1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/assembler.hpp" 27 #include "asm/assembler.inline.hpp" 28 #include "compiler/disassembler.hpp" 29 #include "gc/shared/cardTableModRefBS.hpp" 30 #include "gc/shared/collectedHeap.inline.hpp" 31 #include "interpreter/interpreter.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "memory/universe.hpp" 34 #include "oops/klass.inline.hpp" 35 #include "prims/methodHandles.hpp" 36 #include "runtime/biasedLocking.hpp" 37 #include "runtime/interfaceSupport.hpp" 38 #include "runtime/objectMonitor.hpp" 39 #include "runtime/os.hpp" 40 #include "runtime/sharedRuntime.hpp" 41 #include "runtime/stubRoutines.hpp" 42 #include "runtime/thread.hpp" 43 #include "utilities/macros.hpp" 44 #if INCLUDE_ALL_GCS 45 #include "gc/g1/g1CollectedHeap.inline.hpp" 46 #include "gc/g1/g1SATBCardTableModRefBS.hpp" 47 #include "gc/g1/heapRegion.hpp" 48 #endif // INCLUDE_ALL_GCS 49 #include "crc32c.h" 50 #ifdef COMPILER2 51 #include "opto/intrinsicnode.hpp" 52 #endif 53 54 #ifdef PRODUCT 55 #define BLOCK_COMMENT(str) /* nothing */ 56 #define STOP(error) stop(error) 57 #else 58 #define BLOCK_COMMENT(str) block_comment(str) 59 #define STOP(error) block_comment(error); stop(error) 60 #endif 61 62 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":") 63 64 #ifdef ASSERT 65 bool AbstractAssembler::pd_check_instruction_mark() { return true; } 66 #endif 67 68 static Assembler::Condition reverse[] = { 69 Assembler::noOverflow /* overflow = 0x0 */ , 70 Assembler::overflow /* noOverflow = 0x1 */ , 71 Assembler::aboveEqual /* carrySet = 0x2, below = 0x2 */ , 72 Assembler::below /* aboveEqual = 0x3, carryClear = 0x3 */ , 73 Assembler::notZero /* zero = 0x4, equal = 0x4 */ , 74 Assembler::zero /* notZero = 0x5, notEqual = 0x5 */ , 75 Assembler::above /* belowEqual = 0x6 */ , 76 Assembler::belowEqual /* above = 0x7 */ , 77 Assembler::positive /* negative = 0x8 */ , 78 Assembler::negative /* positive = 0x9 */ , 79 Assembler::noParity /* parity = 0xa */ , 80 Assembler::parity /* noParity = 0xb */ , 81 Assembler::greaterEqual /* less = 0xc */ , 82 Assembler::less /* greaterEqual = 0xd */ , 83 Assembler::greater /* lessEqual = 0xe */ , 84 Assembler::lessEqual /* greater = 0xf, */ 85 86 }; 87 88 89 // Implementation of MacroAssembler 90 91 // First all the versions that have distinct versions depending on 32/64 bit 92 // Unless the difference is trivial (1 line or so). 93 94 #ifndef _LP64 95 96 // 32bit versions 97 98 Address MacroAssembler::as_Address(AddressLiteral adr) { 99 return Address(adr.target(), adr.rspec()); 100 } 101 102 Address MacroAssembler::as_Address(ArrayAddress adr) { 103 return Address::make_array(adr); 104 } 105 106 void MacroAssembler::call_VM_leaf_base(address entry_point, 107 int number_of_arguments) { 108 call(RuntimeAddress(entry_point)); 109 increment(rsp, number_of_arguments * wordSize); 110 } 111 112 void MacroAssembler::cmpklass(Address src1, Metadata* obj) { 113 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 114 } 115 116 void MacroAssembler::cmpklass(Register src1, Metadata* obj) { 117 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 118 } 119 120 void MacroAssembler::cmpoop(Address src1, jobject obj) { 121 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate()); 122 } 123 124 void MacroAssembler::cmpoop(Register src1, jobject obj) { 125 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate()); 126 } 127 128 void MacroAssembler::extend_sign(Register hi, Register lo) { 129 // According to Intel Doc. AP-526, "Integer Divide", p.18. 130 if (VM_Version::is_P6() && hi == rdx && lo == rax) { 131 cdql(); 132 } else { 133 movl(hi, lo); 134 sarl(hi, 31); 135 } 136 } 137 138 void MacroAssembler::jC2(Register tmp, Label& L) { 139 // set parity bit if FPU flag C2 is set (via rax) 140 save_rax(tmp); 141 fwait(); fnstsw_ax(); 142 sahf(); 143 restore_rax(tmp); 144 // branch 145 jcc(Assembler::parity, L); 146 } 147 148 void MacroAssembler::jnC2(Register tmp, Label& L) { 149 // set parity bit if FPU flag C2 is set (via rax) 150 save_rax(tmp); 151 fwait(); fnstsw_ax(); 152 sahf(); 153 restore_rax(tmp); 154 // branch 155 jcc(Assembler::noParity, L); 156 } 157 158 // 32bit can do a case table jump in one instruction but we no longer allow the base 159 // to be installed in the Address class 160 void MacroAssembler::jump(ArrayAddress entry) { 161 jmp(as_Address(entry)); 162 } 163 164 // Note: y_lo will be destroyed 165 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) { 166 // Long compare for Java (semantics as described in JVM spec.) 167 Label high, low, done; 168 169 cmpl(x_hi, y_hi); 170 jcc(Assembler::less, low); 171 jcc(Assembler::greater, high); 172 // x_hi is the return register 173 xorl(x_hi, x_hi); 174 cmpl(x_lo, y_lo); 175 jcc(Assembler::below, low); 176 jcc(Assembler::equal, done); 177 178 bind(high); 179 xorl(x_hi, x_hi); 180 increment(x_hi); 181 jmp(done); 182 183 bind(low); 184 xorl(x_hi, x_hi); 185 decrementl(x_hi); 186 187 bind(done); 188 } 189 190 void MacroAssembler::lea(Register dst, AddressLiteral src) { 191 mov_literal32(dst, (int32_t)src.target(), src.rspec()); 192 } 193 194 void MacroAssembler::lea(Address dst, AddressLiteral adr) { 195 // leal(dst, as_Address(adr)); 196 // see note in movl as to why we must use a move 197 mov_literal32(dst, (int32_t) adr.target(), adr.rspec()); 198 } 199 200 void MacroAssembler::leave() { 201 mov(rsp, rbp); 202 pop(rbp); 203 } 204 205 void MacroAssembler::lmul(int x_rsp_offset, int y_rsp_offset) { 206 // Multiplication of two Java long values stored on the stack 207 // as illustrated below. Result is in rdx:rax. 208 // 209 // rsp ---> [ ?? ] \ \ 210 // .... | y_rsp_offset | 211 // [ y_lo ] / (in bytes) | x_rsp_offset 212 // [ y_hi ] | (in bytes) 213 // .... | 214 // [ x_lo ] / 215 // [ x_hi ] 216 // .... 217 // 218 // Basic idea: lo(result) = lo(x_lo * y_lo) 219 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi) 220 Address x_hi(rsp, x_rsp_offset + wordSize); Address x_lo(rsp, x_rsp_offset); 221 Address y_hi(rsp, y_rsp_offset + wordSize); Address y_lo(rsp, y_rsp_offset); 222 Label quick; 223 // load x_hi, y_hi and check if quick 224 // multiplication is possible 225 movl(rbx, x_hi); 226 movl(rcx, y_hi); 227 movl(rax, rbx); 228 orl(rbx, rcx); // rbx, = 0 <=> x_hi = 0 and y_hi = 0 229 jcc(Assembler::zero, quick); // if rbx, = 0 do quick multiply 230 // do full multiplication 231 // 1st step 232 mull(y_lo); // x_hi * y_lo 233 movl(rbx, rax); // save lo(x_hi * y_lo) in rbx, 234 // 2nd step 235 movl(rax, x_lo); 236 mull(rcx); // x_lo * y_hi 237 addl(rbx, rax); // add lo(x_lo * y_hi) to rbx, 238 // 3rd step 239 bind(quick); // note: rbx, = 0 if quick multiply! 240 movl(rax, x_lo); 241 mull(y_lo); // x_lo * y_lo 242 addl(rdx, rbx); // correct hi(x_lo * y_lo) 243 } 244 245 void MacroAssembler::lneg(Register hi, Register lo) { 246 negl(lo); 247 adcl(hi, 0); 248 negl(hi); 249 } 250 251 void MacroAssembler::lshl(Register hi, Register lo) { 252 // Java shift left long support (semantics as described in JVM spec., p.305) 253 // (basic idea for shift counts s >= n: x << s == (x << n) << (s - n)) 254 // shift value is in rcx ! 255 assert(hi != rcx, "must not use rcx"); 256 assert(lo != rcx, "must not use rcx"); 257 const Register s = rcx; // shift count 258 const int n = BitsPerWord; 259 Label L; 260 andl(s, 0x3f); // s := s & 0x3f (s < 0x40) 261 cmpl(s, n); // if (s < n) 262 jcc(Assembler::less, L); // else (s >= n) 263 movl(hi, lo); // x := x << n 264 xorl(lo, lo); 265 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n! 266 bind(L); // s (mod n) < n 267 shldl(hi, lo); // x := x << s 268 shll(lo); 269 } 270 271 272 void MacroAssembler::lshr(Register hi, Register lo, bool sign_extension) { 273 // Java shift right long support (semantics as described in JVM spec., p.306 & p.310) 274 // (basic idea for shift counts s >= n: x >> s == (x >> n) >> (s - n)) 275 assert(hi != rcx, "must not use rcx"); 276 assert(lo != rcx, "must not use rcx"); 277 const Register s = rcx; // shift count 278 const int n = BitsPerWord; 279 Label L; 280 andl(s, 0x3f); // s := s & 0x3f (s < 0x40) 281 cmpl(s, n); // if (s < n) 282 jcc(Assembler::less, L); // else (s >= n) 283 movl(lo, hi); // x := x >> n 284 if (sign_extension) sarl(hi, 31); 285 else xorl(hi, hi); 286 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n! 287 bind(L); // s (mod n) < n 288 shrdl(lo, hi); // x := x >> s 289 if (sign_extension) sarl(hi); 290 else shrl(hi); 291 } 292 293 void MacroAssembler::movoop(Register dst, jobject obj) { 294 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate()); 295 } 296 297 void MacroAssembler::movoop(Address dst, jobject obj) { 298 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate()); 299 } 300 301 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) { 302 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 303 } 304 305 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) { 306 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 307 } 308 309 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) { 310 // scratch register is not used, 311 // it is defined to match parameters of 64-bit version of this method. 312 if (src.is_lval()) { 313 mov_literal32(dst, (intptr_t)src.target(), src.rspec()); 314 } else { 315 movl(dst, as_Address(src)); 316 } 317 } 318 319 void MacroAssembler::movptr(ArrayAddress dst, Register src) { 320 movl(as_Address(dst), src); 321 } 322 323 void MacroAssembler::movptr(Register dst, ArrayAddress src) { 324 movl(dst, as_Address(src)); 325 } 326 327 // src should NEVER be a real pointer. Use AddressLiteral for true pointers 328 void MacroAssembler::movptr(Address dst, intptr_t src) { 329 movl(dst, src); 330 } 331 332 333 void MacroAssembler::pop_callee_saved_registers() { 334 pop(rcx); 335 pop(rdx); 336 pop(rdi); 337 pop(rsi); 338 } 339 340 void MacroAssembler::pop_fTOS() { 341 fld_d(Address(rsp, 0)); 342 addl(rsp, 2 * wordSize); 343 } 344 345 void MacroAssembler::push_callee_saved_registers() { 346 push(rsi); 347 push(rdi); 348 push(rdx); 349 push(rcx); 350 } 351 352 void MacroAssembler::push_fTOS() { 353 subl(rsp, 2 * wordSize); 354 fstp_d(Address(rsp, 0)); 355 } 356 357 358 void MacroAssembler::pushoop(jobject obj) { 359 push_literal32((int32_t)obj, oop_Relocation::spec_for_immediate()); 360 } 361 362 void MacroAssembler::pushklass(Metadata* obj) { 363 push_literal32((int32_t)obj, metadata_Relocation::spec_for_immediate()); 364 } 365 366 void MacroAssembler::pushptr(AddressLiteral src) { 367 if (src.is_lval()) { 368 push_literal32((int32_t)src.target(), src.rspec()); 369 } else { 370 pushl(as_Address(src)); 371 } 372 } 373 374 void MacroAssembler::set_word_if_not_zero(Register dst) { 375 xorl(dst, dst); 376 set_byte_if_not_zero(dst); 377 } 378 379 static void pass_arg0(MacroAssembler* masm, Register arg) { 380 masm->push(arg); 381 } 382 383 static void pass_arg1(MacroAssembler* masm, Register arg) { 384 masm->push(arg); 385 } 386 387 static void pass_arg2(MacroAssembler* masm, Register arg) { 388 masm->push(arg); 389 } 390 391 static void pass_arg3(MacroAssembler* masm, Register arg) { 392 masm->push(arg); 393 } 394 395 #ifndef PRODUCT 396 extern "C" void findpc(intptr_t x); 397 #endif 398 399 void MacroAssembler::debug32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip, char* msg) { 400 // In order to get locks to work, we need to fake a in_VM state 401 JavaThread* thread = JavaThread::current(); 402 JavaThreadState saved_state = thread->thread_state(); 403 thread->set_thread_state(_thread_in_vm); 404 if (ShowMessageBoxOnError) { 405 JavaThread* thread = JavaThread::current(); 406 JavaThreadState saved_state = thread->thread_state(); 407 thread->set_thread_state(_thread_in_vm); 408 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) { 409 ttyLocker ttyl; 410 BytecodeCounter::print(); 411 } 412 // To see where a verify_oop failed, get $ebx+40/X for this frame. 413 // This is the value of eip which points to where verify_oop will return. 414 if (os::message_box(msg, "Execution stopped, print registers?")) { 415 print_state32(rdi, rsi, rbp, rsp, rbx, rdx, rcx, rax, eip); 416 BREAKPOINT; 417 } 418 } else { 419 ttyLocker ttyl; 420 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", msg); 421 } 422 // Don't assert holding the ttyLock 423 assert(false, "DEBUG MESSAGE: %s", msg); 424 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state); 425 } 426 427 void MacroAssembler::print_state32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip) { 428 ttyLocker ttyl; 429 FlagSetting fs(Debugging, true); 430 tty->print_cr("eip = 0x%08x", eip); 431 #ifndef PRODUCT 432 if ((WizardMode || Verbose) && PrintMiscellaneous) { 433 tty->cr(); 434 findpc(eip); 435 tty->cr(); 436 } 437 #endif 438 #define PRINT_REG(rax) \ 439 { tty->print("%s = ", #rax); os::print_location(tty, rax); } 440 PRINT_REG(rax); 441 PRINT_REG(rbx); 442 PRINT_REG(rcx); 443 PRINT_REG(rdx); 444 PRINT_REG(rdi); 445 PRINT_REG(rsi); 446 PRINT_REG(rbp); 447 PRINT_REG(rsp); 448 #undef PRINT_REG 449 // Print some words near top of staack. 450 int* dump_sp = (int*) rsp; 451 for (int col1 = 0; col1 < 8; col1++) { 452 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 453 os::print_location(tty, *dump_sp++); 454 } 455 for (int row = 0; row < 16; row++) { 456 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 457 for (int col = 0; col < 8; col++) { 458 tty->print(" 0x%08x", *dump_sp++); 459 } 460 tty->cr(); 461 } 462 // Print some instructions around pc: 463 Disassembler::decode((address)eip-64, (address)eip); 464 tty->print_cr("--------"); 465 Disassembler::decode((address)eip, (address)eip+32); 466 } 467 468 void MacroAssembler::stop(const char* msg) { 469 ExternalAddress message((address)msg); 470 // push address of message 471 pushptr(message.addr()); 472 { Label L; call(L, relocInfo::none); bind(L); } // push eip 473 pusha(); // push registers 474 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug32))); 475 hlt(); 476 } 477 478 void MacroAssembler::warn(const char* msg) { 479 push_CPU_state(); 480 481 ExternalAddress message((address) msg); 482 // push address of message 483 pushptr(message.addr()); 484 485 call(RuntimeAddress(CAST_FROM_FN_PTR(address, warning))); 486 addl(rsp, wordSize); // discard argument 487 pop_CPU_state(); 488 } 489 490 void MacroAssembler::print_state() { 491 { Label L; call(L, relocInfo::none); bind(L); } // push eip 492 pusha(); // push registers 493 494 push_CPU_state(); 495 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::print_state32))); 496 pop_CPU_state(); 497 498 popa(); 499 addl(rsp, wordSize); 500 } 501 502 #else // _LP64 503 504 // 64 bit versions 505 506 Address MacroAssembler::as_Address(AddressLiteral adr) { 507 // amd64 always does this as a pc-rel 508 // we can be absolute or disp based on the instruction type 509 // jmp/call are displacements others are absolute 510 assert(!adr.is_lval(), "must be rval"); 511 assert(reachable(adr), "must be"); 512 return Address((int32_t)(intptr_t)(adr.target() - pc()), adr.target(), adr.reloc()); 513 514 } 515 516 Address MacroAssembler::as_Address(ArrayAddress adr) { 517 AddressLiteral base = adr.base(); 518 lea(rscratch1, base); 519 Address index = adr.index(); 520 assert(index._disp == 0, "must not have disp"); // maybe it can? 521 Address array(rscratch1, index._index, index._scale, index._disp); 522 return array; 523 } 524 525 void MacroAssembler::call_VM_leaf_base(address entry_point, int num_args) { 526 Label L, E; 527 528 #ifdef _WIN64 529 // Windows always allocates space for it's register args 530 assert(num_args <= 4, "only register arguments supported"); 531 subq(rsp, frame::arg_reg_save_area_bytes); 532 #endif 533 534 // Align stack if necessary 535 testl(rsp, 15); 536 jcc(Assembler::zero, L); 537 538 subq(rsp, 8); 539 { 540 call(RuntimeAddress(entry_point)); 541 } 542 addq(rsp, 8); 543 jmp(E); 544 545 bind(L); 546 { 547 call(RuntimeAddress(entry_point)); 548 } 549 550 bind(E); 551 552 #ifdef _WIN64 553 // restore stack pointer 554 addq(rsp, frame::arg_reg_save_area_bytes); 555 #endif 556 557 } 558 559 void MacroAssembler::cmp64(Register src1, AddressLiteral src2) { 560 assert(!src2.is_lval(), "should use cmpptr"); 561 562 if (reachable(src2)) { 563 cmpq(src1, as_Address(src2)); 564 } else { 565 lea(rscratch1, src2); 566 Assembler::cmpq(src1, Address(rscratch1, 0)); 567 } 568 } 569 570 int MacroAssembler::corrected_idivq(Register reg) { 571 // Full implementation of Java ldiv and lrem; checks for special 572 // case as described in JVM spec., p.243 & p.271. The function 573 // returns the (pc) offset of the idivl instruction - may be needed 574 // for implicit exceptions. 575 // 576 // normal case special case 577 // 578 // input : rax: dividend min_long 579 // reg: divisor (may not be eax/edx) -1 580 // 581 // output: rax: quotient (= rax idiv reg) min_long 582 // rdx: remainder (= rax irem reg) 0 583 assert(reg != rax && reg != rdx, "reg cannot be rax or rdx register"); 584 static const int64_t min_long = 0x8000000000000000; 585 Label normal_case, special_case; 586 587 // check for special case 588 cmp64(rax, ExternalAddress((address) &min_long)); 589 jcc(Assembler::notEqual, normal_case); 590 xorl(rdx, rdx); // prepare rdx for possible special case (where 591 // remainder = 0) 592 cmpq(reg, -1); 593 jcc(Assembler::equal, special_case); 594 595 // handle normal case 596 bind(normal_case); 597 cdqq(); 598 int idivq_offset = offset(); 599 idivq(reg); 600 601 // normal and special case exit 602 bind(special_case); 603 604 return idivq_offset; 605 } 606 607 void MacroAssembler::decrementq(Register reg, int value) { 608 if (value == min_jint) { subq(reg, value); return; } 609 if (value < 0) { incrementq(reg, -value); return; } 610 if (value == 0) { ; return; } 611 if (value == 1 && UseIncDec) { decq(reg) ; return; } 612 /* else */ { subq(reg, value) ; return; } 613 } 614 615 void MacroAssembler::decrementq(Address dst, int value) { 616 if (value == min_jint) { subq(dst, value); return; } 617 if (value < 0) { incrementq(dst, -value); return; } 618 if (value == 0) { ; return; } 619 if (value == 1 && UseIncDec) { decq(dst) ; return; } 620 /* else */ { subq(dst, value) ; return; } 621 } 622 623 void MacroAssembler::incrementq(AddressLiteral dst) { 624 if (reachable(dst)) { 625 incrementq(as_Address(dst)); 626 } else { 627 lea(rscratch1, dst); 628 incrementq(Address(rscratch1, 0)); 629 } 630 } 631 632 void MacroAssembler::incrementq(Register reg, int value) { 633 if (value == min_jint) { addq(reg, value); return; } 634 if (value < 0) { decrementq(reg, -value); return; } 635 if (value == 0) { ; return; } 636 if (value == 1 && UseIncDec) { incq(reg) ; return; } 637 /* else */ { addq(reg, value) ; return; } 638 } 639 640 void MacroAssembler::incrementq(Address dst, int value) { 641 if (value == min_jint) { addq(dst, value); return; } 642 if (value < 0) { decrementq(dst, -value); return; } 643 if (value == 0) { ; return; } 644 if (value == 1 && UseIncDec) { incq(dst) ; return; } 645 /* else */ { addq(dst, value) ; return; } 646 } 647 648 // 32bit can do a case table jump in one instruction but we no longer allow the base 649 // to be installed in the Address class 650 void MacroAssembler::jump(ArrayAddress entry) { 651 lea(rscratch1, entry.base()); 652 Address dispatch = entry.index(); 653 assert(dispatch._base == noreg, "must be"); 654 dispatch._base = rscratch1; 655 jmp(dispatch); 656 } 657 658 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) { 659 ShouldNotReachHere(); // 64bit doesn't use two regs 660 cmpq(x_lo, y_lo); 661 } 662 663 void MacroAssembler::lea(Register dst, AddressLiteral src) { 664 mov_literal64(dst, (intptr_t)src.target(), src.rspec()); 665 } 666 667 void MacroAssembler::lea(Address dst, AddressLiteral adr) { 668 mov_literal64(rscratch1, (intptr_t)adr.target(), adr.rspec()); 669 movptr(dst, rscratch1); 670 } 671 672 void MacroAssembler::leave() { 673 // %%% is this really better? Why not on 32bit too? 674 emit_int8((unsigned char)0xC9); // LEAVE 675 } 676 677 void MacroAssembler::lneg(Register hi, Register lo) { 678 ShouldNotReachHere(); // 64bit doesn't use two regs 679 negq(lo); 680 } 681 682 void MacroAssembler::movoop(Register dst, jobject obj) { 683 mov_literal64(dst, (intptr_t)obj, oop_Relocation::spec_for_immediate()); 684 } 685 686 void MacroAssembler::movoop(Address dst, jobject obj) { 687 mov_literal64(rscratch1, (intptr_t)obj, oop_Relocation::spec_for_immediate()); 688 movq(dst, rscratch1); 689 } 690 691 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) { 692 mov_literal64(dst, (intptr_t)obj, metadata_Relocation::spec_for_immediate()); 693 } 694 695 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) { 696 mov_literal64(rscratch1, (intptr_t)obj, metadata_Relocation::spec_for_immediate()); 697 movq(dst, rscratch1); 698 } 699 700 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) { 701 if (src.is_lval()) { 702 mov_literal64(dst, (intptr_t)src.target(), src.rspec()); 703 } else { 704 if (reachable(src)) { 705 movq(dst, as_Address(src)); 706 } else { 707 lea(scratch, src); 708 movq(dst, Address(scratch, 0)); 709 } 710 } 711 } 712 713 void MacroAssembler::movptr(ArrayAddress dst, Register src) { 714 movq(as_Address(dst), src); 715 } 716 717 void MacroAssembler::movptr(Register dst, ArrayAddress src) { 718 movq(dst, as_Address(src)); 719 } 720 721 // src should NEVER be a real pointer. Use AddressLiteral for true pointers 722 void MacroAssembler::movptr(Address dst, intptr_t src) { 723 mov64(rscratch1, src); 724 movq(dst, rscratch1); 725 } 726 727 // These are mostly for initializing NULL 728 void MacroAssembler::movptr(Address dst, int32_t src) { 729 movslq(dst, src); 730 } 731 732 void MacroAssembler::movptr(Register dst, int32_t src) { 733 mov64(dst, (intptr_t)src); 734 } 735 736 void MacroAssembler::pushoop(jobject obj) { 737 movoop(rscratch1, obj); 738 push(rscratch1); 739 } 740 741 void MacroAssembler::pushklass(Metadata* obj) { 742 mov_metadata(rscratch1, obj); 743 push(rscratch1); 744 } 745 746 void MacroAssembler::pushptr(AddressLiteral src) { 747 lea(rscratch1, src); 748 if (src.is_lval()) { 749 push(rscratch1); 750 } else { 751 pushq(Address(rscratch1, 0)); 752 } 753 } 754 755 void MacroAssembler::reset_last_Java_frame(bool clear_fp, 756 bool clear_pc) { 757 // we must set sp to zero to clear frame 758 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), NULL_WORD); 759 // must clear fp, so that compiled frames are not confused; it is 760 // possible that we need it only for debugging 761 if (clear_fp) { 762 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD); 763 } 764 765 if (clear_pc) { 766 movptr(Address(r15_thread, JavaThread::last_Java_pc_offset()), NULL_WORD); 767 } 768 } 769 770 void MacroAssembler::set_last_Java_frame(Register last_java_sp, 771 Register last_java_fp, 772 address last_java_pc) { 773 // determine last_java_sp register 774 if (!last_java_sp->is_valid()) { 775 last_java_sp = rsp; 776 } 777 778 // last_java_fp is optional 779 if (last_java_fp->is_valid()) { 780 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), 781 last_java_fp); 782 } 783 784 // last_java_pc is optional 785 if (last_java_pc != NULL) { 786 Address java_pc(r15_thread, 787 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()); 788 lea(rscratch1, InternalAddress(last_java_pc)); 789 movptr(java_pc, rscratch1); 790 } 791 792 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), last_java_sp); 793 } 794 795 static void pass_arg0(MacroAssembler* masm, Register arg) { 796 if (c_rarg0 != arg ) { 797 masm->mov(c_rarg0, arg); 798 } 799 } 800 801 static void pass_arg1(MacroAssembler* masm, Register arg) { 802 if (c_rarg1 != arg ) { 803 masm->mov(c_rarg1, arg); 804 } 805 } 806 807 static void pass_arg2(MacroAssembler* masm, Register arg) { 808 if (c_rarg2 != arg ) { 809 masm->mov(c_rarg2, arg); 810 } 811 } 812 813 static void pass_arg3(MacroAssembler* masm, Register arg) { 814 if (c_rarg3 != arg ) { 815 masm->mov(c_rarg3, arg); 816 } 817 } 818 819 void MacroAssembler::stop(const char* msg) { 820 address rip = pc(); 821 pusha(); // get regs on stack 822 lea(c_rarg0, ExternalAddress((address) msg)); 823 lea(c_rarg1, InternalAddress(rip)); 824 movq(c_rarg2, rsp); // pass pointer to regs array 825 andq(rsp, -16); // align stack as required by ABI 826 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64))); 827 hlt(); 828 } 829 830 void MacroAssembler::warn(const char* msg) { 831 push(rbp); 832 movq(rbp, rsp); 833 andq(rsp, -16); // align stack as required by push_CPU_state and call 834 push_CPU_state(); // keeps alignment at 16 bytes 835 lea(c_rarg0, ExternalAddress((address) msg)); 836 call_VM_leaf(CAST_FROM_FN_PTR(address, warning), c_rarg0); 837 pop_CPU_state(); 838 mov(rsp, rbp); 839 pop(rbp); 840 } 841 842 void MacroAssembler::print_state() { 843 address rip = pc(); 844 pusha(); // get regs on stack 845 push(rbp); 846 movq(rbp, rsp); 847 andq(rsp, -16); // align stack as required by push_CPU_state and call 848 push_CPU_state(); // keeps alignment at 16 bytes 849 850 lea(c_rarg0, InternalAddress(rip)); 851 lea(c_rarg1, Address(rbp, wordSize)); // pass pointer to regs array 852 call_VM_leaf(CAST_FROM_FN_PTR(address, MacroAssembler::print_state64), c_rarg0, c_rarg1); 853 854 pop_CPU_state(); 855 mov(rsp, rbp); 856 pop(rbp); 857 popa(); 858 } 859 860 #ifndef PRODUCT 861 extern "C" void findpc(intptr_t x); 862 #endif 863 864 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[]) { 865 // In order to get locks to work, we need to fake a in_VM state 866 if (ShowMessageBoxOnError) { 867 JavaThread* thread = JavaThread::current(); 868 JavaThreadState saved_state = thread->thread_state(); 869 thread->set_thread_state(_thread_in_vm); 870 #ifndef PRODUCT 871 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) { 872 ttyLocker ttyl; 873 BytecodeCounter::print(); 874 } 875 #endif 876 // To see where a verify_oop failed, get $ebx+40/X for this frame. 877 // XXX correct this offset for amd64 878 // This is the value of eip which points to where verify_oop will return. 879 if (os::message_box(msg, "Execution stopped, print registers?")) { 880 print_state64(pc, regs); 881 BREAKPOINT; 882 assert(false, "start up GDB"); 883 } 884 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state); 885 } else { 886 ttyLocker ttyl; 887 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", 888 msg); 889 assert(false, "DEBUG MESSAGE: %s", msg); 890 } 891 } 892 893 void MacroAssembler::print_state64(int64_t pc, int64_t regs[]) { 894 ttyLocker ttyl; 895 FlagSetting fs(Debugging, true); 896 tty->print_cr("rip = 0x%016lx", pc); 897 #ifndef PRODUCT 898 tty->cr(); 899 findpc(pc); 900 tty->cr(); 901 #endif 902 #define PRINT_REG(rax, value) \ 903 { tty->print("%s = ", #rax); os::print_location(tty, value); } 904 PRINT_REG(rax, regs[15]); 905 PRINT_REG(rbx, regs[12]); 906 PRINT_REG(rcx, regs[14]); 907 PRINT_REG(rdx, regs[13]); 908 PRINT_REG(rdi, regs[8]); 909 PRINT_REG(rsi, regs[9]); 910 PRINT_REG(rbp, regs[10]); 911 PRINT_REG(rsp, regs[11]); 912 PRINT_REG(r8 , regs[7]); 913 PRINT_REG(r9 , regs[6]); 914 PRINT_REG(r10, regs[5]); 915 PRINT_REG(r11, regs[4]); 916 PRINT_REG(r12, regs[3]); 917 PRINT_REG(r13, regs[2]); 918 PRINT_REG(r14, regs[1]); 919 PRINT_REG(r15, regs[0]); 920 #undef PRINT_REG 921 // Print some words near top of staack. 922 int64_t* rsp = (int64_t*) regs[11]; 923 int64_t* dump_sp = rsp; 924 for (int col1 = 0; col1 < 8; col1++) { 925 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (int64_t)dump_sp); 926 os::print_location(tty, *dump_sp++); 927 } 928 for (int row = 0; row < 25; row++) { 929 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (int64_t)dump_sp); 930 for (int col = 0; col < 4; col++) { 931 tty->print(" 0x%016lx", *dump_sp++); 932 } 933 tty->cr(); 934 } 935 // Print some instructions around pc: 936 Disassembler::decode((address)pc-64, (address)pc); 937 tty->print_cr("--------"); 938 Disassembler::decode((address)pc, (address)pc+32); 939 } 940 941 #endif // _LP64 942 943 // Now versions that are common to 32/64 bit 944 945 void MacroAssembler::addptr(Register dst, int32_t imm32) { 946 LP64_ONLY(addq(dst, imm32)) NOT_LP64(addl(dst, imm32)); 947 } 948 949 void MacroAssembler::addptr(Register dst, Register src) { 950 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src)); 951 } 952 953 void MacroAssembler::addptr(Address dst, Register src) { 954 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src)); 955 } 956 957 void MacroAssembler::addsd(XMMRegister dst, AddressLiteral src) { 958 if (reachable(src)) { 959 Assembler::addsd(dst, as_Address(src)); 960 } else { 961 lea(rscratch1, src); 962 Assembler::addsd(dst, Address(rscratch1, 0)); 963 } 964 } 965 966 void MacroAssembler::addss(XMMRegister dst, AddressLiteral src) { 967 if (reachable(src)) { 968 addss(dst, as_Address(src)); 969 } else { 970 lea(rscratch1, src); 971 addss(dst, Address(rscratch1, 0)); 972 } 973 } 974 975 void MacroAssembler::align(int modulus) { 976 align(modulus, offset()); 977 } 978 979 void MacroAssembler::align(int modulus, int target) { 980 if (target % modulus != 0) { 981 nop(modulus - (target % modulus)); 982 } 983 } 984 985 void MacroAssembler::andpd(XMMRegister dst, AddressLiteral src) { 986 // Used in sign-masking with aligned address. 987 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 988 if (reachable(src)) { 989 Assembler::andpd(dst, as_Address(src)); 990 } else { 991 lea(rscratch1, src); 992 Assembler::andpd(dst, Address(rscratch1, 0)); 993 } 994 } 995 996 void MacroAssembler::andps(XMMRegister dst, AddressLiteral src) { 997 // Used in sign-masking with aligned address. 998 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 999 if (reachable(src)) { 1000 Assembler::andps(dst, as_Address(src)); 1001 } else { 1002 lea(rscratch1, src); 1003 Assembler::andps(dst, Address(rscratch1, 0)); 1004 } 1005 } 1006 1007 void MacroAssembler::andptr(Register dst, int32_t imm32) { 1008 LP64_ONLY(andq(dst, imm32)) NOT_LP64(andl(dst, imm32)); 1009 } 1010 1011 void MacroAssembler::atomic_incl(Address counter_addr) { 1012 if (os::is_MP()) 1013 lock(); 1014 incrementl(counter_addr); 1015 } 1016 1017 void MacroAssembler::atomic_incl(AddressLiteral counter_addr, Register scr) { 1018 if (reachable(counter_addr)) { 1019 atomic_incl(as_Address(counter_addr)); 1020 } else { 1021 lea(scr, counter_addr); 1022 atomic_incl(Address(scr, 0)); 1023 } 1024 } 1025 1026 #ifdef _LP64 1027 void MacroAssembler::atomic_incq(Address counter_addr) { 1028 if (os::is_MP()) 1029 lock(); 1030 incrementq(counter_addr); 1031 } 1032 1033 void MacroAssembler::atomic_incq(AddressLiteral counter_addr, Register scr) { 1034 if (reachable(counter_addr)) { 1035 atomic_incq(as_Address(counter_addr)); 1036 } else { 1037 lea(scr, counter_addr); 1038 atomic_incq(Address(scr, 0)); 1039 } 1040 } 1041 #endif 1042 1043 // Writes to stack successive pages until offset reached to check for 1044 // stack overflow + shadow pages. This clobbers tmp. 1045 void MacroAssembler::bang_stack_size(Register size, Register tmp) { 1046 movptr(tmp, rsp); 1047 // Bang stack for total size given plus shadow page size. 1048 // Bang one page at a time because large size can bang beyond yellow and 1049 // red zones. 1050 Label loop; 1051 bind(loop); 1052 movl(Address(tmp, (-os::vm_page_size())), size ); 1053 subptr(tmp, os::vm_page_size()); 1054 subl(size, os::vm_page_size()); 1055 jcc(Assembler::greater, loop); 1056 1057 // Bang down shadow pages too. 1058 // At this point, (tmp-0) is the last address touched, so don't 1059 // touch it again. (It was touched as (tmp-pagesize) but then tmp 1060 // was post-decremented.) Skip this address by starting at i=1, and 1061 // touch a few more pages below. N.B. It is important to touch all 1062 // the way down to and including i=StackShadowPages. 1063 for (int i = 1; i < StackShadowPages; i++) { 1064 // this could be any sized move but this is can be a debugging crumb 1065 // so the bigger the better. 1066 movptr(Address(tmp, (-i*os::vm_page_size())), size ); 1067 } 1068 } 1069 1070 void MacroAssembler::reserved_stack_check() { 1071 // testing if reserved zone needs to be enabled 1072 Label no_reserved_zone_enabling; 1073 Register thread = NOT_LP64(rsi) LP64_ONLY(r15_thread); 1074 NOT_LP64(get_thread(rsi);) 1075 1076 cmpptr(rsp, Address(thread, JavaThread::reserved_stack_activation_offset())); 1077 jcc(Assembler::below, no_reserved_zone_enabling); 1078 1079 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), thread); 1080 jump(RuntimeAddress(StubRoutines::throw_delayed_StackOverflowError_entry())); 1081 should_not_reach_here(); 1082 1083 bind(no_reserved_zone_enabling); 1084 } 1085 1086 int MacroAssembler::biased_locking_enter(Register lock_reg, 1087 Register obj_reg, 1088 Register swap_reg, 1089 Register tmp_reg, 1090 bool swap_reg_contains_mark, 1091 Label& done, 1092 Label* slow_case, 1093 BiasedLockingCounters* counters) { 1094 assert(UseBiasedLocking, "why call this otherwise?"); 1095 assert(swap_reg == rax, "swap_reg must be rax for cmpxchgq"); 1096 assert(tmp_reg != noreg, "tmp_reg must be supplied"); 1097 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg); 1098 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout"); 1099 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes()); 1100 Address saved_mark_addr(lock_reg, 0); 1101 1102 if (PrintBiasedLockingStatistics && counters == NULL) { 1103 counters = BiasedLocking::counters(); 1104 } 1105 // Biased locking 1106 // See whether the lock is currently biased toward our thread and 1107 // whether the epoch is still valid 1108 // Note that the runtime guarantees sufficient alignment of JavaThread 1109 // pointers to allow age to be placed into low bits 1110 // First check to see whether biasing is even enabled for this object 1111 Label cas_label; 1112 int null_check_offset = -1; 1113 if (!swap_reg_contains_mark) { 1114 null_check_offset = offset(); 1115 movptr(swap_reg, mark_addr); 1116 } 1117 movptr(tmp_reg, swap_reg); 1118 andptr(tmp_reg, markOopDesc::biased_lock_mask_in_place); 1119 cmpptr(tmp_reg, markOopDesc::biased_lock_pattern); 1120 jcc(Assembler::notEqual, cas_label); 1121 // The bias pattern is present in the object's header. Need to check 1122 // whether the bias owner and the epoch are both still current. 1123 #ifndef _LP64 1124 // Note that because there is no current thread register on x86_32 we 1125 // need to store off the mark word we read out of the object to 1126 // avoid reloading it and needing to recheck invariants below. This 1127 // store is unfortunate but it makes the overall code shorter and 1128 // simpler. 1129 movptr(saved_mark_addr, swap_reg); 1130 #endif 1131 if (swap_reg_contains_mark) { 1132 null_check_offset = offset(); 1133 } 1134 load_prototype_header(tmp_reg, obj_reg); 1135 #ifdef _LP64 1136 orptr(tmp_reg, r15_thread); 1137 xorptr(tmp_reg, swap_reg); 1138 Register header_reg = tmp_reg; 1139 #else 1140 xorptr(tmp_reg, swap_reg); 1141 get_thread(swap_reg); 1142 xorptr(swap_reg, tmp_reg); 1143 Register header_reg = swap_reg; 1144 #endif 1145 andptr(header_reg, ~((int) markOopDesc::age_mask_in_place)); 1146 if (counters != NULL) { 1147 cond_inc32(Assembler::zero, 1148 ExternalAddress((address) counters->biased_lock_entry_count_addr())); 1149 } 1150 jcc(Assembler::equal, done); 1151 1152 Label try_revoke_bias; 1153 Label try_rebias; 1154 1155 // At this point we know that the header has the bias pattern and 1156 // that we are not the bias owner in the current epoch. We need to 1157 // figure out more details about the state of the header in order to 1158 // know what operations can be legally performed on the object's 1159 // header. 1160 1161 // If the low three bits in the xor result aren't clear, that means 1162 // the prototype header is no longer biased and we have to revoke 1163 // the bias on this object. 1164 testptr(header_reg, markOopDesc::biased_lock_mask_in_place); 1165 jccb(Assembler::notZero, try_revoke_bias); 1166 1167 // Biasing is still enabled for this data type. See whether the 1168 // epoch of the current bias is still valid, meaning that the epoch 1169 // bits of the mark word are equal to the epoch bits of the 1170 // prototype header. (Note that the prototype header's epoch bits 1171 // only change at a safepoint.) If not, attempt to rebias the object 1172 // toward the current thread. Note that we must be absolutely sure 1173 // that the current epoch is invalid in order to do this because 1174 // otherwise the manipulations it performs on the mark word are 1175 // illegal. 1176 testptr(header_reg, markOopDesc::epoch_mask_in_place); 1177 jccb(Assembler::notZero, try_rebias); 1178 1179 // The epoch of the current bias is still valid but we know nothing 1180 // about the owner; it might be set or it might be clear. Try to 1181 // acquire the bias of the object using an atomic operation. If this 1182 // fails we will go in to the runtime to revoke the object's bias. 1183 // Note that we first construct the presumed unbiased header so we 1184 // don't accidentally blow away another thread's valid bias. 1185 NOT_LP64( movptr(swap_reg, saved_mark_addr); ) 1186 andptr(swap_reg, 1187 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place); 1188 #ifdef _LP64 1189 movptr(tmp_reg, swap_reg); 1190 orptr(tmp_reg, r15_thread); 1191 #else 1192 get_thread(tmp_reg); 1193 orptr(tmp_reg, swap_reg); 1194 #endif 1195 if (os::is_MP()) { 1196 lock(); 1197 } 1198 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1199 // If the biasing toward our thread failed, this means that 1200 // another thread succeeded in biasing it toward itself and we 1201 // need to revoke that bias. The revocation will occur in the 1202 // interpreter runtime in the slow case. 1203 if (counters != NULL) { 1204 cond_inc32(Assembler::zero, 1205 ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr())); 1206 } 1207 if (slow_case != NULL) { 1208 jcc(Assembler::notZero, *slow_case); 1209 } 1210 jmp(done); 1211 1212 bind(try_rebias); 1213 // At this point we know the epoch has expired, meaning that the 1214 // current "bias owner", if any, is actually invalid. Under these 1215 // circumstances _only_, we are allowed to use the current header's 1216 // value as the comparison value when doing the cas to acquire the 1217 // bias in the current epoch. In other words, we allow transfer of 1218 // the bias from one thread to another directly in this situation. 1219 // 1220 // FIXME: due to a lack of registers we currently blow away the age 1221 // bits in this situation. Should attempt to preserve them. 1222 load_prototype_header(tmp_reg, obj_reg); 1223 #ifdef _LP64 1224 orptr(tmp_reg, r15_thread); 1225 #else 1226 get_thread(swap_reg); 1227 orptr(tmp_reg, swap_reg); 1228 movptr(swap_reg, saved_mark_addr); 1229 #endif 1230 if (os::is_MP()) { 1231 lock(); 1232 } 1233 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1234 // If the biasing toward our thread failed, then another thread 1235 // succeeded in biasing it toward itself and we need to revoke that 1236 // bias. The revocation will occur in the runtime in the slow case. 1237 if (counters != NULL) { 1238 cond_inc32(Assembler::zero, 1239 ExternalAddress((address) counters->rebiased_lock_entry_count_addr())); 1240 } 1241 if (slow_case != NULL) { 1242 jcc(Assembler::notZero, *slow_case); 1243 } 1244 jmp(done); 1245 1246 bind(try_revoke_bias); 1247 // The prototype mark in the klass doesn't have the bias bit set any 1248 // more, indicating that objects of this data type are not supposed 1249 // to be biased any more. We are going to try to reset the mark of 1250 // this object to the prototype value and fall through to the 1251 // CAS-based locking scheme. Note that if our CAS fails, it means 1252 // that another thread raced us for the privilege of revoking the 1253 // bias of this particular object, so it's okay to continue in the 1254 // normal locking code. 1255 // 1256 // FIXME: due to a lack of registers we currently blow away the age 1257 // bits in this situation. Should attempt to preserve them. 1258 NOT_LP64( movptr(swap_reg, saved_mark_addr); ) 1259 load_prototype_header(tmp_reg, obj_reg); 1260 if (os::is_MP()) { 1261 lock(); 1262 } 1263 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1264 // Fall through to the normal CAS-based lock, because no matter what 1265 // the result of the above CAS, some thread must have succeeded in 1266 // removing the bias bit from the object's header. 1267 if (counters != NULL) { 1268 cond_inc32(Assembler::zero, 1269 ExternalAddress((address) counters->revoked_lock_entry_count_addr())); 1270 } 1271 1272 bind(cas_label); 1273 1274 return null_check_offset; 1275 } 1276 1277 void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) { 1278 assert(UseBiasedLocking, "why call this otherwise?"); 1279 1280 // Check for biased locking unlock case, which is a no-op 1281 // Note: we do not have to check the thread ID for two reasons. 1282 // First, the interpreter checks for IllegalMonitorStateException at 1283 // a higher level. Second, if the bias was revoked while we held the 1284 // lock, the object could not be rebiased toward another thread, so 1285 // the bias bit would be clear. 1286 movptr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 1287 andptr(temp_reg, markOopDesc::biased_lock_mask_in_place); 1288 cmpptr(temp_reg, markOopDesc::biased_lock_pattern); 1289 jcc(Assembler::equal, done); 1290 } 1291 1292 #ifdef COMPILER2 1293 1294 #if INCLUDE_RTM_OPT 1295 1296 // Update rtm_counters based on abort status 1297 // input: abort_status 1298 // rtm_counters (RTMLockingCounters*) 1299 // flags are killed 1300 void MacroAssembler::rtm_counters_update(Register abort_status, Register rtm_counters) { 1301 1302 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abort_count_offset())); 1303 if (PrintPreciseRTMLockingStatistics) { 1304 for (int i = 0; i < RTMLockingCounters::ABORT_STATUS_LIMIT; i++) { 1305 Label check_abort; 1306 testl(abort_status, (1<<i)); 1307 jccb(Assembler::equal, check_abort); 1308 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abortX_count_offset() + (i * sizeof(uintx)))); 1309 bind(check_abort); 1310 } 1311 } 1312 } 1313 1314 // Branch if (random & (count-1) != 0), count is 2^n 1315 // tmp, scr and flags are killed 1316 void MacroAssembler::branch_on_random_using_rdtsc(Register tmp, Register scr, int count, Label& brLabel) { 1317 assert(tmp == rax, ""); 1318 assert(scr == rdx, ""); 1319 rdtsc(); // modifies EDX:EAX 1320 andptr(tmp, count-1); 1321 jccb(Assembler::notZero, brLabel); 1322 } 1323 1324 // Perform abort ratio calculation, set no_rtm bit if high ratio 1325 // input: rtm_counters_Reg (RTMLockingCounters* address) 1326 // tmpReg, rtm_counters_Reg and flags are killed 1327 void MacroAssembler::rtm_abort_ratio_calculation(Register tmpReg, 1328 Register rtm_counters_Reg, 1329 RTMLockingCounters* rtm_counters, 1330 Metadata* method_data) { 1331 Label L_done, L_check_always_rtm1, L_check_always_rtm2; 1332 1333 if (RTMLockingCalculationDelay > 0) { 1334 // Delay calculation 1335 movptr(tmpReg, ExternalAddress((address) RTMLockingCounters::rtm_calculation_flag_addr()), tmpReg); 1336 testptr(tmpReg, tmpReg); 1337 jccb(Assembler::equal, L_done); 1338 } 1339 // Abort ratio calculation only if abort_count > RTMAbortThreshold 1340 // Aborted transactions = abort_count * 100 1341 // All transactions = total_count * RTMTotalCountIncrRate 1342 // Set no_rtm bit if (Aborted transactions >= All transactions * RTMAbortRatio) 1343 1344 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::abort_count_offset())); 1345 cmpptr(tmpReg, RTMAbortThreshold); 1346 jccb(Assembler::below, L_check_always_rtm2); 1347 imulptr(tmpReg, tmpReg, 100); 1348 1349 Register scrReg = rtm_counters_Reg; 1350 movptr(scrReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset())); 1351 imulptr(scrReg, scrReg, RTMTotalCountIncrRate); 1352 imulptr(scrReg, scrReg, RTMAbortRatio); 1353 cmpptr(tmpReg, scrReg); 1354 jccb(Assembler::below, L_check_always_rtm1); 1355 if (method_data != NULL) { 1356 // set rtm_state to "no rtm" in MDO 1357 mov_metadata(tmpReg, method_data); 1358 if (os::is_MP()) { 1359 lock(); 1360 } 1361 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), NoRTM); 1362 } 1363 jmpb(L_done); 1364 bind(L_check_always_rtm1); 1365 // Reload RTMLockingCounters* address 1366 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters)); 1367 bind(L_check_always_rtm2); 1368 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset())); 1369 cmpptr(tmpReg, RTMLockingThreshold / RTMTotalCountIncrRate); 1370 jccb(Assembler::below, L_done); 1371 if (method_data != NULL) { 1372 // set rtm_state to "always rtm" in MDO 1373 mov_metadata(tmpReg, method_data); 1374 if (os::is_MP()) { 1375 lock(); 1376 } 1377 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), UseRTM); 1378 } 1379 bind(L_done); 1380 } 1381 1382 // Update counters and perform abort ratio calculation 1383 // input: abort_status_Reg 1384 // rtm_counters_Reg, flags are killed 1385 void MacroAssembler::rtm_profiling(Register abort_status_Reg, 1386 Register rtm_counters_Reg, 1387 RTMLockingCounters* rtm_counters, 1388 Metadata* method_data, 1389 bool profile_rtm) { 1390 1391 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1392 // update rtm counters based on rax value at abort 1393 // reads abort_status_Reg, updates flags 1394 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters)); 1395 rtm_counters_update(abort_status_Reg, rtm_counters_Reg); 1396 if (profile_rtm) { 1397 // Save abort status because abort_status_Reg is used by following code. 1398 if (RTMRetryCount > 0) { 1399 push(abort_status_Reg); 1400 } 1401 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1402 rtm_abort_ratio_calculation(abort_status_Reg, rtm_counters_Reg, rtm_counters, method_data); 1403 // restore abort status 1404 if (RTMRetryCount > 0) { 1405 pop(abort_status_Reg); 1406 } 1407 } 1408 } 1409 1410 // Retry on abort if abort's status is 0x6: can retry (0x2) | memory conflict (0x4) 1411 // inputs: retry_count_Reg 1412 // : abort_status_Reg 1413 // output: retry_count_Reg decremented by 1 1414 // flags are killed 1415 void MacroAssembler::rtm_retry_lock_on_abort(Register retry_count_Reg, Register abort_status_Reg, Label& retryLabel) { 1416 Label doneRetry; 1417 assert(abort_status_Reg == rax, ""); 1418 // The abort reason bits are in eax (see all states in rtmLocking.hpp) 1419 // 0x6 = conflict on which we can retry (0x2) | memory conflict (0x4) 1420 // if reason is in 0x6 and retry count != 0 then retry 1421 andptr(abort_status_Reg, 0x6); 1422 jccb(Assembler::zero, doneRetry); 1423 testl(retry_count_Reg, retry_count_Reg); 1424 jccb(Assembler::zero, doneRetry); 1425 pause(); 1426 decrementl(retry_count_Reg); 1427 jmp(retryLabel); 1428 bind(doneRetry); 1429 } 1430 1431 // Spin and retry if lock is busy, 1432 // inputs: box_Reg (monitor address) 1433 // : retry_count_Reg 1434 // output: retry_count_Reg decremented by 1 1435 // : clear z flag if retry count exceeded 1436 // tmp_Reg, scr_Reg, flags are killed 1437 void MacroAssembler::rtm_retry_lock_on_busy(Register retry_count_Reg, Register box_Reg, 1438 Register tmp_Reg, Register scr_Reg, Label& retryLabel) { 1439 Label SpinLoop, SpinExit, doneRetry; 1440 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1441 1442 testl(retry_count_Reg, retry_count_Reg); 1443 jccb(Assembler::zero, doneRetry); 1444 decrementl(retry_count_Reg); 1445 movptr(scr_Reg, RTMSpinLoopCount); 1446 1447 bind(SpinLoop); 1448 pause(); 1449 decrementl(scr_Reg); 1450 jccb(Assembler::lessEqual, SpinExit); 1451 movptr(tmp_Reg, Address(box_Reg, owner_offset)); 1452 testptr(tmp_Reg, tmp_Reg); 1453 jccb(Assembler::notZero, SpinLoop); 1454 1455 bind(SpinExit); 1456 jmp(retryLabel); 1457 bind(doneRetry); 1458 incrementl(retry_count_Reg); // clear z flag 1459 } 1460 1461 // Use RTM for normal stack locks 1462 // Input: objReg (object to lock) 1463 void MacroAssembler::rtm_stack_locking(Register objReg, Register tmpReg, Register scrReg, 1464 Register retry_on_abort_count_Reg, 1465 RTMLockingCounters* stack_rtm_counters, 1466 Metadata* method_data, bool profile_rtm, 1467 Label& DONE_LABEL, Label& IsInflated) { 1468 assert(UseRTMForStackLocks, "why call this otherwise?"); 1469 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking"); 1470 assert(tmpReg == rax, ""); 1471 assert(scrReg == rdx, ""); 1472 Label L_rtm_retry, L_decrement_retry, L_on_abort; 1473 1474 if (RTMRetryCount > 0) { 1475 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort 1476 bind(L_rtm_retry); 1477 } 1478 movptr(tmpReg, Address(objReg, 0)); 1479 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased 1480 jcc(Assembler::notZero, IsInflated); 1481 1482 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1483 Label L_noincrement; 1484 if (RTMTotalCountIncrRate > 1) { 1485 // tmpReg, scrReg and flags are killed 1486 branch_on_random_using_rdtsc(tmpReg, scrReg, (int)RTMTotalCountIncrRate, L_noincrement); 1487 } 1488 assert(stack_rtm_counters != NULL, "should not be NULL when profiling RTM"); 1489 atomic_incptr(ExternalAddress((address)stack_rtm_counters->total_count_addr()), scrReg); 1490 bind(L_noincrement); 1491 } 1492 xbegin(L_on_abort); 1493 movptr(tmpReg, Address(objReg, 0)); // fetch markword 1494 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits 1495 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked 1496 jcc(Assembler::equal, DONE_LABEL); // all done if unlocked 1497 1498 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX 1499 if (UseRTMXendForLockBusy) { 1500 xend(); 1501 movptr(abort_status_Reg, 0x2); // Set the abort status to 2 (so we can retry) 1502 jmp(L_decrement_retry); 1503 } 1504 else { 1505 xabort(0); 1506 } 1507 bind(L_on_abort); 1508 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1509 rtm_profiling(abort_status_Reg, scrReg, stack_rtm_counters, method_data, profile_rtm); 1510 } 1511 bind(L_decrement_retry); 1512 if (RTMRetryCount > 0) { 1513 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4) 1514 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry); 1515 } 1516 } 1517 1518 // Use RTM for inflating locks 1519 // inputs: objReg (object to lock) 1520 // boxReg (on-stack box address (displaced header location) - KILLED) 1521 // tmpReg (ObjectMonitor address + markOopDesc::monitor_value) 1522 void MacroAssembler::rtm_inflated_locking(Register objReg, Register boxReg, Register tmpReg, 1523 Register scrReg, Register retry_on_busy_count_Reg, 1524 Register retry_on_abort_count_Reg, 1525 RTMLockingCounters* rtm_counters, 1526 Metadata* method_data, bool profile_rtm, 1527 Label& DONE_LABEL) { 1528 assert(UseRTMLocking, "why call this otherwise?"); 1529 assert(tmpReg == rax, ""); 1530 assert(scrReg == rdx, ""); 1531 Label L_rtm_retry, L_decrement_retry, L_on_abort; 1532 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1533 1534 // Without cast to int32_t a movptr will destroy r10 which is typically obj 1535 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark())); 1536 movptr(boxReg, tmpReg); // Save ObjectMonitor address 1537 1538 if (RTMRetryCount > 0) { 1539 movl(retry_on_busy_count_Reg, RTMRetryCount); // Retry on lock busy 1540 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort 1541 bind(L_rtm_retry); 1542 } 1543 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1544 Label L_noincrement; 1545 if (RTMTotalCountIncrRate > 1) { 1546 // tmpReg, scrReg and flags are killed 1547 branch_on_random_using_rdtsc(tmpReg, scrReg, (int)RTMTotalCountIncrRate, L_noincrement); 1548 } 1549 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1550 atomic_incptr(ExternalAddress((address)rtm_counters->total_count_addr()), scrReg); 1551 bind(L_noincrement); 1552 } 1553 xbegin(L_on_abort); 1554 movptr(tmpReg, Address(objReg, 0)); 1555 movptr(tmpReg, Address(tmpReg, owner_offset)); 1556 testptr(tmpReg, tmpReg); 1557 jcc(Assembler::zero, DONE_LABEL); 1558 if (UseRTMXendForLockBusy) { 1559 xend(); 1560 jmp(L_decrement_retry); 1561 } 1562 else { 1563 xabort(0); 1564 } 1565 bind(L_on_abort); 1566 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX 1567 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1568 rtm_profiling(abort_status_Reg, scrReg, rtm_counters, method_data, profile_rtm); 1569 } 1570 if (RTMRetryCount > 0) { 1571 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4) 1572 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry); 1573 } 1574 1575 movptr(tmpReg, Address(boxReg, owner_offset)) ; 1576 testptr(tmpReg, tmpReg) ; 1577 jccb(Assembler::notZero, L_decrement_retry) ; 1578 1579 // Appears unlocked - try to swing _owner from null to non-null. 1580 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1581 #ifdef _LP64 1582 Register threadReg = r15_thread; 1583 #else 1584 get_thread(scrReg); 1585 Register threadReg = scrReg; 1586 #endif 1587 if (os::is_MP()) { 1588 lock(); 1589 } 1590 cmpxchgptr(threadReg, Address(boxReg, owner_offset)); // Updates tmpReg 1591 1592 if (RTMRetryCount > 0) { 1593 // success done else retry 1594 jccb(Assembler::equal, DONE_LABEL) ; 1595 bind(L_decrement_retry); 1596 // Spin and retry if lock is busy. 1597 rtm_retry_lock_on_busy(retry_on_busy_count_Reg, boxReg, tmpReg, scrReg, L_rtm_retry); 1598 } 1599 else { 1600 bind(L_decrement_retry); 1601 } 1602 } 1603 1604 #endif // INCLUDE_RTM_OPT 1605 1606 // Fast_Lock and Fast_Unlock used by C2 1607 1608 // Because the transitions from emitted code to the runtime 1609 // monitorenter/exit helper stubs are so slow it's critical that 1610 // we inline both the stack-locking fast-path and the inflated fast path. 1611 // 1612 // See also: cmpFastLock and cmpFastUnlock. 1613 // 1614 // What follows is a specialized inline transliteration of the code 1615 // in slow_enter() and slow_exit(). If we're concerned about I$ bloat 1616 // another option would be to emit TrySlowEnter and TrySlowExit methods 1617 // at startup-time. These methods would accept arguments as 1618 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure 1619 // indications in the icc.ZFlag. Fast_Lock and Fast_Unlock would simply 1620 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit. 1621 // In practice, however, the # of lock sites is bounded and is usually small. 1622 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer 1623 // if the processor uses simple bimodal branch predictors keyed by EIP 1624 // Since the helper routines would be called from multiple synchronization 1625 // sites. 1626 // 1627 // An even better approach would be write "MonitorEnter()" and "MonitorExit()" 1628 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites 1629 // to those specialized methods. That'd give us a mostly platform-independent 1630 // implementation that the JITs could optimize and inline at their pleasure. 1631 // Done correctly, the only time we'd need to cross to native could would be 1632 // to park() or unpark() threads. We'd also need a few more unsafe operators 1633 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and 1634 // (b) explicit barriers or fence operations. 1635 // 1636 // TODO: 1637 // 1638 // * Arrange for C2 to pass "Self" into Fast_Lock and Fast_Unlock in one of the registers (scr). 1639 // This avoids manifesting the Self pointer in the Fast_Lock and Fast_Unlock terminals. 1640 // Given TLAB allocation, Self is usually manifested in a register, so passing it into 1641 // the lock operators would typically be faster than reifying Self. 1642 // 1643 // * Ideally I'd define the primitives as: 1644 // fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED. 1645 // fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED 1646 // Unfortunately ADLC bugs prevent us from expressing the ideal form. 1647 // Instead, we're stuck with a rather awkward and brittle register assignments below. 1648 // Furthermore the register assignments are overconstrained, possibly resulting in 1649 // sub-optimal code near the synchronization site. 1650 // 1651 // * Eliminate the sp-proximity tests and just use "== Self" tests instead. 1652 // Alternately, use a better sp-proximity test. 1653 // 1654 // * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value. 1655 // Either one is sufficient to uniquely identify a thread. 1656 // TODO: eliminate use of sp in _owner and use get_thread(tr) instead. 1657 // 1658 // * Intrinsify notify() and notifyAll() for the common cases where the 1659 // object is locked by the calling thread but the waitlist is empty. 1660 // avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll(). 1661 // 1662 // * use jccb and jmpb instead of jcc and jmp to improve code density. 1663 // But beware of excessive branch density on AMD Opterons. 1664 // 1665 // * Both Fast_Lock and Fast_Unlock set the ICC.ZF to indicate success 1666 // or failure of the fast-path. If the fast-path fails then we pass 1667 // control to the slow-path, typically in C. In Fast_Lock and 1668 // Fast_Unlock we often branch to DONE_LABEL, just to find that C2 1669 // will emit a conditional branch immediately after the node. 1670 // So we have branches to branches and lots of ICC.ZF games. 1671 // Instead, it might be better to have C2 pass a "FailureLabel" 1672 // into Fast_Lock and Fast_Unlock. In the case of success, control 1673 // will drop through the node. ICC.ZF is undefined at exit. 1674 // In the case of failure, the node will branch directly to the 1675 // FailureLabel 1676 1677 1678 // obj: object to lock 1679 // box: on-stack box address (displaced header location) - KILLED 1680 // rax,: tmp -- KILLED 1681 // scr: tmp -- KILLED 1682 void MacroAssembler::fast_lock(Register objReg, Register boxReg, Register tmpReg, 1683 Register scrReg, Register cx1Reg, Register cx2Reg, 1684 BiasedLockingCounters* counters, 1685 RTMLockingCounters* rtm_counters, 1686 RTMLockingCounters* stack_rtm_counters, 1687 Metadata* method_data, 1688 bool use_rtm, bool profile_rtm) { 1689 // Ensure the register assignents are disjoint 1690 assert(tmpReg == rax, ""); 1691 1692 if (use_rtm) { 1693 assert_different_registers(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg); 1694 } else { 1695 assert(cx1Reg == noreg, ""); 1696 assert(cx2Reg == noreg, ""); 1697 assert_different_registers(objReg, boxReg, tmpReg, scrReg); 1698 } 1699 1700 if (counters != NULL) { 1701 atomic_incl(ExternalAddress((address)counters->total_entry_count_addr()), scrReg); 1702 } 1703 if (EmitSync & 1) { 1704 // set box->dhw = markOopDesc::unused_mark() 1705 // Force all sync thru slow-path: slow_enter() and slow_exit() 1706 movptr (Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark())); 1707 cmpptr (rsp, (int32_t)NULL_WORD); 1708 } else { 1709 // Possible cases that we'll encounter in fast_lock 1710 // ------------------------------------------------ 1711 // * Inflated 1712 // -- unlocked 1713 // -- Locked 1714 // = by self 1715 // = by other 1716 // * biased 1717 // -- by Self 1718 // -- by other 1719 // * neutral 1720 // * stack-locked 1721 // -- by self 1722 // = sp-proximity test hits 1723 // = sp-proximity test generates false-negative 1724 // -- by other 1725 // 1726 1727 Label IsInflated, DONE_LABEL; 1728 1729 // it's stack-locked, biased or neutral 1730 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage 1731 // order to reduce the number of conditional branches in the most common cases. 1732 // Beware -- there's a subtle invariant that fetch of the markword 1733 // at [FETCH], below, will never observe a biased encoding (*101b). 1734 // If this invariant is not held we risk exclusion (safety) failure. 1735 if (UseBiasedLocking && !UseOptoBiasInlining) { 1736 biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, counters); 1737 } 1738 1739 #if INCLUDE_RTM_OPT 1740 if (UseRTMForStackLocks && use_rtm) { 1741 rtm_stack_locking(objReg, tmpReg, scrReg, cx2Reg, 1742 stack_rtm_counters, method_data, profile_rtm, 1743 DONE_LABEL, IsInflated); 1744 } 1745 #endif // INCLUDE_RTM_OPT 1746 1747 movptr(tmpReg, Address(objReg, 0)); // [FETCH] 1748 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased 1749 jccb(Assembler::notZero, IsInflated); 1750 1751 // Attempt stack-locking ... 1752 orptr (tmpReg, markOopDesc::unlocked_value); 1753 movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS 1754 if (os::is_MP()) { 1755 lock(); 1756 } 1757 cmpxchgptr(boxReg, Address(objReg, 0)); // Updates tmpReg 1758 if (counters != NULL) { 1759 cond_inc32(Assembler::equal, 1760 ExternalAddress((address)counters->fast_path_entry_count_addr())); 1761 } 1762 jcc(Assembler::equal, DONE_LABEL); // Success 1763 1764 // Recursive locking. 1765 // The object is stack-locked: markword contains stack pointer to BasicLock. 1766 // Locked by current thread if difference with current SP is less than one page. 1767 subptr(tmpReg, rsp); 1768 // Next instruction set ZFlag == 1 (Success) if difference is less then one page. 1769 andptr(tmpReg, (int32_t) (NOT_LP64(0xFFFFF003) LP64_ONLY(7 - os::vm_page_size())) ); 1770 movptr(Address(boxReg, 0), tmpReg); 1771 if (counters != NULL) { 1772 cond_inc32(Assembler::equal, 1773 ExternalAddress((address)counters->fast_path_entry_count_addr())); 1774 } 1775 jmp(DONE_LABEL); 1776 1777 bind(IsInflated); 1778 // The object is inflated. tmpReg contains pointer to ObjectMonitor* + markOopDesc::monitor_value 1779 1780 #if INCLUDE_RTM_OPT 1781 // Use the same RTM locking code in 32- and 64-bit VM. 1782 if (use_rtm) { 1783 rtm_inflated_locking(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg, 1784 rtm_counters, method_data, profile_rtm, DONE_LABEL); 1785 } else { 1786 #endif // INCLUDE_RTM_OPT 1787 1788 #ifndef _LP64 1789 // The object is inflated. 1790 1791 // boxReg refers to the on-stack BasicLock in the current frame. 1792 // We'd like to write: 1793 // set box->_displaced_header = markOopDesc::unused_mark(). Any non-0 value suffices. 1794 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers 1795 // additional latency as we have another ST in the store buffer that must drain. 1796 1797 if (EmitSync & 8192) { 1798 movptr(Address(boxReg, 0), 3); // results in ST-before-CAS penalty 1799 get_thread (scrReg); 1800 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2] 1801 movptr(tmpReg, NULL_WORD); // consider: xor vs mov 1802 if (os::is_MP()) { 1803 lock(); 1804 } 1805 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1806 } else 1807 if ((EmitSync & 128) == 0) { // avoid ST-before-CAS 1808 // register juggle because we need tmpReg for cmpxchgptr below 1809 movptr(scrReg, boxReg); 1810 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2] 1811 1812 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes 1813 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) { 1814 // prefetchw [eax + Offset(_owner)-2] 1815 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1816 } 1817 1818 if ((EmitSync & 64) == 0) { 1819 // Optimistic form: consider XORL tmpReg,tmpReg 1820 movptr(tmpReg, NULL_WORD); 1821 } else { 1822 // Can suffer RTS->RTO upgrades on shared or cold $ lines 1823 // Test-And-CAS instead of CAS 1824 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner 1825 testptr(tmpReg, tmpReg); // Locked ? 1826 jccb (Assembler::notZero, DONE_LABEL); 1827 } 1828 1829 // Appears unlocked - try to swing _owner from null to non-null. 1830 // Ideally, I'd manifest "Self" with get_thread and then attempt 1831 // to CAS the register containing Self into m->Owner. 1832 // But we don't have enough registers, so instead we can either try to CAS 1833 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds 1834 // we later store "Self" into m->Owner. Transiently storing a stack address 1835 // (rsp or the address of the box) into m->owner is harmless. 1836 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1837 if (os::is_MP()) { 1838 lock(); 1839 } 1840 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1841 movptr(Address(scrReg, 0), 3); // box->_displaced_header = 3 1842 // If we weren't able to swing _owner from NULL to the BasicLock 1843 // then take the slow path. 1844 jccb (Assembler::notZero, DONE_LABEL); 1845 // update _owner from BasicLock to thread 1846 get_thread (scrReg); // beware: clobbers ICCs 1847 movptr(Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), scrReg); 1848 xorptr(boxReg, boxReg); // set icc.ZFlag = 1 to indicate success 1849 1850 // If the CAS fails we can either retry or pass control to the slow-path. 1851 // We use the latter tactic. 1852 // Pass the CAS result in the icc.ZFlag into DONE_LABEL 1853 // If the CAS was successful ... 1854 // Self has acquired the lock 1855 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it. 1856 // Intentional fall-through into DONE_LABEL ... 1857 } else { 1858 movptr(Address(boxReg, 0), intptr_t(markOopDesc::unused_mark())); // results in ST-before-CAS penalty 1859 movptr(boxReg, tmpReg); 1860 1861 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes 1862 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) { 1863 // prefetchw [eax + Offset(_owner)-2] 1864 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1865 } 1866 1867 if ((EmitSync & 64) == 0) { 1868 // Optimistic form 1869 xorptr (tmpReg, tmpReg); 1870 } else { 1871 // Can suffer RTS->RTO upgrades on shared or cold $ lines 1872 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner 1873 testptr(tmpReg, tmpReg); // Locked ? 1874 jccb (Assembler::notZero, DONE_LABEL); 1875 } 1876 1877 // Appears unlocked - try to swing _owner from null to non-null. 1878 // Use either "Self" (in scr) or rsp as thread identity in _owner. 1879 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1880 get_thread (scrReg); 1881 if (os::is_MP()) { 1882 lock(); 1883 } 1884 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1885 1886 // If the CAS fails we can either retry or pass control to the slow-path. 1887 // We use the latter tactic. 1888 // Pass the CAS result in the icc.ZFlag into DONE_LABEL 1889 // If the CAS was successful ... 1890 // Self has acquired the lock 1891 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it. 1892 // Intentional fall-through into DONE_LABEL ... 1893 } 1894 #else // _LP64 1895 // It's inflated 1896 movq(scrReg, tmpReg); 1897 xorq(tmpReg, tmpReg); 1898 1899 if (os::is_MP()) { 1900 lock(); 1901 } 1902 cmpxchgptr(r15_thread, Address(scrReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1903 // Unconditionally set box->_displaced_header = markOopDesc::unused_mark(). 1904 // Without cast to int32_t movptr will destroy r10 which is typically obj. 1905 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark())); 1906 // Intentional fall-through into DONE_LABEL ... 1907 // Propagate ICC.ZF from CAS above into DONE_LABEL. 1908 #endif // _LP64 1909 #if INCLUDE_RTM_OPT 1910 } // use_rtm() 1911 #endif 1912 // DONE_LABEL is a hot target - we'd really like to place it at the 1913 // start of cache line by padding with NOPs. 1914 // See the AMD and Intel software optimization manuals for the 1915 // most efficient "long" NOP encodings. 1916 // Unfortunately none of our alignment mechanisms suffice. 1917 bind(DONE_LABEL); 1918 1919 // At DONE_LABEL the icc ZFlag is set as follows ... 1920 // Fast_Unlock uses the same protocol. 1921 // ZFlag == 1 -> Success 1922 // ZFlag == 0 -> Failure - force control through the slow-path 1923 } 1924 } 1925 1926 // obj: object to unlock 1927 // box: box address (displaced header location), killed. Must be EAX. 1928 // tmp: killed, cannot be obj nor box. 1929 // 1930 // Some commentary on balanced locking: 1931 // 1932 // Fast_Lock and Fast_Unlock are emitted only for provably balanced lock sites. 1933 // Methods that don't have provably balanced locking are forced to run in the 1934 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock. 1935 // The interpreter provides two properties: 1936 // I1: At return-time the interpreter automatically and quietly unlocks any 1937 // objects acquired the current activation (frame). Recall that the 1938 // interpreter maintains an on-stack list of locks currently held by 1939 // a frame. 1940 // I2: If a method attempts to unlock an object that is not held by the 1941 // the frame the interpreter throws IMSX. 1942 // 1943 // Lets say A(), which has provably balanced locking, acquires O and then calls B(). 1944 // B() doesn't have provably balanced locking so it runs in the interpreter. 1945 // Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O 1946 // is still locked by A(). 1947 // 1948 // The only other source of unbalanced locking would be JNI. The "Java Native Interface: 1949 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter 1950 // should not be unlocked by "normal" java-level locking and vice-versa. The specification 1951 // doesn't specify what will occur if a program engages in such mixed-mode locking, however. 1952 // Arguably given that the spec legislates the JNI case as undefined our implementation 1953 // could reasonably *avoid* checking owner in Fast_Unlock(). 1954 // In the interest of performance we elide m->Owner==Self check in unlock. 1955 // A perfectly viable alternative is to elide the owner check except when 1956 // Xcheck:jni is enabled. 1957 1958 void MacroAssembler::fast_unlock(Register objReg, Register boxReg, Register tmpReg, bool use_rtm) { 1959 assert(boxReg == rax, ""); 1960 assert_different_registers(objReg, boxReg, tmpReg); 1961 1962 if (EmitSync & 4) { 1963 // Disable - inhibit all inlining. Force control through the slow-path 1964 cmpptr (rsp, 0); 1965 } else { 1966 Label DONE_LABEL, Stacked, CheckSucc; 1967 1968 // Critically, the biased locking test must have precedence over 1969 // and appear before the (box->dhw == 0) recursive stack-lock test. 1970 if (UseBiasedLocking && !UseOptoBiasInlining) { 1971 biased_locking_exit(objReg, tmpReg, DONE_LABEL); 1972 } 1973 1974 #if INCLUDE_RTM_OPT 1975 if (UseRTMForStackLocks && use_rtm) { 1976 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking"); 1977 Label L_regular_unlock; 1978 movptr(tmpReg, Address(objReg, 0)); // fetch markword 1979 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits 1980 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked 1981 jccb(Assembler::notEqual, L_regular_unlock); // if !HLE RegularLock 1982 xend(); // otherwise end... 1983 jmp(DONE_LABEL); // ... and we're done 1984 bind(L_regular_unlock); 1985 } 1986 #endif 1987 1988 cmpptr(Address(boxReg, 0), (int32_t)NULL_WORD); // Examine the displaced header 1989 jcc (Assembler::zero, DONE_LABEL); // 0 indicates recursive stack-lock 1990 movptr(tmpReg, Address(objReg, 0)); // Examine the object's markword 1991 testptr(tmpReg, markOopDesc::monitor_value); // Inflated? 1992 jccb (Assembler::zero, Stacked); 1993 1994 // It's inflated. 1995 #if INCLUDE_RTM_OPT 1996 if (use_rtm) { 1997 Label L_regular_inflated_unlock; 1998 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1999 movptr(boxReg, Address(tmpReg, owner_offset)); 2000 testptr(boxReg, boxReg); 2001 jccb(Assembler::notZero, L_regular_inflated_unlock); 2002 xend(); 2003 jmpb(DONE_LABEL); 2004 bind(L_regular_inflated_unlock); 2005 } 2006 #endif 2007 2008 // Despite our balanced locking property we still check that m->_owner == Self 2009 // as java routines or native JNI code called by this thread might 2010 // have released the lock. 2011 // Refer to the comments in synchronizer.cpp for how we might encode extra 2012 // state in _succ so we can avoid fetching EntryList|cxq. 2013 // 2014 // I'd like to add more cases in fast_lock() and fast_unlock() -- 2015 // such as recursive enter and exit -- but we have to be wary of 2016 // I$ bloat, T$ effects and BP$ effects. 2017 // 2018 // If there's no contention try a 1-0 exit. That is, exit without 2019 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how 2020 // we detect and recover from the race that the 1-0 exit admits. 2021 // 2022 // Conceptually Fast_Unlock() must execute a STST|LDST "release" barrier 2023 // before it STs null into _owner, releasing the lock. Updates 2024 // to data protected by the critical section must be visible before 2025 // we drop the lock (and thus before any other thread could acquire 2026 // the lock and observe the fields protected by the lock). 2027 // IA32's memory-model is SPO, so STs are ordered with respect to 2028 // each other and there's no need for an explicit barrier (fence). 2029 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 2030 #ifndef _LP64 2031 get_thread (boxReg); 2032 if ((EmitSync & 4096) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) { 2033 // prefetchw [ebx + Offset(_owner)-2] 2034 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2035 } 2036 2037 // Note that we could employ various encoding schemes to reduce 2038 // the number of loads below (currently 4) to just 2 or 3. 2039 // Refer to the comments in synchronizer.cpp. 2040 // In practice the chain of fetches doesn't seem to impact performance, however. 2041 xorptr(boxReg, boxReg); 2042 if ((EmitSync & 65536) == 0 && (EmitSync & 256)) { 2043 // Attempt to reduce branch density - AMD's branch predictor. 2044 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 2045 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2046 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2047 jccb (Assembler::notZero, DONE_LABEL); 2048 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD); 2049 jmpb (DONE_LABEL); 2050 } else { 2051 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 2052 jccb (Assembler::notZero, DONE_LABEL); 2053 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2054 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2055 jccb (Assembler::notZero, CheckSucc); 2056 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD); 2057 jmpb (DONE_LABEL); 2058 } 2059 2060 // The Following code fragment (EmitSync & 65536) improves the performance of 2061 // contended applications and contended synchronization microbenchmarks. 2062 // Unfortunately the emission of the code - even though not executed - causes regressions 2063 // in scimark and jetstream, evidently because of $ effects. Replacing the code 2064 // with an equal number of never-executed NOPs results in the same regression. 2065 // We leave it off by default. 2066 2067 if ((EmitSync & 65536) != 0) { 2068 Label LSuccess, LGoSlowPath ; 2069 2070 bind (CheckSucc); 2071 2072 // Optional pre-test ... it's safe to elide this 2073 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2074 jccb(Assembler::zero, LGoSlowPath); 2075 2076 // We have a classic Dekker-style idiom: 2077 // ST m->_owner = 0 ; MEMBAR; LD m->_succ 2078 // There are a number of ways to implement the barrier: 2079 // (1) lock:andl &m->_owner, 0 2080 // is fast, but mask doesn't currently support the "ANDL M,IMM32" form. 2081 // LOCK: ANDL [ebx+Offset(_Owner)-2], 0 2082 // Encodes as 81 31 OFF32 IMM32 or 83 63 OFF8 IMM8 2083 // (2) If supported, an explicit MFENCE is appealing. 2084 // In older IA32 processors MFENCE is slower than lock:add or xchg 2085 // particularly if the write-buffer is full as might be the case if 2086 // if stores closely precede the fence or fence-equivalent instruction. 2087 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences 2088 // as the situation has changed with Nehalem and Shanghai. 2089 // (3) In lieu of an explicit fence, use lock:addl to the top-of-stack 2090 // The $lines underlying the top-of-stack should be in M-state. 2091 // The locked add instruction is serializing, of course. 2092 // (4) Use xchg, which is serializing 2093 // mov boxReg, 0; xchgl boxReg, [tmpReg + Offset(_owner)-2] also works 2094 // (5) ST m->_owner = 0 and then execute lock:orl &m->_succ, 0. 2095 // The integer condition codes will tell us if succ was 0. 2096 // Since _succ and _owner should reside in the same $line and 2097 // we just stored into _owner, it's likely that the $line 2098 // remains in M-state for the lock:orl. 2099 // 2100 // We currently use (3), although it's likely that switching to (2) 2101 // is correct for the future. 2102 2103 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD); 2104 if (os::is_MP()) { 2105 lock(); addptr(Address(rsp, 0), 0); 2106 } 2107 // Ratify _succ remains non-null 2108 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), 0); 2109 jccb (Assembler::notZero, LSuccess); 2110 2111 xorptr(boxReg, boxReg); // box is really EAX 2112 if (os::is_MP()) { lock(); } 2113 cmpxchgptr(rsp, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2114 // There's no successor so we tried to regrab the lock with the 2115 // placeholder value. If that didn't work, then another thread 2116 // grabbed the lock so we're done (and exit was a success). 2117 jccb (Assembler::notEqual, LSuccess); 2118 // Since we're low on registers we installed rsp as a placeholding in _owner. 2119 // Now install Self over rsp. This is safe as we're transitioning from 2120 // non-null to non=null 2121 get_thread (boxReg); 2122 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), boxReg); 2123 // Intentional fall-through into LGoSlowPath ... 2124 2125 bind (LGoSlowPath); 2126 orptr(boxReg, 1); // set ICC.ZF=0 to indicate failure 2127 jmpb (DONE_LABEL); 2128 2129 bind (LSuccess); 2130 xorptr(boxReg, boxReg); // set ICC.ZF=1 to indicate success 2131 jmpb (DONE_LABEL); 2132 } 2133 2134 bind (Stacked); 2135 // It's not inflated and it's not recursively stack-locked and it's not biased. 2136 // It must be stack-locked. 2137 // Try to reset the header to displaced header. 2138 // The "box" value on the stack is stable, so we can reload 2139 // and be assured we observe the same value as above. 2140 movptr(tmpReg, Address(boxReg, 0)); 2141 if (os::is_MP()) { 2142 lock(); 2143 } 2144 cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses RAX which is box 2145 // Intention fall-thru into DONE_LABEL 2146 2147 // DONE_LABEL is a hot target - we'd really like to place it at the 2148 // start of cache line by padding with NOPs. 2149 // See the AMD and Intel software optimization manuals for the 2150 // most efficient "long" NOP encodings. 2151 // Unfortunately none of our alignment mechanisms suffice. 2152 if ((EmitSync & 65536) == 0) { 2153 bind (CheckSucc); 2154 } 2155 #else // _LP64 2156 // It's inflated 2157 if (EmitSync & 1024) { 2158 // Emit code to check that _owner == Self 2159 // We could fold the _owner test into subsequent code more efficiently 2160 // than using a stand-alone check, but since _owner checking is off by 2161 // default we don't bother. We also might consider predicating the 2162 // _owner==Self check on Xcheck:jni or running on a debug build. 2163 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2164 xorptr(boxReg, r15_thread); 2165 } else { 2166 xorptr(boxReg, boxReg); 2167 } 2168 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 2169 jccb (Assembler::notZero, DONE_LABEL); 2170 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2171 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2172 jccb (Assembler::notZero, CheckSucc); 2173 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD); 2174 jmpb (DONE_LABEL); 2175 2176 if ((EmitSync & 65536) == 0) { 2177 // Try to avoid passing control into the slow_path ... 2178 Label LSuccess, LGoSlowPath ; 2179 bind (CheckSucc); 2180 2181 // The following optional optimization can be elided if necessary 2182 // Effectively: if (succ == null) goto SlowPath 2183 // The code reduces the window for a race, however, 2184 // and thus benefits performance. 2185 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2186 jccb (Assembler::zero, LGoSlowPath); 2187 2188 if ((EmitSync & 16) && os::is_MP()) { 2189 orptr(boxReg, boxReg); 2190 xchgptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2191 } else { 2192 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD); 2193 if (os::is_MP()) { 2194 // Memory barrier/fence 2195 // Dekker pivot point -- fulcrum : ST Owner; MEMBAR; LD Succ 2196 // Instead of MFENCE we use a dummy locked add of 0 to the top-of-stack. 2197 // This is faster on Nehalem and AMD Shanghai/Barcelona. 2198 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences 2199 // We might also restructure (ST Owner=0;barrier;LD _Succ) to 2200 // (mov box,0; xchgq box, &m->Owner; LD _succ) . 2201 lock(); addl(Address(rsp, 0), 0); 2202 } 2203 } 2204 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2205 jccb (Assembler::notZero, LSuccess); 2206 2207 // Rare inopportune interleaving - race. 2208 // The successor vanished in the small window above. 2209 // The lock is contended -- (cxq|EntryList) != null -- and there's no apparent successor. 2210 // We need to ensure progress and succession. 2211 // Try to reacquire the lock. 2212 // If that fails then the new owner is responsible for succession and this 2213 // thread needs to take no further action and can exit via the fast path (success). 2214 // If the re-acquire succeeds then pass control into the slow path. 2215 // As implemented, this latter mode is horrible because we generated more 2216 // coherence traffic on the lock *and* artifically extended the critical section 2217 // length while by virtue of passing control into the slow path. 2218 2219 // box is really RAX -- the following CMPXCHG depends on that binding 2220 // cmpxchg R,[M] is equivalent to rax = CAS(M,rax,R) 2221 movptr(boxReg, (int32_t)NULL_WORD); 2222 if (os::is_MP()) { lock(); } 2223 cmpxchgptr(r15_thread, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2224 // There's no successor so we tried to regrab the lock. 2225 // If that didn't work, then another thread grabbed the 2226 // lock so we're done (and exit was a success). 2227 jccb (Assembler::notEqual, LSuccess); 2228 // Intentional fall-through into slow-path 2229 2230 bind (LGoSlowPath); 2231 orl (boxReg, 1); // set ICC.ZF=0 to indicate failure 2232 jmpb (DONE_LABEL); 2233 2234 bind (LSuccess); 2235 testl (boxReg, 0); // set ICC.ZF=1 to indicate success 2236 jmpb (DONE_LABEL); 2237 } 2238 2239 bind (Stacked); 2240 movptr(tmpReg, Address (boxReg, 0)); // re-fetch 2241 if (os::is_MP()) { lock(); } 2242 cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses RAX which is box 2243 2244 if (EmitSync & 65536) { 2245 bind (CheckSucc); 2246 } 2247 #endif 2248 bind(DONE_LABEL); 2249 } 2250 } 2251 #endif // COMPILER2 2252 2253 void MacroAssembler::c2bool(Register x) { 2254 // implements x == 0 ? 0 : 1 2255 // note: must only look at least-significant byte of x 2256 // since C-style booleans are stored in one byte 2257 // only! (was bug) 2258 andl(x, 0xFF); 2259 setb(Assembler::notZero, x); 2260 } 2261 2262 // Wouldn't need if AddressLiteral version had new name 2263 void MacroAssembler::call(Label& L, relocInfo::relocType rtype) { 2264 Assembler::call(L, rtype); 2265 } 2266 2267 void MacroAssembler::call(Register entry) { 2268 Assembler::call(entry); 2269 } 2270 2271 void MacroAssembler::call(AddressLiteral entry) { 2272 if (reachable(entry)) { 2273 Assembler::call_literal(entry.target(), entry.rspec()); 2274 } else { 2275 lea(rscratch1, entry); 2276 Assembler::call(rscratch1); 2277 } 2278 } 2279 2280 void MacroAssembler::ic_call(address entry, jint method_index) { 2281 RelocationHolder rh = virtual_call_Relocation::spec(pc(), method_index); 2282 movptr(rax, (intptr_t)Universe::non_oop_word()); 2283 call(AddressLiteral(entry, rh)); 2284 } 2285 2286 // Implementation of call_VM versions 2287 2288 void MacroAssembler::call_VM(Register oop_result, 2289 address entry_point, 2290 bool check_exceptions) { 2291 Label C, E; 2292 call(C, relocInfo::none); 2293 jmp(E); 2294 2295 bind(C); 2296 call_VM_helper(oop_result, entry_point, 0, check_exceptions); 2297 ret(0); 2298 2299 bind(E); 2300 } 2301 2302 void MacroAssembler::call_VM(Register oop_result, 2303 address entry_point, 2304 Register arg_1, 2305 bool check_exceptions) { 2306 Label C, E; 2307 call(C, relocInfo::none); 2308 jmp(E); 2309 2310 bind(C); 2311 pass_arg1(this, arg_1); 2312 call_VM_helper(oop_result, entry_point, 1, check_exceptions); 2313 ret(0); 2314 2315 bind(E); 2316 } 2317 2318 void MacroAssembler::call_VM(Register oop_result, 2319 address entry_point, 2320 Register arg_1, 2321 Register arg_2, 2322 bool check_exceptions) { 2323 Label C, E; 2324 call(C, relocInfo::none); 2325 jmp(E); 2326 2327 bind(C); 2328 2329 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2330 2331 pass_arg2(this, arg_2); 2332 pass_arg1(this, arg_1); 2333 call_VM_helper(oop_result, entry_point, 2, check_exceptions); 2334 ret(0); 2335 2336 bind(E); 2337 } 2338 2339 void MacroAssembler::call_VM(Register oop_result, 2340 address entry_point, 2341 Register arg_1, 2342 Register arg_2, 2343 Register arg_3, 2344 bool check_exceptions) { 2345 Label C, E; 2346 call(C, relocInfo::none); 2347 jmp(E); 2348 2349 bind(C); 2350 2351 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2352 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2353 pass_arg3(this, arg_3); 2354 2355 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2356 pass_arg2(this, arg_2); 2357 2358 pass_arg1(this, arg_1); 2359 call_VM_helper(oop_result, entry_point, 3, check_exceptions); 2360 ret(0); 2361 2362 bind(E); 2363 } 2364 2365 void MacroAssembler::call_VM(Register oop_result, 2366 Register last_java_sp, 2367 address entry_point, 2368 int number_of_arguments, 2369 bool check_exceptions) { 2370 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg); 2371 call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions); 2372 } 2373 2374 void MacroAssembler::call_VM(Register oop_result, 2375 Register last_java_sp, 2376 address entry_point, 2377 Register arg_1, 2378 bool check_exceptions) { 2379 pass_arg1(this, arg_1); 2380 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 2381 } 2382 2383 void MacroAssembler::call_VM(Register oop_result, 2384 Register last_java_sp, 2385 address entry_point, 2386 Register arg_1, 2387 Register arg_2, 2388 bool check_exceptions) { 2389 2390 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2391 pass_arg2(this, arg_2); 2392 pass_arg1(this, arg_1); 2393 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 2394 } 2395 2396 void MacroAssembler::call_VM(Register oop_result, 2397 Register last_java_sp, 2398 address entry_point, 2399 Register arg_1, 2400 Register arg_2, 2401 Register arg_3, 2402 bool check_exceptions) { 2403 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2404 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2405 pass_arg3(this, arg_3); 2406 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2407 pass_arg2(this, arg_2); 2408 pass_arg1(this, arg_1); 2409 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 2410 } 2411 2412 void MacroAssembler::super_call_VM(Register oop_result, 2413 Register last_java_sp, 2414 address entry_point, 2415 int number_of_arguments, 2416 bool check_exceptions) { 2417 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg); 2418 MacroAssembler::call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions); 2419 } 2420 2421 void MacroAssembler::super_call_VM(Register oop_result, 2422 Register last_java_sp, 2423 address entry_point, 2424 Register arg_1, 2425 bool check_exceptions) { 2426 pass_arg1(this, arg_1); 2427 super_call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 2428 } 2429 2430 void MacroAssembler::super_call_VM(Register oop_result, 2431 Register last_java_sp, 2432 address entry_point, 2433 Register arg_1, 2434 Register arg_2, 2435 bool check_exceptions) { 2436 2437 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2438 pass_arg2(this, arg_2); 2439 pass_arg1(this, arg_1); 2440 super_call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 2441 } 2442 2443 void MacroAssembler::super_call_VM(Register oop_result, 2444 Register last_java_sp, 2445 address entry_point, 2446 Register arg_1, 2447 Register arg_2, 2448 Register arg_3, 2449 bool check_exceptions) { 2450 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2451 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2452 pass_arg3(this, arg_3); 2453 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2454 pass_arg2(this, arg_2); 2455 pass_arg1(this, arg_1); 2456 super_call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 2457 } 2458 2459 void MacroAssembler::call_VM_base(Register oop_result, 2460 Register java_thread, 2461 Register last_java_sp, 2462 address entry_point, 2463 int number_of_arguments, 2464 bool check_exceptions) { 2465 // determine java_thread register 2466 if (!java_thread->is_valid()) { 2467 #ifdef _LP64 2468 java_thread = r15_thread; 2469 #else 2470 java_thread = rdi; 2471 get_thread(java_thread); 2472 #endif // LP64 2473 } 2474 // determine last_java_sp register 2475 if (!last_java_sp->is_valid()) { 2476 last_java_sp = rsp; 2477 } 2478 // debugging support 2479 assert(number_of_arguments >= 0 , "cannot have negative number of arguments"); 2480 LP64_ONLY(assert(java_thread == r15_thread, "unexpected register")); 2481 #ifdef ASSERT 2482 // TraceBytecodes does not use r12 but saves it over the call, so don't verify 2483 // r12 is the heapbase. 2484 LP64_ONLY(if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");) 2485 #endif // ASSERT 2486 2487 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result"); 2488 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp"); 2489 2490 // push java thread (becomes first argument of C function) 2491 2492 NOT_LP64(push(java_thread); number_of_arguments++); 2493 LP64_ONLY(mov(c_rarg0, r15_thread)); 2494 2495 // set last Java frame before call 2496 assert(last_java_sp != rbp, "can't use ebp/rbp"); 2497 2498 // Only interpreter should have to set fp 2499 set_last_Java_frame(java_thread, last_java_sp, rbp, NULL); 2500 2501 // do the call, remove parameters 2502 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments); 2503 2504 // restore the thread (cannot use the pushed argument since arguments 2505 // may be overwritten by C code generated by an optimizing compiler); 2506 // however can use the register value directly if it is callee saved. 2507 if (LP64_ONLY(true ||) java_thread == rdi || java_thread == rsi) { 2508 // rdi & rsi (also r15) are callee saved -> nothing to do 2509 #ifdef ASSERT 2510 guarantee(java_thread != rax, "change this code"); 2511 push(rax); 2512 { Label L; 2513 get_thread(rax); 2514 cmpptr(java_thread, rax); 2515 jcc(Assembler::equal, L); 2516 STOP("MacroAssembler::call_VM_base: rdi not callee saved?"); 2517 bind(L); 2518 } 2519 pop(rax); 2520 #endif 2521 } else { 2522 get_thread(java_thread); 2523 } 2524 // reset last Java frame 2525 // Only interpreter should have to clear fp 2526 reset_last_Java_frame(java_thread, true, false); 2527 2528 #ifndef CC_INTERP 2529 // C++ interp handles this in the interpreter 2530 check_and_handle_popframe(java_thread); 2531 check_and_handle_earlyret(java_thread); 2532 #endif /* CC_INTERP */ 2533 2534 if (check_exceptions) { 2535 // check for pending exceptions (java_thread is set upon return) 2536 cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t) NULL_WORD); 2537 #ifndef _LP64 2538 jump_cc(Assembler::notEqual, 2539 RuntimeAddress(StubRoutines::forward_exception_entry())); 2540 #else 2541 // This used to conditionally jump to forward_exception however it is 2542 // possible if we relocate that the branch will not reach. So we must jump 2543 // around so we can always reach 2544 2545 Label ok; 2546 jcc(Assembler::equal, ok); 2547 jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2548 bind(ok); 2549 #endif // LP64 2550 } 2551 2552 // get oop result if there is one and reset the value in the thread 2553 if (oop_result->is_valid()) { 2554 get_vm_result(oop_result, java_thread); 2555 } 2556 } 2557 2558 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) { 2559 2560 // Calculate the value for last_Java_sp 2561 // somewhat subtle. call_VM does an intermediate call 2562 // which places a return address on the stack just under the 2563 // stack pointer as the user finsihed with it. This allows 2564 // use to retrieve last_Java_pc from last_Java_sp[-1]. 2565 // On 32bit we then have to push additional args on the stack to accomplish 2566 // the actual requested call. On 64bit call_VM only can use register args 2567 // so the only extra space is the return address that call_VM created. 2568 // This hopefully explains the calculations here. 2569 2570 #ifdef _LP64 2571 // We've pushed one address, correct last_Java_sp 2572 lea(rax, Address(rsp, wordSize)); 2573 #else 2574 lea(rax, Address(rsp, (1 + number_of_arguments) * wordSize)); 2575 #endif // LP64 2576 2577 call_VM_base(oop_result, noreg, rax, entry_point, number_of_arguments, check_exceptions); 2578 2579 } 2580 2581 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) { 2582 call_VM_leaf_base(entry_point, number_of_arguments); 2583 } 2584 2585 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) { 2586 pass_arg0(this, arg_0); 2587 call_VM_leaf(entry_point, 1); 2588 } 2589 2590 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { 2591 2592 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2593 pass_arg1(this, arg_1); 2594 pass_arg0(this, arg_0); 2595 call_VM_leaf(entry_point, 2); 2596 } 2597 2598 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { 2599 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2600 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2601 pass_arg2(this, arg_2); 2602 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2603 pass_arg1(this, arg_1); 2604 pass_arg0(this, arg_0); 2605 call_VM_leaf(entry_point, 3); 2606 } 2607 2608 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) { 2609 pass_arg0(this, arg_0); 2610 MacroAssembler::call_VM_leaf_base(entry_point, 1); 2611 } 2612 2613 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { 2614 2615 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2616 pass_arg1(this, arg_1); 2617 pass_arg0(this, arg_0); 2618 MacroAssembler::call_VM_leaf_base(entry_point, 2); 2619 } 2620 2621 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { 2622 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2623 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2624 pass_arg2(this, arg_2); 2625 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2626 pass_arg1(this, arg_1); 2627 pass_arg0(this, arg_0); 2628 MacroAssembler::call_VM_leaf_base(entry_point, 3); 2629 } 2630 2631 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) { 2632 LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg")); 2633 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2634 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2635 pass_arg3(this, arg_3); 2636 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2637 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2638 pass_arg2(this, arg_2); 2639 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2640 pass_arg1(this, arg_1); 2641 pass_arg0(this, arg_0); 2642 MacroAssembler::call_VM_leaf_base(entry_point, 4); 2643 } 2644 2645 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) { 2646 movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset())); 2647 movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD); 2648 verify_oop(oop_result, "broken oop in call_VM_base"); 2649 } 2650 2651 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) { 2652 movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset())); 2653 movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD); 2654 } 2655 2656 void MacroAssembler::check_and_handle_earlyret(Register java_thread) { 2657 } 2658 2659 void MacroAssembler::check_and_handle_popframe(Register java_thread) { 2660 } 2661 2662 void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) { 2663 if (reachable(src1)) { 2664 cmpl(as_Address(src1), imm); 2665 } else { 2666 lea(rscratch1, src1); 2667 cmpl(Address(rscratch1, 0), imm); 2668 } 2669 } 2670 2671 void MacroAssembler::cmp32(Register src1, AddressLiteral src2) { 2672 assert(!src2.is_lval(), "use cmpptr"); 2673 if (reachable(src2)) { 2674 cmpl(src1, as_Address(src2)); 2675 } else { 2676 lea(rscratch1, src2); 2677 cmpl(src1, Address(rscratch1, 0)); 2678 } 2679 } 2680 2681 void MacroAssembler::cmp32(Register src1, int32_t imm) { 2682 Assembler::cmpl(src1, imm); 2683 } 2684 2685 void MacroAssembler::cmp32(Register src1, Address src2) { 2686 Assembler::cmpl(src1, src2); 2687 } 2688 2689 void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) { 2690 ucomisd(opr1, opr2); 2691 2692 Label L; 2693 if (unordered_is_less) { 2694 movl(dst, -1); 2695 jcc(Assembler::parity, L); 2696 jcc(Assembler::below , L); 2697 movl(dst, 0); 2698 jcc(Assembler::equal , L); 2699 increment(dst); 2700 } else { // unordered is greater 2701 movl(dst, 1); 2702 jcc(Assembler::parity, L); 2703 jcc(Assembler::above , L); 2704 movl(dst, 0); 2705 jcc(Assembler::equal , L); 2706 decrementl(dst); 2707 } 2708 bind(L); 2709 } 2710 2711 void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) { 2712 ucomiss(opr1, opr2); 2713 2714 Label L; 2715 if (unordered_is_less) { 2716 movl(dst, -1); 2717 jcc(Assembler::parity, L); 2718 jcc(Assembler::below , L); 2719 movl(dst, 0); 2720 jcc(Assembler::equal , L); 2721 increment(dst); 2722 } else { // unordered is greater 2723 movl(dst, 1); 2724 jcc(Assembler::parity, L); 2725 jcc(Assembler::above , L); 2726 movl(dst, 0); 2727 jcc(Assembler::equal , L); 2728 decrementl(dst); 2729 } 2730 bind(L); 2731 } 2732 2733 2734 void MacroAssembler::cmp8(AddressLiteral src1, int imm) { 2735 if (reachable(src1)) { 2736 cmpb(as_Address(src1), imm); 2737 } else { 2738 lea(rscratch1, src1); 2739 cmpb(Address(rscratch1, 0), imm); 2740 } 2741 } 2742 2743 void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) { 2744 #ifdef _LP64 2745 if (src2.is_lval()) { 2746 movptr(rscratch1, src2); 2747 Assembler::cmpq(src1, rscratch1); 2748 } else if (reachable(src2)) { 2749 cmpq(src1, as_Address(src2)); 2750 } else { 2751 lea(rscratch1, src2); 2752 Assembler::cmpq(src1, Address(rscratch1, 0)); 2753 } 2754 #else 2755 if (src2.is_lval()) { 2756 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec()); 2757 } else { 2758 cmpl(src1, as_Address(src2)); 2759 } 2760 #endif // _LP64 2761 } 2762 2763 void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) { 2764 assert(src2.is_lval(), "not a mem-mem compare"); 2765 #ifdef _LP64 2766 // moves src2's literal address 2767 movptr(rscratch1, src2); 2768 Assembler::cmpq(src1, rscratch1); 2769 #else 2770 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec()); 2771 #endif // _LP64 2772 } 2773 2774 void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) { 2775 if (reachable(adr)) { 2776 if (os::is_MP()) 2777 lock(); 2778 cmpxchgptr(reg, as_Address(adr)); 2779 } else { 2780 lea(rscratch1, adr); 2781 if (os::is_MP()) 2782 lock(); 2783 cmpxchgptr(reg, Address(rscratch1, 0)); 2784 } 2785 } 2786 2787 void MacroAssembler::cmpxchgptr(Register reg, Address adr) { 2788 LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr)); 2789 } 2790 2791 void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) { 2792 if (reachable(src)) { 2793 Assembler::comisd(dst, as_Address(src)); 2794 } else { 2795 lea(rscratch1, src); 2796 Assembler::comisd(dst, Address(rscratch1, 0)); 2797 } 2798 } 2799 2800 void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) { 2801 if (reachable(src)) { 2802 Assembler::comiss(dst, as_Address(src)); 2803 } else { 2804 lea(rscratch1, src); 2805 Assembler::comiss(dst, Address(rscratch1, 0)); 2806 } 2807 } 2808 2809 2810 void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) { 2811 Condition negated_cond = negate_condition(cond); 2812 Label L; 2813 jcc(negated_cond, L); 2814 pushf(); // Preserve flags 2815 atomic_incl(counter_addr); 2816 popf(); 2817 bind(L); 2818 } 2819 2820 int MacroAssembler::corrected_idivl(Register reg) { 2821 // Full implementation of Java idiv and irem; checks for 2822 // special case as described in JVM spec., p.243 & p.271. 2823 // The function returns the (pc) offset of the idivl 2824 // instruction - may be needed for implicit exceptions. 2825 // 2826 // normal case special case 2827 // 2828 // input : rax,: dividend min_int 2829 // reg: divisor (may not be rax,/rdx) -1 2830 // 2831 // output: rax,: quotient (= rax, idiv reg) min_int 2832 // rdx: remainder (= rax, irem reg) 0 2833 assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register"); 2834 const int min_int = 0x80000000; 2835 Label normal_case, special_case; 2836 2837 // check for special case 2838 cmpl(rax, min_int); 2839 jcc(Assembler::notEqual, normal_case); 2840 xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0) 2841 cmpl(reg, -1); 2842 jcc(Assembler::equal, special_case); 2843 2844 // handle normal case 2845 bind(normal_case); 2846 cdql(); 2847 int idivl_offset = offset(); 2848 idivl(reg); 2849 2850 // normal and special case exit 2851 bind(special_case); 2852 2853 return idivl_offset; 2854 } 2855 2856 2857 2858 void MacroAssembler::decrementl(Register reg, int value) { 2859 if (value == min_jint) {subl(reg, value) ; return; } 2860 if (value < 0) { incrementl(reg, -value); return; } 2861 if (value == 0) { ; return; } 2862 if (value == 1 && UseIncDec) { decl(reg) ; return; } 2863 /* else */ { subl(reg, value) ; return; } 2864 } 2865 2866 void MacroAssembler::decrementl(Address dst, int value) { 2867 if (value == min_jint) {subl(dst, value) ; return; } 2868 if (value < 0) { incrementl(dst, -value); return; } 2869 if (value == 0) { ; return; } 2870 if (value == 1 && UseIncDec) { decl(dst) ; return; } 2871 /* else */ { subl(dst, value) ; return; } 2872 } 2873 2874 void MacroAssembler::division_with_shift (Register reg, int shift_value) { 2875 assert (shift_value > 0, "illegal shift value"); 2876 Label _is_positive; 2877 testl (reg, reg); 2878 jcc (Assembler::positive, _is_positive); 2879 int offset = (1 << shift_value) - 1 ; 2880 2881 if (offset == 1) { 2882 incrementl(reg); 2883 } else { 2884 addl(reg, offset); 2885 } 2886 2887 bind (_is_positive); 2888 sarl(reg, shift_value); 2889 } 2890 2891 void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) { 2892 if (reachable(src)) { 2893 Assembler::divsd(dst, as_Address(src)); 2894 } else { 2895 lea(rscratch1, src); 2896 Assembler::divsd(dst, Address(rscratch1, 0)); 2897 } 2898 } 2899 2900 void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) { 2901 if (reachable(src)) { 2902 Assembler::divss(dst, as_Address(src)); 2903 } else { 2904 lea(rscratch1, src); 2905 Assembler::divss(dst, Address(rscratch1, 0)); 2906 } 2907 } 2908 2909 // !defined(COMPILER2) is because of stupid core builds 2910 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2) || INCLUDE_JVMCI 2911 void MacroAssembler::empty_FPU_stack() { 2912 if (VM_Version::supports_mmx()) { 2913 emms(); 2914 } else { 2915 for (int i = 8; i-- > 0; ) ffree(i); 2916 } 2917 } 2918 #endif // !LP64 || C1 || !C2 || INCLUDE_JVMCI 2919 2920 2921 // Defines obj, preserves var_size_in_bytes 2922 void MacroAssembler::eden_allocate(Register obj, 2923 Register var_size_in_bytes, 2924 int con_size_in_bytes, 2925 Register t1, 2926 Label& slow_case) { 2927 assert(obj == rax, "obj must be in rax, for cmpxchg"); 2928 assert_different_registers(obj, var_size_in_bytes, t1); 2929 if (!Universe::heap()->supports_inline_contig_alloc()) { 2930 jmp(slow_case); 2931 } else { 2932 Register end = t1; 2933 Label retry; 2934 bind(retry); 2935 ExternalAddress heap_top((address) Universe::heap()->top_addr()); 2936 movptr(obj, heap_top); 2937 if (var_size_in_bytes == noreg) { 2938 lea(end, Address(obj, con_size_in_bytes)); 2939 } else { 2940 lea(end, Address(obj, var_size_in_bytes, Address::times_1)); 2941 } 2942 // if end < obj then we wrapped around => object too long => slow case 2943 cmpptr(end, obj); 2944 jcc(Assembler::below, slow_case); 2945 cmpptr(end, ExternalAddress((address) Universe::heap()->end_addr())); 2946 jcc(Assembler::above, slow_case); 2947 // Compare obj with the top addr, and if still equal, store the new top addr in 2948 // end at the address of the top addr pointer. Sets ZF if was equal, and clears 2949 // it otherwise. Use lock prefix for atomicity on MPs. 2950 locked_cmpxchgptr(end, heap_top); 2951 jcc(Assembler::notEqual, retry); 2952 } 2953 } 2954 2955 void MacroAssembler::enter() { 2956 push(rbp); 2957 mov(rbp, rsp); 2958 } 2959 2960 // A 5 byte nop that is safe for patching (see patch_verified_entry) 2961 void MacroAssembler::fat_nop() { 2962 if (UseAddressNop) { 2963 addr_nop_5(); 2964 } else { 2965 emit_int8(0x26); // es: 2966 emit_int8(0x2e); // cs: 2967 emit_int8(0x64); // fs: 2968 emit_int8(0x65); // gs: 2969 emit_int8((unsigned char)0x90); 2970 } 2971 } 2972 2973 void MacroAssembler::fcmp(Register tmp) { 2974 fcmp(tmp, 1, true, true); 2975 } 2976 2977 void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) { 2978 assert(!pop_right || pop_left, "usage error"); 2979 if (VM_Version::supports_cmov()) { 2980 assert(tmp == noreg, "unneeded temp"); 2981 if (pop_left) { 2982 fucomip(index); 2983 } else { 2984 fucomi(index); 2985 } 2986 if (pop_right) { 2987 fpop(); 2988 } 2989 } else { 2990 assert(tmp != noreg, "need temp"); 2991 if (pop_left) { 2992 if (pop_right) { 2993 fcompp(); 2994 } else { 2995 fcomp(index); 2996 } 2997 } else { 2998 fcom(index); 2999 } 3000 // convert FPU condition into eflags condition via rax, 3001 save_rax(tmp); 3002 fwait(); fnstsw_ax(); 3003 sahf(); 3004 restore_rax(tmp); 3005 } 3006 // condition codes set as follows: 3007 // 3008 // CF (corresponds to C0) if x < y 3009 // PF (corresponds to C2) if unordered 3010 // ZF (corresponds to C3) if x = y 3011 } 3012 3013 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) { 3014 fcmp2int(dst, unordered_is_less, 1, true, true); 3015 } 3016 3017 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) { 3018 fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right); 3019 Label L; 3020 if (unordered_is_less) { 3021 movl(dst, -1); 3022 jcc(Assembler::parity, L); 3023 jcc(Assembler::below , L); 3024 movl(dst, 0); 3025 jcc(Assembler::equal , L); 3026 increment(dst); 3027 } else { // unordered is greater 3028 movl(dst, 1); 3029 jcc(Assembler::parity, L); 3030 jcc(Assembler::above , L); 3031 movl(dst, 0); 3032 jcc(Assembler::equal , L); 3033 decrementl(dst); 3034 } 3035 bind(L); 3036 } 3037 3038 void MacroAssembler::fld_d(AddressLiteral src) { 3039 fld_d(as_Address(src)); 3040 } 3041 3042 void MacroAssembler::fld_s(AddressLiteral src) { 3043 fld_s(as_Address(src)); 3044 } 3045 3046 void MacroAssembler::fld_x(AddressLiteral src) { 3047 Assembler::fld_x(as_Address(src)); 3048 } 3049 3050 void MacroAssembler::fldcw(AddressLiteral src) { 3051 Assembler::fldcw(as_Address(src)); 3052 } 3053 3054 void MacroAssembler::mulpd(XMMRegister dst, AddressLiteral src) { 3055 if (reachable(src)) { 3056 Assembler::mulpd(dst, as_Address(src)); 3057 } else { 3058 lea(rscratch1, src); 3059 Assembler::mulpd(dst, Address(rscratch1, 0)); 3060 } 3061 } 3062 3063 void MacroAssembler::increase_precision() { 3064 subptr(rsp, BytesPerWord); 3065 fnstcw(Address(rsp, 0)); 3066 movl(rax, Address(rsp, 0)); 3067 orl(rax, 0x300); 3068 push(rax); 3069 fldcw(Address(rsp, 0)); 3070 pop(rax); 3071 } 3072 3073 void MacroAssembler::restore_precision() { 3074 fldcw(Address(rsp, 0)); 3075 addptr(rsp, BytesPerWord); 3076 } 3077 3078 void MacroAssembler::fpop() { 3079 ffree(); 3080 fincstp(); 3081 } 3082 3083 void MacroAssembler::load_float(Address src) { 3084 if (UseSSE >= 1) { 3085 movflt(xmm0, src); 3086 } else { 3087 LP64_ONLY(ShouldNotReachHere()); 3088 NOT_LP64(fld_s(src)); 3089 } 3090 } 3091 3092 void MacroAssembler::store_float(Address dst) { 3093 if (UseSSE >= 1) { 3094 movflt(dst, xmm0); 3095 } else { 3096 LP64_ONLY(ShouldNotReachHere()); 3097 NOT_LP64(fstp_s(dst)); 3098 } 3099 } 3100 3101 void MacroAssembler::load_double(Address src) { 3102 if (UseSSE >= 2) { 3103 movdbl(xmm0, src); 3104 } else { 3105 LP64_ONLY(ShouldNotReachHere()); 3106 NOT_LP64(fld_d(src)); 3107 } 3108 } 3109 3110 void MacroAssembler::store_double(Address dst) { 3111 if (UseSSE >= 2) { 3112 movdbl(dst, xmm0); 3113 } else { 3114 LP64_ONLY(ShouldNotReachHere()); 3115 NOT_LP64(fstp_d(dst)); 3116 } 3117 } 3118 3119 void MacroAssembler::fremr(Register tmp) { 3120 save_rax(tmp); 3121 { Label L; 3122 bind(L); 3123 fprem(); 3124 fwait(); fnstsw_ax(); 3125 #ifdef _LP64 3126 testl(rax, 0x400); 3127 jcc(Assembler::notEqual, L); 3128 #else 3129 sahf(); 3130 jcc(Assembler::parity, L); 3131 #endif // _LP64 3132 } 3133 restore_rax(tmp); 3134 // Result is in ST0. 3135 // Note: fxch & fpop to get rid of ST1 3136 // (otherwise FPU stack could overflow eventually) 3137 fxch(1); 3138 fpop(); 3139 } 3140 3141 3142 void MacroAssembler::incrementl(AddressLiteral dst) { 3143 if (reachable(dst)) { 3144 incrementl(as_Address(dst)); 3145 } else { 3146 lea(rscratch1, dst); 3147 incrementl(Address(rscratch1, 0)); 3148 } 3149 } 3150 3151 void MacroAssembler::incrementl(ArrayAddress dst) { 3152 incrementl(as_Address(dst)); 3153 } 3154 3155 void MacroAssembler::incrementl(Register reg, int value) { 3156 if (value == min_jint) {addl(reg, value) ; return; } 3157 if (value < 0) { decrementl(reg, -value); return; } 3158 if (value == 0) { ; return; } 3159 if (value == 1 && UseIncDec) { incl(reg) ; return; } 3160 /* else */ { addl(reg, value) ; return; } 3161 } 3162 3163 void MacroAssembler::incrementl(Address dst, int value) { 3164 if (value == min_jint) {addl(dst, value) ; return; } 3165 if (value < 0) { decrementl(dst, -value); return; } 3166 if (value == 0) { ; return; } 3167 if (value == 1 && UseIncDec) { incl(dst) ; return; } 3168 /* else */ { addl(dst, value) ; return; } 3169 } 3170 3171 void MacroAssembler::jump(AddressLiteral dst) { 3172 if (reachable(dst)) { 3173 jmp_literal(dst.target(), dst.rspec()); 3174 } else { 3175 lea(rscratch1, dst); 3176 jmp(rscratch1); 3177 } 3178 } 3179 3180 void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) { 3181 if (reachable(dst)) { 3182 InstructionMark im(this); 3183 relocate(dst.reloc()); 3184 const int short_size = 2; 3185 const int long_size = 6; 3186 int offs = (intptr_t)dst.target() - ((intptr_t)pc()); 3187 if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) { 3188 // 0111 tttn #8-bit disp 3189 emit_int8(0x70 | cc); 3190 emit_int8((offs - short_size) & 0xFF); 3191 } else { 3192 // 0000 1111 1000 tttn #32-bit disp 3193 emit_int8(0x0F); 3194 emit_int8((unsigned char)(0x80 | cc)); 3195 emit_int32(offs - long_size); 3196 } 3197 } else { 3198 #ifdef ASSERT 3199 warning("reversing conditional branch"); 3200 #endif /* ASSERT */ 3201 Label skip; 3202 jccb(reverse[cc], skip); 3203 lea(rscratch1, dst); 3204 Assembler::jmp(rscratch1); 3205 bind(skip); 3206 } 3207 } 3208 3209 void MacroAssembler::ldmxcsr(AddressLiteral src) { 3210 if (reachable(src)) { 3211 Assembler::ldmxcsr(as_Address(src)); 3212 } else { 3213 lea(rscratch1, src); 3214 Assembler::ldmxcsr(Address(rscratch1, 0)); 3215 } 3216 } 3217 3218 int MacroAssembler::load_signed_byte(Register dst, Address src) { 3219 int off; 3220 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3221 off = offset(); 3222 movsbl(dst, src); // movsxb 3223 } else { 3224 off = load_unsigned_byte(dst, src); 3225 shll(dst, 24); 3226 sarl(dst, 24); 3227 } 3228 return off; 3229 } 3230 3231 // Note: load_signed_short used to be called load_signed_word. 3232 // Although the 'w' in x86 opcodes refers to the term "word" in the assembler 3233 // manual, which means 16 bits, that usage is found nowhere in HotSpot code. 3234 // The term "word" in HotSpot means a 32- or 64-bit machine word. 3235 int MacroAssembler::load_signed_short(Register dst, Address src) { 3236 int off; 3237 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3238 // This is dubious to me since it seems safe to do a signed 16 => 64 bit 3239 // version but this is what 64bit has always done. This seems to imply 3240 // that users are only using 32bits worth. 3241 off = offset(); 3242 movswl(dst, src); // movsxw 3243 } else { 3244 off = load_unsigned_short(dst, src); 3245 shll(dst, 16); 3246 sarl(dst, 16); 3247 } 3248 return off; 3249 } 3250 3251 int MacroAssembler::load_unsigned_byte(Register dst, Address src) { 3252 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16, 3253 // and "3.9 Partial Register Penalties", p. 22). 3254 int off; 3255 if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) { 3256 off = offset(); 3257 movzbl(dst, src); // movzxb 3258 } else { 3259 xorl(dst, dst); 3260 off = offset(); 3261 movb(dst, src); 3262 } 3263 return off; 3264 } 3265 3266 // Note: load_unsigned_short used to be called load_unsigned_word. 3267 int MacroAssembler::load_unsigned_short(Register dst, Address src) { 3268 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16, 3269 // and "3.9 Partial Register Penalties", p. 22). 3270 int off; 3271 if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) { 3272 off = offset(); 3273 movzwl(dst, src); // movzxw 3274 } else { 3275 xorl(dst, dst); 3276 off = offset(); 3277 movw(dst, src); 3278 } 3279 return off; 3280 } 3281 3282 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) { 3283 switch (size_in_bytes) { 3284 #ifndef _LP64 3285 case 8: 3286 assert(dst2 != noreg, "second dest register required"); 3287 movl(dst, src); 3288 movl(dst2, src.plus_disp(BytesPerInt)); 3289 break; 3290 #else 3291 case 8: movq(dst, src); break; 3292 #endif 3293 case 4: movl(dst, src); break; 3294 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break; 3295 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break; 3296 default: ShouldNotReachHere(); 3297 } 3298 } 3299 3300 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) { 3301 switch (size_in_bytes) { 3302 #ifndef _LP64 3303 case 8: 3304 assert(src2 != noreg, "second source register required"); 3305 movl(dst, src); 3306 movl(dst.plus_disp(BytesPerInt), src2); 3307 break; 3308 #else 3309 case 8: movq(dst, src); break; 3310 #endif 3311 case 4: movl(dst, src); break; 3312 case 2: movw(dst, src); break; 3313 case 1: movb(dst, src); break; 3314 default: ShouldNotReachHere(); 3315 } 3316 } 3317 3318 void MacroAssembler::mov32(AddressLiteral dst, Register src) { 3319 if (reachable(dst)) { 3320 movl(as_Address(dst), src); 3321 } else { 3322 lea(rscratch1, dst); 3323 movl(Address(rscratch1, 0), src); 3324 } 3325 } 3326 3327 void MacroAssembler::mov32(Register dst, AddressLiteral src) { 3328 if (reachable(src)) { 3329 movl(dst, as_Address(src)); 3330 } else { 3331 lea(rscratch1, src); 3332 movl(dst, Address(rscratch1, 0)); 3333 } 3334 } 3335 3336 // C++ bool manipulation 3337 3338 void MacroAssembler::movbool(Register dst, Address src) { 3339 if(sizeof(bool) == 1) 3340 movb(dst, src); 3341 else if(sizeof(bool) == 2) 3342 movw(dst, src); 3343 else if(sizeof(bool) == 4) 3344 movl(dst, src); 3345 else 3346 // unsupported 3347 ShouldNotReachHere(); 3348 } 3349 3350 void MacroAssembler::movbool(Address dst, bool boolconst) { 3351 if(sizeof(bool) == 1) 3352 movb(dst, (int) boolconst); 3353 else if(sizeof(bool) == 2) 3354 movw(dst, (int) boolconst); 3355 else if(sizeof(bool) == 4) 3356 movl(dst, (int) boolconst); 3357 else 3358 // unsupported 3359 ShouldNotReachHere(); 3360 } 3361 3362 void MacroAssembler::movbool(Address dst, Register src) { 3363 if(sizeof(bool) == 1) 3364 movb(dst, src); 3365 else if(sizeof(bool) == 2) 3366 movw(dst, src); 3367 else if(sizeof(bool) == 4) 3368 movl(dst, src); 3369 else 3370 // unsupported 3371 ShouldNotReachHere(); 3372 } 3373 3374 void MacroAssembler::movbyte(ArrayAddress dst, int src) { 3375 movb(as_Address(dst), src); 3376 } 3377 3378 void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) { 3379 if (reachable(src)) { 3380 movdl(dst, as_Address(src)); 3381 } else { 3382 lea(rscratch1, src); 3383 movdl(dst, Address(rscratch1, 0)); 3384 } 3385 } 3386 3387 void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) { 3388 if (reachable(src)) { 3389 movq(dst, as_Address(src)); 3390 } else { 3391 lea(rscratch1, src); 3392 movq(dst, Address(rscratch1, 0)); 3393 } 3394 } 3395 3396 void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) { 3397 if (reachable(src)) { 3398 if (UseXmmLoadAndClearUpper) { 3399 movsd (dst, as_Address(src)); 3400 } else { 3401 movlpd(dst, as_Address(src)); 3402 } 3403 } else { 3404 lea(rscratch1, src); 3405 if (UseXmmLoadAndClearUpper) { 3406 movsd (dst, Address(rscratch1, 0)); 3407 } else { 3408 movlpd(dst, Address(rscratch1, 0)); 3409 } 3410 } 3411 } 3412 3413 void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) { 3414 if (reachable(src)) { 3415 movss(dst, as_Address(src)); 3416 } else { 3417 lea(rscratch1, src); 3418 movss(dst, Address(rscratch1, 0)); 3419 } 3420 } 3421 3422 void MacroAssembler::movptr(Register dst, Register src) { 3423 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3424 } 3425 3426 void MacroAssembler::movptr(Register dst, Address src) { 3427 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3428 } 3429 3430 // src should NEVER be a real pointer. Use AddressLiteral for true pointers 3431 void MacroAssembler::movptr(Register dst, intptr_t src) { 3432 LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src)); 3433 } 3434 3435 void MacroAssembler::movptr(Address dst, Register src) { 3436 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3437 } 3438 3439 void MacroAssembler::movdqu(Address dst, XMMRegister src) { 3440 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (src->encoding() > 15)) { 3441 Assembler::vextractf32x4h(dst, src, 0); 3442 } else { 3443 Assembler::movdqu(dst, src); 3444 } 3445 } 3446 3447 void MacroAssembler::movdqu(XMMRegister dst, Address src) { 3448 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (dst->encoding() > 15)) { 3449 Assembler::vinsertf32x4h(dst, src, 0); 3450 } else { 3451 Assembler::movdqu(dst, src); 3452 } 3453 } 3454 3455 void MacroAssembler::movdqu(XMMRegister dst, XMMRegister src) { 3456 if (UseAVX > 2 && !VM_Version::supports_avx512vl()) { 3457 Assembler::evmovdqul(dst, src, Assembler::AVX_512bit); 3458 } else { 3459 Assembler::movdqu(dst, src); 3460 } 3461 } 3462 3463 void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src) { 3464 if (reachable(src)) { 3465 movdqu(dst, as_Address(src)); 3466 } else { 3467 lea(rscratch1, src); 3468 movdqu(dst, Address(rscratch1, 0)); 3469 } 3470 } 3471 3472 void MacroAssembler::vmovdqu(Address dst, XMMRegister src) { 3473 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (src->encoding() > 15)) { 3474 Assembler::vextractf64x4h(dst, src, 0); 3475 } else { 3476 Assembler::vmovdqu(dst, src); 3477 } 3478 } 3479 3480 void MacroAssembler::vmovdqu(XMMRegister dst, Address src) { 3481 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (dst->encoding() > 15)) { 3482 Assembler::vinsertf64x4h(dst, src, 0); 3483 } else { 3484 Assembler::vmovdqu(dst, src); 3485 } 3486 } 3487 3488 void MacroAssembler::vmovdqu(XMMRegister dst, XMMRegister src) { 3489 if (UseAVX > 2 && !VM_Version::supports_avx512vl()) { 3490 Assembler::evmovdqul(dst, src, Assembler::AVX_512bit); 3491 } 3492 else { 3493 Assembler::vmovdqu(dst, src); 3494 } 3495 } 3496 3497 void MacroAssembler::vmovdqu(XMMRegister dst, AddressLiteral src) { 3498 if (reachable(src)) { 3499 vmovdqu(dst, as_Address(src)); 3500 } 3501 else { 3502 lea(rscratch1, src); 3503 vmovdqu(dst, Address(rscratch1, 0)); 3504 } 3505 } 3506 3507 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) { 3508 if (reachable(src)) { 3509 Assembler::movdqa(dst, as_Address(src)); 3510 } else { 3511 lea(rscratch1, src); 3512 Assembler::movdqa(dst, Address(rscratch1, 0)); 3513 } 3514 } 3515 3516 void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) { 3517 if (reachable(src)) { 3518 Assembler::movsd(dst, as_Address(src)); 3519 } else { 3520 lea(rscratch1, src); 3521 Assembler::movsd(dst, Address(rscratch1, 0)); 3522 } 3523 } 3524 3525 void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) { 3526 if (reachable(src)) { 3527 Assembler::movss(dst, as_Address(src)); 3528 } else { 3529 lea(rscratch1, src); 3530 Assembler::movss(dst, Address(rscratch1, 0)); 3531 } 3532 } 3533 3534 void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) { 3535 if (reachable(src)) { 3536 Assembler::mulsd(dst, as_Address(src)); 3537 } else { 3538 lea(rscratch1, src); 3539 Assembler::mulsd(dst, Address(rscratch1, 0)); 3540 } 3541 } 3542 3543 void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) { 3544 if (reachable(src)) { 3545 Assembler::mulss(dst, as_Address(src)); 3546 } else { 3547 lea(rscratch1, src); 3548 Assembler::mulss(dst, Address(rscratch1, 0)); 3549 } 3550 } 3551 3552 void MacroAssembler::null_check(Register reg, int offset) { 3553 if (needs_explicit_null_check(offset)) { 3554 // provoke OS NULL exception if reg = NULL by 3555 // accessing M[reg] w/o changing any (non-CC) registers 3556 // NOTE: cmpl is plenty here to provoke a segv 3557 cmpptr(rax, Address(reg, 0)); 3558 // Note: should probably use testl(rax, Address(reg, 0)); 3559 // may be shorter code (however, this version of 3560 // testl needs to be implemented first) 3561 } else { 3562 // nothing to do, (later) access of M[reg + offset] 3563 // will provoke OS NULL exception if reg = NULL 3564 } 3565 } 3566 3567 void MacroAssembler::os_breakpoint() { 3568 // instead of directly emitting a breakpoint, call os:breakpoint for better debugability 3569 // (e.g., MSVC can't call ps() otherwise) 3570 call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint))); 3571 } 3572 3573 #ifdef _LP64 3574 #define XSTATE_BV 0x200 3575 #endif 3576 3577 void MacroAssembler::pop_CPU_state() { 3578 pop_FPU_state(); 3579 pop_IU_state(); 3580 } 3581 3582 void MacroAssembler::pop_FPU_state() { 3583 #ifndef _LP64 3584 frstor(Address(rsp, 0)); 3585 #else 3586 fxrstor(Address(rsp, 0)); 3587 #endif 3588 addptr(rsp, FPUStateSizeInWords * wordSize); 3589 } 3590 3591 void MacroAssembler::pop_IU_state() { 3592 popa(); 3593 LP64_ONLY(addq(rsp, 8)); 3594 popf(); 3595 } 3596 3597 // Save Integer and Float state 3598 // Warning: Stack must be 16 byte aligned (64bit) 3599 void MacroAssembler::push_CPU_state() { 3600 push_IU_state(); 3601 push_FPU_state(); 3602 } 3603 3604 void MacroAssembler::push_FPU_state() { 3605 subptr(rsp, FPUStateSizeInWords * wordSize); 3606 #ifndef _LP64 3607 fnsave(Address(rsp, 0)); 3608 fwait(); 3609 #else 3610 fxsave(Address(rsp, 0)); 3611 #endif // LP64 3612 } 3613 3614 void MacroAssembler::push_IU_state() { 3615 // Push flags first because pusha kills them 3616 pushf(); 3617 // Make sure rsp stays 16-byte aligned 3618 LP64_ONLY(subq(rsp, 8)); 3619 pusha(); 3620 } 3621 3622 void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp, bool clear_pc) { 3623 // determine java_thread register 3624 if (!java_thread->is_valid()) { 3625 java_thread = rdi; 3626 get_thread(java_thread); 3627 } 3628 // we must set sp to zero to clear frame 3629 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD); 3630 if (clear_fp) { 3631 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD); 3632 } 3633 3634 if (clear_pc) 3635 movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD); 3636 3637 } 3638 3639 void MacroAssembler::restore_rax(Register tmp) { 3640 if (tmp == noreg) pop(rax); 3641 else if (tmp != rax) mov(rax, tmp); 3642 } 3643 3644 void MacroAssembler::round_to(Register reg, int modulus) { 3645 addptr(reg, modulus - 1); 3646 andptr(reg, -modulus); 3647 } 3648 3649 void MacroAssembler::save_rax(Register tmp) { 3650 if (tmp == noreg) push(rax); 3651 else if (tmp != rax) mov(tmp, rax); 3652 } 3653 3654 // Write serialization page so VM thread can do a pseudo remote membar. 3655 // We use the current thread pointer to calculate a thread specific 3656 // offset to write to within the page. This minimizes bus traffic 3657 // due to cache line collision. 3658 void MacroAssembler::serialize_memory(Register thread, Register tmp) { 3659 movl(tmp, thread); 3660 shrl(tmp, os::get_serialize_page_shift_count()); 3661 andl(tmp, (os::vm_page_size() - sizeof(int))); 3662 3663 Address index(noreg, tmp, Address::times_1); 3664 ExternalAddress page(os::get_memory_serialize_page()); 3665 3666 // Size of store must match masking code above 3667 movl(as_Address(ArrayAddress(page, index)), tmp); 3668 } 3669 3670 // Calls to C land 3671 // 3672 // When entering C land, the rbp, & rsp of the last Java frame have to be recorded 3673 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp 3674 // has to be reset to 0. This is required to allow proper stack traversal. 3675 void MacroAssembler::set_last_Java_frame(Register java_thread, 3676 Register last_java_sp, 3677 Register last_java_fp, 3678 address last_java_pc) { 3679 // determine java_thread register 3680 if (!java_thread->is_valid()) { 3681 java_thread = rdi; 3682 get_thread(java_thread); 3683 } 3684 // determine last_java_sp register 3685 if (!last_java_sp->is_valid()) { 3686 last_java_sp = rsp; 3687 } 3688 3689 // last_java_fp is optional 3690 3691 if (last_java_fp->is_valid()) { 3692 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp); 3693 } 3694 3695 // last_java_pc is optional 3696 3697 if (last_java_pc != NULL) { 3698 lea(Address(java_thread, 3699 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()), 3700 InternalAddress(last_java_pc)); 3701 3702 } 3703 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp); 3704 } 3705 3706 void MacroAssembler::shlptr(Register dst, int imm8) { 3707 LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8)); 3708 } 3709 3710 void MacroAssembler::shrptr(Register dst, int imm8) { 3711 LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8)); 3712 } 3713 3714 void MacroAssembler::sign_extend_byte(Register reg) { 3715 if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) { 3716 movsbl(reg, reg); // movsxb 3717 } else { 3718 shll(reg, 24); 3719 sarl(reg, 24); 3720 } 3721 } 3722 3723 void MacroAssembler::sign_extend_short(Register reg) { 3724 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3725 movswl(reg, reg); // movsxw 3726 } else { 3727 shll(reg, 16); 3728 sarl(reg, 16); 3729 } 3730 } 3731 3732 void MacroAssembler::testl(Register dst, AddressLiteral src) { 3733 assert(reachable(src), "Address should be reachable"); 3734 testl(dst, as_Address(src)); 3735 } 3736 3737 void MacroAssembler::pcmpeqb(XMMRegister dst, XMMRegister src) { 3738 int dst_enc = dst->encoding(); 3739 int src_enc = src->encoding(); 3740 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 3741 Assembler::pcmpeqb(dst, src); 3742 } else if ((dst_enc < 16) && (src_enc < 16)) { 3743 Assembler::pcmpeqb(dst, src); 3744 } else if (src_enc < 16) { 3745 subptr(rsp, 64); 3746 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3747 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3748 Assembler::pcmpeqb(xmm0, src); 3749 movdqu(dst, xmm0); 3750 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3751 addptr(rsp, 64); 3752 } else if (dst_enc < 16) { 3753 subptr(rsp, 64); 3754 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3755 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3756 Assembler::pcmpeqb(dst, xmm0); 3757 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3758 addptr(rsp, 64); 3759 } else { 3760 subptr(rsp, 64); 3761 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3762 subptr(rsp, 64); 3763 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 3764 movdqu(xmm0, src); 3765 movdqu(xmm1, dst); 3766 Assembler::pcmpeqb(xmm1, xmm0); 3767 movdqu(dst, xmm1); 3768 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 3769 addptr(rsp, 64); 3770 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3771 addptr(rsp, 64); 3772 } 3773 } 3774 3775 void MacroAssembler::pcmpeqw(XMMRegister dst, XMMRegister src) { 3776 int dst_enc = dst->encoding(); 3777 int src_enc = src->encoding(); 3778 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 3779 Assembler::pcmpeqw(dst, src); 3780 } else if ((dst_enc < 16) && (src_enc < 16)) { 3781 Assembler::pcmpeqw(dst, src); 3782 } else if (src_enc < 16) { 3783 subptr(rsp, 64); 3784 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3785 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3786 Assembler::pcmpeqw(xmm0, src); 3787 movdqu(dst, xmm0); 3788 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3789 addptr(rsp, 64); 3790 } else if (dst_enc < 16) { 3791 subptr(rsp, 64); 3792 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3793 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3794 Assembler::pcmpeqw(dst, xmm0); 3795 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3796 addptr(rsp, 64); 3797 } else { 3798 subptr(rsp, 64); 3799 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3800 subptr(rsp, 64); 3801 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 3802 movdqu(xmm0, src); 3803 movdqu(xmm1, dst); 3804 Assembler::pcmpeqw(xmm1, xmm0); 3805 movdqu(dst, xmm1); 3806 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 3807 addptr(rsp, 64); 3808 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3809 addptr(rsp, 64); 3810 } 3811 } 3812 3813 void MacroAssembler::pcmpestri(XMMRegister dst, Address src, int imm8) { 3814 int dst_enc = dst->encoding(); 3815 if (dst_enc < 16) { 3816 Assembler::pcmpestri(dst, src, imm8); 3817 } else { 3818 subptr(rsp, 64); 3819 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3820 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3821 Assembler::pcmpestri(xmm0, src, imm8); 3822 movdqu(dst, xmm0); 3823 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3824 addptr(rsp, 64); 3825 } 3826 } 3827 3828 void MacroAssembler::pcmpestri(XMMRegister dst, XMMRegister src, int imm8) { 3829 int dst_enc = dst->encoding(); 3830 int src_enc = src->encoding(); 3831 if ((dst_enc < 16) && (src_enc < 16)) { 3832 Assembler::pcmpestri(dst, src, imm8); 3833 } else if (src_enc < 16) { 3834 subptr(rsp, 64); 3835 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3836 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3837 Assembler::pcmpestri(xmm0, src, imm8); 3838 movdqu(dst, xmm0); 3839 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3840 addptr(rsp, 64); 3841 } else if (dst_enc < 16) { 3842 subptr(rsp, 64); 3843 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3844 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3845 Assembler::pcmpestri(dst, xmm0, imm8); 3846 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3847 addptr(rsp, 64); 3848 } else { 3849 subptr(rsp, 64); 3850 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3851 subptr(rsp, 64); 3852 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 3853 movdqu(xmm0, src); 3854 movdqu(xmm1, dst); 3855 Assembler::pcmpestri(xmm1, xmm0, imm8); 3856 movdqu(dst, xmm1); 3857 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 3858 addptr(rsp, 64); 3859 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3860 addptr(rsp, 64); 3861 } 3862 } 3863 3864 void MacroAssembler::pmovzxbw(XMMRegister dst, XMMRegister src) { 3865 int dst_enc = dst->encoding(); 3866 int src_enc = src->encoding(); 3867 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 3868 Assembler::pmovzxbw(dst, src); 3869 } else if ((dst_enc < 16) && (src_enc < 16)) { 3870 Assembler::pmovzxbw(dst, src); 3871 } else if (src_enc < 16) { 3872 subptr(rsp, 64); 3873 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3874 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3875 Assembler::pmovzxbw(xmm0, src); 3876 movdqu(dst, xmm0); 3877 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3878 addptr(rsp, 64); 3879 } else if (dst_enc < 16) { 3880 subptr(rsp, 64); 3881 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3882 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3883 Assembler::pmovzxbw(dst, xmm0); 3884 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3885 addptr(rsp, 64); 3886 } else { 3887 subptr(rsp, 64); 3888 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3889 subptr(rsp, 64); 3890 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 3891 movdqu(xmm0, src); 3892 movdqu(xmm1, dst); 3893 Assembler::pmovzxbw(xmm1, xmm0); 3894 movdqu(dst, xmm1); 3895 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 3896 addptr(rsp, 64); 3897 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3898 addptr(rsp, 64); 3899 } 3900 } 3901 3902 void MacroAssembler::pmovzxbw(XMMRegister dst, Address src) { 3903 int dst_enc = dst->encoding(); 3904 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 3905 Assembler::pmovzxbw(dst, src); 3906 } else if (dst_enc < 16) { 3907 Assembler::pmovzxbw(dst, src); 3908 } else { 3909 subptr(rsp, 64); 3910 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3911 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3912 Assembler::pmovzxbw(xmm0, src); 3913 movdqu(dst, xmm0); 3914 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3915 addptr(rsp, 64); 3916 } 3917 } 3918 3919 void MacroAssembler::pmovmskb(Register dst, XMMRegister src) { 3920 int src_enc = src->encoding(); 3921 if (src_enc < 16) { 3922 Assembler::pmovmskb(dst, src); 3923 } else { 3924 subptr(rsp, 64); 3925 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3926 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3927 Assembler::pmovmskb(dst, xmm0); 3928 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3929 addptr(rsp, 64); 3930 } 3931 } 3932 3933 void MacroAssembler::ptest(XMMRegister dst, XMMRegister src) { 3934 int dst_enc = dst->encoding(); 3935 int src_enc = src->encoding(); 3936 if ((dst_enc < 16) && (src_enc < 16)) { 3937 Assembler::ptest(dst, src); 3938 } else if (src_enc < 16) { 3939 subptr(rsp, 64); 3940 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3941 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 3942 Assembler::ptest(xmm0, src); 3943 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3944 addptr(rsp, 64); 3945 } else if (dst_enc < 16) { 3946 subptr(rsp, 64); 3947 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3948 evmovdqul(xmm0, src, Assembler::AVX_512bit); 3949 Assembler::ptest(dst, xmm0); 3950 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3951 addptr(rsp, 64); 3952 } else { 3953 subptr(rsp, 64); 3954 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 3955 subptr(rsp, 64); 3956 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 3957 movdqu(xmm0, src); 3958 movdqu(xmm1, dst); 3959 Assembler::ptest(xmm1, xmm0); 3960 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 3961 addptr(rsp, 64); 3962 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 3963 addptr(rsp, 64); 3964 } 3965 } 3966 3967 void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) { 3968 if (reachable(src)) { 3969 Assembler::sqrtsd(dst, as_Address(src)); 3970 } else { 3971 lea(rscratch1, src); 3972 Assembler::sqrtsd(dst, Address(rscratch1, 0)); 3973 } 3974 } 3975 3976 void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) { 3977 if (reachable(src)) { 3978 Assembler::sqrtss(dst, as_Address(src)); 3979 } else { 3980 lea(rscratch1, src); 3981 Assembler::sqrtss(dst, Address(rscratch1, 0)); 3982 } 3983 } 3984 3985 void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) { 3986 if (reachable(src)) { 3987 Assembler::subsd(dst, as_Address(src)); 3988 } else { 3989 lea(rscratch1, src); 3990 Assembler::subsd(dst, Address(rscratch1, 0)); 3991 } 3992 } 3993 3994 void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) { 3995 if (reachable(src)) { 3996 Assembler::subss(dst, as_Address(src)); 3997 } else { 3998 lea(rscratch1, src); 3999 Assembler::subss(dst, Address(rscratch1, 0)); 4000 } 4001 } 4002 4003 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) { 4004 if (reachable(src)) { 4005 Assembler::ucomisd(dst, as_Address(src)); 4006 } else { 4007 lea(rscratch1, src); 4008 Assembler::ucomisd(dst, Address(rscratch1, 0)); 4009 } 4010 } 4011 4012 void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) { 4013 if (reachable(src)) { 4014 Assembler::ucomiss(dst, as_Address(src)); 4015 } else { 4016 lea(rscratch1, src); 4017 Assembler::ucomiss(dst, Address(rscratch1, 0)); 4018 } 4019 } 4020 4021 void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src) { 4022 // Used in sign-bit flipping with aligned address. 4023 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 4024 if (reachable(src)) { 4025 Assembler::xorpd(dst, as_Address(src)); 4026 } else { 4027 lea(rscratch1, src); 4028 Assembler::xorpd(dst, Address(rscratch1, 0)); 4029 } 4030 } 4031 4032 void MacroAssembler::xorpd(XMMRegister dst, XMMRegister src) { 4033 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) { 4034 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit); 4035 } 4036 else { 4037 Assembler::xorpd(dst, src); 4038 } 4039 } 4040 4041 void MacroAssembler::xorps(XMMRegister dst, XMMRegister src) { 4042 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) { 4043 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit); 4044 } else { 4045 Assembler::xorps(dst, src); 4046 } 4047 } 4048 4049 void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src) { 4050 // Used in sign-bit flipping with aligned address. 4051 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 4052 if (reachable(src)) { 4053 Assembler::xorps(dst, as_Address(src)); 4054 } else { 4055 lea(rscratch1, src); 4056 Assembler::xorps(dst, Address(rscratch1, 0)); 4057 } 4058 } 4059 4060 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) { 4061 // Used in sign-bit flipping with aligned address. 4062 bool aligned_adr = (((intptr_t)src.target() & 15) == 0); 4063 assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes"); 4064 if (reachable(src)) { 4065 Assembler::pshufb(dst, as_Address(src)); 4066 } else { 4067 lea(rscratch1, src); 4068 Assembler::pshufb(dst, Address(rscratch1, 0)); 4069 } 4070 } 4071 4072 // AVX 3-operands instructions 4073 4074 void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4075 if (reachable(src)) { 4076 vaddsd(dst, nds, as_Address(src)); 4077 } else { 4078 lea(rscratch1, src); 4079 vaddsd(dst, nds, Address(rscratch1, 0)); 4080 } 4081 } 4082 4083 void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4084 if (reachable(src)) { 4085 vaddss(dst, nds, as_Address(src)); 4086 } else { 4087 lea(rscratch1, src); 4088 vaddss(dst, nds, Address(rscratch1, 0)); 4089 } 4090 } 4091 4092 void MacroAssembler::vabsss(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) { 4093 int dst_enc = dst->encoding(); 4094 int nds_enc = nds->encoding(); 4095 int src_enc = src->encoding(); 4096 if ((dst_enc < 16) && (nds_enc < 16)) { 4097 vandps(dst, nds, negate_field, vector_len); 4098 } else if ((src_enc < 16) && (dst_enc < 16)) { 4099 movss(src, nds); 4100 vandps(dst, src, negate_field, vector_len); 4101 } else if (src_enc < 16) { 4102 movss(src, nds); 4103 vandps(src, src, negate_field, vector_len); 4104 movss(dst, src); 4105 } else if (dst_enc < 16) { 4106 movdqu(src, xmm0); 4107 movss(xmm0, nds); 4108 vandps(dst, xmm0, negate_field, vector_len); 4109 movdqu(xmm0, src); 4110 } else if (nds_enc < 16) { 4111 movdqu(src, xmm0); 4112 vandps(xmm0, nds, negate_field, vector_len); 4113 movss(dst, xmm0); 4114 movdqu(xmm0, src); 4115 } else { 4116 movdqu(src, xmm0); 4117 movss(xmm0, nds); 4118 vandps(xmm0, xmm0, negate_field, vector_len); 4119 movss(dst, xmm0); 4120 movdqu(xmm0, src); 4121 } 4122 } 4123 4124 void MacroAssembler::vabssd(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) { 4125 int dst_enc = dst->encoding(); 4126 int nds_enc = nds->encoding(); 4127 int src_enc = src->encoding(); 4128 if ((dst_enc < 16) && (nds_enc < 16)) { 4129 vandpd(dst, nds, negate_field, vector_len); 4130 } else if ((src_enc < 16) && (dst_enc < 16)) { 4131 movsd(src, nds); 4132 vandpd(dst, src, negate_field, vector_len); 4133 } else if (src_enc < 16) { 4134 movsd(src, nds); 4135 vandpd(src, src, negate_field, vector_len); 4136 movsd(dst, src); 4137 } else if (dst_enc < 16) { 4138 movdqu(src, xmm0); 4139 movsd(xmm0, nds); 4140 vandpd(dst, xmm0, negate_field, vector_len); 4141 movdqu(xmm0, src); 4142 } else if (nds_enc < 16) { 4143 movdqu(src, xmm0); 4144 vandpd(xmm0, nds, negate_field, vector_len); 4145 movsd(dst, xmm0); 4146 movdqu(xmm0, src); 4147 } else { 4148 movdqu(src, xmm0); 4149 movsd(xmm0, nds); 4150 vandpd(xmm0, xmm0, negate_field, vector_len); 4151 movsd(dst, xmm0); 4152 movdqu(xmm0, src); 4153 } 4154 } 4155 4156 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4157 int dst_enc = dst->encoding(); 4158 int nds_enc = nds->encoding(); 4159 int src_enc = src->encoding(); 4160 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4161 Assembler::vpaddb(dst, nds, src, vector_len); 4162 } else if ((dst_enc < 16) && (src_enc < 16)) { 4163 Assembler::vpaddb(dst, dst, src, vector_len); 4164 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4165 // use nds as scratch for src 4166 evmovdqul(nds, src, Assembler::AVX_512bit); 4167 Assembler::vpaddb(dst, dst, nds, vector_len); 4168 } else if ((src_enc < 16) && (nds_enc < 16)) { 4169 // use nds as scratch for dst 4170 evmovdqul(nds, dst, Assembler::AVX_512bit); 4171 Assembler::vpaddb(nds, nds, src, vector_len); 4172 evmovdqul(dst, nds, Assembler::AVX_512bit); 4173 } else if (dst_enc < 16) { 4174 // use nds as scatch for xmm0 to hold src 4175 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4176 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4177 Assembler::vpaddb(dst, dst, xmm0, vector_len); 4178 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4179 } else { 4180 // worse case scenario, all regs are in the upper bank 4181 subptr(rsp, 64); 4182 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4183 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4184 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4185 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4186 Assembler::vpaddb(xmm0, xmm0, xmm1, vector_len); 4187 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4188 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4189 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4190 addptr(rsp, 64); 4191 } 4192 } 4193 4194 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4195 int dst_enc = dst->encoding(); 4196 int nds_enc = nds->encoding(); 4197 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4198 Assembler::vpaddb(dst, nds, src, vector_len); 4199 } else if (dst_enc < 16) { 4200 Assembler::vpaddb(dst, dst, src, vector_len); 4201 } else if (nds_enc < 16) { 4202 // implies dst_enc in upper bank with src as scratch 4203 evmovdqul(nds, dst, Assembler::AVX_512bit); 4204 Assembler::vpaddb(nds, nds, src, vector_len); 4205 evmovdqul(dst, nds, Assembler::AVX_512bit); 4206 } else { 4207 // worse case scenario, all regs in upper bank 4208 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4209 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4210 Assembler::vpaddb(xmm0, xmm0, src, vector_len); 4211 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4212 } 4213 } 4214 4215 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4216 int dst_enc = dst->encoding(); 4217 int nds_enc = nds->encoding(); 4218 int src_enc = src->encoding(); 4219 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4220 Assembler::vpaddw(dst, nds, src, vector_len); 4221 } else if ((dst_enc < 16) && (src_enc < 16)) { 4222 Assembler::vpaddw(dst, dst, src, vector_len); 4223 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4224 // use nds as scratch for src 4225 evmovdqul(nds, src, Assembler::AVX_512bit); 4226 Assembler::vpaddw(dst, dst, nds, vector_len); 4227 } else if ((src_enc < 16) && (nds_enc < 16)) { 4228 // use nds as scratch for dst 4229 evmovdqul(nds, dst, Assembler::AVX_512bit); 4230 Assembler::vpaddw(nds, nds, src, vector_len); 4231 evmovdqul(dst, nds, Assembler::AVX_512bit); 4232 } else if (dst_enc < 16) { 4233 // use nds as scatch for xmm0 to hold src 4234 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4235 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4236 Assembler::vpaddw(dst, dst, xmm0, vector_len); 4237 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4238 } else { 4239 // worse case scenario, all regs are in the upper bank 4240 subptr(rsp, 64); 4241 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4242 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4243 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4244 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4245 Assembler::vpaddw(xmm0, xmm0, xmm1, vector_len); 4246 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4247 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4248 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4249 addptr(rsp, 64); 4250 } 4251 } 4252 4253 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4254 int dst_enc = dst->encoding(); 4255 int nds_enc = nds->encoding(); 4256 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4257 Assembler::vpaddw(dst, nds, src, vector_len); 4258 } else if (dst_enc < 16) { 4259 Assembler::vpaddw(dst, dst, src, vector_len); 4260 } else if (nds_enc < 16) { 4261 // implies dst_enc in upper bank with src as scratch 4262 evmovdqul(nds, dst, Assembler::AVX_512bit); 4263 Assembler::vpaddw(nds, nds, src, vector_len); 4264 evmovdqul(dst, nds, Assembler::AVX_512bit); 4265 } else { 4266 // worse case scenario, all regs in upper bank 4267 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4268 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4269 Assembler::vpaddw(xmm0, xmm0, src, vector_len); 4270 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4271 } 4272 } 4273 4274 void MacroAssembler::vpbroadcastw(XMMRegister dst, XMMRegister src) { 4275 int dst_enc = dst->encoding(); 4276 int src_enc = src->encoding(); 4277 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4278 Assembler::vpbroadcastw(dst, src); 4279 } else if ((dst_enc < 16) && (src_enc < 16)) { 4280 Assembler::vpbroadcastw(dst, src); 4281 } else if (src_enc < 16) { 4282 subptr(rsp, 64); 4283 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4284 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4285 Assembler::vpbroadcastw(xmm0, src); 4286 movdqu(dst, xmm0); 4287 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4288 addptr(rsp, 64); 4289 } else if (dst_enc < 16) { 4290 subptr(rsp, 64); 4291 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4292 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4293 Assembler::vpbroadcastw(dst, xmm0); 4294 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4295 addptr(rsp, 64); 4296 } else { 4297 subptr(rsp, 64); 4298 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4299 subptr(rsp, 64); 4300 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4301 movdqu(xmm0, src); 4302 movdqu(xmm1, dst); 4303 Assembler::vpbroadcastw(xmm1, xmm0); 4304 movdqu(dst, xmm1); 4305 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4306 addptr(rsp, 64); 4307 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4308 addptr(rsp, 64); 4309 } 4310 } 4311 4312 void MacroAssembler::vpcmpeqb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4313 int dst_enc = dst->encoding(); 4314 int nds_enc = nds->encoding(); 4315 int src_enc = src->encoding(); 4316 assert(dst_enc == nds_enc, ""); 4317 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4318 Assembler::vpcmpeqb(dst, nds, src, vector_len); 4319 } else if ((dst_enc < 16) && (src_enc < 16)) { 4320 Assembler::vpcmpeqb(dst, nds, src, vector_len); 4321 } else if (src_enc < 16) { 4322 subptr(rsp, 64); 4323 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4324 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4325 Assembler::vpcmpeqb(xmm0, xmm0, src, vector_len); 4326 movdqu(dst, xmm0); 4327 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4328 addptr(rsp, 64); 4329 } else if (dst_enc < 16) { 4330 subptr(rsp, 64); 4331 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4332 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4333 Assembler::vpcmpeqb(dst, dst, xmm0, vector_len); 4334 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4335 addptr(rsp, 64); 4336 } else { 4337 subptr(rsp, 64); 4338 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4339 subptr(rsp, 64); 4340 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4341 movdqu(xmm0, src); 4342 movdqu(xmm1, dst); 4343 Assembler::vpcmpeqb(xmm1, xmm1, xmm0, vector_len); 4344 movdqu(dst, xmm1); 4345 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4346 addptr(rsp, 64); 4347 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4348 addptr(rsp, 64); 4349 } 4350 } 4351 4352 void MacroAssembler::vpcmpeqw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4353 int dst_enc = dst->encoding(); 4354 int nds_enc = nds->encoding(); 4355 int src_enc = src->encoding(); 4356 assert(dst_enc == nds_enc, ""); 4357 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4358 Assembler::vpcmpeqw(dst, nds, src, vector_len); 4359 } else if ((dst_enc < 16) && (src_enc < 16)) { 4360 Assembler::vpcmpeqw(dst, nds, src, vector_len); 4361 } else if (src_enc < 16) { 4362 subptr(rsp, 64); 4363 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4364 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4365 Assembler::vpcmpeqw(xmm0, xmm0, src, vector_len); 4366 movdqu(dst, xmm0); 4367 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4368 addptr(rsp, 64); 4369 } else if (dst_enc < 16) { 4370 subptr(rsp, 64); 4371 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4372 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4373 Assembler::vpcmpeqw(dst, dst, xmm0, vector_len); 4374 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4375 addptr(rsp, 64); 4376 } else { 4377 subptr(rsp, 64); 4378 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4379 subptr(rsp, 64); 4380 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4381 movdqu(xmm0, src); 4382 movdqu(xmm1, dst); 4383 Assembler::vpcmpeqw(xmm1, xmm1, xmm0, vector_len); 4384 movdqu(dst, xmm1); 4385 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4386 addptr(rsp, 64); 4387 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4388 addptr(rsp, 64); 4389 } 4390 } 4391 4392 void MacroAssembler::vpmovzxbw(XMMRegister dst, Address src, int vector_len) { 4393 int dst_enc = dst->encoding(); 4394 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4395 Assembler::vpmovzxbw(dst, src, vector_len); 4396 } else if (dst_enc < 16) { 4397 Assembler::vpmovzxbw(dst, src, vector_len); 4398 } else { 4399 subptr(rsp, 64); 4400 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4401 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4402 Assembler::vpmovzxbw(xmm0, src, vector_len); 4403 movdqu(dst, xmm0); 4404 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4405 addptr(rsp, 64); 4406 } 4407 } 4408 4409 void MacroAssembler::vpmovmskb(Register dst, XMMRegister src) { 4410 int src_enc = src->encoding(); 4411 if (src_enc < 16) { 4412 Assembler::vpmovmskb(dst, src); 4413 } else { 4414 subptr(rsp, 64); 4415 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4416 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4417 Assembler::vpmovmskb(dst, xmm0); 4418 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4419 addptr(rsp, 64); 4420 } 4421 } 4422 4423 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4424 int dst_enc = dst->encoding(); 4425 int nds_enc = nds->encoding(); 4426 int src_enc = src->encoding(); 4427 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4428 Assembler::vpmullw(dst, nds, src, vector_len); 4429 } else if ((dst_enc < 16) && (src_enc < 16)) { 4430 Assembler::vpmullw(dst, dst, src, vector_len); 4431 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4432 // use nds as scratch for src 4433 evmovdqul(nds, src, Assembler::AVX_512bit); 4434 Assembler::vpmullw(dst, dst, nds, vector_len); 4435 } else if ((src_enc < 16) && (nds_enc < 16)) { 4436 // use nds as scratch for dst 4437 evmovdqul(nds, dst, Assembler::AVX_512bit); 4438 Assembler::vpmullw(nds, nds, src, vector_len); 4439 evmovdqul(dst, nds, Assembler::AVX_512bit); 4440 } else if (dst_enc < 16) { 4441 // use nds as scatch for xmm0 to hold src 4442 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4443 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4444 Assembler::vpmullw(dst, dst, xmm0, vector_len); 4445 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4446 } else { 4447 // worse case scenario, all regs are in the upper bank 4448 subptr(rsp, 64); 4449 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4450 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4451 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4452 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4453 Assembler::vpmullw(xmm0, xmm0, xmm1, vector_len); 4454 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4455 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4456 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4457 addptr(rsp, 64); 4458 } 4459 } 4460 4461 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4462 int dst_enc = dst->encoding(); 4463 int nds_enc = nds->encoding(); 4464 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4465 Assembler::vpmullw(dst, nds, src, vector_len); 4466 } else if (dst_enc < 16) { 4467 Assembler::vpmullw(dst, dst, src, vector_len); 4468 } else if (nds_enc < 16) { 4469 // implies dst_enc in upper bank with src as scratch 4470 evmovdqul(nds, dst, Assembler::AVX_512bit); 4471 Assembler::vpmullw(nds, nds, src, vector_len); 4472 evmovdqul(dst, nds, Assembler::AVX_512bit); 4473 } else { 4474 // worse case scenario, all regs in upper bank 4475 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4476 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4477 Assembler::vpmullw(xmm0, xmm0, src, vector_len); 4478 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4479 } 4480 } 4481 4482 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4483 int dst_enc = dst->encoding(); 4484 int nds_enc = nds->encoding(); 4485 int src_enc = src->encoding(); 4486 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4487 Assembler::vpsubb(dst, nds, src, vector_len); 4488 } else if ((dst_enc < 16) && (src_enc < 16)) { 4489 Assembler::vpsubb(dst, dst, src, vector_len); 4490 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4491 // use nds as scratch for src 4492 evmovdqul(nds, src, Assembler::AVX_512bit); 4493 Assembler::vpsubb(dst, dst, nds, vector_len); 4494 } else if ((src_enc < 16) && (nds_enc < 16)) { 4495 // use nds as scratch for dst 4496 evmovdqul(nds, dst, Assembler::AVX_512bit); 4497 Assembler::vpsubb(nds, nds, src, vector_len); 4498 evmovdqul(dst, nds, Assembler::AVX_512bit); 4499 } else if (dst_enc < 16) { 4500 // use nds as scatch for xmm0 to hold src 4501 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4502 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4503 Assembler::vpsubb(dst, dst, xmm0, vector_len); 4504 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4505 } else { 4506 // worse case scenario, all regs are in the upper bank 4507 subptr(rsp, 64); 4508 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4509 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4510 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4511 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4512 Assembler::vpsubb(xmm0, xmm0, xmm1, vector_len); 4513 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4514 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4515 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4516 addptr(rsp, 64); 4517 } 4518 } 4519 4520 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4521 int dst_enc = dst->encoding(); 4522 int nds_enc = nds->encoding(); 4523 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4524 Assembler::vpsubb(dst, nds, src, vector_len); 4525 } else if (dst_enc < 16) { 4526 Assembler::vpsubb(dst, dst, src, vector_len); 4527 } else if (nds_enc < 16) { 4528 // implies dst_enc in upper bank with src as scratch 4529 evmovdqul(nds, dst, Assembler::AVX_512bit); 4530 Assembler::vpsubb(nds, nds, src, vector_len); 4531 evmovdqul(dst, nds, Assembler::AVX_512bit); 4532 } else { 4533 // worse case scenario, all regs in upper bank 4534 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4535 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4536 Assembler::vpsubw(xmm0, xmm0, src, vector_len); 4537 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4538 } 4539 } 4540 4541 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4542 int dst_enc = dst->encoding(); 4543 int nds_enc = nds->encoding(); 4544 int src_enc = src->encoding(); 4545 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4546 Assembler::vpsubw(dst, nds, src, vector_len); 4547 } else if ((dst_enc < 16) && (src_enc < 16)) { 4548 Assembler::vpsubw(dst, dst, src, vector_len); 4549 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4550 // use nds as scratch for src 4551 evmovdqul(nds, src, Assembler::AVX_512bit); 4552 Assembler::vpsubw(dst, dst, nds, vector_len); 4553 } else if ((src_enc < 16) && (nds_enc < 16)) { 4554 // use nds as scratch for dst 4555 evmovdqul(nds, dst, Assembler::AVX_512bit); 4556 Assembler::vpsubw(nds, nds, src, vector_len); 4557 evmovdqul(dst, nds, Assembler::AVX_512bit); 4558 } else if (dst_enc < 16) { 4559 // use nds as scatch for xmm0 to hold src 4560 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4561 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4562 Assembler::vpsubw(dst, dst, xmm0, vector_len); 4563 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4564 } else { 4565 // worse case scenario, all regs are in the upper bank 4566 subptr(rsp, 64); 4567 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4568 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4569 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4570 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4571 Assembler::vpsubw(xmm0, xmm0, xmm1, vector_len); 4572 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4573 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4574 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4575 addptr(rsp, 64); 4576 } 4577 } 4578 4579 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 4580 int dst_enc = dst->encoding(); 4581 int nds_enc = nds->encoding(); 4582 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4583 Assembler::vpsubw(dst, nds, src, vector_len); 4584 } else if (dst_enc < 16) { 4585 Assembler::vpsubw(dst, dst, src, vector_len); 4586 } else if (nds_enc < 16) { 4587 // implies dst_enc in upper bank with src as scratch 4588 evmovdqul(nds, dst, Assembler::AVX_512bit); 4589 Assembler::vpsubw(nds, nds, src, vector_len); 4590 evmovdqul(dst, nds, Assembler::AVX_512bit); 4591 } else { 4592 // worse case scenario, all regs in upper bank 4593 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4594 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4595 Assembler::vpsubw(xmm0, xmm0, src, vector_len); 4596 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4597 } 4598 } 4599 4600 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4601 int dst_enc = dst->encoding(); 4602 int nds_enc = nds->encoding(); 4603 int shift_enc = shift->encoding(); 4604 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4605 Assembler::vpsraw(dst, nds, shift, vector_len); 4606 } else if ((dst_enc < 16) && (shift_enc < 16)) { 4607 Assembler::vpsraw(dst, dst, shift, vector_len); 4608 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4609 // use nds_enc as scratch with shift 4610 evmovdqul(nds, shift, Assembler::AVX_512bit); 4611 Assembler::vpsraw(dst, dst, nds, vector_len); 4612 } else if ((shift_enc < 16) && (nds_enc < 16)) { 4613 // use nds as scratch with dst 4614 evmovdqul(nds, dst, Assembler::AVX_512bit); 4615 Assembler::vpsraw(nds, nds, shift, vector_len); 4616 evmovdqul(dst, nds, Assembler::AVX_512bit); 4617 } else if (dst_enc < 16) { 4618 // use nds to save a copy of xmm0 and hold shift 4619 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4620 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4621 Assembler::vpsraw(dst, dst, xmm0, vector_len); 4622 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4623 } else if (nds_enc < 16) { 4624 // use nds as dest as temps 4625 evmovdqul(nds, dst, Assembler::AVX_512bit); 4626 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4627 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4628 Assembler::vpsraw(nds, nds, xmm0, vector_len); 4629 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4630 evmovdqul(dst, nds, Assembler::AVX_512bit); 4631 } else { 4632 // worse case scenario, all regs are in the upper bank 4633 subptr(rsp, 64); 4634 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4635 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4636 evmovdqul(xmm1, shift, Assembler::AVX_512bit); 4637 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4638 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len); 4639 evmovdqul(xmm1, dst, Assembler::AVX_512bit); 4640 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4641 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4642 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4643 addptr(rsp, 64); 4644 } 4645 } 4646 4647 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 4648 int dst_enc = dst->encoding(); 4649 int nds_enc = nds->encoding(); 4650 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4651 Assembler::vpsraw(dst, nds, shift, vector_len); 4652 } else if (dst_enc < 16) { 4653 Assembler::vpsraw(dst, dst, shift, vector_len); 4654 } else if (nds_enc < 16) { 4655 // use nds as scratch 4656 evmovdqul(nds, dst, Assembler::AVX_512bit); 4657 Assembler::vpsraw(nds, nds, shift, vector_len); 4658 evmovdqul(dst, nds, Assembler::AVX_512bit); 4659 } else { 4660 // use nds as scratch for xmm0 4661 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4662 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4663 Assembler::vpsraw(xmm0, xmm0, shift, vector_len); 4664 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4665 } 4666 } 4667 4668 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4669 int dst_enc = dst->encoding(); 4670 int nds_enc = nds->encoding(); 4671 int shift_enc = shift->encoding(); 4672 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4673 Assembler::vpsrlw(dst, nds, shift, vector_len); 4674 } else if ((dst_enc < 16) && (shift_enc < 16)) { 4675 Assembler::vpsrlw(dst, dst, shift, vector_len); 4676 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4677 // use nds_enc as scratch with shift 4678 evmovdqul(nds, shift, Assembler::AVX_512bit); 4679 Assembler::vpsrlw(dst, dst, nds, vector_len); 4680 } else if ((shift_enc < 16) && (nds_enc < 16)) { 4681 // use nds as scratch with dst 4682 evmovdqul(nds, dst, Assembler::AVX_512bit); 4683 Assembler::vpsrlw(nds, nds, shift, vector_len); 4684 evmovdqul(dst, nds, Assembler::AVX_512bit); 4685 } else if (dst_enc < 16) { 4686 // use nds to save a copy of xmm0 and hold shift 4687 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4688 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4689 Assembler::vpsrlw(dst, dst, xmm0, vector_len); 4690 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4691 } else if (nds_enc < 16) { 4692 // use nds as dest as temps 4693 evmovdqul(nds, dst, Assembler::AVX_512bit); 4694 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4695 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4696 Assembler::vpsrlw(nds, nds, xmm0, vector_len); 4697 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4698 evmovdqul(dst, nds, Assembler::AVX_512bit); 4699 } else { 4700 // worse case scenario, all regs are in the upper bank 4701 subptr(rsp, 64); 4702 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4703 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4704 evmovdqul(xmm1, shift, Assembler::AVX_512bit); 4705 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4706 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len); 4707 evmovdqul(xmm1, dst, Assembler::AVX_512bit); 4708 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4709 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4710 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4711 addptr(rsp, 64); 4712 } 4713 } 4714 4715 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 4716 int dst_enc = dst->encoding(); 4717 int nds_enc = nds->encoding(); 4718 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4719 Assembler::vpsrlw(dst, nds, shift, vector_len); 4720 } else if (dst_enc < 16) { 4721 Assembler::vpsrlw(dst, dst, shift, vector_len); 4722 } else if (nds_enc < 16) { 4723 // use nds as scratch 4724 evmovdqul(nds, dst, Assembler::AVX_512bit); 4725 Assembler::vpsrlw(nds, nds, shift, vector_len); 4726 evmovdqul(dst, nds, Assembler::AVX_512bit); 4727 } else { 4728 // use nds as scratch for xmm0 4729 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4730 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4731 Assembler::vpsrlw(xmm0, xmm0, shift, vector_len); 4732 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4733 } 4734 } 4735 4736 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4737 int dst_enc = dst->encoding(); 4738 int nds_enc = nds->encoding(); 4739 int shift_enc = shift->encoding(); 4740 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4741 Assembler::vpsllw(dst, nds, shift, vector_len); 4742 } else if ((dst_enc < 16) && (shift_enc < 16)) { 4743 Assembler::vpsllw(dst, dst, shift, vector_len); 4744 } else if ((dst_enc < 16) && (nds_enc < 16)) { 4745 // use nds_enc as scratch with shift 4746 evmovdqul(nds, shift, Assembler::AVX_512bit); 4747 Assembler::vpsllw(dst, dst, nds, vector_len); 4748 } else if ((shift_enc < 16) && (nds_enc < 16)) { 4749 // use nds as scratch with dst 4750 evmovdqul(nds, dst, Assembler::AVX_512bit); 4751 Assembler::vpsllw(nds, nds, shift, vector_len); 4752 evmovdqul(dst, nds, Assembler::AVX_512bit); 4753 } else if (dst_enc < 16) { 4754 // use nds to save a copy of xmm0 and hold shift 4755 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4756 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4757 Assembler::vpsllw(dst, dst, xmm0, vector_len); 4758 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4759 } else if (nds_enc < 16) { 4760 // use nds as dest as temps 4761 evmovdqul(nds, dst, Assembler::AVX_512bit); 4762 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4763 evmovdqul(xmm0, shift, Assembler::AVX_512bit); 4764 Assembler::vpsllw(nds, nds, xmm0, vector_len); 4765 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4766 evmovdqul(dst, nds, Assembler::AVX_512bit); 4767 } else { 4768 // worse case scenario, all regs are in the upper bank 4769 subptr(rsp, 64); 4770 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4771 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4772 evmovdqul(xmm1, shift, Assembler::AVX_512bit); 4773 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4774 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len); 4775 evmovdqul(xmm1, dst, Assembler::AVX_512bit); 4776 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4777 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4778 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4779 addptr(rsp, 64); 4780 } 4781 } 4782 4783 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 4784 int dst_enc = dst->encoding(); 4785 int nds_enc = nds->encoding(); 4786 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) { 4787 Assembler::vpsllw(dst, nds, shift, vector_len); 4788 } else if (dst_enc < 16) { 4789 Assembler::vpsllw(dst, dst, shift, vector_len); 4790 } else if (nds_enc < 16) { 4791 // use nds as scratch 4792 evmovdqul(nds, dst, Assembler::AVX_512bit); 4793 Assembler::vpsllw(nds, nds, shift, vector_len); 4794 evmovdqul(dst, nds, Assembler::AVX_512bit); 4795 } else { 4796 // use nds as scratch for xmm0 4797 evmovdqul(nds, xmm0, Assembler::AVX_512bit); 4798 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4799 Assembler::vpsllw(xmm0, xmm0, shift, vector_len); 4800 evmovdqul(xmm0, nds, Assembler::AVX_512bit); 4801 } 4802 } 4803 4804 void MacroAssembler::vptest(XMMRegister dst, XMMRegister src) { 4805 int dst_enc = dst->encoding(); 4806 int src_enc = src->encoding(); 4807 if ((dst_enc < 16) && (src_enc < 16)) { 4808 Assembler::vptest(dst, src); 4809 } else if (src_enc < 16) { 4810 subptr(rsp, 64); 4811 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4812 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4813 Assembler::vptest(xmm0, src); 4814 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4815 addptr(rsp, 64); 4816 } else if (dst_enc < 16) { 4817 subptr(rsp, 64); 4818 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4819 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4820 Assembler::vptest(dst, xmm0); 4821 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4822 addptr(rsp, 64); 4823 } else { 4824 subptr(rsp, 64); 4825 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4826 subptr(rsp, 64); 4827 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4828 movdqu(xmm0, src); 4829 movdqu(xmm1, dst); 4830 Assembler::vptest(xmm1, xmm0); 4831 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4832 addptr(rsp, 64); 4833 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4834 addptr(rsp, 64); 4835 } 4836 } 4837 4838 // This instruction exists within macros, ergo we cannot control its input 4839 // when emitted through those patterns. 4840 void MacroAssembler::punpcklbw(XMMRegister dst, XMMRegister src) { 4841 if (VM_Version::supports_avx512nobw()) { 4842 int dst_enc = dst->encoding(); 4843 int src_enc = src->encoding(); 4844 if (dst_enc == src_enc) { 4845 if (dst_enc < 16) { 4846 Assembler::punpcklbw(dst, src); 4847 } else { 4848 subptr(rsp, 64); 4849 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4850 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4851 Assembler::punpcklbw(xmm0, xmm0); 4852 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4853 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4854 addptr(rsp, 64); 4855 } 4856 } else { 4857 if ((src_enc < 16) && (dst_enc < 16)) { 4858 Assembler::punpcklbw(dst, src); 4859 } else if (src_enc < 16) { 4860 subptr(rsp, 64); 4861 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4862 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4863 Assembler::punpcklbw(xmm0, src); 4864 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4865 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4866 addptr(rsp, 64); 4867 } else if (dst_enc < 16) { 4868 subptr(rsp, 64); 4869 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4870 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4871 Assembler::punpcklbw(dst, xmm0); 4872 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4873 addptr(rsp, 64); 4874 } else { 4875 subptr(rsp, 64); 4876 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4877 subptr(rsp, 64); 4878 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4879 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4880 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4881 Assembler::punpcklbw(xmm0, xmm1); 4882 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4883 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4884 addptr(rsp, 64); 4885 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4886 addptr(rsp, 64); 4887 } 4888 } 4889 } else { 4890 Assembler::punpcklbw(dst, src); 4891 } 4892 } 4893 4894 // This instruction exists within macros, ergo we cannot control its input 4895 // when emitted through those patterns. 4896 void MacroAssembler::pshuflw(XMMRegister dst, XMMRegister src, int mode) { 4897 if (VM_Version::supports_avx512nobw()) { 4898 int dst_enc = dst->encoding(); 4899 int src_enc = src->encoding(); 4900 if (dst_enc == src_enc) { 4901 if (dst_enc < 16) { 4902 Assembler::pshuflw(dst, src, mode); 4903 } else { 4904 subptr(rsp, 64); 4905 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4906 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4907 Assembler::pshuflw(xmm0, xmm0, mode); 4908 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4909 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4910 addptr(rsp, 64); 4911 } 4912 } else { 4913 if ((src_enc < 16) && (dst_enc < 16)) { 4914 Assembler::pshuflw(dst, src, mode); 4915 } else if (src_enc < 16) { 4916 subptr(rsp, 64); 4917 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4918 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4919 Assembler::pshuflw(xmm0, src, mode); 4920 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4921 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4922 addptr(rsp, 64); 4923 } else if (dst_enc < 16) { 4924 subptr(rsp, 64); 4925 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4926 evmovdqul(xmm0, src, Assembler::AVX_512bit); 4927 Assembler::pshuflw(dst, xmm0, mode); 4928 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4929 addptr(rsp, 64); 4930 } else { 4931 subptr(rsp, 64); 4932 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 4933 subptr(rsp, 64); 4934 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit); 4935 evmovdqul(xmm0, dst, Assembler::AVX_512bit); 4936 evmovdqul(xmm1, src, Assembler::AVX_512bit); 4937 Assembler::pshuflw(xmm0, xmm1, mode); 4938 evmovdqul(dst, xmm0, Assembler::AVX_512bit); 4939 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit); 4940 addptr(rsp, 64); 4941 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 4942 addptr(rsp, 64); 4943 } 4944 } 4945 } else { 4946 Assembler::pshuflw(dst, src, mode); 4947 } 4948 } 4949 4950 void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 4951 if (reachable(src)) { 4952 vandpd(dst, nds, as_Address(src), vector_len); 4953 } else { 4954 lea(rscratch1, src); 4955 vandpd(dst, nds, Address(rscratch1, 0), vector_len); 4956 } 4957 } 4958 4959 void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 4960 if (reachable(src)) { 4961 vandps(dst, nds, as_Address(src), vector_len); 4962 } else { 4963 lea(rscratch1, src); 4964 vandps(dst, nds, Address(rscratch1, 0), vector_len); 4965 } 4966 } 4967 4968 void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4969 if (reachable(src)) { 4970 vdivsd(dst, nds, as_Address(src)); 4971 } else { 4972 lea(rscratch1, src); 4973 vdivsd(dst, nds, Address(rscratch1, 0)); 4974 } 4975 } 4976 4977 void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4978 if (reachable(src)) { 4979 vdivss(dst, nds, as_Address(src)); 4980 } else { 4981 lea(rscratch1, src); 4982 vdivss(dst, nds, Address(rscratch1, 0)); 4983 } 4984 } 4985 4986 void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4987 if (reachable(src)) { 4988 vmulsd(dst, nds, as_Address(src)); 4989 } else { 4990 lea(rscratch1, src); 4991 vmulsd(dst, nds, Address(rscratch1, 0)); 4992 } 4993 } 4994 4995 void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4996 if (reachable(src)) { 4997 vmulss(dst, nds, as_Address(src)); 4998 } else { 4999 lea(rscratch1, src); 5000 vmulss(dst, nds, Address(rscratch1, 0)); 5001 } 5002 } 5003 5004 void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5005 if (reachable(src)) { 5006 vsubsd(dst, nds, as_Address(src)); 5007 } else { 5008 lea(rscratch1, src); 5009 vsubsd(dst, nds, Address(rscratch1, 0)); 5010 } 5011 } 5012 5013 void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5014 if (reachable(src)) { 5015 vsubss(dst, nds, as_Address(src)); 5016 } else { 5017 lea(rscratch1, src); 5018 vsubss(dst, nds, Address(rscratch1, 0)); 5019 } 5020 } 5021 5022 void MacroAssembler::vnegatess(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5023 int nds_enc = nds->encoding(); 5024 int dst_enc = dst->encoding(); 5025 bool dst_upper_bank = (dst_enc > 15); 5026 bool nds_upper_bank = (nds_enc > 15); 5027 if (VM_Version::supports_avx512novl() && 5028 (nds_upper_bank || dst_upper_bank)) { 5029 if (dst_upper_bank) { 5030 subptr(rsp, 64); 5031 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5032 movflt(xmm0, nds); 5033 vxorps(xmm0, xmm0, src, Assembler::AVX_128bit); 5034 movflt(dst, xmm0); 5035 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5036 addptr(rsp, 64); 5037 } else { 5038 movflt(dst, nds); 5039 vxorps(dst, dst, src, Assembler::AVX_128bit); 5040 } 5041 } else { 5042 vxorps(dst, nds, src, Assembler::AVX_128bit); 5043 } 5044 } 5045 5046 void MacroAssembler::vnegatesd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 5047 int nds_enc = nds->encoding(); 5048 int dst_enc = dst->encoding(); 5049 bool dst_upper_bank = (dst_enc > 15); 5050 bool nds_upper_bank = (nds_enc > 15); 5051 if (VM_Version::supports_avx512novl() && 5052 (nds_upper_bank || dst_upper_bank)) { 5053 if (dst_upper_bank) { 5054 subptr(rsp, 64); 5055 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit); 5056 movdbl(xmm0, nds); 5057 vxorpd(xmm0, xmm0, src, Assembler::AVX_128bit); 5058 movdbl(dst, xmm0); 5059 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit); 5060 addptr(rsp, 64); 5061 } else { 5062 movdbl(dst, nds); 5063 vxorpd(dst, dst, src, Assembler::AVX_128bit); 5064 } 5065 } else { 5066 vxorpd(dst, nds, src, Assembler::AVX_128bit); 5067 } 5068 } 5069 5070 void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 5071 if (reachable(src)) { 5072 vxorpd(dst, nds, as_Address(src), vector_len); 5073 } else { 5074 lea(rscratch1, src); 5075 vxorpd(dst, nds, Address(rscratch1, 0), vector_len); 5076 } 5077 } 5078 5079 void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 5080 if (reachable(src)) { 5081 vxorps(dst, nds, as_Address(src), vector_len); 5082 } else { 5083 lea(rscratch1, src); 5084 vxorps(dst, nds, Address(rscratch1, 0), vector_len); 5085 } 5086 } 5087 5088 5089 ////////////////////////////////////////////////////////////////////////////////// 5090 #if INCLUDE_ALL_GCS 5091 5092 void MacroAssembler::g1_write_barrier_pre(Register obj, 5093 Register pre_val, 5094 Register thread, 5095 Register tmp, 5096 bool tosca_live, 5097 bool expand_call) { 5098 5099 // If expand_call is true then we expand the call_VM_leaf macro 5100 // directly to skip generating the check by 5101 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp. 5102 5103 #ifdef _LP64 5104 assert(thread == r15_thread, "must be"); 5105 #endif // _LP64 5106 5107 Label done; 5108 Label runtime; 5109 5110 assert(pre_val != noreg, "check this code"); 5111 5112 if (obj != noreg) { 5113 assert_different_registers(obj, pre_val, tmp); 5114 assert(pre_val != rax, "check this code"); 5115 } 5116 5117 Address in_progress(thread, in_bytes(JavaThread::satb_mark_queue_offset() + 5118 SATBMarkQueue::byte_offset_of_active())); 5119 Address index(thread, in_bytes(JavaThread::satb_mark_queue_offset() + 5120 SATBMarkQueue::byte_offset_of_index())); 5121 Address buffer(thread, in_bytes(JavaThread::satb_mark_queue_offset() + 5122 SATBMarkQueue::byte_offset_of_buf())); 5123 5124 5125 // Is marking active? 5126 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) { 5127 cmpl(in_progress, 0); 5128 } else { 5129 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption"); 5130 cmpb(in_progress, 0); 5131 } 5132 jcc(Assembler::equal, done); 5133 5134 // Do we need to load the previous value? 5135 if (obj != noreg) { 5136 load_heap_oop(pre_val, Address(obj, 0)); 5137 } 5138 5139 // Is the previous value null? 5140 cmpptr(pre_val, (int32_t) NULL_WORD); 5141 jcc(Assembler::equal, done); 5142 5143 // Can we store original value in the thread's buffer? 5144 // Is index == 0? 5145 // (The index field is typed as size_t.) 5146 5147 movptr(tmp, index); // tmp := *index_adr 5148 cmpptr(tmp, 0); // tmp == 0? 5149 jcc(Assembler::equal, runtime); // If yes, goto runtime 5150 5151 subptr(tmp, wordSize); // tmp := tmp - wordSize 5152 movptr(index, tmp); // *index_adr := tmp 5153 addptr(tmp, buffer); // tmp := tmp + *buffer_adr 5154 5155 // Record the previous value 5156 movptr(Address(tmp, 0), pre_val); 5157 jmp(done); 5158 5159 bind(runtime); 5160 // save the live input values 5161 if(tosca_live) push(rax); 5162 5163 if (obj != noreg && obj != rax) 5164 push(obj); 5165 5166 if (pre_val != rax) 5167 push(pre_val); 5168 5169 // Calling the runtime using the regular call_VM_leaf mechanism generates 5170 // code (generated by InterpreterMacroAssember::call_VM_leaf_base) 5171 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL. 5172 // 5173 // If we care generating the pre-barrier without a frame (e.g. in the 5174 // intrinsified Reference.get() routine) then ebp might be pointing to 5175 // the caller frame and so this check will most likely fail at runtime. 5176 // 5177 // Expanding the call directly bypasses the generation of the check. 5178 // So when we do not have have a full interpreter frame on the stack 5179 // expand_call should be passed true. 5180 5181 NOT_LP64( push(thread); ) 5182 5183 if (expand_call) { 5184 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); ) 5185 pass_arg1(this, thread); 5186 pass_arg0(this, pre_val); 5187 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), 2); 5188 } else { 5189 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), pre_val, thread); 5190 } 5191 5192 NOT_LP64( pop(thread); ) 5193 5194 // save the live input values 5195 if (pre_val != rax) 5196 pop(pre_val); 5197 5198 if (obj != noreg && obj != rax) 5199 pop(obj); 5200 5201 if(tosca_live) pop(rax); 5202 5203 bind(done); 5204 } 5205 5206 void MacroAssembler::g1_write_barrier_post(Register store_addr, 5207 Register new_val, 5208 Register thread, 5209 Register tmp, 5210 Register tmp2) { 5211 #ifdef _LP64 5212 assert(thread == r15_thread, "must be"); 5213 #endif // _LP64 5214 5215 Address queue_index(thread, in_bytes(JavaThread::dirty_card_queue_offset() + 5216 DirtyCardQueue::byte_offset_of_index())); 5217 Address buffer(thread, in_bytes(JavaThread::dirty_card_queue_offset() + 5218 DirtyCardQueue::byte_offset_of_buf())); 5219 5220 CardTableModRefBS* ct = 5221 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 5222 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code"); 5223 5224 Label done; 5225 Label runtime; 5226 5227 // Does store cross heap regions? 5228 5229 movptr(tmp, store_addr); 5230 xorptr(tmp, new_val); 5231 shrptr(tmp, HeapRegion::LogOfHRGrainBytes); 5232 jcc(Assembler::equal, done); 5233 5234 // crosses regions, storing NULL? 5235 5236 cmpptr(new_val, (int32_t) NULL_WORD); 5237 jcc(Assembler::equal, done); 5238 5239 // storing region crossing non-NULL, is card already dirty? 5240 5241 const Register card_addr = tmp; 5242 const Register cardtable = tmp2; 5243 5244 movptr(card_addr, store_addr); 5245 shrptr(card_addr, CardTableModRefBS::card_shift); 5246 // Do not use ExternalAddress to load 'byte_map_base', since 'byte_map_base' is NOT 5247 // a valid address and therefore is not properly handled by the relocation code. 5248 movptr(cardtable, (intptr_t)ct->byte_map_base); 5249 addptr(card_addr, cardtable); 5250 5251 cmpb(Address(card_addr, 0), (int)G1SATBCardTableModRefBS::g1_young_card_val()); 5252 jcc(Assembler::equal, done); 5253 5254 membar(Assembler::Membar_mask_bits(Assembler::StoreLoad)); 5255 cmpb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val()); 5256 jcc(Assembler::equal, done); 5257 5258 5259 // storing a region crossing, non-NULL oop, card is clean. 5260 // dirty card and log. 5261 5262 movb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val()); 5263 5264 cmpl(queue_index, 0); 5265 jcc(Assembler::equal, runtime); 5266 subl(queue_index, wordSize); 5267 movptr(tmp2, buffer); 5268 #ifdef _LP64 5269 movslq(rscratch1, queue_index); 5270 addq(tmp2, rscratch1); 5271 movq(Address(tmp2, 0), card_addr); 5272 #else 5273 addl(tmp2, queue_index); 5274 movl(Address(tmp2, 0), card_addr); 5275 #endif 5276 jmp(done); 5277 5278 bind(runtime); 5279 // save the live input values 5280 push(store_addr); 5281 push(new_val); 5282 #ifdef _LP64 5283 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, r15_thread); 5284 #else 5285 push(thread); 5286 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, thread); 5287 pop(thread); 5288 #endif 5289 pop(new_val); 5290 pop(store_addr); 5291 5292 bind(done); 5293 } 5294 5295 #endif // INCLUDE_ALL_GCS 5296 ////////////////////////////////////////////////////////////////////////////////// 5297 5298 5299 void MacroAssembler::store_check(Register obj, Address dst) { 5300 store_check(obj); 5301 } 5302 5303 void MacroAssembler::store_check(Register obj) { 5304 // Does a store check for the oop in register obj. The content of 5305 // register obj is destroyed afterwards. 5306 BarrierSet* bs = Universe::heap()->barrier_set(); 5307 assert(bs->kind() == BarrierSet::CardTableForRS || 5308 bs->kind() == BarrierSet::CardTableExtension, 5309 "Wrong barrier set kind"); 5310 5311 CardTableModRefBS* ct = barrier_set_cast<CardTableModRefBS>(bs); 5312 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code"); 5313 5314 shrptr(obj, CardTableModRefBS::card_shift); 5315 5316 Address card_addr; 5317 5318 // The calculation for byte_map_base is as follows: 5319 // byte_map_base = _byte_map - (uintptr_t(low_bound) >> card_shift); 5320 // So this essentially converts an address to a displacement and it will 5321 // never need to be relocated. On 64bit however the value may be too 5322 // large for a 32bit displacement. 5323 intptr_t disp = (intptr_t) ct->byte_map_base; 5324 if (is_simm32(disp)) { 5325 card_addr = Address(noreg, obj, Address::times_1, disp); 5326 } else { 5327 // By doing it as an ExternalAddress 'disp' could be converted to a rip-relative 5328 // displacement and done in a single instruction given favorable mapping and a 5329 // smarter version of as_Address. However, 'ExternalAddress' generates a relocation 5330 // entry and that entry is not properly handled by the relocation code. 5331 AddressLiteral cardtable((address)ct->byte_map_base, relocInfo::none); 5332 Address index(noreg, obj, Address::times_1); 5333 card_addr = as_Address(ArrayAddress(cardtable, index)); 5334 } 5335 5336 int dirty = CardTableModRefBS::dirty_card_val(); 5337 if (UseCondCardMark) { 5338 Label L_already_dirty; 5339 if (UseConcMarkSweepGC) { 5340 membar(Assembler::StoreLoad); 5341 } 5342 cmpb(card_addr, dirty); 5343 jcc(Assembler::equal, L_already_dirty); 5344 movb(card_addr, dirty); 5345 bind(L_already_dirty); 5346 } else { 5347 movb(card_addr, dirty); 5348 } 5349 } 5350 5351 void MacroAssembler::subptr(Register dst, int32_t imm32) { 5352 LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32)); 5353 } 5354 5355 // Force generation of a 4 byte immediate value even if it fits into 8bit 5356 void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) { 5357 LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32)); 5358 } 5359 5360 void MacroAssembler::subptr(Register dst, Register src) { 5361 LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src)); 5362 } 5363 5364 // C++ bool manipulation 5365 void MacroAssembler::testbool(Register dst) { 5366 if(sizeof(bool) == 1) 5367 testb(dst, 0xff); 5368 else if(sizeof(bool) == 2) { 5369 // testw implementation needed for two byte bools 5370 ShouldNotReachHere(); 5371 } else if(sizeof(bool) == 4) 5372 testl(dst, dst); 5373 else 5374 // unsupported 5375 ShouldNotReachHere(); 5376 } 5377 5378 void MacroAssembler::testptr(Register dst, Register src) { 5379 LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src)); 5380 } 5381 5382 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes. 5383 void MacroAssembler::tlab_allocate(Register obj, 5384 Register var_size_in_bytes, 5385 int con_size_in_bytes, 5386 Register t1, 5387 Register t2, 5388 Label& slow_case) { 5389 assert_different_registers(obj, t1, t2); 5390 assert_different_registers(obj, var_size_in_bytes, t1); 5391 Register end = t2; 5392 Register thread = NOT_LP64(t1) LP64_ONLY(r15_thread); 5393 5394 verify_tlab(); 5395 5396 NOT_LP64(get_thread(thread)); 5397 5398 movptr(obj, Address(thread, JavaThread::tlab_top_offset())); 5399 if (var_size_in_bytes == noreg) { 5400 lea(end, Address(obj, con_size_in_bytes)); 5401 } else { 5402 lea(end, Address(obj, var_size_in_bytes, Address::times_1)); 5403 } 5404 cmpptr(end, Address(thread, JavaThread::tlab_end_offset())); 5405 jcc(Assembler::above, slow_case); 5406 5407 // update the tlab top pointer 5408 movptr(Address(thread, JavaThread::tlab_top_offset()), end); 5409 5410 // recover var_size_in_bytes if necessary 5411 if (var_size_in_bytes == end) { 5412 subptr(var_size_in_bytes, obj); 5413 } 5414 verify_tlab(); 5415 } 5416 5417 // Preserves rbx, and rdx. 5418 Register MacroAssembler::tlab_refill(Label& retry, 5419 Label& try_eden, 5420 Label& slow_case) { 5421 Register top = rax; 5422 Register t1 = rcx; 5423 Register t2 = rsi; 5424 Register thread_reg = NOT_LP64(rdi) LP64_ONLY(r15_thread); 5425 assert_different_registers(top, thread_reg, t1, t2, /* preserve: */ rbx, rdx); 5426 Label do_refill, discard_tlab; 5427 5428 if (!Universe::heap()->supports_inline_contig_alloc()) { 5429 // No allocation in the shared eden. 5430 jmp(slow_case); 5431 } 5432 5433 NOT_LP64(get_thread(thread_reg)); 5434 5435 movptr(top, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 5436 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset()))); 5437 5438 // calculate amount of free space 5439 subptr(t1, top); 5440 shrptr(t1, LogHeapWordSize); 5441 5442 // Retain tlab and allocate object in shared space if 5443 // the amount free in the tlab is too large to discard. 5444 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset()))); 5445 jcc(Assembler::lessEqual, discard_tlab); 5446 5447 // Retain 5448 // %%% yuck as movptr... 5449 movptr(t2, (int32_t) ThreadLocalAllocBuffer::refill_waste_limit_increment()); 5450 addptr(Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset())), t2); 5451 if (TLABStats) { 5452 // increment number of slow_allocations 5453 addl(Address(thread_reg, in_bytes(JavaThread::tlab_slow_allocations_offset())), 1); 5454 } 5455 jmp(try_eden); 5456 5457 bind(discard_tlab); 5458 if (TLABStats) { 5459 // increment number of refills 5460 addl(Address(thread_reg, in_bytes(JavaThread::tlab_number_of_refills_offset())), 1); 5461 // accumulate wastage -- t1 is amount free in tlab 5462 addl(Address(thread_reg, in_bytes(JavaThread::tlab_fast_refill_waste_offset())), t1); 5463 } 5464 5465 // if tlab is currently allocated (top or end != null) then 5466 // fill [top, end + alignment_reserve) with array object 5467 testptr(top, top); 5468 jcc(Assembler::zero, do_refill); 5469 5470 // set up the mark word 5471 movptr(Address(top, oopDesc::mark_offset_in_bytes()), (intptr_t)markOopDesc::prototype()->copy_set_hash(0x2)); 5472 // set the length to the remaining space 5473 subptr(t1, typeArrayOopDesc::header_size(T_INT)); 5474 addptr(t1, (int32_t)ThreadLocalAllocBuffer::alignment_reserve()); 5475 shlptr(t1, log2_intptr(HeapWordSize/sizeof(jint))); 5476 movl(Address(top, arrayOopDesc::length_offset_in_bytes()), t1); 5477 // set klass to intArrayKlass 5478 // dubious reloc why not an oop reloc? 5479 movptr(t1, ExternalAddress((address)Universe::intArrayKlassObj_addr())); 5480 // store klass last. concurrent gcs assumes klass length is valid if 5481 // klass field is not null. 5482 store_klass(top, t1); 5483 5484 movptr(t1, top); 5485 subptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset()))); 5486 incr_allocated_bytes(thread_reg, t1, 0); 5487 5488 // refill the tlab with an eden allocation 5489 bind(do_refill); 5490 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset()))); 5491 shlptr(t1, LogHeapWordSize); 5492 // allocate new tlab, address returned in top 5493 eden_allocate(top, t1, 0, t2, slow_case); 5494 5495 // Check that t1 was preserved in eden_allocate. 5496 #ifdef ASSERT 5497 if (UseTLAB) { 5498 Label ok; 5499 Register tsize = rsi; 5500 assert_different_registers(tsize, thread_reg, t1); 5501 push(tsize); 5502 movptr(tsize, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset()))); 5503 shlptr(tsize, LogHeapWordSize); 5504 cmpptr(t1, tsize); 5505 jcc(Assembler::equal, ok); 5506 STOP("assert(t1 != tlab size)"); 5507 should_not_reach_here(); 5508 5509 bind(ok); 5510 pop(tsize); 5511 } 5512 #endif 5513 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())), top); 5514 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())), top); 5515 addptr(top, t1); 5516 subptr(top, (int32_t)ThreadLocalAllocBuffer::alignment_reserve_in_bytes()); 5517 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())), top); 5518 verify_tlab(); 5519 jmp(retry); 5520 5521 return thread_reg; // for use by caller 5522 } 5523 5524 void MacroAssembler::incr_allocated_bytes(Register thread, 5525 Register var_size_in_bytes, 5526 int con_size_in_bytes, 5527 Register t1) { 5528 if (!thread->is_valid()) { 5529 #ifdef _LP64 5530 thread = r15_thread; 5531 #else 5532 assert(t1->is_valid(), "need temp reg"); 5533 thread = t1; 5534 get_thread(thread); 5535 #endif 5536 } 5537 5538 #ifdef _LP64 5539 if (var_size_in_bytes->is_valid()) { 5540 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes); 5541 } else { 5542 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes); 5543 } 5544 #else 5545 if (var_size_in_bytes->is_valid()) { 5546 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes); 5547 } else { 5548 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes); 5549 } 5550 adcl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())+4), 0); 5551 #endif 5552 } 5553 5554 void MacroAssembler::fp_runtime_fallback(address runtime_entry, int nb_args, int num_fpu_regs_in_use) { 5555 pusha(); 5556 5557 // if we are coming from c1, xmm registers may be live 5558 int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8); 5559 if (UseAVX > 2) { 5560 num_xmm_regs = LP64_ONLY(32) NOT_LP64(8); 5561 } 5562 5563 if (UseSSE == 1) { 5564 subptr(rsp, sizeof(jdouble)*8); 5565 for (int n = 0; n < 8; n++) { 5566 movflt(Address(rsp, n*sizeof(jdouble)), as_XMMRegister(n)); 5567 } 5568 } else if (UseSSE >= 2) { 5569 if (UseAVX > 2) { 5570 push(rbx); 5571 movl(rbx, 0xffff); 5572 kmovwl(k1, rbx); 5573 pop(rbx); 5574 } 5575 #ifdef COMPILER2 5576 if (MaxVectorSize > 16) { 5577 if(UseAVX > 2) { 5578 // Save upper half of ZMM registers 5579 subptr(rsp, 32*num_xmm_regs); 5580 for (int n = 0; n < num_xmm_regs; n++) { 5581 vextractf64x4h(Address(rsp, n*32), as_XMMRegister(n), 1); 5582 } 5583 } 5584 assert(UseAVX > 0, "256 bit vectors are supported only with AVX"); 5585 // Save upper half of YMM registers 5586 subptr(rsp, 16*num_xmm_regs); 5587 for (int n = 0; n < num_xmm_regs; n++) { 5588 vextractf128h(Address(rsp, n*16), as_XMMRegister(n)); 5589 } 5590 } 5591 #endif 5592 // Save whole 128bit (16 bytes) XMM registers 5593 subptr(rsp, 16*num_xmm_regs); 5594 #ifdef _LP64 5595 if (VM_Version::supports_evex()) { 5596 for (int n = 0; n < num_xmm_regs; n++) { 5597 vextractf32x4h(Address(rsp, n*16), as_XMMRegister(n), 0); 5598 } 5599 } else { 5600 for (int n = 0; n < num_xmm_regs; n++) { 5601 movdqu(Address(rsp, n*16), as_XMMRegister(n)); 5602 } 5603 } 5604 #else 5605 for (int n = 0; n < num_xmm_regs; n++) { 5606 movdqu(Address(rsp, n*16), as_XMMRegister(n)); 5607 } 5608 #endif 5609 } 5610 5611 // Preserve registers across runtime call 5612 int incoming_argument_and_return_value_offset = -1; 5613 if (num_fpu_regs_in_use > 1) { 5614 // Must preserve all other FPU regs (could alternatively convert 5615 // SharedRuntime::dsin, dcos etc. into assembly routines known not to trash 5616 // FPU state, but can not trust C compiler) 5617 NEEDS_CLEANUP; 5618 // NOTE that in this case we also push the incoming argument(s) to 5619 // the stack and restore it later; we also use this stack slot to 5620 // hold the return value from dsin, dcos etc. 5621 for (int i = 0; i < num_fpu_regs_in_use; i++) { 5622 subptr(rsp, sizeof(jdouble)); 5623 fstp_d(Address(rsp, 0)); 5624 } 5625 incoming_argument_and_return_value_offset = sizeof(jdouble)*(num_fpu_regs_in_use-1); 5626 for (int i = nb_args-1; i >= 0; i--) { 5627 fld_d(Address(rsp, incoming_argument_and_return_value_offset-i*sizeof(jdouble))); 5628 } 5629 } 5630 5631 subptr(rsp, nb_args*sizeof(jdouble)); 5632 for (int i = 0; i < nb_args; i++) { 5633 fstp_d(Address(rsp, i*sizeof(jdouble))); 5634 } 5635 5636 #ifdef _LP64 5637 if (nb_args > 0) { 5638 movdbl(xmm0, Address(rsp, 0)); 5639 } 5640 if (nb_args > 1) { 5641 movdbl(xmm1, Address(rsp, sizeof(jdouble))); 5642 } 5643 assert(nb_args <= 2, "unsupported number of args"); 5644 #endif // _LP64 5645 5646 // NOTE: we must not use call_VM_leaf here because that requires a 5647 // complete interpreter frame in debug mode -- same bug as 4387334 5648 // MacroAssembler::call_VM_leaf_base is perfectly safe and will 5649 // do proper 64bit abi 5650 5651 NEEDS_CLEANUP; 5652 // Need to add stack banging before this runtime call if it needs to 5653 // be taken; however, there is no generic stack banging routine at 5654 // the MacroAssembler level 5655 5656 MacroAssembler::call_VM_leaf_base(runtime_entry, 0); 5657 5658 #ifdef _LP64 5659 movsd(Address(rsp, 0), xmm0); 5660 fld_d(Address(rsp, 0)); 5661 #endif // _LP64 5662 addptr(rsp, sizeof(jdouble)*nb_args); 5663 if (num_fpu_regs_in_use > 1) { 5664 // Must save return value to stack and then restore entire FPU 5665 // stack except incoming arguments 5666 fstp_d(Address(rsp, incoming_argument_and_return_value_offset)); 5667 for (int i = 0; i < num_fpu_regs_in_use - nb_args; i++) { 5668 fld_d(Address(rsp, 0)); 5669 addptr(rsp, sizeof(jdouble)); 5670 } 5671 fld_d(Address(rsp, (nb_args-1)*sizeof(jdouble))); 5672 addptr(rsp, sizeof(jdouble)*nb_args); 5673 } 5674 5675 if (UseSSE == 1) { 5676 for (int n = 0; n < 8; n++) { 5677 movflt(as_XMMRegister(n), Address(rsp, n*sizeof(jdouble))); 5678 } 5679 addptr(rsp, sizeof(jdouble)*8); 5680 } else if (UseSSE >= 2) { 5681 // Restore whole 128bit (16 bytes) XMM registers 5682 #ifdef _LP64 5683 if (VM_Version::supports_evex()) { 5684 for (int n = 0; n < num_xmm_regs; n++) { 5685 vinsertf32x4h(as_XMMRegister(n), Address(rsp, n*16), 0); 5686 } 5687 } else { 5688 for (int n = 0; n < num_xmm_regs; n++) { 5689 movdqu(as_XMMRegister(n), Address(rsp, n*16)); 5690 } 5691 } 5692 #else 5693 for (int n = 0; n < num_xmm_regs; n++) { 5694 movdqu(as_XMMRegister(n), Address(rsp, n*16)); 5695 } 5696 #endif 5697 addptr(rsp, 16*num_xmm_regs); 5698 5699 #ifdef COMPILER2 5700 if (MaxVectorSize > 16) { 5701 // Restore upper half of YMM registers. 5702 for (int n = 0; n < num_xmm_regs; n++) { 5703 vinsertf128h(as_XMMRegister(n), Address(rsp, n*16)); 5704 } 5705 addptr(rsp, 16*num_xmm_regs); 5706 if(UseAVX > 2) { 5707 for (int n = 0; n < num_xmm_regs; n++) { 5708 vinsertf64x4h(as_XMMRegister(n), Address(rsp, n*32), 1); 5709 } 5710 addptr(rsp, 32*num_xmm_regs); 5711 } 5712 } 5713 #endif 5714 } 5715 popa(); 5716 } 5717 5718 static const double pi_4 = 0.7853981633974483; 5719 5720 void MacroAssembler::trigfunc(char trig, int num_fpu_regs_in_use) { 5721 // A hand-coded argument reduction for values in fabs(pi/4, pi/2) 5722 // was attempted in this code; unfortunately it appears that the 5723 // switch to 80-bit precision and back causes this to be 5724 // unprofitable compared with simply performing a runtime call if 5725 // the argument is out of the (-pi/4, pi/4) range. 5726 5727 Register tmp = noreg; 5728 if (!VM_Version::supports_cmov()) { 5729 // fcmp needs a temporary so preserve rbx, 5730 tmp = rbx; 5731 push(tmp); 5732 } 5733 5734 Label slow_case, done; 5735 5736 ExternalAddress pi4_adr = (address)&pi_4; 5737 if (reachable(pi4_adr)) { 5738 // x ?<= pi/4 5739 fld_d(pi4_adr); 5740 fld_s(1); // Stack: X PI/4 X 5741 fabs(); // Stack: |X| PI/4 X 5742 fcmp(tmp); 5743 jcc(Assembler::above, slow_case); 5744 5745 // fastest case: -pi/4 <= x <= pi/4 5746 switch(trig) { 5747 case 's': 5748 fsin(); 5749 break; 5750 case 'c': 5751 fcos(); 5752 break; 5753 case 't': 5754 ftan(); 5755 break; 5756 default: 5757 assert(false, "bad intrinsic"); 5758 break; 5759 } 5760 jmp(done); 5761 } 5762 5763 // slow case: runtime call 5764 bind(slow_case); 5765 5766 switch(trig) { 5767 case 's': 5768 { 5769 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1, num_fpu_regs_in_use); 5770 } 5771 break; 5772 case 'c': 5773 { 5774 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1, num_fpu_regs_in_use); 5775 } 5776 break; 5777 case 't': 5778 { 5779 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1, num_fpu_regs_in_use); 5780 } 5781 break; 5782 default: 5783 assert(false, "bad intrinsic"); 5784 break; 5785 } 5786 5787 // Come here with result in F-TOS 5788 bind(done); 5789 5790 if (tmp != noreg) { 5791 pop(tmp); 5792 } 5793 } 5794 5795 5796 // Look up the method for a megamorphic invokeinterface call. 5797 // The target method is determined by <intf_klass, itable_index>. 5798 // The receiver klass is in recv_klass. 5799 // On success, the result will be in method_result, and execution falls through. 5800 // On failure, execution transfers to the given label. 5801 void MacroAssembler::lookup_interface_method(Register recv_klass, 5802 Register intf_klass, 5803 RegisterOrConstant itable_index, 5804 Register method_result, 5805 Register scan_temp, 5806 Label& L_no_such_interface) { 5807 assert_different_registers(recv_klass, intf_klass, method_result, scan_temp); 5808 assert(itable_index.is_constant() || itable_index.as_register() == method_result, 5809 "caller must use same register for non-constant itable index as for method"); 5810 5811 // Compute start of first itableOffsetEntry (which is at the end of the vtable) 5812 int vtable_base = InstanceKlass::vtable_start_offset() * wordSize; 5813 int itentry_off = itableMethodEntry::method_offset_in_bytes(); 5814 int scan_step = itableOffsetEntry::size() * wordSize; 5815 int vte_size = vtableEntry::size() * wordSize; 5816 Address::ScaleFactor times_vte_scale = Address::times_ptr; 5817 assert(vte_size == wordSize, "else adjust times_vte_scale"); 5818 5819 movl(scan_temp, Address(recv_klass, InstanceKlass::vtable_length_offset() * wordSize)); 5820 5821 // %%% Could store the aligned, prescaled offset in the klassoop. 5822 lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base)); 5823 if (HeapWordsPerLong > 1) { 5824 // Round up to align_object_offset boundary 5825 // see code for InstanceKlass::start_of_itable! 5826 round_to(scan_temp, BytesPerLong); 5827 } 5828 5829 // Adjust recv_klass by scaled itable_index, so we can free itable_index. 5830 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below"); 5831 lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off)); 5832 5833 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) { 5834 // if (scan->interface() == intf) { 5835 // result = (klass + scan->offset() + itable_index); 5836 // } 5837 // } 5838 Label search, found_method; 5839 5840 for (int peel = 1; peel >= 0; peel--) { 5841 movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes())); 5842 cmpptr(intf_klass, method_result); 5843 5844 if (peel) { 5845 jccb(Assembler::equal, found_method); 5846 } else { 5847 jccb(Assembler::notEqual, search); 5848 // (invert the test to fall through to found_method...) 5849 } 5850 5851 if (!peel) break; 5852 5853 bind(search); 5854 5855 // Check that the previous entry is non-null. A null entry means that 5856 // the receiver class doesn't implement the interface, and wasn't the 5857 // same as when the caller was compiled. 5858 testptr(method_result, method_result); 5859 jcc(Assembler::zero, L_no_such_interface); 5860 addptr(scan_temp, scan_step); 5861 } 5862 5863 bind(found_method); 5864 5865 // Got a hit. 5866 movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes())); 5867 movptr(method_result, Address(recv_klass, scan_temp, Address::times_1)); 5868 } 5869 5870 5871 // virtual method calling 5872 void MacroAssembler::lookup_virtual_method(Register recv_klass, 5873 RegisterOrConstant vtable_index, 5874 Register method_result) { 5875 const int base = InstanceKlass::vtable_start_offset() * wordSize; 5876 assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below"); 5877 Address vtable_entry_addr(recv_klass, 5878 vtable_index, Address::times_ptr, 5879 base + vtableEntry::method_offset_in_bytes()); 5880 movptr(method_result, vtable_entry_addr); 5881 } 5882 5883 5884 void MacroAssembler::check_klass_subtype(Register sub_klass, 5885 Register super_klass, 5886 Register temp_reg, 5887 Label& L_success) { 5888 Label L_failure; 5889 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL); 5890 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL); 5891 bind(L_failure); 5892 } 5893 5894 5895 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass, 5896 Register super_klass, 5897 Register temp_reg, 5898 Label* L_success, 5899 Label* L_failure, 5900 Label* L_slow_path, 5901 RegisterOrConstant super_check_offset) { 5902 assert_different_registers(sub_klass, super_klass, temp_reg); 5903 bool must_load_sco = (super_check_offset.constant_or_zero() == -1); 5904 if (super_check_offset.is_register()) { 5905 assert_different_registers(sub_klass, super_klass, 5906 super_check_offset.as_register()); 5907 } else if (must_load_sco) { 5908 assert(temp_reg != noreg, "supply either a temp or a register offset"); 5909 } 5910 5911 Label L_fallthrough; 5912 int label_nulls = 0; 5913 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 5914 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 5915 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; } 5916 assert(label_nulls <= 1, "at most one NULL in the batch"); 5917 5918 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 5919 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 5920 Address super_check_offset_addr(super_klass, sco_offset); 5921 5922 // Hacked jcc, which "knows" that L_fallthrough, at least, is in 5923 // range of a jccb. If this routine grows larger, reconsider at 5924 // least some of these. 5925 #define local_jcc(assembler_cond, label) \ 5926 if (&(label) == &L_fallthrough) jccb(assembler_cond, label); \ 5927 else jcc( assembler_cond, label) /*omit semi*/ 5928 5929 // Hacked jmp, which may only be used just before L_fallthrough. 5930 #define final_jmp(label) \ 5931 if (&(label) == &L_fallthrough) { /*do nothing*/ } \ 5932 else jmp(label) /*omit semi*/ 5933 5934 // If the pointers are equal, we are done (e.g., String[] elements). 5935 // This self-check enables sharing of secondary supertype arrays among 5936 // non-primary types such as array-of-interface. Otherwise, each such 5937 // type would need its own customized SSA. 5938 // We move this check to the front of the fast path because many 5939 // type checks are in fact trivially successful in this manner, 5940 // so we get a nicely predicted branch right at the start of the check. 5941 cmpptr(sub_klass, super_klass); 5942 local_jcc(Assembler::equal, *L_success); 5943 5944 // Check the supertype display: 5945 if (must_load_sco) { 5946 // Positive movl does right thing on LP64. 5947 movl(temp_reg, super_check_offset_addr); 5948 super_check_offset = RegisterOrConstant(temp_reg); 5949 } 5950 Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0); 5951 cmpptr(super_klass, super_check_addr); // load displayed supertype 5952 5953 // This check has worked decisively for primary supers. 5954 // Secondary supers are sought in the super_cache ('super_cache_addr'). 5955 // (Secondary supers are interfaces and very deeply nested subtypes.) 5956 // This works in the same check above because of a tricky aliasing 5957 // between the super_cache and the primary super display elements. 5958 // (The 'super_check_addr' can address either, as the case requires.) 5959 // Note that the cache is updated below if it does not help us find 5960 // what we need immediately. 5961 // So if it was a primary super, we can just fail immediately. 5962 // Otherwise, it's the slow path for us (no success at this point). 5963 5964 if (super_check_offset.is_register()) { 5965 local_jcc(Assembler::equal, *L_success); 5966 cmpl(super_check_offset.as_register(), sc_offset); 5967 if (L_failure == &L_fallthrough) { 5968 local_jcc(Assembler::equal, *L_slow_path); 5969 } else { 5970 local_jcc(Assembler::notEqual, *L_failure); 5971 final_jmp(*L_slow_path); 5972 } 5973 } else if (super_check_offset.as_constant() == sc_offset) { 5974 // Need a slow path; fast failure is impossible. 5975 if (L_slow_path == &L_fallthrough) { 5976 local_jcc(Assembler::equal, *L_success); 5977 } else { 5978 local_jcc(Assembler::notEqual, *L_slow_path); 5979 final_jmp(*L_success); 5980 } 5981 } else { 5982 // No slow path; it's a fast decision. 5983 if (L_failure == &L_fallthrough) { 5984 local_jcc(Assembler::equal, *L_success); 5985 } else { 5986 local_jcc(Assembler::notEqual, *L_failure); 5987 final_jmp(*L_success); 5988 } 5989 } 5990 5991 bind(L_fallthrough); 5992 5993 #undef local_jcc 5994 #undef final_jmp 5995 } 5996 5997 5998 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass, 5999 Register super_klass, 6000 Register temp_reg, 6001 Register temp2_reg, 6002 Label* L_success, 6003 Label* L_failure, 6004 bool set_cond_codes) { 6005 assert_different_registers(sub_klass, super_klass, temp_reg); 6006 if (temp2_reg != noreg) 6007 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg); 6008 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg) 6009 6010 Label L_fallthrough; 6011 int label_nulls = 0; 6012 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 6013 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 6014 assert(label_nulls <= 1, "at most one NULL in the batch"); 6015 6016 // a couple of useful fields in sub_klass: 6017 int ss_offset = in_bytes(Klass::secondary_supers_offset()); 6018 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 6019 Address secondary_supers_addr(sub_klass, ss_offset); 6020 Address super_cache_addr( sub_klass, sc_offset); 6021 6022 // Do a linear scan of the secondary super-klass chain. 6023 // This code is rarely used, so simplicity is a virtue here. 6024 // The repne_scan instruction uses fixed registers, which we must spill. 6025 // Don't worry too much about pre-existing connections with the input regs. 6026 6027 assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super) 6028 assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter) 6029 6030 // Get super_klass value into rax (even if it was in rdi or rcx). 6031 bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false; 6032 if (super_klass != rax || UseCompressedOops) { 6033 if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; } 6034 mov(rax, super_klass); 6035 } 6036 if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; } 6037 if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; } 6038 6039 #ifndef PRODUCT 6040 int* pst_counter = &SharedRuntime::_partial_subtype_ctr; 6041 ExternalAddress pst_counter_addr((address) pst_counter); 6042 NOT_LP64( incrementl(pst_counter_addr) ); 6043 LP64_ONLY( lea(rcx, pst_counter_addr) ); 6044 LP64_ONLY( incrementl(Address(rcx, 0)) ); 6045 #endif //PRODUCT 6046 6047 // We will consult the secondary-super array. 6048 movptr(rdi, secondary_supers_addr); 6049 // Load the array length. (Positive movl does right thing on LP64.) 6050 movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes())); 6051 // Skip to start of data. 6052 addptr(rdi, Array<Klass*>::base_offset_in_bytes()); 6053 6054 // Scan RCX words at [RDI] for an occurrence of RAX. 6055 // Set NZ/Z based on last compare. 6056 // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does 6057 // not change flags (only scas instruction which is repeated sets flags). 6058 // Set Z = 0 (not equal) before 'repne' to indicate that class was not found. 6059 6060 testptr(rax,rax); // Set Z = 0 6061 repne_scan(); 6062 6063 // Unspill the temp. registers: 6064 if (pushed_rdi) pop(rdi); 6065 if (pushed_rcx) pop(rcx); 6066 if (pushed_rax) pop(rax); 6067 6068 if (set_cond_codes) { 6069 // Special hack for the AD files: rdi is guaranteed non-zero. 6070 assert(!pushed_rdi, "rdi must be left non-NULL"); 6071 // Also, the condition codes are properly set Z/NZ on succeed/failure. 6072 } 6073 6074 if (L_failure == &L_fallthrough) 6075 jccb(Assembler::notEqual, *L_failure); 6076 else jcc(Assembler::notEqual, *L_failure); 6077 6078 // Success. Cache the super we found and proceed in triumph. 6079 movptr(super_cache_addr, super_klass); 6080 6081 if (L_success != &L_fallthrough) { 6082 jmp(*L_success); 6083 } 6084 6085 #undef IS_A_TEMP 6086 6087 bind(L_fallthrough); 6088 } 6089 6090 6091 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) { 6092 if (VM_Version::supports_cmov()) { 6093 cmovl(cc, dst, src); 6094 } else { 6095 Label L; 6096 jccb(negate_condition(cc), L); 6097 movl(dst, src); 6098 bind(L); 6099 } 6100 } 6101 6102 void MacroAssembler::cmov32(Condition cc, Register dst, Register src) { 6103 if (VM_Version::supports_cmov()) { 6104 cmovl(cc, dst, src); 6105 } else { 6106 Label L; 6107 jccb(negate_condition(cc), L); 6108 movl(dst, src); 6109 bind(L); 6110 } 6111 } 6112 6113 void MacroAssembler::verify_oop(Register reg, const char* s) { 6114 if (!VerifyOops) return; 6115 6116 // Pass register number to verify_oop_subroutine 6117 const char* b = NULL; 6118 { 6119 ResourceMark rm; 6120 stringStream ss; 6121 ss.print("verify_oop: %s: %s", reg->name(), s); 6122 b = code_string(ss.as_string()); 6123 } 6124 BLOCK_COMMENT("verify_oop {"); 6125 #ifdef _LP64 6126 push(rscratch1); // save r10, trashed by movptr() 6127 #endif 6128 push(rax); // save rax, 6129 push(reg); // pass register argument 6130 ExternalAddress buffer((address) b); 6131 // avoid using pushptr, as it modifies scratch registers 6132 // and our contract is not to modify anything 6133 movptr(rax, buffer.addr()); 6134 push(rax); 6135 // call indirectly to solve generation ordering problem 6136 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); 6137 call(rax); 6138 // Caller pops the arguments (oop, message) and restores rax, r10 6139 BLOCK_COMMENT("} verify_oop"); 6140 } 6141 6142 6143 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr, 6144 Register tmp, 6145 int offset) { 6146 intptr_t value = *delayed_value_addr; 6147 if (value != 0) 6148 return RegisterOrConstant(value + offset); 6149 6150 // load indirectly to solve generation ordering problem 6151 movptr(tmp, ExternalAddress((address) delayed_value_addr)); 6152 6153 #ifdef ASSERT 6154 { Label L; 6155 testptr(tmp, tmp); 6156 if (WizardMode) { 6157 const char* buf = NULL; 6158 { 6159 ResourceMark rm; 6160 stringStream ss; 6161 ss.print("DelayedValue=" INTPTR_FORMAT, delayed_value_addr[1]); 6162 buf = code_string(ss.as_string()); 6163 } 6164 jcc(Assembler::notZero, L); 6165 STOP(buf); 6166 } else { 6167 jccb(Assembler::notZero, L); 6168 hlt(); 6169 } 6170 bind(L); 6171 } 6172 #endif 6173 6174 if (offset != 0) 6175 addptr(tmp, offset); 6176 6177 return RegisterOrConstant(tmp); 6178 } 6179 6180 6181 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot, 6182 int extra_slot_offset) { 6183 // cf. TemplateTable::prepare_invoke(), if (load_receiver). 6184 int stackElementSize = Interpreter::stackElementSize; 6185 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0); 6186 #ifdef ASSERT 6187 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1); 6188 assert(offset1 - offset == stackElementSize, "correct arithmetic"); 6189 #endif 6190 Register scale_reg = noreg; 6191 Address::ScaleFactor scale_factor = Address::no_scale; 6192 if (arg_slot.is_constant()) { 6193 offset += arg_slot.as_constant() * stackElementSize; 6194 } else { 6195 scale_reg = arg_slot.as_register(); 6196 scale_factor = Address::times(stackElementSize); 6197 } 6198 offset += wordSize; // return PC is on stack 6199 return Address(rsp, scale_reg, scale_factor, offset); 6200 } 6201 6202 6203 void MacroAssembler::verify_oop_addr(Address addr, const char* s) { 6204 if (!VerifyOops) return; 6205 6206 // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord); 6207 // Pass register number to verify_oop_subroutine 6208 const char* b = NULL; 6209 { 6210 ResourceMark rm; 6211 stringStream ss; 6212 ss.print("verify_oop_addr: %s", s); 6213 b = code_string(ss.as_string()); 6214 } 6215 #ifdef _LP64 6216 push(rscratch1); // save r10, trashed by movptr() 6217 #endif 6218 push(rax); // save rax, 6219 // addr may contain rsp so we will have to adjust it based on the push 6220 // we just did (and on 64 bit we do two pushes) 6221 // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which 6222 // stores rax into addr which is backwards of what was intended. 6223 if (addr.uses(rsp)) { 6224 lea(rax, addr); 6225 pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord)); 6226 } else { 6227 pushptr(addr); 6228 } 6229 6230 ExternalAddress buffer((address) b); 6231 // pass msg argument 6232 // avoid using pushptr, as it modifies scratch registers 6233 // and our contract is not to modify anything 6234 movptr(rax, buffer.addr()); 6235 push(rax); 6236 6237 // call indirectly to solve generation ordering problem 6238 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); 6239 call(rax); 6240 // Caller pops the arguments (addr, message) and restores rax, r10. 6241 } 6242 6243 void MacroAssembler::verify_tlab() { 6244 #ifdef ASSERT 6245 if (UseTLAB && VerifyOops) { 6246 Label next, ok; 6247 Register t1 = rsi; 6248 Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread); 6249 6250 push(t1); 6251 NOT_LP64(push(thread_reg)); 6252 NOT_LP64(get_thread(thread_reg)); 6253 6254 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 6255 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset()))); 6256 jcc(Assembler::aboveEqual, next); 6257 STOP("assert(top >= start)"); 6258 should_not_reach_here(); 6259 6260 bind(next); 6261 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset()))); 6262 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 6263 jcc(Assembler::aboveEqual, ok); 6264 STOP("assert(top <= end)"); 6265 should_not_reach_here(); 6266 6267 bind(ok); 6268 NOT_LP64(pop(thread_reg)); 6269 pop(t1); 6270 } 6271 #endif 6272 } 6273 6274 class ControlWord { 6275 public: 6276 int32_t _value; 6277 6278 int rounding_control() const { return (_value >> 10) & 3 ; } 6279 int precision_control() const { return (_value >> 8) & 3 ; } 6280 bool precision() const { return ((_value >> 5) & 1) != 0; } 6281 bool underflow() const { return ((_value >> 4) & 1) != 0; } 6282 bool overflow() const { return ((_value >> 3) & 1) != 0; } 6283 bool zero_divide() const { return ((_value >> 2) & 1) != 0; } 6284 bool denormalized() const { return ((_value >> 1) & 1) != 0; } 6285 bool invalid() const { return ((_value >> 0) & 1) != 0; } 6286 6287 void print() const { 6288 // rounding control 6289 const char* rc; 6290 switch (rounding_control()) { 6291 case 0: rc = "round near"; break; 6292 case 1: rc = "round down"; break; 6293 case 2: rc = "round up "; break; 6294 case 3: rc = "chop "; break; 6295 }; 6296 // precision control 6297 const char* pc; 6298 switch (precision_control()) { 6299 case 0: pc = "24 bits "; break; 6300 case 1: pc = "reserved"; break; 6301 case 2: pc = "53 bits "; break; 6302 case 3: pc = "64 bits "; break; 6303 }; 6304 // flags 6305 char f[9]; 6306 f[0] = ' '; 6307 f[1] = ' '; 6308 f[2] = (precision ()) ? 'P' : 'p'; 6309 f[3] = (underflow ()) ? 'U' : 'u'; 6310 f[4] = (overflow ()) ? 'O' : 'o'; 6311 f[5] = (zero_divide ()) ? 'Z' : 'z'; 6312 f[6] = (denormalized()) ? 'D' : 'd'; 6313 f[7] = (invalid ()) ? 'I' : 'i'; 6314 f[8] = '\x0'; 6315 // output 6316 printf("%04x masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc); 6317 } 6318 6319 }; 6320 6321 class StatusWord { 6322 public: 6323 int32_t _value; 6324 6325 bool busy() const { return ((_value >> 15) & 1) != 0; } 6326 bool C3() const { return ((_value >> 14) & 1) != 0; } 6327 bool C2() const { return ((_value >> 10) & 1) != 0; } 6328 bool C1() const { return ((_value >> 9) & 1) != 0; } 6329 bool C0() const { return ((_value >> 8) & 1) != 0; } 6330 int top() const { return (_value >> 11) & 7 ; } 6331 bool error_status() const { return ((_value >> 7) & 1) != 0; } 6332 bool stack_fault() const { return ((_value >> 6) & 1) != 0; } 6333 bool precision() const { return ((_value >> 5) & 1) != 0; } 6334 bool underflow() const { return ((_value >> 4) & 1) != 0; } 6335 bool overflow() const { return ((_value >> 3) & 1) != 0; } 6336 bool zero_divide() const { return ((_value >> 2) & 1) != 0; } 6337 bool denormalized() const { return ((_value >> 1) & 1) != 0; } 6338 bool invalid() const { return ((_value >> 0) & 1) != 0; } 6339 6340 void print() const { 6341 // condition codes 6342 char c[5]; 6343 c[0] = (C3()) ? '3' : '-'; 6344 c[1] = (C2()) ? '2' : '-'; 6345 c[2] = (C1()) ? '1' : '-'; 6346 c[3] = (C0()) ? '0' : '-'; 6347 c[4] = '\x0'; 6348 // flags 6349 char f[9]; 6350 f[0] = (error_status()) ? 'E' : '-'; 6351 f[1] = (stack_fault ()) ? 'S' : '-'; 6352 f[2] = (precision ()) ? 'P' : '-'; 6353 f[3] = (underflow ()) ? 'U' : '-'; 6354 f[4] = (overflow ()) ? 'O' : '-'; 6355 f[5] = (zero_divide ()) ? 'Z' : '-'; 6356 f[6] = (denormalized()) ? 'D' : '-'; 6357 f[7] = (invalid ()) ? 'I' : '-'; 6358 f[8] = '\x0'; 6359 // output 6360 printf("%04x flags = %s, cc = %s, top = %d", _value & 0xFFFF, f, c, top()); 6361 } 6362 6363 }; 6364 6365 class TagWord { 6366 public: 6367 int32_t _value; 6368 6369 int tag_at(int i) const { return (_value >> (i*2)) & 3; } 6370 6371 void print() const { 6372 printf("%04x", _value & 0xFFFF); 6373 } 6374 6375 }; 6376 6377 class FPU_Register { 6378 public: 6379 int32_t _m0; 6380 int32_t _m1; 6381 int16_t _ex; 6382 6383 bool is_indefinite() const { 6384 return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0; 6385 } 6386 6387 void print() const { 6388 char sign = (_ex < 0) ? '-' : '+'; 6389 const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : " "; 6390 printf("%c%04hx.%08x%08x %s", sign, _ex, _m1, _m0, kind); 6391 }; 6392 6393 }; 6394 6395 class FPU_State { 6396 public: 6397 enum { 6398 register_size = 10, 6399 number_of_registers = 8, 6400 register_mask = 7 6401 }; 6402 6403 ControlWord _control_word; 6404 StatusWord _status_word; 6405 TagWord _tag_word; 6406 int32_t _error_offset; 6407 int32_t _error_selector; 6408 int32_t _data_offset; 6409 int32_t _data_selector; 6410 int8_t _register[register_size * number_of_registers]; 6411 6412 int tag_for_st(int i) const { return _tag_word.tag_at((_status_word.top() + i) & register_mask); } 6413 FPU_Register* st(int i) const { return (FPU_Register*)&_register[register_size * i]; } 6414 6415 const char* tag_as_string(int tag) const { 6416 switch (tag) { 6417 case 0: return "valid"; 6418 case 1: return "zero"; 6419 case 2: return "special"; 6420 case 3: return "empty"; 6421 } 6422 ShouldNotReachHere(); 6423 return NULL; 6424 } 6425 6426 void print() const { 6427 // print computation registers 6428 { int t = _status_word.top(); 6429 for (int i = 0; i < number_of_registers; i++) { 6430 int j = (i - t) & register_mask; 6431 printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j); 6432 st(j)->print(); 6433 printf(" %s\n", tag_as_string(_tag_word.tag_at(i))); 6434 } 6435 } 6436 printf("\n"); 6437 // print control registers 6438 printf("ctrl = "); _control_word.print(); printf("\n"); 6439 printf("stat = "); _status_word .print(); printf("\n"); 6440 printf("tags = "); _tag_word .print(); printf("\n"); 6441 } 6442 6443 }; 6444 6445 class Flag_Register { 6446 public: 6447 int32_t _value; 6448 6449 bool overflow() const { return ((_value >> 11) & 1) != 0; } 6450 bool direction() const { return ((_value >> 10) & 1) != 0; } 6451 bool sign() const { return ((_value >> 7) & 1) != 0; } 6452 bool zero() const { return ((_value >> 6) & 1) != 0; } 6453 bool auxiliary_carry() const { return ((_value >> 4) & 1) != 0; } 6454 bool parity() const { return ((_value >> 2) & 1) != 0; } 6455 bool carry() const { return ((_value >> 0) & 1) != 0; } 6456 6457 void print() const { 6458 // flags 6459 char f[8]; 6460 f[0] = (overflow ()) ? 'O' : '-'; 6461 f[1] = (direction ()) ? 'D' : '-'; 6462 f[2] = (sign ()) ? 'S' : '-'; 6463 f[3] = (zero ()) ? 'Z' : '-'; 6464 f[4] = (auxiliary_carry()) ? 'A' : '-'; 6465 f[5] = (parity ()) ? 'P' : '-'; 6466 f[6] = (carry ()) ? 'C' : '-'; 6467 f[7] = '\x0'; 6468 // output 6469 printf("%08x flags = %s", _value, f); 6470 } 6471 6472 }; 6473 6474 class IU_Register { 6475 public: 6476 int32_t _value; 6477 6478 void print() const { 6479 printf("%08x %11d", _value, _value); 6480 } 6481 6482 }; 6483 6484 class IU_State { 6485 public: 6486 Flag_Register _eflags; 6487 IU_Register _rdi; 6488 IU_Register _rsi; 6489 IU_Register _rbp; 6490 IU_Register _rsp; 6491 IU_Register _rbx; 6492 IU_Register _rdx; 6493 IU_Register _rcx; 6494 IU_Register _rax; 6495 6496 void print() const { 6497 // computation registers 6498 printf("rax, = "); _rax.print(); printf("\n"); 6499 printf("rbx, = "); _rbx.print(); printf("\n"); 6500 printf("rcx = "); _rcx.print(); printf("\n"); 6501 printf("rdx = "); _rdx.print(); printf("\n"); 6502 printf("rdi = "); _rdi.print(); printf("\n"); 6503 printf("rsi = "); _rsi.print(); printf("\n"); 6504 printf("rbp, = "); _rbp.print(); printf("\n"); 6505 printf("rsp = "); _rsp.print(); printf("\n"); 6506 printf("\n"); 6507 // control registers 6508 printf("flgs = "); _eflags.print(); printf("\n"); 6509 } 6510 }; 6511 6512 6513 class CPU_State { 6514 public: 6515 FPU_State _fpu_state; 6516 IU_State _iu_state; 6517 6518 void print() const { 6519 printf("--------------------------------------------------\n"); 6520 _iu_state .print(); 6521 printf("\n"); 6522 _fpu_state.print(); 6523 printf("--------------------------------------------------\n"); 6524 } 6525 6526 }; 6527 6528 6529 static void _print_CPU_state(CPU_State* state) { 6530 state->print(); 6531 }; 6532 6533 6534 void MacroAssembler::print_CPU_state() { 6535 push_CPU_state(); 6536 push(rsp); // pass CPU state 6537 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state))); 6538 addptr(rsp, wordSize); // discard argument 6539 pop_CPU_state(); 6540 } 6541 6542 6543 static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) { 6544 static int counter = 0; 6545 FPU_State* fs = &state->_fpu_state; 6546 counter++; 6547 // For leaf calls, only verify that the top few elements remain empty. 6548 // We only need 1 empty at the top for C2 code. 6549 if( stack_depth < 0 ) { 6550 if( fs->tag_for_st(7) != 3 ) { 6551 printf("FPR7 not empty\n"); 6552 state->print(); 6553 assert(false, "error"); 6554 return false; 6555 } 6556 return true; // All other stack states do not matter 6557 } 6558 6559 assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std, 6560 "bad FPU control word"); 6561 6562 // compute stack depth 6563 int i = 0; 6564 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) < 3) i++; 6565 int d = i; 6566 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++; 6567 // verify findings 6568 if (i != FPU_State::number_of_registers) { 6569 // stack not contiguous 6570 printf("%s: stack not contiguous at ST%d\n", s, i); 6571 state->print(); 6572 assert(false, "error"); 6573 return false; 6574 } 6575 // check if computed stack depth corresponds to expected stack depth 6576 if (stack_depth < 0) { 6577 // expected stack depth is -stack_depth or less 6578 if (d > -stack_depth) { 6579 // too many elements on the stack 6580 printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d); 6581 state->print(); 6582 assert(false, "error"); 6583 return false; 6584 } 6585 } else { 6586 // expected stack depth is stack_depth 6587 if (d != stack_depth) { 6588 // wrong stack depth 6589 printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d); 6590 state->print(); 6591 assert(false, "error"); 6592 return false; 6593 } 6594 } 6595 // everything is cool 6596 return true; 6597 } 6598 6599 6600 void MacroAssembler::verify_FPU(int stack_depth, const char* s) { 6601 if (!VerifyFPU) return; 6602 push_CPU_state(); 6603 push(rsp); // pass CPU state 6604 ExternalAddress msg((address) s); 6605 // pass message string s 6606 pushptr(msg.addr()); 6607 push(stack_depth); // pass stack depth 6608 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU))); 6609 addptr(rsp, 3 * wordSize); // discard arguments 6610 // check for error 6611 { Label L; 6612 testl(rax, rax); 6613 jcc(Assembler::notZero, L); 6614 int3(); // break if error condition 6615 bind(L); 6616 } 6617 pop_CPU_state(); 6618 } 6619 6620 void MacroAssembler::restore_cpu_control_state_after_jni() { 6621 // Either restore the MXCSR register after returning from the JNI Call 6622 // or verify that it wasn't changed (with -Xcheck:jni flag). 6623 if (VM_Version::supports_sse()) { 6624 if (RestoreMXCSROnJNICalls) { 6625 ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std())); 6626 } else if (CheckJNICalls) { 6627 call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry())); 6628 } 6629 } 6630 if (VM_Version::supports_avx()) { 6631 // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty. 6632 vzeroupper(); 6633 } 6634 6635 #ifndef _LP64 6636 // Either restore the x87 floating pointer control word after returning 6637 // from the JNI call or verify that it wasn't changed. 6638 if (CheckJNICalls) { 6639 call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry())); 6640 } 6641 #endif // _LP64 6642 } 6643 6644 6645 void MacroAssembler::load_klass(Register dst, Register src) { 6646 #ifdef _LP64 6647 if (UseCompressedClassPointers) { 6648 movl(dst, Address(src, oopDesc::klass_offset_in_bytes())); 6649 decode_klass_not_null(dst); 6650 } else 6651 #endif 6652 movptr(dst, Address(src, oopDesc::klass_offset_in_bytes())); 6653 } 6654 6655 void MacroAssembler::load_prototype_header(Register dst, Register src) { 6656 load_klass(dst, src); 6657 movptr(dst, Address(dst, Klass::prototype_header_offset())); 6658 } 6659 6660 void MacroAssembler::store_klass(Register dst, Register src) { 6661 #ifdef _LP64 6662 if (UseCompressedClassPointers) { 6663 encode_klass_not_null(src); 6664 movl(Address(dst, oopDesc::klass_offset_in_bytes()), src); 6665 } else 6666 #endif 6667 movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src); 6668 } 6669 6670 void MacroAssembler::load_heap_oop(Register dst, Address src) { 6671 #ifdef _LP64 6672 // FIXME: Must change all places where we try to load the klass. 6673 if (UseCompressedOops) { 6674 movl(dst, src); 6675 decode_heap_oop(dst); 6676 } else 6677 #endif 6678 movptr(dst, src); 6679 } 6680 6681 // Doesn't do verfication, generates fixed size code 6682 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src) { 6683 #ifdef _LP64 6684 if (UseCompressedOops) { 6685 movl(dst, src); 6686 decode_heap_oop_not_null(dst); 6687 } else 6688 #endif 6689 movptr(dst, src); 6690 } 6691 6692 void MacroAssembler::store_heap_oop(Address dst, Register src) { 6693 #ifdef _LP64 6694 if (UseCompressedOops) { 6695 assert(!dst.uses(src), "not enough registers"); 6696 encode_heap_oop(src); 6697 movl(dst, src); 6698 } else 6699 #endif 6700 movptr(dst, src); 6701 } 6702 6703 void MacroAssembler::cmp_heap_oop(Register src1, Address src2, Register tmp) { 6704 assert_different_registers(src1, tmp); 6705 #ifdef _LP64 6706 if (UseCompressedOops) { 6707 bool did_push = false; 6708 if (tmp == noreg) { 6709 tmp = rax; 6710 push(tmp); 6711 did_push = true; 6712 assert(!src2.uses(rsp), "can't push"); 6713 } 6714 load_heap_oop(tmp, src2); 6715 cmpptr(src1, tmp); 6716 if (did_push) pop(tmp); 6717 } else 6718 #endif 6719 cmpptr(src1, src2); 6720 } 6721 6722 // Used for storing NULLs. 6723 void MacroAssembler::store_heap_oop_null(Address dst) { 6724 #ifdef _LP64 6725 if (UseCompressedOops) { 6726 movl(dst, (int32_t)NULL_WORD); 6727 } else { 6728 movslq(dst, (int32_t)NULL_WORD); 6729 } 6730 #else 6731 movl(dst, (int32_t)NULL_WORD); 6732 #endif 6733 } 6734 6735 #ifdef _LP64 6736 void MacroAssembler::store_klass_gap(Register dst, Register src) { 6737 if (UseCompressedClassPointers) { 6738 // Store to klass gap in destination 6739 movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src); 6740 } 6741 } 6742 6743 #ifdef ASSERT 6744 void MacroAssembler::verify_heapbase(const char* msg) { 6745 assert (UseCompressedOops, "should be compressed"); 6746 assert (Universe::heap() != NULL, "java heap should be initialized"); 6747 if (CheckCompressedOops) { 6748 Label ok; 6749 push(rscratch1); // cmpptr trashes rscratch1 6750 cmpptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr())); 6751 jcc(Assembler::equal, ok); 6752 STOP(msg); 6753 bind(ok); 6754 pop(rscratch1); 6755 } 6756 } 6757 #endif 6758 6759 // Algorithm must match oop.inline.hpp encode_heap_oop. 6760 void MacroAssembler::encode_heap_oop(Register r) { 6761 #ifdef ASSERT 6762 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?"); 6763 #endif 6764 verify_oop(r, "broken oop in encode_heap_oop"); 6765 if (Universe::narrow_oop_base() == NULL) { 6766 if (Universe::narrow_oop_shift() != 0) { 6767 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6768 shrq(r, LogMinObjAlignmentInBytes); 6769 } 6770 return; 6771 } 6772 testq(r, r); 6773 cmovq(Assembler::equal, r, r12_heapbase); 6774 subq(r, r12_heapbase); 6775 shrq(r, LogMinObjAlignmentInBytes); 6776 } 6777 6778 void MacroAssembler::encode_heap_oop_not_null(Register r) { 6779 #ifdef ASSERT 6780 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?"); 6781 if (CheckCompressedOops) { 6782 Label ok; 6783 testq(r, r); 6784 jcc(Assembler::notEqual, ok); 6785 STOP("null oop passed to encode_heap_oop_not_null"); 6786 bind(ok); 6787 } 6788 #endif 6789 verify_oop(r, "broken oop in encode_heap_oop_not_null"); 6790 if (Universe::narrow_oop_base() != NULL) { 6791 subq(r, r12_heapbase); 6792 } 6793 if (Universe::narrow_oop_shift() != 0) { 6794 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6795 shrq(r, LogMinObjAlignmentInBytes); 6796 } 6797 } 6798 6799 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) { 6800 #ifdef ASSERT 6801 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?"); 6802 if (CheckCompressedOops) { 6803 Label ok; 6804 testq(src, src); 6805 jcc(Assembler::notEqual, ok); 6806 STOP("null oop passed to encode_heap_oop_not_null2"); 6807 bind(ok); 6808 } 6809 #endif 6810 verify_oop(src, "broken oop in encode_heap_oop_not_null2"); 6811 if (dst != src) { 6812 movq(dst, src); 6813 } 6814 if (Universe::narrow_oop_base() != NULL) { 6815 subq(dst, r12_heapbase); 6816 } 6817 if (Universe::narrow_oop_shift() != 0) { 6818 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6819 shrq(dst, LogMinObjAlignmentInBytes); 6820 } 6821 } 6822 6823 void MacroAssembler::decode_heap_oop(Register r) { 6824 #ifdef ASSERT 6825 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?"); 6826 #endif 6827 if (Universe::narrow_oop_base() == NULL) { 6828 if (Universe::narrow_oop_shift() != 0) { 6829 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6830 shlq(r, LogMinObjAlignmentInBytes); 6831 } 6832 } else { 6833 Label done; 6834 shlq(r, LogMinObjAlignmentInBytes); 6835 jccb(Assembler::equal, done); 6836 addq(r, r12_heapbase); 6837 bind(done); 6838 } 6839 verify_oop(r, "broken oop in decode_heap_oop"); 6840 } 6841 6842 void MacroAssembler::decode_heap_oop_not_null(Register r) { 6843 // Note: it will change flags 6844 assert (UseCompressedOops, "should only be used for compressed headers"); 6845 assert (Universe::heap() != NULL, "java heap should be initialized"); 6846 // Cannot assert, unverified entry point counts instructions (see .ad file) 6847 // vtableStubs also counts instructions in pd_code_size_limit. 6848 // Also do not verify_oop as this is called by verify_oop. 6849 if (Universe::narrow_oop_shift() != 0) { 6850 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6851 shlq(r, LogMinObjAlignmentInBytes); 6852 if (Universe::narrow_oop_base() != NULL) { 6853 addq(r, r12_heapbase); 6854 } 6855 } else { 6856 assert (Universe::narrow_oop_base() == NULL, "sanity"); 6857 } 6858 } 6859 6860 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) { 6861 // Note: it will change flags 6862 assert (UseCompressedOops, "should only be used for compressed headers"); 6863 assert (Universe::heap() != NULL, "java heap should be initialized"); 6864 // Cannot assert, unverified entry point counts instructions (see .ad file) 6865 // vtableStubs also counts instructions in pd_code_size_limit. 6866 // Also do not verify_oop as this is called by verify_oop. 6867 if (Universe::narrow_oop_shift() != 0) { 6868 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 6869 if (LogMinObjAlignmentInBytes == Address::times_8) { 6870 leaq(dst, Address(r12_heapbase, src, Address::times_8, 0)); 6871 } else { 6872 if (dst != src) { 6873 movq(dst, src); 6874 } 6875 shlq(dst, LogMinObjAlignmentInBytes); 6876 if (Universe::narrow_oop_base() != NULL) { 6877 addq(dst, r12_heapbase); 6878 } 6879 } 6880 } else { 6881 assert (Universe::narrow_oop_base() == NULL, "sanity"); 6882 if (dst != src) { 6883 movq(dst, src); 6884 } 6885 } 6886 } 6887 6888 void MacroAssembler::encode_klass_not_null(Register r) { 6889 if (Universe::narrow_klass_base() != NULL) { 6890 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base. 6891 assert(r != r12_heapbase, "Encoding a klass in r12"); 6892 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base()); 6893 subq(r, r12_heapbase); 6894 } 6895 if (Universe::narrow_klass_shift() != 0) { 6896 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 6897 shrq(r, LogKlassAlignmentInBytes); 6898 } 6899 if (Universe::narrow_klass_base() != NULL) { 6900 reinit_heapbase(); 6901 } 6902 } 6903 6904 void MacroAssembler::encode_klass_not_null(Register dst, Register src) { 6905 if (dst == src) { 6906 encode_klass_not_null(src); 6907 } else { 6908 if (Universe::narrow_klass_base() != NULL) { 6909 mov64(dst, (int64_t)Universe::narrow_klass_base()); 6910 negq(dst); 6911 addq(dst, src); 6912 } else { 6913 movptr(dst, src); 6914 } 6915 if (Universe::narrow_klass_shift() != 0) { 6916 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 6917 shrq(dst, LogKlassAlignmentInBytes); 6918 } 6919 } 6920 } 6921 6922 // Function instr_size_for_decode_klass_not_null() counts the instructions 6923 // generated by decode_klass_not_null(register r) and reinit_heapbase(), 6924 // when (Universe::heap() != NULL). Hence, if the instructions they 6925 // generate change, then this method needs to be updated. 6926 int MacroAssembler::instr_size_for_decode_klass_not_null() { 6927 assert (UseCompressedClassPointers, "only for compressed klass ptrs"); 6928 if (Universe::narrow_klass_base() != NULL) { 6929 // mov64 + addq + shlq? + mov64 (for reinit_heapbase()). 6930 return (Universe::narrow_klass_shift() == 0 ? 20 : 24); 6931 } else { 6932 // longest load decode klass function, mov64, leaq 6933 return 16; 6934 } 6935 } 6936 6937 // !!! If the instructions that get generated here change then function 6938 // instr_size_for_decode_klass_not_null() needs to get updated. 6939 void MacroAssembler::decode_klass_not_null(Register r) { 6940 // Note: it will change flags 6941 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 6942 assert(r != r12_heapbase, "Decoding a klass in r12"); 6943 // Cannot assert, unverified entry point counts instructions (see .ad file) 6944 // vtableStubs also counts instructions in pd_code_size_limit. 6945 // Also do not verify_oop as this is called by verify_oop. 6946 if (Universe::narrow_klass_shift() != 0) { 6947 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 6948 shlq(r, LogKlassAlignmentInBytes); 6949 } 6950 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base. 6951 if (Universe::narrow_klass_base() != NULL) { 6952 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base()); 6953 addq(r, r12_heapbase); 6954 reinit_heapbase(); 6955 } 6956 } 6957 6958 void MacroAssembler::decode_klass_not_null(Register dst, Register src) { 6959 // Note: it will change flags 6960 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 6961 if (dst == src) { 6962 decode_klass_not_null(dst); 6963 } else { 6964 // Cannot assert, unverified entry point counts instructions (see .ad file) 6965 // vtableStubs also counts instructions in pd_code_size_limit. 6966 // Also do not verify_oop as this is called by verify_oop. 6967 mov64(dst, (int64_t)Universe::narrow_klass_base()); 6968 if (Universe::narrow_klass_shift() != 0) { 6969 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 6970 assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?"); 6971 leaq(dst, Address(dst, src, Address::times_8, 0)); 6972 } else { 6973 addq(dst, src); 6974 } 6975 } 6976 } 6977 6978 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) { 6979 assert (UseCompressedOops, "should only be used for compressed headers"); 6980 assert (Universe::heap() != NULL, "java heap should be initialized"); 6981 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 6982 int oop_index = oop_recorder()->find_index(obj); 6983 RelocationHolder rspec = oop_Relocation::spec(oop_index); 6984 mov_narrow_oop(dst, oop_index, rspec); 6985 } 6986 6987 void MacroAssembler::set_narrow_oop(Address dst, jobject obj) { 6988 assert (UseCompressedOops, "should only be used for compressed headers"); 6989 assert (Universe::heap() != NULL, "java heap should be initialized"); 6990 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 6991 int oop_index = oop_recorder()->find_index(obj); 6992 RelocationHolder rspec = oop_Relocation::spec(oop_index); 6993 mov_narrow_oop(dst, oop_index, rspec); 6994 } 6995 6996 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) { 6997 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 6998 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 6999 int klass_index = oop_recorder()->find_index(k); 7000 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 7001 mov_narrow_oop(dst, Klass::encode_klass(k), rspec); 7002 } 7003 7004 void MacroAssembler::set_narrow_klass(Address dst, Klass* k) { 7005 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7006 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7007 int klass_index = oop_recorder()->find_index(k); 7008 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 7009 mov_narrow_oop(dst, Klass::encode_klass(k), rspec); 7010 } 7011 7012 void MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) { 7013 assert (UseCompressedOops, "should only be used for compressed headers"); 7014 assert (Universe::heap() != NULL, "java heap should be initialized"); 7015 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7016 int oop_index = oop_recorder()->find_index(obj); 7017 RelocationHolder rspec = oop_Relocation::spec(oop_index); 7018 Assembler::cmp_narrow_oop(dst, oop_index, rspec); 7019 } 7020 7021 void MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) { 7022 assert (UseCompressedOops, "should only be used for compressed headers"); 7023 assert (Universe::heap() != NULL, "java heap should be initialized"); 7024 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7025 int oop_index = oop_recorder()->find_index(obj); 7026 RelocationHolder rspec = oop_Relocation::spec(oop_index); 7027 Assembler::cmp_narrow_oop(dst, oop_index, rspec); 7028 } 7029 7030 void MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) { 7031 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7032 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7033 int klass_index = oop_recorder()->find_index(k); 7034 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 7035 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec); 7036 } 7037 7038 void MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) { 7039 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 7040 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 7041 int klass_index = oop_recorder()->find_index(k); 7042 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 7043 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec); 7044 } 7045 7046 void MacroAssembler::reinit_heapbase() { 7047 if (UseCompressedOops || UseCompressedClassPointers) { 7048 if (Universe::heap() != NULL) { 7049 if (Universe::narrow_oop_base() == NULL) { 7050 MacroAssembler::xorptr(r12_heapbase, r12_heapbase); 7051 } else { 7052 mov64(r12_heapbase, (int64_t)Universe::narrow_ptrs_base()); 7053 } 7054 } else { 7055 movptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr())); 7056 } 7057 } 7058 } 7059 7060 #endif // _LP64 7061 7062 7063 // C2 compiled method's prolog code. 7064 void MacroAssembler::verified_entry(int framesize, int stack_bang_size, bool fp_mode_24b) { 7065 7066 // WARNING: Initial instruction MUST be 5 bytes or longer so that 7067 // NativeJump::patch_verified_entry will be able to patch out the entry 7068 // code safely. The push to verify stack depth is ok at 5 bytes, 7069 // the frame allocation can be either 3 or 6 bytes. So if we don't do 7070 // stack bang then we must use the 6 byte frame allocation even if 7071 // we have no frame. :-( 7072 assert(stack_bang_size >= framesize || stack_bang_size <= 0, "stack bang size incorrect"); 7073 7074 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned"); 7075 // Remove word for return addr 7076 framesize -= wordSize; 7077 stack_bang_size -= wordSize; 7078 7079 // Calls to C2R adapters often do not accept exceptional returns. 7080 // We require that their callers must bang for them. But be careful, because 7081 // some VM calls (such as call site linkage) can use several kilobytes of 7082 // stack. But the stack safety zone should account for that. 7083 // See bugs 4446381, 4468289, 4497237. 7084 if (stack_bang_size > 0) { 7085 generate_stack_overflow_check(stack_bang_size); 7086 7087 // We always push rbp, so that on return to interpreter rbp, will be 7088 // restored correctly and we can correct the stack. 7089 push(rbp); 7090 // Save caller's stack pointer into RBP if the frame pointer is preserved. 7091 if (PreserveFramePointer) { 7092 mov(rbp, rsp); 7093 } 7094 // Remove word for ebp 7095 framesize -= wordSize; 7096 7097 // Create frame 7098 if (framesize) { 7099 subptr(rsp, framesize); 7100 } 7101 } else { 7102 // Create frame (force generation of a 4 byte immediate value) 7103 subptr_imm32(rsp, framesize); 7104 7105 // Save RBP register now. 7106 framesize -= wordSize; 7107 movptr(Address(rsp, framesize), rbp); 7108 // Save caller's stack pointer into RBP if the frame pointer is preserved. 7109 if (PreserveFramePointer) { 7110 movptr(rbp, rsp); 7111 if (framesize > 0) { 7112 addptr(rbp, framesize); 7113 } 7114 } 7115 } 7116 7117 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth 7118 framesize -= wordSize; 7119 movptr(Address(rsp, framesize), (int32_t)0xbadb100d); 7120 } 7121 7122 #ifndef _LP64 7123 // If method sets FPU control word do it now 7124 if (fp_mode_24b) { 7125 fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24())); 7126 } 7127 if (UseSSE >= 2 && VerifyFPU) { 7128 verify_FPU(0, "FPU stack must be clean on entry"); 7129 } 7130 #endif 7131 7132 #ifdef ASSERT 7133 if (VerifyStackAtCalls) { 7134 Label L; 7135 push(rax); 7136 mov(rax, rsp); 7137 andptr(rax, StackAlignmentInBytes-1); 7138 cmpptr(rax, StackAlignmentInBytes-wordSize); 7139 pop(rax); 7140 jcc(Assembler::equal, L); 7141 STOP("Stack is not properly aligned!"); 7142 bind(L); 7143 } 7144 #endif 7145 7146 } 7147 7148 void MacroAssembler::clear_mem(Register base, Register cnt, Register tmp) { 7149 // cnt - number of qwords (8-byte words). 7150 // base - start address, qword aligned. 7151 assert(base==rdi, "base register must be edi for rep stos"); 7152 assert(tmp==rax, "tmp register must be eax for rep stos"); 7153 assert(cnt==rcx, "cnt register must be ecx for rep stos"); 7154 7155 xorptr(tmp, tmp); 7156 if (UseFastStosb) { 7157 shlptr(cnt,3); // convert to number of bytes 7158 rep_stosb(); 7159 } else { 7160 NOT_LP64(shlptr(cnt,1);) // convert to number of dwords for 32-bit VM 7161 rep_stos(); 7162 } 7163 } 7164 7165 #ifdef COMPILER2 7166 7167 // IndexOf for constant substrings with size >= 8 chars 7168 // which don't need to be loaded through stack. 7169 void MacroAssembler::string_indexofC8(Register str1, Register str2, 7170 Register cnt1, Register cnt2, 7171 int int_cnt2, Register result, 7172 XMMRegister vec, Register tmp, 7173 int ae) { 7174 ShortBranchVerifier sbv(this); 7175 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 7176 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 7177 assert(ae != StrIntrinsicNode::LU, "Invalid encoding"); 7178 7179 // This method uses the pcmpestri instruction with bound registers 7180 // inputs: 7181 // xmm - substring 7182 // rax - substring length (elements count) 7183 // mem - scanned string 7184 // rdx - string length (elements count) 7185 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts) 7186 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes) 7187 // outputs: 7188 // rcx - matched index in string 7189 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 7190 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts 7191 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8 7192 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2; 7193 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1; 7194 7195 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, 7196 RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR, 7197 MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE; 7198 7199 // Note, inline_string_indexOf() generates checks: 7200 // if (substr.count > string.count) return -1; 7201 // if (substr.count == 0) return 0; 7202 assert(int_cnt2 >= stride, "this code is used only for cnt2 >= 8 chars"); 7203 7204 // Load substring. 7205 if (ae == StrIntrinsicNode::UL) { 7206 pmovzxbw(vec, Address(str2, 0)); 7207 } else { 7208 movdqu(vec, Address(str2, 0)); 7209 } 7210 movl(cnt2, int_cnt2); 7211 movptr(result, str1); // string addr 7212 7213 if (int_cnt2 > stride) { 7214 jmpb(SCAN_TO_SUBSTR); 7215 7216 // Reload substr for rescan, this code 7217 // is executed only for large substrings (> 8 chars) 7218 bind(RELOAD_SUBSTR); 7219 if (ae == StrIntrinsicNode::UL) { 7220 pmovzxbw(vec, Address(str2, 0)); 7221 } else { 7222 movdqu(vec, Address(str2, 0)); 7223 } 7224 negptr(cnt2); // Jumped here with negative cnt2, convert to positive 7225 7226 bind(RELOAD_STR); 7227 // We came here after the beginning of the substring was 7228 // matched but the rest of it was not so we need to search 7229 // again. Start from the next element after the previous match. 7230 7231 // cnt2 is number of substring reminding elements and 7232 // cnt1 is number of string reminding elements when cmp failed. 7233 // Restored cnt1 = cnt1 - cnt2 + int_cnt2 7234 subl(cnt1, cnt2); 7235 addl(cnt1, int_cnt2); 7236 movl(cnt2, int_cnt2); // Now restore cnt2 7237 7238 decrementl(cnt1); // Shift to next element 7239 cmpl(cnt1, cnt2); 7240 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7241 7242 addptr(result, (1<<scale1)); 7243 7244 } // (int_cnt2 > 8) 7245 7246 // Scan string for start of substr in 16-byte vectors 7247 bind(SCAN_TO_SUBSTR); 7248 pcmpestri(vec, Address(result, 0), mode); 7249 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1 7250 subl(cnt1, stride); 7251 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string 7252 cmpl(cnt1, cnt2); 7253 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7254 addptr(result, 16); 7255 jmpb(SCAN_TO_SUBSTR); 7256 7257 // Found a potential substr 7258 bind(FOUND_CANDIDATE); 7259 // Matched whole vector if first element matched (tmp(rcx) == 0). 7260 if (int_cnt2 == stride) { 7261 jccb(Assembler::overflow, RET_FOUND); // OF == 1 7262 } else { // int_cnt2 > 8 7263 jccb(Assembler::overflow, FOUND_SUBSTR); 7264 } 7265 // After pcmpestri tmp(rcx) contains matched element index 7266 // Compute start addr of substr 7267 lea(result, Address(result, tmp, scale1)); 7268 7269 // Make sure string is still long enough 7270 subl(cnt1, tmp); 7271 cmpl(cnt1, cnt2); 7272 if (int_cnt2 == stride) { 7273 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR); 7274 } else { // int_cnt2 > 8 7275 jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD); 7276 } 7277 // Left less then substring. 7278 7279 bind(RET_NOT_FOUND); 7280 movl(result, -1); 7281 jmpb(EXIT); 7282 7283 if (int_cnt2 > stride) { 7284 // This code is optimized for the case when whole substring 7285 // is matched if its head is matched. 7286 bind(MATCH_SUBSTR_HEAD); 7287 pcmpestri(vec, Address(result, 0), mode); 7288 // Reload only string if does not match 7289 jccb(Assembler::noOverflow, RELOAD_STR); // OF == 0 7290 7291 Label CONT_SCAN_SUBSTR; 7292 // Compare the rest of substring (> 8 chars). 7293 bind(FOUND_SUBSTR); 7294 // First 8 chars are already matched. 7295 negptr(cnt2); 7296 addptr(cnt2, stride); 7297 7298 bind(SCAN_SUBSTR); 7299 subl(cnt1, stride); 7300 cmpl(cnt2, -stride); // Do not read beyond substring 7301 jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR); 7302 // Back-up strings to avoid reading beyond substring: 7303 // cnt1 = cnt1 - cnt2 + 8 7304 addl(cnt1, cnt2); // cnt2 is negative 7305 addl(cnt1, stride); 7306 movl(cnt2, stride); negptr(cnt2); 7307 bind(CONT_SCAN_SUBSTR); 7308 if (int_cnt2 < (int)G) { 7309 int tail_off1 = int_cnt2<<scale1; 7310 int tail_off2 = int_cnt2<<scale2; 7311 if (ae == StrIntrinsicNode::UL) { 7312 pmovzxbw(vec, Address(str2, cnt2, scale2, tail_off2)); 7313 } else { 7314 movdqu(vec, Address(str2, cnt2, scale2, tail_off2)); 7315 } 7316 pcmpestri(vec, Address(result, cnt2, scale1, tail_off1), mode); 7317 } else { 7318 // calculate index in register to avoid integer overflow (int_cnt2*2) 7319 movl(tmp, int_cnt2); 7320 addptr(tmp, cnt2); 7321 if (ae == StrIntrinsicNode::UL) { 7322 pmovzxbw(vec, Address(str2, tmp, scale2, 0)); 7323 } else { 7324 movdqu(vec, Address(str2, tmp, scale2, 0)); 7325 } 7326 pcmpestri(vec, Address(result, tmp, scale1, 0), mode); 7327 } 7328 // Need to reload strings pointers if not matched whole vector 7329 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0 7330 addptr(cnt2, stride); 7331 jcc(Assembler::negative, SCAN_SUBSTR); 7332 // Fall through if found full substring 7333 7334 } // (int_cnt2 > 8) 7335 7336 bind(RET_FOUND); 7337 // Found result if we matched full small substring. 7338 // Compute substr offset 7339 subptr(result, str1); 7340 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 7341 shrl(result, 1); // index 7342 } 7343 bind(EXIT); 7344 7345 } // string_indexofC8 7346 7347 // Small strings are loaded through stack if they cross page boundary. 7348 void MacroAssembler::string_indexof(Register str1, Register str2, 7349 Register cnt1, Register cnt2, 7350 int int_cnt2, Register result, 7351 XMMRegister vec, Register tmp, 7352 int ae) { 7353 ShortBranchVerifier sbv(this); 7354 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 7355 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 7356 assert(ae != StrIntrinsicNode::LU, "Invalid encoding"); 7357 7358 // 7359 // int_cnt2 is length of small (< 8 chars) constant substring 7360 // or (-1) for non constant substring in which case its length 7361 // is in cnt2 register. 7362 // 7363 // Note, inline_string_indexOf() generates checks: 7364 // if (substr.count > string.count) return -1; 7365 // if (substr.count == 0) return 0; 7366 // 7367 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8 7368 assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < stride), "should be != 0"); 7369 // This method uses the pcmpestri instruction with bound registers 7370 // inputs: 7371 // xmm - substring 7372 // rax - substring length (elements count) 7373 // mem - scanned string 7374 // rdx - string length (elements count) 7375 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts) 7376 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes) 7377 // outputs: 7378 // rcx - matched index in string 7379 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 7380 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts 7381 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2; 7382 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1; 7383 7384 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR, 7385 RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR, 7386 FOUND_CANDIDATE; 7387 7388 { //======================================================== 7389 // We don't know where these strings are located 7390 // and we can't read beyond them. Load them through stack. 7391 Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR; 7392 7393 movptr(tmp, rsp); // save old SP 7394 7395 if (int_cnt2 > 0) { // small (< 8 chars) constant substring 7396 if (int_cnt2 == (1>>scale2)) { // One byte 7397 assert((ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL), "Only possible for latin1 encoding"); 7398 load_unsigned_byte(result, Address(str2, 0)); 7399 movdl(vec, result); // move 32 bits 7400 } else if (ae == StrIntrinsicNode::LL && int_cnt2 == 3) { // Three bytes 7401 // Not enough header space in 32-bit VM: 12+3 = 15. 7402 movl(result, Address(str2, -1)); 7403 shrl(result, 8); 7404 movdl(vec, result); // move 32 bits 7405 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (2>>scale2)) { // One char 7406 load_unsigned_short(result, Address(str2, 0)); 7407 movdl(vec, result); // move 32 bits 7408 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (4>>scale2)) { // Two chars 7409 movdl(vec, Address(str2, 0)); // move 32 bits 7410 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (8>>scale2)) { // Four chars 7411 movq(vec, Address(str2, 0)); // move 64 bits 7412 } else { // cnt2 = { 3, 5, 6, 7 } || (ae == StrIntrinsicNode::UL && cnt2 ={2, ..., 7}) 7413 // Array header size is 12 bytes in 32-bit VM 7414 // + 6 bytes for 3 chars == 18 bytes, 7415 // enough space to load vec and shift. 7416 assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity"); 7417 if (ae == StrIntrinsicNode::UL) { 7418 int tail_off = int_cnt2-8; 7419 pmovzxbw(vec, Address(str2, tail_off)); 7420 psrldq(vec, -2*tail_off); 7421 } 7422 else { 7423 int tail_off = int_cnt2*(1<<scale2); 7424 movdqu(vec, Address(str2, tail_off-16)); 7425 psrldq(vec, 16-tail_off); 7426 } 7427 } 7428 } else { // not constant substring 7429 cmpl(cnt2, stride); 7430 jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough 7431 7432 // We can read beyond string if srt+16 does not cross page boundary 7433 // since heaps are aligned and mapped by pages. 7434 assert(os::vm_page_size() < (int)G, "default page should be small"); 7435 movl(result, str2); // We need only low 32 bits 7436 andl(result, (os::vm_page_size()-1)); 7437 cmpl(result, (os::vm_page_size()-16)); 7438 jccb(Assembler::belowEqual, CHECK_STR); 7439 7440 // Move small strings to stack to allow load 16 bytes into vec. 7441 subptr(rsp, 16); 7442 int stk_offset = wordSize-(1<<scale2); 7443 push(cnt2); 7444 7445 bind(COPY_SUBSTR); 7446 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL) { 7447 load_unsigned_byte(result, Address(str2, cnt2, scale2, -1)); 7448 movb(Address(rsp, cnt2, scale2, stk_offset), result); 7449 } else if (ae == StrIntrinsicNode::UU) { 7450 load_unsigned_short(result, Address(str2, cnt2, scale2, -2)); 7451 movw(Address(rsp, cnt2, scale2, stk_offset), result); 7452 } 7453 decrement(cnt2); 7454 jccb(Assembler::notZero, COPY_SUBSTR); 7455 7456 pop(cnt2); 7457 movptr(str2, rsp); // New substring address 7458 } // non constant 7459 7460 bind(CHECK_STR); 7461 cmpl(cnt1, stride); 7462 jccb(Assembler::aboveEqual, BIG_STRINGS); 7463 7464 // Check cross page boundary. 7465 movl(result, str1); // We need only low 32 bits 7466 andl(result, (os::vm_page_size()-1)); 7467 cmpl(result, (os::vm_page_size()-16)); 7468 jccb(Assembler::belowEqual, BIG_STRINGS); 7469 7470 subptr(rsp, 16); 7471 int stk_offset = -(1<<scale1); 7472 if (int_cnt2 < 0) { // not constant 7473 push(cnt2); 7474 stk_offset += wordSize; 7475 } 7476 movl(cnt2, cnt1); 7477 7478 bind(COPY_STR); 7479 if (ae == StrIntrinsicNode::LL) { 7480 load_unsigned_byte(result, Address(str1, cnt2, scale1, -1)); 7481 movb(Address(rsp, cnt2, scale1, stk_offset), result); 7482 } else { 7483 load_unsigned_short(result, Address(str1, cnt2, scale1, -2)); 7484 movw(Address(rsp, cnt2, scale1, stk_offset), result); 7485 } 7486 decrement(cnt2); 7487 jccb(Assembler::notZero, COPY_STR); 7488 7489 if (int_cnt2 < 0) { // not constant 7490 pop(cnt2); 7491 } 7492 movptr(str1, rsp); // New string address 7493 7494 bind(BIG_STRINGS); 7495 // Load substring. 7496 if (int_cnt2 < 0) { // -1 7497 if (ae == StrIntrinsicNode::UL) { 7498 pmovzxbw(vec, Address(str2, 0)); 7499 } else { 7500 movdqu(vec, Address(str2, 0)); 7501 } 7502 push(cnt2); // substr count 7503 push(str2); // substr addr 7504 push(str1); // string addr 7505 } else { 7506 // Small (< 8 chars) constant substrings are loaded already. 7507 movl(cnt2, int_cnt2); 7508 } 7509 push(tmp); // original SP 7510 7511 } // Finished loading 7512 7513 //======================================================== 7514 // Start search 7515 // 7516 7517 movptr(result, str1); // string addr 7518 7519 if (int_cnt2 < 0) { // Only for non constant substring 7520 jmpb(SCAN_TO_SUBSTR); 7521 7522 // SP saved at sp+0 7523 // String saved at sp+1*wordSize 7524 // Substr saved at sp+2*wordSize 7525 // Substr count saved at sp+3*wordSize 7526 7527 // Reload substr for rescan, this code 7528 // is executed only for large substrings (> 8 chars) 7529 bind(RELOAD_SUBSTR); 7530 movptr(str2, Address(rsp, 2*wordSize)); 7531 movl(cnt2, Address(rsp, 3*wordSize)); 7532 if (ae == StrIntrinsicNode::UL) { 7533 pmovzxbw(vec, Address(str2, 0)); 7534 } else { 7535 movdqu(vec, Address(str2, 0)); 7536 } 7537 // We came here after the beginning of the substring was 7538 // matched but the rest of it was not so we need to search 7539 // again. Start from the next element after the previous match. 7540 subptr(str1, result); // Restore counter 7541 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 7542 shrl(str1, 1); 7543 } 7544 addl(cnt1, str1); 7545 decrementl(cnt1); // Shift to next element 7546 cmpl(cnt1, cnt2); 7547 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7548 7549 addptr(result, (1<<scale1)); 7550 } // non constant 7551 7552 // Scan string for start of substr in 16-byte vectors 7553 bind(SCAN_TO_SUBSTR); 7554 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 7555 pcmpestri(vec, Address(result, 0), mode); 7556 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1 7557 subl(cnt1, stride); 7558 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string 7559 cmpl(cnt1, cnt2); 7560 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 7561 addptr(result, 16); 7562 7563 bind(ADJUST_STR); 7564 cmpl(cnt1, stride); // Do not read beyond string 7565 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR); 7566 // Back-up string to avoid reading beyond string. 7567 lea(result, Address(result, cnt1, scale1, -16)); 7568 movl(cnt1, stride); 7569 jmpb(SCAN_TO_SUBSTR); 7570 7571 // Found a potential substr 7572 bind(FOUND_CANDIDATE); 7573 // After pcmpestri tmp(rcx) contains matched element index 7574 7575 // Make sure string is still long enough 7576 subl(cnt1, tmp); 7577 cmpl(cnt1, cnt2); 7578 jccb(Assembler::greaterEqual, FOUND_SUBSTR); 7579 // Left less then substring. 7580 7581 bind(RET_NOT_FOUND); 7582 movl(result, -1); 7583 jmpb(CLEANUP); 7584 7585 bind(FOUND_SUBSTR); 7586 // Compute start addr of substr 7587 lea(result, Address(result, tmp, scale1)); 7588 if (int_cnt2 > 0) { // Constant substring 7589 // Repeat search for small substring (< 8 chars) 7590 // from new point without reloading substring. 7591 // Have to check that we don't read beyond string. 7592 cmpl(tmp, stride-int_cnt2); 7593 jccb(Assembler::greater, ADJUST_STR); 7594 // Fall through if matched whole substring. 7595 } else { // non constant 7596 assert(int_cnt2 == -1, "should be != 0"); 7597 7598 addl(tmp, cnt2); 7599 // Found result if we matched whole substring. 7600 cmpl(tmp, stride); 7601 jccb(Assembler::lessEqual, RET_FOUND); 7602 7603 // Repeat search for small substring (<= 8 chars) 7604 // from new point 'str1' without reloading substring. 7605 cmpl(cnt2, stride); 7606 // Have to check that we don't read beyond string. 7607 jccb(Assembler::lessEqual, ADJUST_STR); 7608 7609 Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG; 7610 // Compare the rest of substring (> 8 chars). 7611 movptr(str1, result); 7612 7613 cmpl(tmp, cnt2); 7614 // First 8 chars are already matched. 7615 jccb(Assembler::equal, CHECK_NEXT); 7616 7617 bind(SCAN_SUBSTR); 7618 pcmpestri(vec, Address(str1, 0), mode); 7619 // Need to reload strings pointers if not matched whole vector 7620 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0 7621 7622 bind(CHECK_NEXT); 7623 subl(cnt2, stride); 7624 jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring 7625 addptr(str1, 16); 7626 if (ae == StrIntrinsicNode::UL) { 7627 addptr(str2, 8); 7628 } else { 7629 addptr(str2, 16); 7630 } 7631 subl(cnt1, stride); 7632 cmpl(cnt2, stride); // Do not read beyond substring 7633 jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR); 7634 // Back-up strings to avoid reading beyond substring. 7635 7636 if (ae == StrIntrinsicNode::UL) { 7637 lea(str2, Address(str2, cnt2, scale2, -8)); 7638 lea(str1, Address(str1, cnt2, scale1, -16)); 7639 } else { 7640 lea(str2, Address(str2, cnt2, scale2, -16)); 7641 lea(str1, Address(str1, cnt2, scale1, -16)); 7642 } 7643 subl(cnt1, cnt2); 7644 movl(cnt2, stride); 7645 addl(cnt1, stride); 7646 bind(CONT_SCAN_SUBSTR); 7647 if (ae == StrIntrinsicNode::UL) { 7648 pmovzxbw(vec, Address(str2, 0)); 7649 } else { 7650 movdqu(vec, Address(str2, 0)); 7651 } 7652 jmpb(SCAN_SUBSTR); 7653 7654 bind(RET_FOUND_LONG); 7655 movptr(str1, Address(rsp, wordSize)); 7656 } // non constant 7657 7658 bind(RET_FOUND); 7659 // Compute substr offset 7660 subptr(result, str1); 7661 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 7662 shrl(result, 1); // index 7663 } 7664 bind(CLEANUP); 7665 pop(rsp); // restore SP 7666 7667 } // string_indexof 7668 7669 void MacroAssembler::string_indexof_char(Register str1, Register cnt1, Register ch, Register result, 7670 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp) { 7671 ShortBranchVerifier sbv(this); 7672 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 7673 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 7674 7675 int stride = 8; 7676 7677 Label FOUND_CHAR, SCAN_TO_CHAR, SCAN_TO_CHAR_LOOP, 7678 SCAN_TO_8_CHAR, SCAN_TO_8_CHAR_LOOP, SCAN_TO_16_CHAR_LOOP, 7679 RET_NOT_FOUND, SCAN_TO_8_CHAR_INIT, 7680 FOUND_SEQ_CHAR, DONE_LABEL; 7681 7682 movptr(result, str1); 7683 if (UseAVX >= 2) { 7684 cmpl(cnt1, stride); 7685 jccb(Assembler::less, SCAN_TO_CHAR_LOOP); 7686 cmpl(cnt1, 2*stride); 7687 jccb(Assembler::less, SCAN_TO_8_CHAR_INIT); 7688 movdl(vec1, ch); 7689 vpbroadcastw(vec1, vec1); 7690 vpxor(vec2, vec2); 7691 movl(tmp, cnt1); 7692 andl(tmp, 0xFFFFFFF0); //vector count (in chars) 7693 andl(cnt1,0x0000000F); //tail count (in chars) 7694 7695 bind(SCAN_TO_16_CHAR_LOOP); 7696 vmovdqu(vec3, Address(result, 0)); 7697 vpcmpeqw(vec3, vec3, vec1, 1); 7698 vptest(vec2, vec3); 7699 jcc(Assembler::carryClear, FOUND_CHAR); 7700 addptr(result, 32); 7701 subl(tmp, 2*stride); 7702 jccb(Assembler::notZero, SCAN_TO_16_CHAR_LOOP); 7703 jmp(SCAN_TO_8_CHAR); 7704 bind(SCAN_TO_8_CHAR_INIT); 7705 movdl(vec1, ch); 7706 pshuflw(vec1, vec1, 0x00); 7707 pshufd(vec1, vec1, 0); 7708 pxor(vec2, vec2); 7709 } 7710 bind(SCAN_TO_8_CHAR); 7711 cmpl(cnt1, stride); 7712 if (UseAVX >= 2) { 7713 jccb(Assembler::less, SCAN_TO_CHAR); 7714 } else { 7715 jccb(Assembler::less, SCAN_TO_CHAR_LOOP); 7716 movdl(vec1, ch); 7717 pshuflw(vec1, vec1, 0x00); 7718 pshufd(vec1, vec1, 0); 7719 pxor(vec2, vec2); 7720 } 7721 movl(tmp, cnt1); 7722 andl(tmp, 0xFFFFFFF8); //vector count (in chars) 7723 andl(cnt1,0x00000007); //tail count (in chars) 7724 7725 bind(SCAN_TO_8_CHAR_LOOP); 7726 movdqu(vec3, Address(result, 0)); 7727 pcmpeqw(vec3, vec1); 7728 ptest(vec2, vec3); 7729 jcc(Assembler::carryClear, FOUND_CHAR); 7730 addptr(result, 16); 7731 subl(tmp, stride); 7732 jccb(Assembler::notZero, SCAN_TO_8_CHAR_LOOP); 7733 bind(SCAN_TO_CHAR); 7734 testl(cnt1, cnt1); 7735 jcc(Assembler::zero, RET_NOT_FOUND); 7736 bind(SCAN_TO_CHAR_LOOP); 7737 load_unsigned_short(tmp, Address(result, 0)); 7738 cmpl(ch, tmp); 7739 jccb(Assembler::equal, FOUND_SEQ_CHAR); 7740 addptr(result, 2); 7741 subl(cnt1, 1); 7742 jccb(Assembler::zero, RET_NOT_FOUND); 7743 jmp(SCAN_TO_CHAR_LOOP); 7744 7745 bind(RET_NOT_FOUND); 7746 movl(result, -1); 7747 jmpb(DONE_LABEL); 7748 7749 bind(FOUND_CHAR); 7750 if (UseAVX >= 2) { 7751 vpmovmskb(tmp, vec3); 7752 } else { 7753 pmovmskb(tmp, vec3); 7754 } 7755 bsfl(ch, tmp); 7756 addl(result, ch); 7757 7758 bind(FOUND_SEQ_CHAR); 7759 subptr(result, str1); 7760 shrl(result, 1); 7761 7762 bind(DONE_LABEL); 7763 } // string_indexof_char 7764 7765 // helper function for string_compare 7766 void MacroAssembler::load_next_elements(Register elem1, Register elem2, Register str1, Register str2, 7767 Address::ScaleFactor scale, Address::ScaleFactor scale1, 7768 Address::ScaleFactor scale2, Register index, int ae) { 7769 if (ae == StrIntrinsicNode::LL) { 7770 load_unsigned_byte(elem1, Address(str1, index, scale, 0)); 7771 load_unsigned_byte(elem2, Address(str2, index, scale, 0)); 7772 } else if (ae == StrIntrinsicNode::UU) { 7773 load_unsigned_short(elem1, Address(str1, index, scale, 0)); 7774 load_unsigned_short(elem2, Address(str2, index, scale, 0)); 7775 } else { 7776 load_unsigned_byte(elem1, Address(str1, index, scale1, 0)); 7777 load_unsigned_short(elem2, Address(str2, index, scale2, 0)); 7778 } 7779 } 7780 7781 // Compare strings, used for char[] and byte[]. 7782 void MacroAssembler::string_compare(Register str1, Register str2, 7783 Register cnt1, Register cnt2, Register result, 7784 XMMRegister vec1, int ae) { 7785 ShortBranchVerifier sbv(this); 7786 Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL; 7787 Label COMPARE_WIDE_VECTORS_LOOP_FAILED; // used only _LP64 && AVX3 7788 int stride, stride2, adr_stride, adr_stride1, adr_stride2; 7789 int stride2x2 = 0x40; 7790 Address::ScaleFactor scale, scale1, scale2; 7791 7792 if (ae != StrIntrinsicNode::LL) { 7793 stride2x2 = 0x20; 7794 } 7795 7796 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) { 7797 shrl(cnt2, 1); 7798 } 7799 // Compute the minimum of the string lengths and the 7800 // difference of the string lengths (stack). 7801 // Do the conditional move stuff 7802 movl(result, cnt1); 7803 subl(cnt1, cnt2); 7804 push(cnt1); 7805 cmov32(Assembler::lessEqual, cnt2, result); // cnt2 = min(cnt1, cnt2) 7806 7807 // Is the minimum length zero? 7808 testl(cnt2, cnt2); 7809 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 7810 if (ae == StrIntrinsicNode::LL) { 7811 // Load first bytes 7812 load_unsigned_byte(result, Address(str1, 0)); // result = str1[0] 7813 load_unsigned_byte(cnt1, Address(str2, 0)); // cnt1 = str2[0] 7814 } else if (ae == StrIntrinsicNode::UU) { 7815 // Load first characters 7816 load_unsigned_short(result, Address(str1, 0)); 7817 load_unsigned_short(cnt1, Address(str2, 0)); 7818 } else { 7819 load_unsigned_byte(result, Address(str1, 0)); 7820 load_unsigned_short(cnt1, Address(str2, 0)); 7821 } 7822 subl(result, cnt1); 7823 jcc(Assembler::notZero, POP_LABEL); 7824 7825 if (ae == StrIntrinsicNode::UU) { 7826 // Divide length by 2 to get number of chars 7827 shrl(cnt2, 1); 7828 } 7829 cmpl(cnt2, 1); 7830 jcc(Assembler::equal, LENGTH_DIFF_LABEL); 7831 7832 // Check if the strings start at the same location and setup scale and stride 7833 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7834 cmpptr(str1, str2); 7835 jcc(Assembler::equal, LENGTH_DIFF_LABEL); 7836 if (ae == StrIntrinsicNode::LL) { 7837 scale = Address::times_1; 7838 stride = 16; 7839 } else { 7840 scale = Address::times_2; 7841 stride = 8; 7842 } 7843 } else { 7844 scale = Address::no_scale; // not used 7845 scale1 = Address::times_1; 7846 scale2 = Address::times_2; 7847 stride = 8; 7848 } 7849 7850 if (UseAVX >= 2 && UseSSE42Intrinsics) { 7851 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 7852 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR; 7853 Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR; 7854 Label COMPARE_WIDE_VECTORS_LOOP_AVX2; 7855 Label COMPARE_TAIL_LONG; 7856 Label COMPARE_WIDE_VECTORS_LOOP_AVX3; // used only _LP64 && AVX3 7857 7858 int pcmpmask = 0x19; 7859 if (ae == StrIntrinsicNode::LL) { 7860 pcmpmask &= ~0x01; 7861 } 7862 7863 // Setup to compare 16-chars (32-bytes) vectors, 7864 // start from first character again because it has aligned address. 7865 if (ae == StrIntrinsicNode::LL) { 7866 stride2 = 32; 7867 } else { 7868 stride2 = 16; 7869 } 7870 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7871 adr_stride = stride << scale; 7872 } else { 7873 adr_stride1 = 8; //stride << scale1; 7874 adr_stride2 = 16; //stride << scale2; 7875 } 7876 7877 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri"); 7878 // rax and rdx are used by pcmpestri as elements counters 7879 movl(result, cnt2); 7880 andl(cnt2, ~(stride2-1)); // cnt2 holds the vector count 7881 jcc(Assembler::zero, COMPARE_TAIL_LONG); 7882 7883 // fast path : compare first 2 8-char vectors. 7884 bind(COMPARE_16_CHARS); 7885 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7886 movdqu(vec1, Address(str1, 0)); 7887 } else { 7888 pmovzxbw(vec1, Address(str1, 0)); 7889 } 7890 pcmpestri(vec1, Address(str2, 0), pcmpmask); 7891 jccb(Assembler::below, COMPARE_INDEX_CHAR); 7892 7893 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7894 movdqu(vec1, Address(str1, adr_stride)); 7895 pcmpestri(vec1, Address(str2, adr_stride), pcmpmask); 7896 } else { 7897 pmovzxbw(vec1, Address(str1, adr_stride1)); 7898 pcmpestri(vec1, Address(str2, adr_stride2), pcmpmask); 7899 } 7900 jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS); 7901 addl(cnt1, stride); 7902 7903 // Compare the characters at index in cnt1 7904 bind(COMPARE_INDEX_CHAR); // cnt1 has the offset of the mismatching character 7905 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae); 7906 subl(result, cnt2); 7907 jmp(POP_LABEL); 7908 7909 // Setup the registers to start vector comparison loop 7910 bind(COMPARE_WIDE_VECTORS); 7911 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7912 lea(str1, Address(str1, result, scale)); 7913 lea(str2, Address(str2, result, scale)); 7914 } else { 7915 lea(str1, Address(str1, result, scale1)); 7916 lea(str2, Address(str2, result, scale2)); 7917 } 7918 subl(result, stride2); 7919 subl(cnt2, stride2); 7920 jcc(Assembler::zero, COMPARE_WIDE_TAIL); 7921 negptr(result); 7922 7923 // In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest) 7924 bind(COMPARE_WIDE_VECTORS_LOOP); 7925 7926 #ifdef _LP64 7927 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop 7928 cmpl(cnt2, stride2x2); 7929 jccb(Assembler::below, COMPARE_WIDE_VECTORS_LOOP_AVX2); 7930 testl(cnt2, stride2x2-1); // cnt2 holds the vector count 7931 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX2); // means we cannot subtract by 0x40 7932 7933 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop 7934 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7935 evmovdquq(vec1, Address(str1, result, scale), Assembler::AVX_512bit); 7936 evpcmpeqb(k7, vec1, Address(str2, result, scale), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0 7937 } else { 7938 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_512bit); 7939 evpcmpeqb(k7, vec1, Address(str2, result, scale2), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0 7940 } 7941 kortestql(k7, k7); 7942 jcc(Assembler::aboveEqual, COMPARE_WIDE_VECTORS_LOOP_FAILED); // miscompare 7943 addptr(result, stride2x2); // update since we already compared at this addr 7944 subl(cnt2, stride2x2); // and sub the size too 7945 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX3); 7946 7947 vpxor(vec1, vec1); 7948 jmpb(COMPARE_WIDE_TAIL); 7949 }//if (VM_Version::supports_avx512vlbw()) 7950 #endif // _LP64 7951 7952 7953 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2); 7954 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7955 vmovdqu(vec1, Address(str1, result, scale)); 7956 vpxor(vec1, Address(str2, result, scale)); 7957 } else { 7958 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_256bit); 7959 vpxor(vec1, Address(str2, result, scale2)); 7960 } 7961 vptest(vec1, vec1); 7962 jcc(Assembler::notZero, VECTOR_NOT_EQUAL); 7963 addptr(result, stride2); 7964 subl(cnt2, stride2); 7965 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP); 7966 // clean upper bits of YMM registers 7967 vpxor(vec1, vec1); 7968 7969 // compare wide vectors tail 7970 bind(COMPARE_WIDE_TAIL); 7971 testptr(result, result); 7972 jccb(Assembler::zero, LENGTH_DIFF_LABEL); 7973 7974 movl(result, stride2); 7975 movl(cnt2, result); 7976 negptr(result); 7977 jmp(COMPARE_WIDE_VECTORS_LOOP_AVX2); 7978 7979 // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors. 7980 bind(VECTOR_NOT_EQUAL); 7981 // clean upper bits of YMM registers 7982 vpxor(vec1, vec1); 7983 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7984 lea(str1, Address(str1, result, scale)); 7985 lea(str2, Address(str2, result, scale)); 7986 } else { 7987 lea(str1, Address(str1, result, scale1)); 7988 lea(str2, Address(str2, result, scale2)); 7989 } 7990 jmp(COMPARE_16_CHARS); 7991 7992 // Compare tail chars, length between 1 to 15 chars 7993 bind(COMPARE_TAIL_LONG); 7994 movl(cnt2, result); 7995 cmpl(cnt2, stride); 7996 jccb(Assembler::less, COMPARE_SMALL_STR); 7997 7998 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7999 movdqu(vec1, Address(str1, 0)); 8000 } else { 8001 pmovzxbw(vec1, Address(str1, 0)); 8002 } 8003 pcmpestri(vec1, Address(str2, 0), pcmpmask); 8004 jcc(Assembler::below, COMPARE_INDEX_CHAR); 8005 subptr(cnt2, stride); 8006 jccb(Assembler::zero, LENGTH_DIFF_LABEL); 8007 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8008 lea(str1, Address(str1, result, scale)); 8009 lea(str2, Address(str2, result, scale)); 8010 } else { 8011 lea(str1, Address(str1, result, scale1)); 8012 lea(str2, Address(str2, result, scale2)); 8013 } 8014 negptr(cnt2); 8015 jmpb(WHILE_HEAD_LABEL); 8016 8017 bind(COMPARE_SMALL_STR); 8018 } else if (UseSSE42Intrinsics) { 8019 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 8020 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL; 8021 int pcmpmask = 0x19; 8022 // Setup to compare 8-char (16-byte) vectors, 8023 // start from first character again because it has aligned address. 8024 movl(result, cnt2); 8025 andl(cnt2, ~(stride - 1)); // cnt2 holds the vector count 8026 if (ae == StrIntrinsicNode::LL) { 8027 pcmpmask &= ~0x01; 8028 } 8029 jccb(Assembler::zero, COMPARE_TAIL); 8030 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8031 lea(str1, Address(str1, result, scale)); 8032 lea(str2, Address(str2, result, scale)); 8033 } else { 8034 lea(str1, Address(str1, result, scale1)); 8035 lea(str2, Address(str2, result, scale2)); 8036 } 8037 negptr(result); 8038 8039 // pcmpestri 8040 // inputs: 8041 // vec1- substring 8042 // rax - negative string length (elements count) 8043 // mem - scanned string 8044 // rdx - string length (elements count) 8045 // pcmpmask - cmp mode: 11000 (string compare with negated result) 8046 // + 00 (unsigned bytes) or + 01 (unsigned shorts) 8047 // outputs: 8048 // rcx - first mismatched element index 8049 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri"); 8050 8051 bind(COMPARE_WIDE_VECTORS); 8052 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8053 movdqu(vec1, Address(str1, result, scale)); 8054 pcmpestri(vec1, Address(str2, result, scale), pcmpmask); 8055 } else { 8056 pmovzxbw(vec1, Address(str1, result, scale1)); 8057 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask); 8058 } 8059 // After pcmpestri cnt1(rcx) contains mismatched element index 8060 8061 jccb(Assembler::below, VECTOR_NOT_EQUAL); // CF==1 8062 addptr(result, stride); 8063 subptr(cnt2, stride); 8064 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS); 8065 8066 // compare wide vectors tail 8067 testptr(result, result); 8068 jccb(Assembler::zero, LENGTH_DIFF_LABEL); 8069 8070 movl(cnt2, stride); 8071 movl(result, stride); 8072 negptr(result); 8073 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8074 movdqu(vec1, Address(str1, result, scale)); 8075 pcmpestri(vec1, Address(str2, result, scale), pcmpmask); 8076 } else { 8077 pmovzxbw(vec1, Address(str1, result, scale1)); 8078 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask); 8079 } 8080 jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL); 8081 8082 // Mismatched characters in the vectors 8083 bind(VECTOR_NOT_EQUAL); 8084 addptr(cnt1, result); 8085 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae); 8086 subl(result, cnt2); 8087 jmpb(POP_LABEL); 8088 8089 bind(COMPARE_TAIL); // limit is zero 8090 movl(cnt2, result); 8091 // Fallthru to tail compare 8092 } 8093 // Shift str2 and str1 to the end of the arrays, negate min 8094 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 8095 lea(str1, Address(str1, cnt2, scale)); 8096 lea(str2, Address(str2, cnt2, scale)); 8097 } else { 8098 lea(str1, Address(str1, cnt2, scale1)); 8099 lea(str2, Address(str2, cnt2, scale2)); 8100 } 8101 decrementl(cnt2); // first character was compared already 8102 negptr(cnt2); 8103 8104 // Compare the rest of the elements 8105 bind(WHILE_HEAD_LABEL); 8106 load_next_elements(result, cnt1, str1, str2, scale, scale1, scale2, cnt2, ae); 8107 subl(result, cnt1); 8108 jccb(Assembler::notZero, POP_LABEL); 8109 increment(cnt2); 8110 jccb(Assembler::notZero, WHILE_HEAD_LABEL); 8111 8112 // Strings are equal up to min length. Return the length difference. 8113 bind(LENGTH_DIFF_LABEL); 8114 pop(result); 8115 if (ae == StrIntrinsicNode::UU) { 8116 // Divide diff by 2 to get number of chars 8117 sarl(result, 1); 8118 } 8119 jmpb(DONE_LABEL); 8120 8121 #ifdef _LP64 8122 if (VM_Version::supports_avx512vlbw()) { 8123 8124 bind(COMPARE_WIDE_VECTORS_LOOP_FAILED); 8125 8126 kmovql(cnt1, k7); 8127 notq(cnt1); 8128 bsfq(cnt2, cnt1); 8129 if (ae != StrIntrinsicNode::LL) { 8130 // Divide diff by 2 to get number of chars 8131 sarl(cnt2, 1); 8132 } 8133 addq(result, cnt2); 8134 if (ae == StrIntrinsicNode::LL) { 8135 load_unsigned_byte(cnt1, Address(str2, result)); 8136 load_unsigned_byte(result, Address(str1, result)); 8137 } else if (ae == StrIntrinsicNode::UU) { 8138 load_unsigned_short(cnt1, Address(str2, result, scale)); 8139 load_unsigned_short(result, Address(str1, result, scale)); 8140 } else { 8141 load_unsigned_short(cnt1, Address(str2, result, scale2)); 8142 load_unsigned_byte(result, Address(str1, result, scale1)); 8143 } 8144 subl(result, cnt1); 8145 jmpb(POP_LABEL); 8146 }//if (VM_Version::supports_avx512vlbw()) 8147 #endif // _LP64 8148 8149 // Discard the stored length difference 8150 bind(POP_LABEL); 8151 pop(cnt1); 8152 8153 // That's it 8154 bind(DONE_LABEL); 8155 if(ae == StrIntrinsicNode::UL) { 8156 negl(result); 8157 } 8158 8159 } 8160 8161 // Search for Non-ASCII character (Negative byte value) in a byte array, 8162 // return true if it has any and false otherwise. 8163 void MacroAssembler::has_negatives(Register ary1, Register len, 8164 Register result, Register tmp1, 8165 XMMRegister vec1, XMMRegister vec2) { 8166 8167 // rsi: byte array 8168 // rcx: len 8169 // rax: result 8170 ShortBranchVerifier sbv(this); 8171 assert_different_registers(ary1, len, result, tmp1); 8172 assert_different_registers(vec1, vec2); 8173 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_CHAR, COMPARE_VECTORS, COMPARE_BYTE; 8174 8175 // len == 0 8176 testl(len, len); 8177 jcc(Assembler::zero, FALSE_LABEL); 8178 8179 movl(result, len); // copy 8180 8181 if (UseAVX >= 2 && UseSSE >= 2) { 8182 // With AVX2, use 32-byte vector compare 8183 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8184 8185 // Compare 32-byte vectors 8186 andl(result, 0x0000001f); // tail count (in bytes) 8187 andl(len, 0xffffffe0); // vector count (in bytes) 8188 jccb(Assembler::zero, COMPARE_TAIL); 8189 8190 lea(ary1, Address(ary1, len, Address::times_1)); 8191 negptr(len); 8192 8193 movl(tmp1, 0x80808080); // create mask to test for Unicode chars in vector 8194 movdl(vec2, tmp1); 8195 vpbroadcastd(vec2, vec2); 8196 8197 bind(COMPARE_WIDE_VECTORS); 8198 vmovdqu(vec1, Address(ary1, len, Address::times_1)); 8199 vptest(vec1, vec2); 8200 jccb(Assembler::notZero, TRUE_LABEL); 8201 addptr(len, 32); 8202 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8203 8204 testl(result, result); 8205 jccb(Assembler::zero, FALSE_LABEL); 8206 8207 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32)); 8208 vptest(vec1, vec2); 8209 jccb(Assembler::notZero, TRUE_LABEL); 8210 jmpb(FALSE_LABEL); 8211 8212 bind(COMPARE_TAIL); // len is zero 8213 movl(len, result); 8214 // Fallthru to tail compare 8215 } else if (UseSSE42Intrinsics) { 8216 assert(UseSSE >= 4, "SSE4 must be for SSE4.2 intrinsics to be available"); 8217 // With SSE4.2, use double quad vector compare 8218 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8219 8220 // Compare 16-byte vectors 8221 andl(result, 0x0000000f); // tail count (in bytes) 8222 andl(len, 0xfffffff0); // vector count (in bytes) 8223 jccb(Assembler::zero, COMPARE_TAIL); 8224 8225 lea(ary1, Address(ary1, len, Address::times_1)); 8226 negptr(len); 8227 8228 movl(tmp1, 0x80808080); 8229 movdl(vec2, tmp1); 8230 pshufd(vec2, vec2, 0); 8231 8232 bind(COMPARE_WIDE_VECTORS); 8233 movdqu(vec1, Address(ary1, len, Address::times_1)); 8234 ptest(vec1, vec2); 8235 jccb(Assembler::notZero, TRUE_LABEL); 8236 addptr(len, 16); 8237 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8238 8239 testl(result, result); 8240 jccb(Assembler::zero, FALSE_LABEL); 8241 8242 movdqu(vec1, Address(ary1, result, Address::times_1, -16)); 8243 ptest(vec1, vec2); 8244 jccb(Assembler::notZero, TRUE_LABEL); 8245 jmpb(FALSE_LABEL); 8246 8247 bind(COMPARE_TAIL); // len is zero 8248 movl(len, result); 8249 // Fallthru to tail compare 8250 } 8251 8252 // Compare 4-byte vectors 8253 andl(len, 0xfffffffc); // vector count (in bytes) 8254 jccb(Assembler::zero, COMPARE_CHAR); 8255 8256 lea(ary1, Address(ary1, len, Address::times_1)); 8257 negptr(len); 8258 8259 bind(COMPARE_VECTORS); 8260 movl(tmp1, Address(ary1, len, Address::times_1)); 8261 andl(tmp1, 0x80808080); 8262 jccb(Assembler::notZero, TRUE_LABEL); 8263 addptr(len, 4); 8264 jcc(Assembler::notZero, COMPARE_VECTORS); 8265 8266 // Compare trailing char (final 2 bytes), if any 8267 bind(COMPARE_CHAR); 8268 testl(result, 0x2); // tail char 8269 jccb(Assembler::zero, COMPARE_BYTE); 8270 load_unsigned_short(tmp1, Address(ary1, 0)); 8271 andl(tmp1, 0x00008080); 8272 jccb(Assembler::notZero, TRUE_LABEL); 8273 subptr(result, 2); 8274 lea(ary1, Address(ary1, 2)); 8275 8276 bind(COMPARE_BYTE); 8277 testl(result, 0x1); // tail byte 8278 jccb(Assembler::zero, FALSE_LABEL); 8279 load_unsigned_byte(tmp1, Address(ary1, 0)); 8280 andl(tmp1, 0x00000080); 8281 jccb(Assembler::notEqual, TRUE_LABEL); 8282 jmpb(FALSE_LABEL); 8283 8284 bind(TRUE_LABEL); 8285 movl(result, 1); // return true 8286 jmpb(DONE); 8287 8288 bind(FALSE_LABEL); 8289 xorl(result, result); // return false 8290 8291 // That's it 8292 bind(DONE); 8293 if (UseAVX >= 2 && UseSSE >= 2) { 8294 // clean upper bits of YMM registers 8295 vpxor(vec1, vec1); 8296 vpxor(vec2, vec2); 8297 } 8298 } 8299 8300 // Compare char[] or byte[] arrays aligned to 4 bytes or substrings. 8301 void MacroAssembler::arrays_equals(bool is_array_equ, Register ary1, Register ary2, 8302 Register limit, Register result, Register chr, 8303 XMMRegister vec1, XMMRegister vec2, bool is_char) { 8304 ShortBranchVerifier sbv(this); 8305 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR, COMPARE_BYTE; 8306 8307 int length_offset = arrayOopDesc::length_offset_in_bytes(); 8308 int base_offset = arrayOopDesc::base_offset_in_bytes(is_char ? T_CHAR : T_BYTE); 8309 8310 if (is_array_equ) { 8311 // Check the input args 8312 cmpptr(ary1, ary2); 8313 jcc(Assembler::equal, TRUE_LABEL); 8314 8315 // Need additional checks for arrays_equals. 8316 testptr(ary1, ary1); 8317 jcc(Assembler::zero, FALSE_LABEL); 8318 testptr(ary2, ary2); 8319 jcc(Assembler::zero, FALSE_LABEL); 8320 8321 // Check the lengths 8322 movl(limit, Address(ary1, length_offset)); 8323 cmpl(limit, Address(ary2, length_offset)); 8324 jcc(Assembler::notEqual, FALSE_LABEL); 8325 } 8326 8327 // count == 0 8328 testl(limit, limit); 8329 jcc(Assembler::zero, TRUE_LABEL); 8330 8331 if (is_array_equ) { 8332 // Load array address 8333 lea(ary1, Address(ary1, base_offset)); 8334 lea(ary2, Address(ary2, base_offset)); 8335 } 8336 8337 if (is_array_equ && is_char) { 8338 // arrays_equals when used for char[]. 8339 shll(limit, 1); // byte count != 0 8340 } 8341 movl(result, limit); // copy 8342 8343 if (UseAVX >= 2) { 8344 // With AVX2, use 32-byte vector compare 8345 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8346 8347 // Compare 32-byte vectors 8348 andl(result, 0x0000001f); // tail count (in bytes) 8349 andl(limit, 0xffffffe0); // vector count (in bytes) 8350 jcc(Assembler::zero, COMPARE_TAIL); 8351 8352 lea(ary1, Address(ary1, limit, Address::times_1)); 8353 lea(ary2, Address(ary2, limit, Address::times_1)); 8354 negptr(limit); 8355 8356 bind(COMPARE_WIDE_VECTORS); 8357 8358 #ifdef _LP64 8359 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop 8360 Label COMPARE_WIDE_VECTORS_LOOP_AVX2, COMPARE_WIDE_VECTORS_LOOP_AVX3; 8361 8362 cmpl(limit, -64); 8363 jccb(Assembler::greater, COMPARE_WIDE_VECTORS_LOOP_AVX2); 8364 8365 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop 8366 8367 evmovdquq(vec1, Address(ary1, limit, Address::times_1), Assembler::AVX_512bit); 8368 evpcmpeqb(k7, vec1, Address(ary2, limit, Address::times_1), Assembler::AVX_512bit); 8369 kortestql(k7, k7); 8370 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare 8371 addptr(limit, 64); // update since we already compared at this addr 8372 cmpl(limit, -64); 8373 jccb(Assembler::lessEqual, COMPARE_WIDE_VECTORS_LOOP_AVX3); 8374 8375 // At this point we may still need to compare -limit+result bytes. 8376 // We could execute the next two instruction and just continue via non-wide path: 8377 // cmpl(limit, 0); 8378 // jcc(Assembler::equal, COMPARE_TAIL); // true 8379 // But since we stopped at the points ary{1,2}+limit which are 8380 // not farther than 64 bytes from the ends of arrays ary{1,2}+result 8381 // (|limit| <= 32 and result < 32), 8382 // we may just compare the last 64 bytes. 8383 // 8384 addptr(result, -64); // it is safe, bc we just came from this area 8385 evmovdquq(vec1, Address(ary1, result, Address::times_1), Assembler::AVX_512bit); 8386 evpcmpeqb(k7, vec1, Address(ary2, result, Address::times_1), Assembler::AVX_512bit); 8387 kortestql(k7, k7); 8388 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare 8389 8390 jmp(TRUE_LABEL); 8391 8392 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2); 8393 8394 }//if (VM_Version::supports_avx512vlbw()) 8395 #endif //_LP64 8396 8397 vmovdqu(vec1, Address(ary1, limit, Address::times_1)); 8398 vmovdqu(vec2, Address(ary2, limit, Address::times_1)); 8399 vpxor(vec1, vec2); 8400 8401 vptest(vec1, vec1); 8402 jccb(Assembler::notZero, FALSE_LABEL); 8403 addptr(limit, 32); 8404 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8405 8406 testl(result, result); 8407 jccb(Assembler::zero, TRUE_LABEL); 8408 8409 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32)); 8410 vmovdqu(vec2, Address(ary2, result, Address::times_1, -32)); 8411 vpxor(vec1, vec2); 8412 8413 vptest(vec1, vec1); 8414 jccb(Assembler::notZero, FALSE_LABEL); 8415 jmpb(TRUE_LABEL); 8416 8417 bind(COMPARE_TAIL); // limit is zero 8418 movl(limit, result); 8419 // Fallthru to tail compare 8420 } else if (UseSSE42Intrinsics) { 8421 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 8422 // With SSE4.2, use double quad vector compare 8423 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 8424 8425 // Compare 16-byte vectors 8426 andl(result, 0x0000000f); // tail count (in bytes) 8427 andl(limit, 0xfffffff0); // vector count (in bytes) 8428 jccb(Assembler::zero, COMPARE_TAIL); 8429 8430 lea(ary1, Address(ary1, limit, Address::times_1)); 8431 lea(ary2, Address(ary2, limit, Address::times_1)); 8432 negptr(limit); 8433 8434 bind(COMPARE_WIDE_VECTORS); 8435 movdqu(vec1, Address(ary1, limit, Address::times_1)); 8436 movdqu(vec2, Address(ary2, limit, Address::times_1)); 8437 pxor(vec1, vec2); 8438 8439 ptest(vec1, vec1); 8440 jccb(Assembler::notZero, FALSE_LABEL); 8441 addptr(limit, 16); 8442 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 8443 8444 testl(result, result); 8445 jccb(Assembler::zero, TRUE_LABEL); 8446 8447 movdqu(vec1, Address(ary1, result, Address::times_1, -16)); 8448 movdqu(vec2, Address(ary2, result, Address::times_1, -16)); 8449 pxor(vec1, vec2); 8450 8451 ptest(vec1, vec1); 8452 jccb(Assembler::notZero, FALSE_LABEL); 8453 jmpb(TRUE_LABEL); 8454 8455 bind(COMPARE_TAIL); // limit is zero 8456 movl(limit, result); 8457 // Fallthru to tail compare 8458 } 8459 8460 // Compare 4-byte vectors 8461 andl(limit, 0xfffffffc); // vector count (in bytes) 8462 jccb(Assembler::zero, COMPARE_CHAR); 8463 8464 lea(ary1, Address(ary1, limit, Address::times_1)); 8465 lea(ary2, Address(ary2, limit, Address::times_1)); 8466 negptr(limit); 8467 8468 bind(COMPARE_VECTORS); 8469 movl(chr, Address(ary1, limit, Address::times_1)); 8470 cmpl(chr, Address(ary2, limit, Address::times_1)); 8471 jccb(Assembler::notEqual, FALSE_LABEL); 8472 addptr(limit, 4); 8473 jcc(Assembler::notZero, COMPARE_VECTORS); 8474 8475 // Compare trailing char (final 2 bytes), if any 8476 bind(COMPARE_CHAR); 8477 testl(result, 0x2); // tail char 8478 jccb(Assembler::zero, COMPARE_BYTE); 8479 load_unsigned_short(chr, Address(ary1, 0)); 8480 load_unsigned_short(limit, Address(ary2, 0)); 8481 cmpl(chr, limit); 8482 jccb(Assembler::notEqual, FALSE_LABEL); 8483 8484 if (is_array_equ && is_char) { 8485 bind(COMPARE_BYTE); 8486 } else { 8487 lea(ary1, Address(ary1, 2)); 8488 lea(ary2, Address(ary2, 2)); 8489 8490 bind(COMPARE_BYTE); 8491 testl(result, 0x1); // tail byte 8492 jccb(Assembler::zero, TRUE_LABEL); 8493 load_unsigned_byte(chr, Address(ary1, 0)); 8494 load_unsigned_byte(limit, Address(ary2, 0)); 8495 cmpl(chr, limit); 8496 jccb(Assembler::notEqual, FALSE_LABEL); 8497 } 8498 bind(TRUE_LABEL); 8499 movl(result, 1); // return true 8500 jmpb(DONE); 8501 8502 bind(FALSE_LABEL); 8503 xorl(result, result); // return false 8504 8505 // That's it 8506 bind(DONE); 8507 if (UseAVX >= 2) { 8508 // clean upper bits of YMM registers 8509 vpxor(vec1, vec1); 8510 vpxor(vec2, vec2); 8511 } 8512 } 8513 8514 #endif 8515 8516 void MacroAssembler::generate_fill(BasicType t, bool aligned, 8517 Register to, Register value, Register count, 8518 Register rtmp, XMMRegister xtmp) { 8519 ShortBranchVerifier sbv(this); 8520 assert_different_registers(to, value, count, rtmp); 8521 Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte; 8522 Label L_fill_2_bytes, L_fill_4_bytes; 8523 8524 int shift = -1; 8525 switch (t) { 8526 case T_BYTE: 8527 shift = 2; 8528 break; 8529 case T_SHORT: 8530 shift = 1; 8531 break; 8532 case T_INT: 8533 shift = 0; 8534 break; 8535 default: ShouldNotReachHere(); 8536 } 8537 8538 if (t == T_BYTE) { 8539 andl(value, 0xff); 8540 movl(rtmp, value); 8541 shll(rtmp, 8); 8542 orl(value, rtmp); 8543 } 8544 if (t == T_SHORT) { 8545 andl(value, 0xffff); 8546 } 8547 if (t == T_BYTE || t == T_SHORT) { 8548 movl(rtmp, value); 8549 shll(rtmp, 16); 8550 orl(value, rtmp); 8551 } 8552 8553 cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element 8554 jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp 8555 if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) { 8556 // align source address at 4 bytes address boundary 8557 if (t == T_BYTE) { 8558 // One byte misalignment happens only for byte arrays 8559 testptr(to, 1); 8560 jccb(Assembler::zero, L_skip_align1); 8561 movb(Address(to, 0), value); 8562 increment(to); 8563 decrement(count); 8564 BIND(L_skip_align1); 8565 } 8566 // Two bytes misalignment happens only for byte and short (char) arrays 8567 testptr(to, 2); 8568 jccb(Assembler::zero, L_skip_align2); 8569 movw(Address(to, 0), value); 8570 addptr(to, 2); 8571 subl(count, 1<<(shift-1)); 8572 BIND(L_skip_align2); 8573 } 8574 if (UseSSE < 2) { 8575 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes; 8576 // Fill 32-byte chunks 8577 subl(count, 8 << shift); 8578 jcc(Assembler::less, L_check_fill_8_bytes); 8579 align(16); 8580 8581 BIND(L_fill_32_bytes_loop); 8582 8583 for (int i = 0; i < 32; i += 4) { 8584 movl(Address(to, i), value); 8585 } 8586 8587 addptr(to, 32); 8588 subl(count, 8 << shift); 8589 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop); 8590 BIND(L_check_fill_8_bytes); 8591 addl(count, 8 << shift); 8592 jccb(Assembler::zero, L_exit); 8593 jmpb(L_fill_8_bytes); 8594 8595 // 8596 // length is too short, just fill qwords 8597 // 8598 BIND(L_fill_8_bytes_loop); 8599 movl(Address(to, 0), value); 8600 movl(Address(to, 4), value); 8601 addptr(to, 8); 8602 BIND(L_fill_8_bytes); 8603 subl(count, 1 << (shift + 1)); 8604 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop); 8605 // fall through to fill 4 bytes 8606 } else { 8607 Label L_fill_32_bytes; 8608 if (!UseUnalignedLoadStores) { 8609 // align to 8 bytes, we know we are 4 byte aligned to start 8610 testptr(to, 4); 8611 jccb(Assembler::zero, L_fill_32_bytes); 8612 movl(Address(to, 0), value); 8613 addptr(to, 4); 8614 subl(count, 1<<shift); 8615 } 8616 BIND(L_fill_32_bytes); 8617 { 8618 assert( UseSSE >= 2, "supported cpu only" ); 8619 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes; 8620 if (UseAVX > 2) { 8621 movl(rtmp, 0xffff); 8622 kmovwl(k1, rtmp); 8623 } 8624 movdl(xtmp, value); 8625 if (UseAVX > 2 && UseUnalignedLoadStores) { 8626 // Fill 64-byte chunks 8627 Label L_fill_64_bytes_loop, L_check_fill_32_bytes; 8628 evpbroadcastd(xtmp, xtmp, Assembler::AVX_512bit); 8629 8630 subl(count, 16 << shift); 8631 jcc(Assembler::less, L_check_fill_32_bytes); 8632 align(16); 8633 8634 BIND(L_fill_64_bytes_loop); 8635 evmovdqul(Address(to, 0), xtmp, Assembler::AVX_512bit); 8636 addptr(to, 64); 8637 subl(count, 16 << shift); 8638 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop); 8639 8640 BIND(L_check_fill_32_bytes); 8641 addl(count, 8 << shift); 8642 jccb(Assembler::less, L_check_fill_8_bytes); 8643 vmovdqu(Address(to, 0), xtmp); 8644 addptr(to, 32); 8645 subl(count, 8 << shift); 8646 8647 BIND(L_check_fill_8_bytes); 8648 } else if (UseAVX == 2 && UseUnalignedLoadStores) { 8649 // Fill 64-byte chunks 8650 Label L_fill_64_bytes_loop, L_check_fill_32_bytes; 8651 vpbroadcastd(xtmp, xtmp); 8652 8653 subl(count, 16 << shift); 8654 jcc(Assembler::less, L_check_fill_32_bytes); 8655 align(16); 8656 8657 BIND(L_fill_64_bytes_loop); 8658 vmovdqu(Address(to, 0), xtmp); 8659 vmovdqu(Address(to, 32), xtmp); 8660 addptr(to, 64); 8661 subl(count, 16 << shift); 8662 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop); 8663 8664 BIND(L_check_fill_32_bytes); 8665 addl(count, 8 << shift); 8666 jccb(Assembler::less, L_check_fill_8_bytes); 8667 vmovdqu(Address(to, 0), xtmp); 8668 addptr(to, 32); 8669 subl(count, 8 << shift); 8670 8671 BIND(L_check_fill_8_bytes); 8672 // clean upper bits of YMM registers 8673 movdl(xtmp, value); 8674 pshufd(xtmp, xtmp, 0); 8675 } else { 8676 // Fill 32-byte chunks 8677 pshufd(xtmp, xtmp, 0); 8678 8679 subl(count, 8 << shift); 8680 jcc(Assembler::less, L_check_fill_8_bytes); 8681 align(16); 8682 8683 BIND(L_fill_32_bytes_loop); 8684 8685 if (UseUnalignedLoadStores) { 8686 movdqu(Address(to, 0), xtmp); 8687 movdqu(Address(to, 16), xtmp); 8688 } else { 8689 movq(Address(to, 0), xtmp); 8690 movq(Address(to, 8), xtmp); 8691 movq(Address(to, 16), xtmp); 8692 movq(Address(to, 24), xtmp); 8693 } 8694 8695 addptr(to, 32); 8696 subl(count, 8 << shift); 8697 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop); 8698 8699 BIND(L_check_fill_8_bytes); 8700 } 8701 addl(count, 8 << shift); 8702 jccb(Assembler::zero, L_exit); 8703 jmpb(L_fill_8_bytes); 8704 8705 // 8706 // length is too short, just fill qwords 8707 // 8708 BIND(L_fill_8_bytes_loop); 8709 movq(Address(to, 0), xtmp); 8710 addptr(to, 8); 8711 BIND(L_fill_8_bytes); 8712 subl(count, 1 << (shift + 1)); 8713 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop); 8714 } 8715 } 8716 // fill trailing 4 bytes 8717 BIND(L_fill_4_bytes); 8718 testl(count, 1<<shift); 8719 jccb(Assembler::zero, L_fill_2_bytes); 8720 movl(Address(to, 0), value); 8721 if (t == T_BYTE || t == T_SHORT) { 8722 addptr(to, 4); 8723 BIND(L_fill_2_bytes); 8724 // fill trailing 2 bytes 8725 testl(count, 1<<(shift-1)); 8726 jccb(Assembler::zero, L_fill_byte); 8727 movw(Address(to, 0), value); 8728 if (t == T_BYTE) { 8729 addptr(to, 2); 8730 BIND(L_fill_byte); 8731 // fill trailing byte 8732 testl(count, 1); 8733 jccb(Assembler::zero, L_exit); 8734 movb(Address(to, 0), value); 8735 } else { 8736 BIND(L_fill_byte); 8737 } 8738 } else { 8739 BIND(L_fill_2_bytes); 8740 } 8741 BIND(L_exit); 8742 } 8743 8744 // encode char[] to byte[] in ISO_8859_1 8745 void MacroAssembler::encode_iso_array(Register src, Register dst, Register len, 8746 XMMRegister tmp1Reg, XMMRegister tmp2Reg, 8747 XMMRegister tmp3Reg, XMMRegister tmp4Reg, 8748 Register tmp5, Register result) { 8749 // rsi: src 8750 // rdi: dst 8751 // rdx: len 8752 // rcx: tmp5 8753 // rax: result 8754 ShortBranchVerifier sbv(this); 8755 assert_different_registers(src, dst, len, tmp5, result); 8756 Label L_done, L_copy_1_char, L_copy_1_char_exit; 8757 8758 // set result 8759 xorl(result, result); 8760 // check for zero length 8761 testl(len, len); 8762 jcc(Assembler::zero, L_done); 8763 movl(result, len); 8764 8765 // Setup pointers 8766 lea(src, Address(src, len, Address::times_2)); // char[] 8767 lea(dst, Address(dst, len, Address::times_1)); // byte[] 8768 negptr(len); 8769 8770 if (UseSSE42Intrinsics || UseAVX >= 2) { 8771 assert(UseSSE42Intrinsics ? UseSSE >= 4 : true, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 8772 Label L_chars_8_check, L_copy_8_chars, L_copy_8_chars_exit; 8773 Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit; 8774 8775 if (UseAVX >= 2) { 8776 Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit; 8777 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector 8778 movdl(tmp1Reg, tmp5); 8779 vpbroadcastd(tmp1Reg, tmp1Reg); 8780 jmpb(L_chars_32_check); 8781 8782 bind(L_copy_32_chars); 8783 vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64)); 8784 vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32)); 8785 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1); 8786 vptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 8787 jccb(Assembler::notZero, L_copy_32_chars_exit); 8788 vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1); 8789 vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector_len */ 1); 8790 vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg); 8791 8792 bind(L_chars_32_check); 8793 addptr(len, 32); 8794 jccb(Assembler::lessEqual, L_copy_32_chars); 8795 8796 bind(L_copy_32_chars_exit); 8797 subptr(len, 16); 8798 jccb(Assembler::greater, L_copy_16_chars_exit); 8799 8800 } else if (UseSSE42Intrinsics) { 8801 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector 8802 movdl(tmp1Reg, tmp5); 8803 pshufd(tmp1Reg, tmp1Reg, 0); 8804 jmpb(L_chars_16_check); 8805 } 8806 8807 bind(L_copy_16_chars); 8808 if (UseAVX >= 2) { 8809 vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32)); 8810 vptest(tmp2Reg, tmp1Reg); 8811 jccb(Assembler::notZero, L_copy_16_chars_exit); 8812 vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector_len */ 1); 8813 vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector_len */ 1); 8814 } else { 8815 if (UseAVX > 0) { 8816 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32)); 8817 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16)); 8818 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 0); 8819 } else { 8820 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32)); 8821 por(tmp2Reg, tmp3Reg); 8822 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16)); 8823 por(tmp2Reg, tmp4Reg); 8824 } 8825 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 8826 jccb(Assembler::notZero, L_copy_16_chars_exit); 8827 packuswb(tmp3Reg, tmp4Reg); 8828 } 8829 movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg); 8830 8831 bind(L_chars_16_check); 8832 addptr(len, 16); 8833 jccb(Assembler::lessEqual, L_copy_16_chars); 8834 8835 bind(L_copy_16_chars_exit); 8836 if (UseAVX >= 2) { 8837 // clean upper bits of YMM registers 8838 vpxor(tmp2Reg, tmp2Reg); 8839 vpxor(tmp3Reg, tmp3Reg); 8840 vpxor(tmp4Reg, tmp4Reg); 8841 movdl(tmp1Reg, tmp5); 8842 pshufd(tmp1Reg, tmp1Reg, 0); 8843 } 8844 subptr(len, 8); 8845 jccb(Assembler::greater, L_copy_8_chars_exit); 8846 8847 bind(L_copy_8_chars); 8848 movdqu(tmp3Reg, Address(src, len, Address::times_2, -16)); 8849 ptest(tmp3Reg, tmp1Reg); 8850 jccb(Assembler::notZero, L_copy_8_chars_exit); 8851 packuswb(tmp3Reg, tmp1Reg); 8852 movq(Address(dst, len, Address::times_1, -8), tmp3Reg); 8853 addptr(len, 8); 8854 jccb(Assembler::lessEqual, L_copy_8_chars); 8855 8856 bind(L_copy_8_chars_exit); 8857 subptr(len, 8); 8858 jccb(Assembler::zero, L_done); 8859 } 8860 8861 bind(L_copy_1_char); 8862 load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0)); 8863 testl(tmp5, 0xff00); // check if Unicode char 8864 jccb(Assembler::notZero, L_copy_1_char_exit); 8865 movb(Address(dst, len, Address::times_1, 0), tmp5); 8866 addptr(len, 1); 8867 jccb(Assembler::less, L_copy_1_char); 8868 8869 bind(L_copy_1_char_exit); 8870 addptr(result, len); // len is negative count of not processed elements 8871 bind(L_done); 8872 } 8873 8874 #ifdef _LP64 8875 /** 8876 * Helper for multiply_to_len(). 8877 */ 8878 void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo, Register src1, Register src2) { 8879 addq(dest_lo, src1); 8880 adcq(dest_hi, 0); 8881 addq(dest_lo, src2); 8882 adcq(dest_hi, 0); 8883 } 8884 8885 /** 8886 * Multiply 64 bit by 64 bit first loop. 8887 */ 8888 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart, 8889 Register y, Register y_idx, Register z, 8890 Register carry, Register product, 8891 Register idx, Register kdx) { 8892 // 8893 // jlong carry, x[], y[], z[]; 8894 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) { 8895 // huge_128 product = y[idx] * x[xstart] + carry; 8896 // z[kdx] = (jlong)product; 8897 // carry = (jlong)(product >>> 64); 8898 // } 8899 // z[xstart] = carry; 8900 // 8901 8902 Label L_first_loop, L_first_loop_exit; 8903 Label L_one_x, L_one_y, L_multiply; 8904 8905 decrementl(xstart); 8906 jcc(Assembler::negative, L_one_x); 8907 8908 movq(x_xstart, Address(x, xstart, Address::times_4, 0)); 8909 rorq(x_xstart, 32); // convert big-endian to little-endian 8910 8911 bind(L_first_loop); 8912 decrementl(idx); 8913 jcc(Assembler::negative, L_first_loop_exit); 8914 decrementl(idx); 8915 jcc(Assembler::negative, L_one_y); 8916 movq(y_idx, Address(y, idx, Address::times_4, 0)); 8917 rorq(y_idx, 32); // convert big-endian to little-endian 8918 bind(L_multiply); 8919 movq(product, x_xstart); 8920 mulq(y_idx); // product(rax) * y_idx -> rdx:rax 8921 addq(product, carry); 8922 adcq(rdx, 0); 8923 subl(kdx, 2); 8924 movl(Address(z, kdx, Address::times_4, 4), product); 8925 shrq(product, 32); 8926 movl(Address(z, kdx, Address::times_4, 0), product); 8927 movq(carry, rdx); 8928 jmp(L_first_loop); 8929 8930 bind(L_one_y); 8931 movl(y_idx, Address(y, 0)); 8932 jmp(L_multiply); 8933 8934 bind(L_one_x); 8935 movl(x_xstart, Address(x, 0)); 8936 jmp(L_first_loop); 8937 8938 bind(L_first_loop_exit); 8939 } 8940 8941 /** 8942 * Multiply 64 bit by 64 bit and add 128 bit. 8943 */ 8944 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y, Register z, 8945 Register yz_idx, Register idx, 8946 Register carry, Register product, int offset) { 8947 // huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry; 8948 // z[kdx] = (jlong)product; 8949 8950 movq(yz_idx, Address(y, idx, Address::times_4, offset)); 8951 rorq(yz_idx, 32); // convert big-endian to little-endian 8952 movq(product, x_xstart); 8953 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax) 8954 movq(yz_idx, Address(z, idx, Address::times_4, offset)); 8955 rorq(yz_idx, 32); // convert big-endian to little-endian 8956 8957 add2_with_carry(rdx, product, carry, yz_idx); 8958 8959 movl(Address(z, idx, Address::times_4, offset+4), product); 8960 shrq(product, 32); 8961 movl(Address(z, idx, Address::times_4, offset), product); 8962 8963 } 8964 8965 /** 8966 * Multiply 128 bit by 128 bit. Unrolled inner loop. 8967 */ 8968 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart, Register y, Register z, 8969 Register yz_idx, Register idx, Register jdx, 8970 Register carry, Register product, 8971 Register carry2) { 8972 // jlong carry, x[], y[], z[]; 8973 // int kdx = ystart+1; 8974 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop 8975 // huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry; 8976 // z[kdx+idx+1] = (jlong)product; 8977 // jlong carry2 = (jlong)(product >>> 64); 8978 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry2; 8979 // z[kdx+idx] = (jlong)product; 8980 // carry = (jlong)(product >>> 64); 8981 // } 8982 // idx += 2; 8983 // if (idx > 0) { 8984 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry; 8985 // z[kdx+idx] = (jlong)product; 8986 // carry = (jlong)(product >>> 64); 8987 // } 8988 // 8989 8990 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done; 8991 8992 movl(jdx, idx); 8993 andl(jdx, 0xFFFFFFFC); 8994 shrl(jdx, 2); 8995 8996 bind(L_third_loop); 8997 subl(jdx, 1); 8998 jcc(Assembler::negative, L_third_loop_exit); 8999 subl(idx, 4); 9000 9001 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8); 9002 movq(carry2, rdx); 9003 9004 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0); 9005 movq(carry, rdx); 9006 jmp(L_third_loop); 9007 9008 bind (L_third_loop_exit); 9009 9010 andl (idx, 0x3); 9011 jcc(Assembler::zero, L_post_third_loop_done); 9012 9013 Label L_check_1; 9014 subl(idx, 2); 9015 jcc(Assembler::negative, L_check_1); 9016 9017 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0); 9018 movq(carry, rdx); 9019 9020 bind (L_check_1); 9021 addl (idx, 0x2); 9022 andl (idx, 0x1); 9023 subl(idx, 1); 9024 jcc(Assembler::negative, L_post_third_loop_done); 9025 9026 movl(yz_idx, Address(y, idx, Address::times_4, 0)); 9027 movq(product, x_xstart); 9028 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax) 9029 movl(yz_idx, Address(z, idx, Address::times_4, 0)); 9030 9031 add2_with_carry(rdx, product, yz_idx, carry); 9032 9033 movl(Address(z, idx, Address::times_4, 0), product); 9034 shrq(product, 32); 9035 9036 shlq(rdx, 32); 9037 orq(product, rdx); 9038 movq(carry, product); 9039 9040 bind(L_post_third_loop_done); 9041 } 9042 9043 /** 9044 * Multiply 128 bit by 128 bit using BMI2. Unrolled inner loop. 9045 * 9046 */ 9047 void MacroAssembler::multiply_128_x_128_bmi2_loop(Register y, Register z, 9048 Register carry, Register carry2, 9049 Register idx, Register jdx, 9050 Register yz_idx1, Register yz_idx2, 9051 Register tmp, Register tmp3, Register tmp4) { 9052 assert(UseBMI2Instructions, "should be used only when BMI2 is available"); 9053 9054 // jlong carry, x[], y[], z[]; 9055 // int kdx = ystart+1; 9056 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop 9057 // huge_128 tmp3 = (y[idx+1] * rdx) + z[kdx+idx+1] + carry; 9058 // jlong carry2 = (jlong)(tmp3 >>> 64); 9059 // huge_128 tmp4 = (y[idx] * rdx) + z[kdx+idx] + carry2; 9060 // carry = (jlong)(tmp4 >>> 64); 9061 // z[kdx+idx+1] = (jlong)tmp3; 9062 // z[kdx+idx] = (jlong)tmp4; 9063 // } 9064 // idx += 2; 9065 // if (idx > 0) { 9066 // yz_idx1 = (y[idx] * rdx) + z[kdx+idx] + carry; 9067 // z[kdx+idx] = (jlong)yz_idx1; 9068 // carry = (jlong)(yz_idx1 >>> 64); 9069 // } 9070 // 9071 9072 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done; 9073 9074 movl(jdx, idx); 9075 andl(jdx, 0xFFFFFFFC); 9076 shrl(jdx, 2); 9077 9078 bind(L_third_loop); 9079 subl(jdx, 1); 9080 jcc(Assembler::negative, L_third_loop_exit); 9081 subl(idx, 4); 9082 9083 movq(yz_idx1, Address(y, idx, Address::times_4, 8)); 9084 rorxq(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian 9085 movq(yz_idx2, Address(y, idx, Address::times_4, 0)); 9086 rorxq(yz_idx2, yz_idx2, 32); 9087 9088 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3 9089 mulxq(carry2, tmp, yz_idx2); // yz_idx2 * rdx -> carry2:tmp 9090 9091 movq(yz_idx1, Address(z, idx, Address::times_4, 8)); 9092 rorxq(yz_idx1, yz_idx1, 32); 9093 movq(yz_idx2, Address(z, idx, Address::times_4, 0)); 9094 rorxq(yz_idx2, yz_idx2, 32); 9095 9096 if (VM_Version::supports_adx()) { 9097 adcxq(tmp3, carry); 9098 adoxq(tmp3, yz_idx1); 9099 9100 adcxq(tmp4, tmp); 9101 adoxq(tmp4, yz_idx2); 9102 9103 movl(carry, 0); // does not affect flags 9104 adcxq(carry2, carry); 9105 adoxq(carry2, carry); 9106 } else { 9107 add2_with_carry(tmp4, tmp3, carry, yz_idx1); 9108 add2_with_carry(carry2, tmp4, tmp, yz_idx2); 9109 } 9110 movq(carry, carry2); 9111 9112 movl(Address(z, idx, Address::times_4, 12), tmp3); 9113 shrq(tmp3, 32); 9114 movl(Address(z, idx, Address::times_4, 8), tmp3); 9115 9116 movl(Address(z, idx, Address::times_4, 4), tmp4); 9117 shrq(tmp4, 32); 9118 movl(Address(z, idx, Address::times_4, 0), tmp4); 9119 9120 jmp(L_third_loop); 9121 9122 bind (L_third_loop_exit); 9123 9124 andl (idx, 0x3); 9125 jcc(Assembler::zero, L_post_third_loop_done); 9126 9127 Label L_check_1; 9128 subl(idx, 2); 9129 jcc(Assembler::negative, L_check_1); 9130 9131 movq(yz_idx1, Address(y, idx, Address::times_4, 0)); 9132 rorxq(yz_idx1, yz_idx1, 32); 9133 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3 9134 movq(yz_idx2, Address(z, idx, Address::times_4, 0)); 9135 rorxq(yz_idx2, yz_idx2, 32); 9136 9137 add2_with_carry(tmp4, tmp3, carry, yz_idx2); 9138 9139 movl(Address(z, idx, Address::times_4, 4), tmp3); 9140 shrq(tmp3, 32); 9141 movl(Address(z, idx, Address::times_4, 0), tmp3); 9142 movq(carry, tmp4); 9143 9144 bind (L_check_1); 9145 addl (idx, 0x2); 9146 andl (idx, 0x1); 9147 subl(idx, 1); 9148 jcc(Assembler::negative, L_post_third_loop_done); 9149 movl(tmp4, Address(y, idx, Address::times_4, 0)); 9150 mulxq(carry2, tmp3, tmp4); // tmp4 * rdx -> carry2:tmp3 9151 movl(tmp4, Address(z, idx, Address::times_4, 0)); 9152 9153 add2_with_carry(carry2, tmp3, tmp4, carry); 9154 9155 movl(Address(z, idx, Address::times_4, 0), tmp3); 9156 shrq(tmp3, 32); 9157 9158 shlq(carry2, 32); 9159 orq(tmp3, carry2); 9160 movq(carry, tmp3); 9161 9162 bind(L_post_third_loop_done); 9163 } 9164 9165 /** 9166 * Code for BigInteger::multiplyToLen() instrinsic. 9167 * 9168 * rdi: x 9169 * rax: xlen 9170 * rsi: y 9171 * rcx: ylen 9172 * r8: z 9173 * r11: zlen 9174 * r12: tmp1 9175 * r13: tmp2 9176 * r14: tmp3 9177 * r15: tmp4 9178 * rbx: tmp5 9179 * 9180 */ 9181 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, Register z, Register zlen, 9182 Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5) { 9183 ShortBranchVerifier sbv(this); 9184 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx); 9185 9186 push(tmp1); 9187 push(tmp2); 9188 push(tmp3); 9189 push(tmp4); 9190 push(tmp5); 9191 9192 push(xlen); 9193 push(zlen); 9194 9195 const Register idx = tmp1; 9196 const Register kdx = tmp2; 9197 const Register xstart = tmp3; 9198 9199 const Register y_idx = tmp4; 9200 const Register carry = tmp5; 9201 const Register product = xlen; 9202 const Register x_xstart = zlen; // reuse register 9203 9204 // First Loop. 9205 // 9206 // final static long LONG_MASK = 0xffffffffL; 9207 // int xstart = xlen - 1; 9208 // int ystart = ylen - 1; 9209 // long carry = 0; 9210 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) { 9211 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry; 9212 // z[kdx] = (int)product; 9213 // carry = product >>> 32; 9214 // } 9215 // z[xstart] = (int)carry; 9216 // 9217 9218 movl(idx, ylen); // idx = ylen; 9219 movl(kdx, zlen); // kdx = xlen+ylen; 9220 xorq(carry, carry); // carry = 0; 9221 9222 Label L_done; 9223 9224 movl(xstart, xlen); 9225 decrementl(xstart); 9226 jcc(Assembler::negative, L_done); 9227 9228 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx); 9229 9230 Label L_second_loop; 9231 testl(kdx, kdx); 9232 jcc(Assembler::zero, L_second_loop); 9233 9234 Label L_carry; 9235 subl(kdx, 1); 9236 jcc(Assembler::zero, L_carry); 9237 9238 movl(Address(z, kdx, Address::times_4, 0), carry); 9239 shrq(carry, 32); 9240 subl(kdx, 1); 9241 9242 bind(L_carry); 9243 movl(Address(z, kdx, Address::times_4, 0), carry); 9244 9245 // Second and third (nested) loops. 9246 // 9247 // for (int i = xstart-1; i >= 0; i--) { // Second loop 9248 // carry = 0; 9249 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop 9250 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) + 9251 // (z[k] & LONG_MASK) + carry; 9252 // z[k] = (int)product; 9253 // carry = product >>> 32; 9254 // } 9255 // z[i] = (int)carry; 9256 // } 9257 // 9258 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx 9259 9260 const Register jdx = tmp1; 9261 9262 bind(L_second_loop); 9263 xorl(carry, carry); // carry = 0; 9264 movl(jdx, ylen); // j = ystart+1 9265 9266 subl(xstart, 1); // i = xstart-1; 9267 jcc(Assembler::negative, L_done); 9268 9269 push (z); 9270 9271 Label L_last_x; 9272 lea(z, Address(z, xstart, Address::times_4, 4)); // z = z + k - j 9273 subl(xstart, 1); // i = xstart-1; 9274 jcc(Assembler::negative, L_last_x); 9275 9276 if (UseBMI2Instructions) { 9277 movq(rdx, Address(x, xstart, Address::times_4, 0)); 9278 rorxq(rdx, rdx, 32); // convert big-endian to little-endian 9279 } else { 9280 movq(x_xstart, Address(x, xstart, Address::times_4, 0)); 9281 rorq(x_xstart, 32); // convert big-endian to little-endian 9282 } 9283 9284 Label L_third_loop_prologue; 9285 bind(L_third_loop_prologue); 9286 9287 push (x); 9288 push (xstart); 9289 push (ylen); 9290 9291 9292 if (UseBMI2Instructions) { 9293 multiply_128_x_128_bmi2_loop(y, z, carry, x, jdx, ylen, product, tmp2, x_xstart, tmp3, tmp4); 9294 } else { // !UseBMI2Instructions 9295 multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x); 9296 } 9297 9298 pop(ylen); 9299 pop(xlen); 9300 pop(x); 9301 pop(z); 9302 9303 movl(tmp3, xlen); 9304 addl(tmp3, 1); 9305 movl(Address(z, tmp3, Address::times_4, 0), carry); 9306 subl(tmp3, 1); 9307 jccb(Assembler::negative, L_done); 9308 9309 shrq(carry, 32); 9310 movl(Address(z, tmp3, Address::times_4, 0), carry); 9311 jmp(L_second_loop); 9312 9313 // Next infrequent code is moved outside loops. 9314 bind(L_last_x); 9315 if (UseBMI2Instructions) { 9316 movl(rdx, Address(x, 0)); 9317 } else { 9318 movl(x_xstart, Address(x, 0)); 9319 } 9320 jmp(L_third_loop_prologue); 9321 9322 bind(L_done); 9323 9324 pop(zlen); 9325 pop(xlen); 9326 9327 pop(tmp5); 9328 pop(tmp4); 9329 pop(tmp3); 9330 pop(tmp2); 9331 pop(tmp1); 9332 } 9333 9334 void MacroAssembler::vectorized_mismatch(Register obja, Register objb, Register length, Register log2_array_indxscale, 9335 Register result, Register tmp1, Register tmp2, XMMRegister rymm0, XMMRegister rymm1, XMMRegister rymm2){ 9336 assert(UseSSE42Intrinsics, "SSE4.2 must be enabled."); 9337 Label VECTOR32_LOOP, VECTOR16_LOOP, VECTOR8_LOOP, VECTOR4_LOOP; 9338 Label VECTOR16_TAIL, VECTOR8_TAIL, VECTOR4_TAIL; 9339 Label VECTOR32_NOT_EQUAL, VECTOR16_NOT_EQUAL, VECTOR8_NOT_EQUAL, VECTOR4_NOT_EQUAL; 9340 Label SAME_TILL_END, DONE; 9341 Label BYTES_LOOP, BYTES_TAIL, BYTES_NOT_EQUAL; 9342 9343 //scale is in rcx in both Win64 and Unix 9344 ShortBranchVerifier sbv(this); 9345 9346 shlq(length); 9347 xorq(result, result); 9348 9349 cmpq(length, 8); 9350 jcc(Assembler::equal, VECTOR8_LOOP); 9351 jcc(Assembler::less, VECTOR4_TAIL); 9352 9353 if (UseAVX >= 2){ 9354 9355 cmpq(length, 16); 9356 jcc(Assembler::equal, VECTOR16_LOOP); 9357 jcc(Assembler::less, VECTOR8_LOOP); 9358 9359 cmpq(length, 32); 9360 jccb(Assembler::less, VECTOR16_TAIL); 9361 9362 subq(length, 32); 9363 bind(VECTOR32_LOOP); 9364 vmovdqu(rymm0, Address(obja, result)); 9365 vmovdqu(rymm1, Address(objb, result)); 9366 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_256bit); 9367 vptest(rymm2, rymm2); 9368 jcc(Assembler::notZero, VECTOR32_NOT_EQUAL);//mismatch found 9369 addq(result, 32); 9370 subq(length, 32); 9371 jccb(Assembler::greaterEqual, VECTOR32_LOOP); 9372 addq(length, 32); 9373 jcc(Assembler::equal, SAME_TILL_END); 9374 //falling through if less than 32 bytes left //close the branch here. 9375 9376 bind(VECTOR16_TAIL); 9377 cmpq(length, 16); 9378 jccb(Assembler::less, VECTOR8_TAIL); 9379 bind(VECTOR16_LOOP); 9380 movdqu(rymm0, Address(obja, result)); 9381 movdqu(rymm1, Address(objb, result)); 9382 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_128bit); 9383 ptest(rymm2, rymm2); 9384 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found 9385 addq(result, 16); 9386 subq(length, 16); 9387 jcc(Assembler::equal, SAME_TILL_END); 9388 //falling through if less than 16 bytes left 9389 } else {//regular intrinsics 9390 9391 cmpq(length, 16); 9392 jccb(Assembler::less, VECTOR8_TAIL); 9393 9394 subq(length, 16); 9395 bind(VECTOR16_LOOP); 9396 movdqu(rymm0, Address(obja, result)); 9397 movdqu(rymm1, Address(objb, result)); 9398 pxor(rymm0, rymm1); 9399 ptest(rymm0, rymm0); 9400 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found 9401 addq(result, 16); 9402 subq(length, 16); 9403 jccb(Assembler::greaterEqual, VECTOR16_LOOP); 9404 addq(length, 16); 9405 jcc(Assembler::equal, SAME_TILL_END); 9406 //falling through if less than 16 bytes left 9407 } 9408 9409 bind(VECTOR8_TAIL); 9410 cmpq(length, 8); 9411 jccb(Assembler::less, VECTOR4_TAIL); 9412 bind(VECTOR8_LOOP); 9413 movq(tmp1, Address(obja, result)); 9414 movq(tmp2, Address(objb, result)); 9415 xorq(tmp1, tmp2); 9416 testq(tmp1, tmp1); 9417 jcc(Assembler::notZero, VECTOR8_NOT_EQUAL);//mismatch found 9418 addq(result, 8); 9419 subq(length, 8); 9420 jcc(Assembler::equal, SAME_TILL_END); 9421 //falling through if less than 8 bytes left 9422 9423 bind(VECTOR4_TAIL); 9424 cmpq(length, 4); 9425 jccb(Assembler::less, BYTES_TAIL); 9426 bind(VECTOR4_LOOP); 9427 movl(tmp1, Address(obja, result)); 9428 xorl(tmp1, Address(objb, result)); 9429 testl(tmp1, tmp1); 9430 jcc(Assembler::notZero, VECTOR4_NOT_EQUAL);//mismatch found 9431 addq(result, 4); 9432 subq(length, 4); 9433 jcc(Assembler::equal, SAME_TILL_END); 9434 //falling through if less than 4 bytes left 9435 9436 bind(BYTES_TAIL); 9437 bind(BYTES_LOOP); 9438 load_unsigned_byte(tmp1, Address(obja, result)); 9439 load_unsigned_byte(tmp2, Address(objb, result)); 9440 xorl(tmp1, tmp2); 9441 testl(tmp1, tmp1); 9442 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 9443 decq(length); 9444 jccb(Assembler::zero, SAME_TILL_END); 9445 incq(result); 9446 load_unsigned_byte(tmp1, Address(obja, result)); 9447 load_unsigned_byte(tmp2, Address(objb, result)); 9448 xorl(tmp1, tmp2); 9449 testl(tmp1, tmp1); 9450 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 9451 decq(length); 9452 jccb(Assembler::zero, SAME_TILL_END); 9453 incq(result); 9454 load_unsigned_byte(tmp1, Address(obja, result)); 9455 load_unsigned_byte(tmp2, Address(objb, result)); 9456 xorl(tmp1, tmp2); 9457 testl(tmp1, tmp1); 9458 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 9459 jmpb(SAME_TILL_END); 9460 9461 if (UseAVX >= 2){ 9462 bind(VECTOR32_NOT_EQUAL); 9463 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_256bit); 9464 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_256bit); 9465 vpxor(rymm0, rymm0, rymm2, Assembler::AVX_256bit); 9466 vpmovmskb(tmp1, rymm0); 9467 bsfq(tmp1, tmp1); 9468 addq(result, tmp1); 9469 shrq(result); 9470 jmpb(DONE); 9471 } 9472 9473 bind(VECTOR16_NOT_EQUAL); 9474 if (UseAVX >= 2){ 9475 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_128bit); 9476 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_128bit); 9477 pxor(rymm0, rymm2); 9478 } else { 9479 pcmpeqb(rymm2, rymm2); 9480 pxor(rymm0, rymm1); 9481 pcmpeqb(rymm0, rymm1); 9482 pxor(rymm0, rymm2); 9483 } 9484 pmovmskb(tmp1, rymm0); 9485 bsfq(tmp1, tmp1); 9486 addq(result, tmp1); 9487 shrq(result); 9488 jmpb(DONE); 9489 9490 bind(VECTOR8_NOT_EQUAL); 9491 bind(VECTOR4_NOT_EQUAL); 9492 bsfq(tmp1, tmp1); 9493 shrq(tmp1, 3); 9494 addq(result, tmp1); 9495 bind(BYTES_NOT_EQUAL); 9496 shrq(result); 9497 jmpb(DONE); 9498 9499 bind(SAME_TILL_END); 9500 mov64(result, -1); 9501 9502 bind(DONE); 9503 } 9504 9505 9506 //Helper functions for square_to_len() 9507 9508 /** 9509 * Store the squares of x[], right shifted one bit (divided by 2) into z[] 9510 * Preserves x and z and modifies rest of the registers. 9511 */ 9512 void MacroAssembler::square_rshift(Register x, Register xlen, Register z, Register tmp1, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 9513 // Perform square and right shift by 1 9514 // Handle odd xlen case first, then for even xlen do the following 9515 // jlong carry = 0; 9516 // for (int j=0, i=0; j < xlen; j+=2, i+=4) { 9517 // huge_128 product = x[j:j+1] * x[j:j+1]; 9518 // z[i:i+1] = (carry << 63) | (jlong)(product >>> 65); 9519 // z[i+2:i+3] = (jlong)(product >>> 1); 9520 // carry = (jlong)product; 9521 // } 9522 9523 xorq(tmp5, tmp5); // carry 9524 xorq(rdxReg, rdxReg); 9525 xorl(tmp1, tmp1); // index for x 9526 xorl(tmp4, tmp4); // index for z 9527 9528 Label L_first_loop, L_first_loop_exit; 9529 9530 testl(xlen, 1); 9531 jccb(Assembler::zero, L_first_loop); //jump if xlen is even 9532 9533 // Square and right shift by 1 the odd element using 32 bit multiply 9534 movl(raxReg, Address(x, tmp1, Address::times_4, 0)); 9535 imulq(raxReg, raxReg); 9536 shrq(raxReg, 1); 9537 adcq(tmp5, 0); 9538 movq(Address(z, tmp4, Address::times_4, 0), raxReg); 9539 incrementl(tmp1); 9540 addl(tmp4, 2); 9541 9542 // Square and right shift by 1 the rest using 64 bit multiply 9543 bind(L_first_loop); 9544 cmpptr(tmp1, xlen); 9545 jccb(Assembler::equal, L_first_loop_exit); 9546 9547 // Square 9548 movq(raxReg, Address(x, tmp1, Address::times_4, 0)); 9549 rorq(raxReg, 32); // convert big-endian to little-endian 9550 mulq(raxReg); // 64-bit multiply rax * rax -> rdx:rax 9551 9552 // Right shift by 1 and save carry 9553 shrq(tmp5, 1); // rdx:rax:tmp5 = (tmp5:rdx:rax) >>> 1 9554 rcrq(rdxReg, 1); 9555 rcrq(raxReg, 1); 9556 adcq(tmp5, 0); 9557 9558 // Store result in z 9559 movq(Address(z, tmp4, Address::times_4, 0), rdxReg); 9560 movq(Address(z, tmp4, Address::times_4, 8), raxReg); 9561 9562 // Update indices for x and z 9563 addl(tmp1, 2); 9564 addl(tmp4, 4); 9565 jmp(L_first_loop); 9566 9567 bind(L_first_loop_exit); 9568 } 9569 9570 9571 /** 9572 * Perform the following multiply add operation using BMI2 instructions 9573 * carry:sum = sum + op1*op2 + carry 9574 * op2 should be in rdx 9575 * op2 is preserved, all other registers are modified 9576 */ 9577 void MacroAssembler::multiply_add_64_bmi2(Register sum, Register op1, Register op2, Register carry, Register tmp2) { 9578 // assert op2 is rdx 9579 mulxq(tmp2, op1, op1); // op1 * op2 -> tmp2:op1 9580 addq(sum, carry); 9581 adcq(tmp2, 0); 9582 addq(sum, op1); 9583 adcq(tmp2, 0); 9584 movq(carry, tmp2); 9585 } 9586 9587 /** 9588 * Perform the following multiply add operation: 9589 * carry:sum = sum + op1*op2 + carry 9590 * Preserves op1, op2 and modifies rest of registers 9591 */ 9592 void MacroAssembler::multiply_add_64(Register sum, Register op1, Register op2, Register carry, Register rdxReg, Register raxReg) { 9593 // rdx:rax = op1 * op2 9594 movq(raxReg, op2); 9595 mulq(op1); 9596 9597 // rdx:rax = sum + carry + rdx:rax 9598 addq(sum, carry); 9599 adcq(rdxReg, 0); 9600 addq(sum, raxReg); 9601 adcq(rdxReg, 0); 9602 9603 // carry:sum = rdx:sum 9604 movq(carry, rdxReg); 9605 } 9606 9607 /** 9608 * Add 64 bit long carry into z[] with carry propogation. 9609 * Preserves z and carry register values and modifies rest of registers. 9610 * 9611 */ 9612 void MacroAssembler::add_one_64(Register z, Register zlen, Register carry, Register tmp1) { 9613 Label L_fourth_loop, L_fourth_loop_exit; 9614 9615 movl(tmp1, 1); 9616 subl(zlen, 2); 9617 addq(Address(z, zlen, Address::times_4, 0), carry); 9618 9619 bind(L_fourth_loop); 9620 jccb(Assembler::carryClear, L_fourth_loop_exit); 9621 subl(zlen, 2); 9622 jccb(Assembler::negative, L_fourth_loop_exit); 9623 addq(Address(z, zlen, Address::times_4, 0), tmp1); 9624 jmp(L_fourth_loop); 9625 bind(L_fourth_loop_exit); 9626 } 9627 9628 /** 9629 * Shift z[] left by 1 bit. 9630 * Preserves x, len, z and zlen registers and modifies rest of the registers. 9631 * 9632 */ 9633 void MacroAssembler::lshift_by_1(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4) { 9634 9635 Label L_fifth_loop, L_fifth_loop_exit; 9636 9637 // Fifth loop 9638 // Perform primitiveLeftShift(z, zlen, 1) 9639 9640 const Register prev_carry = tmp1; 9641 const Register new_carry = tmp4; 9642 const Register value = tmp2; 9643 const Register zidx = tmp3; 9644 9645 // int zidx, carry; 9646 // long value; 9647 // carry = 0; 9648 // for (zidx = zlen-2; zidx >=0; zidx -= 2) { 9649 // (carry:value) = (z[i] << 1) | carry ; 9650 // z[i] = value; 9651 // } 9652 9653 movl(zidx, zlen); 9654 xorl(prev_carry, prev_carry); // clear carry flag and prev_carry register 9655 9656 bind(L_fifth_loop); 9657 decl(zidx); // Use decl to preserve carry flag 9658 decl(zidx); 9659 jccb(Assembler::negative, L_fifth_loop_exit); 9660 9661 if (UseBMI2Instructions) { 9662 movq(value, Address(z, zidx, Address::times_4, 0)); 9663 rclq(value, 1); 9664 rorxq(value, value, 32); 9665 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form 9666 } 9667 else { 9668 // clear new_carry 9669 xorl(new_carry, new_carry); 9670 9671 // Shift z[i] by 1, or in previous carry and save new carry 9672 movq(value, Address(z, zidx, Address::times_4, 0)); 9673 shlq(value, 1); 9674 adcl(new_carry, 0); 9675 9676 orq(value, prev_carry); 9677 rorq(value, 0x20); 9678 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form 9679 9680 // Set previous carry = new carry 9681 movl(prev_carry, new_carry); 9682 } 9683 jmp(L_fifth_loop); 9684 9685 bind(L_fifth_loop_exit); 9686 } 9687 9688 9689 /** 9690 * Code for BigInteger::squareToLen() intrinsic 9691 * 9692 * rdi: x 9693 * rsi: len 9694 * r8: z 9695 * rcx: zlen 9696 * r12: tmp1 9697 * r13: tmp2 9698 * r14: tmp3 9699 * r15: tmp4 9700 * rbx: tmp5 9701 * 9702 */ 9703 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) { 9704 9705 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; 9706 push(tmp1); 9707 push(tmp2); 9708 push(tmp3); 9709 push(tmp4); 9710 push(tmp5); 9711 9712 // First loop 9713 // Store the squares, right shifted one bit (i.e., divided by 2). 9714 square_rshift(x, len, z, tmp1, tmp3, tmp4, tmp5, rdxReg, raxReg); 9715 9716 // Add in off-diagonal sums. 9717 // 9718 // Second, third (nested) and fourth loops. 9719 // zlen +=2; 9720 // for (int xidx=len-2,zidx=zlen-4; xidx > 0; xidx-=2,zidx-=4) { 9721 // carry = 0; 9722 // long op2 = x[xidx:xidx+1]; 9723 // for (int j=xidx-2,k=zidx; j >= 0; j-=2) { 9724 // k -= 2; 9725 // long op1 = x[j:j+1]; 9726 // long sum = z[k:k+1]; 9727 // carry:sum = multiply_add_64(sum, op1, op2, carry, tmp_regs); 9728 // z[k:k+1] = sum; 9729 // } 9730 // add_one_64(z, k, carry, tmp_regs); 9731 // } 9732 9733 const Register carry = tmp5; 9734 const Register sum = tmp3; 9735 const Register op1 = tmp4; 9736 Register op2 = tmp2; 9737 9738 push(zlen); 9739 push(len); 9740 addl(zlen,2); 9741 bind(L_second_loop); 9742 xorq(carry, carry); 9743 subl(zlen, 4); 9744 subl(len, 2); 9745 push(zlen); 9746 push(len); 9747 cmpl(len, 0); 9748 jccb(Assembler::lessEqual, L_second_loop_exit); 9749 9750 // Multiply an array by one 64 bit long. 9751 if (UseBMI2Instructions) { 9752 op2 = rdxReg; 9753 movq(op2, Address(x, len, Address::times_4, 0)); 9754 rorxq(op2, op2, 32); 9755 } 9756 else { 9757 movq(op2, Address(x, len, Address::times_4, 0)); 9758 rorq(op2, 32); 9759 } 9760 9761 bind(L_third_loop); 9762 decrementl(len); 9763 jccb(Assembler::negative, L_third_loop_exit); 9764 decrementl(len); 9765 jccb(Assembler::negative, L_last_x); 9766 9767 movq(op1, Address(x, len, Address::times_4, 0)); 9768 rorq(op1, 32); 9769 9770 bind(L_multiply); 9771 subl(zlen, 2); 9772 movq(sum, Address(z, zlen, Address::times_4, 0)); 9773 9774 // Multiply 64 bit by 64 bit and add 64 bits lower half and upper 64 bits as carry. 9775 if (UseBMI2Instructions) { 9776 multiply_add_64_bmi2(sum, op1, op2, carry, tmp2); 9777 } 9778 else { 9779 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9780 } 9781 9782 movq(Address(z, zlen, Address::times_4, 0), sum); 9783 9784 jmp(L_third_loop); 9785 bind(L_third_loop_exit); 9786 9787 // Fourth loop 9788 // Add 64 bit long carry into z with carry propogation. 9789 // Uses offsetted zlen. 9790 add_one_64(z, zlen, carry, tmp1); 9791 9792 pop(len); 9793 pop(zlen); 9794 jmp(L_second_loop); 9795 9796 // Next infrequent code is moved outside loops. 9797 bind(L_last_x); 9798 movl(op1, Address(x, 0)); 9799 jmp(L_multiply); 9800 9801 bind(L_second_loop_exit); 9802 pop(len); 9803 pop(zlen); 9804 pop(len); 9805 pop(zlen); 9806 9807 // Fifth loop 9808 // Shift z left 1 bit. 9809 lshift_by_1(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4); 9810 9811 // z[zlen-1] |= x[len-1] & 1; 9812 movl(tmp3, Address(x, len, Address::times_4, -4)); 9813 andl(tmp3, 1); 9814 orl(Address(z, zlen, Address::times_4, -4), tmp3); 9815 9816 pop(tmp5); 9817 pop(tmp4); 9818 pop(tmp3); 9819 pop(tmp2); 9820 pop(tmp1); 9821 } 9822 9823 /** 9824 * Helper function for mul_add() 9825 * Multiply the in[] by int k and add to out[] starting at offset offs using 9826 * 128 bit by 32 bit multiply and return the carry in tmp5. 9827 * Only quad int aligned length of in[] is operated on in this function. 9828 * k is in rdxReg for BMI2Instructions, for others it is in tmp2. 9829 * This function preserves out, in and k registers. 9830 * len and offset point to the appropriate index in "in" & "out" correspondingly 9831 * tmp5 has the carry. 9832 * other registers are temporary and are modified. 9833 * 9834 */ 9835 void MacroAssembler::mul_add_128_x_32_loop(Register out, Register in, 9836 Register offset, Register len, Register tmp1, Register tmp2, Register tmp3, 9837 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 9838 9839 Label L_first_loop, L_first_loop_exit; 9840 9841 movl(tmp1, len); 9842 shrl(tmp1, 2); 9843 9844 bind(L_first_loop); 9845 subl(tmp1, 1); 9846 jccb(Assembler::negative, L_first_loop_exit); 9847 9848 subl(len, 4); 9849 subl(offset, 4); 9850 9851 Register op2 = tmp2; 9852 const Register sum = tmp3; 9853 const Register op1 = tmp4; 9854 const Register carry = tmp5; 9855 9856 if (UseBMI2Instructions) { 9857 op2 = rdxReg; 9858 } 9859 9860 movq(op1, Address(in, len, Address::times_4, 8)); 9861 rorq(op1, 32); 9862 movq(sum, Address(out, offset, Address::times_4, 8)); 9863 rorq(sum, 32); 9864 if (UseBMI2Instructions) { 9865 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 9866 } 9867 else { 9868 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9869 } 9870 // Store back in big endian from little endian 9871 rorq(sum, 0x20); 9872 movq(Address(out, offset, Address::times_4, 8), sum); 9873 9874 movq(op1, Address(in, len, Address::times_4, 0)); 9875 rorq(op1, 32); 9876 movq(sum, Address(out, offset, Address::times_4, 0)); 9877 rorq(sum, 32); 9878 if (UseBMI2Instructions) { 9879 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 9880 } 9881 else { 9882 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9883 } 9884 // Store back in big endian from little endian 9885 rorq(sum, 0x20); 9886 movq(Address(out, offset, Address::times_4, 0), sum); 9887 9888 jmp(L_first_loop); 9889 bind(L_first_loop_exit); 9890 } 9891 9892 /** 9893 * Code for BigInteger::mulAdd() intrinsic 9894 * 9895 * rdi: out 9896 * rsi: in 9897 * r11: offs (out.length - offset) 9898 * rcx: len 9899 * r8: k 9900 * r12: tmp1 9901 * r13: tmp2 9902 * r14: tmp3 9903 * r15: tmp4 9904 * rbx: tmp5 9905 * Multiply the in[] by word k and add to out[], return the carry in rax 9906 */ 9907 void MacroAssembler::mul_add(Register out, Register in, Register offs, 9908 Register len, Register k, Register tmp1, Register tmp2, Register tmp3, 9909 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 9910 9911 Label L_carry, L_last_in, L_done; 9912 9913 // carry = 0; 9914 // for (int j=len-1; j >= 0; j--) { 9915 // long product = (in[j] & LONG_MASK) * kLong + 9916 // (out[offs] & LONG_MASK) + carry; 9917 // out[offs--] = (int)product; 9918 // carry = product >>> 32; 9919 // } 9920 // 9921 push(tmp1); 9922 push(tmp2); 9923 push(tmp3); 9924 push(tmp4); 9925 push(tmp5); 9926 9927 Register op2 = tmp2; 9928 const Register sum = tmp3; 9929 const Register op1 = tmp4; 9930 const Register carry = tmp5; 9931 9932 if (UseBMI2Instructions) { 9933 op2 = rdxReg; 9934 movl(op2, k); 9935 } 9936 else { 9937 movl(op2, k); 9938 } 9939 9940 xorq(carry, carry); 9941 9942 //First loop 9943 9944 //Multiply in[] by k in a 4 way unrolled loop using 128 bit by 32 bit multiply 9945 //The carry is in tmp5 9946 mul_add_128_x_32_loop(out, in, offs, len, tmp1, tmp2, tmp3, tmp4, tmp5, rdxReg, raxReg); 9947 9948 //Multiply the trailing in[] entry using 64 bit by 32 bit, if any 9949 decrementl(len); 9950 jccb(Assembler::negative, L_carry); 9951 decrementl(len); 9952 jccb(Assembler::negative, L_last_in); 9953 9954 movq(op1, Address(in, len, Address::times_4, 0)); 9955 rorq(op1, 32); 9956 9957 subl(offs, 2); 9958 movq(sum, Address(out, offs, Address::times_4, 0)); 9959 rorq(sum, 32); 9960 9961 if (UseBMI2Instructions) { 9962 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 9963 } 9964 else { 9965 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9966 } 9967 9968 // Store back in big endian from little endian 9969 rorq(sum, 0x20); 9970 movq(Address(out, offs, Address::times_4, 0), sum); 9971 9972 testl(len, len); 9973 jccb(Assembler::zero, L_carry); 9974 9975 //Multiply the last in[] entry, if any 9976 bind(L_last_in); 9977 movl(op1, Address(in, 0)); 9978 movl(sum, Address(out, offs, Address::times_4, -4)); 9979 9980 movl(raxReg, k); 9981 mull(op1); //tmp4 * eax -> edx:eax 9982 addl(sum, carry); 9983 adcl(rdxReg, 0); 9984 addl(sum, raxReg); 9985 adcl(rdxReg, 0); 9986 movl(carry, rdxReg); 9987 9988 movl(Address(out, offs, Address::times_4, -4), sum); 9989 9990 bind(L_carry); 9991 //return tmp5/carry as carry in rax 9992 movl(rax, carry); 9993 9994 bind(L_done); 9995 pop(tmp5); 9996 pop(tmp4); 9997 pop(tmp3); 9998 pop(tmp2); 9999 pop(tmp1); 10000 } 10001 #endif 10002 10003 /** 10004 * Emits code to update CRC-32 with a byte value according to constants in table 10005 * 10006 * @param [in,out]crc Register containing the crc. 10007 * @param [in]val Register containing the byte to fold into the CRC. 10008 * @param [in]table Register containing the table of crc constants. 10009 * 10010 * uint32_t crc; 10011 * val = crc_table[(val ^ crc) & 0xFF]; 10012 * crc = val ^ (crc >> 8); 10013 * 10014 */ 10015 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) { 10016 xorl(val, crc); 10017 andl(val, 0xFF); 10018 shrl(crc, 8); // unsigned shift 10019 xorl(crc, Address(table, val, Address::times_4, 0)); 10020 } 10021 10022 /** 10023 * Fold 128-bit data chunk 10024 */ 10025 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) { 10026 if (UseAVX > 0) { 10027 vpclmulhdq(xtmp, xK, xcrc); // [123:64] 10028 vpclmulldq(xcrc, xK, xcrc); // [63:0] 10029 vpxor(xcrc, xcrc, Address(buf, offset), 0 /* vector_len */); 10030 pxor(xcrc, xtmp); 10031 } else { 10032 movdqa(xtmp, xcrc); 10033 pclmulhdq(xtmp, xK); // [123:64] 10034 pclmulldq(xcrc, xK); // [63:0] 10035 pxor(xcrc, xtmp); 10036 movdqu(xtmp, Address(buf, offset)); 10037 pxor(xcrc, xtmp); 10038 } 10039 } 10040 10041 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) { 10042 if (UseAVX > 0) { 10043 vpclmulhdq(xtmp, xK, xcrc); 10044 vpclmulldq(xcrc, xK, xcrc); 10045 pxor(xcrc, xbuf); 10046 pxor(xcrc, xtmp); 10047 } else { 10048 movdqa(xtmp, xcrc); 10049 pclmulhdq(xtmp, xK); 10050 pclmulldq(xcrc, xK); 10051 pxor(xcrc, xbuf); 10052 pxor(xcrc, xtmp); 10053 } 10054 } 10055 10056 /** 10057 * 8-bit folds to compute 32-bit CRC 10058 * 10059 * uint64_t xcrc; 10060 * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8); 10061 */ 10062 void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) { 10063 movdl(tmp, xcrc); 10064 andl(tmp, 0xFF); 10065 movdl(xtmp, Address(table, tmp, Address::times_4, 0)); 10066 psrldq(xcrc, 1); // unsigned shift one byte 10067 pxor(xcrc, xtmp); 10068 } 10069 10070 /** 10071 * uint32_t crc; 10072 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8); 10073 */ 10074 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) { 10075 movl(tmp, crc); 10076 andl(tmp, 0xFF); 10077 shrl(crc, 8); 10078 xorl(crc, Address(table, tmp, Address::times_4, 0)); 10079 } 10080 10081 /** 10082 * @param crc register containing existing CRC (32-bit) 10083 * @param buf register pointing to input byte buffer (byte*) 10084 * @param len register containing number of bytes 10085 * @param table register that will contain address of CRC table 10086 * @param tmp scratch register 10087 */ 10088 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) { 10089 assert_different_registers(crc, buf, len, table, tmp, rax); 10090 10091 Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned; 10092 Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop; 10093 10094 // For EVEX with VL and BW, provide a standard mask, VL = 128 will guide the merge 10095 // context for the registers used, where all instructions below are using 128-bit mode 10096 // On EVEX without VL and BW, these instructions will all be AVX. 10097 if (VM_Version::supports_avx512vlbw()) { 10098 movl(tmp, 0xffff); 10099 kmovwl(k1, tmp); 10100 } 10101 10102 lea(table, ExternalAddress(StubRoutines::crc_table_addr())); 10103 notl(crc); // ~crc 10104 cmpl(len, 16); 10105 jcc(Assembler::less, L_tail); 10106 10107 // Align buffer to 16 bytes 10108 movl(tmp, buf); 10109 andl(tmp, 0xF); 10110 jccb(Assembler::zero, L_aligned); 10111 subl(tmp, 16); 10112 addl(len, tmp); 10113 10114 align(4); 10115 BIND(L_align_loop); 10116 movsbl(rax, Address(buf, 0)); // load byte with sign extension 10117 update_byte_crc32(crc, rax, table); 10118 increment(buf); 10119 incrementl(tmp); 10120 jccb(Assembler::less, L_align_loop); 10121 10122 BIND(L_aligned); 10123 movl(tmp, len); // save 10124 shrl(len, 4); 10125 jcc(Assembler::zero, L_tail_restore); 10126 10127 // Fold crc into first bytes of vector 10128 movdqa(xmm1, Address(buf, 0)); 10129 movdl(rax, xmm1); 10130 xorl(crc, rax); 10131 pinsrd(xmm1, crc, 0); 10132 addptr(buf, 16); 10133 subl(len, 4); // len > 0 10134 jcc(Assembler::less, L_fold_tail); 10135 10136 movdqa(xmm2, Address(buf, 0)); 10137 movdqa(xmm3, Address(buf, 16)); 10138 movdqa(xmm4, Address(buf, 32)); 10139 addptr(buf, 48); 10140 subl(len, 3); 10141 jcc(Assembler::lessEqual, L_fold_512b); 10142 10143 // Fold total 512 bits of polynomial on each iteration, 10144 // 128 bits per each of 4 parallel streams. 10145 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32)); 10146 10147 align(32); 10148 BIND(L_fold_512b_loop); 10149 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0); 10150 fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16); 10151 fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32); 10152 fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48); 10153 addptr(buf, 64); 10154 subl(len, 4); 10155 jcc(Assembler::greater, L_fold_512b_loop); 10156 10157 // Fold 512 bits to 128 bits. 10158 BIND(L_fold_512b); 10159 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16)); 10160 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2); 10161 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3); 10162 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4); 10163 10164 // Fold the rest of 128 bits data chunks 10165 BIND(L_fold_tail); 10166 addl(len, 3); 10167 jccb(Assembler::lessEqual, L_fold_128b); 10168 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16)); 10169 10170 BIND(L_fold_tail_loop); 10171 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0); 10172 addptr(buf, 16); 10173 decrementl(len); 10174 jccb(Assembler::greater, L_fold_tail_loop); 10175 10176 // Fold 128 bits in xmm1 down into 32 bits in crc register. 10177 BIND(L_fold_128b); 10178 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr())); 10179 if (UseAVX > 0) { 10180 vpclmulqdq(xmm2, xmm0, xmm1, 0x1); 10181 vpand(xmm3, xmm0, xmm2, 0 /* vector_len */); 10182 vpclmulqdq(xmm0, xmm0, xmm3, 0x1); 10183 } else { 10184 movdqa(xmm2, xmm0); 10185 pclmulqdq(xmm2, xmm1, 0x1); 10186 movdqa(xmm3, xmm0); 10187 pand(xmm3, xmm2); 10188 pclmulqdq(xmm0, xmm3, 0x1); 10189 } 10190 psrldq(xmm1, 8); 10191 psrldq(xmm2, 4); 10192 pxor(xmm0, xmm1); 10193 pxor(xmm0, xmm2); 10194 10195 // 8 8-bit folds to compute 32-bit CRC. 10196 for (int j = 0; j < 4; j++) { 10197 fold_8bit_crc32(xmm0, table, xmm1, rax); 10198 } 10199 movdl(crc, xmm0); // mov 32 bits to general register 10200 for (int j = 0; j < 4; j++) { 10201 fold_8bit_crc32(crc, table, rax); 10202 } 10203 10204 BIND(L_tail_restore); 10205 movl(len, tmp); // restore 10206 BIND(L_tail); 10207 andl(len, 0xf); 10208 jccb(Assembler::zero, L_exit); 10209 10210 // Fold the rest of bytes 10211 align(4); 10212 BIND(L_tail_loop); 10213 movsbl(rax, Address(buf, 0)); // load byte with sign extension 10214 update_byte_crc32(crc, rax, table); 10215 increment(buf); 10216 decrementl(len); 10217 jccb(Assembler::greater, L_tail_loop); 10218 10219 BIND(L_exit); 10220 notl(crc); // ~c 10221 } 10222 10223 #ifdef _LP64 10224 // S. Gueron / Information Processing Letters 112 (2012) 184 10225 // Algorithm 4: Computing carry-less multiplication using a precomputed lookup table. 10226 // Input: A 32 bit value B = [byte3, byte2, byte1, byte0]. 10227 // Output: the 64-bit carry-less product of B * CONST 10228 void MacroAssembler::crc32c_ipl_alg4(Register in, uint32_t n, 10229 Register tmp1, Register tmp2, Register tmp3) { 10230 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr())); 10231 if (n > 0) { 10232 addq(tmp3, n * 256 * 8); 10233 } 10234 // Q1 = TABLEExt[n][B & 0xFF]; 10235 movl(tmp1, in); 10236 andl(tmp1, 0x000000FF); 10237 shll(tmp1, 3); 10238 addq(tmp1, tmp3); 10239 movq(tmp1, Address(tmp1, 0)); 10240 10241 // Q2 = TABLEExt[n][B >> 8 & 0xFF]; 10242 movl(tmp2, in); 10243 shrl(tmp2, 8); 10244 andl(tmp2, 0x000000FF); 10245 shll(tmp2, 3); 10246 addq(tmp2, tmp3); 10247 movq(tmp2, Address(tmp2, 0)); 10248 10249 shlq(tmp2, 8); 10250 xorq(tmp1, tmp2); 10251 10252 // Q3 = TABLEExt[n][B >> 16 & 0xFF]; 10253 movl(tmp2, in); 10254 shrl(tmp2, 16); 10255 andl(tmp2, 0x000000FF); 10256 shll(tmp2, 3); 10257 addq(tmp2, tmp3); 10258 movq(tmp2, Address(tmp2, 0)); 10259 10260 shlq(tmp2, 16); 10261 xorq(tmp1, tmp2); 10262 10263 // Q4 = TABLEExt[n][B >> 24 & 0xFF]; 10264 shrl(in, 24); 10265 andl(in, 0x000000FF); 10266 shll(in, 3); 10267 addq(in, tmp3); 10268 movq(in, Address(in, 0)); 10269 10270 shlq(in, 24); 10271 xorq(in, tmp1); 10272 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24; 10273 } 10274 10275 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1, 10276 Register in_out, 10277 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported, 10278 XMMRegister w_xtmp2, 10279 Register tmp1, 10280 Register n_tmp2, Register n_tmp3) { 10281 if (is_pclmulqdq_supported) { 10282 movdl(w_xtmp1, in_out); // modified blindly 10283 10284 movl(tmp1, const_or_pre_comp_const_index); 10285 movdl(w_xtmp2, tmp1); 10286 pclmulqdq(w_xtmp1, w_xtmp2, 0); 10287 10288 movdq(in_out, w_xtmp1); 10289 } else { 10290 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3); 10291 } 10292 } 10293 10294 // Recombination Alternative 2: No bit-reflections 10295 // T1 = (CRC_A * U1) << 1 10296 // T2 = (CRC_B * U2) << 1 10297 // C1 = T1 >> 32 10298 // C2 = T2 >> 32 10299 // T1 = T1 & 0xFFFFFFFF 10300 // T2 = T2 & 0xFFFFFFFF 10301 // T1 = CRC32(0, T1) 10302 // T2 = CRC32(0, T2) 10303 // C1 = C1 ^ T1 10304 // C2 = C2 ^ T2 10305 // CRC = C1 ^ C2 ^ CRC_C 10306 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, 10307 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10308 Register tmp1, Register tmp2, 10309 Register n_tmp3) { 10310 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10311 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10312 shlq(in_out, 1); 10313 movl(tmp1, in_out); 10314 shrq(in_out, 32); 10315 xorl(tmp2, tmp2); 10316 crc32(tmp2, tmp1, 4); 10317 xorl(in_out, tmp2); // we don't care about upper 32 bit contents here 10318 shlq(in1, 1); 10319 movl(tmp1, in1); 10320 shrq(in1, 32); 10321 xorl(tmp2, tmp2); 10322 crc32(tmp2, tmp1, 4); 10323 xorl(in1, tmp2); 10324 xorl(in_out, in1); 10325 xorl(in_out, in2); 10326 } 10327 10328 // Set N to predefined value 10329 // Subtract from a lenght of a buffer 10330 // execute in a loop: 10331 // CRC_A = 0xFFFFFFFF, CRC_B = 0, CRC_C = 0 10332 // for i = 1 to N do 10333 // CRC_A = CRC32(CRC_A, A[i]) 10334 // CRC_B = CRC32(CRC_B, B[i]) 10335 // CRC_C = CRC32(CRC_C, C[i]) 10336 // end for 10337 // Recombine 10338 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, 10339 Register in_out1, Register in_out2, Register in_out3, 10340 Register tmp1, Register tmp2, Register tmp3, 10341 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10342 Register tmp4, Register tmp5, 10343 Register n_tmp6) { 10344 Label L_processPartitions; 10345 Label L_processPartition; 10346 Label L_exit; 10347 10348 bind(L_processPartitions); 10349 cmpl(in_out1, 3 * size); 10350 jcc(Assembler::less, L_exit); 10351 xorl(tmp1, tmp1); 10352 xorl(tmp2, tmp2); 10353 movq(tmp3, in_out2); 10354 addq(tmp3, size); 10355 10356 bind(L_processPartition); 10357 crc32(in_out3, Address(in_out2, 0), 8); 10358 crc32(tmp1, Address(in_out2, size), 8); 10359 crc32(tmp2, Address(in_out2, size * 2), 8); 10360 addq(in_out2, 8); 10361 cmpq(in_out2, tmp3); 10362 jcc(Assembler::less, L_processPartition); 10363 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2, 10364 w_xtmp1, w_xtmp2, w_xtmp3, 10365 tmp4, tmp5, 10366 n_tmp6); 10367 addq(in_out2, 2 * size); 10368 subl(in_out1, 3 * size); 10369 jmp(L_processPartitions); 10370 10371 bind(L_exit); 10372 } 10373 #else 10374 void MacroAssembler::crc32c_ipl_alg4(Register in_out, uint32_t n, 10375 Register tmp1, Register tmp2, Register tmp3, 10376 XMMRegister xtmp1, XMMRegister xtmp2) { 10377 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr())); 10378 if (n > 0) { 10379 addl(tmp3, n * 256 * 8); 10380 } 10381 // Q1 = TABLEExt[n][B & 0xFF]; 10382 movl(tmp1, in_out); 10383 andl(tmp1, 0x000000FF); 10384 shll(tmp1, 3); 10385 addl(tmp1, tmp3); 10386 movq(xtmp1, Address(tmp1, 0)); 10387 10388 // Q2 = TABLEExt[n][B >> 8 & 0xFF]; 10389 movl(tmp2, in_out); 10390 shrl(tmp2, 8); 10391 andl(tmp2, 0x000000FF); 10392 shll(tmp2, 3); 10393 addl(tmp2, tmp3); 10394 movq(xtmp2, Address(tmp2, 0)); 10395 10396 psllq(xtmp2, 8); 10397 pxor(xtmp1, xtmp2); 10398 10399 // Q3 = TABLEExt[n][B >> 16 & 0xFF]; 10400 movl(tmp2, in_out); 10401 shrl(tmp2, 16); 10402 andl(tmp2, 0x000000FF); 10403 shll(tmp2, 3); 10404 addl(tmp2, tmp3); 10405 movq(xtmp2, Address(tmp2, 0)); 10406 10407 psllq(xtmp2, 16); 10408 pxor(xtmp1, xtmp2); 10409 10410 // Q4 = TABLEExt[n][B >> 24 & 0xFF]; 10411 shrl(in_out, 24); 10412 andl(in_out, 0x000000FF); 10413 shll(in_out, 3); 10414 addl(in_out, tmp3); 10415 movq(xtmp2, Address(in_out, 0)); 10416 10417 psllq(xtmp2, 24); 10418 pxor(xtmp1, xtmp2); // Result in CXMM 10419 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24; 10420 } 10421 10422 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1, 10423 Register in_out, 10424 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported, 10425 XMMRegister w_xtmp2, 10426 Register tmp1, 10427 Register n_tmp2, Register n_tmp3) { 10428 if (is_pclmulqdq_supported) { 10429 movdl(w_xtmp1, in_out); 10430 10431 movl(tmp1, const_or_pre_comp_const_index); 10432 movdl(w_xtmp2, tmp1); 10433 pclmulqdq(w_xtmp1, w_xtmp2, 0); 10434 // Keep result in XMM since GPR is 32 bit in length 10435 } else { 10436 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3, w_xtmp1, w_xtmp2); 10437 } 10438 } 10439 10440 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, 10441 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10442 Register tmp1, Register tmp2, 10443 Register n_tmp3) { 10444 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10445 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 10446 10447 psllq(w_xtmp1, 1); 10448 movdl(tmp1, w_xtmp1); 10449 psrlq(w_xtmp1, 32); 10450 movdl(in_out, w_xtmp1); 10451 10452 xorl(tmp2, tmp2); 10453 crc32(tmp2, tmp1, 4); 10454 xorl(in_out, tmp2); 10455 10456 psllq(w_xtmp2, 1); 10457 movdl(tmp1, w_xtmp2); 10458 psrlq(w_xtmp2, 32); 10459 movdl(in1, w_xtmp2); 10460 10461 xorl(tmp2, tmp2); 10462 crc32(tmp2, tmp1, 4); 10463 xorl(in1, tmp2); 10464 xorl(in_out, in1); 10465 xorl(in_out, in2); 10466 } 10467 10468 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, 10469 Register in_out1, Register in_out2, Register in_out3, 10470 Register tmp1, Register tmp2, Register tmp3, 10471 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10472 Register tmp4, Register tmp5, 10473 Register n_tmp6) { 10474 Label L_processPartitions; 10475 Label L_processPartition; 10476 Label L_exit; 10477 10478 bind(L_processPartitions); 10479 cmpl(in_out1, 3 * size); 10480 jcc(Assembler::less, L_exit); 10481 xorl(tmp1, tmp1); 10482 xorl(tmp2, tmp2); 10483 movl(tmp3, in_out2); 10484 addl(tmp3, size); 10485 10486 bind(L_processPartition); 10487 crc32(in_out3, Address(in_out2, 0), 4); 10488 crc32(tmp1, Address(in_out2, size), 4); 10489 crc32(tmp2, Address(in_out2, size*2), 4); 10490 crc32(in_out3, Address(in_out2, 0+4), 4); 10491 crc32(tmp1, Address(in_out2, size+4), 4); 10492 crc32(tmp2, Address(in_out2, size*2+4), 4); 10493 addl(in_out2, 8); 10494 cmpl(in_out2, tmp3); 10495 jcc(Assembler::less, L_processPartition); 10496 10497 push(tmp3); 10498 push(in_out1); 10499 push(in_out2); 10500 tmp4 = tmp3; 10501 tmp5 = in_out1; 10502 n_tmp6 = in_out2; 10503 10504 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2, 10505 w_xtmp1, w_xtmp2, w_xtmp3, 10506 tmp4, tmp5, 10507 n_tmp6); 10508 10509 pop(in_out2); 10510 pop(in_out1); 10511 pop(tmp3); 10512 10513 addl(in_out2, 2 * size); 10514 subl(in_out1, 3 * size); 10515 jmp(L_processPartitions); 10516 10517 bind(L_exit); 10518 } 10519 #endif //LP64 10520 10521 #ifdef _LP64 10522 // Algorithm 2: Pipelined usage of the CRC32 instruction. 10523 // Input: A buffer I of L bytes. 10524 // Output: the CRC32C value of the buffer. 10525 // Notations: 10526 // Write L = 24N + r, with N = floor (L/24). 10527 // r = L mod 24 (0 <= r < 24). 10528 // Consider I as the concatenation of A|B|C|R, where A, B, C, each, 10529 // N quadwords, and R consists of r bytes. 10530 // A[j] = I [8j+7:8j], j= 0, 1, ..., N-1 10531 // B[j] = I [N + 8j+7:N + 8j], j= 0, 1, ..., N-1 10532 // C[j] = I [2N + 8j+7:2N + 8j], j= 0, 1, ..., N-1 10533 // if r > 0 R[j] = I [3N +j], j= 0, 1, ...,r-1 10534 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2, 10535 Register tmp1, Register tmp2, Register tmp3, 10536 Register tmp4, Register tmp5, Register tmp6, 10537 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10538 bool is_pclmulqdq_supported) { 10539 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS]; 10540 Label L_wordByWord; 10541 Label L_byteByByteProlog; 10542 Label L_byteByByte; 10543 Label L_exit; 10544 10545 if (is_pclmulqdq_supported ) { 10546 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr; 10547 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr+1); 10548 10549 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2); 10550 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3); 10551 10552 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4); 10553 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5); 10554 assert((CRC32C_NUM_PRECOMPUTED_CONSTANTS - 1 ) == 5, "Checking whether you declared all of the constants based on the number of \"chunks\""); 10555 } else { 10556 const_or_pre_comp_const_index[0] = 1; 10557 const_or_pre_comp_const_index[1] = 0; 10558 10559 const_or_pre_comp_const_index[2] = 3; 10560 const_or_pre_comp_const_index[3] = 2; 10561 10562 const_or_pre_comp_const_index[4] = 5; 10563 const_or_pre_comp_const_index[5] = 4; 10564 } 10565 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported, 10566 in2, in1, in_out, 10567 tmp1, tmp2, tmp3, 10568 w_xtmp1, w_xtmp2, w_xtmp3, 10569 tmp4, tmp5, 10570 tmp6); 10571 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported, 10572 in2, in1, in_out, 10573 tmp1, tmp2, tmp3, 10574 w_xtmp1, w_xtmp2, w_xtmp3, 10575 tmp4, tmp5, 10576 tmp6); 10577 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported, 10578 in2, in1, in_out, 10579 tmp1, tmp2, tmp3, 10580 w_xtmp1, w_xtmp2, w_xtmp3, 10581 tmp4, tmp5, 10582 tmp6); 10583 movl(tmp1, in2); 10584 andl(tmp1, 0x00000007); 10585 negl(tmp1); 10586 addl(tmp1, in2); 10587 addq(tmp1, in1); 10588 10589 BIND(L_wordByWord); 10590 cmpq(in1, tmp1); 10591 jcc(Assembler::greaterEqual, L_byteByByteProlog); 10592 crc32(in_out, Address(in1, 0), 4); 10593 addq(in1, 4); 10594 jmp(L_wordByWord); 10595 10596 BIND(L_byteByByteProlog); 10597 andl(in2, 0x00000007); 10598 movl(tmp2, 1); 10599 10600 BIND(L_byteByByte); 10601 cmpl(tmp2, in2); 10602 jccb(Assembler::greater, L_exit); 10603 crc32(in_out, Address(in1, 0), 1); 10604 incq(in1); 10605 incl(tmp2); 10606 jmp(L_byteByByte); 10607 10608 BIND(L_exit); 10609 } 10610 #else 10611 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2, 10612 Register tmp1, Register tmp2, Register tmp3, 10613 Register tmp4, Register tmp5, Register tmp6, 10614 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 10615 bool is_pclmulqdq_supported) { 10616 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS]; 10617 Label L_wordByWord; 10618 Label L_byteByByteProlog; 10619 Label L_byteByByte; 10620 Label L_exit; 10621 10622 if (is_pclmulqdq_supported) { 10623 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr; 10624 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 1); 10625 10626 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2); 10627 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3); 10628 10629 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4); 10630 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5); 10631 } else { 10632 const_or_pre_comp_const_index[0] = 1; 10633 const_or_pre_comp_const_index[1] = 0; 10634 10635 const_or_pre_comp_const_index[2] = 3; 10636 const_or_pre_comp_const_index[3] = 2; 10637 10638 const_or_pre_comp_const_index[4] = 5; 10639 const_or_pre_comp_const_index[5] = 4; 10640 } 10641 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported, 10642 in2, in1, in_out, 10643 tmp1, tmp2, tmp3, 10644 w_xtmp1, w_xtmp2, w_xtmp3, 10645 tmp4, tmp5, 10646 tmp6); 10647 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported, 10648 in2, in1, in_out, 10649 tmp1, tmp2, tmp3, 10650 w_xtmp1, w_xtmp2, w_xtmp3, 10651 tmp4, tmp5, 10652 tmp6); 10653 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported, 10654 in2, in1, in_out, 10655 tmp1, tmp2, tmp3, 10656 w_xtmp1, w_xtmp2, w_xtmp3, 10657 tmp4, tmp5, 10658 tmp6); 10659 movl(tmp1, in2); 10660 andl(tmp1, 0x00000007); 10661 negl(tmp1); 10662 addl(tmp1, in2); 10663 addl(tmp1, in1); 10664 10665 BIND(L_wordByWord); 10666 cmpl(in1, tmp1); 10667 jcc(Assembler::greaterEqual, L_byteByByteProlog); 10668 crc32(in_out, Address(in1,0), 4); 10669 addl(in1, 4); 10670 jmp(L_wordByWord); 10671 10672 BIND(L_byteByByteProlog); 10673 andl(in2, 0x00000007); 10674 movl(tmp2, 1); 10675 10676 BIND(L_byteByByte); 10677 cmpl(tmp2, in2); 10678 jccb(Assembler::greater, L_exit); 10679 movb(tmp1, Address(in1, 0)); 10680 crc32(in_out, tmp1, 1); 10681 incl(in1); 10682 incl(tmp2); 10683 jmp(L_byteByByte); 10684 10685 BIND(L_exit); 10686 } 10687 #endif // LP64 10688 #undef BIND 10689 #undef BLOCK_COMMENT 10690 10691 10692 // Compress char[] array to byte[]. 10693 void MacroAssembler::char_array_compress(Register src, Register dst, Register len, 10694 XMMRegister tmp1Reg, XMMRegister tmp2Reg, 10695 XMMRegister tmp3Reg, XMMRegister tmp4Reg, 10696 Register tmp5, Register result) { 10697 Label copy_chars_loop, return_length, return_zero, done; 10698 10699 // rsi: src 10700 // rdi: dst 10701 // rdx: len 10702 // rcx: tmp5 10703 // rax: result 10704 10705 // rsi holds start addr of source char[] to be compressed 10706 // rdi holds start addr of destination byte[] 10707 // rdx holds length 10708 10709 assert(len != result, ""); 10710 10711 // save length for return 10712 push(len); 10713 10714 if (UseSSE42Intrinsics) { 10715 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 10716 Label copy_32_loop, copy_16, copy_tail; 10717 10718 movl(result, len); 10719 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vectors 10720 10721 // vectored compression 10722 andl(len, 0xfffffff0); // vector count (in chars) 10723 andl(result, 0x0000000f); // tail count (in chars) 10724 testl(len, len); 10725 jccb(Assembler::zero, copy_16); 10726 10727 // compress 16 chars per iter 10728 movdl(tmp1Reg, tmp5); 10729 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg 10730 pxor(tmp4Reg, tmp4Reg); 10731 10732 lea(src, Address(src, len, Address::times_2)); 10733 lea(dst, Address(dst, len, Address::times_1)); 10734 negptr(len); 10735 10736 bind(copy_32_loop); 10737 movdqu(tmp2Reg, Address(src, len, Address::times_2)); // load 1st 8 characters 10738 por(tmp4Reg, tmp2Reg); 10739 movdqu(tmp3Reg, Address(src, len, Address::times_2, 16)); // load next 8 characters 10740 por(tmp4Reg, tmp3Reg); 10741 ptest(tmp4Reg, tmp1Reg); // check for Unicode chars in next vector 10742 jcc(Assembler::notZero, return_zero); 10743 packuswb(tmp2Reg, tmp3Reg); // only ASCII chars; compress each to 1 byte 10744 movdqu(Address(dst, len, Address::times_1), tmp2Reg); 10745 addptr(len, 16); 10746 jcc(Assembler::notZero, copy_32_loop); 10747 10748 // compress next vector of 8 chars (if any) 10749 bind(copy_16); 10750 movl(len, result); 10751 andl(len, 0xfffffff8); // vector count (in chars) 10752 andl(result, 0x00000007); // tail count (in chars) 10753 testl(len, len); 10754 jccb(Assembler::zero, copy_tail); 10755 10756 movdl(tmp1Reg, tmp5); 10757 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg 10758 pxor(tmp3Reg, tmp3Reg); 10759 10760 movdqu(tmp2Reg, Address(src, 0)); 10761 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 10762 jccb(Assembler::notZero, return_zero); 10763 packuswb(tmp2Reg, tmp3Reg); // only LATIN1 chars; compress each to 1 byte 10764 movq(Address(dst, 0), tmp2Reg); 10765 addptr(src, 16); 10766 addptr(dst, 8); 10767 10768 bind(copy_tail); 10769 movl(len, result); 10770 } 10771 // compress 1 char per iter 10772 testl(len, len); 10773 jccb(Assembler::zero, return_length); 10774 lea(src, Address(src, len, Address::times_2)); 10775 lea(dst, Address(dst, len, Address::times_1)); 10776 negptr(len); 10777 10778 bind(copy_chars_loop); 10779 load_unsigned_short(result, Address(src, len, Address::times_2)); 10780 testl(result, 0xff00); // check if Unicode char 10781 jccb(Assembler::notZero, return_zero); 10782 movb(Address(dst, len, Address::times_1), result); // ASCII char; compress to 1 byte 10783 increment(len); 10784 jcc(Assembler::notZero, copy_chars_loop); 10785 10786 // if compression succeeded, return length 10787 bind(return_length); 10788 pop(result); 10789 jmpb(done); 10790 10791 // if compression failed, return 0 10792 bind(return_zero); 10793 xorl(result, result); 10794 addptr(rsp, wordSize); 10795 10796 bind(done); 10797 } 10798 10799 // Inflate byte[] array to char[]. 10800 void MacroAssembler::byte_array_inflate(Register src, Register dst, Register len, 10801 XMMRegister tmp1, Register tmp2) { 10802 Label copy_chars_loop, done; 10803 10804 // rsi: src 10805 // rdi: dst 10806 // rdx: len 10807 // rcx: tmp2 10808 10809 // rsi holds start addr of source byte[] to be inflated 10810 // rdi holds start addr of destination char[] 10811 // rdx holds length 10812 assert_different_registers(src, dst, len, tmp2); 10813 10814 if (UseSSE42Intrinsics) { 10815 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available"); 10816 Label copy_8_loop, copy_bytes, copy_tail; 10817 10818 movl(tmp2, len); 10819 andl(tmp2, 0x00000007); // tail count (in chars) 10820 andl(len, 0xfffffff8); // vector count (in chars) 10821 jccb(Assembler::zero, copy_tail); 10822 10823 // vectored inflation 10824 lea(src, Address(src, len, Address::times_1)); 10825 lea(dst, Address(dst, len, Address::times_2)); 10826 negptr(len); 10827 10828 // inflate 8 chars per iter 10829 bind(copy_8_loop); 10830 pmovzxbw(tmp1, Address(src, len, Address::times_1)); // unpack to 8 words 10831 movdqu(Address(dst, len, Address::times_2), tmp1); 10832 addptr(len, 8); 10833 jcc(Assembler::notZero, copy_8_loop); 10834 10835 bind(copy_tail); 10836 movl(len, tmp2); 10837 10838 cmpl(len, 4); 10839 jccb(Assembler::less, copy_bytes); 10840 10841 movdl(tmp1, Address(src, 0)); // load 4 byte chars 10842 pmovzxbw(tmp1, tmp1); 10843 movq(Address(dst, 0), tmp1); 10844 subptr(len, 4); 10845 addptr(src, 4); 10846 addptr(dst, 8); 10847 10848 bind(copy_bytes); 10849 } 10850 testl(len, len); 10851 jccb(Assembler::zero, done); 10852 lea(src, Address(src, len, Address::times_1)); 10853 lea(dst, Address(dst, len, Address::times_2)); 10854 negptr(len); 10855 10856 // inflate 1 char per iter 10857 bind(copy_chars_loop); 10858 load_unsigned_byte(tmp2, Address(src, len, Address::times_1)); // load byte char 10859 movw(Address(dst, len, Address::times_2), tmp2); // inflate byte char to word 10860 increment(len); 10861 jcc(Assembler::notZero, copy_chars_loop); 10862 10863 bind(done); 10864 } 10865 10866 10867 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) { 10868 switch (cond) { 10869 // Note some conditions are synonyms for others 10870 case Assembler::zero: return Assembler::notZero; 10871 case Assembler::notZero: return Assembler::zero; 10872 case Assembler::less: return Assembler::greaterEqual; 10873 case Assembler::lessEqual: return Assembler::greater; 10874 case Assembler::greater: return Assembler::lessEqual; 10875 case Assembler::greaterEqual: return Assembler::less; 10876 case Assembler::below: return Assembler::aboveEqual; 10877 case Assembler::belowEqual: return Assembler::above; 10878 case Assembler::above: return Assembler::belowEqual; 10879 case Assembler::aboveEqual: return Assembler::below; 10880 case Assembler::overflow: return Assembler::noOverflow; 10881 case Assembler::noOverflow: return Assembler::overflow; 10882 case Assembler::negative: return Assembler::positive; 10883 case Assembler::positive: return Assembler::negative; 10884 case Assembler::parity: return Assembler::noParity; 10885 case Assembler::noParity: return Assembler::parity; 10886 } 10887 ShouldNotReachHere(); return Assembler::overflow; 10888 } 10889 10890 SkipIfEqual::SkipIfEqual( 10891 MacroAssembler* masm, const bool* flag_addr, bool value) { 10892 _masm = masm; 10893 _masm->cmp8(ExternalAddress((address)flag_addr), value); 10894 _masm->jcc(Assembler::equal, _label); 10895 } 10896 10897 SkipIfEqual::~SkipIfEqual() { 10898 _masm->bind(_label); 10899 } 10900 10901 // 32-bit Windows has its own fast-path implementation 10902 // of get_thread 10903 #if !defined(WIN32) || defined(_LP64) 10904 10905 // This is simply a call to Thread::current() 10906 void MacroAssembler::get_thread(Register thread) { 10907 if (thread != rax) { 10908 push(rax); 10909 } 10910 LP64_ONLY(push(rdi);) 10911 LP64_ONLY(push(rsi);) 10912 push(rdx); 10913 push(rcx); 10914 #ifdef _LP64 10915 push(r8); 10916 push(r9); 10917 push(r10); 10918 push(r11); 10919 #endif 10920 10921 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, Thread::current), 0); 10922 10923 #ifdef _LP64 10924 pop(r11); 10925 pop(r10); 10926 pop(r9); 10927 pop(r8); 10928 #endif 10929 pop(rcx); 10930 pop(rdx); 10931 LP64_ONLY(pop(rsi);) 10932 LP64_ONLY(pop(rdi);) 10933 if (thread != rax) { 10934 mov(thread, rax); 10935 pop(rax); 10936 } 10937 } 10938 10939 #endif