macroAssembler_x86.cpp revision 9867:3125c4a60cc9
1/* 2 * Copyright (c) 1997, 2015, 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 65bool AbstractAssembler::pd_check_instruction_mark() { return true; } 66#endif 67 68static 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 98Address MacroAssembler::as_Address(AddressLiteral adr) { 99 return Address(adr.target(), adr.rspec()); 100} 101 102Address MacroAssembler::as_Address(ArrayAddress adr) { 103 return Address::make_array(adr); 104} 105 106void 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 112void MacroAssembler::cmpklass(Address src1, Metadata* obj) { 113 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 114} 115 116void MacroAssembler::cmpklass(Register src1, Metadata* obj) { 117 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 118} 119 120void MacroAssembler::cmpoop(Address src1, jobject obj) { 121 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate()); 122} 123 124void MacroAssembler::cmpoop(Register src1, jobject obj) { 125 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate()); 126} 127 128void 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 138void 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 148void 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 160void MacroAssembler::jump(ArrayAddress entry) { 161 jmp(as_Address(entry)); 162} 163 164// Note: y_lo will be destroyed 165void 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 190void MacroAssembler::lea(Register dst, AddressLiteral src) { 191 mov_literal32(dst, (int32_t)src.target(), src.rspec()); 192} 193 194void 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 200void MacroAssembler::leave() { 201 mov(rsp, rbp); 202 pop(rbp); 203} 204 205void 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 245void MacroAssembler::lneg(Register hi, Register lo) { 246 negl(lo); 247 adcl(hi, 0); 248 negl(hi); 249} 250 251void 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 272void 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 293void MacroAssembler::movoop(Register dst, jobject obj) { 294 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate()); 295} 296 297void MacroAssembler::movoop(Address dst, jobject obj) { 298 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate()); 299} 300 301void MacroAssembler::mov_metadata(Register dst, Metadata* obj) { 302 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 303} 304 305void MacroAssembler::mov_metadata(Address dst, Metadata* obj) { 306 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 307} 308 309void 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 319void MacroAssembler::movptr(ArrayAddress dst, Register src) { 320 movl(as_Address(dst), src); 321} 322 323void 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 328void MacroAssembler::movptr(Address dst, intptr_t src) { 329 movl(dst, src); 330} 331 332 333void MacroAssembler::pop_callee_saved_registers() { 334 pop(rcx); 335 pop(rdx); 336 pop(rdi); 337 pop(rsi); 338} 339 340void MacroAssembler::pop_fTOS() { 341 fld_d(Address(rsp, 0)); 342 addl(rsp, 2 * wordSize); 343} 344 345void MacroAssembler::push_callee_saved_registers() { 346 push(rsi); 347 push(rdi); 348 push(rdx); 349 push(rcx); 350} 351 352void MacroAssembler::push_fTOS() { 353 subl(rsp, 2 * wordSize); 354 fstp_d(Address(rsp, 0)); 355} 356 357 358void MacroAssembler::pushoop(jobject obj) { 359 push_literal32((int32_t)obj, oop_Relocation::spec_for_immediate()); 360} 361 362void MacroAssembler::pushklass(Metadata* obj) { 363 push_literal32((int32_t)obj, metadata_Relocation::spec_for_immediate()); 364} 365 366void 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 374void MacroAssembler::set_word_if_not_zero(Register dst) { 375 xorl(dst, dst); 376 set_byte_if_not_zero(dst); 377} 378 379static void pass_arg0(MacroAssembler* masm, Register arg) { 380 masm->push(arg); 381} 382 383static void pass_arg1(MacroAssembler* masm, Register arg) { 384 masm->push(arg); 385} 386 387static void pass_arg2(MacroAssembler* masm, Register arg) { 388 masm->push(arg); 389} 390 391static void pass_arg3(MacroAssembler* masm, Register arg) { 392 masm->push(arg); 393} 394 395#ifndef PRODUCT 396extern "C" void findpc(intptr_t x); 397#endif 398 399void 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 427void 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 468void 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 478void 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 490void 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 506Address 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 516Address 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 525void 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 559void 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 570int 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 607void 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 615void 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 623void 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 632void 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 640void 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 650void 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 658void 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 663void MacroAssembler::lea(Register dst, AddressLiteral src) { 664 mov_literal64(dst, (intptr_t)src.target(), src.rspec()); 665} 666 667void MacroAssembler::lea(Address dst, AddressLiteral adr) { 668 mov_literal64(rscratch1, (intptr_t)adr.target(), adr.rspec()); 669 movptr(dst, rscratch1); 670} 671 672void MacroAssembler::leave() { 673 // %%% is this really better? Why not on 32bit too? 674 emit_int8((unsigned char)0xC9); // LEAVE 675} 676 677void MacroAssembler::lneg(Register hi, Register lo) { 678 ShouldNotReachHere(); // 64bit doesn't use two regs 679 negq(lo); 680} 681 682void MacroAssembler::movoop(Register dst, jobject obj) { 683 mov_literal64(dst, (intptr_t)obj, oop_Relocation::spec_for_immediate()); 684} 685 686void MacroAssembler::movoop(Address dst, jobject obj) { 687 mov_literal64(rscratch1, (intptr_t)obj, oop_Relocation::spec_for_immediate()); 688 movq(dst, rscratch1); 689} 690 691void MacroAssembler::mov_metadata(Register dst, Metadata* obj) { 692 mov_literal64(dst, (intptr_t)obj, metadata_Relocation::spec_for_immediate()); 693} 694 695void 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 700void 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 713void MacroAssembler::movptr(ArrayAddress dst, Register src) { 714 movq(as_Address(dst), src); 715} 716 717void 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 722void MacroAssembler::movptr(Address dst, intptr_t src) { 723 mov64(rscratch1, src); 724 movq(dst, rscratch1); 725} 726 727// These are mostly for initializing NULL 728void MacroAssembler::movptr(Address dst, int32_t src) { 729 movslq(dst, src); 730} 731 732void MacroAssembler::movptr(Register dst, int32_t src) { 733 mov64(dst, (intptr_t)src); 734} 735 736void MacroAssembler::pushoop(jobject obj) { 737 movoop(rscratch1, obj); 738 push(rscratch1); 739} 740 741void MacroAssembler::pushklass(Metadata* obj) { 742 mov_metadata(rscratch1, obj); 743 push(rscratch1); 744} 745 746void 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 755void MacroAssembler::reset_last_Java_frame(bool clear_fp, 756 bool clear_pc) { 757 // we must set sp to zero to clear frame 758 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), NULL_WORD); 759 // must clear fp, so that compiled frames are not confused; it is 760 // possible that we need it only for debugging 761 if (clear_fp) { 762 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD); 763 } 764 765 if (clear_pc) { 766 movptr(Address(r15_thread, JavaThread::last_Java_pc_offset()), NULL_WORD); 767 } 768} 769 770void MacroAssembler::set_last_Java_frame(Register last_java_sp, 771 Register last_java_fp, 772 address last_java_pc) { 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 795static void pass_arg0(MacroAssembler* masm, Register arg) { 796 if (c_rarg0 != arg ) { 797 masm->mov(c_rarg0, arg); 798 } 799} 800 801static void pass_arg1(MacroAssembler* masm, Register arg) { 802 if (c_rarg1 != arg ) { 803 masm->mov(c_rarg1, arg); 804 } 805} 806 807static void pass_arg2(MacroAssembler* masm, Register arg) { 808 if (c_rarg2 != arg ) { 809 masm->mov(c_rarg2, arg); 810 } 811} 812 813static void pass_arg3(MacroAssembler* masm, Register arg) { 814 if (c_rarg3 != arg ) { 815 masm->mov(c_rarg3, arg); 816 } 817} 818 819void 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 830void 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 842void 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 861extern "C" void findpc(intptr_t x); 862#endif 863 864void 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 893void 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 945void MacroAssembler::addptr(Register dst, int32_t imm32) { 946 LP64_ONLY(addq(dst, imm32)) NOT_LP64(addl(dst, imm32)); 947} 948 949void MacroAssembler::addptr(Register dst, Register src) { 950 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src)); 951} 952 953void MacroAssembler::addptr(Address dst, Register src) { 954 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src)); 955} 956 957void 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 966void 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 975void MacroAssembler::align(int modulus) { 976 align(modulus, offset()); 977} 978 979void MacroAssembler::align(int modulus, int target) { 980 if (target % modulus != 0) { 981 nop(modulus - (target % modulus)); 982 } 983} 984 985void MacroAssembler::andpd(XMMRegister dst, AddressLiteral src) { 986 // Used in sign-masking with aligned address. 987 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 988 if (reachable(src)) { 989 Assembler::andpd(dst, as_Address(src)); 990 } else { 991 lea(rscratch1, src); 992 Assembler::andpd(dst, Address(rscratch1, 0)); 993 } 994} 995 996void MacroAssembler::andps(XMMRegister dst, AddressLiteral src) { 997 // Used in sign-masking with aligned address. 998 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 999 if (reachable(src)) { 1000 Assembler::andps(dst, as_Address(src)); 1001 } else { 1002 lea(rscratch1, src); 1003 Assembler::andps(dst, Address(rscratch1, 0)); 1004 } 1005} 1006 1007void MacroAssembler::andptr(Register dst, int32_t imm32) { 1008 LP64_ONLY(andq(dst, imm32)) NOT_LP64(andl(dst, imm32)); 1009} 1010 1011void MacroAssembler::atomic_incl(Address counter_addr) { 1012 if (os::is_MP()) 1013 lock(); 1014 incrementl(counter_addr); 1015} 1016 1017void MacroAssembler::atomic_incl(AddressLiteral counter_addr, Register scr) { 1018 if (reachable(counter_addr)) { 1019 atomic_incl(as_Address(counter_addr)); 1020 } else { 1021 lea(scr, counter_addr); 1022 atomic_incl(Address(scr, 0)); 1023 } 1024} 1025 1026#ifdef _LP64 1027void MacroAssembler::atomic_incq(Address counter_addr) { 1028 if (os::is_MP()) 1029 lock(); 1030 incrementq(counter_addr); 1031} 1032 1033void MacroAssembler::atomic_incq(AddressLiteral counter_addr, Register scr) { 1034 if (reachable(counter_addr)) { 1035 atomic_incq(as_Address(counter_addr)); 1036 } else { 1037 lea(scr, counter_addr); 1038 atomic_incq(Address(scr, 0)); 1039 } 1040} 1041#endif 1042 1043// Writes to stack successive pages until offset reached to check for 1044// stack overflow + shadow pages. This clobbers tmp. 1045void MacroAssembler::bang_stack_size(Register size, Register tmp) { 1046 movptr(tmp, rsp); 1047 // Bang stack for total size given plus shadow page size. 1048 // Bang one page at a time because large size can bang beyond yellow and 1049 // red zones. 1050 Label loop; 1051 bind(loop); 1052 movl(Address(tmp, (-os::vm_page_size())), size ); 1053 subptr(tmp, os::vm_page_size()); 1054 subl(size, os::vm_page_size()); 1055 jcc(Assembler::greater, loop); 1056 1057 // Bang down shadow pages too. 1058 // At this point, (tmp-0) is the last address touched, so don't 1059 // touch it again. (It was touched as (tmp-pagesize) but then tmp 1060 // was post-decremented.) Skip this address by starting at i=1, and 1061 // touch a few more pages below. N.B. It is important to touch all 1062 // the way down including all pages in the shadow zone. 1063 for (int i = 1; i < ((int)JavaThread::stack_shadow_zone_size() / os::vm_page_size()); i++) { 1064 // this could be any sized move but this is can be a debugging crumb 1065 // so the bigger the better. 1066 movptr(Address(tmp, (-i*os::vm_page_size())), size ); 1067 } 1068} 1069 1070void MacroAssembler::reserved_stack_check() { 1071 // testing if reserved zone needs to be enabled 1072 Label no_reserved_zone_enabling; 1073 Register thread = NOT_LP64(rsi) LP64_ONLY(r15_thread); 1074 NOT_LP64(get_thread(rsi);) 1075 1076 cmpptr(rsp, Address(thread, JavaThread::reserved_stack_activation_offset())); 1077 jcc(Assembler::below, no_reserved_zone_enabling); 1078 1079 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), thread); 1080 jump(RuntimeAddress(StubRoutines::throw_delayed_StackOverflowError_entry())); 1081 should_not_reach_here(); 1082 1083 bind(no_reserved_zone_enabling); 1084} 1085 1086int MacroAssembler::biased_locking_enter(Register lock_reg, 1087 Register obj_reg, 1088 Register swap_reg, 1089 Register tmp_reg, 1090 bool swap_reg_contains_mark, 1091 Label& done, 1092 Label* slow_case, 1093 BiasedLockingCounters* counters) { 1094 assert(UseBiasedLocking, "why call this otherwise?"); 1095 assert(swap_reg == rax, "swap_reg must be rax for cmpxchgq"); 1096 assert(tmp_reg != noreg, "tmp_reg must be supplied"); 1097 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg); 1098 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout"); 1099 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes()); 1100 Address saved_mark_addr(lock_reg, 0); 1101 1102 if (PrintBiasedLockingStatistics && counters == NULL) { 1103 counters = BiasedLocking::counters(); 1104 } 1105 // Biased locking 1106 // See whether the lock is currently biased toward our thread and 1107 // whether the epoch is still valid 1108 // Note that the runtime guarantees sufficient alignment of JavaThread 1109 // pointers to allow age to be placed into low bits 1110 // First check to see whether biasing is even enabled for this object 1111 Label cas_label; 1112 int null_check_offset = -1; 1113 if (!swap_reg_contains_mark) { 1114 null_check_offset = offset(); 1115 movptr(swap_reg, mark_addr); 1116 } 1117 movptr(tmp_reg, swap_reg); 1118 andptr(tmp_reg, markOopDesc::biased_lock_mask_in_place); 1119 cmpptr(tmp_reg, markOopDesc::biased_lock_pattern); 1120 jcc(Assembler::notEqual, cas_label); 1121 // The bias pattern is present in the object's header. Need to check 1122 // whether the bias owner and the epoch are both still current. 1123#ifndef _LP64 1124 // Note that because there is no current thread register on x86_32 we 1125 // need to store off the mark word we read out of the object to 1126 // avoid reloading it and needing to recheck invariants below. This 1127 // store is unfortunate but it makes the overall code shorter and 1128 // simpler. 1129 movptr(saved_mark_addr, swap_reg); 1130#endif 1131 if (swap_reg_contains_mark) { 1132 null_check_offset = offset(); 1133 } 1134 load_prototype_header(tmp_reg, obj_reg); 1135#ifdef _LP64 1136 orptr(tmp_reg, r15_thread); 1137 xorptr(tmp_reg, swap_reg); 1138 Register header_reg = tmp_reg; 1139#else 1140 xorptr(tmp_reg, swap_reg); 1141 get_thread(swap_reg); 1142 xorptr(swap_reg, tmp_reg); 1143 Register header_reg = swap_reg; 1144#endif 1145 andptr(header_reg, ~((int) markOopDesc::age_mask_in_place)); 1146 if (counters != NULL) { 1147 cond_inc32(Assembler::zero, 1148 ExternalAddress((address) counters->biased_lock_entry_count_addr())); 1149 } 1150 jcc(Assembler::equal, done); 1151 1152 Label try_revoke_bias; 1153 Label try_rebias; 1154 1155 // At this point we know that the header has the bias pattern and 1156 // that we are not the bias owner in the current epoch. We need to 1157 // figure out more details about the state of the header in order to 1158 // know what operations can be legally performed on the object's 1159 // header. 1160 1161 // If the low three bits in the xor result aren't clear, that means 1162 // the prototype header is no longer biased and we have to revoke 1163 // the bias on this object. 1164 testptr(header_reg, markOopDesc::biased_lock_mask_in_place); 1165 jccb(Assembler::notZero, try_revoke_bias); 1166 1167 // Biasing is still enabled for this data type. See whether the 1168 // epoch of the current bias is still valid, meaning that the epoch 1169 // bits of the mark word are equal to the epoch bits of the 1170 // prototype header. (Note that the prototype header's epoch bits 1171 // only change at a safepoint.) If not, attempt to rebias the object 1172 // toward the current thread. Note that we must be absolutely sure 1173 // that the current epoch is invalid in order to do this because 1174 // otherwise the manipulations it performs on the mark word are 1175 // illegal. 1176 testptr(header_reg, markOopDesc::epoch_mask_in_place); 1177 jccb(Assembler::notZero, try_rebias); 1178 1179 // The epoch of the current bias is still valid but we know nothing 1180 // about the owner; it might be set or it might be clear. Try to 1181 // acquire the bias of the object using an atomic operation. If this 1182 // fails we will go in to the runtime to revoke the object's bias. 1183 // Note that we first construct the presumed unbiased header so we 1184 // don't accidentally blow away another thread's valid bias. 1185 NOT_LP64( movptr(swap_reg, saved_mark_addr); ) 1186 andptr(swap_reg, 1187 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place); 1188#ifdef _LP64 1189 movptr(tmp_reg, swap_reg); 1190 orptr(tmp_reg, r15_thread); 1191#else 1192 get_thread(tmp_reg); 1193 orptr(tmp_reg, swap_reg); 1194#endif 1195 if (os::is_MP()) { 1196 lock(); 1197 } 1198 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1199 // If the biasing toward our thread failed, this means that 1200 // another thread succeeded in biasing it toward itself and we 1201 // need to revoke that bias. The revocation will occur in the 1202 // interpreter runtime in the slow case. 1203 if (counters != NULL) { 1204 cond_inc32(Assembler::zero, 1205 ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr())); 1206 } 1207 if (slow_case != NULL) { 1208 jcc(Assembler::notZero, *slow_case); 1209 } 1210 jmp(done); 1211 1212 bind(try_rebias); 1213 // At this point we know the epoch has expired, meaning that the 1214 // current "bias owner", if any, is actually invalid. Under these 1215 // circumstances _only_, we are allowed to use the current header's 1216 // value as the comparison value when doing the cas to acquire the 1217 // bias in the current epoch. In other words, we allow transfer of 1218 // the bias from one thread to another directly in this situation. 1219 // 1220 // FIXME: due to a lack of registers we currently blow away the age 1221 // bits in this situation. Should attempt to preserve them. 1222 load_prototype_header(tmp_reg, obj_reg); 1223#ifdef _LP64 1224 orptr(tmp_reg, r15_thread); 1225#else 1226 get_thread(swap_reg); 1227 orptr(tmp_reg, swap_reg); 1228 movptr(swap_reg, saved_mark_addr); 1229#endif 1230 if (os::is_MP()) { 1231 lock(); 1232 } 1233 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1234 // If the biasing toward our thread failed, then another thread 1235 // succeeded in biasing it toward itself and we need to revoke that 1236 // bias. The revocation will occur in the runtime in the slow case. 1237 if (counters != NULL) { 1238 cond_inc32(Assembler::zero, 1239 ExternalAddress((address) counters->rebiased_lock_entry_count_addr())); 1240 } 1241 if (slow_case != NULL) { 1242 jcc(Assembler::notZero, *slow_case); 1243 } 1244 jmp(done); 1245 1246 bind(try_revoke_bias); 1247 // The prototype mark in the klass doesn't have the bias bit set any 1248 // more, indicating that objects of this data type are not supposed 1249 // to be biased any more. We are going to try to reset the mark of 1250 // this object to the prototype value and fall through to the 1251 // CAS-based locking scheme. Note that if our CAS fails, it means 1252 // that another thread raced us for the privilege of revoking the 1253 // bias of this particular object, so it's okay to continue in the 1254 // normal locking code. 1255 // 1256 // FIXME: due to a lack of registers we currently blow away the age 1257 // bits in this situation. Should attempt to preserve them. 1258 NOT_LP64( movptr(swap_reg, saved_mark_addr); ) 1259 load_prototype_header(tmp_reg, obj_reg); 1260 if (os::is_MP()) { 1261 lock(); 1262 } 1263 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1264 // Fall through to the normal CAS-based lock, because no matter what 1265 // the result of the above CAS, some thread must have succeeded in 1266 // removing the bias bit from the object's header. 1267 if (counters != NULL) { 1268 cond_inc32(Assembler::zero, 1269 ExternalAddress((address) counters->revoked_lock_entry_count_addr())); 1270 } 1271 1272 bind(cas_label); 1273 1274 return null_check_offset; 1275} 1276 1277void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) { 1278 assert(UseBiasedLocking, "why call this otherwise?"); 1279 1280 // Check for biased locking unlock case, which is a no-op 1281 // Note: we do not have to check the thread ID for two reasons. 1282 // First, the interpreter checks for IllegalMonitorStateException at 1283 // a higher level. Second, if the bias was revoked while we held the 1284 // lock, the object could not be rebiased toward another thread, so 1285 // the bias bit would be clear. 1286 movptr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 1287 andptr(temp_reg, markOopDesc::biased_lock_mask_in_place); 1288 cmpptr(temp_reg, markOopDesc::biased_lock_pattern); 1289 jcc(Assembler::equal, done); 1290} 1291 1292#ifdef COMPILER2 1293 1294#if INCLUDE_RTM_OPT 1295 1296// Update rtm_counters based on abort status 1297// input: abort_status 1298// rtm_counters (RTMLockingCounters*) 1299// flags are killed 1300void MacroAssembler::rtm_counters_update(Register abort_status, Register rtm_counters) { 1301 1302 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abort_count_offset())); 1303 if (PrintPreciseRTMLockingStatistics) { 1304 for (int i = 0; i < RTMLockingCounters::ABORT_STATUS_LIMIT; i++) { 1305 Label check_abort; 1306 testl(abort_status, (1<<i)); 1307 jccb(Assembler::equal, check_abort); 1308 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abortX_count_offset() + (i * sizeof(uintx)))); 1309 bind(check_abort); 1310 } 1311 } 1312} 1313 1314// Branch if (random & (count-1) != 0), count is 2^n 1315// tmp, scr and flags are killed 1316void MacroAssembler::branch_on_random_using_rdtsc(Register tmp, Register scr, int count, Label& brLabel) { 1317 assert(tmp == rax, ""); 1318 assert(scr == rdx, ""); 1319 rdtsc(); // modifies EDX:EAX 1320 andptr(tmp, count-1); 1321 jccb(Assembler::notZero, brLabel); 1322} 1323 1324// Perform abort ratio calculation, set no_rtm bit if high ratio 1325// input: rtm_counters_Reg (RTMLockingCounters* address) 1326// tmpReg, rtm_counters_Reg and flags are killed 1327void MacroAssembler::rtm_abort_ratio_calculation(Register tmpReg, 1328 Register rtm_counters_Reg, 1329 RTMLockingCounters* rtm_counters, 1330 Metadata* method_data) { 1331 Label L_done, L_check_always_rtm1, L_check_always_rtm2; 1332 1333 if (RTMLockingCalculationDelay > 0) { 1334 // Delay calculation 1335 movptr(tmpReg, ExternalAddress((address) RTMLockingCounters::rtm_calculation_flag_addr()), tmpReg); 1336 testptr(tmpReg, tmpReg); 1337 jccb(Assembler::equal, L_done); 1338 } 1339 // Abort ratio calculation only if abort_count > RTMAbortThreshold 1340 // Aborted transactions = abort_count * 100 1341 // All transactions = total_count * RTMTotalCountIncrRate 1342 // Set no_rtm bit if (Aborted transactions >= All transactions * RTMAbortRatio) 1343 1344 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::abort_count_offset())); 1345 cmpptr(tmpReg, RTMAbortThreshold); 1346 jccb(Assembler::below, L_check_always_rtm2); 1347 imulptr(tmpReg, tmpReg, 100); 1348 1349 Register scrReg = rtm_counters_Reg; 1350 movptr(scrReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset())); 1351 imulptr(scrReg, scrReg, RTMTotalCountIncrRate); 1352 imulptr(scrReg, scrReg, RTMAbortRatio); 1353 cmpptr(tmpReg, scrReg); 1354 jccb(Assembler::below, L_check_always_rtm1); 1355 if (method_data != NULL) { 1356 // set rtm_state to "no rtm" in MDO 1357 mov_metadata(tmpReg, method_data); 1358 if (os::is_MP()) { 1359 lock(); 1360 } 1361 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), NoRTM); 1362 } 1363 jmpb(L_done); 1364 bind(L_check_always_rtm1); 1365 // Reload RTMLockingCounters* address 1366 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters)); 1367 bind(L_check_always_rtm2); 1368 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset())); 1369 cmpptr(tmpReg, RTMLockingThreshold / RTMTotalCountIncrRate); 1370 jccb(Assembler::below, L_done); 1371 if (method_data != NULL) { 1372 // set rtm_state to "always rtm" in MDO 1373 mov_metadata(tmpReg, method_data); 1374 if (os::is_MP()) { 1375 lock(); 1376 } 1377 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), UseRTM); 1378 } 1379 bind(L_done); 1380} 1381 1382// Update counters and perform abort ratio calculation 1383// input: abort_status_Reg 1384// rtm_counters_Reg, flags are killed 1385void MacroAssembler::rtm_profiling(Register abort_status_Reg, 1386 Register rtm_counters_Reg, 1387 RTMLockingCounters* rtm_counters, 1388 Metadata* method_data, 1389 bool profile_rtm) { 1390 1391 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1392 // update rtm counters based on rax value at abort 1393 // reads abort_status_Reg, updates flags 1394 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters)); 1395 rtm_counters_update(abort_status_Reg, rtm_counters_Reg); 1396 if (profile_rtm) { 1397 // Save abort status because abort_status_Reg is used by following code. 1398 if (RTMRetryCount > 0) { 1399 push(abort_status_Reg); 1400 } 1401 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1402 rtm_abort_ratio_calculation(abort_status_Reg, rtm_counters_Reg, rtm_counters, method_data); 1403 // restore abort status 1404 if (RTMRetryCount > 0) { 1405 pop(abort_status_Reg); 1406 } 1407 } 1408} 1409 1410// Retry on abort if abort's status is 0x6: can retry (0x2) | memory conflict (0x4) 1411// inputs: retry_count_Reg 1412// : abort_status_Reg 1413// output: retry_count_Reg decremented by 1 1414// flags are killed 1415void MacroAssembler::rtm_retry_lock_on_abort(Register retry_count_Reg, Register abort_status_Reg, Label& retryLabel) { 1416 Label doneRetry; 1417 assert(abort_status_Reg == rax, ""); 1418 // The abort reason bits are in eax (see all states in rtmLocking.hpp) 1419 // 0x6 = conflict on which we can retry (0x2) | memory conflict (0x4) 1420 // if reason is in 0x6 and retry count != 0 then retry 1421 andptr(abort_status_Reg, 0x6); 1422 jccb(Assembler::zero, doneRetry); 1423 testl(retry_count_Reg, retry_count_Reg); 1424 jccb(Assembler::zero, doneRetry); 1425 pause(); 1426 decrementl(retry_count_Reg); 1427 jmp(retryLabel); 1428 bind(doneRetry); 1429} 1430 1431// Spin and retry if lock is busy, 1432// inputs: box_Reg (monitor address) 1433// : retry_count_Reg 1434// output: retry_count_Reg decremented by 1 1435// : clear z flag if retry count exceeded 1436// tmp_Reg, scr_Reg, flags are killed 1437void MacroAssembler::rtm_retry_lock_on_busy(Register retry_count_Reg, Register box_Reg, 1438 Register tmp_Reg, Register scr_Reg, Label& retryLabel) { 1439 Label SpinLoop, SpinExit, doneRetry; 1440 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1441 1442 testl(retry_count_Reg, retry_count_Reg); 1443 jccb(Assembler::zero, doneRetry); 1444 decrementl(retry_count_Reg); 1445 movptr(scr_Reg, RTMSpinLoopCount); 1446 1447 bind(SpinLoop); 1448 pause(); 1449 decrementl(scr_Reg); 1450 jccb(Assembler::lessEqual, SpinExit); 1451 movptr(tmp_Reg, Address(box_Reg, owner_offset)); 1452 testptr(tmp_Reg, tmp_Reg); 1453 jccb(Assembler::notZero, SpinLoop); 1454 1455 bind(SpinExit); 1456 jmp(retryLabel); 1457 bind(doneRetry); 1458 incrementl(retry_count_Reg); // clear z flag 1459} 1460 1461// Use RTM for normal stack locks 1462// Input: objReg (object to lock) 1463void MacroAssembler::rtm_stack_locking(Register objReg, Register tmpReg, Register scrReg, 1464 Register retry_on_abort_count_Reg, 1465 RTMLockingCounters* stack_rtm_counters, 1466 Metadata* method_data, bool profile_rtm, 1467 Label& DONE_LABEL, Label& IsInflated) { 1468 assert(UseRTMForStackLocks, "why call this otherwise?"); 1469 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking"); 1470 assert(tmpReg == rax, ""); 1471 assert(scrReg == rdx, ""); 1472 Label L_rtm_retry, L_decrement_retry, L_on_abort; 1473 1474 if (RTMRetryCount > 0) { 1475 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort 1476 bind(L_rtm_retry); 1477 } 1478 movptr(tmpReg, Address(objReg, 0)); 1479 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased 1480 jcc(Assembler::notZero, IsInflated); 1481 1482 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1483 Label L_noincrement; 1484 if (RTMTotalCountIncrRate > 1) { 1485 // tmpReg, scrReg and flags are killed 1486 branch_on_random_using_rdtsc(tmpReg, scrReg, (int)RTMTotalCountIncrRate, L_noincrement); 1487 } 1488 assert(stack_rtm_counters != NULL, "should not be NULL when profiling RTM"); 1489 atomic_incptr(ExternalAddress((address)stack_rtm_counters->total_count_addr()), scrReg); 1490 bind(L_noincrement); 1491 } 1492 xbegin(L_on_abort); 1493 movptr(tmpReg, Address(objReg, 0)); // fetch markword 1494 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits 1495 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked 1496 jcc(Assembler::equal, DONE_LABEL); // all done if unlocked 1497 1498 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX 1499 if (UseRTMXendForLockBusy) { 1500 xend(); 1501 movptr(abort_status_Reg, 0x2); // Set the abort status to 2 (so we can retry) 1502 jmp(L_decrement_retry); 1503 } 1504 else { 1505 xabort(0); 1506 } 1507 bind(L_on_abort); 1508 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1509 rtm_profiling(abort_status_Reg, scrReg, stack_rtm_counters, method_data, profile_rtm); 1510 } 1511 bind(L_decrement_retry); 1512 if (RTMRetryCount > 0) { 1513 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4) 1514 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry); 1515 } 1516} 1517 1518// Use RTM for inflating locks 1519// inputs: objReg (object to lock) 1520// boxReg (on-stack box address (displaced header location) - KILLED) 1521// tmpReg (ObjectMonitor address + markOopDesc::monitor_value) 1522void MacroAssembler::rtm_inflated_locking(Register objReg, Register boxReg, Register tmpReg, 1523 Register scrReg, Register retry_on_busy_count_Reg, 1524 Register retry_on_abort_count_Reg, 1525 RTMLockingCounters* rtm_counters, 1526 Metadata* method_data, bool profile_rtm, 1527 Label& DONE_LABEL) { 1528 assert(UseRTMLocking, "why call this otherwise?"); 1529 assert(tmpReg == rax, ""); 1530 assert(scrReg == rdx, ""); 1531 Label L_rtm_retry, L_decrement_retry, L_on_abort; 1532 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1533 1534 // Without cast to int32_t a movptr will destroy r10 which is typically obj 1535 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark())); 1536 movptr(boxReg, tmpReg); // Save ObjectMonitor address 1537 1538 if (RTMRetryCount > 0) { 1539 movl(retry_on_busy_count_Reg, RTMRetryCount); // Retry on lock busy 1540 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort 1541 bind(L_rtm_retry); 1542 } 1543 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1544 Label L_noincrement; 1545 if (RTMTotalCountIncrRate > 1) { 1546 // tmpReg, scrReg and flags are killed 1547 branch_on_random_using_rdtsc(tmpReg, scrReg, (int)RTMTotalCountIncrRate, L_noincrement); 1548 } 1549 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1550 atomic_incptr(ExternalAddress((address)rtm_counters->total_count_addr()), scrReg); 1551 bind(L_noincrement); 1552 } 1553 xbegin(L_on_abort); 1554 movptr(tmpReg, Address(objReg, 0)); 1555 movptr(tmpReg, Address(tmpReg, owner_offset)); 1556 testptr(tmpReg, tmpReg); 1557 jcc(Assembler::zero, DONE_LABEL); 1558 if (UseRTMXendForLockBusy) { 1559 xend(); 1560 jmp(L_decrement_retry); 1561 } 1562 else { 1563 xabort(0); 1564 } 1565 bind(L_on_abort); 1566 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX 1567 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1568 rtm_profiling(abort_status_Reg, scrReg, rtm_counters, method_data, profile_rtm); 1569 } 1570 if (RTMRetryCount > 0) { 1571 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4) 1572 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry); 1573 } 1574 1575 movptr(tmpReg, Address(boxReg, owner_offset)) ; 1576 testptr(tmpReg, tmpReg) ; 1577 jccb(Assembler::notZero, L_decrement_retry) ; 1578 1579 // Appears unlocked - try to swing _owner from null to non-null. 1580 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1581#ifdef _LP64 1582 Register threadReg = r15_thread; 1583#else 1584 get_thread(scrReg); 1585 Register threadReg = scrReg; 1586#endif 1587 if (os::is_MP()) { 1588 lock(); 1589 } 1590 cmpxchgptr(threadReg, Address(boxReg, owner_offset)); // Updates tmpReg 1591 1592 if (RTMRetryCount > 0) { 1593 // success done else retry 1594 jccb(Assembler::equal, DONE_LABEL) ; 1595 bind(L_decrement_retry); 1596 // Spin and retry if lock is busy. 1597 rtm_retry_lock_on_busy(retry_on_busy_count_Reg, boxReg, tmpReg, scrReg, L_rtm_retry); 1598 } 1599 else { 1600 bind(L_decrement_retry); 1601 } 1602} 1603 1604#endif // INCLUDE_RTM_OPT 1605 1606// Fast_Lock and Fast_Unlock used by C2 1607 1608// Because the transitions from emitted code to the runtime 1609// monitorenter/exit helper stubs are so slow it's critical that 1610// we inline both the stack-locking fast-path and the inflated fast path. 1611// 1612// See also: cmpFastLock and cmpFastUnlock. 1613// 1614// What follows is a specialized inline transliteration of the code 1615// in slow_enter() and slow_exit(). If we're concerned about I$ bloat 1616// another option would be to emit TrySlowEnter and TrySlowExit methods 1617// at startup-time. These methods would accept arguments as 1618// (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure 1619// indications in the icc.ZFlag. Fast_Lock and Fast_Unlock would simply 1620// marshal the arguments and emit calls to TrySlowEnter and TrySlowExit. 1621// In practice, however, the # of lock sites is bounded and is usually small. 1622// Besides the call overhead, TrySlowEnter and TrySlowExit might suffer 1623// if the processor uses simple bimodal branch predictors keyed by EIP 1624// Since the helper routines would be called from multiple synchronization 1625// sites. 1626// 1627// An even better approach would be write "MonitorEnter()" and "MonitorExit()" 1628// in java - using j.u.c and unsafe - and just bind the lock and unlock sites 1629// to those specialized methods. That'd give us a mostly platform-independent 1630// implementation that the JITs could optimize and inline at their pleasure. 1631// Done correctly, the only time we'd need to cross to native could would be 1632// to park() or unpark() threads. We'd also need a few more unsafe operators 1633// to (a) prevent compiler-JIT reordering of non-volatile accesses, and 1634// (b) explicit barriers or fence operations. 1635// 1636// TODO: 1637// 1638// * Arrange for C2 to pass "Self" into Fast_Lock and Fast_Unlock in one of the registers (scr). 1639// This avoids manifesting the Self pointer in the Fast_Lock and Fast_Unlock terminals. 1640// Given TLAB allocation, Self is usually manifested in a register, so passing it into 1641// the lock operators would typically be faster than reifying Self. 1642// 1643// * Ideally I'd define the primitives as: 1644// fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED. 1645// fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED 1646// Unfortunately ADLC bugs prevent us from expressing the ideal form. 1647// Instead, we're stuck with a rather awkward and brittle register assignments below. 1648// Furthermore the register assignments are overconstrained, possibly resulting in 1649// sub-optimal code near the synchronization site. 1650// 1651// * Eliminate the sp-proximity tests and just use "== Self" tests instead. 1652// Alternately, use a better sp-proximity test. 1653// 1654// * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value. 1655// Either one is sufficient to uniquely identify a thread. 1656// TODO: eliminate use of sp in _owner and use get_thread(tr) instead. 1657// 1658// * Intrinsify notify() and notifyAll() for the common cases where the 1659// object is locked by the calling thread but the waitlist is empty. 1660// avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll(). 1661// 1662// * use jccb and jmpb instead of jcc and jmp to improve code density. 1663// But beware of excessive branch density on AMD Opterons. 1664// 1665// * Both Fast_Lock and Fast_Unlock set the ICC.ZF to indicate success 1666// or failure of the fast-path. If the fast-path fails then we pass 1667// control to the slow-path, typically in C. In Fast_Lock and 1668// Fast_Unlock we often branch to DONE_LABEL, just to find that C2 1669// will emit a conditional branch immediately after the node. 1670// So we have branches to branches and lots of ICC.ZF games. 1671// Instead, it might be better to have C2 pass a "FailureLabel" 1672// into Fast_Lock and Fast_Unlock. In the case of success, control 1673// will drop through the node. ICC.ZF is undefined at exit. 1674// In the case of failure, the node will branch directly to the 1675// FailureLabel 1676 1677 1678// obj: object to lock 1679// box: on-stack box address (displaced header location) - KILLED 1680// rax,: tmp -- KILLED 1681// scr: tmp -- KILLED 1682void MacroAssembler::fast_lock(Register objReg, Register boxReg, Register tmpReg, 1683 Register scrReg, Register cx1Reg, Register cx2Reg, 1684 BiasedLockingCounters* counters, 1685 RTMLockingCounters* rtm_counters, 1686 RTMLockingCounters* stack_rtm_counters, 1687 Metadata* method_data, 1688 bool use_rtm, bool profile_rtm) { 1689 // Ensure the register assignents are disjoint 1690 assert(tmpReg == rax, ""); 1691 1692 if (use_rtm) { 1693 assert_different_registers(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg); 1694 } else { 1695 assert(cx1Reg == noreg, ""); 1696 assert(cx2Reg == noreg, ""); 1697 assert_different_registers(objReg, boxReg, tmpReg, scrReg); 1698 } 1699 1700 if (counters != NULL) { 1701 atomic_incl(ExternalAddress((address)counters->total_entry_count_addr()), scrReg); 1702 } 1703 if (EmitSync & 1) { 1704 // set box->dhw = markOopDesc::unused_mark() 1705 // Force all sync thru slow-path: slow_enter() and slow_exit() 1706 movptr (Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark())); 1707 cmpptr (rsp, (int32_t)NULL_WORD); 1708 } else { 1709 // Possible cases that we'll encounter in fast_lock 1710 // ------------------------------------------------ 1711 // * Inflated 1712 // -- unlocked 1713 // -- Locked 1714 // = by self 1715 // = by other 1716 // * biased 1717 // -- by Self 1718 // -- by other 1719 // * neutral 1720 // * stack-locked 1721 // -- by self 1722 // = sp-proximity test hits 1723 // = sp-proximity test generates false-negative 1724 // -- by other 1725 // 1726 1727 Label IsInflated, DONE_LABEL; 1728 1729 // it's stack-locked, biased or neutral 1730 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage 1731 // order to reduce the number of conditional branches in the most common cases. 1732 // Beware -- there's a subtle invariant that fetch of the markword 1733 // at [FETCH], below, will never observe a biased encoding (*101b). 1734 // If this invariant is not held we risk exclusion (safety) failure. 1735 if (UseBiasedLocking && !UseOptoBiasInlining) { 1736 biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, counters); 1737 } 1738 1739#if INCLUDE_RTM_OPT 1740 if (UseRTMForStackLocks && use_rtm) { 1741 rtm_stack_locking(objReg, tmpReg, scrReg, cx2Reg, 1742 stack_rtm_counters, method_data, profile_rtm, 1743 DONE_LABEL, IsInflated); 1744 } 1745#endif // INCLUDE_RTM_OPT 1746 1747 movptr(tmpReg, Address(objReg, 0)); // [FETCH] 1748 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased 1749 jccb(Assembler::notZero, IsInflated); 1750 1751 // Attempt stack-locking ... 1752 orptr (tmpReg, markOopDesc::unlocked_value); 1753 movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS 1754 if (os::is_MP()) { 1755 lock(); 1756 } 1757 cmpxchgptr(boxReg, Address(objReg, 0)); // Updates tmpReg 1758 if (counters != NULL) { 1759 cond_inc32(Assembler::equal, 1760 ExternalAddress((address)counters->fast_path_entry_count_addr())); 1761 } 1762 jcc(Assembler::equal, DONE_LABEL); // Success 1763 1764 // Recursive locking. 1765 // The object is stack-locked: markword contains stack pointer to BasicLock. 1766 // Locked by current thread if difference with current SP is less than one page. 1767 subptr(tmpReg, rsp); 1768 // Next instruction set ZFlag == 1 (Success) if difference is less then one page. 1769 andptr(tmpReg, (int32_t) (NOT_LP64(0xFFFFF003) LP64_ONLY(7 - os::vm_page_size())) ); 1770 movptr(Address(boxReg, 0), tmpReg); 1771 if (counters != NULL) { 1772 cond_inc32(Assembler::equal, 1773 ExternalAddress((address)counters->fast_path_entry_count_addr())); 1774 } 1775 jmp(DONE_LABEL); 1776 1777 bind(IsInflated); 1778 // The object is inflated. tmpReg contains pointer to ObjectMonitor* + markOopDesc::monitor_value 1779 1780#if INCLUDE_RTM_OPT 1781 // Use the same RTM locking code in 32- and 64-bit VM. 1782 if (use_rtm) { 1783 rtm_inflated_locking(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg, 1784 rtm_counters, method_data, profile_rtm, DONE_LABEL); 1785 } else { 1786#endif // INCLUDE_RTM_OPT 1787 1788#ifndef _LP64 1789 // The object is inflated. 1790 1791 // boxReg refers to the on-stack BasicLock in the current frame. 1792 // We'd like to write: 1793 // set box->_displaced_header = markOopDesc::unused_mark(). Any non-0 value suffices. 1794 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers 1795 // additional latency as we have another ST in the store buffer that must drain. 1796 1797 if (EmitSync & 8192) { 1798 movptr(Address(boxReg, 0), 3); // results in ST-before-CAS penalty 1799 get_thread (scrReg); 1800 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2] 1801 movptr(tmpReg, NULL_WORD); // consider: xor vs mov 1802 if (os::is_MP()) { 1803 lock(); 1804 } 1805 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1806 } else 1807 if ((EmitSync & 128) == 0) { // avoid ST-before-CAS 1808 // register juggle because we need tmpReg for cmpxchgptr below 1809 movptr(scrReg, boxReg); 1810 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2] 1811 1812 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes 1813 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) { 1814 // prefetchw [eax + Offset(_owner)-2] 1815 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1816 } 1817 1818 if ((EmitSync & 64) == 0) { 1819 // Optimistic form: consider XORL tmpReg,tmpReg 1820 movptr(tmpReg, NULL_WORD); 1821 } else { 1822 // Can suffer RTS->RTO upgrades on shared or cold $ lines 1823 // Test-And-CAS instead of CAS 1824 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner 1825 testptr(tmpReg, tmpReg); // Locked ? 1826 jccb (Assembler::notZero, DONE_LABEL); 1827 } 1828 1829 // Appears unlocked - try to swing _owner from null to non-null. 1830 // Ideally, I'd manifest "Self" with get_thread and then attempt 1831 // to CAS the register containing Self into m->Owner. 1832 // But we don't have enough registers, so instead we can either try to CAS 1833 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds 1834 // we later store "Self" into m->Owner. Transiently storing a stack address 1835 // (rsp or the address of the box) into m->owner is harmless. 1836 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1837 if (os::is_MP()) { 1838 lock(); 1839 } 1840 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1841 movptr(Address(scrReg, 0), 3); // box->_displaced_header = 3 1842 // If we weren't able to swing _owner from NULL to the BasicLock 1843 // then take the slow path. 1844 jccb (Assembler::notZero, DONE_LABEL); 1845 // update _owner from BasicLock to thread 1846 get_thread (scrReg); // beware: clobbers ICCs 1847 movptr(Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), scrReg); 1848 xorptr(boxReg, boxReg); // set icc.ZFlag = 1 to indicate success 1849 1850 // If the CAS fails we can either retry or pass control to the slow-path. 1851 // We use the latter tactic. 1852 // Pass the CAS result in the icc.ZFlag into DONE_LABEL 1853 // If the CAS was successful ... 1854 // Self has acquired the lock 1855 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it. 1856 // Intentional fall-through into DONE_LABEL ... 1857 } else { 1858 movptr(Address(boxReg, 0), intptr_t(markOopDesc::unused_mark())); // results in ST-before-CAS penalty 1859 movptr(boxReg, tmpReg); 1860 1861 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes 1862 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) { 1863 // prefetchw [eax + Offset(_owner)-2] 1864 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1865 } 1866 1867 if ((EmitSync & 64) == 0) { 1868 // Optimistic form 1869 xorptr (tmpReg, tmpReg); 1870 } else { 1871 // Can suffer RTS->RTO upgrades on shared or cold $ lines 1872 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner 1873 testptr(tmpReg, tmpReg); // Locked ? 1874 jccb (Assembler::notZero, DONE_LABEL); 1875 } 1876 1877 // Appears unlocked - try to swing _owner from null to non-null. 1878 // Use either "Self" (in scr) or rsp as thread identity in _owner. 1879 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1880 get_thread (scrReg); 1881 if (os::is_MP()) { 1882 lock(); 1883 } 1884 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1885 1886 // If the CAS fails we can either retry or pass control to the slow-path. 1887 // We use the latter tactic. 1888 // Pass the CAS result in the icc.ZFlag into DONE_LABEL 1889 // If the CAS was successful ... 1890 // Self has acquired the lock 1891 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it. 1892 // Intentional fall-through into DONE_LABEL ... 1893 } 1894#else // _LP64 1895 // It's inflated 1896 movq(scrReg, tmpReg); 1897 xorq(tmpReg, tmpReg); 1898 1899 if (os::is_MP()) { 1900 lock(); 1901 } 1902 cmpxchgptr(r15_thread, Address(scrReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1903 // Unconditionally set box->_displaced_header = markOopDesc::unused_mark(). 1904 // Without cast to int32_t movptr will destroy r10 which is typically obj. 1905 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark())); 1906 // Intentional fall-through into DONE_LABEL ... 1907 // Propagate ICC.ZF from CAS above into DONE_LABEL. 1908#endif // _LP64 1909#if INCLUDE_RTM_OPT 1910 } // use_rtm() 1911#endif 1912 // DONE_LABEL is a hot target - we'd really like to place it at the 1913 // start of cache line by padding with NOPs. 1914 // See the AMD and Intel software optimization manuals for the 1915 // most efficient "long" NOP encodings. 1916 // Unfortunately none of our alignment mechanisms suffice. 1917 bind(DONE_LABEL); 1918 1919 // At DONE_LABEL the icc ZFlag is set as follows ... 1920 // Fast_Unlock uses the same protocol. 1921 // ZFlag == 1 -> Success 1922 // ZFlag == 0 -> Failure - force control through the slow-path 1923 } 1924} 1925 1926// obj: object to unlock 1927// box: box address (displaced header location), killed. Must be EAX. 1928// tmp: killed, cannot be obj nor box. 1929// 1930// Some commentary on balanced locking: 1931// 1932// Fast_Lock and Fast_Unlock are emitted only for provably balanced lock sites. 1933// Methods that don't have provably balanced locking are forced to run in the 1934// interpreter - such methods won't be compiled to use fast_lock and fast_unlock. 1935// The interpreter provides two properties: 1936// I1: At return-time the interpreter automatically and quietly unlocks any 1937// objects acquired the current activation (frame). Recall that the 1938// interpreter maintains an on-stack list of locks currently held by 1939// a frame. 1940// I2: If a method attempts to unlock an object that is not held by the 1941// the frame the interpreter throws IMSX. 1942// 1943// Lets say A(), which has provably balanced locking, acquires O and then calls B(). 1944// B() doesn't have provably balanced locking so it runs in the interpreter. 1945// Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O 1946// is still locked by A(). 1947// 1948// The only other source of unbalanced locking would be JNI. The "Java Native Interface: 1949// Programmer's Guide and Specification" claims that an object locked by jni_monitorenter 1950// should not be unlocked by "normal" java-level locking and vice-versa. The specification 1951// doesn't specify what will occur if a program engages in such mixed-mode locking, however. 1952// Arguably given that the spec legislates the JNI case as undefined our implementation 1953// could reasonably *avoid* checking owner in Fast_Unlock(). 1954// In the interest of performance we elide m->Owner==Self check in unlock. 1955// A perfectly viable alternative is to elide the owner check except when 1956// Xcheck:jni is enabled. 1957 1958void MacroAssembler::fast_unlock(Register objReg, Register boxReg, Register tmpReg, bool use_rtm) { 1959 assert(boxReg == rax, ""); 1960 assert_different_registers(objReg, boxReg, tmpReg); 1961 1962 if (EmitSync & 4) { 1963 // Disable - inhibit all inlining. Force control through the slow-path 1964 cmpptr (rsp, 0); 1965 } else { 1966 Label DONE_LABEL, Stacked, CheckSucc; 1967 1968 // Critically, the biased locking test must have precedence over 1969 // and appear before the (box->dhw == 0) recursive stack-lock test. 1970 if (UseBiasedLocking && !UseOptoBiasInlining) { 1971 biased_locking_exit(objReg, tmpReg, DONE_LABEL); 1972 } 1973 1974#if INCLUDE_RTM_OPT 1975 if (UseRTMForStackLocks && use_rtm) { 1976 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking"); 1977 Label L_regular_unlock; 1978 movptr(tmpReg, Address(objReg, 0)); // fetch markword 1979 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits 1980 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked 1981 jccb(Assembler::notEqual, L_regular_unlock); // if !HLE RegularLock 1982 xend(); // otherwise end... 1983 jmp(DONE_LABEL); // ... and we're done 1984 bind(L_regular_unlock); 1985 } 1986#endif 1987 1988 cmpptr(Address(boxReg, 0), (int32_t)NULL_WORD); // Examine the displaced header 1989 jcc (Assembler::zero, DONE_LABEL); // 0 indicates recursive stack-lock 1990 movptr(tmpReg, Address(objReg, 0)); // Examine the object's markword 1991 testptr(tmpReg, markOopDesc::monitor_value); // Inflated? 1992 jccb (Assembler::zero, Stacked); 1993 1994 // It's inflated. 1995#if INCLUDE_RTM_OPT 1996 if (use_rtm) { 1997 Label L_regular_inflated_unlock; 1998 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1999 movptr(boxReg, Address(tmpReg, owner_offset)); 2000 testptr(boxReg, boxReg); 2001 jccb(Assembler::notZero, L_regular_inflated_unlock); 2002 xend(); 2003 jmpb(DONE_LABEL); 2004 bind(L_regular_inflated_unlock); 2005 } 2006#endif 2007 2008 // Despite our balanced locking property we still check that m->_owner == Self 2009 // as java routines or native JNI code called by this thread might 2010 // have released the lock. 2011 // Refer to the comments in synchronizer.cpp for how we might encode extra 2012 // state in _succ so we can avoid fetching EntryList|cxq. 2013 // 2014 // I'd like to add more cases in fast_lock() and fast_unlock() -- 2015 // such as recursive enter and exit -- but we have to be wary of 2016 // I$ bloat, T$ effects and BP$ effects. 2017 // 2018 // If there's no contention try a 1-0 exit. That is, exit without 2019 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how 2020 // we detect and recover from the race that the 1-0 exit admits. 2021 // 2022 // Conceptually Fast_Unlock() must execute a STST|LDST "release" barrier 2023 // before it STs null into _owner, releasing the lock. Updates 2024 // to data protected by the critical section must be visible before 2025 // we drop the lock (and thus before any other thread could acquire 2026 // the lock and observe the fields protected by the lock). 2027 // IA32's memory-model is SPO, so STs are ordered with respect to 2028 // each other and there's no need for an explicit barrier (fence). 2029 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 2030#ifndef _LP64 2031 get_thread (boxReg); 2032 if ((EmitSync & 4096) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) { 2033 // prefetchw [ebx + Offset(_owner)-2] 2034 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2035 } 2036 2037 // Note that we could employ various encoding schemes to reduce 2038 // the number of loads below (currently 4) to just 2 or 3. 2039 // Refer to the comments in synchronizer.cpp. 2040 // In practice the chain of fetches doesn't seem to impact performance, however. 2041 xorptr(boxReg, boxReg); 2042 if ((EmitSync & 65536) == 0 && (EmitSync & 256)) { 2043 // Attempt to reduce branch density - AMD's branch predictor. 2044 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 2045 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2046 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2047 jccb (Assembler::notZero, DONE_LABEL); 2048 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD); 2049 jmpb (DONE_LABEL); 2050 } else { 2051 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 2052 jccb (Assembler::notZero, DONE_LABEL); 2053 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2054 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2055 jccb (Assembler::notZero, CheckSucc); 2056 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD); 2057 jmpb (DONE_LABEL); 2058 } 2059 2060 // The Following code fragment (EmitSync & 65536) improves the performance of 2061 // contended applications and contended synchronization microbenchmarks. 2062 // Unfortunately the emission of the code - even though not executed - causes regressions 2063 // in scimark and jetstream, evidently because of $ effects. Replacing the code 2064 // with an equal number of never-executed NOPs results in the same regression. 2065 // We leave it off by default. 2066 2067 if ((EmitSync & 65536) != 0) { 2068 Label LSuccess, LGoSlowPath ; 2069 2070 bind (CheckSucc); 2071 2072 // Optional pre-test ... it's safe to elide this 2073 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2074 jccb(Assembler::zero, LGoSlowPath); 2075 2076 // We have a classic Dekker-style idiom: 2077 // ST m->_owner = 0 ; MEMBAR; LD m->_succ 2078 // There are a number of ways to implement the barrier: 2079 // (1) lock:andl &m->_owner, 0 2080 // is fast, but mask doesn't currently support the "ANDL M,IMM32" form. 2081 // LOCK: ANDL [ebx+Offset(_Owner)-2], 0 2082 // Encodes as 81 31 OFF32 IMM32 or 83 63 OFF8 IMM8 2083 // (2) If supported, an explicit MFENCE is appealing. 2084 // In older IA32 processors MFENCE is slower than lock:add or xchg 2085 // particularly if the write-buffer is full as might be the case if 2086 // if stores closely precede the fence or fence-equivalent instruction. 2087 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences 2088 // as the situation has changed with Nehalem and Shanghai. 2089 // (3) In lieu of an explicit fence, use lock:addl to the top-of-stack 2090 // The $lines underlying the top-of-stack should be in M-state. 2091 // The locked add instruction is serializing, of course. 2092 // (4) Use xchg, which is serializing 2093 // mov boxReg, 0; xchgl boxReg, [tmpReg + Offset(_owner)-2] also works 2094 // (5) ST m->_owner = 0 and then execute lock:orl &m->_succ, 0. 2095 // The integer condition codes will tell us if succ was 0. 2096 // Since _succ and _owner should reside in the same $line and 2097 // we just stored into _owner, it's likely that the $line 2098 // remains in M-state for the lock:orl. 2099 // 2100 // We currently use (3), although it's likely that switching to (2) 2101 // is correct for the future. 2102 2103 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD); 2104 if (os::is_MP()) { 2105 lock(); addptr(Address(rsp, 0), 0); 2106 } 2107 // Ratify _succ remains non-null 2108 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), 0); 2109 jccb (Assembler::notZero, LSuccess); 2110 2111 xorptr(boxReg, boxReg); // box is really EAX 2112 if (os::is_MP()) { lock(); } 2113 cmpxchgptr(rsp, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2114 // There's no successor so we tried to regrab the lock with the 2115 // placeholder value. If that didn't work, then another thread 2116 // grabbed the lock so we're done (and exit was a success). 2117 jccb (Assembler::notEqual, LSuccess); 2118 // Since we're low on registers we installed rsp as a placeholding in _owner. 2119 // Now install Self over rsp. This is safe as we're transitioning from 2120 // non-null to non=null 2121 get_thread (boxReg); 2122 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), boxReg); 2123 // Intentional fall-through into LGoSlowPath ... 2124 2125 bind (LGoSlowPath); 2126 orptr(boxReg, 1); // set ICC.ZF=0 to indicate failure 2127 jmpb (DONE_LABEL); 2128 2129 bind (LSuccess); 2130 xorptr(boxReg, boxReg); // set ICC.ZF=1 to indicate success 2131 jmpb (DONE_LABEL); 2132 } 2133 2134 bind (Stacked); 2135 // It's not inflated and it's not recursively stack-locked and it's not biased. 2136 // It must be stack-locked. 2137 // Try to reset the header to displaced header. 2138 // The "box" value on the stack is stable, so we can reload 2139 // and be assured we observe the same value as above. 2140 movptr(tmpReg, Address(boxReg, 0)); 2141 if (os::is_MP()) { 2142 lock(); 2143 } 2144 cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses RAX which is box 2145 // Intention fall-thru into DONE_LABEL 2146 2147 // DONE_LABEL is a hot target - we'd really like to place it at the 2148 // start of cache line by padding with NOPs. 2149 // See the AMD and Intel software optimization manuals for the 2150 // most efficient "long" NOP encodings. 2151 // Unfortunately none of our alignment mechanisms suffice. 2152 if ((EmitSync & 65536) == 0) { 2153 bind (CheckSucc); 2154 } 2155#else // _LP64 2156 // It's inflated 2157 if (EmitSync & 1024) { 2158 // Emit code to check that _owner == Self 2159 // We could fold the _owner test into subsequent code more efficiently 2160 // than using a stand-alone check, but since _owner checking is off by 2161 // default we don't bother. We also might consider predicating the 2162 // _owner==Self check on Xcheck:jni or running on a debug build. 2163 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2164 xorptr(boxReg, r15_thread); 2165 } else { 2166 xorptr(boxReg, boxReg); 2167 } 2168 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 2169 jccb (Assembler::notZero, DONE_LABEL); 2170 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2171 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2172 jccb (Assembler::notZero, CheckSucc); 2173 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD); 2174 jmpb (DONE_LABEL); 2175 2176 if ((EmitSync & 65536) == 0) { 2177 // Try to avoid passing control into the slow_path ... 2178 Label LSuccess, LGoSlowPath ; 2179 bind (CheckSucc); 2180 2181 // The following optional optimization can be elided if necessary 2182 // Effectively: if (succ == null) goto SlowPath 2183 // The code reduces the window for a race, however, 2184 // and thus benefits performance. 2185 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2186 jccb (Assembler::zero, LGoSlowPath); 2187 2188 if ((EmitSync & 16) && os::is_MP()) { 2189 orptr(boxReg, boxReg); 2190 xchgptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2191 } else { 2192 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD); 2193 if (os::is_MP()) { 2194 // Memory barrier/fence 2195 // Dekker pivot point -- fulcrum : ST Owner; MEMBAR; LD Succ 2196 // Instead of MFENCE we use a dummy locked add of 0 to the top-of-stack. 2197 // This is faster on Nehalem and AMD Shanghai/Barcelona. 2198 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences 2199 // We might also restructure (ST Owner=0;barrier;LD _Succ) to 2200 // (mov box,0; xchgq box, &m->Owner; LD _succ) . 2201 lock(); addl(Address(rsp, 0), 0); 2202 } 2203 } 2204 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2205 jccb (Assembler::notZero, LSuccess); 2206 2207 // Rare inopportune interleaving - race. 2208 // The successor vanished in the small window above. 2209 // The lock is contended -- (cxq|EntryList) != null -- and there's no apparent successor. 2210 // We need to ensure progress and succession. 2211 // Try to reacquire the lock. 2212 // If that fails then the new owner is responsible for succession and this 2213 // thread needs to take no further action and can exit via the fast path (success). 2214 // If the re-acquire succeeds then pass control into the slow path. 2215 // As implemented, this latter mode is horrible because we generated more 2216 // coherence traffic on the lock *and* artifically extended the critical section 2217 // length while by virtue of passing control into the slow path. 2218 2219 // box is really RAX -- the following CMPXCHG depends on that binding 2220 // cmpxchg R,[M] is equivalent to rax = CAS(M,rax,R) 2221 movptr(boxReg, (int32_t)NULL_WORD); 2222 if (os::is_MP()) { lock(); } 2223 cmpxchgptr(r15_thread, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2224 // There's no successor so we tried to regrab the lock. 2225 // If that didn't work, then another thread grabbed the 2226 // lock so we're done (and exit was a success). 2227 jccb (Assembler::notEqual, LSuccess); 2228 // Intentional fall-through into slow-path 2229 2230 bind (LGoSlowPath); 2231 orl (boxReg, 1); // set ICC.ZF=0 to indicate failure 2232 jmpb (DONE_LABEL); 2233 2234 bind (LSuccess); 2235 testl (boxReg, 0); // set ICC.ZF=1 to indicate success 2236 jmpb (DONE_LABEL); 2237 } 2238 2239 bind (Stacked); 2240 movptr(tmpReg, Address (boxReg, 0)); // re-fetch 2241 if (os::is_MP()) { lock(); } 2242 cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses RAX which is box 2243 2244 if (EmitSync & 65536) { 2245 bind (CheckSucc); 2246 } 2247#endif 2248 bind(DONE_LABEL); 2249 } 2250} 2251#endif // COMPILER2 2252 2253void MacroAssembler::c2bool(Register x) { 2254 // implements x == 0 ? 0 : 1 2255 // note: must only look at least-significant byte of x 2256 // since C-style booleans are stored in one byte 2257 // only! (was bug) 2258 andl(x, 0xFF); 2259 setb(Assembler::notZero, x); 2260} 2261 2262// Wouldn't need if AddressLiteral version had new name 2263void MacroAssembler::call(Label& L, relocInfo::relocType rtype) { 2264 Assembler::call(L, rtype); 2265} 2266 2267void MacroAssembler::call(Register entry) { 2268 Assembler::call(entry); 2269} 2270 2271void MacroAssembler::call(AddressLiteral entry) { 2272 if (reachable(entry)) { 2273 Assembler::call_literal(entry.target(), entry.rspec()); 2274 } else { 2275 lea(rscratch1, entry); 2276 Assembler::call(rscratch1); 2277 } 2278} 2279 2280void MacroAssembler::ic_call(address entry) { 2281 RelocationHolder rh = virtual_call_Relocation::spec(pc()); 2282 movptr(rax, (intptr_t)Universe::non_oop_word()); 2283 call(AddressLiteral(entry, rh)); 2284} 2285 2286// Implementation of call_VM versions 2287 2288void MacroAssembler::call_VM(Register oop_result, 2289 address entry_point, 2290 bool check_exceptions) { 2291 Label C, E; 2292 call(C, relocInfo::none); 2293 jmp(E); 2294 2295 bind(C); 2296 call_VM_helper(oop_result, entry_point, 0, check_exceptions); 2297 ret(0); 2298 2299 bind(E); 2300} 2301 2302void MacroAssembler::call_VM(Register oop_result, 2303 address entry_point, 2304 Register arg_1, 2305 bool check_exceptions) { 2306 Label C, E; 2307 call(C, relocInfo::none); 2308 jmp(E); 2309 2310 bind(C); 2311 pass_arg1(this, arg_1); 2312 call_VM_helper(oop_result, entry_point, 1, check_exceptions); 2313 ret(0); 2314 2315 bind(E); 2316} 2317 2318void MacroAssembler::call_VM(Register oop_result, 2319 address entry_point, 2320 Register arg_1, 2321 Register arg_2, 2322 bool check_exceptions) { 2323 Label C, E; 2324 call(C, relocInfo::none); 2325 jmp(E); 2326 2327 bind(C); 2328 2329 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2330 2331 pass_arg2(this, arg_2); 2332 pass_arg1(this, arg_1); 2333 call_VM_helper(oop_result, entry_point, 2, check_exceptions); 2334 ret(0); 2335 2336 bind(E); 2337} 2338 2339void MacroAssembler::call_VM(Register oop_result, 2340 address entry_point, 2341 Register arg_1, 2342 Register arg_2, 2343 Register arg_3, 2344 bool check_exceptions) { 2345 Label C, E; 2346 call(C, relocInfo::none); 2347 jmp(E); 2348 2349 bind(C); 2350 2351 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2352 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2353 pass_arg3(this, arg_3); 2354 2355 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2356 pass_arg2(this, arg_2); 2357 2358 pass_arg1(this, arg_1); 2359 call_VM_helper(oop_result, entry_point, 3, check_exceptions); 2360 ret(0); 2361 2362 bind(E); 2363} 2364 2365void MacroAssembler::call_VM(Register oop_result, 2366 Register last_java_sp, 2367 address entry_point, 2368 int number_of_arguments, 2369 bool check_exceptions) { 2370 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg); 2371 call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions); 2372} 2373 2374void MacroAssembler::call_VM(Register oop_result, 2375 Register last_java_sp, 2376 address entry_point, 2377 Register arg_1, 2378 bool check_exceptions) { 2379 pass_arg1(this, arg_1); 2380 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 2381} 2382 2383void MacroAssembler::call_VM(Register oop_result, 2384 Register last_java_sp, 2385 address entry_point, 2386 Register arg_1, 2387 Register arg_2, 2388 bool check_exceptions) { 2389 2390 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2391 pass_arg2(this, arg_2); 2392 pass_arg1(this, arg_1); 2393 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 2394} 2395 2396void MacroAssembler::call_VM(Register oop_result, 2397 Register last_java_sp, 2398 address entry_point, 2399 Register arg_1, 2400 Register arg_2, 2401 Register arg_3, 2402 bool check_exceptions) { 2403 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2404 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2405 pass_arg3(this, arg_3); 2406 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2407 pass_arg2(this, arg_2); 2408 pass_arg1(this, arg_1); 2409 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 2410} 2411 2412void MacroAssembler::super_call_VM(Register oop_result, 2413 Register last_java_sp, 2414 address entry_point, 2415 int number_of_arguments, 2416 bool check_exceptions) { 2417 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg); 2418 MacroAssembler::call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions); 2419} 2420 2421void MacroAssembler::super_call_VM(Register oop_result, 2422 Register last_java_sp, 2423 address entry_point, 2424 Register arg_1, 2425 bool check_exceptions) { 2426 pass_arg1(this, arg_1); 2427 super_call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 2428} 2429 2430void MacroAssembler::super_call_VM(Register oop_result, 2431 Register last_java_sp, 2432 address entry_point, 2433 Register arg_1, 2434 Register arg_2, 2435 bool check_exceptions) { 2436 2437 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2438 pass_arg2(this, arg_2); 2439 pass_arg1(this, arg_1); 2440 super_call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 2441} 2442 2443void MacroAssembler::super_call_VM(Register oop_result, 2444 Register last_java_sp, 2445 address entry_point, 2446 Register arg_1, 2447 Register arg_2, 2448 Register arg_3, 2449 bool check_exceptions) { 2450 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2451 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2452 pass_arg3(this, arg_3); 2453 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2454 pass_arg2(this, arg_2); 2455 pass_arg1(this, arg_1); 2456 super_call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 2457} 2458 2459void MacroAssembler::call_VM_base(Register oop_result, 2460 Register java_thread, 2461 Register last_java_sp, 2462 address entry_point, 2463 int number_of_arguments, 2464 bool check_exceptions) { 2465 // determine java_thread register 2466 if (!java_thread->is_valid()) { 2467#ifdef _LP64 2468 java_thread = r15_thread; 2469#else 2470 java_thread = rdi; 2471 get_thread(java_thread); 2472#endif // LP64 2473 } 2474 // determine last_java_sp register 2475 if (!last_java_sp->is_valid()) { 2476 last_java_sp = rsp; 2477 } 2478 // debugging support 2479 assert(number_of_arguments >= 0 , "cannot have negative number of arguments"); 2480 LP64_ONLY(assert(java_thread == r15_thread, "unexpected register")); 2481#ifdef ASSERT 2482 // TraceBytecodes does not use r12 but saves it over the call, so don't verify 2483 // r12 is the heapbase. 2484 LP64_ONLY(if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");) 2485#endif // ASSERT 2486 2487 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result"); 2488 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp"); 2489 2490 // push java thread (becomes first argument of C function) 2491 2492 NOT_LP64(push(java_thread); number_of_arguments++); 2493 LP64_ONLY(mov(c_rarg0, r15_thread)); 2494 2495 // set last Java frame before call 2496 assert(last_java_sp != rbp, "can't use ebp/rbp"); 2497 2498 // Only interpreter should have to set fp 2499 set_last_Java_frame(java_thread, last_java_sp, rbp, NULL); 2500 2501 // do the call, remove parameters 2502 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments); 2503 2504 // restore the thread (cannot use the pushed argument since arguments 2505 // may be overwritten by C code generated by an optimizing compiler); 2506 // however can use the register value directly if it is callee saved. 2507 if (LP64_ONLY(true ||) java_thread == rdi || java_thread == rsi) { 2508 // rdi & rsi (also r15) are callee saved -> nothing to do 2509#ifdef ASSERT 2510 guarantee(java_thread != rax, "change this code"); 2511 push(rax); 2512 { Label L; 2513 get_thread(rax); 2514 cmpptr(java_thread, rax); 2515 jcc(Assembler::equal, L); 2516 STOP("MacroAssembler::call_VM_base: rdi not callee saved?"); 2517 bind(L); 2518 } 2519 pop(rax); 2520#endif 2521 } else { 2522 get_thread(java_thread); 2523 } 2524 // reset last Java frame 2525 // Only interpreter should have to clear fp 2526 reset_last_Java_frame(java_thread, true, false); 2527 2528#ifndef CC_INTERP 2529 // C++ interp handles this in the interpreter 2530 check_and_handle_popframe(java_thread); 2531 check_and_handle_earlyret(java_thread); 2532#endif /* CC_INTERP */ 2533 2534 if (check_exceptions) { 2535 // check for pending exceptions (java_thread is set upon return) 2536 cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t) NULL_WORD); 2537#ifndef _LP64 2538 jump_cc(Assembler::notEqual, 2539 RuntimeAddress(StubRoutines::forward_exception_entry())); 2540#else 2541 // This used to conditionally jump to forward_exception however it is 2542 // possible if we relocate that the branch will not reach. So we must jump 2543 // around so we can always reach 2544 2545 Label ok; 2546 jcc(Assembler::equal, ok); 2547 jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2548 bind(ok); 2549#endif // LP64 2550 } 2551 2552 // get oop result if there is one and reset the value in the thread 2553 if (oop_result->is_valid()) { 2554 get_vm_result(oop_result, java_thread); 2555 } 2556} 2557 2558void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) { 2559 2560 // Calculate the value for last_Java_sp 2561 // somewhat subtle. call_VM does an intermediate call 2562 // which places a return address on the stack just under the 2563 // stack pointer as the user finsihed with it. This allows 2564 // use to retrieve last_Java_pc from last_Java_sp[-1]. 2565 // On 32bit we then have to push additional args on the stack to accomplish 2566 // the actual requested call. On 64bit call_VM only can use register args 2567 // so the only extra space is the return address that call_VM created. 2568 // This hopefully explains the calculations here. 2569 2570#ifdef _LP64 2571 // We've pushed one address, correct last_Java_sp 2572 lea(rax, Address(rsp, wordSize)); 2573#else 2574 lea(rax, Address(rsp, (1 + number_of_arguments) * wordSize)); 2575#endif // LP64 2576 2577 call_VM_base(oop_result, noreg, rax, entry_point, number_of_arguments, check_exceptions); 2578 2579} 2580 2581void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) { 2582 call_VM_leaf_base(entry_point, number_of_arguments); 2583} 2584 2585void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) { 2586 pass_arg0(this, arg_0); 2587 call_VM_leaf(entry_point, 1); 2588} 2589 2590void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { 2591 2592 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2593 pass_arg1(this, arg_1); 2594 pass_arg0(this, arg_0); 2595 call_VM_leaf(entry_point, 2); 2596} 2597 2598void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { 2599 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2600 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2601 pass_arg2(this, arg_2); 2602 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2603 pass_arg1(this, arg_1); 2604 pass_arg0(this, arg_0); 2605 call_VM_leaf(entry_point, 3); 2606} 2607 2608void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) { 2609 pass_arg0(this, arg_0); 2610 MacroAssembler::call_VM_leaf_base(entry_point, 1); 2611} 2612 2613void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { 2614 2615 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2616 pass_arg1(this, arg_1); 2617 pass_arg0(this, arg_0); 2618 MacroAssembler::call_VM_leaf_base(entry_point, 2); 2619} 2620 2621void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { 2622 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2623 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2624 pass_arg2(this, arg_2); 2625 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2626 pass_arg1(this, arg_1); 2627 pass_arg0(this, arg_0); 2628 MacroAssembler::call_VM_leaf_base(entry_point, 3); 2629} 2630 2631void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) { 2632 LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg")); 2633 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2634 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2635 pass_arg3(this, arg_3); 2636 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2637 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2638 pass_arg2(this, arg_2); 2639 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2640 pass_arg1(this, arg_1); 2641 pass_arg0(this, arg_0); 2642 MacroAssembler::call_VM_leaf_base(entry_point, 4); 2643} 2644 2645void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) { 2646 movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset())); 2647 movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD); 2648 verify_oop(oop_result, "broken oop in call_VM_base"); 2649} 2650 2651void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) { 2652 movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset())); 2653 movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD); 2654} 2655 2656void MacroAssembler::check_and_handle_earlyret(Register java_thread) { 2657} 2658 2659void MacroAssembler::check_and_handle_popframe(Register java_thread) { 2660} 2661 2662void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) { 2663 if (reachable(src1)) { 2664 cmpl(as_Address(src1), imm); 2665 } else { 2666 lea(rscratch1, src1); 2667 cmpl(Address(rscratch1, 0), imm); 2668 } 2669} 2670 2671void MacroAssembler::cmp32(Register src1, AddressLiteral src2) { 2672 assert(!src2.is_lval(), "use cmpptr"); 2673 if (reachable(src2)) { 2674 cmpl(src1, as_Address(src2)); 2675 } else { 2676 lea(rscratch1, src2); 2677 cmpl(src1, Address(rscratch1, 0)); 2678 } 2679} 2680 2681void MacroAssembler::cmp32(Register src1, int32_t imm) { 2682 Assembler::cmpl(src1, imm); 2683} 2684 2685void MacroAssembler::cmp32(Register src1, Address src2) { 2686 Assembler::cmpl(src1, src2); 2687} 2688 2689void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) { 2690 ucomisd(opr1, opr2); 2691 2692 Label L; 2693 if (unordered_is_less) { 2694 movl(dst, -1); 2695 jcc(Assembler::parity, L); 2696 jcc(Assembler::below , L); 2697 movl(dst, 0); 2698 jcc(Assembler::equal , L); 2699 increment(dst); 2700 } else { // unordered is greater 2701 movl(dst, 1); 2702 jcc(Assembler::parity, L); 2703 jcc(Assembler::above , L); 2704 movl(dst, 0); 2705 jcc(Assembler::equal , L); 2706 decrementl(dst); 2707 } 2708 bind(L); 2709} 2710 2711void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) { 2712 ucomiss(opr1, opr2); 2713 2714 Label L; 2715 if (unordered_is_less) { 2716 movl(dst, -1); 2717 jcc(Assembler::parity, L); 2718 jcc(Assembler::below , L); 2719 movl(dst, 0); 2720 jcc(Assembler::equal , L); 2721 increment(dst); 2722 } else { // unordered is greater 2723 movl(dst, 1); 2724 jcc(Assembler::parity, L); 2725 jcc(Assembler::above , L); 2726 movl(dst, 0); 2727 jcc(Assembler::equal , L); 2728 decrementl(dst); 2729 } 2730 bind(L); 2731} 2732 2733 2734void MacroAssembler::cmp8(AddressLiteral src1, int imm) { 2735 if (reachable(src1)) { 2736 cmpb(as_Address(src1), imm); 2737 } else { 2738 lea(rscratch1, src1); 2739 cmpb(Address(rscratch1, 0), imm); 2740 } 2741} 2742 2743void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) { 2744#ifdef _LP64 2745 if (src2.is_lval()) { 2746 movptr(rscratch1, src2); 2747 Assembler::cmpq(src1, rscratch1); 2748 } else if (reachable(src2)) { 2749 cmpq(src1, as_Address(src2)); 2750 } else { 2751 lea(rscratch1, src2); 2752 Assembler::cmpq(src1, Address(rscratch1, 0)); 2753 } 2754#else 2755 if (src2.is_lval()) { 2756 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec()); 2757 } else { 2758 cmpl(src1, as_Address(src2)); 2759 } 2760#endif // _LP64 2761} 2762 2763void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) { 2764 assert(src2.is_lval(), "not a mem-mem compare"); 2765#ifdef _LP64 2766 // moves src2's literal address 2767 movptr(rscratch1, src2); 2768 Assembler::cmpq(src1, rscratch1); 2769#else 2770 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec()); 2771#endif // _LP64 2772} 2773 2774void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) { 2775 if (reachable(adr)) { 2776 if (os::is_MP()) 2777 lock(); 2778 cmpxchgptr(reg, as_Address(adr)); 2779 } else { 2780 lea(rscratch1, adr); 2781 if (os::is_MP()) 2782 lock(); 2783 cmpxchgptr(reg, Address(rscratch1, 0)); 2784 } 2785} 2786 2787void MacroAssembler::cmpxchgptr(Register reg, Address adr) { 2788 LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr)); 2789} 2790 2791void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) { 2792 if (reachable(src)) { 2793 Assembler::comisd(dst, as_Address(src)); 2794 } else { 2795 lea(rscratch1, src); 2796 Assembler::comisd(dst, Address(rscratch1, 0)); 2797 } 2798} 2799 2800void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) { 2801 if (reachable(src)) { 2802 Assembler::comiss(dst, as_Address(src)); 2803 } else { 2804 lea(rscratch1, src); 2805 Assembler::comiss(dst, Address(rscratch1, 0)); 2806 } 2807} 2808 2809 2810void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) { 2811 Condition negated_cond = negate_condition(cond); 2812 Label L; 2813 jcc(negated_cond, L); 2814 pushf(); // Preserve flags 2815 atomic_incl(counter_addr); 2816 popf(); 2817 bind(L); 2818} 2819 2820int MacroAssembler::corrected_idivl(Register reg) { 2821 // Full implementation of Java idiv and irem; checks for 2822 // special case as described in JVM spec., p.243 & p.271. 2823 // The function returns the (pc) offset of the idivl 2824 // instruction - may be needed for implicit exceptions. 2825 // 2826 // normal case special case 2827 // 2828 // input : rax,: dividend min_int 2829 // reg: divisor (may not be rax,/rdx) -1 2830 // 2831 // output: rax,: quotient (= rax, idiv reg) min_int 2832 // rdx: remainder (= rax, irem reg) 0 2833 assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register"); 2834 const int min_int = 0x80000000; 2835 Label normal_case, special_case; 2836 2837 // check for special case 2838 cmpl(rax, min_int); 2839 jcc(Assembler::notEqual, normal_case); 2840 xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0) 2841 cmpl(reg, -1); 2842 jcc(Assembler::equal, special_case); 2843 2844 // handle normal case 2845 bind(normal_case); 2846 cdql(); 2847 int idivl_offset = offset(); 2848 idivl(reg); 2849 2850 // normal and special case exit 2851 bind(special_case); 2852 2853 return idivl_offset; 2854} 2855 2856 2857 2858void MacroAssembler::decrementl(Register reg, int value) { 2859 if (value == min_jint) {subl(reg, value) ; return; } 2860 if (value < 0) { incrementl(reg, -value); return; } 2861 if (value == 0) { ; return; } 2862 if (value == 1 && UseIncDec) { decl(reg) ; return; } 2863 /* else */ { subl(reg, value) ; return; } 2864} 2865 2866void MacroAssembler::decrementl(Address dst, int value) { 2867 if (value == min_jint) {subl(dst, value) ; return; } 2868 if (value < 0) { incrementl(dst, -value); return; } 2869 if (value == 0) { ; return; } 2870 if (value == 1 && UseIncDec) { decl(dst) ; return; } 2871 /* else */ { subl(dst, value) ; return; } 2872} 2873 2874void MacroAssembler::division_with_shift (Register reg, int shift_value) { 2875 assert (shift_value > 0, "illegal shift value"); 2876 Label _is_positive; 2877 testl (reg, reg); 2878 jcc (Assembler::positive, _is_positive); 2879 int offset = (1 << shift_value) - 1 ; 2880 2881 if (offset == 1) { 2882 incrementl(reg); 2883 } else { 2884 addl(reg, offset); 2885 } 2886 2887 bind (_is_positive); 2888 sarl(reg, shift_value); 2889} 2890 2891void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) { 2892 if (reachable(src)) { 2893 Assembler::divsd(dst, as_Address(src)); 2894 } else { 2895 lea(rscratch1, src); 2896 Assembler::divsd(dst, Address(rscratch1, 0)); 2897 } 2898} 2899 2900void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) { 2901 if (reachable(src)) { 2902 Assembler::divss(dst, as_Address(src)); 2903 } else { 2904 lea(rscratch1, src); 2905 Assembler::divss(dst, Address(rscratch1, 0)); 2906 } 2907} 2908 2909// !defined(COMPILER2) is because of stupid core builds 2910#if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2) || INCLUDE_JVMCI 2911void MacroAssembler::empty_FPU_stack() { 2912 if (VM_Version::supports_mmx()) { 2913 emms(); 2914 } else { 2915 for (int i = 8; i-- > 0; ) ffree(i); 2916 } 2917} 2918#endif // !LP64 || C1 || !C2 || INCLUDE_JVMCI 2919 2920 2921// Defines obj, preserves var_size_in_bytes 2922void MacroAssembler::eden_allocate(Register obj, 2923 Register var_size_in_bytes, 2924 int con_size_in_bytes, 2925 Register t1, 2926 Label& slow_case) { 2927 assert(obj == rax, "obj must be in rax, for cmpxchg"); 2928 assert_different_registers(obj, var_size_in_bytes, t1); 2929 if (!Universe::heap()->supports_inline_contig_alloc()) { 2930 jmp(slow_case); 2931 } else { 2932 Register end = t1; 2933 Label retry; 2934 bind(retry); 2935 ExternalAddress heap_top((address) Universe::heap()->top_addr()); 2936 movptr(obj, heap_top); 2937 if (var_size_in_bytes == noreg) { 2938 lea(end, Address(obj, con_size_in_bytes)); 2939 } else { 2940 lea(end, Address(obj, var_size_in_bytes, Address::times_1)); 2941 } 2942 // if end < obj then we wrapped around => object too long => slow case 2943 cmpptr(end, obj); 2944 jcc(Assembler::below, slow_case); 2945 cmpptr(end, ExternalAddress((address) Universe::heap()->end_addr())); 2946 jcc(Assembler::above, slow_case); 2947 // Compare obj with the top addr, and if still equal, store the new top addr in 2948 // end at the address of the top addr pointer. Sets ZF if was equal, and clears 2949 // it otherwise. Use lock prefix for atomicity on MPs. 2950 locked_cmpxchgptr(end, heap_top); 2951 jcc(Assembler::notEqual, retry); 2952 } 2953} 2954 2955void MacroAssembler::enter() { 2956 push(rbp); 2957 mov(rbp, rsp); 2958} 2959 2960// A 5 byte nop that is safe for patching (see patch_verified_entry) 2961void MacroAssembler::fat_nop() { 2962 if (UseAddressNop) { 2963 addr_nop_5(); 2964 } else { 2965 emit_int8(0x26); // es: 2966 emit_int8(0x2e); // cs: 2967 emit_int8(0x64); // fs: 2968 emit_int8(0x65); // gs: 2969 emit_int8((unsigned char)0x90); 2970 } 2971} 2972 2973void MacroAssembler::fcmp(Register tmp) { 2974 fcmp(tmp, 1, true, true); 2975} 2976 2977void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) { 2978 assert(!pop_right || pop_left, "usage error"); 2979 if (VM_Version::supports_cmov()) { 2980 assert(tmp == noreg, "unneeded temp"); 2981 if (pop_left) { 2982 fucomip(index); 2983 } else { 2984 fucomi(index); 2985 } 2986 if (pop_right) { 2987 fpop(); 2988 } 2989 } else { 2990 assert(tmp != noreg, "need temp"); 2991 if (pop_left) { 2992 if (pop_right) { 2993 fcompp(); 2994 } else { 2995 fcomp(index); 2996 } 2997 } else { 2998 fcom(index); 2999 } 3000 // convert FPU condition into eflags condition via rax, 3001 save_rax(tmp); 3002 fwait(); fnstsw_ax(); 3003 sahf(); 3004 restore_rax(tmp); 3005 } 3006 // condition codes set as follows: 3007 // 3008 // CF (corresponds to C0) if x < y 3009 // PF (corresponds to C2) if unordered 3010 // ZF (corresponds to C3) if x = y 3011} 3012 3013void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) { 3014 fcmp2int(dst, unordered_is_less, 1, true, true); 3015} 3016 3017void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) { 3018 fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right); 3019 Label L; 3020 if (unordered_is_less) { 3021 movl(dst, -1); 3022 jcc(Assembler::parity, L); 3023 jcc(Assembler::below , L); 3024 movl(dst, 0); 3025 jcc(Assembler::equal , L); 3026 increment(dst); 3027 } else { // unordered is greater 3028 movl(dst, 1); 3029 jcc(Assembler::parity, L); 3030 jcc(Assembler::above , L); 3031 movl(dst, 0); 3032 jcc(Assembler::equal , L); 3033 decrementl(dst); 3034 } 3035 bind(L); 3036} 3037 3038void MacroAssembler::fld_d(AddressLiteral src) { 3039 fld_d(as_Address(src)); 3040} 3041 3042void MacroAssembler::fld_s(AddressLiteral src) { 3043 fld_s(as_Address(src)); 3044} 3045 3046void MacroAssembler::fld_x(AddressLiteral src) { 3047 Assembler::fld_x(as_Address(src)); 3048} 3049 3050void MacroAssembler::fldcw(AddressLiteral src) { 3051 Assembler::fldcw(as_Address(src)); 3052} 3053 3054void MacroAssembler::mulpd(XMMRegister dst, AddressLiteral src) { 3055 if (reachable(src)) { 3056 Assembler::mulpd(dst, as_Address(src)); 3057 } else { 3058 lea(rscratch1, src); 3059 Assembler::mulpd(dst, Address(rscratch1, 0)); 3060 } 3061} 3062 3063void MacroAssembler::pow_exp_core_encoding() { 3064 // kills rax, rcx, rdx 3065 subptr(rsp,sizeof(jdouble)); 3066 // computes 2^X. Stack: X ... 3067 // f2xm1 computes 2^X-1 but only operates on -1<=X<=1. Get int(X) and 3068 // keep it on the thread's stack to compute 2^int(X) later 3069 // then compute 2^(X-int(X)) as (2^(X-int(X)-1+1) 3070 // final result is obtained with: 2^X = 2^int(X) * 2^(X-int(X)) 3071 fld_s(0); // Stack: X X ... 3072 frndint(); // Stack: int(X) X ... 3073 fsuba(1); // Stack: int(X) X-int(X) ... 3074 fistp_s(Address(rsp,0)); // move int(X) as integer to thread's stack. Stack: X-int(X) ... 3075 f2xm1(); // Stack: 2^(X-int(X))-1 ... 3076 fld1(); // Stack: 1 2^(X-int(X))-1 ... 3077 faddp(1); // Stack: 2^(X-int(X)) 3078 // computes 2^(int(X)): add exponent bias (1023) to int(X), then 3079 // shift int(X)+1023 to exponent position. 3080 // Exponent is limited to 11 bits if int(X)+1023 does not fit in 11 3081 // bits, set result to NaN. 0x000 and 0x7FF are reserved exponent 3082 // values so detect them and set result to NaN. 3083 movl(rax,Address(rsp,0)); 3084 movl(rcx, -2048); // 11 bit mask and valid NaN binary encoding 3085 addl(rax, 1023); 3086 movl(rdx,rax); 3087 shll(rax,20); 3088 // Check that 0 < int(X)+1023 < 2047. Otherwise set rax to NaN. 3089 addl(rdx,1); 3090 // Check that 1 < int(X)+1023+1 < 2048 3091 // in 3 steps: 3092 // 1- (int(X)+1023+1)&-2048 == 0 => 0 <= int(X)+1023+1 < 2048 3093 // 2- (int(X)+1023+1)&-2048 != 0 3094 // 3- (int(X)+1023+1)&-2048 != 1 3095 // Do 2- first because addl just updated the flags. 3096 cmov32(Assembler::equal,rax,rcx); 3097 cmpl(rdx,1); 3098 cmov32(Assembler::equal,rax,rcx); 3099 testl(rdx,rcx); 3100 cmov32(Assembler::notEqual,rax,rcx); 3101 movl(Address(rsp,4),rax); 3102 movl(Address(rsp,0),0); 3103 fmul_d(Address(rsp,0)); // Stack: 2^X ... 3104 addptr(rsp,sizeof(jdouble)); 3105} 3106 3107void MacroAssembler::increase_precision() { 3108 subptr(rsp, BytesPerWord); 3109 fnstcw(Address(rsp, 0)); 3110 movl(rax, Address(rsp, 0)); 3111 orl(rax, 0x300); 3112 push(rax); 3113 fldcw(Address(rsp, 0)); 3114 pop(rax); 3115} 3116 3117void MacroAssembler::restore_precision() { 3118 fldcw(Address(rsp, 0)); 3119 addptr(rsp, BytesPerWord); 3120} 3121 3122void MacroAssembler::fast_pow() { 3123 // computes X^Y = 2^(Y * log2(X)) 3124 // if fast computation is not possible, result is NaN. Requires 3125 // fallback from user of this macro. 3126 // increase precision for intermediate steps of the computation 3127 BLOCK_COMMENT("fast_pow {"); 3128 increase_precision(); 3129 fyl2x(); // Stack: (Y*log2(X)) ... 3130 pow_exp_core_encoding(); // Stack: exp(X) ... 3131 restore_precision(); 3132 BLOCK_COMMENT("} fast_pow"); 3133} 3134 3135void MacroAssembler::pow_or_exp(int num_fpu_regs_in_use) { 3136 // kills rax, rcx, rdx 3137 // pow and exp needs 2 extra registers on the fpu stack. 3138 Label slow_case, done; 3139 Register tmp = noreg; 3140 if (!VM_Version::supports_cmov()) { 3141 // fcmp needs a temporary so preserve rdx, 3142 tmp = rdx; 3143 } 3144 Register tmp2 = rax; 3145 Register tmp3 = rcx; 3146 3147 // Stack: X Y 3148 Label x_negative, y_not_2; 3149 3150 static double two = 2.0; 3151 ExternalAddress two_addr((address)&two); 3152 3153 // constant maybe too far on 64 bit 3154 lea(tmp2, two_addr); 3155 fld_d(Address(tmp2, 0)); // Stack: 2 X Y 3156 fcmp(tmp, 2, true, false); // Stack: X Y 3157 jcc(Assembler::parity, y_not_2); 3158 jcc(Assembler::notEqual, y_not_2); 3159 3160 fxch(); fpop(); // Stack: X 3161 fmul(0); // Stack: X*X 3162 3163 jmp(done); 3164 3165 bind(y_not_2); 3166 3167 fldz(); // Stack: 0 X Y 3168 fcmp(tmp, 1, true, false); // Stack: X Y 3169 jcc(Assembler::above, x_negative); 3170 3171 // X >= 0 3172 3173 fld_s(1); // duplicate arguments for runtime call. Stack: Y X Y 3174 fld_s(1); // Stack: X Y X Y 3175 fast_pow(); // Stack: X^Y X Y 3176 fcmp(tmp, 0, false, false); // Stack: X^Y X Y 3177 // X^Y not equal to itself: X^Y is NaN go to slow case. 3178 jcc(Assembler::parity, slow_case); 3179 // get rid of duplicate arguments. Stack: X^Y 3180 if (num_fpu_regs_in_use > 0) { 3181 fxch(); fpop(); 3182 fxch(); fpop(); 3183 } else { 3184 ffree(2); 3185 ffree(1); 3186 } 3187 jmp(done); 3188 3189 // X <= 0 3190 bind(x_negative); 3191 3192 fld_s(1); // Stack: Y X Y 3193 frndint(); // Stack: int(Y) X Y 3194 fcmp(tmp, 2, false, false); // Stack: int(Y) X Y 3195 jcc(Assembler::notEqual, slow_case); 3196 3197 subptr(rsp, 8); 3198 3199 // For X^Y, when X < 0, Y has to be an integer and the final 3200 // result depends on whether it's odd or even. We just checked 3201 // that int(Y) == Y. We move int(Y) to gp registers as a 64 bit 3202 // integer to test its parity. If int(Y) is huge and doesn't fit 3203 // in the 64 bit integer range, the integer indefinite value will 3204 // end up in the gp registers. Huge numbers are all even, the 3205 // integer indefinite number is even so it's fine. 3206 3207#ifdef ASSERT 3208 // Let's check we don't end up with an integer indefinite number 3209 // when not expected. First test for huge numbers: check whether 3210 // int(Y)+1 == int(Y) which is true for very large numbers and 3211 // those are all even. A 64 bit integer is guaranteed to not 3212 // overflow for numbers where y+1 != y (when precision is set to 3213 // double precision). 3214 Label y_not_huge; 3215 3216 fld1(); // Stack: 1 int(Y) X Y 3217 fadd(1); // Stack: 1+int(Y) int(Y) X Y 3218 3219#ifdef _LP64 3220 // trip to memory to force the precision down from double extended 3221 // precision 3222 fstp_d(Address(rsp, 0)); 3223 fld_d(Address(rsp, 0)); 3224#endif 3225 3226 fcmp(tmp, 1, true, false); // Stack: int(Y) X Y 3227#endif 3228 3229 // move int(Y) as 64 bit integer to thread's stack 3230 fistp_d(Address(rsp,0)); // Stack: X Y 3231 3232#ifdef ASSERT 3233 jcc(Assembler::notEqual, y_not_huge); 3234 3235 // Y is huge so we know it's even. It may not fit in a 64 bit 3236 // integer and we don't want the debug code below to see the 3237 // integer indefinite value so overwrite int(Y) on the thread's 3238 // stack with 0. 3239 movl(Address(rsp, 0), 0); 3240 movl(Address(rsp, 4), 0); 3241 3242 bind(y_not_huge); 3243#endif 3244 3245 fld_s(1); // duplicate arguments for runtime call. Stack: Y X Y 3246 fld_s(1); // Stack: X Y X Y 3247 fabs(); // Stack: abs(X) Y X Y 3248 fast_pow(); // Stack: abs(X)^Y X Y 3249 fcmp(tmp, 0, false, false); // Stack: abs(X)^Y X Y 3250 // abs(X)^Y not equal to itself: abs(X)^Y is NaN go to slow case. 3251 3252 pop(tmp2); 3253 NOT_LP64(pop(tmp3)); 3254 jcc(Assembler::parity, slow_case); 3255 3256#ifdef ASSERT 3257 // Check that int(Y) is not integer indefinite value (int 3258 // overflow). Shouldn't happen because for values that would 3259 // overflow, 1+int(Y)==Y which was tested earlier. 3260#ifndef _LP64 3261 { 3262 Label integer; 3263 testl(tmp2, tmp2); 3264 jcc(Assembler::notZero, integer); 3265 cmpl(tmp3, 0x80000000); 3266 jcc(Assembler::notZero, integer); 3267 STOP("integer indefinite value shouldn't be seen here"); 3268 bind(integer); 3269 } 3270#else 3271 { 3272 Label integer; 3273 mov(tmp3, tmp2); // preserve tmp2 for parity check below 3274 shlq(tmp3, 1); 3275 jcc(Assembler::carryClear, integer); 3276 jcc(Assembler::notZero, integer); 3277 STOP("integer indefinite value shouldn't be seen here"); 3278 bind(integer); 3279 } 3280#endif 3281#endif 3282 3283 // get rid of duplicate arguments. Stack: X^Y 3284 if (num_fpu_regs_in_use > 0) { 3285 fxch(); fpop(); 3286 fxch(); fpop(); 3287 } else { 3288 ffree(2); 3289 ffree(1); 3290 } 3291 3292 testl(tmp2, 1); 3293 jcc(Assembler::zero, done); // X <= 0, Y even: X^Y = abs(X)^Y 3294 // X <= 0, Y even: X^Y = -abs(X)^Y 3295 3296 fchs(); // Stack: -abs(X)^Y Y 3297 jmp(done); 3298 3299 // slow case: runtime call 3300 bind(slow_case); 3301 3302 fpop(); // pop incorrect result or int(Y) 3303 3304 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dpow), 2, num_fpu_regs_in_use); 3305 3306 // Come here with result in F-TOS 3307 bind(done); 3308} 3309 3310void MacroAssembler::fpop() { 3311 ffree(); 3312 fincstp(); 3313} 3314 3315void MacroAssembler::load_float(Address src) { 3316 if (UseSSE >= 1) { 3317 movflt(xmm0, src); 3318 } else { 3319 LP64_ONLY(ShouldNotReachHere()); 3320 NOT_LP64(fld_s(src)); 3321 } 3322} 3323 3324void MacroAssembler::store_float(Address dst) { 3325 if (UseSSE >= 1) { 3326 movflt(dst, xmm0); 3327 } else { 3328 LP64_ONLY(ShouldNotReachHere()); 3329 NOT_LP64(fstp_s(dst)); 3330 } 3331} 3332 3333void MacroAssembler::load_double(Address src) { 3334 if (UseSSE >= 2) { 3335 movdbl(xmm0, src); 3336 } else { 3337 LP64_ONLY(ShouldNotReachHere()); 3338 NOT_LP64(fld_d(src)); 3339 } 3340} 3341 3342void MacroAssembler::store_double(Address dst) { 3343 if (UseSSE >= 2) { 3344 movdbl(dst, xmm0); 3345 } else { 3346 LP64_ONLY(ShouldNotReachHere()); 3347 NOT_LP64(fstp_d(dst)); 3348 } 3349} 3350 3351void MacroAssembler::fremr(Register tmp) { 3352 save_rax(tmp); 3353 { Label L; 3354 bind(L); 3355 fprem(); 3356 fwait(); fnstsw_ax(); 3357#ifdef _LP64 3358 testl(rax, 0x400); 3359 jcc(Assembler::notEqual, L); 3360#else 3361 sahf(); 3362 jcc(Assembler::parity, L); 3363#endif // _LP64 3364 } 3365 restore_rax(tmp); 3366 // Result is in ST0. 3367 // Note: fxch & fpop to get rid of ST1 3368 // (otherwise FPU stack could overflow eventually) 3369 fxch(1); 3370 fpop(); 3371} 3372 3373 3374void MacroAssembler::incrementl(AddressLiteral dst) { 3375 if (reachable(dst)) { 3376 incrementl(as_Address(dst)); 3377 } else { 3378 lea(rscratch1, dst); 3379 incrementl(Address(rscratch1, 0)); 3380 } 3381} 3382 3383void MacroAssembler::incrementl(ArrayAddress dst) { 3384 incrementl(as_Address(dst)); 3385} 3386 3387void MacroAssembler::incrementl(Register reg, int value) { 3388 if (value == min_jint) {addl(reg, value) ; return; } 3389 if (value < 0) { decrementl(reg, -value); return; } 3390 if (value == 0) { ; return; } 3391 if (value == 1 && UseIncDec) { incl(reg) ; return; } 3392 /* else */ { addl(reg, value) ; return; } 3393} 3394 3395void MacroAssembler::incrementl(Address dst, int value) { 3396 if (value == min_jint) {addl(dst, value) ; return; } 3397 if (value < 0) { decrementl(dst, -value); return; } 3398 if (value == 0) { ; return; } 3399 if (value == 1 && UseIncDec) { incl(dst) ; return; } 3400 /* else */ { addl(dst, value) ; return; } 3401} 3402 3403void MacroAssembler::jump(AddressLiteral dst) { 3404 if (reachable(dst)) { 3405 jmp_literal(dst.target(), dst.rspec()); 3406 } else { 3407 lea(rscratch1, dst); 3408 jmp(rscratch1); 3409 } 3410} 3411 3412void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) { 3413 if (reachable(dst)) { 3414 InstructionMark im(this); 3415 relocate(dst.reloc()); 3416 const int short_size = 2; 3417 const int long_size = 6; 3418 int offs = (intptr_t)dst.target() - ((intptr_t)pc()); 3419 if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) { 3420 // 0111 tttn #8-bit disp 3421 emit_int8(0x70 | cc); 3422 emit_int8((offs - short_size) & 0xFF); 3423 } else { 3424 // 0000 1111 1000 tttn #32-bit disp 3425 emit_int8(0x0F); 3426 emit_int8((unsigned char)(0x80 | cc)); 3427 emit_int32(offs - long_size); 3428 } 3429 } else { 3430#ifdef ASSERT 3431 warning("reversing conditional branch"); 3432#endif /* ASSERT */ 3433 Label skip; 3434 jccb(reverse[cc], skip); 3435 lea(rscratch1, dst); 3436 Assembler::jmp(rscratch1); 3437 bind(skip); 3438 } 3439} 3440 3441void MacroAssembler::ldmxcsr(AddressLiteral src) { 3442 if (reachable(src)) { 3443 Assembler::ldmxcsr(as_Address(src)); 3444 } else { 3445 lea(rscratch1, src); 3446 Assembler::ldmxcsr(Address(rscratch1, 0)); 3447 } 3448} 3449 3450int MacroAssembler::load_signed_byte(Register dst, Address src) { 3451 int off; 3452 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3453 off = offset(); 3454 movsbl(dst, src); // movsxb 3455 } else { 3456 off = load_unsigned_byte(dst, src); 3457 shll(dst, 24); 3458 sarl(dst, 24); 3459 } 3460 return off; 3461} 3462 3463// Note: load_signed_short used to be called load_signed_word. 3464// Although the 'w' in x86 opcodes refers to the term "word" in the assembler 3465// manual, which means 16 bits, that usage is found nowhere in HotSpot code. 3466// The term "word" in HotSpot means a 32- or 64-bit machine word. 3467int MacroAssembler::load_signed_short(Register dst, Address src) { 3468 int off; 3469 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3470 // This is dubious to me since it seems safe to do a signed 16 => 64 bit 3471 // version but this is what 64bit has always done. This seems to imply 3472 // that users are only using 32bits worth. 3473 off = offset(); 3474 movswl(dst, src); // movsxw 3475 } else { 3476 off = load_unsigned_short(dst, src); 3477 shll(dst, 16); 3478 sarl(dst, 16); 3479 } 3480 return off; 3481} 3482 3483int MacroAssembler::load_unsigned_byte(Register dst, Address src) { 3484 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16, 3485 // and "3.9 Partial Register Penalties", p. 22). 3486 int off; 3487 if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) { 3488 off = offset(); 3489 movzbl(dst, src); // movzxb 3490 } else { 3491 xorl(dst, dst); 3492 off = offset(); 3493 movb(dst, src); 3494 } 3495 return off; 3496} 3497 3498// Note: load_unsigned_short used to be called load_unsigned_word. 3499int MacroAssembler::load_unsigned_short(Register dst, Address src) { 3500 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16, 3501 // and "3.9 Partial Register Penalties", p. 22). 3502 int off; 3503 if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) { 3504 off = offset(); 3505 movzwl(dst, src); // movzxw 3506 } else { 3507 xorl(dst, dst); 3508 off = offset(); 3509 movw(dst, src); 3510 } 3511 return off; 3512} 3513 3514void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) { 3515 switch (size_in_bytes) { 3516#ifndef _LP64 3517 case 8: 3518 assert(dst2 != noreg, "second dest register required"); 3519 movl(dst, src); 3520 movl(dst2, src.plus_disp(BytesPerInt)); 3521 break; 3522#else 3523 case 8: movq(dst, src); break; 3524#endif 3525 case 4: movl(dst, src); break; 3526 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break; 3527 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break; 3528 default: ShouldNotReachHere(); 3529 } 3530} 3531 3532void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) { 3533 switch (size_in_bytes) { 3534#ifndef _LP64 3535 case 8: 3536 assert(src2 != noreg, "second source register required"); 3537 movl(dst, src); 3538 movl(dst.plus_disp(BytesPerInt), src2); 3539 break; 3540#else 3541 case 8: movq(dst, src); break; 3542#endif 3543 case 4: movl(dst, src); break; 3544 case 2: movw(dst, src); break; 3545 case 1: movb(dst, src); break; 3546 default: ShouldNotReachHere(); 3547 } 3548} 3549 3550void MacroAssembler::mov32(AddressLiteral dst, Register src) { 3551 if (reachable(dst)) { 3552 movl(as_Address(dst), src); 3553 } else { 3554 lea(rscratch1, dst); 3555 movl(Address(rscratch1, 0), src); 3556 } 3557} 3558 3559void MacroAssembler::mov32(Register dst, AddressLiteral src) { 3560 if (reachable(src)) { 3561 movl(dst, as_Address(src)); 3562 } else { 3563 lea(rscratch1, src); 3564 movl(dst, Address(rscratch1, 0)); 3565 } 3566} 3567 3568// C++ bool manipulation 3569 3570void MacroAssembler::movbool(Register dst, Address src) { 3571 if(sizeof(bool) == 1) 3572 movb(dst, src); 3573 else if(sizeof(bool) == 2) 3574 movw(dst, src); 3575 else if(sizeof(bool) == 4) 3576 movl(dst, src); 3577 else 3578 // unsupported 3579 ShouldNotReachHere(); 3580} 3581 3582void MacroAssembler::movbool(Address dst, bool boolconst) { 3583 if(sizeof(bool) == 1) 3584 movb(dst, (int) boolconst); 3585 else if(sizeof(bool) == 2) 3586 movw(dst, (int) boolconst); 3587 else if(sizeof(bool) == 4) 3588 movl(dst, (int) boolconst); 3589 else 3590 // unsupported 3591 ShouldNotReachHere(); 3592} 3593 3594void MacroAssembler::movbool(Address dst, Register src) { 3595 if(sizeof(bool) == 1) 3596 movb(dst, src); 3597 else if(sizeof(bool) == 2) 3598 movw(dst, src); 3599 else if(sizeof(bool) == 4) 3600 movl(dst, src); 3601 else 3602 // unsupported 3603 ShouldNotReachHere(); 3604} 3605 3606void MacroAssembler::movbyte(ArrayAddress dst, int src) { 3607 movb(as_Address(dst), src); 3608} 3609 3610void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) { 3611 if (reachable(src)) { 3612 movdl(dst, as_Address(src)); 3613 } else { 3614 lea(rscratch1, src); 3615 movdl(dst, Address(rscratch1, 0)); 3616 } 3617} 3618 3619void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) { 3620 if (reachable(src)) { 3621 movq(dst, as_Address(src)); 3622 } else { 3623 lea(rscratch1, src); 3624 movq(dst, Address(rscratch1, 0)); 3625 } 3626} 3627 3628void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) { 3629 if (reachable(src)) { 3630 if (UseXmmLoadAndClearUpper) { 3631 movsd (dst, as_Address(src)); 3632 } else { 3633 movlpd(dst, as_Address(src)); 3634 } 3635 } else { 3636 lea(rscratch1, src); 3637 if (UseXmmLoadAndClearUpper) { 3638 movsd (dst, Address(rscratch1, 0)); 3639 } else { 3640 movlpd(dst, Address(rscratch1, 0)); 3641 } 3642 } 3643} 3644 3645void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) { 3646 if (reachable(src)) { 3647 movss(dst, as_Address(src)); 3648 } else { 3649 lea(rscratch1, src); 3650 movss(dst, Address(rscratch1, 0)); 3651 } 3652} 3653 3654void MacroAssembler::movptr(Register dst, Register src) { 3655 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3656} 3657 3658void MacroAssembler::movptr(Register dst, Address src) { 3659 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3660} 3661 3662// src should NEVER be a real pointer. Use AddressLiteral for true pointers 3663void MacroAssembler::movptr(Register dst, intptr_t src) { 3664 LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src)); 3665} 3666 3667void MacroAssembler::movptr(Address dst, Register src) { 3668 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3669} 3670 3671void MacroAssembler::movdqu(Address dst, XMMRegister src) { 3672 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (src->encoding() > 15)) { 3673 Assembler::vextractf32x4h(dst, src, 0); 3674 } else { 3675 Assembler::movdqu(dst, src); 3676 } 3677} 3678 3679void MacroAssembler::movdqu(XMMRegister dst, Address src) { 3680 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (dst->encoding() > 15)) { 3681 Assembler::vinsertf32x4h(dst, src, 0); 3682 } else { 3683 Assembler::movdqu(dst, src); 3684 } 3685} 3686 3687void MacroAssembler::movdqu(XMMRegister dst, XMMRegister src) { 3688 if (UseAVX > 2 && !VM_Version::supports_avx512vl()) { 3689 Assembler::evmovdqul(dst, src, Assembler::AVX_512bit); 3690 } else { 3691 Assembler::movdqu(dst, src); 3692 } 3693} 3694 3695void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src) { 3696 if (reachable(src)) { 3697 movdqu(dst, as_Address(src)); 3698 } else { 3699 lea(rscratch1, src); 3700 movdqu(dst, Address(rscratch1, 0)); 3701 } 3702} 3703 3704void MacroAssembler::vmovdqu(Address dst, XMMRegister src) { 3705 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (src->encoding() > 15)) { 3706 Assembler::vextractf64x4h(dst, src, 0); 3707 } else { 3708 Assembler::vmovdqu(dst, src); 3709 } 3710} 3711 3712void MacroAssembler::vmovdqu(XMMRegister dst, Address src) { 3713 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (dst->encoding() > 15)) { 3714 Assembler::vinsertf64x4h(dst, src, 0); 3715 } else { 3716 Assembler::vmovdqu(dst, src); 3717 } 3718} 3719 3720void MacroAssembler::vmovdqu(XMMRegister dst, XMMRegister src) { 3721 if (UseAVX > 2 && !VM_Version::supports_avx512vl()) { 3722 Assembler::evmovdqul(dst, src, Assembler::AVX_512bit); 3723 } 3724 else { 3725 Assembler::vmovdqu(dst, src); 3726 } 3727} 3728 3729void MacroAssembler::vmovdqu(XMMRegister dst, AddressLiteral src) { 3730 if (reachable(src)) { 3731 vmovdqu(dst, as_Address(src)); 3732 } 3733 else { 3734 lea(rscratch1, src); 3735 vmovdqu(dst, Address(rscratch1, 0)); 3736 } 3737} 3738 3739void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) { 3740 if (reachable(src)) { 3741 Assembler::movdqa(dst, as_Address(src)); 3742 } else { 3743 lea(rscratch1, src); 3744 Assembler::movdqa(dst, Address(rscratch1, 0)); 3745 } 3746} 3747 3748void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) { 3749 if (reachable(src)) { 3750 Assembler::movsd(dst, as_Address(src)); 3751 } else { 3752 lea(rscratch1, src); 3753 Assembler::movsd(dst, Address(rscratch1, 0)); 3754 } 3755} 3756 3757void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) { 3758 if (reachable(src)) { 3759 Assembler::movss(dst, as_Address(src)); 3760 } else { 3761 lea(rscratch1, src); 3762 Assembler::movss(dst, Address(rscratch1, 0)); 3763 } 3764} 3765 3766void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) { 3767 if (reachable(src)) { 3768 Assembler::mulsd(dst, as_Address(src)); 3769 } else { 3770 lea(rscratch1, src); 3771 Assembler::mulsd(dst, Address(rscratch1, 0)); 3772 } 3773} 3774 3775void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) { 3776 if (reachable(src)) { 3777 Assembler::mulss(dst, as_Address(src)); 3778 } else { 3779 lea(rscratch1, src); 3780 Assembler::mulss(dst, Address(rscratch1, 0)); 3781 } 3782} 3783 3784void MacroAssembler::null_check(Register reg, int offset) { 3785 if (needs_explicit_null_check(offset)) { 3786 // provoke OS NULL exception if reg = NULL by 3787 // accessing M[reg] w/o changing any (non-CC) registers 3788 // NOTE: cmpl is plenty here to provoke a segv 3789 cmpptr(rax, Address(reg, 0)); 3790 // Note: should probably use testl(rax, Address(reg, 0)); 3791 // may be shorter code (however, this version of 3792 // testl needs to be implemented first) 3793 } else { 3794 // nothing to do, (later) access of M[reg + offset] 3795 // will provoke OS NULL exception if reg = NULL 3796 } 3797} 3798 3799void MacroAssembler::os_breakpoint() { 3800 // instead of directly emitting a breakpoint, call os:breakpoint for better debugability 3801 // (e.g., MSVC can't call ps() otherwise) 3802 call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint))); 3803} 3804 3805#ifdef _LP64 3806#define XSTATE_BV 0x200 3807#endif 3808 3809void MacroAssembler::pop_CPU_state() { 3810 pop_FPU_state(); 3811 pop_IU_state(); 3812} 3813 3814void MacroAssembler::pop_FPU_state() { 3815#ifndef _LP64 3816 frstor(Address(rsp, 0)); 3817#else 3818 fxrstor(Address(rsp, 0)); 3819#endif 3820 addptr(rsp, FPUStateSizeInWords * wordSize); 3821} 3822 3823void MacroAssembler::pop_IU_state() { 3824 popa(); 3825 LP64_ONLY(addq(rsp, 8)); 3826 popf(); 3827} 3828 3829// Save Integer and Float state 3830// Warning: Stack must be 16 byte aligned (64bit) 3831void MacroAssembler::push_CPU_state() { 3832 push_IU_state(); 3833 push_FPU_state(); 3834} 3835 3836void MacroAssembler::push_FPU_state() { 3837 subptr(rsp, FPUStateSizeInWords * wordSize); 3838#ifndef _LP64 3839 fnsave(Address(rsp, 0)); 3840 fwait(); 3841#else 3842 fxsave(Address(rsp, 0)); 3843#endif // LP64 3844} 3845 3846void MacroAssembler::push_IU_state() { 3847 // Push flags first because pusha kills them 3848 pushf(); 3849 // Make sure rsp stays 16-byte aligned 3850 LP64_ONLY(subq(rsp, 8)); 3851 pusha(); 3852} 3853 3854void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp, bool clear_pc) { 3855 // determine java_thread register 3856 if (!java_thread->is_valid()) { 3857 java_thread = rdi; 3858 get_thread(java_thread); 3859 } 3860 // we must set sp to zero to clear frame 3861 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD); 3862 if (clear_fp) { 3863 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD); 3864 } 3865 3866 if (clear_pc) 3867 movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD); 3868 3869} 3870 3871void MacroAssembler::restore_rax(Register tmp) { 3872 if (tmp == noreg) pop(rax); 3873 else if (tmp != rax) mov(rax, tmp); 3874} 3875 3876void MacroAssembler::round_to(Register reg, int modulus) { 3877 addptr(reg, modulus - 1); 3878 andptr(reg, -modulus); 3879} 3880 3881void MacroAssembler::save_rax(Register tmp) { 3882 if (tmp == noreg) push(rax); 3883 else if (tmp != rax) mov(tmp, rax); 3884} 3885 3886// Write serialization page so VM thread can do a pseudo remote membar. 3887// We use the current thread pointer to calculate a thread specific 3888// offset to write to within the page. This minimizes bus traffic 3889// due to cache line collision. 3890void MacroAssembler::serialize_memory(Register thread, Register tmp) { 3891 movl(tmp, thread); 3892 shrl(tmp, os::get_serialize_page_shift_count()); 3893 andl(tmp, (os::vm_page_size() - sizeof(int))); 3894 3895 Address index(noreg, tmp, Address::times_1); 3896 ExternalAddress page(os::get_memory_serialize_page()); 3897 3898 // Size of store must match masking code above 3899 movl(as_Address(ArrayAddress(page, index)), tmp); 3900} 3901 3902// Calls to C land 3903// 3904// When entering C land, the rbp, & rsp of the last Java frame have to be recorded 3905// in the (thread-local) JavaThread object. When leaving C land, the last Java fp 3906// has to be reset to 0. This is required to allow proper stack traversal. 3907void MacroAssembler::set_last_Java_frame(Register java_thread, 3908 Register last_java_sp, 3909 Register last_java_fp, 3910 address last_java_pc) { 3911 // determine java_thread register 3912 if (!java_thread->is_valid()) { 3913 java_thread = rdi; 3914 get_thread(java_thread); 3915 } 3916 // determine last_java_sp register 3917 if (!last_java_sp->is_valid()) { 3918 last_java_sp = rsp; 3919 } 3920 3921 // last_java_fp is optional 3922 3923 if (last_java_fp->is_valid()) { 3924 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp); 3925 } 3926 3927 // last_java_pc is optional 3928 3929 if (last_java_pc != NULL) { 3930 lea(Address(java_thread, 3931 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()), 3932 InternalAddress(last_java_pc)); 3933 3934 } 3935 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp); 3936} 3937 3938void MacroAssembler::shlptr(Register dst, int imm8) { 3939 LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8)); 3940} 3941 3942void MacroAssembler::shrptr(Register dst, int imm8) { 3943 LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8)); 3944} 3945 3946void MacroAssembler::sign_extend_byte(Register reg) { 3947 if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) { 3948 movsbl(reg, reg); // movsxb 3949 } else { 3950 shll(reg, 24); 3951 sarl(reg, 24); 3952 } 3953} 3954 3955void MacroAssembler::sign_extend_short(Register reg) { 3956 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3957 movswl(reg, reg); // movsxw 3958 } else { 3959 shll(reg, 16); 3960 sarl(reg, 16); 3961 } 3962} 3963 3964void MacroAssembler::testl(Register dst, AddressLiteral src) { 3965 assert(reachable(src), "Address should be reachable"); 3966 testl(dst, as_Address(src)); 3967} 3968 3969void MacroAssembler::pcmpeqb(XMMRegister dst, XMMRegister src) { 3970 int dst_enc = dst->encoding(); 3971 int src_enc = src->encoding(); 3972 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 3973 Assembler::pcmpeqb(dst, src); 3974 } else if ((dst_enc < 16) && (src_enc < 16)) { 3975 Assembler::pcmpeqb(dst, src); 3976 } else if (src_enc < 16) { 3977 subptr(rsp, 64); 3978 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3979 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3980 Assembler::pcmpeqb(xmm0, src); 3981 movdqu(dst, xmm0); 3982 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3983 addptr(rsp, 64); 3984 } else if (dst_enc < 16) { 3985 subptr(rsp, 64); 3986 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3987 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3988 Assembler::pcmpeqb(dst, xmm0); 3989 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3990 addptr(rsp, 64); 3991 } else { 3992 subptr(rsp, 64); 3993 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3994 subptr(rsp, 64); 3995 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 3996 movdqu(xmm0, src); 3997 movdqu(xmm1, dst); 3998 Assembler::pcmpeqb(xmm1, xmm0); 3999 movdqu(dst, xmm1); 4000 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4001 addptr(rsp, 64); 4002 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4003 addptr(rsp, 64); 4004 } 4005} 4006 4007void MacroAssembler::pcmpeqw(XMMRegister dst, XMMRegister src) { 4008 int dst_enc = dst->encoding(); 4009 int src_enc = src->encoding(); 4010 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4011 Assembler::pcmpeqw(dst, src); 4012 } else if ((dst_enc < 16) && (src_enc < 16)) { 4013 Assembler::pcmpeqw(dst, src); 4014 } else if (src_enc < 16) { 4015 subptr(rsp, 64); 4016 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4017 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4018 Assembler::pcmpeqw(xmm0, src); 4019 movdqu(dst, xmm0); 4020 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4021 addptr(rsp, 64); 4022 } else if (dst_enc < 16) { 4023 subptr(rsp, 64); 4024 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4025 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4026 Assembler::pcmpeqw(dst, xmm0); 4027 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4028 addptr(rsp, 64); 4029 } else { 4030 subptr(rsp, 64); 4031 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4032 subptr(rsp, 64); 4033 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4034 movdqu(xmm0, src); 4035 movdqu(xmm1, dst); 4036 Assembler::pcmpeqw(xmm1, xmm0); 4037 movdqu(dst, xmm1); 4038 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4039 addptr(rsp, 64); 4040 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4041 addptr(rsp, 64); 4042 } 4043} 4044 4045void MacroAssembler::pcmpestri(XMMRegister dst, Address src, int imm8) { 4046 int dst_enc = dst->encoding(); 4047 if (dst_enc < 16) { 4048 Assembler::pcmpestri(dst, src, imm8); 4049 } else { 4050 subptr(rsp, 64); 4051 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4052 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4053 Assembler::pcmpestri(xmm0, src, imm8); 4054 movdqu(dst, xmm0); 4055 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4056 addptr(rsp, 64); 4057 } 4058} 4059 4060void MacroAssembler::pcmpestri(XMMRegister dst, XMMRegister src, int imm8) { 4061 int dst_enc = dst->encoding(); 4062 int src_enc = src->encoding(); 4063 if ((dst_enc < 16) && (src_enc < 16)) { 4064 Assembler::pcmpestri(dst, src, imm8); 4065 } else if (src_enc < 16) { 4066 subptr(rsp, 64); 4067 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4068 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4069 Assembler::pcmpestri(xmm0, src, imm8); 4070 movdqu(dst, xmm0); 4071 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4072 addptr(rsp, 64); 4073 } else if (dst_enc < 16) { 4074 subptr(rsp, 64); 4075 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4076 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4077 Assembler::pcmpestri(dst, xmm0, imm8); 4078 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4079 addptr(rsp, 64); 4080 } else { 4081 subptr(rsp, 64); 4082 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4083 subptr(rsp, 64); 4084 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4085 movdqu(xmm0, src); 4086 movdqu(xmm1, dst); 4087 Assembler::pcmpestri(xmm1, xmm0, imm8); 4088 movdqu(dst, xmm1); 4089 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4090 addptr(rsp, 64); 4091 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4092 addptr(rsp, 64); 4093 } 4094} 4095 4096void MacroAssembler::pmovzxbw(XMMRegister dst, XMMRegister src) { 4097 int dst_enc = dst->encoding(); 4098 int src_enc = src->encoding(); 4099 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4100 Assembler::pmovzxbw(dst, src); 4101 } else if ((dst_enc < 16) && (src_enc < 16)) { 4102 Assembler::pmovzxbw(dst, src); 4103 } else if (src_enc < 16) { 4104 subptr(rsp, 64); 4105 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4106 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4107 Assembler::pmovzxbw(xmm0, src); 4108 movdqu(dst, xmm0); 4109 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4110 addptr(rsp, 64); 4111 } else if (dst_enc < 16) { 4112 subptr(rsp, 64); 4113 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4114 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4115 Assembler::pmovzxbw(dst, xmm0); 4116 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4117 addptr(rsp, 64); 4118 } else { 4119 subptr(rsp, 64); 4120 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4121 subptr(rsp, 64); 4122 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4123 movdqu(xmm0, src); 4124 movdqu(xmm1, dst); 4125 Assembler::pmovzxbw(xmm1, xmm0); 4126 movdqu(dst, xmm1); 4127 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4128 addptr(rsp, 64); 4129 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4130 addptr(rsp, 64); 4131 } 4132} 4133 4134void MacroAssembler::pmovzxbw(XMMRegister dst, Address src) { 4135 int dst_enc = dst->encoding(); 4136 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4137 Assembler::pmovzxbw(dst, src); 4138 } else if (dst_enc < 16) { 4139 Assembler::pmovzxbw(dst, src); 4140 } else { 4141 subptr(rsp, 64); 4142 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4143 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4144 Assembler::pmovzxbw(xmm0, src); 4145 movdqu(dst, xmm0); 4146 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4147 addptr(rsp, 64); 4148 } 4149} 4150 4151void MacroAssembler::pmovmskb(Register dst, XMMRegister src) { 4152 int src_enc = src->encoding(); 4153 if (src_enc < 16) { 4154 Assembler::pmovmskb(dst, src); 4155 } else { 4156 subptr(rsp, 64); 4157 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4158 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4159 Assembler::pmovmskb(dst, xmm0); 4160 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4161 addptr(rsp, 64); 4162 } 4163} 4164 4165void MacroAssembler::ptest(XMMRegister dst, XMMRegister src) { 4166 int dst_enc = dst->encoding(); 4167 int src_enc = src->encoding(); 4168 if ((dst_enc < 16) && (src_enc < 16)) { 4169 Assembler::ptest(dst, src); 4170 } else if (src_enc < 16) { 4171 subptr(rsp, 64); 4172 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4173 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4174 Assembler::ptest(xmm0, src); 4175 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4176 addptr(rsp, 64); 4177 } else if (dst_enc < 16) { 4178 subptr(rsp, 64); 4179 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4180 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4181 Assembler::ptest(dst, xmm0); 4182 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4183 addptr(rsp, 64); 4184 } else { 4185 subptr(rsp, 64); 4186 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4187 subptr(rsp, 64); 4188 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4189 movdqu(xmm0, src); 4190 movdqu(xmm1, dst); 4191 Assembler::ptest(xmm1, xmm0); 4192 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4193 addptr(rsp, 64); 4194 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4195 addptr(rsp, 64); 4196 } 4197} 4198 4199void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) { 4200 if (reachable(src)) { 4201 Assembler::sqrtsd(dst, as_Address(src)); 4202 } else { 4203 lea(rscratch1, src); 4204 Assembler::sqrtsd(dst, Address(rscratch1, 0)); 4205 } 4206} 4207 4208void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) { 4209 if (reachable(src)) { 4210 Assembler::sqrtss(dst, as_Address(src)); 4211 } else { 4212 lea(rscratch1, src); 4213 Assembler::sqrtss(dst, Address(rscratch1, 0)); 4214 } 4215} 4216 4217void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) { 4218 if (reachable(src)) { 4219 Assembler::subsd(dst, as_Address(src)); 4220 } else { 4221 lea(rscratch1, src); 4222 Assembler::subsd(dst, Address(rscratch1, 0)); 4223 } 4224} 4225 4226void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) { 4227 if (reachable(src)) { 4228 Assembler::subss(dst, as_Address(src)); 4229 } else { 4230 lea(rscratch1, src); 4231 Assembler::subss(dst, Address(rscratch1, 0)); 4232 } 4233} 4234 4235void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) { 4236 if (reachable(src)) { 4237 Assembler::ucomisd(dst, as_Address(src)); 4238 } else { 4239 lea(rscratch1, src); 4240 Assembler::ucomisd(dst, Address(rscratch1, 0)); 4241 } 4242} 4243 4244void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) { 4245 if (reachable(src)) { 4246 Assembler::ucomiss(dst, as_Address(src)); 4247 } else { 4248 lea(rscratch1, src); 4249 Assembler::ucomiss(dst, Address(rscratch1, 0)); 4250 } 4251} 4252 4253void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src) { 4254 // Used in sign-bit flipping with aligned address. 4255 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 4256 if (reachable(src)) { 4257 Assembler::xorpd(dst, as_Address(src)); 4258 } else { 4259 lea(rscratch1, src); 4260 Assembler::xorpd(dst, Address(rscratch1, 0)); 4261 } 4262} 4263 4264void MacroAssembler::xorpd(XMMRegister dst, XMMRegister src) { 4265 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) { 4266 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit); 4267 } 4268 else { 4269 Assembler::xorpd(dst, src); 4270 } 4271} 4272 4273void MacroAssembler::xorps(XMMRegister dst, XMMRegister src) { 4274 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) { 4275 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit); 4276 } else { 4277 Assembler::xorps(dst, src); 4278 } 4279} 4280 4281void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src) { 4282 // Used in sign-bit flipping with aligned address. 4283 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 4284 if (reachable(src)) { 4285 Assembler::xorps(dst, as_Address(src)); 4286 } else { 4287 lea(rscratch1, src); 4288 Assembler::xorps(dst, Address(rscratch1, 0)); 4289 } 4290} 4291 4292void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) { 4293 // Used in sign-bit flipping with aligned address. 4294 bool aligned_adr = (((intptr_t)src.target() & 15) == 0); 4295 assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes"); 4296 if (reachable(src)) { 4297 Assembler::pshufb(dst, as_Address(src)); 4298 } else { 4299 lea(rscratch1, src); 4300 Assembler::pshufb(dst, Address(rscratch1, 0)); 4301 } 4302} 4303 4304// AVX 3-operands instructions 4305 4306void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4307 if (reachable(src)) { 4308 vaddsd(dst, nds, as_Address(src)); 4309 } else { 4310 lea(rscratch1, src); 4311 vaddsd(dst, nds, Address(rscratch1, 0)); 4312 } 4313} 4314 4315void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4316 if (reachable(src)) { 4317 vaddss(dst, nds, as_Address(src)); 4318 } else { 4319 lea(rscratch1, src); 4320 vaddss(dst, nds, Address(rscratch1, 0)); 4321 } 4322} 4323 4324void MacroAssembler::vabsss(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) { 4325 int dst_enc = dst->encoding(); 4326 int nds_enc = nds->encoding(); 4327 int src_enc = src->encoding(); 4328 if ((dst_enc < 16) && (nds_enc < 16)) { 4329 vandps(dst, nds, negate_field, vector_len); 4330 } else if ((src_enc < 16) && (dst_enc < 16)) { 4331 movss(src, nds); 4332 vandps(dst, src, negate_field, vector_len); 4333 } else if (src_enc < 16) { 4334 movss(src, nds); 4335 vandps(src, src, negate_field, vector_len); 4336 movss(dst, src); 4337 } else if (dst_enc < 16) { 4338 movdqu(src, xmm0); 4339 movss(xmm0, nds); 4340 vandps(dst, xmm0, negate_field, vector_len); 4341 movdqu(xmm0, src); 4342 } else if (nds_enc < 16) { 4343 movdqu(src, xmm0); 4344 vandps(xmm0, nds, negate_field, vector_len); 4345 movss(dst, xmm0); 4346 movdqu(xmm0, src); 4347 } else { 4348 movdqu(src, xmm0); 4349 movss(xmm0, nds); 4350 vandps(xmm0, xmm0, negate_field, vector_len); 4351 movss(dst, xmm0); 4352 movdqu(xmm0, src); 4353 } 4354} 4355 4356void MacroAssembler::vabssd(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) { 4357 int dst_enc = dst->encoding(); 4358 int nds_enc = nds->encoding(); 4359 int src_enc = src->encoding(); 4360 if ((dst_enc < 16) && (nds_enc < 16)) { 4361 vandpd(dst, nds, negate_field, vector_len); 4362 } else if ((src_enc < 16) && (dst_enc < 16)) { 4363 movsd(src, nds); 4364 vandpd(dst, src, negate_field, vector_len); 4365 } else if (src_enc < 16) { 4366 movsd(src, nds); 4367 vandpd(src, src, negate_field, vector_len); 4368 movsd(dst, src); 4369 } else if (dst_enc < 16) { 4370 movdqu(src, xmm0); 4371 movsd(xmm0, nds); 4372 vandpd(dst, xmm0, negate_field, vector_len); 4373 movdqu(xmm0, src); 4374 } else if (nds_enc < 16) { 4375 movdqu(src, xmm0); 4376 vandpd(xmm0, nds, negate_field, vector_len); 4377 movsd(dst, xmm0); 4378 movdqu(xmm0, src); 4379 } else { 4380 movdqu(src, xmm0); 4381 movsd(xmm0, nds); 4382 vandpd(xmm0, xmm0, negate_field, vector_len); 4383 movsd(dst, xmm0); 4384 movdqu(xmm0, src); 4385 } 4386} 4387 4388void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4389 int dst_enc = dst->encoding(); 4390 int nds_enc = nds->encoding(); 4391 int src_enc = src->encoding(); 4392 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4393 Assembler::vpaddb(dst, nds, src, vector_len); 4394 } else if ((dst_enc < 16) && (src_enc < 16)) { 4395 Assembler::vpaddb(dst, dst, src, vector_len); 4396 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4397 // use nds as scratch for src 4398 evmovdqul(nds, src, Assembler::AVX_512bit); 4399 Assembler::vpaddb(dst, dst, nds, vector_len); 4400 } else if ((src_enc < 16) && (nds_enc < 16)) { 4401 // use nds as scratch for dst 4402 evmovdqul(nds, dst, Assembler::AVX_512bit); 4403 Assembler::vpaddb(nds, nds, src, vector_len); 4404 evmovdqul(dst, nds, Assembler::AVX_512bit); 4405 } else if (dst_enc < 16) { 4406 // use nds as scatch for xmm0 to hold src 4407 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4408 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4409 Assembler::vpaddb(dst, dst, xmm0, vector_len); 4410 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4411 } else { 4412 // worse case scenario, all regs are in the upper bank 4413 subptr(rsp, 64); 4414 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4415 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4416 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4417 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4418 Assembler::vpaddb(xmm0, xmm0, xmm1, vector_len); 4419 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4420 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4421 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4422 addptr(rsp, 64); 4423 } 4424} 4425 4426void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4427 int dst_enc = dst->encoding(); 4428 int nds_enc = nds->encoding(); 4429 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4430 Assembler::vpaddb(dst, nds, src, vector_len); 4431 } else if (dst_enc < 16) { 4432 Assembler::vpaddb(dst, dst, src, vector_len); 4433 } else if (nds_enc < 16) { 4434 // implies dst_enc in upper bank with src as scratch 4435 evmovdqul(nds, dst, Assembler::AVX_512bit); 4436 Assembler::vpaddb(nds, nds, src, vector_len); 4437 evmovdqul(dst, nds, Assembler::AVX_512bit); 4438 } else { 4439 // worse case scenario, all regs in upper bank 4440 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4441 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4442 Assembler::vpaddb(xmm0, xmm0, src, vector_len); 4443 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4444 } 4445} 4446 4447void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4448 int dst_enc = dst->encoding(); 4449 int nds_enc = nds->encoding(); 4450 int src_enc = src->encoding(); 4451 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4452 Assembler::vpaddw(dst, nds, src, vector_len); 4453 } else if ((dst_enc < 16) && (src_enc < 16)) { 4454 Assembler::vpaddw(dst, dst, src, vector_len); 4455 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4456 // use nds as scratch for src 4457 evmovdqul(nds, src, Assembler::AVX_512bit); 4458 Assembler::vpaddw(dst, dst, nds, vector_len); 4459 } else if ((src_enc < 16) && (nds_enc < 16)) { 4460 // use nds as scratch for dst 4461 evmovdqul(nds, dst, Assembler::AVX_512bit); 4462 Assembler::vpaddw(nds, nds, src, vector_len); 4463 evmovdqul(dst, nds, Assembler::AVX_512bit); 4464 } else if (dst_enc < 16) { 4465 // use nds as scatch for xmm0 to hold src 4466 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4467 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4468 Assembler::vpaddw(dst, dst, xmm0, vector_len); 4469 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4470 } else { 4471 // worse case scenario, all regs are in the upper bank 4472 subptr(rsp, 64); 4473 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4474 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4475 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4476 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4477 Assembler::vpaddw(xmm0, xmm0, xmm1, vector_len); 4478 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4479 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4480 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4481 addptr(rsp, 64); 4482 } 4483} 4484 4485void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4486 int dst_enc = dst->encoding(); 4487 int nds_enc = nds->encoding(); 4488 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4489 Assembler::vpaddw(dst, nds, src, vector_len); 4490 } else if (dst_enc < 16) { 4491 Assembler::vpaddw(dst, dst, src, vector_len); 4492 } else if (nds_enc < 16) { 4493 // implies dst_enc in upper bank with src as scratch 4494 evmovdqul(nds, dst, Assembler::AVX_512bit); 4495 Assembler::vpaddw(nds, nds, src, vector_len); 4496 evmovdqul(dst, nds, Assembler::AVX_512bit); 4497 } else { 4498 // worse case scenario, all regs in upper bank 4499 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4500 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4501 Assembler::vpaddw(xmm0, xmm0, src, vector_len); 4502 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4503 } 4504} 4505 4506void MacroAssembler::vpbroadcastw(XMMRegister dst, XMMRegister src) { 4507 int dst_enc = dst->encoding(); 4508 int src_enc = src->encoding(); 4509 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4510 Assembler::vpbroadcastw(dst, src); 4511 } else if ((dst_enc < 16) && (src_enc < 16)) { 4512 Assembler::vpbroadcastw(dst, src); 4513 } else if (src_enc < 16) { 4514 subptr(rsp, 64); 4515 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4516 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4517 Assembler::vpbroadcastw(xmm0, src); 4518 movdqu(dst, xmm0); 4519 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4520 addptr(rsp, 64); 4521 } else if (dst_enc < 16) { 4522 subptr(rsp, 64); 4523 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4524 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4525 Assembler::vpbroadcastw(dst, xmm0); 4526 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4527 addptr(rsp, 64); 4528 } else { 4529 subptr(rsp, 64); 4530 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4531 subptr(rsp, 64); 4532 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4533 movdqu(xmm0, src); 4534 movdqu(xmm1, dst); 4535 Assembler::vpbroadcastw(xmm1, xmm0); 4536 movdqu(dst, xmm1); 4537 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4538 addptr(rsp, 64); 4539 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4540 addptr(rsp, 64); 4541 } 4542} 4543 4544void MacroAssembler::vpcmpeqb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4545 int dst_enc = dst->encoding(); 4546 int nds_enc = nds->encoding(); 4547 int src_enc = src->encoding(); 4548 assert(dst_enc == nds_enc, ""); 4549 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4550 Assembler::vpcmpeqb(dst, nds, src, vector_len); 4551 } else if ((dst_enc < 16) && (src_enc < 16)) { 4552 Assembler::vpcmpeqb(dst, nds, src, vector_len); 4553 } else if (src_enc < 16) { 4554 subptr(rsp, 64); 4555 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4556 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4557 Assembler::vpcmpeqb(xmm0, xmm0, src, vector_len); 4558 movdqu(dst, xmm0); 4559 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4560 addptr(rsp, 64); 4561 } else if (dst_enc < 16) { 4562 subptr(rsp, 64); 4563 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4564 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4565 Assembler::vpcmpeqb(dst, dst, xmm0, vector_len); 4566 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4567 addptr(rsp, 64); 4568 } else { 4569 subptr(rsp, 64); 4570 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4571 subptr(rsp, 64); 4572 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4573 movdqu(xmm0, src); 4574 movdqu(xmm1, dst); 4575 Assembler::vpcmpeqb(xmm1, xmm1, xmm0, vector_len); 4576 movdqu(dst, xmm1); 4577 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4578 addptr(rsp, 64); 4579 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4580 addptr(rsp, 64); 4581 } 4582} 4583 4584void MacroAssembler::vpcmpeqw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4585 int dst_enc = dst->encoding(); 4586 int nds_enc = nds->encoding(); 4587 int src_enc = src->encoding(); 4588 assert(dst_enc == nds_enc, ""); 4589 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4590 Assembler::vpcmpeqw(dst, nds, src, vector_len); 4591 } else if ((dst_enc < 16) && (src_enc < 16)) { 4592 Assembler::vpcmpeqw(dst, nds, src, vector_len); 4593 } else if (src_enc < 16) { 4594 subptr(rsp, 64); 4595 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4596 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4597 Assembler::vpcmpeqw(xmm0, xmm0, src, vector_len); 4598 movdqu(dst, xmm0); 4599 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4600 addptr(rsp, 64); 4601 } else if (dst_enc < 16) { 4602 subptr(rsp, 64); 4603 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4604 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4605 Assembler::vpcmpeqw(dst, dst, xmm0, vector_len); 4606 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4607 addptr(rsp, 64); 4608 } else { 4609 subptr(rsp, 64); 4610 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4611 subptr(rsp, 64); 4612 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4613 movdqu(xmm0, src); 4614 movdqu(xmm1, dst); 4615 Assembler::vpcmpeqw(xmm1, xmm1, xmm0, vector_len); 4616 movdqu(dst, xmm1); 4617 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4618 addptr(rsp, 64); 4619 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4620 addptr(rsp, 64); 4621 } 4622} 4623 4624void MacroAssembler::vpmovzxbw(XMMRegister dst, Address src, int vector_len) { 4625 int dst_enc = dst->encoding(); 4626 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4627 Assembler::vpmovzxbw(dst, src, vector_len); 4628 } else if (dst_enc < 16) { 4629 Assembler::vpmovzxbw(dst, src, vector_len); 4630 } else { 4631 subptr(rsp, 64); 4632 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4633 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4634 Assembler::vpmovzxbw(xmm0, src, vector_len); 4635 movdqu(dst, xmm0); 4636 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4637 addptr(rsp, 64); 4638 } 4639} 4640 4641void MacroAssembler::vpmovmskb(Register dst, XMMRegister src) { 4642 int src_enc = src->encoding(); 4643 if (src_enc < 16) { 4644 Assembler::vpmovmskb(dst, src); 4645 } else { 4646 subptr(rsp, 64); 4647 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4648 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4649 Assembler::vpmovmskb(dst, xmm0); 4650 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4651 addptr(rsp, 64); 4652 } 4653} 4654 4655void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4656 int dst_enc = dst->encoding(); 4657 int nds_enc = nds->encoding(); 4658 int src_enc = src->encoding(); 4659 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4660 Assembler::vpmullw(dst, nds, src, vector_len); 4661 } else if ((dst_enc < 16) && (src_enc < 16)) { 4662 Assembler::vpmullw(dst, dst, src, vector_len); 4663 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4664 // use nds as scratch for src 4665 evmovdqul(nds, src, Assembler::AVX_512bit); 4666 Assembler::vpmullw(dst, dst, nds, vector_len); 4667 } else if ((src_enc < 16) && (nds_enc < 16)) { 4668 // use nds as scratch for dst 4669 evmovdqul(nds, dst, Assembler::AVX_512bit); 4670 Assembler::vpmullw(nds, nds, src, vector_len); 4671 evmovdqul(dst, nds, Assembler::AVX_512bit); 4672 } else if (dst_enc < 16) { 4673 // use nds as scatch for xmm0 to hold src 4674 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4675 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4676 Assembler::vpmullw(dst, dst, xmm0, vector_len); 4677 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4678 } else { 4679 // worse case scenario, all regs are in the upper bank 4680 subptr(rsp, 64); 4681 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4682 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4683 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4684 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4685 Assembler::vpmullw(xmm0, xmm0, xmm1, vector_len); 4686 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4687 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4688 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4689 addptr(rsp, 64); 4690 } 4691} 4692 4693void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4694 int dst_enc = dst->encoding(); 4695 int nds_enc = nds->encoding(); 4696 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4697 Assembler::vpmullw(dst, nds, src, vector_len); 4698 } else if (dst_enc < 16) { 4699 Assembler::vpmullw(dst, dst, src, vector_len); 4700 } else if (nds_enc < 16) { 4701 // implies dst_enc in upper bank with src as scratch 4702 evmovdqul(nds, dst, Assembler::AVX_512bit); 4703 Assembler::vpmullw(nds, nds, src, vector_len); 4704 evmovdqul(dst, nds, Assembler::AVX_512bit); 4705 } else { 4706 // worse case scenario, all regs in upper bank 4707 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4708 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4709 Assembler::vpmullw(xmm0, xmm0, src, vector_len); 4710 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4711 } 4712} 4713 4714void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4715 int dst_enc = dst->encoding(); 4716 int nds_enc = nds->encoding(); 4717 int src_enc = src->encoding(); 4718 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4719 Assembler::vpsubb(dst, nds, src, vector_len); 4720 } else if ((dst_enc < 16) && (src_enc < 16)) { 4721 Assembler::vpsubb(dst, dst, src, vector_len); 4722 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4723 // use nds as scratch for src 4724 evmovdqul(nds, src, Assembler::AVX_512bit); 4725 Assembler::vpsubb(dst, dst, nds, vector_len); 4726 } else if ((src_enc < 16) && (nds_enc < 16)) { 4727 // use nds as scratch for dst 4728 evmovdqul(nds, dst, Assembler::AVX_512bit); 4729 Assembler::vpsubb(nds, nds, src, vector_len); 4730 evmovdqul(dst, nds, Assembler::AVX_512bit); 4731 } else if (dst_enc < 16) { 4732 // use nds as scatch for xmm0 to hold src 4733 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4734 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4735 Assembler::vpsubb(dst, dst, xmm0, vector_len); 4736 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4737 } else { 4738 // worse case scenario, all regs are in the upper bank 4739 subptr(rsp, 64); 4740 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4741 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4742 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4743 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4744 Assembler::vpsubb(xmm0, xmm0, xmm1, vector_len); 4745 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4746 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4747 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4748 addptr(rsp, 64); 4749 } 4750} 4751 4752void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4753 int dst_enc = dst->encoding(); 4754 int nds_enc = nds->encoding(); 4755 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4756 Assembler::vpsubb(dst, nds, src, vector_len); 4757 } else if (dst_enc < 16) { 4758 Assembler::vpsubb(dst, dst, src, vector_len); 4759 } else if (nds_enc < 16) { 4760 // implies dst_enc in upper bank with src as scratch 4761 evmovdqul(nds, dst, Assembler::AVX_512bit); 4762 Assembler::vpsubb(nds, nds, src, vector_len); 4763 evmovdqul(dst, nds, Assembler::AVX_512bit); 4764 } else { 4765 // worse case scenario, all regs in upper bank 4766 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4767 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4768 Assembler::vpsubw(xmm0, xmm0, src, vector_len); 4769 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4770 } 4771} 4772 4773void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4774 int dst_enc = dst->encoding(); 4775 int nds_enc = nds->encoding(); 4776 int src_enc = src->encoding(); 4777 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4778 Assembler::vpsubw(dst, nds, src, vector_len); 4779 } else if ((dst_enc < 16) && (src_enc < 16)) { 4780 Assembler::vpsubw(dst, dst, src, vector_len); 4781 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4782 // use nds as scratch for src 4783 evmovdqul(nds, src, Assembler::AVX_512bit); 4784 Assembler::vpsubw(dst, dst, nds, vector_len); 4785 } else if ((src_enc < 16) && (nds_enc < 16)) { 4786 // use nds as scratch for dst 4787 evmovdqul(nds, dst, Assembler::AVX_512bit); 4788 Assembler::vpsubw(nds, nds, src, vector_len); 4789 evmovdqul(dst, nds, Assembler::AVX_512bit); 4790 } else if (dst_enc < 16) { 4791 // use nds as scatch for xmm0 to hold src 4792 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4793 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4794 Assembler::vpsubw(dst, dst, xmm0, vector_len); 4795 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4796 } else { 4797 // worse case scenario, all regs are in the upper bank 4798 subptr(rsp, 64); 4799 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4800 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4801 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4802 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4803 Assembler::vpsubw(xmm0, xmm0, xmm1, vector_len); 4804 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4805 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4806 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4807 addptr(rsp, 64); 4808 } 4809} 4810 4811void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4812 int dst_enc = dst->encoding(); 4813 int nds_enc = nds->encoding(); 4814 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4815 Assembler::vpsubw(dst, nds, src, vector_len); 4816 } else if (dst_enc < 16) { 4817 Assembler::vpsubw(dst, dst, src, vector_len); 4818 } else if (nds_enc < 16) { 4819 // implies dst_enc in upper bank with src as scratch 4820 evmovdqul(nds, dst, Assembler::AVX_512bit); 4821 Assembler::vpsubw(nds, nds, src, vector_len); 4822 evmovdqul(dst, nds, Assembler::AVX_512bit); 4823 } else { 4824 // worse case scenario, all regs in upper bank 4825 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4826 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4827 Assembler::vpsubw(xmm0, xmm0, src, vector_len); 4828 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4829 } 4830} 4831 4832void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4833 int dst_enc = dst->encoding(); 4834 int nds_enc = nds->encoding(); 4835 int shift_enc = shift->encoding(); 4836 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4837 Assembler::vpsraw(dst, nds, shift, vector_len); 4838 } else if ((dst_enc < 16) && (shift_enc < 16)) { 4839 Assembler::vpsraw(dst, dst, shift, vector_len); 4840 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4841 // use nds_enc as scratch with shift 4842 evmovdqul(nds, shift, Assembler::AVX_512bit); 4843 Assembler::vpsraw(dst, dst, nds, vector_len); 4844 } else if ((shift_enc < 16) && (nds_enc < 16)) { 4845 // use nds as scratch with dst 4846 evmovdqul(nds, dst, Assembler::AVX_512bit); 4847 Assembler::vpsraw(nds, nds, shift, vector_len); 4848 evmovdqul(dst, nds, Assembler::AVX_512bit); 4849 } else if (dst_enc < 16) { 4850 // use nds to save a copy of xmm0 and hold shift 4851 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4852 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4853 Assembler::vpsraw(dst, dst, xmm0, vector_len); 4854 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4855 } else if (nds_enc < 16) { 4856 // use nds as dest as temps 4857 evmovdqul(nds, dst, Assembler::AVX_512bit); 4858 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4859 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4860 Assembler::vpsraw(nds, nds, xmm0, vector_len); 4861 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4862 evmovdqul(dst, nds, Assembler::AVX_512bit); 4863 } else { 4864 // worse case scenario, all regs are in the upper bank 4865 subptr(rsp, 64); 4866 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4867 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4868 evmovdqul(xmm1, shift, Assembler::AVX_512bit); 4869 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4870 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len); 4871 evmovdqul(xmm1, dst, Assembler::AVX_512bit); 4872 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4873 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4874 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4875 addptr(rsp, 64); 4876 } 4877} 4878 4879void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 4880 int dst_enc = dst->encoding(); 4881 int nds_enc = nds->encoding(); 4882 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4883 Assembler::vpsraw(dst, nds, shift, vector_len); 4884 } else if (dst_enc < 16) { 4885 Assembler::vpsraw(dst, dst, shift, vector_len); 4886 } else if (nds_enc < 16) { 4887 // use nds as scratch 4888 evmovdqul(nds, dst, Assembler::AVX_512bit); 4889 Assembler::vpsraw(nds, nds, shift, vector_len); 4890 evmovdqul(dst, nds, Assembler::AVX_512bit); 4891 } else { 4892 // use nds as scratch for xmm0 4893 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4894 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4895 Assembler::vpsraw(xmm0, xmm0, shift, vector_len); 4896 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4897 } 4898} 4899 4900void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4901 int dst_enc = dst->encoding(); 4902 int nds_enc = nds->encoding(); 4903 int shift_enc = shift->encoding(); 4904 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4905 Assembler::vpsrlw(dst, nds, shift, vector_len); 4906 } else if ((dst_enc < 16) && (shift_enc < 16)) { 4907 Assembler::vpsrlw(dst, dst, shift, vector_len); 4908 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4909 // use nds_enc as scratch with shift 4910 evmovdqul(nds, shift, Assembler::AVX_512bit); 4911 Assembler::vpsrlw(dst, dst, nds, vector_len); 4912 } else if ((shift_enc < 16) && (nds_enc < 16)) { 4913 // use nds as scratch with dst 4914 evmovdqul(nds, dst, Assembler::AVX_512bit); 4915 Assembler::vpsrlw(nds, nds, shift, vector_len); 4916 evmovdqul(dst, nds, Assembler::AVX_512bit); 4917 } else if (dst_enc < 16) { 4918 // use nds to save a copy of xmm0 and hold shift 4919 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4920 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4921 Assembler::vpsrlw(dst, dst, xmm0, vector_len); 4922 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4923 } else if (nds_enc < 16) { 4924 // use nds as dest as temps 4925 evmovdqul(nds, dst, Assembler::AVX_512bit); 4926 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4927 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4928 Assembler::vpsrlw(nds, nds, xmm0, vector_len); 4929 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4930 evmovdqul(dst, nds, Assembler::AVX_512bit); 4931 } else { 4932 // worse case scenario, all regs are in the upper bank 4933 subptr(rsp, 64); 4934 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4935 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4936 evmovdqul(xmm1, shift, Assembler::AVX_512bit); 4937 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4938 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len); 4939 evmovdqul(xmm1, dst, Assembler::AVX_512bit); 4940 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4941 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4942 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4943 addptr(rsp, 64); 4944 } 4945} 4946 4947void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 4948 int dst_enc = dst->encoding(); 4949 int nds_enc = nds->encoding(); 4950 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4951 Assembler::vpsrlw(dst, nds, shift, vector_len); 4952 } else if (dst_enc < 16) { 4953 Assembler::vpsrlw(dst, dst, shift, vector_len); 4954 } else if (nds_enc < 16) { 4955 // use nds as scratch 4956 evmovdqul(nds, dst, Assembler::AVX_512bit); 4957 Assembler::vpsrlw(nds, nds, shift, vector_len); 4958 evmovdqul(dst, nds, Assembler::AVX_512bit); 4959 } else { 4960 // use nds as scratch for xmm0 4961 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4962 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4963 Assembler::vpsrlw(xmm0, xmm0, shift, vector_len); 4964 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4965 } 4966} 4967 4968void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4969 int dst_enc = dst->encoding(); 4970 int nds_enc = nds->encoding(); 4971 int shift_enc = shift->encoding(); 4972 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4973 Assembler::vpsllw(dst, nds, shift, vector_len); 4974 } else if ((dst_enc < 16) && (shift_enc < 16)) { 4975 Assembler::vpsllw(dst, dst, shift, vector_len); 4976 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4977 // use nds_enc as scratch with shift 4978 evmovdqul(nds, shift, Assembler::AVX_512bit); 4979 Assembler::vpsllw(dst, dst, nds, vector_len); 4980 } else if ((shift_enc < 16) && (nds_enc < 16)) { 4981 // use nds as scratch with dst 4982 evmovdqul(nds, dst, Assembler::AVX_512bit); 4983 Assembler::vpsllw(nds, nds, shift, vector_len); 4984 evmovdqul(dst, nds, Assembler::AVX_512bit); 4985 } else if (dst_enc < 16) { 4986 // use nds to save a copy of xmm0 and hold shift 4987 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4988 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4989 Assembler::vpsllw(dst, dst, xmm0, vector_len); 4990 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4991 } else if (nds_enc < 16) { 4992 // use nds as dest as temps 4993 evmovdqul(nds, dst, Assembler::AVX_512bit); 4994 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4995 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4996 Assembler::vpsllw(nds, nds, xmm0, vector_len); 4997 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4998 evmovdqul(dst, nds, Assembler::AVX_512bit); 4999 } else { 5000 // worse case scenario, all regs are in the upper bank 5001 subptr(rsp, 64); 5002 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 5003 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 5004 evmovdqul(xmm1, shift, Assembler::AVX_512bit); 5005 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 5006 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len); 5007 evmovdqul(xmm1, dst, Assembler::AVX_512bit); 5008 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 5009 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 5010 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 5011 addptr(rsp, 64); 5012 } 5013} 5014 5015void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 5016 int dst_enc = dst->encoding(); 5017 int nds_enc = nds->encoding(); 5018 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 5019 Assembler::vpsllw(dst, nds, shift, vector_len); 5020 } else if (dst_enc < 16) { 5021 Assembler::vpsllw(dst, dst, shift, vector_len); 5022 } else if (nds_enc < 16) { 5023 // use nds as scratch 5024 evmovdqul(nds, dst, Assembler::AVX_512bit); 5025 Assembler::vpsllw(nds, nds, shift, vector_len); 5026 evmovdqul(dst, nds, Assembler::AVX_512bit); 5027 } else { 5028 // use nds as scratch for xmm0 5029 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 5030 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 5031 Assembler::vpsllw(xmm0, xmm0, shift, vector_len); 5032 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 5033 } 5034} 5035 5036void MacroAssembler::vptest(XMMRegister dst, XMMRegister src) { 5037 int dst_enc = dst->encoding(); 5038 int src_enc = src->encoding(); 5039 if ((dst_enc < 16) && (src_enc < 16)) { 5040 Assembler::vptest(dst, src); 5041 } else if (src_enc < 16) { 5042 subptr(rsp, 64); 5043 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5044 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 5045 Assembler::vptest(xmm0, src); 5046 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5047 addptr(rsp, 64); 5048 } else if (dst_enc < 16) { 5049 subptr(rsp, 64); 5050 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5051 evmovdqul(xmm0, src, Assembler::AVX_512bit); 5052 Assembler::vptest(dst, xmm0); 5053 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5054 addptr(rsp, 64); 5055 } else { 5056 subptr(rsp, 64); 5057 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5058 subptr(rsp, 64); 5059 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 5060 movdqu(xmm0, src); 5061 movdqu(xmm1, dst); 5062 Assembler::vptest(xmm1, xmm0); 5063 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 5064 addptr(rsp, 64); 5065 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5066 addptr(rsp, 64); 5067 } 5068} 5069 5070// This instruction exists within macros, ergo we cannot control its input 5071// when emitted through those patterns. 5072void MacroAssembler::punpcklbw(XMMRegister dst, XMMRegister src) { 5073 if (VM_Version::supports_avx512nobw()) { 5074 int dst_enc = dst->encoding(); 5075 int src_enc = src->encoding(); 5076 if (dst_enc == src_enc) { 5077 if (dst_enc < 16) { 5078 Assembler::punpcklbw(dst, src); 5079 } else { 5080 subptr(rsp, 64); 5081 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5082 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 5083 Assembler::punpcklbw(xmm0, xmm0); 5084 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 5085 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5086 addptr(rsp, 64); 5087 } 5088 } else { 5089 if ((src_enc < 16) && (dst_enc < 16)) { 5090 Assembler::punpcklbw(dst, src); 5091 } else if (src_enc < 16) { 5092 subptr(rsp, 64); 5093 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5094 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 5095 Assembler::punpcklbw(xmm0, src); 5096 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 5097 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5098 addptr(rsp, 64); 5099 } else if (dst_enc < 16) { 5100 subptr(rsp, 64); 5101 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5102 evmovdqul(xmm0, src, Assembler::AVX_512bit); 5103 Assembler::punpcklbw(dst, xmm0); 5104 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5105 addptr(rsp, 64); 5106 } else { 5107 subptr(rsp, 64); 5108 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5109 subptr(rsp, 64); 5110 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 5111 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 5112 evmovdqul(xmm1, src, Assembler::AVX_512bit); 5113 Assembler::punpcklbw(xmm0, xmm1); 5114 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 5115 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 5116 addptr(rsp, 64); 5117 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5118 addptr(rsp, 64); 5119 } 5120 } 5121 } else { 5122 Assembler::punpcklbw(dst, src); 5123 } 5124} 5125 5126// This instruction exists within macros, ergo we cannot control its input 5127// when emitted through those patterns. 5128void MacroAssembler::pshuflw(XMMRegister dst, XMMRegister src, int mode) { 5129 if (VM_Version::supports_avx512nobw()) { 5130 int dst_enc = dst->encoding(); 5131 int src_enc = src->encoding(); 5132 if (dst_enc == src_enc) { 5133 if (dst_enc < 16) { 5134 Assembler::pshuflw(dst, src, mode); 5135 } else { 5136 subptr(rsp, 64); 5137 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5138 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 5139 Assembler::pshuflw(xmm0, xmm0, mode); 5140 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 5141 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5142 addptr(rsp, 64); 5143 } 5144 } else { 5145 if ((src_enc < 16) && (dst_enc < 16)) { 5146 Assembler::pshuflw(dst, src, mode); 5147 } else if (src_enc < 16) { 5148 subptr(rsp, 64); 5149 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5150 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 5151 Assembler::pshuflw(xmm0, src, mode); 5152 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 5153 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5154 addptr(rsp, 64); 5155 } else if (dst_enc < 16) { 5156 subptr(rsp, 64); 5157 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5158 evmovdqul(xmm0, src, Assembler::AVX_512bit); 5159 Assembler::pshuflw(dst, xmm0, mode); 5160 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5161 addptr(rsp, 64); 5162 } else { 5163 subptr(rsp, 64); 5164 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5165 subptr(rsp, 64); 5166 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 5167 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 5168 evmovdqul(xmm1, src, Assembler::AVX_512bit); 5169 Assembler::pshuflw(xmm0, xmm1, mode); 5170 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 5171 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 5172 addptr(rsp, 64); 5173 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5174 addptr(rsp, 64); 5175 } 5176 } 5177 } else { 5178 Assembler::pshuflw(dst, src, mode); 5179 } 5180} 5181 5182void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 5183 if (reachable(src)) { 5184 vandpd(dst, nds, as_Address(src), vector_len); 5185 } else { 5186 lea(rscratch1, src); 5187 vandpd(dst, nds, Address(rscratch1, 0), vector_len); 5188 } 5189} 5190 5191void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 5192 if (reachable(src)) { 5193 vandps(dst, nds, as_Address(src), vector_len); 5194 } else { 5195 lea(rscratch1, src); 5196 vandps(dst, nds, Address(rscratch1, 0), vector_len); 5197 } 5198} 5199 5200void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5201 if (reachable(src)) { 5202 vdivsd(dst, nds, as_Address(src)); 5203 } else { 5204 lea(rscratch1, src); 5205 vdivsd(dst, nds, Address(rscratch1, 0)); 5206 } 5207} 5208 5209void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5210 if (reachable(src)) { 5211 vdivss(dst, nds, as_Address(src)); 5212 } else { 5213 lea(rscratch1, src); 5214 vdivss(dst, nds, Address(rscratch1, 0)); 5215 } 5216} 5217 5218void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5219 if (reachable(src)) { 5220 vmulsd(dst, nds, as_Address(src)); 5221 } else { 5222 lea(rscratch1, src); 5223 vmulsd(dst, nds, Address(rscratch1, 0)); 5224 } 5225} 5226 5227void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5228 if (reachable(src)) { 5229 vmulss(dst, nds, as_Address(src)); 5230 } else { 5231 lea(rscratch1, src); 5232 vmulss(dst, nds, Address(rscratch1, 0)); 5233 } 5234} 5235 5236void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5237 if (reachable(src)) { 5238 vsubsd(dst, nds, as_Address(src)); 5239 } else { 5240 lea(rscratch1, src); 5241 vsubsd(dst, nds, Address(rscratch1, 0)); 5242 } 5243} 5244 5245void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5246 if (reachable(src)) { 5247 vsubss(dst, nds, as_Address(src)); 5248 } else { 5249 lea(rscratch1, src); 5250 vsubss(dst, nds, Address(rscratch1, 0)); 5251 } 5252} 5253 5254void MacroAssembler::vnegatess(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5255 int nds_enc = nds->encoding(); 5256 int dst_enc = dst->encoding(); 5257 bool dst_upper_bank = (dst_enc > 15); 5258 bool nds_upper_bank = (nds_enc > 15); 5259 if (VM_Version::supports_avx512novl() && 5260 (nds_upper_bank || dst_upper_bank)) { 5261 if (dst_upper_bank) { 5262 subptr(rsp, 64); 5263 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5264 movflt(xmm0, nds); 5265 vxorps(xmm0, xmm0, src, Assembler::AVX_128bit); 5266 movflt(dst, xmm0); 5267 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5268 addptr(rsp, 64); 5269 } else { 5270 movflt(dst, nds); 5271 vxorps(dst, dst, src, Assembler::AVX_128bit); 5272 } 5273 } else { 5274 vxorps(dst, nds, src, Assembler::AVX_128bit); 5275 } 5276} 5277 5278void MacroAssembler::vnegatesd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5279 int nds_enc = nds->encoding(); 5280 int dst_enc = dst->encoding(); 5281 bool dst_upper_bank = (dst_enc > 15); 5282 bool nds_upper_bank = (nds_enc > 15); 5283 if (VM_Version::supports_avx512novl() && 5284 (nds_upper_bank || dst_upper_bank)) { 5285 if (dst_upper_bank) { 5286 subptr(rsp, 64); 5287 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5288 movdbl(xmm0, nds); 5289 vxorpd(xmm0, xmm0, src, Assembler::AVX_128bit); 5290 movdbl(dst, xmm0); 5291 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5292 addptr(rsp, 64); 5293 } else { 5294 movdbl(dst, nds); 5295 vxorpd(dst, dst, src, Assembler::AVX_128bit); 5296 } 5297 } else { 5298 vxorpd(dst, nds, src, Assembler::AVX_128bit); 5299 } 5300} 5301 5302void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 5303 if (reachable(src)) { 5304 vxorpd(dst, nds, as_Address(src), vector_len); 5305 } else { 5306 lea(rscratch1, src); 5307 vxorpd(dst, nds, Address(rscratch1, 0), vector_len); 5308 } 5309} 5310 5311void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 5312 if (reachable(src)) { 5313 vxorps(dst, nds, as_Address(src), vector_len); 5314 } else { 5315 lea(rscratch1, src); 5316 vxorps(dst, nds, Address(rscratch1, 0), vector_len); 5317 } 5318} 5319 5320 5321////////////////////////////////////////////////////////////////////////////////// 5322#if INCLUDE_ALL_GCS 5323 5324void MacroAssembler::g1_write_barrier_pre(Register obj, 5325 Register pre_val, 5326 Register thread, 5327 Register tmp, 5328 bool tosca_live, 5329 bool expand_call) { 5330 5331 // If expand_call is true then we expand the call_VM_leaf macro 5332 // directly to skip generating the check by 5333 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp. 5334 5335#ifdef _LP64 5336 assert(thread == r15_thread, "must be"); 5337#endif // _LP64 5338 5339 Label done; 5340 Label runtime; 5341 5342 assert(pre_val != noreg, "check this code"); 5343 5344 if (obj != noreg) { 5345 assert_different_registers(obj, pre_val, tmp); 5346 assert(pre_val != rax, "check this code"); 5347 } 5348 5349 Address in_progress(thread, in_bytes(JavaThread::satb_mark_queue_offset() + 5350 SATBMarkQueue::byte_offset_of_active())); 5351 Address index(thread, in_bytes(JavaThread::satb_mark_queue_offset() + 5352 SATBMarkQueue::byte_offset_of_index())); 5353 Address buffer(thread, in_bytes(JavaThread::satb_mark_queue_offset() + 5354 SATBMarkQueue::byte_offset_of_buf())); 5355 5356 5357 // Is marking active? 5358 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) { 5359 cmpl(in_progress, 0); 5360 } else { 5361 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption"); 5362 cmpb(in_progress, 0); 5363 } 5364 jcc(Assembler::equal, done); 5365 5366 // Do we need to load the previous value? 5367 if (obj != noreg) { 5368 load_heap_oop(pre_val, Address(obj, 0)); 5369 } 5370 5371 // Is the previous value null? 5372 cmpptr(pre_val, (int32_t) NULL_WORD); 5373 jcc(Assembler::equal, done); 5374 5375 // Can we store original value in the thread's buffer? 5376 // Is index == 0? 5377 // (The index field is typed as size_t.) 5378 5379 movptr(tmp, index); // tmp := *index_adr 5380 cmpptr(tmp, 0); // tmp == 0? 5381 jcc(Assembler::equal, runtime); // If yes, goto runtime 5382 5383 subptr(tmp, wordSize); // tmp := tmp - wordSize 5384 movptr(index, tmp); // *index_adr := tmp 5385 addptr(tmp, buffer); // tmp := tmp + *buffer_adr 5386 5387 // Record the previous value 5388 movptr(Address(tmp, 0), pre_val); 5389 jmp(done); 5390 5391 bind(runtime); 5392 // save the live input values 5393 if(tosca_live) push(rax); 5394 5395 if (obj != noreg && obj != rax) 5396 push(obj); 5397 5398 if (pre_val != rax) 5399 push(pre_val); 5400 5401 // Calling the runtime using the regular call_VM_leaf mechanism generates 5402 // code (generated by InterpreterMacroAssember::call_VM_leaf_base) 5403 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL. 5404 // 5405 // If we care generating the pre-barrier without a frame (e.g. in the 5406 // intrinsified Reference.get() routine) then ebp might be pointing to 5407 // the caller frame and so this check will most likely fail at runtime. 5408 // 5409 // Expanding the call directly bypasses the generation of the check. 5410 // So when we do not have have a full interpreter frame on the stack 5411 // expand_call should be passed true. 5412 5413 NOT_LP64( push(thread); ) 5414 5415 if (expand_call) { 5416 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); ) 5417 pass_arg1(this, thread); 5418 pass_arg0(this, pre_val); 5419 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), 2); 5420 } else { 5421 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), pre_val, thread); 5422 } 5423 5424 NOT_LP64( pop(thread); ) 5425 5426 // save the live input values 5427 if (pre_val != rax) 5428 pop(pre_val); 5429 5430 if (obj != noreg && obj != rax) 5431 pop(obj); 5432 5433 if(tosca_live) pop(rax); 5434 5435 bind(done); 5436} 5437 5438void MacroAssembler::g1_write_barrier_post(Register store_addr, 5439 Register new_val, 5440 Register thread, 5441 Register tmp, 5442 Register tmp2) { 5443#ifdef _LP64 5444 assert(thread == r15_thread, "must be"); 5445#endif // _LP64 5446 5447 Address queue_index(thread, in_bytes(JavaThread::dirty_card_queue_offset() + 5448 DirtyCardQueue::byte_offset_of_index())); 5449 Address buffer(thread, in_bytes(JavaThread::dirty_card_queue_offset() + 5450 DirtyCardQueue::byte_offset_of_buf())); 5451 5452 CardTableModRefBS* ct = 5453 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 5454 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code"); 5455 5456 Label done; 5457 Label runtime; 5458 5459 // Does store cross heap regions? 5460 5461 movptr(tmp, store_addr); 5462 xorptr(tmp, new_val); 5463 shrptr(tmp, HeapRegion::LogOfHRGrainBytes); 5464 jcc(Assembler::equal, done); 5465 5466 // crosses regions, storing NULL? 5467 5468 cmpptr(new_val, (int32_t) NULL_WORD); 5469 jcc(Assembler::equal, done); 5470 5471 // storing region crossing non-NULL, is card already dirty? 5472 5473 const Register card_addr = tmp; 5474 const Register cardtable = tmp2; 5475 5476 movptr(card_addr, store_addr); 5477 shrptr(card_addr, CardTableModRefBS::card_shift); 5478 // Do not use ExternalAddress to load 'byte_map_base', since 'byte_map_base' is NOT 5479 // a valid address and therefore is not properly handled by the relocation code. 5480 movptr(cardtable, (intptr_t)ct->byte_map_base); 5481 addptr(card_addr, cardtable); 5482 5483 cmpb(Address(card_addr, 0), (int)G1SATBCardTableModRefBS::g1_young_card_val()); 5484 jcc(Assembler::equal, done); 5485 5486 membar(Assembler::Membar_mask_bits(Assembler::StoreLoad)); 5487 cmpb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val()); 5488 jcc(Assembler::equal, done); 5489 5490 5491 // storing a region crossing, non-NULL oop, card is clean. 5492 // dirty card and log. 5493 5494 movb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val()); 5495 5496 cmpl(queue_index, 0); 5497 jcc(Assembler::equal, runtime); 5498 subl(queue_index, wordSize); 5499 movptr(tmp2, buffer); 5500#ifdef _LP64 5501 movslq(rscratch1, queue_index); 5502 addq(tmp2, rscratch1); 5503 movq(Address(tmp2, 0), card_addr); 5504#else 5505 addl(tmp2, queue_index); 5506 movl(Address(tmp2, 0), card_addr); 5507#endif 5508 jmp(done); 5509 5510 bind(runtime); 5511 // save the live input values 5512 push(store_addr); 5513 push(new_val); 5514#ifdef _LP64 5515 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, r15_thread); 5516#else 5517 push(thread); 5518 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, thread); 5519 pop(thread); 5520#endif 5521 pop(new_val); 5522 pop(store_addr); 5523 5524 bind(done); 5525} 5526 5527#endif // INCLUDE_ALL_GCS 5528////////////////////////////////////////////////////////////////////////////////// 5529 5530 5531void MacroAssembler::store_check(Register obj, Address dst) { 5532 store_check(obj); 5533} 5534 5535void MacroAssembler::store_check(Register obj) { 5536 // Does a store check for the oop in register obj. The content of 5537 // register obj is destroyed afterwards. 5538 BarrierSet* bs = Universe::heap()->barrier_set(); 5539 assert(bs->kind() == BarrierSet::CardTableForRS || 5540 bs->kind() == BarrierSet::CardTableExtension, 5541 "Wrong barrier set kind"); 5542 5543 CardTableModRefBS* ct = barrier_set_cast<CardTableModRefBS>(bs); 5544 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code"); 5545 5546 shrptr(obj, CardTableModRefBS::card_shift); 5547 5548 Address card_addr; 5549 5550 // The calculation for byte_map_base is as follows: 5551 // byte_map_base = _byte_map - (uintptr_t(low_bound) >> card_shift); 5552 // So this essentially converts an address to a displacement and it will 5553 // never need to be relocated. On 64bit however the value may be too 5554 // large for a 32bit displacement. 5555 intptr_t disp = (intptr_t) ct->byte_map_base; 5556 if (is_simm32(disp)) { 5557 card_addr = Address(noreg, obj, Address::times_1, disp); 5558 } else { 5559 // By doing it as an ExternalAddress 'disp' could be converted to a rip-relative 5560 // displacement and done in a single instruction given favorable mapping and a 5561 // smarter version of as_Address. However, 'ExternalAddress' generates a relocation 5562 // entry and that entry is not properly handled by the relocation code. 5563 AddressLiteral cardtable((address)ct->byte_map_base, relocInfo::none); 5564 Address index(noreg, obj, Address::times_1); 5565 card_addr = as_Address(ArrayAddress(cardtable, index)); 5566 } 5567 5568 int dirty = CardTableModRefBS::dirty_card_val(); 5569 if (UseCondCardMark) { 5570 Label L_already_dirty; 5571 if (UseConcMarkSweepGC) { 5572 membar(Assembler::StoreLoad); 5573 } 5574 cmpb(card_addr, dirty); 5575 jcc(Assembler::equal, L_already_dirty); 5576 movb(card_addr, dirty); 5577 bind(L_already_dirty); 5578 } else { 5579 movb(card_addr, dirty); 5580 } 5581} 5582 5583void MacroAssembler::subptr(Register dst, int32_t imm32) { 5584 LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32)); 5585} 5586 5587// Force generation of a 4 byte immediate value even if it fits into 8bit 5588void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) { 5589 LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32)); 5590} 5591 5592void MacroAssembler::subptr(Register dst, Register src) { 5593 LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src)); 5594} 5595 5596// C++ bool manipulation 5597void MacroAssembler::testbool(Register dst) { 5598 if(sizeof(bool) == 1) 5599 testb(dst, 0xff); 5600 else if(sizeof(bool) == 2) { 5601 // testw implementation needed for two byte bools 5602 ShouldNotReachHere(); 5603 } else if(sizeof(bool) == 4) 5604 testl(dst, dst); 5605 else 5606 // unsupported 5607 ShouldNotReachHere(); 5608} 5609 5610void MacroAssembler::testptr(Register dst, Register src) { 5611 LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src)); 5612} 5613 5614// Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes. 5615void MacroAssembler::tlab_allocate(Register obj, 5616 Register var_size_in_bytes, 5617 int con_size_in_bytes, 5618 Register t1, 5619 Register t2, 5620 Label& slow_case) { 5621 assert_different_registers(obj, t1, t2); 5622 assert_different_registers(obj, var_size_in_bytes, t1); 5623 Register end = t2; 5624 Register thread = NOT_LP64(t1) LP64_ONLY(r15_thread); 5625 5626 verify_tlab(); 5627 5628 NOT_LP64(get_thread(thread)); 5629 5630 movptr(obj, Address(thread, JavaThread::tlab_top_offset())); 5631 if (var_size_in_bytes == noreg) { 5632 lea(end, Address(obj, con_size_in_bytes)); 5633 } else { 5634 lea(end, Address(obj, var_size_in_bytes, Address::times_1)); 5635 } 5636 cmpptr(end, Address(thread, JavaThread::tlab_end_offset())); 5637 jcc(Assembler::above, slow_case); 5638 5639 // update the tlab top pointer 5640 movptr(Address(thread, JavaThread::tlab_top_offset()), end); 5641 5642 // recover var_size_in_bytes if necessary 5643 if (var_size_in_bytes == end) { 5644 subptr(var_size_in_bytes, obj); 5645 } 5646 verify_tlab(); 5647} 5648 5649// Preserves rbx, and rdx. 5650Register MacroAssembler::tlab_refill(Label& retry, 5651 Label& try_eden, 5652 Label& slow_case) { 5653 Register top = rax; 5654 Register t1 = rcx; 5655 Register t2 = rsi; 5656 Register thread_reg = NOT_LP64(rdi) LP64_ONLY(r15_thread); 5657 assert_different_registers(top, thread_reg, t1, t2, /* preserve: */ rbx, rdx); 5658 Label do_refill, discard_tlab; 5659 5660 if (!Universe::heap()->supports_inline_contig_alloc()) { 5661 // No allocation in the shared eden. 5662 jmp(slow_case); 5663 } 5664 5665 NOT_LP64(get_thread(thread_reg)); 5666 5667 movptr(top, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 5668 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset()))); 5669 5670 // calculate amount of free space 5671 subptr(t1, top); 5672 shrptr(t1, LogHeapWordSize); 5673 5674 // Retain tlab and allocate object in shared space if 5675 // the amount free in the tlab is too large to discard. 5676 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset()))); 5677 jcc(Assembler::lessEqual, discard_tlab); 5678 5679 // Retain 5680 // %%% yuck as movptr... 5681 movptr(t2, (int32_t) ThreadLocalAllocBuffer::refill_waste_limit_increment()); 5682 addptr(Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset())), t2); 5683 if (TLABStats) { 5684 // increment number of slow_allocations 5685 addl(Address(thread_reg, in_bytes(JavaThread::tlab_slow_allocations_offset())), 1); 5686 } 5687 jmp(try_eden); 5688 5689 bind(discard_tlab); 5690 if (TLABStats) { 5691 // increment number of refills 5692 addl(Address(thread_reg, in_bytes(JavaThread::tlab_number_of_refills_offset())), 1); 5693 // accumulate wastage -- t1 is amount free in tlab 5694 addl(Address(thread_reg, in_bytes(JavaThread::tlab_fast_refill_waste_offset())), t1); 5695 } 5696 5697 // if tlab is currently allocated (top or end != null) then 5698 // fill [top, end + alignment_reserve) with array object 5699 testptr(top, top); 5700 jcc(Assembler::zero, do_refill); 5701 5702 // set up the mark word 5703 movptr(Address(top, oopDesc::mark_offset_in_bytes()), (intptr_t)markOopDesc::prototype()->copy_set_hash(0x2)); 5704 // set the length to the remaining space 5705 subptr(t1, typeArrayOopDesc::header_size(T_INT)); 5706 addptr(t1, (int32_t)ThreadLocalAllocBuffer::alignment_reserve()); 5707 shlptr(t1, log2_intptr(HeapWordSize/sizeof(jint))); 5708 movl(Address(top, arrayOopDesc::length_offset_in_bytes()), t1); 5709 // set klass to intArrayKlass 5710 // dubious reloc why not an oop reloc? 5711 movptr(t1, ExternalAddress((address)Universe::intArrayKlassObj_addr())); 5712 // store klass last. concurrent gcs assumes klass length is valid if 5713 // klass field is not null. 5714 store_klass(top, t1); 5715 5716 movptr(t1, top); 5717 subptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset()))); 5718 incr_allocated_bytes(thread_reg, t1, 0); 5719 5720 // refill the tlab with an eden allocation 5721 bind(do_refill); 5722 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset()))); 5723 shlptr(t1, LogHeapWordSize); 5724 // allocate new tlab, address returned in top 5725 eden_allocate(top, t1, 0, t2, slow_case); 5726 5727 // Check that t1 was preserved in eden_allocate. 5728#ifdef ASSERT 5729 if (UseTLAB) { 5730 Label ok; 5731 Register tsize = rsi; 5732 assert_different_registers(tsize, thread_reg, t1); 5733 push(tsize); 5734 movptr(tsize, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset()))); 5735 shlptr(tsize, LogHeapWordSize); 5736 cmpptr(t1, tsize); 5737 jcc(Assembler::equal, ok); 5738 STOP("assert(t1 != tlab size)"); 5739 should_not_reach_here(); 5740 5741 bind(ok); 5742 pop(tsize); 5743 } 5744#endif 5745 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())), top); 5746 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())), top); 5747 addptr(top, t1); 5748 subptr(top, (int32_t)ThreadLocalAllocBuffer::alignment_reserve_in_bytes()); 5749 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())), top); 5750 verify_tlab(); 5751 jmp(retry); 5752 5753 return thread_reg; // for use by caller 5754} 5755 5756void MacroAssembler::incr_allocated_bytes(Register thread, 5757 Register var_size_in_bytes, 5758 int con_size_in_bytes, 5759 Register t1) { 5760 if (!thread->is_valid()) { 5761#ifdef _LP64 5762 thread = r15_thread; 5763#else 5764 assert(t1->is_valid(), "need temp reg"); 5765 thread = t1; 5766 get_thread(thread); 5767#endif 5768 } 5769 5770#ifdef _LP64 5771 if (var_size_in_bytes->is_valid()) { 5772 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes); 5773 } else { 5774 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes); 5775 } 5776#else 5777 if (var_size_in_bytes->is_valid()) { 5778 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes); 5779 } else { 5780 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes); 5781 } 5782 adcl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())+4), 0); 5783#endif 5784} 5785 5786void MacroAssembler::fp_runtime_fallback(address runtime_entry, int nb_args, int num_fpu_regs_in_use) { 5787 pusha(); 5788 5789 // if we are coming from c1, xmm registers may be live 5790 int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8); 5791 if (UseAVX > 2) { 5792 num_xmm_regs = LP64_ONLY(32) NOT_LP64(8); 5793 } 5794 5795 if (UseSSE == 1) { 5796 subptr(rsp, sizeof(jdouble)*8); 5797 for (int n = 0; n < 8; n++) { 5798 movflt(Address(rsp, n*sizeof(jdouble)), as_XMMRegister(n)); 5799 } 5800 } else if (UseSSE >= 2) { 5801 if (UseAVX > 2) { 5802 push(rbx); 5803 movl(rbx, 0xffff); 5804 kmovwl(k1, rbx); 5805 pop(rbx); 5806 } 5807#ifdef COMPILER2 5808 if (MaxVectorSize > 16) { 5809 if(UseAVX > 2) { 5810 // Save upper half of ZMM registers 5811 subptr(rsp, 32*num_xmm_regs); 5812 for (int n = 0; n < num_xmm_regs; n++) { 5813 vextractf64x4h(Address(rsp, n*32), as_XMMRegister(n), 1); 5814 } 5815 } 5816 assert(UseAVX > 0, "256 bit vectors are supported only with AVX"); 5817 // Save upper half of YMM registers 5818 subptr(rsp, 16*num_xmm_regs); 5819 for (int n = 0; n < num_xmm_regs; n++) { 5820 vextractf128h(Address(rsp, n*16), as_XMMRegister(n)); 5821 } 5822 } 5823#endif 5824 // Save whole 128bit (16 bytes) XMM registers 5825 subptr(rsp, 16*num_xmm_regs); 5826#ifdef _LP64 5827 if (VM_Version::supports_evex()) { 5828 for (int n = 0; n < num_xmm_regs; n++) { 5829 vextractf32x4h(Address(rsp, n*16), as_XMMRegister(n), 0); 5830 } 5831 } else { 5832 for (int n = 0; n < num_xmm_regs; n++) { 5833 movdqu(Address(rsp, n*16), as_XMMRegister(n)); 5834 } 5835 } 5836#else 5837 for (int n = 0; n < num_xmm_regs; n++) { 5838 movdqu(Address(rsp, n*16), as_XMMRegister(n)); 5839 } 5840#endif 5841 } 5842 5843 // Preserve registers across runtime call 5844 int incoming_argument_and_return_value_offset = -1; 5845 if (num_fpu_regs_in_use > 1) { 5846 // Must preserve all other FPU regs (could alternatively convert 5847 // SharedRuntime::dsin, dcos etc. into assembly routines known not to trash 5848 // FPU state, but can not trust C compiler) 5849 NEEDS_CLEANUP; 5850 // NOTE that in this case we also push the incoming argument(s) to 5851 // the stack and restore it later; we also use this stack slot to 5852 // hold the return value from dsin, dcos etc. 5853 for (int i = 0; i < num_fpu_regs_in_use; i++) { 5854 subptr(rsp, sizeof(jdouble)); 5855 fstp_d(Address(rsp, 0)); 5856 } 5857 incoming_argument_and_return_value_offset = sizeof(jdouble)*(num_fpu_regs_in_use-1); 5858 for (int i = nb_args-1; i >= 0; i--) { 5859 fld_d(Address(rsp, incoming_argument_and_return_value_offset-i*sizeof(jdouble))); 5860 } 5861 } 5862 5863 subptr(rsp, nb_args*sizeof(jdouble)); 5864 for (int i = 0; i < nb_args; i++) { 5865 fstp_d(Address(rsp, i*sizeof(jdouble))); 5866 } 5867 5868#ifdef _LP64 5869 if (nb_args > 0) { 5870 movdbl(xmm0, Address(rsp, 0)); 5871 } 5872 if (nb_args > 1) { 5873 movdbl(xmm1, Address(rsp, sizeof(jdouble))); 5874 } 5875 assert(nb_args <= 2, "unsupported number of args"); 5876#endif // _LP64 5877 5878 // NOTE: we must not use call_VM_leaf here because that requires a 5879 // complete interpreter frame in debug mode -- same bug as 4387334 5880 // MacroAssembler::call_VM_leaf_base is perfectly safe and will 5881 // do proper 64bit abi 5882 5883 NEEDS_CLEANUP; 5884 // Need to add stack banging before this runtime call if it needs to 5885 // be taken; however, there is no generic stack banging routine at 5886 // the MacroAssembler level 5887 5888 MacroAssembler::call_VM_leaf_base(runtime_entry, 0); 5889 5890#ifdef _LP64 5891 movsd(Address(rsp, 0), xmm0); 5892 fld_d(Address(rsp, 0)); 5893#endif // _LP64 5894 addptr(rsp, sizeof(jdouble)*nb_args); 5895 if (num_fpu_regs_in_use > 1) { 5896 // Must save return value to stack and then restore entire FPU 5897 // stack except incoming arguments 5898 fstp_d(Address(rsp, incoming_argument_and_return_value_offset)); 5899 for (int i = 0; i < num_fpu_regs_in_use - nb_args; i++) { 5900 fld_d(Address(rsp, 0)); 5901 addptr(rsp, sizeof(jdouble)); 5902 } 5903 fld_d(Address(rsp, (nb_args-1)*sizeof(jdouble))); 5904 addptr(rsp, sizeof(jdouble)*nb_args); 5905 } 5906 5907 if (UseSSE == 1) { 5908 for (int n = 0; n < 8; n++) { 5909 movflt(as_XMMRegister(n), Address(rsp, n*sizeof(jdouble))); 5910 } 5911 addptr(rsp, sizeof(jdouble)*8); 5912 } else if (UseSSE >= 2) { 5913 // Restore whole 128bit (16 bytes) XMM registers 5914#ifdef _LP64 5915 if (VM_Version::supports_evex()) { 5916 for (int n = 0; n < num_xmm_regs; n++) { 5917 vinsertf32x4h(as_XMMRegister(n), Address(rsp, n*16), 0); 5918 } 5919 } else { 5920 for (int n = 0; n < num_xmm_regs; n++) { 5921 movdqu(as_XMMRegister(n), Address(rsp, n*16)); 5922 } 5923 } 5924#else 5925 for (int n = 0; n < num_xmm_regs; n++) { 5926 movdqu(as_XMMRegister(n), Address(rsp, n*16)); 5927 } 5928#endif 5929 addptr(rsp, 16*num_xmm_regs); 5930 5931#ifdef COMPILER2 5932 if (MaxVectorSize > 16) { 5933 // Restore upper half of YMM registers. 5934 for (int n = 0; n < num_xmm_regs; n++) { 5935 vinsertf128h(as_XMMRegister(n), Address(rsp, n*16)); 5936 } 5937 addptr(rsp, 16*num_xmm_regs); 5938 if(UseAVX > 2) { 5939 for (int n = 0; n < num_xmm_regs; n++) { 5940 vinsertf64x4h(as_XMMRegister(n), Address(rsp, n*32), 1); 5941 } 5942 addptr(rsp, 32*num_xmm_regs); 5943 } 5944 } 5945#endif 5946 } 5947 popa(); 5948} 5949 5950static const double pi_4 = 0.7853981633974483; 5951 5952void MacroAssembler::trigfunc(char trig, int num_fpu_regs_in_use) { 5953 // A hand-coded argument reduction for values in fabs(pi/4, pi/2) 5954 // was attempted in this code; unfortunately it appears that the 5955 // switch to 80-bit precision and back causes this to be 5956 // unprofitable compared with simply performing a runtime call if 5957 // the argument is out of the (-pi/4, pi/4) range. 5958 5959 Register tmp = noreg; 5960 if (!VM_Version::supports_cmov()) { 5961 // fcmp needs a temporary so preserve rbx, 5962 tmp = rbx; 5963 push(tmp); 5964 } 5965 5966 Label slow_case, done; 5967 5968 ExternalAddress pi4_adr = (address)&pi_4; 5969 if (reachable(pi4_adr)) { 5970 // x ?<= pi/4 5971 fld_d(pi4_adr); 5972 fld_s(1); // Stack: X PI/4 X 5973 fabs(); // Stack: |X| PI/4 X 5974 fcmp(tmp); 5975 jcc(Assembler::above, slow_case); 5976 5977 // fastest case: -pi/4 <= x <= pi/4 5978 switch(trig) { 5979 case 's': 5980 fsin(); 5981 break; 5982 case 'c': 5983 fcos(); 5984 break; 5985 case 't': 5986 ftan(); 5987 break; 5988 default: 5989 assert(false, "bad intrinsic"); 5990 break; 5991 } 5992 jmp(done); 5993 } 5994 5995 // slow case: runtime call 5996 bind(slow_case); 5997 5998 switch(trig) { 5999 case 's': 6000 { 6001 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1, num_fpu_regs_in_use); 6002 } 6003 break; 6004 case 'c': 6005 { 6006 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1, num_fpu_regs_in_use); 6007 } 6008 break; 6009 case 't': 6010 { 6011 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1, num_fpu_regs_in_use); 6012 } 6013 break; 6014 default: 6015 assert(false, "bad intrinsic"); 6016 break; 6017 } 6018 6019 // Come here with result in F-TOS 6020 bind(done); 6021 6022 if (tmp != noreg) { 6023 pop(tmp); 6024 } 6025} 6026 6027 6028// Look up the method for a megamorphic invokeinterface call. 6029// The target method is determined by <intf_klass, itable_index>. 6030// The receiver klass is in recv_klass. 6031// On success, the result will be in method_result, and execution falls through. 6032// On failure, execution transfers to the given label. 6033void MacroAssembler::lookup_interface_method(Register recv_klass, 6034 Register intf_klass, 6035 RegisterOrConstant itable_index, 6036 Register method_result, 6037 Register scan_temp, 6038 Label& L_no_such_interface) { 6039 assert_different_registers(recv_klass, intf_klass, method_result, scan_temp); 6040 assert(itable_index.is_constant() || itable_index.as_register() == method_result, 6041 "caller must use same register for non-constant itable index as for method"); 6042 6043 // Compute start of first itableOffsetEntry (which is at the end of the vtable) 6044 int vtable_base = InstanceKlass::vtable_start_offset() * wordSize; 6045 int itentry_off = itableMethodEntry::method_offset_in_bytes(); 6046 int scan_step = itableOffsetEntry::size() * wordSize; 6047 int vte_size = vtableEntry::size() * wordSize; 6048 Address::ScaleFactor times_vte_scale = Address::times_ptr; 6049 assert(vte_size == wordSize, "else adjust times_vte_scale"); 6050 6051 movl(scan_temp, Address(recv_klass, InstanceKlass::vtable_length_offset() * wordSize)); 6052 6053 // %%% Could store the aligned, prescaled offset in the klassoop. 6054 lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base)); 6055 if (HeapWordsPerLong > 1) { 6056 // Round up to align_object_offset boundary 6057 // see code for InstanceKlass::start_of_itable! 6058 round_to(scan_temp, BytesPerLong); 6059 } 6060 6061 // Adjust recv_klass by scaled itable_index, so we can free itable_index. 6062 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below"); 6063 lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off)); 6064 6065 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) { 6066 // if (scan->interface() == intf) { 6067 // result = (klass + scan->offset() + itable_index); 6068 // } 6069 // } 6070 Label search, found_method; 6071 6072 for (int peel = 1; peel >= 0; peel--) { 6073 movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes())); 6074 cmpptr(intf_klass, method_result); 6075 6076 if (peel) { 6077 jccb(Assembler::equal, found_method); 6078 } else { 6079 jccb(Assembler::notEqual, search); 6080 // (invert the test to fall through to found_method...) 6081 } 6082 6083 if (!peel) break; 6084 6085 bind(search); 6086 6087 // Check that the previous entry is non-null. A null entry means that 6088 // the receiver class doesn't implement the interface, and wasn't the 6089 // same as when the caller was compiled. 6090 testptr(method_result, method_result); 6091 jcc(Assembler::zero, L_no_such_interface); 6092 addptr(scan_temp, scan_step); 6093 } 6094 6095 bind(found_method); 6096 6097 // Got a hit. 6098 movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes())); 6099 movptr(method_result, Address(recv_klass, scan_temp, Address::times_1)); 6100} 6101 6102 6103// virtual method calling 6104void MacroAssembler::lookup_virtual_method(Register recv_klass, 6105 RegisterOrConstant vtable_index, 6106 Register method_result) { 6107 const int base = InstanceKlass::vtable_start_offset() * wordSize; 6108 assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below"); 6109 Address vtable_entry_addr(recv_klass, 6110 vtable_index, Address::times_ptr, 6111 base + vtableEntry::method_offset_in_bytes()); 6112 movptr(method_result, vtable_entry_addr); 6113} 6114 6115 6116void MacroAssembler::check_klass_subtype(Register sub_klass, 6117 Register super_klass, 6118 Register temp_reg, 6119 Label& L_success) { 6120 Label L_failure; 6121 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL); 6122 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL); 6123 bind(L_failure); 6124} 6125 6126 6127void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass, 6128 Register super_klass, 6129 Register temp_reg, 6130 Label* L_success, 6131 Label* L_failure, 6132 Label* L_slow_path, 6133 RegisterOrConstant super_check_offset) { 6134 assert_different_registers(sub_klass, super_klass, temp_reg); 6135 bool must_load_sco = (super_check_offset.constant_or_zero() == -1); 6136 if (super_check_offset.is_register()) { 6137 assert_different_registers(sub_klass, super_klass, 6138 super_check_offset.as_register()); 6139 } else if (must_load_sco) { 6140 assert(temp_reg != noreg, "supply either a temp or a register offset"); 6141 } 6142 6143 Label L_fallthrough; 6144 int label_nulls = 0; 6145 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 6146 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 6147 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; } 6148 assert(label_nulls <= 1, "at most one NULL in the batch"); 6149 6150 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 6151 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 6152 Address super_check_offset_addr(super_klass, sco_offset); 6153 6154 // Hacked jcc, which "knows" that L_fallthrough, at least, is in 6155 // range of a jccb. If this routine grows larger, reconsider at 6156 // least some of these. 6157#define local_jcc(assembler_cond, label) \ 6158 if (&(label) == &L_fallthrough) jccb(assembler_cond, label); \ 6159 else jcc( assembler_cond, label) /*omit semi*/ 6160 6161 // Hacked jmp, which may only be used just before L_fallthrough. 6162#define final_jmp(label) \ 6163 if (&(label) == &L_fallthrough) { /*do nothing*/ } \ 6164 else jmp(label) /*omit semi*/ 6165 6166 // If the pointers are equal, we are done (e.g., String[] elements). 6167 // This self-check enables sharing of secondary supertype arrays among 6168 // non-primary types such as array-of-interface. Otherwise, each such 6169 // type would need its own customized SSA. 6170 // We move this check to the front of the fast path because many 6171 // type checks are in fact trivially successful in this manner, 6172 // so we get a nicely predicted branch right at the start of the check. 6173 cmpptr(sub_klass, super_klass); 6174 local_jcc(Assembler::equal, *L_success); 6175 6176 // Check the supertype display: 6177 if (must_load_sco) { 6178 // Positive movl does right thing on LP64. 6179 movl(temp_reg, super_check_offset_addr); 6180 super_check_offset = RegisterOrConstant(temp_reg); 6181 } 6182 Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0); 6183 cmpptr(super_klass, super_check_addr); // load displayed supertype 6184 6185 // This check has worked decisively for primary supers. 6186 // Secondary supers are sought in the super_cache ('super_cache_addr'). 6187 // (Secondary supers are interfaces and very deeply nested subtypes.) 6188 // This works in the same check above because of a tricky aliasing 6189 // between the super_cache and the primary super display elements. 6190 // (The 'super_check_addr' can address either, as the case requires.) 6191 // Note that the cache is updated below if it does not help us find 6192 // what we need immediately. 6193 // So if it was a primary super, we can just fail immediately. 6194 // Otherwise, it's the slow path for us (no success at this point). 6195 6196 if (super_check_offset.is_register()) { 6197 local_jcc(Assembler::equal, *L_success); 6198 cmpl(super_check_offset.as_register(), sc_offset); 6199 if (L_failure == &L_fallthrough) { 6200 local_jcc(Assembler::equal, *L_slow_path); 6201 } else { 6202 local_jcc(Assembler::notEqual, *L_failure); 6203 final_jmp(*L_slow_path); 6204 } 6205 } else if (super_check_offset.as_constant() == sc_offset) { 6206 // Need a slow path; fast failure is impossible. 6207 if (L_slow_path == &L_fallthrough) { 6208 local_jcc(Assembler::equal, *L_success); 6209 } else { 6210 local_jcc(Assembler::notEqual, *L_slow_path); 6211 final_jmp(*L_success); 6212 } 6213 } else { 6214 // No slow path; it's a fast decision. 6215 if (L_failure == &L_fallthrough) { 6216 local_jcc(Assembler::equal, *L_success); 6217 } else { 6218 local_jcc(Assembler::notEqual, *L_failure); 6219 final_jmp(*L_success); 6220 } 6221 } 6222 6223 bind(L_fallthrough); 6224 6225#undef local_jcc 6226#undef final_jmp 6227} 6228 6229 6230void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass, 6231 Register super_klass, 6232 Register temp_reg, 6233 Register temp2_reg, 6234 Label* L_success, 6235 Label* L_failure, 6236 bool set_cond_codes) { 6237 assert_different_registers(sub_klass, super_klass, temp_reg); 6238 if (temp2_reg != noreg) 6239 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg); 6240#define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg) 6241 6242 Label L_fallthrough; 6243 int label_nulls = 0; 6244 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 6245 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 6246 assert(label_nulls <= 1, "at most one NULL in the batch"); 6247 6248 // a couple of useful fields in sub_klass: 6249 int ss_offset = in_bytes(Klass::secondary_supers_offset()); 6250 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 6251 Address secondary_supers_addr(sub_klass, ss_offset); 6252 Address super_cache_addr( sub_klass, sc_offset); 6253 6254 // Do a linear scan of the secondary super-klass chain. 6255 // This code is rarely used, so simplicity is a virtue here. 6256 // The repne_scan instruction uses fixed registers, which we must spill. 6257 // Don't worry too much about pre-existing connections with the input regs. 6258 6259 assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super) 6260 assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter) 6261 6262 // Get super_klass value into rax (even if it was in rdi or rcx). 6263 bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false; 6264 if (super_klass != rax || UseCompressedOops) { 6265 if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; } 6266 mov(rax, super_klass); 6267 } 6268 if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; } 6269 if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; } 6270 6271#ifndef PRODUCT 6272 int* pst_counter = &SharedRuntime::_partial_subtype_ctr; 6273 ExternalAddress pst_counter_addr((address) pst_counter); 6274 NOT_LP64( incrementl(pst_counter_addr) ); 6275 LP64_ONLY( lea(rcx, pst_counter_addr) ); 6276 LP64_ONLY( incrementl(Address(rcx, 0)) ); 6277#endif //PRODUCT 6278 6279 // We will consult the secondary-super array. 6280 movptr(rdi, secondary_supers_addr); 6281 // Load the array length. (Positive movl does right thing on LP64.) 6282 movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes())); 6283 // Skip to start of data. 6284 addptr(rdi, Array<Klass*>::base_offset_in_bytes()); 6285 6286 // Scan RCX words at [RDI] for an occurrence of RAX. 6287 // Set NZ/Z based on last compare. 6288 // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does 6289 // not change flags (only scas instruction which is repeated sets flags). 6290 // Set Z = 0 (not equal) before 'repne' to indicate that class was not found. 6291 6292 testptr(rax,rax); // Set Z = 0 6293 repne_scan(); 6294 6295 // Unspill the temp. registers: 6296 if (pushed_rdi) pop(rdi); 6297 if (pushed_rcx) pop(rcx); 6298 if (pushed_rax) pop(rax); 6299 6300 if (set_cond_codes) { 6301 // Special hack for the AD files: rdi is guaranteed non-zero. 6302 assert(!pushed_rdi, "rdi must be left non-NULL"); 6303 // Also, the condition codes are properly set Z/NZ on succeed/failure. 6304 } 6305 6306 if (L_failure == &L_fallthrough) 6307 jccb(Assembler::notEqual, *L_failure); 6308 else jcc(Assembler::notEqual, *L_failure); 6309 6310 // Success. Cache the super we found and proceed in triumph. 6311 movptr(super_cache_addr, super_klass); 6312 6313 if (L_success != &L_fallthrough) { 6314 jmp(*L_success); 6315 } 6316 6317#undef IS_A_TEMP 6318 6319 bind(L_fallthrough); 6320} 6321 6322 6323void MacroAssembler::cmov32(Condition cc, Register dst, Address src) { 6324 if (VM_Version::supports_cmov()) { 6325 cmovl(cc, dst, src); 6326 } else { 6327 Label L; 6328 jccb(negate_condition(cc), L); 6329 movl(dst, src); 6330 bind(L); 6331 } 6332} 6333 6334void MacroAssembler::cmov32(Condition cc, Register dst, Register src) { 6335 if (VM_Version::supports_cmov()) { 6336 cmovl(cc, dst, src); 6337 } else { 6338 Label L; 6339 jccb(negate_condition(cc), L); 6340 movl(dst, src); 6341 bind(L); 6342 } 6343} 6344 6345void MacroAssembler::verify_oop(Register reg, const char* s) { 6346 if (!VerifyOops) return; 6347 6348 // Pass register number to verify_oop_subroutine 6349 const char* b = NULL; 6350 { 6351 ResourceMark rm; 6352 stringStream ss; 6353 ss.print("verify_oop: %s: %s", reg->name(), s); 6354 b = code_string(ss.as_string()); 6355 } 6356 BLOCK_COMMENT("verify_oop {"); 6357#ifdef _LP64 6358 push(rscratch1); // save r10, trashed by movptr() 6359#endif 6360 push(rax); // save rax, 6361 push(reg); // pass register argument 6362 ExternalAddress buffer((address) b); 6363 // avoid using pushptr, as it modifies scratch registers 6364 // and our contract is not to modify anything 6365 movptr(rax, buffer.addr()); 6366 push(rax); 6367 // call indirectly to solve generation ordering problem 6368 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); 6369 call(rax); 6370 // Caller pops the arguments (oop, message) and restores rax, r10 6371 BLOCK_COMMENT("} verify_oop"); 6372} 6373 6374 6375RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr, 6376 Register tmp, 6377 int offset) { 6378 intptr_t value = *delayed_value_addr; 6379 if (value != 0) 6380 return RegisterOrConstant(value + offset); 6381 6382 // load indirectly to solve generation ordering problem 6383 movptr(tmp, ExternalAddress((address) delayed_value_addr)); 6384 6385#ifdef ASSERT 6386 { Label L; 6387 testptr(tmp, tmp); 6388 if (WizardMode) { 6389 const char* buf = NULL; 6390 { 6391 ResourceMark rm; 6392 stringStream ss; 6393 ss.print("DelayedValue=" INTPTR_FORMAT, delayed_value_addr[1]); 6394 buf = code_string(ss.as_string()); 6395 } 6396 jcc(Assembler::notZero, L); 6397 STOP(buf); 6398 } else { 6399 jccb(Assembler::notZero, L); 6400 hlt(); 6401 } 6402 bind(L); 6403 } 6404#endif 6405 6406 if (offset != 0) 6407 addptr(tmp, offset); 6408 6409 return RegisterOrConstant(tmp); 6410} 6411 6412 6413Address MacroAssembler::argument_address(RegisterOrConstant arg_slot, 6414 int extra_slot_offset) { 6415 // cf. TemplateTable::prepare_invoke(), if (load_receiver). 6416 int stackElementSize = Interpreter::stackElementSize; 6417 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0); 6418#ifdef ASSERT 6419 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1); 6420 assert(offset1 - offset == stackElementSize, "correct arithmetic"); 6421#endif 6422 Register scale_reg = noreg; 6423 Address::ScaleFactor scale_factor = Address::no_scale; 6424 if (arg_slot.is_constant()) { 6425 offset += arg_slot.as_constant() * stackElementSize; 6426 } else { 6427 scale_reg = arg_slot.as_register(); 6428 scale_factor = Address::times(stackElementSize); 6429 } 6430 offset += wordSize; // return PC is on stack 6431 return Address(rsp, scale_reg, scale_factor, offset); 6432} 6433 6434 6435void MacroAssembler::verify_oop_addr(Address addr, const char* s) { 6436 if (!VerifyOops) return; 6437 6438 // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord); 6439 // Pass register number to verify_oop_subroutine 6440 const char* b = NULL; 6441 { 6442 ResourceMark rm; 6443 stringStream ss; 6444 ss.print("verify_oop_addr: %s", s); 6445 b = code_string(ss.as_string()); 6446 } 6447#ifdef _LP64 6448 push(rscratch1); // save r10, trashed by movptr() 6449#endif 6450 push(rax); // save rax, 6451 // addr may contain rsp so we will have to adjust it based on the push 6452 // we just did (and on 64 bit we do two pushes) 6453 // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which 6454 // stores rax into addr which is backwards of what was intended. 6455 if (addr.uses(rsp)) { 6456 lea(rax, addr); 6457 pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord)); 6458 } else { 6459 pushptr(addr); 6460 } 6461 6462 ExternalAddress buffer((address) b); 6463 // pass msg argument 6464 // avoid using pushptr, as it modifies scratch registers 6465 // and our contract is not to modify anything 6466 movptr(rax, buffer.addr()); 6467 push(rax); 6468 6469 // call indirectly to solve generation ordering problem 6470 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); 6471 call(rax); 6472 // Caller pops the arguments (addr, message) and restores rax, r10. 6473} 6474 6475void MacroAssembler::verify_tlab() { 6476#ifdef ASSERT 6477 if (UseTLAB && VerifyOops) { 6478 Label next, ok; 6479 Register t1 = rsi; 6480 Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread); 6481 6482 push(t1); 6483 NOT_LP64(push(thread_reg)); 6484 NOT_LP64(get_thread(thread_reg)); 6485 6486 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 6487 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset()))); 6488 jcc(Assembler::aboveEqual, next); 6489 STOP("assert(top >= start)"); 6490 should_not_reach_here(); 6491 6492 bind(next); 6493 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset()))); 6494 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 6495 jcc(Assembler::aboveEqual, ok); 6496 STOP("assert(top <= end)"); 6497 should_not_reach_here(); 6498 6499 bind(ok); 6500 NOT_LP64(pop(thread_reg)); 6501 pop(t1); 6502 } 6503#endif 6504} 6505 6506class ControlWord { 6507 public: 6508 int32_t _value; 6509 6510 int rounding_control() const { return (_value >> 10) & 3 ; } 6511 int precision_control() const { return (_value >> 8) & 3 ; } 6512 bool precision() const { return ((_value >> 5) & 1) != 0; } 6513 bool underflow() const { return ((_value >> 4) & 1) != 0; } 6514 bool overflow() const { return ((_value >> 3) & 1) != 0; } 6515 bool zero_divide() const { return ((_value >> 2) & 1) != 0; } 6516 bool denormalized() const { return ((_value >> 1) & 1) != 0; } 6517 bool invalid() const { return ((_value >> 0) & 1) != 0; } 6518 6519 void print() const { 6520 // rounding control 6521 const char* rc; 6522 switch (rounding_control()) { 6523 case 0: rc = "round near"; break; 6524 case 1: rc = "round down"; break; 6525 case 2: rc = "round up "; break; 6526 case 3: rc = "chop "; break; 6527 }; 6528 // precision control 6529 const char* pc; 6530 switch (precision_control()) { 6531 case 0: pc = "24 bits "; break; 6532 case 1: pc = "reserved"; break; 6533 case 2: pc = "53 bits "; break; 6534 case 3: pc = "64 bits "; break; 6535 }; 6536 // flags 6537 char f[9]; 6538 f[0] = ' '; 6539 f[1] = ' '; 6540 f[2] = (precision ()) ? 'P' : 'p'; 6541 f[3] = (underflow ()) ? 'U' : 'u'; 6542 f[4] = (overflow ()) ? 'O' : 'o'; 6543 f[5] = (zero_divide ()) ? 'Z' : 'z'; 6544 f[6] = (denormalized()) ? 'D' : 'd'; 6545 f[7] = (invalid ()) ? 'I' : 'i'; 6546 f[8] = '\x0'; 6547 // output 6548 printf("%04x masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc); 6549 } 6550 6551}; 6552 6553class StatusWord { 6554 public: 6555 int32_t _value; 6556 6557 bool busy() const { return ((_value >> 15) & 1) != 0; } 6558 bool C3() const { return ((_value >> 14) & 1) != 0; } 6559 bool C2() const { return ((_value >> 10) & 1) != 0; } 6560 bool C1() const { return ((_value >> 9) & 1) != 0; } 6561 bool C0() const { return ((_value >> 8) & 1) != 0; } 6562 int top() const { return (_value >> 11) & 7 ; } 6563 bool error_status() const { return ((_value >> 7) & 1) != 0; } 6564 bool stack_fault() const { return ((_value >> 6) & 1) != 0; } 6565 bool precision() const { return ((_value >> 5) & 1) != 0; } 6566 bool underflow() const { return ((_value >> 4) & 1) != 0; } 6567 bool overflow() const { return ((_value >> 3) & 1) != 0; } 6568 bool zero_divide() const { return ((_value >> 2) & 1) != 0; } 6569 bool denormalized() const { return ((_value >> 1) & 1) != 0; } 6570 bool invalid() const { return ((_value >> 0) & 1) != 0; } 6571 6572 void print() const { 6573 // condition codes 6574 char c[5]; 6575 c[0] = (C3()) ? '3' : '-'; 6576 c[1] = (C2()) ? '2' : '-'; 6577 c[2] = (C1()) ? '1' : '-'; 6578 c[3] = (C0()) ? '0' : '-'; 6579 c[4] = '\x0'; 6580 // flags 6581 char f[9]; 6582 f[0] = (error_status()) ? 'E' : '-'; 6583 f[1] = (stack_fault ()) ? 'S' : '-'; 6584 f[2] = (precision ()) ? 'P' : '-'; 6585 f[3] = (underflow ()) ? 'U' : '-'; 6586 f[4] = (overflow ()) ? 'O' : '-'; 6587 f[5] = (zero_divide ()) ? 'Z' : '-'; 6588 f[6] = (denormalized()) ? 'D' : '-'; 6589 f[7] = (invalid ()) ? 'I' : '-'; 6590 f[8] = '\x0'; 6591 // output 6592 printf("%04x flags = %s, cc = %s, top = %d", _value & 0xFFFF, f, c, top()); 6593 } 6594 6595}; 6596 6597class TagWord { 6598 public: 6599 int32_t _value; 6600 6601 int tag_at(int i) const { return (_value >> (i*2)) & 3; } 6602 6603 void print() const { 6604 printf("%04x", _value & 0xFFFF); 6605 } 6606 6607}; 6608 6609class FPU_Register { 6610 public: 6611 int32_t _m0; 6612 int32_t _m1; 6613 int16_t _ex; 6614 6615 bool is_indefinite() const { 6616 return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0; 6617 } 6618 6619 void print() const { 6620 char sign = (_ex < 0) ? '-' : '+'; 6621 const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : " "; 6622 printf("%c%04hx.%08x%08x %s", sign, _ex, _m1, _m0, kind); 6623 }; 6624 6625}; 6626 6627class FPU_State { 6628 public: 6629 enum { 6630 register_size = 10, 6631 number_of_registers = 8, 6632 register_mask = 7 6633 }; 6634 6635 ControlWord _control_word; 6636 StatusWord _status_word; 6637 TagWord _tag_word; 6638 int32_t _error_offset; 6639 int32_t _error_selector; 6640 int32_t _data_offset; 6641 int32_t _data_selector; 6642 int8_t _register[register_size * number_of_registers]; 6643 6644 int tag_for_st(int i) const { return _tag_word.tag_at((_status_word.top() + i) & register_mask); } 6645 FPU_Register* st(int i) const { return (FPU_Register*)&_register[register_size * i]; } 6646 6647 const char* tag_as_string(int tag) const { 6648 switch (tag) { 6649 case 0: return "valid"; 6650 case 1: return "zero"; 6651 case 2: return "special"; 6652 case 3: return "empty"; 6653 } 6654 ShouldNotReachHere(); 6655 return NULL; 6656 } 6657 6658 void print() const { 6659 // print computation registers 6660 { int t = _status_word.top(); 6661 for (int i = 0; i < number_of_registers; i++) { 6662 int j = (i - t) & register_mask; 6663 printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j); 6664 st(j)->print(); 6665 printf(" %s\n", tag_as_string(_tag_word.tag_at(i))); 6666 } 6667 } 6668 printf("\n"); 6669 // print control registers 6670 printf("ctrl = "); _control_word.print(); printf("\n"); 6671 printf("stat = "); _status_word .print(); printf("\n"); 6672 printf("tags = "); _tag_word .print(); printf("\n"); 6673 } 6674 6675}; 6676 6677class Flag_Register { 6678 public: 6679 int32_t _value; 6680 6681 bool overflow() const { return ((_value >> 11) & 1) != 0; } 6682 bool direction() const { return ((_value >> 10) & 1) != 0; } 6683 bool sign() const { return ((_value >> 7) & 1) != 0; } 6684 bool zero() const { return ((_value >> 6) & 1) != 0; } 6685 bool auxiliary_carry() const { return ((_value >> 4) & 1) != 0; } 6686 bool parity() const { return ((_value >> 2) & 1) != 0; } 6687 bool carry() const { return ((_value >> 0) & 1) != 0; } 6688 6689 void print() const { 6690 // flags 6691 char f[8]; 6692 f[0] = (overflow ()) ? 'O' : '-'; 6693 f[1] = (direction ()) ? 'D' : '-'; 6694 f[2] = (sign ()) ? 'S' : '-'; 6695 f[3] = (zero ()) ? 'Z' : '-'; 6696 f[4] = (auxiliary_carry()) ? 'A' : '-'; 6697 f[5] = (parity ()) ? 'P' : '-'; 6698 f[6] = (carry ()) ? 'C' : '-'; 6699 f[7] = '\x0'; 6700 // output 6701 printf("%08x flags = %s", _value, f); 6702 } 6703 6704}; 6705 6706class IU_Register { 6707 public: 6708 int32_t _value; 6709 6710 void print() const { 6711 printf("%08x %11d", _value, _value); 6712 } 6713 6714}; 6715 6716class IU_State { 6717 public: 6718 Flag_Register _eflags; 6719 IU_Register _rdi; 6720 IU_Register _rsi; 6721 IU_Register _rbp; 6722 IU_Register _rsp; 6723 IU_Register _rbx; 6724 IU_Register _rdx; 6725 IU_Register _rcx; 6726 IU_Register _rax; 6727 6728 void print() const { 6729 // computation registers 6730 printf("rax, = "); _rax.print(); printf("\n"); 6731 printf("rbx, = "); _rbx.print(); printf("\n"); 6732 printf("rcx = "); _rcx.print(); printf("\n"); 6733 printf("rdx = "); _rdx.print(); printf("\n"); 6734 printf("rdi = "); _rdi.print(); printf("\n"); 6735 printf("rsi = "); _rsi.print(); printf("\n"); 6736 printf("rbp, = "); _rbp.print(); printf("\n"); 6737 printf("rsp = "); _rsp.print(); printf("\n"); 6738 printf("\n"); 6739 // control registers 6740 printf("flgs = "); _eflags.print(); printf("\n"); 6741 } 6742}; 6743 6744 6745class CPU_State { 6746 public: 6747 FPU_State _fpu_state; 6748 IU_State _iu_state; 6749 6750 void print() const { 6751 printf("--------------------------------------------------\n"); 6752 _iu_state .print(); 6753 printf("\n"); 6754 _fpu_state.print(); 6755 printf("--------------------------------------------------\n"); 6756 } 6757 6758}; 6759 6760 6761static void _print_CPU_state(CPU_State* state) { 6762 state->print(); 6763}; 6764 6765 6766void MacroAssembler::print_CPU_state() { 6767 push_CPU_state(); 6768 push(rsp); // pass CPU state 6769 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state))); 6770 addptr(rsp, wordSize); // discard argument 6771 pop_CPU_state(); 6772} 6773 6774 6775static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) { 6776 static int counter = 0; 6777 FPU_State* fs = &state->_fpu_state; 6778 counter++; 6779 // For leaf calls, only verify that the top few elements remain empty. 6780 // We only need 1 empty at the top for C2 code. 6781 if( stack_depth < 0 ) { 6782 if( fs->tag_for_st(7) != 3 ) { 6783 printf("FPR7 not empty\n"); 6784 state->print(); 6785 assert(false, "error"); 6786 return false; 6787 } 6788 return true; // All other stack states do not matter 6789 } 6790 6791 assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std, 6792 "bad FPU control word"); 6793 6794 // compute stack depth 6795 int i = 0; 6796 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) < 3) i++; 6797 int d = i; 6798 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++; 6799 // verify findings 6800 if (i != FPU_State::number_of_registers) { 6801 // stack not contiguous 6802 printf("%s: stack not contiguous at ST%d\n", s, i); 6803 state->print(); 6804 assert(false, "error"); 6805 return false; 6806 } 6807 // check if computed stack depth corresponds to expected stack depth 6808 if (stack_depth < 0) { 6809 // expected stack depth is -stack_depth or less 6810 if (d > -stack_depth) { 6811 // too many elements on the stack 6812 printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d); 6813 state->print(); 6814 assert(false, "error"); 6815 return false; 6816 } 6817 } else { 6818 // expected stack depth is stack_depth 6819 if (d != stack_depth) { 6820 // wrong stack depth 6821 printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d); 6822 state->print(); 6823 assert(false, "error"); 6824 return false; 6825 } 6826 } 6827 // everything is cool 6828 return true; 6829} 6830 6831 6832void MacroAssembler::verify_FPU(int stack_depth, const char* s) { 6833 if (!VerifyFPU) return; 6834 push_CPU_state(); 6835 push(rsp); // pass CPU state 6836 ExternalAddress msg((address) s); 6837 // pass message string s 6838 pushptr(msg.addr()); 6839 push(stack_depth); // pass stack depth 6840 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU))); 6841 addptr(rsp, 3 * wordSize); // discard arguments 6842 // check for error 6843 { Label L; 6844 testl(rax, rax); 6845 jcc(Assembler::notZero, L); 6846 int3(); // break if error condition 6847 bind(L); 6848 } 6849 pop_CPU_state(); 6850} 6851 6852void MacroAssembler::restore_cpu_control_state_after_jni() { 6853 // Either restore the MXCSR register after returning from the JNI Call 6854 // or verify that it wasn't changed (with -Xcheck:jni flag). 6855 if (VM_Version::supports_sse()) { 6856 if (RestoreMXCSROnJNICalls) { 6857 ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std())); 6858 } else if (CheckJNICalls) { 6859 call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry())); 6860 } 6861 } 6862 if (VM_Version::supports_avx()) { 6863 // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty. 6864 vzeroupper(); 6865 } 6866 6867#ifndef _LP64 6868 // Either restore the x87 floating pointer control word after returning 6869 // from the JNI call or verify that it wasn't changed. 6870 if (CheckJNICalls) { 6871 call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry())); 6872 } 6873#endif // _LP64 6874} 6875 6876 6877void MacroAssembler::load_klass(Register dst, Register src) { 6878#ifdef _LP64 6879 if (UseCompressedClassPointers) { 6880 movl(dst, Address(src, oopDesc::klass_offset_in_bytes())); 6881 decode_klass_not_null(dst); 6882 } else 6883#endif 6884 movptr(dst, Address(src, oopDesc::klass_offset_in_bytes())); 6885} 6886 6887void MacroAssembler::load_prototype_header(Register dst, Register src) { 6888 load_klass(dst, src); 6889 movptr(dst, Address(dst, Klass::prototype_header_offset())); 6890} 6891 6892void MacroAssembler::store_klass(Register dst, Register src) { 6893#ifdef _LP64 6894 if (UseCompressedClassPointers) { 6895 encode_klass_not_null(src); 6896 movl(Address(dst, oopDesc::klass_offset_in_bytes()), src); 6897 } else 6898#endif 6899 movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src); 6900} 6901 6902void MacroAssembler::load_heap_oop(Register dst, Address src) { 6903#ifdef _LP64 6904 // FIXME: Must change all places where we try to load the klass. 6905 if (UseCompressedOops) { 6906 movl(dst, src); 6907 decode_heap_oop(dst); 6908 } else 6909#endif 6910 movptr(dst, src); 6911} 6912 6913// Doesn't do verfication, generates fixed size code 6914void MacroAssembler::load_heap_oop_not_null(Register dst, Address src) { 6915#ifdef _LP64 6916 if (UseCompressedOops) { 6917 movl(dst, src); 6918 decode_heap_oop_not_null(dst); 6919 } else 6920#endif 6921 movptr(dst, src); 6922} 6923 6924void MacroAssembler::store_heap_oop(Address dst, Register src) { 6925#ifdef _LP64 6926 if (UseCompressedOops) { 6927 assert(!dst.uses(src), "not enough registers"); 6928 encode_heap_oop(src); 6929 movl(dst, src); 6930 } else 6931#endif 6932 movptr(dst, src); 6933} 6934 6935void MacroAssembler::cmp_heap_oop(Register src1, Address src2, Register tmp) { 6936 assert_different_registers(src1, tmp); 6937#ifdef _LP64 6938 if (UseCompressedOops) { 6939 bool did_push = false; 6940 if (tmp == noreg) { 6941 tmp = rax; 6942 push(tmp); 6943 did_push = true; 6944 assert(!src2.uses(rsp), "can't push"); 6945 } 6946 load_heap_oop(tmp, src2); 6947 cmpptr(src1, tmp); 6948 if (did_push) pop(tmp); 6949 } else 6950#endif 6951 cmpptr(src1, src2); 6952} 6953 6954// Used for storing NULLs. 6955void MacroAssembler::store_heap_oop_null(Address dst) { 6956#ifdef _LP64 6957 if (UseCompressedOops) { 6958 movl(dst, (int32_t)NULL_WORD); 6959 } else { 6960 movslq(dst, (int32_t)NULL_WORD); 6961 } 6962#else 6963 movl(dst, (int32_t)NULL_WORD); 6964#endif 6965} 6966 6967#ifdef _LP64 6968void MacroAssembler::store_klass_gap(Register dst, Register src) { 6969 if (UseCompressedClassPointers) { 6970 // Store to klass gap in destination 6971 movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src); 6972 } 6973} 6974 6975#ifdef ASSERT 6976void MacroAssembler::verify_heapbase(const char* msg) { 6977 assert (UseCompressedOops, "should be compressed"); 6978 assert (Universe::heap() != NULL, "java heap should be initialized"); 6979 if (CheckCompressedOops) { 6980 Label ok; 6981 push(rscratch1); // cmpptr trashes rscratch1 6982 cmpptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr())); 6983 jcc(Assembler::equal, ok); 6984 STOP(msg); 6985 bind(ok); 6986 pop(rscratch1); 6987 } 6988} 6989#endif 6990 6991// Algorithm must match oop.inline.hpp encode_heap_oop. 6992void MacroAssembler::encode_heap_oop(Register r) { 6993#ifdef ASSERT 6994 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?"); 6995#endif 6996 verify_oop(r, "broken oop in encode_heap_oop"); 6997 if (Universe::narrow_oop_base() == NULL) { 6998 if (Universe::narrow_oop_shift() != 0) { 6999 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 7000 shrq(r, LogMinObjAlignmentInBytes); 7001 } 7002 return; 7003 } 7004 testq(r, r); 7005 cmovq(Assembler::equal, r, r12_heapbase); 7006 subq(r, r12_heapbase); 7007 shrq(r, LogMinObjAlignmentInBytes); 7008} 7009 7010void MacroAssembler::encode_heap_oop_not_null(Register r) { 7011#ifdef ASSERT 7012 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?"); 7013 if (CheckCompressedOops) { 7014 Label ok; 7015 testq(r, r); 7016 jcc(Assembler::notEqual, ok); 7017 STOP("null oop passed to encode_heap_oop_not_null"); 7018 bind(ok); 7019 } 7020#endif 7021 verify_oop(r, "broken oop in encode_heap_oop_not_null"); 7022 if (Universe::narrow_oop_base() != NULL) { 7023 subq(r, r12_heapbase); 7024 } 7025 if (Universe::narrow_oop_shift() != 0) { 7026 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 7027 shrq(r, LogMinObjAlignmentInBytes); 7028 } 7029} 7030 7031void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) { 7032#ifdef ASSERT 7033 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?"); 7034 if (CheckCompressedOops) { 7035 Label ok; 7036 testq(src, src); 7037 jcc(Assembler::notEqual, ok); 7038 STOP("null oop passed to encode_heap_oop_not_null2"); 7039 bind(ok); 7040 } 7041#endif 7042 verify_oop(src, "broken oop in encode_heap_oop_not_null2"); 7043 if (dst != src) { 7044 movq(dst, src); 7045 } 7046 if (Universe::narrow_oop_base() != NULL) { 7047 subq(dst, r12_heapbase); 7048 } 7049 if (Universe::narrow_oop_shift() != 0) { 7050 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 7051 shrq(dst, LogMinObjAlignmentInBytes); 7052 } 7053} 7054 7055void MacroAssembler::decode_heap_oop(Register r) { 7056#ifdef ASSERT 7057 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?"); 7058#endif 7059 if (Universe::narrow_oop_base() == NULL) { 7060 if (Universe::narrow_oop_shift() != 0) { 7061 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 7062 shlq(r, LogMinObjAlignmentInBytes); 7063 } 7064 } else { 7065 Label done; 7066 shlq(r, LogMinObjAlignmentInBytes); 7067 jccb(Assembler::equal, done); 7068 addq(r, r12_heapbase); 7069 bind(done); 7070 } 7071 verify_oop(r, "broken oop in decode_heap_oop"); 7072} 7073 7074void MacroAssembler::decode_heap_oop_not_null(Register r) { 7075 // Note: it will change flags 7076 assert (UseCompressedOops, "should only be used for compressed headers"); 7077 assert (Universe::heap() != NULL, "java heap should be initialized"); 7078 // Cannot assert, unverified entry point counts instructions (see .ad file) 7079 // vtableStubs also counts instructions in pd_code_size_limit. 7080 // Also do not verify_oop as this is called by verify_oop. 7081 if (Universe::narrow_oop_shift() != 0) { 7082 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 7083 shlq(r, LogMinObjAlignmentInBytes); 7084 if (Universe::narrow_oop_base() != NULL) { 7085 addq(r, r12_heapbase); 7086 } 7087 } else { 7088 assert (Universe::narrow_oop_base() == NULL, "sanity"); 7089 } 7090} 7091 7092void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) { 7093 // Note: it will change flags 7094 assert (UseCompressedOops, "should only be used for compressed headers"); 7095 assert (Universe::heap() != NULL, "java heap should be initialized"); 7096 // Cannot assert, unverified entry point counts instructions (see .ad file) 7097 // vtableStubs also counts instructions in pd_code_size_limit. 7098 // Also do not verify_oop as this is called by verify_oop. 7099 if (Universe::narrow_oop_shift() != 0) { 7100 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 7101 if (LogMinObjAlignmentInBytes == Address::times_8) { 7102 leaq(dst, Address(r12_heapbase, src, Address::times_8, 0)); 7103 } else { 7104 if (dst != src) { 7105 movq(dst, src); 7106 } 7107 shlq(dst, LogMinObjAlignmentInBytes); 7108 if (Universe::narrow_oop_base() != NULL) { 7109 addq(dst, r12_heapbase); 7110 } 7111 } 7112 } else { 7113 assert (Universe::narrow_oop_base() == NULL, "sanity"); 7114 if (dst != src) { 7115 movq(dst, src); 7116 } 7117 } 7118} 7119 7120void MacroAssembler::encode_klass_not_null(Register r) { 7121 if (Universe::narrow_klass_base() != NULL) { 7122 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base. 7123 assert(r != r12_heapbase, "Encoding a klass in r12"); 7124 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base()); 7125 subq(r, r12_heapbase); 7126 } 7127 if (Universe::narrow_klass_shift() != 0) { 7128 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 7129 shrq(r, LogKlassAlignmentInBytes); 7130 } 7131 if (Universe::narrow_klass_base() != NULL) { 7132 reinit_heapbase(); 7133 } 7134} 7135 7136void MacroAssembler::encode_klass_not_null(Register dst, Register src) { 7137 if (dst == src) { 7138 encode_klass_not_null(src); 7139 } else { 7140 if (Universe::narrow_klass_base() != NULL) { 7141 mov64(dst, (int64_t)Universe::narrow_klass_base()); 7142 negq(dst); 7143 addq(dst, src); 7144 } else { 7145 movptr(dst, src); 7146 } 7147 if (Universe::narrow_klass_shift() != 0) { 7148 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 7149 shrq(dst, LogKlassAlignmentInBytes); 7150 } 7151 } 7152} 7153 7154// Function instr_size_for_decode_klass_not_null() counts the instructions 7155// generated by decode_klass_not_null(register r) and reinit_heapbase(), 7156// when (Universe::heap() != NULL). Hence, if the instructions they 7157// generate change, then this method needs to be updated. 7158int MacroAssembler::instr_size_for_decode_klass_not_null() { 7159 assert (UseCompressedClassPointers, "only for compressed klass ptrs"); 7160 if (Universe::narrow_klass_base() != NULL) { 7161 // mov64 + addq + shlq? + mov64 (for reinit_heapbase()). 7162 return (Universe::narrow_klass_shift() == 0 ? 20 : 24); 7163 } else { 7164 // longest load decode klass function, mov64, leaq 7165 return 16; 7166 } 7167} 7168 7169// !!! If the instructions that get generated here change then function 7170// instr_size_for_decode_klass_not_null() needs to get updated. 7171void MacroAssembler::decode_klass_not_null(Register r) { 7172 // Note: it will change flags 7173 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7174 assert(r != r12_heapbase, "Decoding a klass in r12"); 7175 // Cannot assert, unverified entry point counts instructions (see .ad file) 7176 // vtableStubs also counts instructions in pd_code_size_limit. 7177 // Also do not verify_oop as this is called by verify_oop. 7178 if (Universe::narrow_klass_shift() != 0) { 7179 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 7180 shlq(r, LogKlassAlignmentInBytes); 7181 } 7182 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base. 7183 if (Universe::narrow_klass_base() != NULL) { 7184 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base()); 7185 addq(r, r12_heapbase); 7186 reinit_heapbase(); 7187 } 7188} 7189 7190void MacroAssembler::decode_klass_not_null(Register dst, Register src) { 7191 // Note: it will change flags 7192 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7193 if (dst == src) { 7194 decode_klass_not_null(dst); 7195 } else { 7196 // Cannot assert, unverified entry point counts instructions (see .ad file) 7197 // vtableStubs also counts instructions in pd_code_size_limit. 7198 // Also do not verify_oop as this is called by verify_oop. 7199 mov64(dst, (int64_t)Universe::narrow_klass_base()); 7200 if (Universe::narrow_klass_shift() != 0) { 7201 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 7202 assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?"); 7203 leaq(dst, Address(dst, src, Address::times_8, 0)); 7204 } else { 7205 addq(dst, src); 7206 } 7207 } 7208} 7209 7210void MacroAssembler::set_narrow_oop(Register dst, jobject obj) { 7211 assert (UseCompressedOops, "should only be used for compressed headers"); 7212 assert (Universe::heap() != NULL, "java heap should be initialized"); 7213 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7214 int oop_index = oop_recorder()->find_index(obj); 7215 RelocationHolder rspec = oop_Relocation::spec(oop_index); 7216 mov_narrow_oop(dst, oop_index, rspec); 7217} 7218 7219void MacroAssembler::set_narrow_oop(Address dst, jobject obj) { 7220 assert (UseCompressedOops, "should only be used for compressed headers"); 7221 assert (Universe::heap() != NULL, "java heap should be initialized"); 7222 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7223 int oop_index = oop_recorder()->find_index(obj); 7224 RelocationHolder rspec = oop_Relocation::spec(oop_index); 7225 mov_narrow_oop(dst, oop_index, rspec); 7226} 7227 7228void MacroAssembler::set_narrow_klass(Register dst, Klass* k) { 7229 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7230 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7231 int klass_index = oop_recorder()->find_index(k); 7232 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 7233 mov_narrow_oop(dst, Klass::encode_klass(k), rspec); 7234} 7235 7236void MacroAssembler::set_narrow_klass(Address dst, Klass* k) { 7237 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7238 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7239 int klass_index = oop_recorder()->find_index(k); 7240 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 7241 mov_narrow_oop(dst, Klass::encode_klass(k), rspec); 7242} 7243 7244void MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) { 7245 assert (UseCompressedOops, "should only be used for compressed headers"); 7246 assert (Universe::heap() != NULL, "java heap should be initialized"); 7247 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7248 int oop_index = oop_recorder()->find_index(obj); 7249 RelocationHolder rspec = oop_Relocation::spec(oop_index); 7250 Assembler::cmp_narrow_oop(dst, oop_index, rspec); 7251} 7252 7253void MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) { 7254 assert (UseCompressedOops, "should only be used for compressed headers"); 7255 assert (Universe::heap() != NULL, "java heap should be initialized"); 7256 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7257 int oop_index = oop_recorder()->find_index(obj); 7258 RelocationHolder rspec = oop_Relocation::spec(oop_index); 7259 Assembler::cmp_narrow_oop(dst, oop_index, rspec); 7260} 7261 7262void MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) { 7263 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7264 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7265 int klass_index = oop_recorder()->find_index(k); 7266 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 7267 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec); 7268} 7269 7270void MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) { 7271 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7272 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7273 int klass_index = oop_recorder()->find_index(k); 7274 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 7275 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec); 7276} 7277 7278void MacroAssembler::reinit_heapbase() { 7279 if (UseCompressedOops || UseCompressedClassPointers) { 7280 if (Universe::heap() != NULL) { 7281 if (Universe::narrow_oop_base() == NULL) { 7282 MacroAssembler::xorptr(r12_heapbase, r12_heapbase); 7283 } else { 7284 mov64(r12_heapbase, (int64_t)Universe::narrow_ptrs_base()); 7285 } 7286 } else { 7287 movptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr())); 7288 } 7289 } 7290} 7291 7292#endif // _LP64 7293 7294 7295// C2 compiled method's prolog code. 7296void MacroAssembler::verified_entry(int framesize, int stack_bang_size, bool fp_mode_24b) { 7297 7298 // WARNING: Initial instruction MUST be 5 bytes or longer so that 7299 // NativeJump::patch_verified_entry will be able to patch out the entry 7300 // code safely. The push to verify stack depth is ok at 5 bytes, 7301 // the frame allocation can be either 3 or 6 bytes. So if we don't do 7302 // stack bang then we must use the 6 byte frame allocation even if 7303 // we have no frame. :-( 7304 assert(stack_bang_size >= framesize || stack_bang_size <= 0, "stack bang size incorrect"); 7305 7306 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned"); 7307 // Remove word for return addr 7308 framesize -= wordSize; 7309 stack_bang_size -= wordSize; 7310 7311 // Calls to C2R adapters often do not accept exceptional returns. 7312 // We require that their callers must bang for them. But be careful, because 7313 // some VM calls (such as call site linkage) can use several kilobytes of 7314 // stack. But the stack safety zone should account for that. 7315 // See bugs 4446381, 4468289, 4497237. 7316 if (stack_bang_size > 0) { 7317 generate_stack_overflow_check(stack_bang_size); 7318 7319 // We always push rbp, so that on return to interpreter rbp, will be 7320 // restored correctly and we can correct the stack. 7321 push(rbp); 7322 // Save caller's stack pointer into RBP if the frame pointer is preserved. 7323 if (PreserveFramePointer) { 7324 mov(rbp, rsp); 7325 } 7326 // Remove word for ebp 7327 framesize -= wordSize; 7328 7329 // Create frame 7330 if (framesize) { 7331 subptr(rsp, framesize); 7332 } 7333 } else { 7334 // Create frame (force generation of a 4 byte immediate value) 7335 subptr_imm32(rsp, framesize); 7336 7337 // Save RBP register now. 7338 framesize -= wordSize; 7339 movptr(Address(rsp, framesize), rbp); 7340 // Save caller's stack pointer into RBP if the frame pointer is preserved. 7341 if (PreserveFramePointer) { 7342 movptr(rbp, rsp); 7343 if (framesize > 0) { 7344 addptr(rbp, framesize); 7345 } 7346 } 7347 } 7348 7349 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth 7350 framesize -= wordSize; 7351 movptr(Address(rsp, framesize), (int32_t)0xbadb100d); 7352 } 7353 7354#ifndef _LP64 7355 // If method sets FPU control word do it now 7356 if (fp_mode_24b) { 7357 fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24())); 7358 } 7359 if (UseSSE >= 2 && VerifyFPU) { 7360 verify_FPU(0, "FPU stack must be clean on entry"); 7361 } 7362#endif 7363 7364#ifdef ASSERT 7365 if (VerifyStackAtCalls) { 7366 Label L; 7367 push(rax); 7368 mov(rax, rsp); 7369 andptr(rax, StackAlignmentInBytes-1); 7370 cmpptr(rax, StackAlignmentInBytes-wordSize); 7371 pop(rax); 7372 jcc(Assembler::equal, L); 7373 STOP("Stack is not properly aligned!"); 7374 bind(L); 7375 } 7376#endif 7377 7378} 7379 7380void MacroAssembler::clear_mem(Register base, Register cnt, Register tmp) { 7381 // cnt - number of qwords (8-byte words). 7382 // base - start address, qword aligned. 7383 assert(base==rdi, "base register must be edi for rep stos"); 7384 assert(tmp==rax, "tmp register must be eax for rep stos"); 7385 assert(cnt==rcx, "cnt register must be ecx for rep stos"); 7386 7387 xorptr(tmp, tmp); 7388 if (UseFastStosb) { 7389 shlptr(cnt,3); // convert to number of bytes 7390 rep_stosb(); 7391 } else { 7392 NOT_LP64(shlptr(cnt,1);) // convert to number of dwords for 32-bit VM 7393 rep_stos(); 7394 } 7395} 7396 7397#ifdef COMPILER2 7398 7399// IndexOf for constant substrings with size >= 8 chars 7400// which don't need to be loaded through stack. 7401void MacroAssembler::string_indexofC8(Register str1, Register str2, 7402 Register cnt1, Register cnt2, 7403 int int_cnt2, Register result, 7404 XMMRegister vec, Register tmp, 7405 int ae) { 7406 ShortBranchVerifier sbv(this); 7407 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 7408 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 7409 assert(ae != StrIntrinsicNode::LU, "Invalid encoding"); 7410 7411 // This method uses the pcmpestri instruction with bound registers 7412 // inputs: 7413 // xmm - substring 7414 // rax - substring length (elements count) 7415 // mem - scanned string 7416 // rdx - string length (elements count) 7417 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts) 7418 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes) 7419 // outputs: 7420 // rcx - matched index in string 7421 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 7422 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts 7423 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8 7424 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2; 7425 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1; 7426 7427 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, 7428 RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR, 7429 MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE; 7430 7431 // Note, inline_string_indexOf() generates checks: 7432 // if (substr.count > string.count) return -1; 7433 // if (substr.count == 0) return 0; 7434 assert(int_cnt2 >= stride, "this code is used only for cnt2 >= 8 chars"); 7435 7436 // Load substring. 7437 if (ae == StrIntrinsicNode::UL) { 7438 pmovzxbw(vec, Address(str2, 0)); 7439 } else { 7440 movdqu(vec, Address(str2, 0)); 7441 } 7442 movl(cnt2, int_cnt2); 7443 movptr(result, str1); // string addr 7444 7445 if (int_cnt2 > stride) { 7446 jmpb(SCAN_TO_SUBSTR); 7447 7448 // Reload substr for rescan, this code 7449 // is executed only for large substrings (> 8 chars) 7450 bind(RELOAD_SUBSTR); 7451 if (ae == StrIntrinsicNode::UL) { 7452 pmovzxbw(vec, Address(str2, 0)); 7453 } else { 7454 movdqu(vec, Address(str2, 0)); 7455 } 7456 negptr(cnt2); // Jumped here with negative cnt2, convert to positive 7457 7458 bind(RELOAD_STR); 7459 // We came here after the beginning of the substring was 7460 // matched but the rest of it was not so we need to search 7461 // again. Start from the next element after the previous match. 7462 7463 // cnt2 is number of substring reminding elements and 7464 // cnt1 is number of string reminding elements when cmp failed. 7465 // Restored cnt1 = cnt1 - cnt2 + int_cnt2 7466 subl(cnt1, cnt2); 7467 addl(cnt1, int_cnt2); 7468 movl(cnt2, int_cnt2); // Now restore cnt2 7469 7470 decrementl(cnt1); // Shift to next element 7471 cmpl(cnt1, cnt2); 7472 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7473 7474 addptr(result, (1<<scale1)); 7475 7476 } // (int_cnt2 > 8) 7477 7478 // Scan string for start of substr in 16-byte vectors 7479 bind(SCAN_TO_SUBSTR); 7480 pcmpestri(vec, Address(result, 0), mode); 7481 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1 7482 subl(cnt1, stride); 7483 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string 7484 cmpl(cnt1, cnt2); 7485 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7486 addptr(result, 16); 7487 jmpb(SCAN_TO_SUBSTR); 7488 7489 // Found a potential substr 7490 bind(FOUND_CANDIDATE); 7491 // Matched whole vector if first element matched (tmp(rcx) == 0). 7492 if (int_cnt2 == stride) { 7493 jccb(Assembler::overflow, RET_FOUND); // OF == 1 7494 } else { // int_cnt2 > 8 7495 jccb(Assembler::overflow, FOUND_SUBSTR); 7496 } 7497 // After pcmpestri tmp(rcx) contains matched element index 7498 // Compute start addr of substr 7499 lea(result, Address(result, tmp, scale1)); 7500 7501 // Make sure string is still long enough 7502 subl(cnt1, tmp); 7503 cmpl(cnt1, cnt2); 7504 if (int_cnt2 == stride) { 7505 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR); 7506 } else { // int_cnt2 > 8 7507 jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD); 7508 } 7509 // Left less then substring. 7510 7511 bind(RET_NOT_FOUND); 7512 movl(result, -1); 7513 jmpb(EXIT); 7514 7515 if (int_cnt2 > stride) { 7516 // This code is optimized for the case when whole substring 7517 // is matched if its head is matched. 7518 bind(MATCH_SUBSTR_HEAD); 7519 pcmpestri(vec, Address(result, 0), mode); 7520 // Reload only string if does not match 7521 jccb(Assembler::noOverflow, RELOAD_STR); // OF == 0 7522 7523 Label CONT_SCAN_SUBSTR; 7524 // Compare the rest of substring (> 8 chars). 7525 bind(FOUND_SUBSTR); 7526 // First 8 chars are already matched. 7527 negptr(cnt2); 7528 addptr(cnt2, stride); 7529 7530 bind(SCAN_SUBSTR); 7531 subl(cnt1, stride); 7532 cmpl(cnt2, -stride); // Do not read beyond substring 7533 jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR); 7534 // Back-up strings to avoid reading beyond substring: 7535 // cnt1 = cnt1 - cnt2 + 8 7536 addl(cnt1, cnt2); // cnt2 is negative 7537 addl(cnt1, stride); 7538 movl(cnt2, stride); negptr(cnt2); 7539 bind(CONT_SCAN_SUBSTR); 7540 if (int_cnt2 < (int)G) { 7541 int tail_off1 = int_cnt2<<scale1; 7542 int tail_off2 = int_cnt2<<scale2; 7543 if (ae == StrIntrinsicNode::UL) { 7544 pmovzxbw(vec, Address(str2, cnt2, scale2, tail_off2)); 7545 } else { 7546 movdqu(vec, Address(str2, cnt2, scale2, tail_off2)); 7547 } 7548 pcmpestri(vec, Address(result, cnt2, scale1, tail_off1), mode); 7549 } else { 7550 // calculate index in register to avoid integer overflow (int_cnt2*2) 7551 movl(tmp, int_cnt2); 7552 addptr(tmp, cnt2); 7553 if (ae == StrIntrinsicNode::UL) { 7554 pmovzxbw(vec, Address(str2, tmp, scale2, 0)); 7555 } else { 7556 movdqu(vec, Address(str2, tmp, scale2, 0)); 7557 } 7558 pcmpestri(vec, Address(result, tmp, scale1, 0), mode); 7559 } 7560 // Need to reload strings pointers if not matched whole vector 7561 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0 7562 addptr(cnt2, stride); 7563 jcc(Assembler::negative, SCAN_SUBSTR); 7564 // Fall through if found full substring 7565 7566 } // (int_cnt2 > 8) 7567 7568 bind(RET_FOUND); 7569 // Found result if we matched full small substring. 7570 // Compute substr offset 7571 subptr(result, str1); 7572 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 7573 shrl(result, 1); // index 7574 } 7575 bind(EXIT); 7576 7577} // string_indexofC8 7578 7579// Small strings are loaded through stack if they cross page boundary. 7580void MacroAssembler::string_indexof(Register str1, Register str2, 7581 Register cnt1, Register cnt2, 7582 int int_cnt2, Register result, 7583 XMMRegister vec, Register tmp, 7584 int ae) { 7585 ShortBranchVerifier sbv(this); 7586 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 7587 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 7588 assert(ae != StrIntrinsicNode::LU, "Invalid encoding"); 7589 7590 // 7591 // int_cnt2 is length of small (< 8 chars) constant substring 7592 // or (-1) for non constant substring in which case its length 7593 // is in cnt2 register. 7594 // 7595 // Note, inline_string_indexOf() generates checks: 7596 // if (substr.count > string.count) return -1; 7597 // if (substr.count == 0) return 0; 7598 // 7599 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8 7600 assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < stride), "should be != 0"); 7601 // This method uses the pcmpestri instruction with bound registers 7602 // inputs: 7603 // xmm - substring 7604 // rax - substring length (elements count) 7605 // mem - scanned string 7606 // rdx - string length (elements count) 7607 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts) 7608 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes) 7609 // outputs: 7610 // rcx - matched index in string 7611 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 7612 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts 7613 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2; 7614 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1; 7615 7616 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR, 7617 RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR, 7618 FOUND_CANDIDATE; 7619 7620 { //======================================================== 7621 // We don't know where these strings are located 7622 // and we can't read beyond them. Load them through stack. 7623 Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR; 7624 7625 movptr(tmp, rsp); // save old SP 7626 7627 if (int_cnt2 > 0) { // small (< 8 chars) constant substring 7628 if (int_cnt2 == (1>>scale2)) { // One byte 7629 assert((ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL), "Only possible for latin1 encoding"); 7630 load_unsigned_byte(result, Address(str2, 0)); 7631 movdl(vec, result); // move 32 bits 7632 } else if (ae == StrIntrinsicNode::LL && int_cnt2 == 3) { // Three bytes 7633 // Not enough header space in 32-bit VM: 12+3 = 15. 7634 movl(result, Address(str2, -1)); 7635 shrl(result, 8); 7636 movdl(vec, result); // move 32 bits 7637 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (2>>scale2)) { // One char 7638 load_unsigned_short(result, Address(str2, 0)); 7639 movdl(vec, result); // move 32 bits 7640 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (4>>scale2)) { // Two chars 7641 movdl(vec, Address(str2, 0)); // move 32 bits 7642 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (8>>scale2)) { // Four chars 7643 movq(vec, Address(str2, 0)); // move 64 bits 7644 } else { // cnt2 = { 3, 5, 6, 7 } || (ae == StrIntrinsicNode::UL && cnt2 ={2, ..., 7}) 7645 // Array header size is 12 bytes in 32-bit VM 7646 // + 6 bytes for 3 chars == 18 bytes, 7647 // enough space to load vec and shift. 7648 assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity"); 7649 if (ae == StrIntrinsicNode::UL) { 7650 int tail_off = int_cnt2-8; 7651 pmovzxbw(vec, Address(str2, tail_off)); 7652 psrldq(vec, -2*tail_off); 7653 } 7654 else { 7655 int tail_off = int_cnt2*(1<<scale2); 7656 movdqu(vec, Address(str2, tail_off-16)); 7657 psrldq(vec, 16-tail_off); 7658 } 7659 } 7660 } else { // not constant substring 7661 cmpl(cnt2, stride); 7662 jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough 7663 7664 // We can read beyond string if srt+16 does not cross page boundary 7665 // since heaps are aligned and mapped by pages. 7666 assert(os::vm_page_size() < (int)G, "default page should be small"); 7667 movl(result, str2); // We need only low 32 bits 7668 andl(result, (os::vm_page_size()-1)); 7669 cmpl(result, (os::vm_page_size()-16)); 7670 jccb(Assembler::belowEqual, CHECK_STR); 7671 7672 // Move small strings to stack to allow load 16 bytes into vec. 7673 subptr(rsp, 16); 7674 int stk_offset = wordSize-(1<<scale2); 7675 push(cnt2); 7676 7677 bind(COPY_SUBSTR); 7678 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL) { 7679 load_unsigned_byte(result, Address(str2, cnt2, scale2, -1)); 7680 movb(Address(rsp, cnt2, scale2, stk_offset), result); 7681 } else if (ae == StrIntrinsicNode::UU) { 7682 load_unsigned_short(result, Address(str2, cnt2, scale2, -2)); 7683 movw(Address(rsp, cnt2, scale2, stk_offset), result); 7684 } 7685 decrement(cnt2); 7686 jccb(Assembler::notZero, COPY_SUBSTR); 7687 7688 pop(cnt2); 7689 movptr(str2, rsp); // New substring address 7690 } // non constant 7691 7692 bind(CHECK_STR); 7693 cmpl(cnt1, stride); 7694 jccb(Assembler::aboveEqual, BIG_STRINGS); 7695 7696 // Check cross page boundary. 7697 movl(result, str1); // We need only low 32 bits 7698 andl(result, (os::vm_page_size()-1)); 7699 cmpl(result, (os::vm_page_size()-16)); 7700 jccb(Assembler::belowEqual, BIG_STRINGS); 7701 7702 subptr(rsp, 16); 7703 int stk_offset = -(1<<scale1); 7704 if (int_cnt2 < 0) { // not constant 7705 push(cnt2); 7706 stk_offset += wordSize; 7707 } 7708 movl(cnt2, cnt1); 7709 7710 bind(COPY_STR); 7711 if (ae == StrIntrinsicNode::LL) { 7712 load_unsigned_byte(result, Address(str1, cnt2, scale1, -1)); 7713 movb(Address(rsp, cnt2, scale1, stk_offset), result); 7714 } else { 7715 load_unsigned_short(result, Address(str1, cnt2, scale1, -2)); 7716 movw(Address(rsp, cnt2, scale1, stk_offset), result); 7717 } 7718 decrement(cnt2); 7719 jccb(Assembler::notZero, COPY_STR); 7720 7721 if (int_cnt2 < 0) { // not constant 7722 pop(cnt2); 7723 } 7724 movptr(str1, rsp); // New string address 7725 7726 bind(BIG_STRINGS); 7727 // Load substring. 7728 if (int_cnt2 < 0) { // -1 7729 if (ae == StrIntrinsicNode::UL) { 7730 pmovzxbw(vec, Address(str2, 0)); 7731 } else { 7732 movdqu(vec, Address(str2, 0)); 7733 } 7734 push(cnt2); // substr count 7735 push(str2); // substr addr 7736 push(str1); // string addr 7737 } else { 7738 // Small (< 8 chars) constant substrings are loaded already. 7739 movl(cnt2, int_cnt2); 7740 } 7741 push(tmp); // original SP 7742 7743 } // Finished loading 7744 7745 //======================================================== 7746 // Start search 7747 // 7748 7749 movptr(result, str1); // string addr 7750 7751 if (int_cnt2 < 0) { // Only for non constant substring 7752 jmpb(SCAN_TO_SUBSTR); 7753 7754 // SP saved at sp+0 7755 // String saved at sp+1*wordSize 7756 // Substr saved at sp+2*wordSize 7757 // Substr count saved at sp+3*wordSize 7758 7759 // Reload substr for rescan, this code 7760 // is executed only for large substrings (> 8 chars) 7761 bind(RELOAD_SUBSTR); 7762 movptr(str2, Address(rsp, 2*wordSize)); 7763 movl(cnt2, Address(rsp, 3*wordSize)); 7764 if (ae == StrIntrinsicNode::UL) { 7765 pmovzxbw(vec, Address(str2, 0)); 7766 } else { 7767 movdqu(vec, Address(str2, 0)); 7768 } 7769 // We came here after the beginning of the substring was 7770 // matched but the rest of it was not so we need to search 7771 // again. Start from the next element after the previous match. 7772 subptr(str1, result); // Restore counter 7773 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 7774 shrl(str1, 1); 7775 } 7776 addl(cnt1, str1); 7777 decrementl(cnt1); // Shift to next element 7778 cmpl(cnt1, cnt2); 7779 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7780 7781 addptr(result, (1<<scale1)); 7782 } // non constant 7783 7784 // Scan string for start of substr in 16-byte vectors 7785 bind(SCAN_TO_SUBSTR); 7786 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 7787 pcmpestri(vec, Address(result, 0), mode); 7788 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1 7789 subl(cnt1, stride); 7790 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string 7791 cmpl(cnt1, cnt2); 7792 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7793 addptr(result, 16); 7794 7795 bind(ADJUST_STR); 7796 cmpl(cnt1, stride); // Do not read beyond string 7797 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR); 7798 // Back-up string to avoid reading beyond string. 7799 lea(result, Address(result, cnt1, scale1, -16)); 7800 movl(cnt1, stride); 7801 jmpb(SCAN_TO_SUBSTR); 7802 7803 // Found a potential substr 7804 bind(FOUND_CANDIDATE); 7805 // After pcmpestri tmp(rcx) contains matched element index 7806 7807 // Make sure string is still long enough 7808 subl(cnt1, tmp); 7809 cmpl(cnt1, cnt2); 7810 jccb(Assembler::greaterEqual, FOUND_SUBSTR); 7811 // Left less then substring. 7812 7813 bind(RET_NOT_FOUND); 7814 movl(result, -1); 7815 jmpb(CLEANUP); 7816 7817 bind(FOUND_SUBSTR); 7818 // Compute start addr of substr 7819 lea(result, Address(result, tmp, scale1)); 7820 if (int_cnt2 > 0) { // Constant substring 7821 // Repeat search for small substring (< 8 chars) 7822 // from new point without reloading substring. 7823 // Have to check that we don't read beyond string. 7824 cmpl(tmp, stride-int_cnt2); 7825 jccb(Assembler::greater, ADJUST_STR); 7826 // Fall through if matched whole substring. 7827 } else { // non constant 7828 assert(int_cnt2 == -1, "should be != 0"); 7829 7830 addl(tmp, cnt2); 7831 // Found result if we matched whole substring. 7832 cmpl(tmp, stride); 7833 jccb(Assembler::lessEqual, RET_FOUND); 7834 7835 // Repeat search for small substring (<= 8 chars) 7836 // from new point 'str1' without reloading substring. 7837 cmpl(cnt2, stride); 7838 // Have to check that we don't read beyond string. 7839 jccb(Assembler::lessEqual, ADJUST_STR); 7840 7841 Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG; 7842 // Compare the rest of substring (> 8 chars). 7843 movptr(str1, result); 7844 7845 cmpl(tmp, cnt2); 7846 // First 8 chars are already matched. 7847 jccb(Assembler::equal, CHECK_NEXT); 7848 7849 bind(SCAN_SUBSTR); 7850 pcmpestri(vec, Address(str1, 0), mode); 7851 // Need to reload strings pointers if not matched whole vector 7852 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0 7853 7854 bind(CHECK_NEXT); 7855 subl(cnt2, stride); 7856 jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring 7857 addptr(str1, 16); 7858 if (ae == StrIntrinsicNode::UL) { 7859 addptr(str2, 8); 7860 } else { 7861 addptr(str2, 16); 7862 } 7863 subl(cnt1, stride); 7864 cmpl(cnt2, stride); // Do not read beyond substring 7865 jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR); 7866 // Back-up strings to avoid reading beyond substring. 7867 7868 if (ae == StrIntrinsicNode::UL) { 7869 lea(str2, Address(str2, cnt2, scale2, -8)); 7870 lea(str1, Address(str1, cnt2, scale1, -16)); 7871 } else { 7872 lea(str2, Address(str2, cnt2, scale2, -16)); 7873 lea(str1, Address(str1, cnt2, scale1, -16)); 7874 } 7875 subl(cnt1, cnt2); 7876 movl(cnt2, stride); 7877 addl(cnt1, stride); 7878 bind(CONT_SCAN_SUBSTR); 7879 if (ae == StrIntrinsicNode::UL) { 7880 pmovzxbw(vec, Address(str2, 0)); 7881 } else { 7882 movdqu(vec, Address(str2, 0)); 7883 } 7884 jmpb(SCAN_SUBSTR); 7885 7886 bind(RET_FOUND_LONG); 7887 movptr(str1, Address(rsp, wordSize)); 7888 } // non constant 7889 7890 bind(RET_FOUND); 7891 // Compute substr offset 7892 subptr(result, str1); 7893 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 7894 shrl(result, 1); // index 7895 } 7896 bind(CLEANUP); 7897 pop(rsp); // restore SP 7898 7899} // string_indexof 7900 7901void MacroAssembler::string_indexof_char(Register str1, Register cnt1, Register ch, Register result, 7902 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp) { 7903 ShortBranchVerifier sbv(this); 7904 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 7905 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 7906 7907 int stride = 8; 7908 7909 Label FOUND_CHAR, SCAN_TO_CHAR, SCAN_TO_CHAR_LOOP, 7910 SCAN_TO_8_CHAR, SCAN_TO_8_CHAR_LOOP, SCAN_TO_16_CHAR_LOOP, 7911 RET_NOT_FOUND, SCAN_TO_8_CHAR_INIT, 7912 FOUND_SEQ_CHAR, DONE_LABEL; 7913 7914 movptr(result, str1); 7915 if (UseAVX >= 2) { 7916 cmpl(cnt1, stride); 7917 jccb(Assembler::less, SCAN_TO_CHAR_LOOP); 7918 cmpl(cnt1, 2*stride); 7919 jccb(Assembler::less, SCAN_TO_8_CHAR_INIT); 7920 movdl(vec1, ch); 7921 vpbroadcastw(vec1, vec1); 7922 vpxor(vec2, vec2); 7923 movl(tmp, cnt1); 7924 andl(tmp, 0xFFFFFFF0); //vector count (in chars) 7925 andl(cnt1,0x0000000F); //tail count (in chars) 7926 7927 bind(SCAN_TO_16_CHAR_LOOP); 7928 vmovdqu(vec3, Address(result, 0)); 7929 vpcmpeqw(vec3, vec3, vec1, 1); 7930 vptest(vec2, vec3); 7931 jcc(Assembler::carryClear, FOUND_CHAR); 7932 addptr(result, 32); 7933 subl(tmp, 2*stride); 7934 jccb(Assembler::notZero, SCAN_TO_16_CHAR_LOOP); 7935 jmp(SCAN_TO_8_CHAR); 7936 bind(SCAN_TO_8_CHAR_INIT); 7937 movdl(vec1, ch); 7938 pshuflw(vec1, vec1, 0x00); 7939 pshufd(vec1, vec1, 0); 7940 pxor(vec2, vec2); 7941 } 7942 bind(SCAN_TO_8_CHAR); 7943 cmpl(cnt1, stride); 7944 if (UseAVX >= 2) { 7945 jccb(Assembler::less, SCAN_TO_CHAR); 7946 } else { 7947 jccb(Assembler::less, SCAN_TO_CHAR_LOOP); 7948 movdl(vec1, ch); 7949 pshuflw(vec1, vec1, 0x00); 7950 pshufd(vec1, vec1, 0); 7951 pxor(vec2, vec2); 7952 } 7953 movl(tmp, cnt1); 7954 andl(tmp, 0xFFFFFFF8); //vector count (in chars) 7955 andl(cnt1,0x00000007); //tail count (in chars) 7956 7957 bind(SCAN_TO_8_CHAR_LOOP); 7958 movdqu(vec3, Address(result, 0)); 7959 pcmpeqw(vec3, vec1); 7960 ptest(vec2, vec3); 7961 jcc(Assembler::carryClear, FOUND_CHAR); 7962 addptr(result, 16); 7963 subl(tmp, stride); 7964 jccb(Assembler::notZero, SCAN_TO_8_CHAR_LOOP); 7965 bind(SCAN_TO_CHAR); 7966 testl(cnt1, cnt1); 7967 jcc(Assembler::zero, RET_NOT_FOUND); 7968 bind(SCAN_TO_CHAR_LOOP); 7969 load_unsigned_short(tmp, Address(result, 0)); 7970 cmpl(ch, tmp); 7971 jccb(Assembler::equal, FOUND_SEQ_CHAR); 7972 addptr(result, 2); 7973 subl(cnt1, 1); 7974 jccb(Assembler::zero, RET_NOT_FOUND); 7975 jmp(SCAN_TO_CHAR_LOOP); 7976 7977 bind(RET_NOT_FOUND); 7978 movl(result, -1); 7979 jmpb(DONE_LABEL); 7980 7981 bind(FOUND_CHAR); 7982 if (UseAVX >= 2) { 7983 vpmovmskb(tmp, vec3); 7984 } else { 7985 pmovmskb(tmp, vec3); 7986 } 7987 bsfl(ch, tmp); 7988 addl(result, ch); 7989 7990 bind(FOUND_SEQ_CHAR); 7991 subptr(result, str1); 7992 shrl(result, 1); 7993 7994 bind(DONE_LABEL); 7995} // string_indexof_char 7996 7997// helper function for string_compare 7998void MacroAssembler::load_next_elements(Register elem1, Register elem2, Register str1, Register str2, 7999 Address::ScaleFactor scale, Address::ScaleFactor scale1, 8000 Address::ScaleFactor scale2, Register index, int ae) { 8001 if (ae == StrIntrinsicNode::LL) { 8002 load_unsigned_byte(elem1, Address(str1, index, scale, 0)); 8003 load_unsigned_byte(elem2, Address(str2, index, scale, 0)); 8004 } else if (ae == StrIntrinsicNode::UU) { 8005 load_unsigned_short(elem1, Address(str1, index, scale, 0)); 8006 load_unsigned_short(elem2, Address(str2, index, scale, 0)); 8007 } else { 8008 load_unsigned_byte(elem1, Address(str1, index, scale1, 0)); 8009 load_unsigned_short(elem2, Address(str2, index, scale2, 0)); 8010 } 8011} 8012 8013// Compare strings, used for char[] and byte[]. 8014void MacroAssembler::string_compare(Register str1, Register str2, 8015 Register cnt1, Register cnt2, Register result, 8016 XMMRegister vec1, int ae) { 8017 ShortBranchVerifier sbv(this); 8018 Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL; 8019 int stride, stride2, adr_stride, adr_stride1, adr_stride2; 8020 Address::ScaleFactor scale, scale1, scale2; 8021 8022 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) { 8023 shrl(cnt2, 1); 8024 } 8025 // Compute the minimum of the string lengths and the 8026 // difference of the string lengths (stack). 8027 // Do the conditional move stuff 8028 movl(result, cnt1); 8029 subl(cnt1, cnt2); 8030 push(cnt1); 8031 cmov32(Assembler::lessEqual, cnt2, result); 8032 8033 // Is the minimum length zero? 8034 testl(cnt2, cnt2); 8035 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 8036 if (ae == StrIntrinsicNode::LL) { 8037 // Load first bytes 8038 load_unsigned_byte(result, Address(str1, 0)); 8039 load_unsigned_byte(cnt1, Address(str2, 0)); 8040 } else if (ae == StrIntrinsicNode::UU) { 8041 // Load first characters 8042 load_unsigned_short(result, Address(str1, 0)); 8043 load_unsigned_short(cnt1, Address(str2, 0)); 8044 } else { 8045 load_unsigned_byte(result, Address(str1, 0)); 8046 load_unsigned_short(cnt1, Address(str2, 0)); 8047 } 8048 subl(result, cnt1); 8049 jcc(Assembler::notZero, POP_LABEL); 8050 8051 if (ae == StrIntrinsicNode::UU) { 8052 // Divide length by 2 to get number of chars 8053 shrl(cnt2, 1); 8054 } 8055 cmpl(cnt2, 1); 8056 jcc(Assembler::equal, LENGTH_DIFF_LABEL); 8057 8058 // Check if the strings start at the same location and setup scale and stride 8059 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8060 cmpptr(str1, str2); 8061 jcc(Assembler::equal, LENGTH_DIFF_LABEL); 8062 if (ae == StrIntrinsicNode::LL) { 8063 scale = Address::times_1; 8064 stride = 16; 8065 } else { 8066 scale = Address::times_2; 8067 stride = 8; 8068 } 8069 } else { 8070 scale = Address::no_scale; // not used 8071 scale1 = Address::times_1; 8072 scale2 = Address::times_2; 8073 stride = 8; 8074 } 8075 8076 if (UseAVX >= 2 && UseSSE42Intrinsics) { 8077 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 8078 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR; 8079 Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR; 8080 Label COMPARE_TAIL_LONG; 8081 int pcmpmask = 0x19; 8082 if (ae == StrIntrinsicNode::LL) { 8083 pcmpmask &= ~0x01; 8084 } 8085 8086 // Setup to compare 16-chars (32-bytes) vectors, 8087 // start from first character again because it has aligned address. 8088 if (ae == StrIntrinsicNode::LL) { 8089 stride2 = 32; 8090 } else { 8091 stride2 = 16; 8092 } 8093 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8094 adr_stride = stride << scale; 8095 } else { 8096 adr_stride1 = 8; //stride << scale1; 8097 adr_stride2 = 16; //stride << scale2; 8098 } 8099 8100 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri"); 8101 // rax and rdx are used by pcmpestri as elements counters 8102 movl(result, cnt2); 8103 andl(cnt2, ~(stride2-1)); // cnt2 holds the vector count 8104 jcc(Assembler::zero, COMPARE_TAIL_LONG); 8105 8106 // fast path : compare first 2 8-char vectors. 8107 bind(COMPARE_16_CHARS); 8108 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8109 movdqu(vec1, Address(str1, 0)); 8110 } else { 8111 pmovzxbw(vec1, Address(str1, 0)); 8112 } 8113 pcmpestri(vec1, Address(str2, 0), pcmpmask); 8114 jccb(Assembler::below, COMPARE_INDEX_CHAR); 8115 8116 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8117 movdqu(vec1, Address(str1, adr_stride)); 8118 pcmpestri(vec1, Address(str2, adr_stride), pcmpmask); 8119 } else { 8120 pmovzxbw(vec1, Address(str1, adr_stride1)); 8121 pcmpestri(vec1, Address(str2, adr_stride2), pcmpmask); 8122 } 8123 jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS); 8124 addl(cnt1, stride); 8125 8126 // Compare the characters at index in cnt1 8127 bind(COMPARE_INDEX_CHAR); // cnt1 has the offset of the mismatching character 8128 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae); 8129 subl(result, cnt2); 8130 jmp(POP_LABEL); 8131 8132 // Setup the registers to start vector comparison loop 8133 bind(COMPARE_WIDE_VECTORS); 8134 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8135 lea(str1, Address(str1, result, scale)); 8136 lea(str2, Address(str2, result, scale)); 8137 } else { 8138 lea(str1, Address(str1, result, scale1)); 8139 lea(str2, Address(str2, result, scale2)); 8140 } 8141 subl(result, stride2); 8142 subl(cnt2, stride2); 8143 jccb(Assembler::zero, COMPARE_WIDE_TAIL); 8144 negptr(result); 8145 8146 // In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest) 8147 bind(COMPARE_WIDE_VECTORS_LOOP); 8148 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8149 vmovdqu(vec1, Address(str1, result, scale)); 8150 vpxor(vec1, Address(str2, result, scale)); 8151 } else { 8152 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_256bit); 8153 vpxor(vec1, Address(str2, result, scale2)); 8154 } 8155 vptest(vec1, vec1); 8156 jccb(Assembler::notZero, VECTOR_NOT_EQUAL); 8157 addptr(result, stride2); 8158 subl(cnt2, stride2); 8159 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP); 8160 // clean upper bits of YMM registers 8161 vpxor(vec1, vec1); 8162 8163 // compare wide vectors tail 8164 bind(COMPARE_WIDE_TAIL); 8165 testptr(result, result); 8166 jccb(Assembler::zero, LENGTH_DIFF_LABEL); 8167 8168 movl(result, stride2); 8169 movl(cnt2, result); 8170 negptr(result); 8171 jmpb(COMPARE_WIDE_VECTORS_LOOP); 8172 8173 // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors. 8174 bind(VECTOR_NOT_EQUAL); 8175 // clean upper bits of YMM registers 8176 vpxor(vec1, vec1); 8177 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8178 lea(str1, Address(str1, result, scale)); 8179 lea(str2, Address(str2, result, scale)); 8180 } else { 8181 lea(str1, Address(str1, result, scale1)); 8182 lea(str2, Address(str2, result, scale2)); 8183 } 8184 jmp(COMPARE_16_CHARS); 8185 8186 // Compare tail chars, length between 1 to 15 chars 8187 bind(COMPARE_TAIL_LONG); 8188 movl(cnt2, result); 8189 cmpl(cnt2, stride); 8190 jccb(Assembler::less, COMPARE_SMALL_STR); 8191 8192 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8193 movdqu(vec1, Address(str1, 0)); 8194 } else { 8195 pmovzxbw(vec1, Address(str1, 0)); 8196 } 8197 pcmpestri(vec1, Address(str2, 0), pcmpmask); 8198 jcc(Assembler::below, COMPARE_INDEX_CHAR); 8199 subptr(cnt2, stride); 8200 jccb(Assembler::zero, LENGTH_DIFF_LABEL); 8201 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8202 lea(str1, Address(str1, result, scale)); 8203 lea(str2, Address(str2, result, scale)); 8204 } else { 8205 lea(str1, Address(str1, result, scale1)); 8206 lea(str2, Address(str2, result, scale2)); 8207 } 8208 negptr(cnt2); 8209 jmpb(WHILE_HEAD_LABEL); 8210 8211 bind(COMPARE_SMALL_STR); 8212 } else if (UseSSE42Intrinsics) { 8213 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 8214 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL; 8215 int pcmpmask = 0x19; 8216 // Setup to compare 8-char (16-byte) vectors, 8217 // start from first character again because it has aligned address. 8218 movl(result, cnt2); 8219 andl(cnt2, ~(stride - 1)); // cnt2 holds the vector count 8220 if (ae == StrIntrinsicNode::LL) { 8221 pcmpmask &= ~0x01; 8222 } 8223 jccb(Assembler::zero, COMPARE_TAIL); 8224 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8225 lea(str1, Address(str1, result, scale)); 8226 lea(str2, Address(str2, result, scale)); 8227 } else { 8228 lea(str1, Address(str1, result, scale1)); 8229 lea(str2, Address(str2, result, scale2)); 8230 } 8231 negptr(result); 8232 8233 // pcmpestri 8234 // inputs: 8235 // vec1- substring 8236 // rax - negative string length (elements count) 8237 // mem - scanned string 8238 // rdx - string length (elements count) 8239 // pcmpmask - cmp mode: 11000 (string compare with negated result) 8240 // + 00 (unsigned bytes) or + 01 (unsigned shorts) 8241 // outputs: 8242 // rcx - first mismatched element index 8243 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri"); 8244 8245 bind(COMPARE_WIDE_VECTORS); 8246 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8247 movdqu(vec1, Address(str1, result, scale)); 8248 pcmpestri(vec1, Address(str2, result, scale), pcmpmask); 8249 } else { 8250 pmovzxbw(vec1, Address(str1, result, scale1)); 8251 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask); 8252 } 8253 // After pcmpestri cnt1(rcx) contains mismatched element index 8254 8255 jccb(Assembler::below, VECTOR_NOT_EQUAL); // CF==1 8256 addptr(result, stride); 8257 subptr(cnt2, stride); 8258 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS); 8259 8260 // compare wide vectors tail 8261 testptr(result, result); 8262 jccb(Assembler::zero, LENGTH_DIFF_LABEL); 8263 8264 movl(cnt2, stride); 8265 movl(result, stride); 8266 negptr(result); 8267 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8268 movdqu(vec1, Address(str1, result, scale)); 8269 pcmpestri(vec1, Address(str2, result, scale), pcmpmask); 8270 } else { 8271 pmovzxbw(vec1, Address(str1, result, scale1)); 8272 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask); 8273 } 8274 jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL); 8275 8276 // Mismatched characters in the vectors 8277 bind(VECTOR_NOT_EQUAL); 8278 addptr(cnt1, result); 8279 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae); 8280 subl(result, cnt2); 8281 jmpb(POP_LABEL); 8282 8283 bind(COMPARE_TAIL); // limit is zero 8284 movl(cnt2, result); 8285 // Fallthru to tail compare 8286 } 8287 // Shift str2 and str1 to the end of the arrays, negate min 8288 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8289 lea(str1, Address(str1, cnt2, scale)); 8290 lea(str2, Address(str2, cnt2, scale)); 8291 } else { 8292 lea(str1, Address(str1, cnt2, scale1)); 8293 lea(str2, Address(str2, cnt2, scale2)); 8294 } 8295 decrementl(cnt2); // first character was compared already 8296 negptr(cnt2); 8297 8298 // Compare the rest of the elements 8299 bind(WHILE_HEAD_LABEL); 8300 load_next_elements(result, cnt1, str1, str2, scale, scale1, scale2, cnt2, ae); 8301 subl(result, cnt1); 8302 jccb(Assembler::notZero, POP_LABEL); 8303 increment(cnt2); 8304 jccb(Assembler::notZero, WHILE_HEAD_LABEL); 8305 8306 // Strings are equal up to min length. Return the length difference. 8307 bind(LENGTH_DIFF_LABEL); 8308 pop(result); 8309 if (ae == StrIntrinsicNode::UU) { 8310 // Divide diff by 2 to get number of chars 8311 sarl(result, 1); 8312 } 8313 jmpb(DONE_LABEL); 8314 8315 // Discard the stored length difference 8316 bind(POP_LABEL); 8317 pop(cnt1); 8318 8319 // That's it 8320 bind(DONE_LABEL); 8321 if(ae == StrIntrinsicNode::UL) { 8322 negl(result); 8323 } 8324} 8325 8326// Search for Non-ASCII character (Negative byte value) in a byte array, 8327// return true if it has any and false otherwise. 8328void MacroAssembler::has_negatives(Register ary1, Register len, 8329 Register result, Register tmp1, 8330 XMMRegister vec1, XMMRegister vec2) { 8331 8332 // rsi: byte array 8333 // rcx: len 8334 // rax: result 8335 ShortBranchVerifier sbv(this); 8336 assert_different_registers(ary1, len, result, tmp1); 8337 assert_different_registers(vec1, vec2); 8338 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_CHAR, COMPARE_VECTORS, COMPARE_BYTE; 8339 8340 // len == 0 8341 testl(len, len); 8342 jcc(Assembler::zero, FALSE_LABEL); 8343 8344 movl(result, len); // copy 8345 8346 if (UseAVX >= 2 && UseSSE >= 2) { 8347 // With AVX2, use 32-byte vector compare 8348 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8349 8350 // Compare 32-byte vectors 8351 andl(result, 0x0000001f); // tail count (in bytes) 8352 andl(len, 0xffffffe0); // vector count (in bytes) 8353 jccb(Assembler::zero, COMPARE_TAIL); 8354 8355 lea(ary1, Address(ary1, len, Address::times_1)); 8356 negptr(len); 8357 8358 movl(tmp1, 0x80808080); // create mask to test for Unicode chars in vector 8359 movdl(vec2, tmp1); 8360 vpbroadcastd(vec2, vec2); 8361 8362 bind(COMPARE_WIDE_VECTORS); 8363 vmovdqu(vec1, Address(ary1, len, Address::times_1)); 8364 vptest(vec1, vec2); 8365 jccb(Assembler::notZero, TRUE_LABEL); 8366 addptr(len, 32); 8367 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8368 8369 testl(result, result); 8370 jccb(Assembler::zero, FALSE_LABEL); 8371 8372 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32)); 8373 vptest(vec1, vec2); 8374 jccb(Assembler::notZero, TRUE_LABEL); 8375 jmpb(FALSE_LABEL); 8376 8377 bind(COMPARE_TAIL); // len is zero 8378 movl(len, result); 8379 // Fallthru to tail compare 8380 } else if (UseSSE42Intrinsics) { 8381 assert(UseSSE >= 4, "SSE4 must be for SSE4.2 intrinsics to be available"); 8382 // With SSE4.2, use double quad vector compare 8383 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8384 8385 // Compare 16-byte vectors 8386 andl(result, 0x0000000f); // tail count (in bytes) 8387 andl(len, 0xfffffff0); // vector count (in bytes) 8388 jccb(Assembler::zero, COMPARE_TAIL); 8389 8390 lea(ary1, Address(ary1, len, Address::times_1)); 8391 negptr(len); 8392 8393 movl(tmp1, 0x80808080); 8394 movdl(vec2, tmp1); 8395 pshufd(vec2, vec2, 0); 8396 8397 bind(COMPARE_WIDE_VECTORS); 8398 movdqu(vec1, Address(ary1, len, Address::times_1)); 8399 ptest(vec1, vec2); 8400 jccb(Assembler::notZero, TRUE_LABEL); 8401 addptr(len, 16); 8402 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8403 8404 testl(result, result); 8405 jccb(Assembler::zero, FALSE_LABEL); 8406 8407 movdqu(vec1, Address(ary1, result, Address::times_1, -16)); 8408 ptest(vec1, vec2); 8409 jccb(Assembler::notZero, TRUE_LABEL); 8410 jmpb(FALSE_LABEL); 8411 8412 bind(COMPARE_TAIL); // len is zero 8413 movl(len, result); 8414 // Fallthru to tail compare 8415 } 8416 8417 // Compare 4-byte vectors 8418 andl(len, 0xfffffffc); // vector count (in bytes) 8419 jccb(Assembler::zero, COMPARE_CHAR); 8420 8421 lea(ary1, Address(ary1, len, Address::times_1)); 8422 negptr(len); 8423 8424 bind(COMPARE_VECTORS); 8425 movl(tmp1, Address(ary1, len, Address::times_1)); 8426 andl(tmp1, 0x80808080); 8427 jccb(Assembler::notZero, TRUE_LABEL); 8428 addptr(len, 4); 8429 jcc(Assembler::notZero, COMPARE_VECTORS); 8430 8431 // Compare trailing char (final 2 bytes), if any 8432 bind(COMPARE_CHAR); 8433 testl(result, 0x2); // tail char 8434 jccb(Assembler::zero, COMPARE_BYTE); 8435 load_unsigned_short(tmp1, Address(ary1, 0)); 8436 andl(tmp1, 0x00008080); 8437 jccb(Assembler::notZero, TRUE_LABEL); 8438 subptr(result, 2); 8439 lea(ary1, Address(ary1, 2)); 8440 8441 bind(COMPARE_BYTE); 8442 testl(result, 0x1); // tail byte 8443 jccb(Assembler::zero, FALSE_LABEL); 8444 load_unsigned_byte(tmp1, Address(ary1, 0)); 8445 andl(tmp1, 0x00000080); 8446 jccb(Assembler::notEqual, TRUE_LABEL); 8447 jmpb(FALSE_LABEL); 8448 8449 bind(TRUE_LABEL); 8450 movl(result, 1); // return true 8451 jmpb(DONE); 8452 8453 bind(FALSE_LABEL); 8454 xorl(result, result); // return false 8455 8456 // That's it 8457 bind(DONE); 8458 if (UseAVX >= 2 && UseSSE >= 2) { 8459 // clean upper bits of YMM registers 8460 vpxor(vec1, vec1); 8461 vpxor(vec2, vec2); 8462 } 8463} 8464 8465// Compare char[] or byte[] arrays aligned to 4 bytes or substrings. 8466void MacroAssembler::arrays_equals(bool is_array_equ, Register ary1, Register ary2, 8467 Register limit, Register result, Register chr, 8468 XMMRegister vec1, XMMRegister vec2, bool is_char) { 8469 ShortBranchVerifier sbv(this); 8470 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR, COMPARE_BYTE; 8471 8472 int length_offset = arrayOopDesc::length_offset_in_bytes(); 8473 int base_offset = arrayOopDesc::base_offset_in_bytes(is_char ? T_CHAR : T_BYTE); 8474 8475 if (is_array_equ) { 8476 // Check the input args 8477 cmpptr(ary1, ary2); 8478 jcc(Assembler::equal, TRUE_LABEL); 8479 8480 // Need additional checks for arrays_equals. 8481 testptr(ary1, ary1); 8482 jcc(Assembler::zero, FALSE_LABEL); 8483 testptr(ary2, ary2); 8484 jcc(Assembler::zero, FALSE_LABEL); 8485 8486 // Check the lengths 8487 movl(limit, Address(ary1, length_offset)); 8488 cmpl(limit, Address(ary2, length_offset)); 8489 jcc(Assembler::notEqual, FALSE_LABEL); 8490 } 8491 8492 // count == 0 8493 testl(limit, limit); 8494 jcc(Assembler::zero, TRUE_LABEL); 8495 8496 if (is_array_equ) { 8497 // Load array address 8498 lea(ary1, Address(ary1, base_offset)); 8499 lea(ary2, Address(ary2, base_offset)); 8500 } 8501 8502 if (is_array_equ && is_char) { 8503 // arrays_equals when used for char[]. 8504 shll(limit, 1); // byte count != 0 8505 } 8506 movl(result, limit); // copy 8507 8508 if (UseAVX >= 2) { 8509 // With AVX2, use 32-byte vector compare 8510 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8511 8512 // Compare 32-byte vectors 8513 andl(result, 0x0000001f); // tail count (in bytes) 8514 andl(limit, 0xffffffe0); // vector count (in bytes) 8515 jccb(Assembler::zero, COMPARE_TAIL); 8516 8517 lea(ary1, Address(ary1, limit, Address::times_1)); 8518 lea(ary2, Address(ary2, limit, Address::times_1)); 8519 negptr(limit); 8520 8521 bind(COMPARE_WIDE_VECTORS); 8522 vmovdqu(vec1, Address(ary1, limit, Address::times_1)); 8523 vmovdqu(vec2, Address(ary2, limit, Address::times_1)); 8524 vpxor(vec1, vec2); 8525 8526 vptest(vec1, vec1); 8527 jccb(Assembler::notZero, FALSE_LABEL); 8528 addptr(limit, 32); 8529 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8530 8531 testl(result, result); 8532 jccb(Assembler::zero, TRUE_LABEL); 8533 8534 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32)); 8535 vmovdqu(vec2, Address(ary2, result, Address::times_1, -32)); 8536 vpxor(vec1, vec2); 8537 8538 vptest(vec1, vec1); 8539 jccb(Assembler::notZero, FALSE_LABEL); 8540 jmpb(TRUE_LABEL); 8541 8542 bind(COMPARE_TAIL); // limit is zero 8543 movl(limit, result); 8544 // Fallthru to tail compare 8545 } else if (UseSSE42Intrinsics) { 8546 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 8547 // With SSE4.2, use double quad vector compare 8548 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8549 8550 // Compare 16-byte vectors 8551 andl(result, 0x0000000f); // tail count (in bytes) 8552 andl(limit, 0xfffffff0); // vector count (in bytes) 8553 jccb(Assembler::zero, COMPARE_TAIL); 8554 8555 lea(ary1, Address(ary1, limit, Address::times_1)); 8556 lea(ary2, Address(ary2, limit, Address::times_1)); 8557 negptr(limit); 8558 8559 bind(COMPARE_WIDE_VECTORS); 8560 movdqu(vec1, Address(ary1, limit, Address::times_1)); 8561 movdqu(vec2, Address(ary2, limit, Address::times_1)); 8562 pxor(vec1, vec2); 8563 8564 ptest(vec1, vec1); 8565 jccb(Assembler::notZero, FALSE_LABEL); 8566 addptr(limit, 16); 8567 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8568 8569 testl(result, result); 8570 jccb(Assembler::zero, TRUE_LABEL); 8571 8572 movdqu(vec1, Address(ary1, result, Address::times_1, -16)); 8573 movdqu(vec2, Address(ary2, result, Address::times_1, -16)); 8574 pxor(vec1, vec2); 8575 8576 ptest(vec1, vec1); 8577 jccb(Assembler::notZero, FALSE_LABEL); 8578 jmpb(TRUE_LABEL); 8579 8580 bind(COMPARE_TAIL); // limit is zero 8581 movl(limit, result); 8582 // Fallthru to tail compare 8583 } 8584 8585 // Compare 4-byte vectors 8586 andl(limit, 0xfffffffc); // vector count (in bytes) 8587 jccb(Assembler::zero, COMPARE_CHAR); 8588 8589 lea(ary1, Address(ary1, limit, Address::times_1)); 8590 lea(ary2, Address(ary2, limit, Address::times_1)); 8591 negptr(limit); 8592 8593 bind(COMPARE_VECTORS); 8594 movl(chr, Address(ary1, limit, Address::times_1)); 8595 cmpl(chr, Address(ary2, limit, Address::times_1)); 8596 jccb(Assembler::notEqual, FALSE_LABEL); 8597 addptr(limit, 4); 8598 jcc(Assembler::notZero, COMPARE_VECTORS); 8599 8600 // Compare trailing char (final 2 bytes), if any 8601 bind(COMPARE_CHAR); 8602 testl(result, 0x2); // tail char 8603 jccb(Assembler::zero, COMPARE_BYTE); 8604 load_unsigned_short(chr, Address(ary1, 0)); 8605 load_unsigned_short(limit, Address(ary2, 0)); 8606 cmpl(chr, limit); 8607 jccb(Assembler::notEqual, FALSE_LABEL); 8608 8609 if (is_array_equ && is_char) { 8610 bind(COMPARE_BYTE); 8611 } else { 8612 lea(ary1, Address(ary1, 2)); 8613 lea(ary2, Address(ary2, 2)); 8614 8615 bind(COMPARE_BYTE); 8616 testl(result, 0x1); // tail byte 8617 jccb(Assembler::zero, TRUE_LABEL); 8618 load_unsigned_byte(chr, Address(ary1, 0)); 8619 load_unsigned_byte(limit, Address(ary2, 0)); 8620 cmpl(chr, limit); 8621 jccb(Assembler::notEqual, FALSE_LABEL); 8622 } 8623 bind(TRUE_LABEL); 8624 movl(result, 1); // return true 8625 jmpb(DONE); 8626 8627 bind(FALSE_LABEL); 8628 xorl(result, result); // return false 8629 8630 // That's it 8631 bind(DONE); 8632 if (UseAVX >= 2) { 8633 // clean upper bits of YMM registers 8634 vpxor(vec1, vec1); 8635 vpxor(vec2, vec2); 8636 } 8637} 8638 8639#endif 8640 8641void MacroAssembler::generate_fill(BasicType t, bool aligned, 8642 Register to, Register value, Register count, 8643 Register rtmp, XMMRegister xtmp) { 8644 ShortBranchVerifier sbv(this); 8645 assert_different_registers(to, value, count, rtmp); 8646 Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte; 8647 Label L_fill_2_bytes, L_fill_4_bytes; 8648 8649 int shift = -1; 8650 switch (t) { 8651 case T_BYTE: 8652 shift = 2; 8653 break; 8654 case T_SHORT: 8655 shift = 1; 8656 break; 8657 case T_INT: 8658 shift = 0; 8659 break; 8660 default: ShouldNotReachHere(); 8661 } 8662 8663 if (t == T_BYTE) { 8664 andl(value, 0xff); 8665 movl(rtmp, value); 8666 shll(rtmp, 8); 8667 orl(value, rtmp); 8668 } 8669 if (t == T_SHORT) { 8670 andl(value, 0xffff); 8671 } 8672 if (t == T_BYTE || t == T_SHORT) { 8673 movl(rtmp, value); 8674 shll(rtmp, 16); 8675 orl(value, rtmp); 8676 } 8677 8678 cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element 8679 jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp 8680 if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) { 8681 // align source address at 4 bytes address boundary 8682 if (t == T_BYTE) { 8683 // One byte misalignment happens only for byte arrays 8684 testptr(to, 1); 8685 jccb(Assembler::zero, L_skip_align1); 8686 movb(Address(to, 0), value); 8687 increment(to); 8688 decrement(count); 8689 BIND(L_skip_align1); 8690 } 8691 // Two bytes misalignment happens only for byte and short (char) arrays 8692 testptr(to, 2); 8693 jccb(Assembler::zero, L_skip_align2); 8694 movw(Address(to, 0), value); 8695 addptr(to, 2); 8696 subl(count, 1<<(shift-1)); 8697 BIND(L_skip_align2); 8698 } 8699 if (UseSSE < 2) { 8700 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes; 8701 // Fill 32-byte chunks 8702 subl(count, 8 << shift); 8703 jcc(Assembler::less, L_check_fill_8_bytes); 8704 align(16); 8705 8706 BIND(L_fill_32_bytes_loop); 8707 8708 for (int i = 0; i < 32; i += 4) { 8709 movl(Address(to, i), value); 8710 } 8711 8712 addptr(to, 32); 8713 subl(count, 8 << shift); 8714 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop); 8715 BIND(L_check_fill_8_bytes); 8716 addl(count, 8 << shift); 8717 jccb(Assembler::zero, L_exit); 8718 jmpb(L_fill_8_bytes); 8719 8720 // 8721 // length is too short, just fill qwords 8722 // 8723 BIND(L_fill_8_bytes_loop); 8724 movl(Address(to, 0), value); 8725 movl(Address(to, 4), value); 8726 addptr(to, 8); 8727 BIND(L_fill_8_bytes); 8728 subl(count, 1 << (shift + 1)); 8729 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop); 8730 // fall through to fill 4 bytes 8731 } else { 8732 Label L_fill_32_bytes; 8733 if (!UseUnalignedLoadStores) { 8734 // align to 8 bytes, we know we are 4 byte aligned to start 8735 testptr(to, 4); 8736 jccb(Assembler::zero, L_fill_32_bytes); 8737 movl(Address(to, 0), value); 8738 addptr(to, 4); 8739 subl(count, 1<<shift); 8740 } 8741 BIND(L_fill_32_bytes); 8742 { 8743 assert( UseSSE >= 2, "supported cpu only" ); 8744 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes; 8745 if (UseAVX > 2) { 8746 movl(rtmp, 0xffff); 8747 kmovwl(k1, rtmp); 8748 } 8749 movdl(xtmp, value); 8750 if (UseAVX > 2 && UseUnalignedLoadStores) { 8751 // Fill 64-byte chunks 8752 Label L_fill_64_bytes_loop, L_check_fill_32_bytes; 8753 evpbroadcastd(xtmp, xtmp, Assembler::AVX_512bit); 8754 8755 subl(count, 16 << shift); 8756 jcc(Assembler::less, L_check_fill_32_bytes); 8757 align(16); 8758 8759 BIND(L_fill_64_bytes_loop); 8760 evmovdqul(Address(to, 0), xtmp, Assembler::AVX_512bit); 8761 addptr(to, 64); 8762 subl(count, 16 << shift); 8763 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop); 8764 8765 BIND(L_check_fill_32_bytes); 8766 addl(count, 8 << shift); 8767 jccb(Assembler::less, L_check_fill_8_bytes); 8768 vmovdqu(Address(to, 0), xtmp); 8769 addptr(to, 32); 8770 subl(count, 8 << shift); 8771 8772 BIND(L_check_fill_8_bytes); 8773 } else if (UseAVX == 2 && UseUnalignedLoadStores) { 8774 // Fill 64-byte chunks 8775 Label L_fill_64_bytes_loop, L_check_fill_32_bytes; 8776 vpbroadcastd(xtmp, xtmp); 8777 8778 subl(count, 16 << shift); 8779 jcc(Assembler::less, L_check_fill_32_bytes); 8780 align(16); 8781 8782 BIND(L_fill_64_bytes_loop); 8783 vmovdqu(Address(to, 0), xtmp); 8784 vmovdqu(Address(to, 32), xtmp); 8785 addptr(to, 64); 8786 subl(count, 16 << shift); 8787 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop); 8788 8789 BIND(L_check_fill_32_bytes); 8790 addl(count, 8 << shift); 8791 jccb(Assembler::less, L_check_fill_8_bytes); 8792 vmovdqu(Address(to, 0), xtmp); 8793 addptr(to, 32); 8794 subl(count, 8 << shift); 8795 8796 BIND(L_check_fill_8_bytes); 8797 // clean upper bits of YMM registers 8798 movdl(xtmp, value); 8799 pshufd(xtmp, xtmp, 0); 8800 } else { 8801 // Fill 32-byte chunks 8802 pshufd(xtmp, xtmp, 0); 8803 8804 subl(count, 8 << shift); 8805 jcc(Assembler::less, L_check_fill_8_bytes); 8806 align(16); 8807 8808 BIND(L_fill_32_bytes_loop); 8809 8810 if (UseUnalignedLoadStores) { 8811 movdqu(Address(to, 0), xtmp); 8812 movdqu(Address(to, 16), xtmp); 8813 } else { 8814 movq(Address(to, 0), xtmp); 8815 movq(Address(to, 8), xtmp); 8816 movq(Address(to, 16), xtmp); 8817 movq(Address(to, 24), xtmp); 8818 } 8819 8820 addptr(to, 32); 8821 subl(count, 8 << shift); 8822 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop); 8823 8824 BIND(L_check_fill_8_bytes); 8825 } 8826 addl(count, 8 << shift); 8827 jccb(Assembler::zero, L_exit); 8828 jmpb(L_fill_8_bytes); 8829 8830 // 8831 // length is too short, just fill qwords 8832 // 8833 BIND(L_fill_8_bytes_loop); 8834 movq(Address(to, 0), xtmp); 8835 addptr(to, 8); 8836 BIND(L_fill_8_bytes); 8837 subl(count, 1 << (shift + 1)); 8838 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop); 8839 } 8840 } 8841 // fill trailing 4 bytes 8842 BIND(L_fill_4_bytes); 8843 testl(count, 1<<shift); 8844 jccb(Assembler::zero, L_fill_2_bytes); 8845 movl(Address(to, 0), value); 8846 if (t == T_BYTE || t == T_SHORT) { 8847 addptr(to, 4); 8848 BIND(L_fill_2_bytes); 8849 // fill trailing 2 bytes 8850 testl(count, 1<<(shift-1)); 8851 jccb(Assembler::zero, L_fill_byte); 8852 movw(Address(to, 0), value); 8853 if (t == T_BYTE) { 8854 addptr(to, 2); 8855 BIND(L_fill_byte); 8856 // fill trailing byte 8857 testl(count, 1); 8858 jccb(Assembler::zero, L_exit); 8859 movb(Address(to, 0), value); 8860 } else { 8861 BIND(L_fill_byte); 8862 } 8863 } else { 8864 BIND(L_fill_2_bytes); 8865 } 8866 BIND(L_exit); 8867} 8868 8869// encode char[] to byte[] in ISO_8859_1 8870void MacroAssembler::encode_iso_array(Register src, Register dst, Register len, 8871 XMMRegister tmp1Reg, XMMRegister tmp2Reg, 8872 XMMRegister tmp3Reg, XMMRegister tmp4Reg, 8873 Register tmp5, Register result) { 8874 // rsi: src 8875 // rdi: dst 8876 // rdx: len 8877 // rcx: tmp5 8878 // rax: result 8879 ShortBranchVerifier sbv(this); 8880 assert_different_registers(src, dst, len, tmp5, result); 8881 Label L_done, L_copy_1_char, L_copy_1_char_exit; 8882 8883 // set result 8884 xorl(result, result); 8885 // check for zero length 8886 testl(len, len); 8887 jcc(Assembler::zero, L_done); 8888 movl(result, len); 8889 8890 // Setup pointers 8891 lea(src, Address(src, len, Address::times_2)); // char[] 8892 lea(dst, Address(dst, len, Address::times_1)); // byte[] 8893 negptr(len); 8894 8895 if (UseSSE42Intrinsics || UseAVX >= 2) { 8896 assert(UseSSE42Intrinsics ? UseSSE >= 4 : true, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 8897 Label L_chars_8_check, L_copy_8_chars, L_copy_8_chars_exit; 8898 Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit; 8899 8900 if (UseAVX >= 2) { 8901 Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit; 8902 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector 8903 movdl(tmp1Reg, tmp5); 8904 vpbroadcastd(tmp1Reg, tmp1Reg); 8905 jmpb(L_chars_32_check); 8906 8907 bind(L_copy_32_chars); 8908 vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64)); 8909 vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32)); 8910 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1); 8911 vptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 8912 jccb(Assembler::notZero, L_copy_32_chars_exit); 8913 vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1); 8914 vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector_len */ 1); 8915 vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg); 8916 8917 bind(L_chars_32_check); 8918 addptr(len, 32); 8919 jccb(Assembler::lessEqual, L_copy_32_chars); 8920 8921 bind(L_copy_32_chars_exit); 8922 subptr(len, 16); 8923 jccb(Assembler::greater, L_copy_16_chars_exit); 8924 8925 } else if (UseSSE42Intrinsics) { 8926 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector 8927 movdl(tmp1Reg, tmp5); 8928 pshufd(tmp1Reg, tmp1Reg, 0); 8929 jmpb(L_chars_16_check); 8930 } 8931 8932 bind(L_copy_16_chars); 8933 if (UseAVX >= 2) { 8934 vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32)); 8935 vptest(tmp2Reg, tmp1Reg); 8936 jccb(Assembler::notZero, L_copy_16_chars_exit); 8937 vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector_len */ 1); 8938 vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector_len */ 1); 8939 } else { 8940 if (UseAVX > 0) { 8941 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32)); 8942 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16)); 8943 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 0); 8944 } else { 8945 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32)); 8946 por(tmp2Reg, tmp3Reg); 8947 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16)); 8948 por(tmp2Reg, tmp4Reg); 8949 } 8950 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 8951 jccb(Assembler::notZero, L_copy_16_chars_exit); 8952 packuswb(tmp3Reg, tmp4Reg); 8953 } 8954 movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg); 8955 8956 bind(L_chars_16_check); 8957 addptr(len, 16); 8958 jccb(Assembler::lessEqual, L_copy_16_chars); 8959 8960 bind(L_copy_16_chars_exit); 8961 if (UseAVX >= 2) { 8962 // clean upper bits of YMM registers 8963 vpxor(tmp2Reg, tmp2Reg); 8964 vpxor(tmp3Reg, tmp3Reg); 8965 vpxor(tmp4Reg, tmp4Reg); 8966 movdl(tmp1Reg, tmp5); 8967 pshufd(tmp1Reg, tmp1Reg, 0); 8968 } 8969 subptr(len, 8); 8970 jccb(Assembler::greater, L_copy_8_chars_exit); 8971 8972 bind(L_copy_8_chars); 8973 movdqu(tmp3Reg, Address(src, len, Address::times_2, -16)); 8974 ptest(tmp3Reg, tmp1Reg); 8975 jccb(Assembler::notZero, L_copy_8_chars_exit); 8976 packuswb(tmp3Reg, tmp1Reg); 8977 movq(Address(dst, len, Address::times_1, -8), tmp3Reg); 8978 addptr(len, 8); 8979 jccb(Assembler::lessEqual, L_copy_8_chars); 8980 8981 bind(L_copy_8_chars_exit); 8982 subptr(len, 8); 8983 jccb(Assembler::zero, L_done); 8984 } 8985 8986 bind(L_copy_1_char); 8987 load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0)); 8988 testl(tmp5, 0xff00); // check if Unicode char 8989 jccb(Assembler::notZero, L_copy_1_char_exit); 8990 movb(Address(dst, len, Address::times_1, 0), tmp5); 8991 addptr(len, 1); 8992 jccb(Assembler::less, L_copy_1_char); 8993 8994 bind(L_copy_1_char_exit); 8995 addptr(result, len); // len is negative count of not processed elements 8996 bind(L_done); 8997} 8998 8999#ifdef _LP64 9000/** 9001 * Helper for multiply_to_len(). 9002 */ 9003void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo, Register src1, Register src2) { 9004 addq(dest_lo, src1); 9005 adcq(dest_hi, 0); 9006 addq(dest_lo, src2); 9007 adcq(dest_hi, 0); 9008} 9009 9010/** 9011 * Multiply 64 bit by 64 bit first loop. 9012 */ 9013void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart, 9014 Register y, Register y_idx, Register z, 9015 Register carry, Register product, 9016 Register idx, Register kdx) { 9017 // 9018 // jlong carry, x[], y[], z[]; 9019 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) { 9020 // huge_128 product = y[idx] * x[xstart] + carry; 9021 // z[kdx] = (jlong)product; 9022 // carry = (jlong)(product >>> 64); 9023 // } 9024 // z[xstart] = carry; 9025 // 9026 9027 Label L_first_loop, L_first_loop_exit; 9028 Label L_one_x, L_one_y, L_multiply; 9029 9030 decrementl(xstart); 9031 jcc(Assembler::negative, L_one_x); 9032 9033 movq(x_xstart, Address(x, xstart, Address::times_4, 0)); 9034 rorq(x_xstart, 32); // convert big-endian to little-endian 9035 9036 bind(L_first_loop); 9037 decrementl(idx); 9038 jcc(Assembler::negative, L_first_loop_exit); 9039 decrementl(idx); 9040 jcc(Assembler::negative, L_one_y); 9041 movq(y_idx, Address(y, idx, Address::times_4, 0)); 9042 rorq(y_idx, 32); // convert big-endian to little-endian 9043 bind(L_multiply); 9044 movq(product, x_xstart); 9045 mulq(y_idx); // product(rax) * y_idx -> rdx:rax 9046 addq(product, carry); 9047 adcq(rdx, 0); 9048 subl(kdx, 2); 9049 movl(Address(z, kdx, Address::times_4, 4), product); 9050 shrq(product, 32); 9051 movl(Address(z, kdx, Address::times_4, 0), product); 9052 movq(carry, rdx); 9053 jmp(L_first_loop); 9054 9055 bind(L_one_y); 9056 movl(y_idx, Address(y, 0)); 9057 jmp(L_multiply); 9058 9059 bind(L_one_x); 9060 movl(x_xstart, Address(x, 0)); 9061 jmp(L_first_loop); 9062 9063 bind(L_first_loop_exit); 9064} 9065 9066/** 9067 * Multiply 64 bit by 64 bit and add 128 bit. 9068 */ 9069void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y, Register z, 9070 Register yz_idx, Register idx, 9071 Register carry, Register product, int offset) { 9072 // huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry; 9073 // z[kdx] = (jlong)product; 9074 9075 movq(yz_idx, Address(y, idx, Address::times_4, offset)); 9076 rorq(yz_idx, 32); // convert big-endian to little-endian 9077 movq(product, x_xstart); 9078 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax) 9079 movq(yz_idx, Address(z, idx, Address::times_4, offset)); 9080 rorq(yz_idx, 32); // convert big-endian to little-endian 9081 9082 add2_with_carry(rdx, product, carry, yz_idx); 9083 9084 movl(Address(z, idx, Address::times_4, offset+4), product); 9085 shrq(product, 32); 9086 movl(Address(z, idx, Address::times_4, offset), product); 9087 9088} 9089 9090/** 9091 * Multiply 128 bit by 128 bit. Unrolled inner loop. 9092 */ 9093void MacroAssembler::multiply_128_x_128_loop(Register x_xstart, Register y, Register z, 9094 Register yz_idx, Register idx, Register jdx, 9095 Register carry, Register product, 9096 Register carry2) { 9097 // jlong carry, x[], y[], z[]; 9098 // int kdx = ystart+1; 9099 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop 9100 // huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry; 9101 // z[kdx+idx+1] = (jlong)product; 9102 // jlong carry2 = (jlong)(product >>> 64); 9103 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry2; 9104 // z[kdx+idx] = (jlong)product; 9105 // carry = (jlong)(product >>> 64); 9106 // } 9107 // idx += 2; 9108 // if (idx > 0) { 9109 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry; 9110 // z[kdx+idx] = (jlong)product; 9111 // carry = (jlong)(product >>> 64); 9112 // } 9113 // 9114 9115 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done; 9116 9117 movl(jdx, idx); 9118 andl(jdx, 0xFFFFFFFC); 9119 shrl(jdx, 2); 9120 9121 bind(L_third_loop); 9122 subl(jdx, 1); 9123 jcc(Assembler::negative, L_third_loop_exit); 9124 subl(idx, 4); 9125 9126 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8); 9127 movq(carry2, rdx); 9128 9129 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0); 9130 movq(carry, rdx); 9131 jmp(L_third_loop); 9132 9133 bind (L_third_loop_exit); 9134 9135 andl (idx, 0x3); 9136 jcc(Assembler::zero, L_post_third_loop_done); 9137 9138 Label L_check_1; 9139 subl(idx, 2); 9140 jcc(Assembler::negative, L_check_1); 9141 9142 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0); 9143 movq(carry, rdx); 9144 9145 bind (L_check_1); 9146 addl (idx, 0x2); 9147 andl (idx, 0x1); 9148 subl(idx, 1); 9149 jcc(Assembler::negative, L_post_third_loop_done); 9150 9151 movl(yz_idx, Address(y, idx, Address::times_4, 0)); 9152 movq(product, x_xstart); 9153 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax) 9154 movl(yz_idx, Address(z, idx, Address::times_4, 0)); 9155 9156 add2_with_carry(rdx, product, yz_idx, carry); 9157 9158 movl(Address(z, idx, Address::times_4, 0), product); 9159 shrq(product, 32); 9160 9161 shlq(rdx, 32); 9162 orq(product, rdx); 9163 movq(carry, product); 9164 9165 bind(L_post_third_loop_done); 9166} 9167 9168/** 9169 * Multiply 128 bit by 128 bit using BMI2. Unrolled inner loop. 9170 * 9171 */ 9172void MacroAssembler::multiply_128_x_128_bmi2_loop(Register y, Register z, 9173 Register carry, Register carry2, 9174 Register idx, Register jdx, 9175 Register yz_idx1, Register yz_idx2, 9176 Register tmp, Register tmp3, Register tmp4) { 9177 assert(UseBMI2Instructions, "should be used only when BMI2 is available"); 9178 9179 // jlong carry, x[], y[], z[]; 9180 // int kdx = ystart+1; 9181 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop 9182 // huge_128 tmp3 = (y[idx+1] * rdx) + z[kdx+idx+1] + carry; 9183 // jlong carry2 = (jlong)(tmp3 >>> 64); 9184 // huge_128 tmp4 = (y[idx] * rdx) + z[kdx+idx] + carry2; 9185 // carry = (jlong)(tmp4 >>> 64); 9186 // z[kdx+idx+1] = (jlong)tmp3; 9187 // z[kdx+idx] = (jlong)tmp4; 9188 // } 9189 // idx += 2; 9190 // if (idx > 0) { 9191 // yz_idx1 = (y[idx] * rdx) + z[kdx+idx] + carry; 9192 // z[kdx+idx] = (jlong)yz_idx1; 9193 // carry = (jlong)(yz_idx1 >>> 64); 9194 // } 9195 // 9196 9197 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done; 9198 9199 movl(jdx, idx); 9200 andl(jdx, 0xFFFFFFFC); 9201 shrl(jdx, 2); 9202 9203 bind(L_third_loop); 9204 subl(jdx, 1); 9205 jcc(Assembler::negative, L_third_loop_exit); 9206 subl(idx, 4); 9207 9208 movq(yz_idx1, Address(y, idx, Address::times_4, 8)); 9209 rorxq(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian 9210 movq(yz_idx2, Address(y, idx, Address::times_4, 0)); 9211 rorxq(yz_idx2, yz_idx2, 32); 9212 9213 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3 9214 mulxq(carry2, tmp, yz_idx2); // yz_idx2 * rdx -> carry2:tmp 9215 9216 movq(yz_idx1, Address(z, idx, Address::times_4, 8)); 9217 rorxq(yz_idx1, yz_idx1, 32); 9218 movq(yz_idx2, Address(z, idx, Address::times_4, 0)); 9219 rorxq(yz_idx2, yz_idx2, 32); 9220 9221 if (VM_Version::supports_adx()) { 9222 adcxq(tmp3, carry); 9223 adoxq(tmp3, yz_idx1); 9224 9225 adcxq(tmp4, tmp); 9226 adoxq(tmp4, yz_idx2); 9227 9228 movl(carry, 0); // does not affect flags 9229 adcxq(carry2, carry); 9230 adoxq(carry2, carry); 9231 } else { 9232 add2_with_carry(tmp4, tmp3, carry, yz_idx1); 9233 add2_with_carry(carry2, tmp4, tmp, yz_idx2); 9234 } 9235 movq(carry, carry2); 9236 9237 movl(Address(z, idx, Address::times_4, 12), tmp3); 9238 shrq(tmp3, 32); 9239 movl(Address(z, idx, Address::times_4, 8), tmp3); 9240 9241 movl(Address(z, idx, Address::times_4, 4), tmp4); 9242 shrq(tmp4, 32); 9243 movl(Address(z, idx, Address::times_4, 0), tmp4); 9244 9245 jmp(L_third_loop); 9246 9247 bind (L_third_loop_exit); 9248 9249 andl (idx, 0x3); 9250 jcc(Assembler::zero, L_post_third_loop_done); 9251 9252 Label L_check_1; 9253 subl(idx, 2); 9254 jcc(Assembler::negative, L_check_1); 9255 9256 movq(yz_idx1, Address(y, idx, Address::times_4, 0)); 9257 rorxq(yz_idx1, yz_idx1, 32); 9258 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3 9259 movq(yz_idx2, Address(z, idx, Address::times_4, 0)); 9260 rorxq(yz_idx2, yz_idx2, 32); 9261 9262 add2_with_carry(tmp4, tmp3, carry, yz_idx2); 9263 9264 movl(Address(z, idx, Address::times_4, 4), tmp3); 9265 shrq(tmp3, 32); 9266 movl(Address(z, idx, Address::times_4, 0), tmp3); 9267 movq(carry, tmp4); 9268 9269 bind (L_check_1); 9270 addl (idx, 0x2); 9271 andl (idx, 0x1); 9272 subl(idx, 1); 9273 jcc(Assembler::negative, L_post_third_loop_done); 9274 movl(tmp4, Address(y, idx, Address::times_4, 0)); 9275 mulxq(carry2, tmp3, tmp4); // tmp4 * rdx -> carry2:tmp3 9276 movl(tmp4, Address(z, idx, Address::times_4, 0)); 9277 9278 add2_with_carry(carry2, tmp3, tmp4, carry); 9279 9280 movl(Address(z, idx, Address::times_4, 0), tmp3); 9281 shrq(tmp3, 32); 9282 9283 shlq(carry2, 32); 9284 orq(tmp3, carry2); 9285 movq(carry, tmp3); 9286 9287 bind(L_post_third_loop_done); 9288} 9289 9290/** 9291 * Code for BigInteger::multiplyToLen() instrinsic. 9292 * 9293 * rdi: x 9294 * rax: xlen 9295 * rsi: y 9296 * rcx: ylen 9297 * r8: z 9298 * r11: zlen 9299 * r12: tmp1 9300 * r13: tmp2 9301 * r14: tmp3 9302 * r15: tmp4 9303 * rbx: tmp5 9304 * 9305 */ 9306void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, Register z, Register zlen, 9307 Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5) { 9308 ShortBranchVerifier sbv(this); 9309 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx); 9310 9311 push(tmp1); 9312 push(tmp2); 9313 push(tmp3); 9314 push(tmp4); 9315 push(tmp5); 9316 9317 push(xlen); 9318 push(zlen); 9319 9320 const Register idx = tmp1; 9321 const Register kdx = tmp2; 9322 const Register xstart = tmp3; 9323 9324 const Register y_idx = tmp4; 9325 const Register carry = tmp5; 9326 const Register product = xlen; 9327 const Register x_xstart = zlen; // reuse register 9328 9329 // First Loop. 9330 // 9331 // final static long LONG_MASK = 0xffffffffL; 9332 // int xstart = xlen - 1; 9333 // int ystart = ylen - 1; 9334 // long carry = 0; 9335 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) { 9336 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry; 9337 // z[kdx] = (int)product; 9338 // carry = product >>> 32; 9339 // } 9340 // z[xstart] = (int)carry; 9341 // 9342 9343 movl(idx, ylen); // idx = ylen; 9344 movl(kdx, zlen); // kdx = xlen+ylen; 9345 xorq(carry, carry); // carry = 0; 9346 9347 Label L_done; 9348 9349 movl(xstart, xlen); 9350 decrementl(xstart); 9351 jcc(Assembler::negative, L_done); 9352 9353 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx); 9354 9355 Label L_second_loop; 9356 testl(kdx, kdx); 9357 jcc(Assembler::zero, L_second_loop); 9358 9359 Label L_carry; 9360 subl(kdx, 1); 9361 jcc(Assembler::zero, L_carry); 9362 9363 movl(Address(z, kdx, Address::times_4, 0), carry); 9364 shrq(carry, 32); 9365 subl(kdx, 1); 9366 9367 bind(L_carry); 9368 movl(Address(z, kdx, Address::times_4, 0), carry); 9369 9370 // Second and third (nested) loops. 9371 // 9372 // for (int i = xstart-1; i >= 0; i--) { // Second loop 9373 // carry = 0; 9374 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop 9375 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) + 9376 // (z[k] & LONG_MASK) + carry; 9377 // z[k] = (int)product; 9378 // carry = product >>> 32; 9379 // } 9380 // z[i] = (int)carry; 9381 // } 9382 // 9383 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx 9384 9385 const Register jdx = tmp1; 9386 9387 bind(L_second_loop); 9388 xorl(carry, carry); // carry = 0; 9389 movl(jdx, ylen); // j = ystart+1 9390 9391 subl(xstart, 1); // i = xstart-1; 9392 jcc(Assembler::negative, L_done); 9393 9394 push (z); 9395 9396 Label L_last_x; 9397 lea(z, Address(z, xstart, Address::times_4, 4)); // z = z + k - j 9398 subl(xstart, 1); // i = xstart-1; 9399 jcc(Assembler::negative, L_last_x); 9400 9401 if (UseBMI2Instructions) { 9402 movq(rdx, Address(x, xstart, Address::times_4, 0)); 9403 rorxq(rdx, rdx, 32); // convert big-endian to little-endian 9404 } else { 9405 movq(x_xstart, Address(x, xstart, Address::times_4, 0)); 9406 rorq(x_xstart, 32); // convert big-endian to little-endian 9407 } 9408 9409 Label L_third_loop_prologue; 9410 bind(L_third_loop_prologue); 9411 9412 push (x); 9413 push (xstart); 9414 push (ylen); 9415 9416 9417 if (UseBMI2Instructions) { 9418 multiply_128_x_128_bmi2_loop(y, z, carry, x, jdx, ylen, product, tmp2, x_xstart, tmp3, tmp4); 9419 } else { // !UseBMI2Instructions 9420 multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x); 9421 } 9422 9423 pop(ylen); 9424 pop(xlen); 9425 pop(x); 9426 pop(z); 9427 9428 movl(tmp3, xlen); 9429 addl(tmp3, 1); 9430 movl(Address(z, tmp3, Address::times_4, 0), carry); 9431 subl(tmp3, 1); 9432 jccb(Assembler::negative, L_done); 9433 9434 shrq(carry, 32); 9435 movl(Address(z, tmp3, Address::times_4, 0), carry); 9436 jmp(L_second_loop); 9437 9438 // Next infrequent code is moved outside loops. 9439 bind(L_last_x); 9440 if (UseBMI2Instructions) { 9441 movl(rdx, Address(x, 0)); 9442 } else { 9443 movl(x_xstart, Address(x, 0)); 9444 } 9445 jmp(L_third_loop_prologue); 9446 9447 bind(L_done); 9448 9449 pop(zlen); 9450 pop(xlen); 9451 9452 pop(tmp5); 9453 pop(tmp4); 9454 pop(tmp3); 9455 pop(tmp2); 9456 pop(tmp1); 9457} 9458 9459//Helper functions for square_to_len() 9460 9461/** 9462 * Store the squares of x[], right shifted one bit (divided by 2) into z[] 9463 * Preserves x and z and modifies rest of the registers. 9464 */ 9465 9466void MacroAssembler::square_rshift(Register x, Register xlen, Register z, Register tmp1, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 9467 // Perform square and right shift by 1 9468 // Handle odd xlen case first, then for even xlen do the following 9469 // jlong carry = 0; 9470 // for (int j=0, i=0; j < xlen; j+=2, i+=4) { 9471 // huge_128 product = x[j:j+1] * x[j:j+1]; 9472 // z[i:i+1] = (carry << 63) | (jlong)(product >>> 65); 9473 // z[i+2:i+3] = (jlong)(product >>> 1); 9474 // carry = (jlong)product; 9475 // } 9476 9477 xorq(tmp5, tmp5); // carry 9478 xorq(rdxReg, rdxReg); 9479 xorl(tmp1, tmp1); // index for x 9480 xorl(tmp4, tmp4); // index for z 9481 9482 Label L_first_loop, L_first_loop_exit; 9483 9484 testl(xlen, 1); 9485 jccb(Assembler::zero, L_first_loop); //jump if xlen is even 9486 9487 // Square and right shift by 1 the odd element using 32 bit multiply 9488 movl(raxReg, Address(x, tmp1, Address::times_4, 0)); 9489 imulq(raxReg, raxReg); 9490 shrq(raxReg, 1); 9491 adcq(tmp5, 0); 9492 movq(Address(z, tmp4, Address::times_4, 0), raxReg); 9493 incrementl(tmp1); 9494 addl(tmp4, 2); 9495 9496 // Square and right shift by 1 the rest using 64 bit multiply 9497 bind(L_first_loop); 9498 cmpptr(tmp1, xlen); 9499 jccb(Assembler::equal, L_first_loop_exit); 9500 9501 // Square 9502 movq(raxReg, Address(x, tmp1, Address::times_4, 0)); 9503 rorq(raxReg, 32); // convert big-endian to little-endian 9504 mulq(raxReg); // 64-bit multiply rax * rax -> rdx:rax 9505 9506 // Right shift by 1 and save carry 9507 shrq(tmp5, 1); // rdx:rax:tmp5 = (tmp5:rdx:rax) >>> 1 9508 rcrq(rdxReg, 1); 9509 rcrq(raxReg, 1); 9510 adcq(tmp5, 0); 9511 9512 // Store result in z 9513 movq(Address(z, tmp4, Address::times_4, 0), rdxReg); 9514 movq(Address(z, tmp4, Address::times_4, 8), raxReg); 9515 9516 // Update indices for x and z 9517 addl(tmp1, 2); 9518 addl(tmp4, 4); 9519 jmp(L_first_loop); 9520 9521 bind(L_first_loop_exit); 9522} 9523 9524 9525/** 9526 * Perform the following multiply add operation using BMI2 instructions 9527 * carry:sum = sum + op1*op2 + carry 9528 * op2 should be in rdx 9529 * op2 is preserved, all other registers are modified 9530 */ 9531void MacroAssembler::multiply_add_64_bmi2(Register sum, Register op1, Register op2, Register carry, Register tmp2) { 9532 // assert op2 is rdx 9533 mulxq(tmp2, op1, op1); // op1 * op2 -> tmp2:op1 9534 addq(sum, carry); 9535 adcq(tmp2, 0); 9536 addq(sum, op1); 9537 adcq(tmp2, 0); 9538 movq(carry, tmp2); 9539} 9540 9541/** 9542 * Perform the following multiply add operation: 9543 * carry:sum = sum + op1*op2 + carry 9544 * Preserves op1, op2 and modifies rest of registers 9545 */ 9546void MacroAssembler::multiply_add_64(Register sum, Register op1, Register op2, Register carry, Register rdxReg, Register raxReg) { 9547 // rdx:rax = op1 * op2 9548 movq(raxReg, op2); 9549 mulq(op1); 9550 9551 // rdx:rax = sum + carry + rdx:rax 9552 addq(sum, carry); 9553 adcq(rdxReg, 0); 9554 addq(sum, raxReg); 9555 adcq(rdxReg, 0); 9556 9557 // carry:sum = rdx:sum 9558 movq(carry, rdxReg); 9559} 9560 9561/** 9562 * Add 64 bit long carry into z[] with carry propogation. 9563 * Preserves z and carry register values and modifies rest of registers. 9564 * 9565 */ 9566void MacroAssembler::add_one_64(Register z, Register zlen, Register carry, Register tmp1) { 9567 Label L_fourth_loop, L_fourth_loop_exit; 9568 9569 movl(tmp1, 1); 9570 subl(zlen, 2); 9571 addq(Address(z, zlen, Address::times_4, 0), carry); 9572 9573 bind(L_fourth_loop); 9574 jccb(Assembler::carryClear, L_fourth_loop_exit); 9575 subl(zlen, 2); 9576 jccb(Assembler::negative, L_fourth_loop_exit); 9577 addq(Address(z, zlen, Address::times_4, 0), tmp1); 9578 jmp(L_fourth_loop); 9579 bind(L_fourth_loop_exit); 9580} 9581 9582/** 9583 * Shift z[] left by 1 bit. 9584 * Preserves x, len, z and zlen registers and modifies rest of the registers. 9585 * 9586 */ 9587void MacroAssembler::lshift_by_1(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4) { 9588 9589 Label L_fifth_loop, L_fifth_loop_exit; 9590 9591 // Fifth loop 9592 // Perform primitiveLeftShift(z, zlen, 1) 9593 9594 const Register prev_carry = tmp1; 9595 const Register new_carry = tmp4; 9596 const Register value = tmp2; 9597 const Register zidx = tmp3; 9598 9599 // int zidx, carry; 9600 // long value; 9601 // carry = 0; 9602 // for (zidx = zlen-2; zidx >=0; zidx -= 2) { 9603 // (carry:value) = (z[i] << 1) | carry ; 9604 // z[i] = value; 9605 // } 9606 9607 movl(zidx, zlen); 9608 xorl(prev_carry, prev_carry); // clear carry flag and prev_carry register 9609 9610 bind(L_fifth_loop); 9611 decl(zidx); // Use decl to preserve carry flag 9612 decl(zidx); 9613 jccb(Assembler::negative, L_fifth_loop_exit); 9614 9615 if (UseBMI2Instructions) { 9616 movq(value, Address(z, zidx, Address::times_4, 0)); 9617 rclq(value, 1); 9618 rorxq(value, value, 32); 9619 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form 9620 } 9621 else { 9622 // clear new_carry 9623 xorl(new_carry, new_carry); 9624 9625 // Shift z[i] by 1, or in previous carry and save new carry 9626 movq(value, Address(z, zidx, Address::times_4, 0)); 9627 shlq(value, 1); 9628 adcl(new_carry, 0); 9629 9630 orq(value, prev_carry); 9631 rorq(value, 0x20); 9632 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form 9633 9634 // Set previous carry = new carry 9635 movl(prev_carry, new_carry); 9636 } 9637 jmp(L_fifth_loop); 9638 9639 bind(L_fifth_loop_exit); 9640} 9641 9642 9643/** 9644 * Code for BigInteger::squareToLen() intrinsic 9645 * 9646 * rdi: x 9647 * rsi: len 9648 * r8: z 9649 * rcx: zlen 9650 * r12: tmp1 9651 * r13: tmp2 9652 * r14: tmp3 9653 * r15: tmp4 9654 * rbx: tmp5 9655 * 9656 */ 9657void 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) { 9658 9659 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; 9660 push(tmp1); 9661 push(tmp2); 9662 push(tmp3); 9663 push(tmp4); 9664 push(tmp5); 9665 9666 // First loop 9667 // Store the squares, right shifted one bit (i.e., divided by 2). 9668 square_rshift(x, len, z, tmp1, tmp3, tmp4, tmp5, rdxReg, raxReg); 9669 9670 // Add in off-diagonal sums. 9671 // 9672 // Second, third (nested) and fourth loops. 9673 // zlen +=2; 9674 // for (int xidx=len-2,zidx=zlen-4; xidx > 0; xidx-=2,zidx-=4) { 9675 // carry = 0; 9676 // long op2 = x[xidx:xidx+1]; 9677 // for (int j=xidx-2,k=zidx; j >= 0; j-=2) { 9678 // k -= 2; 9679 // long op1 = x[j:j+1]; 9680 // long sum = z[k:k+1]; 9681 // carry:sum = multiply_add_64(sum, op1, op2, carry, tmp_regs); 9682 // z[k:k+1] = sum; 9683 // } 9684 // add_one_64(z, k, carry, tmp_regs); 9685 // } 9686 9687 const Register carry = tmp5; 9688 const Register sum = tmp3; 9689 const Register op1 = tmp4; 9690 Register op2 = tmp2; 9691 9692 push(zlen); 9693 push(len); 9694 addl(zlen,2); 9695 bind(L_second_loop); 9696 xorq(carry, carry); 9697 subl(zlen, 4); 9698 subl(len, 2); 9699 push(zlen); 9700 push(len); 9701 cmpl(len, 0); 9702 jccb(Assembler::lessEqual, L_second_loop_exit); 9703 9704 // Multiply an array by one 64 bit long. 9705 if (UseBMI2Instructions) { 9706 op2 = rdxReg; 9707 movq(op2, Address(x, len, Address::times_4, 0)); 9708 rorxq(op2, op2, 32); 9709 } 9710 else { 9711 movq(op2, Address(x, len, Address::times_4, 0)); 9712 rorq(op2, 32); 9713 } 9714 9715 bind(L_third_loop); 9716 decrementl(len); 9717 jccb(Assembler::negative, L_third_loop_exit); 9718 decrementl(len); 9719 jccb(Assembler::negative, L_last_x); 9720 9721 movq(op1, Address(x, len, Address::times_4, 0)); 9722 rorq(op1, 32); 9723 9724 bind(L_multiply); 9725 subl(zlen, 2); 9726 movq(sum, Address(z, zlen, Address::times_4, 0)); 9727 9728 // Multiply 64 bit by 64 bit and add 64 bits lower half and upper 64 bits as carry. 9729 if (UseBMI2Instructions) { 9730 multiply_add_64_bmi2(sum, op1, op2, carry, tmp2); 9731 } 9732 else { 9733 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9734 } 9735 9736 movq(Address(z, zlen, Address::times_4, 0), sum); 9737 9738 jmp(L_third_loop); 9739 bind(L_third_loop_exit); 9740 9741 // Fourth loop 9742 // Add 64 bit long carry into z with carry propogation. 9743 // Uses offsetted zlen. 9744 add_one_64(z, zlen, carry, tmp1); 9745 9746 pop(len); 9747 pop(zlen); 9748 jmp(L_second_loop); 9749 9750 // Next infrequent code is moved outside loops. 9751 bind(L_last_x); 9752 movl(op1, Address(x, 0)); 9753 jmp(L_multiply); 9754 9755 bind(L_second_loop_exit); 9756 pop(len); 9757 pop(zlen); 9758 pop(len); 9759 pop(zlen); 9760 9761 // Fifth loop 9762 // Shift z left 1 bit. 9763 lshift_by_1(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4); 9764 9765 // z[zlen-1] |= x[len-1] & 1; 9766 movl(tmp3, Address(x, len, Address::times_4, -4)); 9767 andl(tmp3, 1); 9768 orl(Address(z, zlen, Address::times_4, -4), tmp3); 9769 9770 pop(tmp5); 9771 pop(tmp4); 9772 pop(tmp3); 9773 pop(tmp2); 9774 pop(tmp1); 9775} 9776 9777/** 9778 * Helper function for mul_add() 9779 * Multiply the in[] by int k and add to out[] starting at offset offs using 9780 * 128 bit by 32 bit multiply and return the carry in tmp5. 9781 * Only quad int aligned length of in[] is operated on in this function. 9782 * k is in rdxReg for BMI2Instructions, for others it is in tmp2. 9783 * This function preserves out, in and k registers. 9784 * len and offset point to the appropriate index in "in" & "out" correspondingly 9785 * tmp5 has the carry. 9786 * other registers are temporary and are modified. 9787 * 9788 */ 9789void MacroAssembler::mul_add_128_x_32_loop(Register out, Register in, 9790 Register offset, Register len, Register tmp1, Register tmp2, Register tmp3, 9791 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 9792 9793 Label L_first_loop, L_first_loop_exit; 9794 9795 movl(tmp1, len); 9796 shrl(tmp1, 2); 9797 9798 bind(L_first_loop); 9799 subl(tmp1, 1); 9800 jccb(Assembler::negative, L_first_loop_exit); 9801 9802 subl(len, 4); 9803 subl(offset, 4); 9804 9805 Register op2 = tmp2; 9806 const Register sum = tmp3; 9807 const Register op1 = tmp4; 9808 const Register carry = tmp5; 9809 9810 if (UseBMI2Instructions) { 9811 op2 = rdxReg; 9812 } 9813 9814 movq(op1, Address(in, len, Address::times_4, 8)); 9815 rorq(op1, 32); 9816 movq(sum, Address(out, offset, Address::times_4, 8)); 9817 rorq(sum, 32); 9818 if (UseBMI2Instructions) { 9819 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 9820 } 9821 else { 9822 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9823 } 9824 // Store back in big endian from little endian 9825 rorq(sum, 0x20); 9826 movq(Address(out, offset, Address::times_4, 8), sum); 9827 9828 movq(op1, Address(in, len, Address::times_4, 0)); 9829 rorq(op1, 32); 9830 movq(sum, Address(out, offset, Address::times_4, 0)); 9831 rorq(sum, 32); 9832 if (UseBMI2Instructions) { 9833 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 9834 } 9835 else { 9836 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9837 } 9838 // Store back in big endian from little endian 9839 rorq(sum, 0x20); 9840 movq(Address(out, offset, Address::times_4, 0), sum); 9841 9842 jmp(L_first_loop); 9843 bind(L_first_loop_exit); 9844} 9845 9846/** 9847 * Code for BigInteger::mulAdd() intrinsic 9848 * 9849 * rdi: out 9850 * rsi: in 9851 * r11: offs (out.length - offset) 9852 * rcx: len 9853 * r8: k 9854 * r12: tmp1 9855 * r13: tmp2 9856 * r14: tmp3 9857 * r15: tmp4 9858 * rbx: tmp5 9859 * Multiply the in[] by word k and add to out[], return the carry in rax 9860 */ 9861void MacroAssembler::mul_add(Register out, Register in, Register offs, 9862 Register len, Register k, Register tmp1, Register tmp2, Register tmp3, 9863 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 9864 9865 Label L_carry, L_last_in, L_done; 9866 9867// carry = 0; 9868// for (int j=len-1; j >= 0; j--) { 9869// long product = (in[j] & LONG_MASK) * kLong + 9870// (out[offs] & LONG_MASK) + carry; 9871// out[offs--] = (int)product; 9872// carry = product >>> 32; 9873// } 9874// 9875 push(tmp1); 9876 push(tmp2); 9877 push(tmp3); 9878 push(tmp4); 9879 push(tmp5); 9880 9881 Register op2 = tmp2; 9882 const Register sum = tmp3; 9883 const Register op1 = tmp4; 9884 const Register carry = tmp5; 9885 9886 if (UseBMI2Instructions) { 9887 op2 = rdxReg; 9888 movl(op2, k); 9889 } 9890 else { 9891 movl(op2, k); 9892 } 9893 9894 xorq(carry, carry); 9895 9896 //First loop 9897 9898 //Multiply in[] by k in a 4 way unrolled loop using 128 bit by 32 bit multiply 9899 //The carry is in tmp5 9900 mul_add_128_x_32_loop(out, in, offs, len, tmp1, tmp2, tmp3, tmp4, tmp5, rdxReg, raxReg); 9901 9902 //Multiply the trailing in[] entry using 64 bit by 32 bit, if any 9903 decrementl(len); 9904 jccb(Assembler::negative, L_carry); 9905 decrementl(len); 9906 jccb(Assembler::negative, L_last_in); 9907 9908 movq(op1, Address(in, len, Address::times_4, 0)); 9909 rorq(op1, 32); 9910 9911 subl(offs, 2); 9912 movq(sum, Address(out, offs, Address::times_4, 0)); 9913 rorq(sum, 32); 9914 9915 if (UseBMI2Instructions) { 9916 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 9917 } 9918 else { 9919 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9920 } 9921 9922 // Store back in big endian from little endian 9923 rorq(sum, 0x20); 9924 movq(Address(out, offs, Address::times_4, 0), sum); 9925 9926 testl(len, len); 9927 jccb(Assembler::zero, L_carry); 9928 9929 //Multiply the last in[] entry, if any 9930 bind(L_last_in); 9931 movl(op1, Address(in, 0)); 9932 movl(sum, Address(out, offs, Address::times_4, -4)); 9933 9934 movl(raxReg, k); 9935 mull(op1); //tmp4 * eax -> edx:eax 9936 addl(sum, carry); 9937 adcl(rdxReg, 0); 9938 addl(sum, raxReg); 9939 adcl(rdxReg, 0); 9940 movl(carry, rdxReg); 9941 9942 movl(Address(out, offs, Address::times_4, -4), sum); 9943 9944 bind(L_carry); 9945 //return tmp5/carry as carry in rax 9946 movl(rax, carry); 9947 9948 bind(L_done); 9949 pop(tmp5); 9950 pop(tmp4); 9951 pop(tmp3); 9952 pop(tmp2); 9953 pop(tmp1); 9954} 9955#endif 9956 9957/** 9958 * Emits code to update CRC-32 with a byte value according to constants in table 9959 * 9960 * @param [in,out]crc Register containing the crc. 9961 * @param [in]val Register containing the byte to fold into the CRC. 9962 * @param [in]table Register containing the table of crc constants. 9963 * 9964 * uint32_t crc; 9965 * val = crc_table[(val ^ crc) & 0xFF]; 9966 * crc = val ^ (crc >> 8); 9967 * 9968 */ 9969void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) { 9970 xorl(val, crc); 9971 andl(val, 0xFF); 9972 shrl(crc, 8); // unsigned shift 9973 xorl(crc, Address(table, val, Address::times_4, 0)); 9974} 9975 9976/** 9977 * Fold 128-bit data chunk 9978 */ 9979void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) { 9980 if (UseAVX > 0) { 9981 vpclmulhdq(xtmp, xK, xcrc); // [123:64] 9982 vpclmulldq(xcrc, xK, xcrc); // [63:0] 9983 vpxor(xcrc, xcrc, Address(buf, offset), 0 /* vector_len */); 9984 pxor(xcrc, xtmp); 9985 } else { 9986 movdqa(xtmp, xcrc); 9987 pclmulhdq(xtmp, xK); // [123:64] 9988 pclmulldq(xcrc, xK); // [63:0] 9989 pxor(xcrc, xtmp); 9990 movdqu(xtmp, Address(buf, offset)); 9991 pxor(xcrc, xtmp); 9992 } 9993} 9994 9995void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) { 9996 if (UseAVX > 0) { 9997 vpclmulhdq(xtmp, xK, xcrc); 9998 vpclmulldq(xcrc, xK, xcrc); 9999 pxor(xcrc, xbuf); 10000 pxor(xcrc, xtmp); 10001 } else { 10002 movdqa(xtmp, xcrc); 10003 pclmulhdq(xtmp, xK); 10004 pclmulldq(xcrc, xK); 10005 pxor(xcrc, xbuf); 10006 pxor(xcrc, xtmp); 10007 } 10008} 10009 10010/** 10011 * 8-bit folds to compute 32-bit CRC 10012 * 10013 * uint64_t xcrc; 10014 * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8); 10015 */ 10016void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) { 10017 movdl(tmp, xcrc); 10018 andl(tmp, 0xFF); 10019 movdl(xtmp, Address(table, tmp, Address::times_4, 0)); 10020 psrldq(xcrc, 1); // unsigned shift one byte 10021 pxor(xcrc, xtmp); 10022} 10023 10024/** 10025 * uint32_t crc; 10026 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8); 10027 */ 10028void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) { 10029 movl(tmp, crc); 10030 andl(tmp, 0xFF); 10031 shrl(crc, 8); 10032 xorl(crc, Address(table, tmp, Address::times_4, 0)); 10033} 10034 10035/** 10036 * @param crc register containing existing CRC (32-bit) 10037 * @param buf register pointing to input byte buffer (byte*) 10038 * @param len register containing number of bytes 10039 * @param table register that will contain address of CRC table 10040 * @param tmp scratch register 10041 */ 10042void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) { 10043 assert_different_registers(crc, buf, len, table, tmp, rax); 10044 10045 Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned; 10046 Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop; 10047 10048 // For EVEX with VL and BW, provide a standard mask, VL = 128 will guide the merge 10049 // context for the registers used, where all instructions below are using 128-bit mode 10050 // On EVEX without VL and BW, these instructions will all be AVX. 10051 if (VM_Version::supports_avx512vlbw()) { 10052 movl(tmp, 0xffff); 10053 kmovwl(k1, tmp); 10054 } 10055 10056 lea(table, ExternalAddress(StubRoutines::crc_table_addr())); 10057 notl(crc); // ~crc 10058 cmpl(len, 16); 10059 jcc(Assembler::less, L_tail); 10060 10061 // Align buffer to 16 bytes 10062 movl(tmp, buf); 10063 andl(tmp, 0xF); 10064 jccb(Assembler::zero, L_aligned); 10065 subl(tmp, 16); 10066 addl(len, tmp); 10067 10068 align(4); 10069 BIND(L_align_loop); 10070 movsbl(rax, Address(buf, 0)); // load byte with sign extension 10071 update_byte_crc32(crc, rax, table); 10072 increment(buf); 10073 incrementl(tmp); 10074 jccb(Assembler::less, L_align_loop); 10075 10076 BIND(L_aligned); 10077 movl(tmp, len); // save 10078 shrl(len, 4); 10079 jcc(Assembler::zero, L_tail_restore); 10080 10081 // Fold crc into first bytes of vector 10082 movdqa(xmm1, Address(buf, 0)); 10083 movdl(rax, xmm1); 10084 xorl(crc, rax); 10085 pinsrd(xmm1, crc, 0); 10086 addptr(buf, 16); 10087 subl(len, 4); // len > 0 10088 jcc(Assembler::less, L_fold_tail); 10089 10090 movdqa(xmm2, Address(buf, 0)); 10091 movdqa(xmm3, Address(buf, 16)); 10092 movdqa(xmm4, Address(buf, 32)); 10093 addptr(buf, 48); 10094 subl(len, 3); 10095 jcc(Assembler::lessEqual, L_fold_512b); 10096 10097 // Fold total 512 bits of polynomial on each iteration, 10098 // 128 bits per each of 4 parallel streams. 10099 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32)); 10100 10101 align(32); 10102 BIND(L_fold_512b_loop); 10103 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0); 10104 fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16); 10105 fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32); 10106 fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48); 10107 addptr(buf, 64); 10108 subl(len, 4); 10109 jcc(Assembler::greater, L_fold_512b_loop); 10110 10111 // Fold 512 bits to 128 bits. 10112 BIND(L_fold_512b); 10113 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16)); 10114 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2); 10115 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3); 10116 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4); 10117 10118 // Fold the rest of 128 bits data chunks 10119 BIND(L_fold_tail); 10120 addl(len, 3); 10121 jccb(Assembler::lessEqual, L_fold_128b); 10122 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16)); 10123 10124 BIND(L_fold_tail_loop); 10125 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0); 10126 addptr(buf, 16); 10127 decrementl(len); 10128 jccb(Assembler::greater, L_fold_tail_loop); 10129 10130 // Fold 128 bits in xmm1 down into 32 bits in crc register. 10131 BIND(L_fold_128b); 10132 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr())); 10133 if (UseAVX > 0) { 10134 vpclmulqdq(xmm2, xmm0, xmm1, 0x1); 10135 vpand(xmm3, xmm0, xmm2, 0 /* vector_len */); 10136 vpclmulqdq(xmm0, xmm0, xmm3, 0x1); 10137 } else { 10138 movdqa(xmm2, xmm0); 10139 pclmulqdq(xmm2, xmm1, 0x1); 10140 movdqa(xmm3, xmm0); 10141 pand(xmm3, xmm2); 10142 pclmulqdq(xmm0, xmm3, 0x1); 10143 } 10144 psrldq(xmm1, 8); 10145 psrldq(xmm2, 4); 10146 pxor(xmm0, xmm1); 10147 pxor(xmm0, xmm2); 10148 10149 // 8 8-bit folds to compute 32-bit CRC. 10150 for (int j = 0; j < 4; j++) { 10151 fold_8bit_crc32(xmm0, table, xmm1, rax); 10152 } 10153 movdl(crc, xmm0); // mov 32 bits to general register 10154 for (int j = 0; j < 4; j++) { 10155 fold_8bit_crc32(crc, table, rax); 10156 } 10157 10158 BIND(L_tail_restore); 10159 movl(len, tmp); // restore 10160 BIND(L_tail); 10161 andl(len, 0xf); 10162 jccb(Assembler::zero, L_exit); 10163 10164 // Fold the rest of bytes 10165 align(4); 10166 BIND(L_tail_loop); 10167 movsbl(rax, Address(buf, 0)); // load byte with sign extension 10168 update_byte_crc32(crc, rax, table); 10169 increment(buf); 10170 decrementl(len); 10171 jccb(Assembler::greater, L_tail_loop); 10172 10173 BIND(L_exit); 10174 notl(crc); // ~c 10175} 10176 10177#ifdef _LP64 10178// S. Gueron / Information Processing Letters 112 (2012) 184 10179// Algorithm 4: Computing carry-less multiplication using a precomputed lookup table. 10180// Input: A 32 bit value B = [byte3, byte2, byte1, byte0]. 10181// Output: the 64-bit carry-less product of B * CONST 10182void MacroAssembler::crc32c_ipl_alg4(Register in, uint32_t n, 10183 Register tmp1, Register tmp2, Register tmp3) { 10184 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr())); 10185 if (n > 0) { 10186 addq(tmp3, n * 256 * 8); 10187 } 10188 // Q1 = TABLEExt[n][B & 0xFF]; 10189 movl(tmp1, in); 10190 andl(tmp1, 0x000000FF); 10191 shll(tmp1, 3); 10192 addq(tmp1, tmp3); 10193 movq(tmp1, Address(tmp1, 0)); 10194 10195 // Q2 = TABLEExt[n][B >> 8 & 0xFF]; 10196 movl(tmp2, in); 10197 shrl(tmp2, 8); 10198 andl(tmp2, 0x000000FF); 10199 shll(tmp2, 3); 10200 addq(tmp2, tmp3); 10201 movq(tmp2, Address(tmp2, 0)); 10202 10203 shlq(tmp2, 8); 10204 xorq(tmp1, tmp2); 10205 10206 // Q3 = TABLEExt[n][B >> 16 & 0xFF]; 10207 movl(tmp2, in); 10208 shrl(tmp2, 16); 10209 andl(tmp2, 0x000000FF); 10210 shll(tmp2, 3); 10211 addq(tmp2, tmp3); 10212 movq(tmp2, Address(tmp2, 0)); 10213 10214 shlq(tmp2, 16); 10215 xorq(tmp1, tmp2); 10216 10217 // Q4 = TABLEExt[n][B >> 24 & 0xFF]; 10218 shrl(in, 24); 10219 andl(in, 0x000000FF); 10220 shll(in, 3); 10221 addq(in, tmp3); 10222 movq(in, Address(in, 0)); 10223 10224 shlq(in, 24); 10225 xorq(in, tmp1); 10226 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24; 10227} 10228 10229void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1, 10230 Register in_out, 10231 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported, 10232 XMMRegister w_xtmp2, 10233 Register tmp1, 10234 Register n_tmp2, Register n_tmp3) { 10235 if (is_pclmulqdq_supported) { 10236 movdl(w_xtmp1, in_out); // modified blindly 10237 10238 movl(tmp1, const_or_pre_comp_const_index); 10239 movdl(w_xtmp2, tmp1); 10240 pclmulqdq(w_xtmp1, w_xtmp2, 0); 10241 10242 movdq(in_out, w_xtmp1); 10243 } else { 10244 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3); 10245 } 10246} 10247 10248// Recombination Alternative 2: No bit-reflections 10249// T1 = (CRC_A * U1) << 1 10250// T2 = (CRC_B * U2) << 1 10251// C1 = T1 >> 32 10252// C2 = T2 >> 32 10253// T1 = T1 & 0xFFFFFFFF 10254// T2 = T2 & 0xFFFFFFFF 10255// T1 = CRC32(0, T1) 10256// T2 = CRC32(0, T2) 10257// C1 = C1 ^ T1 10258// C2 = C2 ^ T2 10259// CRC = C1 ^ C2 ^ CRC_C 10260void 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, 10261 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10262 Register tmp1, Register tmp2, 10263 Register n_tmp3) { 10264 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10265 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10266 shlq(in_out, 1); 10267 movl(tmp1, in_out); 10268 shrq(in_out, 32); 10269 xorl(tmp2, tmp2); 10270 crc32(tmp2, tmp1, 4); 10271 xorl(in_out, tmp2); // we don't care about upper 32 bit contents here 10272 shlq(in1, 1); 10273 movl(tmp1, in1); 10274 shrq(in1, 32); 10275 xorl(tmp2, tmp2); 10276 crc32(tmp2, tmp1, 4); 10277 xorl(in1, tmp2); 10278 xorl(in_out, in1); 10279 xorl(in_out, in2); 10280} 10281 10282// Set N to predefined value 10283// Subtract from a lenght of a buffer 10284// execute in a loop: 10285// CRC_A = 0xFFFFFFFF, CRC_B = 0, CRC_C = 0 10286// for i = 1 to N do 10287// CRC_A = CRC32(CRC_A, A[i]) 10288// CRC_B = CRC32(CRC_B, B[i]) 10289// CRC_C = CRC32(CRC_C, C[i]) 10290// end for 10291// Recombine 10292void 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, 10293 Register in_out1, Register in_out2, Register in_out3, 10294 Register tmp1, Register tmp2, Register tmp3, 10295 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10296 Register tmp4, Register tmp5, 10297 Register n_tmp6) { 10298 Label L_processPartitions; 10299 Label L_processPartition; 10300 Label L_exit; 10301 10302 bind(L_processPartitions); 10303 cmpl(in_out1, 3 * size); 10304 jcc(Assembler::less, L_exit); 10305 xorl(tmp1, tmp1); 10306 xorl(tmp2, tmp2); 10307 movq(tmp3, in_out2); 10308 addq(tmp3, size); 10309 10310 bind(L_processPartition); 10311 crc32(in_out3, Address(in_out2, 0), 8); 10312 crc32(tmp1, Address(in_out2, size), 8); 10313 crc32(tmp2, Address(in_out2, size * 2), 8); 10314 addq(in_out2, 8); 10315 cmpq(in_out2, tmp3); 10316 jcc(Assembler::less, L_processPartition); 10317 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2, 10318 w_xtmp1, w_xtmp2, w_xtmp3, 10319 tmp4, tmp5, 10320 n_tmp6); 10321 addq(in_out2, 2 * size); 10322 subl(in_out1, 3 * size); 10323 jmp(L_processPartitions); 10324 10325 bind(L_exit); 10326} 10327#else 10328void MacroAssembler::crc32c_ipl_alg4(Register in_out, uint32_t n, 10329 Register tmp1, Register tmp2, Register tmp3, 10330 XMMRegister xtmp1, XMMRegister xtmp2) { 10331 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr())); 10332 if (n > 0) { 10333 addl(tmp3, n * 256 * 8); 10334 } 10335 // Q1 = TABLEExt[n][B & 0xFF]; 10336 movl(tmp1, in_out); 10337 andl(tmp1, 0x000000FF); 10338 shll(tmp1, 3); 10339 addl(tmp1, tmp3); 10340 movq(xtmp1, Address(tmp1, 0)); 10341 10342 // Q2 = TABLEExt[n][B >> 8 & 0xFF]; 10343 movl(tmp2, in_out); 10344 shrl(tmp2, 8); 10345 andl(tmp2, 0x000000FF); 10346 shll(tmp2, 3); 10347 addl(tmp2, tmp3); 10348 movq(xtmp2, Address(tmp2, 0)); 10349 10350 psllq(xtmp2, 8); 10351 pxor(xtmp1, xtmp2); 10352 10353 // Q3 = TABLEExt[n][B >> 16 & 0xFF]; 10354 movl(tmp2, in_out); 10355 shrl(tmp2, 16); 10356 andl(tmp2, 0x000000FF); 10357 shll(tmp2, 3); 10358 addl(tmp2, tmp3); 10359 movq(xtmp2, Address(tmp2, 0)); 10360 10361 psllq(xtmp2, 16); 10362 pxor(xtmp1, xtmp2); 10363 10364 // Q4 = TABLEExt[n][B >> 24 & 0xFF]; 10365 shrl(in_out, 24); 10366 andl(in_out, 0x000000FF); 10367 shll(in_out, 3); 10368 addl(in_out, tmp3); 10369 movq(xtmp2, Address(in_out, 0)); 10370 10371 psllq(xtmp2, 24); 10372 pxor(xtmp1, xtmp2); // Result in CXMM 10373 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24; 10374} 10375 10376void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1, 10377 Register in_out, 10378 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported, 10379 XMMRegister w_xtmp2, 10380 Register tmp1, 10381 Register n_tmp2, Register n_tmp3) { 10382 if (is_pclmulqdq_supported) { 10383 movdl(w_xtmp1, in_out); 10384 10385 movl(tmp1, const_or_pre_comp_const_index); 10386 movdl(w_xtmp2, tmp1); 10387 pclmulqdq(w_xtmp1, w_xtmp2, 0); 10388 // Keep result in XMM since GPR is 32 bit in length 10389 } else { 10390 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3, w_xtmp1, w_xtmp2); 10391 } 10392} 10393 10394void 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, 10395 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10396 Register tmp1, Register tmp2, 10397 Register n_tmp3) { 10398 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10399 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10400 10401 psllq(w_xtmp1, 1); 10402 movdl(tmp1, w_xtmp1); 10403 psrlq(w_xtmp1, 32); 10404 movdl(in_out, w_xtmp1); 10405 10406 xorl(tmp2, tmp2); 10407 crc32(tmp2, tmp1, 4); 10408 xorl(in_out, tmp2); 10409 10410 psllq(w_xtmp2, 1); 10411 movdl(tmp1, w_xtmp2); 10412 psrlq(w_xtmp2, 32); 10413 movdl(in1, w_xtmp2); 10414 10415 xorl(tmp2, tmp2); 10416 crc32(tmp2, tmp1, 4); 10417 xorl(in1, tmp2); 10418 xorl(in_out, in1); 10419 xorl(in_out, in2); 10420} 10421 10422void 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, 10423 Register in_out1, Register in_out2, Register in_out3, 10424 Register tmp1, Register tmp2, Register tmp3, 10425 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10426 Register tmp4, Register tmp5, 10427 Register n_tmp6) { 10428 Label L_processPartitions; 10429 Label L_processPartition; 10430 Label L_exit; 10431 10432 bind(L_processPartitions); 10433 cmpl(in_out1, 3 * size); 10434 jcc(Assembler::less, L_exit); 10435 xorl(tmp1, tmp1); 10436 xorl(tmp2, tmp2); 10437 movl(tmp3, in_out2); 10438 addl(tmp3, size); 10439 10440 bind(L_processPartition); 10441 crc32(in_out3, Address(in_out2, 0), 4); 10442 crc32(tmp1, Address(in_out2, size), 4); 10443 crc32(tmp2, Address(in_out2, size*2), 4); 10444 crc32(in_out3, Address(in_out2, 0+4), 4); 10445 crc32(tmp1, Address(in_out2, size+4), 4); 10446 crc32(tmp2, Address(in_out2, size*2+4), 4); 10447 addl(in_out2, 8); 10448 cmpl(in_out2, tmp3); 10449 jcc(Assembler::less, L_processPartition); 10450 10451 push(tmp3); 10452 push(in_out1); 10453 push(in_out2); 10454 tmp4 = tmp3; 10455 tmp5 = in_out1; 10456 n_tmp6 = in_out2; 10457 10458 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2, 10459 w_xtmp1, w_xtmp2, w_xtmp3, 10460 tmp4, tmp5, 10461 n_tmp6); 10462 10463 pop(in_out2); 10464 pop(in_out1); 10465 pop(tmp3); 10466 10467 addl(in_out2, 2 * size); 10468 subl(in_out1, 3 * size); 10469 jmp(L_processPartitions); 10470 10471 bind(L_exit); 10472} 10473#endif //LP64 10474 10475#ifdef _LP64 10476// Algorithm 2: Pipelined usage of the CRC32 instruction. 10477// Input: A buffer I of L bytes. 10478// Output: the CRC32C value of the buffer. 10479// Notations: 10480// Write L = 24N + r, with N = floor (L/24). 10481// r = L mod 24 (0 <= r < 24). 10482// Consider I as the concatenation of A|B|C|R, where A, B, C, each, 10483// N quadwords, and R consists of r bytes. 10484// A[j] = I [8j+7:8j], j= 0, 1, ..., N-1 10485// B[j] = I [N + 8j+7:N + 8j], j= 0, 1, ..., N-1 10486// C[j] = I [2N + 8j+7:2N + 8j], j= 0, 1, ..., N-1 10487// if r > 0 R[j] = I [3N +j], j= 0, 1, ...,r-1 10488void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2, 10489 Register tmp1, Register tmp2, Register tmp3, 10490 Register tmp4, Register tmp5, Register tmp6, 10491 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10492 bool is_pclmulqdq_supported) { 10493 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS]; 10494 Label L_wordByWord; 10495 Label L_byteByByteProlog; 10496 Label L_byteByByte; 10497 Label L_exit; 10498 10499 if (is_pclmulqdq_supported ) { 10500 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr; 10501 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr+1); 10502 10503 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2); 10504 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3); 10505 10506 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4); 10507 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5); 10508 assert((CRC32C_NUM_PRECOMPUTED_CONSTANTS - 1 ) == 5, "Checking whether you declared all of the constants based on the number of \"chunks\""); 10509 } else { 10510 const_or_pre_comp_const_index[0] = 1; 10511 const_or_pre_comp_const_index[1] = 0; 10512 10513 const_or_pre_comp_const_index[2] = 3; 10514 const_or_pre_comp_const_index[3] = 2; 10515 10516 const_or_pre_comp_const_index[4] = 5; 10517 const_or_pre_comp_const_index[5] = 4; 10518 } 10519 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported, 10520 in2, in1, in_out, 10521 tmp1, tmp2, tmp3, 10522 w_xtmp1, w_xtmp2, w_xtmp3, 10523 tmp4, tmp5, 10524 tmp6); 10525 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported, 10526 in2, in1, in_out, 10527 tmp1, tmp2, tmp3, 10528 w_xtmp1, w_xtmp2, w_xtmp3, 10529 tmp4, tmp5, 10530 tmp6); 10531 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported, 10532 in2, in1, in_out, 10533 tmp1, tmp2, tmp3, 10534 w_xtmp1, w_xtmp2, w_xtmp3, 10535 tmp4, tmp5, 10536 tmp6); 10537 movl(tmp1, in2); 10538 andl(tmp1, 0x00000007); 10539 negl(tmp1); 10540 addl(tmp1, in2); 10541 addq(tmp1, in1); 10542 10543 BIND(L_wordByWord); 10544 cmpq(in1, tmp1); 10545 jcc(Assembler::greaterEqual, L_byteByByteProlog); 10546 crc32(in_out, Address(in1, 0), 4); 10547 addq(in1, 4); 10548 jmp(L_wordByWord); 10549 10550 BIND(L_byteByByteProlog); 10551 andl(in2, 0x00000007); 10552 movl(tmp2, 1); 10553 10554 BIND(L_byteByByte); 10555 cmpl(tmp2, in2); 10556 jccb(Assembler::greater, L_exit); 10557 crc32(in_out, Address(in1, 0), 1); 10558 incq(in1); 10559 incl(tmp2); 10560 jmp(L_byteByByte); 10561 10562 BIND(L_exit); 10563} 10564#else 10565void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2, 10566 Register tmp1, Register tmp2, Register tmp3, 10567 Register tmp4, Register tmp5, Register tmp6, 10568 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10569 bool is_pclmulqdq_supported) { 10570 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS]; 10571 Label L_wordByWord; 10572 Label L_byteByByteProlog; 10573 Label L_byteByByte; 10574 Label L_exit; 10575 10576 if (is_pclmulqdq_supported) { 10577 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr; 10578 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 1); 10579 10580 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2); 10581 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3); 10582 10583 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4); 10584 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5); 10585 } else { 10586 const_or_pre_comp_const_index[0] = 1; 10587 const_or_pre_comp_const_index[1] = 0; 10588 10589 const_or_pre_comp_const_index[2] = 3; 10590 const_or_pre_comp_const_index[3] = 2; 10591 10592 const_or_pre_comp_const_index[4] = 5; 10593 const_or_pre_comp_const_index[5] = 4; 10594 } 10595 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported, 10596 in2, in1, in_out, 10597 tmp1, tmp2, tmp3, 10598 w_xtmp1, w_xtmp2, w_xtmp3, 10599 tmp4, tmp5, 10600 tmp6); 10601 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported, 10602 in2, in1, in_out, 10603 tmp1, tmp2, tmp3, 10604 w_xtmp1, w_xtmp2, w_xtmp3, 10605 tmp4, tmp5, 10606 tmp6); 10607 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported, 10608 in2, in1, in_out, 10609 tmp1, tmp2, tmp3, 10610 w_xtmp1, w_xtmp2, w_xtmp3, 10611 tmp4, tmp5, 10612 tmp6); 10613 movl(tmp1, in2); 10614 andl(tmp1, 0x00000007); 10615 negl(tmp1); 10616 addl(tmp1, in2); 10617 addl(tmp1, in1); 10618 10619 BIND(L_wordByWord); 10620 cmpl(in1, tmp1); 10621 jcc(Assembler::greaterEqual, L_byteByByteProlog); 10622 crc32(in_out, Address(in1,0), 4); 10623 addl(in1, 4); 10624 jmp(L_wordByWord); 10625 10626 BIND(L_byteByByteProlog); 10627 andl(in2, 0x00000007); 10628 movl(tmp2, 1); 10629 10630 BIND(L_byteByByte); 10631 cmpl(tmp2, in2); 10632 jccb(Assembler::greater, L_exit); 10633 movb(tmp1, Address(in1, 0)); 10634 crc32(in_out, tmp1, 1); 10635 incl(in1); 10636 incl(tmp2); 10637 jmp(L_byteByByte); 10638 10639 BIND(L_exit); 10640} 10641#endif // LP64 10642#undef BIND 10643#undef BLOCK_COMMENT 10644 10645 10646// Compress char[] array to byte[]. 10647void MacroAssembler::char_array_compress(Register src, Register dst, Register len, 10648 XMMRegister tmp1Reg, XMMRegister tmp2Reg, 10649 XMMRegister tmp3Reg, XMMRegister tmp4Reg, 10650 Register tmp5, Register result) { 10651 Label copy_chars_loop, return_length, return_zero, done; 10652 10653 // rsi: src 10654 // rdi: dst 10655 // rdx: len 10656 // rcx: tmp5 10657 // rax: result 10658 10659 // rsi holds start addr of source char[] to be compressed 10660 // rdi holds start addr of destination byte[] 10661 // rdx holds length 10662 10663 assert(len != result, ""); 10664 10665 // save length for return 10666 push(len); 10667 10668 if (UseSSE42Intrinsics) { 10669 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 10670 Label copy_32_loop, copy_16, copy_tail; 10671 10672 movl(result, len); 10673 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vectors 10674 10675 // vectored compression 10676 andl(len, 0xfffffff0); // vector count (in chars) 10677 andl(result, 0x0000000f); // tail count (in chars) 10678 testl(len, len); 10679 jccb(Assembler::zero, copy_16); 10680 10681 // compress 16 chars per iter 10682 movdl(tmp1Reg, tmp5); 10683 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg 10684 pxor(tmp4Reg, tmp4Reg); 10685 10686 lea(src, Address(src, len, Address::times_2)); 10687 lea(dst, Address(dst, len, Address::times_1)); 10688 negptr(len); 10689 10690 bind(copy_32_loop); 10691 movdqu(tmp2Reg, Address(src, len, Address::times_2)); // load 1st 8 characters 10692 por(tmp4Reg, tmp2Reg); 10693 movdqu(tmp3Reg, Address(src, len, Address::times_2, 16)); // load next 8 characters 10694 por(tmp4Reg, tmp3Reg); 10695 ptest(tmp4Reg, tmp1Reg); // check for Unicode chars in next vector 10696 jcc(Assembler::notZero, return_zero); 10697 packuswb(tmp2Reg, tmp3Reg); // only ASCII chars; compress each to 1 byte 10698 movdqu(Address(dst, len, Address::times_1), tmp2Reg); 10699 addptr(len, 16); 10700 jcc(Assembler::notZero, copy_32_loop); 10701 10702 // compress next vector of 8 chars (if any) 10703 bind(copy_16); 10704 movl(len, result); 10705 andl(len, 0xfffffff8); // vector count (in chars) 10706 andl(result, 0x00000007); // tail count (in chars) 10707 testl(len, len); 10708 jccb(Assembler::zero, copy_tail); 10709 10710 movdl(tmp1Reg, tmp5); 10711 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg 10712 pxor(tmp3Reg, tmp3Reg); 10713 10714 movdqu(tmp2Reg, Address(src, 0)); 10715 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 10716 jccb(Assembler::notZero, return_zero); 10717 packuswb(tmp2Reg, tmp3Reg); // only LATIN1 chars; compress each to 1 byte 10718 movq(Address(dst, 0), tmp2Reg); 10719 addptr(src, 16); 10720 addptr(dst, 8); 10721 10722 bind(copy_tail); 10723 movl(len, result); 10724 } 10725 // compress 1 char per iter 10726 testl(len, len); 10727 jccb(Assembler::zero, return_length); 10728 lea(src, Address(src, len, Address::times_2)); 10729 lea(dst, Address(dst, len, Address::times_1)); 10730 negptr(len); 10731 10732 bind(copy_chars_loop); 10733 load_unsigned_short(result, Address(src, len, Address::times_2)); 10734 testl(result, 0xff00); // check if Unicode char 10735 jccb(Assembler::notZero, return_zero); 10736 movb(Address(dst, len, Address::times_1), result); // ASCII char; compress to 1 byte 10737 increment(len); 10738 jcc(Assembler::notZero, copy_chars_loop); 10739 10740 // if compression succeeded, return length 10741 bind(return_length); 10742 pop(result); 10743 jmpb(done); 10744 10745 // if compression failed, return 0 10746 bind(return_zero); 10747 xorl(result, result); 10748 addptr(rsp, wordSize); 10749 10750 bind(done); 10751} 10752 10753// Inflate byte[] array to char[]. 10754void MacroAssembler::byte_array_inflate(Register src, Register dst, Register len, 10755 XMMRegister tmp1, Register tmp2) { 10756 Label copy_chars_loop, done; 10757 10758 // rsi: src 10759 // rdi: dst 10760 // rdx: len 10761 // rcx: tmp2 10762 10763 // rsi holds start addr of source byte[] to be inflated 10764 // rdi holds start addr of destination char[] 10765 // rdx holds length 10766 assert_different_registers(src, dst, len, tmp2); 10767 10768 if (UseSSE42Intrinsics) { 10769 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 10770 Label copy_8_loop, copy_bytes, copy_tail; 10771 10772 movl(tmp2, len); 10773 andl(tmp2, 0x00000007); // tail count (in chars) 10774 andl(len, 0xfffffff8); // vector count (in chars) 10775 jccb(Assembler::zero, copy_tail); 10776 10777 // vectored inflation 10778 lea(src, Address(src, len, Address::times_1)); 10779 lea(dst, Address(dst, len, Address::times_2)); 10780 negptr(len); 10781 10782 // inflate 8 chars per iter 10783 bind(copy_8_loop); 10784 pmovzxbw(tmp1, Address(src, len, Address::times_1)); // unpack to 8 words 10785 movdqu(Address(dst, len, Address::times_2), tmp1); 10786 addptr(len, 8); 10787 jcc(Assembler::notZero, copy_8_loop); 10788 10789 bind(copy_tail); 10790 movl(len, tmp2); 10791 10792 cmpl(len, 4); 10793 jccb(Assembler::less, copy_bytes); 10794 10795 movdl(tmp1, Address(src, 0)); // load 4 byte chars 10796 pmovzxbw(tmp1, tmp1); 10797 movq(Address(dst, 0), tmp1); 10798 subptr(len, 4); 10799 addptr(src, 4); 10800 addptr(dst, 8); 10801 10802 bind(copy_bytes); 10803 } 10804 testl(len, len); 10805 jccb(Assembler::zero, done); 10806 lea(src, Address(src, len, Address::times_1)); 10807 lea(dst, Address(dst, len, Address::times_2)); 10808 negptr(len); 10809 10810 // inflate 1 char per iter 10811 bind(copy_chars_loop); 10812 load_unsigned_byte(tmp2, Address(src, len, Address::times_1)); // load byte char 10813 movw(Address(dst, len, Address::times_2), tmp2); // inflate byte char to word 10814 increment(len); 10815 jcc(Assembler::notZero, copy_chars_loop); 10816 10817 bind(done); 10818} 10819 10820 10821Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) { 10822 switch (cond) { 10823 // Note some conditions are synonyms for others 10824 case Assembler::zero: return Assembler::notZero; 10825 case Assembler::notZero: return Assembler::zero; 10826 case Assembler::less: return Assembler::greaterEqual; 10827 case Assembler::lessEqual: return Assembler::greater; 10828 case Assembler::greater: return Assembler::lessEqual; 10829 case Assembler::greaterEqual: return Assembler::less; 10830 case Assembler::below: return Assembler::aboveEqual; 10831 case Assembler::belowEqual: return Assembler::above; 10832 case Assembler::above: return Assembler::belowEqual; 10833 case Assembler::aboveEqual: return Assembler::below; 10834 case Assembler::overflow: return Assembler::noOverflow; 10835 case Assembler::noOverflow: return Assembler::overflow; 10836 case Assembler::negative: return Assembler::positive; 10837 case Assembler::positive: return Assembler::negative; 10838 case Assembler::parity: return Assembler::noParity; 10839 case Assembler::noParity: return Assembler::parity; 10840 } 10841 ShouldNotReachHere(); return Assembler::overflow; 10842} 10843 10844SkipIfEqual::SkipIfEqual( 10845 MacroAssembler* masm, const bool* flag_addr, bool value) { 10846 _masm = masm; 10847 _masm->cmp8(ExternalAddress((address)flag_addr), value); 10848 _masm->jcc(Assembler::equal, _label); 10849} 10850 10851SkipIfEqual::~SkipIfEqual() { 10852 _masm->bind(_label); 10853} 10854 10855// 32-bit Windows has its own fast-path implementation 10856// of get_thread 10857#if !defined(WIN32) || defined(_LP64) 10858 10859// This is simply a call to Thread::current() 10860void MacroAssembler::get_thread(Register thread) { 10861 if (thread != rax) { 10862 push(rax); 10863 } 10864 LP64_ONLY(push(rdi);) 10865 LP64_ONLY(push(rsi);) 10866 push(rdx); 10867 push(rcx); 10868#ifdef _LP64 10869 push(r8); 10870 push(r9); 10871 push(r10); 10872 push(r11); 10873#endif 10874 10875 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, Thread::current), 0); 10876 10877#ifdef _LP64 10878 pop(r11); 10879 pop(r10); 10880 pop(r9); 10881 pop(r8); 10882#endif 10883 pop(rcx); 10884 pop(rdx); 10885 LP64_ONLY(pop(rsi);) 10886 LP64_ONLY(pop(rdi);) 10887 if (thread != rax) { 10888 mov(thread, rax); 10889 pop(rax); 10890 } 10891} 10892 10893#endif 10894