and there's no apparent successor. 2219 // We need to ensure progress and succession. 2220 // Try to reacquire the lock. 2221 // If that fails then the new owner is responsible for succession and this 2222 // thread needs to take no further action and can exit via the fast path (success). 2223 // If the re-acquire succeeds then pass control into the slow path. 2224 // As implemented, this latter mode is horrible because we generated more 2225 // coherence traffic on the lock *and* artifically extended the critical section 2226 // length while by virtue of passing control into the slow path. 2227 2228 // box is really RAX -- the following CMPXCHG depends on that binding 2229 // cmpxchg R,[M] is equivalent to rax = CAS(M,rax,R) 2230 if (os::is_MP()) { lock(); } 2231 cmpxchgptr(r15_thread, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2232 // There's no successor so we tried to regrab the lock. 2233 // If that didn't work, then another thread grabbed the 2234 // lock so we're done (and exit was a success). 2235 jccb (Assembler::notEqual, LSuccess); 2236 // Intentional fall-through into slow-path 2237 2238 bind (LGoSlowPath); 2239 orl (boxReg, 1); // set ICC.ZF=0 to indicate failure 2240 jmpb (DONE_LABEL); 2241 2242 bind (LSuccess); 2243 testl (boxReg, 0); // set ICC.ZF=1 to indicate success 2244 jmpb (DONE_LABEL); 2245 } 2246 2247 bind (Stacked); 2248 movptr(tmpReg, Address (boxReg, 0)); // re-fetch 2249 if (os::is_MP()) { lock(); } 2250 cmpxchgptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Uses RAX which is box 2251 2252 if (EmitSync & 65536) { 2253 bind (CheckSucc); 2254 } 2255 #endif 2256 bind(DONE_LABEL); 2257 } 2258 } 2259 #endif // COMPILER2 2260 2261 void MacroAssembler::c2bool(Register x) { 2262 // implements x == 0 ? 0 : 1 2263 // note: must only look at least-significant byte of x 2264 // since C-style booleans are stored in one byte 2265 // only! (was bug) 2266 andl(x, 0xFF); 2267 setb(Assembler::notZero, x); 2268 } 2269 2270 // Wouldn't need if AddressLiteral version had new name 2271 void MacroAssembler::call(Label& L, relocInfo::relocType rtype) { 2272 Assembler::call(L, rtype); 2273 } 2274 2275 void MacroAssembler::call(Register entry) { 2276 Assembler::call(entry); 2277 } 2278 2279 void MacroAssembler::call(AddressLiteral entry) { 2280 if (reachable(entry)) { 2281 Assembler::call_literal(entry.target(), entry.rspec()); 2282 } else { 2283 lea(rscratch1, entry); 2284 Assembler::call(rscratch1); 2285 } 2286 } 2287 2288 void MacroAssembler::ic_call(address entry, jint method_index) { 2289 RelocationHolder rh = virtual_call_Relocation::spec(pc(), method_index); 2290 movptr(rax, (intptr_t)Universe::non_oop_word()); 2291 call(AddressLiteral(entry, rh)); 2292 } 2293 2294 // Implementation of call_VM versions 2295 2296 void MacroAssembler::call_VM(Register oop_result, 2297 address entry_point, 2298 bool check_exceptions) { 2299 Label C, E; 2300 call(C, relocInfo::none); 2301 jmp(E); 2302 2303 bind(C); 2304 call_VM_helper(oop_result, entry_point, 0, check_exceptions); 2305 ret(0); 2306 2307 bind(E); 2308 } 2309 2310 void MacroAssembler::call_VM(Register oop_result, 2311 address entry_point, 2312 Register arg_1, 2313 bool check_exceptions) { 2314 Label C, E; 2315 call(C, relocInfo::none); 2316 jmp(E); 2317 2318 bind(C); 2319 pass_arg1(this, arg_1); 2320 call_VM_helper(oop_result, entry_point, 1, check_exceptions); 2321 ret(0); 2322 2323 bind(E); 2324 } 2325 2326 void MacroAssembler::call_VM(Register oop_result, 2327 address entry_point, 2328 Register arg_1, 2329 Register arg_2, 2330 bool check_exceptions) { 2331 Label C, E; 2332 call(C, relocInfo::none); 2333 jmp(E); 2334 2335 bind(C); 2336 2337 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2338 2339 pass_arg2(this, arg_2); 2340 pass_arg1(this, arg_1); 2341 call_VM_helper(oop_result, entry_point, 2, check_exceptions); 2342 ret(0); 2343 2344 bind(E); 2345 } 2346 2347 void MacroAssembler::call_VM(Register oop_result, 2348 address entry_point, 2349 Register arg_1, 2350 Register arg_2, 2351 Register arg_3, 2352 bool check_exceptions) { 2353 Label C, E; 2354 call(C, relocInfo::none); 2355 jmp(E); 2356 2357 bind(C); 2358 2359 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2360 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2361 pass_arg3(this, arg_3); 2362 2363 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2364 pass_arg2(this, arg_2); 2365 2366 pass_arg1(this, arg_1); 2367 call_VM_helper(oop_result, entry_point, 3, check_exceptions); 2368 ret(0); 2369 2370 bind(E); 2371 } 2372 2373 void MacroAssembler::call_VM(Register oop_result, 2374 Register last_java_sp, 2375 address entry_point, 2376 int number_of_arguments, 2377 bool check_exceptions) { 2378 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg); 2379 call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions); 2380 } 2381 2382 void MacroAssembler::call_VM(Register oop_result, 2383 Register last_java_sp, 2384 address entry_point, 2385 Register arg_1, 2386 bool check_exceptions) { 2387 pass_arg1(this, arg_1); 2388 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 2389 } 2390 2391 void MacroAssembler::call_VM(Register oop_result, 2392 Register last_java_sp, 2393 address entry_point, 2394 Register arg_1, 2395 Register arg_2, 2396 bool check_exceptions) { 2397 2398 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2399 pass_arg2(this, arg_2); 2400 pass_arg1(this, arg_1); 2401 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 2402 } 2403 2404 void MacroAssembler::call_VM(Register oop_result, 2405 Register last_java_sp, 2406 address entry_point, 2407 Register arg_1, 2408 Register arg_2, 2409 Register arg_3, 2410 bool check_exceptions) { 2411 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2412 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2413 pass_arg3(this, arg_3); 2414 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2415 pass_arg2(this, arg_2); 2416 pass_arg1(this, arg_1); 2417 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 2418 } 2419 2420 void MacroAssembler::super_call_VM(Register oop_result, 2421 Register last_java_sp, 2422 address entry_point, 2423 int number_of_arguments, 2424 bool check_exceptions) { 2425 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg); 2426 MacroAssembler::call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions); 2427 } 2428 2429 void MacroAssembler::super_call_VM(Register oop_result, 2430 Register last_java_sp, 2431 address entry_point, 2432 Register arg_1, 2433 bool check_exceptions) { 2434 pass_arg1(this, arg_1); 2435 super_call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 2436 } 2437 2438 void MacroAssembler::super_call_VM(Register oop_result, 2439 Register last_java_sp, 2440 address entry_point, 2441 Register arg_1, 2442 Register arg_2, 2443 bool check_exceptions) { 2444 2445 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2446 pass_arg2(this, arg_2); 2447 pass_arg1(this, arg_1); 2448 super_call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 2449 } 2450 2451 void MacroAssembler::super_call_VM(Register oop_result, 2452 Register last_java_sp, 2453 address entry_point, 2454 Register arg_1, 2455 Register arg_2, 2456 Register arg_3, 2457 bool check_exceptions) { 2458 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2459 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2460 pass_arg3(this, arg_3); 2461 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2462 pass_arg2(this, arg_2); 2463 pass_arg1(this, arg_1); 2464 super_call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 2465 } 2466 2467 void MacroAssembler::call_VM_base(Register oop_result, 2468 Register java_thread, 2469 Register last_java_sp, 2470 address entry_point, 2471 int number_of_arguments, 2472 bool check_exceptions) { 2473 // determine java_thread register 2474 if (!java_thread->is_valid()) { 2475 #ifdef _LP64 2476 java_thread = r15_thread; 2477 #else 2478 java_thread = rdi; 2479 get_thread(java_thread); 2480 #endif // LP64 2481 } 2482 // determine last_java_sp register 2483 if (!last_java_sp->is_valid()) { 2484 last_java_sp = rsp; 2485 } 2486 // debugging support 2487 assert(number_of_arguments >= 0 , "cannot have negative number of arguments"); 2488 LP64_ONLY(assert(java_thread == r15_thread, "unexpected register")); 2489 #ifdef ASSERT 2490 // TraceBytecodes does not use r12 but saves it over the call, so don't verify 2491 // r12 is the heapbase. 2492 LP64_ONLY(if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");) 2493 #endif // ASSERT 2494 2495 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result"); 2496 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp"); 2497 2498 // push java thread (becomes first argument of C function) 2499 2500 NOT_LP64(push(java_thread); number_of_arguments++); 2501 LP64_ONLY(mov(c_rarg0, r15_thread)); 2502 2503 // set last Java frame before call 2504 assert(last_java_sp != rbp, "can't use ebp/rbp"); 2505 2506 // Only interpreter should have to set fp 2507 set_last_Java_frame(java_thread, last_java_sp, rbp, NULL); 2508 2509 // do the call, remove parameters 2510 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments); 2511 2512 // restore the thread (cannot use the pushed argument since arguments 2513 // may be overwritten by C code generated by an optimizing compiler); 2514 // however can use the register value directly if it is callee saved. 2515 if (LP64_ONLY(true ||) java_thread == rdi || java_thread == rsi) { 2516 // rdi & rsi (also r15) are callee saved -> nothing to do 2517 #ifdef ASSERT 2518 guarantee(java_thread != rax, "change this code"); 2519 push(rax); 2520 { Label L; 2521 get_thread(rax); 2522 cmpptr(java_thread, rax); 2523 jcc(Assembler::equal, L); 2524 STOP("MacroAssembler::call_VM_base: rdi not callee saved?"); 2525 bind(L); 2526 } 2527 pop(rax); 2528 #endif 2529 } else { 2530 get_thread(java_thread); 2531 } 2532 // reset last Java frame 2533 // Only interpreter should have to clear fp 2534 reset_last_Java_frame(java_thread, true); 2535 2536 // C++ interp handles this in the interpreter 2537 check_and_handle_popframe(java_thread); 2538 check_and_handle_earlyret(java_thread); 2539 2540 if (check_exceptions) { 2541 // check for pending exceptions (java_thread is set upon return) 2542 cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t) NULL_WORD); 2543 #ifndef _LP64 2544 jump_cc(Assembler::notEqual, 2545 RuntimeAddress(StubRoutines::forward_exception_entry())); 2546 #else 2547 // This used to conditionally jump to forward_exception however it is 2548 // possible if we relocate that the branch will not reach. So we must jump 2549 // around so we can always reach 2550 2551 Label ok; 2552 jcc(Assembler::equal, ok); 2553 jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2554 bind(ok); 2555 #endif // LP64 2556 } 2557 2558 // get oop result if there is one and reset the value in the thread 2559 if (oop_result->is_valid()) { 2560 get_vm_result(oop_result, java_thread); 2561 } 2562 } 2563 2564 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) { 2565 2566 // Calculate the value for last_Java_sp 2567 // somewhat subtle. call_VM does an intermediate call 2568 // which places a return address on the stack just under the 2569 // stack pointer as the user finsihed with it. This allows 2570 // use to retrieve last_Java_pc from last_Java_sp[-1]. 2571 // On 32bit we then have to push additional args on the stack to accomplish 2572 // the actual requested call. On 64bit call_VM only can use register args 2573 // so the only extra space is the return address that call_VM created. 2574 // This hopefully explains the calculations here. 2575 2576 #ifdef _LP64 2577 // We've pushed one address, correct last_Java_sp 2578 lea(rax, Address(rsp, wordSize)); 2579 #else 2580 lea(rax, Address(rsp, (1 + number_of_arguments) * wordSize)); 2581 #endif // LP64 2582 2583 call_VM_base(oop_result, noreg, rax, entry_point, number_of_arguments, check_exceptions); 2584 2585 } 2586 2587 // Use this method when MacroAssembler version of call_VM_leaf_base() should be called from Interpreter. 2588 void MacroAssembler::call_VM_leaf0(address entry_point) { 2589 MacroAssembler::call_VM_leaf_base(entry_point, 0); 2590 } 2591 2592 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) { 2593 call_VM_leaf_base(entry_point, number_of_arguments); 2594 } 2595 2596 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) { 2597 pass_arg0(this, arg_0); 2598 call_VM_leaf(entry_point, 1); 2599 } 2600 2601 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { 2602 2603 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2604 pass_arg1(this, arg_1); 2605 pass_arg0(this, arg_0); 2606 call_VM_leaf(entry_point, 2); 2607 } 2608 2609 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { 2610 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2611 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2612 pass_arg2(this, arg_2); 2613 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2614 pass_arg1(this, arg_1); 2615 pass_arg0(this, arg_0); 2616 call_VM_leaf(entry_point, 3); 2617 } 2618 2619 void MacroAssembler::super_call_VM_leaf(address entry_point) { 2620 MacroAssembler::call_VM_leaf_base(entry_point, 1); 2621 } 2622 2623 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) { 2624 pass_arg0(this, arg_0); 2625 MacroAssembler::call_VM_leaf_base(entry_point, 1); 2626 } 2627 2628 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { 2629 2630 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2631 pass_arg1(this, arg_1); 2632 pass_arg0(this, arg_0); 2633 MacroAssembler::call_VM_leaf_base(entry_point, 2); 2634 } 2635 2636 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { 2637 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2638 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2639 pass_arg2(this, arg_2); 2640 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2641 pass_arg1(this, arg_1); 2642 pass_arg0(this, arg_0); 2643 MacroAssembler::call_VM_leaf_base(entry_point, 3); 2644 } 2645 2646 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) { 2647 LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg")); 2648 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2649 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2650 pass_arg3(this, arg_3); 2651 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2652 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2653 pass_arg2(this, arg_2); 2654 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2655 pass_arg1(this, arg_1); 2656 pass_arg0(this, arg_0); 2657 MacroAssembler::call_VM_leaf_base(entry_point, 4); 2658 } 2659 2660 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) { 2661 movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset())); 2662 movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD); 2663 verify_oop(oop_result, "broken oop in call_VM_base"); 2664 } 2665 2666 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) { 2667 movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset())); 2668 movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD); 2669 } 2670 2671 void MacroAssembler::check_and_handle_earlyret(Register java_thread) { 2672 } 2673 2674 void MacroAssembler::check_and_handle_popframe(Register java_thread) { 2675 } 2676 2677 void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) { 2678 if (reachable(src1)) { 2679 cmpl(as_Address(src1), imm); 2680 } else { 2681 lea(rscratch1, src1); 2682 cmpl(Address(rscratch1, 0), imm); 2683 } 2684 } 2685 2686 void MacroAssembler::cmp32(Register src1, AddressLiteral src2) { 2687 assert(!src2.is_lval(), "use cmpptr"); 2688 if (reachable(src2)) { 2689 cmpl(src1, as_Address(src2)); 2690 } else { 2691 lea(rscratch1, src2); 2692 cmpl(src1, Address(rscratch1, 0)); 2693 } 2694 } 2695 2696 void MacroAssembler::cmp32(Register src1, int32_t imm) { 2697 Assembler::cmpl(src1, imm); 2698 } 2699 2700 void MacroAssembler::cmp32(Register src1, Address src2) { 2701 Assembler::cmpl(src1, src2); 2702 } 2703 2704 void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) { 2705 ucomisd(opr1, opr2); 2706 2707 Label L; 2708 if (unordered_is_less) { 2709 movl(dst, -1); 2710 jcc(Assembler::parity, L); 2711 jcc(Assembler::below , L); 2712 movl(dst, 0); 2713 jcc(Assembler::equal , L); 2714 increment(dst); 2715 } else { // unordered is greater 2716 movl(dst, 1); 2717 jcc(Assembler::parity, L); 2718 jcc(Assembler::above , L); 2719 movl(dst, 0); 2720 jcc(Assembler::equal , L); 2721 decrementl(dst); 2722 } 2723 bind(L); 2724 } 2725 2726 void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) { 2727 ucomiss(opr1, opr2); 2728 2729 Label L; 2730 if (unordered_is_less) { 2731 movl(dst, -1); 2732 jcc(Assembler::parity, L); 2733 jcc(Assembler::below , L); 2734 movl(dst, 0); 2735 jcc(Assembler::equal , L); 2736 increment(dst); 2737 } else { // unordered is greater 2738 movl(dst, 1); 2739 jcc(Assembler::parity, L); 2740 jcc(Assembler::above , L); 2741 movl(dst, 0); 2742 jcc(Assembler::equal , L); 2743 decrementl(dst); 2744 } 2745 bind(L); 2746 } 2747 2748 2749 void MacroAssembler::cmp8(AddressLiteral src1, int imm) { 2750 if (reachable(src1)) { 2751 cmpb(as_Address(src1), imm); 2752 } else { 2753 lea(rscratch1, src1); 2754 cmpb(Address(rscratch1, 0), imm); 2755 } 2756 } 2757 2758 void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) { 2759 #ifdef _LP64 2760 if (src2.is_lval()) { 2761 movptr(rscratch1, src2); 2762 Assembler::cmpq(src1, rscratch1); 2763 } else if (reachable(src2)) { 2764 cmpq(src1, as_Address(src2)); 2765 } else { 2766 lea(rscratch1, src2); 2767 Assembler::cmpq(src1, Address(rscratch1, 0)); 2768 } 2769 #else 2770 if (src2.is_lval()) { 2771 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec()); 2772 } else { 2773 cmpl(src1, as_Address(src2)); 2774 } 2775 #endif // _LP64 2776 } 2777 2778 void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) { 2779 assert(src2.is_lval(), "not a mem-mem compare"); 2780 #ifdef _LP64 2781 // moves src2's literal address 2782 movptr(rscratch1, src2); 2783 Assembler::cmpq(src1, rscratch1); 2784 #else 2785 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec()); 2786 #endif // _LP64 2787 } 2788 2789 void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) { 2790 if (reachable(adr)) { 2791 if (os::is_MP()) 2792 lock(); 2793 cmpxchgptr(reg, as_Address(adr)); 2794 } else { 2795 lea(rscratch1, adr); 2796 if (os::is_MP()) 2797 lock(); 2798 cmpxchgptr(reg, Address(rscratch1, 0)); 2799 } 2800 } 2801 2802 void MacroAssembler::cmpxchgptr(Register reg, Address adr) { 2803 LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr)); 2804 } 2805 2806 void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) { 2807 if (reachable(src)) { 2808 Assembler::comisd(dst, as_Address(src)); 2809 } else { 2810 lea(rscratch1, src); 2811 Assembler::comisd(dst, Address(rscratch1, 0)); 2812 } 2813 } 2814 2815 void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) { 2816 if (reachable(src)) { 2817 Assembler::comiss(dst, as_Address(src)); 2818 } else { 2819 lea(rscratch1, src); 2820 Assembler::comiss(dst, Address(rscratch1, 0)); 2821 } 2822 } 2823 2824 2825 void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) { 2826 Condition negated_cond = negate_condition(cond); 2827 Label L; 2828 jcc(negated_cond, L); 2829 pushf(); // Preserve flags 2830 atomic_incl(counter_addr); 2831 popf(); 2832 bind(L); 2833 } 2834 2835 int MacroAssembler::corrected_idivl(Register reg) { 2836 // Full implementation of Java idiv and irem; checks for 2837 // special case as described in JVM spec., p.243 & p.271. 2838 // The function returns the (pc) offset of the idivl 2839 // instruction - may be needed for implicit exceptions. 2840 // 2841 // normal case special case 2842 // 2843 // input : rax,: dividend min_int 2844 // reg: divisor (may not be rax,/rdx) -1 2845 // 2846 // output: rax,: quotient (= rax, idiv reg) min_int 2847 // rdx: remainder (= rax, irem reg) 0 2848 assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register"); 2849 const int min_int = 0x80000000; 2850 Label normal_case, special_case; 2851 2852 // check for special case 2853 cmpl(rax, min_int); 2854 jcc(Assembler::notEqual, normal_case); 2855 xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0) 2856 cmpl(reg, -1); 2857 jcc(Assembler::equal, special_case); 2858 2859 // handle normal case 2860 bind(normal_case); 2861 cdql(); 2862 int idivl_offset = offset(); 2863 idivl(reg); 2864 2865 // normal and special case exit 2866 bind(special_case); 2867 2868 return idivl_offset; 2869 } 2870 2871 2872 2873 void MacroAssembler::decrementl(Register reg, int value) { 2874 if (value == min_jint) {subl(reg, value) ; return; } 2875 if (value < 0) { incrementl(reg, -value); return; } 2876 if (value == 0) { ; return; } 2877 if (value == 1 && UseIncDec) { decl(reg) ; return; } 2878 /* else */ { subl(reg, value) ; return; } 2879 } 2880 2881 void MacroAssembler::decrementl(Address dst, int value) { 2882 if (value == min_jint) {subl(dst, value) ; return; } 2883 if (value < 0) { incrementl(dst, -value); return; } 2884 if (value == 0) { ; return; } 2885 if (value == 1 && UseIncDec) { decl(dst) ; return; } 2886 /* else */ { subl(dst, value) ; return; } 2887 } 2888 2889 void MacroAssembler::division_with_shift (Register reg, int shift_value) { 2890 assert (shift_value > 0, "illegal shift value"); 2891 Label _is_positive; 2892 testl (reg, reg); 2893 jcc (Assembler::positive, _is_positive); 2894 int offset = (1 << shift_value) - 1 ; 2895 2896 if (offset == 1) { 2897 incrementl(reg); 2898 } else { 2899 addl(reg, offset); 2900 } 2901 2902 bind (_is_positive); 2903 sarl(reg, shift_value); 2904 } 2905 2906 void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) { 2907 if (reachable(src)) { 2908 Assembler::divsd(dst, as_Address(src)); 2909 } else { 2910 lea(rscratch1, src); 2911 Assembler::divsd(dst, Address(rscratch1, 0)); 2912 } 2913 } 2914 2915 void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) { 2916 if (reachable(src)) { 2917 Assembler::divss(dst, as_Address(src)); 2918 } else { 2919 lea(rscratch1, src); 2920 Assembler::divss(dst, Address(rscratch1, 0)); 2921 } 2922 } 2923 2924 // !defined(COMPILER2) is because of stupid core builds 2925 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2) || INCLUDE_JVMCI 2926 void MacroAssembler::empty_FPU_stack() { 2927 if (VM_Version::supports_mmx()) { 2928 emms(); 2929 } else { 2930 for (int i = 8; i-- > 0; ) ffree(i); 2931 } 2932 } 2933 #endif // !LP64 || C1 || !C2 || INCLUDE_JVMCI 2934 2935 2936 // Defines obj, preserves var_size_in_bytes 2937 void MacroAssembler::eden_allocate(Register obj, 2938 Register var_size_in_bytes, 2939 int con_size_in_bytes, 2940 Register t1, 2941 Label& slow_case) { 2942 assert(obj == rax, "obj must be in rax, for cmpxchg"); 2943 assert_different_registers(obj, var_size_in_bytes, t1); 2944 if (!Universe::heap()->supports_inline_contig_alloc()) { 2945 jmp(slow_case); 2946 } else { 2947 Register end = t1; 2948 Label retry; 2949 bind(retry); 2950 ExternalAddress heap_top((address) Universe::heap()->top_addr()); 2951 movptr(obj, heap_top); 2952 if (var_size_in_bytes == noreg) { 2953 lea(end, Address(obj, con_size_in_bytes)); 2954 } else { 2955 lea(end, Address(obj, var_size_in_bytes, Address::times_1)); 2956 } 2957 // if end < obj then we wrapped around => object too long => slow case 2958 cmpptr(end, obj); 2959 jcc(Assembler::below, slow_case); 2960 cmpptr(end, ExternalAddress((address) Universe::heap()->end_addr())); 2961 jcc(Assembler::above, slow_case); 2962 // Compare obj with the top addr, and if still equal, store the new top addr in 2963 // end at the address of the top addr pointer. Sets ZF if was equal, and clears 2964 // it otherwise. Use lock prefix for atomicity on MPs. 2965 locked_cmpxchgptr(end, heap_top); 2966 jcc(Assembler::notEqual, retry); 2967 } 2968 } 2969 2970 void MacroAssembler::enter() { 2971 push(rbp); 2972 mov(rbp, rsp); 2973 } 2974 2975 // A 5 byte nop that is safe for patching (see patch_verified_entry) 2976 void MacroAssembler::fat_nop() { 2977 if (UseAddressNop) { 2978 addr_nop_5(); 2979 } else { 2980 emit_int8(0x26); // es: 2981 emit_int8(0x2e); // cs: 2982 emit_int8(0x64); // fs: 2983 emit_int8(0x65); // gs: 2984 emit_int8((unsigned char)0x90); 2985 } 2986 } 2987 2988 void MacroAssembler::fcmp(Register tmp) { 2989 fcmp(tmp, 1, true, true); 2990 } 2991 2992 void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) { 2993 assert(!pop_right || pop_left, "usage error"); 2994 if (VM_Version::supports_cmov()) { 2995 assert(tmp == noreg, "unneeded temp"); 2996 if (pop_left) { 2997 fucomip(index); 2998 } else { 2999 fucomi(index); 3000 } 3001 if (pop_right) { 3002 fpop(); 3003 } 3004 } else { 3005 assert(tmp != noreg, "need temp"); 3006 if (pop_left) { 3007 if (pop_right) { 3008 fcompp(); 3009 } else { 3010 fcomp(index); 3011 } 3012 } else { 3013 fcom(index); 3014 } 3015 // convert FPU condition into eflags condition via rax, 3016 save_rax(tmp); 3017 fwait(); fnstsw_ax(); 3018 sahf(); 3019 restore_rax(tmp); 3020 } 3021 // condition codes set as follows: 3022 // 3023 // CF (corresponds to C0) if x < y 3024 // PF (corresponds to C2) if unordered 3025 // ZF (corresponds to C3) if x = y 3026 } 3027 3028 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) { 3029 fcmp2int(dst, unordered_is_less, 1, true, true); 3030 } 3031 3032 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) { 3033 fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right); 3034 Label L; 3035 if (unordered_is_less) { 3036 movl(dst, -1); 3037 jcc(Assembler::parity, L); 3038 jcc(Assembler::below , L); 3039 movl(dst, 0); 3040 jcc(Assembler::equal , L); 3041 increment(dst); 3042 } else { // unordered is greater 3043 movl(dst, 1); 3044 jcc(Assembler::parity, L); 3045 jcc(Assembler::above , L); 3046 movl(dst, 0); 3047 jcc(Assembler::equal , L); 3048 decrementl(dst); 3049 } 3050 bind(L); 3051 } 3052 3053 void MacroAssembler::fld_d(AddressLiteral src) { 3054 fld_d(as_Address(src)); 3055 } 3056 3057 void MacroAssembler::fld_s(AddressLiteral src) { 3058 fld_s(as_Address(src)); 3059 } 3060 3061 void MacroAssembler::fld_x(AddressLiteral src) { 3062 Assembler::fld_x(as_Address(src)); 3063 } 3064 3065 void MacroAssembler::fldcw(AddressLiteral src) { 3066 Assembler::fldcw(as_Address(src)); 3067 } 3068 3069 void MacroAssembler::mulpd(XMMRegister dst, AddressLiteral src) { 3070 if (reachable(src)) { 3071 Assembler::mulpd(dst, as_Address(src)); 3072 } else { 3073 lea(rscratch1, src); 3074 Assembler::mulpd(dst, Address(rscratch1, 0)); 3075 } 3076 } 3077 3078 void MacroAssembler::increase_precision() { 3079 subptr(rsp, BytesPerWord); 3080 fnstcw(Address(rsp, 0)); 3081 movl(rax, Address(rsp, 0)); 3082 orl(rax, 0x300); 3083 push(rax); 3084 fldcw(Address(rsp, 0)); 3085 pop(rax); 3086 } 3087 3088 void MacroAssembler::restore_precision() { 3089 fldcw(Address(rsp, 0)); 3090 addptr(rsp, BytesPerWord); 3091 } 3092 3093 void MacroAssembler::fpop() { 3094 ffree(); 3095 fincstp(); 3096 } 3097 3098 void MacroAssembler::load_float(Address src) { 3099 if (UseSSE >= 1) { 3100 movflt(xmm0, src); 3101 } else { 3102 LP64_ONLY(ShouldNotReachHere()); 3103 NOT_LP64(fld_s(src)); 3104 } 3105 } 3106 3107 void MacroAssembler::store_float(Address dst) { 3108 if (UseSSE >= 1) { 3109 movflt(dst, xmm0); 3110 } else { 3111 LP64_ONLY(ShouldNotReachHere()); 3112 NOT_LP64(fstp_s(dst)); 3113 } 3114 } 3115 3116 void MacroAssembler::load_double(Address src) { 3117 if (UseSSE >= 2) { 3118 movdbl(xmm0, src); 3119 } else { 3120 LP64_ONLY(ShouldNotReachHere()); 3121 NOT_LP64(fld_d(src)); 3122 } 3123 } 3124 3125 void MacroAssembler::store_double(Address dst) { 3126 if (UseSSE >= 2) { 3127 movdbl(dst, xmm0); 3128 } else { 3129 LP64_ONLY(ShouldNotReachHere()); 3130 NOT_LP64(fstp_d(dst)); 3131 } 3132 } 3133 3134 void MacroAssembler::fremr(Register tmp) { 3135 save_rax(tmp); 3136 { Label L; 3137 bind(L); 3138 fprem(); 3139 fwait(); fnstsw_ax(); 3140 #ifdef _LP64 3141 testl(rax, 0x400); 3142 jcc(Assembler::notEqual, L); 3143 #else 3144 sahf(); 3145 jcc(Assembler::parity, L); 3146 #endif // _LP64 3147 } 3148 restore_rax(tmp); 3149 // Result is in ST0. 3150 // Note: fxch & fpop to get rid of ST1 3151 // (otherwise FPU stack could overflow eventually) 3152 fxch(1); 3153 fpop(); 3154 } 3155 3156 // dst = c = a * b + c 3157 void MacroAssembler::fmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) { 3158 Assembler::vfmadd231sd(c, a, b); 3159 if (dst != c) { 3160 movdbl(dst, c); 3161 } 3162 } 3163 3164 // dst = c = a * b + c 3165 void MacroAssembler::fmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) { 3166 Assembler::vfmadd231ss(c, a, b); 3167 if (dst != c) { 3168 movflt(dst, c); 3169 } 3170 } 3171 3172 3173 3174 3175 void MacroAssembler::incrementl(AddressLiteral dst) { 3176 if (reachable(dst)) { 3177 incrementl(as_Address(dst)); 3178 } else { 3179 lea(rscratch1, dst); 3180 incrementl(Address(rscratch1, 0)); 3181 } 3182 } 3183 3184 void MacroAssembler::incrementl(ArrayAddress dst) { 3185 incrementl(as_Address(dst)); 3186 } 3187 3188 void MacroAssembler::incrementl(Register reg, int value) { 3189 if (value == min_jint) {addl(reg, value) ; return; } 3190 if (value < 0) { decrementl(reg, -value); return; } 3191 if (value == 0) { ; return; } 3192 if (value == 1 && UseIncDec) { incl(reg) ; return; } 3193 /* else */ { addl(reg, value) ; return; } 3194 } 3195 3196 void MacroAssembler::incrementl(Address dst, int value) { 3197 if (value == min_jint) {addl(dst, value) ; return; } 3198 if (value < 0) { decrementl(dst, -value); return; } 3199 if (value == 0) { ; return; } 3200 if (value == 1 && UseIncDec) { incl(dst) ; return; } 3201 /* else */ { addl(dst, value) ; return; } 3202 } 3203 3204 void MacroAssembler::jump(AddressLiteral dst) { 3205 if (reachable(dst)) { 3206 jmp_literal(dst.target(), dst.rspec()); 3207 } else { 3208 lea(rscratch1, dst); 3209 jmp(rscratch1); 3210 } 3211 } 3212 3213 void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) { 3214 if (reachable(dst)) { 3215 InstructionMark im(this); 3216 relocate(dst.reloc()); 3217 const int short_size = 2; 3218 const int long_size = 6; 3219 int offs = (intptr_t)dst.target() - ((intptr_t)pc()); 3220 if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) { 3221 // 0111 tttn #8-bit disp 3222 emit_int8(0x70 | cc); 3223 emit_int8((offs - short_size) & 0xFF); 3224 } else { 3225 // 0000 1111 1000 tttn #32-bit disp 3226 emit_int8(0x0F); 3227 emit_int8((unsigned char)(0x80 | cc)); 3228 emit_int32(offs - long_size); 3229 } 3230 } else { 3231 #ifdef ASSERT 3232 warning("reversing conditional branch"); 3233 #endif /* ASSERT */ 3234 Label skip; 3235 jccb(reverse[cc], skip); 3236 lea(rscratch1, dst); 3237 Assembler::jmp(rscratch1); 3238 bind(skip); 3239 } 3240 } 3241 3242 void MacroAssembler::ldmxcsr(AddressLiteral src) { 3243 if (reachable(src)) { 3244 Assembler::ldmxcsr(as_Address(src)); 3245 } else { 3246 lea(rscratch1, src); 3247 Assembler::ldmxcsr(Address(rscratch1, 0)); 3248 } 3249 } 3250 3251 int MacroAssembler::load_signed_byte(Register dst, Address src) { 3252 int off; 3253 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3254 off = offset(); 3255 movsbl(dst, src); // movsxb 3256 } else { 3257 off = load_unsigned_byte(dst, src); 3258 shll(dst, 24); 3259 sarl(dst, 24); 3260 } 3261 return off; 3262 } 3263 3264 // Note: load_signed_short used to be called load_signed_word. 3265 // Although the 'w' in x86 opcodes refers to the term "word" in the assembler 3266 // manual, which means 16 bits, that usage is found nowhere in HotSpot code. 3267 // The term "word" in HotSpot means a 32- or 64-bit machine word. 3268 int MacroAssembler::load_signed_short(Register dst, Address src) { 3269 int off; 3270 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3271 // This is dubious to me since it seems safe to do a signed 16 => 64 bit 3272 // version but this is what 64bit has always done. This seems to imply 3273 // that users are only using 32bits worth. 3274 off = offset(); 3275 movswl(dst, src); // movsxw 3276 } else { 3277 off = load_unsigned_short(dst, src); 3278 shll(dst, 16); 3279 sarl(dst, 16); 3280 } 3281 return off; 3282 } 3283 3284 int MacroAssembler::load_unsigned_byte(Register dst, Address src) { 3285 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16, 3286 // and "3.9 Partial Register Penalties", p. 22). 3287 int off; 3288 if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) { 3289 off = offset(); 3290 movzbl(dst, src); // movzxb 3291 } else { 3292 xorl(dst, dst); 3293 off = offset(); 3294 movb(dst, src); 3295 } 3296 return off; 3297 } 3298 3299 // Note: load_unsigned_short used to be called load_unsigned_word. 3300 int MacroAssembler::load_unsigned_short(Register dst, Address src) { 3301 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16, 3302 // and "3.9 Partial Register Penalties", p. 22). 3303 int off; 3304 if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) { 3305 off = offset(); 3306 movzwl(dst, src); // movzxw 3307 } else { 3308 xorl(dst, dst); 3309 off = offset(); 3310 movw(dst, src); 3311 } 3312 return off; 3313 } 3314 3315 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) { 3316 switch (size_in_bytes) { 3317 #ifndef _LP64 3318 case 8: 3319 assert(dst2 != noreg, "second dest register required"); 3320 movl(dst, src); 3321 movl(dst2, src.plus_disp(BytesPerInt)); 3322 break; 3323 #else 3324 case 8: movq(dst, src); break; 3325 #endif 3326 case 4: movl(dst, src); break; 3327 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break; 3328 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break; 3329 default: ShouldNotReachHere(); 3330 } 3331 } 3332 3333 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) { 3334 switch (size_in_bytes) { 3335 #ifndef _LP64 3336 case 8: 3337 assert(src2 != noreg, "second source register required"); 3338 movl(dst, src); 3339 movl(dst.plus_disp(BytesPerInt), src2); 3340 break; 3341 #else 3342 case 8: movq(dst, src); break; 3343 #endif 3344 case 4: movl(dst, src); break; 3345 case 2: movw(dst, src); break; 3346 case 1: movb(dst, src); break; 3347 default: ShouldNotReachHere(); 3348 } 3349 } 3350 3351 void MacroAssembler::mov32(AddressLiteral dst, Register src) { 3352 if (reachable(dst)) { 3353 movl(as_Address(dst), src); 3354 } else { 3355 lea(rscratch1, dst); 3356 movl(Address(rscratch1, 0), src); 3357 } 3358 } 3359 3360 void MacroAssembler::mov32(Register dst, AddressLiteral src) { 3361 if (reachable(src)) { 3362 movl(dst, as_Address(src)); 3363 } else { 3364 lea(rscratch1, src); 3365 movl(dst, Address(rscratch1, 0)); 3366 } 3367 } 3368 3369 // C++ bool manipulation 3370 3371 void MacroAssembler::movbool(Register dst, Address src) { 3372 if(sizeof(bool) == 1) 3373 movb(dst, src); 3374 else if(sizeof(bool) == 2) 3375 movw(dst, src); 3376 else if(sizeof(bool) == 4) 3377 movl(dst, src); 3378 else 3379 // unsupported 3380 ShouldNotReachHere(); 3381 } 3382 3383 void MacroAssembler::movbool(Address dst, bool boolconst) { 3384 if(sizeof(bool) == 1) 3385 movb(dst, (int) boolconst); 3386 else if(sizeof(bool) == 2) 3387 movw(dst, (int) boolconst); 3388 else if(sizeof(bool) == 4) 3389 movl(dst, (int) boolconst); 3390 else 3391 // unsupported 3392 ShouldNotReachHere(); 3393 } 3394 3395 void MacroAssembler::movbool(Address dst, Register src) { 3396 if(sizeof(bool) == 1) 3397 movb(dst, src); 3398 else if(sizeof(bool) == 2) 3399 movw(dst, src); 3400 else if(sizeof(bool) == 4) 3401 movl(dst, src); 3402 else 3403 // unsupported 3404 ShouldNotReachHere(); 3405 } 3406 3407 void MacroAssembler::movbyte(ArrayAddress dst, int src) { 3408 movb(as_Address(dst), src); 3409 } 3410 3411 void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) { 3412 if (reachable(src)) { 3413 movdl(dst, as_Address(src)); 3414 } else { 3415 lea(rscratch1, src); 3416 movdl(dst, Address(rscratch1, 0)); 3417 } 3418 } 3419 3420 void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) { 3421 if (reachable(src)) { 3422 movq(dst, as_Address(src)); 3423 } else { 3424 lea(rscratch1, src); 3425 movq(dst, Address(rscratch1, 0)); 3426 } 3427 } 3428 3429 void MacroAssembler::setvectmask(Register dst, Register src) { 3430 Assembler::movl(dst, 1); 3431 Assembler::shlxl(dst, dst, src); 3432 Assembler::decl(dst); 3433 Assembler::kmovdl(k1, dst); 3434 Assembler::movl(dst, src); 3435 } 3436 3437 void MacroAssembler::restorevectmask() { 3438 Assembler::knotwl(k1, k0); 3439 } 3440 3441 void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) { 3442 if (reachable(src)) { 3443 if (UseXmmLoadAndClearUpper) { 3444 movsd (dst, as_Address(src)); 3445 } else { 3446 movlpd(dst, as_Address(src)); 3447 } 3448 } else { 3449 lea(rscratch1, src); 3450 if (UseXmmLoadAndClearUpper) { 3451 movsd (dst, Address(rscratch1, 0)); 3452 } else { 3453 movlpd(dst, Address(rscratch1, 0)); 3454 } 3455 } 3456 } 3457 3458 void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) { 3459 if (reachable(src)) { 3460 movss(dst, as_Address(src)); 3461 } else { 3462 lea(rscratch1, src); 3463 movss(dst, Address(rscratch1, 0)); 3464 } 3465 } 3466 3467 void MacroAssembler::movptr(Register dst, Register src) { 3468 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3469 } 3470 3471 void MacroAssembler::movptr(Register dst, Address src) { 3472 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3473 } 3474 3475 // src should NEVER be a real pointer. Use AddressLiteral for true pointers 3476 void MacroAssembler::movptr(Register dst, intptr_t src) { 3477 LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src)); 3478 } 3479 3480 void MacroAssembler::movptr(Address dst, Register src) { 3481 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3482 } 3483 3484 void MacroAssembler::movdqu(Address dst, XMMRegister src) { 3485 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (src->encoding() > 15)) { 3486 Assembler::vextractf32x4(dst, src, 0); 3487 } else { 3488 Assembler::movdqu(dst, src); 3489 } 3490 } 3491 3492 void MacroAssembler::movdqu(XMMRegister dst, Address src) { 3493 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (dst->encoding() > 15)) { 3494 Assembler::vinsertf32x4(dst, dst, src, 0); 3495 } else { 3496 Assembler::movdqu(dst, src); 3497 } 3498 } 3499 3500 void MacroAssembler::movdqu(XMMRegister dst, XMMRegister src) { 3501 if (UseAVX > 2 && !VM_Version::supports_avx512vl()) { 3502 Assembler::evmovdqul(dst, src, Assembler::AVX_512bit); 3503 } else { 3504 Assembler::movdqu(dst, src); 3505 } 3506 } 3507 3508 void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src, Register scratchReg) { 3509 if (reachable(src)) { 3510 movdqu(dst, as_Address(src)); 3511 } else { 3512 lea(scratchReg, src); 3513 movdqu(dst, Address(scratchReg, 0)); 3514 } 3515 } 3516 3517 void MacroAssembler::vmovdqu(Address dst, XMMRegister src) { 3518 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (src->encoding() > 15)) { 3519 vextractf64x4_low(dst, src); 3520 } else { 3521 Assembler::vmovdqu(dst, src); 3522 } 3523 } 3524 3525 void MacroAssembler::vmovdqu(XMMRegister dst, Address src) { 3526 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (dst->encoding() > 15)) { 3527 vinsertf64x4_low(dst, src); 3528 } else { 3529 Assembler::vmovdqu(dst, src); 3530 } 3531 } 3532 3533 void MacroAssembler::vmovdqu(XMMRegister dst, XMMRegister src) { 3534 if (UseAVX > 2 && !VM_Version::supports_avx512vl()) { 3535 Assembler::evmovdqul(dst, src, Assembler::AVX_512bit); 3536 } 3537 else { 3538 Assembler::vmovdqu(dst, src); 3539 } 3540 } 3541 3542 void MacroAssembler::vmovdqu(XMMRegister dst, AddressLiteral src) { 3543 if (reachable(src)) { 3544 vmovdqu(dst, as_Address(src)); 3545 } 3546 else { 3547 lea(rscratch1, src); 3548 vmovdqu(dst, Address(rscratch1, 0)); 3549 } 3550 } 3551 3552 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) { 3553 if (reachable(src)) { 3554 Assembler::movdqa(dst, as_Address(src)); 3555 } else { 3556 lea(rscratch1, src); 3557 Assembler::movdqa(dst, Address(rscratch1, 0)); 3558 } 3559 } 3560 3561 void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) { 3562 if (reachable(src)) { 3563 Assembler::movsd(dst, as_Address(src)); 3564 } else { 3565 lea(rscratch1, src); 3566 Assembler::movsd(dst, Address(rscratch1, 0)); 3567 } 3568 } 3569 3570 void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) { 3571 if (reachable(src)) { 3572 Assembler::movss(dst, as_Address(src)); 3573 } else { 3574 lea(rscratch1, src); 3575 Assembler::movss(dst, Address(rscratch1, 0)); 3576 } 3577 } 3578 3579 void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) { 3580 if (reachable(src)) { 3581 Assembler::mulsd(dst, as_Address(src)); 3582 } else { 3583 lea(rscratch1, src); 3584 Assembler::mulsd(dst, Address(rscratch1, 0)); 3585 } 3586 } 3587 3588 void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) { 3589 if (reachable(src)) { 3590 Assembler::mulss(dst, as_Address(src)); 3591 } else { 3592 lea(rscratch1, src); 3593 Assembler::mulss(dst, Address(rscratch1, 0)); 3594 } 3595 } 3596 3597 void MacroAssembler::null_check(Register reg, int offset) { 3598 if (needs_explicit_null_check(offset)) { 3599 // provoke OS NULL exception if reg = NULL by 3600 // accessing M[reg] w/o changing any (non-CC) registers 3601 // NOTE: cmpl is plenty here to provoke a segv 3602 cmpptr(rax, Address(reg, 0)); 3603 // Note: should probably use testl(rax, Address(reg, 0)); 3604 // may be shorter code (however, this version of 3605 // testl needs to be implemented first) 3606 } else { 3607 // nothing to do, (later) access of M[reg + offset] 3608 // will provoke OS NULL exception if reg = NULL 3609 } 3610 } 3611 3612 void MacroAssembler::os_breakpoint() { 3613 // instead of directly emitting a breakpoint, call os:breakpoint for better debugability 3614 // (e.g., MSVC can't call ps() otherwise) 3615 call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint))); 3616 } 3617 3618 #ifdef _LP64 3619 #define XSTATE_BV 0x200 3620 #endif 3621 3622 void MacroAssembler::pop_CPU_state() { 3623 pop_FPU_state(); 3624 pop_IU_state(); 3625 } 3626 3627 void MacroAssembler::pop_FPU_state() { 3628 #ifndef _LP64 3629 frstor(Address(rsp, 0)); 3630 #else 3631 fxrstor(Address(rsp, 0)); 3632 #endif 3633 addptr(rsp, FPUStateSizeInWords * wordSize); 3634 } 3635 3636 void MacroAssembler::pop_IU_state() { 3637 popa(); 3638 LP64_ONLY(addq(rsp, 8)); 3639 popf(); 3640 } 3641 3642 // Save Integer and Float state 3643 // Warning: Stack must be 16 byte aligned (64bit) 3644 void MacroAssembler::push_CPU_state() { 3645 push_IU_state(); 3646 push_FPU_state(); 3647 } 3648 3649 void MacroAssembler::push_FPU_state() { 3650 subptr(rsp, FPUStateSizeInWords * wordSize); 3651 #ifndef _LP64 3652 fnsave(Address(rsp, 0)); 3653 fwait(); 3654 #else 3655 fxsave(Address(rsp, 0)); 3656 #endif // LP64 3657 } 3658 3659 void MacroAssembler::push_IU_state() { 3660 // Push flags first because pusha kills them 3661 pushf(); 3662 // Make sure rsp stays 16-byte aligned 3663 LP64_ONLY(subq(rsp, 8)); 3664 pusha(); 3665 } 3666 3667 void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp) { // determine java_thread register 3668 if (!java_thread->is_valid()) { 3669 java_thread = rdi; 3670 get_thread(java_thread); 3671 } 3672 // we must set sp to zero to clear frame 3673 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD); 3674 if (clear_fp) { 3675 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD); 3676 } 3677 3678 // Always clear the pc because it could have been set by make_walkable() 3679 movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD); 3680 3681 vzeroupper(); 3682 } 3683 3684 void MacroAssembler::restore_rax(Register tmp) { 3685 if (tmp == noreg) pop(rax); 3686 else if (tmp != rax) mov(rax, tmp); 3687 } 3688 3689 void MacroAssembler::round_to(Register reg, int modulus) { 3690 addptr(reg, modulus - 1); 3691 andptr(reg, -modulus); 3692 } 3693 3694 void MacroAssembler::save_rax(Register tmp) { 3695 if (tmp == noreg) push(rax); 3696 else if (tmp != rax) mov(tmp, rax); 3697 } 3698 3699 // Write serialization page so VM thread can do a pseudo remote membar. 3700 // We use the current thread pointer to calculate a thread specific 3701 // offset to write to within the page. This minimizes bus traffic 3702 // due to cache line collision. 3703 void MacroAssembler::serialize_memory(Register thread, Register tmp) { 3704 movl(tmp, thread); 3705 shrl(tmp, os::get_serialize_page_shift_count()); 3706 andl(tmp, (os::vm_page_size() - sizeof(int))); 3707 3708 Address index(noreg, tmp, Address::times_1); 3709 ExternalAddress page(os::get_memory_serialize_page()); 3710 3711 // Size of store must match masking code above 3712 movl(as_Address(ArrayAddress(page, index)), tmp); 3713 } 3714 3715 // Calls to C land 3716 // 3717 // When entering C land, the rbp, & rsp of the last Java frame have to be recorded 3718 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp 3719 // has to be reset to 0. This is required to allow proper stack traversal. 3720 void MacroAssembler::set_last_Java_frame(Register java_thread, 3721 Register last_java_sp, 3722 Register last_java_fp, 3723 address last_java_pc) { 3724 vzeroupper(); 3725 // determine java_thread register 3726 if (!java_thread->is_valid()) { 3727 java_thread = rdi; 3728 get_thread(java_thread); 3729 } 3730 // determine last_java_sp register 3731 if (!last_java_sp->is_valid()) { 3732 last_java_sp = rsp; 3733 } 3734 3735 // last_java_fp is optional 3736 3737 if (last_java_fp->is_valid()) { 3738 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp); 3739 } 3740 3741 // last_java_pc is optional 3742 3743 if (last_java_pc != NULL) { 3744 lea(Address(java_thread, 3745 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()), 3746 InternalAddress(last_java_pc)); 3747 3748 } 3749 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp); 3750 } 3751 3752 void MacroAssembler::shlptr(Register dst, int imm8) { 3753 LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8)); 3754 } 3755 3756 void MacroAssembler::shrptr(Register dst, int imm8) { 3757 LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8)); 3758 } 3759 3760 void MacroAssembler::sign_extend_byte(Register reg) { 3761 if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) { 3762 movsbl(reg, reg); // movsxb 3763 } else { 3764 shll(reg, 24); 3765 sarl(reg, 24); 3766 } 3767 } 3768 3769 void MacroAssembler::sign_extend_short(Register reg) { 3770 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3771 movswl(reg, reg); // movsxw 3772 } else { 3773 shll(reg, 16); 3774 sarl(reg, 16); 3775 } 3776 } 3777 3778 void MacroAssembler::testl(Register dst, AddressLiteral src) { 3779 assert(reachable(src), "Address should be reachable"); 3780 testl(dst, as_Address(src)); 3781 } 3782 3783 void MacroAssembler::pcmpeqb(XMMRegister dst, XMMRegister src) { 3784 int dst_enc = dst->encoding(); 3785 int src_enc = src->encoding(); 3786 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 3787 Assembler::pcmpeqb(dst, src); 3788 } else if ((dst_enc < 16) && (src_enc < 16)) { 3789 Assembler::pcmpeqb(dst, src); 3790 } else if (src_enc < 16) { 3791 subptr(rsp, 64); 3792 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3793 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3794 Assembler::pcmpeqb(xmm0, src); 3795 movdqu(dst, xmm0); 3796 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3797 addptr(rsp, 64); 3798 } else if (dst_enc < 16) { 3799 subptr(rsp, 64); 3800 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3801 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3802 Assembler::pcmpeqb(dst, xmm0); 3803 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3804 addptr(rsp, 64); 3805 } else { 3806 subptr(rsp, 64); 3807 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3808 subptr(rsp, 64); 3809 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 3810 movdqu(xmm0, src); 3811 movdqu(xmm1, dst); 3812 Assembler::pcmpeqb(xmm1, xmm0); 3813 movdqu(dst, xmm1); 3814 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 3815 addptr(rsp, 64); 3816 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3817 addptr(rsp, 64); 3818 } 3819 } 3820 3821 void MacroAssembler::pcmpeqw(XMMRegister dst, XMMRegister src) { 3822 int dst_enc = dst->encoding(); 3823 int src_enc = src->encoding(); 3824 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 3825 Assembler::pcmpeqw(dst, src); 3826 } else if ((dst_enc < 16) && (src_enc < 16)) { 3827 Assembler::pcmpeqw(dst, src); 3828 } else if (src_enc < 16) { 3829 subptr(rsp, 64); 3830 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3831 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3832 Assembler::pcmpeqw(xmm0, src); 3833 movdqu(dst, xmm0); 3834 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3835 addptr(rsp, 64); 3836 } else if (dst_enc < 16) { 3837 subptr(rsp, 64); 3838 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3839 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3840 Assembler::pcmpeqw(dst, xmm0); 3841 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3842 addptr(rsp, 64); 3843 } else { 3844 subptr(rsp, 64); 3845 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3846 subptr(rsp, 64); 3847 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 3848 movdqu(xmm0, src); 3849 movdqu(xmm1, dst); 3850 Assembler::pcmpeqw(xmm1, xmm0); 3851 movdqu(dst, xmm1); 3852 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 3853 addptr(rsp, 64); 3854 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3855 addptr(rsp, 64); 3856 } 3857 } 3858 3859 void MacroAssembler::pcmpestri(XMMRegister dst, Address src, int imm8) { 3860 int dst_enc = dst->encoding(); 3861 if (dst_enc < 16) { 3862 Assembler::pcmpestri(dst, src, imm8); 3863 } else { 3864 subptr(rsp, 64); 3865 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3866 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3867 Assembler::pcmpestri(xmm0, src, imm8); 3868 movdqu(dst, xmm0); 3869 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3870 addptr(rsp, 64); 3871 } 3872 } 3873 3874 void MacroAssembler::pcmpestri(XMMRegister dst, XMMRegister src, int imm8) { 3875 int dst_enc = dst->encoding(); 3876 int src_enc = src->encoding(); 3877 if ((dst_enc < 16) && (src_enc < 16)) { 3878 Assembler::pcmpestri(dst, src, imm8); 3879 } else if (src_enc < 16) { 3880 subptr(rsp, 64); 3881 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3882 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3883 Assembler::pcmpestri(xmm0, src, imm8); 3884 movdqu(dst, xmm0); 3885 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3886 addptr(rsp, 64); 3887 } else if (dst_enc < 16) { 3888 subptr(rsp, 64); 3889 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3890 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3891 Assembler::pcmpestri(dst, xmm0, imm8); 3892 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3893 addptr(rsp, 64); 3894 } else { 3895 subptr(rsp, 64); 3896 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3897 subptr(rsp, 64); 3898 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 3899 movdqu(xmm0, src); 3900 movdqu(xmm1, dst); 3901 Assembler::pcmpestri(xmm1, xmm0, imm8); 3902 movdqu(dst, xmm1); 3903 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 3904 addptr(rsp, 64); 3905 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3906 addptr(rsp, 64); 3907 } 3908 } 3909 3910 void MacroAssembler::pmovzxbw(XMMRegister dst, XMMRegister src) { 3911 int dst_enc = dst->encoding(); 3912 int src_enc = src->encoding(); 3913 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 3914 Assembler::pmovzxbw(dst, src); 3915 } else if ((dst_enc < 16) && (src_enc < 16)) { 3916 Assembler::pmovzxbw(dst, src); 3917 } else if (src_enc < 16) { 3918 subptr(rsp, 64); 3919 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3920 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3921 Assembler::pmovzxbw(xmm0, src); 3922 movdqu(dst, xmm0); 3923 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3924 addptr(rsp, 64); 3925 } else if (dst_enc < 16) { 3926 subptr(rsp, 64); 3927 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3928 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3929 Assembler::pmovzxbw(dst, xmm0); 3930 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3931 addptr(rsp, 64); 3932 } else { 3933 subptr(rsp, 64); 3934 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3935 subptr(rsp, 64); 3936 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 3937 movdqu(xmm0, src); 3938 movdqu(xmm1, dst); 3939 Assembler::pmovzxbw(xmm1, xmm0); 3940 movdqu(dst, xmm1); 3941 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 3942 addptr(rsp, 64); 3943 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3944 addptr(rsp, 64); 3945 } 3946 } 3947 3948 void MacroAssembler::pmovzxbw(XMMRegister dst, Address src) { 3949 int dst_enc = dst->encoding(); 3950 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 3951 Assembler::pmovzxbw(dst, src); 3952 } else if (dst_enc < 16) { 3953 Assembler::pmovzxbw(dst, src); 3954 } else { 3955 subptr(rsp, 64); 3956 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3957 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3958 Assembler::pmovzxbw(xmm0, src); 3959 movdqu(dst, xmm0); 3960 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3961 addptr(rsp, 64); 3962 } 3963 } 3964 3965 void MacroAssembler::pmovmskb(Register dst, XMMRegister src) { 3966 int src_enc = src->encoding(); 3967 if (src_enc < 16) { 3968 Assembler::pmovmskb(dst, src); 3969 } else { 3970 subptr(rsp, 64); 3971 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3972 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3973 Assembler::pmovmskb(dst, xmm0); 3974 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3975 addptr(rsp, 64); 3976 } 3977 } 3978 3979 void MacroAssembler::ptest(XMMRegister dst, XMMRegister src) { 3980 int dst_enc = dst->encoding(); 3981 int src_enc = src->encoding(); 3982 if ((dst_enc < 16) && (src_enc < 16)) { 3983 Assembler::ptest(dst, src); 3984 } else if (src_enc < 16) { 3985 subptr(rsp, 64); 3986 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3987 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3988 Assembler::ptest(xmm0, src); 3989 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3990 addptr(rsp, 64); 3991 } else if (dst_enc < 16) { 3992 subptr(rsp, 64); 3993 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3994 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3995 Assembler::ptest(dst, xmm0); 3996 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3997 addptr(rsp, 64); 3998 } else { 3999 subptr(rsp, 64); 4000 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4001 subptr(rsp, 64); 4002 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4003 movdqu(xmm0, src); 4004 movdqu(xmm1, dst); 4005 Assembler::ptest(xmm1, xmm0); 4006 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4007 addptr(rsp, 64); 4008 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4009 addptr(rsp, 64); 4010 } 4011 } 4012 4013 void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) { 4014 if (reachable(src)) { 4015 Assembler::sqrtsd(dst, as_Address(src)); 4016 } else { 4017 lea(rscratch1, src); 4018 Assembler::sqrtsd(dst, Address(rscratch1, 0)); 4019 } 4020 } 4021 4022 void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) { 4023 if (reachable(src)) { 4024 Assembler::sqrtss(dst, as_Address(src)); 4025 } else { 4026 lea(rscratch1, src); 4027 Assembler::sqrtss(dst, Address(rscratch1, 0)); 4028 } 4029 } 4030 4031 void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) { 4032 if (reachable(src)) { 4033 Assembler::subsd(dst, as_Address(src)); 4034 } else { 4035 lea(rscratch1, src); 4036 Assembler::subsd(dst, Address(rscratch1, 0)); 4037 } 4038 } 4039 4040 void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) { 4041 if (reachable(src)) { 4042 Assembler::subss(dst, as_Address(src)); 4043 } else { 4044 lea(rscratch1, src); 4045 Assembler::subss(dst, Address(rscratch1, 0)); 4046 } 4047 } 4048 4049 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) { 4050 if (reachable(src)) { 4051 Assembler::ucomisd(dst, as_Address(src)); 4052 } else { 4053 lea(rscratch1, src); 4054 Assembler::ucomisd(dst, Address(rscratch1, 0)); 4055 } 4056 } 4057 4058 void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) { 4059 if (reachable(src)) { 4060 Assembler::ucomiss(dst, as_Address(src)); 4061 } else { 4062 lea(rscratch1, src); 4063 Assembler::ucomiss(dst, Address(rscratch1, 0)); 4064 } 4065 } 4066 4067 void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src) { 4068 // Used in sign-bit flipping with aligned address. 4069 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 4070 if (reachable(src)) { 4071 Assembler::xorpd(dst, as_Address(src)); 4072 } else { 4073 lea(rscratch1, src); 4074 Assembler::xorpd(dst, Address(rscratch1, 0)); 4075 } 4076 } 4077 4078 void MacroAssembler::xorpd(XMMRegister dst, XMMRegister src) { 4079 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) { 4080 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit); 4081 } 4082 else { 4083 Assembler::xorpd(dst, src); 4084 } 4085 } 4086 4087 void MacroAssembler::xorps(XMMRegister dst, XMMRegister src) { 4088 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) { 4089 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit); 4090 } else { 4091 Assembler::xorps(dst, src); 4092 } 4093 } 4094 4095 void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src) { 4096 // Used in sign-bit flipping with aligned address. 4097 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 4098 if (reachable(src)) { 4099 Assembler::xorps(dst, as_Address(src)); 4100 } else { 4101 lea(rscratch1, src); 4102 Assembler::xorps(dst, Address(rscratch1, 0)); 4103 } 4104 } 4105 4106 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) { 4107 // Used in sign-bit flipping with aligned address. 4108 bool aligned_adr = (((intptr_t)src.target() & 15) == 0); 4109 assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes"); 4110 if (reachable(src)) { 4111 Assembler::pshufb(dst, as_Address(src)); 4112 } else { 4113 lea(rscratch1, src); 4114 Assembler::pshufb(dst, Address(rscratch1, 0)); 4115 } 4116 } 4117 4118 // AVX 3-operands instructions 4119 4120 void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4121 if (reachable(src)) { 4122 vaddsd(dst, nds, as_Address(src)); 4123 } else { 4124 lea(rscratch1, src); 4125 vaddsd(dst, nds, Address(rscratch1, 0)); 4126 } 4127 } 4128 4129 void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4130 if (reachable(src)) { 4131 vaddss(dst, nds, as_Address(src)); 4132 } else { 4133 lea(rscratch1, src); 4134 vaddss(dst, nds, Address(rscratch1, 0)); 4135 } 4136 } 4137 4138 void MacroAssembler::vabsss(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) { 4139 int dst_enc = dst->encoding(); 4140 int nds_enc = nds->encoding(); 4141 int src_enc = src->encoding(); 4142 if ((dst_enc < 16) && (nds_enc < 16)) { 4143 vandps(dst, nds, negate_field, vector_len); 4144 } else if ((src_enc < 16) && (dst_enc < 16)) { 4145 movss(src, nds); 4146 vandps(dst, src, negate_field, vector_len); 4147 } else if (src_enc < 16) { 4148 movss(src, nds); 4149 vandps(src, src, negate_field, vector_len); 4150 movss(dst, src); 4151 } else if (dst_enc < 16) { 4152 movdqu(src, xmm0); 4153 movss(xmm0, nds); 4154 vandps(dst, xmm0, negate_field, vector_len); 4155 movdqu(xmm0, src); 4156 } else if (nds_enc < 16) { 4157 movdqu(src, xmm0); 4158 vandps(xmm0, nds, negate_field, vector_len); 4159 movss(dst, xmm0); 4160 movdqu(xmm0, src); 4161 } else { 4162 movdqu(src, xmm0); 4163 movss(xmm0, nds); 4164 vandps(xmm0, xmm0, negate_field, vector_len); 4165 movss(dst, xmm0); 4166 movdqu(xmm0, src); 4167 } 4168 } 4169 4170 void MacroAssembler::vabssd(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) { 4171 int dst_enc = dst->encoding(); 4172 int nds_enc = nds->encoding(); 4173 int src_enc = src->encoding(); 4174 if ((dst_enc < 16) && (nds_enc < 16)) { 4175 vandpd(dst, nds, negate_field, vector_len); 4176 } else if ((src_enc < 16) && (dst_enc < 16)) { 4177 movsd(src, nds); 4178 vandpd(dst, src, negate_field, vector_len); 4179 } else if (src_enc < 16) { 4180 movsd(src, nds); 4181 vandpd(src, src, negate_field, vector_len); 4182 movsd(dst, src); 4183 } else if (dst_enc < 16) { 4184 movdqu(src, xmm0); 4185 movsd(xmm0, nds); 4186 vandpd(dst, xmm0, negate_field, vector_len); 4187 movdqu(xmm0, src); 4188 } else if (nds_enc < 16) { 4189 movdqu(src, xmm0); 4190 vandpd(xmm0, nds, negate_field, vector_len); 4191 movsd(dst, xmm0); 4192 movdqu(xmm0, src); 4193 } else { 4194 movdqu(src, xmm0); 4195 movsd(xmm0, nds); 4196 vandpd(xmm0, xmm0, negate_field, vector_len); 4197 movsd(dst, xmm0); 4198 movdqu(xmm0, src); 4199 } 4200 } 4201 4202 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4203 int dst_enc = dst->encoding(); 4204 int nds_enc = nds->encoding(); 4205 int src_enc = src->encoding(); 4206 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4207 Assembler::vpaddb(dst, nds, src, vector_len); 4208 } else if ((dst_enc < 16) && (src_enc < 16)) { 4209 Assembler::vpaddb(dst, dst, src, vector_len); 4210 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4211 // use nds as scratch for src 4212 evmovdqul(nds, src, Assembler::AVX_512bit); 4213 Assembler::vpaddb(dst, dst, nds, vector_len); 4214 } else if ((src_enc < 16) && (nds_enc < 16)) { 4215 // use nds as scratch for dst 4216 evmovdqul(nds, dst, Assembler::AVX_512bit); 4217 Assembler::vpaddb(nds, nds, src, vector_len); 4218 evmovdqul(dst, nds, Assembler::AVX_512bit); 4219 } else if (dst_enc < 16) { 4220 // use nds as scatch for xmm0 to hold src 4221 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4222 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4223 Assembler::vpaddb(dst, dst, xmm0, vector_len); 4224 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4225 } else { 4226 // worse case scenario, all regs are in the upper bank 4227 subptr(rsp, 64); 4228 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4229 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4230 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4231 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4232 Assembler::vpaddb(xmm0, xmm0, xmm1, vector_len); 4233 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4234 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4235 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4236 addptr(rsp, 64); 4237 } 4238 } 4239 4240 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4241 int dst_enc = dst->encoding(); 4242 int nds_enc = nds->encoding(); 4243 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4244 Assembler::vpaddb(dst, nds, src, vector_len); 4245 } else if (dst_enc < 16) { 4246 Assembler::vpaddb(dst, dst, src, vector_len); 4247 } else if (nds_enc < 16) { 4248 // implies dst_enc in upper bank with src as scratch 4249 evmovdqul(nds, dst, Assembler::AVX_512bit); 4250 Assembler::vpaddb(nds, nds, src, vector_len); 4251 evmovdqul(dst, nds, Assembler::AVX_512bit); 4252 } else { 4253 // worse case scenario, all regs in upper bank 4254 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4255 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4256 Assembler::vpaddb(xmm0, xmm0, src, vector_len); 4257 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4258 } 4259 } 4260 4261 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4262 int dst_enc = dst->encoding(); 4263 int nds_enc = nds->encoding(); 4264 int src_enc = src->encoding(); 4265 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4266 Assembler::vpaddw(dst, nds, src, vector_len); 4267 } else if ((dst_enc < 16) && (src_enc < 16)) { 4268 Assembler::vpaddw(dst, dst, src, vector_len); 4269 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4270 // use nds as scratch for src 4271 evmovdqul(nds, src, Assembler::AVX_512bit); 4272 Assembler::vpaddw(dst, dst, nds, vector_len); 4273 } else if ((src_enc < 16) && (nds_enc < 16)) { 4274 // use nds as scratch for dst 4275 evmovdqul(nds, dst, Assembler::AVX_512bit); 4276 Assembler::vpaddw(nds, nds, src, vector_len); 4277 evmovdqul(dst, nds, Assembler::AVX_512bit); 4278 } else if (dst_enc < 16) { 4279 // use nds as scatch for xmm0 to hold src 4280 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4281 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4282 Assembler::vpaddw(dst, dst, xmm0, vector_len); 4283 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4284 } else { 4285 // worse case scenario, all regs are in the upper bank 4286 subptr(rsp, 64); 4287 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4288 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4289 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4290 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4291 Assembler::vpaddw(xmm0, xmm0, xmm1, vector_len); 4292 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4293 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4294 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4295 addptr(rsp, 64); 4296 } 4297 } 4298 4299 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4300 int dst_enc = dst->encoding(); 4301 int nds_enc = nds->encoding(); 4302 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4303 Assembler::vpaddw(dst, nds, src, vector_len); 4304 } else if (dst_enc < 16) { 4305 Assembler::vpaddw(dst, dst, src, vector_len); 4306 } else if (nds_enc < 16) { 4307 // implies dst_enc in upper bank with src as scratch 4308 evmovdqul(nds, dst, Assembler::AVX_512bit); 4309 Assembler::vpaddw(nds, nds, src, vector_len); 4310 evmovdqul(dst, nds, Assembler::AVX_512bit); 4311 } else { 4312 // worse case scenario, all regs in upper bank 4313 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4314 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4315 Assembler::vpaddw(xmm0, xmm0, src, vector_len); 4316 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4317 } 4318 } 4319 4320 void MacroAssembler::vpand(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 4321 if (reachable(src)) { 4322 Assembler::vpand(dst, nds, as_Address(src), vector_len); 4323 } else { 4324 lea(rscratch1, src); 4325 Assembler::vpand(dst, nds, Address(rscratch1, 0), vector_len); 4326 } 4327 } 4328 4329 void MacroAssembler::vpbroadcastw(XMMRegister dst, XMMRegister src) { 4330 int dst_enc = dst->encoding(); 4331 int src_enc = src->encoding(); 4332 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4333 Assembler::vpbroadcastw(dst, src); 4334 } else if ((dst_enc < 16) && (src_enc < 16)) { 4335 Assembler::vpbroadcastw(dst, src); 4336 } else if (src_enc < 16) { 4337 subptr(rsp, 64); 4338 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4339 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4340 Assembler::vpbroadcastw(xmm0, src); 4341 movdqu(dst, xmm0); 4342 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4343 addptr(rsp, 64); 4344 } else if (dst_enc < 16) { 4345 subptr(rsp, 64); 4346 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4347 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4348 Assembler::vpbroadcastw(dst, xmm0); 4349 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4350 addptr(rsp, 64); 4351 } else { 4352 subptr(rsp, 64); 4353 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4354 subptr(rsp, 64); 4355 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4356 movdqu(xmm0, src); 4357 movdqu(xmm1, dst); 4358 Assembler::vpbroadcastw(xmm1, xmm0); 4359 movdqu(dst, xmm1); 4360 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4361 addptr(rsp, 64); 4362 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4363 addptr(rsp, 64); 4364 } 4365 } 4366 4367 void MacroAssembler::vpcmpeqb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4368 int dst_enc = dst->encoding(); 4369 int nds_enc = nds->encoding(); 4370 int src_enc = src->encoding(); 4371 assert(dst_enc == nds_enc, ""); 4372 if ((dst_enc < 16) && (src_enc < 16)) { 4373 Assembler::vpcmpeqb(dst, nds, src, vector_len); 4374 } else if (src_enc < 16) { 4375 subptr(rsp, 64); 4376 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4377 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4378 Assembler::vpcmpeqb(xmm0, xmm0, src, vector_len); 4379 movdqu(dst, xmm0); 4380 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4381 addptr(rsp, 64); 4382 } else if (dst_enc < 16) { 4383 subptr(rsp, 64); 4384 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4385 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4386 Assembler::vpcmpeqb(dst, dst, xmm0, vector_len); 4387 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4388 addptr(rsp, 64); 4389 } else { 4390 subptr(rsp, 64); 4391 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4392 subptr(rsp, 64); 4393 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4394 movdqu(xmm0, src); 4395 movdqu(xmm1, dst); 4396 Assembler::vpcmpeqb(xmm1, xmm1, xmm0, vector_len); 4397 movdqu(dst, xmm1); 4398 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4399 addptr(rsp, 64); 4400 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4401 addptr(rsp, 64); 4402 } 4403 } 4404 4405 void MacroAssembler::vpcmpeqw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4406 int dst_enc = dst->encoding(); 4407 int nds_enc = nds->encoding(); 4408 int src_enc = src->encoding(); 4409 assert(dst_enc == nds_enc, ""); 4410 if ((dst_enc < 16) && (src_enc < 16)) { 4411 Assembler::vpcmpeqw(dst, nds, src, vector_len); 4412 } else if (src_enc < 16) { 4413 subptr(rsp, 64); 4414 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4415 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4416 Assembler::vpcmpeqw(xmm0, xmm0, src, vector_len); 4417 movdqu(dst, xmm0); 4418 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4419 addptr(rsp, 64); 4420 } else if (dst_enc < 16) { 4421 subptr(rsp, 64); 4422 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4423 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4424 Assembler::vpcmpeqw(dst, dst, xmm0, vector_len); 4425 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4426 addptr(rsp, 64); 4427 } else { 4428 subptr(rsp, 64); 4429 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4430 subptr(rsp, 64); 4431 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4432 movdqu(xmm0, src); 4433 movdqu(xmm1, dst); 4434 Assembler::vpcmpeqw(xmm1, xmm1, xmm0, vector_len); 4435 movdqu(dst, xmm1); 4436 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4437 addptr(rsp, 64); 4438 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4439 addptr(rsp, 64); 4440 } 4441 } 4442 4443 void MacroAssembler::vpmovzxbw(XMMRegister dst, Address src, int vector_len) { 4444 int dst_enc = dst->encoding(); 4445 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4446 Assembler::vpmovzxbw(dst, src, vector_len); 4447 } else if (dst_enc < 16) { 4448 Assembler::vpmovzxbw(dst, src, vector_len); 4449 } else { 4450 subptr(rsp, 64); 4451 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4452 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4453 Assembler::vpmovzxbw(xmm0, src, vector_len); 4454 movdqu(dst, xmm0); 4455 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4456 addptr(rsp, 64); 4457 } 4458 } 4459 4460 void MacroAssembler::vpmovmskb(Register dst, XMMRegister src) { 4461 int src_enc = src->encoding(); 4462 if (src_enc < 16) { 4463 Assembler::vpmovmskb(dst, src); 4464 } else { 4465 subptr(rsp, 64); 4466 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4467 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4468 Assembler::vpmovmskb(dst, xmm0); 4469 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4470 addptr(rsp, 64); 4471 } 4472 } 4473 4474 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4475 int dst_enc = dst->encoding(); 4476 int nds_enc = nds->encoding(); 4477 int src_enc = src->encoding(); 4478 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4479 Assembler::vpmullw(dst, nds, src, vector_len); 4480 } else if ((dst_enc < 16) && (src_enc < 16)) { 4481 Assembler::vpmullw(dst, dst, src, vector_len); 4482 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4483 // use nds as scratch for src 4484 evmovdqul(nds, src, Assembler::AVX_512bit); 4485 Assembler::vpmullw(dst, dst, nds, vector_len); 4486 } else if ((src_enc < 16) && (nds_enc < 16)) { 4487 // use nds as scratch for dst 4488 evmovdqul(nds, dst, Assembler::AVX_512bit); 4489 Assembler::vpmullw(nds, nds, src, vector_len); 4490 evmovdqul(dst, nds, Assembler::AVX_512bit); 4491 } else if (dst_enc < 16) { 4492 // use nds as scatch for xmm0 to hold src 4493 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4494 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4495 Assembler::vpmullw(dst, dst, xmm0, vector_len); 4496 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4497 } else { 4498 // worse case scenario, all regs are in the upper bank 4499 subptr(rsp, 64); 4500 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4501 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4502 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4503 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4504 Assembler::vpmullw(xmm0, xmm0, xmm1, vector_len); 4505 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4506 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4507 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4508 addptr(rsp, 64); 4509 } 4510 } 4511 4512 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4513 int dst_enc = dst->encoding(); 4514 int nds_enc = nds->encoding(); 4515 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4516 Assembler::vpmullw(dst, nds, src, vector_len); 4517 } else if (dst_enc < 16) { 4518 Assembler::vpmullw(dst, dst, src, vector_len); 4519 } else if (nds_enc < 16) { 4520 // implies dst_enc in upper bank with src as scratch 4521 evmovdqul(nds, dst, Assembler::AVX_512bit); 4522 Assembler::vpmullw(nds, nds, src, vector_len); 4523 evmovdqul(dst, nds, Assembler::AVX_512bit); 4524 } else { 4525 // worse case scenario, all regs in upper bank 4526 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4527 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4528 Assembler::vpmullw(xmm0, xmm0, src, vector_len); 4529 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4530 } 4531 } 4532 4533 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4534 int dst_enc = dst->encoding(); 4535 int nds_enc = nds->encoding(); 4536 int src_enc = src->encoding(); 4537 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4538 Assembler::vpsubb(dst, nds, src, vector_len); 4539 } else if ((dst_enc < 16) && (src_enc < 16)) { 4540 Assembler::vpsubb(dst, dst, src, vector_len); 4541 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4542 // use nds as scratch for src 4543 evmovdqul(nds, src, Assembler::AVX_512bit); 4544 Assembler::vpsubb(dst, dst, nds, vector_len); 4545 } else if ((src_enc < 16) && (nds_enc < 16)) { 4546 // use nds as scratch for dst 4547 evmovdqul(nds, dst, Assembler::AVX_512bit); 4548 Assembler::vpsubb(nds, nds, src, vector_len); 4549 evmovdqul(dst, nds, Assembler::AVX_512bit); 4550 } else if (dst_enc < 16) { 4551 // use nds as scatch for xmm0 to hold src 4552 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4553 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4554 Assembler::vpsubb(dst, dst, xmm0, vector_len); 4555 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4556 } else { 4557 // worse case scenario, all regs are in the upper bank 4558 subptr(rsp, 64); 4559 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4560 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4561 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4562 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4563 Assembler::vpsubb(xmm0, xmm0, xmm1, vector_len); 4564 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4565 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4566 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4567 addptr(rsp, 64); 4568 } 4569 } 4570 4571 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4572 int dst_enc = dst->encoding(); 4573 int nds_enc = nds->encoding(); 4574 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4575 Assembler::vpsubb(dst, nds, src, vector_len); 4576 } else if (dst_enc < 16) { 4577 Assembler::vpsubb(dst, dst, src, vector_len); 4578 } else if (nds_enc < 16) { 4579 // implies dst_enc in upper bank with src as scratch 4580 evmovdqul(nds, dst, Assembler::AVX_512bit); 4581 Assembler::vpsubb(nds, nds, src, vector_len); 4582 evmovdqul(dst, nds, Assembler::AVX_512bit); 4583 } else { 4584 // worse case scenario, all regs in upper bank 4585 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4586 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4587 Assembler::vpsubw(xmm0, xmm0, src, vector_len); 4588 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4589 } 4590 } 4591 4592 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4593 int dst_enc = dst->encoding(); 4594 int nds_enc = nds->encoding(); 4595 int src_enc = src->encoding(); 4596 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4597 Assembler::vpsubw(dst, nds, src, vector_len); 4598 } else if ((dst_enc < 16) && (src_enc < 16)) { 4599 Assembler::vpsubw(dst, dst, src, vector_len); 4600 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4601 // use nds as scratch for src 4602 evmovdqul(nds, src, Assembler::AVX_512bit); 4603 Assembler::vpsubw(dst, dst, nds, vector_len); 4604 } else if ((src_enc < 16) && (nds_enc < 16)) { 4605 // use nds as scratch for dst 4606 evmovdqul(nds, dst, Assembler::AVX_512bit); 4607 Assembler::vpsubw(nds, nds, src, vector_len); 4608 evmovdqul(dst, nds, Assembler::AVX_512bit); 4609 } else if (dst_enc < 16) { 4610 // use nds as scatch for xmm0 to hold src 4611 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4612 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4613 Assembler::vpsubw(dst, dst, xmm0, vector_len); 4614 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4615 } else { 4616 // worse case scenario, all regs are in the upper bank 4617 subptr(rsp, 64); 4618 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4619 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4620 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4621 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4622 Assembler::vpsubw(xmm0, xmm0, xmm1, vector_len); 4623 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4624 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4625 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4626 addptr(rsp, 64); 4627 } 4628 } 4629 4630 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4631 int dst_enc = dst->encoding(); 4632 int nds_enc = nds->encoding(); 4633 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4634 Assembler::vpsubw(dst, nds, src, vector_len); 4635 } else if (dst_enc < 16) { 4636 Assembler::vpsubw(dst, dst, src, vector_len); 4637 } else if (nds_enc < 16) { 4638 // implies dst_enc in upper bank with src as scratch 4639 evmovdqul(nds, dst, Assembler::AVX_512bit); 4640 Assembler::vpsubw(nds, nds, src, vector_len); 4641 evmovdqul(dst, nds, Assembler::AVX_512bit); 4642 } else { 4643 // worse case scenario, all regs in upper bank 4644 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4645 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4646 Assembler::vpsubw(xmm0, xmm0, src, vector_len); 4647 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4648 } 4649 } 4650 4651 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4652 int dst_enc = dst->encoding(); 4653 int nds_enc = nds->encoding(); 4654 int shift_enc = shift->encoding(); 4655 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4656 Assembler::vpsraw(dst, nds, shift, vector_len); 4657 } else if ((dst_enc < 16) && (shift_enc < 16)) { 4658 Assembler::vpsraw(dst, dst, shift, vector_len); 4659 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4660 // use nds_enc as scratch with shift 4661 evmovdqul(nds, shift, Assembler::AVX_512bit); 4662 Assembler::vpsraw(dst, dst, nds, vector_len); 4663 } else if ((shift_enc < 16) && (nds_enc < 16)) { 4664 // use nds as scratch with dst 4665 evmovdqul(nds, dst, Assembler::AVX_512bit); 4666 Assembler::vpsraw(nds, nds, shift, vector_len); 4667 evmovdqul(dst, nds, Assembler::AVX_512bit); 4668 } else if (dst_enc < 16) { 4669 // use nds to save a copy of xmm0 and hold shift 4670 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4671 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4672 Assembler::vpsraw(dst, dst, xmm0, vector_len); 4673 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4674 } else if (nds_enc < 16) { 4675 // use nds as dest as temps 4676 evmovdqul(nds, dst, Assembler::AVX_512bit); 4677 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4678 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4679 Assembler::vpsraw(nds, nds, xmm0, vector_len); 4680 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4681 evmovdqul(dst, nds, Assembler::AVX_512bit); 4682 } else { 4683 // worse case scenario, all regs are in the upper bank 4684 subptr(rsp, 64); 4685 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4686 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4687 evmovdqul(xmm1, shift, Assembler::AVX_512bit); 4688 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4689 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len); 4690 evmovdqul(xmm1, dst, Assembler::AVX_512bit); 4691 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4692 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4693 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4694 addptr(rsp, 64); 4695 } 4696 } 4697 4698 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 4699 int dst_enc = dst->encoding(); 4700 int nds_enc = nds->encoding(); 4701 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4702 Assembler::vpsraw(dst, nds, shift, vector_len); 4703 } else if (dst_enc < 16) { 4704 Assembler::vpsraw(dst, dst, shift, vector_len); 4705 } else if (nds_enc < 16) { 4706 // use nds as scratch 4707 evmovdqul(nds, dst, Assembler::AVX_512bit); 4708 Assembler::vpsraw(nds, nds, shift, vector_len); 4709 evmovdqul(dst, nds, Assembler::AVX_512bit); 4710 } else { 4711 // use nds as scratch for xmm0 4712 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4713 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4714 Assembler::vpsraw(xmm0, xmm0, shift, vector_len); 4715 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4716 } 4717 } 4718 4719 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4720 int dst_enc = dst->encoding(); 4721 int nds_enc = nds->encoding(); 4722 int shift_enc = shift->encoding(); 4723 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4724 Assembler::vpsrlw(dst, nds, shift, vector_len); 4725 } else if ((dst_enc < 16) && (shift_enc < 16)) { 4726 Assembler::vpsrlw(dst, dst, shift, vector_len); 4727 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4728 // use nds_enc as scratch with shift 4729 evmovdqul(nds, shift, Assembler::AVX_512bit); 4730 Assembler::vpsrlw(dst, dst, nds, vector_len); 4731 } else if ((shift_enc < 16) && (nds_enc < 16)) { 4732 // use nds as scratch with dst 4733 evmovdqul(nds, dst, Assembler::AVX_512bit); 4734 Assembler::vpsrlw(nds, nds, shift, vector_len); 4735 evmovdqul(dst, nds, Assembler::AVX_512bit); 4736 } else if (dst_enc < 16) { 4737 // use nds to save a copy of xmm0 and hold shift 4738 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4739 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4740 Assembler::vpsrlw(dst, dst, xmm0, vector_len); 4741 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4742 } else if (nds_enc < 16) { 4743 // use nds as dest as temps 4744 evmovdqul(nds, dst, Assembler::AVX_512bit); 4745 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4746 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4747 Assembler::vpsrlw(nds, nds, xmm0, vector_len); 4748 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4749 evmovdqul(dst, nds, Assembler::AVX_512bit); 4750 } else { 4751 // worse case scenario, all regs are in the upper bank 4752 subptr(rsp, 64); 4753 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4754 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4755 evmovdqul(xmm1, shift, Assembler::AVX_512bit); 4756 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4757 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len); 4758 evmovdqul(xmm1, dst, Assembler::AVX_512bit); 4759 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4760 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4761 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4762 addptr(rsp, 64); 4763 } 4764 } 4765 4766 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 4767 int dst_enc = dst->encoding(); 4768 int nds_enc = nds->encoding(); 4769 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4770 Assembler::vpsrlw(dst, nds, shift, vector_len); 4771 } else if (dst_enc < 16) { 4772 Assembler::vpsrlw(dst, dst, shift, vector_len); 4773 } else if (nds_enc < 16) { 4774 // use nds as scratch 4775 evmovdqul(nds, dst, Assembler::AVX_512bit); 4776 Assembler::vpsrlw(nds, nds, shift, vector_len); 4777 evmovdqul(dst, nds, Assembler::AVX_512bit); 4778 } else { 4779 // use nds as scratch for xmm0 4780 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4781 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4782 Assembler::vpsrlw(xmm0, xmm0, shift, vector_len); 4783 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4784 } 4785 } 4786 4787 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4788 int dst_enc = dst->encoding(); 4789 int nds_enc = nds->encoding(); 4790 int shift_enc = shift->encoding(); 4791 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4792 Assembler::vpsllw(dst, nds, shift, vector_len); 4793 } else if ((dst_enc < 16) && (shift_enc < 16)) { 4794 Assembler::vpsllw(dst, dst, shift, vector_len); 4795 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4796 // use nds_enc as scratch with shift 4797 evmovdqul(nds, shift, Assembler::AVX_512bit); 4798 Assembler::vpsllw(dst, dst, nds, vector_len); 4799 } else if ((shift_enc < 16) && (nds_enc < 16)) { 4800 // use nds as scratch with dst 4801 evmovdqul(nds, dst, Assembler::AVX_512bit); 4802 Assembler::vpsllw(nds, nds, shift, vector_len); 4803 evmovdqul(dst, nds, Assembler::AVX_512bit); 4804 } else if (dst_enc < 16) { 4805 // use nds to save a copy of xmm0 and hold shift 4806 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4807 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4808 Assembler::vpsllw(dst, dst, xmm0, vector_len); 4809 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4810 } else if (nds_enc < 16) { 4811 // use nds as dest as temps 4812 evmovdqul(nds, dst, Assembler::AVX_512bit); 4813 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4814 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4815 Assembler::vpsllw(nds, nds, xmm0, vector_len); 4816 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4817 evmovdqul(dst, nds, Assembler::AVX_512bit); 4818 } else { 4819 // worse case scenario, all regs are in the upper bank 4820 subptr(rsp, 64); 4821 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4822 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4823 evmovdqul(xmm1, shift, Assembler::AVX_512bit); 4824 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4825 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len); 4826 evmovdqul(xmm1, dst, Assembler::AVX_512bit); 4827 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4828 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4829 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4830 addptr(rsp, 64); 4831 } 4832 } 4833 4834 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 4835 int dst_enc = dst->encoding(); 4836 int nds_enc = nds->encoding(); 4837 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4838 Assembler::vpsllw(dst, nds, shift, vector_len); 4839 } else if (dst_enc < 16) { 4840 Assembler::vpsllw(dst, dst, shift, vector_len); 4841 } else if (nds_enc < 16) { 4842 // use nds as scratch 4843 evmovdqul(nds, dst, Assembler::AVX_512bit); 4844 Assembler::vpsllw(nds, nds, shift, vector_len); 4845 evmovdqul(dst, nds, Assembler::AVX_512bit); 4846 } else { 4847 // use nds as scratch for xmm0 4848 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4849 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4850 Assembler::vpsllw(xmm0, xmm0, shift, vector_len); 4851 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4852 } 4853 } 4854 4855 void MacroAssembler::vptest(XMMRegister dst, XMMRegister src) { 4856 int dst_enc = dst->encoding(); 4857 int src_enc = src->encoding(); 4858 if ((dst_enc < 16) && (src_enc < 16)) { 4859 Assembler::vptest(dst, src); 4860 } else if (src_enc < 16) { 4861 subptr(rsp, 64); 4862 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4863 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4864 Assembler::vptest(xmm0, src); 4865 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4866 addptr(rsp, 64); 4867 } else if (dst_enc < 16) { 4868 subptr(rsp, 64); 4869 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4870 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4871 Assembler::vptest(dst, xmm0); 4872 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4873 addptr(rsp, 64); 4874 } else { 4875 subptr(rsp, 64); 4876 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4877 subptr(rsp, 64); 4878 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4879 movdqu(xmm0, src); 4880 movdqu(xmm1, dst); 4881 Assembler::vptest(xmm1, xmm0); 4882 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4883 addptr(rsp, 64); 4884 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4885 addptr(rsp, 64); 4886 } 4887 } 4888 4889 // This instruction exists within macros, ergo we cannot control its input 4890 // when emitted through those patterns. 4891 void MacroAssembler::punpcklbw(XMMRegister dst, XMMRegister src) { 4892 if (VM_Version::supports_avx512nobw()) { 4893 int dst_enc = dst->encoding(); 4894 int src_enc = src->encoding(); 4895 if (dst_enc == src_enc) { 4896 if (dst_enc < 16) { 4897 Assembler::punpcklbw(dst, src); 4898 } else { 4899 subptr(rsp, 64); 4900 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4901 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4902 Assembler::punpcklbw(xmm0, xmm0); 4903 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4904 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4905 addptr(rsp, 64); 4906 } 4907 } else { 4908 if ((src_enc < 16) && (dst_enc < 16)) { 4909 Assembler::punpcklbw(dst, src); 4910 } else if (src_enc < 16) { 4911 subptr(rsp, 64); 4912 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4913 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4914 Assembler::punpcklbw(xmm0, src); 4915 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4916 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4917 addptr(rsp, 64); 4918 } else if (dst_enc < 16) { 4919 subptr(rsp, 64); 4920 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4921 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4922 Assembler::punpcklbw(dst, xmm0); 4923 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4924 addptr(rsp, 64); 4925 } else { 4926 subptr(rsp, 64); 4927 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4928 subptr(rsp, 64); 4929 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4930 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4931 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4932 Assembler::punpcklbw(xmm0, xmm1); 4933 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4934 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4935 addptr(rsp, 64); 4936 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4937 addptr(rsp, 64); 4938 } 4939 } 4940 } else { 4941 Assembler::punpcklbw(dst, src); 4942 } 4943 } 4944 4945 // This instruction exists within macros, ergo we cannot control its input 4946 // when emitted through those patterns. 4947 void MacroAssembler::pshuflw(XMMRegister dst, XMMRegister src, int mode) { 4948 if (VM_Version::supports_avx512nobw()) { 4949 int dst_enc = dst->encoding(); 4950 int src_enc = src->encoding(); 4951 if (dst_enc == src_enc) { 4952 if (dst_enc < 16) { 4953 Assembler::pshuflw(dst, src, mode); 4954 } else { 4955 subptr(rsp, 64); 4956 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4957 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4958 Assembler::pshuflw(xmm0, xmm0, mode); 4959 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4960 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4961 addptr(rsp, 64); 4962 } 4963 } else { 4964 if ((src_enc < 16) && (dst_enc < 16)) { 4965 Assembler::pshuflw(dst, src, mode); 4966 } else if (src_enc < 16) { 4967 subptr(rsp, 64); 4968 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4969 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4970 Assembler::pshuflw(xmm0, src, mode); 4971 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4972 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4973 addptr(rsp, 64); 4974 } else if (dst_enc < 16) { 4975 subptr(rsp, 64); 4976 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4977 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4978 Assembler::pshuflw(dst, xmm0, mode); 4979 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4980 addptr(rsp, 64); 4981 } else { 4982 subptr(rsp, 64); 4983 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4984 subptr(rsp, 64); 4985 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4986 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4987 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4988 Assembler::pshuflw(xmm0, xmm1, mode); 4989 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4990 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4991 addptr(rsp, 64); 4992 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4993 addptr(rsp, 64); 4994 } 4995 } 4996 } else { 4997 Assembler::pshuflw(dst, src, mode); 4998 } 4999 } 5000 5001 void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 5002 if (reachable(src)) { 5003 vandpd(dst, nds, as_Address(src), vector_len); 5004 } else { 5005 lea(rscratch1, src); 5006 vandpd(dst, nds, Address(rscratch1, 0), vector_len); 5007 } 5008 } 5009 5010 void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 5011 if (reachable(src)) { 5012 vandps(dst, nds, as_Address(src), vector_len); 5013 } else { 5014 lea(rscratch1, src); 5015 vandps(dst, nds, Address(rscratch1, 0), vector_len); 5016 } 5017 } 5018 5019 void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5020 if (reachable(src)) { 5021 vdivsd(dst, nds, as_Address(src)); 5022 } else { 5023 lea(rscratch1, src); 5024 vdivsd(dst, nds, Address(rscratch1, 0)); 5025 } 5026 } 5027 5028 void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5029 if (reachable(src)) { 5030 vdivss(dst, nds, as_Address(src)); 5031 } else { 5032 lea(rscratch1, src); 5033 vdivss(dst, nds, Address(rscratch1, 0)); 5034 } 5035 } 5036 5037 void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5038 if (reachable(src)) { 5039 vmulsd(dst, nds, as_Address(src)); 5040 } else { 5041 lea(rscratch1, src); 5042 vmulsd(dst, nds, Address(rscratch1, 0)); 5043 } 5044 } 5045 5046 void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5047 if (reachable(src)) { 5048 vmulss(dst, nds, as_Address(src)); 5049 } else { 5050 lea(rscratch1, src); 5051 vmulss(dst, nds, Address(rscratch1, 0)); 5052 } 5053 } 5054 5055 void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5056 if (reachable(src)) { 5057 vsubsd(dst, nds, as_Address(src)); 5058 } else { 5059 lea(rscratch1, src); 5060 vsubsd(dst, nds, Address(rscratch1, 0)); 5061 } 5062 } 5063 5064 void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5065 if (reachable(src)) { 5066 vsubss(dst, nds, as_Address(src)); 5067 } else { 5068 lea(rscratch1, src); 5069 vsubss(dst, nds, Address(rscratch1, 0)); 5070 } 5071 } 5072 5073 void MacroAssembler::vnegatess(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5074 int nds_enc = nds->encoding(); 5075 int dst_enc = dst->encoding(); 5076 bool dst_upper_bank = (dst_enc > 15); 5077 bool nds_upper_bank = (nds_enc > 15); 5078 if (VM_Version::supports_avx512novl() && 5079 (nds_upper_bank || dst_upper_bank)) { 5080 if (dst_upper_bank) { 5081 subptr(rsp, 64); 5082 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5083 movflt(xmm0, nds); 5084 vxorps(xmm0, xmm0, src, Assembler::AVX_128bit); 5085 movflt(dst, xmm0); 5086 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5087 addptr(rsp, 64); 5088 } else { 5089 movflt(dst, nds); 5090 vxorps(dst, dst, src, Assembler::AVX_128bit); 5091 } 5092 } else { 5093 vxorps(dst, nds, src, Assembler::AVX_128bit); 5094 } 5095 } 5096 5097 void MacroAssembler::vnegatesd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5098 int nds_enc = nds->encoding(); 5099 int dst_enc = dst->encoding(); 5100 bool dst_upper_bank = (dst_enc > 15); 5101 bool nds_upper_bank = (nds_enc > 15); 5102 if (VM_Version::supports_avx512novl() && 5103 (nds_upper_bank || dst_upper_bank)) { 5104 if (dst_upper_bank) { 5105 subptr(rsp, 64); 5106 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5107 movdbl(xmm0, nds); 5108 vxorpd(xmm0, xmm0, src, Assembler::AVX_128bit); 5109 movdbl(dst, xmm0); 5110 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5111 addptr(rsp, 64); 5112 } else { 5113 movdbl(dst, nds); 5114 vxorpd(dst, dst, src, Assembler::AVX_128bit); 5115 } 5116 } else { 5117 vxorpd(dst, nds, src, Assembler::AVX_128bit); 5118 } 5119 } 5120 5121 void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 5122 if (reachable(src)) { 5123 vxorpd(dst, nds, as_Address(src), vector_len); 5124 } else { 5125 lea(rscratch1, src); 5126 vxorpd(dst, nds, Address(rscratch1, 0), vector_len); 5127 } 5128 } 5129 5130 void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 5131 if (reachable(src)) { 5132 vxorps(dst, nds, as_Address(src), vector_len); 5133 } else { 5134 lea(rscratch1, src); 5135 vxorps(dst, nds, Address(rscratch1, 0), vector_len); 5136 } 5137 } 5138 5139 5140 void MacroAssembler::resolve_jobject(Register value, 5141 Register thread, 5142 Register tmp) { 5143 assert_different_registers(value, thread, tmp); 5144 Label done, not_weak; 5145 testptr(value, value); 5146 jcc(Assembler::zero, done); // Use NULL as-is. 5147 testptr(value, JNIHandles::weak_tag_mask); // Test for jweak tag. 5148 jcc(Assembler::zero, not_weak); 5149 // Resolve jweak. 5150 movptr(value, Address(value, -JNIHandles::weak_tag_value)); 5151 verify_oop(value); 5152 #if INCLUDE_ALL_GCS 5153 if (UseG1GC) { 5154 g1_write_barrier_pre(noreg /* obj */, 5155 value /* pre_val */, 5156 thread /* thread */, 5157 tmp /* tmp */, 5158 true /* tosca_live */, 5159 true /* expand_call */); 5160 } 5161 #endif // INCLUDE_ALL_GCS 5162 jmp(done); 5163 bind(not_weak); 5164 // Resolve (untagged) jobject. 5165 movptr(value, Address(value, 0)); 5166 verify_oop(value); 5167 bind(done); 5168 } 5169 5170 void MacroAssembler::clear_jweak_tag(Register possibly_jweak) { 5171 const int32_t inverted_jweak_mask = ~static_cast<int32_t>(JNIHandles::weak_tag_mask); 5172 STATIC_ASSERT(inverted_jweak_mask == -2); // otherwise check this code 5173 // The inverted mask is sign-extended 5174 andptr(possibly_jweak, inverted_jweak_mask); 5175 } 5176 5177 ////////////////////////////////////////////////////////////////////////////////// 5178 #if INCLUDE_ALL_GCS 5179 5180 void MacroAssembler::g1_write_barrier_pre(Register obj, 5181 Register pre_val, 5182 Register thread, 5183 Register tmp, 5184 bool tosca_live, 5185 bool expand_call) { 5186 5187 // If expand_call is true then we expand the call_VM_leaf macro 5188 // directly to skip generating the check by 5189 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp. 5190 5191 #ifdef _LP64 5192 assert(thread == r15_thread, "must be"); 5193 #endif // _LP64 5194 5195 Label done; 5196 Label runtime; 5197 5198 assert(pre_val != noreg, "check this code"); 5199 5200 if (obj != noreg) { 5201 assert_different_registers(obj, pre_val, tmp); 5202 assert(pre_val != rax, "check this code"); 5203 } 5204 5205 Address in_progress(thread, in_bytes(JavaThread::satb_mark_queue_offset() + 5206 SATBMarkQueue::byte_offset_of_active())); 5207 Address index(thread, in_bytes(JavaThread::satb_mark_queue_offset() + 5208 SATBMarkQueue::byte_offset_of_index())); 5209 Address buffer(thread, in_bytes(JavaThread::satb_mark_queue_offset() + 5210 SATBMarkQueue::byte_offset_of_buf())); 5211 5212 5213 // Is marking active? 5214 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) { 5215 cmpl(in_progress, 0); 5216 } else { 5217 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption"); 5218 cmpb(in_progress, 0); 5219 } 5220 jcc(Assembler::equal, done); 5221 5222 // Do we need to load the previous value? 5223 if (obj != noreg) { 5224 load_heap_oop(pre_val, Address(obj, 0)); 5225 } 5226 5227 // Is the previous value null? 5228 cmpptr(pre_val, (int32_t) NULL_WORD); 5229 jcc(Assembler::equal, done); 5230 5231 // Can we store original value in the thread's buffer? 5232 // Is index == 0? 5233 // (The index field is typed as size_t.) 5234 5235 movptr(tmp, index); // tmp := *index_adr 5236 cmpptr(tmp, 0); // tmp == 0? 5237 jcc(Assembler::equal, runtime); // If yes, goto runtime 5238 5239 subptr(tmp, wordSize); // tmp := tmp - wordSize 5240 movptr(index, tmp); // *index_adr := tmp 5241 addptr(tmp, buffer); // tmp := tmp + *buffer_adr 5242 5243 // Record the previous value 5244 movptr(Address(tmp, 0), pre_val); 5245 jmp(done); 5246 5247 bind(runtime); 5248 // save the live input values 5249 if(tosca_live) push(rax); 5250 5251 if (obj != noreg && obj != rax) 5252 push(obj); 5253 5254 if (pre_val != rax) 5255 push(pre_val); 5256 5257 // Calling the runtime using the regular call_VM_leaf mechanism generates 5258 // code (generated by InterpreterMacroAssember::call_VM_leaf_base) 5259 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL. 5260 // 5261 // If we care generating the pre-barrier without a frame (e.g. in the 5262 // intrinsified Reference.get() routine) then ebp might be pointing to 5263 // the caller frame and so this check will most likely fail at runtime. 5264 // 5265 // Expanding the call directly bypasses the generation of the check. 5266 // So when we do not have have a full interpreter frame on the stack 5267 // expand_call should be passed true. 5268 5269 NOT_LP64( push(thread); ) 5270 5271 if (expand_call) { 5272 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); ) 5273 pass_arg1(this, thread); 5274 pass_arg0(this, pre_val); 5275 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), 2); 5276 } else { 5277 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), pre_val, thread); 5278 } 5279 5280 NOT_LP64( pop(thread); ) 5281 5282 // save the live input values 5283 if (pre_val != rax) 5284 pop(pre_val); 5285 5286 if (obj != noreg && obj != rax) 5287 pop(obj); 5288 5289 if(tosca_live) pop(rax); 5290 5291 bind(done); 5292 } 5293 5294 void MacroAssembler::g1_write_barrier_post(Register store_addr, 5295 Register new_val, 5296 Register thread, 5297 Register tmp, 5298 Register tmp2) { 5299 #ifdef _LP64 5300 assert(thread == r15_thread, "must be"); 5301 #endif // _LP64 5302 5303 Address queue_index(thread, in_bytes(JavaThread::dirty_card_queue_offset() + 5304 DirtyCardQueue::byte_offset_of_index())); 5305 Address buffer(thread, in_bytes(JavaThread::dirty_card_queue_offset() + 5306 DirtyCardQueue::byte_offset_of_buf())); 5307 5308 CardTableModRefBS* ct = 5309 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 5310 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code"); 5311 5312 Label done; 5313 Label runtime; 5314 5315 // Does store cross heap regions? 5316 5317 movptr(tmp, store_addr); 5318 xorptr(tmp, new_val); 5319 shrptr(tmp, HeapRegion::LogOfHRGrainBytes); 5320 jcc(Assembler::equal, done); 5321 5322 // crosses regions, storing NULL? 5323 5324 cmpptr(new_val, (int32_t) NULL_WORD); 5325 jcc(Assembler::equal, done); 5326 5327 // storing region crossing non-NULL, is card already dirty? 5328 5329 const Register card_addr = tmp; 5330 const Register cardtable = tmp2; 5331 5332 movptr(card_addr, store_addr); 5333 shrptr(card_addr, CardTableModRefBS::card_shift); 5334 // Do not use ExternalAddress to load 'byte_map_base', since 'byte_map_base' is NOT 5335 // a valid address and therefore is not properly handled by the relocation code. 5336 movptr(cardtable, (intptr_t)ct->byte_map_base); 5337 addptr(card_addr, cardtable); 5338 5339 cmpb(Address(card_addr, 0), (int)G1SATBCardTableModRefBS::g1_young_card_val()); 5340 jcc(Assembler::equal, done); 5341 5342 membar(Assembler::Membar_mask_bits(Assembler::StoreLoad)); 5343 cmpb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val()); 5344 jcc(Assembler::equal, done); 5345 5346 5347 // storing a region crossing, non-NULL oop, card is clean. 5348 // dirty card and log. 5349 5350 movb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val()); 5351 5352 cmpl(queue_index, 0); 5353 jcc(Assembler::equal, runtime); 5354 subl(queue_index, wordSize); 5355 movptr(tmp2, buffer); 5356 #ifdef _LP64 5357 movslq(rscratch1, queue_index); 5358 addq(tmp2, rscratch1); 5359 movq(Address(tmp2, 0), card_addr); 5360 #else 5361 addl(tmp2, queue_index); 5362 movl(Address(tmp2, 0), card_addr); 5363 #endif 5364 jmp(done); 5365 5366 bind(runtime); 5367 // save the live input values 5368 push(store_addr); 5369 push(new_val); 5370 #ifdef _LP64 5371 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, r15_thread); 5372 #else 5373 push(thread); 5374 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, thread); 5375 pop(thread); 5376 #endif 5377 pop(new_val); 5378 pop(store_addr); 5379 5380 bind(done); 5381 } 5382 5383 #endif // INCLUDE_ALL_GCS 5384 ////////////////////////////////////////////////////////////////////////////////// 5385 5386 5387 void MacroAssembler::store_check(Register obj, Address dst) { 5388 store_check(obj); 5389 } 5390 5391 void MacroAssembler::store_check(Register obj) { 5392 // Does a store check for the oop in register obj. The content of 5393 // register obj is destroyed afterwards. 5394 BarrierSet* bs = Universe::heap()->barrier_set(); 5395 assert(bs->kind() == BarrierSet::CardTableForRS || 5396 bs->kind() == BarrierSet::CardTableExtension, 5397 "Wrong barrier set kind"); 5398 5399 CardTableModRefBS* ct = barrier_set_cast<CardTableModRefBS>(bs); 5400 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code"); 5401 5402 shrptr(obj, CardTableModRefBS::card_shift); 5403 5404 Address card_addr; 5405 5406 // The calculation for byte_map_base is as follows: 5407 // byte_map_base = _byte_map - (uintptr_t(low_bound) >> card_shift); 5408 // So this essentially converts an address to a displacement and it will 5409 // never need to be relocated. On 64bit however the value may be too 5410 // large for a 32bit displacement. 5411 intptr_t disp = (intptr_t) ct->byte_map_base; 5412 if (is_simm32(disp)) { 5413 card_addr = Address(noreg, obj, Address::times_1, disp); 5414 } else { 5415 // By doing it as an ExternalAddress 'disp' could be converted to a rip-relative 5416 // displacement and done in a single instruction given favorable mapping and a 5417 // smarter version of as_Address. However, 'ExternalAddress' generates a relocation 5418 // entry and that entry is not properly handled by the relocation code. 5419 AddressLiteral cardtable((address)ct->byte_map_base, relocInfo::none); 5420 Address index(noreg, obj, Address::times_1); 5421 card_addr = as_Address(ArrayAddress(cardtable, index)); 5422 } 5423 5424 int dirty = CardTableModRefBS::dirty_card_val(); 5425 if (UseCondCardMark) { 5426 Label L_already_dirty; 5427 if (UseConcMarkSweepGC) { 5428 membar(Assembler::StoreLoad); 5429 } 5430 cmpb(card_addr, dirty); 5431 jcc(Assembler::equal, L_already_dirty); 5432 movb(card_addr, dirty); 5433 bind(L_already_dirty); 5434 } else { 5435 movb(card_addr, dirty); 5436 } 5437 } 5438 5439 void MacroAssembler::subptr(Register dst, int32_t imm32) { 5440 LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32)); 5441 } 5442 5443 // Force generation of a 4 byte immediate value even if it fits into 8bit 5444 void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) { 5445 LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32)); 5446 } 5447 5448 void MacroAssembler::subptr(Register dst, Register src) { 5449 LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src)); 5450 } 5451 5452 // C++ bool manipulation 5453 void MacroAssembler::testbool(Register dst) { 5454 if(sizeof(bool) == 1) 5455 testb(dst, 0xff); 5456 else if(sizeof(bool) == 2) { 5457 // testw implementation needed for two byte bools 5458 ShouldNotReachHere(); 5459 } else if(sizeof(bool) == 4) 5460 testl(dst, dst); 5461 else 5462 // unsupported 5463 ShouldNotReachHere(); 5464 } 5465 5466 void MacroAssembler::testptr(Register dst, Register src) { 5467 LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src)); 5468 } 5469 5470 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes. 5471 void MacroAssembler::tlab_allocate(Register obj, 5472 Register var_size_in_bytes, 5473 int con_size_in_bytes, 5474 Register t1, 5475 Register t2, 5476 Label& slow_case) { 5477 assert_different_registers(obj, t1, t2); 5478 assert_different_registers(obj, var_size_in_bytes, t1); 5479 Register end = t2; 5480 Register thread = NOT_LP64(t1) LP64_ONLY(r15_thread); 5481 5482 verify_tlab(); 5483 5484 NOT_LP64(get_thread(thread)); 5485 5486 movptr(obj, Address(thread, JavaThread::tlab_top_offset())); 5487 if (var_size_in_bytes == noreg) { 5488 lea(end, Address(obj, con_size_in_bytes)); 5489 } else { 5490 lea(end, Address(obj, var_size_in_bytes, Address::times_1)); 5491 } 5492 cmpptr(end, Address(thread, JavaThread::tlab_end_offset())); 5493 jcc(Assembler::above, slow_case); 5494 5495 // update the tlab top pointer 5496 movptr(Address(thread, JavaThread::tlab_top_offset()), end); 5497 5498 // recover var_size_in_bytes if necessary 5499 if (var_size_in_bytes == end) { 5500 subptr(var_size_in_bytes, obj); 5501 } 5502 verify_tlab(); 5503 } 5504 5505 // Preserves rbx, and rdx. 5506 Register MacroAssembler::tlab_refill(Label& retry, 5507 Label& try_eden, 5508 Label& slow_case) { 5509 Register top = rax; 5510 Register t1 = rcx; // object size 5511 Register t2 = rsi; 5512 Register thread_reg = NOT_LP64(rdi) LP64_ONLY(r15_thread); 5513 assert_different_registers(top, thread_reg, t1, t2, /* preserve: */ rbx, rdx); 5514 Label do_refill, discard_tlab; 5515 5516 if (!Universe::heap()->supports_inline_contig_alloc()) { 5517 // No allocation in the shared eden. 5518 jmp(slow_case); 5519 } 5520 5521 NOT_LP64(get_thread(thread_reg)); 5522 5523 movptr(top, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 5524 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset()))); 5525 5526 // calculate amount of free space 5527 subptr(t1, top); 5528 shrptr(t1, LogHeapWordSize); 5529 5530 // Retain tlab and allocate object in shared space if 5531 // the amount free in the tlab is too large to discard. 5532 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset()))); 5533 jcc(Assembler::lessEqual, discard_tlab); 5534 5535 // Retain 5536 // %%% yuck as movptr... 5537 movptr(t2, (int32_t) ThreadLocalAllocBuffer::refill_waste_limit_increment()); 5538 addptr(Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset())), t2); 5539 if (TLABStats) { 5540 // increment number of slow_allocations 5541 addl(Address(thread_reg, in_bytes(JavaThread::tlab_slow_allocations_offset())), 1); 5542 } 5543 jmp(try_eden); 5544 5545 bind(discard_tlab); 5546 if (TLABStats) { 5547 // increment number of refills 5548 addl(Address(thread_reg, in_bytes(JavaThread::tlab_number_of_refills_offset())), 1); 5549 // accumulate wastage -- t1 is amount free in tlab 5550 addl(Address(thread_reg, in_bytes(JavaThread::tlab_fast_refill_waste_offset())), t1); 5551 } 5552 5553 // if tlab is currently allocated (top or end != null) then 5554 // fill [top, end + alignment_reserve) with array object 5555 testptr(top, top); 5556 jcc(Assembler::zero, do_refill); 5557 5558 // set up the mark word 5559 movptr(Address(top, oopDesc::mark_offset_in_bytes()), (intptr_t)markOopDesc::prototype()->copy_set_hash(0x2)); 5560 // set the length to the remaining space 5561 subptr(t1, typeArrayOopDesc::header_size(T_INT)); 5562 addptr(t1, (int32_t)ThreadLocalAllocBuffer::alignment_reserve()); 5563 shlptr(t1, log2_intptr(HeapWordSize/sizeof(jint))); 5564 movl(Address(top, arrayOopDesc::length_offset_in_bytes()), t1); 5565 // set klass to intArrayKlass 5566 // dubious reloc why not an oop reloc? 5567 movptr(t1, ExternalAddress((address)Universe::intArrayKlassObj_addr())); 5568 // store klass last. concurrent gcs assumes klass length is valid if 5569 // klass field is not null. 5570 store_klass(top, t1); 5571 5572 movptr(t1, top); 5573 subptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset()))); 5574 incr_allocated_bytes(thread_reg, t1, 0); 5575 5576 // refill the tlab with an eden allocation 5577 bind(do_refill); 5578 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset()))); 5579 shlptr(t1, LogHeapWordSize); 5580 // allocate new tlab, address returned in top 5581 eden_allocate(top, t1, 0, t2, slow_case); 5582 5583 // Check that t1 was preserved in eden_allocate. 5584 #ifdef ASSERT 5585 if (UseTLAB) { 5586 Label ok; 5587 Register tsize = rsi; 5588 assert_different_registers(tsize, thread_reg, t1); 5589 push(tsize); 5590 movptr(tsize, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset()))); 5591 shlptr(tsize, LogHeapWordSize); 5592 cmpptr(t1, tsize); 5593 jcc(Assembler::equal, ok); 5594 STOP("assert(t1 != tlab size)"); 5595 should_not_reach_here(); 5596 5597 bind(ok); 5598 pop(tsize); 5599 } 5600 #endif 5601 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())), top); 5602 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())), top); 5603 addptr(top, t1); 5604 subptr(top, (int32_t)ThreadLocalAllocBuffer::alignment_reserve_in_bytes()); 5605 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())), top); 5606 5607 if (ZeroTLAB) { 5608 // This is a fast TLAB refill, therefore the GC is not notified of it. 5609 // So compiled code must fill the new TLAB with zeroes. 5610 movptr(top, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset()))); 5611 zero_memory(top, t1, 0, t2); 5612 } 5613 5614 verify_tlab(); 5615 jmp(retry); 5616 5617 return thread_reg; // for use by caller 5618 } 5619 5620 // Preserves the contents of address, destroys the contents length_in_bytes and temp. 5621 void MacroAssembler::zero_memory(Register address, Register length_in_bytes, int offset_in_bytes, Register temp) { 5622 assert(address != length_in_bytes && address != temp && temp != length_in_bytes, "registers must be different"); 5623 assert((offset_in_bytes & (BytesPerWord - 1)) == 0, "offset must be a multiple of BytesPerWord"); 5624 Label done; 5625 5626 testptr(length_in_bytes, length_in_bytes); 5627 jcc(Assembler::zero, done); 5628 5629 // initialize topmost word, divide index by 2, check if odd and test if zero 5630 // note: for the remaining code to work, index must be a multiple of BytesPerWord 5631 #ifdef ASSERT 5632 { 5633 Label L; 5634 testptr(length_in_bytes, BytesPerWord - 1); 5635 jcc(Assembler::zero, L); 5636 stop("length must be a multiple of BytesPerWord"); 5637 bind(L); 5638 } 5639 #endif 5640 Register index = length_in_bytes; 5641 xorptr(temp, temp); // use _zero reg to clear memory (shorter code) 5642 if (UseIncDec) { 5643 shrptr(index, 3); // divide by 8/16 and set carry flag if bit 2 was set 5644 } else { 5645 shrptr(index, 2); // use 2 instructions to avoid partial flag stall 5646 shrptr(index, 1); 5647 } 5648 #ifndef _LP64 5649 // index could have not been a multiple of 8 (i.e., bit 2 was set) 5650 { 5651 Label even; 5652 // note: if index was a multiple of 8, then it cannot 5653 // be 0 now otherwise it must have been 0 before 5654 // => if it is even, we don't need to check for 0 again 5655 jcc(Assembler::carryClear, even); 5656 // clear topmost word (no jump would be needed if conditional assignment worked here) 5657 movptr(Address(address, index, Address::times_8, offset_in_bytes - 0*BytesPerWord), temp); 5658 // index could be 0 now, must check again 5659 jcc(Assembler::zero, done); 5660 bind(even); 5661 } 5662 #endif // !_LP64 5663 // initialize remaining object fields: index is a multiple of 2 now 5664 { 5665 Label loop; 5666 bind(loop); 5667 movptr(Address(address, index, Address::times_8, offset_in_bytes - 1*BytesPerWord), temp); 5668 NOT_LP64(movptr(Address(address, index, Address::times_8, offset_in_bytes - 2*BytesPerWord), temp);) 5669 decrement(index); 5670 jcc(Assembler::notZero, loop); 5671 } 5672 5673 bind(done); 5674 } 5675 5676 void MacroAssembler::incr_allocated_bytes(Register thread, 5677 Register var_size_in_bytes, 5678 int con_size_in_bytes, 5679 Register t1) { 5680 if (!thread->is_valid()) { 5681 #ifdef _LP64 5682 thread = r15_thread; 5683 #else 5684 assert(t1->is_valid(), "need temp reg"); 5685 thread = t1; 5686 get_thread(thread); 5687 #endif 5688 } 5689 5690 #ifdef _LP64 5691 if (var_size_in_bytes->is_valid()) { 5692 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes); 5693 } else { 5694 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes); 5695 } 5696 #else 5697 if (var_size_in_bytes->is_valid()) { 5698 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes); 5699 } else { 5700 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes); 5701 } 5702 adcl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())+4), 0); 5703 #endif 5704 } 5705 5706 // Look up the method for a megamorphic invokeinterface call. 5707 // The target method is determined by <intf_klass, itable_index>. 5708 // The receiver klass is in recv_klass. 5709 // On success, the result will be in method_result, and execution falls through. 5710 // On failure, execution transfers to the given label. 5711 void MacroAssembler::lookup_interface_method(Register recv_klass, 5712 Register intf_klass, 5713 RegisterOrConstant itable_index, 5714 Register method_result, 5715 Register scan_temp, 5716 Label& L_no_such_interface) { 5717 assert_different_registers(recv_klass, intf_klass, method_result, scan_temp); 5718 assert(itable_index.is_constant() || itable_index.as_register() == method_result, 5719 "caller must use same register for non-constant itable index as for method"); 5720 5721 // Compute start of first itableOffsetEntry (which is at the end of the vtable) 5722 int vtable_base = in_bytes(Klass::vtable_start_offset()); 5723 int itentry_off = itableMethodEntry::method_offset_in_bytes(); 5724 int scan_step = itableOffsetEntry::size() * wordSize; 5725 int vte_size = vtableEntry::size_in_bytes(); 5726 Address::ScaleFactor times_vte_scale = Address::times_ptr; 5727 assert(vte_size == wordSize, "else adjust times_vte_scale"); 5728 5729 movl(scan_temp, Address(recv_klass, Klass::vtable_length_offset())); 5730 5731 // %%% Could store the aligned, prescaled offset in the klassoop. 5732 lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base)); 5733 5734 // Adjust recv_klass by scaled itable_index, so we can free itable_index. 5735 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below"); 5736 lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off)); 5737 5738 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) { 5739 // if (scan->interface() == intf) { 5740 // result = (klass + scan->offset() + itable_index); 5741 // } 5742 // } 5743 Label search, found_method; 5744 5745 for (int peel = 1; peel >= 0; peel--) { 5746 movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes())); 5747 cmpptr(intf_klass, method_result); 5748 5749 if (peel) { 5750 jccb(Assembler::equal, found_method); 5751 } else { 5752 jccb(Assembler::notEqual, search); 5753 // (invert the test to fall through to found_method...) 5754 } 5755 5756 if (!peel) break; 5757 5758 bind(search); 5759 5760 // Check that the previous entry is non-null. A null entry means that 5761 // the receiver class doesn't implement the interface, and wasn't the 5762 // same as when the caller was compiled. 5763 testptr(method_result, method_result); 5764 jcc(Assembler::zero, L_no_such_interface); 5765 addptr(scan_temp, scan_step); 5766 } 5767 5768 bind(found_method); 5769 5770 // Got a hit. 5771 movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes())); 5772 movptr(method_result, Address(recv_klass, scan_temp, Address::times_1)); 5773 } 5774 5775 5776 // virtual method calling 5777 void MacroAssembler::lookup_virtual_method(Register recv_klass, 5778 RegisterOrConstant vtable_index, 5779 Register method_result) { 5780 const int base = in_bytes(Klass::vtable_start_offset()); 5781 assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below"); 5782 Address vtable_entry_addr(recv_klass, 5783 vtable_index, Address::times_ptr, 5784 base + vtableEntry::method_offset_in_bytes()); 5785 movptr(method_result, vtable_entry_addr); 5786 } 5787 5788 5789 void MacroAssembler::check_klass_subtype(Register sub_klass, 5790 Register super_klass, 5791 Register temp_reg, 5792 Label& L_success) { 5793 Label L_failure; 5794 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL); 5795 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL); 5796 bind(L_failure); 5797 } 5798 5799 5800 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass, 5801 Register super_klass, 5802 Register temp_reg, 5803 Label* L_success, 5804 Label* L_failure, 5805 Label* L_slow_path, 5806 RegisterOrConstant super_check_offset) { 5807 assert_different_registers(sub_klass, super_klass, temp_reg); 5808 bool must_load_sco = (super_check_offset.constant_or_zero() == -1); 5809 if (super_check_offset.is_register()) { 5810 assert_different_registers(sub_klass, super_klass, 5811 super_check_offset.as_register()); 5812 } else if (must_load_sco) { 5813 assert(temp_reg != noreg, "supply either a temp or a register offset"); 5814 } 5815 5816 Label L_fallthrough; 5817 int label_nulls = 0; 5818 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 5819 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 5820 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; } 5821 assert(label_nulls <= 1, "at most one NULL in the batch"); 5822 5823 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 5824 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 5825 Address super_check_offset_addr(super_klass, sco_offset); 5826 5827 // Hacked jcc, which "knows" that L_fallthrough, at least, is in 5828 // range of a jccb. If this routine grows larger, reconsider at 5829 // least some of these. 5830 #define local_jcc(assembler_cond, label) \ 5831 if (&(label) == &L_fallthrough) jccb(assembler_cond, label); \ 5832 else jcc( assembler_cond, label) /*omit semi*/ 5833 5834 // Hacked jmp, which may only be used just before L_fallthrough. 5835 #define final_jmp(label) \ 5836 if (&(label) == &L_fallthrough) { /*do nothing*/ } \ 5837 else jmp(label) /*omit semi*/ 5838 5839 // If the pointers are equal, we are done (e.g., String[] elements). 5840 // This self-check enables sharing of secondary supertype arrays among 5841 // non-primary types such as array-of-interface. Otherwise, each such 5842 // type would need its own customized SSA. 5843 // We move this check to the front of the fast path because many 5844 // type checks are in fact trivially successful in this manner, 5845 // so we get a nicely predicted branch right at the start of the check. 5846 cmpptr(sub_klass, super_klass); 5847 local_jcc(Assembler::equal, *L_success); 5848 5849 // Check the supertype display: 5850 if (must_load_sco) { 5851 // Positive movl does right thing on LP64. 5852 movl(temp_reg, super_check_offset_addr); 5853 super_check_offset = RegisterOrConstant(temp_reg); 5854 } 5855 Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0); 5856 cmpptr(super_klass, super_check_addr); // load displayed supertype 5857 5858 // This check has worked decisively for primary supers. 5859 // Secondary supers are sought in the super_cache ('super_cache_addr'). 5860 // (Secondary supers are interfaces and very deeply nested subtypes.) 5861 // This works in the same check above because of a tricky aliasing 5862 // between the super_cache and the primary super display elements. 5863 // (The 'super_check_addr' can address either, as the case requires.) 5864 // Note that the cache is updated below if it does not help us find 5865 // what we need immediately. 5866 // So if it was a primary super, we can just fail immediately. 5867 // Otherwise, it's the slow path for us (no success at this point). 5868 5869 if (super_check_offset.is_register()) { 5870 local_jcc(Assembler::equal, *L_success); 5871 cmpl(super_check_offset.as_register(), sc_offset); 5872 if (L_failure == &L_fallthrough) { 5873 local_jcc(Assembler::equal, *L_slow_path); 5874 } else { 5875 local_jcc(Assembler::notEqual, *L_failure); 5876 final_jmp(*L_slow_path); 5877 } 5878 } else if (super_check_offset.as_constant() == sc_offset) { 5879 // Need a slow path; fast failure is impossible. 5880 if (L_slow_path == &L_fallthrough) { 5881 local_jcc(Assembler::equal, *L_success); 5882 } else { 5883 local_jcc(Assembler::notEqual, *L_slow_path); 5884 final_jmp(*L_success); 5885 } 5886 } else { 5887 // No slow path; it's a fast decision. 5888 if (L_failure == &L_fallthrough) { 5889 local_jcc(Assembler::equal, *L_success); 5890 } else { 5891 local_jcc(Assembler::notEqual, *L_failure); 5892 final_jmp(*L_success); 5893 } 5894 } 5895 5896 bind(L_fallthrough); 5897 5898 #undef local_jcc 5899 #undef final_jmp 5900 } 5901 5902 5903 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass, 5904 Register super_klass, 5905 Register temp_reg, 5906 Register temp2_reg, 5907 Label* L_success, 5908 Label* L_failure, 5909 bool set_cond_codes) { 5910 assert_different_registers(sub_klass, super_klass, temp_reg); 5911 if (temp2_reg != noreg) 5912 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg); 5913 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg) 5914 5915 Label L_fallthrough; 5916 int label_nulls = 0; 5917 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 5918 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 5919 assert(label_nulls <= 1, "at most one NULL in the batch"); 5920 5921 // a couple of useful fields in sub_klass: 5922 int ss_offset = in_bytes(Klass::secondary_supers_offset()); 5923 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 5924 Address secondary_supers_addr(sub_klass, ss_offset); 5925 Address super_cache_addr( sub_klass, sc_offset); 5926 5927 // Do a linear scan of the secondary super-klass chain. 5928 // This code is rarely used, so simplicity is a virtue here. 5929 // The repne_scan instruction uses fixed registers, which we must spill. 5930 // Don't worry too much about pre-existing connections with the input regs. 5931 5932 assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super) 5933 assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter) 5934 5935 // Get super_klass value into rax (even if it was in rdi or rcx). 5936 bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false; 5937 if (super_klass != rax || UseCompressedOops) { 5938 if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; } 5939 mov(rax, super_klass); 5940 } 5941 if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; } 5942 if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; } 5943 5944 #ifndef PRODUCT 5945 int* pst_counter = &SharedRuntime::_partial_subtype_ctr; 5946 ExternalAddress pst_counter_addr((address) pst_counter); 5947 NOT_LP64( incrementl(pst_counter_addr) ); 5948 LP64_ONLY( lea(rcx, pst_counter_addr) ); 5949 LP64_ONLY( incrementl(Address(rcx, 0)) ); 5950 #endif //PRODUCT 5951 5952 // We will consult the secondary-super array. 5953 movptr(rdi, secondary_supers_addr); 5954 // Load the array length. (Positive movl does right thing on LP64.) 5955 movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes())); 5956 // Skip to start of data. 5957 addptr(rdi, Array<Klass*>::base_offset_in_bytes()); 5958 5959 // Scan RCX words at [RDI] for an occurrence of RAX. 5960 // Set NZ/Z based on last compare. 5961 // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does 5962 // not change flags (only scas instruction which is repeated sets flags). 5963 // Set Z = 0 (not equal) before 'repne' to indicate that class was not found. 5964 5965 testptr(rax,rax); // Set Z = 0 5966 repne_scan(); 5967 5968 // Unspill the temp. registers: 5969 if (pushed_rdi) pop(rdi); 5970 if (pushed_rcx) pop(rcx); 5971 if (pushed_rax) pop(rax); 5972 5973 if (set_cond_codes) { 5974 // Special hack for the AD files: rdi is guaranteed non-zero. 5975 assert(!pushed_rdi, "rdi must be left non-NULL"); 5976 // Also, the condition codes are properly set Z/NZ on succeed/failure. 5977 } 5978 5979 if (L_failure == &L_fallthrough) 5980 jccb(Assembler::notEqual, *L_failure); 5981 else jcc(Assembler::notEqual, *L_failure); 5982 5983 // Success. Cache the super we found and proceed in triumph. 5984 movptr(super_cache_addr, super_klass); 5985 5986 if (L_success != &L_fallthrough) { 5987 jmp(*L_success); 5988 } 5989 5990 #undef IS_A_TEMP 5991 5992 bind(L_fallthrough); 5993 } 5994 5995 5996 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) { 5997 if (VM_Version::supports_cmov()) { 5998 cmovl(cc, dst, src); 5999 } else { 6000 Label L; 6001 jccb(negate_condition(cc), L); 6002 movl(dst, src); 6003 bind(L); 6004 } 6005 } 6006 6007 void MacroAssembler::cmov32(Condition cc, Register dst, Register src) { 6008 if (VM_Version::supports_cmov()) { 6009 cmovl(cc, dst, src); 6010 } else { 6011 Label L; 6012 jccb(negate_condition(cc), L); 6013 movl(dst, src); 6014 bind(L); 6015 } 6016 } 6017 6018 void MacroAssembler::verify_oop(Register reg, const char* s) { 6019 if (!VerifyOops || VerifyAdapterSharing) { 6020 // Below address of the code string confuses VerifyAdapterSharing 6021 // because it may differ between otherwise equivalent adapters. 6022 return; 6023 } 6024 6025 // Pass register number to verify_oop_subroutine 6026 const char* b = NULL; 6027 { 6028 ResourceMark rm; 6029 stringStream ss; 6030 ss.print("verify_oop: %s: %s", reg->name(), s); 6031 b = code_string(ss.as_string()); 6032 } 6033 BLOCK_COMMENT("verify_oop {"); 6034 #ifdef _LP64 6035 push(rscratch1); // save r10, trashed by movptr() 6036 #endif 6037 push(rax); // save rax, 6038 push(reg); // pass register argument 6039 ExternalAddress buffer((address) b); 6040 // avoid using pushptr, as it modifies scratch registers 6041 // and our contract is not to modify anything 6042 movptr(rax, buffer.addr()); 6043 push(rax); 6044 // call indirectly to solve generation ordering problem 6045 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); 6046 call(rax); 6047 // Caller pops the arguments (oop, message) and restores rax, r10 6048 BLOCK_COMMENT("} verify_oop"); 6049 } 6050 6051 6052 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr, 6053 Register tmp, 6054 int offset) { 6055 intptr_t value = *delayed_value_addr; 6056 if (value != 0) 6057 return RegisterOrConstant(value + offset); 6058 6059 // load indirectly to solve generation ordering problem 6060 movptr(tmp, ExternalAddress((address) delayed_value_addr)); 6061 6062 #ifdef ASSERT 6063 { Label L; 6064 testptr(tmp, tmp); 6065 if (WizardMode) { 6066 const char* buf = NULL; 6067 { 6068 ResourceMark rm; 6069 stringStream ss; 6070 ss.print("DelayedValue=" INTPTR_FORMAT, delayed_value_addr[1]); 6071 buf = code_string(ss.as_string()); 6072 } 6073 jcc(Assembler::notZero, L); 6074 STOP(buf); 6075 } else { 6076 jccb(Assembler::notZero, L); 6077 hlt(); 6078 } 6079 bind(L); 6080 } 6081 #endif 6082 6083 if (offset != 0) 6084 addptr(tmp, offset); 6085 6086 return RegisterOrConstant(tmp); 6087 } 6088 6089 6090 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot, 6091 int extra_slot_offset) { 6092 // cf. TemplateTable::prepare_invoke(), if (load_receiver). 6093 int stackElementSize = Interpreter::stackElementSize; 6094 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0); 6095 #ifdef ASSERT 6096 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1); 6097 assert(offset1 - offset == stackElementSize, "correct arithmetic"); 6098 #endif 6099 Register scale_reg = noreg; 6100 Address::ScaleFactor scale_factor = Address::no_scale; 6101 if (arg_slot.is_constant()) { 6102 offset += arg_slot.as_constant() * stackElementSize; 6103 } else { 6104 scale_reg = arg_slot.as_register(); 6105 scale_factor = Address::times(stackElementSize); 6106 } 6107 offset += wordSize; // return PC is on stack 6108 return Address(rsp, scale_reg, scale_factor, offset); 6109 } 6110 6111 6112 void MacroAssembler::verify_oop_addr(Address addr, const char* s) { 6113 if (!VerifyOops || VerifyAdapterSharing) { 6114 // Below address of the code string confuses VerifyAdapterSharing 6115 // because it may differ between otherwise equivalent adapters. 6116 return; 6117 } 6118 6119 // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord); 6120 // Pass register number to verify_oop_subroutine 6121 const char* b = NULL; 6122 { 6123 ResourceMark rm; 6124 stringStream ss; 6125 ss.print("verify_oop_addr: %s", s); 6126 b = code_string(ss.as_string()); 6127 } 6128 #ifdef _LP64 6129 push(rscratch1); // save r10, trashed by movptr() 6130 #endif 6131 push(rax); // save rax, 6132 // addr may contain rsp so we will have to adjust it based on the push 6133 // we just did (and on 64 bit we do two pushes) 6134 // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which 6135 // stores rax into addr which is backwards of what was intended. 6136 if (addr.uses(rsp)) { 6137 lea(rax, addr); 6138 pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord)); 6139 } else { 6140 pushptr(addr); 6141 } 6142 6143 ExternalAddress buffer((address) b); 6144 // pass msg argument 6145 // avoid using pushptr, as it modifies scratch registers 6146 // and our contract is not to modify anything 6147 movptr(rax, buffer.addr()); 6148 push(rax); 6149 6150 // call indirectly to solve generation ordering problem 6151 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); 6152 call(rax); 6153 // Caller pops the arguments (addr, message) and restores rax, r10. 6154 } 6155 6156 void MacroAssembler::verify_tlab() { 6157 #ifdef ASSERT 6158 if (UseTLAB && VerifyOops) { 6159 Label next, ok; 6160 Register t1 = rsi; 6161 Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread); 6162 6163 push(t1); 6164 NOT_LP64(push(thread_reg)); 6165 NOT_LP64(get_thread(thread_reg)); 6166 6167 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 6168 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset()))); 6169 jcc(Assembler::aboveEqual, next); 6170 STOP("assert(top >= start)"); 6171 should_not_reach_here(); 6172 6173 bind(next); 6174 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset()))); 6175 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 6176 jcc(Assembler::aboveEqual, ok); 6177 STOP("assert(top <= end)"); 6178 should_not_reach_here(); 6179 6180 bind(ok); 6181 NOT_LP64(pop(thread_reg)); 6182 pop(t1); 6183 } 6184 #endif 6185 } 6186 6187 class ControlWord { 6188 public: 6189 int32_t _value; 6190 6191 int rounding_control() const { return (_value >> 10) & 3 ; } 6192 int precision_control() const { return (_value >> 8) & 3 ; } 6193 bool precision() const { return ((_value >> 5) & 1) != 0; } 6194 bool underflow() const { return ((_value >> 4) & 1) != 0; } 6195 bool overflow() const { return ((_value >> 3) & 1) != 0; } 6196 bool zero_divide() const { return ((_value >> 2) & 1) != 0; } 6197 bool denormalized() const { return ((_value >> 1) & 1) != 0; } 6198 bool invalid() const { return ((_value >> 0) & 1) != 0; } 6199 6200 void print() const { 6201 // rounding control 6202 const char* rc; 6203 switch (rounding_control()) { 6204 case 0: rc = "round near"; break; 6205 case 1: rc = "round down"; break; 6206 case 2: rc = "round up "; break; 6207 case 3: rc = "chop "; break; 6208 }; 6209 // precision control 6210 const char* pc; 6211 switch (precision_control()) { 6212 case 0: pc = "24 bits "; break; 6213 case 1: pc = "reserved"; break; 6214 case 2: pc = "53 bits "; break; 6215 case 3: pc = "64 bits "; break; 6216 }; 6217 // flags 6218 char f[9]; 6219 f[0] = ' '; 6220 f[1] = ' '; 6221 f[2] = (precision ()) ? 'P' : 'p'; 6222 f[3] = (underflow ()) ? 'U' : 'u'; 6223 f[4] = (overflow ()) ? 'O' : 'o'; 6224 f[5] = (zero_divide ()) ? 'Z' : 'z'; 6225 f[6] = (denormalized()) ? 'D' : 'd'; 6226 f[7] = (invalid ()) ? 'I' : 'i'; 6227 f[8] = '\x0'; 6228 // output 6229 printf("%04x masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc); 6230 } 6231 6232 }; 6233 6234 class StatusWord { 6235 public: 6236 int32_t _value; 6237 6238 bool busy() const { return ((_value >> 15) & 1) != 0; } 6239 bool C3() const { return ((_value >> 14) & 1) != 0; } 6240 bool C2() const { return ((_value >> 10) & 1) != 0; } 6241 bool C1() const { return ((_value >> 9) & 1) != 0; } 6242 bool C0() const { return ((_value >> 8) & 1) != 0; } 6243 int top() const { return (_value >> 11) & 7 ; } 6244 bool error_status() const { return ((_value >> 7) & 1) != 0; } 6245 bool stack_fault() const { return ((_value >> 6) & 1) != 0; } 6246 bool precision() const { return ((_value >> 5) & 1) != 0; } 6247 bool underflow() const { return ((_value >> 4) & 1) != 0; } 6248 bool overflow() const { return ((_value >> 3) & 1) != 0; } 6249 bool zero_divide() const { return ((_value >> 2) & 1) != 0; } 6250 bool denormalized() const { return ((_value >> 1) & 1) != 0; } 6251 bool invalid() const { return ((_value >> 0) & 1) != 0; } 6252 6253 void print() const { 6254 // condition codes 6255 char c[5]; 6256 c[0] = (C3()) ? '3' : '-'; 6257 c[1] = (C2()) ? '2' : '-'; 6258 c[2] = (C1()) ? '1' : '-'; 6259 c[3] = (C0()) ? '0' : '-'; 6260 c[4] = '\x0'; 6261 // flags 6262 char f[9]; 6263 f[0] = (error_status()) ? 'E' : '-'; 6264 f[1] = (stack_fault ()) ? 'S' : '-'; 6265 f[2] = (precision ()) ? 'P' : '-'; 6266 f[3] = (underflow ()) ? 'U' : '-'; 6267 f[4] = (overflow ()) ? 'O' : '-'; 6268 f[5] = (zero_divide ()) ? 'Z' : '-'; 6269 f[6] = (denormalized()) ? 'D' : '-'; 6270 f[7] = (invalid ()) ? 'I' : '-'; 6271 f[8] = '\x0'; 6272 // output 6273 printf("%04x flags = %s, cc = %s, top = %d", _value & 0xFFFF, f, c, top()); 6274 } 6275 6276 }; 6277 6278 class TagWord { 6279 public: 6280 int32_t _value; 6281 6282 int tag_at(int i) const { return (_value >> (i*2)) & 3; } 6283 6284 void print() const { 6285 printf("%04x", _value & 0xFFFF); 6286 } 6287 6288 }; 6289 6290 class FPU_Register { 6291 public: 6292 int32_t _m0; 6293 int32_t _m1; 6294 int16_t _ex; 6295 6296 bool is_indefinite() const { 6297 return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0; 6298 } 6299 6300 void print() const { 6301 char sign = (_ex < 0) ? '-' : '+'; 6302 const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : " "; 6303 printf("%c%04hx.%08x%08x %s", sign, _ex, _m1, _m0, kind); 6304 }; 6305 6306 }; 6307 6308 class FPU_State { 6309 public: 6310 enum { 6311 register_size = 10, 6312 number_of_registers = 8, 6313 register_mask = 7 6314 }; 6315 6316 ControlWord _control_word; 6317 StatusWord _status_word; 6318 TagWord _tag_word; 6319 int32_t _error_offset; 6320 int32_t _error_selector; 6321 int32_t _data_offset; 6322 int32_t _data_selector; 6323 int8_t _register[register_size * number_of_registers]; 6324 6325 int tag_for_st(int i) const { return _tag_word.tag_at((_status_word.top() + i) & register_mask); } 6326 FPU_Register* st(int i) const { return (FPU_Register*)&_register[register_size * i]; } 6327 6328 const char* tag_as_string(int tag) const { 6329 switch (tag) { 6330 case 0: return "valid"; 6331 case 1: return "zero"; 6332 case 2: return "special"; 6333 case 3: return "empty"; 6334 } 6335 ShouldNotReachHere(); 6336 return NULL; 6337 } 6338 6339 void print() const { 6340 // print computation registers 6341 { int t = _status_word.top(); 6342 for (int i = 0; i < number_of_registers; i++) { 6343 int j = (i - t) & register_mask; 6344 printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j); 6345 st(j)->print(); 6346 printf(" %s\n", tag_as_string(_tag_word.tag_at(i))); 6347 } 6348 } 6349 printf("\n"); 6350 // print control registers 6351 printf("ctrl = "); _control_word.print(); printf("\n"); 6352 printf("stat = "); _status_word .print(); printf("\n"); 6353 printf("tags = "); _tag_word .print(); printf("\n"); 6354 } 6355 6356 }; 6357 6358 class Flag_Register { 6359 public: 6360 int32_t _value; 6361 6362 bool overflow() const { return ((_value >> 11) & 1) != 0; } 6363 bool direction() const { return ((_value >> 10) & 1) != 0; } 6364 bool sign() const { return ((_value >> 7) & 1) != 0; } 6365 bool zero() const { return ((_value >> 6) & 1) != 0; } 6366 bool auxiliary_carry() const { return ((_value >> 4) & 1) != 0; } 6367 bool parity() const { return ((_value >> 2) & 1) != 0; } 6368 bool carry() const { return ((_value >> 0) & 1) != 0; } 6369 6370 void print() const { 6371 // flags 6372 char f[8]; 6373 f[0] = (overflow ()) ? 'O' : '-'; 6374 f[1] = (direction ()) ? 'D' : '-'; 6375 f[2] = (sign ()) ? 'S' : '-'; 6376 f[3] = (zero ()) ? 'Z' : '-'; 6377 f[4] = (auxiliary_carry()) ? 'A' : '-'; 6378 f[5] = (parity ()) ? 'P' : '-'; 6379 f[6] = (carry ()) ? 'C' : '-'; 6380 f[7] = '\x0'; 6381 // output 6382 printf("%08x flags = %s", _value, f); 6383 } 6384 6385 }; 6386 6387 class IU_Register { 6388 public: 6389 int32_t _value; 6390 6391 void print() const { 6392 printf("%08x %11d", _value, _value); 6393 } 6394 6395 }; 6396 6397 class IU_State { 6398 public: 6399 Flag_Register _eflags; 6400 IU_Register _rdi; 6401 IU_Register _rsi; 6402 IU_Register _rbp; 6403 IU_Register _rsp; 6404 IU_Register _rbx; 6405 IU_Register _rdx; 6406 IU_Register _rcx; 6407 IU_Register _rax; 6408 6409 void print() const { 6410 // computation registers 6411 printf("rax, = "); _rax.print(); printf("\n"); 6412 printf("rbx, = "); _rbx.print(); printf("\n"); 6413 printf("rcx = "); _rcx.print(); printf("\n"); 6414 printf("rdx = "); _rdx.print(); printf("\n"); 6415 printf("rdi = "); _rdi.print(); printf("\n"); 6416 printf("rsi = "); _rsi.print(); printf("\n"); 6417 printf("rbp, = "); _rbp.print(); printf("\n"); 6418 printf("rsp = "); _rsp.print(); printf("\n"); 6419 printf("\n"); 6420 // control registers 6421 printf("flgs = "); _eflags.print(); printf("\n"); 6422 } 6423 }; 6424 6425 6426 class CPU_State { 6427 public: 6428 FPU_State _fpu_state; 6429 IU_State _iu_state; 6430 6431 void print() const { 6432 printf("--------------------------------------------------\n"); 6433 _iu_state .print(); 6434 printf("\n"); 6435 _fpu_state.print(); 6436 printf("--------------------------------------------------\n"); 6437 } 6438 6439 }; 6440 6441 6442 static void _print_CPU_state(CPU_State* state) { 6443 state->print(); 6444 }; 6445 6446 6447 void MacroAssembler::print_CPU_state() { 6448 push_CPU_state(); 6449 push(rsp); // pass CPU state 6450 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state))); 6451 addptr(rsp, wordSize); // discard argument 6452 pop_CPU_state(); 6453 } 6454 6455 6456 static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) { 6457 static int counter = 0; 6458 FPU_State* fs = &state->_fpu_state; 6459 counter++; 6460 // For leaf calls, only verify that the top few elements remain empty. 6461 // We only need 1 empty at the top for C2 code. 6462 if( stack_depth < 0 ) { 6463 if( fs->tag_for_st(7) != 3 ) { 6464 printf("FPR7 not empty\n"); 6465 state->print(); 6466 assert(false, "error"); 6467 return false; 6468 } 6469 return true; // All other stack states do not matter 6470 } 6471 6472 assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std, 6473 "bad FPU control word"); 6474 6475 // compute stack depth 6476 int i = 0; 6477 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) < 3) i++; 6478 int d = i; 6479 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++; 6480 // verify findings 6481 if (i != FPU_State::number_of_registers) { 6482 // stack not contiguous 6483 printf("%s: stack not contiguous at ST%d\n", s, i); 6484 state->print(); 6485 assert(false, "error"); 6486 return false; 6487 } 6488 // check if computed stack depth corresponds to expected stack depth 6489 if (stack_depth < 0) { 6490 // expected stack depth is -stack_depth or less 6491 if (d > -stack_depth) { 6492 // too many elements on the stack 6493 printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d); 6494 state->print(); 6495 assert(false, "error"); 6496 return false; 6497 } 6498 } else { 6499 // expected stack depth is stack_depth 6500 if (d != stack_depth) { 6501 // wrong stack depth 6502 printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d); 6503 state->print(); 6504 assert(false, "error"); 6505 return false; 6506 } 6507 } 6508 // everything is cool 6509 return true; 6510 } 6511 6512 6513 void MacroAssembler::verify_FPU(int stack_depth, const char* s) { 6514 if (!VerifyFPU) return; 6515 push_CPU_state(); 6516 push(rsp); // pass CPU state 6517 ExternalAddress msg((address) s); 6518 // pass message string s 6519 pushptr(msg.addr()); 6520 push(stack_depth); // pass stack depth 6521 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU))); 6522 addptr(rsp, 3 * wordSize); // discard arguments 6523 // check for error 6524 { Label L; 6525 testl(rax, rax); 6526 jcc(Assembler::notZero, L); 6527 int3(); // break if error condition 6528 bind(L); 6529 } 6530 pop_CPU_state(); 6531 } 6532 6533 void MacroAssembler::restore_cpu_control_state_after_jni() { 6534 // Either restore the MXCSR register after returning from the JNI Call 6535 // or verify that it wasn't changed (with -Xcheck:jni flag). 6536 if (VM_Version::supports_sse()) { 6537 if (RestoreMXCSROnJNICalls) { 6538 ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std())); 6539 } else if (CheckJNICalls) { 6540 call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry())); 6541 } 6542 } 6543 // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty. 6544 vzeroupper(); 6545 6546 #ifndef _LP64 6547 // Either restore the x87 floating pointer control word after returning 6548 // from the JNI call or verify that it wasn't changed. 6549 if (CheckJNICalls) { 6550 call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry())); 6551 } 6552 #endif // _LP64 6553 } 6554 6555 void MacroAssembler::load_mirror(Register mirror, Register method) { 6556 // get mirror 6557 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 6558 movptr(mirror, Address(method, Method::const_offset())); 6559 movptr(mirror, Address(mirror, ConstMethod::constants_offset())); 6560 movptr(mirror, Address(mirror, ConstantPool::pool_holder_offset_in_bytes())); 6561 movptr(mirror, Address(mirror, mirror_offset)); 6562 } 6563 6564 void MacroAssembler::load_klass(Register dst, Register src) { 6565 #ifdef _LP64 6566 if (UseCompressedClassPointers) { 6567 movl(dst, Address(src, oopDesc::klass_offset_in_bytes())); 6568 decode_klass_not_null(dst); 6569 } else 6570 #endif 6571 movptr(dst, Address(src, oopDesc::klass_offset_in_bytes())); 6572 } 6573 6574 void MacroAssembler::load_prototype_header(Register dst, Register src) { 6575 load_klass(dst, src); 6576 movptr(dst, Address(dst, Klass::prototype_header_offset())); 6577 } 6578 6579 void MacroAssembler::store_klass(Register dst, Register src) { 6580 #ifdef _LP64 6581 if (UseCompressedClassPointers) { 6582 encode_klass_not_null(src); 6583 movl(Address(dst, oopDesc::klass_offset_in_bytes()), src); 6584 } else 6585 #endif 6586 movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src); 6587 } 6588 6589 void MacroAssembler::load_heap_oop(Register dst, Address src) { 6590 #ifdef _LP64 6591 // FIXME: Must change all places where we try to load the klass. 6592 if (UseCompressedOops) { 6593 movl(dst, src); 6594 decode_heap_oop(dst); 6595 } else 6596 #endif 6597 movptr(dst, src); 6598 } 6599 6600 // Doesn't do verfication, generates fixed size code 6601 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src) { 6602 #ifdef _LP64 6603 if (UseCompressedOops) { 6604 movl(dst, src); 6605 decode_heap_oop_not_null(dst); 6606 } else 6607 #endif 6608 movptr(dst, src); 6609 } 6610 6611 void MacroAssembler::store_heap_oop(Address dst, Register src) { 6612 #ifdef _LP64 6613 if (UseCompressedOops) { 6614 assert(!dst.uses(src), "not enough registers"); 6615 encode_heap_oop(src); 6616 movl(dst, src); 6617 } else 6618 #endif 6619 movptr(dst, src); 6620 } 6621 6622 void MacroAssembler::cmp_heap_oop(Register src1, Address src2, Register tmp) { 6623 assert_different_registers(src1, tmp); 6624 #ifdef _LP64 6625 if (UseCompressedOops) { 6626 bool did_push = false; 6627 if (tmp == noreg) { 6628 tmp = rax; 6629 push(tmp); 6630 did_push = true; 6631 assert(!src2.uses(rsp), "can't push"); 6632 } 6633 load_heap_oop(tmp, src2); 6634 cmpptr(src1, tmp); 6635 if (did_push) pop(tmp); 6636 } else 6637 #endif 6638 cmpptr(src1, src2); 6639 } 6640 6641 // Used for storing NULLs. 6642 void MacroAssembler::store_heap_oop_null(Address dst) { 6643 #ifdef _LP64 6644 if (UseCompressedOops) { 6645 movl(dst, (int32_t)NULL_WORD); 6646 } else { 6647 movslq(dst, (int32_t)NULL_WORD); 6648 } 6649 #else 6650 movl(dst, (int32_t)NULL_WORD); 6651 #endif 6652 } 6653 6654 #ifdef _LP64 6655 void MacroAssembler::store_klass_gap(Register dst, Register src) { 6656 if (UseCompressedClassPointers) { 6657 // Store to klass gap in destination 6658 movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src); 6659 } 6660 } 6661 6662 #ifdef ASSERT 6663 void MacroAssembler::verify_heapbase(const char* msg) { 6664 assert (UseCompressedOops, "should be compressed"); 6665 assert (Universe::heap() != NULL, "java heap should be initialized"); 6666 if (CheckCompressedOops) { 6667 Label ok; 6668 push(rscratch1); // cmpptr trashes rscratch1 6669 cmpptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr())); 6670 jcc(Assembler::equal, ok); 6671 STOP(msg); 6672 bind(ok); 6673 pop(rscratch1); 6674 } 6675 } 6676 #endif 6677 6678 // Algorithm must match oop.inline.hpp encode_heap_oop. 6679 void MacroAssembler::encode_heap_oop(Register r) { 6680 #ifdef ASSERT 6681 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?"); 6682 #endif 6683 verify_oop(r, "broken oop in encode_heap_oop"); 6684 if (Universe::narrow_oop_base() == NULL) { 6685 if (Universe::narrow_oop_shift() != 0) { 6686 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6687 shrq(r, LogMinObjAlignmentInBytes); 6688 } 6689 return; 6690 } 6691 testq(r, r); 6692 cmovq(Assembler::equal, r, r12_heapbase); 6693 subq(r, r12_heapbase); 6694 shrq(r, LogMinObjAlignmentInBytes); 6695 } 6696 6697 void MacroAssembler::encode_heap_oop_not_null(Register r) { 6698 #ifdef ASSERT 6699 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?"); 6700 if (CheckCompressedOops) { 6701 Label ok; 6702 testq(r, r); 6703 jcc(Assembler::notEqual, ok); 6704 STOP("null oop passed to encode_heap_oop_not_null"); 6705 bind(ok); 6706 } 6707 #endif 6708 verify_oop(r, "broken oop in encode_heap_oop_not_null"); 6709 if (Universe::narrow_oop_base() != NULL) { 6710 subq(r, r12_heapbase); 6711 } 6712 if (Universe::narrow_oop_shift() != 0) { 6713 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6714 shrq(r, LogMinObjAlignmentInBytes); 6715 } 6716 } 6717 6718 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) { 6719 #ifdef ASSERT 6720 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?"); 6721 if (CheckCompressedOops) { 6722 Label ok; 6723 testq(src, src); 6724 jcc(Assembler::notEqual, ok); 6725 STOP("null oop passed to encode_heap_oop_not_null2"); 6726 bind(ok); 6727 } 6728 #endif 6729 verify_oop(src, "broken oop in encode_heap_oop_not_null2"); 6730 if (dst != src) { 6731 movq(dst, src); 6732 } 6733 if (Universe::narrow_oop_base() != NULL) { 6734 subq(dst, r12_heapbase); 6735 } 6736 if (Universe::narrow_oop_shift() != 0) { 6737 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6738 shrq(dst, LogMinObjAlignmentInBytes); 6739 } 6740 } 6741 6742 void MacroAssembler::decode_heap_oop(Register r) { 6743 #ifdef ASSERT 6744 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?"); 6745 #endif 6746 if (Universe::narrow_oop_base() == NULL) { 6747 if (Universe::narrow_oop_shift() != 0) { 6748 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6749 shlq(r, LogMinObjAlignmentInBytes); 6750 } 6751 } else { 6752 Label done; 6753 shlq(r, LogMinObjAlignmentInBytes); 6754 jccb(Assembler::equal, done); 6755 addq(r, r12_heapbase); 6756 bind(done); 6757 } 6758 verify_oop(r, "broken oop in decode_heap_oop"); 6759 } 6760 6761 void MacroAssembler::decode_heap_oop_not_null(Register r) { 6762 // Note: it will change flags 6763 assert (UseCompressedOops, "should only be used for compressed headers"); 6764 assert (Universe::heap() != NULL, "java heap should be initialized"); 6765 // Cannot assert, unverified entry point counts instructions (see .ad file) 6766 // vtableStubs also counts instructions in pd_code_size_limit. 6767 // Also do not verify_oop as this is called by verify_oop. 6768 if (Universe::narrow_oop_shift() != 0) { 6769 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6770 shlq(r, LogMinObjAlignmentInBytes); 6771 if (Universe::narrow_oop_base() != NULL) { 6772 addq(r, r12_heapbase); 6773 } 6774 } else { 6775 assert (Universe::narrow_oop_base() == NULL, "sanity"); 6776 } 6777 } 6778 6779 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) { 6780 // Note: it will change flags 6781 assert (UseCompressedOops, "should only be used for compressed headers"); 6782 assert (Universe::heap() != NULL, "java heap should be initialized"); 6783 // Cannot assert, unverified entry point counts instructions (see .ad file) 6784 // vtableStubs also counts instructions in pd_code_size_limit. 6785 // Also do not verify_oop as this is called by verify_oop. 6786 if (Universe::narrow_oop_shift() != 0) { 6787 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6788 if (LogMinObjAlignmentInBytes == Address::times_8) { 6789 leaq(dst, Address(r12_heapbase, src, Address::times_8, 0)); 6790 } else { 6791 if (dst != src) { 6792 movq(dst, src); 6793 } 6794 shlq(dst, LogMinObjAlignmentInBytes); 6795 if (Universe::narrow_oop_base() != NULL) { 6796 addq(dst, r12_heapbase); 6797 } 6798 } 6799 } else { 6800 assert (Universe::narrow_oop_base() == NULL, "sanity"); 6801 if (dst != src) { 6802 movq(dst, src); 6803 } 6804 } 6805 } 6806 6807 void MacroAssembler::encode_klass_not_null(Register r) { 6808 if (Universe::narrow_klass_base() != NULL) { 6809 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base. 6810 assert(r != r12_heapbase, "Encoding a klass in r12"); 6811 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base()); 6812 subq(r, r12_heapbase); 6813 } 6814 if (Universe::narrow_klass_shift() != 0) { 6815 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 6816 shrq(r, LogKlassAlignmentInBytes); 6817 } 6818 if (Universe::narrow_klass_base() != NULL) { 6819 reinit_heapbase(); 6820 } 6821 } 6822 6823 void MacroAssembler::encode_klass_not_null(Register dst, Register src) { 6824 if (dst == src) { 6825 encode_klass_not_null(src); 6826 } else { 6827 if (Universe::narrow_klass_base() != NULL) { 6828 mov64(dst, (int64_t)Universe::narrow_klass_base()); 6829 negq(dst); 6830 addq(dst, src); 6831 } else { 6832 movptr(dst, src); 6833 } 6834 if (Universe::narrow_klass_shift() != 0) { 6835 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 6836 shrq(dst, LogKlassAlignmentInBytes); 6837 } 6838 } 6839 } 6840 6841 // Function instr_size_for_decode_klass_not_null() counts the instructions 6842 // generated by decode_klass_not_null(register r) and reinit_heapbase(), 6843 // when (Universe::heap() != NULL). Hence, if the instructions they 6844 // generate change, then this method needs to be updated. 6845 int MacroAssembler::instr_size_for_decode_klass_not_null() { 6846 assert (UseCompressedClassPointers, "only for compressed klass ptrs"); 6847 if (Universe::narrow_klass_base() != NULL) { 6848 // mov64 + addq + shlq? + mov64 (for reinit_heapbase()). 6849 return (Universe::narrow_klass_shift() == 0 ? 20 : 24); 6850 } else { 6851 // longest load decode klass function, mov64, leaq 6852 return 16; 6853 } 6854 } 6855 6856 // !!! If the instructions that get generated here change then function 6857 // instr_size_for_decode_klass_not_null() needs to get updated. 6858 void MacroAssembler::decode_klass_not_null(Register r) { 6859 // Note: it will change flags 6860 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 6861 assert(r != r12_heapbase, "Decoding a klass in r12"); 6862 // Cannot assert, unverified entry point counts instructions (see .ad file) 6863 // vtableStubs also counts instructions in pd_code_size_limit. 6864 // Also do not verify_oop as this is called by verify_oop. 6865 if (Universe::narrow_klass_shift() != 0) { 6866 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 6867 shlq(r, LogKlassAlignmentInBytes); 6868 } 6869 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base. 6870 if (Universe::narrow_klass_base() != NULL) { 6871 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base()); 6872 addq(r, r12_heapbase); 6873 reinit_heapbase(); 6874 } 6875 } 6876 6877 void MacroAssembler::decode_klass_not_null(Register dst, Register src) { 6878 // Note: it will change flags 6879 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 6880 if (dst == src) { 6881 decode_klass_not_null(dst); 6882 } else { 6883 // Cannot assert, unverified entry point counts instructions (see .ad file) 6884 // vtableStubs also counts instructions in pd_code_size_limit. 6885 // Also do not verify_oop as this is called by verify_oop. 6886 mov64(dst, (int64_t)Universe::narrow_klass_base()); 6887 if (Universe::narrow_klass_shift() != 0) { 6888 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 6889 assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?"); 6890 leaq(dst, Address(dst, src, Address::times_8, 0)); 6891 } else { 6892 addq(dst, src); 6893 } 6894 } 6895 } 6896 6897 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) { 6898 assert (UseCompressedOops, "should only be used for compressed headers"); 6899 assert (Universe::heap() != NULL, "java heap should be initialized"); 6900 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 6901 int oop_index = oop_recorder()->find_index(obj); 6902 RelocationHolder rspec = oop_Relocation::spec(oop_index); 6903 mov_narrow_oop(dst, oop_index, rspec); 6904 } 6905 6906 void MacroAssembler::set_narrow_oop(Address dst, jobject obj) { 6907 assert (UseCompressedOops, "should only be used for compressed headers"); 6908 assert (Universe::heap() != NULL, "java heap should be initialized"); 6909 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 6910 int oop_index = oop_recorder()->find_index(obj); 6911 RelocationHolder rspec = oop_Relocation::spec(oop_index); 6912 mov_narrow_oop(dst, oop_index, rspec); 6913 } 6914 6915 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) { 6916 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 6917 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 6918 int klass_index = oop_recorder()->find_index(k); 6919 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 6920 mov_narrow_oop(dst, Klass::encode_klass(k), rspec); 6921 } 6922 6923 void MacroAssembler::set_narrow_klass(Address dst, Klass* k) { 6924 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 6925 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 6926 int klass_index = oop_recorder()->find_index(k); 6927 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 6928 mov_narrow_oop(dst, Klass::encode_klass(k), rspec); 6929 } 6930 6931 void MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) { 6932 assert (UseCompressedOops, "should only be used for compressed headers"); 6933 assert (Universe::heap() != NULL, "java heap should be initialized"); 6934 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 6935 int oop_index = oop_recorder()->find_index(obj); 6936 RelocationHolder rspec = oop_Relocation::spec(oop_index); 6937 Assembler::cmp_narrow_oop(dst, oop_index, rspec); 6938 } 6939 6940 void MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) { 6941 assert (UseCompressedOops, "should only be used for compressed headers"); 6942 assert (Universe::heap() != NULL, "java heap should be initialized"); 6943 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 6944 int oop_index = oop_recorder()->find_index(obj); 6945 RelocationHolder rspec = oop_Relocation::spec(oop_index); 6946 Assembler::cmp_narrow_oop(dst, oop_index, rspec); 6947 } 6948 6949 void MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) { 6950 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 6951 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 6952 int klass_index = oop_recorder()->find_index(k); 6953 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 6954 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec); 6955 } 6956 6957 void MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) { 6958 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 6959 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 6960 int klass_index = oop_recorder()->find_index(k); 6961 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 6962 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec); 6963 } 6964 6965 void MacroAssembler::reinit_heapbase() { 6966 if (UseCompressedOops || UseCompressedClassPointers) { 6967 if (Universe::heap() != NULL) { 6968 if (Universe::narrow_oop_base() == NULL) { 6969 MacroAssembler::xorptr(r12_heapbase, r12_heapbase); 6970 } else { 6971 mov64(r12_heapbase, (int64_t)Universe::narrow_ptrs_base()); 6972 } 6973 } else { 6974 movptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr())); 6975 } 6976 } 6977 } 6978 6979 #endif // _LP64 6980 6981 6982 // C2 compiled method's prolog code. 6983 void MacroAssembler::verified_entry(int framesize, int stack_bang_size, bool fp_mode_24b) { 6984 6985 // WARNING: Initial instruction MUST be 5 bytes or longer so that 6986 // NativeJump::patch_verified_entry will be able to patch out the entry 6987 // code safely. The push to verify stack depth is ok at 5 bytes, 6988 // the frame allocation can be either 3 or 6 bytes. So if we don't do 6989 // stack bang then we must use the 6 byte frame allocation even if 6990 // we have no frame. :-( 6991 assert(stack_bang_size >= framesize || stack_bang_size <= 0, "stack bang size incorrect"); 6992 6993 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned"); 6994 // Remove word for return addr 6995 framesize -= wordSize; 6996 stack_bang_size -= wordSize; 6997 6998 // Calls to C2R adapters often do not accept exceptional returns. 6999 // We require that their callers must bang for them. But be careful, because 7000 // some VM calls (such as call site linkage) can use several kilobytes of 7001 // stack. But the stack safety zone should account for that. 7002 // See bugs 4446381, 4468289, 4497237. 7003 if (stack_bang_size > 0) { 7004 generate_stack_overflow_check(stack_bang_size); 7005 7006 // We always push rbp, so that on return to interpreter rbp, will be 7007 // restored correctly and we can correct the stack. 7008 push(rbp); 7009 // Save caller's stack pointer into RBP if the frame pointer is preserved. 7010 if (PreserveFramePointer) { 7011 mov(rbp, rsp); 7012 } 7013 // Remove word for ebp 7014 framesize -= wordSize; 7015 7016 // Create frame 7017 if (framesize) { 7018 subptr(rsp, framesize); 7019 } 7020 } else { 7021 // Create frame (force generation of a 4 byte immediate value) 7022 subptr_imm32(rsp, framesize); 7023 7024 // Save RBP register now. 7025 framesize -= wordSize; 7026 movptr(Address(rsp, framesize), rbp); 7027 // Save caller's stack pointer into RBP if the frame pointer is preserved. 7028 if (PreserveFramePointer) { 7029 movptr(rbp, rsp); 7030 if (framesize > 0) { 7031 addptr(rbp, framesize); 7032 } 7033 } 7034 } 7035 7036 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth 7037 framesize -= wordSize; 7038 movptr(Address(rsp, framesize), (int32_t)0xbadb100d); 7039 } 7040 7041 #ifndef _LP64 7042 // If method sets FPU control word do it now 7043 if (fp_mode_24b) { 7044 fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24())); 7045 } 7046 if (UseSSE >= 2 && VerifyFPU) { 7047 verify_FPU(0, "FPU stack must be clean on entry"); 7048 } 7049 #endif 7050 7051 #ifdef ASSERT 7052 if (VerifyStackAtCalls) { 7053 Label L; 7054 push(rax); 7055 mov(rax, rsp); 7056 andptr(rax, StackAlignmentInBytes-1); 7057 cmpptr(rax, StackAlignmentInBytes-wordSize); 7058 pop(rax); 7059 jcc(Assembler::equal, L); 7060 STOP("Stack is not properly aligned!"); 7061 bind(L); 7062 } 7063 #endif 7064 7065 } 7066 7067 void MacroAssembler::clear_mem(Register base, Register cnt, Register tmp, bool is_large) { 7068 // cnt - number of qwords (8-byte words). 7069 // base - start address, qword aligned. 7070 // is_large - if optimizers know cnt is larger than InitArrayShortSize 7071 assert(base==rdi, "base register must be edi for rep stos"); 7072 assert(tmp==rax, "tmp register must be eax for rep stos"); 7073 assert(cnt==rcx, "cnt register must be ecx for rep stos"); 7074 assert(InitArrayShortSize % BytesPerLong == 0, 7075 "InitArrayShortSize should be the multiple of BytesPerLong"); 7076 7077 Label DONE; 7078 7079 xorptr(tmp, tmp); 7080 7081 if (!is_large) { 7082 Label LOOP, LONG; 7083 cmpptr(cnt, InitArrayShortSize/BytesPerLong); 7084 jccb(Assembler::greater, LONG); 7085 7086 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM 7087 7088 decrement(cnt); 7089 jccb(Assembler::negative, DONE); // Zero length 7090 7091 // Use individual pointer-sized stores for small counts: 7092 BIND(LOOP); 7093 movptr(Address(base, cnt, Address::times_ptr), tmp); 7094 decrement(cnt); 7095 jccb(Assembler::greaterEqual, LOOP); 7096 jmpb(DONE); 7097 7098 BIND(LONG); 7099 } 7100 7101 // Use longer rep-prefixed ops for non-small counts: 7102 if (UseFastStosb) { 7103 shlptr(cnt, 3); // convert to number of bytes 7104 rep_stosb(); 7105 } else { 7106 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM 7107 rep_stos(); 7108 } 7109 7110 BIND(DONE); 7111 } 7112 7113 #ifdef COMPILER2 7114 7115 // IndexOf for constant substrings with size >= 8 chars 7116 // which don't need to be loaded through stack. 7117 void MacroAssembler::string_indexofC8(Register str1, Register str2, 7118 Register cnt1, Register cnt2, 7119 int int_cnt2, Register result, 7120 XMMRegister vec, Register tmp, 7121 int ae) { 7122 ShortBranchVerifier sbv(this); 7123 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 7124 assert(ae != StrIntrinsicNode::LU, "Invalid encoding"); 7125 7126 // This method uses the pcmpestri instruction with bound registers 7127 // inputs: 7128 // xmm - substring 7129 // rax - substring length (elements count) 7130 // mem - scanned string 7131 // rdx - string length (elements count) 7132 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts) 7133 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes) 7134 // outputs: 7135 // rcx - matched index in string 7136 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 7137 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts 7138 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8 7139 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2; 7140 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1; 7141 7142 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, 7143 RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR, 7144 MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE; 7145 7146 // Note, inline_string_indexOf() generates checks: 7147 // if (substr.count > string.count) return -1; 7148 // if (substr.count == 0) return 0; 7149 assert(int_cnt2 >= stride, "this code is used only for cnt2 >= 8 chars"); 7150 7151 // Load substring. 7152 if (ae == StrIntrinsicNode::UL) { 7153 pmovzxbw(vec, Address(str2, 0)); 7154 } else { 7155 movdqu(vec, Address(str2, 0)); 7156 } 7157 movl(cnt2, int_cnt2); 7158 movptr(result, str1); // string addr 7159 7160 if (int_cnt2 > stride) { 7161 jmpb(SCAN_TO_SUBSTR); 7162 7163 // Reload substr for rescan, this code 7164 // is executed only for large substrings (> 8 chars) 7165 bind(RELOAD_SUBSTR); 7166 if (ae == StrIntrinsicNode::UL) { 7167 pmovzxbw(vec, Address(str2, 0)); 7168 } else { 7169 movdqu(vec, Address(str2, 0)); 7170 } 7171 negptr(cnt2); // Jumped here with negative cnt2, convert to positive 7172 7173 bind(RELOAD_STR); 7174 // We came here after the beginning of the substring was 7175 // matched but the rest of it was not so we need to search 7176 // again. Start from the next element after the previous match. 7177 7178 // cnt2 is number of substring reminding elements and 7179 // cnt1 is number of string reminding elements when cmp failed. 7180 // Restored cnt1 = cnt1 - cnt2 + int_cnt2 7181 subl(cnt1, cnt2); 7182 addl(cnt1, int_cnt2); 7183 movl(cnt2, int_cnt2); // Now restore cnt2 7184 7185 decrementl(cnt1); // Shift to next element 7186 cmpl(cnt1, cnt2); 7187 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7188 7189 addptr(result, (1<<scale1)); 7190 7191 } // (int_cnt2 > 8) 7192 7193 // Scan string for start of substr in 16-byte vectors 7194 bind(SCAN_TO_SUBSTR); 7195 pcmpestri(vec, Address(result, 0), mode); 7196 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1 7197 subl(cnt1, stride); 7198 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string 7199 cmpl(cnt1, cnt2); 7200 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7201 addptr(result, 16); 7202 jmpb(SCAN_TO_SUBSTR); 7203 7204 // Found a potential substr 7205 bind(FOUND_CANDIDATE); 7206 // Matched whole vector if first element matched (tmp(rcx) == 0). 7207 if (int_cnt2 == stride) { 7208 jccb(Assembler::overflow, RET_FOUND); // OF == 1 7209 } else { // int_cnt2 > 8 7210 jccb(Assembler::overflow, FOUND_SUBSTR); 7211 } 7212 // After pcmpestri tmp(rcx) contains matched element index 7213 // Compute start addr of substr 7214 lea(result, Address(result, tmp, scale1)); 7215 7216 // Make sure string is still long enough 7217 subl(cnt1, tmp); 7218 cmpl(cnt1, cnt2); 7219 if (int_cnt2 == stride) { 7220 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR); 7221 } else { // int_cnt2 > 8 7222 jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD); 7223 } 7224 // Left less then substring. 7225 7226 bind(RET_NOT_FOUND); 7227 movl(result, -1); 7228 jmp(EXIT); 7229 7230 if (int_cnt2 > stride) { 7231 // This code is optimized for the case when whole substring 7232 // is matched if its head is matched. 7233 bind(MATCH_SUBSTR_HEAD); 7234 pcmpestri(vec, Address(result, 0), mode); 7235 // Reload only string if does not match 7236 jcc(Assembler::noOverflow, RELOAD_STR); // OF == 0 7237 7238 Label CONT_SCAN_SUBSTR; 7239 // Compare the rest of substring (> 8 chars). 7240 bind(FOUND_SUBSTR); 7241 // First 8 chars are already matched. 7242 negptr(cnt2); 7243 addptr(cnt2, stride); 7244 7245 bind(SCAN_SUBSTR); 7246 subl(cnt1, stride); 7247 cmpl(cnt2, -stride); // Do not read beyond substring 7248 jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR); 7249 // Back-up strings to avoid reading beyond substring: 7250 // cnt1 = cnt1 - cnt2 + 8 7251 addl(cnt1, cnt2); // cnt2 is negative 7252 addl(cnt1, stride); 7253 movl(cnt2, stride); negptr(cnt2); 7254 bind(CONT_SCAN_SUBSTR); 7255 if (int_cnt2 < (int)G) { 7256 int tail_off1 = int_cnt2<<scale1; 7257 int tail_off2 = int_cnt2<<scale2; 7258 if (ae == StrIntrinsicNode::UL) { 7259 pmovzxbw(vec, Address(str2, cnt2, scale2, tail_off2)); 7260 } else { 7261 movdqu(vec, Address(str2, cnt2, scale2, tail_off2)); 7262 } 7263 pcmpestri(vec, Address(result, cnt2, scale1, tail_off1), mode); 7264 } else { 7265 // calculate index in register to avoid integer overflow (int_cnt2*2) 7266 movl(tmp, int_cnt2); 7267 addptr(tmp, cnt2); 7268 if (ae == StrIntrinsicNode::UL) { 7269 pmovzxbw(vec, Address(str2, tmp, scale2, 0)); 7270 } else { 7271 movdqu(vec, Address(str2, tmp, scale2, 0)); 7272 } 7273 pcmpestri(vec, Address(result, tmp, scale1, 0), mode); 7274 } 7275 // Need to reload strings pointers if not matched whole vector 7276 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0 7277 addptr(cnt2, stride); 7278 jcc(Assembler::negative, SCAN_SUBSTR); 7279 // Fall through if found full substring 7280 7281 } // (int_cnt2 > 8) 7282 7283 bind(RET_FOUND); 7284 // Found result if we matched full small substring. 7285 // Compute substr offset 7286 subptr(result, str1); 7287 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 7288 shrl(result, 1); // index 7289 } 7290 bind(EXIT); 7291 7292 } // string_indexofC8 7293 7294 // Small strings are loaded through stack if they cross page boundary. 7295 void MacroAssembler::string_indexof(Register str1, Register str2, 7296 Register cnt1, Register cnt2, 7297 int int_cnt2, Register result, 7298 XMMRegister vec, Register tmp, 7299 int ae) { 7300 ShortBranchVerifier sbv(this); 7301 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 7302 assert(ae != StrIntrinsicNode::LU, "Invalid encoding"); 7303 7304 // 7305 // int_cnt2 is length of small (< 8 chars) constant substring 7306 // or (-1) for non constant substring in which case its length 7307 // is in cnt2 register. 7308 // 7309 // Note, inline_string_indexOf() generates checks: 7310 // if (substr.count > string.count) return -1; 7311 // if (substr.count == 0) return 0; 7312 // 7313 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8 7314 assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < stride), "should be != 0"); 7315 // This method uses the pcmpestri instruction with bound registers 7316 // inputs: 7317 // xmm - substring 7318 // rax - substring length (elements count) 7319 // mem - scanned string 7320 // rdx - string length (elements count) 7321 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts) 7322 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes) 7323 // outputs: 7324 // rcx - matched index in string 7325 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 7326 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts 7327 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2; 7328 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1; 7329 7330 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR, 7331 RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR, 7332 FOUND_CANDIDATE; 7333 7334 { //======================================================== 7335 // We don't know where these strings are located 7336 // and we can't read beyond them. Load them through stack. 7337 Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR; 7338 7339 movptr(tmp, rsp); // save old SP 7340 7341 if (int_cnt2 > 0) { // small (< 8 chars) constant substring 7342 if (int_cnt2 == (1>>scale2)) { // One byte 7343 assert((ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL), "Only possible for latin1 encoding"); 7344 load_unsigned_byte(result, Address(str2, 0)); 7345 movdl(vec, result); // move 32 bits 7346 } else if (ae == StrIntrinsicNode::LL && int_cnt2 == 3) { // Three bytes 7347 // Not enough header space in 32-bit VM: 12+3 = 15. 7348 movl(result, Address(str2, -1)); 7349 shrl(result, 8); 7350 movdl(vec, result); // move 32 bits 7351 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (2>>scale2)) { // One char 7352 load_unsigned_short(result, Address(str2, 0)); 7353 movdl(vec, result); // move 32 bits 7354 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (4>>scale2)) { // Two chars 7355 movdl(vec, Address(str2, 0)); // move 32 bits 7356 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (8>>scale2)) { // Four chars 7357 movq(vec, Address(str2, 0)); // move 64 bits 7358 } else { // cnt2 = { 3, 5, 6, 7 } || (ae == StrIntrinsicNode::UL && cnt2 ={2, ..., 7}) 7359 // Array header size is 12 bytes in 32-bit VM 7360 // + 6 bytes for 3 chars == 18 bytes, 7361 // enough space to load vec and shift. 7362 assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity"); 7363 if (ae == StrIntrinsicNode::UL) { 7364 int tail_off = int_cnt2-8; 7365 pmovzxbw(vec, Address(str2, tail_off)); 7366 psrldq(vec, -2*tail_off); 7367 } 7368 else { 7369 int tail_off = int_cnt2*(1<<scale2); 7370 movdqu(vec, Address(str2, tail_off-16)); 7371 psrldq(vec, 16-tail_off); 7372 } 7373 } 7374 } else { // not constant substring 7375 cmpl(cnt2, stride); 7376 jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough 7377 7378 // We can read beyond string if srt+16 does not cross page boundary 7379 // since heaps are aligned and mapped by pages. 7380 assert(os::vm_page_size() < (int)G, "default page should be small"); 7381 movl(result, str2); // We need only low 32 bits 7382 andl(result, (os::vm_page_size()-1)); 7383 cmpl(result, (os::vm_page_size()-16)); 7384 jccb(Assembler::belowEqual, CHECK_STR); 7385 7386 // Move small strings to stack to allow load 16 bytes into vec. 7387 subptr(rsp, 16); 7388 int stk_offset = wordSize-(1<<scale2); 7389 push(cnt2); 7390 7391 bind(COPY_SUBSTR); 7392 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL) { 7393 load_unsigned_byte(result, Address(str2, cnt2, scale2, -1)); 7394 movb(Address(rsp, cnt2, scale2, stk_offset), result); 7395 } else if (ae == StrIntrinsicNode::UU) { 7396 load_unsigned_short(result, Address(str2, cnt2, scale2, -2)); 7397 movw(Address(rsp, cnt2, scale2, stk_offset), result); 7398 } 7399 decrement(cnt2); 7400 jccb(Assembler::notZero, COPY_SUBSTR); 7401 7402 pop(cnt2); 7403 movptr(str2, rsp); // New substring address 7404 } // non constant 7405 7406 bind(CHECK_STR); 7407 cmpl(cnt1, stride); 7408 jccb(Assembler::aboveEqual, BIG_STRINGS); 7409 7410 // Check cross page boundary. 7411 movl(result, str1); // We need only low 32 bits 7412 andl(result, (os::vm_page_size()-1)); 7413 cmpl(result, (os::vm_page_size()-16)); 7414 jccb(Assembler::belowEqual, BIG_STRINGS); 7415 7416 subptr(rsp, 16); 7417 int stk_offset = -(1<<scale1); 7418 if (int_cnt2 < 0) { // not constant 7419 push(cnt2); 7420 stk_offset += wordSize; 7421 } 7422 movl(cnt2, cnt1); 7423 7424 bind(COPY_STR); 7425 if (ae == StrIntrinsicNode::LL) { 7426 load_unsigned_byte(result, Address(str1, cnt2, scale1, -1)); 7427 movb(Address(rsp, cnt2, scale1, stk_offset), result); 7428 } else { 7429 load_unsigned_short(result, Address(str1, cnt2, scale1, -2)); 7430 movw(Address(rsp, cnt2, scale1, stk_offset), result); 7431 } 7432 decrement(cnt2); 7433 jccb(Assembler::notZero, COPY_STR); 7434 7435 if (int_cnt2 < 0) { // not constant 7436 pop(cnt2); 7437 } 7438 movptr(str1, rsp); // New string address 7439 7440 bind(BIG_STRINGS); 7441 // Load substring. 7442 if (int_cnt2 < 0) { // -1 7443 if (ae == StrIntrinsicNode::UL) { 7444 pmovzxbw(vec, Address(str2, 0)); 7445 } else { 7446 movdqu(vec, Address(str2, 0)); 7447 } 7448 push(cnt2); // substr count 7449 push(str2); // substr addr 7450 push(str1); // string addr 7451 } else { 7452 // Small (< 8 chars) constant substrings are loaded already. 7453 movl(cnt2, int_cnt2); 7454 } 7455 push(tmp); // original SP 7456 7457 } // Finished loading 7458 7459 //======================================================== 7460 // Start search 7461 // 7462 7463 movptr(result, str1); // string addr 7464 7465 if (int_cnt2 < 0) { // Only for non constant substring 7466 jmpb(SCAN_TO_SUBSTR); 7467 7468 // SP saved at sp+0 7469 // String saved at sp+1*wordSize 7470 // Substr saved at sp+2*wordSize 7471 // Substr count saved at sp+3*wordSize 7472 7473 // Reload substr for rescan, this code 7474 // is executed only for large substrings (> 8 chars) 7475 bind(RELOAD_SUBSTR); 7476 movptr(str2, Address(rsp, 2*wordSize)); 7477 movl(cnt2, Address(rsp, 3*wordSize)); 7478 if (ae == StrIntrinsicNode::UL) { 7479 pmovzxbw(vec, Address(str2, 0)); 7480 } else { 7481 movdqu(vec, Address(str2, 0)); 7482 } 7483 // We came here after the beginning of the substring was 7484 // matched but the rest of it was not so we need to search 7485 // again. Start from the next element after the previous match. 7486 subptr(str1, result); // Restore counter 7487 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 7488 shrl(str1, 1); 7489 } 7490 addl(cnt1, str1); 7491 decrementl(cnt1); // Shift to next element 7492 cmpl(cnt1, cnt2); 7493 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7494 7495 addptr(result, (1<<scale1)); 7496 } // non constant 7497 7498 // Scan string for start of substr in 16-byte vectors 7499 bind(SCAN_TO_SUBSTR); 7500 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 7501 pcmpestri(vec, Address(result, 0), mode); 7502 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1 7503 subl(cnt1, stride); 7504 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string 7505 cmpl(cnt1, cnt2); 7506 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7507 addptr(result, 16); 7508 7509 bind(ADJUST_STR); 7510 cmpl(cnt1, stride); // Do not read beyond string 7511 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR); 7512 // Back-up string to avoid reading beyond string. 7513 lea(result, Address(result, cnt1, scale1, -16)); 7514 movl(cnt1, stride); 7515 jmpb(SCAN_TO_SUBSTR); 7516 7517 // Found a potential substr 7518 bind(FOUND_CANDIDATE); 7519 // After pcmpestri tmp(rcx) contains matched element index 7520 7521 // Make sure string is still long enough 7522 subl(cnt1, tmp); 7523 cmpl(cnt1, cnt2); 7524 jccb(Assembler::greaterEqual, FOUND_SUBSTR); 7525 // Left less then substring. 7526 7527 bind(RET_NOT_FOUND); 7528 movl(result, -1); 7529 jmpb(CLEANUP); 7530 7531 bind(FOUND_SUBSTR); 7532 // Compute start addr of substr 7533 lea(result, Address(result, tmp, scale1)); 7534 if (int_cnt2 > 0) { // Constant substring 7535 // Repeat search for small substring (< 8 chars) 7536 // from new point without reloading substring. 7537 // Have to check that we don't read beyond string. 7538 cmpl(tmp, stride-int_cnt2); 7539 jccb(Assembler::greater, ADJUST_STR); 7540 // Fall through if matched whole substring. 7541 } else { // non constant 7542 assert(int_cnt2 == -1, "should be != 0"); 7543 7544 addl(tmp, cnt2); 7545 // Found result if we matched whole substring. 7546 cmpl(tmp, stride); 7547 jccb(Assembler::lessEqual, RET_FOUND); 7548 7549 // Repeat search for small substring (<= 8 chars) 7550 // from new point 'str1' without reloading substring. 7551 cmpl(cnt2, stride); 7552 // Have to check that we don't read beyond string. 7553 jccb(Assembler::lessEqual, ADJUST_STR); 7554 7555 Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG; 7556 // Compare the rest of substring (> 8 chars). 7557 movptr(str1, result); 7558 7559 cmpl(tmp, cnt2); 7560 // First 8 chars are already matched. 7561 jccb(Assembler::equal, CHECK_NEXT); 7562 7563 bind(SCAN_SUBSTR); 7564 pcmpestri(vec, Address(str1, 0), mode); 7565 // Need to reload strings pointers if not matched whole vector 7566 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0 7567 7568 bind(CHECK_NEXT); 7569 subl(cnt2, stride); 7570 jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring 7571 addptr(str1, 16); 7572 if (ae == StrIntrinsicNode::UL) { 7573 addptr(str2, 8); 7574 } else { 7575 addptr(str2, 16); 7576 } 7577 subl(cnt1, stride); 7578 cmpl(cnt2, stride); // Do not read beyond substring 7579 jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR); 7580 // Back-up strings to avoid reading beyond substring. 7581 7582 if (ae == StrIntrinsicNode::UL) { 7583 lea(str2, Address(str2, cnt2, scale2, -8)); 7584 lea(str1, Address(str1, cnt2, scale1, -16)); 7585 } else { 7586 lea(str2, Address(str2, cnt2, scale2, -16)); 7587 lea(str1, Address(str1, cnt2, scale1, -16)); 7588 } 7589 subl(cnt1, cnt2); 7590 movl(cnt2, stride); 7591 addl(cnt1, stride); 7592 bind(CONT_SCAN_SUBSTR); 7593 if (ae == StrIntrinsicNode::UL) { 7594 pmovzxbw(vec, Address(str2, 0)); 7595 } else { 7596 movdqu(vec, Address(str2, 0)); 7597 } 7598 jmp(SCAN_SUBSTR); 7599 7600 bind(RET_FOUND_LONG); 7601 movptr(str1, Address(rsp, wordSize)); 7602 } // non constant 7603 7604 bind(RET_FOUND); 7605 // Compute substr offset 7606 subptr(result, str1); 7607 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 7608 shrl(result, 1); // index 7609 } 7610 bind(CLEANUP); 7611 pop(rsp); // restore SP 7612 7613 } // string_indexof 7614 7615 void MacroAssembler::string_indexof_char(Register str1, Register cnt1, Register ch, Register result, 7616 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp) { 7617 ShortBranchVerifier sbv(this); 7618 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 7619 7620 int stride = 8; 7621 7622 Label FOUND_CHAR, SCAN_TO_CHAR, SCAN_TO_CHAR_LOOP, 7623 SCAN_TO_8_CHAR, SCAN_TO_8_CHAR_LOOP, SCAN_TO_16_CHAR_LOOP, 7624 RET_NOT_FOUND, SCAN_TO_8_CHAR_INIT, 7625 FOUND_SEQ_CHAR, DONE_LABEL; 7626 7627 movptr(result, str1); 7628 if (UseAVX >= 2) { 7629 cmpl(cnt1, stride); 7630 jcc(Assembler::less, SCAN_TO_CHAR_LOOP); 7631 cmpl(cnt1, 2*stride); 7632 jcc(Assembler::less, SCAN_TO_8_CHAR_INIT); 7633 movdl(vec1, ch); 7634 vpbroadcastw(vec1, vec1); 7635 vpxor(vec2, vec2); 7636 movl(tmp, cnt1); 7637 andl(tmp, 0xFFFFFFF0); //vector count (in chars) 7638 andl(cnt1,0x0000000F); //tail count (in chars) 7639 7640 bind(SCAN_TO_16_CHAR_LOOP); 7641 vmovdqu(vec3, Address(result, 0)); 7642 vpcmpeqw(vec3, vec3, vec1, 1); 7643 vptest(vec2, vec3); 7644 jcc(Assembler::carryClear, FOUND_CHAR); 7645 addptr(result, 32); 7646 subl(tmp, 2*stride); 7647 jccb(Assembler::notZero, SCAN_TO_16_CHAR_LOOP); 7648 jmp(SCAN_TO_8_CHAR); 7649 bind(SCAN_TO_8_CHAR_INIT); 7650 movdl(vec1, ch); 7651 pshuflw(vec1, vec1, 0x00); 7652 pshufd(vec1, vec1, 0); 7653 pxor(vec2, vec2); 7654 } 7655 bind(SCAN_TO_8_CHAR); 7656 cmpl(cnt1, stride); 7657 if (UseAVX >= 2) { 7658 jcc(Assembler::less, SCAN_TO_CHAR); 7659 } else { 7660 jcc(Assembler::less, SCAN_TO_CHAR_LOOP); 7661 movdl(vec1, ch); 7662 pshuflw(vec1, vec1, 0x00); 7663 pshufd(vec1, vec1, 0); 7664 pxor(vec2, vec2); 7665 } 7666 movl(tmp, cnt1); 7667 andl(tmp, 0xFFFFFFF8); //vector count (in chars) 7668 andl(cnt1,0x00000007); //tail count (in chars) 7669 7670 bind(SCAN_TO_8_CHAR_LOOP); 7671 movdqu(vec3, Address(result, 0)); 7672 pcmpeqw(vec3, vec1); 7673 ptest(vec2, vec3); 7674 jcc(Assembler::carryClear, FOUND_CHAR); 7675 addptr(result, 16); 7676 subl(tmp, stride); 7677 jccb(Assembler::notZero, SCAN_TO_8_CHAR_LOOP); 7678 bind(SCAN_TO_CHAR); 7679 testl(cnt1, cnt1); 7680 jcc(Assembler::zero, RET_NOT_FOUND); 7681 bind(SCAN_TO_CHAR_LOOP); 7682 load_unsigned_short(tmp, Address(result, 0)); 7683 cmpl(ch, tmp); 7684 jccb(Assembler::equal, FOUND_SEQ_CHAR); 7685 addptr(result, 2); 7686 subl(cnt1, 1); 7687 jccb(Assembler::zero, RET_NOT_FOUND); 7688 jmp(SCAN_TO_CHAR_LOOP); 7689 7690 bind(RET_NOT_FOUND); 7691 movl(result, -1); 7692 jmpb(DONE_LABEL); 7693 7694 bind(FOUND_CHAR); 7695 if (UseAVX >= 2) { 7696 vpmovmskb(tmp, vec3); 7697 } else { 7698 pmovmskb(tmp, vec3); 7699 } 7700 bsfl(ch, tmp); 7701 addl(result, ch); 7702 7703 bind(FOUND_SEQ_CHAR); 7704 subptr(result, str1); 7705 shrl(result, 1); 7706 7707 bind(DONE_LABEL); 7708 } // string_indexof_char 7709 7710 // helper function for string_compare 7711 void MacroAssembler::load_next_elements(Register elem1, Register elem2, Register str1, Register str2, 7712 Address::ScaleFactor scale, Address::ScaleFactor scale1, 7713 Address::ScaleFactor scale2, Register index, int ae) { 7714 if (ae == StrIntrinsicNode::LL) { 7715 load_unsigned_byte(elem1, Address(str1, index, scale, 0)); 7716 load_unsigned_byte(elem2, Address(str2, index, scale, 0)); 7717 } else if (ae == StrIntrinsicNode::UU) { 7718 load_unsigned_short(elem1, Address(str1, index, scale, 0)); 7719 load_unsigned_short(elem2, Address(str2, index, scale, 0)); 7720 } else { 7721 load_unsigned_byte(elem1, Address(str1, index, scale1, 0)); 7722 load_unsigned_short(elem2, Address(str2, index, scale2, 0)); 7723 } 7724 } 7725 7726 // Compare strings, used for char[] and byte[]. 7727 void MacroAssembler::string_compare(Register str1, Register str2, 7728 Register cnt1, Register cnt2, Register result, 7729 XMMRegister vec1, int ae) { 7730 ShortBranchVerifier sbv(this); 7731 Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL; 7732 Label COMPARE_WIDE_VECTORS_LOOP_FAILED; // used only _LP64 && AVX3 7733 int stride, stride2, adr_stride, adr_stride1, adr_stride2; 7734 int stride2x2 = 0x40; 7735 Address::ScaleFactor scale = Address::no_scale; 7736 Address::ScaleFactor scale1 = Address::no_scale; 7737 Address::ScaleFactor scale2 = Address::no_scale; 7738 7739 if (ae != StrIntrinsicNode::LL) { 7740 stride2x2 = 0x20; 7741 } 7742 7743 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) { 7744 shrl(cnt2, 1); 7745 } 7746 // Compute the minimum of the string lengths and the 7747 // difference of the string lengths (stack). 7748 // Do the conditional move stuff 7749 movl(result, cnt1); 7750 subl(cnt1, cnt2); 7751 push(cnt1); 7752 cmov32(Assembler::lessEqual, cnt2, result); // cnt2 = min(cnt1, cnt2) 7753 7754 // Is the minimum length zero? 7755 testl(cnt2, cnt2); 7756 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 7757 if (ae == StrIntrinsicNode::LL) { 7758 // Load first bytes 7759 load_unsigned_byte(result, Address(str1, 0)); // result = str1[0] 7760 load_unsigned_byte(cnt1, Address(str2, 0)); // cnt1 = str2[0] 7761 } else if (ae == StrIntrinsicNode::UU) { 7762 // Load first characters 7763 load_unsigned_short(result, Address(str1, 0)); 7764 load_unsigned_short(cnt1, Address(str2, 0)); 7765 } else { 7766 load_unsigned_byte(result, Address(str1, 0)); 7767 load_unsigned_short(cnt1, Address(str2, 0)); 7768 } 7769 subl(result, cnt1); 7770 jcc(Assembler::notZero, POP_LABEL); 7771 7772 if (ae == StrIntrinsicNode::UU) { 7773 // Divide length by 2 to get number of chars 7774 shrl(cnt2, 1); 7775 } 7776 cmpl(cnt2, 1); 7777 jcc(Assembler::equal, LENGTH_DIFF_LABEL); 7778 7779 // Check if the strings start at the same location and setup scale and stride 7780 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7781 cmpptr(str1, str2); 7782 jcc(Assembler::equal, LENGTH_DIFF_LABEL); 7783 if (ae == StrIntrinsicNode::LL) { 7784 scale = Address::times_1; 7785 stride = 16; 7786 } else { 7787 scale = Address::times_2; 7788 stride = 8; 7789 } 7790 } else { 7791 scale1 = Address::times_1; 7792 scale2 = Address::times_2; 7793 // scale not used 7794 stride = 8; 7795 } 7796 7797 if (UseAVX >= 2 && UseSSE42Intrinsics) { 7798 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR; 7799 Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR; 7800 Label COMPARE_WIDE_VECTORS_LOOP_AVX2; 7801 Label COMPARE_TAIL_LONG; 7802 Label COMPARE_WIDE_VECTORS_LOOP_AVX3; // used only _LP64 && AVX3 7803 7804 int pcmpmask = 0x19; 7805 if (ae == StrIntrinsicNode::LL) { 7806 pcmpmask &= ~0x01; 7807 } 7808 7809 // Setup to compare 16-chars (32-bytes) vectors, 7810 // start from first character again because it has aligned address. 7811 if (ae == StrIntrinsicNode::LL) { 7812 stride2 = 32; 7813 } else { 7814 stride2 = 16; 7815 } 7816 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7817 adr_stride = stride << scale; 7818 } else { 7819 adr_stride1 = 8; //stride << scale1; 7820 adr_stride2 = 16; //stride << scale2; 7821 } 7822 7823 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri"); 7824 // rax and rdx are used by pcmpestri as elements counters 7825 movl(result, cnt2); 7826 andl(cnt2, ~(stride2-1)); // cnt2 holds the vector count 7827 jcc(Assembler::zero, COMPARE_TAIL_LONG); 7828 7829 // fast path : compare first 2 8-char vectors. 7830 bind(COMPARE_16_CHARS); 7831 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7832 movdqu(vec1, Address(str1, 0)); 7833 } else { 7834 pmovzxbw(vec1, Address(str1, 0)); 7835 } 7836 pcmpestri(vec1, Address(str2, 0), pcmpmask); 7837 jccb(Assembler::below, COMPARE_INDEX_CHAR); 7838 7839 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7840 movdqu(vec1, Address(str1, adr_stride)); 7841 pcmpestri(vec1, Address(str2, adr_stride), pcmpmask); 7842 } else { 7843 pmovzxbw(vec1, Address(str1, adr_stride1)); 7844 pcmpestri(vec1, Address(str2, adr_stride2), pcmpmask); 7845 } 7846 jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS); 7847 addl(cnt1, stride); 7848 7849 // Compare the characters at index in cnt1 7850 bind(COMPARE_INDEX_CHAR); // cnt1 has the offset of the mismatching character 7851 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae); 7852 subl(result, cnt2); 7853 jmp(POP_LABEL); 7854 7855 // Setup the registers to start vector comparison loop 7856 bind(COMPARE_WIDE_VECTORS); 7857 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7858 lea(str1, Address(str1, result, scale)); 7859 lea(str2, Address(str2, result, scale)); 7860 } else { 7861 lea(str1, Address(str1, result, scale1)); 7862 lea(str2, Address(str2, result, scale2)); 7863 } 7864 subl(result, stride2); 7865 subl(cnt2, stride2); 7866 jcc(Assembler::zero, COMPARE_WIDE_TAIL); 7867 negptr(result); 7868 7869 // In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest) 7870 bind(COMPARE_WIDE_VECTORS_LOOP); 7871 7872 #ifdef _LP64 7873 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop 7874 cmpl(cnt2, stride2x2); 7875 jccb(Assembler::below, COMPARE_WIDE_VECTORS_LOOP_AVX2); 7876 testl(cnt2, stride2x2-1); // cnt2 holds the vector count 7877 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX2); // means we cannot subtract by 0x40 7878 7879 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop 7880 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7881 evmovdquq(vec1, Address(str1, result, scale), Assembler::AVX_512bit); 7882 evpcmpeqb(k7, vec1, Address(str2, result, scale), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0 7883 } else { 7884 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_512bit); 7885 evpcmpeqb(k7, vec1, Address(str2, result, scale2), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0 7886 } 7887 kortestql(k7, k7); 7888 jcc(Assembler::aboveEqual, COMPARE_WIDE_VECTORS_LOOP_FAILED); // miscompare 7889 addptr(result, stride2x2); // update since we already compared at this addr 7890 subl(cnt2, stride2x2); // and sub the size too 7891 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX3); 7892 7893 vpxor(vec1, vec1); 7894 jmpb(COMPARE_WIDE_TAIL); 7895 }//if (VM_Version::supports_avx512vlbw()) 7896 #endif // _LP64 7897 7898 7899 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2); 7900 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7901 vmovdqu(vec1, Address(str1, result, scale)); 7902 vpxor(vec1, Address(str2, result, scale)); 7903 } else { 7904 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_256bit); 7905 vpxor(vec1, Address(str2, result, scale2)); 7906 } 7907 vptest(vec1, vec1); 7908 jcc(Assembler::notZero, VECTOR_NOT_EQUAL); 7909 addptr(result, stride2); 7910 subl(cnt2, stride2); 7911 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP); 7912 // clean upper bits of YMM registers 7913 vpxor(vec1, vec1); 7914 7915 // compare wide vectors tail 7916 bind(COMPARE_WIDE_TAIL); 7917 testptr(result, result); 7918 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 7919 7920 movl(result, stride2); 7921 movl(cnt2, result); 7922 negptr(result); 7923 jmp(COMPARE_WIDE_VECTORS_LOOP_AVX2); 7924 7925 // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors. 7926 bind(VECTOR_NOT_EQUAL); 7927 // clean upper bits of YMM registers 7928 vpxor(vec1, vec1); 7929 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7930 lea(str1, Address(str1, result, scale)); 7931 lea(str2, Address(str2, result, scale)); 7932 } else { 7933 lea(str1, Address(str1, result, scale1)); 7934 lea(str2, Address(str2, result, scale2)); 7935 } 7936 jmp(COMPARE_16_CHARS); 7937 7938 // Compare tail chars, length between 1 to 15 chars 7939 bind(COMPARE_TAIL_LONG); 7940 movl(cnt2, result); 7941 cmpl(cnt2, stride); 7942 jcc(Assembler::less, COMPARE_SMALL_STR); 7943 7944 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7945 movdqu(vec1, Address(str1, 0)); 7946 } else { 7947 pmovzxbw(vec1, Address(str1, 0)); 7948 } 7949 pcmpestri(vec1, Address(str2, 0), pcmpmask); 7950 jcc(Assembler::below, COMPARE_INDEX_CHAR); 7951 subptr(cnt2, stride); 7952 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 7953 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7954 lea(str1, Address(str1, result, scale)); 7955 lea(str2, Address(str2, result, scale)); 7956 } else { 7957 lea(str1, Address(str1, result, scale1)); 7958 lea(str2, Address(str2, result, scale2)); 7959 } 7960 negptr(cnt2); 7961 jmpb(WHILE_HEAD_LABEL); 7962 7963 bind(COMPARE_SMALL_STR); 7964 } else if (UseSSE42Intrinsics) { 7965 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL; 7966 int pcmpmask = 0x19; 7967 // Setup to compare 8-char (16-byte) vectors, 7968 // start from first character again because it has aligned address. 7969 movl(result, cnt2); 7970 andl(cnt2, ~(stride - 1)); // cnt2 holds the vector count 7971 if (ae == StrIntrinsicNode::LL) { 7972 pcmpmask &= ~0x01; 7973 } 7974 jcc(Assembler::zero, COMPARE_TAIL); 7975 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7976 lea(str1, Address(str1, result, scale)); 7977 lea(str2, Address(str2, result, scale)); 7978 } else { 7979 lea(str1, Address(str1, result, scale1)); 7980 lea(str2, Address(str2, result, scale2)); 7981 } 7982 negptr(result); 7983 7984 // pcmpestri 7985 // inputs: 7986 // vec1- substring 7987 // rax - negative string length (elements count) 7988 // mem - scanned string 7989 // rdx - string length (elements count) 7990 // pcmpmask - cmp mode: 11000 (string compare with negated result) 7991 // + 00 (unsigned bytes) or + 01 (unsigned shorts) 7992 // outputs: 7993 // rcx - first mismatched element index 7994 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri"); 7995 7996 bind(COMPARE_WIDE_VECTORS); 7997 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7998 movdqu(vec1, Address(str1, result, scale)); 7999 pcmpestri(vec1, Address(str2, result, scale), pcmpmask); 8000 } else { 8001 pmovzxbw(vec1, Address(str1, result, scale1)); 8002 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask); 8003 } 8004 // After pcmpestri cnt1(rcx) contains mismatched element index 8005 8006 jccb(Assembler::below, VECTOR_NOT_EQUAL); // CF==1 8007 addptr(result, stride); 8008 subptr(cnt2, stride); 8009 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS); 8010 8011 // compare wide vectors tail 8012 testptr(result, result); 8013 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 8014 8015 movl(cnt2, stride); 8016 movl(result, stride); 8017 negptr(result); 8018 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8019 movdqu(vec1, Address(str1, result, scale)); 8020 pcmpestri(vec1, Address(str2, result, scale), pcmpmask); 8021 } else { 8022 pmovzxbw(vec1, Address(str1, result, scale1)); 8023 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask); 8024 } 8025 jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL); 8026 8027 // Mismatched characters in the vectors 8028 bind(VECTOR_NOT_EQUAL); 8029 addptr(cnt1, result); 8030 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae); 8031 subl(result, cnt2); 8032 jmpb(POP_LABEL); 8033 8034 bind(COMPARE_TAIL); // limit is zero 8035 movl(cnt2, result); 8036 // Fallthru to tail compare 8037 } 8038 // Shift str2 and str1 to the end of the arrays, negate min 8039 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8040 lea(str1, Address(str1, cnt2, scale)); 8041 lea(str2, Address(str2, cnt2, scale)); 8042 } else { 8043 lea(str1, Address(str1, cnt2, scale1)); 8044 lea(str2, Address(str2, cnt2, scale2)); 8045 } 8046 decrementl(cnt2); // first character was compared already 8047 negptr(cnt2); 8048 8049 // Compare the rest of the elements 8050 bind(WHILE_HEAD_LABEL); 8051 load_next_elements(result, cnt1, str1, str2, scale, scale1, scale2, cnt2, ae); 8052 subl(result, cnt1); 8053 jccb(Assembler::notZero, POP_LABEL); 8054 increment(cnt2); 8055 jccb(Assembler::notZero, WHILE_HEAD_LABEL); 8056 8057 // Strings are equal up to min length. Return the length difference. 8058 bind(LENGTH_DIFF_LABEL); 8059 pop(result); 8060 if (ae == StrIntrinsicNode::UU) { 8061 // Divide diff by 2 to get number of chars 8062 sarl(result, 1); 8063 } 8064 jmpb(DONE_LABEL); 8065 8066 #ifdef _LP64 8067 if (VM_Version::supports_avx512vlbw()) { 8068 8069 bind(COMPARE_WIDE_VECTORS_LOOP_FAILED); 8070 8071 kmovql(cnt1, k7); 8072 notq(cnt1); 8073 bsfq(cnt2, cnt1); 8074 if (ae != StrIntrinsicNode::LL) { 8075 // Divide diff by 2 to get number of chars 8076 sarl(cnt2, 1); 8077 } 8078 addq(result, cnt2); 8079 if (ae == StrIntrinsicNode::LL) { 8080 load_unsigned_byte(cnt1, Address(str2, result)); 8081 load_unsigned_byte(result, Address(str1, result)); 8082 } else if (ae == StrIntrinsicNode::UU) { 8083 load_unsigned_short(cnt1, Address(str2, result, scale)); 8084 load_unsigned_short(result, Address(str1, result, scale)); 8085 } else { 8086 load_unsigned_short(cnt1, Address(str2, result, scale2)); 8087 load_unsigned_byte(result, Address(str1, result, scale1)); 8088 } 8089 subl(result, cnt1); 8090 jmpb(POP_LABEL); 8091 }//if (VM_Version::supports_avx512vlbw()) 8092 #endif // _LP64 8093 8094 // Discard the stored length difference 8095 bind(POP_LABEL); 8096 pop(cnt1); 8097 8098 // That's it 8099 bind(DONE_LABEL); 8100 if(ae == StrIntrinsicNode::UL) { 8101 negl(result); 8102 } 8103 8104 } 8105 8106 // Search for Non-ASCII character (Negative byte value) in a byte array, 8107 // return true if it has any and false otherwise. 8108 // ..\jdk\src\java.base\share\classes\java\lang\StringCoding.java 8109 // @HotSpotIntrinsicCandidate 8110 // private static boolean hasNegatives(byte[] ba, int off, int len) { 8111 // for (int i = off; i < off + len; i++) { 8112 // if (ba[i] < 0) { 8113 // return true; 8114 // } 8115 // } 8116 // return false; 8117 // } 8118 void MacroAssembler::has_negatives(Register ary1, Register len, 8119 Register result, Register tmp1, 8120 XMMRegister vec1, XMMRegister vec2) { 8121 // rsi: byte array 8122 // rcx: len 8123 // rax: result 8124 ShortBranchVerifier sbv(this); 8125 assert_different_registers(ary1, len, result, tmp1); 8126 assert_different_registers(vec1, vec2); 8127 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_CHAR, COMPARE_VECTORS, COMPARE_BYTE; 8128 8129 // len == 0 8130 testl(len, len); 8131 jcc(Assembler::zero, FALSE_LABEL); 8132 8133 if ((UseAVX > 2) && // AVX512 8134 VM_Version::supports_avx512vlbw() && 8135 VM_Version::supports_bmi2()) { 8136 8137 set_vector_masking(); // opening of the stub context for programming mask registers 8138 8139 Label test_64_loop, test_tail; 8140 Register tmp3_aliased = len; 8141 8142 movl(tmp1, len); 8143 vpxor(vec2, vec2, vec2, Assembler::AVX_512bit); 8144 8145 andl(tmp1, 64 - 1); // tail count (in chars) 0x3F 8146 andl(len, ~(64 - 1)); // vector count (in chars) 8147 jccb(Assembler::zero, test_tail); 8148 8149 lea(ary1, Address(ary1, len, Address::times_1)); 8150 negptr(len); 8151 8152 bind(test_64_loop); 8153 // Check whether our 64 elements of size byte contain negatives 8154 evpcmpgtb(k2, vec2, Address(ary1, len, Address::times_1), Assembler::AVX_512bit); 8155 kortestql(k2, k2); 8156 jcc(Assembler::notZero, TRUE_LABEL); 8157 8158 addptr(len, 64); 8159 jccb(Assembler::notZero, test_64_loop); 8160 8161 8162 bind(test_tail); 8163 // bail out when there is nothing to be done 8164 testl(tmp1, -1); 8165 jcc(Assembler::zero, FALSE_LABEL); 8166 8167 // Save k1 8168 kmovql(k3, k1); 8169 8170 // ~(~0 << len) applied up to two times (for 32-bit scenario) 8171 #ifdef _LP64 8172 mov64(tmp3_aliased, 0xFFFFFFFFFFFFFFFF); 8173 shlxq(tmp3_aliased, tmp3_aliased, tmp1); 8174 notq(tmp3_aliased); 8175 kmovql(k1, tmp3_aliased); 8176 #else 8177 Label k_init; 8178 jmp(k_init); 8179 8180 // We could not read 64-bits from a general purpose register thus we move 8181 // data required to compose 64 1's to the instruction stream 8182 // We emit 64 byte wide series of elements from 0..63 which later on would 8183 // be used as a compare targets with tail count contained in tmp1 register. 8184 // Result would be a k1 register having tmp1 consecutive number or 1 8185 // counting from least significant bit. 8186 address tmp = pc(); 8187 emit_int64(0x0706050403020100); 8188 emit_int64(0x0F0E0D0C0B0A0908); 8189 emit_int64(0x1716151413121110); 8190 emit_int64(0x1F1E1D1C1B1A1918); 8191 emit_int64(0x2726252423222120); 8192 emit_int64(0x2F2E2D2C2B2A2928); 8193 emit_int64(0x3736353433323130); 8194 emit_int64(0x3F3E3D3C3B3A3938); 8195 8196 bind(k_init); 8197 lea(len, InternalAddress(tmp)); 8198 // create mask to test for negative byte inside a vector 8199 evpbroadcastb(vec1, tmp1, Assembler::AVX_512bit); 8200 evpcmpgtb(k1, vec1, Address(len, 0), Assembler::AVX_512bit); 8201 8202 #endif 8203 evpcmpgtb(k2, k1, vec2, Address(ary1, 0), Assembler::AVX_512bit); 8204 ktestq(k2, k1); 8205 // Restore k1 8206 kmovql(k1, k3); 8207 jcc(Assembler::notZero, TRUE_LABEL); 8208 8209 jmp(FALSE_LABEL); 8210 8211 clear_vector_masking(); // closing of the stub context for programming mask registers 8212 } else { 8213 movl(result, len); // copy 8214 8215 if (UseAVX == 2 && UseSSE >= 2) { 8216 // With AVX2, use 32-byte vector compare 8217 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8218 8219 // Compare 32-byte vectors 8220 andl(result, 0x0000001f); // tail count (in bytes) 8221 andl(len, 0xffffffe0); // vector count (in bytes) 8222 jccb(Assembler::zero, COMPARE_TAIL); 8223 8224 lea(ary1, Address(ary1, len, Address::times_1)); 8225 negptr(len); 8226 8227 movl(tmp1, 0x80808080); // create mask to test for Unicode chars in vector 8228 movdl(vec2, tmp1); 8229 vpbroadcastd(vec2, vec2); 8230 8231 bind(COMPARE_WIDE_VECTORS); 8232 vmovdqu(vec1, Address(ary1, len, Address::times_1)); 8233 vptest(vec1, vec2); 8234 jccb(Assembler::notZero, TRUE_LABEL); 8235 addptr(len, 32); 8236 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8237 8238 testl(result, result); 8239 jccb(Assembler::zero, FALSE_LABEL); 8240 8241 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32)); 8242 vptest(vec1, vec2); 8243 jccb(Assembler::notZero, TRUE_LABEL); 8244 jmpb(FALSE_LABEL); 8245 8246 bind(COMPARE_TAIL); // len is zero 8247 movl(len, result); 8248 // Fallthru to tail compare 8249 } else if (UseSSE42Intrinsics) { 8250 // With SSE4.2, use double quad vector compare 8251 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8252 8253 // Compare 16-byte vectors 8254 andl(result, 0x0000000f); // tail count (in bytes) 8255 andl(len, 0xfffffff0); // vector count (in bytes) 8256 jccb(Assembler::zero, COMPARE_TAIL); 8257 8258 lea(ary1, Address(ary1, len, Address::times_1)); 8259 negptr(len); 8260 8261 movl(tmp1, 0x80808080); 8262 movdl(vec2, tmp1); 8263 pshufd(vec2, vec2, 0); 8264 8265 bind(COMPARE_WIDE_VECTORS); 8266 movdqu(vec1, Address(ary1, len, Address::times_1)); 8267 ptest(vec1, vec2); 8268 jccb(Assembler::notZero, TRUE_LABEL); 8269 addptr(len, 16); 8270 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8271 8272 testl(result, result); 8273 jccb(Assembler::zero, FALSE_LABEL); 8274 8275 movdqu(vec1, Address(ary1, result, Address::times_1, -16)); 8276 ptest(vec1, vec2); 8277 jccb(Assembler::notZero, TRUE_LABEL); 8278 jmpb(FALSE_LABEL); 8279 8280 bind(COMPARE_TAIL); // len is zero 8281 movl(len, result); 8282 // Fallthru to tail compare 8283 } 8284 } 8285 // Compare 4-byte vectors 8286 andl(len, 0xfffffffc); // vector count (in bytes) 8287 jccb(Assembler::zero, COMPARE_CHAR); 8288 8289 lea(ary1, Address(ary1, len, Address::times_1)); 8290 negptr(len); 8291 8292 bind(COMPARE_VECTORS); 8293 movl(tmp1, Address(ary1, len, Address::times_1)); 8294 andl(tmp1, 0x80808080); 8295 jccb(Assembler::notZero, TRUE_LABEL); 8296 addptr(len, 4); 8297 jcc(Assembler::notZero, COMPARE_VECTORS); 8298 8299 // Compare trailing char (final 2 bytes), if any 8300 bind(COMPARE_CHAR); 8301 testl(result, 0x2); // tail char 8302 jccb(Assembler::zero, COMPARE_BYTE); 8303 load_unsigned_short(tmp1, Address(ary1, 0)); 8304 andl(tmp1, 0x00008080); 8305 jccb(Assembler::notZero, TRUE_LABEL); 8306 subptr(result, 2); 8307 lea(ary1, Address(ary1, 2)); 8308 8309 bind(COMPARE_BYTE); 8310 testl(result, 0x1); // tail byte 8311 jccb(Assembler::zero, FALSE_LABEL); 8312 load_unsigned_byte(tmp1, Address(ary1, 0)); 8313 andl(tmp1, 0x00000080); 8314 jccb(Assembler::notEqual, TRUE_LABEL); 8315 jmpb(FALSE_LABEL); 8316 8317 bind(TRUE_LABEL); 8318 movl(result, 1); // return true 8319 jmpb(DONE); 8320 8321 bind(FALSE_LABEL); 8322 xorl(result, result); // return false 8323 8324 // That's it 8325 bind(DONE); 8326 if (UseAVX >= 2 && UseSSE >= 2) { 8327 // clean upper bits of YMM registers 8328 vpxor(vec1, vec1); 8329 vpxor(vec2, vec2); 8330 } 8331 } 8332 // Compare char[] or byte[] arrays aligned to 4 bytes or substrings. 8333 void MacroAssembler::arrays_equals(bool is_array_equ, Register ary1, Register ary2, 8334 Register limit, Register result, Register chr, 8335 XMMRegister vec1, XMMRegister vec2, bool is_char) { 8336 ShortBranchVerifier sbv(this); 8337 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR, COMPARE_BYTE; 8338 8339 int length_offset = arrayOopDesc::length_offset_in_bytes(); 8340 int base_offset = arrayOopDesc::base_offset_in_bytes(is_char ? T_CHAR : T_BYTE); 8341 8342 if (is_array_equ) { 8343 // Check the input args 8344 cmpptr(ary1, ary2); 8345 jcc(Assembler::equal, TRUE_LABEL); 8346 8347 // Need additional checks for arrays_equals. 8348 testptr(ary1, ary1); 8349 jcc(Assembler::zero, FALSE_LABEL); 8350 testptr(ary2, ary2); 8351 jcc(Assembler::zero, FALSE_LABEL); 8352 8353 // Check the lengths 8354 movl(limit, Address(ary1, length_offset)); 8355 cmpl(limit, Address(ary2, length_offset)); 8356 jcc(Assembler::notEqual, FALSE_LABEL); 8357 } 8358 8359 // count == 0 8360 testl(limit, limit); 8361 jcc(Assembler::zero, TRUE_LABEL); 8362 8363 if (is_array_equ) { 8364 // Load array address 8365 lea(ary1, Address(ary1, base_offset)); 8366 lea(ary2, Address(ary2, base_offset)); 8367 } 8368 8369 if (is_array_equ && is_char) { 8370 // arrays_equals when used for char[]. 8371 shll(limit, 1); // byte count != 0 8372 } 8373 movl(result, limit); // copy 8374 8375 if (UseAVX >= 2) { 8376 // With AVX2, use 32-byte vector compare 8377 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8378 8379 // Compare 32-byte vectors 8380 andl(result, 0x0000001f); // tail count (in bytes) 8381 andl(limit, 0xffffffe0); // vector count (in bytes) 8382 jcc(Assembler::zero, COMPARE_TAIL); 8383 8384 lea(ary1, Address(ary1, limit, Address::times_1)); 8385 lea(ary2, Address(ary2, limit, Address::times_1)); 8386 negptr(limit); 8387 8388 bind(COMPARE_WIDE_VECTORS); 8389 8390 #ifdef _LP64 8391 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop 8392 Label COMPARE_WIDE_VECTORS_LOOP_AVX2, COMPARE_WIDE_VECTORS_LOOP_AVX3; 8393 8394 cmpl(limit, -64); 8395 jccb(Assembler::greater, COMPARE_WIDE_VECTORS_LOOP_AVX2); 8396 8397 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop 8398 8399 evmovdquq(vec1, Address(ary1, limit, Address::times_1), Assembler::AVX_512bit); 8400 evpcmpeqb(k7, vec1, Address(ary2, limit, Address::times_1), Assembler::AVX_512bit); 8401 kortestql(k7, k7); 8402 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare 8403 addptr(limit, 64); // update since we already compared at this addr 8404 cmpl(limit, -64); 8405 jccb(Assembler::lessEqual, COMPARE_WIDE_VECTORS_LOOP_AVX3); 8406 8407 // At this point we may still need to compare -limit+result bytes. 8408 // We could execute the next two instruction and just continue via non-wide path: 8409 // cmpl(limit, 0); 8410 // jcc(Assembler::equal, COMPARE_TAIL); // true 8411 // But since we stopped at the points ary{1,2}+limit which are 8412 // not farther than 64 bytes from the ends of arrays ary{1,2}+result 8413 // (|limit| <= 32 and result < 32), 8414 // we may just compare the last 64 bytes. 8415 // 8416 addptr(result, -64); // it is safe, bc we just came from this area 8417 evmovdquq(vec1, Address(ary1, result, Address::times_1), Assembler::AVX_512bit); 8418 evpcmpeqb(k7, vec1, Address(ary2, result, Address::times_1), Assembler::AVX_512bit); 8419 kortestql(k7, k7); 8420 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare 8421 8422 jmp(TRUE_LABEL); 8423 8424 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2); 8425 8426 }//if (VM_Version::supports_avx512vlbw()) 8427 #endif //_LP64 8428 8429 vmovdqu(vec1, Address(ary1, limit, Address::times_1)); 8430 vmovdqu(vec2, Address(ary2, limit, Address::times_1)); 8431 vpxor(vec1, vec2); 8432 8433 vptest(vec1, vec1); 8434 jcc(Assembler::notZero, FALSE_LABEL); 8435 addptr(limit, 32); 8436 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8437 8438 testl(result, result); 8439 jcc(Assembler::zero, TRUE_LABEL); 8440 8441 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32)); 8442 vmovdqu(vec2, Address(ary2, result, Address::times_1, -32)); 8443 vpxor(vec1, vec2); 8444 8445 vptest(vec1, vec1); 8446 jccb(Assembler::notZero, FALSE_LABEL); 8447 jmpb(TRUE_LABEL); 8448 8449 bind(COMPARE_TAIL); // limit is zero 8450 movl(limit, result); 8451 // Fallthru to tail compare 8452 } else if (UseSSE42Intrinsics) { 8453 // With SSE4.2, use double quad vector compare 8454 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8455 8456 // Compare 16-byte vectors 8457 andl(result, 0x0000000f); // tail count (in bytes) 8458 andl(limit, 0xfffffff0); // vector count (in bytes) 8459 jcc(Assembler::zero, COMPARE_TAIL); 8460 8461 lea(ary1, Address(ary1, limit, Address::times_1)); 8462 lea(ary2, Address(ary2, limit, Address::times_1)); 8463 negptr(limit); 8464 8465 bind(COMPARE_WIDE_VECTORS); 8466 movdqu(vec1, Address(ary1, limit, Address::times_1)); 8467 movdqu(vec2, Address(ary2, limit, Address::times_1)); 8468 pxor(vec1, vec2); 8469 8470 ptest(vec1, vec1); 8471 jcc(Assembler::notZero, FALSE_LABEL); 8472 addptr(limit, 16); 8473 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8474 8475 testl(result, result); 8476 jcc(Assembler::zero, TRUE_LABEL); 8477 8478 movdqu(vec1, Address(ary1, result, Address::times_1, -16)); 8479 movdqu(vec2, Address(ary2, result, Address::times_1, -16)); 8480 pxor(vec1, vec2); 8481 8482 ptest(vec1, vec1); 8483 jccb(Assembler::notZero, FALSE_LABEL); 8484 jmpb(TRUE_LABEL); 8485 8486 bind(COMPARE_TAIL); // limit is zero 8487 movl(limit, result); 8488 // Fallthru to tail compare 8489 } 8490 8491 // Compare 4-byte vectors 8492 andl(limit, 0xfffffffc); // vector count (in bytes) 8493 jccb(Assembler::zero, COMPARE_CHAR); 8494 8495 lea(ary1, Address(ary1, limit, Address::times_1)); 8496 lea(ary2, Address(ary2, limit, Address::times_1)); 8497 negptr(limit); 8498 8499 bind(COMPARE_VECTORS); 8500 movl(chr, Address(ary1, limit, Address::times_1)); 8501 cmpl(chr, Address(ary2, limit, Address::times_1)); 8502 jccb(Assembler::notEqual, FALSE_LABEL); 8503 addptr(limit, 4); 8504 jcc(Assembler::notZero, COMPARE_VECTORS); 8505 8506 // Compare trailing char (final 2 bytes), if any 8507 bind(COMPARE_CHAR); 8508 testl(result, 0x2); // tail char 8509 jccb(Assembler::zero, COMPARE_BYTE); 8510 load_unsigned_short(chr, Address(ary1, 0)); 8511 load_unsigned_short(limit, Address(ary2, 0)); 8512 cmpl(chr, limit); 8513 jccb(Assembler::notEqual, FALSE_LABEL); 8514 8515 if (is_array_equ && is_char) { 8516 bind(COMPARE_BYTE); 8517 } else { 8518 lea(ary1, Address(ary1, 2)); 8519 lea(ary2, Address(ary2, 2)); 8520 8521 bind(COMPARE_BYTE); 8522 testl(result, 0x1); // tail byte 8523 jccb(Assembler::zero, TRUE_LABEL); 8524 load_unsigned_byte(chr, Address(ary1, 0)); 8525 load_unsigned_byte(limit, Address(ary2, 0)); 8526 cmpl(chr, limit); 8527 jccb(Assembler::notEqual, FALSE_LABEL); 8528 } 8529 bind(TRUE_LABEL); 8530 movl(result, 1); // return true 8531 jmpb(DONE); 8532 8533 bind(FALSE_LABEL); 8534 xorl(result, result); // return false 8535 8536 // That's it 8537 bind(DONE); 8538 if (UseAVX >= 2) { 8539 // clean upper bits of YMM registers 8540 vpxor(vec1, vec1); 8541 vpxor(vec2, vec2); 8542 } 8543 } 8544 8545 #endif 8546 8547 void MacroAssembler::generate_fill(BasicType t, bool aligned, 8548 Register to, Register value, Register count, 8549 Register rtmp, XMMRegister xtmp) { 8550 ShortBranchVerifier sbv(this); 8551 assert_different_registers(to, value, count, rtmp); 8552 Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte; 8553 Label L_fill_2_bytes, L_fill_4_bytes; 8554 8555 int shift = -1; 8556 switch (t) { 8557 case T_BYTE: 8558 shift = 2; 8559 break; 8560 case T_SHORT: 8561 shift = 1; 8562 break; 8563 case T_INT: 8564 shift = 0; 8565 break; 8566 default: ShouldNotReachHere(); 8567 } 8568 8569 if (t == T_BYTE) { 8570 andl(value, 0xff); 8571 movl(rtmp, value); 8572 shll(rtmp, 8); 8573 orl(value, rtmp); 8574 } 8575 if (t == T_SHORT) { 8576 andl(value, 0xffff); 8577 } 8578 if (t == T_BYTE || t == T_SHORT) { 8579 movl(rtmp, value); 8580 shll(rtmp, 16); 8581 orl(value, rtmp); 8582 } 8583 8584 cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element 8585 jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp 8586 if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) { 8587 // align source address at 4 bytes address boundary 8588 if (t == T_BYTE) { 8589 // One byte misalignment happens only for byte arrays 8590 testptr(to, 1); 8591 jccb(Assembler::zero, L_skip_align1); 8592 movb(Address(to, 0), value); 8593 increment(to); 8594 decrement(count); 8595 BIND(L_skip_align1); 8596 } 8597 // Two bytes misalignment happens only for byte and short (char) arrays 8598 testptr(to, 2); 8599 jccb(Assembler::zero, L_skip_align2); 8600 movw(Address(to, 0), value); 8601 addptr(to, 2); 8602 subl(count, 1<<(shift-1)); 8603 BIND(L_skip_align2); 8604 } 8605 if (UseSSE < 2) { 8606 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes; 8607 // Fill 32-byte chunks 8608 subl(count, 8 << shift); 8609 jcc(Assembler::less, L_check_fill_8_bytes); 8610 align(16); 8611 8612 BIND(L_fill_32_bytes_loop); 8613 8614 for (int i = 0; i < 32; i += 4) { 8615 movl(Address(to, i), value); 8616 } 8617 8618 addptr(to, 32); 8619 subl(count, 8 << shift); 8620 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop); 8621 BIND(L_check_fill_8_bytes); 8622 addl(count, 8 << shift); 8623 jccb(Assembler::zero, L_exit); 8624 jmpb(L_fill_8_bytes); 8625 8626 // 8627 // length is too short, just fill qwords 8628 // 8629 BIND(L_fill_8_bytes_loop); 8630 movl(Address(to, 0), value); 8631 movl(Address(to, 4), value); 8632 addptr(to, 8); 8633 BIND(L_fill_8_bytes); 8634 subl(count, 1 << (shift + 1)); 8635 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop); 8636 // fall through to fill 4 bytes 8637 } else { 8638 Label L_fill_32_bytes; 8639 if (!UseUnalignedLoadStores) { 8640 // align to 8 bytes, we know we are 4 byte aligned to start 8641 testptr(to, 4); 8642 jccb(Assembler::zero, L_fill_32_bytes); 8643 movl(Address(to, 0), value); 8644 addptr(to, 4); 8645 subl(count, 1<<shift); 8646 } 8647 BIND(L_fill_32_bytes); 8648 { 8649 assert( UseSSE >= 2, "supported cpu only" ); 8650 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes; 8651 if (UseAVX > 2) { 8652 movl(rtmp, 0xffff); 8653 kmovwl(k1, rtmp); 8654 } 8655 movdl(xtmp, value); 8656 if (UseAVX > 2 && UseUnalignedLoadStores) { 8657 // Fill 64-byte chunks 8658 Label L_fill_64_bytes_loop, L_check_fill_32_bytes; 8659 evpbroadcastd(xtmp, xtmp, Assembler::AVX_512bit); 8660 8661 subl(count, 16 << shift); 8662 jcc(Assembler::less, L_check_fill_32_bytes); 8663 align(16); 8664 8665 BIND(L_fill_64_bytes_loop); 8666 evmovdqul(Address(to, 0), xtmp, Assembler::AVX_512bit); 8667 addptr(to, 64); 8668 subl(count, 16 << shift); 8669 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop); 8670 8671 BIND(L_check_fill_32_bytes); 8672 addl(count, 8 << shift); 8673 jccb(Assembler::less, L_check_fill_8_bytes); 8674 vmovdqu(Address(to, 0), xtmp); 8675 addptr(to, 32); 8676 subl(count, 8 << shift); 8677 8678 BIND(L_check_fill_8_bytes); 8679 } else if (UseAVX == 2 && UseUnalignedLoadStores) { 8680 // Fill 64-byte chunks 8681 Label L_fill_64_bytes_loop, L_check_fill_32_bytes; 8682 vpbroadcastd(xtmp, xtmp); 8683 8684 subl(count, 16 << shift); 8685 jcc(Assembler::less, L_check_fill_32_bytes); 8686 align(16); 8687 8688 BIND(L_fill_64_bytes_loop); 8689 vmovdqu(Address(to, 0), xtmp); 8690 vmovdqu(Address(to, 32), xtmp); 8691 addptr(to, 64); 8692 subl(count, 16 << shift); 8693 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop); 8694 8695 BIND(L_check_fill_32_bytes); 8696 addl(count, 8 << shift); 8697 jccb(Assembler::less, L_check_fill_8_bytes); 8698 vmovdqu(Address(to, 0), xtmp); 8699 addptr(to, 32); 8700 subl(count, 8 << shift); 8701 8702 BIND(L_check_fill_8_bytes); 8703 // clean upper bits of YMM registers 8704 movdl(xtmp, value); 8705 pshufd(xtmp, xtmp, 0); 8706 } else { 8707 // Fill 32-byte chunks 8708 pshufd(xtmp, xtmp, 0); 8709 8710 subl(count, 8 << shift); 8711 jcc(Assembler::less, L_check_fill_8_bytes); 8712 align(16); 8713 8714 BIND(L_fill_32_bytes_loop); 8715 8716 if (UseUnalignedLoadStores) { 8717 movdqu(Address(to, 0), xtmp); 8718 movdqu(Address(to, 16), xtmp); 8719 } else { 8720 movq(Address(to, 0), xtmp); 8721 movq(Address(to, 8), xtmp); 8722 movq(Address(to, 16), xtmp); 8723 movq(Address(to, 24), xtmp); 8724 } 8725 8726 addptr(to, 32); 8727 subl(count, 8 << shift); 8728 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop); 8729 8730 BIND(L_check_fill_8_bytes); 8731 } 8732 addl(count, 8 << shift); 8733 jccb(Assembler::zero, L_exit); 8734 jmpb(L_fill_8_bytes); 8735 8736 // 8737 // length is too short, just fill qwords 8738 // 8739 BIND(L_fill_8_bytes_loop); 8740 movq(Address(to, 0), xtmp); 8741 addptr(to, 8); 8742 BIND(L_fill_8_bytes); 8743 subl(count, 1 << (shift + 1)); 8744 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop); 8745 } 8746 } 8747 // fill trailing 4 bytes 8748 BIND(L_fill_4_bytes); 8749 testl(count, 1<<shift); 8750 jccb(Assembler::zero, L_fill_2_bytes); 8751 movl(Address(to, 0), value); 8752 if (t == T_BYTE || t == T_SHORT) { 8753 addptr(to, 4); 8754 BIND(L_fill_2_bytes); 8755 // fill trailing 2 bytes 8756 testl(count, 1<<(shift-1)); 8757 jccb(Assembler::zero, L_fill_byte); 8758 movw(Address(to, 0), value); 8759 if (t == T_BYTE) { 8760 addptr(to, 2); 8761 BIND(L_fill_byte); 8762 // fill trailing byte 8763 testl(count, 1); 8764 jccb(Assembler::zero, L_exit); 8765 movb(Address(to, 0), value); 8766 } else { 8767 BIND(L_fill_byte); 8768 } 8769 } else { 8770 BIND(L_fill_2_bytes); 8771 } 8772 BIND(L_exit); 8773 } 8774 8775 // encode char[] to byte[] in ISO_8859_1 8776 //@HotSpotIntrinsicCandidate 8777 //private static int implEncodeISOArray(byte[] sa, int sp, 8778 //byte[] da, int dp, int len) { 8779 // int i = 0; 8780 // for (; i < len; i++) { 8781 // char c = StringUTF16.getChar(sa, sp++); 8782 // if (c > '\u00FF') 8783 // break; 8784 // da[dp++] = (byte)c; 8785 // } 8786 // return i; 8787 //} 8788 void MacroAssembler::encode_iso_array(Register src, Register dst, Register len, 8789 XMMRegister tmp1Reg, XMMRegister tmp2Reg, 8790 XMMRegister tmp3Reg, XMMRegister tmp4Reg, 8791 Register tmp5, Register result) { 8792 8793 // rsi: src 8794 // rdi: dst 8795 // rdx: len 8796 // rcx: tmp5 8797 // rax: result 8798 ShortBranchVerifier sbv(this); 8799 assert_different_registers(src, dst, len, tmp5, result); 8800 Label L_done, L_copy_1_char, L_copy_1_char_exit; 8801 8802 // set result 8803 xorl(result, result); 8804 // check for zero length 8805 testl(len, len); 8806 jcc(Assembler::zero, L_done); 8807 8808 movl(result, len); 8809 8810 // Setup pointers 8811 lea(src, Address(src, len, Address::times_2)); // char[] 8812 lea(dst, Address(dst, len, Address::times_1)); // byte[] 8813 negptr(len); 8814 8815 if (UseSSE42Intrinsics || UseAVX >= 2) { 8816 Label L_chars_8_check, L_copy_8_chars, L_copy_8_chars_exit; 8817 Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit; 8818 8819 if (UseAVX >= 2) { 8820 Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit; 8821 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector 8822 movdl(tmp1Reg, tmp5); 8823 vpbroadcastd(tmp1Reg, tmp1Reg); 8824 jmp(L_chars_32_check); 8825 8826 bind(L_copy_32_chars); 8827 vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64)); 8828 vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32)); 8829 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1); 8830 vptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 8831 jccb(Assembler::notZero, L_copy_32_chars_exit); 8832 vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1); 8833 vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector_len */ 1); 8834 vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg); 8835 8836 bind(L_chars_32_check); 8837 addptr(len, 32); 8838 jcc(Assembler::lessEqual, L_copy_32_chars); 8839 8840 bind(L_copy_32_chars_exit); 8841 subptr(len, 16); 8842 jccb(Assembler::greater, L_copy_16_chars_exit); 8843 8844 } else if (UseSSE42Intrinsics) { 8845 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector 8846 movdl(tmp1Reg, tmp5); 8847 pshufd(tmp1Reg, tmp1Reg, 0); 8848 jmpb(L_chars_16_check); 8849 } 8850 8851 bind(L_copy_16_chars); 8852 if (UseAVX >= 2) { 8853 vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32)); 8854 vptest(tmp2Reg, tmp1Reg); 8855 jcc(Assembler::notZero, L_copy_16_chars_exit); 8856 vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector_len */ 1); 8857 vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector_len */ 1); 8858 } else { 8859 if (UseAVX > 0) { 8860 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32)); 8861 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16)); 8862 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 0); 8863 } else { 8864 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32)); 8865 por(tmp2Reg, tmp3Reg); 8866 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16)); 8867 por(tmp2Reg, tmp4Reg); 8868 } 8869 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 8870 jccb(Assembler::notZero, L_copy_16_chars_exit); 8871 packuswb(tmp3Reg, tmp4Reg); 8872 } 8873 movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg); 8874 8875 bind(L_chars_16_check); 8876 addptr(len, 16); 8877 jcc(Assembler::lessEqual, L_copy_16_chars); 8878 8879 bind(L_copy_16_chars_exit); 8880 if (UseAVX >= 2) { 8881 // clean upper bits of YMM registers 8882 vpxor(tmp2Reg, tmp2Reg); 8883 vpxor(tmp3Reg, tmp3Reg); 8884 vpxor(tmp4Reg, tmp4Reg); 8885 movdl(tmp1Reg, tmp5); 8886 pshufd(tmp1Reg, tmp1Reg, 0); 8887 } 8888 subptr(len, 8); 8889 jccb(Assembler::greater, L_copy_8_chars_exit); 8890 8891 bind(L_copy_8_chars); 8892 movdqu(tmp3Reg, Address(src, len, Address::times_2, -16)); 8893 ptest(tmp3Reg, tmp1Reg); 8894 jccb(Assembler::notZero, L_copy_8_chars_exit); 8895 packuswb(tmp3Reg, tmp1Reg); 8896 movq(Address(dst, len, Address::times_1, -8), tmp3Reg); 8897 addptr(len, 8); 8898 jccb(Assembler::lessEqual, L_copy_8_chars); 8899 8900 bind(L_copy_8_chars_exit); 8901 subptr(len, 8); 8902 jccb(Assembler::zero, L_done); 8903 } 8904 8905 bind(L_copy_1_char); 8906 load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0)); 8907 testl(tmp5, 0xff00); // check if Unicode char 8908 jccb(Assembler::notZero, L_copy_1_char_exit); 8909 movb(Address(dst, len, Address::times_1, 0), tmp5); 8910 addptr(len, 1); 8911 jccb(Assembler::less, L_copy_1_char); 8912 8913 bind(L_copy_1_char_exit); 8914 addptr(result, len); // len is negative count of not processed elements 8915 8916 bind(L_done); 8917 } 8918 8919 #ifdef _LP64 8920 /** 8921 * Helper for multiply_to_len(). 8922 */ 8923 void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo, Register src1, Register src2) { 8924 addq(dest_lo, src1); 8925 adcq(dest_hi, 0); 8926 addq(dest_lo, src2); 8927 adcq(dest_hi, 0); 8928 } 8929 8930 /** 8931 * Multiply 64 bit by 64 bit first loop. 8932 */ 8933 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart, 8934 Register y, Register y_idx, Register z, 8935 Register carry, Register product, 8936 Register idx, Register kdx) { 8937 // 8938 // jlong carry, x[], y[], z[]; 8939 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) { 8940 // huge_128 product = y[idx] * x[xstart] + carry; 8941 // z[kdx] = (jlong)product; 8942 // carry = (jlong)(product >>> 64); 8943 // } 8944 // z[xstart] = carry; 8945 // 8946 8947 Label L_first_loop, L_first_loop_exit; 8948 Label L_one_x, L_one_y, L_multiply; 8949 8950 decrementl(xstart); 8951 jcc(Assembler::negative, L_one_x); 8952 8953 movq(x_xstart, Address(x, xstart, Address::times_4, 0)); 8954 rorq(x_xstart, 32); // convert big-endian to little-endian 8955 8956 bind(L_first_loop); 8957 decrementl(idx); 8958 jcc(Assembler::negative, L_first_loop_exit); 8959 decrementl(idx); 8960 jcc(Assembler::negative, L_one_y); 8961 movq(y_idx, Address(y, idx, Address::times_4, 0)); 8962 rorq(y_idx, 32); // convert big-endian to little-endian 8963 bind(L_multiply); 8964 movq(product, x_xstart); 8965 mulq(y_idx); // product(rax) * y_idx -> rdx:rax 8966 addq(product, carry); 8967 adcq(rdx, 0); 8968 subl(kdx, 2); 8969 movl(Address(z, kdx, Address::times_4, 4), product); 8970 shrq(product, 32); 8971 movl(Address(z, kdx, Address::times_4, 0), product); 8972 movq(carry, rdx); 8973 jmp(L_first_loop); 8974 8975 bind(L_one_y); 8976 movl(y_idx, Address(y, 0)); 8977 jmp(L_multiply); 8978 8979 bind(L_one_x); 8980 movl(x_xstart, Address(x, 0)); 8981 jmp(L_first_loop); 8982 8983 bind(L_first_loop_exit); 8984 } 8985 8986 /** 8987 * Multiply 64 bit by 64 bit and add 128 bit. 8988 */ 8989 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y, Register z, 8990 Register yz_idx, Register idx, 8991 Register carry, Register product, int offset) { 8992 // huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry; 8993 // z[kdx] = (jlong)product; 8994 8995 movq(yz_idx, Address(y, idx, Address::times_4, offset)); 8996 rorq(yz_idx, 32); // convert big-endian to little-endian 8997 movq(product, x_xstart); 8998 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax) 8999 movq(yz_idx, Address(z, idx, Address::times_4, offset)); 9000 rorq(yz_idx, 32); // convert big-endian to little-endian 9001 9002 add2_with_carry(rdx, product, carry, yz_idx); 9003 9004 movl(Address(z, idx, Address::times_4, offset+4), product); 9005 shrq(product, 32); 9006 movl(Address(z, idx, Address::times_4, offset), product); 9007 9008 } 9009 9010 /** 9011 * Multiply 128 bit by 128 bit. Unrolled inner loop. 9012 */ 9013 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart, Register y, Register z, 9014 Register yz_idx, Register idx, Register jdx, 9015 Register carry, Register product, 9016 Register carry2) { 9017 // jlong carry, x[], y[], z[]; 9018 // int kdx = ystart+1; 9019 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop 9020 // huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry; 9021 // z[kdx+idx+1] = (jlong)product; 9022 // jlong carry2 = (jlong)(product >>> 64); 9023 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry2; 9024 // z[kdx+idx] = (jlong)product; 9025 // carry = (jlong)(product >>> 64); 9026 // } 9027 // idx += 2; 9028 // if (idx > 0) { 9029 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry; 9030 // z[kdx+idx] = (jlong)product; 9031 // carry = (jlong)(product >>> 64); 9032 // } 9033 // 9034 9035 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done; 9036 9037 movl(jdx, idx); 9038 andl(jdx, 0xFFFFFFFC); 9039 shrl(jdx, 2); 9040 9041 bind(L_third_loop); 9042 subl(jdx, 1); 9043 jcc(Assembler::negative, L_third_loop_exit); 9044 subl(idx, 4); 9045 9046 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8); 9047 movq(carry2, rdx); 9048 9049 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0); 9050 movq(carry, rdx); 9051 jmp(L_third_loop); 9052 9053 bind (L_third_loop_exit); 9054 9055 andl (idx, 0x3); 9056 jcc(Assembler::zero, L_post_third_loop_done); 9057 9058 Label L_check_1; 9059 subl(idx, 2); 9060 jcc(Assembler::negative, L_check_1); 9061 9062 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0); 9063 movq(carry, rdx); 9064 9065 bind (L_check_1); 9066 addl (idx, 0x2); 9067 andl (idx, 0x1); 9068 subl(idx, 1); 9069 jcc(Assembler::negative, L_post_third_loop_done); 9070 9071 movl(yz_idx, Address(y, idx, Address::times_4, 0)); 9072 movq(product, x_xstart); 9073 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax) 9074 movl(yz_idx, Address(z, idx, Address::times_4, 0)); 9075 9076 add2_with_carry(rdx, product, yz_idx, carry); 9077 9078 movl(Address(z, idx, Address::times_4, 0), product); 9079 shrq(product, 32); 9080 9081 shlq(rdx, 32); 9082 orq(product, rdx); 9083 movq(carry, product); 9084 9085 bind(L_post_third_loop_done); 9086 } 9087 9088 /** 9089 * Multiply 128 bit by 128 bit using BMI2. Unrolled inner loop. 9090 * 9091 */ 9092 void MacroAssembler::multiply_128_x_128_bmi2_loop(Register y, Register z, 9093 Register carry, Register carry2, 9094 Register idx, Register jdx, 9095 Register yz_idx1, Register yz_idx2, 9096 Register tmp, Register tmp3, Register tmp4) { 9097 assert(UseBMI2Instructions, "should be used only when BMI2 is available"); 9098 9099 // jlong carry, x[], y[], z[]; 9100 // int kdx = ystart+1; 9101 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop 9102 // huge_128 tmp3 = (y[idx+1] * rdx) + z[kdx+idx+1] + carry; 9103 // jlong carry2 = (jlong)(tmp3 >>> 64); 9104 // huge_128 tmp4 = (y[idx] * rdx) + z[kdx+idx] + carry2; 9105 // carry = (jlong)(tmp4 >>> 64); 9106 // z[kdx+idx+1] = (jlong)tmp3; 9107 // z[kdx+idx] = (jlong)tmp4; 9108 // } 9109 // idx += 2; 9110 // if (idx > 0) { 9111 // yz_idx1 = (y[idx] * rdx) + z[kdx+idx] + carry; 9112 // z[kdx+idx] = (jlong)yz_idx1; 9113 // carry = (jlong)(yz_idx1 >>> 64); 9114 // } 9115 // 9116 9117 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done; 9118 9119 movl(jdx, idx); 9120 andl(jdx, 0xFFFFFFFC); 9121 shrl(jdx, 2); 9122 9123 bind(L_third_loop); 9124 subl(jdx, 1); 9125 jcc(Assembler::negative, L_third_loop_exit); 9126 subl(idx, 4); 9127 9128 movq(yz_idx1, Address(y, idx, Address::times_4, 8)); 9129 rorxq(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian 9130 movq(yz_idx2, Address(y, idx, Address::times_4, 0)); 9131 rorxq(yz_idx2, yz_idx2, 32); 9132 9133 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3 9134 mulxq(carry2, tmp, yz_idx2); // yz_idx2 * rdx -> carry2:tmp 9135 9136 movq(yz_idx1, Address(z, idx, Address::times_4, 8)); 9137 rorxq(yz_idx1, yz_idx1, 32); 9138 movq(yz_idx2, Address(z, idx, Address::times_4, 0)); 9139 rorxq(yz_idx2, yz_idx2, 32); 9140 9141 if (VM_Version::supports_adx()) { 9142 adcxq(tmp3, carry); 9143 adoxq(tmp3, yz_idx1); 9144 9145 adcxq(tmp4, tmp); 9146 adoxq(tmp4, yz_idx2); 9147 9148 movl(carry, 0); // does not affect flags 9149 adcxq(carry2, carry); 9150 adoxq(carry2, carry); 9151 } else { 9152 add2_with_carry(tmp4, tmp3, carry, yz_idx1); 9153 add2_with_carry(carry2, tmp4, tmp, yz_idx2); 9154 } 9155 movq(carry, carry2); 9156 9157 movl(Address(z, idx, Address::times_4, 12), tmp3); 9158 shrq(tmp3, 32); 9159 movl(Address(z, idx, Address::times_4, 8), tmp3); 9160 9161 movl(Address(z, idx, Address::times_4, 4), tmp4); 9162 shrq(tmp4, 32); 9163 movl(Address(z, idx, Address::times_4, 0), tmp4); 9164 9165 jmp(L_third_loop); 9166 9167 bind (L_third_loop_exit); 9168 9169 andl (idx, 0x3); 9170 jcc(Assembler::zero, L_post_third_loop_done); 9171 9172 Label L_check_1; 9173 subl(idx, 2); 9174 jcc(Assembler::negative, L_check_1); 9175 9176 movq(yz_idx1, Address(y, idx, Address::times_4, 0)); 9177 rorxq(yz_idx1, yz_idx1, 32); 9178 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3 9179 movq(yz_idx2, Address(z, idx, Address::times_4, 0)); 9180 rorxq(yz_idx2, yz_idx2, 32); 9181 9182 add2_with_carry(tmp4, tmp3, carry, yz_idx2); 9183 9184 movl(Address(z, idx, Address::times_4, 4), tmp3); 9185 shrq(tmp3, 32); 9186 movl(Address(z, idx, Address::times_4, 0), tmp3); 9187 movq(carry, tmp4); 9188 9189 bind (L_check_1); 9190 addl (idx, 0x2); 9191 andl (idx, 0x1); 9192 subl(idx, 1); 9193 jcc(Assembler::negative, L_post_third_loop_done); 9194 movl(tmp4, Address(y, idx, Address::times_4, 0)); 9195 mulxq(carry2, tmp3, tmp4); // tmp4 * rdx -> carry2:tmp3 9196 movl(tmp4, Address(z, idx, Address::times_4, 0)); 9197 9198 add2_with_carry(carry2, tmp3, tmp4, carry); 9199 9200 movl(Address(z, idx, Address::times_4, 0), tmp3); 9201 shrq(tmp3, 32); 9202 9203 shlq(carry2, 32); 9204 orq(tmp3, carry2); 9205 movq(carry, tmp3); 9206 9207 bind(L_post_third_loop_done); 9208 } 9209 9210 /** 9211 * Code for BigInteger::multiplyToLen() instrinsic. 9212 * 9213 * rdi: x 9214 * rax: xlen 9215 * rsi: y 9216 * rcx: ylen 9217 * r8: z 9218 * r11: zlen 9219 * r12: tmp1 9220 * r13: tmp2 9221 * r14: tmp3 9222 * r15: tmp4 9223 * rbx: tmp5 9224 * 9225 */ 9226 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, Register z, Register zlen, 9227 Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5) { 9228 ShortBranchVerifier sbv(this); 9229 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx); 9230 9231 push(tmp1); 9232 push(tmp2); 9233 push(tmp3); 9234 push(tmp4); 9235 push(tmp5); 9236 9237 push(xlen); 9238 push(zlen); 9239 9240 const Register idx = tmp1; 9241 const Register kdx = tmp2; 9242 const Register xstart = tmp3; 9243 9244 const Register y_idx = tmp4; 9245 const Register carry = tmp5; 9246 const Register product = xlen; 9247 const Register x_xstart = zlen; // reuse register 9248 9249 // First Loop. 9250 // 9251 // final static long LONG_MASK = 0xffffffffL; 9252 // int xstart = xlen - 1; 9253 // int ystart = ylen - 1; 9254 // long carry = 0; 9255 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) { 9256 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry; 9257 // z[kdx] = (int)product; 9258 // carry = product >>> 32; 9259 // } 9260 // z[xstart] = (int)carry; 9261 // 9262 9263 movl(idx, ylen); // idx = ylen; 9264 movl(kdx, zlen); // kdx = xlen+ylen; 9265 xorq(carry, carry); // carry = 0; 9266 9267 Label L_done; 9268 9269 movl(xstart, xlen); 9270 decrementl(xstart); 9271 jcc(Assembler::negative, L_done); 9272 9273 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx); 9274 9275 Label L_second_loop; 9276 testl(kdx, kdx); 9277 jcc(Assembler::zero, L_second_loop); 9278 9279 Label L_carry; 9280 subl(kdx, 1); 9281 jcc(Assembler::zero, L_carry); 9282 9283 movl(Address(z, kdx, Address::times_4, 0), carry); 9284 shrq(carry, 32); 9285 subl(kdx, 1); 9286 9287 bind(L_carry); 9288 movl(Address(z, kdx, Address::times_4, 0), carry); 9289 9290 // Second and third (nested) loops. 9291 // 9292 // for (int i = xstart-1; i >= 0; i--) { // Second loop 9293 // carry = 0; 9294 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop 9295 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) + 9296 // (z[k] & LONG_MASK) + carry; 9297 // z[k] = (int)product; 9298 // carry = product >>> 32; 9299 // } 9300 // z[i] = (int)carry; 9301 // } 9302 // 9303 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx 9304 9305 const Register jdx = tmp1; 9306 9307 bind(L_second_loop); 9308 xorl(carry, carry); // carry = 0; 9309 movl(jdx, ylen); // j = ystart+1 9310 9311 subl(xstart, 1); // i = xstart-1; 9312 jcc(Assembler::negative, L_done); 9313 9314 push (z); 9315 9316 Label L_last_x; 9317 lea(z, Address(z, xstart, Address::times_4, 4)); // z = z + k - j 9318 subl(xstart, 1); // i = xstart-1; 9319 jcc(Assembler::negative, L_last_x); 9320 9321 if (UseBMI2Instructions) { 9322 movq(rdx, Address(x, xstart, Address::times_4, 0)); 9323 rorxq(rdx, rdx, 32); // convert big-endian to little-endian 9324 } else { 9325 movq(x_xstart, Address(x, xstart, Address::times_4, 0)); 9326 rorq(x_xstart, 32); // convert big-endian to little-endian 9327 } 9328 9329 Label L_third_loop_prologue; 9330 bind(L_third_loop_prologue); 9331 9332 push (x); 9333 push (xstart); 9334 push (ylen); 9335 9336 9337 if (UseBMI2Instructions) { 9338 multiply_128_x_128_bmi2_loop(y, z, carry, x, jdx, ylen, product, tmp2, x_xstart, tmp3, tmp4); 9339 } else { // !UseBMI2Instructions 9340 multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x); 9341 } 9342 9343 pop(ylen); 9344 pop(xlen); 9345 pop(x); 9346 pop(z); 9347 9348 movl(tmp3, xlen); 9349 addl(tmp3, 1); 9350 movl(Address(z, tmp3, Address::times_4, 0), carry); 9351 subl(tmp3, 1); 9352 jccb(Assembler::negative, L_done); 9353 9354 shrq(carry, 32); 9355 movl(Address(z, tmp3, Address::times_4, 0), carry); 9356 jmp(L_second_loop); 9357 9358 // Next infrequent code is moved outside loops. 9359 bind(L_last_x); 9360 if (UseBMI2Instructions) { 9361 movl(rdx, Address(x, 0)); 9362 } else { 9363 movl(x_xstart, Address(x, 0)); 9364 } 9365 jmp(L_third_loop_prologue); 9366 9367 bind(L_done); 9368 9369 pop(zlen); 9370 pop(xlen); 9371 9372 pop(tmp5); 9373 pop(tmp4); 9374 pop(tmp3); 9375 pop(tmp2); 9376 pop(tmp1); 9377 } 9378 9379 void MacroAssembler::vectorized_mismatch(Register obja, Register objb, Register length, Register log2_array_indxscale, 9380 Register result, Register tmp1, Register tmp2, XMMRegister rymm0, XMMRegister rymm1, XMMRegister rymm2){ 9381 assert(UseSSE42Intrinsics, "SSE4.2 must be enabled."); 9382 Label VECTOR64_LOOP, VECTOR64_TAIL, VECTOR64_NOT_EQUAL, VECTOR32_TAIL; 9383 Label VECTOR32_LOOP, VECTOR16_LOOP, VECTOR8_LOOP, VECTOR4_LOOP; 9384 Label VECTOR16_TAIL, VECTOR8_TAIL, VECTOR4_TAIL; 9385 Label VECTOR32_NOT_EQUAL, VECTOR16_NOT_EQUAL, VECTOR8_NOT_EQUAL, VECTOR4_NOT_EQUAL; 9386 Label SAME_TILL_END, DONE; 9387 Label BYTES_LOOP, BYTES_TAIL, BYTES_NOT_EQUAL; 9388 9389 //scale is in rcx in both Win64 and Unix 9390 ShortBranchVerifier sbv(this); 9391 9392 shlq(length); 9393 xorq(result, result); 9394 9395 if ((UseAVX > 2) && 9396 VM_Version::supports_avx512vlbw()) { 9397 set_vector_masking(); // opening of the stub context for programming mask registers 9398 cmpq(length, 64); 9399 jcc(Assembler::less, VECTOR32_TAIL); 9400 movq(tmp1, length); 9401 andq(tmp1, 0x3F); // tail count 9402 andq(length, ~(0x3F)); //vector count 9403 9404 bind(VECTOR64_LOOP); 9405 // AVX512 code to compare 64 byte vectors. 9406 evmovdqub(rymm0, Address(obja, result), Assembler::AVX_512bit); 9407 evpcmpeqb(k7, rymm0, Address(objb, result), Assembler::AVX_512bit); 9408 kortestql(k7, k7); 9409 jcc(Assembler::aboveEqual, VECTOR64_NOT_EQUAL); // mismatch 9410 addq(result, 64); 9411 subq(length, 64); 9412 jccb(Assembler::notZero, VECTOR64_LOOP); 9413 9414 //bind(VECTOR64_TAIL); 9415 testq(tmp1, tmp1); 9416 jcc(Assembler::zero, SAME_TILL_END); 9417 9418 bind(VECTOR64_TAIL); 9419 // AVX512 code to compare upto 63 byte vectors. 9420 // Save k1 9421 kmovql(k3, k1); 9422 mov64(tmp2, 0xFFFFFFFFFFFFFFFF); 9423 shlxq(tmp2, tmp2, tmp1); 9424 notq(tmp2); 9425 kmovql(k1, tmp2); 9426 9427 evmovdqub(rymm0, k1, Address(obja, result), Assembler::AVX_512bit); 9428 evpcmpeqb(k7, k1, rymm0, Address(objb, result), Assembler::AVX_512bit); 9429 9430 ktestql(k7, k1); 9431 // Restore k1 9432 kmovql(k1, k3); 9433 jcc(Assembler::below, SAME_TILL_END); // not mismatch 9434 9435 bind(VECTOR64_NOT_EQUAL); 9436 kmovql(tmp1, k7); 9437 notq(tmp1); 9438 tzcntq(tmp1, tmp1); 9439 addq(result, tmp1); 9440 shrq(result); 9441 jmp(DONE); 9442 bind(VECTOR32_TAIL); 9443 clear_vector_masking(); // closing of the stub context for programming mask registers 9444 } 9445 9446 cmpq(length, 8); 9447 jcc(Assembler::equal, VECTOR8_LOOP); 9448 jcc(Assembler::less, VECTOR4_TAIL); 9449 9450 if (UseAVX >= 2) { 9451 9452 cmpq(length, 16); 9453 jcc(Assembler::equal, VECTOR16_LOOP); 9454 jcc(Assembler::less, VECTOR8_LOOP); 9455 9456 cmpq(length, 32); 9457 jccb(Assembler::less, VECTOR16_TAIL); 9458 9459 subq(length, 32); 9460 bind(VECTOR32_LOOP); 9461 vmovdqu(rymm0, Address(obja, result)); 9462 vmovdqu(rymm1, Address(objb, result)); 9463 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_256bit); 9464 vptest(rymm2, rymm2); 9465 jcc(Assembler::notZero, VECTOR32_NOT_EQUAL);//mismatch found 9466 addq(result, 32); 9467 subq(length, 32); 9468 jccb(Assembler::greaterEqual, VECTOR32_LOOP); 9469 addq(length, 32); 9470 jcc(Assembler::equal, SAME_TILL_END); 9471 //falling through if less than 32 bytes left //close the branch here. 9472 9473 bind(VECTOR16_TAIL); 9474 cmpq(length, 16); 9475 jccb(Assembler::less, VECTOR8_TAIL); 9476 bind(VECTOR16_LOOP); 9477 movdqu(rymm0, Address(obja, result)); 9478 movdqu(rymm1, Address(objb, result)); 9479 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_128bit); 9480 ptest(rymm2, rymm2); 9481 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found 9482 addq(result, 16); 9483 subq(length, 16); 9484 jcc(Assembler::equal, SAME_TILL_END); 9485 //falling through if less than 16 bytes left 9486 } else {//regular intrinsics 9487 9488 cmpq(length, 16); 9489 jccb(Assembler::less, VECTOR8_TAIL); 9490 9491 subq(length, 16); 9492 bind(VECTOR16_LOOP); 9493 movdqu(rymm0, Address(obja, result)); 9494 movdqu(rymm1, Address(objb, result)); 9495 pxor(rymm0, rymm1); 9496 ptest(rymm0, rymm0); 9497 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found 9498 addq(result, 16); 9499 subq(length, 16); 9500 jccb(Assembler::greaterEqual, VECTOR16_LOOP); 9501 addq(length, 16); 9502 jcc(Assembler::equal, SAME_TILL_END); 9503 //falling through if less than 16 bytes left 9504 } 9505 9506 bind(VECTOR8_TAIL); 9507 cmpq(length, 8); 9508 jccb(Assembler::less, VECTOR4_TAIL); 9509 bind(VECTOR8_LOOP); 9510 movq(tmp1, Address(obja, result)); 9511 movq(tmp2, Address(objb, result)); 9512 xorq(tmp1, tmp2); 9513 testq(tmp1, tmp1); 9514 jcc(Assembler::notZero, VECTOR8_NOT_EQUAL);//mismatch found 9515 addq(result, 8); 9516 subq(length, 8); 9517 jcc(Assembler::equal, SAME_TILL_END); 9518 //falling through if less than 8 bytes left 9519 9520 bind(VECTOR4_TAIL); 9521 cmpq(length, 4); 9522 jccb(Assembler::less, BYTES_TAIL); 9523 bind(VECTOR4_LOOP); 9524 movl(tmp1, Address(obja, result)); 9525 xorl(tmp1, Address(objb, result)); 9526 testl(tmp1, tmp1); 9527 jcc(Assembler::notZero, VECTOR4_NOT_EQUAL);//mismatch found 9528 addq(result, 4); 9529 subq(length, 4); 9530 jcc(Assembler::equal, SAME_TILL_END); 9531 //falling through if less than 4 bytes left 9532 9533 bind(BYTES_TAIL); 9534 bind(BYTES_LOOP); 9535 load_unsigned_byte(tmp1, Address(obja, result)); 9536 load_unsigned_byte(tmp2, Address(objb, result)); 9537 xorl(tmp1, tmp2); 9538 testl(tmp1, tmp1); 9539 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 9540 decq(length); 9541 jccb(Assembler::zero, SAME_TILL_END); 9542 incq(result); 9543 load_unsigned_byte(tmp1, Address(obja, result)); 9544 load_unsigned_byte(tmp2, Address(objb, result)); 9545 xorl(tmp1, tmp2); 9546 testl(tmp1, tmp1); 9547 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 9548 decq(length); 9549 jccb(Assembler::zero, SAME_TILL_END); 9550 incq(result); 9551 load_unsigned_byte(tmp1, Address(obja, result)); 9552 load_unsigned_byte(tmp2, Address(objb, result)); 9553 xorl(tmp1, tmp2); 9554 testl(tmp1, tmp1); 9555 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 9556 jmpb(SAME_TILL_END); 9557 9558 if (UseAVX >= 2) { 9559 bind(VECTOR32_NOT_EQUAL); 9560 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_256bit); 9561 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_256bit); 9562 vpxor(rymm0, rymm0, rymm2, Assembler::AVX_256bit); 9563 vpmovmskb(tmp1, rymm0); 9564 bsfq(tmp1, tmp1); 9565 addq(result, tmp1); 9566 shrq(result); 9567 jmpb(DONE); 9568 } 9569 9570 bind(VECTOR16_NOT_EQUAL); 9571 if (UseAVX >= 2) { 9572 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_128bit); 9573 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_128bit); 9574 pxor(rymm0, rymm2); 9575 } else { 9576 pcmpeqb(rymm2, rymm2); 9577 pxor(rymm0, rymm1); 9578 pcmpeqb(rymm0, rymm1); 9579 pxor(rymm0, rymm2); 9580 } 9581 pmovmskb(tmp1, rymm0); 9582 bsfq(tmp1, tmp1); 9583 addq(result, tmp1); 9584 shrq(result); 9585 jmpb(DONE); 9586 9587 bind(VECTOR8_NOT_EQUAL); 9588 bind(VECTOR4_NOT_EQUAL); 9589 bsfq(tmp1, tmp1); 9590 shrq(tmp1, 3); 9591 addq(result, tmp1); 9592 bind(BYTES_NOT_EQUAL); 9593 shrq(result); 9594 jmpb(DONE); 9595 9596 bind(SAME_TILL_END); 9597 mov64(result, -1); 9598 9599 bind(DONE); 9600 } 9601 9602 //Helper functions for square_to_len() 9603 9604 /** 9605 * Store the squares of x[], right shifted one bit (divided by 2) into z[] 9606 * Preserves x and z and modifies rest of the registers. 9607 */ 9608 void MacroAssembler::square_rshift(Register x, Register xlen, Register z, Register tmp1, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 9609 // Perform square and right shift by 1 9610 // Handle odd xlen case first, then for even xlen do the following 9611 // jlong carry = 0; 9612 // for (int j=0, i=0; j < xlen; j+=2, i+=4) { 9613 // huge_128 product = x[j:j+1] * x[j:j+1]; 9614 // z[i:i+1] = (carry << 63) | (jlong)(product >>> 65); 9615 // z[i+2:i+3] = (jlong)(product >>> 1); 9616 // carry = (jlong)product; 9617 // } 9618 9619 xorq(tmp5, tmp5); // carry 9620 xorq(rdxReg, rdxReg); 9621 xorl(tmp1, tmp1); // index for x 9622 xorl(tmp4, tmp4); // index for z 9623 9624 Label L_first_loop, L_first_loop_exit; 9625 9626 testl(xlen, 1); 9627 jccb(Assembler::zero, L_first_loop); //jump if xlen is even 9628 9629 // Square and right shift by 1 the odd element using 32 bit multiply 9630 movl(raxReg, Address(x, tmp1, Address::times_4, 0)); 9631 imulq(raxReg, raxReg); 9632 shrq(raxReg, 1); 9633 adcq(tmp5, 0); 9634 movq(Address(z, tmp4, Address::times_4, 0), raxReg); 9635 incrementl(tmp1); 9636 addl(tmp4, 2); 9637 9638 // Square and right shift by 1 the rest using 64 bit multiply 9639 bind(L_first_loop); 9640 cmpptr(tmp1, xlen); 9641 jccb(Assembler::equal, L_first_loop_exit); 9642 9643 // Square 9644 movq(raxReg, Address(x, tmp1, Address::times_4, 0)); 9645 rorq(raxReg, 32); // convert big-endian to little-endian 9646 mulq(raxReg); // 64-bit multiply rax * rax -> rdx:rax 9647 9648 // Right shift by 1 and save carry 9649 shrq(tmp5, 1); // rdx:rax:tmp5 = (tmp5:rdx:rax) >>> 1 9650 rcrq(rdxReg, 1); 9651 rcrq(raxReg, 1); 9652 adcq(tmp5, 0); 9653 9654 // Store result in z 9655 movq(Address(z, tmp4, Address::times_4, 0), rdxReg); 9656 movq(Address(z, tmp4, Address::times_4, 8), raxReg); 9657 9658 // Update indices for x and z 9659 addl(tmp1, 2); 9660 addl(tmp4, 4); 9661 jmp(L_first_loop); 9662 9663 bind(L_first_loop_exit); 9664 } 9665 9666 9667 /** 9668 * Perform the following multiply add operation using BMI2 instructions 9669 * carry:sum = sum + op1*op2 + carry 9670 * op2 should be in rdx 9671 * op2 is preserved, all other registers are modified 9672 */ 9673 void MacroAssembler::multiply_add_64_bmi2(Register sum, Register op1, Register op2, Register carry, Register tmp2) { 9674 // assert op2 is rdx 9675 mulxq(tmp2, op1, op1); // op1 * op2 -> tmp2:op1 9676 addq(sum, carry); 9677 adcq(tmp2, 0); 9678 addq(sum, op1); 9679 adcq(tmp2, 0); 9680 movq(carry, tmp2); 9681 } 9682 9683 /** 9684 * Perform the following multiply add operation: 9685 * carry:sum = sum + op1*op2 + carry 9686 * Preserves op1, op2 and modifies rest of registers 9687 */ 9688 void MacroAssembler::multiply_add_64(Register sum, Register op1, Register op2, Register carry, Register rdxReg, Register raxReg) { 9689 // rdx:rax = op1 * op2 9690 movq(raxReg, op2); 9691 mulq(op1); 9692 9693 // rdx:rax = sum + carry + rdx:rax 9694 addq(sum, carry); 9695 adcq(rdxReg, 0); 9696 addq(sum, raxReg); 9697 adcq(rdxReg, 0); 9698 9699 // carry:sum = rdx:sum 9700 movq(carry, rdxReg); 9701 } 9702 9703 /** 9704 * Add 64 bit long carry into z[] with carry propogation. 9705 * Preserves z and carry register values and modifies rest of registers. 9706 * 9707 */ 9708 void MacroAssembler::add_one_64(Register z, Register zlen, Register carry, Register tmp1) { 9709 Label L_fourth_loop, L_fourth_loop_exit; 9710 9711 movl(tmp1, 1); 9712 subl(zlen, 2); 9713 addq(Address(z, zlen, Address::times_4, 0), carry); 9714 9715 bind(L_fourth_loop); 9716 jccb(Assembler::carryClear, L_fourth_loop_exit); 9717 subl(zlen, 2); 9718 jccb(Assembler::negative, L_fourth_loop_exit); 9719 addq(Address(z, zlen, Address::times_4, 0), tmp1); 9720 jmp(L_fourth_loop); 9721 bind(L_fourth_loop_exit); 9722 } 9723 9724 /** 9725 * Shift z[] left by 1 bit. 9726 * Preserves x, len, z and zlen registers and modifies rest of the registers. 9727 * 9728 */ 9729 void MacroAssembler::lshift_by_1(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4) { 9730 9731 Label L_fifth_loop, L_fifth_loop_exit; 9732 9733 // Fifth loop 9734 // Perform primitiveLeftShift(z, zlen, 1) 9735 9736 const Register prev_carry = tmp1; 9737 const Register new_carry = tmp4; 9738 const Register value = tmp2; 9739 const Register zidx = tmp3; 9740 9741 // int zidx, carry; 9742 // long value; 9743 // carry = 0; 9744 // for (zidx = zlen-2; zidx >=0; zidx -= 2) { 9745 // (carry:value) = (z[i] << 1) | carry ; 9746 // z[i] = value; 9747 // } 9748 9749 movl(zidx, zlen); 9750 xorl(prev_carry, prev_carry); // clear carry flag and prev_carry register 9751 9752 bind(L_fifth_loop); 9753 decl(zidx); // Use decl to preserve carry flag 9754 decl(zidx); 9755 jccb(Assembler::negative, L_fifth_loop_exit); 9756 9757 if (UseBMI2Instructions) { 9758 movq(value, Address(z, zidx, Address::times_4, 0)); 9759 rclq(value, 1); 9760 rorxq(value, value, 32); 9761 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form 9762 } 9763 else { 9764 // clear new_carry 9765 xorl(new_carry, new_carry); 9766 9767 // Shift z[i] by 1, or in previous carry and save new carry 9768 movq(value, Address(z, zidx, Address::times_4, 0)); 9769 shlq(value, 1); 9770 adcl(new_carry, 0); 9771 9772 orq(value, prev_carry); 9773 rorq(value, 0x20); 9774 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form 9775 9776 // Set previous carry = new carry 9777 movl(prev_carry, new_carry); 9778 } 9779 jmp(L_fifth_loop); 9780 9781 bind(L_fifth_loop_exit); 9782 } 9783 9784 9785 /** 9786 * Code for BigInteger::squareToLen() intrinsic 9787 * 9788 * rdi: x 9789 * rsi: len 9790 * r8: z 9791 * rcx: zlen 9792 * r12: tmp1 9793 * r13: tmp2 9794 * r14: tmp3 9795 * r15: tmp4 9796 * rbx: tmp5 9797 * 9798 */ 9799 void MacroAssembler::square_to_len(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 9800 9801 Label L_second_loop, L_second_loop_exit, L_third_loop, L_third_loop_exit, fifth_loop, fifth_loop_exit, L_last_x, L_multiply; 9802 push(tmp1); 9803 push(tmp2); 9804 push(tmp3); 9805 push(tmp4); 9806 push(tmp5); 9807 9808 // First loop 9809 // Store the squares, right shifted one bit (i.e., divided by 2). 9810 square_rshift(x, len, z, tmp1, tmp3, tmp4, tmp5, rdxReg, raxReg); 9811 9812 // Add in off-diagonal sums. 9813 // 9814 // Second, third (nested) and fourth loops. 9815 // zlen +=2; 9816 // for (int xidx=len-2,zidx=zlen-4; xidx > 0; xidx-=2,zidx-=4) { 9817 // carry = 0; 9818 // long op2 = x[xidx:xidx+1]; 9819 // for (int j=xidx-2,k=zidx; j >= 0; j-=2) { 9820 // k -= 2; 9821 // long op1 = x[j:j+1]; 9822 // long sum = z[k:k+1]; 9823 // carry:sum = multiply_add_64(sum, op1, op2, carry, tmp_regs); 9824 // z[k:k+1] = sum; 9825 // } 9826 // add_one_64(z, k, carry, tmp_regs); 9827 // } 9828 9829 const Register carry = tmp5; 9830 const Register sum = tmp3; 9831 const Register op1 = tmp4; 9832 Register op2 = tmp2; 9833 9834 push(zlen); 9835 push(len); 9836 addl(zlen,2); 9837 bind(L_second_loop); 9838 xorq(carry, carry); 9839 subl(zlen, 4); 9840 subl(len, 2); 9841 push(zlen); 9842 push(len); 9843 cmpl(len, 0); 9844 jccb(Assembler::lessEqual, L_second_loop_exit); 9845 9846 // Multiply an array by one 64 bit long. 9847 if (UseBMI2Instructions) { 9848 op2 = rdxReg; 9849 movq(op2, Address(x, len, Address::times_4, 0)); 9850 rorxq(op2, op2, 32); 9851 } 9852 else { 9853 movq(op2, Address(x, len, Address::times_4, 0)); 9854 rorq(op2, 32); 9855 } 9856 9857 bind(L_third_loop); 9858 decrementl(len); 9859 jccb(Assembler::negative, L_third_loop_exit); 9860 decrementl(len); 9861 jccb(Assembler::negative, L_last_x); 9862 9863 movq(op1, Address(x, len, Address::times_4, 0)); 9864 rorq(op1, 32); 9865 9866 bind(L_multiply); 9867 subl(zlen, 2); 9868 movq(sum, Address(z, zlen, Address::times_4, 0)); 9869 9870 // Multiply 64 bit by 64 bit and add 64 bits lower half and upper 64 bits as carry. 9871 if (UseBMI2Instructions) { 9872 multiply_add_64_bmi2(sum, op1, op2, carry, tmp2); 9873 } 9874 else { 9875 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9876 } 9877 9878 movq(Address(z, zlen, Address::times_4, 0), sum); 9879 9880 jmp(L_third_loop); 9881 bind(L_third_loop_exit); 9882 9883 // Fourth loop 9884 // Add 64 bit long carry into z with carry propogation. 9885 // Uses offsetted zlen. 9886 add_one_64(z, zlen, carry, tmp1); 9887 9888 pop(len); 9889 pop(zlen); 9890 jmp(L_second_loop); 9891 9892 // Next infrequent code is moved outside loops. 9893 bind(L_last_x); 9894 movl(op1, Address(x, 0)); 9895 jmp(L_multiply); 9896 9897 bind(L_second_loop_exit); 9898 pop(len); 9899 pop(zlen); 9900 pop(len); 9901 pop(zlen); 9902 9903 // Fifth loop 9904 // Shift z left 1 bit. 9905 lshift_by_1(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4); 9906 9907 // z[zlen-1] |= x[len-1] & 1; 9908 movl(tmp3, Address(x, len, Address::times_4, -4)); 9909 andl(tmp3, 1); 9910 orl(Address(z, zlen, Address::times_4, -4), tmp3); 9911 9912 pop(tmp5); 9913 pop(tmp4); 9914 pop(tmp3); 9915 pop(tmp2); 9916 pop(tmp1); 9917 } 9918 9919 /** 9920 * Helper function for mul_add() 9921 * Multiply the in[] by int k and add to out[] starting at offset offs using 9922 * 128 bit by 32 bit multiply and return the carry in tmp5. 9923 * Only quad int aligned length of in[] is operated on in this function. 9924 * k is in rdxReg for BMI2Instructions, for others it is in tmp2. 9925 * This function preserves out, in and k registers. 9926 * len and offset point to the appropriate index in "in" & "out" correspondingly 9927 * tmp5 has the carry. 9928 * other registers are temporary and are modified. 9929 * 9930 */ 9931 void MacroAssembler::mul_add_128_x_32_loop(Register out, Register in, 9932 Register offset, Register len, Register tmp1, Register tmp2, Register tmp3, 9933 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 9934 9935 Label L_first_loop, L_first_loop_exit; 9936 9937 movl(tmp1, len); 9938 shrl(tmp1, 2); 9939 9940 bind(L_first_loop); 9941 subl(tmp1, 1); 9942 jccb(Assembler::negative, L_first_loop_exit); 9943 9944 subl(len, 4); 9945 subl(offset, 4); 9946 9947 Register op2 = tmp2; 9948 const Register sum = tmp3; 9949 const Register op1 = tmp4; 9950 const Register carry = tmp5; 9951 9952 if (UseBMI2Instructions) { 9953 op2 = rdxReg; 9954 } 9955 9956 movq(op1, Address(in, len, Address::times_4, 8)); 9957 rorq(op1, 32); 9958 movq(sum, Address(out, offset, Address::times_4, 8)); 9959 rorq(sum, 32); 9960 if (UseBMI2Instructions) { 9961 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 9962 } 9963 else { 9964 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9965 } 9966 // Store back in big endian from little endian 9967 rorq(sum, 0x20); 9968 movq(Address(out, offset, Address::times_4, 8), sum); 9969 9970 movq(op1, Address(in, len, Address::times_4, 0)); 9971 rorq(op1, 32); 9972 movq(sum, Address(out, offset, Address::times_4, 0)); 9973 rorq(sum, 32); 9974 if (UseBMI2Instructions) { 9975 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 9976 } 9977 else { 9978 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9979 } 9980 // Store back in big endian from little endian 9981 rorq(sum, 0x20); 9982 movq(Address(out, offset, Address::times_4, 0), sum); 9983 9984 jmp(L_first_loop); 9985 bind(L_first_loop_exit); 9986 } 9987 9988 /** 9989 * Code for BigInteger::mulAdd() intrinsic 9990 * 9991 * rdi: out 9992 * rsi: in 9993 * r11: offs (out.length - offset) 9994 * rcx: len 9995 * r8: k 9996 * r12: tmp1 9997 * r13: tmp2 9998 * r14: tmp3 9999 * r15: tmp4 10000 * rbx: tmp5 10001 * Multiply the in[] by word k and add to out[], return the carry in rax 10002 */ 10003 void MacroAssembler::mul_add(Register out, Register in, Register offs, 10004 Register len, Register k, Register tmp1, Register tmp2, Register tmp3, 10005 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 10006 10007 Label L_carry, L_last_in, L_done; 10008 10009 // carry = 0; 10010 // for (int j=len-1; j >= 0; j--) { 10011 // long product = (in[j] & LONG_MASK) * kLong + 10012 // (out[offs] & LONG_MASK) + carry; 10013 // out[offs--] = (int)product; 10014 // carry = product >>> 32; 10015 // } 10016 // 10017 push(tmp1); 10018 push(tmp2); 10019 push(tmp3); 10020 push(tmp4); 10021 push(tmp5); 10022 10023 Register op2 = tmp2; 10024 const Register sum = tmp3; 10025 const Register op1 = tmp4; 10026 const Register carry = tmp5; 10027 10028 if (UseBMI2Instructions) { 10029 op2 = rdxReg; 10030 movl(op2, k); 10031 } 10032 else { 10033 movl(op2, k); 10034 } 10035 10036 xorq(carry, carry); 10037 10038 //First loop 10039 10040 //Multiply in[] by k in a 4 way unrolled loop using 128 bit by 32 bit multiply 10041 //The carry is in tmp5 10042 mul_add_128_x_32_loop(out, in, offs, len, tmp1, tmp2, tmp3, tmp4, tmp5, rdxReg, raxReg); 10043 10044 //Multiply the trailing in[] entry using 64 bit by 32 bit, if any 10045 decrementl(len); 10046 jccb(Assembler::negative, L_carry); 10047 decrementl(len); 10048 jccb(Assembler::negative, L_last_in); 10049 10050 movq(op1, Address(in, len, Address::times_4, 0)); 10051 rorq(op1, 32); 10052 10053 subl(offs, 2); 10054 movq(sum, Address(out, offs, Address::times_4, 0)); 10055 rorq(sum, 32); 10056 10057 if (UseBMI2Instructions) { 10058 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 10059 } 10060 else { 10061 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 10062 } 10063 10064 // Store back in big endian from little endian 10065 rorq(sum, 0x20); 10066 movq(Address(out, offs, Address::times_4, 0), sum); 10067 10068 testl(len, len); 10069 jccb(Assembler::zero, L_carry); 10070 10071 //Multiply the last in[] entry, if any 10072 bind(L_last_in); 10073 movl(op1, Address(in, 0)); 10074 movl(sum, Address(out, offs, Address::times_4, -4)); 10075 10076 movl(raxReg, k); 10077 mull(op1); //tmp4 * eax -> edx:eax 10078 addl(sum, carry); 10079 adcl(rdxReg, 0); 10080 addl(sum, raxReg); 10081 adcl(rdxReg, 0); 10082 movl(carry, rdxReg); 10083 10084 movl(Address(out, offs, Address::times_4, -4), sum); 10085 10086 bind(L_carry); 10087 //return tmp5/carry as carry in rax 10088 movl(rax, carry); 10089 10090 bind(L_done); 10091 pop(tmp5); 10092 pop(tmp4); 10093 pop(tmp3); 10094 pop(tmp2); 10095 pop(tmp1); 10096 } 10097 #endif 10098 10099 /** 10100 * Emits code to update CRC-32 with a byte value according to constants in table 10101 * 10102 * @param [in,out]crc Register containing the crc. 10103 * @param [in]val Register containing the byte to fold into the CRC. 10104 * @param [in]table Register containing the table of crc constants. 10105 * 10106 * uint32_t crc; 10107 * val = crc_table[(val ^ crc) & 0xFF]; 10108 * crc = val ^ (crc >> 8); 10109 * 10110 */ 10111 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) { 10112 xorl(val, crc); 10113 andl(val, 0xFF); 10114 shrl(crc, 8); // unsigned shift 10115 xorl(crc, Address(table, val, Address::times_4, 0)); 10116 } 10117 10118 /** 10119 * Fold 128-bit data chunk 10120 */ 10121 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) { 10122 if (UseAVX > 0) { 10123 vpclmulhdq(xtmp, xK, xcrc); // [123:64] 10124 vpclmulldq(xcrc, xK, xcrc); // [63:0] 10125 vpxor(xcrc, xcrc, Address(buf, offset), 0 /* vector_len */); 10126 pxor(xcrc, xtmp); 10127 } else { 10128 movdqa(xtmp, xcrc); 10129 pclmulhdq(xtmp, xK); // [123:64] 10130 pclmulldq(xcrc, xK); // [63:0] 10131 pxor(xcrc, xtmp); 10132 movdqu(xtmp, Address(buf, offset)); 10133 pxor(xcrc, xtmp); 10134 } 10135 } 10136 10137 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) { 10138 if (UseAVX > 0) { 10139 vpclmulhdq(xtmp, xK, xcrc); 10140 vpclmulldq(xcrc, xK, xcrc); 10141 pxor(xcrc, xbuf); 10142 pxor(xcrc, xtmp); 10143 } else { 10144 movdqa(xtmp, xcrc); 10145 pclmulhdq(xtmp, xK); 10146 pclmulldq(xcrc, xK); 10147 pxor(xcrc, xbuf); 10148 pxor(xcrc, xtmp); 10149 } 10150 } 10151 10152 /** 10153 * 8-bit folds to compute 32-bit CRC 10154 * 10155 * uint64_t xcrc; 10156 * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8); 10157 */ 10158 void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) { 10159 movdl(tmp, xcrc); 10160 andl(tmp, 0xFF); 10161 movdl(xtmp, Address(table, tmp, Address::times_4, 0)); 10162 psrldq(xcrc, 1); // unsigned shift one byte 10163 pxor(xcrc, xtmp); 10164 } 10165 10166 /** 10167 * uint32_t crc; 10168 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8); 10169 */ 10170 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) { 10171 movl(tmp, crc); 10172 andl(tmp, 0xFF); 10173 shrl(crc, 8); 10174 xorl(crc, Address(table, tmp, Address::times_4, 0)); 10175 } 10176 10177 /** 10178 * @param crc register containing existing CRC (32-bit) 10179 * @param buf register pointing to input byte buffer (byte*) 10180 * @param len register containing number of bytes 10181 * @param table register that will contain address of CRC table 10182 * @param tmp scratch register 10183 */ 10184 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) { 10185 assert_different_registers(crc, buf, len, table, tmp, rax); 10186 10187 Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned; 10188 Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop; 10189 10190 // For EVEX with VL and BW, provide a standard mask, VL = 128 will guide the merge 10191 // context for the registers used, where all instructions below are using 128-bit mode 10192 // On EVEX without VL and BW, these instructions will all be AVX. 10193 if (VM_Version::supports_avx512vlbw()) { 10194 movl(tmp, 0xffff); 10195 kmovwl(k1, tmp); 10196 } 10197 10198 lea(table, ExternalAddress(StubRoutines::crc_table_addr())); 10199 notl(crc); // ~crc 10200 cmpl(len, 16); 10201 jcc(Assembler::less, L_tail); 10202 10203 // Align buffer to 16 bytes 10204 movl(tmp, buf); 10205 andl(tmp, 0xF); 10206 jccb(Assembler::zero, L_aligned); 10207 subl(tmp, 16); 10208 addl(len, tmp); 10209 10210 align(4); 10211 BIND(L_align_loop); 10212 movsbl(rax, Address(buf, 0)); // load byte with sign extension 10213 update_byte_crc32(crc, rax, table); 10214 increment(buf); 10215 incrementl(tmp); 10216 jccb(Assembler::less, L_align_loop); 10217 10218 BIND(L_aligned); 10219 movl(tmp, len); // save 10220 shrl(len, 4); 10221 jcc(Assembler::zero, L_tail_restore); 10222 10223 // Fold crc into first bytes of vector 10224 movdqa(xmm1, Address(buf, 0)); 10225 movdl(rax, xmm1); 10226 xorl(crc, rax); 10227 if (VM_Version::supports_sse4_1()) { 10228 pinsrd(xmm1, crc, 0); 10229 } else { 10230 pinsrw(xmm1, crc, 0); 10231 shrl(crc, 16); 10232 pinsrw(xmm1, crc, 1); 10233 } 10234 addptr(buf, 16); 10235 subl(len, 4); // len > 0 10236 jcc(Assembler::less, L_fold_tail); 10237 10238 movdqa(xmm2, Address(buf, 0)); 10239 movdqa(xmm3, Address(buf, 16)); 10240 movdqa(xmm4, Address(buf, 32)); 10241 addptr(buf, 48); 10242 subl(len, 3); 10243 jcc(Assembler::lessEqual, L_fold_512b); 10244 10245 // Fold total 512 bits of polynomial on each iteration, 10246 // 128 bits per each of 4 parallel streams. 10247 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32)); 10248 10249 align(32); 10250 BIND(L_fold_512b_loop); 10251 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0); 10252 fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16); 10253 fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32); 10254 fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48); 10255 addptr(buf, 64); 10256 subl(len, 4); 10257 jcc(Assembler::greater, L_fold_512b_loop); 10258 10259 // Fold 512 bits to 128 bits. 10260 BIND(L_fold_512b); 10261 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16)); 10262 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2); 10263 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3); 10264 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4); 10265 10266 // Fold the rest of 128 bits data chunks 10267 BIND(L_fold_tail); 10268 addl(len, 3); 10269 jccb(Assembler::lessEqual, L_fold_128b); 10270 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16)); 10271 10272 BIND(L_fold_tail_loop); 10273 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0); 10274 addptr(buf, 16); 10275 decrementl(len); 10276 jccb(Assembler::greater, L_fold_tail_loop); 10277 10278 // Fold 128 bits in xmm1 down into 32 bits in crc register. 10279 BIND(L_fold_128b); 10280 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr())); 10281 if (UseAVX > 0) { 10282 vpclmulqdq(xmm2, xmm0, xmm1, 0x1); 10283 vpand(xmm3, xmm0, xmm2, 0 /* vector_len */); 10284 vpclmulqdq(xmm0, xmm0, xmm3, 0x1); 10285 } else { 10286 movdqa(xmm2, xmm0); 10287 pclmulqdq(xmm2, xmm1, 0x1); 10288 movdqa(xmm3, xmm0); 10289 pand(xmm3, xmm2); 10290 pclmulqdq(xmm0, xmm3, 0x1); 10291 } 10292 psrldq(xmm1, 8); 10293 psrldq(xmm2, 4); 10294 pxor(xmm0, xmm1); 10295 pxor(xmm0, xmm2); 10296 10297 // 8 8-bit folds to compute 32-bit CRC. 10298 for (int j = 0; j < 4; j++) { 10299 fold_8bit_crc32(xmm0, table, xmm1, rax); 10300 } 10301 movdl(crc, xmm0); // mov 32 bits to general register 10302 for (int j = 0; j < 4; j++) { 10303 fold_8bit_crc32(crc, table, rax); 10304 } 10305 10306 BIND(L_tail_restore); 10307 movl(len, tmp); // restore 10308 BIND(L_tail); 10309 andl(len, 0xf); 10310 jccb(Assembler::zero, L_exit); 10311 10312 // Fold the rest of bytes 10313 align(4); 10314 BIND(L_tail_loop); 10315 movsbl(rax, Address(buf, 0)); // load byte with sign extension 10316 update_byte_crc32(crc, rax, table); 10317 increment(buf); 10318 decrementl(len); 10319 jccb(Assembler::greater, L_tail_loop); 10320 10321 BIND(L_exit); 10322 notl(crc); // ~c 10323 } 10324 10325 #ifdef _LP64 10326 // S. Gueron / Information Processing Letters 112 (2012) 184 10327 // Algorithm 4: Computing carry-less multiplication using a precomputed lookup table. 10328 // Input: A 32 bit value B = [byte3, byte2, byte1, byte0]. 10329 // Output: the 64-bit carry-less product of B * CONST 10330 void MacroAssembler::crc32c_ipl_alg4(Register in, uint32_t n, 10331 Register tmp1, Register tmp2, Register tmp3) { 10332 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr())); 10333 if (n > 0) { 10334 addq(tmp3, n * 256 * 8); 10335 } 10336 // Q1 = TABLEExt[n][B & 0xFF]; 10337 movl(tmp1, in); 10338 andl(tmp1, 0x000000FF); 10339 shll(tmp1, 3); 10340 addq(tmp1, tmp3); 10341 movq(tmp1, Address(tmp1, 0)); 10342 10343 // Q2 = TABLEExt[n][B >> 8 & 0xFF]; 10344 movl(tmp2, in); 10345 shrl(tmp2, 8); 10346 andl(tmp2, 0x000000FF); 10347 shll(tmp2, 3); 10348 addq(tmp2, tmp3); 10349 movq(tmp2, Address(tmp2, 0)); 10350 10351 shlq(tmp2, 8); 10352 xorq(tmp1, tmp2); 10353 10354 // Q3 = TABLEExt[n][B >> 16 & 0xFF]; 10355 movl(tmp2, in); 10356 shrl(tmp2, 16); 10357 andl(tmp2, 0x000000FF); 10358 shll(tmp2, 3); 10359 addq(tmp2, tmp3); 10360 movq(tmp2, Address(tmp2, 0)); 10361 10362 shlq(tmp2, 16); 10363 xorq(tmp1, tmp2); 10364 10365 // Q4 = TABLEExt[n][B >> 24 & 0xFF]; 10366 shrl(in, 24); 10367 andl(in, 0x000000FF); 10368 shll(in, 3); 10369 addq(in, tmp3); 10370 movq(in, Address(in, 0)); 10371 10372 shlq(in, 24); 10373 xorq(in, tmp1); 10374 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24; 10375 } 10376 10377 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1, 10378 Register in_out, 10379 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported, 10380 XMMRegister w_xtmp2, 10381 Register tmp1, 10382 Register n_tmp2, Register n_tmp3) { 10383 if (is_pclmulqdq_supported) { 10384 movdl(w_xtmp1, in_out); // modified blindly 10385 10386 movl(tmp1, const_or_pre_comp_const_index); 10387 movdl(w_xtmp2, tmp1); 10388 pclmulqdq(w_xtmp1, w_xtmp2, 0); 10389 10390 movdq(in_out, w_xtmp1); 10391 } else { 10392 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3); 10393 } 10394 } 10395 10396 // Recombination Alternative 2: No bit-reflections 10397 // T1 = (CRC_A * U1) << 1 10398 // T2 = (CRC_B * U2) << 1 10399 // C1 = T1 >> 32 10400 // C2 = T2 >> 32 10401 // T1 = T1 & 0xFFFFFFFF 10402 // T2 = T2 & 0xFFFFFFFF 10403 // T1 = CRC32(0, T1) 10404 // T2 = CRC32(0, T2) 10405 // C1 = C1 ^ T1 10406 // C2 = C2 ^ T2 10407 // CRC = C1 ^ C2 ^ CRC_C 10408 void MacroAssembler::crc32c_rec_alt2(uint32_t const_or_pre_comp_const_index_u1, uint32_t const_or_pre_comp_const_index_u2, bool is_pclmulqdq_supported, Register in_out, Register in1, Register in2, 10409 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10410 Register tmp1, Register tmp2, 10411 Register n_tmp3) { 10412 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10413 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10414 shlq(in_out, 1); 10415 movl(tmp1, in_out); 10416 shrq(in_out, 32); 10417 xorl(tmp2, tmp2); 10418 crc32(tmp2, tmp1, 4); 10419 xorl(in_out, tmp2); // we don't care about upper 32 bit contents here 10420 shlq(in1, 1); 10421 movl(tmp1, in1); 10422 shrq(in1, 32); 10423 xorl(tmp2, tmp2); 10424 crc32(tmp2, tmp1, 4); 10425 xorl(in1, tmp2); 10426 xorl(in_out, in1); 10427 xorl(in_out, in2); 10428 } 10429 10430 // Set N to predefined value 10431 // Subtract from a lenght of a buffer 10432 // execute in a loop: 10433 // CRC_A = 0xFFFFFFFF, CRC_B = 0, CRC_C = 0 10434 // for i = 1 to N do 10435 // CRC_A = CRC32(CRC_A, A[i]) 10436 // CRC_B = CRC32(CRC_B, B[i]) 10437 // CRC_C = CRC32(CRC_C, C[i]) 10438 // end for 10439 // Recombine 10440 void MacroAssembler::crc32c_proc_chunk(uint32_t size, uint32_t const_or_pre_comp_const_index_u1, uint32_t const_or_pre_comp_const_index_u2, bool is_pclmulqdq_supported, 10441 Register in_out1, Register in_out2, Register in_out3, 10442 Register tmp1, Register tmp2, Register tmp3, 10443 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10444 Register tmp4, Register tmp5, 10445 Register n_tmp6) { 10446 Label L_processPartitions; 10447 Label L_processPartition; 10448 Label L_exit; 10449 10450 bind(L_processPartitions); 10451 cmpl(in_out1, 3 * size); 10452 jcc(Assembler::less, L_exit); 10453 xorl(tmp1, tmp1); 10454 xorl(tmp2, tmp2); 10455 movq(tmp3, in_out2); 10456 addq(tmp3, size); 10457 10458 bind(L_processPartition); 10459 crc32(in_out3, Address(in_out2, 0), 8); 10460 crc32(tmp1, Address(in_out2, size), 8); 10461 crc32(tmp2, Address(in_out2, size * 2), 8); 10462 addq(in_out2, 8); 10463 cmpq(in_out2, tmp3); 10464 jcc(Assembler::less, L_processPartition); 10465 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2, 10466 w_xtmp1, w_xtmp2, w_xtmp3, 10467 tmp4, tmp5, 10468 n_tmp6); 10469 addq(in_out2, 2 * size); 10470 subl(in_out1, 3 * size); 10471 jmp(L_processPartitions); 10472 10473 bind(L_exit); 10474 } 10475 #else 10476 void MacroAssembler::crc32c_ipl_alg4(Register in_out, uint32_t n, 10477 Register tmp1, Register tmp2, Register tmp3, 10478 XMMRegister xtmp1, XMMRegister xtmp2) { 10479 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr())); 10480 if (n > 0) { 10481 addl(tmp3, n * 256 * 8); 10482 } 10483 // Q1 = TABLEExt[n][B & 0xFF]; 10484 movl(tmp1, in_out); 10485 andl(tmp1, 0x000000FF); 10486 shll(tmp1, 3); 10487 addl(tmp1, tmp3); 10488 movq(xtmp1, Address(tmp1, 0)); 10489 10490 // Q2 = TABLEExt[n][B >> 8 & 0xFF]; 10491 movl(tmp2, in_out); 10492 shrl(tmp2, 8); 10493 andl(tmp2, 0x000000FF); 10494 shll(tmp2, 3); 10495 addl(tmp2, tmp3); 10496 movq(xtmp2, Address(tmp2, 0)); 10497 10498 psllq(xtmp2, 8); 10499 pxor(xtmp1, xtmp2); 10500 10501 // Q3 = TABLEExt[n][B >> 16 & 0xFF]; 10502 movl(tmp2, in_out); 10503 shrl(tmp2, 16); 10504 andl(tmp2, 0x000000FF); 10505 shll(tmp2, 3); 10506 addl(tmp2, tmp3); 10507 movq(xtmp2, Address(tmp2, 0)); 10508 10509 psllq(xtmp2, 16); 10510 pxor(xtmp1, xtmp2); 10511 10512 // Q4 = TABLEExt[n][B >> 24 & 0xFF]; 10513 shrl(in_out, 24); 10514 andl(in_out, 0x000000FF); 10515 shll(in_out, 3); 10516 addl(in_out, tmp3); 10517 movq(xtmp2, Address(in_out, 0)); 10518 10519 psllq(xtmp2, 24); 10520 pxor(xtmp1, xtmp2); // Result in CXMM 10521 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24; 10522 } 10523 10524 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1, 10525 Register in_out, 10526 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported, 10527 XMMRegister w_xtmp2, 10528 Register tmp1, 10529 Register n_tmp2, Register n_tmp3) { 10530 if (is_pclmulqdq_supported) { 10531 movdl(w_xtmp1, in_out); 10532 10533 movl(tmp1, const_or_pre_comp_const_index); 10534 movdl(w_xtmp2, tmp1); 10535 pclmulqdq(w_xtmp1, w_xtmp2, 0); 10536 // Keep result in XMM since GPR is 32 bit in length 10537 } else { 10538 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3, w_xtmp1, w_xtmp2); 10539 } 10540 } 10541 10542 void MacroAssembler::crc32c_rec_alt2(uint32_t const_or_pre_comp_const_index_u1, uint32_t const_or_pre_comp_const_index_u2, bool is_pclmulqdq_supported, Register in_out, Register in1, Register in2, 10543 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10544 Register tmp1, Register tmp2, 10545 Register n_tmp3) { 10546 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10547 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10548 10549 psllq(w_xtmp1, 1); 10550 movdl(tmp1, w_xtmp1); 10551 psrlq(w_xtmp1, 32); 10552 movdl(in_out, w_xtmp1); 10553 10554 xorl(tmp2, tmp2); 10555 crc32(tmp2, tmp1, 4); 10556 xorl(in_out, tmp2); 10557 10558 psllq(w_xtmp2, 1); 10559 movdl(tmp1, w_xtmp2); 10560 psrlq(w_xtmp2, 32); 10561 movdl(in1, w_xtmp2); 10562 10563 xorl(tmp2, tmp2); 10564 crc32(tmp2, tmp1, 4); 10565 xorl(in1, tmp2); 10566 xorl(in_out, in1); 10567 xorl(in_out, in2); 10568 } 10569 10570 void MacroAssembler::crc32c_proc_chunk(uint32_t size, uint32_t const_or_pre_comp_const_index_u1, uint32_t const_or_pre_comp_const_index_u2, bool is_pclmulqdq_supported, 10571 Register in_out1, Register in_out2, Register in_out3, 10572 Register tmp1, Register tmp2, Register tmp3, 10573 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10574 Register tmp4, Register tmp5, 10575 Register n_tmp6) { 10576 Label L_processPartitions; 10577 Label L_processPartition; 10578 Label L_exit; 10579 10580 bind(L_processPartitions); 10581 cmpl(in_out1, 3 * size); 10582 jcc(Assembler::less, L_exit); 10583 xorl(tmp1, tmp1); 10584 xorl(tmp2, tmp2); 10585 movl(tmp3, in_out2); 10586 addl(tmp3, size); 10587 10588 bind(L_processPartition); 10589 crc32(in_out3, Address(in_out2, 0), 4); 10590 crc32(tmp1, Address(in_out2, size), 4); 10591 crc32(tmp2, Address(in_out2, size*2), 4); 10592 crc32(in_out3, Address(in_out2, 0+4), 4); 10593 crc32(tmp1, Address(in_out2, size+4), 4); 10594 crc32(tmp2, Address(in_out2, size*2+4), 4); 10595 addl(in_out2, 8); 10596 cmpl(in_out2, tmp3); 10597 jcc(Assembler::less, L_processPartition); 10598 10599 push(tmp3); 10600 push(in_out1); 10601 push(in_out2); 10602 tmp4 = tmp3; 10603 tmp5 = in_out1; 10604 n_tmp6 = in_out2; 10605 10606 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2, 10607 w_xtmp1, w_xtmp2, w_xtmp3, 10608 tmp4, tmp5, 10609 n_tmp6); 10610 10611 pop(in_out2); 10612 pop(in_out1); 10613 pop(tmp3); 10614 10615 addl(in_out2, 2 * size); 10616 subl(in_out1, 3 * size); 10617 jmp(L_processPartitions); 10618 10619 bind(L_exit); 10620 } 10621 #endif //LP64 10622 10623 #ifdef _LP64 10624 // Algorithm 2: Pipelined usage of the CRC32 instruction. 10625 // Input: A buffer I of L bytes. 10626 // Output: the CRC32C value of the buffer. 10627 // Notations: 10628 // Write L = 24N + r, with N = floor (L/24). 10629 // r = L mod 24 (0 <= r < 24). 10630 // Consider I as the concatenation of A|B|C|R, where A, B, C, each, 10631 // N quadwords, and R consists of r bytes. 10632 // A[j] = I [8j+7:8j], j= 0, 1, ..., N-1 10633 // B[j] = I [N + 8j+7:N + 8j], j= 0, 1, ..., N-1 10634 // C[j] = I [2N + 8j+7:2N + 8j], j= 0, 1, ..., N-1 10635 // if r > 0 R[j] = I [3N +j], j= 0, 1, ...,r-1 10636 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2, 10637 Register tmp1, Register tmp2, Register tmp3, 10638 Register tmp4, Register tmp5, Register tmp6, 10639 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10640 bool is_pclmulqdq_supported) { 10641 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS]; 10642 Label L_wordByWord; 10643 Label L_byteByByteProlog; 10644 Label L_byteByByte; 10645 Label L_exit; 10646 10647 if (is_pclmulqdq_supported ) { 10648 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr; 10649 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr+1); 10650 10651 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2); 10652 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3); 10653 10654 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4); 10655 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5); 10656 assert((CRC32C_NUM_PRECOMPUTED_CONSTANTS - 1 ) == 5, "Checking whether you declared all of the constants based on the number of \"chunks\""); 10657 } else { 10658 const_or_pre_comp_const_index[0] = 1; 10659 const_or_pre_comp_const_index[1] = 0; 10660 10661 const_or_pre_comp_const_index[2] = 3; 10662 const_or_pre_comp_const_index[3] = 2; 10663 10664 const_or_pre_comp_const_index[4] = 5; 10665 const_or_pre_comp_const_index[5] = 4; 10666 } 10667 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported, 10668 in2, in1, in_out, 10669 tmp1, tmp2, tmp3, 10670 w_xtmp1, w_xtmp2, w_xtmp3, 10671 tmp4, tmp5, 10672 tmp6); 10673 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported, 10674 in2, in1, in_out, 10675 tmp1, tmp2, tmp3, 10676 w_xtmp1, w_xtmp2, w_xtmp3, 10677 tmp4, tmp5, 10678 tmp6); 10679 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported, 10680 in2, in1, in_out, 10681 tmp1, tmp2, tmp3, 10682 w_xtmp1, w_xtmp2, w_xtmp3, 10683 tmp4, tmp5, 10684 tmp6); 10685 movl(tmp1, in2); 10686 andl(tmp1, 0x00000007); 10687 negl(tmp1); 10688 addl(tmp1, in2); 10689 addq(tmp1, in1); 10690 10691 BIND(L_wordByWord); 10692 cmpq(in1, tmp1); 10693 jcc(Assembler::greaterEqual, L_byteByByteProlog); 10694 crc32(in_out, Address(in1, 0), 4); 10695 addq(in1, 4); 10696 jmp(L_wordByWord); 10697 10698 BIND(L_byteByByteProlog); 10699 andl(in2, 0x00000007); 10700 movl(tmp2, 1); 10701 10702 BIND(L_byteByByte); 10703 cmpl(tmp2, in2); 10704 jccb(Assembler::greater, L_exit); 10705 crc32(in_out, Address(in1, 0), 1); 10706 incq(in1); 10707 incl(tmp2); 10708 jmp(L_byteByByte); 10709 10710 BIND(L_exit); 10711 } 10712 #else 10713 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2, 10714 Register tmp1, Register tmp2, Register tmp3, 10715 Register tmp4, Register tmp5, Register tmp6, 10716 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10717 bool is_pclmulqdq_supported) { 10718 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS]; 10719 Label L_wordByWord; 10720 Label L_byteByByteProlog; 10721 Label L_byteByByte; 10722 Label L_exit; 10723 10724 if (is_pclmulqdq_supported) { 10725 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr; 10726 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 1); 10727 10728 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2); 10729 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3); 10730 10731 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4); 10732 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5); 10733 } else { 10734 const_or_pre_comp_const_index[0] = 1; 10735 const_or_pre_comp_const_index[1] = 0; 10736 10737 const_or_pre_comp_const_index[2] = 3; 10738 const_or_pre_comp_const_index[3] = 2; 10739 10740 const_or_pre_comp_const_index[4] = 5; 10741 const_or_pre_comp_const_index[5] = 4; 10742 } 10743 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported, 10744 in2, in1, in_out, 10745 tmp1, tmp2, tmp3, 10746 w_xtmp1, w_xtmp2, w_xtmp3, 10747 tmp4, tmp5, 10748 tmp6); 10749 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported, 10750 in2, in1, in_out, 10751 tmp1, tmp2, tmp3, 10752 w_xtmp1, w_xtmp2, w_xtmp3, 10753 tmp4, tmp5, 10754 tmp6); 10755 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported, 10756 in2, in1, in_out, 10757 tmp1, tmp2, tmp3, 10758 w_xtmp1, w_xtmp2, w_xtmp3, 10759 tmp4, tmp5, 10760 tmp6); 10761 movl(tmp1, in2); 10762 andl(tmp1, 0x00000007); 10763 negl(tmp1); 10764 addl(tmp1, in2); 10765 addl(tmp1, in1); 10766 10767 BIND(L_wordByWord); 10768 cmpl(in1, tmp1); 10769 jcc(Assembler::greaterEqual, L_byteByByteProlog); 10770 crc32(in_out, Address(in1,0), 4); 10771 addl(in1, 4); 10772 jmp(L_wordByWord); 10773 10774 BIND(L_byteByByteProlog); 10775 andl(in2, 0x00000007); 10776 movl(tmp2, 1); 10777 10778 BIND(L_byteByByte); 10779 cmpl(tmp2, in2); 10780 jccb(Assembler::greater, L_exit); 10781 movb(tmp1, Address(in1, 0)); 10782 crc32(in_out, tmp1, 1); 10783 incl(in1); 10784 incl(tmp2); 10785 jmp(L_byteByByte); 10786 10787 BIND(L_exit); 10788 } 10789 #endif // LP64 10790 #undef BIND 10791 #undef BLOCK_COMMENT 10792 10793 // Compress char[] array to byte[]. 10794 // ..\jdk\src\java.base\share\classes\java\lang\StringUTF16.java 10795 // @HotSpotIntrinsicCandidate 10796 // private static int compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) { 10797 // for (int i = 0; i < len; i++) { 10798 // int c = src[srcOff++]; 10799 // if (c >>> 8 != 0) { 10800 // return 0; 10801 // } 10802 // dst[dstOff++] = (byte)c; 10803 // } 10804 // return len; 10805 // } 10806 void MacroAssembler::char_array_compress(Register src, Register dst, Register len, 10807 XMMRegister tmp1Reg, XMMRegister tmp2Reg, 10808 XMMRegister tmp3Reg, XMMRegister tmp4Reg, 10809 Register tmp5, Register result) { 10810 Label copy_chars_loop, return_length, return_zero, done, below_threshold; 10811 10812 // rsi: src 10813 // rdi: dst 10814 // rdx: len 10815 // rcx: tmp5 10816 // rax: result 10817 10818 // rsi holds start addr of source char[] to be compressed 10819 // rdi holds start addr of destination byte[] 10820 // rdx holds length 10821 10822 assert(len != result, ""); 10823 10824 // save length for return 10825 push(len); 10826 10827 if ((UseAVX > 2) && // AVX512 10828 VM_Version::supports_avx512vlbw() && 10829 VM_Version::supports_bmi2()) { 10830 10831 set_vector_masking(); // opening of the stub context for programming mask registers 10832 10833 Label copy_32_loop, copy_loop_tail, restore_k1_return_zero; 10834 10835 // alignement 10836 Label post_alignement; 10837 10838 // if length of the string is less than 16, handle it in an old fashioned 10839 // way 10840 testl(len, -32); 10841 jcc(Assembler::zero, below_threshold); 10842 10843 // First check whether a character is compressable ( <= 0xFF). 10844 // Create mask to test for Unicode chars inside zmm vector 10845 movl(result, 0x00FF); 10846 evpbroadcastw(tmp2Reg, result, Assembler::AVX_512bit); 10847 10848 // Save k1 10849 kmovql(k3, k1); 10850 10851 testl(len, -64); 10852 jcc(Assembler::zero, post_alignement); 10853 10854 movl(tmp5, dst); 10855 andl(tmp5, (32 - 1)); 10856 negl(tmp5); 10857 andl(tmp5, (32 - 1)); 10858 10859 // bail out when there is nothing to be done 10860 testl(tmp5, 0xFFFFFFFF); 10861 jcc(Assembler::zero, post_alignement); 10862 10863 // ~(~0 << len), where len is the # of remaining elements to process 10864 movl(result, 0xFFFFFFFF); 10865 shlxl(result, result, tmp5); 10866 notl(result); 10867 kmovdl(k1, result); 10868 10869 evmovdquw(tmp1Reg, k1, Address(src, 0), Assembler::AVX_512bit); 10870 evpcmpuw(k2, k1, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit); 10871 ktestd(k2, k1); 10872 jcc(Assembler::carryClear, restore_k1_return_zero); 10873 10874 evpmovwb(Address(dst, 0), k1, tmp1Reg, Assembler::AVX_512bit); 10875 10876 addptr(src, tmp5); 10877 addptr(src, tmp5); 10878 addptr(dst, tmp5); 10879 subl(len, tmp5); 10880 10881 bind(post_alignement); 10882 // end of alignement 10883 10884 movl(tmp5, len); 10885 andl(tmp5, (32 - 1)); // tail count (in chars) 10886 andl(len, ~(32 - 1)); // vector count (in chars) 10887 jcc(Assembler::zero, copy_loop_tail); 10888 10889 lea(src, Address(src, len, Address::times_2)); 10890 lea(dst, Address(dst, len, Address::times_1)); 10891 negptr(len); 10892 10893 bind(copy_32_loop); 10894 evmovdquw(tmp1Reg, Address(src, len, Address::times_2), Assembler::AVX_512bit); 10895 evpcmpuw(k2, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit); 10896 kortestdl(k2, k2); 10897 jcc(Assembler::carryClear, restore_k1_return_zero); 10898 10899 // All elements in current processed chunk are valid candidates for 10900 // compression. Write a truncated byte elements to the memory. 10901 evpmovwb(Address(dst, len, Address::times_1), tmp1Reg, Assembler::AVX_512bit); 10902 addptr(len, 32); 10903 jcc(Assembler::notZero, copy_32_loop); 10904 10905 bind(copy_loop_tail); 10906 // bail out when there is nothing to be done 10907 testl(tmp5, 0xFFFFFFFF); 10908 // Restore k1 10909 kmovql(k1, k3); 10910 jcc(Assembler::zero, return_length); 10911 10912 movl(len, tmp5); 10913 10914 // ~(~0 << len), where len is the # of remaining elements to process 10915 movl(result, 0xFFFFFFFF); 10916 shlxl(result, result, len); 10917 notl(result); 10918 10919 kmovdl(k1, result); 10920 10921 evmovdquw(tmp1Reg, k1, Address(src, 0), Assembler::AVX_512bit); 10922 evpcmpuw(k2, k1, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit); 10923 ktestd(k2, k1); 10924 jcc(Assembler::carryClear, restore_k1_return_zero); 10925 10926 evpmovwb(Address(dst, 0), k1, tmp1Reg, Assembler::AVX_512bit); 10927 // Restore k1 10928 kmovql(k1, k3); 10929 jmp(return_length); 10930 10931 bind(restore_k1_return_zero); 10932 // Restore k1 10933 kmovql(k1, k3); 10934 jmp(return_zero); 10935 10936 clear_vector_masking(); // closing of the stub context for programming mask registers 10937 } 10938 if (UseSSE42Intrinsics) { 10939 Label copy_32_loop, copy_16, copy_tail; 10940 10941 bind(below_threshold); 10942 10943 movl(result, len); 10944 10945 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vectors 10946 10947 // vectored compression 10948 andl(len, 0xfffffff0); // vector count (in chars) 10949 andl(result, 0x0000000f); // tail count (in chars) 10950 testl(len, len); 10951 jccb(Assembler::zero, copy_16); 10952 10953 // compress 16 chars per iter 10954 movdl(tmp1Reg, tmp5); 10955 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg 10956 pxor(tmp4Reg, tmp4Reg); 10957 10958 lea(src, Address(src, len, Address::times_2)); 10959 lea(dst, Address(dst, len, Address::times_1)); 10960 negptr(len); 10961 10962 bind(copy_32_loop); 10963 movdqu(tmp2Reg, Address(src, len, Address::times_2)); // load 1st 8 characters 10964 por(tmp4Reg, tmp2Reg); 10965 movdqu(tmp3Reg, Address(src, len, Address::times_2, 16)); // load next 8 characters 10966 por(tmp4Reg, tmp3Reg); 10967 ptest(tmp4Reg, tmp1Reg); // check for Unicode chars in next vector 10968 jcc(Assembler::notZero, return_zero); 10969 packuswb(tmp2Reg, tmp3Reg); // only ASCII chars; compress each to 1 byte 10970 movdqu(Address(dst, len, Address::times_1), tmp2Reg); 10971 addptr(len, 16); 10972 jcc(Assembler::notZero, copy_32_loop); 10973 10974 // compress next vector of 8 chars (if any) 10975 bind(copy_16); 10976 movl(len, result); 10977 andl(len, 0xfffffff8); // vector count (in chars) 10978 andl(result, 0x00000007); // tail count (in chars) 10979 testl(len, len); 10980 jccb(Assembler::zero, copy_tail); 10981 10982 movdl(tmp1Reg, tmp5); 10983 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg 10984 pxor(tmp3Reg, tmp3Reg); 10985 10986 movdqu(tmp2Reg, Address(src, 0)); 10987 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 10988 jccb(Assembler::notZero, return_zero); 10989 packuswb(tmp2Reg, tmp3Reg); // only LATIN1 chars; compress each to 1 byte 10990 movq(Address(dst, 0), tmp2Reg); 10991 addptr(src, 16); 10992 addptr(dst, 8); 10993 10994 bind(copy_tail); 10995 movl(len, result); 10996 } 10997 // compress 1 char per iter 10998 testl(len, len); 10999 jccb(Assembler::zero, return_length); 11000 lea(src, Address(src, len, Address::times_2)); 11001 lea(dst, Address(dst, len, Address::times_1)); 11002 negptr(len); 11003 11004 bind(copy_chars_loop); 11005 load_unsigned_short(result, Address(src, len, Address::times_2)); 11006 testl(result, 0xff00); // check if Unicode char 11007 jccb(Assembler::notZero, return_zero); 11008 movb(Address(dst, len, Address::times_1), result); // ASCII char; compress to 1 byte 11009 increment(len); 11010 jcc(Assembler::notZero, copy_chars_loop); 11011 11012 // if compression succeeded, return length 11013 bind(return_length); 11014 pop(result); 11015 jmpb(done); 11016 11017 // if compression failed, return 0 11018 bind(return_zero); 11019 xorl(result, result); 11020 addptr(rsp, wordSize); 11021 11022 bind(done); 11023 } 11024 11025 // Inflate byte[] array to char[]. 11026 // ..\jdk\src\java.base\share\classes\java\lang\StringLatin1.java 11027 // @HotSpotIntrinsicCandidate 11028 // private static void inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) { 11029 // for (int i = 0; i < len; i++) { 11030 // dst[dstOff++] = (char)(src[srcOff++] & 0xff); 11031 // } 11032 // } 11033 void MacroAssembler::byte_array_inflate(Register src, Register dst, Register len, 11034 XMMRegister tmp1, Register tmp2) { 11035 Label copy_chars_loop, done, below_threshold; 11036 // rsi: src 11037 // rdi: dst 11038 // rdx: len 11039 // rcx: tmp2 11040 11041 // rsi holds start addr of source byte[] to be inflated 11042 // rdi holds start addr of destination char[] 11043 // rdx holds length 11044 assert_different_registers(src, dst, len, tmp2); 11045 11046 if ((UseAVX > 2) && // AVX512 11047 VM_Version::supports_avx512vlbw() && 11048 VM_Version::supports_bmi2()) { 11049 11050 set_vector_masking(); // opening of the stub context for programming mask registers 11051 11052 Label copy_32_loop, copy_tail; 11053 Register tmp3_aliased = len; 11054 11055 // if length of the string is less than 16, handle it in an old fashioned 11056 // way 11057 testl(len, -16); 11058 jcc(Assembler::zero, below_threshold); 11059 11060 // In order to use only one arithmetic operation for the main loop we use 11061 // this pre-calculation 11062 movl(tmp2, len); 11063 andl(tmp2, (32 - 1)); // tail count (in chars), 32 element wide loop 11064 andl(len, -32); // vector count 11065 jccb(Assembler::zero, copy_tail); 11066 11067 lea(src, Address(src, len, Address::times_1)); 11068 lea(dst, Address(dst, len, Address::times_2)); 11069 negptr(len); 11070 11071 11072 // inflate 32 chars per iter 11073 bind(copy_32_loop); 11074 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_512bit); 11075 evmovdquw(Address(dst, len, Address::times_2), tmp1, Assembler::AVX_512bit); 11076 addptr(len, 32); 11077 jcc(Assembler::notZero, copy_32_loop); 11078 11079 bind(copy_tail); 11080 // bail out when there is nothing to be done 11081 testl(tmp2, -1); // we don't destroy the contents of tmp2 here 11082 jcc(Assembler::zero, done); 11083 11084 // Save k1 11085 kmovql(k2, k1); 11086 11087 // ~(~0 << length), where length is the # of remaining elements to process 11088 movl(tmp3_aliased, -1); 11089 shlxl(tmp3_aliased, tmp3_aliased, tmp2); 11090 notl(tmp3_aliased); 11091 kmovdl(k1, tmp3_aliased); 11092 evpmovzxbw(tmp1, k1, Address(src, 0), Assembler::AVX_512bit); 11093 evmovdquw(Address(dst, 0), k1, tmp1, Assembler::AVX_512bit); 11094 11095 // Restore k1 11096 kmovql(k1, k2); 11097 jmp(done); 11098 11099 clear_vector_masking(); // closing of the stub context for programming mask registers 11100 } 11101 if (UseSSE42Intrinsics) { 11102 Label copy_16_loop, copy_8_loop, copy_bytes, copy_new_tail, copy_tail; 11103 11104 movl(tmp2, len); 11105 11106 if (UseAVX > 1) { 11107 andl(tmp2, (16 - 1)); 11108 andl(len, -16); 11109 jccb(Assembler::zero, copy_new_tail); 11110 } else { 11111 andl(tmp2, 0x00000007); // tail count (in chars) 11112 andl(len, 0xfffffff8); // vector count (in chars) 11113 jccb(Assembler::zero, copy_tail); 11114 } 11115 11116 // vectored inflation 11117 lea(src, Address(src, len, Address::times_1)); 11118 lea(dst, Address(dst, len, Address::times_2)); 11119 negptr(len); 11120 11121 if (UseAVX > 1) { 11122 bind(copy_16_loop); 11123 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_256bit); 11124 vmovdqu(Address(dst, len, Address::times_2), tmp1); 11125 addptr(len, 16); 11126 jcc(Assembler::notZero, copy_16_loop); 11127 11128 bind(below_threshold); 11129 bind(copy_new_tail); 11130 if ((UseAVX > 2) && 11131 VM_Version::supports_avx512vlbw() && 11132 VM_Version::supports_bmi2()) { 11133 movl(tmp2, len); 11134 } else { 11135 movl(len, tmp2); 11136 } 11137 andl(tmp2, 0x00000007); 11138 andl(len, 0xFFFFFFF8); 11139 jccb(Assembler::zero, copy_tail); 11140 11141 pmovzxbw(tmp1, Address(src, 0)); 11142 movdqu(Address(dst, 0), tmp1); 11143 addptr(src, 8); 11144 addptr(dst, 2 * 8); 11145 11146 jmp(copy_tail, true); 11147 } 11148 11149 // inflate 8 chars per iter 11150 bind(copy_8_loop); 11151 pmovzxbw(tmp1, Address(src, len, Address::times_1)); // unpack to 8 words 11152 movdqu(Address(dst, len, Address::times_2), tmp1); 11153 addptr(len, 8); 11154 jcc(Assembler::notZero, copy_8_loop); 11155 11156 bind(copy_tail); 11157 movl(len, tmp2); 11158 11159 cmpl(len, 4); 11160 jccb(Assembler::less, copy_bytes); 11161 11162 movdl(tmp1, Address(src, 0)); // load 4 byte chars 11163 pmovzxbw(tmp1, tmp1); 11164 movq(Address(dst, 0), tmp1); 11165 subptr(len, 4); 11166 addptr(src, 4); 11167 addptr(dst, 8); 11168 11169 bind(copy_bytes); 11170 } 11171 testl(len, len); 11172 jccb(Assembler::zero, done); 11173 lea(src, Address(src, len, Address::times_1)); 11174 lea(dst, Address(dst, len, Address::times_2)); 11175 negptr(len); 11176 11177 // inflate 1 char per iter 11178 bind(copy_chars_loop); 11179 load_unsigned_byte(tmp2, Address(src, len, Address::times_1)); // load byte char 11180 movw(Address(dst, len, Address::times_2), tmp2); // inflate byte char to word 11181 increment(len); 11182 jcc(Assembler::notZero, copy_chars_loop); 11183 11184 bind(done); 11185 } 11186 11187 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) { 11188 switch (cond) { 11189 // Note some conditions are synonyms for others 11190 case Assembler::zero: return Assembler::notZero; 11191 case Assembler::notZero: return Assembler::zero; 11192 case Assembler::less: return Assembler::greaterEqual; 11193 case Assembler::lessEqual: return Assembler::greater; 11194 case Assembler::greater: return Assembler::lessEqual; 11195 case Assembler::greaterEqual: return Assembler::less; 11196 case Assembler::below: return Assembler::aboveEqual; 11197 case Assembler::belowEqual: return Assembler::above; 11198 case Assembler::above: return Assembler::belowEqual; 11199 case Assembler::aboveEqual: return Assembler::below; 11200 case Assembler::overflow: return Assembler::noOverflow; 11201 case Assembler::noOverflow: return Assembler::overflow; 11202 case Assembler::negative: return Assembler::positive; 11203 case Assembler::positive: return Assembler::negative; 11204 case Assembler::parity: return Assembler::noParity; 11205 case Assembler::noParity: return Assembler::parity; 11206 } 11207 ShouldNotReachHere(); return Assembler::overflow; 11208 } 11209 11210 SkipIfEqual::SkipIfEqual( 11211 MacroAssembler* masm, const bool* flag_addr, bool value) { 11212 _masm = masm; 11213 _masm->cmp8(ExternalAddress((address)flag_addr), value); 11214 _masm->jcc(Assembler::equal, _label); 11215 } 11216 11217 SkipIfEqual::~SkipIfEqual() { 11218 _masm->bind(_label); 11219 } 11220 11221 // 32-bit Windows has its own fast-path implementation 11222 // of get_thread 11223 #if !defined(WIN32) || defined(_LP64) 11224 11225 // This is simply a call to Thread::current() 11226 void MacroAssembler::get_thread(Register thread) { 11227 if (thread != rax) { 11228 push(rax); 11229 } 11230 LP64_ONLY(push(rdi);) 11231 LP64_ONLY(push(rsi);) 11232 push(rdx); 11233 push(rcx); 11234 #ifdef _LP64 11235 push(r8); 11236 push(r9); 11237 push(r10); 11238 push(r11); 11239 #endif 11240 11241 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, Thread::current), 0); 11242 11243 #ifdef _LP64 11244 pop(r11); 11245 pop(r10); 11246 pop(r9); 11247 pop(r8); 11248 #endif 11249 pop(rcx); 11250 pop(rdx); 11251 LP64_ONLY(pop(rsi);) 11252 LP64_ONLY(pop(rdi);) 11253 if (thread != rax) { 11254 mov(thread, rax); 11255 pop(rax); 11256 } 11257 } 11258 11259 #endif