1 /* 2 * Copyright (c) 1997, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "jvm.h" 27 #include "asm/assembler.hpp" 28 #include "asm/assembler.inline.hpp" 29 #include "compiler/disassembler.hpp" 30 #include "gc/shared/barrierSet.hpp" 31 #include "gc/shared/barrierSetAssembler.hpp" 32 #include "gc/shared/collectedHeap.inline.hpp" 33 #include "interpreter/interpreter.hpp" 34 #include "memory/resourceArea.hpp" 35 #include "memory/universe.hpp" 36 #include "oops/accessDecorators.hpp" 37 #include "oops/compressedOops.inline.hpp" 38 #include "oops/klass.inline.hpp" 39 #include "prims/methodHandles.hpp" 40 #include "runtime/biasedLocking.hpp" 41 #include "runtime/flags/flagSetting.hpp" 42 #include "runtime/interfaceSupport.inline.hpp" 43 #include "runtime/objectMonitor.hpp" 44 #include "runtime/os.hpp" 45 #include "runtime/safepoint.hpp" 46 #include "runtime/safepointMechanism.hpp" 47 #include "runtime/sharedRuntime.hpp" 48 #include "runtime/stubRoutines.hpp" 49 #include "runtime/thread.hpp" 50 #include "utilities/macros.hpp" 51 #include "crc32c.h" 52 #ifdef COMPILER2 53 #include "opto/intrinsicnode.hpp" 54 #endif 55 56 #ifdef PRODUCT 57 #define BLOCK_COMMENT(str) /* nothing */ 58 #define STOP(error) stop(error) 59 #else 60 #define BLOCK_COMMENT(str) block_comment(str) 61 #define STOP(error) block_comment(error); stop(error) 62 #endif 63 64 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":") 65 66 #ifdef ASSERT 67 bool AbstractAssembler::pd_check_instruction_mark() { return true; } 68 #endif 69 70 static Assembler::Condition reverse[] = { 71 Assembler::noOverflow /* overflow = 0x0 */ , 72 Assembler::overflow /* noOverflow = 0x1 */ , 73 Assembler::aboveEqual /* carrySet = 0x2, below = 0x2 */ , 74 Assembler::below /* aboveEqual = 0x3, carryClear = 0x3 */ , 75 Assembler::notZero /* zero = 0x4, equal = 0x4 */ , 76 Assembler::zero /* notZero = 0x5, notEqual = 0x5 */ , 77 Assembler::above /* belowEqual = 0x6 */ , 78 Assembler::belowEqual /* above = 0x7 */ , 79 Assembler::positive /* negative = 0x8 */ , 80 Assembler::negative /* positive = 0x9 */ , 81 Assembler::noParity /* parity = 0xa */ , 82 Assembler::parity /* noParity = 0xb */ , 83 Assembler::greaterEqual /* less = 0xc */ , 84 Assembler::less /* greaterEqual = 0xd */ , 85 Assembler::greater /* lessEqual = 0xe */ , 86 Assembler::lessEqual /* greater = 0xf, */ 87 88 }; 89 90 91 // Implementation of MacroAssembler 92 93 // First all the versions that have distinct versions depending on 32/64 bit 94 // Unless the difference is trivial (1 line or so). 95 96 #ifndef _LP64 97 98 // 32bit versions 99 100 Address MacroAssembler::as_Address(AddressLiteral adr) { 101 return Address(adr.target(), adr.rspec()); 102 } 103 104 Address MacroAssembler::as_Address(ArrayAddress adr) { 105 return Address::make_array(adr); 106 } 107 108 void MacroAssembler::call_VM_leaf_base(address entry_point, 109 int number_of_arguments) { 110 call(RuntimeAddress(entry_point)); 111 increment(rsp, number_of_arguments * wordSize); 112 } 113 114 void MacroAssembler::cmpklass(Address src1, Metadata* obj) { 115 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 116 } 117 118 void MacroAssembler::cmpklass(Register src1, Metadata* obj) { 119 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 120 } 121 122 void MacroAssembler::cmpoop_raw(Address src1, jobject obj) { 123 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate()); 124 } 125 126 void MacroAssembler::cmpoop_raw(Register src1, jobject obj) { 127 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate()); 128 } 129 130 void MacroAssembler::cmpoop(Address src1, jobject obj) { 131 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 132 bs->obj_equals(this, src1, obj); 133 } 134 135 void MacroAssembler::cmpoop(Register src1, jobject obj) { 136 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 137 bs->obj_equals(this, src1, obj); 138 } 139 140 void MacroAssembler::extend_sign(Register hi, Register lo) { 141 // According to Intel Doc. AP-526, "Integer Divide", p.18. 142 if (VM_Version::is_P6() && hi == rdx && lo == rax) { 143 cdql(); 144 } else { 145 movl(hi, lo); 146 sarl(hi, 31); 147 } 148 } 149 150 void MacroAssembler::jC2(Register tmp, Label& L) { 151 // set parity bit if FPU flag C2 is set (via rax) 152 save_rax(tmp); 153 fwait(); fnstsw_ax(); 154 sahf(); 155 restore_rax(tmp); 156 // branch 157 jcc(Assembler::parity, L); 158 } 159 160 void MacroAssembler::jnC2(Register tmp, Label& L) { 161 // set parity bit if FPU flag C2 is set (via rax) 162 save_rax(tmp); 163 fwait(); fnstsw_ax(); 164 sahf(); 165 restore_rax(tmp); 166 // branch 167 jcc(Assembler::noParity, L); 168 } 169 170 // 32bit can do a case table jump in one instruction but we no longer allow the base 171 // to be installed in the Address class 172 void MacroAssembler::jump(ArrayAddress entry) { 173 jmp(as_Address(entry)); 174 } 175 176 // Note: y_lo will be destroyed 177 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) { 178 // Long compare for Java (semantics as described in JVM spec.) 179 Label high, low, done; 180 181 cmpl(x_hi, y_hi); 182 jcc(Assembler::less, low); 183 jcc(Assembler::greater, high); 184 // x_hi is the return register 185 xorl(x_hi, x_hi); 186 cmpl(x_lo, y_lo); 187 jcc(Assembler::below, low); 188 jcc(Assembler::equal, done); 189 190 bind(high); 191 xorl(x_hi, x_hi); 192 increment(x_hi); 193 jmp(done); 194 195 bind(low); 196 xorl(x_hi, x_hi); 197 decrementl(x_hi); 198 199 bind(done); 200 } 201 202 void MacroAssembler::lea(Register dst, AddressLiteral src) { 203 mov_literal32(dst, (int32_t)src.target(), src.rspec()); 204 } 205 206 void MacroAssembler::lea(Address dst, AddressLiteral adr) { 207 // leal(dst, as_Address(adr)); 208 // see note in movl as to why we must use a move 209 mov_literal32(dst, (int32_t) adr.target(), adr.rspec()); 210 } 211 212 void MacroAssembler::leave() { 213 mov(rsp, rbp); 214 pop(rbp); 215 } 216 217 void MacroAssembler::lmul(int x_rsp_offset, int y_rsp_offset) { 218 // Multiplication of two Java long values stored on the stack 219 // as illustrated below. Result is in rdx:rax. 220 // 221 // rsp ---> [ ?? ] \ \ 222 // .... | y_rsp_offset | 223 // [ y_lo ] / (in bytes) | x_rsp_offset 224 // [ y_hi ] | (in bytes) 225 // .... | 226 // [ x_lo ] / 227 // [ x_hi ] 228 // .... 229 // 230 // Basic idea: lo(result) = lo(x_lo * y_lo) 231 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi) 232 Address x_hi(rsp, x_rsp_offset + wordSize); Address x_lo(rsp, x_rsp_offset); 233 Address y_hi(rsp, y_rsp_offset + wordSize); Address y_lo(rsp, y_rsp_offset); 234 Label quick; 235 // load x_hi, y_hi and check if quick 236 // multiplication is possible 237 movl(rbx, x_hi); 238 movl(rcx, y_hi); 239 movl(rax, rbx); 240 orl(rbx, rcx); // rbx, = 0 <=> x_hi = 0 and y_hi = 0 241 jcc(Assembler::zero, quick); // if rbx, = 0 do quick multiply 242 // do full multiplication 243 // 1st step 244 mull(y_lo); // x_hi * y_lo 245 movl(rbx, rax); // save lo(x_hi * y_lo) in rbx, 246 // 2nd step 247 movl(rax, x_lo); 248 mull(rcx); // x_lo * y_hi 249 addl(rbx, rax); // add lo(x_lo * y_hi) to rbx, 250 // 3rd step 251 bind(quick); // note: rbx, = 0 if quick multiply! 252 movl(rax, x_lo); 253 mull(y_lo); // x_lo * y_lo 254 addl(rdx, rbx); // correct hi(x_lo * y_lo) 255 } 256 257 void MacroAssembler::lneg(Register hi, Register lo) { 258 negl(lo); 259 adcl(hi, 0); 260 negl(hi); 261 } 262 263 void MacroAssembler::lshl(Register hi, Register lo) { 264 // Java shift left long support (semantics as described in JVM spec., p.305) 265 // (basic idea for shift counts s >= n: x << s == (x << n) << (s - n)) 266 // shift value is in rcx ! 267 assert(hi != rcx, "must not use rcx"); 268 assert(lo != rcx, "must not use rcx"); 269 const Register s = rcx; // shift count 270 const int n = BitsPerWord; 271 Label L; 272 andl(s, 0x3f); // s := s & 0x3f (s < 0x40) 273 cmpl(s, n); // if (s < n) 274 jcc(Assembler::less, L); // else (s >= n) 275 movl(hi, lo); // x := x << n 276 xorl(lo, lo); 277 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n! 278 bind(L); // s (mod n) < n 279 shldl(hi, lo); // x := x << s 280 shll(lo); 281 } 282 283 284 void MacroAssembler::lshr(Register hi, Register lo, bool sign_extension) { 285 // Java shift right long support (semantics as described in JVM spec., p.306 & p.310) 286 // (basic idea for shift counts s >= n: x >> s == (x >> n) >> (s - n)) 287 assert(hi != rcx, "must not use rcx"); 288 assert(lo != rcx, "must not use rcx"); 289 const Register s = rcx; // shift count 290 const int n = BitsPerWord; 291 Label L; 292 andl(s, 0x3f); // s := s & 0x3f (s < 0x40) 293 cmpl(s, n); // if (s < n) 294 jcc(Assembler::less, L); // else (s >= n) 295 movl(lo, hi); // x := x >> n 296 if (sign_extension) sarl(hi, 31); 297 else xorl(hi, hi); 298 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n! 299 bind(L); // s (mod n) < n 300 shrdl(lo, hi); // x := x >> s 301 if (sign_extension) sarl(hi); 302 else shrl(hi); 303 } 304 305 void MacroAssembler::movoop(Register dst, jobject obj) { 306 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate()); 307 } 308 309 void MacroAssembler::movoop(Address dst, jobject obj) { 310 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate()); 311 } 312 313 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) { 314 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 315 } 316 317 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) { 318 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate()); 319 } 320 321 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) { 322 // scratch register is not used, 323 // it is defined to match parameters of 64-bit version of this method. 324 if (src.is_lval()) { 325 mov_literal32(dst, (intptr_t)src.target(), src.rspec()); 326 } else { 327 movl(dst, as_Address(src)); 328 } 329 } 330 331 void MacroAssembler::movptr(ArrayAddress dst, Register src) { 332 movl(as_Address(dst), src); 333 } 334 335 void MacroAssembler::movptr(Register dst, ArrayAddress src) { 336 movl(dst, as_Address(src)); 337 } 338 339 // src should NEVER be a real pointer. Use AddressLiteral for true pointers 340 void MacroAssembler::movptr(Address dst, intptr_t src) { 341 movl(dst, src); 342 } 343 344 345 void MacroAssembler::pop_callee_saved_registers() { 346 pop(rcx); 347 pop(rdx); 348 pop(rdi); 349 pop(rsi); 350 } 351 352 void MacroAssembler::pop_fTOS() { 353 fld_d(Address(rsp, 0)); 354 addl(rsp, 2 * wordSize); 355 } 356 357 void MacroAssembler::push_callee_saved_registers() { 358 push(rsi); 359 push(rdi); 360 push(rdx); 361 push(rcx); 362 } 363 364 void MacroAssembler::push_fTOS() { 365 subl(rsp, 2 * wordSize); 366 fstp_d(Address(rsp, 0)); 367 } 368 369 370 void MacroAssembler::pushoop(jobject obj) { 371 push_literal32((int32_t)obj, oop_Relocation::spec_for_immediate()); 372 } 373 374 void MacroAssembler::pushklass(Metadata* obj) { 375 push_literal32((int32_t)obj, metadata_Relocation::spec_for_immediate()); 376 } 377 378 void MacroAssembler::pushptr(AddressLiteral src) { 379 if (src.is_lval()) { 380 push_literal32((int32_t)src.target(), src.rspec()); 381 } else { 382 pushl(as_Address(src)); 383 } 384 } 385 386 void MacroAssembler::set_word_if_not_zero(Register dst) { 387 xorl(dst, dst); 388 set_byte_if_not_zero(dst); 389 } 390 391 static void pass_arg0(MacroAssembler* masm, Register arg) { 392 masm->push(arg); 393 } 394 395 static void pass_arg1(MacroAssembler* masm, Register arg) { 396 masm->push(arg); 397 } 398 399 static void pass_arg2(MacroAssembler* masm, Register arg) { 400 masm->push(arg); 401 } 402 403 static void pass_arg3(MacroAssembler* masm, Register arg) { 404 masm->push(arg); 405 } 406 407 #ifndef PRODUCT 408 extern "C" void findpc(intptr_t x); 409 #endif 410 411 void MacroAssembler::debug32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip, char* msg) { 412 // In order to get locks to work, we need to fake a in_VM state 413 JavaThread* thread = JavaThread::current(); 414 JavaThreadState saved_state = thread->thread_state(); 415 thread->set_thread_state(_thread_in_vm); 416 if (ShowMessageBoxOnError) { 417 JavaThread* thread = JavaThread::current(); 418 JavaThreadState saved_state = thread->thread_state(); 419 thread->set_thread_state(_thread_in_vm); 420 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) { 421 ttyLocker ttyl; 422 BytecodeCounter::print(); 423 } 424 // To see where a verify_oop failed, get $ebx+40/X for this frame. 425 // This is the value of eip which points to where verify_oop will return. 426 if (os::message_box(msg, "Execution stopped, print registers?")) { 427 print_state32(rdi, rsi, rbp, rsp, rbx, rdx, rcx, rax, eip); 428 BREAKPOINT; 429 } 430 } 431 fatal("DEBUG MESSAGE: %s", msg); 432 } 433 434 void MacroAssembler::print_state32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip) { 435 ttyLocker ttyl; 436 FlagSetting fs(Debugging, true); 437 tty->print_cr("eip = 0x%08x", eip); 438 #ifndef PRODUCT 439 if ((WizardMode || Verbose) && PrintMiscellaneous) { 440 tty->cr(); 441 findpc(eip); 442 tty->cr(); 443 } 444 #endif 445 #define PRINT_REG(rax) \ 446 { tty->print("%s = ", #rax); os::print_location(tty, rax); } 447 PRINT_REG(rax); 448 PRINT_REG(rbx); 449 PRINT_REG(rcx); 450 PRINT_REG(rdx); 451 PRINT_REG(rdi); 452 PRINT_REG(rsi); 453 PRINT_REG(rbp); 454 PRINT_REG(rsp); 455 #undef PRINT_REG 456 // Print some words near top of staack. 457 int* dump_sp = (int*) rsp; 458 for (int col1 = 0; col1 < 8; col1++) { 459 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 460 os::print_location(tty, *dump_sp++); 461 } 462 for (int row = 0; row < 16; row++) { 463 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 464 for (int col = 0; col < 8; col++) { 465 tty->print(" 0x%08x", *dump_sp++); 466 } 467 tty->cr(); 468 } 469 // Print some instructions around pc: 470 Disassembler::decode((address)eip-64, (address)eip); 471 tty->print_cr("--------"); 472 Disassembler::decode((address)eip, (address)eip+32); 473 } 474 475 void MacroAssembler::stop(const char* msg) { 476 ExternalAddress message((address)msg); 477 // push address of message 478 pushptr(message.addr()); 479 { Label L; call(L, relocInfo::none); bind(L); } // push eip 480 pusha(); // push registers 481 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug32))); 482 hlt(); 483 } 484 485 void MacroAssembler::warn(const char* msg) { 486 push_CPU_state(); 487 488 ExternalAddress message((address) msg); 489 // push address of message 490 pushptr(message.addr()); 491 492 call(RuntimeAddress(CAST_FROM_FN_PTR(address, warning))); 493 addl(rsp, wordSize); // discard argument 494 pop_CPU_state(); 495 } 496 497 void MacroAssembler::print_state() { 498 { Label L; call(L, relocInfo::none); bind(L); } // push eip 499 pusha(); // push registers 500 501 push_CPU_state(); 502 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::print_state32))); 503 pop_CPU_state(); 504 505 popa(); 506 addl(rsp, wordSize); 507 } 508 509 #else // _LP64 510 511 // 64 bit versions 512 513 Address MacroAssembler::as_Address(AddressLiteral adr) { 514 // amd64 always does this as a pc-rel 515 // we can be absolute or disp based on the instruction type 516 // jmp/call are displacements others are absolute 517 assert(!adr.is_lval(), "must be rval"); 518 assert(reachable(adr), "must be"); 519 return Address((int32_t)(intptr_t)(adr.target() - pc()), adr.target(), adr.reloc()); 520 521 } 522 523 Address MacroAssembler::as_Address(ArrayAddress adr) { 524 AddressLiteral base = adr.base(); 525 lea(rscratch1, base); 526 Address index = adr.index(); 527 assert(index._disp == 0, "must not have disp"); // maybe it can? 528 Address array(rscratch1, index._index, index._scale, index._disp); 529 return array; 530 } 531 532 void MacroAssembler::call_VM_leaf_base(address entry_point, int num_args) { 533 Label L, E; 534 535 #ifdef _WIN64 536 // Windows always allocates space for it's register args 537 assert(num_args <= 4, "only register arguments supported"); 538 subq(rsp, frame::arg_reg_save_area_bytes); 539 #endif 540 541 // Align stack if necessary 542 testl(rsp, 15); 543 jcc(Assembler::zero, L); 544 545 subq(rsp, 8); 546 { 547 call(RuntimeAddress(entry_point)); 548 } 549 addq(rsp, 8); 550 jmp(E); 551 552 bind(L); 553 { 554 call(RuntimeAddress(entry_point)); 555 } 556 557 bind(E); 558 559 #ifdef _WIN64 560 // restore stack pointer 561 addq(rsp, frame::arg_reg_save_area_bytes); 562 #endif 563 564 } 565 566 void MacroAssembler::cmp64(Register src1, AddressLiteral src2) { 567 assert(!src2.is_lval(), "should use cmpptr"); 568 569 if (reachable(src2)) { 570 cmpq(src1, as_Address(src2)); 571 } else { 572 lea(rscratch1, src2); 573 Assembler::cmpq(src1, Address(rscratch1, 0)); 574 } 575 } 576 577 int MacroAssembler::corrected_idivq(Register reg) { 578 // Full implementation of Java ldiv and lrem; checks for special 579 // case as described in JVM spec., p.243 & p.271. The function 580 // returns the (pc) offset of the idivl instruction - may be needed 581 // for implicit exceptions. 582 // 583 // normal case special case 584 // 585 // input : rax: dividend min_long 586 // reg: divisor (may not be eax/edx) -1 587 // 588 // output: rax: quotient (= rax idiv reg) min_long 589 // rdx: remainder (= rax irem reg) 0 590 assert(reg != rax && reg != rdx, "reg cannot be rax or rdx register"); 591 static const int64_t min_long = 0x8000000000000000; 592 Label normal_case, special_case; 593 594 // check for special case 595 cmp64(rax, ExternalAddress((address) &min_long)); 596 jcc(Assembler::notEqual, normal_case); 597 xorl(rdx, rdx); // prepare rdx for possible special case (where 598 // remainder = 0) 599 cmpq(reg, -1); 600 jcc(Assembler::equal, special_case); 601 602 // handle normal case 603 bind(normal_case); 604 cdqq(); 605 int idivq_offset = offset(); 606 idivq(reg); 607 608 // normal and special case exit 609 bind(special_case); 610 611 return idivq_offset; 612 } 613 614 void MacroAssembler::decrementq(Register reg, int value) { 615 if (value == min_jint) { subq(reg, value); return; } 616 if (value < 0) { incrementq(reg, -value); return; } 617 if (value == 0) { ; return; } 618 if (value == 1 && UseIncDec) { decq(reg) ; return; } 619 /* else */ { subq(reg, value) ; return; } 620 } 621 622 void MacroAssembler::decrementq(Address dst, int value) { 623 if (value == min_jint) { subq(dst, value); return; } 624 if (value < 0) { incrementq(dst, -value); return; } 625 if (value == 0) { ; return; } 626 if (value == 1 && UseIncDec) { decq(dst) ; return; } 627 /* else */ { subq(dst, value) ; return; } 628 } 629 630 void MacroAssembler::incrementq(AddressLiteral dst) { 631 if (reachable(dst)) { 632 incrementq(as_Address(dst)); 633 } else { 634 lea(rscratch1, dst); 635 incrementq(Address(rscratch1, 0)); 636 } 637 } 638 639 void MacroAssembler::incrementq(Register reg, int value) { 640 if (value == min_jint) { addq(reg, value); return; } 641 if (value < 0) { decrementq(reg, -value); return; } 642 if (value == 0) { ; return; } 643 if (value == 1 && UseIncDec) { incq(reg) ; return; } 644 /* else */ { addq(reg, value) ; return; } 645 } 646 647 void MacroAssembler::incrementq(Address dst, int value) { 648 if (value == min_jint) { addq(dst, value); return; } 649 if (value < 0) { decrementq(dst, -value); return; } 650 if (value == 0) { ; return; } 651 if (value == 1 && UseIncDec) { incq(dst) ; return; } 652 /* else */ { addq(dst, value) ; return; } 653 } 654 655 // 32bit can do a case table jump in one instruction but we no longer allow the base 656 // to be installed in the Address class 657 void MacroAssembler::jump(ArrayAddress entry) { 658 lea(rscratch1, entry.base()); 659 Address dispatch = entry.index(); 660 assert(dispatch._base == noreg, "must be"); 661 dispatch._base = rscratch1; 662 jmp(dispatch); 663 } 664 665 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) { 666 ShouldNotReachHere(); // 64bit doesn't use two regs 667 cmpq(x_lo, y_lo); 668 } 669 670 void MacroAssembler::lea(Register dst, AddressLiteral src) { 671 mov_literal64(dst, (intptr_t)src.target(), src.rspec()); 672 } 673 674 void MacroAssembler::lea(Address dst, AddressLiteral adr) { 675 mov_literal64(rscratch1, (intptr_t)adr.target(), adr.rspec()); 676 movptr(dst, rscratch1); 677 } 678 679 void MacroAssembler::leave() { 680 // %%% is this really better? Why not on 32bit too? 681 emit_int8((unsigned char)0xC9); // LEAVE 682 } 683 684 void MacroAssembler::lneg(Register hi, Register lo) { 685 ShouldNotReachHere(); // 64bit doesn't use two regs 686 negq(lo); 687 } 688 689 void MacroAssembler::movoop(Register dst, jobject obj) { 690 mov_literal64(dst, (intptr_t)obj, oop_Relocation::spec_for_immediate()); 691 } 692 693 void MacroAssembler::movoop(Address dst, jobject obj) { 694 mov_literal64(rscratch1, (intptr_t)obj, oop_Relocation::spec_for_immediate()); 695 movq(dst, rscratch1); 696 } 697 698 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) { 699 mov_literal64(dst, (intptr_t)obj, metadata_Relocation::spec_for_immediate()); 700 } 701 702 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) { 703 mov_literal64(rscratch1, (intptr_t)obj, metadata_Relocation::spec_for_immediate()); 704 movq(dst, rscratch1); 705 } 706 707 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) { 708 if (src.is_lval()) { 709 mov_literal64(dst, (intptr_t)src.target(), src.rspec()); 710 } else { 711 if (reachable(src)) { 712 movq(dst, as_Address(src)); 713 } else { 714 lea(scratch, src); 715 movq(dst, Address(scratch, 0)); 716 } 717 } 718 } 719 720 void MacroAssembler::movptr(ArrayAddress dst, Register src) { 721 movq(as_Address(dst), src); 722 } 723 724 void MacroAssembler::movptr(Register dst, ArrayAddress src) { 725 movq(dst, as_Address(src)); 726 } 727 728 // src should NEVER be a real pointer. Use AddressLiteral for true pointers 729 void MacroAssembler::movptr(Address dst, intptr_t src) { 730 mov64(rscratch1, src); 731 movq(dst, rscratch1); 732 } 733 734 // These are mostly for initializing NULL 735 void MacroAssembler::movptr(Address dst, int32_t src) { 736 movslq(dst, src); 737 } 738 739 void MacroAssembler::movptr(Register dst, int32_t src) { 740 mov64(dst, (intptr_t)src); 741 } 742 743 void MacroAssembler::pushoop(jobject obj) { 744 movoop(rscratch1, obj); 745 push(rscratch1); 746 } 747 748 void MacroAssembler::pushklass(Metadata* obj) { 749 mov_metadata(rscratch1, obj); 750 push(rscratch1); 751 } 752 753 void MacroAssembler::pushptr(AddressLiteral src) { 754 lea(rscratch1, src); 755 if (src.is_lval()) { 756 push(rscratch1); 757 } else { 758 pushq(Address(rscratch1, 0)); 759 } 760 } 761 762 void MacroAssembler::reset_last_Java_frame(bool clear_fp) { 763 // we must set sp to zero to clear frame 764 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), NULL_WORD); 765 // must clear fp, so that compiled frames are not confused; it is 766 // possible that we need it only for debugging 767 if (clear_fp) { 768 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD); 769 } 770 771 // Always clear the pc because it could have been set by make_walkable() 772 movptr(Address(r15_thread, JavaThread::last_Java_pc_offset()), NULL_WORD); 773 vzeroupper(); 774 } 775 776 void MacroAssembler::set_last_Java_frame(Register last_java_sp, 777 Register last_java_fp, 778 address last_java_pc) { 779 vzeroupper(); 780 // determine last_java_sp register 781 if (!last_java_sp->is_valid()) { 782 last_java_sp = rsp; 783 } 784 785 // last_java_fp is optional 786 if (last_java_fp->is_valid()) { 787 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), 788 last_java_fp); 789 } 790 791 // last_java_pc is optional 792 if (last_java_pc != NULL) { 793 Address java_pc(r15_thread, 794 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()); 795 lea(rscratch1, InternalAddress(last_java_pc)); 796 movptr(java_pc, rscratch1); 797 } 798 799 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), last_java_sp); 800 } 801 802 static void pass_arg0(MacroAssembler* masm, Register arg) { 803 if (c_rarg0 != arg ) { 804 masm->mov(c_rarg0, arg); 805 } 806 } 807 808 static void pass_arg1(MacroAssembler* masm, Register arg) { 809 if (c_rarg1 != arg ) { 810 masm->mov(c_rarg1, arg); 811 } 812 } 813 814 static void pass_arg2(MacroAssembler* masm, Register arg) { 815 if (c_rarg2 != arg ) { 816 masm->mov(c_rarg2, arg); 817 } 818 } 819 820 static void pass_arg3(MacroAssembler* masm, Register arg) { 821 if (c_rarg3 != arg ) { 822 masm->mov(c_rarg3, arg); 823 } 824 } 825 826 void MacroAssembler::stop(const char* msg) { 827 if (ShowMessageBoxOnError) { 828 address rip = pc(); 829 pusha(); // get regs on stack 830 lea(c_rarg1, InternalAddress(rip)); 831 movq(c_rarg2, rsp); // pass pointer to regs array 832 } 833 lea(c_rarg0, ExternalAddress((address) msg)); 834 andq(rsp, -16); // align stack as required by ABI 835 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64))); 836 hlt(); 837 } 838 839 void MacroAssembler::warn(const char* msg) { 840 push(rbp); 841 movq(rbp, rsp); 842 andq(rsp, -16); // align stack as required by push_CPU_state and call 843 push_CPU_state(); // keeps alignment at 16 bytes 844 lea(c_rarg0, ExternalAddress((address) msg)); 845 lea(rax, ExternalAddress(CAST_FROM_FN_PTR(address, warning))); 846 call(rax); 847 pop_CPU_state(); 848 mov(rsp, rbp); 849 pop(rbp); 850 } 851 852 void MacroAssembler::print_state() { 853 address rip = pc(); 854 pusha(); // get regs on stack 855 push(rbp); 856 movq(rbp, rsp); 857 andq(rsp, -16); // align stack as required by push_CPU_state and call 858 push_CPU_state(); // keeps alignment at 16 bytes 859 860 lea(c_rarg0, InternalAddress(rip)); 861 lea(c_rarg1, Address(rbp, wordSize)); // pass pointer to regs array 862 call_VM_leaf(CAST_FROM_FN_PTR(address, MacroAssembler::print_state64), c_rarg0, c_rarg1); 863 864 pop_CPU_state(); 865 mov(rsp, rbp); 866 pop(rbp); 867 popa(); 868 } 869 870 #ifndef PRODUCT 871 extern "C" void findpc(intptr_t x); 872 #endif 873 874 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[]) { 875 // In order to get locks to work, we need to fake a in_VM state 876 if (ShowMessageBoxOnError) { 877 JavaThread* thread = JavaThread::current(); 878 JavaThreadState saved_state = thread->thread_state(); 879 thread->set_thread_state(_thread_in_vm); 880 #ifndef PRODUCT 881 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) { 882 ttyLocker ttyl; 883 BytecodeCounter::print(); 884 } 885 #endif 886 // To see where a verify_oop failed, get $ebx+40/X for this frame. 887 // XXX correct this offset for amd64 888 // This is the value of eip which points to where verify_oop will return. 889 if (os::message_box(msg, "Execution stopped, print registers?")) { 890 print_state64(pc, regs); 891 BREAKPOINT; 892 } 893 } 894 fatal("DEBUG MESSAGE: %s", msg); 895 } 896 897 void MacroAssembler::print_state64(int64_t pc, int64_t regs[]) { 898 ttyLocker ttyl; 899 FlagSetting fs(Debugging, true); 900 tty->print_cr("rip = 0x%016lx", (intptr_t)pc); 901 #ifndef PRODUCT 902 tty->cr(); 903 findpc(pc); 904 tty->cr(); 905 #endif 906 #define PRINT_REG(rax, value) \ 907 { tty->print("%s = ", #rax); os::print_location(tty, value); } 908 PRINT_REG(rax, regs[15]); 909 PRINT_REG(rbx, regs[12]); 910 PRINT_REG(rcx, regs[14]); 911 PRINT_REG(rdx, regs[13]); 912 PRINT_REG(rdi, regs[8]); 913 PRINT_REG(rsi, regs[9]); 914 PRINT_REG(rbp, regs[10]); 915 PRINT_REG(rsp, regs[11]); 916 PRINT_REG(r8 , regs[7]); 917 PRINT_REG(r9 , regs[6]); 918 PRINT_REG(r10, regs[5]); 919 PRINT_REG(r11, regs[4]); 920 PRINT_REG(r12, regs[3]); 921 PRINT_REG(r13, regs[2]); 922 PRINT_REG(r14, regs[1]); 923 PRINT_REG(r15, regs[0]); 924 #undef PRINT_REG 925 // Print some words near top of staack. 926 int64_t* rsp = (int64_t*) regs[11]; 927 int64_t* dump_sp = rsp; 928 for (int col1 = 0; col1 < 8; col1++) { 929 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 930 os::print_location(tty, *dump_sp++); 931 } 932 for (int row = 0; row < 25; row++) { 933 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 934 for (int col = 0; col < 4; col++) { 935 tty->print(" 0x%016lx", (intptr_t)*dump_sp++); 936 } 937 tty->cr(); 938 } 939 // Print some instructions around pc: 940 Disassembler::decode((address)pc-64, (address)pc); 941 tty->print_cr("--------"); 942 Disassembler::decode((address)pc, (address)pc+32); 943 } 944 945 #endif // _LP64 946 947 // Now versions that are common to 32/64 bit 948 949 void MacroAssembler::addptr(Register dst, int32_t imm32) { 950 LP64_ONLY(addq(dst, imm32)) NOT_LP64(addl(dst, imm32)); 951 } 952 953 void MacroAssembler::addptr(Register dst, Register src) { 954 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src)); 955 } 956 957 void MacroAssembler::addptr(Address dst, Register src) { 958 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src)); 959 } 960 961 void MacroAssembler::addsd(XMMRegister dst, AddressLiteral src) { 962 if (reachable(src)) { 963 Assembler::addsd(dst, as_Address(src)); 964 } else { 965 lea(rscratch1, src); 966 Assembler::addsd(dst, Address(rscratch1, 0)); 967 } 968 } 969 970 void MacroAssembler::addss(XMMRegister dst, AddressLiteral src) { 971 if (reachable(src)) { 972 addss(dst, as_Address(src)); 973 } else { 974 lea(rscratch1, src); 975 addss(dst, Address(rscratch1, 0)); 976 } 977 } 978 979 void MacroAssembler::addpd(XMMRegister dst, AddressLiteral src) { 980 if (reachable(src)) { 981 Assembler::addpd(dst, as_Address(src)); 982 } else { 983 lea(rscratch1, src); 984 Assembler::addpd(dst, Address(rscratch1, 0)); 985 } 986 } 987 988 void MacroAssembler::align(int modulus) { 989 align(modulus, offset()); 990 } 991 992 void MacroAssembler::align(int modulus, int target) { 993 if (target % modulus != 0) { 994 nop(modulus - (target % modulus)); 995 } 996 } 997 998 void MacroAssembler::andpd(XMMRegister dst, AddressLiteral src, Register scratch_reg) { 999 // Used in sign-masking with aligned address. 1000 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 1001 if (reachable(src)) { 1002 Assembler::andpd(dst, as_Address(src)); 1003 } else { 1004 lea(scratch_reg, src); 1005 Assembler::andpd(dst, Address(scratch_reg, 0)); 1006 } 1007 } 1008 1009 void MacroAssembler::andps(XMMRegister dst, AddressLiteral src, Register scratch_reg) { 1010 // Used in sign-masking with aligned address. 1011 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 1012 if (reachable(src)) { 1013 Assembler::andps(dst, as_Address(src)); 1014 } else { 1015 lea(scratch_reg, src); 1016 Assembler::andps(dst, Address(scratch_reg, 0)); 1017 } 1018 } 1019 1020 void MacroAssembler::andptr(Register dst, int32_t imm32) { 1021 LP64_ONLY(andq(dst, imm32)) NOT_LP64(andl(dst, imm32)); 1022 } 1023 1024 void MacroAssembler::atomic_incl(Address counter_addr) { 1025 lock(); 1026 incrementl(counter_addr); 1027 } 1028 1029 void MacroAssembler::atomic_incl(AddressLiteral counter_addr, Register scr) { 1030 if (reachable(counter_addr)) { 1031 atomic_incl(as_Address(counter_addr)); 1032 } else { 1033 lea(scr, counter_addr); 1034 atomic_incl(Address(scr, 0)); 1035 } 1036 } 1037 1038 #ifdef _LP64 1039 void MacroAssembler::atomic_incq(Address counter_addr) { 1040 lock(); 1041 incrementq(counter_addr); 1042 } 1043 1044 void MacroAssembler::atomic_incq(AddressLiteral counter_addr, Register scr) { 1045 if (reachable(counter_addr)) { 1046 atomic_incq(as_Address(counter_addr)); 1047 } else { 1048 lea(scr, counter_addr); 1049 atomic_incq(Address(scr, 0)); 1050 } 1051 } 1052 #endif 1053 1054 // Writes to stack successive pages until offset reached to check for 1055 // stack overflow + shadow pages. This clobbers tmp. 1056 void MacroAssembler::bang_stack_size(Register size, Register tmp) { 1057 movptr(tmp, rsp); 1058 // Bang stack for total size given plus shadow page size. 1059 // Bang one page at a time because large size can bang beyond yellow and 1060 // red zones. 1061 Label loop; 1062 bind(loop); 1063 movl(Address(tmp, (-os::vm_page_size())), size ); 1064 subptr(tmp, os::vm_page_size()); 1065 subl(size, os::vm_page_size()); 1066 jcc(Assembler::greater, loop); 1067 1068 // Bang down shadow pages too. 1069 // At this point, (tmp-0) is the last address touched, so don't 1070 // touch it again. (It was touched as (tmp-pagesize) but then tmp 1071 // was post-decremented.) Skip this address by starting at i=1, and 1072 // touch a few more pages below. N.B. It is important to touch all 1073 // the way down including all pages in the shadow zone. 1074 for (int i = 1; i < ((int)JavaThread::stack_shadow_zone_size() / os::vm_page_size()); i++) { 1075 // this could be any sized move but this is can be a debugging crumb 1076 // so the bigger the better. 1077 movptr(Address(tmp, (-i*os::vm_page_size())), size ); 1078 } 1079 } 1080 1081 void MacroAssembler::reserved_stack_check() { 1082 // testing if reserved zone needs to be enabled 1083 Label no_reserved_zone_enabling; 1084 Register thread = NOT_LP64(rsi) LP64_ONLY(r15_thread); 1085 NOT_LP64(get_thread(rsi);) 1086 1087 cmpptr(rsp, Address(thread, JavaThread::reserved_stack_activation_offset())); 1088 jcc(Assembler::below, no_reserved_zone_enabling); 1089 1090 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), thread); 1091 jump(RuntimeAddress(StubRoutines::throw_delayed_StackOverflowError_entry())); 1092 should_not_reach_here(); 1093 1094 bind(no_reserved_zone_enabling); 1095 } 1096 1097 int MacroAssembler::biased_locking_enter(Register lock_reg, 1098 Register obj_reg, 1099 Register swap_reg, 1100 Register tmp_reg, 1101 bool swap_reg_contains_mark, 1102 Label& done, 1103 Label* slow_case, 1104 BiasedLockingCounters* counters) { 1105 assert(UseBiasedLocking, "why call this otherwise?"); 1106 assert(swap_reg == rax, "swap_reg must be rax for cmpxchgq"); 1107 assert(tmp_reg != noreg, "tmp_reg must be supplied"); 1108 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg); 1109 assert(markWord::age_shift == markWord::lock_bits + markWord::biased_lock_bits, "biased locking makes assumptions about bit layout"); 1110 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes()); 1111 NOT_LP64( Address saved_mark_addr(lock_reg, 0); ) 1112 1113 if (PrintBiasedLockingStatistics && counters == NULL) { 1114 counters = BiasedLocking::counters(); 1115 } 1116 // Biased locking 1117 // See whether the lock is currently biased toward our thread and 1118 // whether the epoch is still valid 1119 // Note that the runtime guarantees sufficient alignment of JavaThread 1120 // pointers to allow age to be placed into low bits 1121 // First check to see whether biasing is even enabled for this object 1122 Label cas_label; 1123 int null_check_offset = -1; 1124 if (!swap_reg_contains_mark) { 1125 null_check_offset = offset(); 1126 movptr(swap_reg, mark_addr); 1127 } 1128 movptr(tmp_reg, swap_reg); 1129 andptr(tmp_reg, markWord::biased_lock_mask_in_place); 1130 cmpptr(tmp_reg, markWord::biased_lock_pattern); 1131 jcc(Assembler::notEqual, cas_label); 1132 // The bias pattern is present in the object's header. Need to check 1133 // whether the bias owner and the epoch are both still current. 1134 #ifndef _LP64 1135 // Note that because there is no current thread register on x86_32 we 1136 // need to store off the mark word we read out of the object to 1137 // avoid reloading it and needing to recheck invariants below. This 1138 // store is unfortunate but it makes the overall code shorter and 1139 // simpler. 1140 movptr(saved_mark_addr, swap_reg); 1141 #endif 1142 if (swap_reg_contains_mark) { 1143 null_check_offset = offset(); 1144 } 1145 load_prototype_header(tmp_reg, obj_reg); 1146 #ifdef _LP64 1147 orptr(tmp_reg, r15_thread); 1148 xorptr(tmp_reg, swap_reg); 1149 Register header_reg = tmp_reg; 1150 #else 1151 xorptr(tmp_reg, swap_reg); 1152 get_thread(swap_reg); 1153 xorptr(swap_reg, tmp_reg); 1154 Register header_reg = swap_reg; 1155 #endif 1156 andptr(header_reg, ~((int) markWord::age_mask_in_place)); 1157 if (counters != NULL) { 1158 cond_inc32(Assembler::zero, 1159 ExternalAddress((address) counters->biased_lock_entry_count_addr())); 1160 } 1161 jcc(Assembler::equal, done); 1162 1163 Label try_revoke_bias; 1164 Label try_rebias; 1165 1166 // At this point we know that the header has the bias pattern and 1167 // that we are not the bias owner in the current epoch. We need to 1168 // figure out more details about the state of the header in order to 1169 // know what operations can be legally performed on the object's 1170 // header. 1171 1172 // If the low three bits in the xor result aren't clear, that means 1173 // the prototype header is no longer biased and we have to revoke 1174 // the bias on this object. 1175 testptr(header_reg, markWord::biased_lock_mask_in_place); 1176 jccb(Assembler::notZero, try_revoke_bias); 1177 1178 // Biasing is still enabled for this data type. See whether the 1179 // epoch of the current bias is still valid, meaning that the epoch 1180 // bits of the mark word are equal to the epoch bits of the 1181 // prototype header. (Note that the prototype header's epoch bits 1182 // only change at a safepoint.) If not, attempt to rebias the object 1183 // toward the current thread. Note that we must be absolutely sure 1184 // that the current epoch is invalid in order to do this because 1185 // otherwise the manipulations it performs on the mark word are 1186 // illegal. 1187 testptr(header_reg, markWord::epoch_mask_in_place); 1188 jccb(Assembler::notZero, try_rebias); 1189 1190 // The epoch of the current bias is still valid but we know nothing 1191 // about the owner; it might be set or it might be clear. Try to 1192 // acquire the bias of the object using an atomic operation. If this 1193 // fails we will go in to the runtime to revoke the object's bias. 1194 // Note that we first construct the presumed unbiased header so we 1195 // don't accidentally blow away another thread's valid bias. 1196 NOT_LP64( movptr(swap_reg, saved_mark_addr); ) 1197 andptr(swap_reg, 1198 markWord::biased_lock_mask_in_place | markWord::age_mask_in_place | markWord::epoch_mask_in_place); 1199 #ifdef _LP64 1200 movptr(tmp_reg, swap_reg); 1201 orptr(tmp_reg, r15_thread); 1202 #else 1203 get_thread(tmp_reg); 1204 orptr(tmp_reg, swap_reg); 1205 #endif 1206 lock(); 1207 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1208 // If the biasing toward our thread failed, this means that 1209 // another thread succeeded in biasing it toward itself and we 1210 // need to revoke that bias. The revocation will occur in the 1211 // interpreter runtime in the slow case. 1212 if (counters != NULL) { 1213 cond_inc32(Assembler::zero, 1214 ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr())); 1215 } 1216 if (slow_case != NULL) { 1217 jcc(Assembler::notZero, *slow_case); 1218 } 1219 jmp(done); 1220 1221 bind(try_rebias); 1222 // At this point we know the epoch has expired, meaning that the 1223 // current "bias owner", if any, is actually invalid. Under these 1224 // circumstances _only_, we are allowed to use the current header's 1225 // value as the comparison value when doing the cas to acquire the 1226 // bias in the current epoch. In other words, we allow transfer of 1227 // the bias from one thread to another directly in this situation. 1228 // 1229 // FIXME: due to a lack of registers we currently blow away the age 1230 // bits in this situation. Should attempt to preserve them. 1231 load_prototype_header(tmp_reg, obj_reg); 1232 #ifdef _LP64 1233 orptr(tmp_reg, r15_thread); 1234 #else 1235 get_thread(swap_reg); 1236 orptr(tmp_reg, swap_reg); 1237 movptr(swap_reg, saved_mark_addr); 1238 #endif 1239 lock(); 1240 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1241 // If the biasing toward our thread failed, then another thread 1242 // succeeded in biasing it toward itself and we need to revoke that 1243 // bias. The revocation will occur in the runtime in the slow case. 1244 if (counters != NULL) { 1245 cond_inc32(Assembler::zero, 1246 ExternalAddress((address) counters->rebiased_lock_entry_count_addr())); 1247 } 1248 if (slow_case != NULL) { 1249 jcc(Assembler::notZero, *slow_case); 1250 } 1251 jmp(done); 1252 1253 bind(try_revoke_bias); 1254 // The prototype mark in the klass doesn't have the bias bit set any 1255 // more, indicating that objects of this data type are not supposed 1256 // to be biased any more. We are going to try to reset the mark of 1257 // this object to the prototype value and fall through to the 1258 // CAS-based locking scheme. Note that if our CAS fails, it means 1259 // that another thread raced us for the privilege of revoking the 1260 // bias of this particular object, so it's okay to continue in the 1261 // normal locking code. 1262 // 1263 // FIXME: due to a lack of registers we currently blow away the age 1264 // bits in this situation. Should attempt to preserve them. 1265 NOT_LP64( movptr(swap_reg, saved_mark_addr); ) 1266 load_prototype_header(tmp_reg, obj_reg); 1267 lock(); 1268 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1269 // Fall through to the normal CAS-based lock, because no matter what 1270 // the result of the above CAS, some thread must have succeeded in 1271 // removing the bias bit from the object's header. 1272 if (counters != NULL) { 1273 cond_inc32(Assembler::zero, 1274 ExternalAddress((address) counters->revoked_lock_entry_count_addr())); 1275 } 1276 1277 bind(cas_label); 1278 1279 return null_check_offset; 1280 } 1281 1282 void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) { 1283 assert(UseBiasedLocking, "why call this otherwise?"); 1284 1285 // Check for biased locking unlock case, which is a no-op 1286 // Note: we do not have to check the thread ID for two reasons. 1287 // First, the interpreter checks for IllegalMonitorStateException at 1288 // a higher level. Second, if the bias was revoked while we held the 1289 // lock, the object could not be rebiased toward another thread, so 1290 // the bias bit would be clear. 1291 movptr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 1292 andptr(temp_reg, markWord::biased_lock_mask_in_place); 1293 cmpptr(temp_reg, markWord::biased_lock_pattern); 1294 jcc(Assembler::equal, done); 1295 } 1296 1297 #ifdef COMPILER2 1298 1299 // Increment the ObjectMonitor's ref_count for safety or force a branch 1300 // to 'done' with ICC.ZF=0 to indicate failure/take the slow path. 1301 void MacroAssembler::inc_om_ref_count(Register obj_reg, Register om_reg, Register tmp_reg, Label& done) { 1302 atomic_incl(Address(om_reg, OM_OFFSET_NO_MONITOR_VALUE_TAG(ref_count))); 1303 1304 Label LGoSlowPath; 1305 if (AsyncDeflateIdleMonitors) { 1306 // Race here if monitor is not owned! The above ref_count bump 1307 // will cause subsequent async deflation to skip it. However, 1308 // previous or concurrent async deflation is a race. 1309 1310 // First check: if the owner field == DEFLATER_MARKER: 1311 movptr(tmp_reg, Address(om_reg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1312 // DEFLATER_MARKER == reinterpret_cast<void*>(-1) so the compiler 1313 // doesn't like to use the define here: 1314 cmpptr(tmp_reg, -1); 1315 // If marked for async deflation, then take the slow path. This is a 1316 // simpler check than what ObjectMonitorHandle::save_om_ptr() does 1317 // so ObjectMonitor::install_displaced_markword_in_object() doesn't 1318 // have to be implemented in macro assembler. 1319 jccb(Assembler::equal, LGoSlowPath); 1320 1321 // Second check: if ref_count field <= 0: 1322 movptr(tmp_reg, Address(om_reg, OM_OFFSET_NO_MONITOR_VALUE_TAG(ref_count))); 1323 cmpptr(tmp_reg, 0); 1324 // If async deflation is in the process of bailing out, but has not 1325 // yet restored the ref_count field, then we take the slow path. We 1326 // want a stable ref_count value for the fast path. 1327 jccb(Assembler::lessEqual, LGoSlowPath); 1328 1329 // Final check: if object field == obj_reg: 1330 cmpptr(obj_reg, Address(om_reg, OM_OFFSET_NO_MONITOR_VALUE_TAG(object))); 1331 // If the ObjectMonitor has been deflated and recycled, then take 1332 // the slow path. 1333 jccb(Assembler::notEqual, LGoSlowPath); 1334 } 1335 1336 Label LRetToCaller; 1337 // We leave the ref_count incremented to protect the caller's code 1338 // paths against async deflation. 1339 jmpb(LRetToCaller); 1340 1341 bind(LGoSlowPath); 1342 lock(); 1343 decrementl(Address(om_reg, OM_OFFSET_NO_MONITOR_VALUE_TAG(ref_count))); 1344 // Jump to 'done' with ICC.ZF=0 to indicate failure/take the slow path. 1345 orl(tmp_reg, 1); 1346 jmp(done); 1347 1348 bind(LRetToCaller); 1349 } 1350 1351 #if INCLUDE_RTM_OPT 1352 1353 // Update rtm_counters based on abort status 1354 // input: abort_status 1355 // rtm_counters (RTMLockingCounters*) 1356 // flags are killed 1357 void MacroAssembler::rtm_counters_update(Register abort_status, Register rtm_counters) { 1358 1359 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abort_count_offset())); 1360 if (PrintPreciseRTMLockingStatistics) { 1361 for (int i = 0; i < RTMLockingCounters::ABORT_STATUS_LIMIT; i++) { 1362 Label check_abort; 1363 testl(abort_status, (1<<i)); 1364 jccb(Assembler::equal, check_abort); 1365 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abortX_count_offset() + (i * sizeof(uintx)))); 1366 bind(check_abort); 1367 } 1368 } 1369 } 1370 1371 // Branch if (random & (count-1) != 0), count is 2^n 1372 // tmp, scr and flags are killed 1373 void MacroAssembler::branch_on_random_using_rdtsc(Register tmp, Register scr, int count, Label& brLabel) { 1374 assert(tmp == rax, ""); 1375 assert(scr == rdx, ""); 1376 rdtsc(); // modifies EDX:EAX 1377 andptr(tmp, count-1); 1378 jccb(Assembler::notZero, brLabel); 1379 } 1380 1381 // Perform abort ratio calculation, set no_rtm bit if high ratio 1382 // input: rtm_counters_Reg (RTMLockingCounters* address) 1383 // tmpReg, rtm_counters_Reg and flags are killed 1384 void MacroAssembler::rtm_abort_ratio_calculation(Register tmpReg, 1385 Register rtm_counters_Reg, 1386 RTMLockingCounters* rtm_counters, 1387 Metadata* method_data) { 1388 Label L_done, L_check_always_rtm1, L_check_always_rtm2; 1389 1390 if (RTMLockingCalculationDelay > 0) { 1391 // Delay calculation 1392 movptr(tmpReg, ExternalAddress((address) RTMLockingCounters::rtm_calculation_flag_addr()), tmpReg); 1393 testptr(tmpReg, tmpReg); 1394 jccb(Assembler::equal, L_done); 1395 } 1396 // Abort ratio calculation only if abort_count > RTMAbortThreshold 1397 // Aborted transactions = abort_count * 100 1398 // All transactions = total_count * RTMTotalCountIncrRate 1399 // Set no_rtm bit if (Aborted transactions >= All transactions * RTMAbortRatio) 1400 1401 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::abort_count_offset())); 1402 cmpptr(tmpReg, RTMAbortThreshold); 1403 jccb(Assembler::below, L_check_always_rtm2); 1404 imulptr(tmpReg, tmpReg, 100); 1405 1406 Register scrReg = rtm_counters_Reg; 1407 movptr(scrReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset())); 1408 imulptr(scrReg, scrReg, RTMTotalCountIncrRate); 1409 imulptr(scrReg, scrReg, RTMAbortRatio); 1410 cmpptr(tmpReg, scrReg); 1411 jccb(Assembler::below, L_check_always_rtm1); 1412 if (method_data != NULL) { 1413 // set rtm_state to "no rtm" in MDO 1414 mov_metadata(tmpReg, method_data); 1415 lock(); 1416 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), NoRTM); 1417 } 1418 jmpb(L_done); 1419 bind(L_check_always_rtm1); 1420 // Reload RTMLockingCounters* address 1421 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters)); 1422 bind(L_check_always_rtm2); 1423 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset())); 1424 cmpptr(tmpReg, RTMLockingThreshold / RTMTotalCountIncrRate); 1425 jccb(Assembler::below, L_done); 1426 if (method_data != NULL) { 1427 // set rtm_state to "always rtm" in MDO 1428 mov_metadata(tmpReg, method_data); 1429 lock(); 1430 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), UseRTM); 1431 } 1432 bind(L_done); 1433 } 1434 1435 // Update counters and perform abort ratio calculation 1436 // input: abort_status_Reg 1437 // rtm_counters_Reg, flags are killed 1438 void MacroAssembler::rtm_profiling(Register abort_status_Reg, 1439 Register rtm_counters_Reg, 1440 RTMLockingCounters* rtm_counters, 1441 Metadata* method_data, 1442 bool profile_rtm) { 1443 1444 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1445 // update rtm counters based on rax value at abort 1446 // reads abort_status_Reg, updates flags 1447 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters)); 1448 rtm_counters_update(abort_status_Reg, rtm_counters_Reg); 1449 if (profile_rtm) { 1450 // Save abort status because abort_status_Reg is used by following code. 1451 if (RTMRetryCount > 0) { 1452 push(abort_status_Reg); 1453 } 1454 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1455 rtm_abort_ratio_calculation(abort_status_Reg, rtm_counters_Reg, rtm_counters, method_data); 1456 // restore abort status 1457 if (RTMRetryCount > 0) { 1458 pop(abort_status_Reg); 1459 } 1460 } 1461 } 1462 1463 // Retry on abort if abort's status is 0x6: can retry (0x2) | memory conflict (0x4) 1464 // inputs: retry_count_Reg 1465 // : abort_status_Reg 1466 // output: retry_count_Reg decremented by 1 1467 // flags are killed 1468 void MacroAssembler::rtm_retry_lock_on_abort(Register retry_count_Reg, Register abort_status_Reg, Label& retryLabel) { 1469 Label doneRetry; 1470 assert(abort_status_Reg == rax, ""); 1471 // The abort reason bits are in eax (see all states in rtmLocking.hpp) 1472 // 0x6 = conflict on which we can retry (0x2) | memory conflict (0x4) 1473 // if reason is in 0x6 and retry count != 0 then retry 1474 andptr(abort_status_Reg, 0x6); 1475 jccb(Assembler::zero, doneRetry); 1476 testl(retry_count_Reg, retry_count_Reg); 1477 jccb(Assembler::zero, doneRetry); 1478 pause(); 1479 decrementl(retry_count_Reg); 1480 jmp(retryLabel); 1481 bind(doneRetry); 1482 } 1483 1484 // Spin and retry if lock is busy, 1485 // inputs: box_Reg (monitor address) 1486 // : retry_count_Reg 1487 // output: retry_count_Reg decremented by 1 1488 // : clear z flag if retry count exceeded 1489 // tmp_Reg, scr_Reg, flags are killed 1490 void MacroAssembler::rtm_retry_lock_on_busy(Register retry_count_Reg, Register box_Reg, 1491 Register tmp_Reg, Register scr_Reg, Label& retryLabel) { 1492 Label SpinLoop, SpinExit, doneRetry; 1493 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1494 1495 testl(retry_count_Reg, retry_count_Reg); 1496 jccb(Assembler::zero, doneRetry); 1497 decrementl(retry_count_Reg); 1498 movptr(scr_Reg, RTMSpinLoopCount); 1499 1500 bind(SpinLoop); 1501 pause(); 1502 decrementl(scr_Reg); 1503 jccb(Assembler::lessEqual, SpinExit); 1504 movptr(tmp_Reg, Address(box_Reg, owner_offset)); 1505 testptr(tmp_Reg, tmp_Reg); 1506 jccb(Assembler::notZero, SpinLoop); 1507 1508 bind(SpinExit); 1509 jmp(retryLabel); 1510 bind(doneRetry); 1511 incrementl(retry_count_Reg); // clear z flag 1512 } 1513 1514 // Use RTM for normal stack locks 1515 // Input: objReg (object to lock) 1516 void MacroAssembler::rtm_stack_locking(Register objReg, Register tmpReg, Register scrReg, 1517 Register retry_on_abort_count_Reg, 1518 RTMLockingCounters* stack_rtm_counters, 1519 Metadata* method_data, bool profile_rtm, 1520 Label& DONE_LABEL, Label& IsInflated) { 1521 assert(UseRTMForStackLocks, "why call this otherwise?"); 1522 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking"); 1523 assert(tmpReg == rax, ""); 1524 assert(scrReg == rdx, ""); 1525 Label L_rtm_retry, L_decrement_retry, L_on_abort; 1526 1527 if (RTMRetryCount > 0) { 1528 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort 1529 bind(L_rtm_retry); 1530 } 1531 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); 1532 testptr(tmpReg, markWord::monitor_value); // inflated vs stack-locked|neutral|biased 1533 jcc(Assembler::notZero, IsInflated); 1534 1535 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1536 Label L_noincrement; 1537 if (RTMTotalCountIncrRate > 1) { 1538 // tmpReg, scrReg and flags are killed 1539 branch_on_random_using_rdtsc(tmpReg, scrReg, RTMTotalCountIncrRate, L_noincrement); 1540 } 1541 assert(stack_rtm_counters != NULL, "should not be NULL when profiling RTM"); 1542 atomic_incptr(ExternalAddress((address)stack_rtm_counters->total_count_addr()), scrReg); 1543 bind(L_noincrement); 1544 } 1545 xbegin(L_on_abort); 1546 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // fetch markword 1547 andptr(tmpReg, markWord::biased_lock_mask_in_place); // look at 3 lock bits 1548 cmpptr(tmpReg, markWord::unlocked_value); // bits = 001 unlocked 1549 jcc(Assembler::equal, DONE_LABEL); // all done if unlocked 1550 1551 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX 1552 if (UseRTMXendForLockBusy) { 1553 xend(); 1554 movptr(abort_status_Reg, 0x2); // Set the abort status to 2 (so we can retry) 1555 jmp(L_decrement_retry); 1556 } 1557 else { 1558 xabort(0); 1559 } 1560 bind(L_on_abort); 1561 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1562 rtm_profiling(abort_status_Reg, scrReg, stack_rtm_counters, method_data, profile_rtm); 1563 } 1564 bind(L_decrement_retry); 1565 if (RTMRetryCount > 0) { 1566 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4) 1567 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry); 1568 } 1569 } 1570 1571 // Use RTM for inflating locks 1572 // inputs: objReg (object to lock) 1573 // boxReg (on-stack box address (displaced header location) - KILLED) 1574 // tmpReg (ObjectMonitor address + markWord::monitor_value) 1575 void MacroAssembler::rtm_inflated_locking(Register objReg, Register boxReg, Register tmpReg, 1576 Register scrReg, Register retry_on_busy_count_Reg, 1577 Register retry_on_abort_count_Reg, 1578 RTMLockingCounters* rtm_counters, 1579 Metadata* method_data, bool profile_rtm, 1580 Label& DONE_LABEL) { 1581 assert(UseRTMLocking, "why call this otherwise?"); 1582 assert(tmpReg == rax, ""); 1583 assert(scrReg == rdx, ""); 1584 Label L_rtm_retry, L_decrement_retry, L_on_abort, L_local_done; 1585 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1586 1587 // Without cast to int32_t this style of movptr will destroy r10 which is typically obj. 1588 movptr(Address(boxReg, 0), (int32_t)intptr_t(markWord::unused_mark().value())); 1589 1590 if (!HandshakeAfterDeflateIdleMonitors) { 1591 // Increment the ObjectMonitor's ref_count for safety or force the 1592 // enter slow path via DONE_LABEL. 1593 // In rtm_inflated_locking(), initially tmpReg contains the object's 1594 // mark word which, in this case, is the (ObjectMonitor* | monitor_value). 1595 // Also this code uses scrReg as its temporary register. 1596 inc_om_ref_count(objReg, tmpReg /* om_reg */, scrReg /* tmp_reg */, DONE_LABEL); 1597 } 1598 1599 movptr(boxReg, tmpReg); // Save ObjectMonitor address 1600 1601 if (RTMRetryCount > 0) { 1602 movl(retry_on_busy_count_Reg, RTMRetryCount); // Retry on lock busy 1603 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort 1604 bind(L_rtm_retry); 1605 } 1606 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1607 Label L_noincrement; 1608 if (RTMTotalCountIncrRate > 1) { 1609 // tmpReg, scrReg and flags are killed 1610 branch_on_random_using_rdtsc(tmpReg, scrReg, RTMTotalCountIncrRate, L_noincrement); 1611 } 1612 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1613 atomic_incptr(ExternalAddress((address)rtm_counters->total_count_addr()), scrReg); 1614 bind(L_noincrement); 1615 } 1616 xbegin(L_on_abort); 1617 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); 1618 movptr(tmpReg, Address(tmpReg, owner_offset)); 1619 testptr(tmpReg, tmpReg); 1620 jcc(Assembler::zero, L_local_done); 1621 if (UseRTMXendForLockBusy) { 1622 xend(); 1623 jmp(L_decrement_retry); 1624 } 1625 else { 1626 xabort(0); 1627 } 1628 bind(L_on_abort); 1629 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX 1630 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1631 rtm_profiling(abort_status_Reg, scrReg, rtm_counters, method_data, profile_rtm); 1632 } 1633 if (RTMRetryCount > 0) { 1634 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4) 1635 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry); 1636 } 1637 1638 movptr(tmpReg, Address(boxReg, owner_offset)) ; 1639 testptr(tmpReg, tmpReg) ; 1640 jccb(Assembler::notZero, L_decrement_retry) ; 1641 1642 // Appears unlocked - try to swing _owner from null to non-null. 1643 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1644 #ifdef _LP64 1645 Register threadReg = r15_thread; 1646 #else 1647 get_thread(scrReg); 1648 Register threadReg = scrReg; 1649 #endif 1650 lock(); 1651 cmpxchgptr(threadReg, Address(boxReg, owner_offset)); // Updates tmpReg 1652 1653 if (RTMRetryCount > 0) { 1654 // success done else retry 1655 jccb(Assembler::equal, L_local_done); 1656 bind(L_decrement_retry); 1657 // Spin and retry if lock is busy. 1658 rtm_retry_lock_on_busy(retry_on_busy_count_Reg, boxReg, tmpReg, scrReg, L_rtm_retry); 1659 } 1660 else { 1661 bind(L_decrement_retry); 1662 } 1663 1664 // rtm_inflated_locking() exit paths come here except for a failed 1665 // inc_om_ref_count() which goes directly to DONE_LABEL. 1666 bind(L_local_done); 1667 if (!HandshakeAfterDeflateIdleMonitors) { 1668 pushf(); // Preserve flags. 1669 // Decrement the ObjectMonitor's ref_count. 1670 lock(); 1671 decrementl(Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(ref_count))); 1672 popf(); // Restore flags so we have the proper ICC.ZF value. 1673 } 1674 1675 jmp(DONE_LABEL) ; 1676 } 1677 1678 #endif // INCLUDE_RTM_OPT 1679 1680 // fast_lock and fast_unlock used by C2 1681 1682 // Because the transitions from emitted code to the runtime 1683 // monitorenter/exit helper stubs are so slow it's critical that 1684 // we inline both the stack-locking fast path and the inflated fast path. 1685 // 1686 // See also: cmpFastLock and cmpFastUnlock. 1687 // 1688 // What follows is a specialized inline transliteration of the code 1689 // in enter() and exit(). If we're concerned about I$ bloat another 1690 // option would be to emit TrySlowEnter and TrySlowExit methods 1691 // at startup-time. These methods would accept arguments as 1692 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure 1693 // indications in the icc.ZFlag. fast_lock and fast_unlock would simply 1694 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit. 1695 // In practice, however, the # of lock sites is bounded and is usually small. 1696 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer 1697 // if the processor uses simple bimodal branch predictors keyed by EIP 1698 // Since the helper routines would be called from multiple synchronization 1699 // sites. 1700 // 1701 // An even better approach would be write "MonitorEnter()" and "MonitorExit()" 1702 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites 1703 // to those specialized methods. That'd give us a mostly platform-independent 1704 // implementation that the JITs could optimize and inline at their pleasure. 1705 // Done correctly, the only time we'd need to cross to native could would be 1706 // to park() or unpark() threads. We'd also need a few more unsafe operators 1707 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and 1708 // (b) explicit barriers or fence operations. 1709 // 1710 // TODO: 1711 // 1712 // * Arrange for C2 to pass "Self" into fast_lock and fast_unlock in one of the registers (scr). 1713 // This avoids manifesting the Self pointer in the fast_lock and fast_unlock terminals. 1714 // Given TLAB allocation, Self is usually manifested in a register, so passing it into 1715 // the lock operators would typically be faster than reifying Self. 1716 // 1717 // * Ideally I'd define the primitives as: 1718 // fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED. 1719 // fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED 1720 // Unfortunately ADLC bugs prevent us from expressing the ideal form. 1721 // Instead, we're stuck with a rather awkward and brittle register assignments below. 1722 // Furthermore the register assignments are overconstrained, possibly resulting in 1723 // sub-optimal code near the synchronization site. 1724 // 1725 // * Eliminate the sp-proximity tests and just use "== Self" tests instead. 1726 // Alternately, use a better sp-proximity test. 1727 // 1728 // * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value. 1729 // Either one is sufficient to uniquely identify a thread. 1730 // TODO: eliminate use of sp in _owner and use get_thread(tr) instead. 1731 // 1732 // * Intrinsify notify() and notifyAll() for the common cases where the 1733 // object is locked by the calling thread but the waitlist is empty. 1734 // avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll(). 1735 // 1736 // * use jccb and jmpb instead of jcc and jmp to improve code density. 1737 // But beware of excessive branch density on AMD Opterons. 1738 // 1739 // * Both fast_lock and fast_unlock set the ICC.ZF to indicate success 1740 // or failure of the fast path. If the fast path fails then we pass 1741 // control to the slow path, typically in C. In fast_lock and 1742 // fast_unlock we often branch to DONE_LABEL, just to find that C2 1743 // will emit a conditional branch immediately after the node. 1744 // So we have branches to branches and lots of ICC.ZF games. 1745 // Instead, it might be better to have C2 pass a "FailureLabel" 1746 // into fast_lock and fast_unlock. In the case of success, control 1747 // will drop through the node. ICC.ZF is undefined at exit. 1748 // In the case of failure, the node will branch directly to the 1749 // FailureLabel 1750 1751 1752 // obj: object to lock 1753 // box: on-stack box address (displaced header location) - KILLED 1754 // rax,: tmp -- KILLED 1755 // scr: tmp -- KILLED 1756 void MacroAssembler::fast_lock(Register objReg, Register boxReg, Register tmpReg, 1757 Register scrReg, Register cx1Reg, Register cx2Reg, 1758 BiasedLockingCounters* counters, 1759 RTMLockingCounters* rtm_counters, 1760 RTMLockingCounters* stack_rtm_counters, 1761 Metadata* method_data, 1762 bool use_rtm, bool profile_rtm) { 1763 // Ensure the register assignments are disjoint 1764 assert(tmpReg == rax, ""); 1765 1766 if (use_rtm) { 1767 assert_different_registers(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg); 1768 } else { 1769 assert(cx1Reg == noreg, ""); 1770 assert(cx2Reg == noreg, ""); 1771 assert_different_registers(objReg, boxReg, tmpReg, scrReg); 1772 } 1773 1774 if (counters != NULL) { 1775 atomic_incl(ExternalAddress((address)counters->total_entry_count_addr()), scrReg); 1776 } 1777 1778 // Possible cases that we'll encounter in fast_lock 1779 // ------------------------------------------------ 1780 // * Inflated 1781 // -- unlocked 1782 // -- Locked 1783 // = by self 1784 // = by other 1785 // * biased 1786 // -- by Self 1787 // -- by other 1788 // * neutral 1789 // * stack-locked 1790 // -- by self 1791 // = sp-proximity test hits 1792 // = sp-proximity test generates false-negative 1793 // -- by other 1794 // 1795 1796 Label IsInflated, DONE_LABEL; 1797 1798 // it's stack-locked, biased or neutral 1799 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage 1800 // order to reduce the number of conditional branches in the most common cases. 1801 // Beware -- there's a subtle invariant that fetch of the markword 1802 // at [FETCH], below, will never observe a biased encoding (*101b). 1803 // If this invariant is not held we risk exclusion (safety) failure. 1804 if (UseBiasedLocking && !UseOptoBiasInlining) { 1805 biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, counters); 1806 } 1807 1808 #if INCLUDE_RTM_OPT 1809 if (UseRTMForStackLocks && use_rtm) { 1810 rtm_stack_locking(objReg, tmpReg, scrReg, cx2Reg, 1811 stack_rtm_counters, method_data, profile_rtm, 1812 DONE_LABEL, IsInflated); 1813 } 1814 #endif // INCLUDE_RTM_OPT 1815 1816 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // [FETCH] 1817 testptr(tmpReg, markWord::monitor_value); // inflated vs stack-locked|neutral|biased 1818 jccb(Assembler::notZero, IsInflated); 1819 1820 // Attempt stack-locking ... 1821 orptr (tmpReg, markWord::unlocked_value); 1822 movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS 1823 lock(); 1824 cmpxchgptr(boxReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Updates tmpReg 1825 if (counters != NULL) { 1826 cond_inc32(Assembler::equal, 1827 ExternalAddress((address)counters->fast_path_entry_count_addr())); 1828 } 1829 jcc(Assembler::equal, DONE_LABEL); // Success 1830 1831 // Recursive locking. 1832 // The object is stack-locked: markword contains stack pointer to BasicLock. 1833 // Locked by current thread if difference with current SP is less than one page. 1834 subptr(tmpReg, rsp); 1835 // Next instruction set ZFlag == 1 (Success) if difference is less then one page. 1836 andptr(tmpReg, (int32_t) (NOT_LP64(0xFFFFF003) LP64_ONLY(7 - os::vm_page_size())) ); 1837 movptr(Address(boxReg, 0), tmpReg); 1838 if (counters != NULL) { 1839 cond_inc32(Assembler::equal, 1840 ExternalAddress((address)counters->fast_path_entry_count_addr())); 1841 } 1842 jmp(DONE_LABEL); 1843 1844 bind(IsInflated); 1845 // The object is inflated. tmpReg contains pointer to ObjectMonitor* + markWord::monitor_value 1846 1847 #if INCLUDE_RTM_OPT 1848 // Use the same RTM locking code in 32- and 64-bit VM. 1849 if (use_rtm) { 1850 rtm_inflated_locking(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg, 1851 rtm_counters, method_data, profile_rtm, DONE_LABEL); 1852 } else { 1853 #endif // INCLUDE_RTM_OPT 1854 1855 #ifndef _LP64 1856 // The object is inflated. 1857 1858 // boxReg refers to the on-stack BasicLock in the current frame. 1859 // We'd like to write: 1860 // set box->_displaced_header = markWord::unused_mark(). Any non-0 value suffices. 1861 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers 1862 // additional latency as we have another ST in the store buffer that must drain. 1863 1864 // avoid ST-before-CAS 1865 // register juggle because we need tmpReg for cmpxchgptr below 1866 movptr(scrReg, boxReg); 1867 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2] 1868 1869 // Optimistic form: consider XORL tmpReg,tmpReg 1870 movptr(tmpReg, NULL_WORD); 1871 1872 // Appears unlocked - try to swing _owner from null to non-null. 1873 // Ideally, I'd manifest "Self" with get_thread and then attempt 1874 // to CAS the register containing Self into m->Owner. 1875 // But we don't have enough registers, so instead we can either try to CAS 1876 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds 1877 // we later store "Self" into m->Owner. Transiently storing a stack address 1878 // (rsp or the address of the box) into m->owner is harmless. 1879 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1880 lock(); 1881 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1882 movptr(Address(scrReg, 0), 3); // box->_displaced_header = 3 1883 // If we weren't able to swing _owner from NULL to the BasicLock 1884 // then take the slow path. 1885 jccb (Assembler::notZero, DONE_LABEL); 1886 // update _owner from BasicLock to thread 1887 get_thread (scrReg); // beware: clobbers ICCs 1888 movptr(Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), scrReg); 1889 xorptr(boxReg, boxReg); // set icc.ZFlag = 1 to indicate success 1890 1891 // If the CAS fails we can either retry or pass control to the slow path. 1892 // We use the latter tactic. 1893 // Pass the CAS result in the icc.ZFlag into DONE_LABEL 1894 // If the CAS was successful ... 1895 // Self has acquired the lock 1896 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it. 1897 // Intentional fall-through into DONE_LABEL ... 1898 #else // _LP64 1899 // It's inflated and we use scrReg for ObjectMonitor* in this section. 1900 movq(scrReg, tmpReg); 1901 1902 // Unconditionally set box->_displaced_header = markWord::unused_mark(). 1903 // Without cast to int32_t this style of movptr will destroy r10 which is typically obj. 1904 movptr(Address(boxReg, 0), (int32_t)intptr_t(markWord::unused_mark().value())); 1905 1906 if (!HandshakeAfterDeflateIdleMonitors) { 1907 // Increment the ObjectMonitor's ref_count for safety or force the 1908 // enter slow path via DONE_LABEL. 1909 // In fast_lock(), scrReg contains the object's mark word which, 1910 // in this case, is the (ObjectMonitor* | monitor_value). Also this 1911 // code uses tmpReg as its temporary register. 1912 inc_om_ref_count(objReg, scrReg /* om_reg */, tmpReg /* tmp_reg */, DONE_LABEL); 1913 } 1914 1915 xorq(tmpReg, tmpReg); 1916 lock(); 1917 cmpxchgptr(r15_thread, Address(scrReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1918 // Intentional fall-through into DONE_LABEL ... 1919 // Propagate ICC.ZF from CAS above into DONE_LABEL. 1920 1921 if (!HandshakeAfterDeflateIdleMonitors) { 1922 pushf(); // Preserve flags. 1923 // Decrement the ObjectMonitor's ref_count. 1924 lock(); 1925 decrementl(Address(scrReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(ref_count))); 1926 popf(); // Restore flags so we have the proper ICC.ZF value. 1927 } 1928 #endif // _LP64 1929 #if INCLUDE_RTM_OPT 1930 } // use_rtm() 1931 #endif 1932 // DONE_LABEL is a hot target - we'd really like to place it at the 1933 // start of cache line by padding with NOPs. 1934 // See the AMD and Intel software optimization manuals for the 1935 // most efficient "long" NOP encodings. 1936 // Unfortunately none of our alignment mechanisms suffice. 1937 bind(DONE_LABEL); 1938 1939 // At DONE_LABEL the icc ZFlag is set as follows ... 1940 // fast_unlock uses the same protocol. 1941 // ZFlag == 1 -> Success 1942 // ZFlag == 0 -> Failure - force control through the slow path 1943 } 1944 1945 // obj: object to unlock 1946 // box: box address (displaced header location), killed. Must be EAX. 1947 // tmp: killed, cannot be obj nor box. 1948 // 1949 // Some commentary on balanced locking: 1950 // 1951 // fast_lock and fast_unlock are emitted only for provably balanced lock sites. 1952 // Methods that don't have provably balanced locking are forced to run in the 1953 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock. 1954 // The interpreter provides two properties: 1955 // I1: At return-time the interpreter automatically and quietly unlocks any 1956 // objects acquired the current activation (frame). Recall that the 1957 // interpreter maintains an on-stack list of locks currently held by 1958 // a frame. 1959 // I2: If a method attempts to unlock an object that is not held by the 1960 // the frame the interpreter throws IMSX. 1961 // 1962 // Lets say A(), which has provably balanced locking, acquires O and then calls B(). 1963 // B() doesn't have provably balanced locking so it runs in the interpreter. 1964 // Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O 1965 // is still locked by A(). 1966 // 1967 // The only other source of unbalanced locking would be JNI. The "Java Native Interface: 1968 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter 1969 // should not be unlocked by "normal" java-level locking and vice-versa. The specification 1970 // doesn't specify what will occur if a program engages in such mixed-mode locking, however. 1971 // Arguably given that the spec legislates the JNI case as undefined our implementation 1972 // could reasonably *avoid* checking owner in fast_unlock(). 1973 // In the interest of performance we elide m->Owner==Self check in unlock. 1974 // A perfectly viable alternative is to elide the owner check except when 1975 // Xcheck:jni is enabled. 1976 1977 void MacroAssembler::fast_unlock(Register objReg, Register boxReg, Register tmpReg, bool use_rtm) { 1978 assert(boxReg == rax, ""); 1979 assert_different_registers(objReg, boxReg, tmpReg); 1980 1981 Label DONE_LABEL, Stacked, CheckSucc; 1982 1983 // Critically, the biased locking test must have precedence over 1984 // and appear before the (box->dhw == 0) recursive stack-lock test. 1985 if (UseBiasedLocking && !UseOptoBiasInlining) { 1986 biased_locking_exit(objReg, tmpReg, DONE_LABEL); 1987 } 1988 1989 #if INCLUDE_RTM_OPT 1990 if (UseRTMForStackLocks && use_rtm) { 1991 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking"); 1992 Label L_regular_unlock; 1993 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // fetch markword 1994 andptr(tmpReg, markWord::biased_lock_mask_in_place); // look at 3 lock bits 1995 cmpptr(tmpReg, markWord::unlocked_value); // bits = 001 unlocked 1996 jccb(Assembler::notEqual, L_regular_unlock); // if !HLE RegularLock 1997 xend(); // otherwise end... 1998 jmp(DONE_LABEL); // ... and we're done 1999 bind(L_regular_unlock); 2000 } 2001 #endif 2002 2003 cmpptr(Address(boxReg, 0), (int32_t)NULL_WORD); // Examine the displaced header 2004 jcc (Assembler::zero, DONE_LABEL); // 0 indicates recursive stack-lock 2005 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Examine the object's markword 2006 testptr(tmpReg, markWord::monitor_value); // Inflated? 2007 jcc (Assembler::zero, Stacked); 2008 2009 // It's inflated. 2010 #if INCLUDE_RTM_OPT 2011 if (use_rtm) { 2012 Label L_regular_inflated_unlock; 2013 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 2014 movptr(boxReg, Address(tmpReg, owner_offset)); 2015 testptr(boxReg, boxReg); 2016 jccb(Assembler::notZero, L_regular_inflated_unlock); 2017 xend(); 2018 jmp(DONE_LABEL); 2019 bind(L_regular_inflated_unlock); 2020 } 2021 #endif 2022 2023 // Despite our balanced locking property we still check that m->_owner == Self 2024 // as java routines or native JNI code called by this thread might 2025 // have released the lock. 2026 // Refer to the comments in synchronizer.cpp for how we might encode extra 2027 // state in _succ so we can avoid fetching EntryList|cxq. 2028 // 2029 // I'd like to add more cases in fast_lock() and fast_unlock() -- 2030 // such as recursive enter and exit -- but we have to be wary of 2031 // I$ bloat, T$ effects and BP$ effects. 2032 // 2033 // If there's no contention try a 1-0 exit. That is, exit without 2034 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how 2035 // we detect and recover from the race that the 1-0 exit admits. 2036 // 2037 // Conceptually fast_unlock() must execute a STST|LDST "release" barrier 2038 // before it STs null into _owner, releasing the lock. Updates 2039 // to data protected by the critical section must be visible before 2040 // we drop the lock (and thus before any other thread could acquire 2041 // the lock and observe the fields protected by the lock). 2042 // IA32's memory-model is SPO, so STs are ordered with respect to 2043 // each other and there's no need for an explicit barrier (fence). 2044 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 2045 #ifndef _LP64 2046 get_thread (boxReg); 2047 2048 // Note that we could employ various encoding schemes to reduce 2049 // the number of loads below (currently 4) to just 2 or 3. 2050 // Refer to the comments in synchronizer.cpp. 2051 // In practice the chain of fetches doesn't seem to impact performance, however. 2052 xorptr(boxReg, boxReg); 2053 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 2054 jccb (Assembler::notZero, DONE_LABEL); 2055 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2056 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2057 jccb (Assembler::notZero, CheckSucc); 2058 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD); 2059 jmpb (DONE_LABEL); 2060 2061 bind (Stacked); 2062 // It's not inflated and it's not recursively stack-locked and it's not biased. 2063 // It must be stack-locked. 2064 // Try to reset the header to displaced header. 2065 // The "box" value on the stack is stable, so we can reload 2066 // and be assured we observe the same value as above. 2067 movptr(tmpReg, Address(boxReg, 0)); 2068 lock(); 2069 cmpxchgptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Uses RAX which is box 2070 // Intention fall-thru into DONE_LABEL 2071 2072 // DONE_LABEL is a hot target - we'd really like to place it at the 2073 // start of cache line by padding with NOPs. 2074 // See the AMD and Intel software optimization manuals for the 2075 // most efficient "long" NOP encodings. 2076 // Unfortunately none of our alignment mechanisms suffice. 2077 bind (CheckSucc); 2078 #else // _LP64 2079 // It's inflated 2080 2081 if (!HandshakeAfterDeflateIdleMonitors) { 2082 // Increment the ObjectMonitor's ref_count for safety or force the 2083 // exit slow path via DONE_LABEL. 2084 // In fast_unlock(), tmpReg contains the object's mark word which, 2085 // in this case, is the (ObjectMonitor* | monitor_value). Also this 2086 // code uses boxReg as its temporary register. 2087 inc_om_ref_count(objReg, tmpReg /* om_reg */, boxReg /* tmp_reg */, DONE_LABEL); 2088 } 2089 2090 // Try to avoid passing control into the slow path ... 2091 Label LSuccess, LGoSlowPath; 2092 xorptr(boxReg, boxReg); 2093 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 2094 jccb(Assembler::notZero, LGoSlowPath); 2095 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2096 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2097 jccb (Assembler::notZero, CheckSucc); 2098 // Without cast to int32_t this style of movptr will destroy r10 which is typically obj. 2099 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD); 2100 jmpb(LSuccess); 2101 2102 bind (CheckSucc); 2103 2104 // The following optional optimization can be elided if necessary 2105 // Effectively: if (succ == null) goto slow path 2106 // The code reduces the window for a race, however, 2107 // and thus benefits performance. 2108 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2109 jccb (Assembler::zero, LGoSlowPath); 2110 2111 xorptr(boxReg, boxReg); 2112 // Without cast to int32_t this style of movptr will destroy r10 which is typically obj. 2113 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD); 2114 2115 // Memory barrier/fence 2116 // Dekker pivot point -- fulcrum : ST Owner; MEMBAR; LD Succ 2117 // Instead of MFENCE we use a dummy locked add of 0 to the top-of-stack. 2118 // This is faster on Nehalem and AMD Shanghai/Barcelona. 2119 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences 2120 // We might also restructure (ST Owner=0;barrier;LD _Succ) to 2121 // (mov box,0; xchgq box, &m->Owner; LD _succ) . 2122 lock(); addl(Address(rsp, 0), 0); 2123 2124 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2125 jccb (Assembler::notZero, LSuccess); 2126 2127 // Rare inopportune interleaving - race. 2128 // The successor vanished in the small window above. 2129 // The lock is contended -- (cxq|EntryList) != null -- and there's no apparent successor. 2130 // We need to ensure progress and succession. 2131 // Try to reacquire the lock. 2132 // If that fails then the new owner is responsible for succession and this 2133 // thread needs to take no further action and can exit via the fast path (success). 2134 // If the re-acquire succeeds then pass control into the slow path. 2135 // As implemented, this latter mode is horrible because we generated more 2136 // coherence traffic on the lock *and* artifically extended the critical section 2137 // length while by virtue of passing control into the slow path. 2138 2139 // box is really RAX -- the following CMPXCHG depends on that binding 2140 // cmpxchg R,[M] is equivalent to rax = CAS(M,rax,R) 2141 lock(); 2142 cmpxchgptr(r15_thread, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2143 // There's no successor so we tried to regrab the lock. 2144 // If that didn't work, then another thread grabbed the 2145 // lock so we're done (and exit was a success). 2146 jccb (Assembler::notEqual, LSuccess); 2147 // Intentional fall-through into slow path 2148 2149 bind (LGoSlowPath); 2150 if (!HandshakeAfterDeflateIdleMonitors) { 2151 lock(); 2152 decrementl(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(ref_count))); 2153 } 2154 orl (boxReg, 1); // set ICC.ZF=0 to indicate failure 2155 jmpb (DONE_LABEL); 2156 2157 bind (LSuccess); 2158 if (!HandshakeAfterDeflateIdleMonitors) { 2159 lock(); 2160 decrementl(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(ref_count))); 2161 } 2162 testl (boxReg, 0); // set ICC.ZF=1 to indicate success 2163 jmpb (DONE_LABEL); 2164 2165 bind (Stacked); 2166 movptr(tmpReg, Address (boxReg, 0)); // re-fetch 2167 lock(); 2168 cmpxchgptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Uses RAX which is box 2169 2170 #endif 2171 bind(DONE_LABEL); 2172 } 2173 #endif // COMPILER2 2174 2175 void MacroAssembler::c2bool(Register x) { 2176 // implements x == 0 ? 0 : 1 2177 // note: must only look at least-significant byte of x 2178 // since C-style booleans are stored in one byte 2179 // only! (was bug) 2180 andl(x, 0xFF); 2181 setb(Assembler::notZero, x); 2182 } 2183 2184 // Wouldn't need if AddressLiteral version had new name 2185 void MacroAssembler::call(Label& L, relocInfo::relocType rtype) { 2186 Assembler::call(L, rtype); 2187 } 2188 2189 void MacroAssembler::call(Register entry) { 2190 Assembler::call(entry); 2191 } 2192 2193 void MacroAssembler::call(AddressLiteral entry) { 2194 if (reachable(entry)) { 2195 Assembler::call_literal(entry.target(), entry.rspec()); 2196 } else { 2197 lea(rscratch1, entry); 2198 Assembler::call(rscratch1); 2199 } 2200 } 2201 2202 void MacroAssembler::ic_call(address entry, jint method_index) { 2203 RelocationHolder rh = virtual_call_Relocation::spec(pc(), method_index); 2204 movptr(rax, (intptr_t)Universe::non_oop_word()); 2205 call(AddressLiteral(entry, rh)); 2206 } 2207 2208 // Implementation of call_VM versions 2209 2210 void MacroAssembler::call_VM(Register oop_result, 2211 address entry_point, 2212 bool check_exceptions) { 2213 Label C, E; 2214 call(C, relocInfo::none); 2215 jmp(E); 2216 2217 bind(C); 2218 call_VM_helper(oop_result, entry_point, 0, check_exceptions); 2219 ret(0); 2220 2221 bind(E); 2222 } 2223 2224 void MacroAssembler::call_VM(Register oop_result, 2225 address entry_point, 2226 Register arg_1, 2227 bool check_exceptions) { 2228 Label C, E; 2229 call(C, relocInfo::none); 2230 jmp(E); 2231 2232 bind(C); 2233 pass_arg1(this, arg_1); 2234 call_VM_helper(oop_result, entry_point, 1, check_exceptions); 2235 ret(0); 2236 2237 bind(E); 2238 } 2239 2240 void MacroAssembler::call_VM(Register oop_result, 2241 address entry_point, 2242 Register arg_1, 2243 Register arg_2, 2244 bool check_exceptions) { 2245 Label C, E; 2246 call(C, relocInfo::none); 2247 jmp(E); 2248 2249 bind(C); 2250 2251 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2252 2253 pass_arg2(this, arg_2); 2254 pass_arg1(this, arg_1); 2255 call_VM_helper(oop_result, entry_point, 2, check_exceptions); 2256 ret(0); 2257 2258 bind(E); 2259 } 2260 2261 void MacroAssembler::call_VM(Register oop_result, 2262 address entry_point, 2263 Register arg_1, 2264 Register arg_2, 2265 Register arg_3, 2266 bool check_exceptions) { 2267 Label C, E; 2268 call(C, relocInfo::none); 2269 jmp(E); 2270 2271 bind(C); 2272 2273 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2274 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2275 pass_arg3(this, arg_3); 2276 2277 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2278 pass_arg2(this, arg_2); 2279 2280 pass_arg1(this, arg_1); 2281 call_VM_helper(oop_result, entry_point, 3, check_exceptions); 2282 ret(0); 2283 2284 bind(E); 2285 } 2286 2287 void MacroAssembler::call_VM(Register oop_result, 2288 Register last_java_sp, 2289 address entry_point, 2290 int number_of_arguments, 2291 bool check_exceptions) { 2292 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg); 2293 call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions); 2294 } 2295 2296 void MacroAssembler::call_VM(Register oop_result, 2297 Register last_java_sp, 2298 address entry_point, 2299 Register arg_1, 2300 bool check_exceptions) { 2301 pass_arg1(this, arg_1); 2302 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 2303 } 2304 2305 void MacroAssembler::call_VM(Register oop_result, 2306 Register last_java_sp, 2307 address entry_point, 2308 Register arg_1, 2309 Register arg_2, 2310 bool check_exceptions) { 2311 2312 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2313 pass_arg2(this, arg_2); 2314 pass_arg1(this, arg_1); 2315 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 2316 } 2317 2318 void MacroAssembler::call_VM(Register oop_result, 2319 Register last_java_sp, 2320 address entry_point, 2321 Register arg_1, 2322 Register arg_2, 2323 Register arg_3, 2324 bool check_exceptions) { 2325 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2326 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2327 pass_arg3(this, arg_3); 2328 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2329 pass_arg2(this, arg_2); 2330 pass_arg1(this, arg_1); 2331 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 2332 } 2333 2334 void MacroAssembler::super_call_VM(Register oop_result, 2335 Register last_java_sp, 2336 address entry_point, 2337 int number_of_arguments, 2338 bool check_exceptions) { 2339 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg); 2340 MacroAssembler::call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions); 2341 } 2342 2343 void MacroAssembler::super_call_VM(Register oop_result, 2344 Register last_java_sp, 2345 address entry_point, 2346 Register arg_1, 2347 bool check_exceptions) { 2348 pass_arg1(this, arg_1); 2349 super_call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 2350 } 2351 2352 void MacroAssembler::super_call_VM(Register oop_result, 2353 Register last_java_sp, 2354 address entry_point, 2355 Register arg_1, 2356 Register arg_2, 2357 bool check_exceptions) { 2358 2359 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2360 pass_arg2(this, arg_2); 2361 pass_arg1(this, arg_1); 2362 super_call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 2363 } 2364 2365 void MacroAssembler::super_call_VM(Register oop_result, 2366 Register last_java_sp, 2367 address entry_point, 2368 Register arg_1, 2369 Register arg_2, 2370 Register arg_3, 2371 bool check_exceptions) { 2372 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2373 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2374 pass_arg3(this, arg_3); 2375 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2376 pass_arg2(this, arg_2); 2377 pass_arg1(this, arg_1); 2378 super_call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 2379 } 2380 2381 void MacroAssembler::call_VM_base(Register oop_result, 2382 Register java_thread, 2383 Register last_java_sp, 2384 address entry_point, 2385 int number_of_arguments, 2386 bool check_exceptions) { 2387 // determine java_thread register 2388 if (!java_thread->is_valid()) { 2389 #ifdef _LP64 2390 java_thread = r15_thread; 2391 #else 2392 java_thread = rdi; 2393 get_thread(java_thread); 2394 #endif // LP64 2395 } 2396 // determine last_java_sp register 2397 if (!last_java_sp->is_valid()) { 2398 last_java_sp = rsp; 2399 } 2400 // debugging support 2401 assert(number_of_arguments >= 0 , "cannot have negative number of arguments"); 2402 LP64_ONLY(assert(java_thread == r15_thread, "unexpected register")); 2403 #ifdef ASSERT 2404 // TraceBytecodes does not use r12 but saves it over the call, so don't verify 2405 // r12 is the heapbase. 2406 LP64_ONLY(if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");) 2407 #endif // ASSERT 2408 2409 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result"); 2410 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp"); 2411 2412 // push java thread (becomes first argument of C function) 2413 2414 NOT_LP64(push(java_thread); number_of_arguments++); 2415 LP64_ONLY(mov(c_rarg0, r15_thread)); 2416 2417 // set last Java frame before call 2418 assert(last_java_sp != rbp, "can't use ebp/rbp"); 2419 2420 // Only interpreter should have to set fp 2421 set_last_Java_frame(java_thread, last_java_sp, rbp, NULL); 2422 2423 // do the call, remove parameters 2424 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments); 2425 2426 // restore the thread (cannot use the pushed argument since arguments 2427 // may be overwritten by C code generated by an optimizing compiler); 2428 // however can use the register value directly if it is callee saved. 2429 if (LP64_ONLY(true ||) java_thread == rdi || java_thread == rsi) { 2430 // rdi & rsi (also r15) are callee saved -> nothing to do 2431 #ifdef ASSERT 2432 guarantee(java_thread != rax, "change this code"); 2433 push(rax); 2434 { Label L; 2435 get_thread(rax); 2436 cmpptr(java_thread, rax); 2437 jcc(Assembler::equal, L); 2438 STOP("MacroAssembler::call_VM_base: rdi not callee saved?"); 2439 bind(L); 2440 } 2441 pop(rax); 2442 #endif 2443 } else { 2444 get_thread(java_thread); 2445 } 2446 // reset last Java frame 2447 // Only interpreter should have to clear fp 2448 reset_last_Java_frame(java_thread, true); 2449 2450 // C++ interp handles this in the interpreter 2451 check_and_handle_popframe(java_thread); 2452 check_and_handle_earlyret(java_thread); 2453 2454 if (check_exceptions) { 2455 // check for pending exceptions (java_thread is set upon return) 2456 cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t) NULL_WORD); 2457 #ifndef _LP64 2458 jump_cc(Assembler::notEqual, 2459 RuntimeAddress(StubRoutines::forward_exception_entry())); 2460 #else 2461 // This used to conditionally jump to forward_exception however it is 2462 // possible if we relocate that the branch will not reach. So we must jump 2463 // around so we can always reach 2464 2465 Label ok; 2466 jcc(Assembler::equal, ok); 2467 jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2468 bind(ok); 2469 #endif // LP64 2470 } 2471 2472 // get oop result if there is one and reset the value in the thread 2473 if (oop_result->is_valid()) { 2474 get_vm_result(oop_result, java_thread); 2475 } 2476 } 2477 2478 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) { 2479 2480 // Calculate the value for last_Java_sp 2481 // somewhat subtle. call_VM does an intermediate call 2482 // which places a return address on the stack just under the 2483 // stack pointer as the user finsihed with it. This allows 2484 // use to retrieve last_Java_pc from last_Java_sp[-1]. 2485 // On 32bit we then have to push additional args on the stack to accomplish 2486 // the actual requested call. On 64bit call_VM only can use register args 2487 // so the only extra space is the return address that call_VM created. 2488 // This hopefully explains the calculations here. 2489 2490 #ifdef _LP64 2491 // We've pushed one address, correct last_Java_sp 2492 lea(rax, Address(rsp, wordSize)); 2493 #else 2494 lea(rax, Address(rsp, (1 + number_of_arguments) * wordSize)); 2495 #endif // LP64 2496 2497 call_VM_base(oop_result, noreg, rax, entry_point, number_of_arguments, check_exceptions); 2498 2499 } 2500 2501 // Use this method when MacroAssembler version of call_VM_leaf_base() should be called from Interpreter. 2502 void MacroAssembler::call_VM_leaf0(address entry_point) { 2503 MacroAssembler::call_VM_leaf_base(entry_point, 0); 2504 } 2505 2506 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) { 2507 call_VM_leaf_base(entry_point, number_of_arguments); 2508 } 2509 2510 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) { 2511 pass_arg0(this, arg_0); 2512 call_VM_leaf(entry_point, 1); 2513 } 2514 2515 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { 2516 2517 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2518 pass_arg1(this, arg_1); 2519 pass_arg0(this, arg_0); 2520 call_VM_leaf(entry_point, 2); 2521 } 2522 2523 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { 2524 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2525 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2526 pass_arg2(this, arg_2); 2527 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2528 pass_arg1(this, arg_1); 2529 pass_arg0(this, arg_0); 2530 call_VM_leaf(entry_point, 3); 2531 } 2532 2533 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) { 2534 pass_arg0(this, arg_0); 2535 MacroAssembler::call_VM_leaf_base(entry_point, 1); 2536 } 2537 2538 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { 2539 2540 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2541 pass_arg1(this, arg_1); 2542 pass_arg0(this, arg_0); 2543 MacroAssembler::call_VM_leaf_base(entry_point, 2); 2544 } 2545 2546 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { 2547 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2548 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2549 pass_arg2(this, arg_2); 2550 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2551 pass_arg1(this, arg_1); 2552 pass_arg0(this, arg_0); 2553 MacroAssembler::call_VM_leaf_base(entry_point, 3); 2554 } 2555 2556 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) { 2557 LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg")); 2558 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2559 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2560 pass_arg3(this, arg_3); 2561 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2562 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2563 pass_arg2(this, arg_2); 2564 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2565 pass_arg1(this, arg_1); 2566 pass_arg0(this, arg_0); 2567 MacroAssembler::call_VM_leaf_base(entry_point, 4); 2568 } 2569 2570 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) { 2571 movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset())); 2572 movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD); 2573 verify_oop(oop_result, "broken oop in call_VM_base"); 2574 } 2575 2576 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) { 2577 movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset())); 2578 movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD); 2579 } 2580 2581 void MacroAssembler::check_and_handle_earlyret(Register java_thread) { 2582 } 2583 2584 void MacroAssembler::check_and_handle_popframe(Register java_thread) { 2585 } 2586 2587 void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) { 2588 if (reachable(src1)) { 2589 cmpl(as_Address(src1), imm); 2590 } else { 2591 lea(rscratch1, src1); 2592 cmpl(Address(rscratch1, 0), imm); 2593 } 2594 } 2595 2596 void MacroAssembler::cmp32(Register src1, AddressLiteral src2) { 2597 assert(!src2.is_lval(), "use cmpptr"); 2598 if (reachable(src2)) { 2599 cmpl(src1, as_Address(src2)); 2600 } else { 2601 lea(rscratch1, src2); 2602 cmpl(src1, Address(rscratch1, 0)); 2603 } 2604 } 2605 2606 void MacroAssembler::cmp32(Register src1, int32_t imm) { 2607 Assembler::cmpl(src1, imm); 2608 } 2609 2610 void MacroAssembler::cmp32(Register src1, Address src2) { 2611 Assembler::cmpl(src1, src2); 2612 } 2613 2614 void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) { 2615 ucomisd(opr1, opr2); 2616 2617 Label L; 2618 if (unordered_is_less) { 2619 movl(dst, -1); 2620 jcc(Assembler::parity, L); 2621 jcc(Assembler::below , L); 2622 movl(dst, 0); 2623 jcc(Assembler::equal , L); 2624 increment(dst); 2625 } else { // unordered is greater 2626 movl(dst, 1); 2627 jcc(Assembler::parity, L); 2628 jcc(Assembler::above , L); 2629 movl(dst, 0); 2630 jcc(Assembler::equal , L); 2631 decrementl(dst); 2632 } 2633 bind(L); 2634 } 2635 2636 void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) { 2637 ucomiss(opr1, opr2); 2638 2639 Label L; 2640 if (unordered_is_less) { 2641 movl(dst, -1); 2642 jcc(Assembler::parity, L); 2643 jcc(Assembler::below , L); 2644 movl(dst, 0); 2645 jcc(Assembler::equal , L); 2646 increment(dst); 2647 } else { // unordered is greater 2648 movl(dst, 1); 2649 jcc(Assembler::parity, L); 2650 jcc(Assembler::above , L); 2651 movl(dst, 0); 2652 jcc(Assembler::equal , L); 2653 decrementl(dst); 2654 } 2655 bind(L); 2656 } 2657 2658 2659 void MacroAssembler::cmp8(AddressLiteral src1, int imm) { 2660 if (reachable(src1)) { 2661 cmpb(as_Address(src1), imm); 2662 } else { 2663 lea(rscratch1, src1); 2664 cmpb(Address(rscratch1, 0), imm); 2665 } 2666 } 2667 2668 void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) { 2669 #ifdef _LP64 2670 if (src2.is_lval()) { 2671 movptr(rscratch1, src2); 2672 Assembler::cmpq(src1, rscratch1); 2673 } else if (reachable(src2)) { 2674 cmpq(src1, as_Address(src2)); 2675 } else { 2676 lea(rscratch1, src2); 2677 Assembler::cmpq(src1, Address(rscratch1, 0)); 2678 } 2679 #else 2680 if (src2.is_lval()) { 2681 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec()); 2682 } else { 2683 cmpl(src1, as_Address(src2)); 2684 } 2685 #endif // _LP64 2686 } 2687 2688 void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) { 2689 assert(src2.is_lval(), "not a mem-mem compare"); 2690 #ifdef _LP64 2691 // moves src2's literal address 2692 movptr(rscratch1, src2); 2693 Assembler::cmpq(src1, rscratch1); 2694 #else 2695 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec()); 2696 #endif // _LP64 2697 } 2698 2699 void MacroAssembler::cmpoop(Register src1, Register src2) { 2700 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 2701 bs->obj_equals(this, src1, src2); 2702 } 2703 2704 void MacroAssembler::cmpoop(Register src1, Address src2) { 2705 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 2706 bs->obj_equals(this, src1, src2); 2707 } 2708 2709 #ifdef _LP64 2710 void MacroAssembler::cmpoop(Register src1, jobject src2) { 2711 movoop(rscratch1, src2); 2712 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 2713 bs->obj_equals(this, src1, rscratch1); 2714 } 2715 #endif 2716 2717 void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) { 2718 if (reachable(adr)) { 2719 lock(); 2720 cmpxchgptr(reg, as_Address(adr)); 2721 } else { 2722 lea(rscratch1, adr); 2723 lock(); 2724 cmpxchgptr(reg, Address(rscratch1, 0)); 2725 } 2726 } 2727 2728 void MacroAssembler::cmpxchgptr(Register reg, Address adr) { 2729 LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr)); 2730 } 2731 2732 void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) { 2733 if (reachable(src)) { 2734 Assembler::comisd(dst, as_Address(src)); 2735 } else { 2736 lea(rscratch1, src); 2737 Assembler::comisd(dst, Address(rscratch1, 0)); 2738 } 2739 } 2740 2741 void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) { 2742 if (reachable(src)) { 2743 Assembler::comiss(dst, as_Address(src)); 2744 } else { 2745 lea(rscratch1, src); 2746 Assembler::comiss(dst, Address(rscratch1, 0)); 2747 } 2748 } 2749 2750 2751 void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) { 2752 Condition negated_cond = negate_condition(cond); 2753 Label L; 2754 jcc(negated_cond, L); 2755 pushf(); // Preserve flags 2756 atomic_incl(counter_addr); 2757 popf(); 2758 bind(L); 2759 } 2760 2761 int MacroAssembler::corrected_idivl(Register reg) { 2762 // Full implementation of Java idiv and irem; checks for 2763 // special case as described in JVM spec., p.243 & p.271. 2764 // The function returns the (pc) offset of the idivl 2765 // instruction - may be needed for implicit exceptions. 2766 // 2767 // normal case special case 2768 // 2769 // input : rax,: dividend min_int 2770 // reg: divisor (may not be rax,/rdx) -1 2771 // 2772 // output: rax,: quotient (= rax, idiv reg) min_int 2773 // rdx: remainder (= rax, irem reg) 0 2774 assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register"); 2775 const int min_int = 0x80000000; 2776 Label normal_case, special_case; 2777 2778 // check for special case 2779 cmpl(rax, min_int); 2780 jcc(Assembler::notEqual, normal_case); 2781 xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0) 2782 cmpl(reg, -1); 2783 jcc(Assembler::equal, special_case); 2784 2785 // handle normal case 2786 bind(normal_case); 2787 cdql(); 2788 int idivl_offset = offset(); 2789 idivl(reg); 2790 2791 // normal and special case exit 2792 bind(special_case); 2793 2794 return idivl_offset; 2795 } 2796 2797 2798 2799 void MacroAssembler::decrementl(Register reg, int value) { 2800 if (value == min_jint) {subl(reg, value) ; return; } 2801 if (value < 0) { incrementl(reg, -value); return; } 2802 if (value == 0) { ; return; } 2803 if (value == 1 && UseIncDec) { decl(reg) ; return; } 2804 /* else */ { subl(reg, value) ; return; } 2805 } 2806 2807 void MacroAssembler::decrementl(Address dst, int value) { 2808 if (value == min_jint) {subl(dst, value) ; return; } 2809 if (value < 0) { incrementl(dst, -value); return; } 2810 if (value == 0) { ; return; } 2811 if (value == 1 && UseIncDec) { decl(dst) ; return; } 2812 /* else */ { subl(dst, value) ; return; } 2813 } 2814 2815 void MacroAssembler::division_with_shift (Register reg, int shift_value) { 2816 assert (shift_value > 0, "illegal shift value"); 2817 Label _is_positive; 2818 testl (reg, reg); 2819 jcc (Assembler::positive, _is_positive); 2820 int offset = (1 << shift_value) - 1 ; 2821 2822 if (offset == 1) { 2823 incrementl(reg); 2824 } else { 2825 addl(reg, offset); 2826 } 2827 2828 bind (_is_positive); 2829 sarl(reg, shift_value); 2830 } 2831 2832 void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) { 2833 if (reachable(src)) { 2834 Assembler::divsd(dst, as_Address(src)); 2835 } else { 2836 lea(rscratch1, src); 2837 Assembler::divsd(dst, Address(rscratch1, 0)); 2838 } 2839 } 2840 2841 void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) { 2842 if (reachable(src)) { 2843 Assembler::divss(dst, as_Address(src)); 2844 } else { 2845 lea(rscratch1, src); 2846 Assembler::divss(dst, Address(rscratch1, 0)); 2847 } 2848 } 2849 2850 // !defined(COMPILER2) is because of stupid core builds 2851 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2) || INCLUDE_JVMCI 2852 void MacroAssembler::empty_FPU_stack() { 2853 if (VM_Version::supports_mmx()) { 2854 emms(); 2855 } else { 2856 for (int i = 8; i-- > 0; ) ffree(i); 2857 } 2858 } 2859 #endif // !LP64 || C1 || !C2 || INCLUDE_JVMCI 2860 2861 2862 void MacroAssembler::enter() { 2863 push(rbp); 2864 mov(rbp, rsp); 2865 } 2866 2867 // A 5 byte nop that is safe for patching (see patch_verified_entry) 2868 void MacroAssembler::fat_nop() { 2869 if (UseAddressNop) { 2870 addr_nop_5(); 2871 } else { 2872 emit_int8(0x26); // es: 2873 emit_int8(0x2e); // cs: 2874 emit_int8(0x64); // fs: 2875 emit_int8(0x65); // gs: 2876 emit_int8((unsigned char)0x90); 2877 } 2878 } 2879 2880 void MacroAssembler::fcmp(Register tmp) { 2881 fcmp(tmp, 1, true, true); 2882 } 2883 2884 void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) { 2885 assert(!pop_right || pop_left, "usage error"); 2886 if (VM_Version::supports_cmov()) { 2887 assert(tmp == noreg, "unneeded temp"); 2888 if (pop_left) { 2889 fucomip(index); 2890 } else { 2891 fucomi(index); 2892 } 2893 if (pop_right) { 2894 fpop(); 2895 } 2896 } else { 2897 assert(tmp != noreg, "need temp"); 2898 if (pop_left) { 2899 if (pop_right) { 2900 fcompp(); 2901 } else { 2902 fcomp(index); 2903 } 2904 } else { 2905 fcom(index); 2906 } 2907 // convert FPU condition into eflags condition via rax, 2908 save_rax(tmp); 2909 fwait(); fnstsw_ax(); 2910 sahf(); 2911 restore_rax(tmp); 2912 } 2913 // condition codes set as follows: 2914 // 2915 // CF (corresponds to C0) if x < y 2916 // PF (corresponds to C2) if unordered 2917 // ZF (corresponds to C3) if x = y 2918 } 2919 2920 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) { 2921 fcmp2int(dst, unordered_is_less, 1, true, true); 2922 } 2923 2924 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) { 2925 fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right); 2926 Label L; 2927 if (unordered_is_less) { 2928 movl(dst, -1); 2929 jcc(Assembler::parity, L); 2930 jcc(Assembler::below , L); 2931 movl(dst, 0); 2932 jcc(Assembler::equal , L); 2933 increment(dst); 2934 } else { // unordered is greater 2935 movl(dst, 1); 2936 jcc(Assembler::parity, L); 2937 jcc(Assembler::above , L); 2938 movl(dst, 0); 2939 jcc(Assembler::equal , L); 2940 decrementl(dst); 2941 } 2942 bind(L); 2943 } 2944 2945 void MacroAssembler::fld_d(AddressLiteral src) { 2946 fld_d(as_Address(src)); 2947 } 2948 2949 void MacroAssembler::fld_s(AddressLiteral src) { 2950 fld_s(as_Address(src)); 2951 } 2952 2953 void MacroAssembler::fld_x(AddressLiteral src) { 2954 Assembler::fld_x(as_Address(src)); 2955 } 2956 2957 void MacroAssembler::fldcw(AddressLiteral src) { 2958 Assembler::fldcw(as_Address(src)); 2959 } 2960 2961 void MacroAssembler::mulpd(XMMRegister dst, AddressLiteral src) { 2962 if (reachable(src)) { 2963 Assembler::mulpd(dst, as_Address(src)); 2964 } else { 2965 lea(rscratch1, src); 2966 Assembler::mulpd(dst, Address(rscratch1, 0)); 2967 } 2968 } 2969 2970 void MacroAssembler::increase_precision() { 2971 subptr(rsp, BytesPerWord); 2972 fnstcw(Address(rsp, 0)); 2973 movl(rax, Address(rsp, 0)); 2974 orl(rax, 0x300); 2975 push(rax); 2976 fldcw(Address(rsp, 0)); 2977 pop(rax); 2978 } 2979 2980 void MacroAssembler::restore_precision() { 2981 fldcw(Address(rsp, 0)); 2982 addptr(rsp, BytesPerWord); 2983 } 2984 2985 void MacroAssembler::fpop() { 2986 ffree(); 2987 fincstp(); 2988 } 2989 2990 void MacroAssembler::load_float(Address src) { 2991 if (UseSSE >= 1) { 2992 movflt(xmm0, src); 2993 } else { 2994 LP64_ONLY(ShouldNotReachHere()); 2995 NOT_LP64(fld_s(src)); 2996 } 2997 } 2998 2999 void MacroAssembler::store_float(Address dst) { 3000 if (UseSSE >= 1) { 3001 movflt(dst, xmm0); 3002 } else { 3003 LP64_ONLY(ShouldNotReachHere()); 3004 NOT_LP64(fstp_s(dst)); 3005 } 3006 } 3007 3008 void MacroAssembler::load_double(Address src) { 3009 if (UseSSE >= 2) { 3010 movdbl(xmm0, src); 3011 } else { 3012 LP64_ONLY(ShouldNotReachHere()); 3013 NOT_LP64(fld_d(src)); 3014 } 3015 } 3016 3017 void MacroAssembler::store_double(Address dst) { 3018 if (UseSSE >= 2) { 3019 movdbl(dst, xmm0); 3020 } else { 3021 LP64_ONLY(ShouldNotReachHere()); 3022 NOT_LP64(fstp_d(dst)); 3023 } 3024 } 3025 3026 void MacroAssembler::fremr(Register tmp) { 3027 save_rax(tmp); 3028 { Label L; 3029 bind(L); 3030 fprem(); 3031 fwait(); fnstsw_ax(); 3032 #ifdef _LP64 3033 testl(rax, 0x400); 3034 jcc(Assembler::notEqual, L); 3035 #else 3036 sahf(); 3037 jcc(Assembler::parity, L); 3038 #endif // _LP64 3039 } 3040 restore_rax(tmp); 3041 // Result is in ST0. 3042 // Note: fxch & fpop to get rid of ST1 3043 // (otherwise FPU stack could overflow eventually) 3044 fxch(1); 3045 fpop(); 3046 } 3047 3048 // dst = c = a * b + c 3049 void MacroAssembler::fmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) { 3050 Assembler::vfmadd231sd(c, a, b); 3051 if (dst != c) { 3052 movdbl(dst, c); 3053 } 3054 } 3055 3056 // dst = c = a * b + c 3057 void MacroAssembler::fmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) { 3058 Assembler::vfmadd231ss(c, a, b); 3059 if (dst != c) { 3060 movflt(dst, c); 3061 } 3062 } 3063 3064 // dst = c = a * b + c 3065 void MacroAssembler::vfmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c, int vector_len) { 3066 Assembler::vfmadd231pd(c, a, b, vector_len); 3067 if (dst != c) { 3068 vmovdqu(dst, c); 3069 } 3070 } 3071 3072 // dst = c = a * b + c 3073 void MacroAssembler::vfmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c, int vector_len) { 3074 Assembler::vfmadd231ps(c, a, b, vector_len); 3075 if (dst != c) { 3076 vmovdqu(dst, c); 3077 } 3078 } 3079 3080 // dst = c = a * b + c 3081 void MacroAssembler::vfmad(XMMRegister dst, XMMRegister a, Address b, XMMRegister c, int vector_len) { 3082 Assembler::vfmadd231pd(c, a, b, vector_len); 3083 if (dst != c) { 3084 vmovdqu(dst, c); 3085 } 3086 } 3087 3088 // dst = c = a * b + c 3089 void MacroAssembler::vfmaf(XMMRegister dst, XMMRegister a, Address b, XMMRegister c, int vector_len) { 3090 Assembler::vfmadd231ps(c, a, b, vector_len); 3091 if (dst != c) { 3092 vmovdqu(dst, c); 3093 } 3094 } 3095 3096 void MacroAssembler::incrementl(AddressLiteral dst) { 3097 if (reachable(dst)) { 3098 incrementl(as_Address(dst)); 3099 } else { 3100 lea(rscratch1, dst); 3101 incrementl(Address(rscratch1, 0)); 3102 } 3103 } 3104 3105 void MacroAssembler::incrementl(ArrayAddress dst) { 3106 incrementl(as_Address(dst)); 3107 } 3108 3109 void MacroAssembler::incrementl(Register reg, int value) { 3110 if (value == min_jint) {addl(reg, value) ; return; } 3111 if (value < 0) { decrementl(reg, -value); return; } 3112 if (value == 0) { ; return; } 3113 if (value == 1 && UseIncDec) { incl(reg) ; return; } 3114 /* else */ { addl(reg, value) ; return; } 3115 } 3116 3117 void MacroAssembler::incrementl(Address dst, int value) { 3118 if (value == min_jint) {addl(dst, value) ; return; } 3119 if (value < 0) { decrementl(dst, -value); return; } 3120 if (value == 0) { ; return; } 3121 if (value == 1 && UseIncDec) { incl(dst) ; return; } 3122 /* else */ { addl(dst, value) ; return; } 3123 } 3124 3125 void MacroAssembler::jump(AddressLiteral dst) { 3126 if (reachable(dst)) { 3127 jmp_literal(dst.target(), dst.rspec()); 3128 } else { 3129 lea(rscratch1, dst); 3130 jmp(rscratch1); 3131 } 3132 } 3133 3134 void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) { 3135 if (reachable(dst)) { 3136 InstructionMark im(this); 3137 relocate(dst.reloc()); 3138 const int short_size = 2; 3139 const int long_size = 6; 3140 int offs = (intptr_t)dst.target() - ((intptr_t)pc()); 3141 if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) { 3142 // 0111 tttn #8-bit disp 3143 emit_int8(0x70 | cc); 3144 emit_int8((offs - short_size) & 0xFF); 3145 } else { 3146 // 0000 1111 1000 tttn #32-bit disp 3147 emit_int8(0x0F); 3148 emit_int8((unsigned char)(0x80 | cc)); 3149 emit_int32(offs - long_size); 3150 } 3151 } else { 3152 #ifdef ASSERT 3153 warning("reversing conditional branch"); 3154 #endif /* ASSERT */ 3155 Label skip; 3156 jccb(reverse[cc], skip); 3157 lea(rscratch1, dst); 3158 Assembler::jmp(rscratch1); 3159 bind(skip); 3160 } 3161 } 3162 3163 void MacroAssembler::ldmxcsr(AddressLiteral src) { 3164 if (reachable(src)) { 3165 Assembler::ldmxcsr(as_Address(src)); 3166 } else { 3167 lea(rscratch1, src); 3168 Assembler::ldmxcsr(Address(rscratch1, 0)); 3169 } 3170 } 3171 3172 int MacroAssembler::load_signed_byte(Register dst, Address src) { 3173 int off; 3174 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3175 off = offset(); 3176 movsbl(dst, src); // movsxb 3177 } else { 3178 off = load_unsigned_byte(dst, src); 3179 shll(dst, 24); 3180 sarl(dst, 24); 3181 } 3182 return off; 3183 } 3184 3185 // Note: load_signed_short used to be called load_signed_word. 3186 // Although the 'w' in x86 opcodes refers to the term "word" in the assembler 3187 // manual, which means 16 bits, that usage is found nowhere in HotSpot code. 3188 // The term "word" in HotSpot means a 32- or 64-bit machine word. 3189 int MacroAssembler::load_signed_short(Register dst, Address src) { 3190 int off; 3191 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3192 // This is dubious to me since it seems safe to do a signed 16 => 64 bit 3193 // version but this is what 64bit has always done. This seems to imply 3194 // that users are only using 32bits worth. 3195 off = offset(); 3196 movswl(dst, src); // movsxw 3197 } else { 3198 off = load_unsigned_short(dst, src); 3199 shll(dst, 16); 3200 sarl(dst, 16); 3201 } 3202 return off; 3203 } 3204 3205 int MacroAssembler::load_unsigned_byte(Register dst, Address src) { 3206 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16, 3207 // and "3.9 Partial Register Penalties", p. 22). 3208 int off; 3209 if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) { 3210 off = offset(); 3211 movzbl(dst, src); // movzxb 3212 } else { 3213 xorl(dst, dst); 3214 off = offset(); 3215 movb(dst, src); 3216 } 3217 return off; 3218 } 3219 3220 // Note: load_unsigned_short used to be called load_unsigned_word. 3221 int MacroAssembler::load_unsigned_short(Register dst, Address src) { 3222 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16, 3223 // and "3.9 Partial Register Penalties", p. 22). 3224 int off; 3225 if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) { 3226 off = offset(); 3227 movzwl(dst, src); // movzxw 3228 } else { 3229 xorl(dst, dst); 3230 off = offset(); 3231 movw(dst, src); 3232 } 3233 return off; 3234 } 3235 3236 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) { 3237 switch (size_in_bytes) { 3238 #ifndef _LP64 3239 case 8: 3240 assert(dst2 != noreg, "second dest register required"); 3241 movl(dst, src); 3242 movl(dst2, src.plus_disp(BytesPerInt)); 3243 break; 3244 #else 3245 case 8: movq(dst, src); break; 3246 #endif 3247 case 4: movl(dst, src); break; 3248 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break; 3249 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break; 3250 default: ShouldNotReachHere(); 3251 } 3252 } 3253 3254 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) { 3255 switch (size_in_bytes) { 3256 #ifndef _LP64 3257 case 8: 3258 assert(src2 != noreg, "second source register required"); 3259 movl(dst, src); 3260 movl(dst.plus_disp(BytesPerInt), src2); 3261 break; 3262 #else 3263 case 8: movq(dst, src); break; 3264 #endif 3265 case 4: movl(dst, src); break; 3266 case 2: movw(dst, src); break; 3267 case 1: movb(dst, src); break; 3268 default: ShouldNotReachHere(); 3269 } 3270 } 3271 3272 void MacroAssembler::mov32(AddressLiteral dst, Register src) { 3273 if (reachable(dst)) { 3274 movl(as_Address(dst), src); 3275 } else { 3276 lea(rscratch1, dst); 3277 movl(Address(rscratch1, 0), src); 3278 } 3279 } 3280 3281 void MacroAssembler::mov32(Register dst, AddressLiteral src) { 3282 if (reachable(src)) { 3283 movl(dst, as_Address(src)); 3284 } else { 3285 lea(rscratch1, src); 3286 movl(dst, Address(rscratch1, 0)); 3287 } 3288 } 3289 3290 // C++ bool manipulation 3291 3292 void MacroAssembler::movbool(Register dst, Address src) { 3293 if(sizeof(bool) == 1) 3294 movb(dst, src); 3295 else if(sizeof(bool) == 2) 3296 movw(dst, src); 3297 else if(sizeof(bool) == 4) 3298 movl(dst, src); 3299 else 3300 // unsupported 3301 ShouldNotReachHere(); 3302 } 3303 3304 void MacroAssembler::movbool(Address dst, bool boolconst) { 3305 if(sizeof(bool) == 1) 3306 movb(dst, (int) boolconst); 3307 else if(sizeof(bool) == 2) 3308 movw(dst, (int) boolconst); 3309 else if(sizeof(bool) == 4) 3310 movl(dst, (int) boolconst); 3311 else 3312 // unsupported 3313 ShouldNotReachHere(); 3314 } 3315 3316 void MacroAssembler::movbool(Address dst, Register src) { 3317 if(sizeof(bool) == 1) 3318 movb(dst, src); 3319 else if(sizeof(bool) == 2) 3320 movw(dst, src); 3321 else if(sizeof(bool) == 4) 3322 movl(dst, src); 3323 else 3324 // unsupported 3325 ShouldNotReachHere(); 3326 } 3327 3328 void MacroAssembler::movbyte(ArrayAddress dst, int src) { 3329 movb(as_Address(dst), src); 3330 } 3331 3332 void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) { 3333 if (reachable(src)) { 3334 movdl(dst, as_Address(src)); 3335 } else { 3336 lea(rscratch1, src); 3337 movdl(dst, Address(rscratch1, 0)); 3338 } 3339 } 3340 3341 void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) { 3342 if (reachable(src)) { 3343 movq(dst, as_Address(src)); 3344 } else { 3345 lea(rscratch1, src); 3346 movq(dst, Address(rscratch1, 0)); 3347 } 3348 } 3349 3350 #ifdef COMPILER2 3351 void MacroAssembler::setvectmask(Register dst, Register src) { 3352 guarantee(PostLoopMultiversioning, "must be"); 3353 Assembler::movl(dst, 1); 3354 Assembler::shlxl(dst, dst, src); 3355 Assembler::decl(dst); 3356 Assembler::kmovdl(k1, dst); 3357 Assembler::movl(dst, src); 3358 } 3359 3360 void MacroAssembler::restorevectmask() { 3361 guarantee(PostLoopMultiversioning, "must be"); 3362 Assembler::knotwl(k1, k0); 3363 } 3364 #endif // COMPILER2 3365 3366 void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) { 3367 if (reachable(src)) { 3368 if (UseXmmLoadAndClearUpper) { 3369 movsd (dst, as_Address(src)); 3370 } else { 3371 movlpd(dst, as_Address(src)); 3372 } 3373 } else { 3374 lea(rscratch1, src); 3375 if (UseXmmLoadAndClearUpper) { 3376 movsd (dst, Address(rscratch1, 0)); 3377 } else { 3378 movlpd(dst, Address(rscratch1, 0)); 3379 } 3380 } 3381 } 3382 3383 void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) { 3384 if (reachable(src)) { 3385 movss(dst, as_Address(src)); 3386 } else { 3387 lea(rscratch1, src); 3388 movss(dst, Address(rscratch1, 0)); 3389 } 3390 } 3391 3392 void MacroAssembler::movptr(Register dst, Register src) { 3393 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3394 } 3395 3396 void MacroAssembler::movptr(Register dst, Address src) { 3397 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3398 } 3399 3400 // src should NEVER be a real pointer. Use AddressLiteral for true pointers 3401 void MacroAssembler::movptr(Register dst, intptr_t src) { 3402 LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src)); 3403 } 3404 3405 void MacroAssembler::movptr(Address dst, Register src) { 3406 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3407 } 3408 3409 void MacroAssembler::movdqu(Address dst, XMMRegister src) { 3410 assert(((src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15"); 3411 Assembler::movdqu(dst, src); 3412 } 3413 3414 void MacroAssembler::movdqu(XMMRegister dst, Address src) { 3415 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15"); 3416 Assembler::movdqu(dst, src); 3417 } 3418 3419 void MacroAssembler::movdqu(XMMRegister dst, XMMRegister src) { 3420 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15"); 3421 Assembler::movdqu(dst, src); 3422 } 3423 3424 void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src, Register scratchReg) { 3425 if (reachable(src)) { 3426 movdqu(dst, as_Address(src)); 3427 } else { 3428 lea(scratchReg, src); 3429 movdqu(dst, Address(scratchReg, 0)); 3430 } 3431 } 3432 3433 void MacroAssembler::vmovdqu(Address dst, XMMRegister src) { 3434 assert(((src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15"); 3435 Assembler::vmovdqu(dst, src); 3436 } 3437 3438 void MacroAssembler::vmovdqu(XMMRegister dst, Address src) { 3439 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15"); 3440 Assembler::vmovdqu(dst, src); 3441 } 3442 3443 void MacroAssembler::vmovdqu(XMMRegister dst, XMMRegister src) { 3444 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15"); 3445 Assembler::vmovdqu(dst, src); 3446 } 3447 3448 void MacroAssembler::vmovdqu(XMMRegister dst, AddressLiteral src, Register scratch_reg) { 3449 if (reachable(src)) { 3450 vmovdqu(dst, as_Address(src)); 3451 } 3452 else { 3453 lea(scratch_reg, src); 3454 vmovdqu(dst, Address(scratch_reg, 0)); 3455 } 3456 } 3457 3458 void MacroAssembler::evmovdquq(XMMRegister dst, AddressLiteral src, int vector_len, Register rscratch) { 3459 if (reachable(src)) { 3460 Assembler::evmovdquq(dst, as_Address(src), vector_len); 3461 } else { 3462 lea(rscratch, src); 3463 Assembler::evmovdquq(dst, Address(rscratch, 0), vector_len); 3464 } 3465 } 3466 3467 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) { 3468 if (reachable(src)) { 3469 Assembler::movdqa(dst, as_Address(src)); 3470 } else { 3471 lea(rscratch1, src); 3472 Assembler::movdqa(dst, Address(rscratch1, 0)); 3473 } 3474 } 3475 3476 void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) { 3477 if (reachable(src)) { 3478 Assembler::movsd(dst, as_Address(src)); 3479 } else { 3480 lea(rscratch1, src); 3481 Assembler::movsd(dst, Address(rscratch1, 0)); 3482 } 3483 } 3484 3485 void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) { 3486 if (reachable(src)) { 3487 Assembler::movss(dst, as_Address(src)); 3488 } else { 3489 lea(rscratch1, src); 3490 Assembler::movss(dst, Address(rscratch1, 0)); 3491 } 3492 } 3493 3494 void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) { 3495 if (reachable(src)) { 3496 Assembler::mulsd(dst, as_Address(src)); 3497 } else { 3498 lea(rscratch1, src); 3499 Assembler::mulsd(dst, Address(rscratch1, 0)); 3500 } 3501 } 3502 3503 void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) { 3504 if (reachable(src)) { 3505 Assembler::mulss(dst, as_Address(src)); 3506 } else { 3507 lea(rscratch1, src); 3508 Assembler::mulss(dst, Address(rscratch1, 0)); 3509 } 3510 } 3511 3512 void MacroAssembler::null_check(Register reg, int offset) { 3513 if (needs_explicit_null_check(offset)) { 3514 // provoke OS NULL exception if reg = NULL by 3515 // accessing M[reg] w/o changing any (non-CC) registers 3516 // NOTE: cmpl is plenty here to provoke a segv 3517 cmpptr(rax, Address(reg, 0)); 3518 // Note: should probably use testl(rax, Address(reg, 0)); 3519 // may be shorter code (however, this version of 3520 // testl needs to be implemented first) 3521 } else { 3522 // nothing to do, (later) access of M[reg + offset] 3523 // will provoke OS NULL exception if reg = NULL 3524 } 3525 } 3526 3527 void MacroAssembler::os_breakpoint() { 3528 // instead of directly emitting a breakpoint, call os:breakpoint for better debugability 3529 // (e.g., MSVC can't call ps() otherwise) 3530 call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint))); 3531 } 3532 3533 void MacroAssembler::unimplemented(const char* what) { 3534 const char* buf = NULL; 3535 { 3536 ResourceMark rm; 3537 stringStream ss; 3538 ss.print("unimplemented: %s", what); 3539 buf = code_string(ss.as_string()); 3540 } 3541 stop(buf); 3542 } 3543 3544 #ifdef _LP64 3545 #define XSTATE_BV 0x200 3546 #endif 3547 3548 void MacroAssembler::pop_CPU_state() { 3549 pop_FPU_state(); 3550 pop_IU_state(); 3551 } 3552 3553 void MacroAssembler::pop_FPU_state() { 3554 #ifndef _LP64 3555 frstor(Address(rsp, 0)); 3556 #else 3557 fxrstor(Address(rsp, 0)); 3558 #endif 3559 addptr(rsp, FPUStateSizeInWords * wordSize); 3560 } 3561 3562 void MacroAssembler::pop_IU_state() { 3563 popa(); 3564 LP64_ONLY(addq(rsp, 8)); 3565 popf(); 3566 } 3567 3568 // Save Integer and Float state 3569 // Warning: Stack must be 16 byte aligned (64bit) 3570 void MacroAssembler::push_CPU_state() { 3571 push_IU_state(); 3572 push_FPU_state(); 3573 } 3574 3575 void MacroAssembler::push_FPU_state() { 3576 subptr(rsp, FPUStateSizeInWords * wordSize); 3577 #ifndef _LP64 3578 fnsave(Address(rsp, 0)); 3579 fwait(); 3580 #else 3581 fxsave(Address(rsp, 0)); 3582 #endif // LP64 3583 } 3584 3585 void MacroAssembler::push_IU_state() { 3586 // Push flags first because pusha kills them 3587 pushf(); 3588 // Make sure rsp stays 16-byte aligned 3589 LP64_ONLY(subq(rsp, 8)); 3590 pusha(); 3591 } 3592 3593 void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp) { // determine java_thread register 3594 if (!java_thread->is_valid()) { 3595 java_thread = rdi; 3596 get_thread(java_thread); 3597 } 3598 // we must set sp to zero to clear frame 3599 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD); 3600 if (clear_fp) { 3601 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD); 3602 } 3603 3604 // Always clear the pc because it could have been set by make_walkable() 3605 movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD); 3606 3607 vzeroupper(); 3608 } 3609 3610 void MacroAssembler::restore_rax(Register tmp) { 3611 if (tmp == noreg) pop(rax); 3612 else if (tmp != rax) mov(rax, tmp); 3613 } 3614 3615 void MacroAssembler::round_to(Register reg, int modulus) { 3616 addptr(reg, modulus - 1); 3617 andptr(reg, -modulus); 3618 } 3619 3620 void MacroAssembler::save_rax(Register tmp) { 3621 if (tmp == noreg) push(rax); 3622 else if (tmp != rax) mov(tmp, rax); 3623 } 3624 3625 void MacroAssembler::safepoint_poll(Label& slow_path, Register thread_reg, Register temp_reg) { 3626 if (SafepointMechanism::uses_thread_local_poll()) { 3627 #ifdef _LP64 3628 assert(thread_reg == r15_thread, "should be"); 3629 #else 3630 if (thread_reg == noreg) { 3631 thread_reg = temp_reg; 3632 get_thread(thread_reg); 3633 } 3634 #endif 3635 testb(Address(thread_reg, Thread::polling_page_offset()), SafepointMechanism::poll_bit()); 3636 jcc(Assembler::notZero, slow_path); // handshake bit set implies poll 3637 } else { 3638 cmp32(ExternalAddress(SafepointSynchronize::address_of_state()), 3639 SafepointSynchronize::_not_synchronized); 3640 jcc(Assembler::notEqual, slow_path); 3641 } 3642 } 3643 3644 // Calls to C land 3645 // 3646 // When entering C land, the rbp, & rsp of the last Java frame have to be recorded 3647 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp 3648 // has to be reset to 0. This is required to allow proper stack traversal. 3649 void MacroAssembler::set_last_Java_frame(Register java_thread, 3650 Register last_java_sp, 3651 Register last_java_fp, 3652 address last_java_pc) { 3653 vzeroupper(); 3654 // determine java_thread register 3655 if (!java_thread->is_valid()) { 3656 java_thread = rdi; 3657 get_thread(java_thread); 3658 } 3659 // determine last_java_sp register 3660 if (!last_java_sp->is_valid()) { 3661 last_java_sp = rsp; 3662 } 3663 3664 // last_java_fp is optional 3665 3666 if (last_java_fp->is_valid()) { 3667 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp); 3668 } 3669 3670 // last_java_pc is optional 3671 3672 if (last_java_pc != NULL) { 3673 lea(Address(java_thread, 3674 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()), 3675 InternalAddress(last_java_pc)); 3676 3677 } 3678 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp); 3679 } 3680 3681 void MacroAssembler::shlptr(Register dst, int imm8) { 3682 LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8)); 3683 } 3684 3685 void MacroAssembler::shrptr(Register dst, int imm8) { 3686 LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8)); 3687 } 3688 3689 void MacroAssembler::sign_extend_byte(Register reg) { 3690 if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) { 3691 movsbl(reg, reg); // movsxb 3692 } else { 3693 shll(reg, 24); 3694 sarl(reg, 24); 3695 } 3696 } 3697 3698 void MacroAssembler::sign_extend_short(Register reg) { 3699 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3700 movswl(reg, reg); // movsxw 3701 } else { 3702 shll(reg, 16); 3703 sarl(reg, 16); 3704 } 3705 } 3706 3707 void MacroAssembler::testl(Register dst, AddressLiteral src) { 3708 assert(reachable(src), "Address should be reachable"); 3709 testl(dst, as_Address(src)); 3710 } 3711 3712 void MacroAssembler::pcmpeqb(XMMRegister dst, XMMRegister src) { 3713 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3714 Assembler::pcmpeqb(dst, src); 3715 } 3716 3717 void MacroAssembler::pcmpeqw(XMMRegister dst, XMMRegister src) { 3718 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3719 Assembler::pcmpeqw(dst, src); 3720 } 3721 3722 void MacroAssembler::pcmpestri(XMMRegister dst, Address src, int imm8) { 3723 assert((dst->encoding() < 16),"XMM register should be 0-15"); 3724 Assembler::pcmpestri(dst, src, imm8); 3725 } 3726 3727 void MacroAssembler::pcmpestri(XMMRegister dst, XMMRegister src, int imm8) { 3728 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15"); 3729 Assembler::pcmpestri(dst, src, imm8); 3730 } 3731 3732 void MacroAssembler::pmovzxbw(XMMRegister dst, XMMRegister src) { 3733 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3734 Assembler::pmovzxbw(dst, src); 3735 } 3736 3737 void MacroAssembler::pmovzxbw(XMMRegister dst, Address src) { 3738 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3739 Assembler::pmovzxbw(dst, src); 3740 } 3741 3742 void MacroAssembler::pmovmskb(Register dst, XMMRegister src) { 3743 assert((src->encoding() < 16),"XMM register should be 0-15"); 3744 Assembler::pmovmskb(dst, src); 3745 } 3746 3747 void MacroAssembler::ptest(XMMRegister dst, XMMRegister src) { 3748 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15"); 3749 Assembler::ptest(dst, src); 3750 } 3751 3752 void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) { 3753 if (reachable(src)) { 3754 Assembler::sqrtsd(dst, as_Address(src)); 3755 } else { 3756 lea(rscratch1, src); 3757 Assembler::sqrtsd(dst, Address(rscratch1, 0)); 3758 } 3759 } 3760 3761 void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) { 3762 if (reachable(src)) { 3763 Assembler::sqrtss(dst, as_Address(src)); 3764 } else { 3765 lea(rscratch1, src); 3766 Assembler::sqrtss(dst, Address(rscratch1, 0)); 3767 } 3768 } 3769 3770 void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) { 3771 if (reachable(src)) { 3772 Assembler::subsd(dst, as_Address(src)); 3773 } else { 3774 lea(rscratch1, src); 3775 Assembler::subsd(dst, Address(rscratch1, 0)); 3776 } 3777 } 3778 3779 void MacroAssembler::roundsd(XMMRegister dst, AddressLiteral src, int32_t rmode, Register scratch_reg) { 3780 if (reachable(src)) { 3781 Assembler::roundsd(dst, as_Address(src), rmode); 3782 } else { 3783 lea(scratch_reg, src); 3784 Assembler::roundsd(dst, Address(scratch_reg, 0), rmode); 3785 } 3786 } 3787 3788 void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) { 3789 if (reachable(src)) { 3790 Assembler::subss(dst, as_Address(src)); 3791 } else { 3792 lea(rscratch1, src); 3793 Assembler::subss(dst, Address(rscratch1, 0)); 3794 } 3795 } 3796 3797 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) { 3798 if (reachable(src)) { 3799 Assembler::ucomisd(dst, as_Address(src)); 3800 } else { 3801 lea(rscratch1, src); 3802 Assembler::ucomisd(dst, Address(rscratch1, 0)); 3803 } 3804 } 3805 3806 void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) { 3807 if (reachable(src)) { 3808 Assembler::ucomiss(dst, as_Address(src)); 3809 } else { 3810 lea(rscratch1, src); 3811 Assembler::ucomiss(dst, Address(rscratch1, 0)); 3812 } 3813 } 3814 3815 void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src, Register scratch_reg) { 3816 // Used in sign-bit flipping with aligned address. 3817 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 3818 if (reachable(src)) { 3819 Assembler::xorpd(dst, as_Address(src)); 3820 } else { 3821 lea(scratch_reg, src); 3822 Assembler::xorpd(dst, Address(scratch_reg, 0)); 3823 } 3824 } 3825 3826 void MacroAssembler::xorpd(XMMRegister dst, XMMRegister src) { 3827 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) { 3828 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit); 3829 } 3830 else { 3831 Assembler::xorpd(dst, src); 3832 } 3833 } 3834 3835 void MacroAssembler::xorps(XMMRegister dst, XMMRegister src) { 3836 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) { 3837 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit); 3838 } else { 3839 Assembler::xorps(dst, src); 3840 } 3841 } 3842 3843 void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src, Register scratch_reg) { 3844 // Used in sign-bit flipping with aligned address. 3845 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 3846 if (reachable(src)) { 3847 Assembler::xorps(dst, as_Address(src)); 3848 } else { 3849 lea(scratch_reg, src); 3850 Assembler::xorps(dst, Address(scratch_reg, 0)); 3851 } 3852 } 3853 3854 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) { 3855 // Used in sign-bit flipping with aligned address. 3856 bool aligned_adr = (((intptr_t)src.target() & 15) == 0); 3857 assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes"); 3858 if (reachable(src)) { 3859 Assembler::pshufb(dst, as_Address(src)); 3860 } else { 3861 lea(rscratch1, src); 3862 Assembler::pshufb(dst, Address(rscratch1, 0)); 3863 } 3864 } 3865 3866 // AVX 3-operands instructions 3867 3868 void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 3869 if (reachable(src)) { 3870 vaddsd(dst, nds, as_Address(src)); 3871 } else { 3872 lea(rscratch1, src); 3873 vaddsd(dst, nds, Address(rscratch1, 0)); 3874 } 3875 } 3876 3877 void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 3878 if (reachable(src)) { 3879 vaddss(dst, nds, as_Address(src)); 3880 } else { 3881 lea(rscratch1, src); 3882 vaddss(dst, nds, Address(rscratch1, 0)); 3883 } 3884 } 3885 3886 void MacroAssembler::vpaddd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register rscratch) { 3887 assert(UseAVX > 0, "requires some form of AVX"); 3888 if (reachable(src)) { 3889 Assembler::vpaddd(dst, nds, as_Address(src), vector_len); 3890 } else { 3891 lea(rscratch, src); 3892 Assembler::vpaddd(dst, nds, Address(rscratch, 0), vector_len); 3893 } 3894 } 3895 3896 void MacroAssembler::vabsss(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) { 3897 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15"); 3898 vandps(dst, nds, negate_field, vector_len); 3899 } 3900 3901 void MacroAssembler::vabssd(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) { 3902 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15"); 3903 vandpd(dst, nds, negate_field, vector_len); 3904 } 3905 3906 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 3907 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3908 Assembler::vpaddb(dst, nds, src, vector_len); 3909 } 3910 3911 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 3912 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3913 Assembler::vpaddb(dst, nds, src, vector_len); 3914 } 3915 3916 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 3917 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3918 Assembler::vpaddw(dst, nds, src, vector_len); 3919 } 3920 3921 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 3922 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3923 Assembler::vpaddw(dst, nds, src, vector_len); 3924 } 3925 3926 void MacroAssembler::vpand(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) { 3927 if (reachable(src)) { 3928 Assembler::vpand(dst, nds, as_Address(src), vector_len); 3929 } else { 3930 lea(scratch_reg, src); 3931 Assembler::vpand(dst, nds, Address(scratch_reg, 0), vector_len); 3932 } 3933 } 3934 3935 void MacroAssembler::vpbroadcastw(XMMRegister dst, XMMRegister src, int vector_len) { 3936 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3937 Assembler::vpbroadcastw(dst, src, vector_len); 3938 } 3939 3940 void MacroAssembler::vpcmpeqb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 3941 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3942 Assembler::vpcmpeqb(dst, nds, src, vector_len); 3943 } 3944 3945 void MacroAssembler::vpcmpeqw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 3946 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3947 Assembler::vpcmpeqw(dst, nds, src, vector_len); 3948 } 3949 3950 void MacroAssembler::vpmovzxbw(XMMRegister dst, Address src, int vector_len) { 3951 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3952 Assembler::vpmovzxbw(dst, src, vector_len); 3953 } 3954 3955 void MacroAssembler::vpmovmskb(Register dst, XMMRegister src) { 3956 assert((src->encoding() < 16),"XMM register should be 0-15"); 3957 Assembler::vpmovmskb(dst, src); 3958 } 3959 3960 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 3961 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3962 Assembler::vpmullw(dst, nds, src, vector_len); 3963 } 3964 3965 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 3966 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3967 Assembler::vpmullw(dst, nds, src, vector_len); 3968 } 3969 3970 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 3971 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3972 Assembler::vpsubb(dst, nds, src, vector_len); 3973 } 3974 3975 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 3976 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3977 Assembler::vpsubb(dst, nds, src, vector_len); 3978 } 3979 3980 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 3981 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3982 Assembler::vpsubw(dst, nds, src, vector_len); 3983 } 3984 3985 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 3986 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3987 Assembler::vpsubw(dst, nds, src, vector_len); 3988 } 3989 3990 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 3991 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3992 Assembler::vpsraw(dst, nds, shift, vector_len); 3993 } 3994 3995 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 3996 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3997 Assembler::vpsraw(dst, nds, shift, vector_len); 3998 } 3999 4000 void MacroAssembler::evpsraq(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4001 assert(UseAVX > 2,""); 4002 if (!VM_Version::supports_avx512vl() && vector_len < 2) { 4003 vector_len = 2; 4004 } 4005 Assembler::evpsraq(dst, nds, shift, vector_len); 4006 } 4007 4008 void MacroAssembler::evpsraq(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 4009 assert(UseAVX > 2,""); 4010 if (!VM_Version::supports_avx512vl() && vector_len < 2) { 4011 vector_len = 2; 4012 } 4013 Assembler::evpsraq(dst, nds, shift, vector_len); 4014 } 4015 4016 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4017 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 4018 Assembler::vpsrlw(dst, nds, shift, vector_len); 4019 } 4020 4021 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 4022 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 4023 Assembler::vpsrlw(dst, nds, shift, vector_len); 4024 } 4025 4026 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 4027 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 4028 Assembler::vpsllw(dst, nds, shift, vector_len); 4029 } 4030 4031 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 4032 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 4033 Assembler::vpsllw(dst, nds, shift, vector_len); 4034 } 4035 4036 void MacroAssembler::vptest(XMMRegister dst, XMMRegister src) { 4037 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15"); 4038 Assembler::vptest(dst, src); 4039 } 4040 4041 void MacroAssembler::punpcklbw(XMMRegister dst, XMMRegister src) { 4042 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 4043 Assembler::punpcklbw(dst, src); 4044 } 4045 4046 void MacroAssembler::pshufd(XMMRegister dst, Address src, int mode) { 4047 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15"); 4048 Assembler::pshufd(dst, src, mode); 4049 } 4050 4051 void MacroAssembler::pshuflw(XMMRegister dst, XMMRegister src, int mode) { 4052 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 4053 Assembler::pshuflw(dst, src, mode); 4054 } 4055 4056 void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) { 4057 if (reachable(src)) { 4058 vandpd(dst, nds, as_Address(src), vector_len); 4059 } else { 4060 lea(scratch_reg, src); 4061 vandpd(dst, nds, Address(scratch_reg, 0), vector_len); 4062 } 4063 } 4064 4065 void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) { 4066 if (reachable(src)) { 4067 vandps(dst, nds, as_Address(src), vector_len); 4068 } else { 4069 lea(scratch_reg, src); 4070 vandps(dst, nds, Address(scratch_reg, 0), vector_len); 4071 } 4072 } 4073 4074 void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4075 if (reachable(src)) { 4076 vdivsd(dst, nds, as_Address(src)); 4077 } else { 4078 lea(rscratch1, src); 4079 vdivsd(dst, nds, Address(rscratch1, 0)); 4080 } 4081 } 4082 4083 void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4084 if (reachable(src)) { 4085 vdivss(dst, nds, as_Address(src)); 4086 } else { 4087 lea(rscratch1, src); 4088 vdivss(dst, nds, Address(rscratch1, 0)); 4089 } 4090 } 4091 4092 void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4093 if (reachable(src)) { 4094 vmulsd(dst, nds, as_Address(src)); 4095 } else { 4096 lea(rscratch1, src); 4097 vmulsd(dst, nds, Address(rscratch1, 0)); 4098 } 4099 } 4100 4101 void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4102 if (reachable(src)) { 4103 vmulss(dst, nds, as_Address(src)); 4104 } else { 4105 lea(rscratch1, src); 4106 vmulss(dst, nds, Address(rscratch1, 0)); 4107 } 4108 } 4109 4110 void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4111 if (reachable(src)) { 4112 vsubsd(dst, nds, as_Address(src)); 4113 } else { 4114 lea(rscratch1, src); 4115 vsubsd(dst, nds, Address(rscratch1, 0)); 4116 } 4117 } 4118 4119 void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4120 if (reachable(src)) { 4121 vsubss(dst, nds, as_Address(src)); 4122 } else { 4123 lea(rscratch1, src); 4124 vsubss(dst, nds, Address(rscratch1, 0)); 4125 } 4126 } 4127 4128 void MacroAssembler::vnegatess(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4129 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15"); 4130 vxorps(dst, nds, src, Assembler::AVX_128bit); 4131 } 4132 4133 void MacroAssembler::vnegatesd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4134 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15"); 4135 vxorpd(dst, nds, src, Assembler::AVX_128bit); 4136 } 4137 4138 void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) { 4139 if (reachable(src)) { 4140 vxorpd(dst, nds, as_Address(src), vector_len); 4141 } else { 4142 lea(scratch_reg, src); 4143 vxorpd(dst, nds, Address(scratch_reg, 0), vector_len); 4144 } 4145 } 4146 4147 void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) { 4148 if (reachable(src)) { 4149 vxorps(dst, nds, as_Address(src), vector_len); 4150 } else { 4151 lea(scratch_reg, src); 4152 vxorps(dst, nds, Address(scratch_reg, 0), vector_len); 4153 } 4154 } 4155 4156 void MacroAssembler::vpxor(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) { 4157 if (UseAVX > 1 || (vector_len < 1)) { 4158 if (reachable(src)) { 4159 Assembler::vpxor(dst, nds, as_Address(src), vector_len); 4160 } else { 4161 lea(scratch_reg, src); 4162 Assembler::vpxor(dst, nds, Address(scratch_reg, 0), vector_len); 4163 } 4164 } 4165 else { 4166 MacroAssembler::vxorpd(dst, nds, src, vector_len, scratch_reg); 4167 } 4168 } 4169 4170 //------------------------------------------------------------------------------------------- 4171 #ifdef COMPILER2 4172 // Generic instructions support for use in .ad files C2 code generation 4173 4174 void MacroAssembler::vabsnegd(int opcode, XMMRegister dst, XMMRegister src, Register scr) { 4175 if (dst != src) { 4176 movdqu(dst, src); 4177 } 4178 if (opcode == Op_AbsVD) { 4179 andpd(dst, ExternalAddress(StubRoutines::x86::vector_double_sign_mask()), scr); 4180 } else { 4181 assert((opcode == Op_NegVD),"opcode should be Op_NegD"); 4182 xorpd(dst, ExternalAddress(StubRoutines::x86::vector_double_sign_flip()), scr); 4183 } 4184 } 4185 4186 void MacroAssembler::vabsnegd(int opcode, XMMRegister dst, XMMRegister src, int vector_len, Register scr) { 4187 if (opcode == Op_AbsVD) { 4188 vandpd(dst, src, ExternalAddress(StubRoutines::x86::vector_double_sign_mask()), vector_len, scr); 4189 } else { 4190 assert((opcode == Op_NegVD),"opcode should be Op_NegD"); 4191 vxorpd(dst, src, ExternalAddress(StubRoutines::x86::vector_double_sign_flip()), vector_len, scr); 4192 } 4193 } 4194 4195 void MacroAssembler::vabsnegf(int opcode, XMMRegister dst, XMMRegister src, Register scr) { 4196 if (dst != src) { 4197 movdqu(dst, src); 4198 } 4199 if (opcode == Op_AbsVF) { 4200 andps(dst, ExternalAddress(StubRoutines::x86::vector_float_sign_mask()), scr); 4201 } else { 4202 assert((opcode == Op_NegVF),"opcode should be Op_NegF"); 4203 xorps(dst, ExternalAddress(StubRoutines::x86::vector_float_sign_flip()), scr); 4204 } 4205 } 4206 4207 void MacroAssembler::vabsnegf(int opcode, XMMRegister dst, XMMRegister src, int vector_len, Register scr) { 4208 if (opcode == Op_AbsVF) { 4209 vandps(dst, src, ExternalAddress(StubRoutines::x86::vector_float_sign_mask()), vector_len, scr); 4210 } else { 4211 assert((opcode == Op_NegVF),"opcode should be Op_NegF"); 4212 vxorps(dst, src, ExternalAddress(StubRoutines::x86::vector_float_sign_flip()), vector_len, scr); 4213 } 4214 } 4215 4216 void MacroAssembler::vextendbw(bool sign, XMMRegister dst, XMMRegister src) { 4217 if (sign) { 4218 pmovsxbw(dst, src); 4219 } else { 4220 pmovzxbw(dst, src); 4221 } 4222 } 4223 4224 void MacroAssembler::vextendbw(bool sign, XMMRegister dst, XMMRegister src, int vector_len) { 4225 if (sign) { 4226 vpmovsxbw(dst, src, vector_len); 4227 } else { 4228 vpmovzxbw(dst, src, vector_len); 4229 } 4230 } 4231 4232 void MacroAssembler::vshiftd(int opcode, XMMRegister dst, XMMRegister src) { 4233 if (opcode == Op_RShiftVI) { 4234 psrad(dst, src); 4235 } else if (opcode == Op_LShiftVI) { 4236 pslld(dst, src); 4237 } else { 4238 assert((opcode == Op_URShiftVI),"opcode should be Op_URShiftVI"); 4239 psrld(dst, src); 4240 } 4241 } 4242 4243 void MacroAssembler::vshiftd(int opcode, XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4244 if (opcode == Op_RShiftVI) { 4245 vpsrad(dst, nds, src, vector_len); 4246 } else if (opcode == Op_LShiftVI) { 4247 vpslld(dst, nds, src, vector_len); 4248 } else { 4249 assert((opcode == Op_URShiftVI),"opcode should be Op_URShiftVI"); 4250 vpsrld(dst, nds, src, vector_len); 4251 } 4252 } 4253 4254 void MacroAssembler::vshiftw(int opcode, XMMRegister dst, XMMRegister src) { 4255 if ((opcode == Op_RShiftVS) || (opcode == Op_RShiftVB)) { 4256 psraw(dst, src); 4257 } else if ((opcode == Op_LShiftVS) || (opcode == Op_LShiftVB)) { 4258 psllw(dst, src); 4259 } else { 4260 assert(((opcode == Op_URShiftVS) || (opcode == Op_URShiftVB)),"opcode should be one of Op_URShiftVS or Op_URShiftVB"); 4261 psrlw(dst, src); 4262 } 4263 } 4264 4265 void MacroAssembler::vshiftw(int opcode, XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4266 if ((opcode == Op_RShiftVS) || (opcode == Op_RShiftVB)) { 4267 vpsraw(dst, nds, src, vector_len); 4268 } else if ((opcode == Op_LShiftVS) || (opcode == Op_LShiftVB)) { 4269 vpsllw(dst, nds, src, vector_len); 4270 } else { 4271 assert(((opcode == Op_URShiftVS) || (opcode == Op_URShiftVB)),"opcode should be one of Op_URShiftVS or Op_URShiftVB"); 4272 vpsrlw(dst, nds, src, vector_len); 4273 } 4274 } 4275 4276 void MacroAssembler::vshiftq(int opcode, XMMRegister dst, XMMRegister src) { 4277 if (opcode == Op_RShiftVL) { 4278 psrlq(dst, src); // using srl to implement sra on pre-avs512 systems 4279 } else if (opcode == Op_LShiftVL) { 4280 psllq(dst, src); 4281 } else { 4282 assert((opcode == Op_URShiftVL),"opcode should be Op_URShiftVL"); 4283 psrlq(dst, src); 4284 } 4285 } 4286 4287 void MacroAssembler::vshiftq(int opcode, XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 4288 if (opcode == Op_RShiftVL) { 4289 evpsraq(dst, nds, src, vector_len); 4290 } else if (opcode == Op_LShiftVL) { 4291 vpsllq(dst, nds, src, vector_len); 4292 } else { 4293 assert((opcode == Op_URShiftVL),"opcode should be Op_URShiftVL"); 4294 vpsrlq(dst, nds, src, vector_len); 4295 } 4296 } 4297 #endif 4298 //------------------------------------------------------------------------------------------- 4299 4300 void MacroAssembler::clear_jweak_tag(Register possibly_jweak) { 4301 const int32_t inverted_jweak_mask = ~static_cast<int32_t>(JNIHandles::weak_tag_mask); 4302 STATIC_ASSERT(inverted_jweak_mask == -2); // otherwise check this code 4303 // The inverted mask is sign-extended 4304 andptr(possibly_jweak, inverted_jweak_mask); 4305 } 4306 4307 void MacroAssembler::resolve_jobject(Register value, 4308 Register thread, 4309 Register tmp) { 4310 assert_different_registers(value, thread, tmp); 4311 Label done, not_weak; 4312 testptr(value, value); 4313 jcc(Assembler::zero, done); // Use NULL as-is. 4314 testptr(value, JNIHandles::weak_tag_mask); // Test for jweak tag. 4315 jcc(Assembler::zero, not_weak); 4316 // Resolve jweak. 4317 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF, 4318 value, Address(value, -JNIHandles::weak_tag_value), tmp, thread); 4319 verify_oop(value); 4320 jmp(done); 4321 bind(not_weak); 4322 // Resolve (untagged) jobject. 4323 access_load_at(T_OBJECT, IN_NATIVE, value, Address(value, 0), tmp, thread); 4324 verify_oop(value); 4325 bind(done); 4326 } 4327 4328 void MacroAssembler::subptr(Register dst, int32_t imm32) { 4329 LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32)); 4330 } 4331 4332 // Force generation of a 4 byte immediate value even if it fits into 8bit 4333 void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) { 4334 LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32)); 4335 } 4336 4337 void MacroAssembler::subptr(Register dst, Register src) { 4338 LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src)); 4339 } 4340 4341 // C++ bool manipulation 4342 void MacroAssembler::testbool(Register dst) { 4343 if(sizeof(bool) == 1) 4344 testb(dst, 0xff); 4345 else if(sizeof(bool) == 2) { 4346 // testw implementation needed for two byte bools 4347 ShouldNotReachHere(); 4348 } else if(sizeof(bool) == 4) 4349 testl(dst, dst); 4350 else 4351 // unsupported 4352 ShouldNotReachHere(); 4353 } 4354 4355 void MacroAssembler::testptr(Register dst, Register src) { 4356 LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src)); 4357 } 4358 4359 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes. 4360 void MacroAssembler::tlab_allocate(Register thread, Register obj, 4361 Register var_size_in_bytes, 4362 int con_size_in_bytes, 4363 Register t1, 4364 Register t2, 4365 Label& slow_case) { 4366 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 4367 bs->tlab_allocate(this, thread, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case); 4368 } 4369 4370 // Defines obj, preserves var_size_in_bytes 4371 void MacroAssembler::eden_allocate(Register thread, Register obj, 4372 Register var_size_in_bytes, 4373 int con_size_in_bytes, 4374 Register t1, 4375 Label& slow_case) { 4376 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 4377 bs->eden_allocate(this, thread, obj, var_size_in_bytes, con_size_in_bytes, t1, slow_case); 4378 } 4379 4380 // Preserves the contents of address, destroys the contents length_in_bytes and temp. 4381 void MacroAssembler::zero_memory(Register address, Register length_in_bytes, int offset_in_bytes, Register temp) { 4382 assert(address != length_in_bytes && address != temp && temp != length_in_bytes, "registers must be different"); 4383 assert((offset_in_bytes & (BytesPerWord - 1)) == 0, "offset must be a multiple of BytesPerWord"); 4384 Label done; 4385 4386 testptr(length_in_bytes, length_in_bytes); 4387 jcc(Assembler::zero, done); 4388 4389 // initialize topmost word, divide index by 2, check if odd and test if zero 4390 // note: for the remaining code to work, index must be a multiple of BytesPerWord 4391 #ifdef ASSERT 4392 { 4393 Label L; 4394 testptr(length_in_bytes, BytesPerWord - 1); 4395 jcc(Assembler::zero, L); 4396 stop("length must be a multiple of BytesPerWord"); 4397 bind(L); 4398 } 4399 #endif 4400 Register index = length_in_bytes; 4401 xorptr(temp, temp); // use _zero reg to clear memory (shorter code) 4402 if (UseIncDec) { 4403 shrptr(index, 3); // divide by 8/16 and set carry flag if bit 2 was set 4404 } else { 4405 shrptr(index, 2); // use 2 instructions to avoid partial flag stall 4406 shrptr(index, 1); 4407 } 4408 #ifndef _LP64 4409 // index could have not been a multiple of 8 (i.e., bit 2 was set) 4410 { 4411 Label even; 4412 // note: if index was a multiple of 8, then it cannot 4413 // be 0 now otherwise it must have been 0 before 4414 // => if it is even, we don't need to check for 0 again 4415 jcc(Assembler::carryClear, even); 4416 // clear topmost word (no jump would be needed if conditional assignment worked here) 4417 movptr(Address(address, index, Address::times_8, offset_in_bytes - 0*BytesPerWord), temp); 4418 // index could be 0 now, must check again 4419 jcc(Assembler::zero, done); 4420 bind(even); 4421 } 4422 #endif // !_LP64 4423 // initialize remaining object fields: index is a multiple of 2 now 4424 { 4425 Label loop; 4426 bind(loop); 4427 movptr(Address(address, index, Address::times_8, offset_in_bytes - 1*BytesPerWord), temp); 4428 NOT_LP64(movptr(Address(address, index, Address::times_8, offset_in_bytes - 2*BytesPerWord), temp);) 4429 decrement(index); 4430 jcc(Assembler::notZero, loop); 4431 } 4432 4433 bind(done); 4434 } 4435 4436 // Look up the method for a megamorphic invokeinterface call. 4437 // The target method is determined by <intf_klass, itable_index>. 4438 // The receiver klass is in recv_klass. 4439 // On success, the result will be in method_result, and execution falls through. 4440 // On failure, execution transfers to the given label. 4441 void MacroAssembler::lookup_interface_method(Register recv_klass, 4442 Register intf_klass, 4443 RegisterOrConstant itable_index, 4444 Register method_result, 4445 Register scan_temp, 4446 Label& L_no_such_interface, 4447 bool return_method) { 4448 assert_different_registers(recv_klass, intf_klass, scan_temp); 4449 assert_different_registers(method_result, intf_klass, scan_temp); 4450 assert(recv_klass != method_result || !return_method, 4451 "recv_klass can be destroyed when method isn't needed"); 4452 4453 assert(itable_index.is_constant() || itable_index.as_register() == method_result, 4454 "caller must use same register for non-constant itable index as for method"); 4455 4456 // Compute start of first itableOffsetEntry (which is at the end of the vtable) 4457 int vtable_base = in_bytes(Klass::vtable_start_offset()); 4458 int itentry_off = itableMethodEntry::method_offset_in_bytes(); 4459 int scan_step = itableOffsetEntry::size() * wordSize; 4460 int vte_size = vtableEntry::size_in_bytes(); 4461 Address::ScaleFactor times_vte_scale = Address::times_ptr; 4462 assert(vte_size == wordSize, "else adjust times_vte_scale"); 4463 4464 movl(scan_temp, Address(recv_klass, Klass::vtable_length_offset())); 4465 4466 // %%% Could store the aligned, prescaled offset in the klassoop. 4467 lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base)); 4468 4469 if (return_method) { 4470 // Adjust recv_klass by scaled itable_index, so we can free itable_index. 4471 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below"); 4472 lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off)); 4473 } 4474 4475 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) { 4476 // if (scan->interface() == intf) { 4477 // result = (klass + scan->offset() + itable_index); 4478 // } 4479 // } 4480 Label search, found_method; 4481 4482 for (int peel = 1; peel >= 0; peel--) { 4483 movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes())); 4484 cmpptr(intf_klass, method_result); 4485 4486 if (peel) { 4487 jccb(Assembler::equal, found_method); 4488 } else { 4489 jccb(Assembler::notEqual, search); 4490 // (invert the test to fall through to found_method...) 4491 } 4492 4493 if (!peel) break; 4494 4495 bind(search); 4496 4497 // Check that the previous entry is non-null. A null entry means that 4498 // the receiver class doesn't implement the interface, and wasn't the 4499 // same as when the caller was compiled. 4500 testptr(method_result, method_result); 4501 jcc(Assembler::zero, L_no_such_interface); 4502 addptr(scan_temp, scan_step); 4503 } 4504 4505 bind(found_method); 4506 4507 if (return_method) { 4508 // Got a hit. 4509 movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes())); 4510 movptr(method_result, Address(recv_klass, scan_temp, Address::times_1)); 4511 } 4512 } 4513 4514 4515 // virtual method calling 4516 void MacroAssembler::lookup_virtual_method(Register recv_klass, 4517 RegisterOrConstant vtable_index, 4518 Register method_result) { 4519 const int base = in_bytes(Klass::vtable_start_offset()); 4520 assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below"); 4521 Address vtable_entry_addr(recv_klass, 4522 vtable_index, Address::times_ptr, 4523 base + vtableEntry::method_offset_in_bytes()); 4524 movptr(method_result, vtable_entry_addr); 4525 } 4526 4527 4528 void MacroAssembler::check_klass_subtype(Register sub_klass, 4529 Register super_klass, 4530 Register temp_reg, 4531 Label& L_success) { 4532 Label L_failure; 4533 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL); 4534 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL); 4535 bind(L_failure); 4536 } 4537 4538 4539 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass, 4540 Register super_klass, 4541 Register temp_reg, 4542 Label* L_success, 4543 Label* L_failure, 4544 Label* L_slow_path, 4545 RegisterOrConstant super_check_offset) { 4546 assert_different_registers(sub_klass, super_klass, temp_reg); 4547 bool must_load_sco = (super_check_offset.constant_or_zero() == -1); 4548 if (super_check_offset.is_register()) { 4549 assert_different_registers(sub_klass, super_klass, 4550 super_check_offset.as_register()); 4551 } else if (must_load_sco) { 4552 assert(temp_reg != noreg, "supply either a temp or a register offset"); 4553 } 4554 4555 Label L_fallthrough; 4556 int label_nulls = 0; 4557 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 4558 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 4559 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; } 4560 assert(label_nulls <= 1, "at most one NULL in the batch"); 4561 4562 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 4563 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 4564 Address super_check_offset_addr(super_klass, sco_offset); 4565 4566 // Hacked jcc, which "knows" that L_fallthrough, at least, is in 4567 // range of a jccb. If this routine grows larger, reconsider at 4568 // least some of these. 4569 #define local_jcc(assembler_cond, label) \ 4570 if (&(label) == &L_fallthrough) jccb(assembler_cond, label); \ 4571 else jcc( assembler_cond, label) /*omit semi*/ 4572 4573 // Hacked jmp, which may only be used just before L_fallthrough. 4574 #define final_jmp(label) \ 4575 if (&(label) == &L_fallthrough) { /*do nothing*/ } \ 4576 else jmp(label) /*omit semi*/ 4577 4578 // If the pointers are equal, we are done (e.g., String[] elements). 4579 // This self-check enables sharing of secondary supertype arrays among 4580 // non-primary types such as array-of-interface. Otherwise, each such 4581 // type would need its own customized SSA. 4582 // We move this check to the front of the fast path because many 4583 // type checks are in fact trivially successful in this manner, 4584 // so we get a nicely predicted branch right at the start of the check. 4585 cmpptr(sub_klass, super_klass); 4586 local_jcc(Assembler::equal, *L_success); 4587 4588 // Check the supertype display: 4589 if (must_load_sco) { 4590 // Positive movl does right thing on LP64. 4591 movl(temp_reg, super_check_offset_addr); 4592 super_check_offset = RegisterOrConstant(temp_reg); 4593 } 4594 Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0); 4595 cmpptr(super_klass, super_check_addr); // load displayed supertype 4596 4597 // This check has worked decisively for primary supers. 4598 // Secondary supers are sought in the super_cache ('super_cache_addr'). 4599 // (Secondary supers are interfaces and very deeply nested subtypes.) 4600 // This works in the same check above because of a tricky aliasing 4601 // between the super_cache and the primary super display elements. 4602 // (The 'super_check_addr' can address either, as the case requires.) 4603 // Note that the cache is updated below if it does not help us find 4604 // what we need immediately. 4605 // So if it was a primary super, we can just fail immediately. 4606 // Otherwise, it's the slow path for us (no success at this point). 4607 4608 if (super_check_offset.is_register()) { 4609 local_jcc(Assembler::equal, *L_success); 4610 cmpl(super_check_offset.as_register(), sc_offset); 4611 if (L_failure == &L_fallthrough) { 4612 local_jcc(Assembler::equal, *L_slow_path); 4613 } else { 4614 local_jcc(Assembler::notEqual, *L_failure); 4615 final_jmp(*L_slow_path); 4616 } 4617 } else if (super_check_offset.as_constant() == sc_offset) { 4618 // Need a slow path; fast failure is impossible. 4619 if (L_slow_path == &L_fallthrough) { 4620 local_jcc(Assembler::equal, *L_success); 4621 } else { 4622 local_jcc(Assembler::notEqual, *L_slow_path); 4623 final_jmp(*L_success); 4624 } 4625 } else { 4626 // No slow path; it's a fast decision. 4627 if (L_failure == &L_fallthrough) { 4628 local_jcc(Assembler::equal, *L_success); 4629 } else { 4630 local_jcc(Assembler::notEqual, *L_failure); 4631 final_jmp(*L_success); 4632 } 4633 } 4634 4635 bind(L_fallthrough); 4636 4637 #undef local_jcc 4638 #undef final_jmp 4639 } 4640 4641 4642 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass, 4643 Register super_klass, 4644 Register temp_reg, 4645 Register temp2_reg, 4646 Label* L_success, 4647 Label* L_failure, 4648 bool set_cond_codes) { 4649 assert_different_registers(sub_klass, super_klass, temp_reg); 4650 if (temp2_reg != noreg) 4651 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg); 4652 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg) 4653 4654 Label L_fallthrough; 4655 int label_nulls = 0; 4656 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 4657 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 4658 assert(label_nulls <= 1, "at most one NULL in the batch"); 4659 4660 // a couple of useful fields in sub_klass: 4661 int ss_offset = in_bytes(Klass::secondary_supers_offset()); 4662 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 4663 Address secondary_supers_addr(sub_klass, ss_offset); 4664 Address super_cache_addr( sub_klass, sc_offset); 4665 4666 // Do a linear scan of the secondary super-klass chain. 4667 // This code is rarely used, so simplicity is a virtue here. 4668 // The repne_scan instruction uses fixed registers, which we must spill. 4669 // Don't worry too much about pre-existing connections with the input regs. 4670 4671 assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super) 4672 assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter) 4673 4674 // Get super_klass value into rax (even if it was in rdi or rcx). 4675 bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false; 4676 if (super_klass != rax || UseCompressedOops) { 4677 if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; } 4678 mov(rax, super_klass); 4679 } 4680 if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; } 4681 if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; } 4682 4683 #ifndef PRODUCT 4684 int* pst_counter = &SharedRuntime::_partial_subtype_ctr; 4685 ExternalAddress pst_counter_addr((address) pst_counter); 4686 NOT_LP64( incrementl(pst_counter_addr) ); 4687 LP64_ONLY( lea(rcx, pst_counter_addr) ); 4688 LP64_ONLY( incrementl(Address(rcx, 0)) ); 4689 #endif //PRODUCT 4690 4691 // We will consult the secondary-super array. 4692 movptr(rdi, secondary_supers_addr); 4693 // Load the array length. (Positive movl does right thing on LP64.) 4694 movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes())); 4695 // Skip to start of data. 4696 addptr(rdi, Array<Klass*>::base_offset_in_bytes()); 4697 4698 // Scan RCX words at [RDI] for an occurrence of RAX. 4699 // Set NZ/Z based on last compare. 4700 // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does 4701 // not change flags (only scas instruction which is repeated sets flags). 4702 // Set Z = 0 (not equal) before 'repne' to indicate that class was not found. 4703 4704 testptr(rax,rax); // Set Z = 0 4705 repne_scan(); 4706 4707 // Unspill the temp. registers: 4708 if (pushed_rdi) pop(rdi); 4709 if (pushed_rcx) pop(rcx); 4710 if (pushed_rax) pop(rax); 4711 4712 if (set_cond_codes) { 4713 // Special hack for the AD files: rdi is guaranteed non-zero. 4714 assert(!pushed_rdi, "rdi must be left non-NULL"); 4715 // Also, the condition codes are properly set Z/NZ on succeed/failure. 4716 } 4717 4718 if (L_failure == &L_fallthrough) 4719 jccb(Assembler::notEqual, *L_failure); 4720 else jcc(Assembler::notEqual, *L_failure); 4721 4722 // Success. Cache the super we found and proceed in triumph. 4723 movptr(super_cache_addr, super_klass); 4724 4725 if (L_success != &L_fallthrough) { 4726 jmp(*L_success); 4727 } 4728 4729 #undef IS_A_TEMP 4730 4731 bind(L_fallthrough); 4732 } 4733 4734 void MacroAssembler::clinit_barrier(Register klass, Register thread, Label* L_fast_path, Label* L_slow_path) { 4735 assert(L_fast_path != NULL || L_slow_path != NULL, "at least one is required"); 4736 4737 Label L_fallthrough; 4738 if (L_fast_path == NULL) { 4739 L_fast_path = &L_fallthrough; 4740 } else if (L_slow_path == NULL) { 4741 L_slow_path = &L_fallthrough; 4742 } 4743 4744 // Fast path check: class is fully initialized 4745 cmpb(Address(klass, InstanceKlass::init_state_offset()), InstanceKlass::fully_initialized); 4746 jcc(Assembler::equal, *L_fast_path); 4747 4748 // Fast path check: current thread is initializer thread 4749 cmpptr(thread, Address(klass, InstanceKlass::init_thread_offset())); 4750 if (L_slow_path == &L_fallthrough) { 4751 jcc(Assembler::equal, *L_fast_path); 4752 bind(*L_slow_path); 4753 } else if (L_fast_path == &L_fallthrough) { 4754 jcc(Assembler::notEqual, *L_slow_path); 4755 bind(*L_fast_path); 4756 } else { 4757 Unimplemented(); 4758 } 4759 } 4760 4761 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) { 4762 if (VM_Version::supports_cmov()) { 4763 cmovl(cc, dst, src); 4764 } else { 4765 Label L; 4766 jccb(negate_condition(cc), L); 4767 movl(dst, src); 4768 bind(L); 4769 } 4770 } 4771 4772 void MacroAssembler::cmov32(Condition cc, Register dst, Register src) { 4773 if (VM_Version::supports_cmov()) { 4774 cmovl(cc, dst, src); 4775 } else { 4776 Label L; 4777 jccb(negate_condition(cc), L); 4778 movl(dst, src); 4779 bind(L); 4780 } 4781 } 4782 4783 void MacroAssembler::verify_oop(Register reg, const char* s) { 4784 if (!VerifyOops) return; 4785 4786 // Pass register number to verify_oop_subroutine 4787 const char* b = NULL; 4788 { 4789 ResourceMark rm; 4790 stringStream ss; 4791 ss.print("verify_oop: %s: %s", reg->name(), s); 4792 b = code_string(ss.as_string()); 4793 } 4794 BLOCK_COMMENT("verify_oop {"); 4795 #ifdef _LP64 4796 push(rscratch1); // save r10, trashed by movptr() 4797 #endif 4798 push(rax); // save rax, 4799 push(reg); // pass register argument 4800 ExternalAddress buffer((address) b); 4801 // avoid using pushptr, as it modifies scratch registers 4802 // and our contract is not to modify anything 4803 movptr(rax, buffer.addr()); 4804 push(rax); 4805 // call indirectly to solve generation ordering problem 4806 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); 4807 call(rax); 4808 // Caller pops the arguments (oop, message) and restores rax, r10 4809 BLOCK_COMMENT("} verify_oop"); 4810 } 4811 4812 4813 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr, 4814 Register tmp, 4815 int offset) { 4816 intptr_t value = *delayed_value_addr; 4817 if (value != 0) 4818 return RegisterOrConstant(value + offset); 4819 4820 // load indirectly to solve generation ordering problem 4821 movptr(tmp, ExternalAddress((address) delayed_value_addr)); 4822 4823 #ifdef ASSERT 4824 { Label L; 4825 testptr(tmp, tmp); 4826 if (WizardMode) { 4827 const char* buf = NULL; 4828 { 4829 ResourceMark rm; 4830 stringStream ss; 4831 ss.print("DelayedValue=" INTPTR_FORMAT, delayed_value_addr[1]); 4832 buf = code_string(ss.as_string()); 4833 } 4834 jcc(Assembler::notZero, L); 4835 STOP(buf); 4836 } else { 4837 jccb(Assembler::notZero, L); 4838 hlt(); 4839 } 4840 bind(L); 4841 } 4842 #endif 4843 4844 if (offset != 0) 4845 addptr(tmp, offset); 4846 4847 return RegisterOrConstant(tmp); 4848 } 4849 4850 4851 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot, 4852 int extra_slot_offset) { 4853 // cf. TemplateTable::prepare_invoke(), if (load_receiver). 4854 int stackElementSize = Interpreter::stackElementSize; 4855 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0); 4856 #ifdef ASSERT 4857 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1); 4858 assert(offset1 - offset == stackElementSize, "correct arithmetic"); 4859 #endif 4860 Register scale_reg = noreg; 4861 Address::ScaleFactor scale_factor = Address::no_scale; 4862 if (arg_slot.is_constant()) { 4863 offset += arg_slot.as_constant() * stackElementSize; 4864 } else { 4865 scale_reg = arg_slot.as_register(); 4866 scale_factor = Address::times(stackElementSize); 4867 } 4868 offset += wordSize; // return PC is on stack 4869 return Address(rsp, scale_reg, scale_factor, offset); 4870 } 4871 4872 4873 void MacroAssembler::verify_oop_addr(Address addr, const char* s) { 4874 if (!VerifyOops) return; 4875 4876 // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord); 4877 // Pass register number to verify_oop_subroutine 4878 const char* b = NULL; 4879 { 4880 ResourceMark rm; 4881 stringStream ss; 4882 ss.print("verify_oop_addr: %s", s); 4883 b = code_string(ss.as_string()); 4884 } 4885 #ifdef _LP64 4886 push(rscratch1); // save r10, trashed by movptr() 4887 #endif 4888 push(rax); // save rax, 4889 // addr may contain rsp so we will have to adjust it based on the push 4890 // we just did (and on 64 bit we do two pushes) 4891 // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which 4892 // stores rax into addr which is backwards of what was intended. 4893 if (addr.uses(rsp)) { 4894 lea(rax, addr); 4895 pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord)); 4896 } else { 4897 pushptr(addr); 4898 } 4899 4900 ExternalAddress buffer((address) b); 4901 // pass msg argument 4902 // avoid using pushptr, as it modifies scratch registers 4903 // and our contract is not to modify anything 4904 movptr(rax, buffer.addr()); 4905 push(rax); 4906 4907 // call indirectly to solve generation ordering problem 4908 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); 4909 call(rax); 4910 // Caller pops the arguments (addr, message) and restores rax, r10. 4911 } 4912 4913 void MacroAssembler::verify_tlab() { 4914 #ifdef ASSERT 4915 if (UseTLAB && VerifyOops) { 4916 Label next, ok; 4917 Register t1 = rsi; 4918 Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread); 4919 4920 push(t1); 4921 NOT_LP64(push(thread_reg)); 4922 NOT_LP64(get_thread(thread_reg)); 4923 4924 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 4925 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset()))); 4926 jcc(Assembler::aboveEqual, next); 4927 STOP("assert(top >= start)"); 4928 should_not_reach_here(); 4929 4930 bind(next); 4931 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset()))); 4932 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 4933 jcc(Assembler::aboveEqual, ok); 4934 STOP("assert(top <= end)"); 4935 should_not_reach_here(); 4936 4937 bind(ok); 4938 NOT_LP64(pop(thread_reg)); 4939 pop(t1); 4940 } 4941 #endif 4942 } 4943 4944 class ControlWord { 4945 public: 4946 int32_t _value; 4947 4948 int rounding_control() const { return (_value >> 10) & 3 ; } 4949 int precision_control() const { return (_value >> 8) & 3 ; } 4950 bool precision() const { return ((_value >> 5) & 1) != 0; } 4951 bool underflow() const { return ((_value >> 4) & 1) != 0; } 4952 bool overflow() const { return ((_value >> 3) & 1) != 0; } 4953 bool zero_divide() const { return ((_value >> 2) & 1) != 0; } 4954 bool denormalized() const { return ((_value >> 1) & 1) != 0; } 4955 bool invalid() const { return ((_value >> 0) & 1) != 0; } 4956 4957 void print() const { 4958 // rounding control 4959 const char* rc; 4960 switch (rounding_control()) { 4961 case 0: rc = "round near"; break; 4962 case 1: rc = "round down"; break; 4963 case 2: rc = "round up "; break; 4964 case 3: rc = "chop "; break; 4965 }; 4966 // precision control 4967 const char* pc; 4968 switch (precision_control()) { 4969 case 0: pc = "24 bits "; break; 4970 case 1: pc = "reserved"; break; 4971 case 2: pc = "53 bits "; break; 4972 case 3: pc = "64 bits "; break; 4973 }; 4974 // flags 4975 char f[9]; 4976 f[0] = ' '; 4977 f[1] = ' '; 4978 f[2] = (precision ()) ? 'P' : 'p'; 4979 f[3] = (underflow ()) ? 'U' : 'u'; 4980 f[4] = (overflow ()) ? 'O' : 'o'; 4981 f[5] = (zero_divide ()) ? 'Z' : 'z'; 4982 f[6] = (denormalized()) ? 'D' : 'd'; 4983 f[7] = (invalid ()) ? 'I' : 'i'; 4984 f[8] = '\x0'; 4985 // output 4986 printf("%04x masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc); 4987 } 4988 4989 }; 4990 4991 class StatusWord { 4992 public: 4993 int32_t _value; 4994 4995 bool busy() const { return ((_value >> 15) & 1) != 0; } 4996 bool C3() const { return ((_value >> 14) & 1) != 0; } 4997 bool C2() const { return ((_value >> 10) & 1) != 0; } 4998 bool C1() const { return ((_value >> 9) & 1) != 0; } 4999 bool C0() const { return ((_value >> 8) & 1) != 0; } 5000 int top() const { return (_value >> 11) & 7 ; } 5001 bool error_status() const { return ((_value >> 7) & 1) != 0; } 5002 bool stack_fault() const { return ((_value >> 6) & 1) != 0; } 5003 bool precision() const { return ((_value >> 5) & 1) != 0; } 5004 bool underflow() const { return ((_value >> 4) & 1) != 0; } 5005 bool overflow() const { return ((_value >> 3) & 1) != 0; } 5006 bool zero_divide() const { return ((_value >> 2) & 1) != 0; } 5007 bool denormalized() const { return ((_value >> 1) & 1) != 0; } 5008 bool invalid() const { return ((_value >> 0) & 1) != 0; } 5009 5010 void print() const { 5011 // condition codes 5012 char c[5]; 5013 c[0] = (C3()) ? '3' : '-'; 5014 c[1] = (C2()) ? '2' : '-'; 5015 c[2] = (C1()) ? '1' : '-'; 5016 c[3] = (C0()) ? '0' : '-'; 5017 c[4] = '\x0'; 5018 // flags 5019 char f[9]; 5020 f[0] = (error_status()) ? 'E' : '-'; 5021 f[1] = (stack_fault ()) ? 'S' : '-'; 5022 f[2] = (precision ()) ? 'P' : '-'; 5023 f[3] = (underflow ()) ? 'U' : '-'; 5024 f[4] = (overflow ()) ? 'O' : '-'; 5025 f[5] = (zero_divide ()) ? 'Z' : '-'; 5026 f[6] = (denormalized()) ? 'D' : '-'; 5027 f[7] = (invalid ()) ? 'I' : '-'; 5028 f[8] = '\x0'; 5029 // output 5030 printf("%04x flags = %s, cc = %s, top = %d", _value & 0xFFFF, f, c, top()); 5031 } 5032 5033 }; 5034 5035 class TagWord { 5036 public: 5037 int32_t _value; 5038 5039 int tag_at(int i) const { return (_value >> (i*2)) & 3; } 5040 5041 void print() const { 5042 printf("%04x", _value & 0xFFFF); 5043 } 5044 5045 }; 5046 5047 class FPU_Register { 5048 public: 5049 int32_t _m0; 5050 int32_t _m1; 5051 int16_t _ex; 5052 5053 bool is_indefinite() const { 5054 return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0; 5055 } 5056 5057 void print() const { 5058 char sign = (_ex < 0) ? '-' : '+'; 5059 const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : " "; 5060 printf("%c%04hx.%08x%08x %s", sign, _ex, _m1, _m0, kind); 5061 }; 5062 5063 }; 5064 5065 class FPU_State { 5066 public: 5067 enum { 5068 register_size = 10, 5069 number_of_registers = 8, 5070 register_mask = 7 5071 }; 5072 5073 ControlWord _control_word; 5074 StatusWord _status_word; 5075 TagWord _tag_word; 5076 int32_t _error_offset; 5077 int32_t _error_selector; 5078 int32_t _data_offset; 5079 int32_t _data_selector; 5080 int8_t _register[register_size * number_of_registers]; 5081 5082 int tag_for_st(int i) const { return _tag_word.tag_at((_status_word.top() + i) & register_mask); } 5083 FPU_Register* st(int i) const { return (FPU_Register*)&_register[register_size * i]; } 5084 5085 const char* tag_as_string(int tag) const { 5086 switch (tag) { 5087 case 0: return "valid"; 5088 case 1: return "zero"; 5089 case 2: return "special"; 5090 case 3: return "empty"; 5091 } 5092 ShouldNotReachHere(); 5093 return NULL; 5094 } 5095 5096 void print() const { 5097 // print computation registers 5098 { int t = _status_word.top(); 5099 for (int i = 0; i < number_of_registers; i++) { 5100 int j = (i - t) & register_mask; 5101 printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j); 5102 st(j)->print(); 5103 printf(" %s\n", tag_as_string(_tag_word.tag_at(i))); 5104 } 5105 } 5106 printf("\n"); 5107 // print control registers 5108 printf("ctrl = "); _control_word.print(); printf("\n"); 5109 printf("stat = "); _status_word .print(); printf("\n"); 5110 printf("tags = "); _tag_word .print(); printf("\n"); 5111 } 5112 5113 }; 5114 5115 class Flag_Register { 5116 public: 5117 int32_t _value; 5118 5119 bool overflow() const { return ((_value >> 11) & 1) != 0; } 5120 bool direction() const { return ((_value >> 10) & 1) != 0; } 5121 bool sign() const { return ((_value >> 7) & 1) != 0; } 5122 bool zero() const { return ((_value >> 6) & 1) != 0; } 5123 bool auxiliary_carry() const { return ((_value >> 4) & 1) != 0; } 5124 bool parity() const { return ((_value >> 2) & 1) != 0; } 5125 bool carry() const { return ((_value >> 0) & 1) != 0; } 5126 5127 void print() const { 5128 // flags 5129 char f[8]; 5130 f[0] = (overflow ()) ? 'O' : '-'; 5131 f[1] = (direction ()) ? 'D' : '-'; 5132 f[2] = (sign ()) ? 'S' : '-'; 5133 f[3] = (zero ()) ? 'Z' : '-'; 5134 f[4] = (auxiliary_carry()) ? 'A' : '-'; 5135 f[5] = (parity ()) ? 'P' : '-'; 5136 f[6] = (carry ()) ? 'C' : '-'; 5137 f[7] = '\x0'; 5138 // output 5139 printf("%08x flags = %s", _value, f); 5140 } 5141 5142 }; 5143 5144 class IU_Register { 5145 public: 5146 int32_t _value; 5147 5148 void print() const { 5149 printf("%08x %11d", _value, _value); 5150 } 5151 5152 }; 5153 5154 class IU_State { 5155 public: 5156 Flag_Register _eflags; 5157 IU_Register _rdi; 5158 IU_Register _rsi; 5159 IU_Register _rbp; 5160 IU_Register _rsp; 5161 IU_Register _rbx; 5162 IU_Register _rdx; 5163 IU_Register _rcx; 5164 IU_Register _rax; 5165 5166 void print() const { 5167 // computation registers 5168 printf("rax, = "); _rax.print(); printf("\n"); 5169 printf("rbx, = "); _rbx.print(); printf("\n"); 5170 printf("rcx = "); _rcx.print(); printf("\n"); 5171 printf("rdx = "); _rdx.print(); printf("\n"); 5172 printf("rdi = "); _rdi.print(); printf("\n"); 5173 printf("rsi = "); _rsi.print(); printf("\n"); 5174 printf("rbp, = "); _rbp.print(); printf("\n"); 5175 printf("rsp = "); _rsp.print(); printf("\n"); 5176 printf("\n"); 5177 // control registers 5178 printf("flgs = "); _eflags.print(); printf("\n"); 5179 } 5180 }; 5181 5182 5183 class CPU_State { 5184 public: 5185 FPU_State _fpu_state; 5186 IU_State _iu_state; 5187 5188 void print() const { 5189 printf("--------------------------------------------------\n"); 5190 _iu_state .print(); 5191 printf("\n"); 5192 _fpu_state.print(); 5193 printf("--------------------------------------------------\n"); 5194 } 5195 5196 }; 5197 5198 5199 static void _print_CPU_state(CPU_State* state) { 5200 state->print(); 5201 }; 5202 5203 5204 void MacroAssembler::print_CPU_state() { 5205 push_CPU_state(); 5206 push(rsp); // pass CPU state 5207 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state))); 5208 addptr(rsp, wordSize); // discard argument 5209 pop_CPU_state(); 5210 } 5211 5212 5213 static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) { 5214 static int counter = 0; 5215 FPU_State* fs = &state->_fpu_state; 5216 counter++; 5217 // For leaf calls, only verify that the top few elements remain empty. 5218 // We only need 1 empty at the top for C2 code. 5219 if( stack_depth < 0 ) { 5220 if( fs->tag_for_st(7) != 3 ) { 5221 printf("FPR7 not empty\n"); 5222 state->print(); 5223 assert(false, "error"); 5224 return false; 5225 } 5226 return true; // All other stack states do not matter 5227 } 5228 5229 assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std, 5230 "bad FPU control word"); 5231 5232 // compute stack depth 5233 int i = 0; 5234 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) < 3) i++; 5235 int d = i; 5236 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++; 5237 // verify findings 5238 if (i != FPU_State::number_of_registers) { 5239 // stack not contiguous 5240 printf("%s: stack not contiguous at ST%d\n", s, i); 5241 state->print(); 5242 assert(false, "error"); 5243 return false; 5244 } 5245 // check if computed stack depth corresponds to expected stack depth 5246 if (stack_depth < 0) { 5247 // expected stack depth is -stack_depth or less 5248 if (d > -stack_depth) { 5249 // too many elements on the stack 5250 printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d); 5251 state->print(); 5252 assert(false, "error"); 5253 return false; 5254 } 5255 } else { 5256 // expected stack depth is stack_depth 5257 if (d != stack_depth) { 5258 // wrong stack depth 5259 printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d); 5260 state->print(); 5261 assert(false, "error"); 5262 return false; 5263 } 5264 } 5265 // everything is cool 5266 return true; 5267 } 5268 5269 5270 void MacroAssembler::verify_FPU(int stack_depth, const char* s) { 5271 if (!VerifyFPU) return; 5272 push_CPU_state(); 5273 push(rsp); // pass CPU state 5274 ExternalAddress msg((address) s); 5275 // pass message string s 5276 pushptr(msg.addr()); 5277 push(stack_depth); // pass stack depth 5278 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU))); 5279 addptr(rsp, 3 * wordSize); // discard arguments 5280 // check for error 5281 { Label L; 5282 testl(rax, rax); 5283 jcc(Assembler::notZero, L); 5284 int3(); // break if error condition 5285 bind(L); 5286 } 5287 pop_CPU_state(); 5288 } 5289 5290 void MacroAssembler::restore_cpu_control_state_after_jni() { 5291 // Either restore the MXCSR register after returning from the JNI Call 5292 // or verify that it wasn't changed (with -Xcheck:jni flag). 5293 if (VM_Version::supports_sse()) { 5294 if (RestoreMXCSROnJNICalls) { 5295 ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std())); 5296 } else if (CheckJNICalls) { 5297 call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry())); 5298 } 5299 } 5300 // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty. 5301 vzeroupper(); 5302 // Reset k1 to 0xffff. 5303 5304 #ifdef COMPILER2 5305 if (PostLoopMultiversioning && VM_Version::supports_evex()) { 5306 push(rcx); 5307 movl(rcx, 0xffff); 5308 kmovwl(k1, rcx); 5309 pop(rcx); 5310 } 5311 #endif // COMPILER2 5312 5313 #ifndef _LP64 5314 // Either restore the x87 floating pointer control word after returning 5315 // from the JNI call or verify that it wasn't changed. 5316 if (CheckJNICalls) { 5317 call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry())); 5318 } 5319 #endif // _LP64 5320 } 5321 5322 // ((OopHandle)result).resolve(); 5323 void MacroAssembler::resolve_oop_handle(Register result, Register tmp) { 5324 assert_different_registers(result, tmp); 5325 5326 // Only 64 bit platforms support GCs that require a tmp register 5327 // Only IN_HEAP loads require a thread_tmp register 5328 // OopHandle::resolve is an indirection like jobject. 5329 access_load_at(T_OBJECT, IN_NATIVE, 5330 result, Address(result, 0), tmp, /*tmp_thread*/noreg); 5331 } 5332 5333 // ((WeakHandle)result).resolve(); 5334 void MacroAssembler::resolve_weak_handle(Register rresult, Register rtmp) { 5335 assert_different_registers(rresult, rtmp); 5336 Label resolved; 5337 5338 // A null weak handle resolves to null. 5339 cmpptr(rresult, 0); 5340 jcc(Assembler::equal, resolved); 5341 5342 // Only 64 bit platforms support GCs that require a tmp register 5343 // Only IN_HEAP loads require a thread_tmp register 5344 // WeakHandle::resolve is an indirection like jweak. 5345 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF, 5346 rresult, Address(rresult, 0), rtmp, /*tmp_thread*/noreg); 5347 bind(resolved); 5348 } 5349 5350 void MacroAssembler::load_mirror(Register mirror, Register method, Register tmp) { 5351 // get mirror 5352 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 5353 load_method_holder(mirror, method); 5354 movptr(mirror, Address(mirror, mirror_offset)); 5355 resolve_oop_handle(mirror, tmp); 5356 } 5357 5358 void MacroAssembler::load_method_holder_cld(Register rresult, Register rmethod) { 5359 load_method_holder(rresult, rmethod); 5360 movptr(rresult, Address(rresult, InstanceKlass::class_loader_data_offset())); 5361 } 5362 5363 void MacroAssembler::load_method_holder(Register holder, Register method) { 5364 movptr(holder, Address(method, Method::const_offset())); // ConstMethod* 5365 movptr(holder, Address(holder, ConstMethod::constants_offset())); // ConstantPool* 5366 movptr(holder, Address(holder, ConstantPool::pool_holder_offset_in_bytes())); // InstanceKlass* 5367 } 5368 5369 void MacroAssembler::load_klass(Register dst, Register src) { 5370 #ifdef _LP64 5371 if (UseCompressedClassPointers) { 5372 movl(dst, Address(src, oopDesc::klass_offset_in_bytes())); 5373 decode_klass_not_null(dst); 5374 } else 5375 #endif 5376 movptr(dst, Address(src, oopDesc::klass_offset_in_bytes())); 5377 } 5378 5379 void MacroAssembler::load_prototype_header(Register dst, Register src) { 5380 load_klass(dst, src); 5381 movptr(dst, Address(dst, Klass::prototype_header_offset())); 5382 } 5383 5384 void MacroAssembler::store_klass(Register dst, Register src) { 5385 #ifdef _LP64 5386 if (UseCompressedClassPointers) { 5387 encode_klass_not_null(src); 5388 movl(Address(dst, oopDesc::klass_offset_in_bytes()), src); 5389 } else 5390 #endif 5391 movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src); 5392 } 5393 5394 void MacroAssembler::access_load_at(BasicType type, DecoratorSet decorators, Register dst, Address src, 5395 Register tmp1, Register thread_tmp) { 5396 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 5397 decorators = AccessInternal::decorator_fixup(decorators); 5398 bool as_raw = (decorators & AS_RAW) != 0; 5399 if (as_raw) { 5400 bs->BarrierSetAssembler::load_at(this, decorators, type, dst, src, tmp1, thread_tmp); 5401 } else { 5402 bs->load_at(this, decorators, type, dst, src, tmp1, thread_tmp); 5403 } 5404 } 5405 5406 void MacroAssembler::access_store_at(BasicType type, DecoratorSet decorators, Address dst, Register src, 5407 Register tmp1, Register tmp2) { 5408 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 5409 decorators = AccessInternal::decorator_fixup(decorators); 5410 bool as_raw = (decorators & AS_RAW) != 0; 5411 if (as_raw) { 5412 bs->BarrierSetAssembler::store_at(this, decorators, type, dst, src, tmp1, tmp2); 5413 } else { 5414 bs->store_at(this, decorators, type, dst, src, tmp1, tmp2); 5415 } 5416 } 5417 5418 void MacroAssembler::resolve(DecoratorSet decorators, Register obj) { 5419 // Use stronger ACCESS_WRITE|ACCESS_READ by default. 5420 if ((decorators & (ACCESS_READ | ACCESS_WRITE)) == 0) { 5421 decorators |= ACCESS_READ | ACCESS_WRITE; 5422 } 5423 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 5424 return bs->resolve(this, decorators, obj); 5425 } 5426 5427 void MacroAssembler::load_heap_oop(Register dst, Address src, Register tmp1, 5428 Register thread_tmp, DecoratorSet decorators) { 5429 access_load_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, thread_tmp); 5430 } 5431 5432 // Doesn't do verfication, generates fixed size code 5433 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src, Register tmp1, 5434 Register thread_tmp, DecoratorSet decorators) { 5435 access_load_at(T_OBJECT, IN_HEAP | IS_NOT_NULL | decorators, dst, src, tmp1, thread_tmp); 5436 } 5437 5438 void MacroAssembler::store_heap_oop(Address dst, Register src, Register tmp1, 5439 Register tmp2, DecoratorSet decorators) { 5440 access_store_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, tmp2); 5441 } 5442 5443 // Used for storing NULLs. 5444 void MacroAssembler::store_heap_oop_null(Address dst) { 5445 access_store_at(T_OBJECT, IN_HEAP, dst, noreg, noreg, noreg); 5446 } 5447 5448 #ifdef _LP64 5449 void MacroAssembler::store_klass_gap(Register dst, Register src) { 5450 if (UseCompressedClassPointers) { 5451 // Store to klass gap in destination 5452 movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src); 5453 } 5454 } 5455 5456 #ifdef ASSERT 5457 void MacroAssembler::verify_heapbase(const char* msg) { 5458 assert (UseCompressedOops, "should be compressed"); 5459 assert (Universe::heap() != NULL, "java heap should be initialized"); 5460 if (CheckCompressedOops) { 5461 Label ok; 5462 push(rscratch1); // cmpptr trashes rscratch1 5463 cmpptr(r12_heapbase, ExternalAddress((address)CompressedOops::ptrs_base_addr())); 5464 jcc(Assembler::equal, ok); 5465 STOP(msg); 5466 bind(ok); 5467 pop(rscratch1); 5468 } 5469 } 5470 #endif 5471 5472 // Algorithm must match oop.inline.hpp encode_heap_oop. 5473 void MacroAssembler::encode_heap_oop(Register r) { 5474 #ifdef ASSERT 5475 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?"); 5476 #endif 5477 verify_oop(r, "broken oop in encode_heap_oop"); 5478 if (CompressedOops::base() == NULL) { 5479 if (CompressedOops::shift() != 0) { 5480 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong"); 5481 shrq(r, LogMinObjAlignmentInBytes); 5482 } 5483 return; 5484 } 5485 testq(r, r); 5486 cmovq(Assembler::equal, r, r12_heapbase); 5487 subq(r, r12_heapbase); 5488 shrq(r, LogMinObjAlignmentInBytes); 5489 } 5490 5491 void MacroAssembler::encode_heap_oop_not_null(Register r) { 5492 #ifdef ASSERT 5493 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?"); 5494 if (CheckCompressedOops) { 5495 Label ok; 5496 testq(r, r); 5497 jcc(Assembler::notEqual, ok); 5498 STOP("null oop passed to encode_heap_oop_not_null"); 5499 bind(ok); 5500 } 5501 #endif 5502 verify_oop(r, "broken oop in encode_heap_oop_not_null"); 5503 if (CompressedOops::base() != NULL) { 5504 subq(r, r12_heapbase); 5505 } 5506 if (CompressedOops::shift() != 0) { 5507 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong"); 5508 shrq(r, LogMinObjAlignmentInBytes); 5509 } 5510 } 5511 5512 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) { 5513 #ifdef ASSERT 5514 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?"); 5515 if (CheckCompressedOops) { 5516 Label ok; 5517 testq(src, src); 5518 jcc(Assembler::notEqual, ok); 5519 STOP("null oop passed to encode_heap_oop_not_null2"); 5520 bind(ok); 5521 } 5522 #endif 5523 verify_oop(src, "broken oop in encode_heap_oop_not_null2"); 5524 if (dst != src) { 5525 movq(dst, src); 5526 } 5527 if (CompressedOops::base() != NULL) { 5528 subq(dst, r12_heapbase); 5529 } 5530 if (CompressedOops::shift() != 0) { 5531 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong"); 5532 shrq(dst, LogMinObjAlignmentInBytes); 5533 } 5534 } 5535 5536 void MacroAssembler::decode_heap_oop(Register r) { 5537 #ifdef ASSERT 5538 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?"); 5539 #endif 5540 if (CompressedOops::base() == NULL) { 5541 if (CompressedOops::shift() != 0) { 5542 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong"); 5543 shlq(r, LogMinObjAlignmentInBytes); 5544 } 5545 } else { 5546 Label done; 5547 shlq(r, LogMinObjAlignmentInBytes); 5548 jccb(Assembler::equal, done); 5549 addq(r, r12_heapbase); 5550 bind(done); 5551 } 5552 verify_oop(r, "broken oop in decode_heap_oop"); 5553 } 5554 5555 void MacroAssembler::decode_heap_oop_not_null(Register r) { 5556 // Note: it will change flags 5557 assert (UseCompressedOops, "should only be used for compressed headers"); 5558 assert (Universe::heap() != NULL, "java heap should be initialized"); 5559 // Cannot assert, unverified entry point counts instructions (see .ad file) 5560 // vtableStubs also counts instructions in pd_code_size_limit. 5561 // Also do not verify_oop as this is called by verify_oop. 5562 if (CompressedOops::shift() != 0) { 5563 assert(LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong"); 5564 shlq(r, LogMinObjAlignmentInBytes); 5565 if (CompressedOops::base() != NULL) { 5566 addq(r, r12_heapbase); 5567 } 5568 } else { 5569 assert (CompressedOops::base() == NULL, "sanity"); 5570 } 5571 } 5572 5573 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) { 5574 // Note: it will change flags 5575 assert (UseCompressedOops, "should only be used for compressed headers"); 5576 assert (Universe::heap() != NULL, "java heap should be initialized"); 5577 // Cannot assert, unverified entry point counts instructions (see .ad file) 5578 // vtableStubs also counts instructions in pd_code_size_limit. 5579 // Also do not verify_oop as this is called by verify_oop. 5580 if (CompressedOops::shift() != 0) { 5581 assert(LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong"); 5582 if (LogMinObjAlignmentInBytes == Address::times_8) { 5583 leaq(dst, Address(r12_heapbase, src, Address::times_8, 0)); 5584 } else { 5585 if (dst != src) { 5586 movq(dst, src); 5587 } 5588 shlq(dst, LogMinObjAlignmentInBytes); 5589 if (CompressedOops::base() != NULL) { 5590 addq(dst, r12_heapbase); 5591 } 5592 } 5593 } else { 5594 assert (CompressedOops::base() == NULL, "sanity"); 5595 if (dst != src) { 5596 movq(dst, src); 5597 } 5598 } 5599 } 5600 5601 void MacroAssembler::encode_klass_not_null(Register r) { 5602 if (CompressedKlassPointers::base() != NULL) { 5603 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base. 5604 assert(r != r12_heapbase, "Encoding a klass in r12"); 5605 mov64(r12_heapbase, (int64_t)CompressedKlassPointers::base()); 5606 subq(r, r12_heapbase); 5607 } 5608 if (CompressedKlassPointers::shift() != 0) { 5609 assert (LogKlassAlignmentInBytes == CompressedKlassPointers::shift(), "decode alg wrong"); 5610 shrq(r, LogKlassAlignmentInBytes); 5611 } 5612 if (CompressedKlassPointers::base() != NULL) { 5613 reinit_heapbase(); 5614 } 5615 } 5616 5617 void MacroAssembler::encode_klass_not_null(Register dst, Register src) { 5618 if (dst == src) { 5619 encode_klass_not_null(src); 5620 } else { 5621 if (CompressedKlassPointers::base() != NULL) { 5622 mov64(dst, (int64_t)CompressedKlassPointers::base()); 5623 negq(dst); 5624 addq(dst, src); 5625 } else { 5626 movptr(dst, src); 5627 } 5628 if (CompressedKlassPointers::shift() != 0) { 5629 assert (LogKlassAlignmentInBytes == CompressedKlassPointers::shift(), "decode alg wrong"); 5630 shrq(dst, LogKlassAlignmentInBytes); 5631 } 5632 } 5633 } 5634 5635 // Function instr_size_for_decode_klass_not_null() counts the instructions 5636 // generated by decode_klass_not_null(register r) and reinit_heapbase(), 5637 // when (Universe::heap() != NULL). Hence, if the instructions they 5638 // generate change, then this method needs to be updated. 5639 int MacroAssembler::instr_size_for_decode_klass_not_null() { 5640 assert (UseCompressedClassPointers, "only for compressed klass ptrs"); 5641 if (CompressedKlassPointers::base() != NULL) { 5642 // mov64 + addq + shlq? + mov64 (for reinit_heapbase()). 5643 return (CompressedKlassPointers::shift() == 0 ? 20 : 24); 5644 } else { 5645 // longest load decode klass function, mov64, leaq 5646 return 16; 5647 } 5648 } 5649 5650 // !!! If the instructions that get generated here change then function 5651 // instr_size_for_decode_klass_not_null() needs to get updated. 5652 void MacroAssembler::decode_klass_not_null(Register r) { 5653 // Note: it will change flags 5654 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 5655 assert(r != r12_heapbase, "Decoding a klass in r12"); 5656 // Cannot assert, unverified entry point counts instructions (see .ad file) 5657 // vtableStubs also counts instructions in pd_code_size_limit. 5658 // Also do not verify_oop as this is called by verify_oop. 5659 if (CompressedKlassPointers::shift() != 0) { 5660 assert(LogKlassAlignmentInBytes == CompressedKlassPointers::shift(), "decode alg wrong"); 5661 shlq(r, LogKlassAlignmentInBytes); 5662 } 5663 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base. 5664 if (CompressedKlassPointers::base() != NULL) { 5665 mov64(r12_heapbase, (int64_t)CompressedKlassPointers::base()); 5666 addq(r, r12_heapbase); 5667 reinit_heapbase(); 5668 } 5669 } 5670 5671 void MacroAssembler::decode_klass_not_null(Register dst, Register src) { 5672 // Note: it will change flags 5673 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 5674 if (dst == src) { 5675 decode_klass_not_null(dst); 5676 } else { 5677 // Cannot assert, unverified entry point counts instructions (see .ad file) 5678 // vtableStubs also counts instructions in pd_code_size_limit. 5679 // Also do not verify_oop as this is called by verify_oop. 5680 mov64(dst, (int64_t)CompressedKlassPointers::base()); 5681 if (CompressedKlassPointers::shift() != 0) { 5682 assert(LogKlassAlignmentInBytes == CompressedKlassPointers::shift(), "decode alg wrong"); 5683 assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?"); 5684 leaq(dst, Address(dst, src, Address::times_8, 0)); 5685 } else { 5686 addq(dst, src); 5687 } 5688 } 5689 } 5690 5691 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) { 5692 assert (UseCompressedOops, "should only be used for compressed headers"); 5693 assert (Universe::heap() != NULL, "java heap should be initialized"); 5694 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5695 int oop_index = oop_recorder()->find_index(obj); 5696 RelocationHolder rspec = oop_Relocation::spec(oop_index); 5697 mov_narrow_oop(dst, oop_index, rspec); 5698 } 5699 5700 void MacroAssembler::set_narrow_oop(Address dst, jobject obj) { 5701 assert (UseCompressedOops, "should only be used for compressed headers"); 5702 assert (Universe::heap() != NULL, "java heap should be initialized"); 5703 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5704 int oop_index = oop_recorder()->find_index(obj); 5705 RelocationHolder rspec = oop_Relocation::spec(oop_index); 5706 mov_narrow_oop(dst, oop_index, rspec); 5707 } 5708 5709 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) { 5710 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 5711 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5712 int klass_index = oop_recorder()->find_index(k); 5713 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 5714 mov_narrow_oop(dst, CompressedKlassPointers::encode(k), rspec); 5715 } 5716 5717 void MacroAssembler::set_narrow_klass(Address dst, Klass* k) { 5718 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 5719 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5720 int klass_index = oop_recorder()->find_index(k); 5721 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 5722 mov_narrow_oop(dst, CompressedKlassPointers::encode(k), rspec); 5723 } 5724 5725 void MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) { 5726 assert (UseCompressedOops, "should only be used for compressed headers"); 5727 assert (Universe::heap() != NULL, "java heap should be initialized"); 5728 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5729 int oop_index = oop_recorder()->find_index(obj); 5730 RelocationHolder rspec = oop_Relocation::spec(oop_index); 5731 Assembler::cmp_narrow_oop(dst, oop_index, rspec); 5732 } 5733 5734 void MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) { 5735 assert (UseCompressedOops, "should only be used for compressed headers"); 5736 assert (Universe::heap() != NULL, "java heap should be initialized"); 5737 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5738 int oop_index = oop_recorder()->find_index(obj); 5739 RelocationHolder rspec = oop_Relocation::spec(oop_index); 5740 Assembler::cmp_narrow_oop(dst, oop_index, rspec); 5741 } 5742 5743 void MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) { 5744 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 5745 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5746 int klass_index = oop_recorder()->find_index(k); 5747 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 5748 Assembler::cmp_narrow_oop(dst, CompressedKlassPointers::encode(k), rspec); 5749 } 5750 5751 void MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) { 5752 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 5753 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5754 int klass_index = oop_recorder()->find_index(k); 5755 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 5756 Assembler::cmp_narrow_oop(dst, CompressedKlassPointers::encode(k), rspec); 5757 } 5758 5759 void MacroAssembler::reinit_heapbase() { 5760 if (UseCompressedOops || UseCompressedClassPointers) { 5761 if (Universe::heap() != NULL) { 5762 if (CompressedOops::base() == NULL) { 5763 MacroAssembler::xorptr(r12_heapbase, r12_heapbase); 5764 } else { 5765 mov64(r12_heapbase, (int64_t)CompressedOops::ptrs_base()); 5766 } 5767 } else { 5768 movptr(r12_heapbase, ExternalAddress((address)CompressedOops::ptrs_base_addr())); 5769 } 5770 } 5771 } 5772 5773 #endif // _LP64 5774 5775 // C2 compiled method's prolog code. 5776 void MacroAssembler::verified_entry(int framesize, int stack_bang_size, bool fp_mode_24b, bool is_stub) { 5777 5778 // WARNING: Initial instruction MUST be 5 bytes or longer so that 5779 // NativeJump::patch_verified_entry will be able to patch out the entry 5780 // code safely. The push to verify stack depth is ok at 5 bytes, 5781 // the frame allocation can be either 3 or 6 bytes. So if we don't do 5782 // stack bang then we must use the 6 byte frame allocation even if 5783 // we have no frame. :-( 5784 assert(stack_bang_size >= framesize || stack_bang_size <= 0, "stack bang size incorrect"); 5785 5786 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned"); 5787 // Remove word for return addr 5788 framesize -= wordSize; 5789 stack_bang_size -= wordSize; 5790 5791 // Calls to C2R adapters often do not accept exceptional returns. 5792 // We require that their callers must bang for them. But be careful, because 5793 // some VM calls (such as call site linkage) can use several kilobytes of 5794 // stack. But the stack safety zone should account for that. 5795 // See bugs 4446381, 4468289, 4497237. 5796 if (stack_bang_size > 0) { 5797 generate_stack_overflow_check(stack_bang_size); 5798 5799 // We always push rbp, so that on return to interpreter rbp, will be 5800 // restored correctly and we can correct the stack. 5801 push(rbp); 5802 // Save caller's stack pointer into RBP if the frame pointer is preserved. 5803 if (PreserveFramePointer) { 5804 mov(rbp, rsp); 5805 } 5806 // Remove word for ebp 5807 framesize -= wordSize; 5808 5809 // Create frame 5810 if (framesize) { 5811 subptr(rsp, framesize); 5812 } 5813 } else { 5814 // Create frame (force generation of a 4 byte immediate value) 5815 subptr_imm32(rsp, framesize); 5816 5817 // Save RBP register now. 5818 framesize -= wordSize; 5819 movptr(Address(rsp, framesize), rbp); 5820 // Save caller's stack pointer into RBP if the frame pointer is preserved. 5821 if (PreserveFramePointer) { 5822 movptr(rbp, rsp); 5823 if (framesize > 0) { 5824 addptr(rbp, framesize); 5825 } 5826 } 5827 } 5828 5829 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth 5830 framesize -= wordSize; 5831 movptr(Address(rsp, framesize), (int32_t)0xbadb100d); 5832 } 5833 5834 #ifndef _LP64 5835 // If method sets FPU control word do it now 5836 if (fp_mode_24b) { 5837 fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24())); 5838 } 5839 if (UseSSE >= 2 && VerifyFPU) { 5840 verify_FPU(0, "FPU stack must be clean on entry"); 5841 } 5842 #endif 5843 5844 #ifdef ASSERT 5845 if (VerifyStackAtCalls) { 5846 Label L; 5847 push(rax); 5848 mov(rax, rsp); 5849 andptr(rax, StackAlignmentInBytes-1); 5850 cmpptr(rax, StackAlignmentInBytes-wordSize); 5851 pop(rax); 5852 jcc(Assembler::equal, L); 5853 STOP("Stack is not properly aligned!"); 5854 bind(L); 5855 } 5856 #endif 5857 5858 if (!is_stub) { 5859 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 5860 bs->nmethod_entry_barrier(this); 5861 } 5862 } 5863 5864 // clear memory of size 'cnt' qwords, starting at 'base' using XMM/YMM registers 5865 void MacroAssembler::xmm_clear_mem(Register base, Register cnt, XMMRegister xtmp) { 5866 // cnt - number of qwords (8-byte words). 5867 // base - start address, qword aligned. 5868 Label L_zero_64_bytes, L_loop, L_sloop, L_tail, L_end; 5869 if (UseAVX >= 2) { 5870 vpxor(xtmp, xtmp, xtmp, AVX_256bit); 5871 } else { 5872 pxor(xtmp, xtmp); 5873 } 5874 jmp(L_zero_64_bytes); 5875 5876 BIND(L_loop); 5877 if (UseAVX >= 2) { 5878 vmovdqu(Address(base, 0), xtmp); 5879 vmovdqu(Address(base, 32), xtmp); 5880 } else { 5881 movdqu(Address(base, 0), xtmp); 5882 movdqu(Address(base, 16), xtmp); 5883 movdqu(Address(base, 32), xtmp); 5884 movdqu(Address(base, 48), xtmp); 5885 } 5886 addptr(base, 64); 5887 5888 BIND(L_zero_64_bytes); 5889 subptr(cnt, 8); 5890 jccb(Assembler::greaterEqual, L_loop); 5891 addptr(cnt, 4); 5892 jccb(Assembler::less, L_tail); 5893 // Copy trailing 32 bytes 5894 if (UseAVX >= 2) { 5895 vmovdqu(Address(base, 0), xtmp); 5896 } else { 5897 movdqu(Address(base, 0), xtmp); 5898 movdqu(Address(base, 16), xtmp); 5899 } 5900 addptr(base, 32); 5901 subptr(cnt, 4); 5902 5903 BIND(L_tail); 5904 addptr(cnt, 4); 5905 jccb(Assembler::lessEqual, L_end); 5906 decrement(cnt); 5907 5908 BIND(L_sloop); 5909 movq(Address(base, 0), xtmp); 5910 addptr(base, 8); 5911 decrement(cnt); 5912 jccb(Assembler::greaterEqual, L_sloop); 5913 BIND(L_end); 5914 } 5915 5916 void MacroAssembler::clear_mem(Register base, Register cnt, Register tmp, XMMRegister xtmp, bool is_large) { 5917 // cnt - number of qwords (8-byte words). 5918 // base - start address, qword aligned. 5919 // is_large - if optimizers know cnt is larger than InitArrayShortSize 5920 assert(base==rdi, "base register must be edi for rep stos"); 5921 assert(tmp==rax, "tmp register must be eax for rep stos"); 5922 assert(cnt==rcx, "cnt register must be ecx for rep stos"); 5923 assert(InitArrayShortSize % BytesPerLong == 0, 5924 "InitArrayShortSize should be the multiple of BytesPerLong"); 5925 5926 Label DONE; 5927 5928 if (!is_large || !UseXMMForObjInit) { 5929 xorptr(tmp, tmp); 5930 } 5931 5932 if (!is_large) { 5933 Label LOOP, LONG; 5934 cmpptr(cnt, InitArrayShortSize/BytesPerLong); 5935 jccb(Assembler::greater, LONG); 5936 5937 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM 5938 5939 decrement(cnt); 5940 jccb(Assembler::negative, DONE); // Zero length 5941 5942 // Use individual pointer-sized stores for small counts: 5943 BIND(LOOP); 5944 movptr(Address(base, cnt, Address::times_ptr), tmp); 5945 decrement(cnt); 5946 jccb(Assembler::greaterEqual, LOOP); 5947 jmpb(DONE); 5948 5949 BIND(LONG); 5950 } 5951 5952 // Use longer rep-prefixed ops for non-small counts: 5953 if (UseFastStosb) { 5954 shlptr(cnt, 3); // convert to number of bytes 5955 rep_stosb(); 5956 } else if (UseXMMForObjInit) { 5957 movptr(tmp, base); 5958 xmm_clear_mem(tmp, cnt, xtmp); 5959 } else { 5960 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM 5961 rep_stos(); 5962 } 5963 5964 BIND(DONE); 5965 } 5966 5967 #ifdef COMPILER2 5968 5969 // IndexOf for constant substrings with size >= 8 chars 5970 // which don't need to be loaded through stack. 5971 void MacroAssembler::string_indexofC8(Register str1, Register str2, 5972 Register cnt1, Register cnt2, 5973 int int_cnt2, Register result, 5974 XMMRegister vec, Register tmp, 5975 int ae) { 5976 ShortBranchVerifier sbv(this); 5977 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 5978 assert(ae != StrIntrinsicNode::LU, "Invalid encoding"); 5979 5980 // This method uses the pcmpestri instruction with bound registers 5981 // inputs: 5982 // xmm - substring 5983 // rax - substring length (elements count) 5984 // mem - scanned string 5985 // rdx - string length (elements count) 5986 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts) 5987 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes) 5988 // outputs: 5989 // rcx - matched index in string 5990 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 5991 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts 5992 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8 5993 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2; 5994 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1; 5995 5996 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, 5997 RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR, 5998 MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE; 5999 6000 // Note, inline_string_indexOf() generates checks: 6001 // if (substr.count > string.count) return -1; 6002 // if (substr.count == 0) return 0; 6003 assert(int_cnt2 >= stride, "this code is used only for cnt2 >= 8 chars"); 6004 6005 // Load substring. 6006 if (ae == StrIntrinsicNode::UL) { 6007 pmovzxbw(vec, Address(str2, 0)); 6008 } else { 6009 movdqu(vec, Address(str2, 0)); 6010 } 6011 movl(cnt2, int_cnt2); 6012 movptr(result, str1); // string addr 6013 6014 if (int_cnt2 > stride) { 6015 jmpb(SCAN_TO_SUBSTR); 6016 6017 // Reload substr for rescan, this code 6018 // is executed only for large substrings (> 8 chars) 6019 bind(RELOAD_SUBSTR); 6020 if (ae == StrIntrinsicNode::UL) { 6021 pmovzxbw(vec, Address(str2, 0)); 6022 } else { 6023 movdqu(vec, Address(str2, 0)); 6024 } 6025 negptr(cnt2); // Jumped here with negative cnt2, convert to positive 6026 6027 bind(RELOAD_STR); 6028 // We came here after the beginning of the substring was 6029 // matched but the rest of it was not so we need to search 6030 // again. Start from the next element after the previous match. 6031 6032 // cnt2 is number of substring reminding elements and 6033 // cnt1 is number of string reminding elements when cmp failed. 6034 // Restored cnt1 = cnt1 - cnt2 + int_cnt2 6035 subl(cnt1, cnt2); 6036 addl(cnt1, int_cnt2); 6037 movl(cnt2, int_cnt2); // Now restore cnt2 6038 6039 decrementl(cnt1); // Shift to next element 6040 cmpl(cnt1, cnt2); 6041 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring 6042 6043 addptr(result, (1<<scale1)); 6044 6045 } // (int_cnt2 > 8) 6046 6047 // Scan string for start of substr in 16-byte vectors 6048 bind(SCAN_TO_SUBSTR); 6049 pcmpestri(vec, Address(result, 0), mode); 6050 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1 6051 subl(cnt1, stride); 6052 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string 6053 cmpl(cnt1, cnt2); 6054 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 6055 addptr(result, 16); 6056 jmpb(SCAN_TO_SUBSTR); 6057 6058 // Found a potential substr 6059 bind(FOUND_CANDIDATE); 6060 // Matched whole vector if first element matched (tmp(rcx) == 0). 6061 if (int_cnt2 == stride) { 6062 jccb(Assembler::overflow, RET_FOUND); // OF == 1 6063 } else { // int_cnt2 > 8 6064 jccb(Assembler::overflow, FOUND_SUBSTR); 6065 } 6066 // After pcmpestri tmp(rcx) contains matched element index 6067 // Compute start addr of substr 6068 lea(result, Address(result, tmp, scale1)); 6069 6070 // Make sure string is still long enough 6071 subl(cnt1, tmp); 6072 cmpl(cnt1, cnt2); 6073 if (int_cnt2 == stride) { 6074 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR); 6075 } else { // int_cnt2 > 8 6076 jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD); 6077 } 6078 // Left less then substring. 6079 6080 bind(RET_NOT_FOUND); 6081 movl(result, -1); 6082 jmp(EXIT); 6083 6084 if (int_cnt2 > stride) { 6085 // This code is optimized for the case when whole substring 6086 // is matched if its head is matched. 6087 bind(MATCH_SUBSTR_HEAD); 6088 pcmpestri(vec, Address(result, 0), mode); 6089 // Reload only string if does not match 6090 jcc(Assembler::noOverflow, RELOAD_STR); // OF == 0 6091 6092 Label CONT_SCAN_SUBSTR; 6093 // Compare the rest of substring (> 8 chars). 6094 bind(FOUND_SUBSTR); 6095 // First 8 chars are already matched. 6096 negptr(cnt2); 6097 addptr(cnt2, stride); 6098 6099 bind(SCAN_SUBSTR); 6100 subl(cnt1, stride); 6101 cmpl(cnt2, -stride); // Do not read beyond substring 6102 jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR); 6103 // Back-up strings to avoid reading beyond substring: 6104 // cnt1 = cnt1 - cnt2 + 8 6105 addl(cnt1, cnt2); // cnt2 is negative 6106 addl(cnt1, stride); 6107 movl(cnt2, stride); negptr(cnt2); 6108 bind(CONT_SCAN_SUBSTR); 6109 if (int_cnt2 < (int)G) { 6110 int tail_off1 = int_cnt2<<scale1; 6111 int tail_off2 = int_cnt2<<scale2; 6112 if (ae == StrIntrinsicNode::UL) { 6113 pmovzxbw(vec, Address(str2, cnt2, scale2, tail_off2)); 6114 } else { 6115 movdqu(vec, Address(str2, cnt2, scale2, tail_off2)); 6116 } 6117 pcmpestri(vec, Address(result, cnt2, scale1, tail_off1), mode); 6118 } else { 6119 // calculate index in register to avoid integer overflow (int_cnt2*2) 6120 movl(tmp, int_cnt2); 6121 addptr(tmp, cnt2); 6122 if (ae == StrIntrinsicNode::UL) { 6123 pmovzxbw(vec, Address(str2, tmp, scale2, 0)); 6124 } else { 6125 movdqu(vec, Address(str2, tmp, scale2, 0)); 6126 } 6127 pcmpestri(vec, Address(result, tmp, scale1, 0), mode); 6128 } 6129 // Need to reload strings pointers if not matched whole vector 6130 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0 6131 addptr(cnt2, stride); 6132 jcc(Assembler::negative, SCAN_SUBSTR); 6133 // Fall through if found full substring 6134 6135 } // (int_cnt2 > 8) 6136 6137 bind(RET_FOUND); 6138 // Found result if we matched full small substring. 6139 // Compute substr offset 6140 subptr(result, str1); 6141 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 6142 shrl(result, 1); // index 6143 } 6144 bind(EXIT); 6145 6146 } // string_indexofC8 6147 6148 // Small strings are loaded through stack if they cross page boundary. 6149 void MacroAssembler::string_indexof(Register str1, Register str2, 6150 Register cnt1, Register cnt2, 6151 int int_cnt2, Register result, 6152 XMMRegister vec, Register tmp, 6153 int ae) { 6154 ShortBranchVerifier sbv(this); 6155 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 6156 assert(ae != StrIntrinsicNode::LU, "Invalid encoding"); 6157 6158 // 6159 // int_cnt2 is length of small (< 8 chars) constant substring 6160 // or (-1) for non constant substring in which case its length 6161 // is in cnt2 register. 6162 // 6163 // Note, inline_string_indexOf() generates checks: 6164 // if (substr.count > string.count) return -1; 6165 // if (substr.count == 0) return 0; 6166 // 6167 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8 6168 assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < stride), "should be != 0"); 6169 // This method uses the pcmpestri instruction with bound registers 6170 // inputs: 6171 // xmm - substring 6172 // rax - substring length (elements count) 6173 // mem - scanned string 6174 // rdx - string length (elements count) 6175 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts) 6176 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes) 6177 // outputs: 6178 // rcx - matched index in string 6179 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 6180 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts 6181 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2; 6182 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1; 6183 6184 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR, 6185 RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR, 6186 FOUND_CANDIDATE; 6187 6188 { //======================================================== 6189 // We don't know where these strings are located 6190 // and we can't read beyond them. Load them through stack. 6191 Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR; 6192 6193 movptr(tmp, rsp); // save old SP 6194 6195 if (int_cnt2 > 0) { // small (< 8 chars) constant substring 6196 if (int_cnt2 == (1>>scale2)) { // One byte 6197 assert((ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL), "Only possible for latin1 encoding"); 6198 load_unsigned_byte(result, Address(str2, 0)); 6199 movdl(vec, result); // move 32 bits 6200 } else if (ae == StrIntrinsicNode::LL && int_cnt2 == 3) { // Three bytes 6201 // Not enough header space in 32-bit VM: 12+3 = 15. 6202 movl(result, Address(str2, -1)); 6203 shrl(result, 8); 6204 movdl(vec, result); // move 32 bits 6205 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (2>>scale2)) { // One char 6206 load_unsigned_short(result, Address(str2, 0)); 6207 movdl(vec, result); // move 32 bits 6208 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (4>>scale2)) { // Two chars 6209 movdl(vec, Address(str2, 0)); // move 32 bits 6210 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (8>>scale2)) { // Four chars 6211 movq(vec, Address(str2, 0)); // move 64 bits 6212 } else { // cnt2 = { 3, 5, 6, 7 } || (ae == StrIntrinsicNode::UL && cnt2 ={2, ..., 7}) 6213 // Array header size is 12 bytes in 32-bit VM 6214 // + 6 bytes for 3 chars == 18 bytes, 6215 // enough space to load vec and shift. 6216 assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity"); 6217 if (ae == StrIntrinsicNode::UL) { 6218 int tail_off = int_cnt2-8; 6219 pmovzxbw(vec, Address(str2, tail_off)); 6220 psrldq(vec, -2*tail_off); 6221 } 6222 else { 6223 int tail_off = int_cnt2*(1<<scale2); 6224 movdqu(vec, Address(str2, tail_off-16)); 6225 psrldq(vec, 16-tail_off); 6226 } 6227 } 6228 } else { // not constant substring 6229 cmpl(cnt2, stride); 6230 jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough 6231 6232 // We can read beyond string if srt+16 does not cross page boundary 6233 // since heaps are aligned and mapped by pages. 6234 assert(os::vm_page_size() < (int)G, "default page should be small"); 6235 movl(result, str2); // We need only low 32 bits 6236 andl(result, (os::vm_page_size()-1)); 6237 cmpl(result, (os::vm_page_size()-16)); 6238 jccb(Assembler::belowEqual, CHECK_STR); 6239 6240 // Move small strings to stack to allow load 16 bytes into vec. 6241 subptr(rsp, 16); 6242 int stk_offset = wordSize-(1<<scale2); 6243 push(cnt2); 6244 6245 bind(COPY_SUBSTR); 6246 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL) { 6247 load_unsigned_byte(result, Address(str2, cnt2, scale2, -1)); 6248 movb(Address(rsp, cnt2, scale2, stk_offset), result); 6249 } else if (ae == StrIntrinsicNode::UU) { 6250 load_unsigned_short(result, Address(str2, cnt2, scale2, -2)); 6251 movw(Address(rsp, cnt2, scale2, stk_offset), result); 6252 } 6253 decrement(cnt2); 6254 jccb(Assembler::notZero, COPY_SUBSTR); 6255 6256 pop(cnt2); 6257 movptr(str2, rsp); // New substring address 6258 } // non constant 6259 6260 bind(CHECK_STR); 6261 cmpl(cnt1, stride); 6262 jccb(Assembler::aboveEqual, BIG_STRINGS); 6263 6264 // Check cross page boundary. 6265 movl(result, str1); // We need only low 32 bits 6266 andl(result, (os::vm_page_size()-1)); 6267 cmpl(result, (os::vm_page_size()-16)); 6268 jccb(Assembler::belowEqual, BIG_STRINGS); 6269 6270 subptr(rsp, 16); 6271 int stk_offset = -(1<<scale1); 6272 if (int_cnt2 < 0) { // not constant 6273 push(cnt2); 6274 stk_offset += wordSize; 6275 } 6276 movl(cnt2, cnt1); 6277 6278 bind(COPY_STR); 6279 if (ae == StrIntrinsicNode::LL) { 6280 load_unsigned_byte(result, Address(str1, cnt2, scale1, -1)); 6281 movb(Address(rsp, cnt2, scale1, stk_offset), result); 6282 } else { 6283 load_unsigned_short(result, Address(str1, cnt2, scale1, -2)); 6284 movw(Address(rsp, cnt2, scale1, stk_offset), result); 6285 } 6286 decrement(cnt2); 6287 jccb(Assembler::notZero, COPY_STR); 6288 6289 if (int_cnt2 < 0) { // not constant 6290 pop(cnt2); 6291 } 6292 movptr(str1, rsp); // New string address 6293 6294 bind(BIG_STRINGS); 6295 // Load substring. 6296 if (int_cnt2 < 0) { // -1 6297 if (ae == StrIntrinsicNode::UL) { 6298 pmovzxbw(vec, Address(str2, 0)); 6299 } else { 6300 movdqu(vec, Address(str2, 0)); 6301 } 6302 push(cnt2); // substr count 6303 push(str2); // substr addr 6304 push(str1); // string addr 6305 } else { 6306 // Small (< 8 chars) constant substrings are loaded already. 6307 movl(cnt2, int_cnt2); 6308 } 6309 push(tmp); // original SP 6310 6311 } // Finished loading 6312 6313 //======================================================== 6314 // Start search 6315 // 6316 6317 movptr(result, str1); // string addr 6318 6319 if (int_cnt2 < 0) { // Only for non constant substring 6320 jmpb(SCAN_TO_SUBSTR); 6321 6322 // SP saved at sp+0 6323 // String saved at sp+1*wordSize 6324 // Substr saved at sp+2*wordSize 6325 // Substr count saved at sp+3*wordSize 6326 6327 // Reload substr for rescan, this code 6328 // is executed only for large substrings (> 8 chars) 6329 bind(RELOAD_SUBSTR); 6330 movptr(str2, Address(rsp, 2*wordSize)); 6331 movl(cnt2, Address(rsp, 3*wordSize)); 6332 if (ae == StrIntrinsicNode::UL) { 6333 pmovzxbw(vec, Address(str2, 0)); 6334 } else { 6335 movdqu(vec, Address(str2, 0)); 6336 } 6337 // We came here after the beginning of the substring was 6338 // matched but the rest of it was not so we need to search 6339 // again. Start from the next element after the previous match. 6340 subptr(str1, result); // Restore counter 6341 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 6342 shrl(str1, 1); 6343 } 6344 addl(cnt1, str1); 6345 decrementl(cnt1); // Shift to next element 6346 cmpl(cnt1, cnt2); 6347 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring 6348 6349 addptr(result, (1<<scale1)); 6350 } // non constant 6351 6352 // Scan string for start of substr in 16-byte vectors 6353 bind(SCAN_TO_SUBSTR); 6354 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 6355 pcmpestri(vec, Address(result, 0), mode); 6356 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1 6357 subl(cnt1, stride); 6358 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string 6359 cmpl(cnt1, cnt2); 6360 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 6361 addptr(result, 16); 6362 6363 bind(ADJUST_STR); 6364 cmpl(cnt1, stride); // Do not read beyond string 6365 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR); 6366 // Back-up string to avoid reading beyond string. 6367 lea(result, Address(result, cnt1, scale1, -16)); 6368 movl(cnt1, stride); 6369 jmpb(SCAN_TO_SUBSTR); 6370 6371 // Found a potential substr 6372 bind(FOUND_CANDIDATE); 6373 // After pcmpestri tmp(rcx) contains matched element index 6374 6375 // Make sure string is still long enough 6376 subl(cnt1, tmp); 6377 cmpl(cnt1, cnt2); 6378 jccb(Assembler::greaterEqual, FOUND_SUBSTR); 6379 // Left less then substring. 6380 6381 bind(RET_NOT_FOUND); 6382 movl(result, -1); 6383 jmp(CLEANUP); 6384 6385 bind(FOUND_SUBSTR); 6386 // Compute start addr of substr 6387 lea(result, Address(result, tmp, scale1)); 6388 if (int_cnt2 > 0) { // Constant substring 6389 // Repeat search for small substring (< 8 chars) 6390 // from new point without reloading substring. 6391 // Have to check that we don't read beyond string. 6392 cmpl(tmp, stride-int_cnt2); 6393 jccb(Assembler::greater, ADJUST_STR); 6394 // Fall through if matched whole substring. 6395 } else { // non constant 6396 assert(int_cnt2 == -1, "should be != 0"); 6397 6398 addl(tmp, cnt2); 6399 // Found result if we matched whole substring. 6400 cmpl(tmp, stride); 6401 jcc(Assembler::lessEqual, RET_FOUND); 6402 6403 // Repeat search for small substring (<= 8 chars) 6404 // from new point 'str1' without reloading substring. 6405 cmpl(cnt2, stride); 6406 // Have to check that we don't read beyond string. 6407 jccb(Assembler::lessEqual, ADJUST_STR); 6408 6409 Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG; 6410 // Compare the rest of substring (> 8 chars). 6411 movptr(str1, result); 6412 6413 cmpl(tmp, cnt2); 6414 // First 8 chars are already matched. 6415 jccb(Assembler::equal, CHECK_NEXT); 6416 6417 bind(SCAN_SUBSTR); 6418 pcmpestri(vec, Address(str1, 0), mode); 6419 // Need to reload strings pointers if not matched whole vector 6420 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0 6421 6422 bind(CHECK_NEXT); 6423 subl(cnt2, stride); 6424 jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring 6425 addptr(str1, 16); 6426 if (ae == StrIntrinsicNode::UL) { 6427 addptr(str2, 8); 6428 } else { 6429 addptr(str2, 16); 6430 } 6431 subl(cnt1, stride); 6432 cmpl(cnt2, stride); // Do not read beyond substring 6433 jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR); 6434 // Back-up strings to avoid reading beyond substring. 6435 6436 if (ae == StrIntrinsicNode::UL) { 6437 lea(str2, Address(str2, cnt2, scale2, -8)); 6438 lea(str1, Address(str1, cnt2, scale1, -16)); 6439 } else { 6440 lea(str2, Address(str2, cnt2, scale2, -16)); 6441 lea(str1, Address(str1, cnt2, scale1, -16)); 6442 } 6443 subl(cnt1, cnt2); 6444 movl(cnt2, stride); 6445 addl(cnt1, stride); 6446 bind(CONT_SCAN_SUBSTR); 6447 if (ae == StrIntrinsicNode::UL) { 6448 pmovzxbw(vec, Address(str2, 0)); 6449 } else { 6450 movdqu(vec, Address(str2, 0)); 6451 } 6452 jmp(SCAN_SUBSTR); 6453 6454 bind(RET_FOUND_LONG); 6455 movptr(str1, Address(rsp, wordSize)); 6456 } // non constant 6457 6458 bind(RET_FOUND); 6459 // Compute substr offset 6460 subptr(result, str1); 6461 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 6462 shrl(result, 1); // index 6463 } 6464 bind(CLEANUP); 6465 pop(rsp); // restore SP 6466 6467 } // string_indexof 6468 6469 void MacroAssembler::string_indexof_char(Register str1, Register cnt1, Register ch, Register result, 6470 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp) { 6471 ShortBranchVerifier sbv(this); 6472 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 6473 6474 int stride = 8; 6475 6476 Label FOUND_CHAR, SCAN_TO_CHAR, SCAN_TO_CHAR_LOOP, 6477 SCAN_TO_8_CHAR, SCAN_TO_8_CHAR_LOOP, SCAN_TO_16_CHAR_LOOP, 6478 RET_NOT_FOUND, SCAN_TO_8_CHAR_INIT, 6479 FOUND_SEQ_CHAR, DONE_LABEL; 6480 6481 movptr(result, str1); 6482 if (UseAVX >= 2) { 6483 cmpl(cnt1, stride); 6484 jcc(Assembler::less, SCAN_TO_CHAR); 6485 cmpl(cnt1, 2*stride); 6486 jcc(Assembler::less, SCAN_TO_8_CHAR_INIT); 6487 movdl(vec1, ch); 6488 vpbroadcastw(vec1, vec1, Assembler::AVX_256bit); 6489 vpxor(vec2, vec2); 6490 movl(tmp, cnt1); 6491 andl(tmp, 0xFFFFFFF0); //vector count (in chars) 6492 andl(cnt1,0x0000000F); //tail count (in chars) 6493 6494 bind(SCAN_TO_16_CHAR_LOOP); 6495 vmovdqu(vec3, Address(result, 0)); 6496 vpcmpeqw(vec3, vec3, vec1, 1); 6497 vptest(vec2, vec3); 6498 jcc(Assembler::carryClear, FOUND_CHAR); 6499 addptr(result, 32); 6500 subl(tmp, 2*stride); 6501 jcc(Assembler::notZero, SCAN_TO_16_CHAR_LOOP); 6502 jmp(SCAN_TO_8_CHAR); 6503 bind(SCAN_TO_8_CHAR_INIT); 6504 movdl(vec1, ch); 6505 pshuflw(vec1, vec1, 0x00); 6506 pshufd(vec1, vec1, 0); 6507 pxor(vec2, vec2); 6508 } 6509 bind(SCAN_TO_8_CHAR); 6510 cmpl(cnt1, stride); 6511 jcc(Assembler::less, SCAN_TO_CHAR); 6512 if (UseAVX < 2) { 6513 movdl(vec1, ch); 6514 pshuflw(vec1, vec1, 0x00); 6515 pshufd(vec1, vec1, 0); 6516 pxor(vec2, vec2); 6517 } 6518 movl(tmp, cnt1); 6519 andl(tmp, 0xFFFFFFF8); //vector count (in chars) 6520 andl(cnt1,0x00000007); //tail count (in chars) 6521 6522 bind(SCAN_TO_8_CHAR_LOOP); 6523 movdqu(vec3, Address(result, 0)); 6524 pcmpeqw(vec3, vec1); 6525 ptest(vec2, vec3); 6526 jcc(Assembler::carryClear, FOUND_CHAR); 6527 addptr(result, 16); 6528 subl(tmp, stride); 6529 jcc(Assembler::notZero, SCAN_TO_8_CHAR_LOOP); 6530 bind(SCAN_TO_CHAR); 6531 testl(cnt1, cnt1); 6532 jcc(Assembler::zero, RET_NOT_FOUND); 6533 bind(SCAN_TO_CHAR_LOOP); 6534 load_unsigned_short(tmp, Address(result, 0)); 6535 cmpl(ch, tmp); 6536 jccb(Assembler::equal, FOUND_SEQ_CHAR); 6537 addptr(result, 2); 6538 subl(cnt1, 1); 6539 jccb(Assembler::zero, RET_NOT_FOUND); 6540 jmp(SCAN_TO_CHAR_LOOP); 6541 6542 bind(RET_NOT_FOUND); 6543 movl(result, -1); 6544 jmpb(DONE_LABEL); 6545 6546 bind(FOUND_CHAR); 6547 if (UseAVX >= 2) { 6548 vpmovmskb(tmp, vec3); 6549 } else { 6550 pmovmskb(tmp, vec3); 6551 } 6552 bsfl(ch, tmp); 6553 addl(result, ch); 6554 6555 bind(FOUND_SEQ_CHAR); 6556 subptr(result, str1); 6557 shrl(result, 1); 6558 6559 bind(DONE_LABEL); 6560 } // string_indexof_char 6561 6562 // helper function for string_compare 6563 void MacroAssembler::load_next_elements(Register elem1, Register elem2, Register str1, Register str2, 6564 Address::ScaleFactor scale, Address::ScaleFactor scale1, 6565 Address::ScaleFactor scale2, Register index, int ae) { 6566 if (ae == StrIntrinsicNode::LL) { 6567 load_unsigned_byte(elem1, Address(str1, index, scale, 0)); 6568 load_unsigned_byte(elem2, Address(str2, index, scale, 0)); 6569 } else if (ae == StrIntrinsicNode::UU) { 6570 load_unsigned_short(elem1, Address(str1, index, scale, 0)); 6571 load_unsigned_short(elem2, Address(str2, index, scale, 0)); 6572 } else { 6573 load_unsigned_byte(elem1, Address(str1, index, scale1, 0)); 6574 load_unsigned_short(elem2, Address(str2, index, scale2, 0)); 6575 } 6576 } 6577 6578 // Compare strings, used for char[] and byte[]. 6579 void MacroAssembler::string_compare(Register str1, Register str2, 6580 Register cnt1, Register cnt2, Register result, 6581 XMMRegister vec1, int ae) { 6582 ShortBranchVerifier sbv(this); 6583 Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL; 6584 Label COMPARE_WIDE_VECTORS_LOOP_FAILED; // used only _LP64 && AVX3 6585 int stride, stride2, adr_stride, adr_stride1, adr_stride2; 6586 int stride2x2 = 0x40; 6587 Address::ScaleFactor scale = Address::no_scale; 6588 Address::ScaleFactor scale1 = Address::no_scale; 6589 Address::ScaleFactor scale2 = Address::no_scale; 6590 6591 if (ae != StrIntrinsicNode::LL) { 6592 stride2x2 = 0x20; 6593 } 6594 6595 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) { 6596 shrl(cnt2, 1); 6597 } 6598 // Compute the minimum of the string lengths and the 6599 // difference of the string lengths (stack). 6600 // Do the conditional move stuff 6601 movl(result, cnt1); 6602 subl(cnt1, cnt2); 6603 push(cnt1); 6604 cmov32(Assembler::lessEqual, cnt2, result); // cnt2 = min(cnt1, cnt2) 6605 6606 // Is the minimum length zero? 6607 testl(cnt2, cnt2); 6608 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 6609 if (ae == StrIntrinsicNode::LL) { 6610 // Load first bytes 6611 load_unsigned_byte(result, Address(str1, 0)); // result = str1[0] 6612 load_unsigned_byte(cnt1, Address(str2, 0)); // cnt1 = str2[0] 6613 } else if (ae == StrIntrinsicNode::UU) { 6614 // Load first characters 6615 load_unsigned_short(result, Address(str1, 0)); 6616 load_unsigned_short(cnt1, Address(str2, 0)); 6617 } else { 6618 load_unsigned_byte(result, Address(str1, 0)); 6619 load_unsigned_short(cnt1, Address(str2, 0)); 6620 } 6621 subl(result, cnt1); 6622 jcc(Assembler::notZero, POP_LABEL); 6623 6624 if (ae == StrIntrinsicNode::UU) { 6625 // Divide length by 2 to get number of chars 6626 shrl(cnt2, 1); 6627 } 6628 cmpl(cnt2, 1); 6629 jcc(Assembler::equal, LENGTH_DIFF_LABEL); 6630 6631 // Check if the strings start at the same location and setup scale and stride 6632 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6633 cmpptr(str1, str2); 6634 jcc(Assembler::equal, LENGTH_DIFF_LABEL); 6635 if (ae == StrIntrinsicNode::LL) { 6636 scale = Address::times_1; 6637 stride = 16; 6638 } else { 6639 scale = Address::times_2; 6640 stride = 8; 6641 } 6642 } else { 6643 scale1 = Address::times_1; 6644 scale2 = Address::times_2; 6645 // scale not used 6646 stride = 8; 6647 } 6648 6649 if (UseAVX >= 2 && UseSSE42Intrinsics) { 6650 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR; 6651 Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR; 6652 Label COMPARE_WIDE_VECTORS_LOOP_AVX2; 6653 Label COMPARE_TAIL_LONG; 6654 Label COMPARE_WIDE_VECTORS_LOOP_AVX3; // used only _LP64 && AVX3 6655 6656 int pcmpmask = 0x19; 6657 if (ae == StrIntrinsicNode::LL) { 6658 pcmpmask &= ~0x01; 6659 } 6660 6661 // Setup to compare 16-chars (32-bytes) vectors, 6662 // start from first character again because it has aligned address. 6663 if (ae == StrIntrinsicNode::LL) { 6664 stride2 = 32; 6665 } else { 6666 stride2 = 16; 6667 } 6668 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6669 adr_stride = stride << scale; 6670 } else { 6671 adr_stride1 = 8; //stride << scale1; 6672 adr_stride2 = 16; //stride << scale2; 6673 } 6674 6675 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri"); 6676 // rax and rdx are used by pcmpestri as elements counters 6677 movl(result, cnt2); 6678 andl(cnt2, ~(stride2-1)); // cnt2 holds the vector count 6679 jcc(Assembler::zero, COMPARE_TAIL_LONG); 6680 6681 // fast path : compare first 2 8-char vectors. 6682 bind(COMPARE_16_CHARS); 6683 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6684 movdqu(vec1, Address(str1, 0)); 6685 } else { 6686 pmovzxbw(vec1, Address(str1, 0)); 6687 } 6688 pcmpestri(vec1, Address(str2, 0), pcmpmask); 6689 jccb(Assembler::below, COMPARE_INDEX_CHAR); 6690 6691 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6692 movdqu(vec1, Address(str1, adr_stride)); 6693 pcmpestri(vec1, Address(str2, adr_stride), pcmpmask); 6694 } else { 6695 pmovzxbw(vec1, Address(str1, adr_stride1)); 6696 pcmpestri(vec1, Address(str2, adr_stride2), pcmpmask); 6697 } 6698 jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS); 6699 addl(cnt1, stride); 6700 6701 // Compare the characters at index in cnt1 6702 bind(COMPARE_INDEX_CHAR); // cnt1 has the offset of the mismatching character 6703 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae); 6704 subl(result, cnt2); 6705 jmp(POP_LABEL); 6706 6707 // Setup the registers to start vector comparison loop 6708 bind(COMPARE_WIDE_VECTORS); 6709 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6710 lea(str1, Address(str1, result, scale)); 6711 lea(str2, Address(str2, result, scale)); 6712 } else { 6713 lea(str1, Address(str1, result, scale1)); 6714 lea(str2, Address(str2, result, scale2)); 6715 } 6716 subl(result, stride2); 6717 subl(cnt2, stride2); 6718 jcc(Assembler::zero, COMPARE_WIDE_TAIL); 6719 negptr(result); 6720 6721 // In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest) 6722 bind(COMPARE_WIDE_VECTORS_LOOP); 6723 6724 #ifdef _LP64 6725 if ((AVX3Threshold == 0) && VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop 6726 cmpl(cnt2, stride2x2); 6727 jccb(Assembler::below, COMPARE_WIDE_VECTORS_LOOP_AVX2); 6728 testl(cnt2, stride2x2-1); // cnt2 holds the vector count 6729 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX2); // means we cannot subtract by 0x40 6730 6731 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop 6732 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6733 evmovdquq(vec1, Address(str1, result, scale), Assembler::AVX_512bit); 6734 evpcmpeqb(k7, vec1, Address(str2, result, scale), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0 6735 } else { 6736 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_512bit); 6737 evpcmpeqb(k7, vec1, Address(str2, result, scale2), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0 6738 } 6739 kortestql(k7, k7); 6740 jcc(Assembler::aboveEqual, COMPARE_WIDE_VECTORS_LOOP_FAILED); // miscompare 6741 addptr(result, stride2x2); // update since we already compared at this addr 6742 subl(cnt2, stride2x2); // and sub the size too 6743 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX3); 6744 6745 vpxor(vec1, vec1); 6746 jmpb(COMPARE_WIDE_TAIL); 6747 }//if (VM_Version::supports_avx512vlbw()) 6748 #endif // _LP64 6749 6750 6751 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2); 6752 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6753 vmovdqu(vec1, Address(str1, result, scale)); 6754 vpxor(vec1, Address(str2, result, scale)); 6755 } else { 6756 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_256bit); 6757 vpxor(vec1, Address(str2, result, scale2)); 6758 } 6759 vptest(vec1, vec1); 6760 jcc(Assembler::notZero, VECTOR_NOT_EQUAL); 6761 addptr(result, stride2); 6762 subl(cnt2, stride2); 6763 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP); 6764 // clean upper bits of YMM registers 6765 vpxor(vec1, vec1); 6766 6767 // compare wide vectors tail 6768 bind(COMPARE_WIDE_TAIL); 6769 testptr(result, result); 6770 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 6771 6772 movl(result, stride2); 6773 movl(cnt2, result); 6774 negptr(result); 6775 jmp(COMPARE_WIDE_VECTORS_LOOP_AVX2); 6776 6777 // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors. 6778 bind(VECTOR_NOT_EQUAL); 6779 // clean upper bits of YMM registers 6780 vpxor(vec1, vec1); 6781 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6782 lea(str1, Address(str1, result, scale)); 6783 lea(str2, Address(str2, result, scale)); 6784 } else { 6785 lea(str1, Address(str1, result, scale1)); 6786 lea(str2, Address(str2, result, scale2)); 6787 } 6788 jmp(COMPARE_16_CHARS); 6789 6790 // Compare tail chars, length between 1 to 15 chars 6791 bind(COMPARE_TAIL_LONG); 6792 movl(cnt2, result); 6793 cmpl(cnt2, stride); 6794 jcc(Assembler::less, COMPARE_SMALL_STR); 6795 6796 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6797 movdqu(vec1, Address(str1, 0)); 6798 } else { 6799 pmovzxbw(vec1, Address(str1, 0)); 6800 } 6801 pcmpestri(vec1, Address(str2, 0), pcmpmask); 6802 jcc(Assembler::below, COMPARE_INDEX_CHAR); 6803 subptr(cnt2, stride); 6804 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 6805 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6806 lea(str1, Address(str1, result, scale)); 6807 lea(str2, Address(str2, result, scale)); 6808 } else { 6809 lea(str1, Address(str1, result, scale1)); 6810 lea(str2, Address(str2, result, scale2)); 6811 } 6812 negptr(cnt2); 6813 jmpb(WHILE_HEAD_LABEL); 6814 6815 bind(COMPARE_SMALL_STR); 6816 } else if (UseSSE42Intrinsics) { 6817 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL; 6818 int pcmpmask = 0x19; 6819 // Setup to compare 8-char (16-byte) vectors, 6820 // start from first character again because it has aligned address. 6821 movl(result, cnt2); 6822 andl(cnt2, ~(stride - 1)); // cnt2 holds the vector count 6823 if (ae == StrIntrinsicNode::LL) { 6824 pcmpmask &= ~0x01; 6825 } 6826 jcc(Assembler::zero, COMPARE_TAIL); 6827 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6828 lea(str1, Address(str1, result, scale)); 6829 lea(str2, Address(str2, result, scale)); 6830 } else { 6831 lea(str1, Address(str1, result, scale1)); 6832 lea(str2, Address(str2, result, scale2)); 6833 } 6834 negptr(result); 6835 6836 // pcmpestri 6837 // inputs: 6838 // vec1- substring 6839 // rax - negative string length (elements count) 6840 // mem - scanned string 6841 // rdx - string length (elements count) 6842 // pcmpmask - cmp mode: 11000 (string compare with negated result) 6843 // + 00 (unsigned bytes) or + 01 (unsigned shorts) 6844 // outputs: 6845 // rcx - first mismatched element index 6846 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri"); 6847 6848 bind(COMPARE_WIDE_VECTORS); 6849 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6850 movdqu(vec1, Address(str1, result, scale)); 6851 pcmpestri(vec1, Address(str2, result, scale), pcmpmask); 6852 } else { 6853 pmovzxbw(vec1, Address(str1, result, scale1)); 6854 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask); 6855 } 6856 // After pcmpestri cnt1(rcx) contains mismatched element index 6857 6858 jccb(Assembler::below, VECTOR_NOT_EQUAL); // CF==1 6859 addptr(result, stride); 6860 subptr(cnt2, stride); 6861 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS); 6862 6863 // compare wide vectors tail 6864 testptr(result, result); 6865 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 6866 6867 movl(cnt2, stride); 6868 movl(result, stride); 6869 negptr(result); 6870 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6871 movdqu(vec1, Address(str1, result, scale)); 6872 pcmpestri(vec1, Address(str2, result, scale), pcmpmask); 6873 } else { 6874 pmovzxbw(vec1, Address(str1, result, scale1)); 6875 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask); 6876 } 6877 jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL); 6878 6879 // Mismatched characters in the vectors 6880 bind(VECTOR_NOT_EQUAL); 6881 addptr(cnt1, result); 6882 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae); 6883 subl(result, cnt2); 6884 jmpb(POP_LABEL); 6885 6886 bind(COMPARE_TAIL); // limit is zero 6887 movl(cnt2, result); 6888 // Fallthru to tail compare 6889 } 6890 // Shift str2 and str1 to the end of the arrays, negate min 6891 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6892 lea(str1, Address(str1, cnt2, scale)); 6893 lea(str2, Address(str2, cnt2, scale)); 6894 } else { 6895 lea(str1, Address(str1, cnt2, scale1)); 6896 lea(str2, Address(str2, cnt2, scale2)); 6897 } 6898 decrementl(cnt2); // first character was compared already 6899 negptr(cnt2); 6900 6901 // Compare the rest of the elements 6902 bind(WHILE_HEAD_LABEL); 6903 load_next_elements(result, cnt1, str1, str2, scale, scale1, scale2, cnt2, ae); 6904 subl(result, cnt1); 6905 jccb(Assembler::notZero, POP_LABEL); 6906 increment(cnt2); 6907 jccb(Assembler::notZero, WHILE_HEAD_LABEL); 6908 6909 // Strings are equal up to min length. Return the length difference. 6910 bind(LENGTH_DIFF_LABEL); 6911 pop(result); 6912 if (ae == StrIntrinsicNode::UU) { 6913 // Divide diff by 2 to get number of chars 6914 sarl(result, 1); 6915 } 6916 jmpb(DONE_LABEL); 6917 6918 #ifdef _LP64 6919 if (VM_Version::supports_avx512vlbw()) { 6920 6921 bind(COMPARE_WIDE_VECTORS_LOOP_FAILED); 6922 6923 kmovql(cnt1, k7); 6924 notq(cnt1); 6925 bsfq(cnt2, cnt1); 6926 if (ae != StrIntrinsicNode::LL) { 6927 // Divide diff by 2 to get number of chars 6928 sarl(cnt2, 1); 6929 } 6930 addq(result, cnt2); 6931 if (ae == StrIntrinsicNode::LL) { 6932 load_unsigned_byte(cnt1, Address(str2, result)); 6933 load_unsigned_byte(result, Address(str1, result)); 6934 } else if (ae == StrIntrinsicNode::UU) { 6935 load_unsigned_short(cnt1, Address(str2, result, scale)); 6936 load_unsigned_short(result, Address(str1, result, scale)); 6937 } else { 6938 load_unsigned_short(cnt1, Address(str2, result, scale2)); 6939 load_unsigned_byte(result, Address(str1, result, scale1)); 6940 } 6941 subl(result, cnt1); 6942 jmpb(POP_LABEL); 6943 }//if (VM_Version::supports_avx512vlbw()) 6944 #endif // _LP64 6945 6946 // Discard the stored length difference 6947 bind(POP_LABEL); 6948 pop(cnt1); 6949 6950 // That's it 6951 bind(DONE_LABEL); 6952 if(ae == StrIntrinsicNode::UL) { 6953 negl(result); 6954 } 6955 6956 } 6957 6958 // Search for Non-ASCII character (Negative byte value) in a byte array, 6959 // return true if it has any and false otherwise. 6960 // ..\jdk\src\java.base\share\classes\java\lang\StringCoding.java 6961 // @HotSpotIntrinsicCandidate 6962 // private static boolean hasNegatives(byte[] ba, int off, int len) { 6963 // for (int i = off; i < off + len; i++) { 6964 // if (ba[i] < 0) { 6965 // return true; 6966 // } 6967 // } 6968 // return false; 6969 // } 6970 void MacroAssembler::has_negatives(Register ary1, Register len, 6971 Register result, Register tmp1, 6972 XMMRegister vec1, XMMRegister vec2) { 6973 // rsi: byte array 6974 // rcx: len 6975 // rax: result 6976 ShortBranchVerifier sbv(this); 6977 assert_different_registers(ary1, len, result, tmp1); 6978 assert_different_registers(vec1, vec2); 6979 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_CHAR, COMPARE_VECTORS, COMPARE_BYTE; 6980 6981 // len == 0 6982 testl(len, len); 6983 jcc(Assembler::zero, FALSE_LABEL); 6984 6985 if ((AVX3Threshold == 0) && (UseAVX > 2) && // AVX512 6986 VM_Version::supports_avx512vlbw() && 6987 VM_Version::supports_bmi2()) { 6988 6989 Label test_64_loop, test_tail; 6990 Register tmp3_aliased = len; 6991 6992 movl(tmp1, len); 6993 vpxor(vec2, vec2, vec2, Assembler::AVX_512bit); 6994 6995 andl(tmp1, 64 - 1); // tail count (in chars) 0x3F 6996 andl(len, ~(64 - 1)); // vector count (in chars) 6997 jccb(Assembler::zero, test_tail); 6998 6999 lea(ary1, Address(ary1, len, Address::times_1)); 7000 negptr(len); 7001 7002 bind(test_64_loop); 7003 // Check whether our 64 elements of size byte contain negatives 7004 evpcmpgtb(k2, vec2, Address(ary1, len, Address::times_1), Assembler::AVX_512bit); 7005 kortestql(k2, k2); 7006 jcc(Assembler::notZero, TRUE_LABEL); 7007 7008 addptr(len, 64); 7009 jccb(Assembler::notZero, test_64_loop); 7010 7011 7012 bind(test_tail); 7013 // bail out when there is nothing to be done 7014 testl(tmp1, -1); 7015 jcc(Assembler::zero, FALSE_LABEL); 7016 7017 // ~(~0 << len) applied up to two times (for 32-bit scenario) 7018 #ifdef _LP64 7019 mov64(tmp3_aliased, 0xFFFFFFFFFFFFFFFF); 7020 shlxq(tmp3_aliased, tmp3_aliased, tmp1); 7021 notq(tmp3_aliased); 7022 kmovql(k3, tmp3_aliased); 7023 #else 7024 Label k_init; 7025 jmp(k_init); 7026 7027 // We could not read 64-bits from a general purpose register thus we move 7028 // data required to compose 64 1's to the instruction stream 7029 // We emit 64 byte wide series of elements from 0..63 which later on would 7030 // be used as a compare targets with tail count contained in tmp1 register. 7031 // Result would be a k register having tmp1 consecutive number or 1 7032 // counting from least significant bit. 7033 address tmp = pc(); 7034 emit_int64(0x0706050403020100); 7035 emit_int64(0x0F0E0D0C0B0A0908); 7036 emit_int64(0x1716151413121110); 7037 emit_int64(0x1F1E1D1C1B1A1918); 7038 emit_int64(0x2726252423222120); 7039 emit_int64(0x2F2E2D2C2B2A2928); 7040 emit_int64(0x3736353433323130); 7041 emit_int64(0x3F3E3D3C3B3A3938); 7042 7043 bind(k_init); 7044 lea(len, InternalAddress(tmp)); 7045 // create mask to test for negative byte inside a vector 7046 evpbroadcastb(vec1, tmp1, Assembler::AVX_512bit); 7047 evpcmpgtb(k3, vec1, Address(len, 0), Assembler::AVX_512bit); 7048 7049 #endif 7050 evpcmpgtb(k2, k3, vec2, Address(ary1, 0), Assembler::AVX_512bit); 7051 ktestq(k2, k3); 7052 jcc(Assembler::notZero, TRUE_LABEL); 7053 7054 jmp(FALSE_LABEL); 7055 } else { 7056 movl(result, len); // copy 7057 7058 if (UseAVX >= 2 && UseSSE >= 2) { 7059 // With AVX2, use 32-byte vector compare 7060 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 7061 7062 // Compare 32-byte vectors 7063 andl(result, 0x0000001f); // tail count (in bytes) 7064 andl(len, 0xffffffe0); // vector count (in bytes) 7065 jccb(Assembler::zero, COMPARE_TAIL); 7066 7067 lea(ary1, Address(ary1, len, Address::times_1)); 7068 negptr(len); 7069 7070 movl(tmp1, 0x80808080); // create mask to test for Unicode chars in vector 7071 movdl(vec2, tmp1); 7072 vpbroadcastd(vec2, vec2, Assembler::AVX_256bit); 7073 7074 bind(COMPARE_WIDE_VECTORS); 7075 vmovdqu(vec1, Address(ary1, len, Address::times_1)); 7076 vptest(vec1, vec2); 7077 jccb(Assembler::notZero, TRUE_LABEL); 7078 addptr(len, 32); 7079 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 7080 7081 testl(result, result); 7082 jccb(Assembler::zero, FALSE_LABEL); 7083 7084 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32)); 7085 vptest(vec1, vec2); 7086 jccb(Assembler::notZero, TRUE_LABEL); 7087 jmpb(FALSE_LABEL); 7088 7089 bind(COMPARE_TAIL); // len is zero 7090 movl(len, result); 7091 // Fallthru to tail compare 7092 } else if (UseSSE42Intrinsics) { 7093 // With SSE4.2, use double quad vector compare 7094 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 7095 7096 // Compare 16-byte vectors 7097 andl(result, 0x0000000f); // tail count (in bytes) 7098 andl(len, 0xfffffff0); // vector count (in bytes) 7099 jcc(Assembler::zero, COMPARE_TAIL); 7100 7101 lea(ary1, Address(ary1, len, Address::times_1)); 7102 negptr(len); 7103 7104 movl(tmp1, 0x80808080); 7105 movdl(vec2, tmp1); 7106 pshufd(vec2, vec2, 0); 7107 7108 bind(COMPARE_WIDE_VECTORS); 7109 movdqu(vec1, Address(ary1, len, Address::times_1)); 7110 ptest(vec1, vec2); 7111 jcc(Assembler::notZero, TRUE_LABEL); 7112 addptr(len, 16); 7113 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 7114 7115 testl(result, result); 7116 jcc(Assembler::zero, FALSE_LABEL); 7117 7118 movdqu(vec1, Address(ary1, result, Address::times_1, -16)); 7119 ptest(vec1, vec2); 7120 jccb(Assembler::notZero, TRUE_LABEL); 7121 jmpb(FALSE_LABEL); 7122 7123 bind(COMPARE_TAIL); // len is zero 7124 movl(len, result); 7125 // Fallthru to tail compare 7126 } 7127 } 7128 // Compare 4-byte vectors 7129 andl(len, 0xfffffffc); // vector count (in bytes) 7130 jccb(Assembler::zero, COMPARE_CHAR); 7131 7132 lea(ary1, Address(ary1, len, Address::times_1)); 7133 negptr(len); 7134 7135 bind(COMPARE_VECTORS); 7136 movl(tmp1, Address(ary1, len, Address::times_1)); 7137 andl(tmp1, 0x80808080); 7138 jccb(Assembler::notZero, TRUE_LABEL); 7139 addptr(len, 4); 7140 jcc(Assembler::notZero, COMPARE_VECTORS); 7141 7142 // Compare trailing char (final 2 bytes), if any 7143 bind(COMPARE_CHAR); 7144 testl(result, 0x2); // tail char 7145 jccb(Assembler::zero, COMPARE_BYTE); 7146 load_unsigned_short(tmp1, Address(ary1, 0)); 7147 andl(tmp1, 0x00008080); 7148 jccb(Assembler::notZero, TRUE_LABEL); 7149 subptr(result, 2); 7150 lea(ary1, Address(ary1, 2)); 7151 7152 bind(COMPARE_BYTE); 7153 testl(result, 0x1); // tail byte 7154 jccb(Assembler::zero, FALSE_LABEL); 7155 load_unsigned_byte(tmp1, Address(ary1, 0)); 7156 andl(tmp1, 0x00000080); 7157 jccb(Assembler::notEqual, TRUE_LABEL); 7158 jmpb(FALSE_LABEL); 7159 7160 bind(TRUE_LABEL); 7161 movl(result, 1); // return true 7162 jmpb(DONE); 7163 7164 bind(FALSE_LABEL); 7165 xorl(result, result); // return false 7166 7167 // That's it 7168 bind(DONE); 7169 if (UseAVX >= 2 && UseSSE >= 2) { 7170 // clean upper bits of YMM registers 7171 vpxor(vec1, vec1); 7172 vpxor(vec2, vec2); 7173 } 7174 } 7175 // Compare char[] or byte[] arrays aligned to 4 bytes or substrings. 7176 void MacroAssembler::arrays_equals(bool is_array_equ, Register ary1, Register ary2, 7177 Register limit, Register result, Register chr, 7178 XMMRegister vec1, XMMRegister vec2, bool is_char) { 7179 ShortBranchVerifier sbv(this); 7180 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR, COMPARE_BYTE; 7181 7182 int length_offset = arrayOopDesc::length_offset_in_bytes(); 7183 int base_offset = arrayOopDesc::base_offset_in_bytes(is_char ? T_CHAR : T_BYTE); 7184 7185 if (is_array_equ) { 7186 // Check the input args 7187 cmpoop(ary1, ary2); 7188 jcc(Assembler::equal, TRUE_LABEL); 7189 7190 // Need additional checks for arrays_equals. 7191 testptr(ary1, ary1); 7192 jcc(Assembler::zero, FALSE_LABEL); 7193 testptr(ary2, ary2); 7194 jcc(Assembler::zero, FALSE_LABEL); 7195 7196 // Check the lengths 7197 movl(limit, Address(ary1, length_offset)); 7198 cmpl(limit, Address(ary2, length_offset)); 7199 jcc(Assembler::notEqual, FALSE_LABEL); 7200 } 7201 7202 // count == 0 7203 testl(limit, limit); 7204 jcc(Assembler::zero, TRUE_LABEL); 7205 7206 if (is_array_equ) { 7207 // Load array address 7208 lea(ary1, Address(ary1, base_offset)); 7209 lea(ary2, Address(ary2, base_offset)); 7210 } 7211 7212 if (is_array_equ && is_char) { 7213 // arrays_equals when used for char[]. 7214 shll(limit, 1); // byte count != 0 7215 } 7216 movl(result, limit); // copy 7217 7218 if (UseAVX >= 2) { 7219 // With AVX2, use 32-byte vector compare 7220 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 7221 7222 // Compare 32-byte vectors 7223 andl(result, 0x0000001f); // tail count (in bytes) 7224 andl(limit, 0xffffffe0); // vector count (in bytes) 7225 jcc(Assembler::zero, COMPARE_TAIL); 7226 7227 lea(ary1, Address(ary1, limit, Address::times_1)); 7228 lea(ary2, Address(ary2, limit, Address::times_1)); 7229 negptr(limit); 7230 7231 #ifdef _LP64 7232 if ((AVX3Threshold == 0) && VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop 7233 Label COMPARE_WIDE_VECTORS_LOOP_AVX2, COMPARE_WIDE_VECTORS_LOOP_AVX3; 7234 7235 cmpl(limit, -64); 7236 jcc(Assembler::greater, COMPARE_WIDE_VECTORS_LOOP_AVX2); 7237 7238 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop 7239 7240 evmovdquq(vec1, Address(ary1, limit, Address::times_1), Assembler::AVX_512bit); 7241 evpcmpeqb(k7, vec1, Address(ary2, limit, Address::times_1), Assembler::AVX_512bit); 7242 kortestql(k7, k7); 7243 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare 7244 addptr(limit, 64); // update since we already compared at this addr 7245 cmpl(limit, -64); 7246 jccb(Assembler::lessEqual, COMPARE_WIDE_VECTORS_LOOP_AVX3); 7247 7248 // At this point we may still need to compare -limit+result bytes. 7249 // We could execute the next two instruction and just continue via non-wide path: 7250 // cmpl(limit, 0); 7251 // jcc(Assembler::equal, COMPARE_TAIL); // true 7252 // But since we stopped at the points ary{1,2}+limit which are 7253 // not farther than 64 bytes from the ends of arrays ary{1,2}+result 7254 // (|limit| <= 32 and result < 32), 7255 // we may just compare the last 64 bytes. 7256 // 7257 addptr(result, -64); // it is safe, bc we just came from this area 7258 evmovdquq(vec1, Address(ary1, result, Address::times_1), Assembler::AVX_512bit); 7259 evpcmpeqb(k7, vec1, Address(ary2, result, Address::times_1), Assembler::AVX_512bit); 7260 kortestql(k7, k7); 7261 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare 7262 7263 jmp(TRUE_LABEL); 7264 7265 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2); 7266 7267 }//if (VM_Version::supports_avx512vlbw()) 7268 #endif //_LP64 7269 bind(COMPARE_WIDE_VECTORS); 7270 vmovdqu(vec1, Address(ary1, limit, Address::times_1)); 7271 vmovdqu(vec2, Address(ary2, limit, Address::times_1)); 7272 vpxor(vec1, vec2); 7273 7274 vptest(vec1, vec1); 7275 jcc(Assembler::notZero, FALSE_LABEL); 7276 addptr(limit, 32); 7277 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 7278 7279 testl(result, result); 7280 jcc(Assembler::zero, TRUE_LABEL); 7281 7282 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32)); 7283 vmovdqu(vec2, Address(ary2, result, Address::times_1, -32)); 7284 vpxor(vec1, vec2); 7285 7286 vptest(vec1, vec1); 7287 jccb(Assembler::notZero, FALSE_LABEL); 7288 jmpb(TRUE_LABEL); 7289 7290 bind(COMPARE_TAIL); // limit is zero 7291 movl(limit, result); 7292 // Fallthru to tail compare 7293 } else if (UseSSE42Intrinsics) { 7294 // With SSE4.2, use double quad vector compare 7295 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 7296 7297 // Compare 16-byte vectors 7298 andl(result, 0x0000000f); // tail count (in bytes) 7299 andl(limit, 0xfffffff0); // vector count (in bytes) 7300 jcc(Assembler::zero, COMPARE_TAIL); 7301 7302 lea(ary1, Address(ary1, limit, Address::times_1)); 7303 lea(ary2, Address(ary2, limit, Address::times_1)); 7304 negptr(limit); 7305 7306 bind(COMPARE_WIDE_VECTORS); 7307 movdqu(vec1, Address(ary1, limit, Address::times_1)); 7308 movdqu(vec2, Address(ary2, limit, Address::times_1)); 7309 pxor(vec1, vec2); 7310 7311 ptest(vec1, vec1); 7312 jcc(Assembler::notZero, FALSE_LABEL); 7313 addptr(limit, 16); 7314 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 7315 7316 testl(result, result); 7317 jcc(Assembler::zero, TRUE_LABEL); 7318 7319 movdqu(vec1, Address(ary1, result, Address::times_1, -16)); 7320 movdqu(vec2, Address(ary2, result, Address::times_1, -16)); 7321 pxor(vec1, vec2); 7322 7323 ptest(vec1, vec1); 7324 jccb(Assembler::notZero, FALSE_LABEL); 7325 jmpb(TRUE_LABEL); 7326 7327 bind(COMPARE_TAIL); // limit is zero 7328 movl(limit, result); 7329 // Fallthru to tail compare 7330 } 7331 7332 // Compare 4-byte vectors 7333 andl(limit, 0xfffffffc); // vector count (in bytes) 7334 jccb(Assembler::zero, COMPARE_CHAR); 7335 7336 lea(ary1, Address(ary1, limit, Address::times_1)); 7337 lea(ary2, Address(ary2, limit, Address::times_1)); 7338 negptr(limit); 7339 7340 bind(COMPARE_VECTORS); 7341 movl(chr, Address(ary1, limit, Address::times_1)); 7342 cmpl(chr, Address(ary2, limit, Address::times_1)); 7343 jccb(Assembler::notEqual, FALSE_LABEL); 7344 addptr(limit, 4); 7345 jcc(Assembler::notZero, COMPARE_VECTORS); 7346 7347 // Compare trailing char (final 2 bytes), if any 7348 bind(COMPARE_CHAR); 7349 testl(result, 0x2); // tail char 7350 jccb(Assembler::zero, COMPARE_BYTE); 7351 load_unsigned_short(chr, Address(ary1, 0)); 7352 load_unsigned_short(limit, Address(ary2, 0)); 7353 cmpl(chr, limit); 7354 jccb(Assembler::notEqual, FALSE_LABEL); 7355 7356 if (is_array_equ && is_char) { 7357 bind(COMPARE_BYTE); 7358 } else { 7359 lea(ary1, Address(ary1, 2)); 7360 lea(ary2, Address(ary2, 2)); 7361 7362 bind(COMPARE_BYTE); 7363 testl(result, 0x1); // tail byte 7364 jccb(Assembler::zero, TRUE_LABEL); 7365 load_unsigned_byte(chr, Address(ary1, 0)); 7366 load_unsigned_byte(limit, Address(ary2, 0)); 7367 cmpl(chr, limit); 7368 jccb(Assembler::notEqual, FALSE_LABEL); 7369 } 7370 bind(TRUE_LABEL); 7371 movl(result, 1); // return true 7372 jmpb(DONE); 7373 7374 bind(FALSE_LABEL); 7375 xorl(result, result); // return false 7376 7377 // That's it 7378 bind(DONE); 7379 if (UseAVX >= 2) { 7380 // clean upper bits of YMM registers 7381 vpxor(vec1, vec1); 7382 vpxor(vec2, vec2); 7383 } 7384 } 7385 7386 #endif 7387 7388 void MacroAssembler::generate_fill(BasicType t, bool aligned, 7389 Register to, Register value, Register count, 7390 Register rtmp, XMMRegister xtmp) { 7391 ShortBranchVerifier sbv(this); 7392 assert_different_registers(to, value, count, rtmp); 7393 Label L_exit; 7394 Label L_fill_2_bytes, L_fill_4_bytes; 7395 7396 int shift = -1; 7397 switch (t) { 7398 case T_BYTE: 7399 shift = 2; 7400 break; 7401 case T_SHORT: 7402 shift = 1; 7403 break; 7404 case T_INT: 7405 shift = 0; 7406 break; 7407 default: ShouldNotReachHere(); 7408 } 7409 7410 if (t == T_BYTE) { 7411 andl(value, 0xff); 7412 movl(rtmp, value); 7413 shll(rtmp, 8); 7414 orl(value, rtmp); 7415 } 7416 if (t == T_SHORT) { 7417 andl(value, 0xffff); 7418 } 7419 if (t == T_BYTE || t == T_SHORT) { 7420 movl(rtmp, value); 7421 shll(rtmp, 16); 7422 orl(value, rtmp); 7423 } 7424 7425 cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element 7426 jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp 7427 if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) { 7428 Label L_skip_align2; 7429 // align source address at 4 bytes address boundary 7430 if (t == T_BYTE) { 7431 Label L_skip_align1; 7432 // One byte misalignment happens only for byte arrays 7433 testptr(to, 1); 7434 jccb(Assembler::zero, L_skip_align1); 7435 movb(Address(to, 0), value); 7436 increment(to); 7437 decrement(count); 7438 BIND(L_skip_align1); 7439 } 7440 // Two bytes misalignment happens only for byte and short (char) arrays 7441 testptr(to, 2); 7442 jccb(Assembler::zero, L_skip_align2); 7443 movw(Address(to, 0), value); 7444 addptr(to, 2); 7445 subl(count, 1<<(shift-1)); 7446 BIND(L_skip_align2); 7447 } 7448 if (UseSSE < 2) { 7449 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes; 7450 // Fill 32-byte chunks 7451 subl(count, 8 << shift); 7452 jcc(Assembler::less, L_check_fill_8_bytes); 7453 align(16); 7454 7455 BIND(L_fill_32_bytes_loop); 7456 7457 for (int i = 0; i < 32; i += 4) { 7458 movl(Address(to, i), value); 7459 } 7460 7461 addptr(to, 32); 7462 subl(count, 8 << shift); 7463 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop); 7464 BIND(L_check_fill_8_bytes); 7465 addl(count, 8 << shift); 7466 jccb(Assembler::zero, L_exit); 7467 jmpb(L_fill_8_bytes); 7468 7469 // 7470 // length is too short, just fill qwords 7471 // 7472 BIND(L_fill_8_bytes_loop); 7473 movl(Address(to, 0), value); 7474 movl(Address(to, 4), value); 7475 addptr(to, 8); 7476 BIND(L_fill_8_bytes); 7477 subl(count, 1 << (shift + 1)); 7478 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop); 7479 // fall through to fill 4 bytes 7480 } else { 7481 Label L_fill_32_bytes; 7482 if (!UseUnalignedLoadStores) { 7483 // align to 8 bytes, we know we are 4 byte aligned to start 7484 testptr(to, 4); 7485 jccb(Assembler::zero, L_fill_32_bytes); 7486 movl(Address(to, 0), value); 7487 addptr(to, 4); 7488 subl(count, 1<<shift); 7489 } 7490 BIND(L_fill_32_bytes); 7491 { 7492 assert( UseSSE >= 2, "supported cpu only" ); 7493 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes; 7494 movdl(xtmp, value); 7495 if (UseAVX >= 2 && UseUnalignedLoadStores) { 7496 Label L_check_fill_32_bytes; 7497 if (UseAVX > 2) { 7498 // Fill 64-byte chunks 7499 Label L_fill_64_bytes_loop_avx3, L_check_fill_64_bytes_avx2; 7500 7501 // If number of bytes to fill < AVX3Threshold, perform fill using AVX2 7502 cmpl(count, AVX3Threshold); 7503 jccb(Assembler::below, L_check_fill_64_bytes_avx2); 7504 7505 vpbroadcastd(xtmp, xtmp, Assembler::AVX_512bit); 7506 7507 subl(count, 16 << shift); 7508 jccb(Assembler::less, L_check_fill_32_bytes); 7509 align(16); 7510 7511 BIND(L_fill_64_bytes_loop_avx3); 7512 evmovdqul(Address(to, 0), xtmp, Assembler::AVX_512bit); 7513 addptr(to, 64); 7514 subl(count, 16 << shift); 7515 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop_avx3); 7516 jmpb(L_check_fill_32_bytes); 7517 7518 BIND(L_check_fill_64_bytes_avx2); 7519 } 7520 // Fill 64-byte chunks 7521 Label L_fill_64_bytes_loop; 7522 vpbroadcastd(xtmp, xtmp, Assembler::AVX_256bit); 7523 7524 subl(count, 16 << shift); 7525 jcc(Assembler::less, L_check_fill_32_bytes); 7526 align(16); 7527 7528 BIND(L_fill_64_bytes_loop); 7529 vmovdqu(Address(to, 0), xtmp); 7530 vmovdqu(Address(to, 32), xtmp); 7531 addptr(to, 64); 7532 subl(count, 16 << shift); 7533 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop); 7534 7535 BIND(L_check_fill_32_bytes); 7536 addl(count, 8 << shift); 7537 jccb(Assembler::less, L_check_fill_8_bytes); 7538 vmovdqu(Address(to, 0), xtmp); 7539 addptr(to, 32); 7540 subl(count, 8 << shift); 7541 7542 BIND(L_check_fill_8_bytes); 7543 // clean upper bits of YMM registers 7544 movdl(xtmp, value); 7545 pshufd(xtmp, xtmp, 0); 7546 } else { 7547 // Fill 32-byte chunks 7548 pshufd(xtmp, xtmp, 0); 7549 7550 subl(count, 8 << shift); 7551 jcc(Assembler::less, L_check_fill_8_bytes); 7552 align(16); 7553 7554 BIND(L_fill_32_bytes_loop); 7555 7556 if (UseUnalignedLoadStores) { 7557 movdqu(Address(to, 0), xtmp); 7558 movdqu(Address(to, 16), xtmp); 7559 } else { 7560 movq(Address(to, 0), xtmp); 7561 movq(Address(to, 8), xtmp); 7562 movq(Address(to, 16), xtmp); 7563 movq(Address(to, 24), xtmp); 7564 } 7565 7566 addptr(to, 32); 7567 subl(count, 8 << shift); 7568 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop); 7569 7570 BIND(L_check_fill_8_bytes); 7571 } 7572 addl(count, 8 << shift); 7573 jccb(Assembler::zero, L_exit); 7574 jmpb(L_fill_8_bytes); 7575 7576 // 7577 // length is too short, just fill qwords 7578 // 7579 BIND(L_fill_8_bytes_loop); 7580 movq(Address(to, 0), xtmp); 7581 addptr(to, 8); 7582 BIND(L_fill_8_bytes); 7583 subl(count, 1 << (shift + 1)); 7584 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop); 7585 } 7586 } 7587 // fill trailing 4 bytes 7588 BIND(L_fill_4_bytes); 7589 testl(count, 1<<shift); 7590 jccb(Assembler::zero, L_fill_2_bytes); 7591 movl(Address(to, 0), value); 7592 if (t == T_BYTE || t == T_SHORT) { 7593 Label L_fill_byte; 7594 addptr(to, 4); 7595 BIND(L_fill_2_bytes); 7596 // fill trailing 2 bytes 7597 testl(count, 1<<(shift-1)); 7598 jccb(Assembler::zero, L_fill_byte); 7599 movw(Address(to, 0), value); 7600 if (t == T_BYTE) { 7601 addptr(to, 2); 7602 BIND(L_fill_byte); 7603 // fill trailing byte 7604 testl(count, 1); 7605 jccb(Assembler::zero, L_exit); 7606 movb(Address(to, 0), value); 7607 } else { 7608 BIND(L_fill_byte); 7609 } 7610 } else { 7611 BIND(L_fill_2_bytes); 7612 } 7613 BIND(L_exit); 7614 } 7615 7616 // encode char[] to byte[] in ISO_8859_1 7617 //@HotSpotIntrinsicCandidate 7618 //private static int implEncodeISOArray(byte[] sa, int sp, 7619 //byte[] da, int dp, int len) { 7620 // int i = 0; 7621 // for (; i < len; i++) { 7622 // char c = StringUTF16.getChar(sa, sp++); 7623 // if (c > '\u00FF') 7624 // break; 7625 // da[dp++] = (byte)c; 7626 // } 7627 // return i; 7628 //} 7629 void MacroAssembler::encode_iso_array(Register src, Register dst, Register len, 7630 XMMRegister tmp1Reg, XMMRegister tmp2Reg, 7631 XMMRegister tmp3Reg, XMMRegister tmp4Reg, 7632 Register tmp5, Register result) { 7633 7634 // rsi: src 7635 // rdi: dst 7636 // rdx: len 7637 // rcx: tmp5 7638 // rax: result 7639 ShortBranchVerifier sbv(this); 7640 assert_different_registers(src, dst, len, tmp5, result); 7641 Label L_done, L_copy_1_char, L_copy_1_char_exit; 7642 7643 // set result 7644 xorl(result, result); 7645 // check for zero length 7646 testl(len, len); 7647 jcc(Assembler::zero, L_done); 7648 7649 movl(result, len); 7650 7651 // Setup pointers 7652 lea(src, Address(src, len, Address::times_2)); // char[] 7653 lea(dst, Address(dst, len, Address::times_1)); // byte[] 7654 negptr(len); 7655 7656 if (UseSSE42Intrinsics || UseAVX >= 2) { 7657 Label L_copy_8_chars, L_copy_8_chars_exit; 7658 Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit; 7659 7660 if (UseAVX >= 2) { 7661 Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit; 7662 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector 7663 movdl(tmp1Reg, tmp5); 7664 vpbroadcastd(tmp1Reg, tmp1Reg, Assembler::AVX_256bit); 7665 jmp(L_chars_32_check); 7666 7667 bind(L_copy_32_chars); 7668 vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64)); 7669 vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32)); 7670 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1); 7671 vptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 7672 jccb(Assembler::notZero, L_copy_32_chars_exit); 7673 vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1); 7674 vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector_len */ 1); 7675 vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg); 7676 7677 bind(L_chars_32_check); 7678 addptr(len, 32); 7679 jcc(Assembler::lessEqual, L_copy_32_chars); 7680 7681 bind(L_copy_32_chars_exit); 7682 subptr(len, 16); 7683 jccb(Assembler::greater, L_copy_16_chars_exit); 7684 7685 } else if (UseSSE42Intrinsics) { 7686 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector 7687 movdl(tmp1Reg, tmp5); 7688 pshufd(tmp1Reg, tmp1Reg, 0); 7689 jmpb(L_chars_16_check); 7690 } 7691 7692 bind(L_copy_16_chars); 7693 if (UseAVX >= 2) { 7694 vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32)); 7695 vptest(tmp2Reg, tmp1Reg); 7696 jcc(Assembler::notZero, L_copy_16_chars_exit); 7697 vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector_len */ 1); 7698 vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector_len */ 1); 7699 } else { 7700 if (UseAVX > 0) { 7701 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32)); 7702 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16)); 7703 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 0); 7704 } else { 7705 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32)); 7706 por(tmp2Reg, tmp3Reg); 7707 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16)); 7708 por(tmp2Reg, tmp4Reg); 7709 } 7710 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 7711 jccb(Assembler::notZero, L_copy_16_chars_exit); 7712 packuswb(tmp3Reg, tmp4Reg); 7713 } 7714 movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg); 7715 7716 bind(L_chars_16_check); 7717 addptr(len, 16); 7718 jcc(Assembler::lessEqual, L_copy_16_chars); 7719 7720 bind(L_copy_16_chars_exit); 7721 if (UseAVX >= 2) { 7722 // clean upper bits of YMM registers 7723 vpxor(tmp2Reg, tmp2Reg); 7724 vpxor(tmp3Reg, tmp3Reg); 7725 vpxor(tmp4Reg, tmp4Reg); 7726 movdl(tmp1Reg, tmp5); 7727 pshufd(tmp1Reg, tmp1Reg, 0); 7728 } 7729 subptr(len, 8); 7730 jccb(Assembler::greater, L_copy_8_chars_exit); 7731 7732 bind(L_copy_8_chars); 7733 movdqu(tmp3Reg, Address(src, len, Address::times_2, -16)); 7734 ptest(tmp3Reg, tmp1Reg); 7735 jccb(Assembler::notZero, L_copy_8_chars_exit); 7736 packuswb(tmp3Reg, tmp1Reg); 7737 movq(Address(dst, len, Address::times_1, -8), tmp3Reg); 7738 addptr(len, 8); 7739 jccb(Assembler::lessEqual, L_copy_8_chars); 7740 7741 bind(L_copy_8_chars_exit); 7742 subptr(len, 8); 7743 jccb(Assembler::zero, L_done); 7744 } 7745 7746 bind(L_copy_1_char); 7747 load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0)); 7748 testl(tmp5, 0xff00); // check if Unicode char 7749 jccb(Assembler::notZero, L_copy_1_char_exit); 7750 movb(Address(dst, len, Address::times_1, 0), tmp5); 7751 addptr(len, 1); 7752 jccb(Assembler::less, L_copy_1_char); 7753 7754 bind(L_copy_1_char_exit); 7755 addptr(result, len); // len is negative count of not processed elements 7756 7757 bind(L_done); 7758 } 7759 7760 #ifdef _LP64 7761 /** 7762 * Helper for multiply_to_len(). 7763 */ 7764 void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo, Register src1, Register src2) { 7765 addq(dest_lo, src1); 7766 adcq(dest_hi, 0); 7767 addq(dest_lo, src2); 7768 adcq(dest_hi, 0); 7769 } 7770 7771 /** 7772 * Multiply 64 bit by 64 bit first loop. 7773 */ 7774 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart, 7775 Register y, Register y_idx, Register z, 7776 Register carry, Register product, 7777 Register idx, Register kdx) { 7778 // 7779 // jlong carry, x[], y[], z[]; 7780 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) { 7781 // huge_128 product = y[idx] * x[xstart] + carry; 7782 // z[kdx] = (jlong)product; 7783 // carry = (jlong)(product >>> 64); 7784 // } 7785 // z[xstart] = carry; 7786 // 7787 7788 Label L_first_loop, L_first_loop_exit; 7789 Label L_one_x, L_one_y, L_multiply; 7790 7791 decrementl(xstart); 7792 jcc(Assembler::negative, L_one_x); 7793 7794 movq(x_xstart, Address(x, xstart, Address::times_4, 0)); 7795 rorq(x_xstart, 32); // convert big-endian to little-endian 7796 7797 bind(L_first_loop); 7798 decrementl(idx); 7799 jcc(Assembler::negative, L_first_loop_exit); 7800 decrementl(idx); 7801 jcc(Assembler::negative, L_one_y); 7802 movq(y_idx, Address(y, idx, Address::times_4, 0)); 7803 rorq(y_idx, 32); // convert big-endian to little-endian 7804 bind(L_multiply); 7805 movq(product, x_xstart); 7806 mulq(y_idx); // product(rax) * y_idx -> rdx:rax 7807 addq(product, carry); 7808 adcq(rdx, 0); 7809 subl(kdx, 2); 7810 movl(Address(z, kdx, Address::times_4, 4), product); 7811 shrq(product, 32); 7812 movl(Address(z, kdx, Address::times_4, 0), product); 7813 movq(carry, rdx); 7814 jmp(L_first_loop); 7815 7816 bind(L_one_y); 7817 movl(y_idx, Address(y, 0)); 7818 jmp(L_multiply); 7819 7820 bind(L_one_x); 7821 movl(x_xstart, Address(x, 0)); 7822 jmp(L_first_loop); 7823 7824 bind(L_first_loop_exit); 7825 } 7826 7827 /** 7828 * Multiply 64 bit by 64 bit and add 128 bit. 7829 */ 7830 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y, Register z, 7831 Register yz_idx, Register idx, 7832 Register carry, Register product, int offset) { 7833 // huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry; 7834 // z[kdx] = (jlong)product; 7835 7836 movq(yz_idx, Address(y, idx, Address::times_4, offset)); 7837 rorq(yz_idx, 32); // convert big-endian to little-endian 7838 movq(product, x_xstart); 7839 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax) 7840 movq(yz_idx, Address(z, idx, Address::times_4, offset)); 7841 rorq(yz_idx, 32); // convert big-endian to little-endian 7842 7843 add2_with_carry(rdx, product, carry, yz_idx); 7844 7845 movl(Address(z, idx, Address::times_4, offset+4), product); 7846 shrq(product, 32); 7847 movl(Address(z, idx, Address::times_4, offset), product); 7848 7849 } 7850 7851 /** 7852 * Multiply 128 bit by 128 bit. Unrolled inner loop. 7853 */ 7854 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart, Register y, Register z, 7855 Register yz_idx, Register idx, Register jdx, 7856 Register carry, Register product, 7857 Register carry2) { 7858 // jlong carry, x[], y[], z[]; 7859 // int kdx = ystart+1; 7860 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop 7861 // huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry; 7862 // z[kdx+idx+1] = (jlong)product; 7863 // jlong carry2 = (jlong)(product >>> 64); 7864 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry2; 7865 // z[kdx+idx] = (jlong)product; 7866 // carry = (jlong)(product >>> 64); 7867 // } 7868 // idx += 2; 7869 // if (idx > 0) { 7870 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry; 7871 // z[kdx+idx] = (jlong)product; 7872 // carry = (jlong)(product >>> 64); 7873 // } 7874 // 7875 7876 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done; 7877 7878 movl(jdx, idx); 7879 andl(jdx, 0xFFFFFFFC); 7880 shrl(jdx, 2); 7881 7882 bind(L_third_loop); 7883 subl(jdx, 1); 7884 jcc(Assembler::negative, L_third_loop_exit); 7885 subl(idx, 4); 7886 7887 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8); 7888 movq(carry2, rdx); 7889 7890 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0); 7891 movq(carry, rdx); 7892 jmp(L_third_loop); 7893 7894 bind (L_third_loop_exit); 7895 7896 andl (idx, 0x3); 7897 jcc(Assembler::zero, L_post_third_loop_done); 7898 7899 Label L_check_1; 7900 subl(idx, 2); 7901 jcc(Assembler::negative, L_check_1); 7902 7903 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0); 7904 movq(carry, rdx); 7905 7906 bind (L_check_1); 7907 addl (idx, 0x2); 7908 andl (idx, 0x1); 7909 subl(idx, 1); 7910 jcc(Assembler::negative, L_post_third_loop_done); 7911 7912 movl(yz_idx, Address(y, idx, Address::times_4, 0)); 7913 movq(product, x_xstart); 7914 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax) 7915 movl(yz_idx, Address(z, idx, Address::times_4, 0)); 7916 7917 add2_with_carry(rdx, product, yz_idx, carry); 7918 7919 movl(Address(z, idx, Address::times_4, 0), product); 7920 shrq(product, 32); 7921 7922 shlq(rdx, 32); 7923 orq(product, rdx); 7924 movq(carry, product); 7925 7926 bind(L_post_third_loop_done); 7927 } 7928 7929 /** 7930 * Multiply 128 bit by 128 bit using BMI2. Unrolled inner loop. 7931 * 7932 */ 7933 void MacroAssembler::multiply_128_x_128_bmi2_loop(Register y, Register z, 7934 Register carry, Register carry2, 7935 Register idx, Register jdx, 7936 Register yz_idx1, Register yz_idx2, 7937 Register tmp, Register tmp3, Register tmp4) { 7938 assert(UseBMI2Instructions, "should be used only when BMI2 is available"); 7939 7940 // jlong carry, x[], y[], z[]; 7941 // int kdx = ystart+1; 7942 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop 7943 // huge_128 tmp3 = (y[idx+1] * rdx) + z[kdx+idx+1] + carry; 7944 // jlong carry2 = (jlong)(tmp3 >>> 64); 7945 // huge_128 tmp4 = (y[idx] * rdx) + z[kdx+idx] + carry2; 7946 // carry = (jlong)(tmp4 >>> 64); 7947 // z[kdx+idx+1] = (jlong)tmp3; 7948 // z[kdx+idx] = (jlong)tmp4; 7949 // } 7950 // idx += 2; 7951 // if (idx > 0) { 7952 // yz_idx1 = (y[idx] * rdx) + z[kdx+idx] + carry; 7953 // z[kdx+idx] = (jlong)yz_idx1; 7954 // carry = (jlong)(yz_idx1 >>> 64); 7955 // } 7956 // 7957 7958 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done; 7959 7960 movl(jdx, idx); 7961 andl(jdx, 0xFFFFFFFC); 7962 shrl(jdx, 2); 7963 7964 bind(L_third_loop); 7965 subl(jdx, 1); 7966 jcc(Assembler::negative, L_third_loop_exit); 7967 subl(idx, 4); 7968 7969 movq(yz_idx1, Address(y, idx, Address::times_4, 8)); 7970 rorxq(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian 7971 movq(yz_idx2, Address(y, idx, Address::times_4, 0)); 7972 rorxq(yz_idx2, yz_idx2, 32); 7973 7974 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3 7975 mulxq(carry2, tmp, yz_idx2); // yz_idx2 * rdx -> carry2:tmp 7976 7977 movq(yz_idx1, Address(z, idx, Address::times_4, 8)); 7978 rorxq(yz_idx1, yz_idx1, 32); 7979 movq(yz_idx2, Address(z, idx, Address::times_4, 0)); 7980 rorxq(yz_idx2, yz_idx2, 32); 7981 7982 if (VM_Version::supports_adx()) { 7983 adcxq(tmp3, carry); 7984 adoxq(tmp3, yz_idx1); 7985 7986 adcxq(tmp4, tmp); 7987 adoxq(tmp4, yz_idx2); 7988 7989 movl(carry, 0); // does not affect flags 7990 adcxq(carry2, carry); 7991 adoxq(carry2, carry); 7992 } else { 7993 add2_with_carry(tmp4, tmp3, carry, yz_idx1); 7994 add2_with_carry(carry2, tmp4, tmp, yz_idx2); 7995 } 7996 movq(carry, carry2); 7997 7998 movl(Address(z, idx, Address::times_4, 12), tmp3); 7999 shrq(tmp3, 32); 8000 movl(Address(z, idx, Address::times_4, 8), tmp3); 8001 8002 movl(Address(z, idx, Address::times_4, 4), tmp4); 8003 shrq(tmp4, 32); 8004 movl(Address(z, idx, Address::times_4, 0), tmp4); 8005 8006 jmp(L_third_loop); 8007 8008 bind (L_third_loop_exit); 8009 8010 andl (idx, 0x3); 8011 jcc(Assembler::zero, L_post_third_loop_done); 8012 8013 Label L_check_1; 8014 subl(idx, 2); 8015 jcc(Assembler::negative, L_check_1); 8016 8017 movq(yz_idx1, Address(y, idx, Address::times_4, 0)); 8018 rorxq(yz_idx1, yz_idx1, 32); 8019 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3 8020 movq(yz_idx2, Address(z, idx, Address::times_4, 0)); 8021 rorxq(yz_idx2, yz_idx2, 32); 8022 8023 add2_with_carry(tmp4, tmp3, carry, yz_idx2); 8024 8025 movl(Address(z, idx, Address::times_4, 4), tmp3); 8026 shrq(tmp3, 32); 8027 movl(Address(z, idx, Address::times_4, 0), tmp3); 8028 movq(carry, tmp4); 8029 8030 bind (L_check_1); 8031 addl (idx, 0x2); 8032 andl (idx, 0x1); 8033 subl(idx, 1); 8034 jcc(Assembler::negative, L_post_third_loop_done); 8035 movl(tmp4, Address(y, idx, Address::times_4, 0)); 8036 mulxq(carry2, tmp3, tmp4); // tmp4 * rdx -> carry2:tmp3 8037 movl(tmp4, Address(z, idx, Address::times_4, 0)); 8038 8039 add2_with_carry(carry2, tmp3, tmp4, carry); 8040 8041 movl(Address(z, idx, Address::times_4, 0), tmp3); 8042 shrq(tmp3, 32); 8043 8044 shlq(carry2, 32); 8045 orq(tmp3, carry2); 8046 movq(carry, tmp3); 8047 8048 bind(L_post_third_loop_done); 8049 } 8050 8051 /** 8052 * Code for BigInteger::multiplyToLen() instrinsic. 8053 * 8054 * rdi: x 8055 * rax: xlen 8056 * rsi: y 8057 * rcx: ylen 8058 * r8: z 8059 * r11: zlen 8060 * r12: tmp1 8061 * r13: tmp2 8062 * r14: tmp3 8063 * r15: tmp4 8064 * rbx: tmp5 8065 * 8066 */ 8067 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, Register z, Register zlen, 8068 Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5) { 8069 ShortBranchVerifier sbv(this); 8070 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx); 8071 8072 push(tmp1); 8073 push(tmp2); 8074 push(tmp3); 8075 push(tmp4); 8076 push(tmp5); 8077 8078 push(xlen); 8079 push(zlen); 8080 8081 const Register idx = tmp1; 8082 const Register kdx = tmp2; 8083 const Register xstart = tmp3; 8084 8085 const Register y_idx = tmp4; 8086 const Register carry = tmp5; 8087 const Register product = xlen; 8088 const Register x_xstart = zlen; // reuse register 8089 8090 // First Loop. 8091 // 8092 // final static long LONG_MASK = 0xffffffffL; 8093 // int xstart = xlen - 1; 8094 // int ystart = ylen - 1; 8095 // long carry = 0; 8096 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) { 8097 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry; 8098 // z[kdx] = (int)product; 8099 // carry = product >>> 32; 8100 // } 8101 // z[xstart] = (int)carry; 8102 // 8103 8104 movl(idx, ylen); // idx = ylen; 8105 movl(kdx, zlen); // kdx = xlen+ylen; 8106 xorq(carry, carry); // carry = 0; 8107 8108 Label L_done; 8109 8110 movl(xstart, xlen); 8111 decrementl(xstart); 8112 jcc(Assembler::negative, L_done); 8113 8114 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx); 8115 8116 Label L_second_loop; 8117 testl(kdx, kdx); 8118 jcc(Assembler::zero, L_second_loop); 8119 8120 Label L_carry; 8121 subl(kdx, 1); 8122 jcc(Assembler::zero, L_carry); 8123 8124 movl(Address(z, kdx, Address::times_4, 0), carry); 8125 shrq(carry, 32); 8126 subl(kdx, 1); 8127 8128 bind(L_carry); 8129 movl(Address(z, kdx, Address::times_4, 0), carry); 8130 8131 // Second and third (nested) loops. 8132 // 8133 // for (int i = xstart-1; i >= 0; i--) { // Second loop 8134 // carry = 0; 8135 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop 8136 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) + 8137 // (z[k] & LONG_MASK) + carry; 8138 // z[k] = (int)product; 8139 // carry = product >>> 32; 8140 // } 8141 // z[i] = (int)carry; 8142 // } 8143 // 8144 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx 8145 8146 const Register jdx = tmp1; 8147 8148 bind(L_second_loop); 8149 xorl(carry, carry); // carry = 0; 8150 movl(jdx, ylen); // j = ystart+1 8151 8152 subl(xstart, 1); // i = xstart-1; 8153 jcc(Assembler::negative, L_done); 8154 8155 push (z); 8156 8157 Label L_last_x; 8158 lea(z, Address(z, xstart, Address::times_4, 4)); // z = z + k - j 8159 subl(xstart, 1); // i = xstart-1; 8160 jcc(Assembler::negative, L_last_x); 8161 8162 if (UseBMI2Instructions) { 8163 movq(rdx, Address(x, xstart, Address::times_4, 0)); 8164 rorxq(rdx, rdx, 32); // convert big-endian to little-endian 8165 } else { 8166 movq(x_xstart, Address(x, xstart, Address::times_4, 0)); 8167 rorq(x_xstart, 32); // convert big-endian to little-endian 8168 } 8169 8170 Label L_third_loop_prologue; 8171 bind(L_third_loop_prologue); 8172 8173 push (x); 8174 push (xstart); 8175 push (ylen); 8176 8177 8178 if (UseBMI2Instructions) { 8179 multiply_128_x_128_bmi2_loop(y, z, carry, x, jdx, ylen, product, tmp2, x_xstart, tmp3, tmp4); 8180 } else { // !UseBMI2Instructions 8181 multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x); 8182 } 8183 8184 pop(ylen); 8185 pop(xlen); 8186 pop(x); 8187 pop(z); 8188 8189 movl(tmp3, xlen); 8190 addl(tmp3, 1); 8191 movl(Address(z, tmp3, Address::times_4, 0), carry); 8192 subl(tmp3, 1); 8193 jccb(Assembler::negative, L_done); 8194 8195 shrq(carry, 32); 8196 movl(Address(z, tmp3, Address::times_4, 0), carry); 8197 jmp(L_second_loop); 8198 8199 // Next infrequent code is moved outside loops. 8200 bind(L_last_x); 8201 if (UseBMI2Instructions) { 8202 movl(rdx, Address(x, 0)); 8203 } else { 8204 movl(x_xstart, Address(x, 0)); 8205 } 8206 jmp(L_third_loop_prologue); 8207 8208 bind(L_done); 8209 8210 pop(zlen); 8211 pop(xlen); 8212 8213 pop(tmp5); 8214 pop(tmp4); 8215 pop(tmp3); 8216 pop(tmp2); 8217 pop(tmp1); 8218 } 8219 8220 void MacroAssembler::vectorized_mismatch(Register obja, Register objb, Register length, Register log2_array_indxscale, 8221 Register result, Register tmp1, Register tmp2, XMMRegister rymm0, XMMRegister rymm1, XMMRegister rymm2){ 8222 assert(UseSSE42Intrinsics, "SSE4.2 must be enabled."); 8223 Label VECTOR16_LOOP, VECTOR8_LOOP, VECTOR4_LOOP; 8224 Label VECTOR8_TAIL, VECTOR4_TAIL; 8225 Label VECTOR32_NOT_EQUAL, VECTOR16_NOT_EQUAL, VECTOR8_NOT_EQUAL, VECTOR4_NOT_EQUAL; 8226 Label SAME_TILL_END, DONE; 8227 Label BYTES_LOOP, BYTES_TAIL, BYTES_NOT_EQUAL; 8228 8229 //scale is in rcx in both Win64 and Unix 8230 ShortBranchVerifier sbv(this); 8231 8232 shlq(length); 8233 xorq(result, result); 8234 8235 if ((AVX3Threshold == 0) && (UseAVX > 2) && 8236 VM_Version::supports_avx512vlbw()) { 8237 Label VECTOR64_LOOP, VECTOR64_NOT_EQUAL, VECTOR32_TAIL; 8238 8239 cmpq(length, 64); 8240 jcc(Assembler::less, VECTOR32_TAIL); 8241 8242 movq(tmp1, length); 8243 andq(tmp1, 0x3F); // tail count 8244 andq(length, ~(0x3F)); //vector count 8245 8246 bind(VECTOR64_LOOP); 8247 // AVX512 code to compare 64 byte vectors. 8248 evmovdqub(rymm0, Address(obja, result), Assembler::AVX_512bit); 8249 evpcmpeqb(k7, rymm0, Address(objb, result), Assembler::AVX_512bit); 8250 kortestql(k7, k7); 8251 jcc(Assembler::aboveEqual, VECTOR64_NOT_EQUAL); // mismatch 8252 addq(result, 64); 8253 subq(length, 64); 8254 jccb(Assembler::notZero, VECTOR64_LOOP); 8255 8256 //bind(VECTOR64_TAIL); 8257 testq(tmp1, tmp1); 8258 jcc(Assembler::zero, SAME_TILL_END); 8259 8260 //bind(VECTOR64_TAIL); 8261 // AVX512 code to compare upto 63 byte vectors. 8262 mov64(tmp2, 0xFFFFFFFFFFFFFFFF); 8263 shlxq(tmp2, tmp2, tmp1); 8264 notq(tmp2); 8265 kmovql(k3, tmp2); 8266 8267 evmovdqub(rymm0, k3, Address(obja, result), Assembler::AVX_512bit); 8268 evpcmpeqb(k7, k3, rymm0, Address(objb, result), Assembler::AVX_512bit); 8269 8270 ktestql(k7, k3); 8271 jcc(Assembler::below, SAME_TILL_END); // not mismatch 8272 8273 bind(VECTOR64_NOT_EQUAL); 8274 kmovql(tmp1, k7); 8275 notq(tmp1); 8276 tzcntq(tmp1, tmp1); 8277 addq(result, tmp1); 8278 shrq(result); 8279 jmp(DONE); 8280 bind(VECTOR32_TAIL); 8281 } 8282 8283 cmpq(length, 8); 8284 jcc(Assembler::equal, VECTOR8_LOOP); 8285 jcc(Assembler::less, VECTOR4_TAIL); 8286 8287 if (UseAVX >= 2) { 8288 Label VECTOR16_TAIL, VECTOR32_LOOP; 8289 8290 cmpq(length, 16); 8291 jcc(Assembler::equal, VECTOR16_LOOP); 8292 jcc(Assembler::less, VECTOR8_LOOP); 8293 8294 cmpq(length, 32); 8295 jccb(Assembler::less, VECTOR16_TAIL); 8296 8297 subq(length, 32); 8298 bind(VECTOR32_LOOP); 8299 vmovdqu(rymm0, Address(obja, result)); 8300 vmovdqu(rymm1, Address(objb, result)); 8301 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_256bit); 8302 vptest(rymm2, rymm2); 8303 jcc(Assembler::notZero, VECTOR32_NOT_EQUAL);//mismatch found 8304 addq(result, 32); 8305 subq(length, 32); 8306 jcc(Assembler::greaterEqual, VECTOR32_LOOP); 8307 addq(length, 32); 8308 jcc(Assembler::equal, SAME_TILL_END); 8309 //falling through if less than 32 bytes left //close the branch here. 8310 8311 bind(VECTOR16_TAIL); 8312 cmpq(length, 16); 8313 jccb(Assembler::less, VECTOR8_TAIL); 8314 bind(VECTOR16_LOOP); 8315 movdqu(rymm0, Address(obja, result)); 8316 movdqu(rymm1, Address(objb, result)); 8317 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_128bit); 8318 ptest(rymm2, rymm2); 8319 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found 8320 addq(result, 16); 8321 subq(length, 16); 8322 jcc(Assembler::equal, SAME_TILL_END); 8323 //falling through if less than 16 bytes left 8324 } else {//regular intrinsics 8325 8326 cmpq(length, 16); 8327 jccb(Assembler::less, VECTOR8_TAIL); 8328 8329 subq(length, 16); 8330 bind(VECTOR16_LOOP); 8331 movdqu(rymm0, Address(obja, result)); 8332 movdqu(rymm1, Address(objb, result)); 8333 pxor(rymm0, rymm1); 8334 ptest(rymm0, rymm0); 8335 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found 8336 addq(result, 16); 8337 subq(length, 16); 8338 jccb(Assembler::greaterEqual, VECTOR16_LOOP); 8339 addq(length, 16); 8340 jcc(Assembler::equal, SAME_TILL_END); 8341 //falling through if less than 16 bytes left 8342 } 8343 8344 bind(VECTOR8_TAIL); 8345 cmpq(length, 8); 8346 jccb(Assembler::less, VECTOR4_TAIL); 8347 bind(VECTOR8_LOOP); 8348 movq(tmp1, Address(obja, result)); 8349 movq(tmp2, Address(objb, result)); 8350 xorq(tmp1, tmp2); 8351 testq(tmp1, tmp1); 8352 jcc(Assembler::notZero, VECTOR8_NOT_EQUAL);//mismatch found 8353 addq(result, 8); 8354 subq(length, 8); 8355 jcc(Assembler::equal, SAME_TILL_END); 8356 //falling through if less than 8 bytes left 8357 8358 bind(VECTOR4_TAIL); 8359 cmpq(length, 4); 8360 jccb(Assembler::less, BYTES_TAIL); 8361 bind(VECTOR4_LOOP); 8362 movl(tmp1, Address(obja, result)); 8363 xorl(tmp1, Address(objb, result)); 8364 testl(tmp1, tmp1); 8365 jcc(Assembler::notZero, VECTOR4_NOT_EQUAL);//mismatch found 8366 addq(result, 4); 8367 subq(length, 4); 8368 jcc(Assembler::equal, SAME_TILL_END); 8369 //falling through if less than 4 bytes left 8370 8371 bind(BYTES_TAIL); 8372 bind(BYTES_LOOP); 8373 load_unsigned_byte(tmp1, Address(obja, result)); 8374 load_unsigned_byte(tmp2, Address(objb, result)); 8375 xorl(tmp1, tmp2); 8376 testl(tmp1, tmp1); 8377 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 8378 decq(length); 8379 jcc(Assembler::zero, SAME_TILL_END); 8380 incq(result); 8381 load_unsigned_byte(tmp1, Address(obja, result)); 8382 load_unsigned_byte(tmp2, Address(objb, result)); 8383 xorl(tmp1, tmp2); 8384 testl(tmp1, tmp1); 8385 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 8386 decq(length); 8387 jcc(Assembler::zero, SAME_TILL_END); 8388 incq(result); 8389 load_unsigned_byte(tmp1, Address(obja, result)); 8390 load_unsigned_byte(tmp2, Address(objb, result)); 8391 xorl(tmp1, tmp2); 8392 testl(tmp1, tmp1); 8393 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 8394 jmp(SAME_TILL_END); 8395 8396 if (UseAVX >= 2) { 8397 bind(VECTOR32_NOT_EQUAL); 8398 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_256bit); 8399 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_256bit); 8400 vpxor(rymm0, rymm0, rymm2, Assembler::AVX_256bit); 8401 vpmovmskb(tmp1, rymm0); 8402 bsfq(tmp1, tmp1); 8403 addq(result, tmp1); 8404 shrq(result); 8405 jmp(DONE); 8406 } 8407 8408 bind(VECTOR16_NOT_EQUAL); 8409 if (UseAVX >= 2) { 8410 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_128bit); 8411 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_128bit); 8412 pxor(rymm0, rymm2); 8413 } else { 8414 pcmpeqb(rymm2, rymm2); 8415 pxor(rymm0, rymm1); 8416 pcmpeqb(rymm0, rymm1); 8417 pxor(rymm0, rymm2); 8418 } 8419 pmovmskb(tmp1, rymm0); 8420 bsfq(tmp1, tmp1); 8421 addq(result, tmp1); 8422 shrq(result); 8423 jmpb(DONE); 8424 8425 bind(VECTOR8_NOT_EQUAL); 8426 bind(VECTOR4_NOT_EQUAL); 8427 bsfq(tmp1, tmp1); 8428 shrq(tmp1, 3); 8429 addq(result, tmp1); 8430 bind(BYTES_NOT_EQUAL); 8431 shrq(result); 8432 jmpb(DONE); 8433 8434 bind(SAME_TILL_END); 8435 mov64(result, -1); 8436 8437 bind(DONE); 8438 } 8439 8440 //Helper functions for square_to_len() 8441 8442 /** 8443 * Store the squares of x[], right shifted one bit (divided by 2) into z[] 8444 * Preserves x and z and modifies rest of the registers. 8445 */ 8446 void MacroAssembler::square_rshift(Register x, Register xlen, Register z, Register tmp1, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 8447 // Perform square and right shift by 1 8448 // Handle odd xlen case first, then for even xlen do the following 8449 // jlong carry = 0; 8450 // for (int j=0, i=0; j < xlen; j+=2, i+=4) { 8451 // huge_128 product = x[j:j+1] * x[j:j+1]; 8452 // z[i:i+1] = (carry << 63) | (jlong)(product >>> 65); 8453 // z[i+2:i+3] = (jlong)(product >>> 1); 8454 // carry = (jlong)product; 8455 // } 8456 8457 xorq(tmp5, tmp5); // carry 8458 xorq(rdxReg, rdxReg); 8459 xorl(tmp1, tmp1); // index for x 8460 xorl(tmp4, tmp4); // index for z 8461 8462 Label L_first_loop, L_first_loop_exit; 8463 8464 testl(xlen, 1); 8465 jccb(Assembler::zero, L_first_loop); //jump if xlen is even 8466 8467 // Square and right shift by 1 the odd element using 32 bit multiply 8468 movl(raxReg, Address(x, tmp1, Address::times_4, 0)); 8469 imulq(raxReg, raxReg); 8470 shrq(raxReg, 1); 8471 adcq(tmp5, 0); 8472 movq(Address(z, tmp4, Address::times_4, 0), raxReg); 8473 incrementl(tmp1); 8474 addl(tmp4, 2); 8475 8476 // Square and right shift by 1 the rest using 64 bit multiply 8477 bind(L_first_loop); 8478 cmpptr(tmp1, xlen); 8479 jccb(Assembler::equal, L_first_loop_exit); 8480 8481 // Square 8482 movq(raxReg, Address(x, tmp1, Address::times_4, 0)); 8483 rorq(raxReg, 32); // convert big-endian to little-endian 8484 mulq(raxReg); // 64-bit multiply rax * rax -> rdx:rax 8485 8486 // Right shift by 1 and save carry 8487 shrq(tmp5, 1); // rdx:rax:tmp5 = (tmp5:rdx:rax) >>> 1 8488 rcrq(rdxReg, 1); 8489 rcrq(raxReg, 1); 8490 adcq(tmp5, 0); 8491 8492 // Store result in z 8493 movq(Address(z, tmp4, Address::times_4, 0), rdxReg); 8494 movq(Address(z, tmp4, Address::times_4, 8), raxReg); 8495 8496 // Update indices for x and z 8497 addl(tmp1, 2); 8498 addl(tmp4, 4); 8499 jmp(L_first_loop); 8500 8501 bind(L_first_loop_exit); 8502 } 8503 8504 8505 /** 8506 * Perform the following multiply add operation using BMI2 instructions 8507 * carry:sum = sum + op1*op2 + carry 8508 * op2 should be in rdx 8509 * op2 is preserved, all other registers are modified 8510 */ 8511 void MacroAssembler::multiply_add_64_bmi2(Register sum, Register op1, Register op2, Register carry, Register tmp2) { 8512 // assert op2 is rdx 8513 mulxq(tmp2, op1, op1); // op1 * op2 -> tmp2:op1 8514 addq(sum, carry); 8515 adcq(tmp2, 0); 8516 addq(sum, op1); 8517 adcq(tmp2, 0); 8518 movq(carry, tmp2); 8519 } 8520 8521 /** 8522 * Perform the following multiply add operation: 8523 * carry:sum = sum + op1*op2 + carry 8524 * Preserves op1, op2 and modifies rest of registers 8525 */ 8526 void MacroAssembler::multiply_add_64(Register sum, Register op1, Register op2, Register carry, Register rdxReg, Register raxReg) { 8527 // rdx:rax = op1 * op2 8528 movq(raxReg, op2); 8529 mulq(op1); 8530 8531 // rdx:rax = sum + carry + rdx:rax 8532 addq(sum, carry); 8533 adcq(rdxReg, 0); 8534 addq(sum, raxReg); 8535 adcq(rdxReg, 0); 8536 8537 // carry:sum = rdx:sum 8538 movq(carry, rdxReg); 8539 } 8540 8541 /** 8542 * Add 64 bit long carry into z[] with carry propogation. 8543 * Preserves z and carry register values and modifies rest of registers. 8544 * 8545 */ 8546 void MacroAssembler::add_one_64(Register z, Register zlen, Register carry, Register tmp1) { 8547 Label L_fourth_loop, L_fourth_loop_exit; 8548 8549 movl(tmp1, 1); 8550 subl(zlen, 2); 8551 addq(Address(z, zlen, Address::times_4, 0), carry); 8552 8553 bind(L_fourth_loop); 8554 jccb(Assembler::carryClear, L_fourth_loop_exit); 8555 subl(zlen, 2); 8556 jccb(Assembler::negative, L_fourth_loop_exit); 8557 addq(Address(z, zlen, Address::times_4, 0), tmp1); 8558 jmp(L_fourth_loop); 8559 bind(L_fourth_loop_exit); 8560 } 8561 8562 /** 8563 * Shift z[] left by 1 bit. 8564 * Preserves x, len, z and zlen registers and modifies rest of the registers. 8565 * 8566 */ 8567 void MacroAssembler::lshift_by_1(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4) { 8568 8569 Label L_fifth_loop, L_fifth_loop_exit; 8570 8571 // Fifth loop 8572 // Perform primitiveLeftShift(z, zlen, 1) 8573 8574 const Register prev_carry = tmp1; 8575 const Register new_carry = tmp4; 8576 const Register value = tmp2; 8577 const Register zidx = tmp3; 8578 8579 // int zidx, carry; 8580 // long value; 8581 // carry = 0; 8582 // for (zidx = zlen-2; zidx >=0; zidx -= 2) { 8583 // (carry:value) = (z[i] << 1) | carry ; 8584 // z[i] = value; 8585 // } 8586 8587 movl(zidx, zlen); 8588 xorl(prev_carry, prev_carry); // clear carry flag and prev_carry register 8589 8590 bind(L_fifth_loop); 8591 decl(zidx); // Use decl to preserve carry flag 8592 decl(zidx); 8593 jccb(Assembler::negative, L_fifth_loop_exit); 8594 8595 if (UseBMI2Instructions) { 8596 movq(value, Address(z, zidx, Address::times_4, 0)); 8597 rclq(value, 1); 8598 rorxq(value, value, 32); 8599 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form 8600 } 8601 else { 8602 // clear new_carry 8603 xorl(new_carry, new_carry); 8604 8605 // Shift z[i] by 1, or in previous carry and save new carry 8606 movq(value, Address(z, zidx, Address::times_4, 0)); 8607 shlq(value, 1); 8608 adcl(new_carry, 0); 8609 8610 orq(value, prev_carry); 8611 rorq(value, 0x20); 8612 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form 8613 8614 // Set previous carry = new carry 8615 movl(prev_carry, new_carry); 8616 } 8617 jmp(L_fifth_loop); 8618 8619 bind(L_fifth_loop_exit); 8620 } 8621 8622 8623 /** 8624 * Code for BigInteger::squareToLen() intrinsic 8625 * 8626 * rdi: x 8627 * rsi: len 8628 * r8: z 8629 * rcx: zlen 8630 * r12: tmp1 8631 * r13: tmp2 8632 * r14: tmp3 8633 * r15: tmp4 8634 * rbx: tmp5 8635 * 8636 */ 8637 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) { 8638 8639 Label L_second_loop, L_second_loop_exit, L_third_loop, L_third_loop_exit, L_last_x, L_multiply; 8640 push(tmp1); 8641 push(tmp2); 8642 push(tmp3); 8643 push(tmp4); 8644 push(tmp5); 8645 8646 // First loop 8647 // Store the squares, right shifted one bit (i.e., divided by 2). 8648 square_rshift(x, len, z, tmp1, tmp3, tmp4, tmp5, rdxReg, raxReg); 8649 8650 // Add in off-diagonal sums. 8651 // 8652 // Second, third (nested) and fourth loops. 8653 // zlen +=2; 8654 // for (int xidx=len-2,zidx=zlen-4; xidx > 0; xidx-=2,zidx-=4) { 8655 // carry = 0; 8656 // long op2 = x[xidx:xidx+1]; 8657 // for (int j=xidx-2,k=zidx; j >= 0; j-=2) { 8658 // k -= 2; 8659 // long op1 = x[j:j+1]; 8660 // long sum = z[k:k+1]; 8661 // carry:sum = multiply_add_64(sum, op1, op2, carry, tmp_regs); 8662 // z[k:k+1] = sum; 8663 // } 8664 // add_one_64(z, k, carry, tmp_regs); 8665 // } 8666 8667 const Register carry = tmp5; 8668 const Register sum = tmp3; 8669 const Register op1 = tmp4; 8670 Register op2 = tmp2; 8671 8672 push(zlen); 8673 push(len); 8674 addl(zlen,2); 8675 bind(L_second_loop); 8676 xorq(carry, carry); 8677 subl(zlen, 4); 8678 subl(len, 2); 8679 push(zlen); 8680 push(len); 8681 cmpl(len, 0); 8682 jccb(Assembler::lessEqual, L_second_loop_exit); 8683 8684 // Multiply an array by one 64 bit long. 8685 if (UseBMI2Instructions) { 8686 op2 = rdxReg; 8687 movq(op2, Address(x, len, Address::times_4, 0)); 8688 rorxq(op2, op2, 32); 8689 } 8690 else { 8691 movq(op2, Address(x, len, Address::times_4, 0)); 8692 rorq(op2, 32); 8693 } 8694 8695 bind(L_third_loop); 8696 decrementl(len); 8697 jccb(Assembler::negative, L_third_loop_exit); 8698 decrementl(len); 8699 jccb(Assembler::negative, L_last_x); 8700 8701 movq(op1, Address(x, len, Address::times_4, 0)); 8702 rorq(op1, 32); 8703 8704 bind(L_multiply); 8705 subl(zlen, 2); 8706 movq(sum, Address(z, zlen, Address::times_4, 0)); 8707 8708 // Multiply 64 bit by 64 bit and add 64 bits lower half and upper 64 bits as carry. 8709 if (UseBMI2Instructions) { 8710 multiply_add_64_bmi2(sum, op1, op2, carry, tmp2); 8711 } 8712 else { 8713 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 8714 } 8715 8716 movq(Address(z, zlen, Address::times_4, 0), sum); 8717 8718 jmp(L_third_loop); 8719 bind(L_third_loop_exit); 8720 8721 // Fourth loop 8722 // Add 64 bit long carry into z with carry propogation. 8723 // Uses offsetted zlen. 8724 add_one_64(z, zlen, carry, tmp1); 8725 8726 pop(len); 8727 pop(zlen); 8728 jmp(L_second_loop); 8729 8730 // Next infrequent code is moved outside loops. 8731 bind(L_last_x); 8732 movl(op1, Address(x, 0)); 8733 jmp(L_multiply); 8734 8735 bind(L_second_loop_exit); 8736 pop(len); 8737 pop(zlen); 8738 pop(len); 8739 pop(zlen); 8740 8741 // Fifth loop 8742 // Shift z left 1 bit. 8743 lshift_by_1(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4); 8744 8745 // z[zlen-1] |= x[len-1] & 1; 8746 movl(tmp3, Address(x, len, Address::times_4, -4)); 8747 andl(tmp3, 1); 8748 orl(Address(z, zlen, Address::times_4, -4), tmp3); 8749 8750 pop(tmp5); 8751 pop(tmp4); 8752 pop(tmp3); 8753 pop(tmp2); 8754 pop(tmp1); 8755 } 8756 8757 /** 8758 * Helper function for mul_add() 8759 * Multiply the in[] by int k and add to out[] starting at offset offs using 8760 * 128 bit by 32 bit multiply and return the carry in tmp5. 8761 * Only quad int aligned length of in[] is operated on in this function. 8762 * k is in rdxReg for BMI2Instructions, for others it is in tmp2. 8763 * This function preserves out, in and k registers. 8764 * len and offset point to the appropriate index in "in" & "out" correspondingly 8765 * tmp5 has the carry. 8766 * other registers are temporary and are modified. 8767 * 8768 */ 8769 void MacroAssembler::mul_add_128_x_32_loop(Register out, Register in, 8770 Register offset, Register len, Register tmp1, Register tmp2, Register tmp3, 8771 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 8772 8773 Label L_first_loop, L_first_loop_exit; 8774 8775 movl(tmp1, len); 8776 shrl(tmp1, 2); 8777 8778 bind(L_first_loop); 8779 subl(tmp1, 1); 8780 jccb(Assembler::negative, L_first_loop_exit); 8781 8782 subl(len, 4); 8783 subl(offset, 4); 8784 8785 Register op2 = tmp2; 8786 const Register sum = tmp3; 8787 const Register op1 = tmp4; 8788 const Register carry = tmp5; 8789 8790 if (UseBMI2Instructions) { 8791 op2 = rdxReg; 8792 } 8793 8794 movq(op1, Address(in, len, Address::times_4, 8)); 8795 rorq(op1, 32); 8796 movq(sum, Address(out, offset, Address::times_4, 8)); 8797 rorq(sum, 32); 8798 if (UseBMI2Instructions) { 8799 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 8800 } 8801 else { 8802 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 8803 } 8804 // Store back in big endian from little endian 8805 rorq(sum, 0x20); 8806 movq(Address(out, offset, Address::times_4, 8), sum); 8807 8808 movq(op1, Address(in, len, Address::times_4, 0)); 8809 rorq(op1, 32); 8810 movq(sum, Address(out, offset, Address::times_4, 0)); 8811 rorq(sum, 32); 8812 if (UseBMI2Instructions) { 8813 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 8814 } 8815 else { 8816 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 8817 } 8818 // Store back in big endian from little endian 8819 rorq(sum, 0x20); 8820 movq(Address(out, offset, Address::times_4, 0), sum); 8821 8822 jmp(L_first_loop); 8823 bind(L_first_loop_exit); 8824 } 8825 8826 /** 8827 * Code for BigInteger::mulAdd() intrinsic 8828 * 8829 * rdi: out 8830 * rsi: in 8831 * r11: offs (out.length - offset) 8832 * rcx: len 8833 * r8: k 8834 * r12: tmp1 8835 * r13: tmp2 8836 * r14: tmp3 8837 * r15: tmp4 8838 * rbx: tmp5 8839 * Multiply the in[] by word k and add to out[], return the carry in rax 8840 */ 8841 void MacroAssembler::mul_add(Register out, Register in, Register offs, 8842 Register len, Register k, Register tmp1, Register tmp2, Register tmp3, 8843 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 8844 8845 Label L_carry, L_last_in, L_done; 8846 8847 // carry = 0; 8848 // for (int j=len-1; j >= 0; j--) { 8849 // long product = (in[j] & LONG_MASK) * kLong + 8850 // (out[offs] & LONG_MASK) + carry; 8851 // out[offs--] = (int)product; 8852 // carry = product >>> 32; 8853 // } 8854 // 8855 push(tmp1); 8856 push(tmp2); 8857 push(tmp3); 8858 push(tmp4); 8859 push(tmp5); 8860 8861 Register op2 = tmp2; 8862 const Register sum = tmp3; 8863 const Register op1 = tmp4; 8864 const Register carry = tmp5; 8865 8866 if (UseBMI2Instructions) { 8867 op2 = rdxReg; 8868 movl(op2, k); 8869 } 8870 else { 8871 movl(op2, k); 8872 } 8873 8874 xorq(carry, carry); 8875 8876 //First loop 8877 8878 //Multiply in[] by k in a 4 way unrolled loop using 128 bit by 32 bit multiply 8879 //The carry is in tmp5 8880 mul_add_128_x_32_loop(out, in, offs, len, tmp1, tmp2, tmp3, tmp4, tmp5, rdxReg, raxReg); 8881 8882 //Multiply the trailing in[] entry using 64 bit by 32 bit, if any 8883 decrementl(len); 8884 jccb(Assembler::negative, L_carry); 8885 decrementl(len); 8886 jccb(Assembler::negative, L_last_in); 8887 8888 movq(op1, Address(in, len, Address::times_4, 0)); 8889 rorq(op1, 32); 8890 8891 subl(offs, 2); 8892 movq(sum, Address(out, offs, Address::times_4, 0)); 8893 rorq(sum, 32); 8894 8895 if (UseBMI2Instructions) { 8896 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 8897 } 8898 else { 8899 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 8900 } 8901 8902 // Store back in big endian from little endian 8903 rorq(sum, 0x20); 8904 movq(Address(out, offs, Address::times_4, 0), sum); 8905 8906 testl(len, len); 8907 jccb(Assembler::zero, L_carry); 8908 8909 //Multiply the last in[] entry, if any 8910 bind(L_last_in); 8911 movl(op1, Address(in, 0)); 8912 movl(sum, Address(out, offs, Address::times_4, -4)); 8913 8914 movl(raxReg, k); 8915 mull(op1); //tmp4 * eax -> edx:eax 8916 addl(sum, carry); 8917 adcl(rdxReg, 0); 8918 addl(sum, raxReg); 8919 adcl(rdxReg, 0); 8920 movl(carry, rdxReg); 8921 8922 movl(Address(out, offs, Address::times_4, -4), sum); 8923 8924 bind(L_carry); 8925 //return tmp5/carry as carry in rax 8926 movl(rax, carry); 8927 8928 bind(L_done); 8929 pop(tmp5); 8930 pop(tmp4); 8931 pop(tmp3); 8932 pop(tmp2); 8933 pop(tmp1); 8934 } 8935 #endif 8936 8937 /** 8938 * Emits code to update CRC-32 with a byte value according to constants in table 8939 * 8940 * @param [in,out]crc Register containing the crc. 8941 * @param [in]val Register containing the byte to fold into the CRC. 8942 * @param [in]table Register containing the table of crc constants. 8943 * 8944 * uint32_t crc; 8945 * val = crc_table[(val ^ crc) & 0xFF]; 8946 * crc = val ^ (crc >> 8); 8947 * 8948 */ 8949 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) { 8950 xorl(val, crc); 8951 andl(val, 0xFF); 8952 shrl(crc, 8); // unsigned shift 8953 xorl(crc, Address(table, val, Address::times_4, 0)); 8954 } 8955 8956 /** 8957 * Fold four 128-bit data chunks 8958 */ 8959 void MacroAssembler::fold_128bit_crc32_avx512(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) { 8960 evpclmulhdq(xtmp, xK, xcrc, Assembler::AVX_512bit); // [123:64] 8961 evpclmulldq(xcrc, xK, xcrc, Assembler::AVX_512bit); // [63:0] 8962 evpxorq(xcrc, xcrc, Address(buf, offset), Assembler::AVX_512bit /* vector_len */); 8963 evpxorq(xcrc, xcrc, xtmp, Assembler::AVX_512bit /* vector_len */); 8964 } 8965 8966 /** 8967 * Fold 128-bit data chunk 8968 */ 8969 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) { 8970 if (UseAVX > 0) { 8971 vpclmulhdq(xtmp, xK, xcrc); // [123:64] 8972 vpclmulldq(xcrc, xK, xcrc); // [63:0] 8973 vpxor(xcrc, xcrc, Address(buf, offset), 0 /* vector_len */); 8974 pxor(xcrc, xtmp); 8975 } else { 8976 movdqa(xtmp, xcrc); 8977 pclmulhdq(xtmp, xK); // [123:64] 8978 pclmulldq(xcrc, xK); // [63:0] 8979 pxor(xcrc, xtmp); 8980 movdqu(xtmp, Address(buf, offset)); 8981 pxor(xcrc, xtmp); 8982 } 8983 } 8984 8985 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) { 8986 if (UseAVX > 0) { 8987 vpclmulhdq(xtmp, xK, xcrc); 8988 vpclmulldq(xcrc, xK, xcrc); 8989 pxor(xcrc, xbuf); 8990 pxor(xcrc, xtmp); 8991 } else { 8992 movdqa(xtmp, xcrc); 8993 pclmulhdq(xtmp, xK); 8994 pclmulldq(xcrc, xK); 8995 pxor(xcrc, xbuf); 8996 pxor(xcrc, xtmp); 8997 } 8998 } 8999 9000 /** 9001 * 8-bit folds to compute 32-bit CRC 9002 * 9003 * uint64_t xcrc; 9004 * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8); 9005 */ 9006 void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) { 9007 movdl(tmp, xcrc); 9008 andl(tmp, 0xFF); 9009 movdl(xtmp, Address(table, tmp, Address::times_4, 0)); 9010 psrldq(xcrc, 1); // unsigned shift one byte 9011 pxor(xcrc, xtmp); 9012 } 9013 9014 /** 9015 * uint32_t crc; 9016 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8); 9017 */ 9018 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) { 9019 movl(tmp, crc); 9020 andl(tmp, 0xFF); 9021 shrl(crc, 8); 9022 xorl(crc, Address(table, tmp, Address::times_4, 0)); 9023 } 9024 9025 /** 9026 * @param crc register containing existing CRC (32-bit) 9027 * @param buf register pointing to input byte buffer (byte*) 9028 * @param len register containing number of bytes 9029 * @param table register that will contain address of CRC table 9030 * @param tmp scratch register 9031 */ 9032 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) { 9033 assert_different_registers(crc, buf, len, table, tmp, rax); 9034 9035 Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned; 9036 Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop; 9037 9038 // For EVEX with VL and BW, provide a standard mask, VL = 128 will guide the merge 9039 // context for the registers used, where all instructions below are using 128-bit mode 9040 // On EVEX without VL and BW, these instructions will all be AVX. 9041 lea(table, ExternalAddress(StubRoutines::crc_table_addr())); 9042 notl(crc); // ~crc 9043 cmpl(len, 16); 9044 jcc(Assembler::less, L_tail); 9045 9046 // Align buffer to 16 bytes 9047 movl(tmp, buf); 9048 andl(tmp, 0xF); 9049 jccb(Assembler::zero, L_aligned); 9050 subl(tmp, 16); 9051 addl(len, tmp); 9052 9053 align(4); 9054 BIND(L_align_loop); 9055 movsbl(rax, Address(buf, 0)); // load byte with sign extension 9056 update_byte_crc32(crc, rax, table); 9057 increment(buf); 9058 incrementl(tmp); 9059 jccb(Assembler::less, L_align_loop); 9060 9061 BIND(L_aligned); 9062 movl(tmp, len); // save 9063 shrl(len, 4); 9064 jcc(Assembler::zero, L_tail_restore); 9065 9066 // Fold crc into first bytes of vector 9067 movdqa(xmm1, Address(buf, 0)); 9068 movdl(rax, xmm1); 9069 xorl(crc, rax); 9070 if (VM_Version::supports_sse4_1()) { 9071 pinsrd(xmm1, crc, 0); 9072 } else { 9073 pinsrw(xmm1, crc, 0); 9074 shrl(crc, 16); 9075 pinsrw(xmm1, crc, 1); 9076 } 9077 addptr(buf, 16); 9078 subl(len, 4); // len > 0 9079 jcc(Assembler::less, L_fold_tail); 9080 9081 movdqa(xmm2, Address(buf, 0)); 9082 movdqa(xmm3, Address(buf, 16)); 9083 movdqa(xmm4, Address(buf, 32)); 9084 addptr(buf, 48); 9085 subl(len, 3); 9086 jcc(Assembler::lessEqual, L_fold_512b); 9087 9088 // Fold total 512 bits of polynomial on each iteration, 9089 // 128 bits per each of 4 parallel streams. 9090 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32)); 9091 9092 align(32); 9093 BIND(L_fold_512b_loop); 9094 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0); 9095 fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16); 9096 fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32); 9097 fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48); 9098 addptr(buf, 64); 9099 subl(len, 4); 9100 jcc(Assembler::greater, L_fold_512b_loop); 9101 9102 // Fold 512 bits to 128 bits. 9103 BIND(L_fold_512b); 9104 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16)); 9105 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2); 9106 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3); 9107 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4); 9108 9109 // Fold the rest of 128 bits data chunks 9110 BIND(L_fold_tail); 9111 addl(len, 3); 9112 jccb(Assembler::lessEqual, L_fold_128b); 9113 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16)); 9114 9115 BIND(L_fold_tail_loop); 9116 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0); 9117 addptr(buf, 16); 9118 decrementl(len); 9119 jccb(Assembler::greater, L_fold_tail_loop); 9120 9121 // Fold 128 bits in xmm1 down into 32 bits in crc register. 9122 BIND(L_fold_128b); 9123 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr())); 9124 if (UseAVX > 0) { 9125 vpclmulqdq(xmm2, xmm0, xmm1, 0x1); 9126 vpand(xmm3, xmm0, xmm2, 0 /* vector_len */); 9127 vpclmulqdq(xmm0, xmm0, xmm3, 0x1); 9128 } else { 9129 movdqa(xmm2, xmm0); 9130 pclmulqdq(xmm2, xmm1, 0x1); 9131 movdqa(xmm3, xmm0); 9132 pand(xmm3, xmm2); 9133 pclmulqdq(xmm0, xmm3, 0x1); 9134 } 9135 psrldq(xmm1, 8); 9136 psrldq(xmm2, 4); 9137 pxor(xmm0, xmm1); 9138 pxor(xmm0, xmm2); 9139 9140 // 8 8-bit folds to compute 32-bit CRC. 9141 for (int j = 0; j < 4; j++) { 9142 fold_8bit_crc32(xmm0, table, xmm1, rax); 9143 } 9144 movdl(crc, xmm0); // mov 32 bits to general register 9145 for (int j = 0; j < 4; j++) { 9146 fold_8bit_crc32(crc, table, rax); 9147 } 9148 9149 BIND(L_tail_restore); 9150 movl(len, tmp); // restore 9151 BIND(L_tail); 9152 andl(len, 0xf); 9153 jccb(Assembler::zero, L_exit); 9154 9155 // Fold the rest of bytes 9156 align(4); 9157 BIND(L_tail_loop); 9158 movsbl(rax, Address(buf, 0)); // load byte with sign extension 9159 update_byte_crc32(crc, rax, table); 9160 increment(buf); 9161 decrementl(len); 9162 jccb(Assembler::greater, L_tail_loop); 9163 9164 BIND(L_exit); 9165 notl(crc); // ~c 9166 } 9167 9168 #ifdef _LP64 9169 // S. Gueron / Information Processing Letters 112 (2012) 184 9170 // Algorithm 4: Computing carry-less multiplication using a precomputed lookup table. 9171 // Input: A 32 bit value B = [byte3, byte2, byte1, byte0]. 9172 // Output: the 64-bit carry-less product of B * CONST 9173 void MacroAssembler::crc32c_ipl_alg4(Register in, uint32_t n, 9174 Register tmp1, Register tmp2, Register tmp3) { 9175 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr())); 9176 if (n > 0) { 9177 addq(tmp3, n * 256 * 8); 9178 } 9179 // Q1 = TABLEExt[n][B & 0xFF]; 9180 movl(tmp1, in); 9181 andl(tmp1, 0x000000FF); 9182 shll(tmp1, 3); 9183 addq(tmp1, tmp3); 9184 movq(tmp1, Address(tmp1, 0)); 9185 9186 // Q2 = TABLEExt[n][B >> 8 & 0xFF]; 9187 movl(tmp2, in); 9188 shrl(tmp2, 8); 9189 andl(tmp2, 0x000000FF); 9190 shll(tmp2, 3); 9191 addq(tmp2, tmp3); 9192 movq(tmp2, Address(tmp2, 0)); 9193 9194 shlq(tmp2, 8); 9195 xorq(tmp1, tmp2); 9196 9197 // Q3 = TABLEExt[n][B >> 16 & 0xFF]; 9198 movl(tmp2, in); 9199 shrl(tmp2, 16); 9200 andl(tmp2, 0x000000FF); 9201 shll(tmp2, 3); 9202 addq(tmp2, tmp3); 9203 movq(tmp2, Address(tmp2, 0)); 9204 9205 shlq(tmp2, 16); 9206 xorq(tmp1, tmp2); 9207 9208 // Q4 = TABLEExt[n][B >> 24 & 0xFF]; 9209 shrl(in, 24); 9210 andl(in, 0x000000FF); 9211 shll(in, 3); 9212 addq(in, tmp3); 9213 movq(in, Address(in, 0)); 9214 9215 shlq(in, 24); 9216 xorq(in, tmp1); 9217 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24; 9218 } 9219 9220 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1, 9221 Register in_out, 9222 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported, 9223 XMMRegister w_xtmp2, 9224 Register tmp1, 9225 Register n_tmp2, Register n_tmp3) { 9226 if (is_pclmulqdq_supported) { 9227 movdl(w_xtmp1, in_out); // modified blindly 9228 9229 movl(tmp1, const_or_pre_comp_const_index); 9230 movdl(w_xtmp2, tmp1); 9231 pclmulqdq(w_xtmp1, w_xtmp2, 0); 9232 9233 movdq(in_out, w_xtmp1); 9234 } else { 9235 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3); 9236 } 9237 } 9238 9239 // Recombination Alternative 2: No bit-reflections 9240 // T1 = (CRC_A * U1) << 1 9241 // T2 = (CRC_B * U2) << 1 9242 // C1 = T1 >> 32 9243 // C2 = T2 >> 32 9244 // T1 = T1 & 0xFFFFFFFF 9245 // T2 = T2 & 0xFFFFFFFF 9246 // T1 = CRC32(0, T1) 9247 // T2 = CRC32(0, T2) 9248 // C1 = C1 ^ T1 9249 // C2 = C2 ^ T2 9250 // CRC = C1 ^ C2 ^ CRC_C 9251 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, 9252 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 9253 Register tmp1, Register tmp2, 9254 Register n_tmp3) { 9255 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 9256 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 9257 shlq(in_out, 1); 9258 movl(tmp1, in_out); 9259 shrq(in_out, 32); 9260 xorl(tmp2, tmp2); 9261 crc32(tmp2, tmp1, 4); 9262 xorl(in_out, tmp2); // we don't care about upper 32 bit contents here 9263 shlq(in1, 1); 9264 movl(tmp1, in1); 9265 shrq(in1, 32); 9266 xorl(tmp2, tmp2); 9267 crc32(tmp2, tmp1, 4); 9268 xorl(in1, tmp2); 9269 xorl(in_out, in1); 9270 xorl(in_out, in2); 9271 } 9272 9273 // Set N to predefined value 9274 // Subtract from a lenght of a buffer 9275 // execute in a loop: 9276 // CRC_A = 0xFFFFFFFF, CRC_B = 0, CRC_C = 0 9277 // for i = 1 to N do 9278 // CRC_A = CRC32(CRC_A, A[i]) 9279 // CRC_B = CRC32(CRC_B, B[i]) 9280 // CRC_C = CRC32(CRC_C, C[i]) 9281 // end for 9282 // Recombine 9283 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, 9284 Register in_out1, Register in_out2, Register in_out3, 9285 Register tmp1, Register tmp2, Register tmp3, 9286 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 9287 Register tmp4, Register tmp5, 9288 Register n_tmp6) { 9289 Label L_processPartitions; 9290 Label L_processPartition; 9291 Label L_exit; 9292 9293 bind(L_processPartitions); 9294 cmpl(in_out1, 3 * size); 9295 jcc(Assembler::less, L_exit); 9296 xorl(tmp1, tmp1); 9297 xorl(tmp2, tmp2); 9298 movq(tmp3, in_out2); 9299 addq(tmp3, size); 9300 9301 bind(L_processPartition); 9302 crc32(in_out3, Address(in_out2, 0), 8); 9303 crc32(tmp1, Address(in_out2, size), 8); 9304 crc32(tmp2, Address(in_out2, size * 2), 8); 9305 addq(in_out2, 8); 9306 cmpq(in_out2, tmp3); 9307 jcc(Assembler::less, L_processPartition); 9308 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2, 9309 w_xtmp1, w_xtmp2, w_xtmp3, 9310 tmp4, tmp5, 9311 n_tmp6); 9312 addq(in_out2, 2 * size); 9313 subl(in_out1, 3 * size); 9314 jmp(L_processPartitions); 9315 9316 bind(L_exit); 9317 } 9318 #else 9319 void MacroAssembler::crc32c_ipl_alg4(Register in_out, uint32_t n, 9320 Register tmp1, Register tmp2, Register tmp3, 9321 XMMRegister xtmp1, XMMRegister xtmp2) { 9322 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr())); 9323 if (n > 0) { 9324 addl(tmp3, n * 256 * 8); 9325 } 9326 // Q1 = TABLEExt[n][B & 0xFF]; 9327 movl(tmp1, in_out); 9328 andl(tmp1, 0x000000FF); 9329 shll(tmp1, 3); 9330 addl(tmp1, tmp3); 9331 movq(xtmp1, Address(tmp1, 0)); 9332 9333 // Q2 = TABLEExt[n][B >> 8 & 0xFF]; 9334 movl(tmp2, in_out); 9335 shrl(tmp2, 8); 9336 andl(tmp2, 0x000000FF); 9337 shll(tmp2, 3); 9338 addl(tmp2, tmp3); 9339 movq(xtmp2, Address(tmp2, 0)); 9340 9341 psllq(xtmp2, 8); 9342 pxor(xtmp1, xtmp2); 9343 9344 // Q3 = TABLEExt[n][B >> 16 & 0xFF]; 9345 movl(tmp2, in_out); 9346 shrl(tmp2, 16); 9347 andl(tmp2, 0x000000FF); 9348 shll(tmp2, 3); 9349 addl(tmp2, tmp3); 9350 movq(xtmp2, Address(tmp2, 0)); 9351 9352 psllq(xtmp2, 16); 9353 pxor(xtmp1, xtmp2); 9354 9355 // Q4 = TABLEExt[n][B >> 24 & 0xFF]; 9356 shrl(in_out, 24); 9357 andl(in_out, 0x000000FF); 9358 shll(in_out, 3); 9359 addl(in_out, tmp3); 9360 movq(xtmp2, Address(in_out, 0)); 9361 9362 psllq(xtmp2, 24); 9363 pxor(xtmp1, xtmp2); // Result in CXMM 9364 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24; 9365 } 9366 9367 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1, 9368 Register in_out, 9369 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported, 9370 XMMRegister w_xtmp2, 9371 Register tmp1, 9372 Register n_tmp2, Register n_tmp3) { 9373 if (is_pclmulqdq_supported) { 9374 movdl(w_xtmp1, in_out); 9375 9376 movl(tmp1, const_or_pre_comp_const_index); 9377 movdl(w_xtmp2, tmp1); 9378 pclmulqdq(w_xtmp1, w_xtmp2, 0); 9379 // Keep result in XMM since GPR is 32 bit in length 9380 } else { 9381 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3, w_xtmp1, w_xtmp2); 9382 } 9383 } 9384 9385 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, 9386 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 9387 Register tmp1, Register tmp2, 9388 Register n_tmp3) { 9389 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 9390 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 9391 9392 psllq(w_xtmp1, 1); 9393 movdl(tmp1, w_xtmp1); 9394 psrlq(w_xtmp1, 32); 9395 movdl(in_out, w_xtmp1); 9396 9397 xorl(tmp2, tmp2); 9398 crc32(tmp2, tmp1, 4); 9399 xorl(in_out, tmp2); 9400 9401 psllq(w_xtmp2, 1); 9402 movdl(tmp1, w_xtmp2); 9403 psrlq(w_xtmp2, 32); 9404 movdl(in1, w_xtmp2); 9405 9406 xorl(tmp2, tmp2); 9407 crc32(tmp2, tmp1, 4); 9408 xorl(in1, tmp2); 9409 xorl(in_out, in1); 9410 xorl(in_out, in2); 9411 } 9412 9413 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, 9414 Register in_out1, Register in_out2, Register in_out3, 9415 Register tmp1, Register tmp2, Register tmp3, 9416 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 9417 Register tmp4, Register tmp5, 9418 Register n_tmp6) { 9419 Label L_processPartitions; 9420 Label L_processPartition; 9421 Label L_exit; 9422 9423 bind(L_processPartitions); 9424 cmpl(in_out1, 3 * size); 9425 jcc(Assembler::less, L_exit); 9426 xorl(tmp1, tmp1); 9427 xorl(tmp2, tmp2); 9428 movl(tmp3, in_out2); 9429 addl(tmp3, size); 9430 9431 bind(L_processPartition); 9432 crc32(in_out3, Address(in_out2, 0), 4); 9433 crc32(tmp1, Address(in_out2, size), 4); 9434 crc32(tmp2, Address(in_out2, size*2), 4); 9435 crc32(in_out3, Address(in_out2, 0+4), 4); 9436 crc32(tmp1, Address(in_out2, size+4), 4); 9437 crc32(tmp2, Address(in_out2, size*2+4), 4); 9438 addl(in_out2, 8); 9439 cmpl(in_out2, tmp3); 9440 jcc(Assembler::less, L_processPartition); 9441 9442 push(tmp3); 9443 push(in_out1); 9444 push(in_out2); 9445 tmp4 = tmp3; 9446 tmp5 = in_out1; 9447 n_tmp6 = in_out2; 9448 9449 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2, 9450 w_xtmp1, w_xtmp2, w_xtmp3, 9451 tmp4, tmp5, 9452 n_tmp6); 9453 9454 pop(in_out2); 9455 pop(in_out1); 9456 pop(tmp3); 9457 9458 addl(in_out2, 2 * size); 9459 subl(in_out1, 3 * size); 9460 jmp(L_processPartitions); 9461 9462 bind(L_exit); 9463 } 9464 #endif //LP64 9465 9466 #ifdef _LP64 9467 // Algorithm 2: Pipelined usage of the CRC32 instruction. 9468 // Input: A buffer I of L bytes. 9469 // Output: the CRC32C value of the buffer. 9470 // Notations: 9471 // Write L = 24N + r, with N = floor (L/24). 9472 // r = L mod 24 (0 <= r < 24). 9473 // Consider I as the concatenation of A|B|C|R, where A, B, C, each, 9474 // N quadwords, and R consists of r bytes. 9475 // A[j] = I [8j+7:8j], j= 0, 1, ..., N-1 9476 // B[j] = I [N + 8j+7:N + 8j], j= 0, 1, ..., N-1 9477 // C[j] = I [2N + 8j+7:2N + 8j], j= 0, 1, ..., N-1 9478 // if r > 0 R[j] = I [3N +j], j= 0, 1, ...,r-1 9479 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2, 9480 Register tmp1, Register tmp2, Register tmp3, 9481 Register tmp4, Register tmp5, Register tmp6, 9482 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 9483 bool is_pclmulqdq_supported) { 9484 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS]; 9485 Label L_wordByWord; 9486 Label L_byteByByteProlog; 9487 Label L_byteByByte; 9488 Label L_exit; 9489 9490 if (is_pclmulqdq_supported ) { 9491 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr; 9492 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr+1); 9493 9494 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2); 9495 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3); 9496 9497 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4); 9498 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5); 9499 assert((CRC32C_NUM_PRECOMPUTED_CONSTANTS - 1 ) == 5, "Checking whether you declared all of the constants based on the number of \"chunks\""); 9500 } else { 9501 const_or_pre_comp_const_index[0] = 1; 9502 const_or_pre_comp_const_index[1] = 0; 9503 9504 const_or_pre_comp_const_index[2] = 3; 9505 const_or_pre_comp_const_index[3] = 2; 9506 9507 const_or_pre_comp_const_index[4] = 5; 9508 const_or_pre_comp_const_index[5] = 4; 9509 } 9510 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported, 9511 in2, in1, in_out, 9512 tmp1, tmp2, tmp3, 9513 w_xtmp1, w_xtmp2, w_xtmp3, 9514 tmp4, tmp5, 9515 tmp6); 9516 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported, 9517 in2, in1, in_out, 9518 tmp1, tmp2, tmp3, 9519 w_xtmp1, w_xtmp2, w_xtmp3, 9520 tmp4, tmp5, 9521 tmp6); 9522 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported, 9523 in2, in1, in_out, 9524 tmp1, tmp2, tmp3, 9525 w_xtmp1, w_xtmp2, w_xtmp3, 9526 tmp4, tmp5, 9527 tmp6); 9528 movl(tmp1, in2); 9529 andl(tmp1, 0x00000007); 9530 negl(tmp1); 9531 addl(tmp1, in2); 9532 addq(tmp1, in1); 9533 9534 BIND(L_wordByWord); 9535 cmpq(in1, tmp1); 9536 jcc(Assembler::greaterEqual, L_byteByByteProlog); 9537 crc32(in_out, Address(in1, 0), 4); 9538 addq(in1, 4); 9539 jmp(L_wordByWord); 9540 9541 BIND(L_byteByByteProlog); 9542 andl(in2, 0x00000007); 9543 movl(tmp2, 1); 9544 9545 BIND(L_byteByByte); 9546 cmpl(tmp2, in2); 9547 jccb(Assembler::greater, L_exit); 9548 crc32(in_out, Address(in1, 0), 1); 9549 incq(in1); 9550 incl(tmp2); 9551 jmp(L_byteByByte); 9552 9553 BIND(L_exit); 9554 } 9555 #else 9556 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2, 9557 Register tmp1, Register tmp2, Register tmp3, 9558 Register tmp4, Register tmp5, Register tmp6, 9559 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 9560 bool is_pclmulqdq_supported) { 9561 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS]; 9562 Label L_wordByWord; 9563 Label L_byteByByteProlog; 9564 Label L_byteByByte; 9565 Label L_exit; 9566 9567 if (is_pclmulqdq_supported) { 9568 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr; 9569 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 1); 9570 9571 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2); 9572 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3); 9573 9574 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4); 9575 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5); 9576 } else { 9577 const_or_pre_comp_const_index[0] = 1; 9578 const_or_pre_comp_const_index[1] = 0; 9579 9580 const_or_pre_comp_const_index[2] = 3; 9581 const_or_pre_comp_const_index[3] = 2; 9582 9583 const_or_pre_comp_const_index[4] = 5; 9584 const_or_pre_comp_const_index[5] = 4; 9585 } 9586 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported, 9587 in2, in1, in_out, 9588 tmp1, tmp2, tmp3, 9589 w_xtmp1, w_xtmp2, w_xtmp3, 9590 tmp4, tmp5, 9591 tmp6); 9592 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported, 9593 in2, in1, in_out, 9594 tmp1, tmp2, tmp3, 9595 w_xtmp1, w_xtmp2, w_xtmp3, 9596 tmp4, tmp5, 9597 tmp6); 9598 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported, 9599 in2, in1, in_out, 9600 tmp1, tmp2, tmp3, 9601 w_xtmp1, w_xtmp2, w_xtmp3, 9602 tmp4, tmp5, 9603 tmp6); 9604 movl(tmp1, in2); 9605 andl(tmp1, 0x00000007); 9606 negl(tmp1); 9607 addl(tmp1, in2); 9608 addl(tmp1, in1); 9609 9610 BIND(L_wordByWord); 9611 cmpl(in1, tmp1); 9612 jcc(Assembler::greaterEqual, L_byteByByteProlog); 9613 crc32(in_out, Address(in1,0), 4); 9614 addl(in1, 4); 9615 jmp(L_wordByWord); 9616 9617 BIND(L_byteByByteProlog); 9618 andl(in2, 0x00000007); 9619 movl(tmp2, 1); 9620 9621 BIND(L_byteByByte); 9622 cmpl(tmp2, in2); 9623 jccb(Assembler::greater, L_exit); 9624 movb(tmp1, Address(in1, 0)); 9625 crc32(in_out, tmp1, 1); 9626 incl(in1); 9627 incl(tmp2); 9628 jmp(L_byteByByte); 9629 9630 BIND(L_exit); 9631 } 9632 #endif // LP64 9633 #undef BIND 9634 #undef BLOCK_COMMENT 9635 9636 // Compress char[] array to byte[]. 9637 // ..\jdk\src\java.base\share\classes\java\lang\StringUTF16.java 9638 // @HotSpotIntrinsicCandidate 9639 // private static int compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) { 9640 // for (int i = 0; i < len; i++) { 9641 // int c = src[srcOff++]; 9642 // if (c >>> 8 != 0) { 9643 // return 0; 9644 // } 9645 // dst[dstOff++] = (byte)c; 9646 // } 9647 // return len; 9648 // } 9649 void MacroAssembler::char_array_compress(Register src, Register dst, Register len, 9650 XMMRegister tmp1Reg, XMMRegister tmp2Reg, 9651 XMMRegister tmp3Reg, XMMRegister tmp4Reg, 9652 Register tmp5, Register result) { 9653 Label copy_chars_loop, return_length, return_zero, done; 9654 9655 // rsi: src 9656 // rdi: dst 9657 // rdx: len 9658 // rcx: tmp5 9659 // rax: result 9660 9661 // rsi holds start addr of source char[] to be compressed 9662 // rdi holds start addr of destination byte[] 9663 // rdx holds length 9664 9665 assert(len != result, ""); 9666 9667 // save length for return 9668 push(len); 9669 9670 if ((AVX3Threshold == 0) && (UseAVX > 2) && // AVX512 9671 VM_Version::supports_avx512vlbw() && 9672 VM_Version::supports_bmi2()) { 9673 9674 Label copy_32_loop, copy_loop_tail, below_threshold; 9675 9676 // alignment 9677 Label post_alignment; 9678 9679 // if length of the string is less than 16, handle it in an old fashioned way 9680 testl(len, -32); 9681 jcc(Assembler::zero, below_threshold); 9682 9683 // First check whether a character is compressable ( <= 0xFF). 9684 // Create mask to test for Unicode chars inside zmm vector 9685 movl(result, 0x00FF); 9686 evpbroadcastw(tmp2Reg, result, Assembler::AVX_512bit); 9687 9688 testl(len, -64); 9689 jcc(Assembler::zero, post_alignment); 9690 9691 movl(tmp5, dst); 9692 andl(tmp5, (32 - 1)); 9693 negl(tmp5); 9694 andl(tmp5, (32 - 1)); 9695 9696 // bail out when there is nothing to be done 9697 testl(tmp5, 0xFFFFFFFF); 9698 jcc(Assembler::zero, post_alignment); 9699 9700 // ~(~0 << len), where len is the # of remaining elements to process 9701 movl(result, 0xFFFFFFFF); 9702 shlxl(result, result, tmp5); 9703 notl(result); 9704 kmovdl(k3, result); 9705 9706 evmovdquw(tmp1Reg, k3, Address(src, 0), Assembler::AVX_512bit); 9707 evpcmpuw(k2, k3, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit); 9708 ktestd(k2, k3); 9709 jcc(Assembler::carryClear, return_zero); 9710 9711 evpmovwb(Address(dst, 0), k3, tmp1Reg, Assembler::AVX_512bit); 9712 9713 addptr(src, tmp5); 9714 addptr(src, tmp5); 9715 addptr(dst, tmp5); 9716 subl(len, tmp5); 9717 9718 bind(post_alignment); 9719 // end of alignment 9720 9721 movl(tmp5, len); 9722 andl(tmp5, (32 - 1)); // tail count (in chars) 9723 andl(len, ~(32 - 1)); // vector count (in chars) 9724 jcc(Assembler::zero, copy_loop_tail); 9725 9726 lea(src, Address(src, len, Address::times_2)); 9727 lea(dst, Address(dst, len, Address::times_1)); 9728 negptr(len); 9729 9730 bind(copy_32_loop); 9731 evmovdquw(tmp1Reg, Address(src, len, Address::times_2), Assembler::AVX_512bit); 9732 evpcmpuw(k2, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit); 9733 kortestdl(k2, k2); 9734 jcc(Assembler::carryClear, return_zero); 9735 9736 // All elements in current processed chunk are valid candidates for 9737 // compression. Write a truncated byte elements to the memory. 9738 evpmovwb(Address(dst, len, Address::times_1), tmp1Reg, Assembler::AVX_512bit); 9739 addptr(len, 32); 9740 jcc(Assembler::notZero, copy_32_loop); 9741 9742 bind(copy_loop_tail); 9743 // bail out when there is nothing to be done 9744 testl(tmp5, 0xFFFFFFFF); 9745 jcc(Assembler::zero, return_length); 9746 9747 movl(len, tmp5); 9748 9749 // ~(~0 << len), where len is the # of remaining elements to process 9750 movl(result, 0xFFFFFFFF); 9751 shlxl(result, result, len); 9752 notl(result); 9753 9754 kmovdl(k3, result); 9755 9756 evmovdquw(tmp1Reg, k3, Address(src, 0), Assembler::AVX_512bit); 9757 evpcmpuw(k2, k3, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit); 9758 ktestd(k2, k3); 9759 jcc(Assembler::carryClear, return_zero); 9760 9761 evpmovwb(Address(dst, 0), k3, tmp1Reg, Assembler::AVX_512bit); 9762 jmp(return_length); 9763 9764 bind(below_threshold); 9765 } 9766 9767 if (UseSSE42Intrinsics) { 9768 Label copy_32_loop, copy_16, copy_tail; 9769 9770 movl(result, len); 9771 9772 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vectors 9773 9774 // vectored compression 9775 andl(len, 0xfffffff0); // vector count (in chars) 9776 andl(result, 0x0000000f); // tail count (in chars) 9777 testl(len, len); 9778 jcc(Assembler::zero, copy_16); 9779 9780 // compress 16 chars per iter 9781 movdl(tmp1Reg, tmp5); 9782 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg 9783 pxor(tmp4Reg, tmp4Reg); 9784 9785 lea(src, Address(src, len, Address::times_2)); 9786 lea(dst, Address(dst, len, Address::times_1)); 9787 negptr(len); 9788 9789 bind(copy_32_loop); 9790 movdqu(tmp2Reg, Address(src, len, Address::times_2)); // load 1st 8 characters 9791 por(tmp4Reg, tmp2Reg); 9792 movdqu(tmp3Reg, Address(src, len, Address::times_2, 16)); // load next 8 characters 9793 por(tmp4Reg, tmp3Reg); 9794 ptest(tmp4Reg, tmp1Reg); // check for Unicode chars in next vector 9795 jcc(Assembler::notZero, return_zero); 9796 packuswb(tmp2Reg, tmp3Reg); // only ASCII chars; compress each to 1 byte 9797 movdqu(Address(dst, len, Address::times_1), tmp2Reg); 9798 addptr(len, 16); 9799 jcc(Assembler::notZero, copy_32_loop); 9800 9801 // compress next vector of 8 chars (if any) 9802 bind(copy_16); 9803 movl(len, result); 9804 andl(len, 0xfffffff8); // vector count (in chars) 9805 andl(result, 0x00000007); // tail count (in chars) 9806 testl(len, len); 9807 jccb(Assembler::zero, copy_tail); 9808 9809 movdl(tmp1Reg, tmp5); 9810 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg 9811 pxor(tmp3Reg, tmp3Reg); 9812 9813 movdqu(tmp2Reg, Address(src, 0)); 9814 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 9815 jccb(Assembler::notZero, return_zero); 9816 packuswb(tmp2Reg, tmp3Reg); // only LATIN1 chars; compress each to 1 byte 9817 movq(Address(dst, 0), tmp2Reg); 9818 addptr(src, 16); 9819 addptr(dst, 8); 9820 9821 bind(copy_tail); 9822 movl(len, result); 9823 } 9824 // compress 1 char per iter 9825 testl(len, len); 9826 jccb(Assembler::zero, return_length); 9827 lea(src, Address(src, len, Address::times_2)); 9828 lea(dst, Address(dst, len, Address::times_1)); 9829 negptr(len); 9830 9831 bind(copy_chars_loop); 9832 load_unsigned_short(result, Address(src, len, Address::times_2)); 9833 testl(result, 0xff00); // check if Unicode char 9834 jccb(Assembler::notZero, return_zero); 9835 movb(Address(dst, len, Address::times_1), result); // ASCII char; compress to 1 byte 9836 increment(len); 9837 jcc(Assembler::notZero, copy_chars_loop); 9838 9839 // if compression succeeded, return length 9840 bind(return_length); 9841 pop(result); 9842 jmpb(done); 9843 9844 // if compression failed, return 0 9845 bind(return_zero); 9846 xorl(result, result); 9847 addptr(rsp, wordSize); 9848 9849 bind(done); 9850 } 9851 9852 // Inflate byte[] array to char[]. 9853 // ..\jdk\src\java.base\share\classes\java\lang\StringLatin1.java 9854 // @HotSpotIntrinsicCandidate 9855 // private static void inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) { 9856 // for (int i = 0; i < len; i++) { 9857 // dst[dstOff++] = (char)(src[srcOff++] & 0xff); 9858 // } 9859 // } 9860 void MacroAssembler::byte_array_inflate(Register src, Register dst, Register len, 9861 XMMRegister tmp1, Register tmp2) { 9862 Label copy_chars_loop, done, below_threshold, avx3_threshold; 9863 // rsi: src 9864 // rdi: dst 9865 // rdx: len 9866 // rcx: tmp2 9867 9868 // rsi holds start addr of source byte[] to be inflated 9869 // rdi holds start addr of destination char[] 9870 // rdx holds length 9871 assert_different_registers(src, dst, len, tmp2); 9872 movl(tmp2, len); 9873 if ((UseAVX > 2) && // AVX512 9874 VM_Version::supports_avx512vlbw() && 9875 VM_Version::supports_bmi2()) { 9876 9877 Label copy_32_loop, copy_tail; 9878 Register tmp3_aliased = len; 9879 9880 // if length of the string is less than 16, handle it in an old fashioned way 9881 testl(len, -16); 9882 jcc(Assembler::zero, below_threshold); 9883 9884 testl(len, -1 * AVX3Threshold); 9885 jcc(Assembler::zero, avx3_threshold); 9886 9887 // In order to use only one arithmetic operation for the main loop we use 9888 // this pre-calculation 9889 andl(tmp2, (32 - 1)); // tail count (in chars), 32 element wide loop 9890 andl(len, -32); // vector count 9891 jccb(Assembler::zero, copy_tail); 9892 9893 lea(src, Address(src, len, Address::times_1)); 9894 lea(dst, Address(dst, len, Address::times_2)); 9895 negptr(len); 9896 9897 9898 // inflate 32 chars per iter 9899 bind(copy_32_loop); 9900 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_512bit); 9901 evmovdquw(Address(dst, len, Address::times_2), tmp1, Assembler::AVX_512bit); 9902 addptr(len, 32); 9903 jcc(Assembler::notZero, copy_32_loop); 9904 9905 bind(copy_tail); 9906 // bail out when there is nothing to be done 9907 testl(tmp2, -1); // we don't destroy the contents of tmp2 here 9908 jcc(Assembler::zero, done); 9909 9910 // ~(~0 << length), where length is the # of remaining elements to process 9911 movl(tmp3_aliased, -1); 9912 shlxl(tmp3_aliased, tmp3_aliased, tmp2); 9913 notl(tmp3_aliased); 9914 kmovdl(k2, tmp3_aliased); 9915 evpmovzxbw(tmp1, k2, Address(src, 0), Assembler::AVX_512bit); 9916 evmovdquw(Address(dst, 0), k2, tmp1, Assembler::AVX_512bit); 9917 9918 jmp(done); 9919 bind(avx3_threshold); 9920 } 9921 if (UseSSE42Intrinsics) { 9922 Label copy_16_loop, copy_8_loop, copy_bytes, copy_new_tail, copy_tail; 9923 9924 if (UseAVX > 1) { 9925 andl(tmp2, (16 - 1)); 9926 andl(len, -16); 9927 jccb(Assembler::zero, copy_new_tail); 9928 } else { 9929 andl(tmp2, 0x00000007); // tail count (in chars) 9930 andl(len, 0xfffffff8); // vector count (in chars) 9931 jccb(Assembler::zero, copy_tail); 9932 } 9933 9934 // vectored inflation 9935 lea(src, Address(src, len, Address::times_1)); 9936 lea(dst, Address(dst, len, Address::times_2)); 9937 negptr(len); 9938 9939 if (UseAVX > 1) { 9940 bind(copy_16_loop); 9941 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_256bit); 9942 vmovdqu(Address(dst, len, Address::times_2), tmp1); 9943 addptr(len, 16); 9944 jcc(Assembler::notZero, copy_16_loop); 9945 9946 bind(below_threshold); 9947 bind(copy_new_tail); 9948 movl(len, tmp2); 9949 andl(tmp2, 0x00000007); 9950 andl(len, 0xFFFFFFF8); 9951 jccb(Assembler::zero, copy_tail); 9952 9953 pmovzxbw(tmp1, Address(src, 0)); 9954 movdqu(Address(dst, 0), tmp1); 9955 addptr(src, 8); 9956 addptr(dst, 2 * 8); 9957 9958 jmp(copy_tail, true); 9959 } 9960 9961 // inflate 8 chars per iter 9962 bind(copy_8_loop); 9963 pmovzxbw(tmp1, Address(src, len, Address::times_1)); // unpack to 8 words 9964 movdqu(Address(dst, len, Address::times_2), tmp1); 9965 addptr(len, 8); 9966 jcc(Assembler::notZero, copy_8_loop); 9967 9968 bind(copy_tail); 9969 movl(len, tmp2); 9970 9971 cmpl(len, 4); 9972 jccb(Assembler::less, copy_bytes); 9973 9974 movdl(tmp1, Address(src, 0)); // load 4 byte chars 9975 pmovzxbw(tmp1, tmp1); 9976 movq(Address(dst, 0), tmp1); 9977 subptr(len, 4); 9978 addptr(src, 4); 9979 addptr(dst, 8); 9980 9981 bind(copy_bytes); 9982 } else { 9983 bind(below_threshold); 9984 } 9985 9986 testl(len, len); 9987 jccb(Assembler::zero, done); 9988 lea(src, Address(src, len, Address::times_1)); 9989 lea(dst, Address(dst, len, Address::times_2)); 9990 negptr(len); 9991 9992 // inflate 1 char per iter 9993 bind(copy_chars_loop); 9994 load_unsigned_byte(tmp2, Address(src, len, Address::times_1)); // load byte char 9995 movw(Address(dst, len, Address::times_2), tmp2); // inflate byte char to word 9996 increment(len); 9997 jcc(Assembler::notZero, copy_chars_loop); 9998 9999 bind(done); 10000 } 10001 10002 #ifdef _LP64 10003 void MacroAssembler::cache_wb(Address line) 10004 { 10005 // 64 bit cpus always support clflush 10006 assert(VM_Version::supports_clflush(), "clflush should be available"); 10007 bool optimized = VM_Version::supports_clflushopt(); 10008 bool no_evict = VM_Version::supports_clwb(); 10009 10010 // prefer clwb (writeback without evict) otherwise 10011 // prefer clflushopt (potentially parallel writeback with evict) 10012 // otherwise fallback on clflush (serial writeback with evict) 10013 10014 if (optimized) { 10015 if (no_evict) { 10016 clwb(line); 10017 } else { 10018 clflushopt(line); 10019 } 10020 } else { 10021 // no need for fence when using CLFLUSH 10022 clflush(line); 10023 } 10024 } 10025 10026 void MacroAssembler::cache_wbsync(bool is_pre) 10027 { 10028 assert(VM_Version::supports_clflush(), "clflush should be available"); 10029 bool optimized = VM_Version::supports_clflushopt(); 10030 bool no_evict = VM_Version::supports_clwb(); 10031 10032 // pick the correct implementation 10033 10034 if (!is_pre && (optimized || no_evict)) { 10035 // need an sfence for post flush when using clflushopt or clwb 10036 // otherwise no no need for any synchroniaztion 10037 10038 sfence(); 10039 } 10040 } 10041 #endif // _LP64 10042 10043 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) { 10044 switch (cond) { 10045 // Note some conditions are synonyms for others 10046 case Assembler::zero: return Assembler::notZero; 10047 case Assembler::notZero: return Assembler::zero; 10048 case Assembler::less: return Assembler::greaterEqual; 10049 case Assembler::lessEqual: return Assembler::greater; 10050 case Assembler::greater: return Assembler::lessEqual; 10051 case Assembler::greaterEqual: return Assembler::less; 10052 case Assembler::below: return Assembler::aboveEqual; 10053 case Assembler::belowEqual: return Assembler::above; 10054 case Assembler::above: return Assembler::belowEqual; 10055 case Assembler::aboveEqual: return Assembler::below; 10056 case Assembler::overflow: return Assembler::noOverflow; 10057 case Assembler::noOverflow: return Assembler::overflow; 10058 case Assembler::negative: return Assembler::positive; 10059 case Assembler::positive: return Assembler::negative; 10060 case Assembler::parity: return Assembler::noParity; 10061 case Assembler::noParity: return Assembler::parity; 10062 } 10063 ShouldNotReachHere(); return Assembler::overflow; 10064 } 10065 10066 SkipIfEqual::SkipIfEqual( 10067 MacroAssembler* masm, const bool* flag_addr, bool value) { 10068 _masm = masm; 10069 _masm->cmp8(ExternalAddress((address)flag_addr), value); 10070 _masm->jcc(Assembler::equal, _label); 10071 } 10072 10073 SkipIfEqual::~SkipIfEqual() { 10074 _masm->bind(_label); 10075 } 10076 10077 // 32-bit Windows has its own fast-path implementation 10078 // of get_thread 10079 #if !defined(WIN32) || defined(_LP64) 10080 10081 // This is simply a call to Thread::current() 10082 void MacroAssembler::get_thread(Register thread) { 10083 if (thread != rax) { 10084 push(rax); 10085 } 10086 LP64_ONLY(push(rdi);) 10087 LP64_ONLY(push(rsi);) 10088 push(rdx); 10089 push(rcx); 10090 #ifdef _LP64 10091 push(r8); 10092 push(r9); 10093 push(r10); 10094 push(r11); 10095 #endif 10096 10097 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, Thread::current), 0); 10098 10099 #ifdef _LP64 10100 pop(r11); 10101 pop(r10); 10102 pop(r9); 10103 pop(r8); 10104 #endif 10105 pop(rcx); 10106 pop(rdx); 10107 LP64_ONLY(pop(rsi);) 10108 LP64_ONLY(pop(rdi);) 10109 if (thread != rax) { 10110 mov(thread, rax); 10111 pop(rax); 10112 } 10113 } 10114 10115 #endif