1 /* 2 * Copyright (c) 1997, 2018, 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 "jvm.h" 27 #include "asm/assembler.hpp" 28 #include "asm/assembler.inline.hpp" 29 #include "compiler/disassembler.hpp" 30 #include "gc/shared/cardTable.hpp" 31 #include "gc/shared/cardTableModRefBS.hpp" 32 #include "gc/shared/collectedHeap.inline.hpp" 33 #include "interpreter/interpreter.hpp" 34 #include "memory/resourceArea.hpp" 35 #include "memory/universe.hpp" 36 #include "oops/klass.inline.hpp" 37 #include "prims/methodHandles.hpp" 38 #include "runtime/biasedLocking.hpp" 39 #include "runtime/interfaceSupport.hpp" 40 #include "runtime/objectMonitor.hpp" 41 #include "runtime/os.hpp" 42 #include "runtime/safepoint.hpp" 43 #include "runtime/safepointMechanism.hpp" 44 #include "runtime/sharedRuntime.hpp" 45 #include "runtime/stubRoutines.hpp" 46 #include "runtime/thread.hpp" 47 #include "utilities/macros.hpp" 48 #if INCLUDE_ALL_GCS 49 #include "gc/g1/g1BarrierSet.hpp" 50 #include "gc/g1/g1CardTable.hpp" 51 #include "gc/g1/g1CollectedHeap.inline.hpp" 52 #include "gc/g1/heapRegion.hpp" 53 #endif // INCLUDE_ALL_GCS 54 #include "crc32c.h" 55 #ifdef COMPILER2 56 #include "opto/intrinsicnode.hpp" 57 #endif 58 59 #ifdef PRODUCT 60 #define BLOCK_COMMENT(str) /* nothing */ 61 #define STOP(error) stop(error) 62 #else 63 #define BLOCK_COMMENT(str) block_comment(str) 64 #define STOP(error) block_comment(error); stop(error) 65 #endif 66 67 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":") 68 69 #ifdef ASSERT 70 bool AbstractAssembler::pd_check_instruction_mark() { return true; } 71 #endif 72 73 static Assembler::Condition reverse[] = { 74 Assembler::noOverflow /* overflow = 0x0 */ , 75 Assembler::overflow /* noOverflow = 0x1 */ , 76 Assembler::aboveEqual /* carrySet = 0x2, below = 0x2 */ , 77 Assembler::below /* aboveEqual = 0x3, carryClear = 0x3 */ , 78 Assembler::notZero /* zero = 0x4, equal = 0x4 */ , 79 Assembler::zero /* notZero = 0x5, notEqual = 0x5 */ , 80 Assembler::above /* belowEqual = 0x6 */ , 81 Assembler::belowEqual /* above = 0x7 */ , 82 Assembler::positive /* negative = 0x8 */ , 83 Assembler::negative /* positive = 0x9 */ , 84 Assembler::noParity /* parity = 0xa */ , 85 Assembler::parity /* noParity = 0xb */ , 86 Assembler::greaterEqual /* less = 0xc */ , 87 Assembler::less /* greaterEqual = 0xd */ , 88 Assembler::greater /* lessEqual = 0xe */ , 89 Assembler::lessEqual /* greater = 0xf, */ 90 91 }; 92 93 94 // Implementation of MacroAssembler 95 96 // First all the versions that have distinct versions depending on 32/64 bit 97 // Unless the difference is trivial (1 line or so). 98 99 #ifndef _LP64 100 101 // 32bit versions 102 103 Address MacroAssembler::as_Address(AddressLiteral adr) { 104 return Address(adr.target(), adr.rspec()); 105 } 106 107 Address MacroAssembler::as_Address(ArrayAddress adr) { 108 return Address::make_array(adr); 109 } 110 111 void MacroAssembler::call_VM_leaf_base(address entry_point, 112 int number_of_arguments) { 113 call(RuntimeAddress(entry_point)); 114 increment(rsp, number_of_arguments * wordSize); 115 } 116 117 void MacroAssembler::cmpklass(Address src1, Metadata* obj) { 118 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 119 } 120 121 void MacroAssembler::cmpklass(Register src1, Metadata* obj) { 122 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 123 } 124 125 void MacroAssembler::cmpoop(Address src1, jobject obj) { 126 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate()); 127 } 128 129 void MacroAssembler::cmpoop(Register src1, jobject obj) { 130 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate()); 131 } 132 133 void MacroAssembler::extend_sign(Register hi, Register lo) { 134 // According to Intel Doc. AP-526, "Integer Divide", p.18. 135 if (VM_Version::is_P6() && hi == rdx && lo == rax) { 136 cdql(); 137 } else { 138 movl(hi, lo); 139 sarl(hi, 31); 140 } 141 } 142 143 void MacroAssembler::jC2(Register tmp, Label& L) { 144 // set parity bit if FPU flag C2 is set (via rax) 145 save_rax(tmp); 146 fwait(); fnstsw_ax(); 147 sahf(); 148 restore_rax(tmp); 149 // branch 150 jcc(Assembler::parity, L); 151 } 152 153 void MacroAssembler::jnC2(Register tmp, Label& L) { 154 // set parity bit if FPU flag C2 is set (via rax) 155 save_rax(tmp); 156 fwait(); fnstsw_ax(); 157 sahf(); 158 restore_rax(tmp); 159 // branch 160 jcc(Assembler::noParity, L); 161 } 162 163 // 32bit can do a case table jump in one instruction but we no longer allow the base 164 // to be installed in the Address class 165 void MacroAssembler::jump(ArrayAddress entry) { 166 jmp(as_Address(entry)); 167 } 168 169 // Note: y_lo will be destroyed 170 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) { 171 // Long compare for Java (semantics as described in JVM spec.) 172 Label high, low, done; 173 174 cmpl(x_hi, y_hi); 175 jcc(Assembler::less, low); 176 jcc(Assembler::greater, high); 177 // x_hi is the return register 178 xorl(x_hi, x_hi); 179 cmpl(x_lo, y_lo); 180 jcc(Assembler::below, low); 181 jcc(Assembler::equal, done); 182 183 bind(high); 184 xorl(x_hi, x_hi); 185 increment(x_hi); 186 jmp(done); 187 188 bind(low); 189 xorl(x_hi, x_hi); 190 decrementl(x_hi); 191 192 bind(done); 193 } 194 195 void MacroAssembler::lea(Register dst, AddressLiteral src) { 196 mov_literal32(dst, (int32_t)src.target(), src.rspec()); 197 } 198 199 void MacroAssembler::lea(Address dst, AddressLiteral adr) { 200 // leal(dst, as_Address(adr)); 201 // see note in movl as to why we must use a move 202 mov_literal32(dst, (int32_t) adr.target(), adr.rspec()); 203 } 204 205 void MacroAssembler::leave() { 206 mov(rsp, rbp); 207 pop(rbp); 208 } 209 210 void MacroAssembler::lmul(int x_rsp_offset, int y_rsp_offset) { 211 // Multiplication of two Java long values stored on the stack 212 // as illustrated below. Result is in rdx:rax. 213 // 214 // rsp ---> [ ?? ] \ \ 215 // .... | y_rsp_offset | 216 // [ y_lo ] / (in bytes) | x_rsp_offset 217 // [ y_hi ] | (in bytes) 218 // .... | 219 // [ x_lo ] / 220 // [ x_hi ] 221 // .... 222 // 223 // Basic idea: lo(result) = lo(x_lo * y_lo) 224 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi) 225 Address x_hi(rsp, x_rsp_offset + wordSize); Address x_lo(rsp, x_rsp_offset); 226 Address y_hi(rsp, y_rsp_offset + wordSize); Address y_lo(rsp, y_rsp_offset); 227 Label quick; 228 // load x_hi, y_hi and check if quick 229 // multiplication is possible 230 movl(rbx, x_hi); 231 movl(rcx, y_hi); 232 movl(rax, rbx); 233 orl(rbx, rcx); // rbx, = 0 <=> x_hi = 0 and y_hi = 0 234 jcc(Assembler::zero, quick); // if rbx, = 0 do quick multiply 235 // do full multiplication 236 // 1st step 237 mull(y_lo); // x_hi * y_lo 238 movl(rbx, rax); // save lo(x_hi * y_lo) in rbx, 239 // 2nd step 240 movl(rax, x_lo); 241 mull(rcx); // x_lo * y_hi 242 addl(rbx, rax); // add lo(x_lo * y_hi) to rbx, 243 // 3rd step 244 bind(quick); // note: rbx, = 0 if quick multiply! 245 movl(rax, x_lo); 246 mull(y_lo); // x_lo * y_lo 247 addl(rdx, rbx); // correct hi(x_lo * y_lo) 248 } 249 250 void MacroAssembler::lneg(Register hi, Register lo) { 251 negl(lo); 252 adcl(hi, 0); 253 negl(hi); 254 } 255 256 void MacroAssembler::lshl(Register hi, Register lo) { 257 // Java shift left long support (semantics as described in JVM spec., p.305) 258 // (basic idea for shift counts s >= n: x << s == (x << n) << (s - n)) 259 // shift value is in rcx ! 260 assert(hi != rcx, "must not use rcx"); 261 assert(lo != rcx, "must not use rcx"); 262 const Register s = rcx; // shift count 263 const int n = BitsPerWord; 264 Label L; 265 andl(s, 0x3f); // s := s & 0x3f (s < 0x40) 266 cmpl(s, n); // if (s < n) 267 jcc(Assembler::less, L); // else (s >= n) 268 movl(hi, lo); // x := x << n 269 xorl(lo, lo); 270 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n! 271 bind(L); // s (mod n) < n 272 shldl(hi, lo); // x := x << s 273 shll(lo); 274 } 275 276 277 void MacroAssembler::lshr(Register hi, Register lo, bool sign_extension) { 278 // Java shift right long support (semantics as described in JVM spec., p.306 & p.310) 279 // (basic idea for shift counts s >= n: x >> s == (x >> n) >> (s - n)) 280 assert(hi != rcx, "must not use rcx"); 281 assert(lo != rcx, "must not use rcx"); 282 const Register s = rcx; // shift count 283 const int n = BitsPerWord; 284 Label L; 285 andl(s, 0x3f); // s := s & 0x3f (s < 0x40) 286 cmpl(s, n); // if (s < n) 287 jcc(Assembler::less, L); // else (s >= n) 288 movl(lo, hi); // x := x >> n 289 if (sign_extension) sarl(hi, 31); 290 else xorl(hi, hi); 291 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n! 292 bind(L); // s (mod n) < n 293 shrdl(lo, hi); // x := x >> s 294 if (sign_extension) sarl(hi); 295 else shrl(hi); 296 } 297 298 void MacroAssembler::movoop(Register dst, jobject obj) { 299 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate()); 300 } 301 302 void MacroAssembler::movoop(Address dst, jobject obj) { 303 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate()); 304 } 305 306 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) { 307 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 308 } 309 310 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) { 311 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 312 } 313 314 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) { 315 // scratch register is not used, 316 // it is defined to match parameters of 64-bit version of this method. 317 if (src.is_lval()) { 318 mov_literal32(dst, (intptr_t)src.target(), src.rspec()); 319 } else { 320 movl(dst, as_Address(src)); 321 } 322 } 323 324 void MacroAssembler::movptr(ArrayAddress dst, Register src) { 325 movl(as_Address(dst), src); 326 } 327 328 void MacroAssembler::movptr(Register dst, ArrayAddress src) { 329 movl(dst, as_Address(src)); 330 } 331 332 // src should NEVER be a real pointer. Use AddressLiteral for true pointers 333 void MacroAssembler::movptr(Address dst, intptr_t src) { 334 movl(dst, src); 335 } 336 337 338 void MacroAssembler::pop_callee_saved_registers() { 339 pop(rcx); 340 pop(rdx); 341 pop(rdi); 342 pop(rsi); 343 } 344 345 void MacroAssembler::pop_fTOS() { 346 fld_d(Address(rsp, 0)); 347 addl(rsp, 2 * wordSize); 348 } 349 350 void MacroAssembler::push_callee_saved_registers() { 351 push(rsi); 352 push(rdi); 353 push(rdx); 354 push(rcx); 355 } 356 357 void MacroAssembler::push_fTOS() { 358 subl(rsp, 2 * wordSize); 359 fstp_d(Address(rsp, 0)); 360 } 361 362 363 void MacroAssembler::pushoop(jobject obj) { 364 push_literal32((int32_t)obj, oop_Relocation::spec_for_immediate()); 365 } 366 367 void MacroAssembler::pushklass(Metadata* obj) { 368 push_literal32((int32_t)obj, metadata_Relocation::spec_for_immediate()); 369 } 370 371 void MacroAssembler::pushptr(AddressLiteral src) { 372 if (src.is_lval()) { 373 push_literal32((int32_t)src.target(), src.rspec()); 374 } else { 375 pushl(as_Address(src)); 376 } 377 } 378 379 void MacroAssembler::set_word_if_not_zero(Register dst) { 380 xorl(dst, dst); 381 set_byte_if_not_zero(dst); 382 } 383 384 static void pass_arg0(MacroAssembler* masm, Register arg) { 385 masm->push(arg); 386 } 387 388 static void pass_arg1(MacroAssembler* masm, Register arg) { 389 masm->push(arg); 390 } 391 392 static void pass_arg2(MacroAssembler* masm, Register arg) { 393 masm->push(arg); 394 } 395 396 static void pass_arg3(MacroAssembler* masm, Register arg) { 397 masm->push(arg); 398 } 399 400 #ifndef PRODUCT 401 extern "C" void findpc(intptr_t x); 402 #endif 403 404 void MacroAssembler::debug32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip, char* msg) { 405 // In order to get locks to work, we need to fake a in_VM state 406 JavaThread* thread = JavaThread::current(); 407 JavaThreadState saved_state = thread->thread_state(); 408 thread->set_thread_state(_thread_in_vm); 409 if (ShowMessageBoxOnError) { 410 JavaThread* thread = JavaThread::current(); 411 JavaThreadState saved_state = thread->thread_state(); 412 thread->set_thread_state(_thread_in_vm); 413 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) { 414 ttyLocker ttyl; 415 BytecodeCounter::print(); 416 } 417 // To see where a verify_oop failed, get $ebx+40/X for this frame. 418 // This is the value of eip which points to where verify_oop will return. 419 if (os::message_box(msg, "Execution stopped, print registers?")) { 420 print_state32(rdi, rsi, rbp, rsp, rbx, rdx, rcx, rax, eip); 421 BREAKPOINT; 422 } 423 } else { 424 ttyLocker ttyl; 425 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", msg); 426 } 427 // Don't assert holding the ttyLock 428 assert(false, "DEBUG MESSAGE: %s", msg); 429 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state); 430 } 431 432 void MacroAssembler::print_state32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip) { 433 ttyLocker ttyl; 434 FlagSetting fs(Debugging, true); 435 tty->print_cr("eip = 0x%08x", eip); 436 #ifndef PRODUCT 437 if ((WizardMode || Verbose) && PrintMiscellaneous) { 438 tty->cr(); 439 findpc(eip); 440 tty->cr(); 441 } 442 #endif 443 #define PRINT_REG(rax) \ 444 { tty->print("%s = ", #rax); os::print_location(tty, rax); } 445 PRINT_REG(rax); 446 PRINT_REG(rbx); 447 PRINT_REG(rcx); 448 PRINT_REG(rdx); 449 PRINT_REG(rdi); 450 PRINT_REG(rsi); 451 PRINT_REG(rbp); 452 PRINT_REG(rsp); 453 #undef PRINT_REG 454 // Print some words near top of staack. 455 int* dump_sp = (int*) rsp; 456 for (int col1 = 0; col1 < 8; col1++) { 457 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 458 os::print_location(tty, *dump_sp++); 459 } 460 for (int row = 0; row < 16; row++) { 461 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 462 for (int col = 0; col < 8; col++) { 463 tty->print(" 0x%08x", *dump_sp++); 464 } 465 tty->cr(); 466 } 467 // Print some instructions around pc: 468 Disassembler::decode((address)eip-64, (address)eip); 469 tty->print_cr("--------"); 470 Disassembler::decode((address)eip, (address)eip+32); 471 } 472 473 void MacroAssembler::stop(const char* msg) { 474 ExternalAddress message((address)msg); 475 // push address of message 476 pushptr(message.addr()); 477 { Label L; call(L, relocInfo::none); bind(L); } // push eip 478 pusha(); // push registers 479 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug32))); 480 hlt(); 481 } 482 483 void MacroAssembler::warn(const char* msg) { 484 push_CPU_state(); 485 486 ExternalAddress message((address) msg); 487 // push address of message 488 pushptr(message.addr()); 489 490 call(RuntimeAddress(CAST_FROM_FN_PTR(address, warning))); 491 addl(rsp, wordSize); // discard argument 492 pop_CPU_state(); 493 } 494 495 void MacroAssembler::print_state() { 496 { Label L; call(L, relocInfo::none); bind(L); } // push eip 497 pusha(); // push registers 498 499 push_CPU_state(); 500 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::print_state32))); 501 pop_CPU_state(); 502 503 popa(); 504 addl(rsp, wordSize); 505 } 506 507 #else // _LP64 508 509 // 64 bit versions 510 511 Address MacroAssembler::as_Address(AddressLiteral adr) { 512 // amd64 always does this as a pc-rel 513 // we can be absolute or disp based on the instruction type 514 // jmp/call are displacements others are absolute 515 assert(!adr.is_lval(), "must be rval"); 516 assert(reachable(adr), "must be"); 517 return Address((int32_t)(intptr_t)(adr.target() - pc()), adr.target(), adr.reloc()); 518 519 } 520 521 Address MacroAssembler::as_Address(ArrayAddress adr) { 522 AddressLiteral base = adr.base(); 523 lea(rscratch1, base); 524 Address index = adr.index(); 525 assert(index._disp == 0, "must not have disp"); // maybe it can? 526 Address array(rscratch1, index._index, index._scale, index._disp); 527 return array; 528 } 529 530 void MacroAssembler::call_VM_leaf_base(address entry_point, int num_args) { 531 Label L, E; 532 533 #ifdef _WIN64 534 // Windows always allocates space for it's register args 535 assert(num_args <= 4, "only register arguments supported"); 536 subq(rsp, frame::arg_reg_save_area_bytes); 537 #endif 538 539 // Align stack if necessary 540 testl(rsp, 15); 541 jcc(Assembler::zero, L); 542 543 subq(rsp, 8); 544 { 545 call(RuntimeAddress(entry_point)); 546 } 547 addq(rsp, 8); 548 jmp(E); 549 550 bind(L); 551 { 552 call(RuntimeAddress(entry_point)); 553 } 554 555 bind(E); 556 557 #ifdef _WIN64 558 // restore stack pointer 559 addq(rsp, frame::arg_reg_save_area_bytes); 560 #endif 561 562 } 563 564 void MacroAssembler::cmp64(Register src1, AddressLiteral src2) { 565 assert(!src2.is_lval(), "should use cmpptr"); 566 567 if (reachable(src2)) { 568 cmpq(src1, as_Address(src2)); 569 } else { 570 lea(rscratch1, src2); 571 Assembler::cmpq(src1, Address(rscratch1, 0)); 572 } 573 } 574 575 int MacroAssembler::corrected_idivq(Register reg) { 576 // Full implementation of Java ldiv and lrem; checks for special 577 // case as described in JVM spec., p.243 & p.271. The function 578 // returns the (pc) offset of the idivl instruction - may be needed 579 // for implicit exceptions. 580 // 581 // normal case special case 582 // 583 // input : rax: dividend min_long 584 // reg: divisor (may not be eax/edx) -1 585 // 586 // output: rax: quotient (= rax idiv reg) min_long 587 // rdx: remainder (= rax irem reg) 0 588 assert(reg != rax && reg != rdx, "reg cannot be rax or rdx register"); 589 static const int64_t min_long = 0x8000000000000000; 590 Label normal_case, special_case; 591 592 // check for special case 593 cmp64(rax, ExternalAddress((address) &min_long)); 594 jcc(Assembler::notEqual, normal_case); 595 xorl(rdx, rdx); // prepare rdx for possible special case (where 596 // remainder = 0) 597 cmpq(reg, -1); 598 jcc(Assembler::equal, special_case); 599 600 // handle normal case 601 bind(normal_case); 602 cdqq(); 603 int idivq_offset = offset(); 604 idivq(reg); 605 606 // normal and special case exit 607 bind(special_case); 608 609 return idivq_offset; 610 } 611 612 void MacroAssembler::decrementq(Register reg, int value) { 613 if (value == min_jint) { subq(reg, value); return; } 614 if (value < 0) { incrementq(reg, -value); return; } 615 if (value == 0) { ; return; } 616 if (value == 1 && UseIncDec) { decq(reg) ; return; } 617 /* else */ { subq(reg, value) ; return; } 618 } 619 620 void MacroAssembler::decrementq(Address dst, int value) { 621 if (value == min_jint) { subq(dst, value); return; } 622 if (value < 0) { incrementq(dst, -value); return; } 623 if (value == 0) { ; return; } 624 if (value == 1 && UseIncDec) { decq(dst) ; return; } 625 /* else */ { subq(dst, value) ; return; } 626 } 627 628 void MacroAssembler::incrementq(AddressLiteral dst) { 629 if (reachable(dst)) { 630 incrementq(as_Address(dst)); 631 } else { 632 lea(rscratch1, dst); 633 incrementq(Address(rscratch1, 0)); 634 } 635 } 636 637 void MacroAssembler::incrementq(Register reg, int value) { 638 if (value == min_jint) { addq(reg, value); return; } 639 if (value < 0) { decrementq(reg, -value); return; } 640 if (value == 0) { ; return; } 641 if (value == 1 && UseIncDec) { incq(reg) ; return; } 642 /* else */ { addq(reg, value) ; return; } 643 } 644 645 void MacroAssembler::incrementq(Address dst, int value) { 646 if (value == min_jint) { addq(dst, value); return; } 647 if (value < 0) { decrementq(dst, -value); return; } 648 if (value == 0) { ; return; } 649 if (value == 1 && UseIncDec) { incq(dst) ; return; } 650 /* else */ { addq(dst, value) ; return; } 651 } 652 653 // 32bit can do a case table jump in one instruction but we no longer allow the base 654 // to be installed in the Address class 655 void MacroAssembler::jump(ArrayAddress entry) { 656 lea(rscratch1, entry.base()); 657 Address dispatch = entry.index(); 658 assert(dispatch._base == noreg, "must be"); 659 dispatch._base = rscratch1; 660 jmp(dispatch); 661 } 662 663 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) { 664 ShouldNotReachHere(); // 64bit doesn't use two regs 665 cmpq(x_lo, y_lo); 666 } 667 668 void MacroAssembler::lea(Register dst, AddressLiteral src) { 669 mov_literal64(dst, (intptr_t)src.target(), src.rspec()); 670 } 671 672 void MacroAssembler::lea(Address dst, AddressLiteral adr) { 673 mov_literal64(rscratch1, (intptr_t)adr.target(), adr.rspec()); 674 movptr(dst, rscratch1); 675 } 676 677 void MacroAssembler::leave() { 678 // %%% is this really better? Why not on 32bit too? 679 emit_int8((unsigned char)0xC9); // LEAVE 680 } 681 682 void MacroAssembler::lneg(Register hi, Register lo) { 683 ShouldNotReachHere(); // 64bit doesn't use two regs 684 negq(lo); 685 } 686 687 void MacroAssembler::movoop(Register dst, jobject obj) { 688 mov_literal64(dst, (intptr_t)obj, oop_Relocation::spec_for_immediate()); 689 } 690 691 void MacroAssembler::movoop(Address dst, jobject obj) { 692 mov_literal64(rscratch1, (intptr_t)obj, oop_Relocation::spec_for_immediate()); 693 movq(dst, rscratch1); 694 } 695 696 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) { 697 mov_literal64(dst, (intptr_t)obj, metadata_Relocation::spec_for_immediate()); 698 } 699 700 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) { 701 mov_literal64(rscratch1, (intptr_t)obj, metadata_Relocation::spec_for_immediate()); 702 movq(dst, rscratch1); 703 } 704 705 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) { 706 if (src.is_lval()) { 707 mov_literal64(dst, (intptr_t)src.target(), src.rspec()); 708 } else { 709 if (reachable(src)) { 710 movq(dst, as_Address(src)); 711 } else { 712 lea(scratch, src); 713 movq(dst, Address(scratch, 0)); 714 } 715 } 716 } 717 718 void MacroAssembler::movptr(ArrayAddress dst, Register src) { 719 movq(as_Address(dst), src); 720 } 721 722 void MacroAssembler::movptr(Register dst, ArrayAddress src) { 723 movq(dst, as_Address(src)); 724 } 725 726 // src should NEVER be a real pointer. Use AddressLiteral for true pointers 727 void MacroAssembler::movptr(Address dst, intptr_t src) { 728 mov64(rscratch1, src); 729 movq(dst, rscratch1); 730 } 731 732 // These are mostly for initializing NULL 733 void MacroAssembler::movptr(Address dst, int32_t src) { 734 movslq(dst, src); 735 } 736 737 void MacroAssembler::movptr(Register dst, int32_t src) { 738 mov64(dst, (intptr_t)src); 739 } 740 741 void MacroAssembler::pushoop(jobject obj) { 742 movoop(rscratch1, obj); 743 push(rscratch1); 744 } 745 746 void MacroAssembler::pushklass(Metadata* obj) { 747 mov_metadata(rscratch1, obj); 748 push(rscratch1); 749 } 750 751 void MacroAssembler::pushptr(AddressLiteral src) { 752 lea(rscratch1, src); 753 if (src.is_lval()) { 754 push(rscratch1); 755 } else { 756 pushq(Address(rscratch1, 0)); 757 } 758 } 759 760 void MacroAssembler::reset_last_Java_frame(bool clear_fp) { 761 // we must set sp to zero to clear frame 762 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), NULL_WORD); 763 // must clear fp, so that compiled frames are not confused; it is 764 // possible that we need it only for debugging 765 if (clear_fp) { 766 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD); 767 } 768 769 // Always clear the pc because it could have been set by make_walkable() 770 movptr(Address(r15_thread, JavaThread::last_Java_pc_offset()), NULL_WORD); 771 vzeroupper(); 772 } 773 774 void MacroAssembler::set_last_Java_frame(Register last_java_sp, 775 Register last_java_fp, 776 address last_java_pc) { 777 vzeroupper(); 778 // determine last_java_sp register 779 if (!last_java_sp->is_valid()) { 780 last_java_sp = rsp; 781 } 782 783 // last_java_fp is optional 784 if (last_java_fp->is_valid()) { 785 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), 786 last_java_fp); 787 } 788 789 // last_java_pc is optional 790 if (last_java_pc != NULL) { 791 Address java_pc(r15_thread, 792 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()); 793 lea(rscratch1, InternalAddress(last_java_pc)); 794 movptr(java_pc, rscratch1); 795 } 796 797 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), last_java_sp); 798 } 799 800 static void pass_arg0(MacroAssembler* masm, Register arg) { 801 if (c_rarg0 != arg ) { 802 masm->mov(c_rarg0, arg); 803 } 804 } 805 806 static void pass_arg1(MacroAssembler* masm, Register arg) { 807 if (c_rarg1 != arg ) { 808 masm->mov(c_rarg1, arg); 809 } 810 } 811 812 static void pass_arg2(MacroAssembler* masm, Register arg) { 813 if (c_rarg2 != arg ) { 814 masm->mov(c_rarg2, arg); 815 } 816 } 817 818 static void pass_arg3(MacroAssembler* masm, Register arg) { 819 if (c_rarg3 != arg ) { 820 masm->mov(c_rarg3, arg); 821 } 822 } 823 824 void MacroAssembler::stop(const char* msg) { 825 address rip = pc(); 826 pusha(); // get regs on stack 827 lea(c_rarg0, ExternalAddress((address) msg)); 828 lea(c_rarg1, InternalAddress(rip)); 829 movq(c_rarg2, rsp); // pass pointer to regs array 830 andq(rsp, -16); // align stack as required by ABI 831 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64))); 832 hlt(); 833 } 834 835 void MacroAssembler::warn(const char* msg) { 836 push(rbp); 837 movq(rbp, rsp); 838 andq(rsp, -16); // align stack as required by push_CPU_state and call 839 push_CPU_state(); // keeps alignment at 16 bytes 840 lea(c_rarg0, ExternalAddress((address) msg)); 841 lea(rax, ExternalAddress(CAST_FROM_FN_PTR(address, warning))); 842 call(rax); 843 pop_CPU_state(); 844 mov(rsp, rbp); 845 pop(rbp); 846 } 847 848 void MacroAssembler::print_state() { 849 address rip = pc(); 850 pusha(); // get regs on stack 851 push(rbp); 852 movq(rbp, rsp); 853 andq(rsp, -16); // align stack as required by push_CPU_state and call 854 push_CPU_state(); // keeps alignment at 16 bytes 855 856 lea(c_rarg0, InternalAddress(rip)); 857 lea(c_rarg1, Address(rbp, wordSize)); // pass pointer to regs array 858 call_VM_leaf(CAST_FROM_FN_PTR(address, MacroAssembler::print_state64), c_rarg0, c_rarg1); 859 860 pop_CPU_state(); 861 mov(rsp, rbp); 862 pop(rbp); 863 popa(); 864 } 865 866 #ifndef PRODUCT 867 extern "C" void findpc(intptr_t x); 868 #endif 869 870 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[]) { 871 // In order to get locks to work, we need to fake a in_VM state 872 if (ShowMessageBoxOnError) { 873 JavaThread* thread = JavaThread::current(); 874 JavaThreadState saved_state = thread->thread_state(); 875 thread->set_thread_state(_thread_in_vm); 876 #ifndef PRODUCT 877 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) { 878 ttyLocker ttyl; 879 BytecodeCounter::print(); 880 } 881 #endif 882 // To see where a verify_oop failed, get $ebx+40/X for this frame. 883 // XXX correct this offset for amd64 884 // This is the value of eip which points to where verify_oop will return. 885 if (os::message_box(msg, "Execution stopped, print registers?")) { 886 print_state64(pc, regs); 887 BREAKPOINT; 888 assert(false, "start up GDB"); 889 } 890 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state); 891 } else { 892 ttyLocker ttyl; 893 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", 894 msg); 895 assert(false, "DEBUG MESSAGE: %s", msg); 896 } 897 } 898 899 void MacroAssembler::print_state64(int64_t pc, int64_t regs[]) { 900 ttyLocker ttyl; 901 FlagSetting fs(Debugging, true); 902 tty->print_cr("rip = 0x%016lx", (intptr_t)pc); 903 #ifndef PRODUCT 904 tty->cr(); 905 findpc(pc); 906 tty->cr(); 907 #endif 908 #define PRINT_REG(rax, value) \ 909 { tty->print("%s = ", #rax); os::print_location(tty, value); } 910 PRINT_REG(rax, regs[15]); 911 PRINT_REG(rbx, regs[12]); 912 PRINT_REG(rcx, regs[14]); 913 PRINT_REG(rdx, regs[13]); 914 PRINT_REG(rdi, regs[8]); 915 PRINT_REG(rsi, regs[9]); 916 PRINT_REG(rbp, regs[10]); 917 PRINT_REG(rsp, regs[11]); 918 PRINT_REG(r8 , regs[7]); 919 PRINT_REG(r9 , regs[6]); 920 PRINT_REG(r10, regs[5]); 921 PRINT_REG(r11, regs[4]); 922 PRINT_REG(r12, regs[3]); 923 PRINT_REG(r13, regs[2]); 924 PRINT_REG(r14, regs[1]); 925 PRINT_REG(r15, regs[0]); 926 #undef PRINT_REG 927 // Print some words near top of staack. 928 int64_t* rsp = (int64_t*) regs[11]; 929 int64_t* dump_sp = rsp; 930 for (int col1 = 0; col1 < 8; col1++) { 931 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 932 os::print_location(tty, *dump_sp++); 933 } 934 for (int row = 0; row < 25; row++) { 935 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 936 for (int col = 0; col < 4; col++) { 937 tty->print(" 0x%016lx", (intptr_t)*dump_sp++); 938 } 939 tty->cr(); 940 } 941 // Print some instructions around pc: 942 Disassembler::decode((address)pc-64, (address)pc); 943 tty->print_cr("--------"); 944 Disassembler::decode((address)pc, (address)pc+32); 945 } 946 947 #endif // _LP64 948 949 // Now versions that are common to 32/64 bit 950 951 void MacroAssembler::addptr(Register dst, int32_t imm32) { 952 LP64_ONLY(addq(dst, imm32)) NOT_LP64(addl(dst, imm32)); 953 } 954 955 void MacroAssembler::addptr(Register dst, Register src) { 956 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src)); 957 } 958 959 void MacroAssembler::addptr(Address dst, Register src) { 960 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src)); 961 } 962 963 void MacroAssembler::addsd(XMMRegister dst, AddressLiteral src) { 964 if (reachable(src)) { 965 Assembler::addsd(dst, as_Address(src)); 966 } else { 967 lea(rscratch1, src); 968 Assembler::addsd(dst, Address(rscratch1, 0)); 969 } 970 } 971 972 void MacroAssembler::addss(XMMRegister dst, AddressLiteral src) { 973 if (reachable(src)) { 974 addss(dst, as_Address(src)); 975 } else { 976 lea(rscratch1, src); 977 addss(dst, Address(rscratch1, 0)); 978 } 979 } 980 981 void MacroAssembler::addpd(XMMRegister dst, AddressLiteral src) { 982 if (reachable(src)) { 983 Assembler::addpd(dst, as_Address(src)); 984 } else { 985 lea(rscratch1, src); 986 Assembler::addpd(dst, Address(rscratch1, 0)); 987 } 988 } 989 990 void MacroAssembler::align(int modulus) { 991 align(modulus, offset()); 992 } 993 994 void MacroAssembler::align(int modulus, int target) { 995 if (target % modulus != 0) { 996 nop(modulus - (target % modulus)); 997 } 998 } 999 1000 void MacroAssembler::andpd(XMMRegister dst, AddressLiteral src) { 1001 // Used in sign-masking with aligned address. 1002 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 1003 if (reachable(src)) { 1004 Assembler::andpd(dst, as_Address(src)); 1005 } else { 1006 lea(rscratch1, src); 1007 Assembler::andpd(dst, Address(rscratch1, 0)); 1008 } 1009 } 1010 1011 void MacroAssembler::andps(XMMRegister dst, AddressLiteral src) { 1012 // Used in sign-masking with aligned address. 1013 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 1014 if (reachable(src)) { 1015 Assembler::andps(dst, as_Address(src)); 1016 } else { 1017 lea(rscratch1, src); 1018 Assembler::andps(dst, Address(rscratch1, 0)); 1019 } 1020 } 1021 1022 void MacroAssembler::andptr(Register dst, int32_t imm32) { 1023 LP64_ONLY(andq(dst, imm32)) NOT_LP64(andl(dst, imm32)); 1024 } 1025 1026 void MacroAssembler::atomic_incl(Address counter_addr) { 1027 if (os::is_MP()) 1028 lock(); 1029 incrementl(counter_addr); 1030 } 1031 1032 void MacroAssembler::atomic_incl(AddressLiteral counter_addr, Register scr) { 1033 if (reachable(counter_addr)) { 1034 atomic_incl(as_Address(counter_addr)); 1035 } else { 1036 lea(scr, counter_addr); 1037 atomic_incl(Address(scr, 0)); 1038 } 1039 } 1040 1041 #ifdef _LP64 1042 void MacroAssembler::atomic_incq(Address counter_addr) { 1043 if (os::is_MP()) 1044 lock(); 1045 incrementq(counter_addr); 1046 } 1047 1048 void MacroAssembler::atomic_incq(AddressLiteral counter_addr, Register scr) { 1049 if (reachable(counter_addr)) { 1050 atomic_incq(as_Address(counter_addr)); 1051 } else { 1052 lea(scr, counter_addr); 1053 atomic_incq(Address(scr, 0)); 1054 } 1055 } 1056 #endif 1057 1058 // Writes to stack successive pages until offset reached to check for 1059 // stack overflow + shadow pages. This clobbers tmp. 1060 void MacroAssembler::bang_stack_size(Register size, Register tmp) { 1061 movptr(tmp, rsp); 1062 // Bang stack for total size given plus shadow page size. 1063 // Bang one page at a time because large size can bang beyond yellow and 1064 // red zones. 1065 Label loop; 1066 bind(loop); 1067 movl(Address(tmp, (-os::vm_page_size())), size ); 1068 subptr(tmp, os::vm_page_size()); 1069 subl(size, os::vm_page_size()); 1070 jcc(Assembler::greater, loop); 1071 1072 // Bang down shadow pages too. 1073 // At this point, (tmp-0) is the last address touched, so don't 1074 // touch it again. (It was touched as (tmp-pagesize) but then tmp 1075 // was post-decremented.) Skip this address by starting at i=1, and 1076 // touch a few more pages below. N.B. It is important to touch all 1077 // the way down including all pages in the shadow zone. 1078 for (int i = 1; i < ((int)JavaThread::stack_shadow_zone_size() / os::vm_page_size()); i++) { 1079 // this could be any sized move but this is can be a debugging crumb 1080 // so the bigger the better. 1081 movptr(Address(tmp, (-i*os::vm_page_size())), size ); 1082 } 1083 } 1084 1085 void MacroAssembler::reserved_stack_check() { 1086 // testing if reserved zone needs to be enabled 1087 Label no_reserved_zone_enabling; 1088 Register thread = NOT_LP64(rsi) LP64_ONLY(r15_thread); 1089 NOT_LP64(get_thread(rsi);) 1090 1091 cmpptr(rsp, Address(thread, JavaThread::reserved_stack_activation_offset())); 1092 jcc(Assembler::below, no_reserved_zone_enabling); 1093 1094 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), thread); 1095 jump(RuntimeAddress(StubRoutines::throw_delayed_StackOverflowError_entry())); 1096 should_not_reach_here(); 1097 1098 bind(no_reserved_zone_enabling); 1099 } 1100 1101 int MacroAssembler::biased_locking_enter(Register lock_reg, 1102 Register obj_reg, 1103 Register swap_reg, 1104 Register tmp_reg, 1105 bool swap_reg_contains_mark, 1106 Label& done, 1107 Label* slow_case, 1108 BiasedLockingCounters* counters) { 1109 assert(UseBiasedLocking, "why call this otherwise?"); 1110 assert(swap_reg == rax, "swap_reg must be rax for cmpxchgq"); 1111 assert(tmp_reg != noreg, "tmp_reg must be supplied"); 1112 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg); 1113 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout"); 1114 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes()); 1115 NOT_LP64( Address saved_mark_addr(lock_reg, 0); ) 1116 1117 if (PrintBiasedLockingStatistics && counters == NULL) { 1118 counters = BiasedLocking::counters(); 1119 } 1120 // Biased locking 1121 // See whether the lock is currently biased toward our thread and 1122 // whether the epoch is still valid 1123 // Note that the runtime guarantees sufficient alignment of JavaThread 1124 // pointers to allow age to be placed into low bits 1125 // First check to see whether biasing is even enabled for this object 1126 Label cas_label; 1127 int null_check_offset = -1; 1128 if (!swap_reg_contains_mark) { 1129 null_check_offset = offset(); 1130 movptr(swap_reg, mark_addr); 1131 } 1132 movptr(tmp_reg, swap_reg); 1133 andptr(tmp_reg, markOopDesc::biased_lock_mask_in_place); 1134 cmpptr(tmp_reg, markOopDesc::biased_lock_pattern); 1135 jcc(Assembler::notEqual, cas_label); 1136 // The bias pattern is present in the object's header. Need to check 1137 // whether the bias owner and the epoch are both still current. 1138 #ifndef _LP64 1139 // Note that because there is no current thread register on x86_32 we 1140 // need to store off the mark word we read out of the object to 1141 // avoid reloading it and needing to recheck invariants below. This 1142 // store is unfortunate but it makes the overall code shorter and 1143 // simpler. 1144 movptr(saved_mark_addr, swap_reg); 1145 #endif 1146 if (swap_reg_contains_mark) { 1147 null_check_offset = offset(); 1148 } 1149 load_prototype_header(tmp_reg, obj_reg); 1150 #ifdef _LP64 1151 orptr(tmp_reg, r15_thread); 1152 xorptr(tmp_reg, swap_reg); 1153 Register header_reg = tmp_reg; 1154 #else 1155 xorptr(tmp_reg, swap_reg); 1156 get_thread(swap_reg); 1157 xorptr(swap_reg, tmp_reg); 1158 Register header_reg = swap_reg; 1159 #endif 1160 andptr(header_reg, ~((int) markOopDesc::age_mask_in_place)); 1161 if (counters != NULL) { 1162 cond_inc32(Assembler::zero, 1163 ExternalAddress((address) counters->biased_lock_entry_count_addr())); 1164 } 1165 jcc(Assembler::equal, done); 1166 1167 Label try_revoke_bias; 1168 Label try_rebias; 1169 1170 // At this point we know that the header has the bias pattern and 1171 // that we are not the bias owner in the current epoch. We need to 1172 // figure out more details about the state of the header in order to 1173 // know what operations can be legally performed on the object's 1174 // header. 1175 1176 // If the low three bits in the xor result aren't clear, that means 1177 // the prototype header is no longer biased and we have to revoke 1178 // the bias on this object. 1179 testptr(header_reg, markOopDesc::biased_lock_mask_in_place); 1180 jccb(Assembler::notZero, try_revoke_bias); 1181 1182 // Biasing is still enabled for this data type. See whether the 1183 // epoch of the current bias is still valid, meaning that the epoch 1184 // bits of the mark word are equal to the epoch bits of the 1185 // prototype header. (Note that the prototype header's epoch bits 1186 // only change at a safepoint.) If not, attempt to rebias the object 1187 // toward the current thread. Note that we must be absolutely sure 1188 // that the current epoch is invalid in order to do this because 1189 // otherwise the manipulations it performs on the mark word are 1190 // illegal. 1191 testptr(header_reg, markOopDesc::epoch_mask_in_place); 1192 jccb(Assembler::notZero, try_rebias); 1193 1194 // The epoch of the current bias is still valid but we know nothing 1195 // about the owner; it might be set or it might be clear. Try to 1196 // acquire the bias of the object using an atomic operation. If this 1197 // fails we will go in to the runtime to revoke the object's bias. 1198 // Note that we first construct the presumed unbiased header so we 1199 // don't accidentally blow away another thread's valid bias. 1200 NOT_LP64( movptr(swap_reg, saved_mark_addr); ) 1201 andptr(swap_reg, 1202 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place); 1203 #ifdef _LP64 1204 movptr(tmp_reg, swap_reg); 1205 orptr(tmp_reg, r15_thread); 1206 #else 1207 get_thread(tmp_reg); 1208 orptr(tmp_reg, swap_reg); 1209 #endif 1210 if (os::is_MP()) { 1211 lock(); 1212 } 1213 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1214 // If the biasing toward our thread failed, this means that 1215 // another thread succeeded in biasing it toward itself and we 1216 // need to revoke that bias. The revocation will occur in the 1217 // interpreter runtime in the slow case. 1218 if (counters != NULL) { 1219 cond_inc32(Assembler::zero, 1220 ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr())); 1221 } 1222 if (slow_case != NULL) { 1223 jcc(Assembler::notZero, *slow_case); 1224 } 1225 jmp(done); 1226 1227 bind(try_rebias); 1228 // At this point we know the epoch has expired, meaning that the 1229 // current "bias owner", if any, is actually invalid. Under these 1230 // circumstances _only_, we are allowed to use the current header's 1231 // value as the comparison value when doing the cas to acquire the 1232 // bias in the current epoch. In other words, we allow transfer of 1233 // the bias from one thread to another directly in this situation. 1234 // 1235 // FIXME: due to a lack of registers we currently blow away the age 1236 // bits in this situation. Should attempt to preserve them. 1237 load_prototype_header(tmp_reg, obj_reg); 1238 #ifdef _LP64 1239 orptr(tmp_reg, r15_thread); 1240 #else 1241 get_thread(swap_reg); 1242 orptr(tmp_reg, swap_reg); 1243 movptr(swap_reg, saved_mark_addr); 1244 #endif 1245 if (os::is_MP()) { 1246 lock(); 1247 } 1248 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1249 // If the biasing toward our thread failed, then another thread 1250 // succeeded in biasing it toward itself and we need to revoke that 1251 // bias. The revocation will occur in the runtime in the slow case. 1252 if (counters != NULL) { 1253 cond_inc32(Assembler::zero, 1254 ExternalAddress((address) counters->rebiased_lock_entry_count_addr())); 1255 } 1256 if (slow_case != NULL) { 1257 jcc(Assembler::notZero, *slow_case); 1258 } 1259 jmp(done); 1260 1261 bind(try_revoke_bias); 1262 // The prototype mark in the klass doesn't have the bias bit set any 1263 // more, indicating that objects of this data type are not supposed 1264 // to be biased any more. We are going to try to reset the mark of 1265 // this object to the prototype value and fall through to the 1266 // CAS-based locking scheme. Note that if our CAS fails, it means 1267 // that another thread raced us for the privilege of revoking the 1268 // bias of this particular object, so it's okay to continue in the 1269 // normal locking code. 1270 // 1271 // FIXME: due to a lack of registers we currently blow away the age 1272 // bits in this situation. Should attempt to preserve them. 1273 NOT_LP64( movptr(swap_reg, saved_mark_addr); ) 1274 load_prototype_header(tmp_reg, obj_reg); 1275 if (os::is_MP()) { 1276 lock(); 1277 } 1278 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1279 // Fall through to the normal CAS-based lock, because no matter what 1280 // the result of the above CAS, some thread must have succeeded in 1281 // removing the bias bit from the object's header. 1282 if (counters != NULL) { 1283 cond_inc32(Assembler::zero, 1284 ExternalAddress((address) counters->revoked_lock_entry_count_addr())); 1285 } 1286 1287 bind(cas_label); 1288 1289 return null_check_offset; 1290 } 1291 1292 void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) { 1293 assert(UseBiasedLocking, "why call this otherwise?"); 1294 1295 // Check for biased locking unlock case, which is a no-op 1296 // Note: we do not have to check the thread ID for two reasons. 1297 // First, the interpreter checks for IllegalMonitorStateException at 1298 // a higher level. Second, if the bias was revoked while we held the 1299 // lock, the object could not be rebiased toward another thread, so 1300 // the bias bit would be clear. 1301 movptr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 1302 andptr(temp_reg, markOopDesc::biased_lock_mask_in_place); 1303 cmpptr(temp_reg, markOopDesc::biased_lock_pattern); 1304 jcc(Assembler::equal, done); 1305 } 1306 1307 #ifdef COMPILER2 1308 1309 #if INCLUDE_RTM_OPT 1310 1311 // Update rtm_counters based on abort status 1312 // input: abort_status 1313 // rtm_counters (RTMLockingCounters*) 1314 // flags are killed 1315 void MacroAssembler::rtm_counters_update(Register abort_status, Register rtm_counters) { 1316 1317 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abort_count_offset())); 1318 if (PrintPreciseRTMLockingStatistics) { 1319 for (int i = 0; i < RTMLockingCounters::ABORT_STATUS_LIMIT; i++) { 1320 Label check_abort; 1321 testl(abort_status, (1<<i)); 1322 jccb(Assembler::equal, check_abort); 1323 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abortX_count_offset() + (i * sizeof(uintx)))); 1324 bind(check_abort); 1325 } 1326 } 1327 } 1328 1329 // Branch if (random & (count-1) != 0), count is 2^n 1330 // tmp, scr and flags are killed 1331 void MacroAssembler::branch_on_random_using_rdtsc(Register tmp, Register scr, int count, Label& brLabel) { 1332 assert(tmp == rax, ""); 1333 assert(scr == rdx, ""); 1334 rdtsc(); // modifies EDX:EAX 1335 andptr(tmp, count-1); 1336 jccb(Assembler::notZero, brLabel); 1337 } 1338 1339 // Perform abort ratio calculation, set no_rtm bit if high ratio 1340 // input: rtm_counters_Reg (RTMLockingCounters* address) 1341 // tmpReg, rtm_counters_Reg and flags are killed 1342 void MacroAssembler::rtm_abort_ratio_calculation(Register tmpReg, 1343 Register rtm_counters_Reg, 1344 RTMLockingCounters* rtm_counters, 1345 Metadata* method_data) { 1346 Label L_done, L_check_always_rtm1, L_check_always_rtm2; 1347 1348 if (RTMLockingCalculationDelay > 0) { 1349 // Delay calculation 1350 movptr(tmpReg, ExternalAddress((address) RTMLockingCounters::rtm_calculation_flag_addr()), tmpReg); 1351 testptr(tmpReg, tmpReg); 1352 jccb(Assembler::equal, L_done); 1353 } 1354 // Abort ratio calculation only if abort_count > RTMAbortThreshold 1355 // Aborted transactions = abort_count * 100 1356 // All transactions = total_count * RTMTotalCountIncrRate 1357 // Set no_rtm bit if (Aborted transactions >= All transactions * RTMAbortRatio) 1358 1359 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::abort_count_offset())); 1360 cmpptr(tmpReg, RTMAbortThreshold); 1361 jccb(Assembler::below, L_check_always_rtm2); 1362 imulptr(tmpReg, tmpReg, 100); 1363 1364 Register scrReg = rtm_counters_Reg; 1365 movptr(scrReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset())); 1366 imulptr(scrReg, scrReg, RTMTotalCountIncrRate); 1367 imulptr(scrReg, scrReg, RTMAbortRatio); 1368 cmpptr(tmpReg, scrReg); 1369 jccb(Assembler::below, L_check_always_rtm1); 1370 if (method_data != NULL) { 1371 // set rtm_state to "no rtm" in MDO 1372 mov_metadata(tmpReg, method_data); 1373 if (os::is_MP()) { 1374 lock(); 1375 } 1376 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), NoRTM); 1377 } 1378 jmpb(L_done); 1379 bind(L_check_always_rtm1); 1380 // Reload RTMLockingCounters* address 1381 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters)); 1382 bind(L_check_always_rtm2); 1383 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset())); 1384 cmpptr(tmpReg, RTMLockingThreshold / RTMTotalCountIncrRate); 1385 jccb(Assembler::below, L_done); 1386 if (method_data != NULL) { 1387 // set rtm_state to "always rtm" in MDO 1388 mov_metadata(tmpReg, method_data); 1389 if (os::is_MP()) { 1390 lock(); 1391 } 1392 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), UseRTM); 1393 } 1394 bind(L_done); 1395 } 1396 1397 // Update counters and perform abort ratio calculation 1398 // input: abort_status_Reg 1399 // rtm_counters_Reg, flags are killed 1400 void MacroAssembler::rtm_profiling(Register abort_status_Reg, 1401 Register rtm_counters_Reg, 1402 RTMLockingCounters* rtm_counters, 1403 Metadata* method_data, 1404 bool profile_rtm) { 1405 1406 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1407 // update rtm counters based on rax value at abort 1408 // reads abort_status_Reg, updates flags 1409 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters)); 1410 rtm_counters_update(abort_status_Reg, rtm_counters_Reg); 1411 if (profile_rtm) { 1412 // Save abort status because abort_status_Reg is used by following code. 1413 if (RTMRetryCount > 0) { 1414 push(abort_status_Reg); 1415 } 1416 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1417 rtm_abort_ratio_calculation(abort_status_Reg, rtm_counters_Reg, rtm_counters, method_data); 1418 // restore abort status 1419 if (RTMRetryCount > 0) { 1420 pop(abort_status_Reg); 1421 } 1422 } 1423 } 1424 1425 // Retry on abort if abort's status is 0x6: can retry (0x2) | memory conflict (0x4) 1426 // inputs: retry_count_Reg 1427 // : abort_status_Reg 1428 // output: retry_count_Reg decremented by 1 1429 // flags are killed 1430 void MacroAssembler::rtm_retry_lock_on_abort(Register retry_count_Reg, Register abort_status_Reg, Label& retryLabel) { 1431 Label doneRetry; 1432 assert(abort_status_Reg == rax, ""); 1433 // The abort reason bits are in eax (see all states in rtmLocking.hpp) 1434 // 0x6 = conflict on which we can retry (0x2) | memory conflict (0x4) 1435 // if reason is in 0x6 and retry count != 0 then retry 1436 andptr(abort_status_Reg, 0x6); 1437 jccb(Assembler::zero, doneRetry); 1438 testl(retry_count_Reg, retry_count_Reg); 1439 jccb(Assembler::zero, doneRetry); 1440 pause(); 1441 decrementl(retry_count_Reg); 1442 jmp(retryLabel); 1443 bind(doneRetry); 1444 } 1445 1446 // Spin and retry if lock is busy, 1447 // inputs: box_Reg (monitor address) 1448 // : retry_count_Reg 1449 // output: retry_count_Reg decremented by 1 1450 // : clear z flag if retry count exceeded 1451 // tmp_Reg, scr_Reg, flags are killed 1452 void MacroAssembler::rtm_retry_lock_on_busy(Register retry_count_Reg, Register box_Reg, 1453 Register tmp_Reg, Register scr_Reg, Label& retryLabel) { 1454 Label SpinLoop, SpinExit, doneRetry; 1455 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1456 1457 testl(retry_count_Reg, retry_count_Reg); 1458 jccb(Assembler::zero, doneRetry); 1459 decrementl(retry_count_Reg); 1460 movptr(scr_Reg, RTMSpinLoopCount); 1461 1462 bind(SpinLoop); 1463 pause(); 1464 decrementl(scr_Reg); 1465 jccb(Assembler::lessEqual, SpinExit); 1466 movptr(tmp_Reg, Address(box_Reg, owner_offset)); 1467 testptr(tmp_Reg, tmp_Reg); 1468 jccb(Assembler::notZero, SpinLoop); 1469 1470 bind(SpinExit); 1471 jmp(retryLabel); 1472 bind(doneRetry); 1473 incrementl(retry_count_Reg); // clear z flag 1474 } 1475 1476 // Use RTM for normal stack locks 1477 // Input: objReg (object to lock) 1478 void MacroAssembler::rtm_stack_locking(Register objReg, Register tmpReg, Register scrReg, 1479 Register retry_on_abort_count_Reg, 1480 RTMLockingCounters* stack_rtm_counters, 1481 Metadata* method_data, bool profile_rtm, 1482 Label& DONE_LABEL, Label& IsInflated) { 1483 assert(UseRTMForStackLocks, "why call this otherwise?"); 1484 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking"); 1485 assert(tmpReg == rax, ""); 1486 assert(scrReg == rdx, ""); 1487 Label L_rtm_retry, L_decrement_retry, L_on_abort; 1488 1489 if (RTMRetryCount > 0) { 1490 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort 1491 bind(L_rtm_retry); 1492 } 1493 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); 1494 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased 1495 jcc(Assembler::notZero, IsInflated); 1496 1497 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1498 Label L_noincrement; 1499 if (RTMTotalCountIncrRate > 1) { 1500 // tmpReg, scrReg and flags are killed 1501 branch_on_random_using_rdtsc(tmpReg, scrReg, RTMTotalCountIncrRate, L_noincrement); 1502 } 1503 assert(stack_rtm_counters != NULL, "should not be NULL when profiling RTM"); 1504 atomic_incptr(ExternalAddress((address)stack_rtm_counters->total_count_addr()), scrReg); 1505 bind(L_noincrement); 1506 } 1507 xbegin(L_on_abort); 1508 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // fetch markword 1509 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits 1510 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked 1511 jcc(Assembler::equal, DONE_LABEL); // all done if unlocked 1512 1513 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX 1514 if (UseRTMXendForLockBusy) { 1515 xend(); 1516 movptr(abort_status_Reg, 0x2); // Set the abort status to 2 (so we can retry) 1517 jmp(L_decrement_retry); 1518 } 1519 else { 1520 xabort(0); 1521 } 1522 bind(L_on_abort); 1523 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1524 rtm_profiling(abort_status_Reg, scrReg, stack_rtm_counters, method_data, profile_rtm); 1525 } 1526 bind(L_decrement_retry); 1527 if (RTMRetryCount > 0) { 1528 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4) 1529 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry); 1530 } 1531 } 1532 1533 // Use RTM for inflating locks 1534 // inputs: objReg (object to lock) 1535 // boxReg (on-stack box address (displaced header location) - KILLED) 1536 // tmpReg (ObjectMonitor address + markOopDesc::monitor_value) 1537 void MacroAssembler::rtm_inflated_locking(Register objReg, Register boxReg, Register tmpReg, 1538 Register scrReg, Register retry_on_busy_count_Reg, 1539 Register retry_on_abort_count_Reg, 1540 RTMLockingCounters* rtm_counters, 1541 Metadata* method_data, bool profile_rtm, 1542 Label& DONE_LABEL) { 1543 assert(UseRTMLocking, "why call this otherwise?"); 1544 assert(tmpReg == rax, ""); 1545 assert(scrReg == rdx, ""); 1546 Label L_rtm_retry, L_decrement_retry, L_on_abort; 1547 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1548 1549 // Without cast to int32_t a movptr will destroy r10 which is typically obj 1550 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark())); 1551 movptr(boxReg, tmpReg); // Save ObjectMonitor address 1552 1553 if (RTMRetryCount > 0) { 1554 movl(retry_on_busy_count_Reg, RTMRetryCount); // Retry on lock busy 1555 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort 1556 bind(L_rtm_retry); 1557 } 1558 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1559 Label L_noincrement; 1560 if (RTMTotalCountIncrRate > 1) { 1561 // tmpReg, scrReg and flags are killed 1562 branch_on_random_using_rdtsc(tmpReg, scrReg, RTMTotalCountIncrRate, L_noincrement); 1563 } 1564 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1565 atomic_incptr(ExternalAddress((address)rtm_counters->total_count_addr()), scrReg); 1566 bind(L_noincrement); 1567 } 1568 xbegin(L_on_abort); 1569 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); 1570 movptr(tmpReg, Address(tmpReg, owner_offset)); 1571 testptr(tmpReg, tmpReg); 1572 jcc(Assembler::zero, DONE_LABEL); 1573 if (UseRTMXendForLockBusy) { 1574 xend(); 1575 jmp(L_decrement_retry); 1576 } 1577 else { 1578 xabort(0); 1579 } 1580 bind(L_on_abort); 1581 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX 1582 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1583 rtm_profiling(abort_status_Reg, scrReg, rtm_counters, method_data, profile_rtm); 1584 } 1585 if (RTMRetryCount > 0) { 1586 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4) 1587 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry); 1588 } 1589 1590 movptr(tmpReg, Address(boxReg, owner_offset)) ; 1591 testptr(tmpReg, tmpReg) ; 1592 jccb(Assembler::notZero, L_decrement_retry) ; 1593 1594 // Appears unlocked - try to swing _owner from null to non-null. 1595 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1596 #ifdef _LP64 1597 Register threadReg = r15_thread; 1598 #else 1599 get_thread(scrReg); 1600 Register threadReg = scrReg; 1601 #endif 1602 if (os::is_MP()) { 1603 lock(); 1604 } 1605 cmpxchgptr(threadReg, Address(boxReg, owner_offset)); // Updates tmpReg 1606 1607 if (RTMRetryCount > 0) { 1608 // success done else retry 1609 jccb(Assembler::equal, DONE_LABEL) ; 1610 bind(L_decrement_retry); 1611 // Spin and retry if lock is busy. 1612 rtm_retry_lock_on_busy(retry_on_busy_count_Reg, boxReg, tmpReg, scrReg, L_rtm_retry); 1613 } 1614 else { 1615 bind(L_decrement_retry); 1616 } 1617 } 1618 1619 #endif // INCLUDE_RTM_OPT 1620 1621 // Fast_Lock and Fast_Unlock used by C2 1622 1623 // Because the transitions from emitted code to the runtime 1624 // monitorenter/exit helper stubs are so slow it's critical that 1625 // we inline both the stack-locking fast-path and the inflated fast path. 1626 // 1627 // See also: cmpFastLock and cmpFastUnlock. 1628 // 1629 // What follows is a specialized inline transliteration of the code 1630 // in slow_enter() and slow_exit(). If we're concerned about I$ bloat 1631 // another option would be to emit TrySlowEnter and TrySlowExit methods 1632 // at startup-time. These methods would accept arguments as 1633 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure 1634 // indications in the icc.ZFlag. Fast_Lock and Fast_Unlock would simply 1635 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit. 1636 // In practice, however, the # of lock sites is bounded and is usually small. 1637 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer 1638 // if the processor uses simple bimodal branch predictors keyed by EIP 1639 // Since the helper routines would be called from multiple synchronization 1640 // sites. 1641 // 1642 // An even better approach would be write "MonitorEnter()" and "MonitorExit()" 1643 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites 1644 // to those specialized methods. That'd give us a mostly platform-independent 1645 // implementation that the JITs could optimize and inline at their pleasure. 1646 // Done correctly, the only time we'd need to cross to native could would be 1647 // to park() or unpark() threads. We'd also need a few more unsafe operators 1648 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and 1649 // (b) explicit barriers or fence operations. 1650 // 1651 // TODO: 1652 // 1653 // * Arrange for C2 to pass "Self" into Fast_Lock and Fast_Unlock in one of the registers (scr). 1654 // This avoids manifesting the Self pointer in the Fast_Lock and Fast_Unlock terminals. 1655 // Given TLAB allocation, Self is usually manifested in a register, so passing it into 1656 // the lock operators would typically be faster than reifying Self. 1657 // 1658 // * Ideally I'd define the primitives as: 1659 // fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED. 1660 // fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED 1661 // Unfortunately ADLC bugs prevent us from expressing the ideal form. 1662 // Instead, we're stuck with a rather awkward and brittle register assignments below. 1663 // Furthermore the register assignments are overconstrained, possibly resulting in 1664 // sub-optimal code near the synchronization site. 1665 // 1666 // * Eliminate the sp-proximity tests and just use "== Self" tests instead. 1667 // Alternately, use a better sp-proximity test. 1668 // 1669 // * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value. 1670 // Either one is sufficient to uniquely identify a thread. 1671 // TODO: eliminate use of sp in _owner and use get_thread(tr) instead. 1672 // 1673 // * Intrinsify notify() and notifyAll() for the common cases where the 1674 // object is locked by the calling thread but the waitlist is empty. 1675 // avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll(). 1676 // 1677 // * use jccb and jmpb instead of jcc and jmp to improve code density. 1678 // But beware of excessive branch density on AMD Opterons. 1679 // 1680 // * Both Fast_Lock and Fast_Unlock set the ICC.ZF to indicate success 1681 // or failure of the fast-path. If the fast-path fails then we pass 1682 // control to the slow-path, typically in C. In Fast_Lock and 1683 // Fast_Unlock we often branch to DONE_LABEL, just to find that C2 1684 // will emit a conditional branch immediately after the node. 1685 // So we have branches to branches and lots of ICC.ZF games. 1686 // Instead, it might be better to have C2 pass a "FailureLabel" 1687 // into Fast_Lock and Fast_Unlock. In the case of success, control 1688 // will drop through the node. ICC.ZF is undefined at exit. 1689 // In the case of failure, the node will branch directly to the 1690 // FailureLabel 1691 1692 1693 // obj: object to lock 1694 // box: on-stack box address (displaced header location) - KILLED 1695 // rax,: tmp -- KILLED 1696 // scr: tmp -- KILLED 1697 void MacroAssembler::fast_lock(Register objReg, Register boxReg, Register tmpReg, 1698 Register scrReg, Register cx1Reg, Register cx2Reg, 1699 BiasedLockingCounters* counters, 1700 RTMLockingCounters* rtm_counters, 1701 RTMLockingCounters* stack_rtm_counters, 1702 Metadata* method_data, 1703 bool use_rtm, bool profile_rtm) { 1704 // Ensure the register assignments are disjoint 1705 assert(tmpReg == rax, ""); 1706 1707 if (use_rtm) { 1708 assert_different_registers(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg); 1709 } else { 1710 assert(cx1Reg == noreg, ""); 1711 assert(cx2Reg == noreg, ""); 1712 assert_different_registers(objReg, boxReg, tmpReg, scrReg); 1713 } 1714 1715 if (counters != NULL) { 1716 atomic_incl(ExternalAddress((address)counters->total_entry_count_addr()), scrReg); 1717 } 1718 if (EmitSync & 1) { 1719 // set box->dhw = markOopDesc::unused_mark() 1720 // Force all sync thru slow-path: slow_enter() and slow_exit() 1721 movptr (Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark())); 1722 cmpptr (rsp, (int32_t)NULL_WORD); 1723 } else { 1724 // Possible cases that we'll encounter in fast_lock 1725 // ------------------------------------------------ 1726 // * Inflated 1727 // -- unlocked 1728 // -- Locked 1729 // = by self 1730 // = by other 1731 // * biased 1732 // -- by Self 1733 // -- by other 1734 // * neutral 1735 // * stack-locked 1736 // -- by self 1737 // = sp-proximity test hits 1738 // = sp-proximity test generates false-negative 1739 // -- by other 1740 // 1741 1742 Label IsInflated, DONE_LABEL; 1743 1744 // it's stack-locked, biased or neutral 1745 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage 1746 // order to reduce the number of conditional branches in the most common cases. 1747 // Beware -- there's a subtle invariant that fetch of the markword 1748 // at [FETCH], below, will never observe a biased encoding (*101b). 1749 // If this invariant is not held we risk exclusion (safety) failure. 1750 if (UseBiasedLocking && !UseOptoBiasInlining) { 1751 biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, counters); 1752 } 1753 1754 #if INCLUDE_RTM_OPT 1755 if (UseRTMForStackLocks && use_rtm) { 1756 rtm_stack_locking(objReg, tmpReg, scrReg, cx2Reg, 1757 stack_rtm_counters, method_data, profile_rtm, 1758 DONE_LABEL, IsInflated); 1759 } 1760 #endif // INCLUDE_RTM_OPT 1761 1762 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // [FETCH] 1763 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased 1764 jccb(Assembler::notZero, IsInflated); 1765 1766 // Attempt stack-locking ... 1767 orptr (tmpReg, markOopDesc::unlocked_value); 1768 movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS 1769 if (os::is_MP()) { 1770 lock(); 1771 } 1772 cmpxchgptr(boxReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Updates tmpReg 1773 if (counters != NULL) { 1774 cond_inc32(Assembler::equal, 1775 ExternalAddress((address)counters->fast_path_entry_count_addr())); 1776 } 1777 jcc(Assembler::equal, DONE_LABEL); // Success 1778 1779 // Recursive locking. 1780 // The object is stack-locked: markword contains stack pointer to BasicLock. 1781 // Locked by current thread if difference with current SP is less than one page. 1782 subptr(tmpReg, rsp); 1783 // Next instruction set ZFlag == 1 (Success) if difference is less then one page. 1784 andptr(tmpReg, (int32_t) (NOT_LP64(0xFFFFF003) LP64_ONLY(7 - os::vm_page_size())) ); 1785 movptr(Address(boxReg, 0), tmpReg); 1786 if (counters != NULL) { 1787 cond_inc32(Assembler::equal, 1788 ExternalAddress((address)counters->fast_path_entry_count_addr())); 1789 } 1790 jmp(DONE_LABEL); 1791 1792 bind(IsInflated); 1793 // The object is inflated. tmpReg contains pointer to ObjectMonitor* + markOopDesc::monitor_value 1794 1795 #if INCLUDE_RTM_OPT 1796 // Use the same RTM locking code in 32- and 64-bit VM. 1797 if (use_rtm) { 1798 rtm_inflated_locking(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg, 1799 rtm_counters, method_data, profile_rtm, DONE_LABEL); 1800 } else { 1801 #endif // INCLUDE_RTM_OPT 1802 1803 #ifndef _LP64 1804 // The object is inflated. 1805 1806 // boxReg refers to the on-stack BasicLock in the current frame. 1807 // We'd like to write: 1808 // set box->_displaced_header = markOopDesc::unused_mark(). Any non-0 value suffices. 1809 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers 1810 // additional latency as we have another ST in the store buffer that must drain. 1811 1812 if (EmitSync & 8192) { 1813 movptr(Address(boxReg, 0), 3); // results in ST-before-CAS penalty 1814 get_thread (scrReg); 1815 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2] 1816 movptr(tmpReg, NULL_WORD); // consider: xor vs mov 1817 if (os::is_MP()) { 1818 lock(); 1819 } 1820 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1821 } else 1822 if ((EmitSync & 128) == 0) { // avoid ST-before-CAS 1823 // register juggle because we need tmpReg for cmpxchgptr below 1824 movptr(scrReg, boxReg); 1825 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2] 1826 1827 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes 1828 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) { 1829 // prefetchw [eax + Offset(_owner)-2] 1830 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1831 } 1832 1833 if ((EmitSync & 64) == 0) { 1834 // Optimistic form: consider XORL tmpReg,tmpReg 1835 movptr(tmpReg, NULL_WORD); 1836 } else { 1837 // Can suffer RTS->RTO upgrades on shared or cold $ lines 1838 // Test-And-CAS instead of CAS 1839 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner 1840 testptr(tmpReg, tmpReg); // Locked ? 1841 jccb (Assembler::notZero, DONE_LABEL); 1842 } 1843 1844 // Appears unlocked - try to swing _owner from null to non-null. 1845 // Ideally, I'd manifest "Self" with get_thread and then attempt 1846 // to CAS the register containing Self into m->Owner. 1847 // But we don't have enough registers, so instead we can either try to CAS 1848 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds 1849 // we later store "Self" into m->Owner. Transiently storing a stack address 1850 // (rsp or the address of the box) into m->owner is harmless. 1851 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1852 if (os::is_MP()) { 1853 lock(); 1854 } 1855 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1856 movptr(Address(scrReg, 0), 3); // box->_displaced_header = 3 1857 // If we weren't able to swing _owner from NULL to the BasicLock 1858 // then take the slow path. 1859 jccb (Assembler::notZero, DONE_LABEL); 1860 // update _owner from BasicLock to thread 1861 get_thread (scrReg); // beware: clobbers ICCs 1862 movptr(Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), scrReg); 1863 xorptr(boxReg, boxReg); // set icc.ZFlag = 1 to indicate success 1864 1865 // If the CAS fails we can either retry or pass control to the slow-path. 1866 // We use the latter tactic. 1867 // Pass the CAS result in the icc.ZFlag into DONE_LABEL 1868 // If the CAS was successful ... 1869 // Self has acquired the lock 1870 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it. 1871 // Intentional fall-through into DONE_LABEL ... 1872 } else { 1873 movptr(Address(boxReg, 0), intptr_t(markOopDesc::unused_mark())); // results in ST-before-CAS penalty 1874 movptr(boxReg, tmpReg); 1875 1876 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes 1877 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) { 1878 // prefetchw [eax + Offset(_owner)-2] 1879 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1880 } 1881 1882 if ((EmitSync & 64) == 0) { 1883 // Optimistic form 1884 xorptr (tmpReg, tmpReg); 1885 } else { 1886 // Can suffer RTS->RTO upgrades on shared or cold $ lines 1887 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner 1888 testptr(tmpReg, tmpReg); // Locked ? 1889 jccb (Assembler::notZero, DONE_LABEL); 1890 } 1891 1892 // Appears unlocked - try to swing _owner from null to non-null. 1893 // Use either "Self" (in scr) or rsp as thread identity in _owner. 1894 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1895 get_thread (scrReg); 1896 if (os::is_MP()) { 1897 lock(); 1898 } 1899 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1900 1901 // If the CAS fails we can either retry or pass control to the slow-path. 1902 // We use the latter tactic. 1903 // Pass the CAS result in the icc.ZFlag into DONE_LABEL 1904 // If the CAS was successful ... 1905 // Self has acquired the lock 1906 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it. 1907 // Intentional fall-through into DONE_LABEL ... 1908 } 1909 #else // _LP64 1910 // It's inflated 1911 movq(scrReg, tmpReg); 1912 xorq(tmpReg, tmpReg); 1913 1914 if (os::is_MP()) { 1915 lock(); 1916 } 1917 cmpxchgptr(r15_thread, Address(scrReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1918 // Unconditionally set box->_displaced_header = markOopDesc::unused_mark(). 1919 // Without cast to int32_t movptr will destroy r10 which is typically obj. 1920 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark())); 1921 // Intentional fall-through into DONE_LABEL ... 1922 // Propagate ICC.ZF from CAS above into DONE_LABEL. 1923 #endif // _LP64 1924 #if INCLUDE_RTM_OPT 1925 } // use_rtm() 1926 #endif 1927 // DONE_LABEL is a hot target - we'd really like to place it at the 1928 // start of cache line by padding with NOPs. 1929 // See the AMD and Intel software optimization manuals for the 1930 // most efficient "long" NOP encodings. 1931 // Unfortunately none of our alignment mechanisms suffice. 1932 bind(DONE_LABEL); 1933 1934 // At DONE_LABEL the icc ZFlag is set as follows ... 1935 // Fast_Unlock uses the same protocol. 1936 // ZFlag == 1 -> Success 1937 // ZFlag == 0 -> Failure - force control through the slow-path 1938 } 1939 } 1940 1941 // obj: object to unlock 1942 // box: box address (displaced header location), killed. Must be EAX. 1943 // tmp: killed, cannot be obj nor box. 1944 // 1945 // Some commentary on balanced locking: 1946 // 1947 // Fast_Lock and Fast_Unlock are emitted only for provably balanced lock sites. 1948 // Methods that don't have provably balanced locking are forced to run in the 1949 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock. 1950 // The interpreter provides two properties: 1951 // I1: At return-time the interpreter automatically and quietly unlocks any 1952 // objects acquired the current activation (frame). Recall that the 1953 // interpreter maintains an on-stack list of locks currently held by 1954 // a frame. 1955 // I2: If a method attempts to unlock an object that is not held by the 1956 // the frame the interpreter throws IMSX. 1957 // 1958 // Lets say A(), which has provably balanced locking, acquires O and then calls B(). 1959 // B() doesn't have provably balanced locking so it runs in the interpreter. 1960 // Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O 1961 // is still locked by A(). 1962 // 1963 // The only other source of unbalanced locking would be JNI. The "Java Native Interface: 1964 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter 1965 // should not be unlocked by "normal" java-level locking and vice-versa. The specification 1966 // doesn't specify what will occur if a program engages in such mixed-mode locking, however. 1967 // Arguably given that the spec legislates the JNI case as undefined our implementation 1968 // could reasonably *avoid* checking owner in Fast_Unlock(). 1969 // In the interest of performance we elide m->Owner==Self check in unlock. 1970 // A perfectly viable alternative is to elide the owner check except when 1971 // Xcheck:jni is enabled. 1972 1973 void MacroAssembler::fast_unlock(Register objReg, Register boxReg, Register tmpReg, bool use_rtm) { 1974 assert(boxReg == rax, ""); 1975 assert_different_registers(objReg, boxReg, tmpReg); 1976 1977 if (EmitSync & 4) { 1978 // Disable - inhibit all inlining. Force control through the slow-path 1979 cmpptr (rsp, 0); 1980 } else { 1981 Label DONE_LABEL, Stacked, CheckSucc; 1982 1983 // Critically, the biased locking test must have precedence over 1984 // and appear before the (box->dhw == 0) recursive stack-lock test. 1985 if (UseBiasedLocking && !UseOptoBiasInlining) { 1986 biased_locking_exit(objReg, tmpReg, DONE_LABEL); 1987 } 1988 1989 #if INCLUDE_RTM_OPT 1990 if (UseRTMForStackLocks && use_rtm) { 1991 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking"); 1992 Label L_regular_unlock; 1993 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // fetch markword 1994 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits 1995 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked 1996 jccb(Assembler::notEqual, L_regular_unlock); // if !HLE RegularLock 1997 xend(); // otherwise end... 1998 jmp(DONE_LABEL); // ... and we're done 1999 bind(L_regular_unlock); 2000 } 2001 #endif 2002 2003 cmpptr(Address(boxReg, 0), (int32_t)NULL_WORD); // Examine the displaced header 2004 jcc (Assembler::zero, DONE_LABEL); // 0 indicates recursive stack-lock 2005 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Examine the object's markword 2006 testptr(tmpReg, markOopDesc::monitor_value); // Inflated? 2007 jccb (Assembler::zero, Stacked); 2008 2009 // It's inflated. 2010 #if INCLUDE_RTM_OPT 2011 if (use_rtm) { 2012 Label L_regular_inflated_unlock; 2013 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 2014 movptr(boxReg, Address(tmpReg, owner_offset)); 2015 testptr(boxReg, boxReg); 2016 jccb(Assembler::notZero, L_regular_inflated_unlock); 2017 xend(); 2018 jmpb(DONE_LABEL); 2019 bind(L_regular_inflated_unlock); 2020 } 2021 #endif 2022 2023 // Despite our balanced locking property we still check that m->_owner == Self 2024 // as java routines or native JNI code called by this thread might 2025 // have released the lock. 2026 // Refer to the comments in synchronizer.cpp for how we might encode extra 2027 // state in _succ so we can avoid fetching EntryList|cxq. 2028 // 2029 // I'd like to add more cases in fast_lock() and fast_unlock() -- 2030 // such as recursive enter and exit -- but we have to be wary of 2031 // I$ bloat, T$ effects and BP$ effects. 2032 // 2033 // If there's no contention try a 1-0 exit. That is, exit without 2034 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how 2035 // we detect and recover from the race that the 1-0 exit admits. 2036 // 2037 // Conceptually Fast_Unlock() must execute a STST|LDST "release" barrier 2038 // before it STs null into _owner, releasing the lock. Updates 2039 // to data protected by the critical section must be visible before 2040 // we drop the lock (and thus before any other thread could acquire 2041 // the lock and observe the fields protected by the lock). 2042 // IA32's memory-model is SPO, so STs are ordered with respect to 2043 // each other and there's no need for an explicit barrier (fence). 2044 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 2045 #ifndef _LP64 2046 get_thread (boxReg); 2047 if ((EmitSync & 4096) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) { 2048 // prefetchw [ebx + Offset(_owner)-2] 2049 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2050 } 2051 2052 // Note that we could employ various encoding schemes to reduce 2053 // the number of loads below (currently 4) to just 2 or 3. 2054 // Refer to the comments in synchronizer.cpp. 2055 // In practice the chain of fetches doesn't seem to impact performance, however. 2056 xorptr(boxReg, boxReg); 2057 if ((EmitSync & 65536) == 0 && (EmitSync & 256)) { 2058 // Attempt to reduce branch density - AMD's branch predictor. 2059 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 2060 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2061 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2062 jccb (Assembler::notZero, DONE_LABEL); 2063 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD); 2064 jmpb (DONE_LABEL); 2065 } else { 2066 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 2067 jccb (Assembler::notZero, DONE_LABEL); 2068 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2069 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2070 jccb (Assembler::notZero, CheckSucc); 2071 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD); 2072 jmpb (DONE_LABEL); 2073 } 2074 2075 // The Following code fragment (EmitSync & 65536) improves the performance of 2076 // contended applications and contended synchronization microbenchmarks. 2077 // Unfortunately the emission of the code - even though not executed - causes regressions 2078 // in scimark and jetstream, evidently because of $ effects. Replacing the code 2079 // with an equal number of never-executed NOPs results in the same regression. 2080 // We leave it off by default. 2081 2082 if ((EmitSync & 65536) != 0) { 2083 Label LSuccess, LGoSlowPath ; 2084 2085 bind (CheckSucc); 2086 2087 // Optional pre-test ... it's safe to elide this 2088 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2089 jccb(Assembler::zero, LGoSlowPath); 2090 2091 // We have a classic Dekker-style idiom: 2092 // ST m->_owner = 0 ; MEMBAR; LD m->_succ 2093 // There are a number of ways to implement the barrier: 2094 // (1) lock:andl &m->_owner, 0 2095 // is fast, but mask doesn't currently support the "ANDL M,IMM32" form. 2096 // LOCK: ANDL [ebx+Offset(_Owner)-2], 0 2097 // Encodes as 81 31 OFF32 IMM32 or 83 63 OFF8 IMM8 2098 // (2) If supported, an explicit MFENCE is appealing. 2099 // In older IA32 processors MFENCE is slower than lock:add or xchg 2100 // particularly if the write-buffer is full as might be the case if 2101 // if stores closely precede the fence or fence-equivalent instruction. 2102 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences 2103 // as the situation has changed with Nehalem and Shanghai. 2104 // (3) In lieu of an explicit fence, use lock:addl to the top-of-stack 2105 // The $lines underlying the top-of-stack should be in M-state. 2106 // The locked add instruction is serializing, of course. 2107 // (4) Use xchg, which is serializing 2108 // mov boxReg, 0; xchgl boxReg, [tmpReg + Offset(_owner)-2] also works 2109 // (5) ST m->_owner = 0 and then execute lock:orl &m->_succ, 0. 2110 // The integer condition codes will tell us if succ was 0. 2111 // Since _succ and _owner should reside in the same $line and 2112 // we just stored into _owner, it's likely that the $line 2113 // remains in M-state for the lock:orl. 2114 // 2115 // We currently use (3), although it's likely that switching to (2) 2116 // is correct for the future. 2117 2118 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD); 2119 if (os::is_MP()) { 2120 lock(); addptr(Address(rsp, 0), 0); 2121 } 2122 // Ratify _succ remains non-null 2123 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), 0); 2124 jccb (Assembler::notZero, LSuccess); 2125 2126 xorptr(boxReg, boxReg); // box is really EAX 2127 if (os::is_MP()) { lock(); } 2128 cmpxchgptr(rsp, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2129 // There's no successor so we tried to regrab the lock with the 2130 // placeholder value. If that didn't work, then another thread 2131 // grabbed the lock so we're done (and exit was a success). 2132 jccb (Assembler::notEqual, LSuccess); 2133 // Since we're low on registers we installed rsp as a placeholding in _owner. 2134 // Now install Self over rsp. This is safe as we're transitioning from 2135 // non-null to non=null 2136 get_thread (boxReg); 2137 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), boxReg); 2138 // Intentional fall-through into LGoSlowPath ... 2139 2140 bind (LGoSlowPath); 2141 orptr(boxReg, 1); // set ICC.ZF=0 to indicate failure 2142 jmpb (DONE_LABEL); 2143 2144 bind (LSuccess); 2145 xorptr(boxReg, boxReg); // set ICC.ZF=1 to indicate success 2146 jmpb (DONE_LABEL); 2147 } 2148 2149 bind (Stacked); 2150 // It's not inflated and it's not recursively stack-locked and it's not biased. 2151 // It must be stack-locked. 2152 // Try to reset the header to displaced header. 2153 // The "box" value on the stack is stable, so we can reload 2154 // and be assured we observe the same value as above. 2155 movptr(tmpReg, Address(boxReg, 0)); 2156 if (os::is_MP()) { 2157 lock(); 2158 } 2159 cmpxchgptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Uses RAX which is box 2160 // Intention fall-thru into DONE_LABEL 2161 2162 // DONE_LABEL is a hot target - we'd really like to place it at the 2163 // start of cache line by padding with NOPs. 2164 // See the AMD and Intel software optimization manuals for the 2165 // most efficient "long" NOP encodings. 2166 // Unfortunately none of our alignment mechanisms suffice. 2167 if ((EmitSync & 65536) == 0) { 2168 bind (CheckSucc); 2169 } 2170 #else // _LP64 2171 // It's inflated 2172 if (EmitSync & 1024) { 2173 // Emit code to check that _owner == Self 2174 // We could fold the _owner test into subsequent code more efficiently 2175 // than using a stand-alone check, but since _owner checking is off by 2176 // default we don't bother. We also might consider predicating the 2177 // _owner==Self check on Xcheck:jni or running on a debug build. 2178 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2179 xorptr(boxReg, r15_thread); 2180 } else { 2181 xorptr(boxReg, boxReg); 2182 } 2183 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 2184 jccb (Assembler::notZero, DONE_LABEL); 2185 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2186 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2187 jccb (Assembler::notZero, CheckSucc); 2188 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD); 2189 jmpb (DONE_LABEL); 2190 2191 if ((EmitSync & 65536) == 0) { 2192 // Try to avoid passing control into the slow_path ... 2193 Label LSuccess, LGoSlowPath ; 2194 bind (CheckSucc); 2195 2196 // The following optional optimization can be elided if necessary 2197 // Effectively: if (succ == null) goto SlowPath 2198 // The code reduces the window for a race, however, 2199 // and thus benefits performance. 2200 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2201 jccb (Assembler::zero, LGoSlowPath); 2202 2203 xorptr(boxReg, boxReg); 2204 if ((EmitSync & 16) && os::is_MP()) { 2205 xchgptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2206 } else { 2207 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD); 2208 if (os::is_MP()) { 2209 // Memory barrier/fence 2210 // Dekker pivot point -- fulcrum : ST Owner; MEMBAR; LD Succ 2211 // Instead of MFENCE we use a dummy locked add of 0 to the top-of-stack. 2212 // This is faster on Nehalem and AMD Shanghai/Barcelona. 2213 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences 2214 // We might also restructure (ST Owner=0;barrier;LD _Succ) to 2215 // (mov box,0; xchgq box, &m->Owner; LD _succ) . 2216 lock(); addl(Address(rsp, 0), 0); 2217 } 2218 } 2219 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2220 jccb (Assembler::notZero, LSuccess); 2221 2222 // Rare inopportune interleaving - race. 2223 // The successor vanished in the small window above. 2224 // The lock is contended -- (cxq|EntryList) != null -- and there's no apparent successor. 2225 // We need to ensure progress and succession. 2226 // Try to reacquire the lock. 2227 // If that fails then the new owner is responsible for succession and this 2228 // thread needs to take no further action and can exit via the fast path (success). 2229 // If the re-acquire succeeds then pass control into the slow path. 2230 // As implemented, this latter mode is horrible because we generated more 2231 // coherence traffic on the lock *and* artifically extended the critical section 2232 // length while by virtue of passing control into the slow path. 2233 2234 // box is really RAX -- the following CMPXCHG depends on that binding 2235 // cmpxchg R,[M] is equivalent to rax = CAS(M,rax,R) 2236 if (os::is_MP()) { lock(); } 2237 cmpxchgptr(r15_thread, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2238 // There's no successor so we tried to regrab the lock. 2239 // If that didn't work, then another thread grabbed the 2240 // lock so we're done (and exit was a success). 2241 jccb (Assembler::notEqual, LSuccess); 2242 // Intentional fall-through into slow-path 2243 2244 bind (LGoSlowPath); 2245 orl (boxReg, 1); // set ICC.ZF=0 to indicate failure 2246 jmpb (DONE_LABEL); 2247 2248 bind (LSuccess); 2249 testl (boxReg, 0); // set ICC.ZF=1 to indicate success 2250 jmpb (DONE_LABEL); 2251 } 2252 2253 bind (Stacked); 2254 movptr(tmpReg, Address (boxReg, 0)); // re-fetch 2255 if (os::is_MP()) { lock(); } 2256 cmpxchgptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Uses RAX which is box 2257 2258 if (EmitSync & 65536) { 2259 bind (CheckSucc); 2260 } 2261 #endif 2262 bind(DONE_LABEL); 2263 } 2264 } 2265 #endif // COMPILER2 2266 2267 void MacroAssembler::c2bool(Register x) { 2268 // implements x == 0 ? 0 : 1 2269 // note: must only look at least-significant byte of x 2270 // since C-style booleans are stored in one byte 2271 // only! (was bug) 2272 andl(x, 0xFF); 2273 setb(Assembler::notZero, x); 2274 } 2275 2276 // Wouldn't need if AddressLiteral version had new name 2277 void MacroAssembler::call(Label& L, relocInfo::relocType rtype) { 2278 Assembler::call(L, rtype); 2279 } 2280 2281 void MacroAssembler::call(Register entry) { 2282 Assembler::call(entry); 2283 } 2284 2285 void MacroAssembler::call(AddressLiteral entry) { 2286 if (reachable(entry)) { 2287 Assembler::call_literal(entry.target(), entry.rspec()); 2288 } else { 2289 lea(rscratch1, entry); 2290 Assembler::call(rscratch1); 2291 } 2292 } 2293 2294 void MacroAssembler::ic_call(address entry, jint method_index) { 2295 RelocationHolder rh = virtual_call_Relocation::spec(pc(), method_index); 2296 movptr(rax, (intptr_t)Universe::non_oop_word()); 2297 call(AddressLiteral(entry, rh)); 2298 } 2299 2300 // Implementation of call_VM versions 2301 2302 void MacroAssembler::call_VM(Register oop_result, 2303 address entry_point, 2304 bool check_exceptions) { 2305 Label C, E; 2306 call(C, relocInfo::none); 2307 jmp(E); 2308 2309 bind(C); 2310 call_VM_helper(oop_result, entry_point, 0, check_exceptions); 2311 ret(0); 2312 2313 bind(E); 2314 } 2315 2316 void MacroAssembler::call_VM(Register oop_result, 2317 address entry_point, 2318 Register arg_1, 2319 bool check_exceptions) { 2320 Label C, E; 2321 call(C, relocInfo::none); 2322 jmp(E); 2323 2324 bind(C); 2325 pass_arg1(this, arg_1); 2326 call_VM_helper(oop_result, entry_point, 1, check_exceptions); 2327 ret(0); 2328 2329 bind(E); 2330 } 2331 2332 void MacroAssembler::call_VM(Register oop_result, 2333 address entry_point, 2334 Register arg_1, 2335 Register arg_2, 2336 bool check_exceptions) { 2337 Label C, E; 2338 call(C, relocInfo::none); 2339 jmp(E); 2340 2341 bind(C); 2342 2343 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2344 2345 pass_arg2(this, arg_2); 2346 pass_arg1(this, arg_1); 2347 call_VM_helper(oop_result, entry_point, 2, check_exceptions); 2348 ret(0); 2349 2350 bind(E); 2351 } 2352 2353 void MacroAssembler::call_VM(Register oop_result, 2354 address entry_point, 2355 Register arg_1, 2356 Register arg_2, 2357 Register arg_3, 2358 bool check_exceptions) { 2359 Label C, E; 2360 call(C, relocInfo::none); 2361 jmp(E); 2362 2363 bind(C); 2364 2365 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2366 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2367 pass_arg3(this, arg_3); 2368 2369 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2370 pass_arg2(this, arg_2); 2371 2372 pass_arg1(this, arg_1); 2373 call_VM_helper(oop_result, entry_point, 3, check_exceptions); 2374 ret(0); 2375 2376 bind(E); 2377 } 2378 2379 void MacroAssembler::call_VM(Register oop_result, 2380 Register last_java_sp, 2381 address entry_point, 2382 int number_of_arguments, 2383 bool check_exceptions) { 2384 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg); 2385 call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions); 2386 } 2387 2388 void MacroAssembler::call_VM(Register oop_result, 2389 Register last_java_sp, 2390 address entry_point, 2391 Register arg_1, 2392 bool check_exceptions) { 2393 pass_arg1(this, arg_1); 2394 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 2395 } 2396 2397 void MacroAssembler::call_VM(Register oop_result, 2398 Register last_java_sp, 2399 address entry_point, 2400 Register arg_1, 2401 Register arg_2, 2402 bool check_exceptions) { 2403 2404 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2405 pass_arg2(this, arg_2); 2406 pass_arg1(this, arg_1); 2407 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 2408 } 2409 2410 void MacroAssembler::call_VM(Register oop_result, 2411 Register last_java_sp, 2412 address entry_point, 2413 Register arg_1, 2414 Register arg_2, 2415 Register arg_3, 2416 bool check_exceptions) { 2417 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2418 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2419 pass_arg3(this, arg_3); 2420 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2421 pass_arg2(this, arg_2); 2422 pass_arg1(this, arg_1); 2423 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 2424 } 2425 2426 void MacroAssembler::super_call_VM(Register oop_result, 2427 Register last_java_sp, 2428 address entry_point, 2429 int number_of_arguments, 2430 bool check_exceptions) { 2431 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg); 2432 MacroAssembler::call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions); 2433 } 2434 2435 void MacroAssembler::super_call_VM(Register oop_result, 2436 Register last_java_sp, 2437 address entry_point, 2438 Register arg_1, 2439 bool check_exceptions) { 2440 pass_arg1(this, arg_1); 2441 super_call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 2442 } 2443 2444 void MacroAssembler::super_call_VM(Register oop_result, 2445 Register last_java_sp, 2446 address entry_point, 2447 Register arg_1, 2448 Register arg_2, 2449 bool check_exceptions) { 2450 2451 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2452 pass_arg2(this, arg_2); 2453 pass_arg1(this, arg_1); 2454 super_call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 2455 } 2456 2457 void MacroAssembler::super_call_VM(Register oop_result, 2458 Register last_java_sp, 2459 address entry_point, 2460 Register arg_1, 2461 Register arg_2, 2462 Register arg_3, 2463 bool check_exceptions) { 2464 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2465 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2466 pass_arg3(this, arg_3); 2467 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2468 pass_arg2(this, arg_2); 2469 pass_arg1(this, arg_1); 2470 super_call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 2471 } 2472 2473 void MacroAssembler::call_VM_base(Register oop_result, 2474 Register java_thread, 2475 Register last_java_sp, 2476 address entry_point, 2477 int number_of_arguments, 2478 bool check_exceptions) { 2479 // determine java_thread register 2480 if (!java_thread->is_valid()) { 2481 #ifdef _LP64 2482 java_thread = r15_thread; 2483 #else 2484 java_thread = rdi; 2485 get_thread(java_thread); 2486 #endif // LP64 2487 } 2488 // determine last_java_sp register 2489 if (!last_java_sp->is_valid()) { 2490 last_java_sp = rsp; 2491 } 2492 // debugging support 2493 assert(number_of_arguments >= 0 , "cannot have negative number of arguments"); 2494 LP64_ONLY(assert(java_thread == r15_thread, "unexpected register")); 2495 #ifdef ASSERT 2496 // TraceBytecodes does not use r12 but saves it over the call, so don't verify 2497 // r12 is the heapbase. 2498 LP64_ONLY(if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");) 2499 #endif // ASSERT 2500 2501 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result"); 2502 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp"); 2503 2504 // push java thread (becomes first argument of C function) 2505 2506 NOT_LP64(push(java_thread); number_of_arguments++); 2507 LP64_ONLY(mov(c_rarg0, r15_thread)); 2508 2509 // set last Java frame before call 2510 assert(last_java_sp != rbp, "can't use ebp/rbp"); 2511 2512 // Only interpreter should have to set fp 2513 set_last_Java_frame(java_thread, last_java_sp, rbp, NULL); 2514 2515 // do the call, remove parameters 2516 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments); 2517 2518 // restore the thread (cannot use the pushed argument since arguments 2519 // may be overwritten by C code generated by an optimizing compiler); 2520 // however can use the register value directly if it is callee saved. 2521 if (LP64_ONLY(true ||) java_thread == rdi || java_thread == rsi) { 2522 // rdi & rsi (also r15) are callee saved -> nothing to do 2523 #ifdef ASSERT 2524 guarantee(java_thread != rax, "change this code"); 2525 push(rax); 2526 { Label L; 2527 get_thread(rax); 2528 cmpptr(java_thread, rax); 2529 jcc(Assembler::equal, L); 2530 STOP("MacroAssembler::call_VM_base: rdi not callee saved?"); 2531 bind(L); 2532 } 2533 pop(rax); 2534 #endif 2535 } else { 2536 get_thread(java_thread); 2537 } 2538 // reset last Java frame 2539 // Only interpreter should have to clear fp 2540 reset_last_Java_frame(java_thread, true); 2541 2542 // C++ interp handles this in the interpreter 2543 check_and_handle_popframe(java_thread); 2544 check_and_handle_earlyret(java_thread); 2545 2546 if (check_exceptions) { 2547 // check for pending exceptions (java_thread is set upon return) 2548 cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t) NULL_WORD); 2549 #ifndef _LP64 2550 jump_cc(Assembler::notEqual, 2551 RuntimeAddress(StubRoutines::forward_exception_entry())); 2552 #else 2553 // This used to conditionally jump to forward_exception however it is 2554 // possible if we relocate that the branch will not reach. So we must jump 2555 // around so we can always reach 2556 2557 Label ok; 2558 jcc(Assembler::equal, ok); 2559 jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2560 bind(ok); 2561 #endif // LP64 2562 } 2563 2564 // get oop result if there is one and reset the value in the thread 2565 if (oop_result->is_valid()) { 2566 get_vm_result(oop_result, java_thread); 2567 } 2568 } 2569 2570 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) { 2571 2572 // Calculate the value for last_Java_sp 2573 // somewhat subtle. call_VM does an intermediate call 2574 // which places a return address on the stack just under the 2575 // stack pointer as the user finsihed with it. This allows 2576 // use to retrieve last_Java_pc from last_Java_sp[-1]. 2577 // On 32bit we then have to push additional args on the stack to accomplish 2578 // the actual requested call. On 64bit call_VM only can use register args 2579 // so the only extra space is the return address that call_VM created. 2580 // This hopefully explains the calculations here. 2581 2582 #ifdef _LP64 2583 // We've pushed one address, correct last_Java_sp 2584 lea(rax, Address(rsp, wordSize)); 2585 #else 2586 lea(rax, Address(rsp, (1 + number_of_arguments) * wordSize)); 2587 #endif // LP64 2588 2589 call_VM_base(oop_result, noreg, rax, entry_point, number_of_arguments, check_exceptions); 2590 2591 } 2592 2593 // Use this method when MacroAssembler version of call_VM_leaf_base() should be called from Interpreter. 2594 void MacroAssembler::call_VM_leaf0(address entry_point) { 2595 MacroAssembler::call_VM_leaf_base(entry_point, 0); 2596 } 2597 2598 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) { 2599 call_VM_leaf_base(entry_point, number_of_arguments); 2600 } 2601 2602 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) { 2603 pass_arg0(this, arg_0); 2604 call_VM_leaf(entry_point, 1); 2605 } 2606 2607 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { 2608 2609 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2610 pass_arg1(this, arg_1); 2611 pass_arg0(this, arg_0); 2612 call_VM_leaf(entry_point, 2); 2613 } 2614 2615 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { 2616 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2617 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2618 pass_arg2(this, arg_2); 2619 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2620 pass_arg1(this, arg_1); 2621 pass_arg0(this, arg_0); 2622 call_VM_leaf(entry_point, 3); 2623 } 2624 2625 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) { 2626 pass_arg0(this, arg_0); 2627 MacroAssembler::call_VM_leaf_base(entry_point, 1); 2628 } 2629 2630 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { 2631 2632 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2633 pass_arg1(this, arg_1); 2634 pass_arg0(this, arg_0); 2635 MacroAssembler::call_VM_leaf_base(entry_point, 2); 2636 } 2637 2638 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { 2639 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2640 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2641 pass_arg2(this, arg_2); 2642 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2643 pass_arg1(this, arg_1); 2644 pass_arg0(this, arg_0); 2645 MacroAssembler::call_VM_leaf_base(entry_point, 3); 2646 } 2647 2648 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) { 2649 LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg")); 2650 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2651 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2652 pass_arg3(this, arg_3); 2653 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2654 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2655 pass_arg2(this, arg_2); 2656 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2657 pass_arg1(this, arg_1); 2658 pass_arg0(this, arg_0); 2659 MacroAssembler::call_VM_leaf_base(entry_point, 4); 2660 } 2661 2662 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) { 2663 movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset())); 2664 movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD); 2665 verify_oop(oop_result, "broken oop in call_VM_base"); 2666 } 2667 2668 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) { 2669 movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset())); 2670 movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD); 2671 } 2672 2673 void MacroAssembler::check_and_handle_earlyret(Register java_thread) { 2674 } 2675 2676 void MacroAssembler::check_and_handle_popframe(Register java_thread) { 2677 } 2678 2679 void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) { 2680 if (reachable(src1)) { 2681 cmpl(as_Address(src1), imm); 2682 } else { 2683 lea(rscratch1, src1); 2684 cmpl(Address(rscratch1, 0), imm); 2685 } 2686 } 2687 2688 void MacroAssembler::cmp32(Register src1, AddressLiteral src2) { 2689 assert(!src2.is_lval(), "use cmpptr"); 2690 if (reachable(src2)) { 2691 cmpl(src1, as_Address(src2)); 2692 } else { 2693 lea(rscratch1, src2); 2694 cmpl(src1, Address(rscratch1, 0)); 2695 } 2696 } 2697 2698 void MacroAssembler::cmp32(Register src1, int32_t imm) { 2699 Assembler::cmpl(src1, imm); 2700 } 2701 2702 void MacroAssembler::cmp32(Register src1, Address src2) { 2703 Assembler::cmpl(src1, src2); 2704 } 2705 2706 void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) { 2707 ucomisd(opr1, opr2); 2708 2709 Label L; 2710 if (unordered_is_less) { 2711 movl(dst, -1); 2712 jcc(Assembler::parity, L); 2713 jcc(Assembler::below , L); 2714 movl(dst, 0); 2715 jcc(Assembler::equal , L); 2716 increment(dst); 2717 } else { // unordered is greater 2718 movl(dst, 1); 2719 jcc(Assembler::parity, L); 2720 jcc(Assembler::above , L); 2721 movl(dst, 0); 2722 jcc(Assembler::equal , L); 2723 decrementl(dst); 2724 } 2725 bind(L); 2726 } 2727 2728 void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) { 2729 ucomiss(opr1, opr2); 2730 2731 Label L; 2732 if (unordered_is_less) { 2733 movl(dst, -1); 2734 jcc(Assembler::parity, L); 2735 jcc(Assembler::below , L); 2736 movl(dst, 0); 2737 jcc(Assembler::equal , L); 2738 increment(dst); 2739 } else { // unordered is greater 2740 movl(dst, 1); 2741 jcc(Assembler::parity, L); 2742 jcc(Assembler::above , L); 2743 movl(dst, 0); 2744 jcc(Assembler::equal , L); 2745 decrementl(dst); 2746 } 2747 bind(L); 2748 } 2749 2750 2751 void MacroAssembler::cmp8(AddressLiteral src1, int imm) { 2752 if (reachable(src1)) { 2753 cmpb(as_Address(src1), imm); 2754 } else { 2755 lea(rscratch1, src1); 2756 cmpb(Address(rscratch1, 0), imm); 2757 } 2758 } 2759 2760 void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) { 2761 #ifdef _LP64 2762 if (src2.is_lval()) { 2763 movptr(rscratch1, src2); 2764 Assembler::cmpq(src1, rscratch1); 2765 } else if (reachable(src2)) { 2766 cmpq(src1, as_Address(src2)); 2767 } else { 2768 lea(rscratch1, src2); 2769 Assembler::cmpq(src1, Address(rscratch1, 0)); 2770 } 2771 #else 2772 if (src2.is_lval()) { 2773 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec()); 2774 } else { 2775 cmpl(src1, as_Address(src2)); 2776 } 2777 #endif // _LP64 2778 } 2779 2780 void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) { 2781 assert(src2.is_lval(), "not a mem-mem compare"); 2782 #ifdef _LP64 2783 // moves src2's literal address 2784 movptr(rscratch1, src2); 2785 Assembler::cmpq(src1, rscratch1); 2786 #else 2787 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec()); 2788 #endif // _LP64 2789 } 2790 2791 void MacroAssembler::cmpoop(Register src1, Register src2) { 2792 cmpptr(src1, src2); 2793 } 2794 2795 void MacroAssembler::cmpoop(Register src1, Address src2) { 2796 cmpptr(src1, src2); 2797 } 2798 2799 #ifdef _LP64 2800 void MacroAssembler::cmpoop(Register src1, jobject src2) { 2801 movoop(rscratch1, src2); 2802 cmpptr(src1, rscratch1); 2803 } 2804 #endif 2805 2806 void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) { 2807 if (reachable(adr)) { 2808 if (os::is_MP()) 2809 lock(); 2810 cmpxchgptr(reg, as_Address(adr)); 2811 } else { 2812 lea(rscratch1, adr); 2813 if (os::is_MP()) 2814 lock(); 2815 cmpxchgptr(reg, Address(rscratch1, 0)); 2816 } 2817 } 2818 2819 void MacroAssembler::cmpxchgptr(Register reg, Address adr) { 2820 LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr)); 2821 } 2822 2823 void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) { 2824 if (reachable(src)) { 2825 Assembler::comisd(dst, as_Address(src)); 2826 } else { 2827 lea(rscratch1, src); 2828 Assembler::comisd(dst, Address(rscratch1, 0)); 2829 } 2830 } 2831 2832 void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) { 2833 if (reachable(src)) { 2834 Assembler::comiss(dst, as_Address(src)); 2835 } else { 2836 lea(rscratch1, src); 2837 Assembler::comiss(dst, Address(rscratch1, 0)); 2838 } 2839 } 2840 2841 2842 void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) { 2843 Condition negated_cond = negate_condition(cond); 2844 Label L; 2845 jcc(negated_cond, L); 2846 pushf(); // Preserve flags 2847 atomic_incl(counter_addr); 2848 popf(); 2849 bind(L); 2850 } 2851 2852 int MacroAssembler::corrected_idivl(Register reg) { 2853 // Full implementation of Java idiv and irem; checks for 2854 // special case as described in JVM spec., p.243 & p.271. 2855 // The function returns the (pc) offset of the idivl 2856 // instruction - may be needed for implicit exceptions. 2857 // 2858 // normal case special case 2859 // 2860 // input : rax,: dividend min_int 2861 // reg: divisor (may not be rax,/rdx) -1 2862 // 2863 // output: rax,: quotient (= rax, idiv reg) min_int 2864 // rdx: remainder (= rax, irem reg) 0 2865 assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register"); 2866 const int min_int = 0x80000000; 2867 Label normal_case, special_case; 2868 2869 // check for special case 2870 cmpl(rax, min_int); 2871 jcc(Assembler::notEqual, normal_case); 2872 xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0) 2873 cmpl(reg, -1); 2874 jcc(Assembler::equal, special_case); 2875 2876 // handle normal case 2877 bind(normal_case); 2878 cdql(); 2879 int idivl_offset = offset(); 2880 idivl(reg); 2881 2882 // normal and special case exit 2883 bind(special_case); 2884 2885 return idivl_offset; 2886 } 2887 2888 2889 2890 void MacroAssembler::decrementl(Register reg, int value) { 2891 if (value == min_jint) {subl(reg, value) ; return; } 2892 if (value < 0) { incrementl(reg, -value); return; } 2893 if (value == 0) { ; return; } 2894 if (value == 1 && UseIncDec) { decl(reg) ; return; } 2895 /* else */ { subl(reg, value) ; return; } 2896 } 2897 2898 void MacroAssembler::decrementl(Address dst, int value) { 2899 if (value == min_jint) {subl(dst, value) ; return; } 2900 if (value < 0) { incrementl(dst, -value); return; } 2901 if (value == 0) { ; return; } 2902 if (value == 1 && UseIncDec) { decl(dst) ; return; } 2903 /* else */ { subl(dst, value) ; return; } 2904 } 2905 2906 void MacroAssembler::division_with_shift (Register reg, int shift_value) { 2907 assert (shift_value > 0, "illegal shift value"); 2908 Label _is_positive; 2909 testl (reg, reg); 2910 jcc (Assembler::positive, _is_positive); 2911 int offset = (1 << shift_value) - 1 ; 2912 2913 if (offset == 1) { 2914 incrementl(reg); 2915 } else { 2916 addl(reg, offset); 2917 } 2918 2919 bind (_is_positive); 2920 sarl(reg, shift_value); 2921 } 2922 2923 void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) { 2924 if (reachable(src)) { 2925 Assembler::divsd(dst, as_Address(src)); 2926 } else { 2927 lea(rscratch1, src); 2928 Assembler::divsd(dst, Address(rscratch1, 0)); 2929 } 2930 } 2931 2932 void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) { 2933 if (reachable(src)) { 2934 Assembler::divss(dst, as_Address(src)); 2935 } else { 2936 lea(rscratch1, src); 2937 Assembler::divss(dst, Address(rscratch1, 0)); 2938 } 2939 } 2940 2941 // !defined(COMPILER2) is because of stupid core builds 2942 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2) || INCLUDE_JVMCI 2943 void MacroAssembler::empty_FPU_stack() { 2944 if (VM_Version::supports_mmx()) { 2945 emms(); 2946 } else { 2947 for (int i = 8; i-- > 0; ) ffree(i); 2948 } 2949 } 2950 #endif // !LP64 || C1 || !C2 || INCLUDE_JVMCI 2951 2952 2953 // Defines obj, preserves var_size_in_bytes 2954 void MacroAssembler::eden_allocate(Register obj, 2955 Register var_size_in_bytes, 2956 int con_size_in_bytes, 2957 Register t1, 2958 Label& slow_case) { 2959 assert(obj == rax, "obj must be in rax, for cmpxchg"); 2960 assert_different_registers(obj, var_size_in_bytes, t1); 2961 if (!Universe::heap()->supports_inline_contig_alloc()) { 2962 jmp(slow_case); 2963 } else { 2964 Register end = t1; 2965 Label retry; 2966 bind(retry); 2967 ExternalAddress heap_top((address) Universe::heap()->top_addr()); 2968 movptr(obj, heap_top); 2969 if (var_size_in_bytes == noreg) { 2970 lea(end, Address(obj, con_size_in_bytes)); 2971 } else { 2972 lea(end, Address(obj, var_size_in_bytes, Address::times_1)); 2973 } 2974 // if end < obj then we wrapped around => object too long => slow case 2975 cmpptr(end, obj); 2976 jcc(Assembler::below, slow_case); 2977 cmpptr(end, ExternalAddress((address) Universe::heap()->end_addr())); 2978 jcc(Assembler::above, slow_case); 2979 // Compare obj with the top addr, and if still equal, store the new top addr in 2980 // end at the address of the top addr pointer. Sets ZF if was equal, and clears 2981 // it otherwise. Use lock prefix for atomicity on MPs. 2982 locked_cmpxchgptr(end, heap_top); 2983 jcc(Assembler::notEqual, retry); 2984 } 2985 } 2986 2987 void MacroAssembler::enter() { 2988 push(rbp); 2989 mov(rbp, rsp); 2990 } 2991 2992 // A 5 byte nop that is safe for patching (see patch_verified_entry) 2993 void MacroAssembler::fat_nop() { 2994 if (UseAddressNop) { 2995 addr_nop_5(); 2996 } else { 2997 emit_int8(0x26); // es: 2998 emit_int8(0x2e); // cs: 2999 emit_int8(0x64); // fs: 3000 emit_int8(0x65); // gs: 3001 emit_int8((unsigned char)0x90); 3002 } 3003 } 3004 3005 void MacroAssembler::fcmp(Register tmp) { 3006 fcmp(tmp, 1, true, true); 3007 } 3008 3009 void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) { 3010 assert(!pop_right || pop_left, "usage error"); 3011 if (VM_Version::supports_cmov()) { 3012 assert(tmp == noreg, "unneeded temp"); 3013 if (pop_left) { 3014 fucomip(index); 3015 } else { 3016 fucomi(index); 3017 } 3018 if (pop_right) { 3019 fpop(); 3020 } 3021 } else { 3022 assert(tmp != noreg, "need temp"); 3023 if (pop_left) { 3024 if (pop_right) { 3025 fcompp(); 3026 } else { 3027 fcomp(index); 3028 } 3029 } else { 3030 fcom(index); 3031 } 3032 // convert FPU condition into eflags condition via rax, 3033 save_rax(tmp); 3034 fwait(); fnstsw_ax(); 3035 sahf(); 3036 restore_rax(tmp); 3037 } 3038 // condition codes set as follows: 3039 // 3040 // CF (corresponds to C0) if x < y 3041 // PF (corresponds to C2) if unordered 3042 // ZF (corresponds to C3) if x = y 3043 } 3044 3045 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) { 3046 fcmp2int(dst, unordered_is_less, 1, true, true); 3047 } 3048 3049 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) { 3050 fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right); 3051 Label L; 3052 if (unordered_is_less) { 3053 movl(dst, -1); 3054 jcc(Assembler::parity, L); 3055 jcc(Assembler::below , L); 3056 movl(dst, 0); 3057 jcc(Assembler::equal , L); 3058 increment(dst); 3059 } else { // unordered is greater 3060 movl(dst, 1); 3061 jcc(Assembler::parity, L); 3062 jcc(Assembler::above , L); 3063 movl(dst, 0); 3064 jcc(Assembler::equal , L); 3065 decrementl(dst); 3066 } 3067 bind(L); 3068 } 3069 3070 void MacroAssembler::fld_d(AddressLiteral src) { 3071 fld_d(as_Address(src)); 3072 } 3073 3074 void MacroAssembler::fld_s(AddressLiteral src) { 3075 fld_s(as_Address(src)); 3076 } 3077 3078 void MacroAssembler::fld_x(AddressLiteral src) { 3079 Assembler::fld_x(as_Address(src)); 3080 } 3081 3082 void MacroAssembler::fldcw(AddressLiteral src) { 3083 Assembler::fldcw(as_Address(src)); 3084 } 3085 3086 void MacroAssembler::mulpd(XMMRegister dst, AddressLiteral src) { 3087 if (reachable(src)) { 3088 Assembler::mulpd(dst, as_Address(src)); 3089 } else { 3090 lea(rscratch1, src); 3091 Assembler::mulpd(dst, Address(rscratch1, 0)); 3092 } 3093 } 3094 3095 void MacroAssembler::increase_precision() { 3096 subptr(rsp, BytesPerWord); 3097 fnstcw(Address(rsp, 0)); 3098 movl(rax, Address(rsp, 0)); 3099 orl(rax, 0x300); 3100 push(rax); 3101 fldcw(Address(rsp, 0)); 3102 pop(rax); 3103 } 3104 3105 void MacroAssembler::restore_precision() { 3106 fldcw(Address(rsp, 0)); 3107 addptr(rsp, BytesPerWord); 3108 } 3109 3110 void MacroAssembler::fpop() { 3111 ffree(); 3112 fincstp(); 3113 } 3114 3115 void MacroAssembler::load_float(Address src) { 3116 if (UseSSE >= 1) { 3117 movflt(xmm0, src); 3118 } else { 3119 LP64_ONLY(ShouldNotReachHere()); 3120 NOT_LP64(fld_s(src)); 3121 } 3122 } 3123 3124 void MacroAssembler::store_float(Address dst) { 3125 if (UseSSE >= 1) { 3126 movflt(dst, xmm0); 3127 } else { 3128 LP64_ONLY(ShouldNotReachHere()); 3129 NOT_LP64(fstp_s(dst)); 3130 } 3131 } 3132 3133 void MacroAssembler::load_double(Address src) { 3134 if (UseSSE >= 2) { 3135 movdbl(xmm0, src); 3136 } else { 3137 LP64_ONLY(ShouldNotReachHere()); 3138 NOT_LP64(fld_d(src)); 3139 } 3140 } 3141 3142 void MacroAssembler::store_double(Address dst) { 3143 if (UseSSE >= 2) { 3144 movdbl(dst, xmm0); 3145 } else { 3146 LP64_ONLY(ShouldNotReachHere()); 3147 NOT_LP64(fstp_d(dst)); 3148 } 3149 } 3150 3151 void MacroAssembler::fremr(Register tmp) { 3152 save_rax(tmp); 3153 { Label L; 3154 bind(L); 3155 fprem(); 3156 fwait(); fnstsw_ax(); 3157 #ifdef _LP64 3158 testl(rax, 0x400); 3159 jcc(Assembler::notEqual, L); 3160 #else 3161 sahf(); 3162 jcc(Assembler::parity, L); 3163 #endif // _LP64 3164 } 3165 restore_rax(tmp); 3166 // Result is in ST0. 3167 // Note: fxch & fpop to get rid of ST1 3168 // (otherwise FPU stack could overflow eventually) 3169 fxch(1); 3170 fpop(); 3171 } 3172 3173 // dst = c = a * b + c 3174 void MacroAssembler::fmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) { 3175 Assembler::vfmadd231sd(c, a, b); 3176 if (dst != c) { 3177 movdbl(dst, c); 3178 } 3179 } 3180 3181 // dst = c = a * b + c 3182 void MacroAssembler::fmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) { 3183 Assembler::vfmadd231ss(c, a, b); 3184 if (dst != c) { 3185 movflt(dst, c); 3186 } 3187 } 3188 3189 // dst = c = a * b + c 3190 void MacroAssembler::vfmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c, int vector_len) { 3191 Assembler::vfmadd231pd(c, a, b, vector_len); 3192 if (dst != c) { 3193 vmovdqu(dst, c); 3194 } 3195 } 3196 3197 // dst = c = a * b + c 3198 void MacroAssembler::vfmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c, int vector_len) { 3199 Assembler::vfmadd231ps(c, a, b, vector_len); 3200 if (dst != c) { 3201 vmovdqu(dst, c); 3202 } 3203 } 3204 3205 // dst = c = a * b + c 3206 void MacroAssembler::vfmad(XMMRegister dst, XMMRegister a, Address b, XMMRegister c, int vector_len) { 3207 Assembler::vfmadd231pd(c, a, b, vector_len); 3208 if (dst != c) { 3209 vmovdqu(dst, c); 3210 } 3211 } 3212 3213 // dst = c = a * b + c 3214 void MacroAssembler::vfmaf(XMMRegister dst, XMMRegister a, Address b, XMMRegister c, int vector_len) { 3215 Assembler::vfmadd231ps(c, a, b, vector_len); 3216 if (dst != c) { 3217 vmovdqu(dst, c); 3218 } 3219 } 3220 3221 void MacroAssembler::incrementl(AddressLiteral dst) { 3222 if (reachable(dst)) { 3223 incrementl(as_Address(dst)); 3224 } else { 3225 lea(rscratch1, dst); 3226 incrementl(Address(rscratch1, 0)); 3227 } 3228 } 3229 3230 void MacroAssembler::incrementl(ArrayAddress dst) { 3231 incrementl(as_Address(dst)); 3232 } 3233 3234 void MacroAssembler::incrementl(Register reg, int value) { 3235 if (value == min_jint) {addl(reg, value) ; return; } 3236 if (value < 0) { decrementl(reg, -value); return; } 3237 if (value == 0) { ; return; } 3238 if (value == 1 && UseIncDec) { incl(reg) ; return; } 3239 /* else */ { addl(reg, value) ; return; } 3240 } 3241 3242 void MacroAssembler::incrementl(Address dst, int value) { 3243 if (value == min_jint) {addl(dst, value) ; return; } 3244 if (value < 0) { decrementl(dst, -value); return; } 3245 if (value == 0) { ; return; } 3246 if (value == 1 && UseIncDec) { incl(dst) ; return; } 3247 /* else */ { addl(dst, value) ; return; } 3248 } 3249 3250 void MacroAssembler::jump(AddressLiteral dst) { 3251 if (reachable(dst)) { 3252 jmp_literal(dst.target(), dst.rspec()); 3253 } else { 3254 lea(rscratch1, dst); 3255 jmp(rscratch1); 3256 } 3257 } 3258 3259 void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) { 3260 if (reachable(dst)) { 3261 InstructionMark im(this); 3262 relocate(dst.reloc()); 3263 const int short_size = 2; 3264 const int long_size = 6; 3265 int offs = (intptr_t)dst.target() - ((intptr_t)pc()); 3266 if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) { 3267 // 0111 tttn #8-bit disp 3268 emit_int8(0x70 | cc); 3269 emit_int8((offs - short_size) & 0xFF); 3270 } else { 3271 // 0000 1111 1000 tttn #32-bit disp 3272 emit_int8(0x0F); 3273 emit_int8((unsigned char)(0x80 | cc)); 3274 emit_int32(offs - long_size); 3275 } 3276 } else { 3277 #ifdef ASSERT 3278 warning("reversing conditional branch"); 3279 #endif /* ASSERT */ 3280 Label skip; 3281 jccb(reverse[cc], skip); 3282 lea(rscratch1, dst); 3283 Assembler::jmp(rscratch1); 3284 bind(skip); 3285 } 3286 } 3287 3288 void MacroAssembler::ldmxcsr(AddressLiteral src) { 3289 if (reachable(src)) { 3290 Assembler::ldmxcsr(as_Address(src)); 3291 } else { 3292 lea(rscratch1, src); 3293 Assembler::ldmxcsr(Address(rscratch1, 0)); 3294 } 3295 } 3296 3297 int MacroAssembler::load_signed_byte(Register dst, Address src) { 3298 int off; 3299 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3300 off = offset(); 3301 movsbl(dst, src); // movsxb 3302 } else { 3303 off = load_unsigned_byte(dst, src); 3304 shll(dst, 24); 3305 sarl(dst, 24); 3306 } 3307 return off; 3308 } 3309 3310 // Note: load_signed_short used to be called load_signed_word. 3311 // Although the 'w' in x86 opcodes refers to the term "word" in the assembler 3312 // manual, which means 16 bits, that usage is found nowhere in HotSpot code. 3313 // The term "word" in HotSpot means a 32- or 64-bit machine word. 3314 int MacroAssembler::load_signed_short(Register dst, Address src) { 3315 int off; 3316 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3317 // This is dubious to me since it seems safe to do a signed 16 => 64 bit 3318 // version but this is what 64bit has always done. This seems to imply 3319 // that users are only using 32bits worth. 3320 off = offset(); 3321 movswl(dst, src); // movsxw 3322 } else { 3323 off = load_unsigned_short(dst, src); 3324 shll(dst, 16); 3325 sarl(dst, 16); 3326 } 3327 return off; 3328 } 3329 3330 int MacroAssembler::load_unsigned_byte(Register dst, Address src) { 3331 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16, 3332 // and "3.9 Partial Register Penalties", p. 22). 3333 int off; 3334 if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) { 3335 off = offset(); 3336 movzbl(dst, src); // movzxb 3337 } else { 3338 xorl(dst, dst); 3339 off = offset(); 3340 movb(dst, src); 3341 } 3342 return off; 3343 } 3344 3345 // Note: load_unsigned_short used to be called load_unsigned_word. 3346 int MacroAssembler::load_unsigned_short(Register dst, Address src) { 3347 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16, 3348 // and "3.9 Partial Register Penalties", p. 22). 3349 int off; 3350 if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) { 3351 off = offset(); 3352 movzwl(dst, src); // movzxw 3353 } else { 3354 xorl(dst, dst); 3355 off = offset(); 3356 movw(dst, src); 3357 } 3358 return off; 3359 } 3360 3361 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) { 3362 switch (size_in_bytes) { 3363 #ifndef _LP64 3364 case 8: 3365 assert(dst2 != noreg, "second dest register required"); 3366 movl(dst, src); 3367 movl(dst2, src.plus_disp(BytesPerInt)); 3368 break; 3369 #else 3370 case 8: movq(dst, src); break; 3371 #endif 3372 case 4: movl(dst, src); break; 3373 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break; 3374 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break; 3375 default: ShouldNotReachHere(); 3376 } 3377 } 3378 3379 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) { 3380 switch (size_in_bytes) { 3381 #ifndef _LP64 3382 case 8: 3383 assert(src2 != noreg, "second source register required"); 3384 movl(dst, src); 3385 movl(dst.plus_disp(BytesPerInt), src2); 3386 break; 3387 #else 3388 case 8: movq(dst, src); break; 3389 #endif 3390 case 4: movl(dst, src); break; 3391 case 2: movw(dst, src); break; 3392 case 1: movb(dst, src); break; 3393 default: ShouldNotReachHere(); 3394 } 3395 } 3396 3397 void MacroAssembler::mov32(AddressLiteral dst, Register src) { 3398 if (reachable(dst)) { 3399 movl(as_Address(dst), src); 3400 } else { 3401 lea(rscratch1, dst); 3402 movl(Address(rscratch1, 0), src); 3403 } 3404 } 3405 3406 void MacroAssembler::mov32(Register dst, AddressLiteral src) { 3407 if (reachable(src)) { 3408 movl(dst, as_Address(src)); 3409 } else { 3410 lea(rscratch1, src); 3411 movl(dst, Address(rscratch1, 0)); 3412 } 3413 } 3414 3415 // C++ bool manipulation 3416 3417 void MacroAssembler::movbool(Register dst, Address src) { 3418 if(sizeof(bool) == 1) 3419 movb(dst, src); 3420 else if(sizeof(bool) == 2) 3421 movw(dst, src); 3422 else if(sizeof(bool) == 4) 3423 movl(dst, src); 3424 else 3425 // unsupported 3426 ShouldNotReachHere(); 3427 } 3428 3429 void MacroAssembler::movbool(Address dst, bool boolconst) { 3430 if(sizeof(bool) == 1) 3431 movb(dst, (int) boolconst); 3432 else if(sizeof(bool) == 2) 3433 movw(dst, (int) boolconst); 3434 else if(sizeof(bool) == 4) 3435 movl(dst, (int) boolconst); 3436 else 3437 // unsupported 3438 ShouldNotReachHere(); 3439 } 3440 3441 void MacroAssembler::movbool(Address dst, Register src) { 3442 if(sizeof(bool) == 1) 3443 movb(dst, src); 3444 else if(sizeof(bool) == 2) 3445 movw(dst, src); 3446 else if(sizeof(bool) == 4) 3447 movl(dst, src); 3448 else 3449 // unsupported 3450 ShouldNotReachHere(); 3451 } 3452 3453 void MacroAssembler::movbyte(ArrayAddress dst, int src) { 3454 movb(as_Address(dst), src); 3455 } 3456 3457 void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) { 3458 if (reachable(src)) { 3459 movdl(dst, as_Address(src)); 3460 } else { 3461 lea(rscratch1, src); 3462 movdl(dst, Address(rscratch1, 0)); 3463 } 3464 } 3465 3466 void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) { 3467 if (reachable(src)) { 3468 movq(dst, as_Address(src)); 3469 } else { 3470 lea(rscratch1, src); 3471 movq(dst, Address(rscratch1, 0)); 3472 } 3473 } 3474 3475 void MacroAssembler::setvectmask(Register dst, Register src) { 3476 Assembler::movl(dst, 1); 3477 Assembler::shlxl(dst, dst, src); 3478 Assembler::decl(dst); 3479 Assembler::kmovdl(k1, dst); 3480 Assembler::movl(dst, src); 3481 } 3482 3483 void MacroAssembler::restorevectmask() { 3484 Assembler::knotwl(k1, k0); 3485 } 3486 3487 void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) { 3488 if (reachable(src)) { 3489 if (UseXmmLoadAndClearUpper) { 3490 movsd (dst, as_Address(src)); 3491 } else { 3492 movlpd(dst, as_Address(src)); 3493 } 3494 } else { 3495 lea(rscratch1, src); 3496 if (UseXmmLoadAndClearUpper) { 3497 movsd (dst, Address(rscratch1, 0)); 3498 } else { 3499 movlpd(dst, Address(rscratch1, 0)); 3500 } 3501 } 3502 } 3503 3504 void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) { 3505 if (reachable(src)) { 3506 movss(dst, as_Address(src)); 3507 } else { 3508 lea(rscratch1, src); 3509 movss(dst, Address(rscratch1, 0)); 3510 } 3511 } 3512 3513 void MacroAssembler::movptr(Register dst, Register src) { 3514 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3515 } 3516 3517 void MacroAssembler::movptr(Register dst, Address src) { 3518 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3519 } 3520 3521 // src should NEVER be a real pointer. Use AddressLiteral for true pointers 3522 void MacroAssembler::movptr(Register dst, intptr_t src) { 3523 LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src)); 3524 } 3525 3526 void MacroAssembler::movptr(Address dst, Register src) { 3527 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3528 } 3529 3530 void MacroAssembler::movdqu(Address dst, XMMRegister src) { 3531 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (src->encoding() > 15)) { 3532 Assembler::vextractf32x4(dst, src, 0); 3533 } else { 3534 Assembler::movdqu(dst, src); 3535 } 3536 } 3537 3538 void MacroAssembler::movdqu(XMMRegister dst, Address src) { 3539 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (dst->encoding() > 15)) { 3540 Assembler::vinsertf32x4(dst, dst, src, 0); 3541 } else { 3542 Assembler::movdqu(dst, src); 3543 } 3544 } 3545 3546 void MacroAssembler::movdqu(XMMRegister dst, XMMRegister src) { 3547 if (UseAVX > 2 && !VM_Version::supports_avx512vl()) { 3548 Assembler::evmovdqul(dst, src, Assembler::AVX_512bit); 3549 } else { 3550 Assembler::movdqu(dst, src); 3551 } 3552 } 3553 3554 void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src, Register scratchReg) { 3555 if (reachable(src)) { 3556 movdqu(dst, as_Address(src)); 3557 } else { 3558 lea(scratchReg, src); 3559 movdqu(dst, Address(scratchReg, 0)); 3560 } 3561 } 3562 3563 void MacroAssembler::vmovdqu(Address dst, XMMRegister src) { 3564 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (src->encoding() > 15)) { 3565 vextractf64x4_low(dst, src); 3566 } else { 3567 Assembler::vmovdqu(dst, src); 3568 } 3569 } 3570 3571 void MacroAssembler::vmovdqu(XMMRegister dst, Address src) { 3572 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (dst->encoding() > 15)) { 3573 vinsertf64x4_low(dst, src); 3574 } else { 3575 Assembler::vmovdqu(dst, src); 3576 } 3577 } 3578 3579 void MacroAssembler::vmovdqu(XMMRegister dst, XMMRegister src) { 3580 if (UseAVX > 2 && !VM_Version::supports_avx512vl()) { 3581 Assembler::evmovdqul(dst, src, Assembler::AVX_512bit); 3582 } 3583 else { 3584 Assembler::vmovdqu(dst, src); 3585 } 3586 } 3587 3588 void MacroAssembler::vmovdqu(XMMRegister dst, AddressLiteral src) { 3589 if (reachable(src)) { 3590 vmovdqu(dst, as_Address(src)); 3591 } 3592 else { 3593 lea(rscratch1, src); 3594 vmovdqu(dst, Address(rscratch1, 0)); 3595 } 3596 } 3597 3598 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) { 3599 if (reachable(src)) { 3600 Assembler::movdqa(dst, as_Address(src)); 3601 } else { 3602 lea(rscratch1, src); 3603 Assembler::movdqa(dst, Address(rscratch1, 0)); 3604 } 3605 } 3606 3607 void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) { 3608 if (reachable(src)) { 3609 Assembler::movsd(dst, as_Address(src)); 3610 } else { 3611 lea(rscratch1, src); 3612 Assembler::movsd(dst, Address(rscratch1, 0)); 3613 } 3614 } 3615 3616 void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) { 3617 if (reachable(src)) { 3618 Assembler::movss(dst, as_Address(src)); 3619 } else { 3620 lea(rscratch1, src); 3621 Assembler::movss(dst, Address(rscratch1, 0)); 3622 } 3623 } 3624 3625 void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) { 3626 if (reachable(src)) { 3627 Assembler::mulsd(dst, as_Address(src)); 3628 } else { 3629 lea(rscratch1, src); 3630 Assembler::mulsd(dst, Address(rscratch1, 0)); 3631 } 3632 } 3633 3634 void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) { 3635 if (reachable(src)) { 3636 Assembler::mulss(dst, as_Address(src)); 3637 } else { 3638 lea(rscratch1, src); 3639 Assembler::mulss(dst, Address(rscratch1, 0)); 3640 } 3641 } 3642 3643 void MacroAssembler::null_check(Register reg, int offset) { 3644 if (needs_explicit_null_check(offset)) { 3645 // provoke OS NULL exception if reg = NULL by 3646 // accessing M[reg] w/o changing any (non-CC) registers 3647 // NOTE: cmpl is plenty here to provoke a segv 3648 cmpptr(rax, Address(reg, 0)); 3649 // Note: should probably use testl(rax, Address(reg, 0)); 3650 // may be shorter code (however, this version of 3651 // testl needs to be implemented first) 3652 } else { 3653 // nothing to do, (later) access of M[reg + offset] 3654 // will provoke OS NULL exception if reg = NULL 3655 } 3656 } 3657 3658 void MacroAssembler::os_breakpoint() { 3659 // instead of directly emitting a breakpoint, call os:breakpoint for better debugability 3660 // (e.g., MSVC can't call ps() otherwise) 3661 call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint))); 3662 } 3663 3664 void MacroAssembler::unimplemented(const char* what) { 3665 const char* buf = NULL; 3666 { 3667 ResourceMark rm; 3668 stringStream ss; 3669 ss.print("unimplemented: %s", what); 3670 buf = code_string(ss.as_string()); 3671 } 3672 stop(buf); 3673 } 3674 3675 #ifdef _LP64 3676 #define XSTATE_BV 0x200 3677 #endif 3678 3679 void MacroAssembler::pop_CPU_state() { 3680 pop_FPU_state(); 3681 pop_IU_state(); 3682 } 3683 3684 void MacroAssembler::pop_FPU_state() { 3685 #ifndef _LP64 3686 frstor(Address(rsp, 0)); 3687 #else 3688 fxrstor(Address(rsp, 0)); 3689 #endif 3690 addptr(rsp, FPUStateSizeInWords * wordSize); 3691 } 3692 3693 void MacroAssembler::pop_IU_state() { 3694 popa(); 3695 LP64_ONLY(addq(rsp, 8)); 3696 popf(); 3697 } 3698 3699 // Save Integer and Float state 3700 // Warning: Stack must be 16 byte aligned (64bit) 3701 void MacroAssembler::push_CPU_state() { 3702 push_IU_state(); 3703 push_FPU_state(); 3704 } 3705 3706 void MacroAssembler::push_FPU_state() { 3707 subptr(rsp, FPUStateSizeInWords * wordSize); 3708 #ifndef _LP64 3709 fnsave(Address(rsp, 0)); 3710 fwait(); 3711 #else 3712 fxsave(Address(rsp, 0)); 3713 #endif // LP64 3714 } 3715 3716 void MacroAssembler::push_IU_state() { 3717 // Push flags first because pusha kills them 3718 pushf(); 3719 // Make sure rsp stays 16-byte aligned 3720 LP64_ONLY(subq(rsp, 8)); 3721 pusha(); 3722 } 3723 3724 void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp) { // determine java_thread register 3725 if (!java_thread->is_valid()) { 3726 java_thread = rdi; 3727 get_thread(java_thread); 3728 } 3729 // we must set sp to zero to clear frame 3730 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD); 3731 if (clear_fp) { 3732 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD); 3733 } 3734 3735 // Always clear the pc because it could have been set by make_walkable() 3736 movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD); 3737 3738 vzeroupper(); 3739 } 3740 3741 void MacroAssembler::restore_rax(Register tmp) { 3742 if (tmp == noreg) pop(rax); 3743 else if (tmp != rax) mov(rax, tmp); 3744 } 3745 3746 void MacroAssembler::round_to(Register reg, int modulus) { 3747 addptr(reg, modulus - 1); 3748 andptr(reg, -modulus); 3749 } 3750 3751 void MacroAssembler::save_rax(Register tmp) { 3752 if (tmp == noreg) push(rax); 3753 else if (tmp != rax) mov(tmp, rax); 3754 } 3755 3756 // Write serialization page so VM thread can do a pseudo remote membar. 3757 // We use the current thread pointer to calculate a thread specific 3758 // offset to write to within the page. This minimizes bus traffic 3759 // due to cache line collision. 3760 void MacroAssembler::serialize_memory(Register thread, Register tmp) { 3761 movl(tmp, thread); 3762 shrl(tmp, os::get_serialize_page_shift_count()); 3763 andl(tmp, (os::vm_page_size() - sizeof(int))); 3764 3765 Address index(noreg, tmp, Address::times_1); 3766 ExternalAddress page(os::get_memory_serialize_page()); 3767 3768 // Size of store must match masking code above 3769 movl(as_Address(ArrayAddress(page, index)), tmp); 3770 } 3771 3772 void MacroAssembler::safepoint_poll(Label& slow_path, Register thread_reg, Register temp_reg) { 3773 if (SafepointMechanism::uses_thread_local_poll()) { 3774 #ifdef _LP64 3775 assert(thread_reg == r15_thread, "should be"); 3776 #else 3777 if (thread_reg == noreg) { 3778 thread_reg = temp_reg; 3779 get_thread(thread_reg); 3780 } 3781 #endif 3782 testb(Address(thread_reg, Thread::polling_page_offset()), SafepointMechanism::poll_bit()); 3783 jcc(Assembler::notZero, slow_path); // handshake bit set implies poll 3784 } else { 3785 cmp32(ExternalAddress(SafepointSynchronize::address_of_state()), 3786 SafepointSynchronize::_not_synchronized); 3787 jcc(Assembler::notEqual, slow_path); 3788 } 3789 } 3790 3791 // Calls to C land 3792 // 3793 // When entering C land, the rbp, & rsp of the last Java frame have to be recorded 3794 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp 3795 // has to be reset to 0. This is required to allow proper stack traversal. 3796 void MacroAssembler::set_last_Java_frame(Register java_thread, 3797 Register last_java_sp, 3798 Register last_java_fp, 3799 address last_java_pc) { 3800 vzeroupper(); 3801 // determine java_thread register 3802 if (!java_thread->is_valid()) { 3803 java_thread = rdi; 3804 get_thread(java_thread); 3805 } 3806 // determine last_java_sp register 3807 if (!last_java_sp->is_valid()) { 3808 last_java_sp = rsp; 3809 } 3810 3811 // last_java_fp is optional 3812 3813 if (last_java_fp->is_valid()) { 3814 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp); 3815 } 3816 3817 // last_java_pc is optional 3818 3819 if (last_java_pc != NULL) { 3820 lea(Address(java_thread, 3821 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()), 3822 InternalAddress(last_java_pc)); 3823 3824 } 3825 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp); 3826 } 3827 3828 void MacroAssembler::shlptr(Register dst, int imm8) { 3829 LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8)); 3830 } 3831 3832 void MacroAssembler::shrptr(Register dst, int imm8) { 3833 LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8)); 3834 } 3835 3836 void MacroAssembler::sign_extend_byte(Register reg) { 3837 if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) { 3838 movsbl(reg, reg); // movsxb 3839 } else { 3840 shll(reg, 24); 3841 sarl(reg, 24); 3842 } 3843 } 3844 3845 void MacroAssembler::sign_extend_short(Register reg) { 3846 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3847 movswl(reg, reg); // movsxw 3848 } else { 3849 shll(reg, 16); 3850 sarl(reg, 16); 3851 } 3852 } 3853 3854 void MacroAssembler::testl(Register dst, AddressLiteral src) { 3855 assert(reachable(src), "Address should be reachable"); 3856 testl(dst, as_Address(src)); 3857 } 3858 3859 void MacroAssembler::pcmpeqb(XMMRegister dst, XMMRegister src) { 3860 int dst_enc = dst->encoding(); 3861 int src_enc = src->encoding(); 3862 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 3863 Assembler::pcmpeqb(dst, src); 3864 } else if ((dst_enc < 16) && (src_enc < 16)) { 3865 Assembler::pcmpeqb(dst, src); 3866 } else if (src_enc < 16) { 3867 subptr(rsp, 64); 3868 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3869 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3870 Assembler::pcmpeqb(xmm0, src); 3871 movdqu(dst, xmm0); 3872 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3873 addptr(rsp, 64); 3874 } else if (dst_enc < 16) { 3875 subptr(rsp, 64); 3876 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3877 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3878 Assembler::pcmpeqb(dst, xmm0); 3879 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3880 addptr(rsp, 64); 3881 } else { 3882 subptr(rsp, 64); 3883 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3884 subptr(rsp, 64); 3885 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 3886 movdqu(xmm0, src); 3887 movdqu(xmm1, dst); 3888 Assembler::pcmpeqb(xmm1, xmm0); 3889 movdqu(dst, xmm1); 3890 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 3891 addptr(rsp, 64); 3892 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3893 addptr(rsp, 64); 3894 } 3895 } 3896 3897 void MacroAssembler::pcmpeqw(XMMRegister dst, XMMRegister src) { 3898 int dst_enc = dst->encoding(); 3899 int src_enc = src->encoding(); 3900 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 3901 Assembler::pcmpeqw(dst, src); 3902 } else if ((dst_enc < 16) && (src_enc < 16)) { 3903 Assembler::pcmpeqw(dst, src); 3904 } else if (src_enc < 16) { 3905 subptr(rsp, 64); 3906 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3907 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3908 Assembler::pcmpeqw(xmm0, src); 3909 movdqu(dst, xmm0); 3910 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3911 addptr(rsp, 64); 3912 } else if (dst_enc < 16) { 3913 subptr(rsp, 64); 3914 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3915 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3916 Assembler::pcmpeqw(dst, xmm0); 3917 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3918 addptr(rsp, 64); 3919 } else { 3920 subptr(rsp, 64); 3921 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3922 subptr(rsp, 64); 3923 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 3924 movdqu(xmm0, src); 3925 movdqu(xmm1, dst); 3926 Assembler::pcmpeqw(xmm1, xmm0); 3927 movdqu(dst, xmm1); 3928 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 3929 addptr(rsp, 64); 3930 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3931 addptr(rsp, 64); 3932 } 3933 } 3934 3935 void MacroAssembler::pcmpestri(XMMRegister dst, Address src, int imm8) { 3936 int dst_enc = dst->encoding(); 3937 if (dst_enc < 16) { 3938 Assembler::pcmpestri(dst, src, imm8); 3939 } else { 3940 subptr(rsp, 64); 3941 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3942 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3943 Assembler::pcmpestri(xmm0, src, imm8); 3944 movdqu(dst, xmm0); 3945 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3946 addptr(rsp, 64); 3947 } 3948 } 3949 3950 void MacroAssembler::pcmpestri(XMMRegister dst, XMMRegister src, int imm8) { 3951 int dst_enc = dst->encoding(); 3952 int src_enc = src->encoding(); 3953 if ((dst_enc < 16) && (src_enc < 16)) { 3954 Assembler::pcmpestri(dst, src, imm8); 3955 } else if (src_enc < 16) { 3956 subptr(rsp, 64); 3957 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3958 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3959 Assembler::pcmpestri(xmm0, src, imm8); 3960 movdqu(dst, xmm0); 3961 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3962 addptr(rsp, 64); 3963 } else if (dst_enc < 16) { 3964 subptr(rsp, 64); 3965 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3966 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3967 Assembler::pcmpestri(dst, xmm0, imm8); 3968 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3969 addptr(rsp, 64); 3970 } else { 3971 subptr(rsp, 64); 3972 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3973 subptr(rsp, 64); 3974 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 3975 movdqu(xmm0, src); 3976 movdqu(xmm1, dst); 3977 Assembler::pcmpestri(xmm1, xmm0, imm8); 3978 movdqu(dst, xmm1); 3979 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 3980 addptr(rsp, 64); 3981 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3982 addptr(rsp, 64); 3983 } 3984 } 3985 3986 void MacroAssembler::pmovzxbw(XMMRegister dst, XMMRegister src) { 3987 int dst_enc = dst->encoding(); 3988 int src_enc = src->encoding(); 3989 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 3990 Assembler::pmovzxbw(dst, src); 3991 } else if ((dst_enc < 16) && (src_enc < 16)) { 3992 Assembler::pmovzxbw(dst, src); 3993 } else if (src_enc < 16) { 3994 subptr(rsp, 64); 3995 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3996 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3997 Assembler::pmovzxbw(xmm0, src); 3998 movdqu(dst, xmm0); 3999 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4000 addptr(rsp, 64); 4001 } else if (dst_enc < 16) { 4002 subptr(rsp, 64); 4003 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4004 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4005 Assembler::pmovzxbw(dst, xmm0); 4006 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4007 addptr(rsp, 64); 4008 } else { 4009 subptr(rsp, 64); 4010 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4011 subptr(rsp, 64); 4012 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4013 movdqu(xmm0, src); 4014 movdqu(xmm1, dst); 4015 Assembler::pmovzxbw(xmm1, xmm0); 4016 movdqu(dst, xmm1); 4017 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4018 addptr(rsp, 64); 4019 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4020 addptr(rsp, 64); 4021 } 4022 } 4023 4024 void MacroAssembler::pmovzxbw(XMMRegister dst, Address src) { 4025 int dst_enc = dst->encoding(); 4026 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4027 Assembler::pmovzxbw(dst, src); 4028 } else if (dst_enc < 16) { 4029 Assembler::pmovzxbw(dst, src); 4030 } else { 4031 subptr(rsp, 64); 4032 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4033 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4034 Assembler::pmovzxbw(xmm0, src); 4035 movdqu(dst, xmm0); 4036 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4037 addptr(rsp, 64); 4038 } 4039 } 4040 4041 void MacroAssembler::pmovmskb(Register dst, XMMRegister src) { 4042 int src_enc = src->encoding(); 4043 if (src_enc < 16) { 4044 Assembler::pmovmskb(dst, src); 4045 } else { 4046 subptr(rsp, 64); 4047 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4048 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4049 Assembler::pmovmskb(dst, xmm0); 4050 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4051 addptr(rsp, 64); 4052 } 4053 } 4054 4055 void MacroAssembler::ptest(XMMRegister dst, XMMRegister src) { 4056 int dst_enc = dst->encoding(); 4057 int src_enc = src->encoding(); 4058 if ((dst_enc < 16) && (src_enc < 16)) { 4059 Assembler::ptest(dst, src); 4060 } else if (src_enc < 16) { 4061 subptr(rsp, 64); 4062 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4063 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4064 Assembler::ptest(xmm0, src); 4065 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4066 addptr(rsp, 64); 4067 } else if (dst_enc < 16) { 4068 subptr(rsp, 64); 4069 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4070 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4071 Assembler::ptest(dst, xmm0); 4072 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4073 addptr(rsp, 64); 4074 } else { 4075 subptr(rsp, 64); 4076 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4077 subptr(rsp, 64); 4078 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4079 movdqu(xmm0, src); 4080 movdqu(xmm1, dst); 4081 Assembler::ptest(xmm1, xmm0); 4082 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4083 addptr(rsp, 64); 4084 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4085 addptr(rsp, 64); 4086 } 4087 } 4088 4089 void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) { 4090 if (reachable(src)) { 4091 Assembler::sqrtsd(dst, as_Address(src)); 4092 } else { 4093 lea(rscratch1, src); 4094 Assembler::sqrtsd(dst, Address(rscratch1, 0)); 4095 } 4096 } 4097 4098 void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) { 4099 if (reachable(src)) { 4100 Assembler::sqrtss(dst, as_Address(src)); 4101 } else { 4102 lea(rscratch1, src); 4103 Assembler::sqrtss(dst, Address(rscratch1, 0)); 4104 } 4105 } 4106 4107 void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) { 4108 if (reachable(src)) { 4109 Assembler::subsd(dst, as_Address(src)); 4110 } else { 4111 lea(rscratch1, src); 4112 Assembler::subsd(dst, Address(rscratch1, 0)); 4113 } 4114 } 4115 4116 void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) { 4117 if (reachable(src)) { 4118 Assembler::subss(dst, as_Address(src)); 4119 } else { 4120 lea(rscratch1, src); 4121 Assembler::subss(dst, Address(rscratch1, 0)); 4122 } 4123 } 4124 4125 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) { 4126 if (reachable(src)) { 4127 Assembler::ucomisd(dst, as_Address(src)); 4128 } else { 4129 lea(rscratch1, src); 4130 Assembler::ucomisd(dst, Address(rscratch1, 0)); 4131 } 4132 } 4133 4134 void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) { 4135 if (reachable(src)) { 4136 Assembler::ucomiss(dst, as_Address(src)); 4137 } else { 4138 lea(rscratch1, src); 4139 Assembler::ucomiss(dst, Address(rscratch1, 0)); 4140 } 4141 } 4142 4143 void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src) { 4144 // Used in sign-bit flipping with aligned address. 4145 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 4146 if (reachable(src)) { 4147 Assembler::xorpd(dst, as_Address(src)); 4148 } else { 4149 lea(rscratch1, src); 4150 Assembler::xorpd(dst, Address(rscratch1, 0)); 4151 } 4152 } 4153 4154 void MacroAssembler::xorpd(XMMRegister dst, XMMRegister src) { 4155 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) { 4156 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit); 4157 } 4158 else { 4159 Assembler::xorpd(dst, src); 4160 } 4161 } 4162 4163 void MacroAssembler::xorps(XMMRegister dst, XMMRegister src) { 4164 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) { 4165 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit); 4166 } else { 4167 Assembler::xorps(dst, src); 4168 } 4169 } 4170 4171 void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src) { 4172 // Used in sign-bit flipping with aligned address. 4173 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 4174 if (reachable(src)) { 4175 Assembler::xorps(dst, as_Address(src)); 4176 } else { 4177 lea(rscratch1, src); 4178 Assembler::xorps(dst, Address(rscratch1, 0)); 4179 } 4180 } 4181 4182 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) { 4183 // Used in sign-bit flipping with aligned address. 4184 bool aligned_adr = (((intptr_t)src.target() & 15) == 0); 4185 assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes"); 4186 if (reachable(src)) { 4187 Assembler::pshufb(dst, as_Address(src)); 4188 } else { 4189 lea(rscratch1, src); 4190 Assembler::pshufb(dst, Address(rscratch1, 0)); 4191 } 4192 } 4193 4194 // AVX 3-operands instructions 4195 4196 void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4197 if (reachable(src)) { 4198 vaddsd(dst, nds, as_Address(src)); 4199 } else { 4200 lea(rscratch1, src); 4201 vaddsd(dst, nds, Address(rscratch1, 0)); 4202 } 4203 } 4204 4205 void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4206 if (reachable(src)) { 4207 vaddss(dst, nds, as_Address(src)); 4208 } else { 4209 lea(rscratch1, src); 4210 vaddss(dst, nds, Address(rscratch1, 0)); 4211 } 4212 } 4213 4214 void MacroAssembler::vabsss(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) { 4215 int dst_enc = dst->encoding(); 4216 int nds_enc = nds->encoding(); 4217 int src_enc = src->encoding(); 4218 if ((dst_enc < 16) && (nds_enc < 16)) { 4219 vandps(dst, nds, negate_field, vector_len); 4220 } else if ((src_enc < 16) && (dst_enc < 16)) { 4221 evmovdqul(src, nds, Assembler::AVX_512bit); 4222 vandps(dst, src, negate_field, vector_len); 4223 } else if (src_enc < 16) { 4224 evmovdqul(src, nds, Assembler::AVX_512bit); 4225 vandps(src, src, negate_field, vector_len); 4226 evmovdqul(dst, src, Assembler::AVX_512bit); 4227 } else if (dst_enc < 16) { 4228 evmovdqul(src, xmm0, Assembler::AVX_512bit); 4229 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4230 vandps(dst, xmm0, negate_field, vector_len); 4231 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4232 } else { 4233 if (src_enc != dst_enc) { 4234 evmovdqul(src, xmm0, Assembler::AVX_512bit); 4235 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4236 vandps(xmm0, xmm0, negate_field, vector_len); 4237 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4238 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4239 } else { 4240 subptr(rsp, 64); 4241 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4242 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4243 vandps(xmm0, xmm0, negate_field, vector_len); 4244 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4245 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4246 addptr(rsp, 64); 4247 } 4248 } 4249 } 4250 4251 void MacroAssembler::vabssd(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) { 4252 int dst_enc = dst->encoding(); 4253 int nds_enc = nds->encoding(); 4254 int src_enc = src->encoding(); 4255 if ((dst_enc < 16) && (nds_enc < 16)) { 4256 vandpd(dst, nds, negate_field, vector_len); 4257 } else if ((src_enc < 16) && (dst_enc < 16)) { 4258 evmovdqul(src, nds, Assembler::AVX_512bit); 4259 vandpd(dst, src, negate_field, vector_len); 4260 } else if (src_enc < 16) { 4261 evmovdqul(src, nds, Assembler::AVX_512bit); 4262 vandpd(src, src, negate_field, vector_len); 4263 evmovdqul(dst, src, Assembler::AVX_512bit); 4264 } else if (dst_enc < 16) { 4265 evmovdqul(src, xmm0, Assembler::AVX_512bit); 4266 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4267 vandpd(dst, xmm0, negate_field, vector_len); 4268 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4269 } else { 4270 if (src_enc != dst_enc) { 4271 evmovdqul(src, xmm0, Assembler::AVX_512bit); 4272 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4273 vandpd(xmm0, xmm0, negate_field, vector_len); 4274 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4275 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4276 } else { 4277 subptr(rsp, 64); 4278 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4279 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4280 vandpd(xmm0, xmm0, negate_field, vector_len); 4281 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4282 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4283 addptr(rsp, 64); 4284 } 4285 } 4286 } 4287 4288 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4289 int dst_enc = dst->encoding(); 4290 int nds_enc = nds->encoding(); 4291 int src_enc = src->encoding(); 4292 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4293 Assembler::vpaddb(dst, nds, src, vector_len); 4294 } else if ((dst_enc < 16) && (src_enc < 16)) { 4295 Assembler::vpaddb(dst, dst, src, vector_len); 4296 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4297 // use nds as scratch for src 4298 evmovdqul(nds, src, Assembler::AVX_512bit); 4299 Assembler::vpaddb(dst, dst, nds, vector_len); 4300 } else if ((src_enc < 16) && (nds_enc < 16)) { 4301 // use nds as scratch for dst 4302 evmovdqul(nds, dst, Assembler::AVX_512bit); 4303 Assembler::vpaddb(nds, nds, src, vector_len); 4304 evmovdqul(dst, nds, Assembler::AVX_512bit); 4305 } else if (dst_enc < 16) { 4306 // use nds as scatch for xmm0 to hold src 4307 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4308 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4309 Assembler::vpaddb(dst, dst, xmm0, vector_len); 4310 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4311 } else { 4312 // worse case scenario, all regs are in the upper bank 4313 subptr(rsp, 64); 4314 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4315 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4316 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4317 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4318 Assembler::vpaddb(xmm0, xmm0, xmm1, vector_len); 4319 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4320 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4321 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4322 addptr(rsp, 64); 4323 } 4324 } 4325 4326 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4327 int dst_enc = dst->encoding(); 4328 int nds_enc = nds->encoding(); 4329 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4330 Assembler::vpaddb(dst, nds, src, vector_len); 4331 } else if (dst_enc < 16) { 4332 Assembler::vpaddb(dst, dst, src, vector_len); 4333 } else if (nds_enc < 16) { 4334 // implies dst_enc in upper bank with src as scratch 4335 evmovdqul(nds, dst, Assembler::AVX_512bit); 4336 Assembler::vpaddb(nds, nds, src, vector_len); 4337 evmovdqul(dst, nds, Assembler::AVX_512bit); 4338 } else { 4339 // worse case scenario, all regs in upper bank 4340 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4341 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4342 Assembler::vpaddb(xmm0, xmm0, src, vector_len); 4343 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4344 } 4345 } 4346 4347 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4348 int dst_enc = dst->encoding(); 4349 int nds_enc = nds->encoding(); 4350 int src_enc = src->encoding(); 4351 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4352 Assembler::vpaddw(dst, nds, src, vector_len); 4353 } else if ((dst_enc < 16) && (src_enc < 16)) { 4354 Assembler::vpaddw(dst, dst, src, vector_len); 4355 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4356 // use nds as scratch for src 4357 evmovdqul(nds, src, Assembler::AVX_512bit); 4358 Assembler::vpaddw(dst, dst, nds, vector_len); 4359 } else if ((src_enc < 16) && (nds_enc < 16)) { 4360 // use nds as scratch for dst 4361 evmovdqul(nds, dst, Assembler::AVX_512bit); 4362 Assembler::vpaddw(nds, nds, src, vector_len); 4363 evmovdqul(dst, nds, Assembler::AVX_512bit); 4364 } else if (dst_enc < 16) { 4365 // use nds as scatch for xmm0 to hold src 4366 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4367 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4368 Assembler::vpaddw(dst, dst, xmm0, vector_len); 4369 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4370 } else { 4371 // worse case scenario, all regs are in the upper bank 4372 subptr(rsp, 64); 4373 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4374 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4375 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4376 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4377 Assembler::vpaddw(xmm0, xmm0, xmm1, vector_len); 4378 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4379 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4380 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4381 addptr(rsp, 64); 4382 } 4383 } 4384 4385 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4386 int dst_enc = dst->encoding(); 4387 int nds_enc = nds->encoding(); 4388 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4389 Assembler::vpaddw(dst, nds, src, vector_len); 4390 } else if (dst_enc < 16) { 4391 Assembler::vpaddw(dst, dst, src, vector_len); 4392 } else if (nds_enc < 16) { 4393 // implies dst_enc in upper bank with src as scratch 4394 evmovdqul(nds, dst, Assembler::AVX_512bit); 4395 Assembler::vpaddw(nds, nds, src, vector_len); 4396 evmovdqul(dst, nds, Assembler::AVX_512bit); 4397 } else { 4398 // worse case scenario, all regs in upper bank 4399 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4400 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4401 Assembler::vpaddw(xmm0, xmm0, src, vector_len); 4402 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4403 } 4404 } 4405 4406 void MacroAssembler::vpand(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 4407 if (reachable(src)) { 4408 Assembler::vpand(dst, nds, as_Address(src), vector_len); 4409 } else { 4410 lea(rscratch1, src); 4411 Assembler::vpand(dst, nds, Address(rscratch1, 0), vector_len); 4412 } 4413 } 4414 4415 void MacroAssembler::vpbroadcastw(XMMRegister dst, XMMRegister src) { 4416 int dst_enc = dst->encoding(); 4417 int src_enc = src->encoding(); 4418 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4419 Assembler::vpbroadcastw(dst, src); 4420 } else if ((dst_enc < 16) && (src_enc < 16)) { 4421 Assembler::vpbroadcastw(dst, src); 4422 } else if (src_enc < 16) { 4423 subptr(rsp, 64); 4424 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4425 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4426 Assembler::vpbroadcastw(xmm0, src); 4427 movdqu(dst, xmm0); 4428 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4429 addptr(rsp, 64); 4430 } else if (dst_enc < 16) { 4431 subptr(rsp, 64); 4432 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4433 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4434 Assembler::vpbroadcastw(dst, xmm0); 4435 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4436 addptr(rsp, 64); 4437 } else { 4438 subptr(rsp, 64); 4439 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4440 subptr(rsp, 64); 4441 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4442 movdqu(xmm0, src); 4443 movdqu(xmm1, dst); 4444 Assembler::vpbroadcastw(xmm1, xmm0); 4445 movdqu(dst, xmm1); 4446 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4447 addptr(rsp, 64); 4448 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4449 addptr(rsp, 64); 4450 } 4451 } 4452 4453 void MacroAssembler::vpcmpeqb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4454 int dst_enc = dst->encoding(); 4455 int nds_enc = nds->encoding(); 4456 int src_enc = src->encoding(); 4457 assert(dst_enc == nds_enc, ""); 4458 if ((dst_enc < 16) && (src_enc < 16)) { 4459 Assembler::vpcmpeqb(dst, nds, src, vector_len); 4460 } else if (src_enc < 16) { 4461 subptr(rsp, 64); 4462 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4463 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4464 Assembler::vpcmpeqb(xmm0, xmm0, src, vector_len); 4465 movdqu(dst, xmm0); 4466 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4467 addptr(rsp, 64); 4468 } else if (dst_enc < 16) { 4469 subptr(rsp, 64); 4470 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4471 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4472 Assembler::vpcmpeqb(dst, dst, xmm0, vector_len); 4473 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4474 addptr(rsp, 64); 4475 } else { 4476 subptr(rsp, 64); 4477 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4478 subptr(rsp, 64); 4479 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4480 movdqu(xmm0, src); 4481 movdqu(xmm1, dst); 4482 Assembler::vpcmpeqb(xmm1, xmm1, xmm0, vector_len); 4483 movdqu(dst, xmm1); 4484 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4485 addptr(rsp, 64); 4486 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4487 addptr(rsp, 64); 4488 } 4489 } 4490 4491 void MacroAssembler::vpcmpeqw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4492 int dst_enc = dst->encoding(); 4493 int nds_enc = nds->encoding(); 4494 int src_enc = src->encoding(); 4495 assert(dst_enc == nds_enc, ""); 4496 if ((dst_enc < 16) && (src_enc < 16)) { 4497 Assembler::vpcmpeqw(dst, nds, src, vector_len); 4498 } else if (src_enc < 16) { 4499 subptr(rsp, 64); 4500 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4501 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4502 Assembler::vpcmpeqw(xmm0, xmm0, src, vector_len); 4503 movdqu(dst, xmm0); 4504 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4505 addptr(rsp, 64); 4506 } else if (dst_enc < 16) { 4507 subptr(rsp, 64); 4508 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4509 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4510 Assembler::vpcmpeqw(dst, dst, xmm0, vector_len); 4511 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4512 addptr(rsp, 64); 4513 } else { 4514 subptr(rsp, 64); 4515 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4516 subptr(rsp, 64); 4517 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4518 movdqu(xmm0, src); 4519 movdqu(xmm1, dst); 4520 Assembler::vpcmpeqw(xmm1, xmm1, xmm0, vector_len); 4521 movdqu(dst, xmm1); 4522 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4523 addptr(rsp, 64); 4524 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4525 addptr(rsp, 64); 4526 } 4527 } 4528 4529 void MacroAssembler::vpmovzxbw(XMMRegister dst, Address src, int vector_len) { 4530 int dst_enc = dst->encoding(); 4531 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4532 Assembler::vpmovzxbw(dst, src, vector_len); 4533 } else if (dst_enc < 16) { 4534 Assembler::vpmovzxbw(dst, src, vector_len); 4535 } else { 4536 subptr(rsp, 64); 4537 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4538 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4539 Assembler::vpmovzxbw(xmm0, src, vector_len); 4540 movdqu(dst, xmm0); 4541 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4542 addptr(rsp, 64); 4543 } 4544 } 4545 4546 void MacroAssembler::vpmovmskb(Register dst, XMMRegister src) { 4547 int src_enc = src->encoding(); 4548 if (src_enc < 16) { 4549 Assembler::vpmovmskb(dst, src); 4550 } else { 4551 subptr(rsp, 64); 4552 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4553 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4554 Assembler::vpmovmskb(dst, xmm0); 4555 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4556 addptr(rsp, 64); 4557 } 4558 } 4559 4560 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4561 int dst_enc = dst->encoding(); 4562 int nds_enc = nds->encoding(); 4563 int src_enc = src->encoding(); 4564 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4565 Assembler::vpmullw(dst, nds, src, vector_len); 4566 } else if ((dst_enc < 16) && (src_enc < 16)) { 4567 Assembler::vpmullw(dst, dst, src, vector_len); 4568 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4569 // use nds as scratch for src 4570 evmovdqul(nds, src, Assembler::AVX_512bit); 4571 Assembler::vpmullw(dst, dst, nds, vector_len); 4572 } else if ((src_enc < 16) && (nds_enc < 16)) { 4573 // use nds as scratch for dst 4574 evmovdqul(nds, dst, Assembler::AVX_512bit); 4575 Assembler::vpmullw(nds, nds, src, vector_len); 4576 evmovdqul(dst, nds, Assembler::AVX_512bit); 4577 } else if (dst_enc < 16) { 4578 // use nds as scatch for xmm0 to hold src 4579 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4580 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4581 Assembler::vpmullw(dst, dst, xmm0, vector_len); 4582 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4583 } else { 4584 // worse case scenario, all regs are in the upper bank 4585 subptr(rsp, 64); 4586 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4587 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4588 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4589 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4590 Assembler::vpmullw(xmm0, xmm0, xmm1, vector_len); 4591 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4592 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4593 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4594 addptr(rsp, 64); 4595 } 4596 } 4597 4598 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4599 int dst_enc = dst->encoding(); 4600 int nds_enc = nds->encoding(); 4601 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4602 Assembler::vpmullw(dst, nds, src, vector_len); 4603 } else if (dst_enc < 16) { 4604 Assembler::vpmullw(dst, dst, src, vector_len); 4605 } else if (nds_enc < 16) { 4606 // implies dst_enc in upper bank with src as scratch 4607 evmovdqul(nds, dst, Assembler::AVX_512bit); 4608 Assembler::vpmullw(nds, nds, src, vector_len); 4609 evmovdqul(dst, nds, Assembler::AVX_512bit); 4610 } else { 4611 // worse case scenario, all regs in upper bank 4612 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4613 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4614 Assembler::vpmullw(xmm0, xmm0, src, vector_len); 4615 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4616 } 4617 } 4618 4619 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4620 int dst_enc = dst->encoding(); 4621 int nds_enc = nds->encoding(); 4622 int src_enc = src->encoding(); 4623 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4624 Assembler::vpsubb(dst, nds, src, vector_len); 4625 } else if ((dst_enc < 16) && (src_enc < 16)) { 4626 Assembler::vpsubb(dst, dst, src, vector_len); 4627 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4628 // use nds as scratch for src 4629 evmovdqul(nds, src, Assembler::AVX_512bit); 4630 Assembler::vpsubb(dst, dst, nds, vector_len); 4631 } else if ((src_enc < 16) && (nds_enc < 16)) { 4632 // use nds as scratch for dst 4633 evmovdqul(nds, dst, Assembler::AVX_512bit); 4634 Assembler::vpsubb(nds, nds, src, vector_len); 4635 evmovdqul(dst, nds, Assembler::AVX_512bit); 4636 } else if (dst_enc < 16) { 4637 // use nds as scatch for xmm0 to hold src 4638 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4639 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4640 Assembler::vpsubb(dst, dst, xmm0, vector_len); 4641 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4642 } else { 4643 // worse case scenario, all regs are in the upper bank 4644 subptr(rsp, 64); 4645 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4646 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4647 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4648 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4649 Assembler::vpsubb(xmm0, xmm0, xmm1, vector_len); 4650 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4651 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4652 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4653 addptr(rsp, 64); 4654 } 4655 } 4656 4657 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4658 int dst_enc = dst->encoding(); 4659 int nds_enc = nds->encoding(); 4660 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4661 Assembler::vpsubb(dst, nds, src, vector_len); 4662 } else if (dst_enc < 16) { 4663 Assembler::vpsubb(dst, dst, src, vector_len); 4664 } else if (nds_enc < 16) { 4665 // implies dst_enc in upper bank with src as scratch 4666 evmovdqul(nds, dst, Assembler::AVX_512bit); 4667 Assembler::vpsubb(nds, nds, src, vector_len); 4668 evmovdqul(dst, nds, Assembler::AVX_512bit); 4669 } else { 4670 // worse case scenario, all regs in upper bank 4671 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4672 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4673 Assembler::vpsubw(xmm0, xmm0, src, vector_len); 4674 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4675 } 4676 } 4677 4678 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4679 int dst_enc = dst->encoding(); 4680 int nds_enc = nds->encoding(); 4681 int src_enc = src->encoding(); 4682 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4683 Assembler::vpsubw(dst, nds, src, vector_len); 4684 } else if ((dst_enc < 16) && (src_enc < 16)) { 4685 Assembler::vpsubw(dst, dst, src, vector_len); 4686 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4687 // use nds as scratch for src 4688 evmovdqul(nds, src, Assembler::AVX_512bit); 4689 Assembler::vpsubw(dst, dst, nds, vector_len); 4690 } else if ((src_enc < 16) && (nds_enc < 16)) { 4691 // use nds as scratch for dst 4692 evmovdqul(nds, dst, Assembler::AVX_512bit); 4693 Assembler::vpsubw(nds, nds, src, vector_len); 4694 evmovdqul(dst, nds, Assembler::AVX_512bit); 4695 } else if (dst_enc < 16) { 4696 // use nds as scatch for xmm0 to hold src 4697 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4698 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4699 Assembler::vpsubw(dst, dst, xmm0, vector_len); 4700 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4701 } else { 4702 // worse case scenario, all regs are in the upper bank 4703 subptr(rsp, 64); 4704 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4705 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4706 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4707 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4708 Assembler::vpsubw(xmm0, xmm0, xmm1, vector_len); 4709 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4710 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4711 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4712 addptr(rsp, 64); 4713 } 4714 } 4715 4716 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4717 int dst_enc = dst->encoding(); 4718 int nds_enc = nds->encoding(); 4719 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4720 Assembler::vpsubw(dst, nds, src, vector_len); 4721 } else if (dst_enc < 16) { 4722 Assembler::vpsubw(dst, dst, src, vector_len); 4723 } else if (nds_enc < 16) { 4724 // implies dst_enc in upper bank with src as scratch 4725 evmovdqul(nds, dst, Assembler::AVX_512bit); 4726 Assembler::vpsubw(nds, nds, src, vector_len); 4727 evmovdqul(dst, nds, Assembler::AVX_512bit); 4728 } else { 4729 // worse case scenario, all regs in upper bank 4730 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4731 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4732 Assembler::vpsubw(xmm0, xmm0, src, vector_len); 4733 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4734 } 4735 } 4736 4737 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4738 int dst_enc = dst->encoding(); 4739 int nds_enc = nds->encoding(); 4740 int shift_enc = shift->encoding(); 4741 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4742 Assembler::vpsraw(dst, nds, shift, vector_len); 4743 } else if ((dst_enc < 16) && (shift_enc < 16)) { 4744 Assembler::vpsraw(dst, dst, shift, vector_len); 4745 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4746 // use nds_enc as scratch with shift 4747 evmovdqul(nds, shift, Assembler::AVX_512bit); 4748 Assembler::vpsraw(dst, dst, nds, vector_len); 4749 } else if ((shift_enc < 16) && (nds_enc < 16)) { 4750 // use nds as scratch with dst 4751 evmovdqul(nds, dst, Assembler::AVX_512bit); 4752 Assembler::vpsraw(nds, nds, shift, vector_len); 4753 evmovdqul(dst, nds, Assembler::AVX_512bit); 4754 } else if (dst_enc < 16) { 4755 // use nds to save a copy of xmm0 and hold shift 4756 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4757 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4758 Assembler::vpsraw(dst, dst, xmm0, vector_len); 4759 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4760 } else if (nds_enc < 16) { 4761 // use nds as dest as temps 4762 evmovdqul(nds, dst, Assembler::AVX_512bit); 4763 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4764 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4765 Assembler::vpsraw(nds, nds, xmm0, vector_len); 4766 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4767 evmovdqul(dst, nds, Assembler::AVX_512bit); 4768 } else { 4769 // worse case scenario, all regs are in the upper bank 4770 subptr(rsp, 64); 4771 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4772 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4773 evmovdqul(xmm1, shift, Assembler::AVX_512bit); 4774 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4775 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len); 4776 evmovdqul(xmm1, dst, Assembler::AVX_512bit); 4777 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4778 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4779 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4780 addptr(rsp, 64); 4781 } 4782 } 4783 4784 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 4785 int dst_enc = dst->encoding(); 4786 int nds_enc = nds->encoding(); 4787 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4788 Assembler::vpsraw(dst, nds, shift, vector_len); 4789 } else if (dst_enc < 16) { 4790 Assembler::vpsraw(dst, dst, shift, vector_len); 4791 } else if (nds_enc < 16) { 4792 // use nds as scratch 4793 evmovdqul(nds, dst, Assembler::AVX_512bit); 4794 Assembler::vpsraw(nds, nds, shift, vector_len); 4795 evmovdqul(dst, nds, Assembler::AVX_512bit); 4796 } else { 4797 // use nds as scratch for xmm0 4798 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4799 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4800 Assembler::vpsraw(xmm0, xmm0, shift, vector_len); 4801 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4802 } 4803 } 4804 4805 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4806 int dst_enc = dst->encoding(); 4807 int nds_enc = nds->encoding(); 4808 int shift_enc = shift->encoding(); 4809 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4810 Assembler::vpsrlw(dst, nds, shift, vector_len); 4811 } else if ((dst_enc < 16) && (shift_enc < 16)) { 4812 Assembler::vpsrlw(dst, dst, shift, vector_len); 4813 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4814 // use nds_enc as scratch with shift 4815 evmovdqul(nds, shift, Assembler::AVX_512bit); 4816 Assembler::vpsrlw(dst, dst, nds, vector_len); 4817 } else if ((shift_enc < 16) && (nds_enc < 16)) { 4818 // use nds as scratch with dst 4819 evmovdqul(nds, dst, Assembler::AVX_512bit); 4820 Assembler::vpsrlw(nds, nds, shift, vector_len); 4821 evmovdqul(dst, nds, Assembler::AVX_512bit); 4822 } else if (dst_enc < 16) { 4823 // use nds to save a copy of xmm0 and hold shift 4824 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4825 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4826 Assembler::vpsrlw(dst, dst, xmm0, vector_len); 4827 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4828 } else if (nds_enc < 16) { 4829 // use nds as dest as temps 4830 evmovdqul(nds, dst, Assembler::AVX_512bit); 4831 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4832 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4833 Assembler::vpsrlw(nds, nds, xmm0, vector_len); 4834 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4835 evmovdqul(dst, nds, Assembler::AVX_512bit); 4836 } else { 4837 // worse case scenario, all regs are in the upper bank 4838 subptr(rsp, 64); 4839 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4840 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4841 evmovdqul(xmm1, shift, Assembler::AVX_512bit); 4842 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4843 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len); 4844 evmovdqul(xmm1, dst, Assembler::AVX_512bit); 4845 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4846 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4847 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4848 addptr(rsp, 64); 4849 } 4850 } 4851 4852 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 4853 int dst_enc = dst->encoding(); 4854 int nds_enc = nds->encoding(); 4855 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4856 Assembler::vpsrlw(dst, nds, shift, vector_len); 4857 } else if (dst_enc < 16) { 4858 Assembler::vpsrlw(dst, dst, shift, vector_len); 4859 } else if (nds_enc < 16) { 4860 // use nds as scratch 4861 evmovdqul(nds, dst, Assembler::AVX_512bit); 4862 Assembler::vpsrlw(nds, nds, shift, vector_len); 4863 evmovdqul(dst, nds, Assembler::AVX_512bit); 4864 } else { 4865 // use nds as scratch for xmm0 4866 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4867 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4868 Assembler::vpsrlw(xmm0, xmm0, shift, vector_len); 4869 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4870 } 4871 } 4872 4873 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4874 int dst_enc = dst->encoding(); 4875 int nds_enc = nds->encoding(); 4876 int shift_enc = shift->encoding(); 4877 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4878 Assembler::vpsllw(dst, nds, shift, vector_len); 4879 } else if ((dst_enc < 16) && (shift_enc < 16)) { 4880 Assembler::vpsllw(dst, dst, shift, vector_len); 4881 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4882 // use nds_enc as scratch with shift 4883 evmovdqul(nds, shift, Assembler::AVX_512bit); 4884 Assembler::vpsllw(dst, dst, nds, vector_len); 4885 } else if ((shift_enc < 16) && (nds_enc < 16)) { 4886 // use nds as scratch with dst 4887 evmovdqul(nds, dst, Assembler::AVX_512bit); 4888 Assembler::vpsllw(nds, nds, shift, vector_len); 4889 evmovdqul(dst, nds, Assembler::AVX_512bit); 4890 } else if (dst_enc < 16) { 4891 // use nds to save a copy of xmm0 and hold shift 4892 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4893 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4894 Assembler::vpsllw(dst, dst, xmm0, vector_len); 4895 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4896 } else if (nds_enc < 16) { 4897 // use nds as dest as temps 4898 evmovdqul(nds, dst, Assembler::AVX_512bit); 4899 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4900 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4901 Assembler::vpsllw(nds, nds, xmm0, vector_len); 4902 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4903 evmovdqul(dst, nds, Assembler::AVX_512bit); 4904 } else { 4905 // worse case scenario, all regs are in the upper bank 4906 subptr(rsp, 64); 4907 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4908 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4909 evmovdqul(xmm1, shift, Assembler::AVX_512bit); 4910 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4911 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len); 4912 evmovdqul(xmm1, dst, Assembler::AVX_512bit); 4913 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4914 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4915 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4916 addptr(rsp, 64); 4917 } 4918 } 4919 4920 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 4921 int dst_enc = dst->encoding(); 4922 int nds_enc = nds->encoding(); 4923 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4924 Assembler::vpsllw(dst, nds, shift, vector_len); 4925 } else if (dst_enc < 16) { 4926 Assembler::vpsllw(dst, dst, shift, vector_len); 4927 } else if (nds_enc < 16) { 4928 // use nds as scratch 4929 evmovdqul(nds, dst, Assembler::AVX_512bit); 4930 Assembler::vpsllw(nds, nds, shift, vector_len); 4931 evmovdqul(dst, nds, Assembler::AVX_512bit); 4932 } else { 4933 // use nds as scratch for xmm0 4934 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4935 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4936 Assembler::vpsllw(xmm0, xmm0, shift, vector_len); 4937 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4938 } 4939 } 4940 4941 void MacroAssembler::vptest(XMMRegister dst, XMMRegister src) { 4942 int dst_enc = dst->encoding(); 4943 int src_enc = src->encoding(); 4944 if ((dst_enc < 16) && (src_enc < 16)) { 4945 Assembler::vptest(dst, src); 4946 } else if (src_enc < 16) { 4947 subptr(rsp, 64); 4948 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4949 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4950 Assembler::vptest(xmm0, src); 4951 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4952 addptr(rsp, 64); 4953 } else if (dst_enc < 16) { 4954 subptr(rsp, 64); 4955 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4956 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4957 Assembler::vptest(dst, xmm0); 4958 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4959 addptr(rsp, 64); 4960 } else { 4961 subptr(rsp, 64); 4962 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4963 subptr(rsp, 64); 4964 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4965 movdqu(xmm0, src); 4966 movdqu(xmm1, dst); 4967 Assembler::vptest(xmm1, xmm0); 4968 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4969 addptr(rsp, 64); 4970 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4971 addptr(rsp, 64); 4972 } 4973 } 4974 4975 // This instruction exists within macros, ergo we cannot control its input 4976 // when emitted through those patterns. 4977 void MacroAssembler::punpcklbw(XMMRegister dst, XMMRegister src) { 4978 if (VM_Version::supports_avx512nobw()) { 4979 int dst_enc = dst->encoding(); 4980 int src_enc = src->encoding(); 4981 if (dst_enc == src_enc) { 4982 if (dst_enc < 16) { 4983 Assembler::punpcklbw(dst, src); 4984 } else { 4985 subptr(rsp, 64); 4986 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4987 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4988 Assembler::punpcklbw(xmm0, xmm0); 4989 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4990 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4991 addptr(rsp, 64); 4992 } 4993 } else { 4994 if ((src_enc < 16) && (dst_enc < 16)) { 4995 Assembler::punpcklbw(dst, src); 4996 } else if (src_enc < 16) { 4997 subptr(rsp, 64); 4998 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4999 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 5000 Assembler::punpcklbw(xmm0, src); 5001 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 5002 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5003 addptr(rsp, 64); 5004 } else if (dst_enc < 16) { 5005 subptr(rsp, 64); 5006 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5007 evmovdqul(xmm0, src, Assembler::AVX_512bit); 5008 Assembler::punpcklbw(dst, xmm0); 5009 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5010 addptr(rsp, 64); 5011 } else { 5012 subptr(rsp, 64); 5013 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5014 subptr(rsp, 64); 5015 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 5016 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 5017 evmovdqul(xmm1, src, Assembler::AVX_512bit); 5018 Assembler::punpcklbw(xmm0, xmm1); 5019 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 5020 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 5021 addptr(rsp, 64); 5022 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5023 addptr(rsp, 64); 5024 } 5025 } 5026 } else { 5027 Assembler::punpcklbw(dst, src); 5028 } 5029 } 5030 5031 void MacroAssembler::pshufd(XMMRegister dst, Address src, int mode) { 5032 if (VM_Version::supports_avx512vl()) { 5033 Assembler::pshufd(dst, src, mode); 5034 } else { 5035 int dst_enc = dst->encoding(); 5036 if (dst_enc < 16) { 5037 Assembler::pshufd(dst, src, mode); 5038 } else { 5039 subptr(rsp, 64); 5040 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5041 Assembler::pshufd(xmm0, src, mode); 5042 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 5043 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5044 addptr(rsp, 64); 5045 } 5046 } 5047 } 5048 5049 // This instruction exists within macros, ergo we cannot control its input 5050 // when emitted through those patterns. 5051 void MacroAssembler::pshuflw(XMMRegister dst, XMMRegister src, int mode) { 5052 if (VM_Version::supports_avx512nobw()) { 5053 int dst_enc = dst->encoding(); 5054 int src_enc = src->encoding(); 5055 if (dst_enc == src_enc) { 5056 if (dst_enc < 16) { 5057 Assembler::pshuflw(dst, src, mode); 5058 } else { 5059 subptr(rsp, 64); 5060 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5061 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 5062 Assembler::pshuflw(xmm0, xmm0, mode); 5063 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 5064 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5065 addptr(rsp, 64); 5066 } 5067 } else { 5068 if ((src_enc < 16) && (dst_enc < 16)) { 5069 Assembler::pshuflw(dst, src, mode); 5070 } else if (src_enc < 16) { 5071 subptr(rsp, 64); 5072 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5073 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 5074 Assembler::pshuflw(xmm0, src, mode); 5075 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 5076 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5077 addptr(rsp, 64); 5078 } else if (dst_enc < 16) { 5079 subptr(rsp, 64); 5080 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5081 evmovdqul(xmm0, src, Assembler::AVX_512bit); 5082 Assembler::pshuflw(dst, xmm0, mode); 5083 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5084 addptr(rsp, 64); 5085 } else { 5086 subptr(rsp, 64); 5087 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5088 subptr(rsp, 64); 5089 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 5090 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 5091 evmovdqul(xmm1, src, Assembler::AVX_512bit); 5092 Assembler::pshuflw(xmm0, xmm1, mode); 5093 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 5094 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 5095 addptr(rsp, 64); 5096 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5097 addptr(rsp, 64); 5098 } 5099 } 5100 } else { 5101 Assembler::pshuflw(dst, src, mode); 5102 } 5103 } 5104 5105 void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 5106 if (reachable(src)) { 5107 vandpd(dst, nds, as_Address(src), vector_len); 5108 } else { 5109 lea(rscratch1, src); 5110 vandpd(dst, nds, Address(rscratch1, 0), vector_len); 5111 } 5112 } 5113 5114 void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 5115 if (reachable(src)) { 5116 vandps(dst, nds, as_Address(src), vector_len); 5117 } else { 5118 lea(rscratch1, src); 5119 vandps(dst, nds, Address(rscratch1, 0), vector_len); 5120 } 5121 } 5122 5123 void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5124 if (reachable(src)) { 5125 vdivsd(dst, nds, as_Address(src)); 5126 } else { 5127 lea(rscratch1, src); 5128 vdivsd(dst, nds, Address(rscratch1, 0)); 5129 } 5130 } 5131 5132 void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5133 if (reachable(src)) { 5134 vdivss(dst, nds, as_Address(src)); 5135 } else { 5136 lea(rscratch1, src); 5137 vdivss(dst, nds, Address(rscratch1, 0)); 5138 } 5139 } 5140 5141 void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5142 if (reachable(src)) { 5143 vmulsd(dst, nds, as_Address(src)); 5144 } else { 5145 lea(rscratch1, src); 5146 vmulsd(dst, nds, Address(rscratch1, 0)); 5147 } 5148 } 5149 5150 void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5151 if (reachable(src)) { 5152 vmulss(dst, nds, as_Address(src)); 5153 } else { 5154 lea(rscratch1, src); 5155 vmulss(dst, nds, Address(rscratch1, 0)); 5156 } 5157 } 5158 5159 void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5160 if (reachable(src)) { 5161 vsubsd(dst, nds, as_Address(src)); 5162 } else { 5163 lea(rscratch1, src); 5164 vsubsd(dst, nds, Address(rscratch1, 0)); 5165 } 5166 } 5167 5168 void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5169 if (reachable(src)) { 5170 vsubss(dst, nds, as_Address(src)); 5171 } else { 5172 lea(rscratch1, src); 5173 vsubss(dst, nds, Address(rscratch1, 0)); 5174 } 5175 } 5176 5177 void MacroAssembler::vnegatess(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5178 int nds_enc = nds->encoding(); 5179 int dst_enc = dst->encoding(); 5180 bool dst_upper_bank = (dst_enc > 15); 5181 bool nds_upper_bank = (nds_enc > 15); 5182 if (VM_Version::supports_avx512novl() && 5183 (nds_upper_bank || dst_upper_bank)) { 5184 if (dst_upper_bank) { 5185 subptr(rsp, 64); 5186 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5187 movflt(xmm0, nds); 5188 vxorps(xmm0, xmm0, src, Assembler::AVX_128bit); 5189 movflt(dst, xmm0); 5190 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5191 addptr(rsp, 64); 5192 } else { 5193 movflt(dst, nds); 5194 vxorps(dst, dst, src, Assembler::AVX_128bit); 5195 } 5196 } else { 5197 vxorps(dst, nds, src, Assembler::AVX_128bit); 5198 } 5199 } 5200 5201 void MacroAssembler::vnegatesd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5202 int nds_enc = nds->encoding(); 5203 int dst_enc = dst->encoding(); 5204 bool dst_upper_bank = (dst_enc > 15); 5205 bool nds_upper_bank = (nds_enc > 15); 5206 if (VM_Version::supports_avx512novl() && 5207 (nds_upper_bank || dst_upper_bank)) { 5208 if (dst_upper_bank) { 5209 subptr(rsp, 64); 5210 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5211 movdbl(xmm0, nds); 5212 vxorpd(xmm0, xmm0, src, Assembler::AVX_128bit); 5213 movdbl(dst, xmm0); 5214 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5215 addptr(rsp, 64); 5216 } else { 5217 movdbl(dst, nds); 5218 vxorpd(dst, dst, src, Assembler::AVX_128bit); 5219 } 5220 } else { 5221 vxorpd(dst, nds, src, Assembler::AVX_128bit); 5222 } 5223 } 5224 5225 void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 5226 if (reachable(src)) { 5227 vxorpd(dst, nds, as_Address(src), vector_len); 5228 } else { 5229 lea(rscratch1, src); 5230 vxorpd(dst, nds, Address(rscratch1, 0), vector_len); 5231 } 5232 } 5233 5234 void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 5235 if (reachable(src)) { 5236 vxorps(dst, nds, as_Address(src), vector_len); 5237 } else { 5238 lea(rscratch1, src); 5239 vxorps(dst, nds, Address(rscratch1, 0), vector_len); 5240 } 5241 } 5242 5243 5244 void MacroAssembler::resolve_jobject(Register value, 5245 Register thread, 5246 Register tmp) { 5247 assert_different_registers(value, thread, tmp); 5248 Label done, not_weak; 5249 testptr(value, value); 5250 jcc(Assembler::zero, done); // Use NULL as-is. 5251 testptr(value, JNIHandles::weak_tag_mask); // Test for jweak tag. 5252 jcc(Assembler::zero, not_weak); 5253 // Resolve jweak. 5254 #if INCLUDE_ALL_GCS 5255 if (UseLoadBarrier) { 5256 load_barrier(value, Address(value, -JNIHandles::weak_tag_value), false /* expand call */, LoadBarrierOnPhantomOopRef); 5257 } else 5258 #endif 5259 { 5260 movptr(value, Address(value, -JNIHandles::weak_tag_value)); 5261 } 5262 verify_oop(value); 5263 #if INCLUDE_ALL_GCS 5264 if (UseG1GC) { 5265 g1_write_barrier_pre(noreg /* obj */, 5266 value /* pre_val */, 5267 thread /* thread */, 5268 tmp /* tmp */, 5269 true /* tosca_live */, 5270 true /* expand_call */); 5271 } 5272 #endif // INCLUDE_ALL_GCS 5273 jmp(done); 5274 bind(not_weak); 5275 // Resolve (untagged) jobject. 5276 movptr(value, Address(value, 0)); 5277 verify_oop(value); 5278 bind(done); 5279 } 5280 5281 void MacroAssembler::clear_jweak_tag(Register possibly_jweak) { 5282 const int32_t inverted_jweak_mask = ~static_cast<int32_t>(JNIHandles::weak_tag_mask); 5283 STATIC_ASSERT(inverted_jweak_mask == -2); // otherwise check this code 5284 // The inverted mask is sign-extended 5285 andptr(possibly_jweak, inverted_jweak_mask); 5286 } 5287 5288 ////////////////////////////////////////////////////////////////////////////////// 5289 #if INCLUDE_ALL_GCS 5290 5291 void MacroAssembler::g1_write_barrier_pre(Register obj, 5292 Register pre_val, 5293 Register thread, 5294 Register tmp, 5295 bool tosca_live, 5296 bool expand_call) { 5297 5298 // If expand_call is true then we expand the call_VM_leaf macro 5299 // directly to skip generating the check by 5300 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp. 5301 5302 #ifdef _LP64 5303 assert(thread == r15_thread, "must be"); 5304 #endif // _LP64 5305 5306 Label done; 5307 Label runtime; 5308 5309 assert(pre_val != noreg, "check this code"); 5310 5311 if (obj != noreg) { 5312 assert_different_registers(obj, pre_val, tmp); 5313 assert(pre_val != rax, "check this code"); 5314 } 5315 5316 Address in_progress(thread, in_bytes(JavaThread::satb_mark_queue_offset() + 5317 SATBMarkQueue::byte_offset_of_active())); 5318 Address index(thread, in_bytes(JavaThread::satb_mark_queue_offset() + 5319 SATBMarkQueue::byte_offset_of_index())); 5320 Address buffer(thread, in_bytes(JavaThread::satb_mark_queue_offset() + 5321 SATBMarkQueue::byte_offset_of_buf())); 5322 5323 5324 // Is marking active? 5325 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) { 5326 cmpl(in_progress, 0); 5327 } else { 5328 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption"); 5329 cmpb(in_progress, 0); 5330 } 5331 jcc(Assembler::equal, done); 5332 5333 // Do we need to load the previous value? 5334 if (obj != noreg) { 5335 load_heap_oop(pre_val, Address(obj, 0)); 5336 } 5337 5338 // Is the previous value null? 5339 cmpptr(pre_val, (int32_t) NULL_WORD); 5340 jcc(Assembler::equal, done); 5341 5342 // Can we store original value in the thread's buffer? 5343 // Is index == 0? 5344 // (The index field is typed as size_t.) 5345 5346 movptr(tmp, index); // tmp := *index_adr 5347 cmpptr(tmp, 0); // tmp == 0? 5348 jcc(Assembler::equal, runtime); // If yes, goto runtime 5349 5350 subptr(tmp, wordSize); // tmp := tmp - wordSize 5351 movptr(index, tmp); // *index_adr := tmp 5352 addptr(tmp, buffer); // tmp := tmp + *buffer_adr 5353 5354 // Record the previous value 5355 movptr(Address(tmp, 0), pre_val); 5356 jmp(done); 5357 5358 bind(runtime); 5359 // save the live input values 5360 if(tosca_live) push(rax); 5361 5362 if (obj != noreg && obj != rax) 5363 push(obj); 5364 5365 if (pre_val != rax) 5366 push(pre_val); 5367 5368 // Calling the runtime using the regular call_VM_leaf mechanism generates 5369 // code (generated by InterpreterMacroAssember::call_VM_leaf_base) 5370 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL. 5371 // 5372 // If we care generating the pre-barrier without a frame (e.g. in the 5373 // intrinsified Reference.get() routine) then ebp might be pointing to 5374 // the caller frame and so this check will most likely fail at runtime. 5375 // 5376 // Expanding the call directly bypasses the generation of the check. 5377 // So when we do not have have a full interpreter frame on the stack 5378 // expand_call should be passed true. 5379 5380 NOT_LP64( push(thread); ) 5381 5382 if (expand_call) { 5383 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); ) 5384 pass_arg1(this, thread); 5385 pass_arg0(this, pre_val); 5386 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), 2); 5387 } else { 5388 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), pre_val, thread); 5389 } 5390 5391 NOT_LP64( pop(thread); ) 5392 5393 // save the live input values 5394 if (pre_val != rax) 5395 pop(pre_val); 5396 5397 if (obj != noreg && obj != rax) 5398 pop(obj); 5399 5400 if(tosca_live) pop(rax); 5401 5402 bind(done); 5403 } 5404 5405 void MacroAssembler::g1_write_barrier_post(Register store_addr, 5406 Register new_val, 5407 Register thread, 5408 Register tmp, 5409 Register tmp2) { 5410 #ifdef _LP64 5411 assert(thread == r15_thread, "must be"); 5412 #endif // _LP64 5413 5414 Address queue_index(thread, in_bytes(JavaThread::dirty_card_queue_offset() + 5415 DirtyCardQueue::byte_offset_of_index())); 5416 Address buffer(thread, in_bytes(JavaThread::dirty_card_queue_offset() + 5417 DirtyCardQueue::byte_offset_of_buf())); 5418 5419 CardTableModRefBS* ctbs = 5420 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 5421 CardTable* ct = ctbs->card_table(); 5422 assert(sizeof(*ct->byte_map_base()) == sizeof(jbyte), "adjust this code"); 5423 5424 Label done; 5425 Label runtime; 5426 5427 // Does store cross heap regions? 5428 5429 movptr(tmp, store_addr); 5430 xorptr(tmp, new_val); 5431 shrptr(tmp, HeapRegion::LogOfHRGrainBytes); 5432 jcc(Assembler::equal, done); 5433 5434 // crosses regions, storing NULL? 5435 5436 cmpptr(new_val, (int32_t) NULL_WORD); 5437 jcc(Assembler::equal, done); 5438 5439 // storing region crossing non-NULL, is card already dirty? 5440 5441 const Register card_addr = tmp; 5442 const Register cardtable = tmp2; 5443 5444 movptr(card_addr, store_addr); 5445 shrptr(card_addr, CardTable::card_shift); 5446 // Do not use ExternalAddress to load 'byte_map_base', since 'byte_map_base' is NOT 5447 // a valid address and therefore is not properly handled by the relocation code. 5448 movptr(cardtable, (intptr_t)ct->byte_map_base()); 5449 addptr(card_addr, cardtable); 5450 5451 cmpb(Address(card_addr, 0), (int)G1CardTable::g1_young_card_val()); 5452 jcc(Assembler::equal, done); 5453 5454 membar(Assembler::Membar_mask_bits(Assembler::StoreLoad)); 5455 cmpb(Address(card_addr, 0), (int)CardTable::dirty_card_val()); 5456 jcc(Assembler::equal, done); 5457 5458 5459 // storing a region crossing, non-NULL oop, card is clean. 5460 // dirty card and log. 5461 5462 movb(Address(card_addr, 0), (int)CardTable::dirty_card_val()); 5463 5464 cmpl(queue_index, 0); 5465 jcc(Assembler::equal, runtime); 5466 subl(queue_index, wordSize); 5467 movptr(tmp2, buffer); 5468 #ifdef _LP64 5469 movslq(rscratch1, queue_index); 5470 addq(tmp2, rscratch1); 5471 movq(Address(tmp2, 0), card_addr); 5472 #else 5473 addl(tmp2, queue_index); 5474 movl(Address(tmp2, 0), card_addr); 5475 #endif 5476 jmp(done); 5477 5478 bind(runtime); 5479 // save the live input values 5480 push(store_addr); 5481 push(new_val); 5482 #ifdef _LP64 5483 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, r15_thread); 5484 #else 5485 push(thread); 5486 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, thread); 5487 pop(thread); 5488 #endif 5489 pop(new_val); 5490 pop(store_addr); 5491 5492 bind(done); 5493 } 5494 5495 #endif // INCLUDE_ALL_GCS 5496 ////////////////////////////////////////////////////////////////////////////////// 5497 5498 5499 void MacroAssembler::store_check(Register obj, Address dst) { 5500 store_check(obj); 5501 } 5502 5503 void MacroAssembler::store_check(Register obj) { 5504 // Does a store check for the oop in register obj. The content of 5505 // register obj is destroyed afterwards. 5506 BarrierSet* bs = Universe::heap()->barrier_set(); 5507 assert(bs->kind() == BarrierSet::CardTableModRef, 5508 "Wrong barrier set kind"); 5509 5510 CardTableModRefBS* ctbs = barrier_set_cast<CardTableModRefBS>(bs); 5511 CardTable* ct = ctbs->card_table(); 5512 assert(sizeof(*ct->byte_map_base()) == sizeof(jbyte), "adjust this code"); 5513 5514 shrptr(obj, CardTable::card_shift); 5515 5516 Address card_addr; 5517 5518 // The calculation for byte_map_base is as follows: 5519 // byte_map_base = _byte_map - (uintptr_t(low_bound) >> card_shift); 5520 // So this essentially converts an address to a displacement and it will 5521 // never need to be relocated. On 64bit however the value may be too 5522 // large for a 32bit displacement. 5523 intptr_t disp = (intptr_t) ct->byte_map_base(); 5524 if (is_simm32(disp)) { 5525 card_addr = Address(noreg, obj, Address::times_1, disp); 5526 } else { 5527 // By doing it as an ExternalAddress 'disp' could be converted to a rip-relative 5528 // displacement and done in a single instruction given favorable mapping and a 5529 // smarter version of as_Address. However, 'ExternalAddress' generates a relocation 5530 // entry and that entry is not properly handled by the relocation code. 5531 AddressLiteral cardtable((address)ct->byte_map_base(), relocInfo::none); 5532 Address index(noreg, obj, Address::times_1); 5533 card_addr = as_Address(ArrayAddress(cardtable, index)); 5534 } 5535 5536 int dirty = CardTable::dirty_card_val(); 5537 if (UseCondCardMark) { 5538 Label L_already_dirty; 5539 if (UseConcMarkSweepGC) { 5540 membar(Assembler::StoreLoad); 5541 } 5542 cmpb(card_addr, dirty); 5543 jcc(Assembler::equal, L_already_dirty); 5544 movb(card_addr, dirty); 5545 bind(L_already_dirty); 5546 } else { 5547 movb(card_addr, dirty); 5548 } 5549 } 5550 5551 void MacroAssembler::subptr(Register dst, int32_t imm32) { 5552 LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32)); 5553 } 5554 5555 // Force generation of a 4 byte immediate value even if it fits into 8bit 5556 void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) { 5557 LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32)); 5558 } 5559 5560 void MacroAssembler::subptr(Register dst, Register src) { 5561 LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src)); 5562 } 5563 5564 // C++ bool manipulation 5565 void MacroAssembler::testbool(Register dst) { 5566 if(sizeof(bool) == 1) 5567 testb(dst, 0xff); 5568 else if(sizeof(bool) == 2) { 5569 // testw implementation needed for two byte bools 5570 ShouldNotReachHere(); 5571 } else if(sizeof(bool) == 4) 5572 testl(dst, dst); 5573 else 5574 // unsupported 5575 ShouldNotReachHere(); 5576 } 5577 5578 void MacroAssembler::testptr(Register dst, Register src) { 5579 LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src)); 5580 } 5581 5582 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes. 5583 void MacroAssembler::tlab_allocate(Register obj, 5584 Register var_size_in_bytes, 5585 int con_size_in_bytes, 5586 Register t1, 5587 Register t2, 5588 Label& slow_case) { 5589 assert_different_registers(obj, t1, t2); 5590 assert_different_registers(obj, var_size_in_bytes, t1); 5591 Register end = t2; 5592 Register thread = NOT_LP64(t1) LP64_ONLY(r15_thread); 5593 5594 verify_tlab(); 5595 5596 NOT_LP64(get_thread(thread)); 5597 5598 movptr(obj, Address(thread, JavaThread::tlab_top_offset())); 5599 if (var_size_in_bytes == noreg) { 5600 lea(end, Address(obj, con_size_in_bytes)); 5601 } else { 5602 lea(end, Address(obj, var_size_in_bytes, Address::times_1)); 5603 } 5604 cmpptr(end, Address(thread, JavaThread::tlab_end_offset())); 5605 jcc(Assembler::above, slow_case); 5606 5607 // update the tlab top pointer 5608 movptr(Address(thread, JavaThread::tlab_top_offset()), end); 5609 5610 // recover var_size_in_bytes if necessary 5611 if (var_size_in_bytes == end) { 5612 subptr(var_size_in_bytes, obj); 5613 } 5614 verify_tlab(); 5615 } 5616 5617 // Preserves the contents of address, destroys the contents length_in_bytes and temp. 5618 void MacroAssembler::zero_memory(Register address, Register length_in_bytes, int offset_in_bytes, Register temp) { 5619 assert(address != length_in_bytes && address != temp && temp != length_in_bytes, "registers must be different"); 5620 assert((offset_in_bytes & (BytesPerWord - 1)) == 0, "offset must be a multiple of BytesPerWord"); 5621 Label done; 5622 5623 testptr(length_in_bytes, length_in_bytes); 5624 jcc(Assembler::zero, done); 5625 5626 // initialize topmost word, divide index by 2, check if odd and test if zero 5627 // note: for the remaining code to work, index must be a multiple of BytesPerWord 5628 #ifdef ASSERT 5629 { 5630 Label L; 5631 testptr(length_in_bytes, BytesPerWord - 1); 5632 jcc(Assembler::zero, L); 5633 stop("length must be a multiple of BytesPerWord"); 5634 bind(L); 5635 } 5636 #endif 5637 Register index = length_in_bytes; 5638 xorptr(temp, temp); // use _zero reg to clear memory (shorter code) 5639 if (UseIncDec) { 5640 shrptr(index, 3); // divide by 8/16 and set carry flag if bit 2 was set 5641 } else { 5642 shrptr(index, 2); // use 2 instructions to avoid partial flag stall 5643 shrptr(index, 1); 5644 } 5645 #ifndef _LP64 5646 // index could have not been a multiple of 8 (i.e., bit 2 was set) 5647 { 5648 Label even; 5649 // note: if index was a multiple of 8, then it cannot 5650 // be 0 now otherwise it must have been 0 before 5651 // => if it is even, we don't need to check for 0 again 5652 jcc(Assembler::carryClear, even); 5653 // clear topmost word (no jump would be needed if conditional assignment worked here) 5654 movptr(Address(address, index, Address::times_8, offset_in_bytes - 0*BytesPerWord), temp); 5655 // index could be 0 now, must check again 5656 jcc(Assembler::zero, done); 5657 bind(even); 5658 } 5659 #endif // !_LP64 5660 // initialize remaining object fields: index is a multiple of 2 now 5661 { 5662 Label loop; 5663 bind(loop); 5664 movptr(Address(address, index, Address::times_8, offset_in_bytes - 1*BytesPerWord), temp); 5665 NOT_LP64(movptr(Address(address, index, Address::times_8, offset_in_bytes - 2*BytesPerWord), temp);) 5666 decrement(index); 5667 jcc(Assembler::notZero, loop); 5668 } 5669 5670 bind(done); 5671 } 5672 5673 void MacroAssembler::incr_allocated_bytes(Register thread, 5674 Register var_size_in_bytes, 5675 int con_size_in_bytes, 5676 Register t1) { 5677 if (!thread->is_valid()) { 5678 #ifdef _LP64 5679 thread = r15_thread; 5680 #else 5681 assert(t1->is_valid(), "need temp reg"); 5682 thread = t1; 5683 get_thread(thread); 5684 #endif 5685 } 5686 5687 #ifdef _LP64 5688 if (var_size_in_bytes->is_valid()) { 5689 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes); 5690 } else { 5691 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes); 5692 } 5693 #else 5694 if (var_size_in_bytes->is_valid()) { 5695 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes); 5696 } else { 5697 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes); 5698 } 5699 adcl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())+4), 0); 5700 #endif 5701 } 5702 5703 // Look up the method for a megamorphic invokeinterface call. 5704 // The target method is determined by <intf_klass, itable_index>. 5705 // The receiver klass is in recv_klass. 5706 // On success, the result will be in method_result, and execution falls through. 5707 // On failure, execution transfers to the given label. 5708 void MacroAssembler::lookup_interface_method(Register recv_klass, 5709 Register intf_klass, 5710 RegisterOrConstant itable_index, 5711 Register method_result, 5712 Register scan_temp, 5713 Label& L_no_such_interface, 5714 bool return_method) { 5715 assert_different_registers(recv_klass, intf_klass, scan_temp); 5716 assert_different_registers(method_result, intf_klass, scan_temp); 5717 assert(recv_klass != method_result || !return_method, 5718 "recv_klass can be destroyed when method isn't needed"); 5719 5720 assert(itable_index.is_constant() || itable_index.as_register() == method_result, 5721 "caller must use same register for non-constant itable index as for method"); 5722 5723 // Compute start of first itableOffsetEntry (which is at the end of the vtable) 5724 int vtable_base = in_bytes(Klass::vtable_start_offset()); 5725 int itentry_off = itableMethodEntry::method_offset_in_bytes(); 5726 int scan_step = itableOffsetEntry::size() * wordSize; 5727 int vte_size = vtableEntry::size_in_bytes(); 5728 Address::ScaleFactor times_vte_scale = Address::times_ptr; 5729 assert(vte_size == wordSize, "else adjust times_vte_scale"); 5730 5731 movl(scan_temp, Address(recv_klass, Klass::vtable_length_offset())); 5732 5733 // %%% Could store the aligned, prescaled offset in the klassoop. 5734 lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base)); 5735 5736 if (return_method) { 5737 // Adjust recv_klass by scaled itable_index, so we can free itable_index. 5738 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below"); 5739 lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off)); 5740 } 5741 5742 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) { 5743 // if (scan->interface() == intf) { 5744 // result = (klass + scan->offset() + itable_index); 5745 // } 5746 // } 5747 Label search, found_method; 5748 5749 for (int peel = 1; peel >= 0; peel--) { 5750 movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes())); 5751 cmpptr(intf_klass, method_result); 5752 5753 if (peel) { 5754 jccb(Assembler::equal, found_method); 5755 } else { 5756 jccb(Assembler::notEqual, search); 5757 // (invert the test to fall through to found_method...) 5758 } 5759 5760 if (!peel) break; 5761 5762 bind(search); 5763 5764 // Check that the previous entry is non-null. A null entry means that 5765 // the receiver class doesn't implement the interface, and wasn't the 5766 // same as when the caller was compiled. 5767 testptr(method_result, method_result); 5768 jcc(Assembler::zero, L_no_such_interface); 5769 addptr(scan_temp, scan_step); 5770 } 5771 5772 bind(found_method); 5773 5774 if (return_method) { 5775 // Got a hit. 5776 movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes())); 5777 movptr(method_result, Address(recv_klass, scan_temp, Address::times_1)); 5778 } 5779 } 5780 5781 5782 // virtual method calling 5783 void MacroAssembler::lookup_virtual_method(Register recv_klass, 5784 RegisterOrConstant vtable_index, 5785 Register method_result) { 5786 const int base = in_bytes(Klass::vtable_start_offset()); 5787 assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below"); 5788 Address vtable_entry_addr(recv_klass, 5789 vtable_index, Address::times_ptr, 5790 base + vtableEntry::method_offset_in_bytes()); 5791 movptr(method_result, vtable_entry_addr); 5792 } 5793 5794 5795 void MacroAssembler::check_klass_subtype(Register sub_klass, 5796 Register super_klass, 5797 Register temp_reg, 5798 Label& L_success) { 5799 Label L_failure; 5800 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL); 5801 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL); 5802 bind(L_failure); 5803 } 5804 5805 5806 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass, 5807 Register super_klass, 5808 Register temp_reg, 5809 Label* L_success, 5810 Label* L_failure, 5811 Label* L_slow_path, 5812 RegisterOrConstant super_check_offset) { 5813 assert_different_registers(sub_klass, super_klass, temp_reg); 5814 bool must_load_sco = (super_check_offset.constant_or_zero() == -1); 5815 if (super_check_offset.is_register()) { 5816 assert_different_registers(sub_klass, super_klass, 5817 super_check_offset.as_register()); 5818 } else if (must_load_sco) { 5819 assert(temp_reg != noreg, "supply either a temp or a register offset"); 5820 } 5821 5822 Label L_fallthrough; 5823 int label_nulls = 0; 5824 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 5825 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 5826 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; } 5827 assert(label_nulls <= 1, "at most one NULL in the batch"); 5828 5829 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 5830 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 5831 Address super_check_offset_addr(super_klass, sco_offset); 5832 5833 // Hacked jcc, which "knows" that L_fallthrough, at least, is in 5834 // range of a jccb. If this routine grows larger, reconsider at 5835 // least some of these. 5836 #define local_jcc(assembler_cond, label) \ 5837 if (&(label) == &L_fallthrough) jccb(assembler_cond, label); \ 5838 else jcc( assembler_cond, label) /*omit semi*/ 5839 5840 // Hacked jmp, which may only be used just before L_fallthrough. 5841 #define final_jmp(label) \ 5842 if (&(label) == &L_fallthrough) { /*do nothing*/ } \ 5843 else jmp(label) /*omit semi*/ 5844 5845 // If the pointers are equal, we are done (e.g., String[] elements). 5846 // This self-check enables sharing of secondary supertype arrays among 5847 // non-primary types such as array-of-interface. Otherwise, each such 5848 // type would need its own customized SSA. 5849 // We move this check to the front of the fast path because many 5850 // type checks are in fact trivially successful in this manner, 5851 // so we get a nicely predicted branch right at the start of the check. 5852 cmpptr(sub_klass, super_klass); 5853 local_jcc(Assembler::equal, *L_success); 5854 5855 // Check the supertype display: 5856 if (must_load_sco) { 5857 // Positive movl does right thing on LP64. 5858 movl(temp_reg, super_check_offset_addr); 5859 super_check_offset = RegisterOrConstant(temp_reg); 5860 } 5861 Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0); 5862 cmpptr(super_klass, super_check_addr); // load displayed supertype 5863 5864 // This check has worked decisively for primary supers. 5865 // Secondary supers are sought in the super_cache ('super_cache_addr'). 5866 // (Secondary supers are interfaces and very deeply nested subtypes.) 5867 // This works in the same check above because of a tricky aliasing 5868 // between the super_cache and the primary super display elements. 5869 // (The 'super_check_addr' can address either, as the case requires.) 5870 // Note that the cache is updated below if it does not help us find 5871 // what we need immediately. 5872 // So if it was a primary super, we can just fail immediately. 5873 // Otherwise, it's the slow path for us (no success at this point). 5874 5875 if (super_check_offset.is_register()) { 5876 local_jcc(Assembler::equal, *L_success); 5877 cmpl(super_check_offset.as_register(), sc_offset); 5878 if (L_failure == &L_fallthrough) { 5879 local_jcc(Assembler::equal, *L_slow_path); 5880 } else { 5881 local_jcc(Assembler::notEqual, *L_failure); 5882 final_jmp(*L_slow_path); 5883 } 5884 } else if (super_check_offset.as_constant() == sc_offset) { 5885 // Need a slow path; fast failure is impossible. 5886 if (L_slow_path == &L_fallthrough) { 5887 local_jcc(Assembler::equal, *L_success); 5888 } else { 5889 local_jcc(Assembler::notEqual, *L_slow_path); 5890 final_jmp(*L_success); 5891 } 5892 } else { 5893 // No slow path; it's a fast decision. 5894 if (L_failure == &L_fallthrough) { 5895 local_jcc(Assembler::equal, *L_success); 5896 } else { 5897 local_jcc(Assembler::notEqual, *L_failure); 5898 final_jmp(*L_success); 5899 } 5900 } 5901 5902 bind(L_fallthrough); 5903 5904 #undef local_jcc 5905 #undef final_jmp 5906 } 5907 5908 5909 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass, 5910 Register super_klass, 5911 Register temp_reg, 5912 Register temp2_reg, 5913 Label* L_success, 5914 Label* L_failure, 5915 bool set_cond_codes) { 5916 assert_different_registers(sub_klass, super_klass, temp_reg); 5917 if (temp2_reg != noreg) 5918 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg); 5919 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg) 5920 5921 Label L_fallthrough; 5922 int label_nulls = 0; 5923 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 5924 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 5925 assert(label_nulls <= 1, "at most one NULL in the batch"); 5926 5927 // a couple of useful fields in sub_klass: 5928 int ss_offset = in_bytes(Klass::secondary_supers_offset()); 5929 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 5930 Address secondary_supers_addr(sub_klass, ss_offset); 5931 Address super_cache_addr( sub_klass, sc_offset); 5932 5933 // Do a linear scan of the secondary super-klass chain. 5934 // This code is rarely used, so simplicity is a virtue here. 5935 // The repne_scan instruction uses fixed registers, which we must spill. 5936 // Don't worry too much about pre-existing connections with the input regs. 5937 5938 assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super) 5939 assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter) 5940 5941 // Get super_klass value into rax (even if it was in rdi or rcx). 5942 bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false; 5943 if (super_klass != rax || UseCompressedOops) { 5944 if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; } 5945 mov(rax, super_klass); 5946 } 5947 if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; } 5948 if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; } 5949 5950 #ifndef PRODUCT 5951 int* pst_counter = &SharedRuntime::_partial_subtype_ctr; 5952 ExternalAddress pst_counter_addr((address) pst_counter); 5953 NOT_LP64( incrementl(pst_counter_addr) ); 5954 LP64_ONLY( lea(rcx, pst_counter_addr) ); 5955 LP64_ONLY( incrementl(Address(rcx, 0)) ); 5956 #endif //PRODUCT 5957 5958 // We will consult the secondary-super array. 5959 movptr(rdi, secondary_supers_addr); 5960 // Load the array length. (Positive movl does right thing on LP64.) 5961 movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes())); 5962 // Skip to start of data. 5963 addptr(rdi, Array<Klass*>::base_offset_in_bytes()); 5964 5965 // Scan RCX words at [RDI] for an occurrence of RAX. 5966 // Set NZ/Z based on last compare. 5967 // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does 5968 // not change flags (only scas instruction which is repeated sets flags). 5969 // Set Z = 0 (not equal) before 'repne' to indicate that class was not found. 5970 5971 testptr(rax,rax); // Set Z = 0 5972 repne_scan(); 5973 5974 // Unspill the temp. registers: 5975 if (pushed_rdi) pop(rdi); 5976 if (pushed_rcx) pop(rcx); 5977 if (pushed_rax) pop(rax); 5978 5979 if (set_cond_codes) { 5980 // Special hack for the AD files: rdi is guaranteed non-zero. 5981 assert(!pushed_rdi, "rdi must be left non-NULL"); 5982 // Also, the condition codes are properly set Z/NZ on succeed/failure. 5983 } 5984 5985 if (L_failure == &L_fallthrough) 5986 jccb(Assembler::notEqual, *L_failure); 5987 else jcc(Assembler::notEqual, *L_failure); 5988 5989 // Success. Cache the super we found and proceed in triumph. 5990 movptr(super_cache_addr, super_klass); 5991 5992 if (L_success != &L_fallthrough) { 5993 jmp(*L_success); 5994 } 5995 5996 #undef IS_A_TEMP 5997 5998 bind(L_fallthrough); 5999 } 6000 6001 6002 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) { 6003 if (VM_Version::supports_cmov()) { 6004 cmovl(cc, dst, src); 6005 } else { 6006 Label L; 6007 jccb(negate_condition(cc), L); 6008 movl(dst, src); 6009 bind(L); 6010 } 6011 } 6012 6013 void MacroAssembler::cmov32(Condition cc, Register dst, Register src) { 6014 if (VM_Version::supports_cmov()) { 6015 cmovl(cc, dst, src); 6016 } else { 6017 Label L; 6018 jccb(negate_condition(cc), L); 6019 movl(dst, src); 6020 bind(L); 6021 } 6022 } 6023 6024 void MacroAssembler::verify_oop(Register reg, const char* s) { 6025 if (!VerifyOops) return; 6026 6027 // Pass register number to verify_oop_subroutine 6028 const char* b = NULL; 6029 { 6030 ResourceMark rm; 6031 stringStream ss; 6032 ss.print("verify_oop: %s: %s", reg->name(), s); 6033 b = code_string(ss.as_string()); 6034 } 6035 BLOCK_COMMENT("verify_oop {"); 6036 #ifdef _LP64 6037 push(rscratch1); // save r10, trashed by movptr() 6038 #endif 6039 push(rax); // save rax, 6040 push(reg); // pass register argument 6041 ExternalAddress buffer((address) b); 6042 // avoid using pushptr, as it modifies scratch registers 6043 // and our contract is not to modify anything 6044 movptr(rax, buffer.addr()); 6045 push(rax); 6046 // call indirectly to solve generation ordering problem 6047 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); 6048 call(rax); 6049 // Caller pops the arguments (oop, message) and restores rax, r10 6050 BLOCK_COMMENT("} verify_oop"); 6051 } 6052 6053 6054 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr, 6055 Register tmp, 6056 int offset) { 6057 intptr_t value = *delayed_value_addr; 6058 if (value != 0) 6059 return RegisterOrConstant(value + offset); 6060 6061 // load indirectly to solve generation ordering problem 6062 movptr(tmp, ExternalAddress((address) delayed_value_addr)); 6063 6064 #ifdef ASSERT 6065 { Label L; 6066 testptr(tmp, tmp); 6067 if (WizardMode) { 6068 const char* buf = NULL; 6069 { 6070 ResourceMark rm; 6071 stringStream ss; 6072 ss.print("DelayedValue=" INTPTR_FORMAT, delayed_value_addr[1]); 6073 buf = code_string(ss.as_string()); 6074 } 6075 jcc(Assembler::notZero, L); 6076 STOP(buf); 6077 } else { 6078 jccb(Assembler::notZero, L); 6079 hlt(); 6080 } 6081 bind(L); 6082 } 6083 #endif 6084 6085 if (offset != 0) 6086 addptr(tmp, offset); 6087 6088 return RegisterOrConstant(tmp); 6089 } 6090 6091 6092 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot, 6093 int extra_slot_offset) { 6094 // cf. TemplateTable::prepare_invoke(), if (load_receiver). 6095 int stackElementSize = Interpreter::stackElementSize; 6096 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0); 6097 #ifdef ASSERT 6098 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1); 6099 assert(offset1 - offset == stackElementSize, "correct arithmetic"); 6100 #endif 6101 Register scale_reg = noreg; 6102 Address::ScaleFactor scale_factor = Address::no_scale; 6103 if (arg_slot.is_constant()) { 6104 offset += arg_slot.as_constant() * stackElementSize; 6105 } else { 6106 scale_reg = arg_slot.as_register(); 6107 scale_factor = Address::times(stackElementSize); 6108 } 6109 offset += wordSize; // return PC is on stack 6110 return Address(rsp, scale_reg, scale_factor, offset); 6111 } 6112 6113 6114 void MacroAssembler::verify_oop_addr(Address addr, const char* s) { 6115 if (!VerifyOops) return; 6116 6117 // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord); 6118 // Pass register number to verify_oop_subroutine 6119 const char* b = NULL; 6120 { 6121 ResourceMark rm; 6122 stringStream ss; 6123 ss.print("verify_oop_addr: %s", s); 6124 b = code_string(ss.as_string()); 6125 } 6126 #ifdef _LP64 6127 push(rscratch1); // save r10, trashed by movptr() 6128 #endif 6129 push(rax); // save rax, 6130 // addr may contain rsp so we will have to adjust it based on the push 6131 // we just did (and on 64 bit we do two pushes) 6132 // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which 6133 // stores rax into addr which is backwards of what was intended. 6134 if (addr.uses(rsp)) { 6135 lea(rax, addr); 6136 pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord)); 6137 } else { 6138 pushptr(addr); 6139 } 6140 6141 ExternalAddress buffer((address) b); 6142 // pass msg argument 6143 // avoid using pushptr, as it modifies scratch registers 6144 // and our contract is not to modify anything 6145 movptr(rax, buffer.addr()); 6146 push(rax); 6147 6148 // call indirectly to solve generation ordering problem 6149 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); 6150 call(rax); 6151 // Caller pops the arguments (addr, message) and restores rax, r10. 6152 } 6153 6154 void MacroAssembler::verify_tlab() { 6155 #ifdef ASSERT 6156 if (UseTLAB && VerifyOops) { 6157 Label next, ok; 6158 Register t1 = rsi; 6159 Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread); 6160 6161 push(t1); 6162 NOT_LP64(push(thread_reg)); 6163 NOT_LP64(get_thread(thread_reg)); 6164 6165 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 6166 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset()))); 6167 jcc(Assembler::aboveEqual, next); 6168 STOP("assert(top >= start)"); 6169 should_not_reach_here(); 6170 6171 bind(next); 6172 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset()))); 6173 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 6174 jcc(Assembler::aboveEqual, ok); 6175 STOP("assert(top <= end)"); 6176 should_not_reach_here(); 6177 6178 bind(ok); 6179 NOT_LP64(pop(thread_reg)); 6180 pop(t1); 6181 } 6182 #endif 6183 } 6184 6185 class ControlWord { 6186 public: 6187 int32_t _value; 6188 6189 int rounding_control() const { return (_value >> 10) & 3 ; } 6190 int precision_control() const { return (_value >> 8) & 3 ; } 6191 bool precision() const { return ((_value >> 5) & 1) != 0; } 6192 bool underflow() const { return ((_value >> 4) & 1) != 0; } 6193 bool overflow() const { return ((_value >> 3) & 1) != 0; } 6194 bool zero_divide() const { return ((_value >> 2) & 1) != 0; } 6195 bool denormalized() const { return ((_value >> 1) & 1) != 0; } 6196 bool invalid() const { return ((_value >> 0) & 1) != 0; } 6197 6198 void print() const { 6199 // rounding control 6200 const char* rc; 6201 switch (rounding_control()) { 6202 case 0: rc = "round near"; break; 6203 case 1: rc = "round down"; break; 6204 case 2: rc = "round up "; break; 6205 case 3: rc = "chop "; break; 6206 }; 6207 // precision control 6208 const char* pc; 6209 switch (precision_control()) { 6210 case 0: pc = "24 bits "; break; 6211 case 1: pc = "reserved"; break; 6212 case 2: pc = "53 bits "; break; 6213 case 3: pc = "64 bits "; break; 6214 }; 6215 // flags 6216 char f[9]; 6217 f[0] = ' '; 6218 f[1] = ' '; 6219 f[2] = (precision ()) ? 'P' : 'p'; 6220 f[3] = (underflow ()) ? 'U' : 'u'; 6221 f[4] = (overflow ()) ? 'O' : 'o'; 6222 f[5] = (zero_divide ()) ? 'Z' : 'z'; 6223 f[6] = (denormalized()) ? 'D' : 'd'; 6224 f[7] = (invalid ()) ? 'I' : 'i'; 6225 f[8] = '\x0'; 6226 // output 6227 printf("%04x masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc); 6228 } 6229 6230 }; 6231 6232 class StatusWord { 6233 public: 6234 int32_t _value; 6235 6236 bool busy() const { return ((_value >> 15) & 1) != 0; } 6237 bool C3() const { return ((_value >> 14) & 1) != 0; } 6238 bool C2() const { return ((_value >> 10) & 1) != 0; } 6239 bool C1() const { return ((_value >> 9) & 1) != 0; } 6240 bool C0() const { return ((_value >> 8) & 1) != 0; } 6241 int top() const { return (_value >> 11) & 7 ; } 6242 bool error_status() const { return ((_value >> 7) & 1) != 0; } 6243 bool stack_fault() const { return ((_value >> 6) & 1) != 0; } 6244 bool precision() const { return ((_value >> 5) & 1) != 0; } 6245 bool underflow() const { return ((_value >> 4) & 1) != 0; } 6246 bool overflow() const { return ((_value >> 3) & 1) != 0; } 6247 bool zero_divide() const { return ((_value >> 2) & 1) != 0; } 6248 bool denormalized() const { return ((_value >> 1) & 1) != 0; } 6249 bool invalid() const { return ((_value >> 0) & 1) != 0; } 6250 6251 void print() const { 6252 // condition codes 6253 char c[5]; 6254 c[0] = (C3()) ? '3' : '-'; 6255 c[1] = (C2()) ? '2' : '-'; 6256 c[2] = (C1()) ? '1' : '-'; 6257 c[3] = (C0()) ? '0' : '-'; 6258 c[4] = '\x0'; 6259 // flags 6260 char f[9]; 6261 f[0] = (error_status()) ? 'E' : '-'; 6262 f[1] = (stack_fault ()) ? 'S' : '-'; 6263 f[2] = (precision ()) ? 'P' : '-'; 6264 f[3] = (underflow ()) ? 'U' : '-'; 6265 f[4] = (overflow ()) ? 'O' : '-'; 6266 f[5] = (zero_divide ()) ? 'Z' : '-'; 6267 f[6] = (denormalized()) ? 'D' : '-'; 6268 f[7] = (invalid ()) ? 'I' : '-'; 6269 f[8] = '\x0'; 6270 // output 6271 printf("%04x flags = %s, cc = %s, top = %d", _value & 0xFFFF, f, c, top()); 6272 } 6273 6274 }; 6275 6276 class TagWord { 6277 public: 6278 int32_t _value; 6279 6280 int tag_at(int i) const { return (_value >> (i*2)) & 3; } 6281 6282 void print() const { 6283 printf("%04x", _value & 0xFFFF); 6284 } 6285 6286 }; 6287 6288 class FPU_Register { 6289 public: 6290 int32_t _m0; 6291 int32_t _m1; 6292 int16_t _ex; 6293 6294 bool is_indefinite() const { 6295 return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0; 6296 } 6297 6298 void print() const { 6299 char sign = (_ex < 0) ? '-' : '+'; 6300 const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : " "; 6301 printf("%c%04hx.%08x%08x %s", sign, _ex, _m1, _m0, kind); 6302 }; 6303 6304 }; 6305 6306 class FPU_State { 6307 public: 6308 enum { 6309 register_size = 10, 6310 number_of_registers = 8, 6311 register_mask = 7 6312 }; 6313 6314 ControlWord _control_word; 6315 StatusWord _status_word; 6316 TagWord _tag_word; 6317 int32_t _error_offset; 6318 int32_t _error_selector; 6319 int32_t _data_offset; 6320 int32_t _data_selector; 6321 int8_t _register[register_size * number_of_registers]; 6322 6323 int tag_for_st(int i) const { return _tag_word.tag_at((_status_word.top() + i) & register_mask); } 6324 FPU_Register* st(int i) const { return (FPU_Register*)&_register[register_size * i]; } 6325 6326 const char* tag_as_string(int tag) const { 6327 switch (tag) { 6328 case 0: return "valid"; 6329 case 1: return "zero"; 6330 case 2: return "special"; 6331 case 3: return "empty"; 6332 } 6333 ShouldNotReachHere(); 6334 return NULL; 6335 } 6336 6337 void print() const { 6338 // print computation registers 6339 { int t = _status_word.top(); 6340 for (int i = 0; i < number_of_registers; i++) { 6341 int j = (i - t) & register_mask; 6342 printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j); 6343 st(j)->print(); 6344 printf(" %s\n", tag_as_string(_tag_word.tag_at(i))); 6345 } 6346 } 6347 printf("\n"); 6348 // print control registers 6349 printf("ctrl = "); _control_word.print(); printf("\n"); 6350 printf("stat = "); _status_word .print(); printf("\n"); 6351 printf("tags = "); _tag_word .print(); printf("\n"); 6352 } 6353 6354 }; 6355 6356 class Flag_Register { 6357 public: 6358 int32_t _value; 6359 6360 bool overflow() const { return ((_value >> 11) & 1) != 0; } 6361 bool direction() const { return ((_value >> 10) & 1) != 0; } 6362 bool sign() const { return ((_value >> 7) & 1) != 0; } 6363 bool zero() const { return ((_value >> 6) & 1) != 0; } 6364 bool auxiliary_carry() const { return ((_value >> 4) & 1) != 0; } 6365 bool parity() const { return ((_value >> 2) & 1) != 0; } 6366 bool carry() const { return ((_value >> 0) & 1) != 0; } 6367 6368 void print() const { 6369 // flags 6370 char f[8]; 6371 f[0] = (overflow ()) ? 'O' : '-'; 6372 f[1] = (direction ()) ? 'D' : '-'; 6373 f[2] = (sign ()) ? 'S' : '-'; 6374 f[3] = (zero ()) ? 'Z' : '-'; 6375 f[4] = (auxiliary_carry()) ? 'A' : '-'; 6376 f[5] = (parity ()) ? 'P' : '-'; 6377 f[6] = (carry ()) ? 'C' : '-'; 6378 f[7] = '\x0'; 6379 // output 6380 printf("%08x flags = %s", _value, f); 6381 } 6382 6383 }; 6384 6385 class IU_Register { 6386 public: 6387 int32_t _value; 6388 6389 void print() const { 6390 printf("%08x %11d", _value, _value); 6391 } 6392 6393 }; 6394 6395 class IU_State { 6396 public: 6397 Flag_Register _eflags; 6398 IU_Register _rdi; 6399 IU_Register _rsi; 6400 IU_Register _rbp; 6401 IU_Register _rsp; 6402 IU_Register _rbx; 6403 IU_Register _rdx; 6404 IU_Register _rcx; 6405 IU_Register _rax; 6406 6407 void print() const { 6408 // computation registers 6409 printf("rax, = "); _rax.print(); printf("\n"); 6410 printf("rbx, = "); _rbx.print(); printf("\n"); 6411 printf("rcx = "); _rcx.print(); printf("\n"); 6412 printf("rdx = "); _rdx.print(); printf("\n"); 6413 printf("rdi = "); _rdi.print(); printf("\n"); 6414 printf("rsi = "); _rsi.print(); printf("\n"); 6415 printf("rbp, = "); _rbp.print(); printf("\n"); 6416 printf("rsp = "); _rsp.print(); printf("\n"); 6417 printf("\n"); 6418 // control registers 6419 printf("flgs = "); _eflags.print(); printf("\n"); 6420 } 6421 }; 6422 6423 6424 class CPU_State { 6425 public: 6426 FPU_State _fpu_state; 6427 IU_State _iu_state; 6428 6429 void print() const { 6430 printf("--------------------------------------------------\n"); 6431 _iu_state .print(); 6432 printf("\n"); 6433 _fpu_state.print(); 6434 printf("--------------------------------------------------\n"); 6435 } 6436 6437 }; 6438 6439 6440 static void _print_CPU_state(CPU_State* state) { 6441 state->print(); 6442 }; 6443 6444 6445 void MacroAssembler::print_CPU_state() { 6446 push_CPU_state(); 6447 push(rsp); // pass CPU state 6448 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state))); 6449 addptr(rsp, wordSize); // discard argument 6450 pop_CPU_state(); 6451 } 6452 6453 6454 static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) { 6455 static int counter = 0; 6456 FPU_State* fs = &state->_fpu_state; 6457 counter++; 6458 // For leaf calls, only verify that the top few elements remain empty. 6459 // We only need 1 empty at the top for C2 code. 6460 if( stack_depth < 0 ) { 6461 if( fs->tag_for_st(7) != 3 ) { 6462 printf("FPR7 not empty\n"); 6463 state->print(); 6464 assert(false, "error"); 6465 return false; 6466 } 6467 return true; // All other stack states do not matter 6468 } 6469 6470 assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std, 6471 "bad FPU control word"); 6472 6473 // compute stack depth 6474 int i = 0; 6475 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) < 3) i++; 6476 int d = i; 6477 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++; 6478 // verify findings 6479 if (i != FPU_State::number_of_registers) { 6480 // stack not contiguous 6481 printf("%s: stack not contiguous at ST%d\n", s, i); 6482 state->print(); 6483 assert(false, "error"); 6484 return false; 6485 } 6486 // check if computed stack depth corresponds to expected stack depth 6487 if (stack_depth < 0) { 6488 // expected stack depth is -stack_depth or less 6489 if (d > -stack_depth) { 6490 // too many elements on the stack 6491 printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d); 6492 state->print(); 6493 assert(false, "error"); 6494 return false; 6495 } 6496 } else { 6497 // expected stack depth is stack_depth 6498 if (d != stack_depth) { 6499 // wrong stack depth 6500 printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d); 6501 state->print(); 6502 assert(false, "error"); 6503 return false; 6504 } 6505 } 6506 // everything is cool 6507 return true; 6508 } 6509 6510 6511 void MacroAssembler::verify_FPU(int stack_depth, const char* s) { 6512 if (!VerifyFPU) return; 6513 push_CPU_state(); 6514 push(rsp); // pass CPU state 6515 ExternalAddress msg((address) s); 6516 // pass message string s 6517 pushptr(msg.addr()); 6518 push(stack_depth); // pass stack depth 6519 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU))); 6520 addptr(rsp, 3 * wordSize); // discard arguments 6521 // check for error 6522 { Label L; 6523 testl(rax, rax); 6524 jcc(Assembler::notZero, L); 6525 int3(); // break if error condition 6526 bind(L); 6527 } 6528 pop_CPU_state(); 6529 } 6530 6531 void MacroAssembler::restore_cpu_control_state_after_jni() { 6532 // Either restore the MXCSR register after returning from the JNI Call 6533 // or verify that it wasn't changed (with -Xcheck:jni flag). 6534 if (VM_Version::supports_sse()) { 6535 if (RestoreMXCSROnJNICalls) { 6536 ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std())); 6537 } else if (CheckJNICalls) { 6538 call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry())); 6539 } 6540 } 6541 // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty. 6542 vzeroupper(); 6543 // Reset k1 to 0xffff. 6544 if (VM_Version::supports_evex()) { 6545 push(rcx); 6546 movl(rcx, 0xffff); 6547 kmovwl(k1, rcx); 6548 pop(rcx); 6549 } 6550 6551 #ifndef _LP64 6552 // Either restore the x87 floating pointer control word after returning 6553 // from the JNI call or verify that it wasn't changed. 6554 if (CheckJNICalls) { 6555 call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry())); 6556 } 6557 #endif // _LP64 6558 } 6559 6560 // ((OopHandle)result).resolve(); 6561 void MacroAssembler::resolve_oop_handle(Register result) { 6562 // OopHandle::resolve is an indirection. 6563 movptr(result, Address(result, 0)); 6564 } 6565 6566 void MacroAssembler::load_mirror(Register mirror, Register method) { 6567 // get mirror 6568 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 6569 movptr(mirror, Address(method, Method::const_offset())); 6570 movptr(mirror, Address(mirror, ConstMethod::constants_offset())); 6571 movptr(mirror, Address(mirror, ConstantPool::pool_holder_offset_in_bytes())); 6572 movptr(mirror, Address(mirror, mirror_offset)); 6573 resolve_oop_handle(mirror); 6574 } 6575 6576 void MacroAssembler::load_klass(Register dst, Register src) { 6577 #ifdef _LP64 6578 if (UseCompressedClassPointers) { 6579 movl(dst, Address(src, oopDesc::klass_offset_in_bytes())); 6580 decode_klass_not_null(dst); 6581 } else 6582 #endif 6583 movptr(dst, Address(src, oopDesc::klass_offset_in_bytes())); 6584 } 6585 6586 void MacroAssembler::load_prototype_header(Register dst, Register src) { 6587 load_klass(dst, src); 6588 movptr(dst, Address(dst, Klass::prototype_header_offset())); 6589 } 6590 6591 void MacroAssembler::store_klass(Register dst, Register src) { 6592 #ifdef _LP64 6593 if (UseCompressedClassPointers) { 6594 encode_klass_not_null(src); 6595 movl(Address(dst, oopDesc::klass_offset_in_bytes()), src); 6596 } else 6597 #endif 6598 movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src); 6599 } 6600 6601 #if INCLUDE_ALL_GCS && defined(_LP64) 6602 6603 void MacroAssembler::load_barrier(Register ref, Address ref_addr, bool expand_call, LoadBarrierOn on) { 6604 Label done; 6605 const Register resolved_ref_addr = rsi; 6606 assert_different_registers(ref, resolved_ref_addr); 6607 6608 BLOCK_COMMENT("load_barrier {"); 6609 6610 // Save temp register 6611 push(resolved_ref_addr); 6612 6613 // Resolve reference address now, ref_addr might use the same register as ref, 6614 // which means it gets killed when we write to ref. 6615 lea(resolved_ref_addr, ref_addr); 6616 6617 // Load reference 6618 movptr(ref, Address(resolved_ref_addr, 0)); 6619 6620 // Check if mask is not bad, which includes an implicit null check. 6621 testptr(ref, Address(r15_thread, JavaThread::zaddress_bad_mask_offset())); 6622 jcc(Assembler::zero, done); 6623 6624 // Save live registers 6625 push(rax); 6626 push(rcx); 6627 push(rdx); 6628 push(rdi); 6629 push(r8); 6630 push(r9); 6631 push(r10); 6632 push(r11); 6633 6634 // We may end up here from generate_native_wrapper, then the method may have 6635 // floats as arguments, and we must spill them before calling the VM runtime 6636 // leaf. From the interpreter all floats are passed on the stack. 6637 assert(Argument::n_float_register_parameters_j == 8, "Found %d float regs", Argument::n_float_register_parameters_j); 6638 int f_spill_size = Argument::n_float_register_parameters_j * wordSize * 2; 6639 subptr(rsp, f_spill_size); 6640 movdqu(Address(rsp, 14 * wordSize), xmm7); 6641 movdqu(Address(rsp, 12 * wordSize), xmm6); 6642 movdqu(Address(rsp, 10 * wordSize), xmm5); 6643 movdqu(Address(rsp, 8 * wordSize), xmm4); 6644 movdqu(Address(rsp, 6 * wordSize), xmm3); 6645 movdqu(Address(rsp, 4 * wordSize), xmm2); 6646 movdqu(Address(rsp, 2 * wordSize), xmm1); 6647 movdqu(Address(rsp, 0 * wordSize), xmm0); 6648 6649 // Call into VM to handle the slow path 6650 if (expand_call) { 6651 assert(ref != c_rarg1, "smashed arg"); 6652 pass_arg1(this, resolved_ref_addr); 6653 pass_arg0(this, ref); 6654 switch (on) { 6655 case LoadBarrierOnStrongOopRef: 6656 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, SharedRuntime::z_load_barrier_on_oop_field_preloaded), 2); 6657 break; 6658 case LoadBarrierOnWeakOopRef: 6659 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, SharedRuntime::z_load_barrier_on_weak_oop_field_preloaded), 2); 6660 break; 6661 case LoadBarrierOnPhantomOopRef: 6662 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, SharedRuntime::z_load_barrier_on_phantom_oop_field_preloaded), 2); 6663 break; 6664 default: 6665 fatal("Unknown strength: %d", on); 6666 break; 6667 } 6668 } else { 6669 switch (on) { 6670 case LoadBarrierOnStrongOopRef: 6671 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::z_load_barrier_on_oop_field_preloaded), ref, resolved_ref_addr); 6672 break; 6673 case LoadBarrierOnWeakOopRef: 6674 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::z_load_barrier_on_weak_oop_field_preloaded), ref, resolved_ref_addr); 6675 break; 6676 case LoadBarrierOnPhantomOopRef: 6677 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::z_load_barrier_on_phantom_oop_field_preloaded), ref, resolved_ref_addr); 6678 break; 6679 default: fatal("Unknown strength: %d", on); 6680 break; 6681 } 6682 } 6683 6684 // Restore live registers 6685 movdqu(xmm0, Address(rsp, 0 * wordSize)); 6686 movdqu(xmm1, Address(rsp, 2 * wordSize)); 6687 movdqu(xmm2, Address(rsp, 4 * wordSize)); 6688 movdqu(xmm3, Address(rsp, 6 * wordSize)); 6689 movdqu(xmm4, Address(rsp, 8 * wordSize)); 6690 movdqu(xmm5, Address(rsp, 10 * wordSize)); 6691 movdqu(xmm6, Address(rsp, 12 * wordSize)); 6692 movdqu(xmm7, Address(rsp, 14 * wordSize)); 6693 addptr(rsp, f_spill_size); 6694 6695 pop(r11); 6696 pop(r10); 6697 pop(r9); 6698 pop(r8); 6699 pop(rdi); 6700 pop(rdx); 6701 pop(rcx); 6702 6703 if (ref == rax) { 6704 addptr(rsp, wordSize); 6705 } else { 6706 movptr(ref, rax); 6707 pop(rax); 6708 } 6709 6710 bind(done); 6711 6712 // Restore temp register 6713 pop(resolved_ref_addr); 6714 6715 BLOCK_COMMENT("} load_barrier"); 6716 } 6717 6718 #endif 6719 6720 void MacroAssembler::load_heap_oop(Register dst, Address src, bool expand_call, LoadBarrierOn on) { 6721 #ifdef _LP64 6722 #if INCLUDE_ALL_GCS 6723 if (UseLoadBarrier) { 6724 load_barrier(dst, src, expand_call, on); 6725 } else 6726 #endif 6727 if (UseCompressedOops) { 6728 movl(dst, src); 6729 decode_heap_oop(dst); 6730 } else 6731 #endif 6732 movptr(dst, src); 6733 } 6734 6735 // Doesn't do verfication, generates fixed size code 6736 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src) { 6737 #ifdef _LP64 6738 if (UseCompressedOops) { 6739 movl(dst, src); 6740 decode_heap_oop_not_null(dst); 6741 } else 6742 #endif 6743 movptr(dst, src); 6744 } 6745 6746 void MacroAssembler::store_heap_oop(Address dst, Register src) { 6747 #ifdef ASSERT 6748 if (VerifyOops && UseLoadBarrier) { 6749 // Check if mask is good 6750 Label done; 6751 testptr(src, Address(r15_thread, JavaThread::zaddress_bad_mask_offset())); 6752 jcc(Assembler::zero, done); 6753 STOP("Writing broken oop"); 6754 should_not_reach_here(); 6755 bind(done); 6756 } 6757 #endif 6758 6759 #ifdef _LP64 6760 if (UseCompressedOops) { 6761 assert(!dst.uses(src), "not enough registers"); 6762 encode_heap_oop(src); 6763 movl(dst, src); 6764 } else 6765 #endif 6766 movptr(dst, src); 6767 } 6768 6769 void MacroAssembler::cmp_heap_oop(Register src1, Address src2, Register tmp) { 6770 assert_different_registers(src1, tmp); 6771 #ifdef _LP64 6772 if (UseCompressedOops) { 6773 bool did_push = false; 6774 if (tmp == noreg) { 6775 tmp = rax; 6776 push(tmp); 6777 did_push = true; 6778 assert(!src2.uses(rsp), "can't push"); 6779 } 6780 load_heap_oop(tmp, src2); 6781 cmpptr(src1, tmp); 6782 if (did_push) pop(tmp); 6783 } else 6784 #endif 6785 cmpptr(src1, src2); 6786 } 6787 6788 // Used for storing NULLs. 6789 void MacroAssembler::store_heap_oop_null(Address dst) { 6790 #ifdef _LP64 6791 if (UseCompressedOops) { 6792 movl(dst, (int32_t)NULL_WORD); 6793 } else { 6794 movslq(dst, (int32_t)NULL_WORD); 6795 } 6796 #else 6797 movl(dst, (int32_t)NULL_WORD); 6798 #endif 6799 } 6800 6801 #ifdef _LP64 6802 void MacroAssembler::store_klass_gap(Register dst, Register src) { 6803 if (UseCompressedClassPointers) { 6804 // Store to klass gap in destination 6805 movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src); 6806 } 6807 } 6808 6809 #ifdef ASSERT 6810 void MacroAssembler::verify_heapbase(const char* msg) { 6811 assert (UseCompressedOops, "should be compressed"); 6812 assert (Universe::heap() != NULL, "java heap should be initialized"); 6813 if (CheckCompressedOops) { 6814 Label ok; 6815 push(rscratch1); // cmpptr trashes rscratch1 6816 cmpptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr())); 6817 jcc(Assembler::equal, ok); 6818 STOP(msg); 6819 bind(ok); 6820 pop(rscratch1); 6821 } 6822 } 6823 #endif 6824 6825 // Algorithm must match oop.inline.hpp encode_heap_oop. 6826 void MacroAssembler::encode_heap_oop(Register r) { 6827 #ifdef ASSERT 6828 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?"); 6829 #endif 6830 verify_oop(r, "broken oop in encode_heap_oop"); 6831 if (Universe::narrow_oop_base() == NULL) { 6832 if (Universe::narrow_oop_shift() != 0) { 6833 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6834 shrq(r, LogMinObjAlignmentInBytes); 6835 } 6836 return; 6837 } 6838 testq(r, r); 6839 cmovq(Assembler::equal, r, r12_heapbase); 6840 subq(r, r12_heapbase); 6841 shrq(r, LogMinObjAlignmentInBytes); 6842 } 6843 6844 void MacroAssembler::encode_heap_oop_not_null(Register r) { 6845 #ifdef ASSERT 6846 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?"); 6847 if (CheckCompressedOops) { 6848 Label ok; 6849 testq(r, r); 6850 jcc(Assembler::notEqual, ok); 6851 STOP("null oop passed to encode_heap_oop_not_null"); 6852 bind(ok); 6853 } 6854 #endif 6855 verify_oop(r, "broken oop in encode_heap_oop_not_null"); 6856 if (Universe::narrow_oop_base() != NULL) { 6857 subq(r, r12_heapbase); 6858 } 6859 if (Universe::narrow_oop_shift() != 0) { 6860 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6861 shrq(r, LogMinObjAlignmentInBytes); 6862 } 6863 } 6864 6865 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) { 6866 #ifdef ASSERT 6867 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?"); 6868 if (CheckCompressedOops) { 6869 Label ok; 6870 testq(src, src); 6871 jcc(Assembler::notEqual, ok); 6872 STOP("null oop passed to encode_heap_oop_not_null2"); 6873 bind(ok); 6874 } 6875 #endif 6876 verify_oop(src, "broken oop in encode_heap_oop_not_null2"); 6877 if (dst != src) { 6878 movq(dst, src); 6879 } 6880 if (Universe::narrow_oop_base() != NULL) { 6881 subq(dst, r12_heapbase); 6882 } 6883 if (Universe::narrow_oop_shift() != 0) { 6884 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6885 shrq(dst, LogMinObjAlignmentInBytes); 6886 } 6887 } 6888 6889 void MacroAssembler::decode_heap_oop(Register r) { 6890 #ifdef ASSERT 6891 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?"); 6892 #endif 6893 if (Universe::narrow_oop_base() == NULL) { 6894 if (Universe::narrow_oop_shift() != 0) { 6895 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6896 shlq(r, LogMinObjAlignmentInBytes); 6897 } 6898 } else { 6899 Label done; 6900 shlq(r, LogMinObjAlignmentInBytes); 6901 jccb(Assembler::equal, done); 6902 addq(r, r12_heapbase); 6903 bind(done); 6904 } 6905 verify_oop(r, "broken oop in decode_heap_oop"); 6906 } 6907 6908 void MacroAssembler::decode_heap_oop_not_null(Register r) { 6909 // Note: it will change flags 6910 assert (UseCompressedOops, "should only be used for compressed headers"); 6911 assert (Universe::heap() != NULL, "java heap should be initialized"); 6912 // Cannot assert, unverified entry point counts instructions (see .ad file) 6913 // vtableStubs also counts instructions in pd_code_size_limit. 6914 // Also do not verify_oop as this is called by verify_oop. 6915 if (Universe::narrow_oop_shift() != 0) { 6916 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6917 shlq(r, LogMinObjAlignmentInBytes); 6918 if (Universe::narrow_oop_base() != NULL) { 6919 addq(r, r12_heapbase); 6920 } 6921 } else { 6922 assert (Universe::narrow_oop_base() == NULL, "sanity"); 6923 } 6924 } 6925 6926 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) { 6927 // Note: it will change flags 6928 assert (UseCompressedOops, "should only be used for compressed headers"); 6929 assert (Universe::heap() != NULL, "java heap should be initialized"); 6930 // Cannot assert, unverified entry point counts instructions (see .ad file) 6931 // vtableStubs also counts instructions in pd_code_size_limit. 6932 // Also do not verify_oop as this is called by verify_oop. 6933 if (Universe::narrow_oop_shift() != 0) { 6934 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6935 if (LogMinObjAlignmentInBytes == Address::times_8) { 6936 leaq(dst, Address(r12_heapbase, src, Address::times_8, 0)); 6937 } else { 6938 if (dst != src) { 6939 movq(dst, src); 6940 } 6941 shlq(dst, LogMinObjAlignmentInBytes); 6942 if (Universe::narrow_oop_base() != NULL) { 6943 addq(dst, r12_heapbase); 6944 } 6945 } 6946 } else { 6947 assert (Universe::narrow_oop_base() == NULL, "sanity"); 6948 if (dst != src) { 6949 movq(dst, src); 6950 } 6951 } 6952 } 6953 6954 void MacroAssembler::encode_klass_not_null(Register r) { 6955 if (Universe::narrow_klass_base() != NULL) { 6956 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base. 6957 assert(r != r12_heapbase, "Encoding a klass in r12"); 6958 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base()); 6959 subq(r, r12_heapbase); 6960 } 6961 if (Universe::narrow_klass_shift() != 0) { 6962 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 6963 shrq(r, LogKlassAlignmentInBytes); 6964 } 6965 if (Universe::narrow_klass_base() != NULL) { 6966 reinit_heapbase(); 6967 } 6968 } 6969 6970 void MacroAssembler::encode_klass_not_null(Register dst, Register src) { 6971 if (dst == src) { 6972 encode_klass_not_null(src); 6973 } else { 6974 if (Universe::narrow_klass_base() != NULL) { 6975 mov64(dst, (int64_t)Universe::narrow_klass_base()); 6976 negq(dst); 6977 addq(dst, src); 6978 } else { 6979 movptr(dst, src); 6980 } 6981 if (Universe::narrow_klass_shift() != 0) { 6982 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 6983 shrq(dst, LogKlassAlignmentInBytes); 6984 } 6985 } 6986 } 6987 6988 // Function instr_size_for_decode_klass_not_null() counts the instructions 6989 // generated by decode_klass_not_null(register r) and reinit_heapbase(), 6990 // when (Universe::heap() != NULL). Hence, if the instructions they 6991 // generate change, then this method needs to be updated. 6992 int MacroAssembler::instr_size_for_decode_klass_not_null() { 6993 assert (UseCompressedClassPointers, "only for compressed klass ptrs"); 6994 if (Universe::narrow_klass_base() != NULL) { 6995 // mov64 + addq + shlq? + mov64 (for reinit_heapbase()). 6996 return (Universe::narrow_klass_shift() == 0 ? 20 : 24); 6997 } else { 6998 // longest load decode klass function, mov64, leaq 6999 return 16; 7000 } 7001 } 7002 7003 // !!! If the instructions that get generated here change then function 7004 // instr_size_for_decode_klass_not_null() needs to get updated. 7005 void MacroAssembler::decode_klass_not_null(Register r) { 7006 // Note: it will change flags 7007 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7008 assert(r != r12_heapbase, "Decoding a klass in r12"); 7009 // Cannot assert, unverified entry point counts instructions (see .ad file) 7010 // vtableStubs also counts instructions in pd_code_size_limit. 7011 // Also do not verify_oop as this is called by verify_oop. 7012 if (Universe::narrow_klass_shift() != 0) { 7013 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 7014 shlq(r, LogKlassAlignmentInBytes); 7015 } 7016 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base. 7017 if (Universe::narrow_klass_base() != NULL) { 7018 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base()); 7019 addq(r, r12_heapbase); 7020 reinit_heapbase(); 7021 } 7022 } 7023 7024 void MacroAssembler::decode_klass_not_null(Register dst, Register src) { 7025 // Note: it will change flags 7026 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7027 if (dst == src) { 7028 decode_klass_not_null(dst); 7029 } else { 7030 // Cannot assert, unverified entry point counts instructions (see .ad file) 7031 // vtableStubs also counts instructions in pd_code_size_limit. 7032 // Also do not verify_oop as this is called by verify_oop. 7033 mov64(dst, (int64_t)Universe::narrow_klass_base()); 7034 if (Universe::narrow_klass_shift() != 0) { 7035 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 7036 assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?"); 7037 leaq(dst, Address(dst, src, Address::times_8, 0)); 7038 } else { 7039 addq(dst, src); 7040 } 7041 } 7042 } 7043 7044 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) { 7045 assert (UseCompressedOops, "should only be used for compressed headers"); 7046 assert (Universe::heap() != NULL, "java heap should be initialized"); 7047 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7048 int oop_index = oop_recorder()->find_index(obj); 7049 RelocationHolder rspec = oop_Relocation::spec(oop_index); 7050 mov_narrow_oop(dst, oop_index, rspec); 7051 } 7052 7053 void MacroAssembler::set_narrow_oop(Address dst, jobject obj) { 7054 assert (UseCompressedOops, "should only be used for compressed headers"); 7055 assert (Universe::heap() != NULL, "java heap should be initialized"); 7056 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7057 int oop_index = oop_recorder()->find_index(obj); 7058 RelocationHolder rspec = oop_Relocation::spec(oop_index); 7059 mov_narrow_oop(dst, oop_index, rspec); 7060 } 7061 7062 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) { 7063 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7064 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7065 int klass_index = oop_recorder()->find_index(k); 7066 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 7067 mov_narrow_oop(dst, Klass::encode_klass(k), rspec); 7068 } 7069 7070 void MacroAssembler::set_narrow_klass(Address dst, Klass* k) { 7071 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7072 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7073 int klass_index = oop_recorder()->find_index(k); 7074 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 7075 mov_narrow_oop(dst, Klass::encode_klass(k), rspec); 7076 } 7077 7078 void MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) { 7079 assert (UseCompressedOops, "should only be used for compressed headers"); 7080 assert (Universe::heap() != NULL, "java heap should be initialized"); 7081 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7082 int oop_index = oop_recorder()->find_index(obj); 7083 RelocationHolder rspec = oop_Relocation::spec(oop_index); 7084 Assembler::cmp_narrow_oop(dst, oop_index, rspec); 7085 } 7086 7087 void MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) { 7088 assert (UseCompressedOops, "should only be used for compressed headers"); 7089 assert (Universe::heap() != NULL, "java heap should be initialized"); 7090 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7091 int oop_index = oop_recorder()->find_index(obj); 7092 RelocationHolder rspec = oop_Relocation::spec(oop_index); 7093 Assembler::cmp_narrow_oop(dst, oop_index, rspec); 7094 } 7095 7096 void MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) { 7097 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7098 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7099 int klass_index = oop_recorder()->find_index(k); 7100 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 7101 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec); 7102 } 7103 7104 void MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) { 7105 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7106 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7107 int klass_index = oop_recorder()->find_index(k); 7108 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 7109 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec); 7110 } 7111 7112 void MacroAssembler::reinit_heapbase() { 7113 if (UseCompressedOops || UseCompressedClassPointers) { 7114 if (Universe::heap() != NULL) { 7115 if (Universe::narrow_oop_base() == NULL) { 7116 MacroAssembler::xorptr(r12_heapbase, r12_heapbase); 7117 } else { 7118 mov64(r12_heapbase, (int64_t)Universe::narrow_ptrs_base()); 7119 } 7120 } else { 7121 movptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr())); 7122 } 7123 } 7124 } 7125 7126 #endif // _LP64 7127 7128 // C2 compiled method's prolog code. 7129 void MacroAssembler::verified_entry(int framesize, int stack_bang_size, bool fp_mode_24b) { 7130 7131 // WARNING: Initial instruction MUST be 5 bytes or longer so that 7132 // NativeJump::patch_verified_entry will be able to patch out the entry 7133 // code safely. The push to verify stack depth is ok at 5 bytes, 7134 // the frame allocation can be either 3 or 6 bytes. So if we don't do 7135 // stack bang then we must use the 6 byte frame allocation even if 7136 // we have no frame. :-( 7137 assert(stack_bang_size >= framesize || stack_bang_size <= 0, "stack bang size incorrect"); 7138 7139 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned"); 7140 // Remove word for return addr 7141 framesize -= wordSize; 7142 stack_bang_size -= wordSize; 7143 7144 // Calls to C2R adapters often do not accept exceptional returns. 7145 // We require that their callers must bang for them. But be careful, because 7146 // some VM calls (such as call site linkage) can use several kilobytes of 7147 // stack. But the stack safety zone should account for that. 7148 // See bugs 4446381, 4468289, 4497237. 7149 if (stack_bang_size > 0) { 7150 generate_stack_overflow_check(stack_bang_size); 7151 7152 // We always push rbp, so that on return to interpreter rbp, will be 7153 // restored correctly and we can correct the stack. 7154 push(rbp); 7155 // Save caller's stack pointer into RBP if the frame pointer is preserved. 7156 if (PreserveFramePointer) { 7157 mov(rbp, rsp); 7158 } 7159 // Remove word for ebp 7160 framesize -= wordSize; 7161 7162 // Create frame 7163 if (framesize) { 7164 subptr(rsp, framesize); 7165 } 7166 } else { 7167 // Create frame (force generation of a 4 byte immediate value) 7168 subptr_imm32(rsp, framesize); 7169 7170 // Save RBP register now. 7171 framesize -= wordSize; 7172 movptr(Address(rsp, framesize), rbp); 7173 // Save caller's stack pointer into RBP if the frame pointer is preserved. 7174 if (PreserveFramePointer) { 7175 movptr(rbp, rsp); 7176 if (framesize > 0) { 7177 addptr(rbp, framesize); 7178 } 7179 } 7180 } 7181 7182 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth 7183 framesize -= wordSize; 7184 movptr(Address(rsp, framesize), (int32_t)0xbadb100d); 7185 } 7186 7187 #ifndef _LP64 7188 // If method sets FPU control word do it now 7189 if (fp_mode_24b) { 7190 fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24())); 7191 } 7192 if (UseSSE >= 2 && VerifyFPU) { 7193 verify_FPU(0, "FPU stack must be clean on entry"); 7194 } 7195 #endif 7196 7197 #ifdef ASSERT 7198 if (VerifyStackAtCalls) { 7199 Label L; 7200 push(rax); 7201 mov(rax, rsp); 7202 andptr(rax, StackAlignmentInBytes-1); 7203 cmpptr(rax, StackAlignmentInBytes-wordSize); 7204 pop(rax); 7205 jcc(Assembler::equal, L); 7206 STOP("Stack is not properly aligned!"); 7207 bind(L); 7208 } 7209 #endif 7210 7211 } 7212 7213 void MacroAssembler::clear_mem(Register base, Register cnt, Register tmp, bool is_large) { 7214 // cnt - number of qwords (8-byte words). 7215 // base - start address, qword aligned. 7216 // is_large - if optimizers know cnt is larger than InitArrayShortSize 7217 assert(base==rdi, "base register must be edi for rep stos"); 7218 assert(tmp==rax, "tmp register must be eax for rep stos"); 7219 assert(cnt==rcx, "cnt register must be ecx for rep stos"); 7220 assert(InitArrayShortSize % BytesPerLong == 0, 7221 "InitArrayShortSize should be the multiple of BytesPerLong"); 7222 7223 Label DONE; 7224 7225 xorptr(tmp, tmp); 7226 7227 if (!is_large) { 7228 Label LOOP, LONG; 7229 cmpptr(cnt, InitArrayShortSize/BytesPerLong); 7230 jccb(Assembler::greater, LONG); 7231 7232 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM 7233 7234 decrement(cnt); 7235 jccb(Assembler::negative, DONE); // Zero length 7236 7237 // Use individual pointer-sized stores for small counts: 7238 BIND(LOOP); 7239 movptr(Address(base, cnt, Address::times_ptr), tmp); 7240 decrement(cnt); 7241 jccb(Assembler::greaterEqual, LOOP); 7242 jmpb(DONE); 7243 7244 BIND(LONG); 7245 } 7246 7247 // Use longer rep-prefixed ops for non-small counts: 7248 if (UseFastStosb) { 7249 shlptr(cnt, 3); // convert to number of bytes 7250 rep_stosb(); 7251 } else { 7252 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM 7253 rep_stos(); 7254 } 7255 7256 BIND(DONE); 7257 } 7258 7259 #ifdef COMPILER2 7260 7261 // IndexOf for constant substrings with size >= 8 chars 7262 // which don't need to be loaded through stack. 7263 void MacroAssembler::string_indexofC8(Register str1, Register str2, 7264 Register cnt1, Register cnt2, 7265 int int_cnt2, Register result, 7266 XMMRegister vec, Register tmp, 7267 int ae) { 7268 ShortBranchVerifier sbv(this); 7269 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 7270 assert(ae != StrIntrinsicNode::LU, "Invalid encoding"); 7271 7272 // This method uses the pcmpestri instruction with bound registers 7273 // inputs: 7274 // xmm - substring 7275 // rax - substring length (elements count) 7276 // mem - scanned string 7277 // rdx - string length (elements count) 7278 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts) 7279 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes) 7280 // outputs: 7281 // rcx - matched index in string 7282 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 7283 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts 7284 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8 7285 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2; 7286 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1; 7287 7288 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, 7289 RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR, 7290 MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE; 7291 7292 // Note, inline_string_indexOf() generates checks: 7293 // if (substr.count > string.count) return -1; 7294 // if (substr.count == 0) return 0; 7295 assert(int_cnt2 >= stride, "this code is used only for cnt2 >= 8 chars"); 7296 7297 // Load substring. 7298 if (ae == StrIntrinsicNode::UL) { 7299 pmovzxbw(vec, Address(str2, 0)); 7300 } else { 7301 movdqu(vec, Address(str2, 0)); 7302 } 7303 movl(cnt2, int_cnt2); 7304 movptr(result, str1); // string addr 7305 7306 if (int_cnt2 > stride) { 7307 jmpb(SCAN_TO_SUBSTR); 7308 7309 // Reload substr for rescan, this code 7310 // is executed only for large substrings (> 8 chars) 7311 bind(RELOAD_SUBSTR); 7312 if (ae == StrIntrinsicNode::UL) { 7313 pmovzxbw(vec, Address(str2, 0)); 7314 } else { 7315 movdqu(vec, Address(str2, 0)); 7316 } 7317 negptr(cnt2); // Jumped here with negative cnt2, convert to positive 7318 7319 bind(RELOAD_STR); 7320 // We came here after the beginning of the substring was 7321 // matched but the rest of it was not so we need to search 7322 // again. Start from the next element after the previous match. 7323 7324 // cnt2 is number of substring reminding elements and 7325 // cnt1 is number of string reminding elements when cmp failed. 7326 // Restored cnt1 = cnt1 - cnt2 + int_cnt2 7327 subl(cnt1, cnt2); 7328 addl(cnt1, int_cnt2); 7329 movl(cnt2, int_cnt2); // Now restore cnt2 7330 7331 decrementl(cnt1); // Shift to next element 7332 cmpl(cnt1, cnt2); 7333 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7334 7335 addptr(result, (1<<scale1)); 7336 7337 } // (int_cnt2 > 8) 7338 7339 // Scan string for start of substr in 16-byte vectors 7340 bind(SCAN_TO_SUBSTR); 7341 pcmpestri(vec, Address(result, 0), mode); 7342 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1 7343 subl(cnt1, stride); 7344 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string 7345 cmpl(cnt1, cnt2); 7346 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7347 addptr(result, 16); 7348 jmpb(SCAN_TO_SUBSTR); 7349 7350 // Found a potential substr 7351 bind(FOUND_CANDIDATE); 7352 // Matched whole vector if first element matched (tmp(rcx) == 0). 7353 if (int_cnt2 == stride) { 7354 jccb(Assembler::overflow, RET_FOUND); // OF == 1 7355 } else { // int_cnt2 > 8 7356 jccb(Assembler::overflow, FOUND_SUBSTR); 7357 } 7358 // After pcmpestri tmp(rcx) contains matched element index 7359 // Compute start addr of substr 7360 lea(result, Address(result, tmp, scale1)); 7361 7362 // Make sure string is still long enough 7363 subl(cnt1, tmp); 7364 cmpl(cnt1, cnt2); 7365 if (int_cnt2 == stride) { 7366 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR); 7367 } else { // int_cnt2 > 8 7368 jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD); 7369 } 7370 // Left less then substring. 7371 7372 bind(RET_NOT_FOUND); 7373 movl(result, -1); 7374 jmp(EXIT); 7375 7376 if (int_cnt2 > stride) { 7377 // This code is optimized for the case when whole substring 7378 // is matched if its head is matched. 7379 bind(MATCH_SUBSTR_HEAD); 7380 pcmpestri(vec, Address(result, 0), mode); 7381 // Reload only string if does not match 7382 jcc(Assembler::noOverflow, RELOAD_STR); // OF == 0 7383 7384 Label CONT_SCAN_SUBSTR; 7385 // Compare the rest of substring (> 8 chars). 7386 bind(FOUND_SUBSTR); 7387 // First 8 chars are already matched. 7388 negptr(cnt2); 7389 addptr(cnt2, stride); 7390 7391 bind(SCAN_SUBSTR); 7392 subl(cnt1, stride); 7393 cmpl(cnt2, -stride); // Do not read beyond substring 7394 jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR); 7395 // Back-up strings to avoid reading beyond substring: 7396 // cnt1 = cnt1 - cnt2 + 8 7397 addl(cnt1, cnt2); // cnt2 is negative 7398 addl(cnt1, stride); 7399 movl(cnt2, stride); negptr(cnt2); 7400 bind(CONT_SCAN_SUBSTR); 7401 if (int_cnt2 < (int)G) { 7402 int tail_off1 = int_cnt2<<scale1; 7403 int tail_off2 = int_cnt2<<scale2; 7404 if (ae == StrIntrinsicNode::UL) { 7405 pmovzxbw(vec, Address(str2, cnt2, scale2, tail_off2)); 7406 } else { 7407 movdqu(vec, Address(str2, cnt2, scale2, tail_off2)); 7408 } 7409 pcmpestri(vec, Address(result, cnt2, scale1, tail_off1), mode); 7410 } else { 7411 // calculate index in register to avoid integer overflow (int_cnt2*2) 7412 movl(tmp, int_cnt2); 7413 addptr(tmp, cnt2); 7414 if (ae == StrIntrinsicNode::UL) { 7415 pmovzxbw(vec, Address(str2, tmp, scale2, 0)); 7416 } else { 7417 movdqu(vec, Address(str2, tmp, scale2, 0)); 7418 } 7419 pcmpestri(vec, Address(result, tmp, scale1, 0), mode); 7420 } 7421 // Need to reload strings pointers if not matched whole vector 7422 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0 7423 addptr(cnt2, stride); 7424 jcc(Assembler::negative, SCAN_SUBSTR); 7425 // Fall through if found full substring 7426 7427 } // (int_cnt2 > 8) 7428 7429 bind(RET_FOUND); 7430 // Found result if we matched full small substring. 7431 // Compute substr offset 7432 subptr(result, str1); 7433 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 7434 shrl(result, 1); // index 7435 } 7436 bind(EXIT); 7437 7438 } // string_indexofC8 7439 7440 // Small strings are loaded through stack if they cross page boundary. 7441 void MacroAssembler::string_indexof(Register str1, Register str2, 7442 Register cnt1, Register cnt2, 7443 int int_cnt2, Register result, 7444 XMMRegister vec, Register tmp, 7445 int ae) { 7446 ShortBranchVerifier sbv(this); 7447 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 7448 assert(ae != StrIntrinsicNode::LU, "Invalid encoding"); 7449 7450 // 7451 // int_cnt2 is length of small (< 8 chars) constant substring 7452 // or (-1) for non constant substring in which case its length 7453 // is in cnt2 register. 7454 // 7455 // Note, inline_string_indexOf() generates checks: 7456 // if (substr.count > string.count) return -1; 7457 // if (substr.count == 0) return 0; 7458 // 7459 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8 7460 assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < stride), "should be != 0"); 7461 // This method uses the pcmpestri instruction with bound registers 7462 // inputs: 7463 // xmm - substring 7464 // rax - substring length (elements count) 7465 // mem - scanned string 7466 // rdx - string length (elements count) 7467 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts) 7468 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes) 7469 // outputs: 7470 // rcx - matched index in string 7471 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 7472 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts 7473 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2; 7474 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1; 7475 7476 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR, 7477 RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR, 7478 FOUND_CANDIDATE; 7479 7480 { //======================================================== 7481 // We don't know where these strings are located 7482 // and we can't read beyond them. Load them through stack. 7483 Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR; 7484 7485 movptr(tmp, rsp); // save old SP 7486 7487 if (int_cnt2 > 0) { // small (< 8 chars) constant substring 7488 if (int_cnt2 == (1>>scale2)) { // One byte 7489 assert((ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL), "Only possible for latin1 encoding"); 7490 load_unsigned_byte(result, Address(str2, 0)); 7491 movdl(vec, result); // move 32 bits 7492 } else if (ae == StrIntrinsicNode::LL && int_cnt2 == 3) { // Three bytes 7493 // Not enough header space in 32-bit VM: 12+3 = 15. 7494 movl(result, Address(str2, -1)); 7495 shrl(result, 8); 7496 movdl(vec, result); // move 32 bits 7497 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (2>>scale2)) { // One char 7498 load_unsigned_short(result, Address(str2, 0)); 7499 movdl(vec, result); // move 32 bits 7500 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (4>>scale2)) { // Two chars 7501 movdl(vec, Address(str2, 0)); // move 32 bits 7502 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (8>>scale2)) { // Four chars 7503 movq(vec, Address(str2, 0)); // move 64 bits 7504 } else { // cnt2 = { 3, 5, 6, 7 } || (ae == StrIntrinsicNode::UL && cnt2 ={2, ..., 7}) 7505 // Array header size is 12 bytes in 32-bit VM 7506 // + 6 bytes for 3 chars == 18 bytes, 7507 // enough space to load vec and shift. 7508 assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity"); 7509 if (ae == StrIntrinsicNode::UL) { 7510 int tail_off = int_cnt2-8; 7511 pmovzxbw(vec, Address(str2, tail_off)); 7512 psrldq(vec, -2*tail_off); 7513 } 7514 else { 7515 int tail_off = int_cnt2*(1<<scale2); 7516 movdqu(vec, Address(str2, tail_off-16)); 7517 psrldq(vec, 16-tail_off); 7518 } 7519 } 7520 } else { // not constant substring 7521 cmpl(cnt2, stride); 7522 jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough 7523 7524 // We can read beyond string if srt+16 does not cross page boundary 7525 // since heaps are aligned and mapped by pages. 7526 assert(os::vm_page_size() < (int)G, "default page should be small"); 7527 movl(result, str2); // We need only low 32 bits 7528 andl(result, (os::vm_page_size()-1)); 7529 cmpl(result, (os::vm_page_size()-16)); 7530 jccb(Assembler::belowEqual, CHECK_STR); 7531 7532 // Move small strings to stack to allow load 16 bytes into vec. 7533 subptr(rsp, 16); 7534 int stk_offset = wordSize-(1<<scale2); 7535 push(cnt2); 7536 7537 bind(COPY_SUBSTR); 7538 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL) { 7539 load_unsigned_byte(result, Address(str2, cnt2, scale2, -1)); 7540 movb(Address(rsp, cnt2, scale2, stk_offset), result); 7541 } else if (ae == StrIntrinsicNode::UU) { 7542 load_unsigned_short(result, Address(str2, cnt2, scale2, -2)); 7543 movw(Address(rsp, cnt2, scale2, stk_offset), result); 7544 } 7545 decrement(cnt2); 7546 jccb(Assembler::notZero, COPY_SUBSTR); 7547 7548 pop(cnt2); 7549 movptr(str2, rsp); // New substring address 7550 } // non constant 7551 7552 bind(CHECK_STR); 7553 cmpl(cnt1, stride); 7554 jccb(Assembler::aboveEqual, BIG_STRINGS); 7555 7556 // Check cross page boundary. 7557 movl(result, str1); // We need only low 32 bits 7558 andl(result, (os::vm_page_size()-1)); 7559 cmpl(result, (os::vm_page_size()-16)); 7560 jccb(Assembler::belowEqual, BIG_STRINGS); 7561 7562 subptr(rsp, 16); 7563 int stk_offset = -(1<<scale1); 7564 if (int_cnt2 < 0) { // not constant 7565 push(cnt2); 7566 stk_offset += wordSize; 7567 } 7568 movl(cnt2, cnt1); 7569 7570 bind(COPY_STR); 7571 if (ae == StrIntrinsicNode::LL) { 7572 load_unsigned_byte(result, Address(str1, cnt2, scale1, -1)); 7573 movb(Address(rsp, cnt2, scale1, stk_offset), result); 7574 } else { 7575 load_unsigned_short(result, Address(str1, cnt2, scale1, -2)); 7576 movw(Address(rsp, cnt2, scale1, stk_offset), result); 7577 } 7578 decrement(cnt2); 7579 jccb(Assembler::notZero, COPY_STR); 7580 7581 if (int_cnt2 < 0) { // not constant 7582 pop(cnt2); 7583 } 7584 movptr(str1, rsp); // New string address 7585 7586 bind(BIG_STRINGS); 7587 // Load substring. 7588 if (int_cnt2 < 0) { // -1 7589 if (ae == StrIntrinsicNode::UL) { 7590 pmovzxbw(vec, Address(str2, 0)); 7591 } else { 7592 movdqu(vec, Address(str2, 0)); 7593 } 7594 push(cnt2); // substr count 7595 push(str2); // substr addr 7596 push(str1); // string addr 7597 } else { 7598 // Small (< 8 chars) constant substrings are loaded already. 7599 movl(cnt2, int_cnt2); 7600 } 7601 push(tmp); // original SP 7602 7603 } // Finished loading 7604 7605 //======================================================== 7606 // Start search 7607 // 7608 7609 movptr(result, str1); // string addr 7610 7611 if (int_cnt2 < 0) { // Only for non constant substring 7612 jmpb(SCAN_TO_SUBSTR); 7613 7614 // SP saved at sp+0 7615 // String saved at sp+1*wordSize 7616 // Substr saved at sp+2*wordSize 7617 // Substr count saved at sp+3*wordSize 7618 7619 // Reload substr for rescan, this code 7620 // is executed only for large substrings (> 8 chars) 7621 bind(RELOAD_SUBSTR); 7622 movptr(str2, Address(rsp, 2*wordSize)); 7623 movl(cnt2, Address(rsp, 3*wordSize)); 7624 if (ae == StrIntrinsicNode::UL) { 7625 pmovzxbw(vec, Address(str2, 0)); 7626 } else { 7627 movdqu(vec, Address(str2, 0)); 7628 } 7629 // We came here after the beginning of the substring was 7630 // matched but the rest of it was not so we need to search 7631 // again. Start from the next element after the previous match. 7632 subptr(str1, result); // Restore counter 7633 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 7634 shrl(str1, 1); 7635 } 7636 addl(cnt1, str1); 7637 decrementl(cnt1); // Shift to next element 7638 cmpl(cnt1, cnt2); 7639 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7640 7641 addptr(result, (1<<scale1)); 7642 } // non constant 7643 7644 // Scan string for start of substr in 16-byte vectors 7645 bind(SCAN_TO_SUBSTR); 7646 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 7647 pcmpestri(vec, Address(result, 0), mode); 7648 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1 7649 subl(cnt1, stride); 7650 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string 7651 cmpl(cnt1, cnt2); 7652 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7653 addptr(result, 16); 7654 7655 bind(ADJUST_STR); 7656 cmpl(cnt1, stride); // Do not read beyond string 7657 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR); 7658 // Back-up string to avoid reading beyond string. 7659 lea(result, Address(result, cnt1, scale1, -16)); 7660 movl(cnt1, stride); 7661 jmpb(SCAN_TO_SUBSTR); 7662 7663 // Found a potential substr 7664 bind(FOUND_CANDIDATE); 7665 // After pcmpestri tmp(rcx) contains matched element index 7666 7667 // Make sure string is still long enough 7668 subl(cnt1, tmp); 7669 cmpl(cnt1, cnt2); 7670 jccb(Assembler::greaterEqual, FOUND_SUBSTR); 7671 // Left less then substring. 7672 7673 bind(RET_NOT_FOUND); 7674 movl(result, -1); 7675 jmpb(CLEANUP); 7676 7677 bind(FOUND_SUBSTR); 7678 // Compute start addr of substr 7679 lea(result, Address(result, tmp, scale1)); 7680 if (int_cnt2 > 0) { // Constant substring 7681 // Repeat search for small substring (< 8 chars) 7682 // from new point without reloading substring. 7683 // Have to check that we don't read beyond string. 7684 cmpl(tmp, stride-int_cnt2); 7685 jccb(Assembler::greater, ADJUST_STR); 7686 // Fall through if matched whole substring. 7687 } else { // non constant 7688 assert(int_cnt2 == -1, "should be != 0"); 7689 7690 addl(tmp, cnt2); 7691 // Found result if we matched whole substring. 7692 cmpl(tmp, stride); 7693 jccb(Assembler::lessEqual, RET_FOUND); 7694 7695 // Repeat search for small substring (<= 8 chars) 7696 // from new point 'str1' without reloading substring. 7697 cmpl(cnt2, stride); 7698 // Have to check that we don't read beyond string. 7699 jccb(Assembler::lessEqual, ADJUST_STR); 7700 7701 Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG; 7702 // Compare the rest of substring (> 8 chars). 7703 movptr(str1, result); 7704 7705 cmpl(tmp, cnt2); 7706 // First 8 chars are already matched. 7707 jccb(Assembler::equal, CHECK_NEXT); 7708 7709 bind(SCAN_SUBSTR); 7710 pcmpestri(vec, Address(str1, 0), mode); 7711 // Need to reload strings pointers if not matched whole vector 7712 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0 7713 7714 bind(CHECK_NEXT); 7715 subl(cnt2, stride); 7716 jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring 7717 addptr(str1, 16); 7718 if (ae == StrIntrinsicNode::UL) { 7719 addptr(str2, 8); 7720 } else { 7721 addptr(str2, 16); 7722 } 7723 subl(cnt1, stride); 7724 cmpl(cnt2, stride); // Do not read beyond substring 7725 jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR); 7726 // Back-up strings to avoid reading beyond substring. 7727 7728 if (ae == StrIntrinsicNode::UL) { 7729 lea(str2, Address(str2, cnt2, scale2, -8)); 7730 lea(str1, Address(str1, cnt2, scale1, -16)); 7731 } else { 7732 lea(str2, Address(str2, cnt2, scale2, -16)); 7733 lea(str1, Address(str1, cnt2, scale1, -16)); 7734 } 7735 subl(cnt1, cnt2); 7736 movl(cnt2, stride); 7737 addl(cnt1, stride); 7738 bind(CONT_SCAN_SUBSTR); 7739 if (ae == StrIntrinsicNode::UL) { 7740 pmovzxbw(vec, Address(str2, 0)); 7741 } else { 7742 movdqu(vec, Address(str2, 0)); 7743 } 7744 jmp(SCAN_SUBSTR); 7745 7746 bind(RET_FOUND_LONG); 7747 movptr(str1, Address(rsp, wordSize)); 7748 } // non constant 7749 7750 bind(RET_FOUND); 7751 // Compute substr offset 7752 subptr(result, str1); 7753 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 7754 shrl(result, 1); // index 7755 } 7756 bind(CLEANUP); 7757 pop(rsp); // restore SP 7758 7759 } // string_indexof 7760 7761 void MacroAssembler::string_indexof_char(Register str1, Register cnt1, Register ch, Register result, 7762 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp) { 7763 ShortBranchVerifier sbv(this); 7764 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 7765 7766 int stride = 8; 7767 7768 Label FOUND_CHAR, SCAN_TO_CHAR, SCAN_TO_CHAR_LOOP, 7769 SCAN_TO_8_CHAR, SCAN_TO_8_CHAR_LOOP, SCAN_TO_16_CHAR_LOOP, 7770 RET_NOT_FOUND, SCAN_TO_8_CHAR_INIT, 7771 FOUND_SEQ_CHAR, DONE_LABEL; 7772 7773 movptr(result, str1); 7774 if (UseAVX >= 2) { 7775 cmpl(cnt1, stride); 7776 jcc(Assembler::less, SCAN_TO_CHAR_LOOP); 7777 cmpl(cnt1, 2*stride); 7778 jcc(Assembler::less, SCAN_TO_8_CHAR_INIT); 7779 movdl(vec1, ch); 7780 vpbroadcastw(vec1, vec1); 7781 vpxor(vec2, vec2); 7782 movl(tmp, cnt1); 7783 andl(tmp, 0xFFFFFFF0); //vector count (in chars) 7784 andl(cnt1,0x0000000F); //tail count (in chars) 7785 7786 bind(SCAN_TO_16_CHAR_LOOP); 7787 vmovdqu(vec3, Address(result, 0)); 7788 vpcmpeqw(vec3, vec3, vec1, 1); 7789 vptest(vec2, vec3); 7790 jcc(Assembler::carryClear, FOUND_CHAR); 7791 addptr(result, 32); 7792 subl(tmp, 2*stride); 7793 jccb(Assembler::notZero, SCAN_TO_16_CHAR_LOOP); 7794 jmp(SCAN_TO_8_CHAR); 7795 bind(SCAN_TO_8_CHAR_INIT); 7796 movdl(vec1, ch); 7797 pshuflw(vec1, vec1, 0x00); 7798 pshufd(vec1, vec1, 0); 7799 pxor(vec2, vec2); 7800 } 7801 bind(SCAN_TO_8_CHAR); 7802 cmpl(cnt1, stride); 7803 if (UseAVX >= 2) { 7804 jcc(Assembler::less, SCAN_TO_CHAR); 7805 } else { 7806 jcc(Assembler::less, SCAN_TO_CHAR_LOOP); 7807 movdl(vec1, ch); 7808 pshuflw(vec1, vec1, 0x00); 7809 pshufd(vec1, vec1, 0); 7810 pxor(vec2, vec2); 7811 } 7812 movl(tmp, cnt1); 7813 andl(tmp, 0xFFFFFFF8); //vector count (in chars) 7814 andl(cnt1,0x00000007); //tail count (in chars) 7815 7816 bind(SCAN_TO_8_CHAR_LOOP); 7817 movdqu(vec3, Address(result, 0)); 7818 pcmpeqw(vec3, vec1); 7819 ptest(vec2, vec3); 7820 jcc(Assembler::carryClear, FOUND_CHAR); 7821 addptr(result, 16); 7822 subl(tmp, stride); 7823 jccb(Assembler::notZero, SCAN_TO_8_CHAR_LOOP); 7824 bind(SCAN_TO_CHAR); 7825 testl(cnt1, cnt1); 7826 jcc(Assembler::zero, RET_NOT_FOUND); 7827 bind(SCAN_TO_CHAR_LOOP); 7828 load_unsigned_short(tmp, Address(result, 0)); 7829 cmpl(ch, tmp); 7830 jccb(Assembler::equal, FOUND_SEQ_CHAR); 7831 addptr(result, 2); 7832 subl(cnt1, 1); 7833 jccb(Assembler::zero, RET_NOT_FOUND); 7834 jmp(SCAN_TO_CHAR_LOOP); 7835 7836 bind(RET_NOT_FOUND); 7837 movl(result, -1); 7838 jmpb(DONE_LABEL); 7839 7840 bind(FOUND_CHAR); 7841 if (UseAVX >= 2) { 7842 vpmovmskb(tmp, vec3); 7843 } else { 7844 pmovmskb(tmp, vec3); 7845 } 7846 bsfl(ch, tmp); 7847 addl(result, ch); 7848 7849 bind(FOUND_SEQ_CHAR); 7850 subptr(result, str1); 7851 shrl(result, 1); 7852 7853 bind(DONE_LABEL); 7854 } // string_indexof_char 7855 7856 // helper function for string_compare 7857 void MacroAssembler::load_next_elements(Register elem1, Register elem2, Register str1, Register str2, 7858 Address::ScaleFactor scale, Address::ScaleFactor scale1, 7859 Address::ScaleFactor scale2, Register index, int ae) { 7860 if (ae == StrIntrinsicNode::LL) { 7861 load_unsigned_byte(elem1, Address(str1, index, scale, 0)); 7862 load_unsigned_byte(elem2, Address(str2, index, scale, 0)); 7863 } else if (ae == StrIntrinsicNode::UU) { 7864 load_unsigned_short(elem1, Address(str1, index, scale, 0)); 7865 load_unsigned_short(elem2, Address(str2, index, scale, 0)); 7866 } else { 7867 load_unsigned_byte(elem1, Address(str1, index, scale1, 0)); 7868 load_unsigned_short(elem2, Address(str2, index, scale2, 0)); 7869 } 7870 } 7871 7872 // Compare strings, used for char[] and byte[]. 7873 void MacroAssembler::string_compare(Register str1, Register str2, 7874 Register cnt1, Register cnt2, Register result, 7875 XMMRegister vec1, int ae) { 7876 ShortBranchVerifier sbv(this); 7877 Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL; 7878 Label COMPARE_WIDE_VECTORS_LOOP_FAILED; // used only _LP64 && AVX3 7879 int stride, stride2, adr_stride, adr_stride1, adr_stride2; 7880 int stride2x2 = 0x40; 7881 Address::ScaleFactor scale = Address::no_scale; 7882 Address::ScaleFactor scale1 = Address::no_scale; 7883 Address::ScaleFactor scale2 = Address::no_scale; 7884 7885 if (ae != StrIntrinsicNode::LL) { 7886 stride2x2 = 0x20; 7887 } 7888 7889 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) { 7890 shrl(cnt2, 1); 7891 } 7892 // Compute the minimum of the string lengths and the 7893 // difference of the string lengths (stack). 7894 // Do the conditional move stuff 7895 movl(result, cnt1); 7896 subl(cnt1, cnt2); 7897 push(cnt1); 7898 cmov32(Assembler::lessEqual, cnt2, result); // cnt2 = min(cnt1, cnt2) 7899 7900 // Is the minimum length zero? 7901 testl(cnt2, cnt2); 7902 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 7903 if (ae == StrIntrinsicNode::LL) { 7904 // Load first bytes 7905 load_unsigned_byte(result, Address(str1, 0)); // result = str1[0] 7906 load_unsigned_byte(cnt1, Address(str2, 0)); // cnt1 = str2[0] 7907 } else if (ae == StrIntrinsicNode::UU) { 7908 // Load first characters 7909 load_unsigned_short(result, Address(str1, 0)); 7910 load_unsigned_short(cnt1, Address(str2, 0)); 7911 } else { 7912 load_unsigned_byte(result, Address(str1, 0)); 7913 load_unsigned_short(cnt1, Address(str2, 0)); 7914 } 7915 subl(result, cnt1); 7916 jcc(Assembler::notZero, POP_LABEL); 7917 7918 if (ae == StrIntrinsicNode::UU) { 7919 // Divide length by 2 to get number of chars 7920 shrl(cnt2, 1); 7921 } 7922 cmpl(cnt2, 1); 7923 jcc(Assembler::equal, LENGTH_DIFF_LABEL); 7924 7925 // Check if the strings start at the same location and setup scale and stride 7926 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7927 cmpptr(str1, str2); 7928 jcc(Assembler::equal, LENGTH_DIFF_LABEL); 7929 if (ae == StrIntrinsicNode::LL) { 7930 scale = Address::times_1; 7931 stride = 16; 7932 } else { 7933 scale = Address::times_2; 7934 stride = 8; 7935 } 7936 } else { 7937 scale1 = Address::times_1; 7938 scale2 = Address::times_2; 7939 // scale not used 7940 stride = 8; 7941 } 7942 7943 if (UseAVX >= 2 && UseSSE42Intrinsics) { 7944 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR; 7945 Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR; 7946 Label COMPARE_WIDE_VECTORS_LOOP_AVX2; 7947 Label COMPARE_TAIL_LONG; 7948 Label COMPARE_WIDE_VECTORS_LOOP_AVX3; // used only _LP64 && AVX3 7949 7950 int pcmpmask = 0x19; 7951 if (ae == StrIntrinsicNode::LL) { 7952 pcmpmask &= ~0x01; 7953 } 7954 7955 // Setup to compare 16-chars (32-bytes) vectors, 7956 // start from first character again because it has aligned address. 7957 if (ae == StrIntrinsicNode::LL) { 7958 stride2 = 32; 7959 } else { 7960 stride2 = 16; 7961 } 7962 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7963 adr_stride = stride << scale; 7964 } else { 7965 adr_stride1 = 8; //stride << scale1; 7966 adr_stride2 = 16; //stride << scale2; 7967 } 7968 7969 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri"); 7970 // rax and rdx are used by pcmpestri as elements counters 7971 movl(result, cnt2); 7972 andl(cnt2, ~(stride2-1)); // cnt2 holds the vector count 7973 jcc(Assembler::zero, COMPARE_TAIL_LONG); 7974 7975 // fast path : compare first 2 8-char vectors. 7976 bind(COMPARE_16_CHARS); 7977 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7978 movdqu(vec1, Address(str1, 0)); 7979 } else { 7980 pmovzxbw(vec1, Address(str1, 0)); 7981 } 7982 pcmpestri(vec1, Address(str2, 0), pcmpmask); 7983 jccb(Assembler::below, COMPARE_INDEX_CHAR); 7984 7985 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7986 movdqu(vec1, Address(str1, adr_stride)); 7987 pcmpestri(vec1, Address(str2, adr_stride), pcmpmask); 7988 } else { 7989 pmovzxbw(vec1, Address(str1, adr_stride1)); 7990 pcmpestri(vec1, Address(str2, adr_stride2), pcmpmask); 7991 } 7992 jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS); 7993 addl(cnt1, stride); 7994 7995 // Compare the characters at index in cnt1 7996 bind(COMPARE_INDEX_CHAR); // cnt1 has the offset of the mismatching character 7997 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae); 7998 subl(result, cnt2); 7999 jmp(POP_LABEL); 8000 8001 // Setup the registers to start vector comparison loop 8002 bind(COMPARE_WIDE_VECTORS); 8003 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8004 lea(str1, Address(str1, result, scale)); 8005 lea(str2, Address(str2, result, scale)); 8006 } else { 8007 lea(str1, Address(str1, result, scale1)); 8008 lea(str2, Address(str2, result, scale2)); 8009 } 8010 subl(result, stride2); 8011 subl(cnt2, stride2); 8012 jcc(Assembler::zero, COMPARE_WIDE_TAIL); 8013 negptr(result); 8014 8015 // In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest) 8016 bind(COMPARE_WIDE_VECTORS_LOOP); 8017 8018 #ifdef _LP64 8019 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop 8020 cmpl(cnt2, stride2x2); 8021 jccb(Assembler::below, COMPARE_WIDE_VECTORS_LOOP_AVX2); 8022 testl(cnt2, stride2x2-1); // cnt2 holds the vector count 8023 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX2); // means we cannot subtract by 0x40 8024 8025 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop 8026 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8027 evmovdquq(vec1, Address(str1, result, scale), Assembler::AVX_512bit); 8028 evpcmpeqb(k7, vec1, Address(str2, result, scale), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0 8029 } else { 8030 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_512bit); 8031 evpcmpeqb(k7, vec1, Address(str2, result, scale2), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0 8032 } 8033 kortestql(k7, k7); 8034 jcc(Assembler::aboveEqual, COMPARE_WIDE_VECTORS_LOOP_FAILED); // miscompare 8035 addptr(result, stride2x2); // update since we already compared at this addr 8036 subl(cnt2, stride2x2); // and sub the size too 8037 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX3); 8038 8039 vpxor(vec1, vec1); 8040 jmpb(COMPARE_WIDE_TAIL); 8041 }//if (VM_Version::supports_avx512vlbw()) 8042 #endif // _LP64 8043 8044 8045 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2); 8046 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8047 vmovdqu(vec1, Address(str1, result, scale)); 8048 vpxor(vec1, Address(str2, result, scale)); 8049 } else { 8050 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_256bit); 8051 vpxor(vec1, Address(str2, result, scale2)); 8052 } 8053 vptest(vec1, vec1); 8054 jcc(Assembler::notZero, VECTOR_NOT_EQUAL); 8055 addptr(result, stride2); 8056 subl(cnt2, stride2); 8057 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP); 8058 // clean upper bits of YMM registers 8059 vpxor(vec1, vec1); 8060 8061 // compare wide vectors tail 8062 bind(COMPARE_WIDE_TAIL); 8063 testptr(result, result); 8064 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 8065 8066 movl(result, stride2); 8067 movl(cnt2, result); 8068 negptr(result); 8069 jmp(COMPARE_WIDE_VECTORS_LOOP_AVX2); 8070 8071 // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors. 8072 bind(VECTOR_NOT_EQUAL); 8073 // clean upper bits of YMM registers 8074 vpxor(vec1, vec1); 8075 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8076 lea(str1, Address(str1, result, scale)); 8077 lea(str2, Address(str2, result, scale)); 8078 } else { 8079 lea(str1, Address(str1, result, scale1)); 8080 lea(str2, Address(str2, result, scale2)); 8081 } 8082 jmp(COMPARE_16_CHARS); 8083 8084 // Compare tail chars, length between 1 to 15 chars 8085 bind(COMPARE_TAIL_LONG); 8086 movl(cnt2, result); 8087 cmpl(cnt2, stride); 8088 jcc(Assembler::less, COMPARE_SMALL_STR); 8089 8090 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8091 movdqu(vec1, Address(str1, 0)); 8092 } else { 8093 pmovzxbw(vec1, Address(str1, 0)); 8094 } 8095 pcmpestri(vec1, Address(str2, 0), pcmpmask); 8096 jcc(Assembler::below, COMPARE_INDEX_CHAR); 8097 subptr(cnt2, stride); 8098 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 8099 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8100 lea(str1, Address(str1, result, scale)); 8101 lea(str2, Address(str2, result, scale)); 8102 } else { 8103 lea(str1, Address(str1, result, scale1)); 8104 lea(str2, Address(str2, result, scale2)); 8105 } 8106 negptr(cnt2); 8107 jmpb(WHILE_HEAD_LABEL); 8108 8109 bind(COMPARE_SMALL_STR); 8110 } else if (UseSSE42Intrinsics) { 8111 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL; 8112 int pcmpmask = 0x19; 8113 // Setup to compare 8-char (16-byte) vectors, 8114 // start from first character again because it has aligned address. 8115 movl(result, cnt2); 8116 andl(cnt2, ~(stride - 1)); // cnt2 holds the vector count 8117 if (ae == StrIntrinsicNode::LL) { 8118 pcmpmask &= ~0x01; 8119 } 8120 jcc(Assembler::zero, COMPARE_TAIL); 8121 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8122 lea(str1, Address(str1, result, scale)); 8123 lea(str2, Address(str2, result, scale)); 8124 } else { 8125 lea(str1, Address(str1, result, scale1)); 8126 lea(str2, Address(str2, result, scale2)); 8127 } 8128 negptr(result); 8129 8130 // pcmpestri 8131 // inputs: 8132 // vec1- substring 8133 // rax - negative string length (elements count) 8134 // mem - scanned string 8135 // rdx - string length (elements count) 8136 // pcmpmask - cmp mode: 11000 (string compare with negated result) 8137 // + 00 (unsigned bytes) or + 01 (unsigned shorts) 8138 // outputs: 8139 // rcx - first mismatched element index 8140 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri"); 8141 8142 bind(COMPARE_WIDE_VECTORS); 8143 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8144 movdqu(vec1, Address(str1, result, scale)); 8145 pcmpestri(vec1, Address(str2, result, scale), pcmpmask); 8146 } else { 8147 pmovzxbw(vec1, Address(str1, result, scale1)); 8148 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask); 8149 } 8150 // After pcmpestri cnt1(rcx) contains mismatched element index 8151 8152 jccb(Assembler::below, VECTOR_NOT_EQUAL); // CF==1 8153 addptr(result, stride); 8154 subptr(cnt2, stride); 8155 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS); 8156 8157 // compare wide vectors tail 8158 testptr(result, result); 8159 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 8160 8161 movl(cnt2, stride); 8162 movl(result, stride); 8163 negptr(result); 8164 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8165 movdqu(vec1, Address(str1, result, scale)); 8166 pcmpestri(vec1, Address(str2, result, scale), pcmpmask); 8167 } else { 8168 pmovzxbw(vec1, Address(str1, result, scale1)); 8169 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask); 8170 } 8171 jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL); 8172 8173 // Mismatched characters in the vectors 8174 bind(VECTOR_NOT_EQUAL); 8175 addptr(cnt1, result); 8176 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae); 8177 subl(result, cnt2); 8178 jmpb(POP_LABEL); 8179 8180 bind(COMPARE_TAIL); // limit is zero 8181 movl(cnt2, result); 8182 // Fallthru to tail compare 8183 } 8184 // Shift str2 and str1 to the end of the arrays, negate min 8185 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8186 lea(str1, Address(str1, cnt2, scale)); 8187 lea(str2, Address(str2, cnt2, scale)); 8188 } else { 8189 lea(str1, Address(str1, cnt2, scale1)); 8190 lea(str2, Address(str2, cnt2, scale2)); 8191 } 8192 decrementl(cnt2); // first character was compared already 8193 negptr(cnt2); 8194 8195 // Compare the rest of the elements 8196 bind(WHILE_HEAD_LABEL); 8197 load_next_elements(result, cnt1, str1, str2, scale, scale1, scale2, cnt2, ae); 8198 subl(result, cnt1); 8199 jccb(Assembler::notZero, POP_LABEL); 8200 increment(cnt2); 8201 jccb(Assembler::notZero, WHILE_HEAD_LABEL); 8202 8203 // Strings are equal up to min length. Return the length difference. 8204 bind(LENGTH_DIFF_LABEL); 8205 pop(result); 8206 if (ae == StrIntrinsicNode::UU) { 8207 // Divide diff by 2 to get number of chars 8208 sarl(result, 1); 8209 } 8210 jmpb(DONE_LABEL); 8211 8212 #ifdef _LP64 8213 if (VM_Version::supports_avx512vlbw()) { 8214 8215 bind(COMPARE_WIDE_VECTORS_LOOP_FAILED); 8216 8217 kmovql(cnt1, k7); 8218 notq(cnt1); 8219 bsfq(cnt2, cnt1); 8220 if (ae != StrIntrinsicNode::LL) { 8221 // Divide diff by 2 to get number of chars 8222 sarl(cnt2, 1); 8223 } 8224 addq(result, cnt2); 8225 if (ae == StrIntrinsicNode::LL) { 8226 load_unsigned_byte(cnt1, Address(str2, result)); 8227 load_unsigned_byte(result, Address(str1, result)); 8228 } else if (ae == StrIntrinsicNode::UU) { 8229 load_unsigned_short(cnt1, Address(str2, result, scale)); 8230 load_unsigned_short(result, Address(str1, result, scale)); 8231 } else { 8232 load_unsigned_short(cnt1, Address(str2, result, scale2)); 8233 load_unsigned_byte(result, Address(str1, result, scale1)); 8234 } 8235 subl(result, cnt1); 8236 jmpb(POP_LABEL); 8237 }//if (VM_Version::supports_avx512vlbw()) 8238 #endif // _LP64 8239 8240 // Discard the stored length difference 8241 bind(POP_LABEL); 8242 pop(cnt1); 8243 8244 // That's it 8245 bind(DONE_LABEL); 8246 if(ae == StrIntrinsicNode::UL) { 8247 negl(result); 8248 } 8249 8250 } 8251 8252 // Search for Non-ASCII character (Negative byte value) in a byte array, 8253 // return true if it has any and false otherwise. 8254 // ..\jdk\src\java.base\share\classes\java\lang\StringCoding.java 8255 // @HotSpotIntrinsicCandidate 8256 // private static boolean hasNegatives(byte[] ba, int off, int len) { 8257 // for (int i = off; i < off + len; i++) { 8258 // if (ba[i] < 0) { 8259 // return true; 8260 // } 8261 // } 8262 // return false; 8263 // } 8264 void MacroAssembler::has_negatives(Register ary1, Register len, 8265 Register result, Register tmp1, 8266 XMMRegister vec1, XMMRegister vec2) { 8267 // rsi: byte array 8268 // rcx: len 8269 // rax: result 8270 ShortBranchVerifier sbv(this); 8271 assert_different_registers(ary1, len, result, tmp1); 8272 assert_different_registers(vec1, vec2); 8273 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_CHAR, COMPARE_VECTORS, COMPARE_BYTE; 8274 8275 // len == 0 8276 testl(len, len); 8277 jcc(Assembler::zero, FALSE_LABEL); 8278 8279 if ((UseAVX > 2) && // AVX512 8280 VM_Version::supports_avx512vlbw() && 8281 VM_Version::supports_bmi2()) { 8282 8283 set_vector_masking(); // opening of the stub context for programming mask registers 8284 8285 Label test_64_loop, test_tail; 8286 Register tmp3_aliased = len; 8287 8288 movl(tmp1, len); 8289 vpxor(vec2, vec2, vec2, Assembler::AVX_512bit); 8290 8291 andl(tmp1, 64 - 1); // tail count (in chars) 0x3F 8292 andl(len, ~(64 - 1)); // vector count (in chars) 8293 jccb(Assembler::zero, test_tail); 8294 8295 lea(ary1, Address(ary1, len, Address::times_1)); 8296 negptr(len); 8297 8298 bind(test_64_loop); 8299 // Check whether our 64 elements of size byte contain negatives 8300 evpcmpgtb(k2, vec2, Address(ary1, len, Address::times_1), Assembler::AVX_512bit); 8301 kortestql(k2, k2); 8302 jcc(Assembler::notZero, TRUE_LABEL); 8303 8304 addptr(len, 64); 8305 jccb(Assembler::notZero, test_64_loop); 8306 8307 8308 bind(test_tail); 8309 // bail out when there is nothing to be done 8310 testl(tmp1, -1); 8311 jcc(Assembler::zero, FALSE_LABEL); 8312 8313 // Save k1 8314 kmovql(k3, k1); 8315 8316 // ~(~0 << len) applied up to two times (for 32-bit scenario) 8317 #ifdef _LP64 8318 mov64(tmp3_aliased, 0xFFFFFFFFFFFFFFFF); 8319 shlxq(tmp3_aliased, tmp3_aliased, tmp1); 8320 notq(tmp3_aliased); 8321 kmovql(k1, tmp3_aliased); 8322 #else 8323 Label k_init; 8324 jmp(k_init); 8325 8326 // We could not read 64-bits from a general purpose register thus we move 8327 // data required to compose 64 1's to the instruction stream 8328 // We emit 64 byte wide series of elements from 0..63 which later on would 8329 // be used as a compare targets with tail count contained in tmp1 register. 8330 // Result would be a k1 register having tmp1 consecutive number or 1 8331 // counting from least significant bit. 8332 address tmp = pc(); 8333 emit_int64(0x0706050403020100); 8334 emit_int64(0x0F0E0D0C0B0A0908); 8335 emit_int64(0x1716151413121110); 8336 emit_int64(0x1F1E1D1C1B1A1918); 8337 emit_int64(0x2726252423222120); 8338 emit_int64(0x2F2E2D2C2B2A2928); 8339 emit_int64(0x3736353433323130); 8340 emit_int64(0x3F3E3D3C3B3A3938); 8341 8342 bind(k_init); 8343 lea(len, InternalAddress(tmp)); 8344 // create mask to test for negative byte inside a vector 8345 evpbroadcastb(vec1, tmp1, Assembler::AVX_512bit); 8346 evpcmpgtb(k1, vec1, Address(len, 0), Assembler::AVX_512bit); 8347 8348 #endif 8349 evpcmpgtb(k2, k1, vec2, Address(ary1, 0), Assembler::AVX_512bit); 8350 ktestq(k2, k1); 8351 // Restore k1 8352 kmovql(k1, k3); 8353 jcc(Assembler::notZero, TRUE_LABEL); 8354 8355 jmp(FALSE_LABEL); 8356 8357 clear_vector_masking(); // closing of the stub context for programming mask registers 8358 } else { 8359 movl(result, len); // copy 8360 8361 if (UseAVX == 2 && UseSSE >= 2) { 8362 // With AVX2, use 32-byte vector compare 8363 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8364 8365 // Compare 32-byte vectors 8366 andl(result, 0x0000001f); // tail count (in bytes) 8367 andl(len, 0xffffffe0); // vector count (in bytes) 8368 jccb(Assembler::zero, COMPARE_TAIL); 8369 8370 lea(ary1, Address(ary1, len, Address::times_1)); 8371 negptr(len); 8372 8373 movl(tmp1, 0x80808080); // create mask to test for Unicode chars in vector 8374 movdl(vec2, tmp1); 8375 vpbroadcastd(vec2, vec2); 8376 8377 bind(COMPARE_WIDE_VECTORS); 8378 vmovdqu(vec1, Address(ary1, len, Address::times_1)); 8379 vptest(vec1, vec2); 8380 jccb(Assembler::notZero, TRUE_LABEL); 8381 addptr(len, 32); 8382 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8383 8384 testl(result, result); 8385 jccb(Assembler::zero, FALSE_LABEL); 8386 8387 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32)); 8388 vptest(vec1, vec2); 8389 jccb(Assembler::notZero, TRUE_LABEL); 8390 jmpb(FALSE_LABEL); 8391 8392 bind(COMPARE_TAIL); // len is zero 8393 movl(len, result); 8394 // Fallthru to tail compare 8395 } else if (UseSSE42Intrinsics) { 8396 // With SSE4.2, use double quad vector compare 8397 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8398 8399 // Compare 16-byte vectors 8400 andl(result, 0x0000000f); // tail count (in bytes) 8401 andl(len, 0xfffffff0); // vector count (in bytes) 8402 jccb(Assembler::zero, COMPARE_TAIL); 8403 8404 lea(ary1, Address(ary1, len, Address::times_1)); 8405 negptr(len); 8406 8407 movl(tmp1, 0x80808080); 8408 movdl(vec2, tmp1); 8409 pshufd(vec2, vec2, 0); 8410 8411 bind(COMPARE_WIDE_VECTORS); 8412 movdqu(vec1, Address(ary1, len, Address::times_1)); 8413 ptest(vec1, vec2); 8414 jccb(Assembler::notZero, TRUE_LABEL); 8415 addptr(len, 16); 8416 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8417 8418 testl(result, result); 8419 jccb(Assembler::zero, FALSE_LABEL); 8420 8421 movdqu(vec1, Address(ary1, result, Address::times_1, -16)); 8422 ptest(vec1, vec2); 8423 jccb(Assembler::notZero, TRUE_LABEL); 8424 jmpb(FALSE_LABEL); 8425 8426 bind(COMPARE_TAIL); // len is zero 8427 movl(len, result); 8428 // Fallthru to tail compare 8429 } 8430 } 8431 // Compare 4-byte vectors 8432 andl(len, 0xfffffffc); // vector count (in bytes) 8433 jccb(Assembler::zero, COMPARE_CHAR); 8434 8435 lea(ary1, Address(ary1, len, Address::times_1)); 8436 negptr(len); 8437 8438 bind(COMPARE_VECTORS); 8439 movl(tmp1, Address(ary1, len, Address::times_1)); 8440 andl(tmp1, 0x80808080); 8441 jccb(Assembler::notZero, TRUE_LABEL); 8442 addptr(len, 4); 8443 jcc(Assembler::notZero, COMPARE_VECTORS); 8444 8445 // Compare trailing char (final 2 bytes), if any 8446 bind(COMPARE_CHAR); 8447 testl(result, 0x2); // tail char 8448 jccb(Assembler::zero, COMPARE_BYTE); 8449 load_unsigned_short(tmp1, Address(ary1, 0)); 8450 andl(tmp1, 0x00008080); 8451 jccb(Assembler::notZero, TRUE_LABEL); 8452 subptr(result, 2); 8453 lea(ary1, Address(ary1, 2)); 8454 8455 bind(COMPARE_BYTE); 8456 testl(result, 0x1); // tail byte 8457 jccb(Assembler::zero, FALSE_LABEL); 8458 load_unsigned_byte(tmp1, Address(ary1, 0)); 8459 andl(tmp1, 0x00000080); 8460 jccb(Assembler::notEqual, TRUE_LABEL); 8461 jmpb(FALSE_LABEL); 8462 8463 bind(TRUE_LABEL); 8464 movl(result, 1); // return true 8465 jmpb(DONE); 8466 8467 bind(FALSE_LABEL); 8468 xorl(result, result); // return false 8469 8470 // That's it 8471 bind(DONE); 8472 if (UseAVX >= 2 && UseSSE >= 2) { 8473 // clean upper bits of YMM registers 8474 vpxor(vec1, vec1); 8475 vpxor(vec2, vec2); 8476 } 8477 } 8478 // Compare char[] or byte[] arrays aligned to 4 bytes or substrings. 8479 void MacroAssembler::arrays_equals(bool is_array_equ, Register ary1, Register ary2, 8480 Register limit, Register result, Register chr, 8481 XMMRegister vec1, XMMRegister vec2, bool is_char) { 8482 ShortBranchVerifier sbv(this); 8483 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR, COMPARE_BYTE; 8484 8485 int length_offset = arrayOopDesc::length_offset_in_bytes(); 8486 int base_offset = arrayOopDesc::base_offset_in_bytes(is_char ? T_CHAR : T_BYTE); 8487 8488 if (is_array_equ) { 8489 // Check the input args 8490 cmpoop(ary1, ary2); 8491 jcc(Assembler::equal, TRUE_LABEL); 8492 8493 // Need additional checks for arrays_equals. 8494 testptr(ary1, ary1); 8495 jcc(Assembler::zero, FALSE_LABEL); 8496 testptr(ary2, ary2); 8497 jcc(Assembler::zero, FALSE_LABEL); 8498 8499 // Check the lengths 8500 movl(limit, Address(ary1, length_offset)); 8501 cmpl(limit, Address(ary2, length_offset)); 8502 jcc(Assembler::notEqual, FALSE_LABEL); 8503 } 8504 8505 // count == 0 8506 testl(limit, limit); 8507 jcc(Assembler::zero, TRUE_LABEL); 8508 8509 if (is_array_equ) { 8510 // Load array address 8511 lea(ary1, Address(ary1, base_offset)); 8512 lea(ary2, Address(ary2, base_offset)); 8513 } 8514 8515 if (is_array_equ && is_char) { 8516 // arrays_equals when used for char[]. 8517 shll(limit, 1); // byte count != 0 8518 } 8519 movl(result, limit); // copy 8520 8521 if (UseAVX >= 2) { 8522 // With AVX2, use 32-byte vector compare 8523 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8524 8525 // Compare 32-byte vectors 8526 andl(result, 0x0000001f); // tail count (in bytes) 8527 andl(limit, 0xffffffe0); // vector count (in bytes) 8528 jcc(Assembler::zero, COMPARE_TAIL); 8529 8530 lea(ary1, Address(ary1, limit, Address::times_1)); 8531 lea(ary2, Address(ary2, limit, Address::times_1)); 8532 negptr(limit); 8533 8534 bind(COMPARE_WIDE_VECTORS); 8535 8536 #ifdef _LP64 8537 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop 8538 Label COMPARE_WIDE_VECTORS_LOOP_AVX2, COMPARE_WIDE_VECTORS_LOOP_AVX3; 8539 8540 cmpl(limit, -64); 8541 jccb(Assembler::greater, COMPARE_WIDE_VECTORS_LOOP_AVX2); 8542 8543 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop 8544 8545 evmovdquq(vec1, Address(ary1, limit, Address::times_1), Assembler::AVX_512bit); 8546 evpcmpeqb(k7, vec1, Address(ary2, limit, Address::times_1), Assembler::AVX_512bit); 8547 kortestql(k7, k7); 8548 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare 8549 addptr(limit, 64); // update since we already compared at this addr 8550 cmpl(limit, -64); 8551 jccb(Assembler::lessEqual, COMPARE_WIDE_VECTORS_LOOP_AVX3); 8552 8553 // At this point we may still need to compare -limit+result bytes. 8554 // We could execute the next two instruction and just continue via non-wide path: 8555 // cmpl(limit, 0); 8556 // jcc(Assembler::equal, COMPARE_TAIL); // true 8557 // But since we stopped at the points ary{1,2}+limit which are 8558 // not farther than 64 bytes from the ends of arrays ary{1,2}+result 8559 // (|limit| <= 32 and result < 32), 8560 // we may just compare the last 64 bytes. 8561 // 8562 addptr(result, -64); // it is safe, bc we just came from this area 8563 evmovdquq(vec1, Address(ary1, result, Address::times_1), Assembler::AVX_512bit); 8564 evpcmpeqb(k7, vec1, Address(ary2, result, Address::times_1), Assembler::AVX_512bit); 8565 kortestql(k7, k7); 8566 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare 8567 8568 jmp(TRUE_LABEL); 8569 8570 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2); 8571 8572 }//if (VM_Version::supports_avx512vlbw()) 8573 #endif //_LP64 8574 8575 vmovdqu(vec1, Address(ary1, limit, Address::times_1)); 8576 vmovdqu(vec2, Address(ary2, limit, Address::times_1)); 8577 vpxor(vec1, vec2); 8578 8579 vptest(vec1, vec1); 8580 jcc(Assembler::notZero, FALSE_LABEL); 8581 addptr(limit, 32); 8582 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8583 8584 testl(result, result); 8585 jcc(Assembler::zero, TRUE_LABEL); 8586 8587 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32)); 8588 vmovdqu(vec2, Address(ary2, result, Address::times_1, -32)); 8589 vpxor(vec1, vec2); 8590 8591 vptest(vec1, vec1); 8592 jccb(Assembler::notZero, FALSE_LABEL); 8593 jmpb(TRUE_LABEL); 8594 8595 bind(COMPARE_TAIL); // limit is zero 8596 movl(limit, result); 8597 // Fallthru to tail compare 8598 } else if (UseSSE42Intrinsics) { 8599 // With SSE4.2, use double quad vector compare 8600 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8601 8602 // Compare 16-byte vectors 8603 andl(result, 0x0000000f); // tail count (in bytes) 8604 andl(limit, 0xfffffff0); // vector count (in bytes) 8605 jcc(Assembler::zero, COMPARE_TAIL); 8606 8607 lea(ary1, Address(ary1, limit, Address::times_1)); 8608 lea(ary2, Address(ary2, limit, Address::times_1)); 8609 negptr(limit); 8610 8611 bind(COMPARE_WIDE_VECTORS); 8612 movdqu(vec1, Address(ary1, limit, Address::times_1)); 8613 movdqu(vec2, Address(ary2, limit, Address::times_1)); 8614 pxor(vec1, vec2); 8615 8616 ptest(vec1, vec1); 8617 jcc(Assembler::notZero, FALSE_LABEL); 8618 addptr(limit, 16); 8619 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8620 8621 testl(result, result); 8622 jcc(Assembler::zero, TRUE_LABEL); 8623 8624 movdqu(vec1, Address(ary1, result, Address::times_1, -16)); 8625 movdqu(vec2, Address(ary2, result, Address::times_1, -16)); 8626 pxor(vec1, vec2); 8627 8628 ptest(vec1, vec1); 8629 jccb(Assembler::notZero, FALSE_LABEL); 8630 jmpb(TRUE_LABEL); 8631 8632 bind(COMPARE_TAIL); // limit is zero 8633 movl(limit, result); 8634 // Fallthru to tail compare 8635 } 8636 8637 // Compare 4-byte vectors 8638 andl(limit, 0xfffffffc); // vector count (in bytes) 8639 jccb(Assembler::zero, COMPARE_CHAR); 8640 8641 lea(ary1, Address(ary1, limit, Address::times_1)); 8642 lea(ary2, Address(ary2, limit, Address::times_1)); 8643 negptr(limit); 8644 8645 bind(COMPARE_VECTORS); 8646 movl(chr, Address(ary1, limit, Address::times_1)); 8647 cmpl(chr, Address(ary2, limit, Address::times_1)); 8648 jccb(Assembler::notEqual, FALSE_LABEL); 8649 addptr(limit, 4); 8650 jcc(Assembler::notZero, COMPARE_VECTORS); 8651 8652 // Compare trailing char (final 2 bytes), if any 8653 bind(COMPARE_CHAR); 8654 testl(result, 0x2); // tail char 8655 jccb(Assembler::zero, COMPARE_BYTE); 8656 load_unsigned_short(chr, Address(ary1, 0)); 8657 load_unsigned_short(limit, Address(ary2, 0)); 8658 cmpl(chr, limit); 8659 jccb(Assembler::notEqual, FALSE_LABEL); 8660 8661 if (is_array_equ && is_char) { 8662 bind(COMPARE_BYTE); 8663 } else { 8664 lea(ary1, Address(ary1, 2)); 8665 lea(ary2, Address(ary2, 2)); 8666 8667 bind(COMPARE_BYTE); 8668 testl(result, 0x1); // tail byte 8669 jccb(Assembler::zero, TRUE_LABEL); 8670 load_unsigned_byte(chr, Address(ary1, 0)); 8671 load_unsigned_byte(limit, Address(ary2, 0)); 8672 cmpl(chr, limit); 8673 jccb(Assembler::notEqual, FALSE_LABEL); 8674 } 8675 bind(TRUE_LABEL); 8676 movl(result, 1); // return true 8677 jmpb(DONE); 8678 8679 bind(FALSE_LABEL); 8680 xorl(result, result); // return false 8681 8682 // That's it 8683 bind(DONE); 8684 if (UseAVX >= 2) { 8685 // clean upper bits of YMM registers 8686 vpxor(vec1, vec1); 8687 vpxor(vec2, vec2); 8688 } 8689 } 8690 8691 #endif 8692 8693 void MacroAssembler::generate_fill(BasicType t, bool aligned, 8694 Register to, Register value, Register count, 8695 Register rtmp, XMMRegister xtmp) { 8696 ShortBranchVerifier sbv(this); 8697 assert_different_registers(to, value, count, rtmp); 8698 Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte; 8699 Label L_fill_2_bytes, L_fill_4_bytes; 8700 8701 int shift = -1; 8702 switch (t) { 8703 case T_BYTE: 8704 shift = 2; 8705 break; 8706 case T_SHORT: 8707 shift = 1; 8708 break; 8709 case T_INT: 8710 shift = 0; 8711 break; 8712 default: ShouldNotReachHere(); 8713 } 8714 8715 if (t == T_BYTE) { 8716 andl(value, 0xff); 8717 movl(rtmp, value); 8718 shll(rtmp, 8); 8719 orl(value, rtmp); 8720 } 8721 if (t == T_SHORT) { 8722 andl(value, 0xffff); 8723 } 8724 if (t == T_BYTE || t == T_SHORT) { 8725 movl(rtmp, value); 8726 shll(rtmp, 16); 8727 orl(value, rtmp); 8728 } 8729 8730 cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element 8731 jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp 8732 if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) { 8733 // align source address at 4 bytes address boundary 8734 if (t == T_BYTE) { 8735 // One byte misalignment happens only for byte arrays 8736 testptr(to, 1); 8737 jccb(Assembler::zero, L_skip_align1); 8738 movb(Address(to, 0), value); 8739 increment(to); 8740 decrement(count); 8741 BIND(L_skip_align1); 8742 } 8743 // Two bytes misalignment happens only for byte and short (char) arrays 8744 testptr(to, 2); 8745 jccb(Assembler::zero, L_skip_align2); 8746 movw(Address(to, 0), value); 8747 addptr(to, 2); 8748 subl(count, 1<<(shift-1)); 8749 BIND(L_skip_align2); 8750 } 8751 if (UseSSE < 2) { 8752 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes; 8753 // Fill 32-byte chunks 8754 subl(count, 8 << shift); 8755 jcc(Assembler::less, L_check_fill_8_bytes); 8756 align(16); 8757 8758 BIND(L_fill_32_bytes_loop); 8759 8760 for (int i = 0; i < 32; i += 4) { 8761 movl(Address(to, i), value); 8762 } 8763 8764 addptr(to, 32); 8765 subl(count, 8 << shift); 8766 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop); 8767 BIND(L_check_fill_8_bytes); 8768 addl(count, 8 << shift); 8769 jccb(Assembler::zero, L_exit); 8770 jmpb(L_fill_8_bytes); 8771 8772 // 8773 // length is too short, just fill qwords 8774 // 8775 BIND(L_fill_8_bytes_loop); 8776 movl(Address(to, 0), value); 8777 movl(Address(to, 4), value); 8778 addptr(to, 8); 8779 BIND(L_fill_8_bytes); 8780 subl(count, 1 << (shift + 1)); 8781 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop); 8782 // fall through to fill 4 bytes 8783 } else { 8784 Label L_fill_32_bytes; 8785 if (!UseUnalignedLoadStores) { 8786 // align to 8 bytes, we know we are 4 byte aligned to start 8787 testptr(to, 4); 8788 jccb(Assembler::zero, L_fill_32_bytes); 8789 movl(Address(to, 0), value); 8790 addptr(to, 4); 8791 subl(count, 1<<shift); 8792 } 8793 BIND(L_fill_32_bytes); 8794 { 8795 assert( UseSSE >= 2, "supported cpu only" ); 8796 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes; 8797 if (UseAVX > 2) { 8798 movl(rtmp, 0xffff); 8799 kmovwl(k1, rtmp); 8800 } 8801 movdl(xtmp, value); 8802 if (UseAVX > 2 && UseUnalignedLoadStores) { 8803 // Fill 64-byte chunks 8804 Label L_fill_64_bytes_loop, L_check_fill_32_bytes; 8805 evpbroadcastd(xtmp, xtmp, Assembler::AVX_512bit); 8806 8807 subl(count, 16 << shift); 8808 jcc(Assembler::less, L_check_fill_32_bytes); 8809 align(16); 8810 8811 BIND(L_fill_64_bytes_loop); 8812 evmovdqul(Address(to, 0), xtmp, Assembler::AVX_512bit); 8813 addptr(to, 64); 8814 subl(count, 16 << shift); 8815 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop); 8816 8817 BIND(L_check_fill_32_bytes); 8818 addl(count, 8 << shift); 8819 jccb(Assembler::less, L_check_fill_8_bytes); 8820 vmovdqu(Address(to, 0), xtmp); 8821 addptr(to, 32); 8822 subl(count, 8 << shift); 8823 8824 BIND(L_check_fill_8_bytes); 8825 } else if (UseAVX == 2 && UseUnalignedLoadStores) { 8826 // Fill 64-byte chunks 8827 Label L_fill_64_bytes_loop, L_check_fill_32_bytes; 8828 vpbroadcastd(xtmp, xtmp); 8829 8830 subl(count, 16 << shift); 8831 jcc(Assembler::less, L_check_fill_32_bytes); 8832 align(16); 8833 8834 BIND(L_fill_64_bytes_loop); 8835 vmovdqu(Address(to, 0), xtmp); 8836 vmovdqu(Address(to, 32), xtmp); 8837 addptr(to, 64); 8838 subl(count, 16 << shift); 8839 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop); 8840 8841 BIND(L_check_fill_32_bytes); 8842 addl(count, 8 << shift); 8843 jccb(Assembler::less, L_check_fill_8_bytes); 8844 vmovdqu(Address(to, 0), xtmp); 8845 addptr(to, 32); 8846 subl(count, 8 << shift); 8847 8848 BIND(L_check_fill_8_bytes); 8849 // clean upper bits of YMM registers 8850 movdl(xtmp, value); 8851 pshufd(xtmp, xtmp, 0); 8852 } else { 8853 // Fill 32-byte chunks 8854 pshufd(xtmp, xtmp, 0); 8855 8856 subl(count, 8 << shift); 8857 jcc(Assembler::less, L_check_fill_8_bytes); 8858 align(16); 8859 8860 BIND(L_fill_32_bytes_loop); 8861 8862 if (UseUnalignedLoadStores) { 8863 movdqu(Address(to, 0), xtmp); 8864 movdqu(Address(to, 16), xtmp); 8865 } else { 8866 movq(Address(to, 0), xtmp); 8867 movq(Address(to, 8), xtmp); 8868 movq(Address(to, 16), xtmp); 8869 movq(Address(to, 24), xtmp); 8870 } 8871 8872 addptr(to, 32); 8873 subl(count, 8 << shift); 8874 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop); 8875 8876 BIND(L_check_fill_8_bytes); 8877 } 8878 addl(count, 8 << shift); 8879 jccb(Assembler::zero, L_exit); 8880 jmpb(L_fill_8_bytes); 8881 8882 // 8883 // length is too short, just fill qwords 8884 // 8885 BIND(L_fill_8_bytes_loop); 8886 movq(Address(to, 0), xtmp); 8887 addptr(to, 8); 8888 BIND(L_fill_8_bytes); 8889 subl(count, 1 << (shift + 1)); 8890 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop); 8891 } 8892 } 8893 // fill trailing 4 bytes 8894 BIND(L_fill_4_bytes); 8895 testl(count, 1<<shift); 8896 jccb(Assembler::zero, L_fill_2_bytes); 8897 movl(Address(to, 0), value); 8898 if (t == T_BYTE || t == T_SHORT) { 8899 addptr(to, 4); 8900 BIND(L_fill_2_bytes); 8901 // fill trailing 2 bytes 8902 testl(count, 1<<(shift-1)); 8903 jccb(Assembler::zero, L_fill_byte); 8904 movw(Address(to, 0), value); 8905 if (t == T_BYTE) { 8906 addptr(to, 2); 8907 BIND(L_fill_byte); 8908 // fill trailing byte 8909 testl(count, 1); 8910 jccb(Assembler::zero, L_exit); 8911 movb(Address(to, 0), value); 8912 } else { 8913 BIND(L_fill_byte); 8914 } 8915 } else { 8916 BIND(L_fill_2_bytes); 8917 } 8918 BIND(L_exit); 8919 } 8920 8921 // encode char[] to byte[] in ISO_8859_1 8922 //@HotSpotIntrinsicCandidate 8923 //private static int implEncodeISOArray(byte[] sa, int sp, 8924 //byte[] da, int dp, int len) { 8925 // int i = 0; 8926 // for (; i < len; i++) { 8927 // char c = StringUTF16.getChar(sa, sp++); 8928 // if (c > '\u00FF') 8929 // break; 8930 // da[dp++] = (byte)c; 8931 // } 8932 // return i; 8933 //} 8934 void MacroAssembler::encode_iso_array(Register src, Register dst, Register len, 8935 XMMRegister tmp1Reg, XMMRegister tmp2Reg, 8936 XMMRegister tmp3Reg, XMMRegister tmp4Reg, 8937 Register tmp5, Register result) { 8938 8939 // rsi: src 8940 // rdi: dst 8941 // rdx: len 8942 // rcx: tmp5 8943 // rax: result 8944 ShortBranchVerifier sbv(this); 8945 assert_different_registers(src, dst, len, tmp5, result); 8946 Label L_done, L_copy_1_char, L_copy_1_char_exit; 8947 8948 // set result 8949 xorl(result, result); 8950 // check for zero length 8951 testl(len, len); 8952 jcc(Assembler::zero, L_done); 8953 8954 movl(result, len); 8955 8956 // Setup pointers 8957 lea(src, Address(src, len, Address::times_2)); // char[] 8958 lea(dst, Address(dst, len, Address::times_1)); // byte[] 8959 negptr(len); 8960 8961 if (UseSSE42Intrinsics || UseAVX >= 2) { 8962 Label L_chars_8_check, L_copy_8_chars, L_copy_8_chars_exit; 8963 Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit; 8964 8965 if (UseAVX >= 2) { 8966 Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit; 8967 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector 8968 movdl(tmp1Reg, tmp5); 8969 vpbroadcastd(tmp1Reg, tmp1Reg); 8970 jmp(L_chars_32_check); 8971 8972 bind(L_copy_32_chars); 8973 vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64)); 8974 vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32)); 8975 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1); 8976 vptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 8977 jccb(Assembler::notZero, L_copy_32_chars_exit); 8978 vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1); 8979 vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector_len */ 1); 8980 vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg); 8981 8982 bind(L_chars_32_check); 8983 addptr(len, 32); 8984 jcc(Assembler::lessEqual, L_copy_32_chars); 8985 8986 bind(L_copy_32_chars_exit); 8987 subptr(len, 16); 8988 jccb(Assembler::greater, L_copy_16_chars_exit); 8989 8990 } else if (UseSSE42Intrinsics) { 8991 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector 8992 movdl(tmp1Reg, tmp5); 8993 pshufd(tmp1Reg, tmp1Reg, 0); 8994 jmpb(L_chars_16_check); 8995 } 8996 8997 bind(L_copy_16_chars); 8998 if (UseAVX >= 2) { 8999 vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32)); 9000 vptest(tmp2Reg, tmp1Reg); 9001 jcc(Assembler::notZero, L_copy_16_chars_exit); 9002 vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector_len */ 1); 9003 vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector_len */ 1); 9004 } else { 9005 if (UseAVX > 0) { 9006 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32)); 9007 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16)); 9008 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 0); 9009 } else { 9010 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32)); 9011 por(tmp2Reg, tmp3Reg); 9012 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16)); 9013 por(tmp2Reg, tmp4Reg); 9014 } 9015 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 9016 jccb(Assembler::notZero, L_copy_16_chars_exit); 9017 packuswb(tmp3Reg, tmp4Reg); 9018 } 9019 movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg); 9020 9021 bind(L_chars_16_check); 9022 addptr(len, 16); 9023 jcc(Assembler::lessEqual, L_copy_16_chars); 9024 9025 bind(L_copy_16_chars_exit); 9026 if (UseAVX >= 2) { 9027 // clean upper bits of YMM registers 9028 vpxor(tmp2Reg, tmp2Reg); 9029 vpxor(tmp3Reg, tmp3Reg); 9030 vpxor(tmp4Reg, tmp4Reg); 9031 movdl(tmp1Reg, tmp5); 9032 pshufd(tmp1Reg, tmp1Reg, 0); 9033 } 9034 subptr(len, 8); 9035 jccb(Assembler::greater, L_copy_8_chars_exit); 9036 9037 bind(L_copy_8_chars); 9038 movdqu(tmp3Reg, Address(src, len, Address::times_2, -16)); 9039 ptest(tmp3Reg, tmp1Reg); 9040 jccb(Assembler::notZero, L_copy_8_chars_exit); 9041 packuswb(tmp3Reg, tmp1Reg); 9042 movq(Address(dst, len, Address::times_1, -8), tmp3Reg); 9043 addptr(len, 8); 9044 jccb(Assembler::lessEqual, L_copy_8_chars); 9045 9046 bind(L_copy_8_chars_exit); 9047 subptr(len, 8); 9048 jccb(Assembler::zero, L_done); 9049 } 9050 9051 bind(L_copy_1_char); 9052 load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0)); 9053 testl(tmp5, 0xff00); // check if Unicode char 9054 jccb(Assembler::notZero, L_copy_1_char_exit); 9055 movb(Address(dst, len, Address::times_1, 0), tmp5); 9056 addptr(len, 1); 9057 jccb(Assembler::less, L_copy_1_char); 9058 9059 bind(L_copy_1_char_exit); 9060 addptr(result, len); // len is negative count of not processed elements 9061 9062 bind(L_done); 9063 } 9064 9065 #ifdef _LP64 9066 /** 9067 * Helper for multiply_to_len(). 9068 */ 9069 void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo, Register src1, Register src2) { 9070 addq(dest_lo, src1); 9071 adcq(dest_hi, 0); 9072 addq(dest_lo, src2); 9073 adcq(dest_hi, 0); 9074 } 9075 9076 /** 9077 * Multiply 64 bit by 64 bit first loop. 9078 */ 9079 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart, 9080 Register y, Register y_idx, Register z, 9081 Register carry, Register product, 9082 Register idx, Register kdx) { 9083 // 9084 // jlong carry, x[], y[], z[]; 9085 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) { 9086 // huge_128 product = y[idx] * x[xstart] + carry; 9087 // z[kdx] = (jlong)product; 9088 // carry = (jlong)(product >>> 64); 9089 // } 9090 // z[xstart] = carry; 9091 // 9092 9093 Label L_first_loop, L_first_loop_exit; 9094 Label L_one_x, L_one_y, L_multiply; 9095 9096 decrementl(xstart); 9097 jcc(Assembler::negative, L_one_x); 9098 9099 movq(x_xstart, Address(x, xstart, Address::times_4, 0)); 9100 rorq(x_xstart, 32); // convert big-endian to little-endian 9101 9102 bind(L_first_loop); 9103 decrementl(idx); 9104 jcc(Assembler::negative, L_first_loop_exit); 9105 decrementl(idx); 9106 jcc(Assembler::negative, L_one_y); 9107 movq(y_idx, Address(y, idx, Address::times_4, 0)); 9108 rorq(y_idx, 32); // convert big-endian to little-endian 9109 bind(L_multiply); 9110 movq(product, x_xstart); 9111 mulq(y_idx); // product(rax) * y_idx -> rdx:rax 9112 addq(product, carry); 9113 adcq(rdx, 0); 9114 subl(kdx, 2); 9115 movl(Address(z, kdx, Address::times_4, 4), product); 9116 shrq(product, 32); 9117 movl(Address(z, kdx, Address::times_4, 0), product); 9118 movq(carry, rdx); 9119 jmp(L_first_loop); 9120 9121 bind(L_one_y); 9122 movl(y_idx, Address(y, 0)); 9123 jmp(L_multiply); 9124 9125 bind(L_one_x); 9126 movl(x_xstart, Address(x, 0)); 9127 jmp(L_first_loop); 9128 9129 bind(L_first_loop_exit); 9130 } 9131 9132 /** 9133 * Multiply 64 bit by 64 bit and add 128 bit. 9134 */ 9135 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y, Register z, 9136 Register yz_idx, Register idx, 9137 Register carry, Register product, int offset) { 9138 // huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry; 9139 // z[kdx] = (jlong)product; 9140 9141 movq(yz_idx, Address(y, idx, Address::times_4, offset)); 9142 rorq(yz_idx, 32); // convert big-endian to little-endian 9143 movq(product, x_xstart); 9144 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax) 9145 movq(yz_idx, Address(z, idx, Address::times_4, offset)); 9146 rorq(yz_idx, 32); // convert big-endian to little-endian 9147 9148 add2_with_carry(rdx, product, carry, yz_idx); 9149 9150 movl(Address(z, idx, Address::times_4, offset+4), product); 9151 shrq(product, 32); 9152 movl(Address(z, idx, Address::times_4, offset), product); 9153 9154 } 9155 9156 /** 9157 * Multiply 128 bit by 128 bit. Unrolled inner loop. 9158 */ 9159 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart, Register y, Register z, 9160 Register yz_idx, Register idx, Register jdx, 9161 Register carry, Register product, 9162 Register carry2) { 9163 // jlong carry, x[], y[], z[]; 9164 // int kdx = ystart+1; 9165 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop 9166 // huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry; 9167 // z[kdx+idx+1] = (jlong)product; 9168 // jlong carry2 = (jlong)(product >>> 64); 9169 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry2; 9170 // z[kdx+idx] = (jlong)product; 9171 // carry = (jlong)(product >>> 64); 9172 // } 9173 // idx += 2; 9174 // if (idx > 0) { 9175 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry; 9176 // z[kdx+idx] = (jlong)product; 9177 // carry = (jlong)(product >>> 64); 9178 // } 9179 // 9180 9181 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done; 9182 9183 movl(jdx, idx); 9184 andl(jdx, 0xFFFFFFFC); 9185 shrl(jdx, 2); 9186 9187 bind(L_third_loop); 9188 subl(jdx, 1); 9189 jcc(Assembler::negative, L_third_loop_exit); 9190 subl(idx, 4); 9191 9192 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8); 9193 movq(carry2, rdx); 9194 9195 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0); 9196 movq(carry, rdx); 9197 jmp(L_third_loop); 9198 9199 bind (L_third_loop_exit); 9200 9201 andl (idx, 0x3); 9202 jcc(Assembler::zero, L_post_third_loop_done); 9203 9204 Label L_check_1; 9205 subl(idx, 2); 9206 jcc(Assembler::negative, L_check_1); 9207 9208 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0); 9209 movq(carry, rdx); 9210 9211 bind (L_check_1); 9212 addl (idx, 0x2); 9213 andl (idx, 0x1); 9214 subl(idx, 1); 9215 jcc(Assembler::negative, L_post_third_loop_done); 9216 9217 movl(yz_idx, Address(y, idx, Address::times_4, 0)); 9218 movq(product, x_xstart); 9219 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax) 9220 movl(yz_idx, Address(z, idx, Address::times_4, 0)); 9221 9222 add2_with_carry(rdx, product, yz_idx, carry); 9223 9224 movl(Address(z, idx, Address::times_4, 0), product); 9225 shrq(product, 32); 9226 9227 shlq(rdx, 32); 9228 orq(product, rdx); 9229 movq(carry, product); 9230 9231 bind(L_post_third_loop_done); 9232 } 9233 9234 /** 9235 * Multiply 128 bit by 128 bit using BMI2. Unrolled inner loop. 9236 * 9237 */ 9238 void MacroAssembler::multiply_128_x_128_bmi2_loop(Register y, Register z, 9239 Register carry, Register carry2, 9240 Register idx, Register jdx, 9241 Register yz_idx1, Register yz_idx2, 9242 Register tmp, Register tmp3, Register tmp4) { 9243 assert(UseBMI2Instructions, "should be used only when BMI2 is available"); 9244 9245 // jlong carry, x[], y[], z[]; 9246 // int kdx = ystart+1; 9247 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop 9248 // huge_128 tmp3 = (y[idx+1] * rdx) + z[kdx+idx+1] + carry; 9249 // jlong carry2 = (jlong)(tmp3 >>> 64); 9250 // huge_128 tmp4 = (y[idx] * rdx) + z[kdx+idx] + carry2; 9251 // carry = (jlong)(tmp4 >>> 64); 9252 // z[kdx+idx+1] = (jlong)tmp3; 9253 // z[kdx+idx] = (jlong)tmp4; 9254 // } 9255 // idx += 2; 9256 // if (idx > 0) { 9257 // yz_idx1 = (y[idx] * rdx) + z[kdx+idx] + carry; 9258 // z[kdx+idx] = (jlong)yz_idx1; 9259 // carry = (jlong)(yz_idx1 >>> 64); 9260 // } 9261 // 9262 9263 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done; 9264 9265 movl(jdx, idx); 9266 andl(jdx, 0xFFFFFFFC); 9267 shrl(jdx, 2); 9268 9269 bind(L_third_loop); 9270 subl(jdx, 1); 9271 jcc(Assembler::negative, L_third_loop_exit); 9272 subl(idx, 4); 9273 9274 movq(yz_idx1, Address(y, idx, Address::times_4, 8)); 9275 rorxq(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian 9276 movq(yz_idx2, Address(y, idx, Address::times_4, 0)); 9277 rorxq(yz_idx2, yz_idx2, 32); 9278 9279 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3 9280 mulxq(carry2, tmp, yz_idx2); // yz_idx2 * rdx -> carry2:tmp 9281 9282 movq(yz_idx1, Address(z, idx, Address::times_4, 8)); 9283 rorxq(yz_idx1, yz_idx1, 32); 9284 movq(yz_idx2, Address(z, idx, Address::times_4, 0)); 9285 rorxq(yz_idx2, yz_idx2, 32); 9286 9287 if (VM_Version::supports_adx()) { 9288 adcxq(tmp3, carry); 9289 adoxq(tmp3, yz_idx1); 9290 9291 adcxq(tmp4, tmp); 9292 adoxq(tmp4, yz_idx2); 9293 9294 movl(carry, 0); // does not affect flags 9295 adcxq(carry2, carry); 9296 adoxq(carry2, carry); 9297 } else { 9298 add2_with_carry(tmp4, tmp3, carry, yz_idx1); 9299 add2_with_carry(carry2, tmp4, tmp, yz_idx2); 9300 } 9301 movq(carry, carry2); 9302 9303 movl(Address(z, idx, Address::times_4, 12), tmp3); 9304 shrq(tmp3, 32); 9305 movl(Address(z, idx, Address::times_4, 8), tmp3); 9306 9307 movl(Address(z, idx, Address::times_4, 4), tmp4); 9308 shrq(tmp4, 32); 9309 movl(Address(z, idx, Address::times_4, 0), tmp4); 9310 9311 jmp(L_third_loop); 9312 9313 bind (L_third_loop_exit); 9314 9315 andl (idx, 0x3); 9316 jcc(Assembler::zero, L_post_third_loop_done); 9317 9318 Label L_check_1; 9319 subl(idx, 2); 9320 jcc(Assembler::negative, L_check_1); 9321 9322 movq(yz_idx1, Address(y, idx, Address::times_4, 0)); 9323 rorxq(yz_idx1, yz_idx1, 32); 9324 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3 9325 movq(yz_idx2, Address(z, idx, Address::times_4, 0)); 9326 rorxq(yz_idx2, yz_idx2, 32); 9327 9328 add2_with_carry(tmp4, tmp3, carry, yz_idx2); 9329 9330 movl(Address(z, idx, Address::times_4, 4), tmp3); 9331 shrq(tmp3, 32); 9332 movl(Address(z, idx, Address::times_4, 0), tmp3); 9333 movq(carry, tmp4); 9334 9335 bind (L_check_1); 9336 addl (idx, 0x2); 9337 andl (idx, 0x1); 9338 subl(idx, 1); 9339 jcc(Assembler::negative, L_post_third_loop_done); 9340 movl(tmp4, Address(y, idx, Address::times_4, 0)); 9341 mulxq(carry2, tmp3, tmp4); // tmp4 * rdx -> carry2:tmp3 9342 movl(tmp4, Address(z, idx, Address::times_4, 0)); 9343 9344 add2_with_carry(carry2, tmp3, tmp4, carry); 9345 9346 movl(Address(z, idx, Address::times_4, 0), tmp3); 9347 shrq(tmp3, 32); 9348 9349 shlq(carry2, 32); 9350 orq(tmp3, carry2); 9351 movq(carry, tmp3); 9352 9353 bind(L_post_third_loop_done); 9354 } 9355 9356 /** 9357 * Code for BigInteger::multiplyToLen() instrinsic. 9358 * 9359 * rdi: x 9360 * rax: xlen 9361 * rsi: y 9362 * rcx: ylen 9363 * r8: z 9364 * r11: zlen 9365 * r12: tmp1 9366 * r13: tmp2 9367 * r14: tmp3 9368 * r15: tmp4 9369 * rbx: tmp5 9370 * 9371 */ 9372 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, Register z, Register zlen, 9373 Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5) { 9374 ShortBranchVerifier sbv(this); 9375 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx); 9376 9377 push(tmp1); 9378 push(tmp2); 9379 push(tmp3); 9380 push(tmp4); 9381 push(tmp5); 9382 9383 push(xlen); 9384 push(zlen); 9385 9386 const Register idx = tmp1; 9387 const Register kdx = tmp2; 9388 const Register xstart = tmp3; 9389 9390 const Register y_idx = tmp4; 9391 const Register carry = tmp5; 9392 const Register product = xlen; 9393 const Register x_xstart = zlen; // reuse register 9394 9395 // First Loop. 9396 // 9397 // final static long LONG_MASK = 0xffffffffL; 9398 // int xstart = xlen - 1; 9399 // int ystart = ylen - 1; 9400 // long carry = 0; 9401 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) { 9402 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry; 9403 // z[kdx] = (int)product; 9404 // carry = product >>> 32; 9405 // } 9406 // z[xstart] = (int)carry; 9407 // 9408 9409 movl(idx, ylen); // idx = ylen; 9410 movl(kdx, zlen); // kdx = xlen+ylen; 9411 xorq(carry, carry); // carry = 0; 9412 9413 Label L_done; 9414 9415 movl(xstart, xlen); 9416 decrementl(xstart); 9417 jcc(Assembler::negative, L_done); 9418 9419 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx); 9420 9421 Label L_second_loop; 9422 testl(kdx, kdx); 9423 jcc(Assembler::zero, L_second_loop); 9424 9425 Label L_carry; 9426 subl(kdx, 1); 9427 jcc(Assembler::zero, L_carry); 9428 9429 movl(Address(z, kdx, Address::times_4, 0), carry); 9430 shrq(carry, 32); 9431 subl(kdx, 1); 9432 9433 bind(L_carry); 9434 movl(Address(z, kdx, Address::times_4, 0), carry); 9435 9436 // Second and third (nested) loops. 9437 // 9438 // for (int i = xstart-1; i >= 0; i--) { // Second loop 9439 // carry = 0; 9440 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop 9441 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) + 9442 // (z[k] & LONG_MASK) + carry; 9443 // z[k] = (int)product; 9444 // carry = product >>> 32; 9445 // } 9446 // z[i] = (int)carry; 9447 // } 9448 // 9449 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx 9450 9451 const Register jdx = tmp1; 9452 9453 bind(L_second_loop); 9454 xorl(carry, carry); // carry = 0; 9455 movl(jdx, ylen); // j = ystart+1 9456 9457 subl(xstart, 1); // i = xstart-1; 9458 jcc(Assembler::negative, L_done); 9459 9460 push (z); 9461 9462 Label L_last_x; 9463 lea(z, Address(z, xstart, Address::times_4, 4)); // z = z + k - j 9464 subl(xstart, 1); // i = xstart-1; 9465 jcc(Assembler::negative, L_last_x); 9466 9467 if (UseBMI2Instructions) { 9468 movq(rdx, Address(x, xstart, Address::times_4, 0)); 9469 rorxq(rdx, rdx, 32); // convert big-endian to little-endian 9470 } else { 9471 movq(x_xstart, Address(x, xstart, Address::times_4, 0)); 9472 rorq(x_xstart, 32); // convert big-endian to little-endian 9473 } 9474 9475 Label L_third_loop_prologue; 9476 bind(L_third_loop_prologue); 9477 9478 push (x); 9479 push (xstart); 9480 push (ylen); 9481 9482 9483 if (UseBMI2Instructions) { 9484 multiply_128_x_128_bmi2_loop(y, z, carry, x, jdx, ylen, product, tmp2, x_xstart, tmp3, tmp4); 9485 } else { // !UseBMI2Instructions 9486 multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x); 9487 } 9488 9489 pop(ylen); 9490 pop(xlen); 9491 pop(x); 9492 pop(z); 9493 9494 movl(tmp3, xlen); 9495 addl(tmp3, 1); 9496 movl(Address(z, tmp3, Address::times_4, 0), carry); 9497 subl(tmp3, 1); 9498 jccb(Assembler::negative, L_done); 9499 9500 shrq(carry, 32); 9501 movl(Address(z, tmp3, Address::times_4, 0), carry); 9502 jmp(L_second_loop); 9503 9504 // Next infrequent code is moved outside loops. 9505 bind(L_last_x); 9506 if (UseBMI2Instructions) { 9507 movl(rdx, Address(x, 0)); 9508 } else { 9509 movl(x_xstart, Address(x, 0)); 9510 } 9511 jmp(L_third_loop_prologue); 9512 9513 bind(L_done); 9514 9515 pop(zlen); 9516 pop(xlen); 9517 9518 pop(tmp5); 9519 pop(tmp4); 9520 pop(tmp3); 9521 pop(tmp2); 9522 pop(tmp1); 9523 } 9524 9525 void MacroAssembler::vectorized_mismatch(Register obja, Register objb, Register length, Register log2_array_indxscale, 9526 Register result, Register tmp1, Register tmp2, XMMRegister rymm0, XMMRegister rymm1, XMMRegister rymm2){ 9527 assert(UseSSE42Intrinsics, "SSE4.2 must be enabled."); 9528 Label VECTOR64_LOOP, VECTOR64_TAIL, VECTOR64_NOT_EQUAL, VECTOR32_TAIL; 9529 Label VECTOR32_LOOP, VECTOR16_LOOP, VECTOR8_LOOP, VECTOR4_LOOP; 9530 Label VECTOR16_TAIL, VECTOR8_TAIL, VECTOR4_TAIL; 9531 Label VECTOR32_NOT_EQUAL, VECTOR16_NOT_EQUAL, VECTOR8_NOT_EQUAL, VECTOR4_NOT_EQUAL; 9532 Label SAME_TILL_END, DONE; 9533 Label BYTES_LOOP, BYTES_TAIL, BYTES_NOT_EQUAL; 9534 9535 //scale is in rcx in both Win64 and Unix 9536 ShortBranchVerifier sbv(this); 9537 9538 shlq(length); 9539 xorq(result, result); 9540 9541 if ((UseAVX > 2) && 9542 VM_Version::supports_avx512vlbw()) { 9543 set_vector_masking(); // opening of the stub context for programming mask registers 9544 cmpq(length, 64); 9545 jcc(Assembler::less, VECTOR32_TAIL); 9546 movq(tmp1, length); 9547 andq(tmp1, 0x3F); // tail count 9548 andq(length, ~(0x3F)); //vector count 9549 9550 bind(VECTOR64_LOOP); 9551 // AVX512 code to compare 64 byte vectors. 9552 evmovdqub(rymm0, Address(obja, result), Assembler::AVX_512bit); 9553 evpcmpeqb(k7, rymm0, Address(objb, result), Assembler::AVX_512bit); 9554 kortestql(k7, k7); 9555 jcc(Assembler::aboveEqual, VECTOR64_NOT_EQUAL); // mismatch 9556 addq(result, 64); 9557 subq(length, 64); 9558 jccb(Assembler::notZero, VECTOR64_LOOP); 9559 9560 //bind(VECTOR64_TAIL); 9561 testq(tmp1, tmp1); 9562 jcc(Assembler::zero, SAME_TILL_END); 9563 9564 bind(VECTOR64_TAIL); 9565 // AVX512 code to compare upto 63 byte vectors. 9566 // Save k1 9567 kmovql(k3, k1); 9568 mov64(tmp2, 0xFFFFFFFFFFFFFFFF); 9569 shlxq(tmp2, tmp2, tmp1); 9570 notq(tmp2); 9571 kmovql(k1, tmp2); 9572 9573 evmovdqub(rymm0, k1, Address(obja, result), Assembler::AVX_512bit); 9574 evpcmpeqb(k7, k1, rymm0, Address(objb, result), Assembler::AVX_512bit); 9575 9576 ktestql(k7, k1); 9577 // Restore k1 9578 kmovql(k1, k3); 9579 jcc(Assembler::below, SAME_TILL_END); // not mismatch 9580 9581 bind(VECTOR64_NOT_EQUAL); 9582 kmovql(tmp1, k7); 9583 notq(tmp1); 9584 tzcntq(tmp1, tmp1); 9585 addq(result, tmp1); 9586 shrq(result); 9587 jmp(DONE); 9588 bind(VECTOR32_TAIL); 9589 clear_vector_masking(); // closing of the stub context for programming mask registers 9590 } 9591 9592 cmpq(length, 8); 9593 jcc(Assembler::equal, VECTOR8_LOOP); 9594 jcc(Assembler::less, VECTOR4_TAIL); 9595 9596 if (UseAVX >= 2) { 9597 9598 cmpq(length, 16); 9599 jcc(Assembler::equal, VECTOR16_LOOP); 9600 jcc(Assembler::less, VECTOR8_LOOP); 9601 9602 cmpq(length, 32); 9603 jccb(Assembler::less, VECTOR16_TAIL); 9604 9605 subq(length, 32); 9606 bind(VECTOR32_LOOP); 9607 vmovdqu(rymm0, Address(obja, result)); 9608 vmovdqu(rymm1, Address(objb, result)); 9609 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_256bit); 9610 vptest(rymm2, rymm2); 9611 jcc(Assembler::notZero, VECTOR32_NOT_EQUAL);//mismatch found 9612 addq(result, 32); 9613 subq(length, 32); 9614 jccb(Assembler::greaterEqual, VECTOR32_LOOP); 9615 addq(length, 32); 9616 jcc(Assembler::equal, SAME_TILL_END); 9617 //falling through if less than 32 bytes left //close the branch here. 9618 9619 bind(VECTOR16_TAIL); 9620 cmpq(length, 16); 9621 jccb(Assembler::less, VECTOR8_TAIL); 9622 bind(VECTOR16_LOOP); 9623 movdqu(rymm0, Address(obja, result)); 9624 movdqu(rymm1, Address(objb, result)); 9625 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_128bit); 9626 ptest(rymm2, rymm2); 9627 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found 9628 addq(result, 16); 9629 subq(length, 16); 9630 jcc(Assembler::equal, SAME_TILL_END); 9631 //falling through if less than 16 bytes left 9632 } else {//regular intrinsics 9633 9634 cmpq(length, 16); 9635 jccb(Assembler::less, VECTOR8_TAIL); 9636 9637 subq(length, 16); 9638 bind(VECTOR16_LOOP); 9639 movdqu(rymm0, Address(obja, result)); 9640 movdqu(rymm1, Address(objb, result)); 9641 pxor(rymm0, rymm1); 9642 ptest(rymm0, rymm0); 9643 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found 9644 addq(result, 16); 9645 subq(length, 16); 9646 jccb(Assembler::greaterEqual, VECTOR16_LOOP); 9647 addq(length, 16); 9648 jcc(Assembler::equal, SAME_TILL_END); 9649 //falling through if less than 16 bytes left 9650 } 9651 9652 bind(VECTOR8_TAIL); 9653 cmpq(length, 8); 9654 jccb(Assembler::less, VECTOR4_TAIL); 9655 bind(VECTOR8_LOOP); 9656 movq(tmp1, Address(obja, result)); 9657 movq(tmp2, Address(objb, result)); 9658 xorq(tmp1, tmp2); 9659 testq(tmp1, tmp1); 9660 jcc(Assembler::notZero, VECTOR8_NOT_EQUAL);//mismatch found 9661 addq(result, 8); 9662 subq(length, 8); 9663 jcc(Assembler::equal, SAME_TILL_END); 9664 //falling through if less than 8 bytes left 9665 9666 bind(VECTOR4_TAIL); 9667 cmpq(length, 4); 9668 jccb(Assembler::less, BYTES_TAIL); 9669 bind(VECTOR4_LOOP); 9670 movl(tmp1, Address(obja, result)); 9671 xorl(tmp1, Address(objb, result)); 9672 testl(tmp1, tmp1); 9673 jcc(Assembler::notZero, VECTOR4_NOT_EQUAL);//mismatch found 9674 addq(result, 4); 9675 subq(length, 4); 9676 jcc(Assembler::equal, SAME_TILL_END); 9677 //falling through if less than 4 bytes left 9678 9679 bind(BYTES_TAIL); 9680 bind(BYTES_LOOP); 9681 load_unsigned_byte(tmp1, Address(obja, result)); 9682 load_unsigned_byte(tmp2, Address(objb, result)); 9683 xorl(tmp1, tmp2); 9684 testl(tmp1, tmp1); 9685 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 9686 decq(length); 9687 jccb(Assembler::zero, SAME_TILL_END); 9688 incq(result); 9689 load_unsigned_byte(tmp1, Address(obja, result)); 9690 load_unsigned_byte(tmp2, Address(objb, result)); 9691 xorl(tmp1, tmp2); 9692 testl(tmp1, tmp1); 9693 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 9694 decq(length); 9695 jccb(Assembler::zero, SAME_TILL_END); 9696 incq(result); 9697 load_unsigned_byte(tmp1, Address(obja, result)); 9698 load_unsigned_byte(tmp2, Address(objb, result)); 9699 xorl(tmp1, tmp2); 9700 testl(tmp1, tmp1); 9701 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 9702 jmpb(SAME_TILL_END); 9703 9704 if (UseAVX >= 2) { 9705 bind(VECTOR32_NOT_EQUAL); 9706 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_256bit); 9707 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_256bit); 9708 vpxor(rymm0, rymm0, rymm2, Assembler::AVX_256bit); 9709 vpmovmskb(tmp1, rymm0); 9710 bsfq(tmp1, tmp1); 9711 addq(result, tmp1); 9712 shrq(result); 9713 jmpb(DONE); 9714 } 9715 9716 bind(VECTOR16_NOT_EQUAL); 9717 if (UseAVX >= 2) { 9718 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_128bit); 9719 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_128bit); 9720 pxor(rymm0, rymm2); 9721 } else { 9722 pcmpeqb(rymm2, rymm2); 9723 pxor(rymm0, rymm1); 9724 pcmpeqb(rymm0, rymm1); 9725 pxor(rymm0, rymm2); 9726 } 9727 pmovmskb(tmp1, rymm0); 9728 bsfq(tmp1, tmp1); 9729 addq(result, tmp1); 9730 shrq(result); 9731 jmpb(DONE); 9732 9733 bind(VECTOR8_NOT_EQUAL); 9734 bind(VECTOR4_NOT_EQUAL); 9735 bsfq(tmp1, tmp1); 9736 shrq(tmp1, 3); 9737 addq(result, tmp1); 9738 bind(BYTES_NOT_EQUAL); 9739 shrq(result); 9740 jmpb(DONE); 9741 9742 bind(SAME_TILL_END); 9743 mov64(result, -1); 9744 9745 bind(DONE); 9746 } 9747 9748 //Helper functions for square_to_len() 9749 9750 /** 9751 * Store the squares of x[], right shifted one bit (divided by 2) into z[] 9752 * Preserves x and z and modifies rest of the registers. 9753 */ 9754 void MacroAssembler::square_rshift(Register x, Register xlen, Register z, Register tmp1, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 9755 // Perform square and right shift by 1 9756 // Handle odd xlen case first, then for even xlen do the following 9757 // jlong carry = 0; 9758 // for (int j=0, i=0; j < xlen; j+=2, i+=4) { 9759 // huge_128 product = x[j:j+1] * x[j:j+1]; 9760 // z[i:i+1] = (carry << 63) | (jlong)(product >>> 65); 9761 // z[i+2:i+3] = (jlong)(product >>> 1); 9762 // carry = (jlong)product; 9763 // } 9764 9765 xorq(tmp5, tmp5); // carry 9766 xorq(rdxReg, rdxReg); 9767 xorl(tmp1, tmp1); // index for x 9768 xorl(tmp4, tmp4); // index for z 9769 9770 Label L_first_loop, L_first_loop_exit; 9771 9772 testl(xlen, 1); 9773 jccb(Assembler::zero, L_first_loop); //jump if xlen is even 9774 9775 // Square and right shift by 1 the odd element using 32 bit multiply 9776 movl(raxReg, Address(x, tmp1, Address::times_4, 0)); 9777 imulq(raxReg, raxReg); 9778 shrq(raxReg, 1); 9779 adcq(tmp5, 0); 9780 movq(Address(z, tmp4, Address::times_4, 0), raxReg); 9781 incrementl(tmp1); 9782 addl(tmp4, 2); 9783 9784 // Square and right shift by 1 the rest using 64 bit multiply 9785 bind(L_first_loop); 9786 cmpptr(tmp1, xlen); 9787 jccb(Assembler::equal, L_first_loop_exit); 9788 9789 // Square 9790 movq(raxReg, Address(x, tmp1, Address::times_4, 0)); 9791 rorq(raxReg, 32); // convert big-endian to little-endian 9792 mulq(raxReg); // 64-bit multiply rax * rax -> rdx:rax 9793 9794 // Right shift by 1 and save carry 9795 shrq(tmp5, 1); // rdx:rax:tmp5 = (tmp5:rdx:rax) >>> 1 9796 rcrq(rdxReg, 1); 9797 rcrq(raxReg, 1); 9798 adcq(tmp5, 0); 9799 9800 // Store result in z 9801 movq(Address(z, tmp4, Address::times_4, 0), rdxReg); 9802 movq(Address(z, tmp4, Address::times_4, 8), raxReg); 9803 9804 // Update indices for x and z 9805 addl(tmp1, 2); 9806 addl(tmp4, 4); 9807 jmp(L_first_loop); 9808 9809 bind(L_first_loop_exit); 9810 } 9811 9812 9813 /** 9814 * Perform the following multiply add operation using BMI2 instructions 9815 * carry:sum = sum + op1*op2 + carry 9816 * op2 should be in rdx 9817 * op2 is preserved, all other registers are modified 9818 */ 9819 void MacroAssembler::multiply_add_64_bmi2(Register sum, Register op1, Register op2, Register carry, Register tmp2) { 9820 // assert op2 is rdx 9821 mulxq(tmp2, op1, op1); // op1 * op2 -> tmp2:op1 9822 addq(sum, carry); 9823 adcq(tmp2, 0); 9824 addq(sum, op1); 9825 adcq(tmp2, 0); 9826 movq(carry, tmp2); 9827 } 9828 9829 /** 9830 * Perform the following multiply add operation: 9831 * carry:sum = sum + op1*op2 + carry 9832 * Preserves op1, op2 and modifies rest of registers 9833 */ 9834 void MacroAssembler::multiply_add_64(Register sum, Register op1, Register op2, Register carry, Register rdxReg, Register raxReg) { 9835 // rdx:rax = op1 * op2 9836 movq(raxReg, op2); 9837 mulq(op1); 9838 9839 // rdx:rax = sum + carry + rdx:rax 9840 addq(sum, carry); 9841 adcq(rdxReg, 0); 9842 addq(sum, raxReg); 9843 adcq(rdxReg, 0); 9844 9845 // carry:sum = rdx:sum 9846 movq(carry, rdxReg); 9847 } 9848 9849 /** 9850 * Add 64 bit long carry into z[] with carry propogation. 9851 * Preserves z and carry register values and modifies rest of registers. 9852 * 9853 */ 9854 void MacroAssembler::add_one_64(Register z, Register zlen, Register carry, Register tmp1) { 9855 Label L_fourth_loop, L_fourth_loop_exit; 9856 9857 movl(tmp1, 1); 9858 subl(zlen, 2); 9859 addq(Address(z, zlen, Address::times_4, 0), carry); 9860 9861 bind(L_fourth_loop); 9862 jccb(Assembler::carryClear, L_fourth_loop_exit); 9863 subl(zlen, 2); 9864 jccb(Assembler::negative, L_fourth_loop_exit); 9865 addq(Address(z, zlen, Address::times_4, 0), tmp1); 9866 jmp(L_fourth_loop); 9867 bind(L_fourth_loop_exit); 9868 } 9869 9870 /** 9871 * Shift z[] left by 1 bit. 9872 * Preserves x, len, z and zlen registers and modifies rest of the registers. 9873 * 9874 */ 9875 void MacroAssembler::lshift_by_1(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4) { 9876 9877 Label L_fifth_loop, L_fifth_loop_exit; 9878 9879 // Fifth loop 9880 // Perform primitiveLeftShift(z, zlen, 1) 9881 9882 const Register prev_carry = tmp1; 9883 const Register new_carry = tmp4; 9884 const Register value = tmp2; 9885 const Register zidx = tmp3; 9886 9887 // int zidx, carry; 9888 // long value; 9889 // carry = 0; 9890 // for (zidx = zlen-2; zidx >=0; zidx -= 2) { 9891 // (carry:value) = (z[i] << 1) | carry ; 9892 // z[i] = value; 9893 // } 9894 9895 movl(zidx, zlen); 9896 xorl(prev_carry, prev_carry); // clear carry flag and prev_carry register 9897 9898 bind(L_fifth_loop); 9899 decl(zidx); // Use decl to preserve carry flag 9900 decl(zidx); 9901 jccb(Assembler::negative, L_fifth_loop_exit); 9902 9903 if (UseBMI2Instructions) { 9904 movq(value, Address(z, zidx, Address::times_4, 0)); 9905 rclq(value, 1); 9906 rorxq(value, value, 32); 9907 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form 9908 } 9909 else { 9910 // clear new_carry 9911 xorl(new_carry, new_carry); 9912 9913 // Shift z[i] by 1, or in previous carry and save new carry 9914 movq(value, Address(z, zidx, Address::times_4, 0)); 9915 shlq(value, 1); 9916 adcl(new_carry, 0); 9917 9918 orq(value, prev_carry); 9919 rorq(value, 0x20); 9920 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form 9921 9922 // Set previous carry = new carry 9923 movl(prev_carry, new_carry); 9924 } 9925 jmp(L_fifth_loop); 9926 9927 bind(L_fifth_loop_exit); 9928 } 9929 9930 9931 /** 9932 * Code for BigInteger::squareToLen() intrinsic 9933 * 9934 * rdi: x 9935 * rsi: len 9936 * r8: z 9937 * rcx: zlen 9938 * r12: tmp1 9939 * r13: tmp2 9940 * r14: tmp3 9941 * r15: tmp4 9942 * rbx: tmp5 9943 * 9944 */ 9945 void MacroAssembler::square_to_len(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 9946 9947 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; 9948 push(tmp1); 9949 push(tmp2); 9950 push(tmp3); 9951 push(tmp4); 9952 push(tmp5); 9953 9954 // First loop 9955 // Store the squares, right shifted one bit (i.e., divided by 2). 9956 square_rshift(x, len, z, tmp1, tmp3, tmp4, tmp5, rdxReg, raxReg); 9957 9958 // Add in off-diagonal sums. 9959 // 9960 // Second, third (nested) and fourth loops. 9961 // zlen +=2; 9962 // for (int xidx=len-2,zidx=zlen-4; xidx > 0; xidx-=2,zidx-=4) { 9963 // carry = 0; 9964 // long op2 = x[xidx:xidx+1]; 9965 // for (int j=xidx-2,k=zidx; j >= 0; j-=2) { 9966 // k -= 2; 9967 // long op1 = x[j:j+1]; 9968 // long sum = z[k:k+1]; 9969 // carry:sum = multiply_add_64(sum, op1, op2, carry, tmp_regs); 9970 // z[k:k+1] = sum; 9971 // } 9972 // add_one_64(z, k, carry, tmp_regs); 9973 // } 9974 9975 const Register carry = tmp5; 9976 const Register sum = tmp3; 9977 const Register op1 = tmp4; 9978 Register op2 = tmp2; 9979 9980 push(zlen); 9981 push(len); 9982 addl(zlen,2); 9983 bind(L_second_loop); 9984 xorq(carry, carry); 9985 subl(zlen, 4); 9986 subl(len, 2); 9987 push(zlen); 9988 push(len); 9989 cmpl(len, 0); 9990 jccb(Assembler::lessEqual, L_second_loop_exit); 9991 9992 // Multiply an array by one 64 bit long. 9993 if (UseBMI2Instructions) { 9994 op2 = rdxReg; 9995 movq(op2, Address(x, len, Address::times_4, 0)); 9996 rorxq(op2, op2, 32); 9997 } 9998 else { 9999 movq(op2, Address(x, len, Address::times_4, 0)); 10000 rorq(op2, 32); 10001 } 10002 10003 bind(L_third_loop); 10004 decrementl(len); 10005 jccb(Assembler::negative, L_third_loop_exit); 10006 decrementl(len); 10007 jccb(Assembler::negative, L_last_x); 10008 10009 movq(op1, Address(x, len, Address::times_4, 0)); 10010 rorq(op1, 32); 10011 10012 bind(L_multiply); 10013 subl(zlen, 2); 10014 movq(sum, Address(z, zlen, Address::times_4, 0)); 10015 10016 // Multiply 64 bit by 64 bit and add 64 bits lower half and upper 64 bits as carry. 10017 if (UseBMI2Instructions) { 10018 multiply_add_64_bmi2(sum, op1, op2, carry, tmp2); 10019 } 10020 else { 10021 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 10022 } 10023 10024 movq(Address(z, zlen, Address::times_4, 0), sum); 10025 10026 jmp(L_third_loop); 10027 bind(L_third_loop_exit); 10028 10029 // Fourth loop 10030 // Add 64 bit long carry into z with carry propogation. 10031 // Uses offsetted zlen. 10032 add_one_64(z, zlen, carry, tmp1); 10033 10034 pop(len); 10035 pop(zlen); 10036 jmp(L_second_loop); 10037 10038 // Next infrequent code is moved outside loops. 10039 bind(L_last_x); 10040 movl(op1, Address(x, 0)); 10041 jmp(L_multiply); 10042 10043 bind(L_second_loop_exit); 10044 pop(len); 10045 pop(zlen); 10046 pop(len); 10047 pop(zlen); 10048 10049 // Fifth loop 10050 // Shift z left 1 bit. 10051 lshift_by_1(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4); 10052 10053 // z[zlen-1] |= x[len-1] & 1; 10054 movl(tmp3, Address(x, len, Address::times_4, -4)); 10055 andl(tmp3, 1); 10056 orl(Address(z, zlen, Address::times_4, -4), tmp3); 10057 10058 pop(tmp5); 10059 pop(tmp4); 10060 pop(tmp3); 10061 pop(tmp2); 10062 pop(tmp1); 10063 } 10064 10065 /** 10066 * Helper function for mul_add() 10067 * Multiply the in[] by int k and add to out[] starting at offset offs using 10068 * 128 bit by 32 bit multiply and return the carry in tmp5. 10069 * Only quad int aligned length of in[] is operated on in this function. 10070 * k is in rdxReg for BMI2Instructions, for others it is in tmp2. 10071 * This function preserves out, in and k registers. 10072 * len and offset point to the appropriate index in "in" & "out" correspondingly 10073 * tmp5 has the carry. 10074 * other registers are temporary and are modified. 10075 * 10076 */ 10077 void MacroAssembler::mul_add_128_x_32_loop(Register out, Register in, 10078 Register offset, Register len, Register tmp1, Register tmp2, Register tmp3, 10079 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 10080 10081 Label L_first_loop, L_first_loop_exit; 10082 10083 movl(tmp1, len); 10084 shrl(tmp1, 2); 10085 10086 bind(L_first_loop); 10087 subl(tmp1, 1); 10088 jccb(Assembler::negative, L_first_loop_exit); 10089 10090 subl(len, 4); 10091 subl(offset, 4); 10092 10093 Register op2 = tmp2; 10094 const Register sum = tmp3; 10095 const Register op1 = tmp4; 10096 const Register carry = tmp5; 10097 10098 if (UseBMI2Instructions) { 10099 op2 = rdxReg; 10100 } 10101 10102 movq(op1, Address(in, len, Address::times_4, 8)); 10103 rorq(op1, 32); 10104 movq(sum, Address(out, offset, Address::times_4, 8)); 10105 rorq(sum, 32); 10106 if (UseBMI2Instructions) { 10107 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 10108 } 10109 else { 10110 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 10111 } 10112 // Store back in big endian from little endian 10113 rorq(sum, 0x20); 10114 movq(Address(out, offset, Address::times_4, 8), sum); 10115 10116 movq(op1, Address(in, len, Address::times_4, 0)); 10117 rorq(op1, 32); 10118 movq(sum, Address(out, offset, Address::times_4, 0)); 10119 rorq(sum, 32); 10120 if (UseBMI2Instructions) { 10121 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 10122 } 10123 else { 10124 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 10125 } 10126 // Store back in big endian from little endian 10127 rorq(sum, 0x20); 10128 movq(Address(out, offset, Address::times_4, 0), sum); 10129 10130 jmp(L_first_loop); 10131 bind(L_first_loop_exit); 10132 } 10133 10134 /** 10135 * Code for BigInteger::mulAdd() intrinsic 10136 * 10137 * rdi: out 10138 * rsi: in 10139 * r11: offs (out.length - offset) 10140 * rcx: len 10141 * r8: k 10142 * r12: tmp1 10143 * r13: tmp2 10144 * r14: tmp3 10145 * r15: tmp4 10146 * rbx: tmp5 10147 * Multiply the in[] by word k and add to out[], return the carry in rax 10148 */ 10149 void MacroAssembler::mul_add(Register out, Register in, Register offs, 10150 Register len, Register k, Register tmp1, Register tmp2, Register tmp3, 10151 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 10152 10153 Label L_carry, L_last_in, L_done; 10154 10155 // carry = 0; 10156 // for (int j=len-1; j >= 0; j--) { 10157 // long product = (in[j] & LONG_MASK) * kLong + 10158 // (out[offs] & LONG_MASK) + carry; 10159 // out[offs--] = (int)product; 10160 // carry = product >>> 32; 10161 // } 10162 // 10163 push(tmp1); 10164 push(tmp2); 10165 push(tmp3); 10166 push(tmp4); 10167 push(tmp5); 10168 10169 Register op2 = tmp2; 10170 const Register sum = tmp3; 10171 const Register op1 = tmp4; 10172 const Register carry = tmp5; 10173 10174 if (UseBMI2Instructions) { 10175 op2 = rdxReg; 10176 movl(op2, k); 10177 } 10178 else { 10179 movl(op2, k); 10180 } 10181 10182 xorq(carry, carry); 10183 10184 //First loop 10185 10186 //Multiply in[] by k in a 4 way unrolled loop using 128 bit by 32 bit multiply 10187 //The carry is in tmp5 10188 mul_add_128_x_32_loop(out, in, offs, len, tmp1, tmp2, tmp3, tmp4, tmp5, rdxReg, raxReg); 10189 10190 //Multiply the trailing in[] entry using 64 bit by 32 bit, if any 10191 decrementl(len); 10192 jccb(Assembler::negative, L_carry); 10193 decrementl(len); 10194 jccb(Assembler::negative, L_last_in); 10195 10196 movq(op1, Address(in, len, Address::times_4, 0)); 10197 rorq(op1, 32); 10198 10199 subl(offs, 2); 10200 movq(sum, Address(out, offs, Address::times_4, 0)); 10201 rorq(sum, 32); 10202 10203 if (UseBMI2Instructions) { 10204 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 10205 } 10206 else { 10207 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 10208 } 10209 10210 // Store back in big endian from little endian 10211 rorq(sum, 0x20); 10212 movq(Address(out, offs, Address::times_4, 0), sum); 10213 10214 testl(len, len); 10215 jccb(Assembler::zero, L_carry); 10216 10217 //Multiply the last in[] entry, if any 10218 bind(L_last_in); 10219 movl(op1, Address(in, 0)); 10220 movl(sum, Address(out, offs, Address::times_4, -4)); 10221 10222 movl(raxReg, k); 10223 mull(op1); //tmp4 * eax -> edx:eax 10224 addl(sum, carry); 10225 adcl(rdxReg, 0); 10226 addl(sum, raxReg); 10227 adcl(rdxReg, 0); 10228 movl(carry, rdxReg); 10229 10230 movl(Address(out, offs, Address::times_4, -4), sum); 10231 10232 bind(L_carry); 10233 //return tmp5/carry as carry in rax 10234 movl(rax, carry); 10235 10236 bind(L_done); 10237 pop(tmp5); 10238 pop(tmp4); 10239 pop(tmp3); 10240 pop(tmp2); 10241 pop(tmp1); 10242 } 10243 #endif 10244 10245 /** 10246 * Emits code to update CRC-32 with a byte value according to constants in table 10247 * 10248 * @param [in,out]crc Register containing the crc. 10249 * @param [in]val Register containing the byte to fold into the CRC. 10250 * @param [in]table Register containing the table of crc constants. 10251 * 10252 * uint32_t crc; 10253 * val = crc_table[(val ^ crc) & 0xFF]; 10254 * crc = val ^ (crc >> 8); 10255 * 10256 */ 10257 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) { 10258 xorl(val, crc); 10259 andl(val, 0xFF); 10260 shrl(crc, 8); // unsigned shift 10261 xorl(crc, Address(table, val, Address::times_4, 0)); 10262 } 10263 10264 /** 10265 * Fold 128-bit data chunk 10266 */ 10267 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) { 10268 if (UseAVX > 0) { 10269 vpclmulhdq(xtmp, xK, xcrc); // [123:64] 10270 vpclmulldq(xcrc, xK, xcrc); // [63:0] 10271 vpxor(xcrc, xcrc, Address(buf, offset), 0 /* vector_len */); 10272 pxor(xcrc, xtmp); 10273 } else { 10274 movdqa(xtmp, xcrc); 10275 pclmulhdq(xtmp, xK); // [123:64] 10276 pclmulldq(xcrc, xK); // [63:0] 10277 pxor(xcrc, xtmp); 10278 movdqu(xtmp, Address(buf, offset)); 10279 pxor(xcrc, xtmp); 10280 } 10281 } 10282 10283 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) { 10284 if (UseAVX > 0) { 10285 vpclmulhdq(xtmp, xK, xcrc); 10286 vpclmulldq(xcrc, xK, xcrc); 10287 pxor(xcrc, xbuf); 10288 pxor(xcrc, xtmp); 10289 } else { 10290 movdqa(xtmp, xcrc); 10291 pclmulhdq(xtmp, xK); 10292 pclmulldq(xcrc, xK); 10293 pxor(xcrc, xbuf); 10294 pxor(xcrc, xtmp); 10295 } 10296 } 10297 10298 /** 10299 * 8-bit folds to compute 32-bit CRC 10300 * 10301 * uint64_t xcrc; 10302 * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8); 10303 */ 10304 void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) { 10305 movdl(tmp, xcrc); 10306 andl(tmp, 0xFF); 10307 movdl(xtmp, Address(table, tmp, Address::times_4, 0)); 10308 psrldq(xcrc, 1); // unsigned shift one byte 10309 pxor(xcrc, xtmp); 10310 } 10311 10312 /** 10313 * uint32_t crc; 10314 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8); 10315 */ 10316 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) { 10317 movl(tmp, crc); 10318 andl(tmp, 0xFF); 10319 shrl(crc, 8); 10320 xorl(crc, Address(table, tmp, Address::times_4, 0)); 10321 } 10322 10323 /** 10324 * @param crc register containing existing CRC (32-bit) 10325 * @param buf register pointing to input byte buffer (byte*) 10326 * @param len register containing number of bytes 10327 * @param table register that will contain address of CRC table 10328 * @param tmp scratch register 10329 */ 10330 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) { 10331 assert_different_registers(crc, buf, len, table, tmp, rax); 10332 10333 Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned; 10334 Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop; 10335 10336 // For EVEX with VL and BW, provide a standard mask, VL = 128 will guide the merge 10337 // context for the registers used, where all instructions below are using 128-bit mode 10338 // On EVEX without VL and BW, these instructions will all be AVX. 10339 if (VM_Version::supports_avx512vlbw()) { 10340 movl(tmp, 0xffff); 10341 kmovwl(k1, tmp); 10342 } 10343 10344 lea(table, ExternalAddress(StubRoutines::crc_table_addr())); 10345 notl(crc); // ~crc 10346 cmpl(len, 16); 10347 jcc(Assembler::less, L_tail); 10348 10349 // Align buffer to 16 bytes 10350 movl(tmp, buf); 10351 andl(tmp, 0xF); 10352 jccb(Assembler::zero, L_aligned); 10353 subl(tmp, 16); 10354 addl(len, tmp); 10355 10356 align(4); 10357 BIND(L_align_loop); 10358 movsbl(rax, Address(buf, 0)); // load byte with sign extension 10359 update_byte_crc32(crc, rax, table); 10360 increment(buf); 10361 incrementl(tmp); 10362 jccb(Assembler::less, L_align_loop); 10363 10364 BIND(L_aligned); 10365 movl(tmp, len); // save 10366 shrl(len, 4); 10367 jcc(Assembler::zero, L_tail_restore); 10368 10369 // Fold crc into first bytes of vector 10370 movdqa(xmm1, Address(buf, 0)); 10371 movdl(rax, xmm1); 10372 xorl(crc, rax); 10373 if (VM_Version::supports_sse4_1()) { 10374 pinsrd(xmm1, crc, 0); 10375 } else { 10376 pinsrw(xmm1, crc, 0); 10377 shrl(crc, 16); 10378 pinsrw(xmm1, crc, 1); 10379 } 10380 addptr(buf, 16); 10381 subl(len, 4); // len > 0 10382 jcc(Assembler::less, L_fold_tail); 10383 10384 movdqa(xmm2, Address(buf, 0)); 10385 movdqa(xmm3, Address(buf, 16)); 10386 movdqa(xmm4, Address(buf, 32)); 10387 addptr(buf, 48); 10388 subl(len, 3); 10389 jcc(Assembler::lessEqual, L_fold_512b); 10390 10391 // Fold total 512 bits of polynomial on each iteration, 10392 // 128 bits per each of 4 parallel streams. 10393 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32)); 10394 10395 align(32); 10396 BIND(L_fold_512b_loop); 10397 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0); 10398 fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16); 10399 fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32); 10400 fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48); 10401 addptr(buf, 64); 10402 subl(len, 4); 10403 jcc(Assembler::greater, L_fold_512b_loop); 10404 10405 // Fold 512 bits to 128 bits. 10406 BIND(L_fold_512b); 10407 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16)); 10408 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2); 10409 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3); 10410 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4); 10411 10412 // Fold the rest of 128 bits data chunks 10413 BIND(L_fold_tail); 10414 addl(len, 3); 10415 jccb(Assembler::lessEqual, L_fold_128b); 10416 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16)); 10417 10418 BIND(L_fold_tail_loop); 10419 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0); 10420 addptr(buf, 16); 10421 decrementl(len); 10422 jccb(Assembler::greater, L_fold_tail_loop); 10423 10424 // Fold 128 bits in xmm1 down into 32 bits in crc register. 10425 BIND(L_fold_128b); 10426 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr())); 10427 if (UseAVX > 0) { 10428 vpclmulqdq(xmm2, xmm0, xmm1, 0x1); 10429 vpand(xmm3, xmm0, xmm2, 0 /* vector_len */); 10430 vpclmulqdq(xmm0, xmm0, xmm3, 0x1); 10431 } else { 10432 movdqa(xmm2, xmm0); 10433 pclmulqdq(xmm2, xmm1, 0x1); 10434 movdqa(xmm3, xmm0); 10435 pand(xmm3, xmm2); 10436 pclmulqdq(xmm0, xmm3, 0x1); 10437 } 10438 psrldq(xmm1, 8); 10439 psrldq(xmm2, 4); 10440 pxor(xmm0, xmm1); 10441 pxor(xmm0, xmm2); 10442 10443 // 8 8-bit folds to compute 32-bit CRC. 10444 for (int j = 0; j < 4; j++) { 10445 fold_8bit_crc32(xmm0, table, xmm1, rax); 10446 } 10447 movdl(crc, xmm0); // mov 32 bits to general register 10448 for (int j = 0; j < 4; j++) { 10449 fold_8bit_crc32(crc, table, rax); 10450 } 10451 10452 BIND(L_tail_restore); 10453 movl(len, tmp); // restore 10454 BIND(L_tail); 10455 andl(len, 0xf); 10456 jccb(Assembler::zero, L_exit); 10457 10458 // Fold the rest of bytes 10459 align(4); 10460 BIND(L_tail_loop); 10461 movsbl(rax, Address(buf, 0)); // load byte with sign extension 10462 update_byte_crc32(crc, rax, table); 10463 increment(buf); 10464 decrementl(len); 10465 jccb(Assembler::greater, L_tail_loop); 10466 10467 BIND(L_exit); 10468 notl(crc); // ~c 10469 } 10470 10471 #ifdef _LP64 10472 // S. Gueron / Information Processing Letters 112 (2012) 184 10473 // Algorithm 4: Computing carry-less multiplication using a precomputed lookup table. 10474 // Input: A 32 bit value B = [byte3, byte2, byte1, byte0]. 10475 // Output: the 64-bit carry-less product of B * CONST 10476 void MacroAssembler::crc32c_ipl_alg4(Register in, uint32_t n, 10477 Register tmp1, Register tmp2, Register tmp3) { 10478 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr())); 10479 if (n > 0) { 10480 addq(tmp3, n * 256 * 8); 10481 } 10482 // Q1 = TABLEExt[n][B & 0xFF]; 10483 movl(tmp1, in); 10484 andl(tmp1, 0x000000FF); 10485 shll(tmp1, 3); 10486 addq(tmp1, tmp3); 10487 movq(tmp1, Address(tmp1, 0)); 10488 10489 // Q2 = TABLEExt[n][B >> 8 & 0xFF]; 10490 movl(tmp2, in); 10491 shrl(tmp2, 8); 10492 andl(tmp2, 0x000000FF); 10493 shll(tmp2, 3); 10494 addq(tmp2, tmp3); 10495 movq(tmp2, Address(tmp2, 0)); 10496 10497 shlq(tmp2, 8); 10498 xorq(tmp1, tmp2); 10499 10500 // Q3 = TABLEExt[n][B >> 16 & 0xFF]; 10501 movl(tmp2, in); 10502 shrl(tmp2, 16); 10503 andl(tmp2, 0x000000FF); 10504 shll(tmp2, 3); 10505 addq(tmp2, tmp3); 10506 movq(tmp2, Address(tmp2, 0)); 10507 10508 shlq(tmp2, 16); 10509 xorq(tmp1, tmp2); 10510 10511 // Q4 = TABLEExt[n][B >> 24 & 0xFF]; 10512 shrl(in, 24); 10513 andl(in, 0x000000FF); 10514 shll(in, 3); 10515 addq(in, tmp3); 10516 movq(in, Address(in, 0)); 10517 10518 shlq(in, 24); 10519 xorq(in, tmp1); 10520 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24; 10521 } 10522 10523 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1, 10524 Register in_out, 10525 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported, 10526 XMMRegister w_xtmp2, 10527 Register tmp1, 10528 Register n_tmp2, Register n_tmp3) { 10529 if (is_pclmulqdq_supported) { 10530 movdl(w_xtmp1, in_out); // modified blindly 10531 10532 movl(tmp1, const_or_pre_comp_const_index); 10533 movdl(w_xtmp2, tmp1); 10534 pclmulqdq(w_xtmp1, w_xtmp2, 0); 10535 10536 movdq(in_out, w_xtmp1); 10537 } else { 10538 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3); 10539 } 10540 } 10541 10542 // Recombination Alternative 2: No bit-reflections 10543 // T1 = (CRC_A * U1) << 1 10544 // T2 = (CRC_B * U2) << 1 10545 // C1 = T1 >> 32 10546 // C2 = T2 >> 32 10547 // T1 = T1 & 0xFFFFFFFF 10548 // T2 = T2 & 0xFFFFFFFF 10549 // T1 = CRC32(0, T1) 10550 // T2 = CRC32(0, T2) 10551 // C1 = C1 ^ T1 10552 // C2 = C2 ^ T2 10553 // CRC = C1 ^ C2 ^ CRC_C 10554 void MacroAssembler::crc32c_rec_alt2(uint32_t const_or_pre_comp_const_index_u1, uint32_t const_or_pre_comp_const_index_u2, bool is_pclmulqdq_supported, Register in_out, Register in1, Register in2, 10555 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10556 Register tmp1, Register tmp2, 10557 Register n_tmp3) { 10558 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10559 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10560 shlq(in_out, 1); 10561 movl(tmp1, in_out); 10562 shrq(in_out, 32); 10563 xorl(tmp2, tmp2); 10564 crc32(tmp2, tmp1, 4); 10565 xorl(in_out, tmp2); // we don't care about upper 32 bit contents here 10566 shlq(in1, 1); 10567 movl(tmp1, in1); 10568 shrq(in1, 32); 10569 xorl(tmp2, tmp2); 10570 crc32(tmp2, tmp1, 4); 10571 xorl(in1, tmp2); 10572 xorl(in_out, in1); 10573 xorl(in_out, in2); 10574 } 10575 10576 // Set N to predefined value 10577 // Subtract from a lenght of a buffer 10578 // execute in a loop: 10579 // CRC_A = 0xFFFFFFFF, CRC_B = 0, CRC_C = 0 10580 // for i = 1 to N do 10581 // CRC_A = CRC32(CRC_A, A[i]) 10582 // CRC_B = CRC32(CRC_B, B[i]) 10583 // CRC_C = CRC32(CRC_C, C[i]) 10584 // end for 10585 // Recombine 10586 void MacroAssembler::crc32c_proc_chunk(uint32_t size, uint32_t const_or_pre_comp_const_index_u1, uint32_t const_or_pre_comp_const_index_u2, bool is_pclmulqdq_supported, 10587 Register in_out1, Register in_out2, Register in_out3, 10588 Register tmp1, Register tmp2, Register tmp3, 10589 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10590 Register tmp4, Register tmp5, 10591 Register n_tmp6) { 10592 Label L_processPartitions; 10593 Label L_processPartition; 10594 Label L_exit; 10595 10596 bind(L_processPartitions); 10597 cmpl(in_out1, 3 * size); 10598 jcc(Assembler::less, L_exit); 10599 xorl(tmp1, tmp1); 10600 xorl(tmp2, tmp2); 10601 movq(tmp3, in_out2); 10602 addq(tmp3, size); 10603 10604 bind(L_processPartition); 10605 crc32(in_out3, Address(in_out2, 0), 8); 10606 crc32(tmp1, Address(in_out2, size), 8); 10607 crc32(tmp2, Address(in_out2, size * 2), 8); 10608 addq(in_out2, 8); 10609 cmpq(in_out2, tmp3); 10610 jcc(Assembler::less, L_processPartition); 10611 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2, 10612 w_xtmp1, w_xtmp2, w_xtmp3, 10613 tmp4, tmp5, 10614 n_tmp6); 10615 addq(in_out2, 2 * size); 10616 subl(in_out1, 3 * size); 10617 jmp(L_processPartitions); 10618 10619 bind(L_exit); 10620 } 10621 #else 10622 void MacroAssembler::crc32c_ipl_alg4(Register in_out, uint32_t n, 10623 Register tmp1, Register tmp2, Register tmp3, 10624 XMMRegister xtmp1, XMMRegister xtmp2) { 10625 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr())); 10626 if (n > 0) { 10627 addl(tmp3, n * 256 * 8); 10628 } 10629 // Q1 = TABLEExt[n][B & 0xFF]; 10630 movl(tmp1, in_out); 10631 andl(tmp1, 0x000000FF); 10632 shll(tmp1, 3); 10633 addl(tmp1, tmp3); 10634 movq(xtmp1, Address(tmp1, 0)); 10635 10636 // Q2 = TABLEExt[n][B >> 8 & 0xFF]; 10637 movl(tmp2, in_out); 10638 shrl(tmp2, 8); 10639 andl(tmp2, 0x000000FF); 10640 shll(tmp2, 3); 10641 addl(tmp2, tmp3); 10642 movq(xtmp2, Address(tmp2, 0)); 10643 10644 psllq(xtmp2, 8); 10645 pxor(xtmp1, xtmp2); 10646 10647 // Q3 = TABLEExt[n][B >> 16 & 0xFF]; 10648 movl(tmp2, in_out); 10649 shrl(tmp2, 16); 10650 andl(tmp2, 0x000000FF); 10651 shll(tmp2, 3); 10652 addl(tmp2, tmp3); 10653 movq(xtmp2, Address(tmp2, 0)); 10654 10655 psllq(xtmp2, 16); 10656 pxor(xtmp1, xtmp2); 10657 10658 // Q4 = TABLEExt[n][B >> 24 & 0xFF]; 10659 shrl(in_out, 24); 10660 andl(in_out, 0x000000FF); 10661 shll(in_out, 3); 10662 addl(in_out, tmp3); 10663 movq(xtmp2, Address(in_out, 0)); 10664 10665 psllq(xtmp2, 24); 10666 pxor(xtmp1, xtmp2); // Result in CXMM 10667 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24; 10668 } 10669 10670 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1, 10671 Register in_out, 10672 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported, 10673 XMMRegister w_xtmp2, 10674 Register tmp1, 10675 Register n_tmp2, Register n_tmp3) { 10676 if (is_pclmulqdq_supported) { 10677 movdl(w_xtmp1, in_out); 10678 10679 movl(tmp1, const_or_pre_comp_const_index); 10680 movdl(w_xtmp2, tmp1); 10681 pclmulqdq(w_xtmp1, w_xtmp2, 0); 10682 // Keep result in XMM since GPR is 32 bit in length 10683 } else { 10684 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3, w_xtmp1, w_xtmp2); 10685 } 10686 } 10687 10688 void MacroAssembler::crc32c_rec_alt2(uint32_t const_or_pre_comp_const_index_u1, uint32_t const_or_pre_comp_const_index_u2, bool is_pclmulqdq_supported, Register in_out, Register in1, Register in2, 10689 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10690 Register tmp1, Register tmp2, 10691 Register n_tmp3) { 10692 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10693 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10694 10695 psllq(w_xtmp1, 1); 10696 movdl(tmp1, w_xtmp1); 10697 psrlq(w_xtmp1, 32); 10698 movdl(in_out, w_xtmp1); 10699 10700 xorl(tmp2, tmp2); 10701 crc32(tmp2, tmp1, 4); 10702 xorl(in_out, tmp2); 10703 10704 psllq(w_xtmp2, 1); 10705 movdl(tmp1, w_xtmp2); 10706 psrlq(w_xtmp2, 32); 10707 movdl(in1, w_xtmp2); 10708 10709 xorl(tmp2, tmp2); 10710 crc32(tmp2, tmp1, 4); 10711 xorl(in1, tmp2); 10712 xorl(in_out, in1); 10713 xorl(in_out, in2); 10714 } 10715 10716 void MacroAssembler::crc32c_proc_chunk(uint32_t size, uint32_t const_or_pre_comp_const_index_u1, uint32_t const_or_pre_comp_const_index_u2, bool is_pclmulqdq_supported, 10717 Register in_out1, Register in_out2, Register in_out3, 10718 Register tmp1, Register tmp2, Register tmp3, 10719 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10720 Register tmp4, Register tmp5, 10721 Register n_tmp6) { 10722 Label L_processPartitions; 10723 Label L_processPartition; 10724 Label L_exit; 10725 10726 bind(L_processPartitions); 10727 cmpl(in_out1, 3 * size); 10728 jcc(Assembler::less, L_exit); 10729 xorl(tmp1, tmp1); 10730 xorl(tmp2, tmp2); 10731 movl(tmp3, in_out2); 10732 addl(tmp3, size); 10733 10734 bind(L_processPartition); 10735 crc32(in_out3, Address(in_out2, 0), 4); 10736 crc32(tmp1, Address(in_out2, size), 4); 10737 crc32(tmp2, Address(in_out2, size*2), 4); 10738 crc32(in_out3, Address(in_out2, 0+4), 4); 10739 crc32(tmp1, Address(in_out2, size+4), 4); 10740 crc32(tmp2, Address(in_out2, size*2+4), 4); 10741 addl(in_out2, 8); 10742 cmpl(in_out2, tmp3); 10743 jcc(Assembler::less, L_processPartition); 10744 10745 push(tmp3); 10746 push(in_out1); 10747 push(in_out2); 10748 tmp4 = tmp3; 10749 tmp5 = in_out1; 10750 n_tmp6 = in_out2; 10751 10752 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2, 10753 w_xtmp1, w_xtmp2, w_xtmp3, 10754 tmp4, tmp5, 10755 n_tmp6); 10756 10757 pop(in_out2); 10758 pop(in_out1); 10759 pop(tmp3); 10760 10761 addl(in_out2, 2 * size); 10762 subl(in_out1, 3 * size); 10763 jmp(L_processPartitions); 10764 10765 bind(L_exit); 10766 } 10767 #endif //LP64 10768 10769 #ifdef _LP64 10770 // Algorithm 2: Pipelined usage of the CRC32 instruction. 10771 // Input: A buffer I of L bytes. 10772 // Output: the CRC32C value of the buffer. 10773 // Notations: 10774 // Write L = 24N + r, with N = floor (L/24). 10775 // r = L mod 24 (0 <= r < 24). 10776 // Consider I as the concatenation of A|B|C|R, where A, B, C, each, 10777 // N quadwords, and R consists of r bytes. 10778 // A[j] = I [8j+7:8j], j= 0, 1, ..., N-1 10779 // B[j] = I [N + 8j+7:N + 8j], j= 0, 1, ..., N-1 10780 // C[j] = I [2N + 8j+7:2N + 8j], j= 0, 1, ..., N-1 10781 // if r > 0 R[j] = I [3N +j], j= 0, 1, ...,r-1 10782 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2, 10783 Register tmp1, Register tmp2, Register tmp3, 10784 Register tmp4, Register tmp5, Register tmp6, 10785 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10786 bool is_pclmulqdq_supported) { 10787 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS]; 10788 Label L_wordByWord; 10789 Label L_byteByByteProlog; 10790 Label L_byteByByte; 10791 Label L_exit; 10792 10793 if (is_pclmulqdq_supported ) { 10794 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr; 10795 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr+1); 10796 10797 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2); 10798 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3); 10799 10800 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4); 10801 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5); 10802 assert((CRC32C_NUM_PRECOMPUTED_CONSTANTS - 1 ) == 5, "Checking whether you declared all of the constants based on the number of \"chunks\""); 10803 } else { 10804 const_or_pre_comp_const_index[0] = 1; 10805 const_or_pre_comp_const_index[1] = 0; 10806 10807 const_or_pre_comp_const_index[2] = 3; 10808 const_or_pre_comp_const_index[3] = 2; 10809 10810 const_or_pre_comp_const_index[4] = 5; 10811 const_or_pre_comp_const_index[5] = 4; 10812 } 10813 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported, 10814 in2, in1, in_out, 10815 tmp1, tmp2, tmp3, 10816 w_xtmp1, w_xtmp2, w_xtmp3, 10817 tmp4, tmp5, 10818 tmp6); 10819 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported, 10820 in2, in1, in_out, 10821 tmp1, tmp2, tmp3, 10822 w_xtmp1, w_xtmp2, w_xtmp3, 10823 tmp4, tmp5, 10824 tmp6); 10825 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported, 10826 in2, in1, in_out, 10827 tmp1, tmp2, tmp3, 10828 w_xtmp1, w_xtmp2, w_xtmp3, 10829 tmp4, tmp5, 10830 tmp6); 10831 movl(tmp1, in2); 10832 andl(tmp1, 0x00000007); 10833 negl(tmp1); 10834 addl(tmp1, in2); 10835 addq(tmp1, in1); 10836 10837 BIND(L_wordByWord); 10838 cmpq(in1, tmp1); 10839 jcc(Assembler::greaterEqual, L_byteByByteProlog); 10840 crc32(in_out, Address(in1, 0), 4); 10841 addq(in1, 4); 10842 jmp(L_wordByWord); 10843 10844 BIND(L_byteByByteProlog); 10845 andl(in2, 0x00000007); 10846 movl(tmp2, 1); 10847 10848 BIND(L_byteByByte); 10849 cmpl(tmp2, in2); 10850 jccb(Assembler::greater, L_exit); 10851 crc32(in_out, Address(in1, 0), 1); 10852 incq(in1); 10853 incl(tmp2); 10854 jmp(L_byteByByte); 10855 10856 BIND(L_exit); 10857 } 10858 #else 10859 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2, 10860 Register tmp1, Register tmp2, Register tmp3, 10861 Register tmp4, Register tmp5, Register tmp6, 10862 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10863 bool is_pclmulqdq_supported) { 10864 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS]; 10865 Label L_wordByWord; 10866 Label L_byteByByteProlog; 10867 Label L_byteByByte; 10868 Label L_exit; 10869 10870 if (is_pclmulqdq_supported) { 10871 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr; 10872 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 1); 10873 10874 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2); 10875 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3); 10876 10877 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4); 10878 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5); 10879 } else { 10880 const_or_pre_comp_const_index[0] = 1; 10881 const_or_pre_comp_const_index[1] = 0; 10882 10883 const_or_pre_comp_const_index[2] = 3; 10884 const_or_pre_comp_const_index[3] = 2; 10885 10886 const_or_pre_comp_const_index[4] = 5; 10887 const_or_pre_comp_const_index[5] = 4; 10888 } 10889 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported, 10890 in2, in1, in_out, 10891 tmp1, tmp2, tmp3, 10892 w_xtmp1, w_xtmp2, w_xtmp3, 10893 tmp4, tmp5, 10894 tmp6); 10895 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported, 10896 in2, in1, in_out, 10897 tmp1, tmp2, tmp3, 10898 w_xtmp1, w_xtmp2, w_xtmp3, 10899 tmp4, tmp5, 10900 tmp6); 10901 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported, 10902 in2, in1, in_out, 10903 tmp1, tmp2, tmp3, 10904 w_xtmp1, w_xtmp2, w_xtmp3, 10905 tmp4, tmp5, 10906 tmp6); 10907 movl(tmp1, in2); 10908 andl(tmp1, 0x00000007); 10909 negl(tmp1); 10910 addl(tmp1, in2); 10911 addl(tmp1, in1); 10912 10913 BIND(L_wordByWord); 10914 cmpl(in1, tmp1); 10915 jcc(Assembler::greaterEqual, L_byteByByteProlog); 10916 crc32(in_out, Address(in1,0), 4); 10917 addl(in1, 4); 10918 jmp(L_wordByWord); 10919 10920 BIND(L_byteByByteProlog); 10921 andl(in2, 0x00000007); 10922 movl(tmp2, 1); 10923 10924 BIND(L_byteByByte); 10925 cmpl(tmp2, in2); 10926 jccb(Assembler::greater, L_exit); 10927 movb(tmp1, Address(in1, 0)); 10928 crc32(in_out, tmp1, 1); 10929 incl(in1); 10930 incl(tmp2); 10931 jmp(L_byteByByte); 10932 10933 BIND(L_exit); 10934 } 10935 #endif // LP64 10936 #undef BIND 10937 #undef BLOCK_COMMENT 10938 10939 // Compress char[] array to byte[]. 10940 // ..\jdk\src\java.base\share\classes\java\lang\StringUTF16.java 10941 // @HotSpotIntrinsicCandidate 10942 // private static int compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) { 10943 // for (int i = 0; i < len; i++) { 10944 // int c = src[srcOff++]; 10945 // if (c >>> 8 != 0) { 10946 // return 0; 10947 // } 10948 // dst[dstOff++] = (byte)c; 10949 // } 10950 // return len; 10951 // } 10952 void MacroAssembler::char_array_compress(Register src, Register dst, Register len, 10953 XMMRegister tmp1Reg, XMMRegister tmp2Reg, 10954 XMMRegister tmp3Reg, XMMRegister tmp4Reg, 10955 Register tmp5, Register result) { 10956 Label copy_chars_loop, return_length, return_zero, done, below_threshold; 10957 10958 // rsi: src 10959 // rdi: dst 10960 // rdx: len 10961 // rcx: tmp5 10962 // rax: result 10963 10964 // rsi holds start addr of source char[] to be compressed 10965 // rdi holds start addr of destination byte[] 10966 // rdx holds length 10967 10968 assert(len != result, ""); 10969 10970 // save length for return 10971 push(len); 10972 10973 if ((UseAVX > 2) && // AVX512 10974 VM_Version::supports_avx512vlbw() && 10975 VM_Version::supports_bmi2()) { 10976 10977 set_vector_masking(); // opening of the stub context for programming mask registers 10978 10979 Label copy_32_loop, copy_loop_tail, restore_k1_return_zero; 10980 10981 // alignement 10982 Label post_alignement; 10983 10984 // if length of the string is less than 16, handle it in an old fashioned 10985 // way 10986 testl(len, -32); 10987 jcc(Assembler::zero, below_threshold); 10988 10989 // First check whether a character is compressable ( <= 0xFF). 10990 // Create mask to test for Unicode chars inside zmm vector 10991 movl(result, 0x00FF); 10992 evpbroadcastw(tmp2Reg, result, Assembler::AVX_512bit); 10993 10994 // Save k1 10995 kmovql(k3, k1); 10996 10997 testl(len, -64); 10998 jcc(Assembler::zero, post_alignement); 10999 11000 movl(tmp5, dst); 11001 andl(tmp5, (32 - 1)); 11002 negl(tmp5); 11003 andl(tmp5, (32 - 1)); 11004 11005 // bail out when there is nothing to be done 11006 testl(tmp5, 0xFFFFFFFF); 11007 jcc(Assembler::zero, post_alignement); 11008 11009 // ~(~0 << len), where len is the # of remaining elements to process 11010 movl(result, 0xFFFFFFFF); 11011 shlxl(result, result, tmp5); 11012 notl(result); 11013 kmovdl(k1, result); 11014 11015 evmovdquw(tmp1Reg, k1, Address(src, 0), Assembler::AVX_512bit); 11016 evpcmpuw(k2, k1, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit); 11017 ktestd(k2, k1); 11018 jcc(Assembler::carryClear, restore_k1_return_zero); 11019 11020 evpmovwb(Address(dst, 0), k1, tmp1Reg, Assembler::AVX_512bit); 11021 11022 addptr(src, tmp5); 11023 addptr(src, tmp5); 11024 addptr(dst, tmp5); 11025 subl(len, tmp5); 11026 11027 bind(post_alignement); 11028 // end of alignement 11029 11030 movl(tmp5, len); 11031 andl(tmp5, (32 - 1)); // tail count (in chars) 11032 andl(len, ~(32 - 1)); // vector count (in chars) 11033 jcc(Assembler::zero, copy_loop_tail); 11034 11035 lea(src, Address(src, len, Address::times_2)); 11036 lea(dst, Address(dst, len, Address::times_1)); 11037 negptr(len); 11038 11039 bind(copy_32_loop); 11040 evmovdquw(tmp1Reg, Address(src, len, Address::times_2), Assembler::AVX_512bit); 11041 evpcmpuw(k2, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit); 11042 kortestdl(k2, k2); 11043 jcc(Assembler::carryClear, restore_k1_return_zero); 11044 11045 // All elements in current processed chunk are valid candidates for 11046 // compression. Write a truncated byte elements to the memory. 11047 evpmovwb(Address(dst, len, Address::times_1), tmp1Reg, Assembler::AVX_512bit); 11048 addptr(len, 32); 11049 jcc(Assembler::notZero, copy_32_loop); 11050 11051 bind(copy_loop_tail); 11052 // bail out when there is nothing to be done 11053 testl(tmp5, 0xFFFFFFFF); 11054 // Restore k1 11055 kmovql(k1, k3); 11056 jcc(Assembler::zero, return_length); 11057 11058 movl(len, tmp5); 11059 11060 // ~(~0 << len), where len is the # of remaining elements to process 11061 movl(result, 0xFFFFFFFF); 11062 shlxl(result, result, len); 11063 notl(result); 11064 11065 kmovdl(k1, result); 11066 11067 evmovdquw(tmp1Reg, k1, Address(src, 0), Assembler::AVX_512bit); 11068 evpcmpuw(k2, k1, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit); 11069 ktestd(k2, k1); 11070 jcc(Assembler::carryClear, restore_k1_return_zero); 11071 11072 evpmovwb(Address(dst, 0), k1, tmp1Reg, Assembler::AVX_512bit); 11073 // Restore k1 11074 kmovql(k1, k3); 11075 jmp(return_length); 11076 11077 bind(restore_k1_return_zero); 11078 // Restore k1 11079 kmovql(k1, k3); 11080 jmp(return_zero); 11081 11082 clear_vector_masking(); // closing of the stub context for programming mask registers 11083 } 11084 if (UseSSE42Intrinsics) { 11085 Label copy_32_loop, copy_16, copy_tail; 11086 11087 bind(below_threshold); 11088 11089 movl(result, len); 11090 11091 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vectors 11092 11093 // vectored compression 11094 andl(len, 0xfffffff0); // vector count (in chars) 11095 andl(result, 0x0000000f); // tail count (in chars) 11096 testl(len, len); 11097 jccb(Assembler::zero, copy_16); 11098 11099 // compress 16 chars per iter 11100 movdl(tmp1Reg, tmp5); 11101 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg 11102 pxor(tmp4Reg, tmp4Reg); 11103 11104 lea(src, Address(src, len, Address::times_2)); 11105 lea(dst, Address(dst, len, Address::times_1)); 11106 negptr(len); 11107 11108 bind(copy_32_loop); 11109 movdqu(tmp2Reg, Address(src, len, Address::times_2)); // load 1st 8 characters 11110 por(tmp4Reg, tmp2Reg); 11111 movdqu(tmp3Reg, Address(src, len, Address::times_2, 16)); // load next 8 characters 11112 por(tmp4Reg, tmp3Reg); 11113 ptest(tmp4Reg, tmp1Reg); // check for Unicode chars in next vector 11114 jcc(Assembler::notZero, return_zero); 11115 packuswb(tmp2Reg, tmp3Reg); // only ASCII chars; compress each to 1 byte 11116 movdqu(Address(dst, len, Address::times_1), tmp2Reg); 11117 addptr(len, 16); 11118 jcc(Assembler::notZero, copy_32_loop); 11119 11120 // compress next vector of 8 chars (if any) 11121 bind(copy_16); 11122 movl(len, result); 11123 andl(len, 0xfffffff8); // vector count (in chars) 11124 andl(result, 0x00000007); // tail count (in chars) 11125 testl(len, len); 11126 jccb(Assembler::zero, copy_tail); 11127 11128 movdl(tmp1Reg, tmp5); 11129 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg 11130 pxor(tmp3Reg, tmp3Reg); 11131 11132 movdqu(tmp2Reg, Address(src, 0)); 11133 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 11134 jccb(Assembler::notZero, return_zero); 11135 packuswb(tmp2Reg, tmp3Reg); // only LATIN1 chars; compress each to 1 byte 11136 movq(Address(dst, 0), tmp2Reg); 11137 addptr(src, 16); 11138 addptr(dst, 8); 11139 11140 bind(copy_tail); 11141 movl(len, result); 11142 } 11143 // compress 1 char per iter 11144 testl(len, len); 11145 jccb(Assembler::zero, return_length); 11146 lea(src, Address(src, len, Address::times_2)); 11147 lea(dst, Address(dst, len, Address::times_1)); 11148 negptr(len); 11149 11150 bind(copy_chars_loop); 11151 load_unsigned_short(result, Address(src, len, Address::times_2)); 11152 testl(result, 0xff00); // check if Unicode char 11153 jccb(Assembler::notZero, return_zero); 11154 movb(Address(dst, len, Address::times_1), result); // ASCII char; compress to 1 byte 11155 increment(len); 11156 jcc(Assembler::notZero, copy_chars_loop); 11157 11158 // if compression succeeded, return length 11159 bind(return_length); 11160 pop(result); 11161 jmpb(done); 11162 11163 // if compression failed, return 0 11164 bind(return_zero); 11165 xorl(result, result); 11166 addptr(rsp, wordSize); 11167 11168 bind(done); 11169 } 11170 11171 // Inflate byte[] array to char[]. 11172 // ..\jdk\src\java.base\share\classes\java\lang\StringLatin1.java 11173 // @HotSpotIntrinsicCandidate 11174 // private static void inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) { 11175 // for (int i = 0; i < len; i++) { 11176 // dst[dstOff++] = (char)(src[srcOff++] & 0xff); 11177 // } 11178 // } 11179 void MacroAssembler::byte_array_inflate(Register src, Register dst, Register len, 11180 XMMRegister tmp1, Register tmp2) { 11181 Label copy_chars_loop, done, below_threshold; 11182 // rsi: src 11183 // rdi: dst 11184 // rdx: len 11185 // rcx: tmp2 11186 11187 // rsi holds start addr of source byte[] to be inflated 11188 // rdi holds start addr of destination char[] 11189 // rdx holds length 11190 assert_different_registers(src, dst, len, tmp2); 11191 11192 if ((UseAVX > 2) && // AVX512 11193 VM_Version::supports_avx512vlbw() && 11194 VM_Version::supports_bmi2()) { 11195 11196 set_vector_masking(); // opening of the stub context for programming mask registers 11197 11198 Label copy_32_loop, copy_tail; 11199 Register tmp3_aliased = len; 11200 11201 // if length of the string is less than 16, handle it in an old fashioned 11202 // way 11203 testl(len, -16); 11204 jcc(Assembler::zero, below_threshold); 11205 11206 // In order to use only one arithmetic operation for the main loop we use 11207 // this pre-calculation 11208 movl(tmp2, len); 11209 andl(tmp2, (32 - 1)); // tail count (in chars), 32 element wide loop 11210 andl(len, -32); // vector count 11211 jccb(Assembler::zero, copy_tail); 11212 11213 lea(src, Address(src, len, Address::times_1)); 11214 lea(dst, Address(dst, len, Address::times_2)); 11215 negptr(len); 11216 11217 11218 // inflate 32 chars per iter 11219 bind(copy_32_loop); 11220 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_512bit); 11221 evmovdquw(Address(dst, len, Address::times_2), tmp1, Assembler::AVX_512bit); 11222 addptr(len, 32); 11223 jcc(Assembler::notZero, copy_32_loop); 11224 11225 bind(copy_tail); 11226 // bail out when there is nothing to be done 11227 testl(tmp2, -1); // we don't destroy the contents of tmp2 here 11228 jcc(Assembler::zero, done); 11229 11230 // Save k1 11231 kmovql(k2, k1); 11232 11233 // ~(~0 << length), where length is the # of remaining elements to process 11234 movl(tmp3_aliased, -1); 11235 shlxl(tmp3_aliased, tmp3_aliased, tmp2); 11236 notl(tmp3_aliased); 11237 kmovdl(k1, tmp3_aliased); 11238 evpmovzxbw(tmp1, k1, Address(src, 0), Assembler::AVX_512bit); 11239 evmovdquw(Address(dst, 0), k1, tmp1, Assembler::AVX_512bit); 11240 11241 // Restore k1 11242 kmovql(k1, k2); 11243 jmp(done); 11244 11245 clear_vector_masking(); // closing of the stub context for programming mask registers 11246 } 11247 if (UseSSE42Intrinsics) { 11248 Label copy_16_loop, copy_8_loop, copy_bytes, copy_new_tail, copy_tail; 11249 11250 movl(tmp2, len); 11251 11252 if (UseAVX > 1) { 11253 andl(tmp2, (16 - 1)); 11254 andl(len, -16); 11255 jccb(Assembler::zero, copy_new_tail); 11256 } else { 11257 andl(tmp2, 0x00000007); // tail count (in chars) 11258 andl(len, 0xfffffff8); // vector count (in chars) 11259 jccb(Assembler::zero, copy_tail); 11260 } 11261 11262 // vectored inflation 11263 lea(src, Address(src, len, Address::times_1)); 11264 lea(dst, Address(dst, len, Address::times_2)); 11265 negptr(len); 11266 11267 if (UseAVX > 1) { 11268 bind(copy_16_loop); 11269 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_256bit); 11270 vmovdqu(Address(dst, len, Address::times_2), tmp1); 11271 addptr(len, 16); 11272 jcc(Assembler::notZero, copy_16_loop); 11273 11274 bind(below_threshold); 11275 bind(copy_new_tail); 11276 if ((UseAVX > 2) && 11277 VM_Version::supports_avx512vlbw() && 11278 VM_Version::supports_bmi2()) { 11279 movl(tmp2, len); 11280 } else { 11281 movl(len, tmp2); 11282 } 11283 andl(tmp2, 0x00000007); 11284 andl(len, 0xFFFFFFF8); 11285 jccb(Assembler::zero, copy_tail); 11286 11287 pmovzxbw(tmp1, Address(src, 0)); 11288 movdqu(Address(dst, 0), tmp1); 11289 addptr(src, 8); 11290 addptr(dst, 2 * 8); 11291 11292 jmp(copy_tail, true); 11293 } 11294 11295 // inflate 8 chars per iter 11296 bind(copy_8_loop); 11297 pmovzxbw(tmp1, Address(src, len, Address::times_1)); // unpack to 8 words 11298 movdqu(Address(dst, len, Address::times_2), tmp1); 11299 addptr(len, 8); 11300 jcc(Assembler::notZero, copy_8_loop); 11301 11302 bind(copy_tail); 11303 movl(len, tmp2); 11304 11305 cmpl(len, 4); 11306 jccb(Assembler::less, copy_bytes); 11307 11308 movdl(tmp1, Address(src, 0)); // load 4 byte chars 11309 pmovzxbw(tmp1, tmp1); 11310 movq(Address(dst, 0), tmp1); 11311 subptr(len, 4); 11312 addptr(src, 4); 11313 addptr(dst, 8); 11314 11315 bind(copy_bytes); 11316 } 11317 testl(len, len); 11318 jccb(Assembler::zero, done); 11319 lea(src, Address(src, len, Address::times_1)); 11320 lea(dst, Address(dst, len, Address::times_2)); 11321 negptr(len); 11322 11323 // inflate 1 char per iter 11324 bind(copy_chars_loop); 11325 load_unsigned_byte(tmp2, Address(src, len, Address::times_1)); // load byte char 11326 movw(Address(dst, len, Address::times_2), tmp2); // inflate byte char to word 11327 increment(len); 11328 jcc(Assembler::notZero, copy_chars_loop); 11329 11330 bind(done); 11331 } 11332 11333 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) { 11334 switch (cond) { 11335 // Note some conditions are synonyms for others 11336 case Assembler::zero: return Assembler::notZero; 11337 case Assembler::notZero: return Assembler::zero; 11338 case Assembler::less: return Assembler::greaterEqual; 11339 case Assembler::lessEqual: return Assembler::greater; 11340 case Assembler::greater: return Assembler::lessEqual; 11341 case Assembler::greaterEqual: return Assembler::less; 11342 case Assembler::below: return Assembler::aboveEqual; 11343 case Assembler::belowEqual: return Assembler::above; 11344 case Assembler::above: return Assembler::belowEqual; 11345 case Assembler::aboveEqual: return Assembler::below; 11346 case Assembler::overflow: return Assembler::noOverflow; 11347 case Assembler::noOverflow: return Assembler::overflow; 11348 case Assembler::negative: return Assembler::positive; 11349 case Assembler::positive: return Assembler::negative; 11350 case Assembler::parity: return Assembler::noParity; 11351 case Assembler::noParity: return Assembler::parity; 11352 } 11353 ShouldNotReachHere(); return Assembler::overflow; 11354 } 11355 11356 SkipIfEqual::SkipIfEqual( 11357 MacroAssembler* masm, const bool* flag_addr, bool value) { 11358 _masm = masm; 11359 _masm->cmp8(ExternalAddress((address)flag_addr), value); 11360 _masm->jcc(Assembler::equal, _label); 11361 } 11362 11363 SkipIfEqual::~SkipIfEqual() { 11364 _masm->bind(_label); 11365 } 11366 11367 // 32-bit Windows has its own fast-path implementation 11368 // of get_thread 11369 #if !defined(WIN32) || defined(_LP64) 11370 11371 // This is simply a call to Thread::current() 11372 void MacroAssembler::get_thread(Register thread) { 11373 if (thread != rax) { 11374 push(rax); 11375 } 11376 LP64_ONLY(push(rdi);) 11377 LP64_ONLY(push(rsi);) 11378 push(rdx); 11379 push(rcx); 11380 #ifdef _LP64 11381 push(r8); 11382 push(r9); 11383 push(r10); 11384 push(r11); 11385 #endif 11386 11387 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, Thread::current), 0); 11388 11389 #ifdef _LP64 11390 pop(r11); 11391 pop(r10); 11392 pop(r9); 11393 pop(r8); 11394 #endif 11395 pop(rcx); 11396 pop(rdx); 11397 LP64_ONLY(pop(rsi);) 11398 LP64_ONLY(pop(rdi);) 11399 if (thread != rax) { 11400 mov(thread, rax); 11401 pop(rax); 11402 } 11403 } 11404 11405 #endif