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/klass.inline.hpp" 38 #include "prims/methodHandles.hpp" 39 #include "runtime/biasedLocking.hpp" 40 #include "runtime/flags/flagSetting.hpp" 41 #include "runtime/interfaceSupport.inline.hpp" 42 #include "runtime/objectMonitor.hpp" 43 #include "runtime/os.hpp" 44 #include "runtime/safepoint.hpp" 45 #include "runtime/safepointMechanism.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 #include "runtime/stubRoutines.hpp" 48 #include "runtime/thread.hpp" 49 #include "utilities/macros.hpp" 50 #include "vmreg_x86.inline.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 } else { 431 ttyLocker ttyl; 432 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", msg); 433 } 434 // Don't assert holding the ttyLock 435 assert(false, "DEBUG MESSAGE: %s", msg); 436 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state); 437 } 438 439 void MacroAssembler::print_state32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip) { 440 ttyLocker ttyl; 441 FlagSetting fs(Debugging, true); 442 tty->print_cr("eip = 0x%08x", eip); 443 #ifndef PRODUCT 444 if ((WizardMode || Verbose) && PrintMiscellaneous) { 445 tty->cr(); 446 findpc(eip); 447 tty->cr(); 448 } 449 #endif 450 #define PRINT_REG(rax) \ 451 { tty->print("%s = ", #rax); os::print_location(tty, rax); } 452 PRINT_REG(rax); 453 PRINT_REG(rbx); 454 PRINT_REG(rcx); 455 PRINT_REG(rdx); 456 PRINT_REG(rdi); 457 PRINT_REG(rsi); 458 PRINT_REG(rbp); 459 PRINT_REG(rsp); 460 #undef PRINT_REG 461 // Print some words near top of staack. 462 int* dump_sp = (int*) rsp; 463 for (int col1 = 0; col1 < 8; col1++) { 464 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 465 os::print_location(tty, *dump_sp++); 466 } 467 for (int row = 0; row < 16; row++) { 468 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 469 for (int col = 0; col < 8; col++) { 470 tty->print(" 0x%08x", *dump_sp++); 471 } 472 tty->cr(); 473 } 474 // Print some instructions around pc: 475 Disassembler::decode((address)eip-64, (address)eip); 476 tty->print_cr("--------"); 477 Disassembler::decode((address)eip, (address)eip+32); 478 } 479 480 void MacroAssembler::stop(const char* msg) { 481 ExternalAddress message((address)msg); 482 // push address of message 483 pushptr(message.addr()); 484 { Label L; call(L, relocInfo::none); bind(L); } // push eip 485 pusha(); // push registers 486 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug32))); 487 hlt(); 488 } 489 490 void MacroAssembler::warn(const char* msg) { 491 push_CPU_state(); 492 493 ExternalAddress message((address) msg); 494 // push address of message 495 pushptr(message.addr()); 496 497 call(RuntimeAddress(CAST_FROM_FN_PTR(address, warning))); 498 addl(rsp, wordSize); // discard argument 499 pop_CPU_state(); 500 } 501 502 void MacroAssembler::print_state() { 503 { Label L; call(L, relocInfo::none); bind(L); } // push eip 504 pusha(); // push registers 505 506 push_CPU_state(); 507 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::print_state32))); 508 pop_CPU_state(); 509 510 popa(); 511 addl(rsp, wordSize); 512 } 513 514 #else // _LP64 515 516 // 64 bit versions 517 518 Address MacroAssembler::as_Address(AddressLiteral adr) { 519 // amd64 always does this as a pc-rel 520 // we can be absolute or disp based on the instruction type 521 // jmp/call are displacements others are absolute 522 assert(!adr.is_lval(), "must be rval"); 523 assert(reachable(adr), "must be"); 524 return Address((int32_t)(intptr_t)(adr.target() - pc()), adr.target(), adr.reloc()); 525 526 } 527 528 Address MacroAssembler::as_Address(ArrayAddress adr) { 529 AddressLiteral base = adr.base(); 530 lea(rscratch1, base); 531 Address index = adr.index(); 532 assert(index._disp == 0, "must not have disp"); // maybe it can? 533 Address array(rscratch1, index._index, index._scale, index._disp); 534 return array; 535 } 536 537 void MacroAssembler::call_VM_leaf_base(address entry_point, int num_args) { 538 Label L, E; 539 540 #ifdef _WIN64 541 // Windows always allocates space for it's register args 542 assert(num_args <= 4, "only register arguments supported"); 543 subq(rsp, frame::arg_reg_save_area_bytes); 544 #endif 545 546 // Align stack if necessary 547 testl(rsp, 15); 548 jcc(Assembler::zero, L); 549 550 subq(rsp, 8); 551 { 552 call(RuntimeAddress(entry_point)); 553 } 554 addq(rsp, 8); 555 jmp(E); 556 557 bind(L); 558 { 559 call(RuntimeAddress(entry_point)); 560 } 561 562 bind(E); 563 564 #ifdef _WIN64 565 // restore stack pointer 566 addq(rsp, frame::arg_reg_save_area_bytes); 567 #endif 568 569 } 570 571 void MacroAssembler::cmp64(Register src1, AddressLiteral src2) { 572 assert(!src2.is_lval(), "should use cmpptr"); 573 574 if (reachable(src2)) { 575 cmpq(src1, as_Address(src2)); 576 } else { 577 lea(rscratch1, src2); 578 Assembler::cmpq(src1, Address(rscratch1, 0)); 579 } 580 } 581 582 int MacroAssembler::corrected_idivq(Register reg) { 583 // Full implementation of Java ldiv and lrem; checks for special 584 // case as described in JVM spec., p.243 & p.271. The function 585 // returns the (pc) offset of the idivl instruction - may be needed 586 // for implicit exceptions. 587 // 588 // normal case special case 589 // 590 // input : rax: dividend min_long 591 // reg: divisor (may not be eax/edx) -1 592 // 593 // output: rax: quotient (= rax idiv reg) min_long 594 // rdx: remainder (= rax irem reg) 0 595 assert(reg != rax && reg != rdx, "reg cannot be rax or rdx register"); 596 static const int64_t min_long = 0x8000000000000000; 597 Label normal_case, special_case; 598 599 // check for special case 600 cmp64(rax, ExternalAddress((address) &min_long)); 601 jcc(Assembler::notEqual, normal_case); 602 xorl(rdx, rdx); // prepare rdx for possible special case (where 603 // remainder = 0) 604 cmpq(reg, -1); 605 jcc(Assembler::equal, special_case); 606 607 // handle normal case 608 bind(normal_case); 609 cdqq(); 610 int idivq_offset = offset(); 611 idivq(reg); 612 613 // normal and special case exit 614 bind(special_case); 615 616 return idivq_offset; 617 } 618 619 void MacroAssembler::decrementq(Register reg, int value) { 620 if (value == min_jint) { subq(reg, value); return; } 621 if (value < 0) { incrementq(reg, -value); return; } 622 if (value == 0) { ; return; } 623 if (value == 1 && UseIncDec) { decq(reg) ; return; } 624 /* else */ { subq(reg, value) ; return; } 625 } 626 627 void MacroAssembler::decrementq(Address dst, int value) { 628 if (value == min_jint) { subq(dst, value); return; } 629 if (value < 0) { incrementq(dst, -value); return; } 630 if (value == 0) { ; return; } 631 if (value == 1 && UseIncDec) { decq(dst) ; return; } 632 /* else */ { subq(dst, value) ; return; } 633 } 634 635 void MacroAssembler::incrementq(AddressLiteral dst) { 636 if (reachable(dst)) { 637 incrementq(as_Address(dst)); 638 } else { 639 lea(rscratch1, dst); 640 incrementq(Address(rscratch1, 0)); 641 } 642 } 643 644 void MacroAssembler::incrementq(Register reg, int value) { 645 if (value == min_jint) { addq(reg, value); return; } 646 if (value < 0) { decrementq(reg, -value); return; } 647 if (value == 0) { ; return; } 648 if (value == 1 && UseIncDec) { incq(reg) ; return; } 649 /* else */ { addq(reg, value) ; return; } 650 } 651 652 void MacroAssembler::incrementq(Address dst, int value) { 653 if (value == min_jint) { addq(dst, value); return; } 654 if (value < 0) { decrementq(dst, -value); return; } 655 if (value == 0) { ; return; } 656 if (value == 1 && UseIncDec) { incq(dst) ; return; } 657 /* else */ { addq(dst, value) ; return; } 658 } 659 660 // 32bit can do a case table jump in one instruction but we no longer allow the base 661 // to be installed in the Address class 662 void MacroAssembler::jump(ArrayAddress entry) { 663 lea(rscratch1, entry.base()); 664 Address dispatch = entry.index(); 665 assert(dispatch._base == noreg, "must be"); 666 dispatch._base = rscratch1; 667 jmp(dispatch); 668 } 669 670 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) { 671 ShouldNotReachHere(); // 64bit doesn't use two regs 672 cmpq(x_lo, y_lo); 673 } 674 675 void MacroAssembler::lea(Register dst, AddressLiteral src) { 676 mov_literal64(dst, (intptr_t)src.target(), src.rspec()); 677 } 678 679 void MacroAssembler::lea(Address dst, AddressLiteral adr) { 680 mov_literal64(rscratch1, (intptr_t)adr.target(), adr.rspec()); 681 movptr(dst, rscratch1); 682 } 683 684 void MacroAssembler::leave() { 685 // %%% is this really better? Why not on 32bit too? 686 emit_int8((unsigned char)0xC9); // LEAVE 687 } 688 689 void MacroAssembler::lneg(Register hi, Register lo) { 690 ShouldNotReachHere(); // 64bit doesn't use two regs 691 negq(lo); 692 } 693 694 void MacroAssembler::movoop(Register dst, jobject obj) { 695 mov_literal64(dst, (intptr_t)obj, oop_Relocation::spec_for_immediate()); 696 } 697 698 void MacroAssembler::movoop(Address dst, jobject obj) { 699 mov_literal64(rscratch1, (intptr_t)obj, oop_Relocation::spec_for_immediate()); 700 movq(dst, rscratch1); 701 } 702 703 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) { 704 mov_literal64(dst, (intptr_t)obj, metadata_Relocation::spec_for_immediate()); 705 } 706 707 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) { 708 mov_literal64(rscratch1, (intptr_t)obj, metadata_Relocation::spec_for_immediate()); 709 movq(dst, rscratch1); 710 } 711 712 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) { 713 if (src.is_lval()) { 714 mov_literal64(dst, (intptr_t)src.target(), src.rspec()); 715 } else { 716 if (reachable(src)) { 717 movq(dst, as_Address(src)); 718 } else { 719 lea(scratch, src); 720 movq(dst, Address(scratch, 0)); 721 } 722 } 723 } 724 725 void MacroAssembler::movptr(ArrayAddress dst, Register src) { 726 movq(as_Address(dst), src); 727 } 728 729 void MacroAssembler::movptr(Register dst, ArrayAddress src) { 730 movq(dst, as_Address(src)); 731 } 732 733 // src should NEVER be a real pointer. Use AddressLiteral for true pointers 734 void MacroAssembler::movptr(Address dst, intptr_t src) { 735 mov64(rscratch1, src); 736 movq(dst, rscratch1); 737 } 738 739 // These are mostly for initializing NULL 740 void MacroAssembler::movptr(Address dst, int32_t src) { 741 movslq(dst, src); 742 } 743 744 void MacroAssembler::movptr(Register dst, int32_t src) { 745 mov64(dst, (intptr_t)src); 746 } 747 748 void MacroAssembler::pushoop(jobject obj) { 749 movoop(rscratch1, obj); 750 push(rscratch1); 751 } 752 753 void MacroAssembler::pushklass(Metadata* obj) { 754 mov_metadata(rscratch1, obj); 755 push(rscratch1); 756 } 757 758 void MacroAssembler::pushptr(AddressLiteral src) { 759 lea(rscratch1, src); 760 if (src.is_lval()) { 761 push(rscratch1); 762 } else { 763 pushq(Address(rscratch1, 0)); 764 } 765 } 766 767 void MacroAssembler::reset_last_Java_frame(bool clear_fp) { 768 // we must set sp to zero to clear frame 769 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), NULL_WORD); 770 // must clear fp, so that compiled frames are not confused; it is 771 // possible that we need it only for debugging 772 if (clear_fp) { 773 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD); 774 } 775 776 // Always clear the pc because it could have been set by make_walkable() 777 movptr(Address(r15_thread, JavaThread::last_Java_pc_offset()), NULL_WORD); 778 vzeroupper(); 779 } 780 781 void MacroAssembler::set_last_Java_frame(Register last_java_sp, 782 Register last_java_fp, 783 address last_java_pc) { 784 vzeroupper(); 785 // determine last_java_sp register 786 if (!last_java_sp->is_valid()) { 787 last_java_sp = rsp; 788 } 789 790 // last_java_fp is optional 791 if (last_java_fp->is_valid()) { 792 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), 793 last_java_fp); 794 } 795 796 // last_java_pc is optional 797 if (last_java_pc != NULL) { 798 Address java_pc(r15_thread, 799 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()); 800 lea(rscratch1, InternalAddress(last_java_pc)); 801 movptr(java_pc, rscratch1); 802 } 803 804 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), last_java_sp); 805 } 806 807 static void pass_arg0(MacroAssembler* masm, Register arg) { 808 if (c_rarg0 != arg ) { 809 masm->mov(c_rarg0, arg); 810 } 811 } 812 813 static void pass_arg1(MacroAssembler* masm, Register arg) { 814 if (c_rarg1 != arg ) { 815 masm->mov(c_rarg1, arg); 816 } 817 } 818 819 static void pass_arg2(MacroAssembler* masm, Register arg) { 820 if (c_rarg2 != arg ) { 821 masm->mov(c_rarg2, arg); 822 } 823 } 824 825 static void pass_arg3(MacroAssembler* masm, Register arg) { 826 if (c_rarg3 != arg ) { 827 masm->mov(c_rarg3, arg); 828 } 829 } 830 831 void MacroAssembler::stop(const char* msg) { 832 address rip = pc(); 833 pusha(); // get regs on stack 834 lea(c_rarg0, ExternalAddress((address) msg)); 835 lea(c_rarg1, InternalAddress(rip)); 836 movq(c_rarg2, rsp); // pass pointer to regs array 837 andq(rsp, -16); // align stack as required by ABI 838 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64))); 839 hlt(); 840 } 841 842 void MacroAssembler::warn(const char* msg) { 843 push(rbp); 844 movq(rbp, rsp); 845 andq(rsp, -16); // align stack as required by push_CPU_state and call 846 push_CPU_state(); // keeps alignment at 16 bytes 847 lea(c_rarg0, ExternalAddress((address) msg)); 848 lea(rax, ExternalAddress(CAST_FROM_FN_PTR(address, warning))); 849 call(rax); 850 pop_CPU_state(); 851 mov(rsp, rbp); 852 pop(rbp); 853 } 854 855 void MacroAssembler::print_state() { 856 address rip = pc(); 857 pusha(); // get regs on stack 858 push(rbp); 859 movq(rbp, rsp); 860 andq(rsp, -16); // align stack as required by push_CPU_state and call 861 push_CPU_state(); // keeps alignment at 16 bytes 862 863 lea(c_rarg0, InternalAddress(rip)); 864 lea(c_rarg1, Address(rbp, wordSize)); // pass pointer to regs array 865 call_VM_leaf(CAST_FROM_FN_PTR(address, MacroAssembler::print_state64), c_rarg0, c_rarg1); 866 867 pop_CPU_state(); 868 mov(rsp, rbp); 869 pop(rbp); 870 popa(); 871 } 872 873 #ifndef PRODUCT 874 extern "C" void findpc(intptr_t x); 875 #endif 876 877 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[]) { 878 // In order to get locks to work, we need to fake a in_VM state 879 if (ShowMessageBoxOnError) { 880 JavaThread* thread = JavaThread::current(); 881 JavaThreadState saved_state = thread->thread_state(); 882 thread->set_thread_state(_thread_in_vm); 883 #ifndef PRODUCT 884 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) { 885 ttyLocker ttyl; 886 BytecodeCounter::print(); 887 } 888 #endif 889 // To see where a verify_oop failed, get $ebx+40/X for this frame. 890 // XXX correct this offset for amd64 891 // This is the value of eip which points to where verify_oop will return. 892 if (os::message_box(msg, "Execution stopped, print registers?")) { 893 print_state64(pc, regs); 894 BREAKPOINT; 895 assert(false, "start up GDB"); 896 } 897 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state); 898 } else { 899 ttyLocker ttyl; 900 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", 901 msg); 902 assert(false, "DEBUG MESSAGE: %s", msg); 903 } 904 } 905 906 void MacroAssembler::print_state64(int64_t pc, int64_t regs[]) { 907 ttyLocker ttyl; 908 FlagSetting fs(Debugging, true); 909 tty->print_cr("rip = 0x%016lx", (intptr_t)pc); 910 #ifndef PRODUCT 911 tty->cr(); 912 findpc(pc); 913 tty->cr(); 914 #endif 915 #define PRINT_REG(rax, value) \ 916 { tty->print("%s = ", #rax); os::print_location(tty, value); } 917 PRINT_REG(rax, regs[15]); 918 PRINT_REG(rbx, regs[12]); 919 PRINT_REG(rcx, regs[14]); 920 PRINT_REG(rdx, regs[13]); 921 PRINT_REG(rdi, regs[8]); 922 PRINT_REG(rsi, regs[9]); 923 PRINT_REG(rbp, regs[10]); 924 PRINT_REG(rsp, regs[11]); 925 PRINT_REG(r8 , regs[7]); 926 PRINT_REG(r9 , regs[6]); 927 PRINT_REG(r10, regs[5]); 928 PRINT_REG(r11, regs[4]); 929 PRINT_REG(r12, regs[3]); 930 PRINT_REG(r13, regs[2]); 931 PRINT_REG(r14, regs[1]); 932 PRINT_REG(r15, regs[0]); 933 #undef PRINT_REG 934 // Print some words near top of staack. 935 int64_t* rsp = (int64_t*) regs[11]; 936 int64_t* dump_sp = rsp; 937 for (int col1 = 0; col1 < 8; col1++) { 938 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 939 os::print_location(tty, *dump_sp++); 940 } 941 for (int row = 0; row < 25; row++) { 942 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); 943 for (int col = 0; col < 4; col++) { 944 tty->print(" 0x%016lx", (intptr_t)*dump_sp++); 945 } 946 tty->cr(); 947 } 948 // Print some instructions around pc: 949 Disassembler::decode((address)pc-64, (address)pc); 950 tty->print_cr("--------"); 951 Disassembler::decode((address)pc, (address)pc+32); 952 } 953 954 #endif // _LP64 955 956 // Now versions that are common to 32/64 bit 957 958 void MacroAssembler::addptr(Register dst, int32_t imm32) { 959 LP64_ONLY(addq(dst, imm32)) NOT_LP64(addl(dst, imm32)); 960 } 961 962 void MacroAssembler::addptr(Register dst, Register src) { 963 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src)); 964 } 965 966 void MacroAssembler::addptr(Address dst, Register src) { 967 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src)); 968 } 969 970 void MacroAssembler::addsd(XMMRegister dst, AddressLiteral src) { 971 if (reachable(src)) { 972 Assembler::addsd(dst, as_Address(src)); 973 } else { 974 lea(rscratch1, src); 975 Assembler::addsd(dst, Address(rscratch1, 0)); 976 } 977 } 978 979 void MacroAssembler::addss(XMMRegister dst, AddressLiteral src) { 980 if (reachable(src)) { 981 addss(dst, as_Address(src)); 982 } else { 983 lea(rscratch1, src); 984 addss(dst, Address(rscratch1, 0)); 985 } 986 } 987 988 void MacroAssembler::addpd(XMMRegister dst, AddressLiteral src) { 989 if (reachable(src)) { 990 Assembler::addpd(dst, as_Address(src)); 991 } else { 992 lea(rscratch1, src); 993 Assembler::addpd(dst, Address(rscratch1, 0)); 994 } 995 } 996 997 void MacroAssembler::align(int modulus) { 998 align(modulus, offset()); 999 } 1000 1001 void MacroAssembler::align(int modulus, int target) { 1002 if (target % modulus != 0) { 1003 nop(modulus - (target % modulus)); 1004 } 1005 } 1006 1007 void MacroAssembler::andpd(XMMRegister dst, AddressLiteral src) { 1008 // Used in sign-masking with aligned address. 1009 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 1010 if (reachable(src)) { 1011 Assembler::andpd(dst, as_Address(src)); 1012 } else { 1013 lea(rscratch1, src); 1014 Assembler::andpd(dst, Address(rscratch1, 0)); 1015 } 1016 } 1017 1018 void MacroAssembler::andps(XMMRegister dst, AddressLiteral src) { 1019 // Used in sign-masking with aligned address. 1020 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 1021 if (reachable(src)) { 1022 Assembler::andps(dst, as_Address(src)); 1023 } else { 1024 lea(rscratch1, src); 1025 Assembler::andps(dst, Address(rscratch1, 0)); 1026 } 1027 } 1028 1029 void MacroAssembler::andptr(Register dst, int32_t imm32) { 1030 LP64_ONLY(andq(dst, imm32)) NOT_LP64(andl(dst, imm32)); 1031 } 1032 1033 void MacroAssembler::atomic_incl(Address counter_addr) { 1034 lock(); 1035 incrementl(counter_addr); 1036 } 1037 1038 void MacroAssembler::atomic_incl(AddressLiteral counter_addr, Register scr) { 1039 if (reachable(counter_addr)) { 1040 atomic_incl(as_Address(counter_addr)); 1041 } else { 1042 lea(scr, counter_addr); 1043 atomic_incl(Address(scr, 0)); 1044 } 1045 } 1046 1047 #ifdef _LP64 1048 void MacroAssembler::atomic_incq(Address counter_addr) { 1049 lock(); 1050 incrementq(counter_addr); 1051 } 1052 1053 void MacroAssembler::atomic_incq(AddressLiteral counter_addr, Register scr) { 1054 if (reachable(counter_addr)) { 1055 atomic_incq(as_Address(counter_addr)); 1056 } else { 1057 lea(scr, counter_addr); 1058 atomic_incq(Address(scr, 0)); 1059 } 1060 } 1061 #endif 1062 1063 // Writes to stack successive pages until offset reached to check for 1064 // stack overflow + shadow pages. This clobbers tmp. 1065 void MacroAssembler::bang_stack_size(Register size, Register tmp) { 1066 movptr(tmp, rsp); 1067 // Bang stack for total size given plus shadow page size. 1068 // Bang one page at a time because large size can bang beyond yellow and 1069 // red zones. 1070 Label loop; 1071 bind(loop); 1072 movl(Address(tmp, (-os::vm_page_size())), size ); 1073 subptr(tmp, os::vm_page_size()); 1074 subl(size, os::vm_page_size()); 1075 jcc(Assembler::greater, loop); 1076 1077 // Bang down shadow pages too. 1078 // At this point, (tmp-0) is the last address touched, so don't 1079 // touch it again. (It was touched as (tmp-pagesize) but then tmp 1080 // was post-decremented.) Skip this address by starting at i=1, and 1081 // touch a few more pages below. N.B. It is important to touch all 1082 // the way down including all pages in the shadow zone. 1083 for (int i = 1; i < ((int)JavaThread::stack_shadow_zone_size() / os::vm_page_size()); i++) { 1084 // this could be any sized move but this is can be a debugging crumb 1085 // so the bigger the better. 1086 movptr(Address(tmp, (-i*os::vm_page_size())), size ); 1087 } 1088 } 1089 1090 void MacroAssembler::reserved_stack_check() { 1091 // testing if reserved zone needs to be enabled 1092 Label no_reserved_zone_enabling; 1093 Register thread = NOT_LP64(rsi) LP64_ONLY(r15_thread); 1094 NOT_LP64(get_thread(rsi);) 1095 1096 cmpptr(rsp, Address(thread, JavaThread::reserved_stack_activation_offset())); 1097 jcc(Assembler::below, no_reserved_zone_enabling); 1098 1099 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), thread); 1100 jump(RuntimeAddress(StubRoutines::throw_delayed_StackOverflowError_entry())); 1101 should_not_reach_here(); 1102 1103 bind(no_reserved_zone_enabling); 1104 } 1105 1106 int MacroAssembler::biased_locking_enter(Register lock_reg, 1107 Register obj_reg, 1108 Register swap_reg, 1109 Register tmp_reg, 1110 bool swap_reg_contains_mark, 1111 Label& done, 1112 Label* slow_case, 1113 BiasedLockingCounters* counters) { 1114 assert(UseBiasedLocking, "why call this otherwise?"); 1115 assert(swap_reg == rax, "swap_reg must be rax for cmpxchgq"); 1116 assert(tmp_reg != noreg, "tmp_reg must be supplied"); 1117 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg); 1118 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout"); 1119 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes()); 1120 NOT_LP64( Address saved_mark_addr(lock_reg, 0); ) 1121 1122 if (PrintBiasedLockingStatistics && counters == NULL) { 1123 counters = BiasedLocking::counters(); 1124 } 1125 // Biased locking 1126 // See whether the lock is currently biased toward our thread and 1127 // whether the epoch is still valid 1128 // Note that the runtime guarantees sufficient alignment of JavaThread 1129 // pointers to allow age to be placed into low bits 1130 // First check to see whether biasing is even enabled for this object 1131 Label cas_label; 1132 int null_check_offset = -1; 1133 if (!swap_reg_contains_mark) { 1134 null_check_offset = offset(); 1135 movptr(swap_reg, mark_addr); 1136 } 1137 movptr(tmp_reg, swap_reg); 1138 andptr(tmp_reg, markOopDesc::biased_lock_mask_in_place); 1139 cmpptr(tmp_reg, markOopDesc::biased_lock_pattern); 1140 jcc(Assembler::notEqual, cas_label); 1141 // The bias pattern is present in the object's header. Need to check 1142 // whether the bias owner and the epoch are both still current. 1143 #ifndef _LP64 1144 // Note that because there is no current thread register on x86_32 we 1145 // need to store off the mark word we read out of the object to 1146 // avoid reloading it and needing to recheck invariants below. This 1147 // store is unfortunate but it makes the overall code shorter and 1148 // simpler. 1149 movptr(saved_mark_addr, swap_reg); 1150 #endif 1151 if (swap_reg_contains_mark) { 1152 null_check_offset = offset(); 1153 } 1154 load_prototype_header(tmp_reg, obj_reg); 1155 #ifdef _LP64 1156 orptr(tmp_reg, r15_thread); 1157 xorptr(tmp_reg, swap_reg); 1158 Register header_reg = tmp_reg; 1159 #else 1160 xorptr(tmp_reg, swap_reg); 1161 get_thread(swap_reg); 1162 xorptr(swap_reg, tmp_reg); 1163 Register header_reg = swap_reg; 1164 #endif 1165 andptr(header_reg, ~((int) markOopDesc::age_mask_in_place)); 1166 if (counters != NULL) { 1167 cond_inc32(Assembler::zero, 1168 ExternalAddress((address) counters->biased_lock_entry_count_addr())); 1169 } 1170 jcc(Assembler::equal, done); 1171 1172 Label try_revoke_bias; 1173 Label try_rebias; 1174 1175 // At this point we know that the header has the bias pattern and 1176 // that we are not the bias owner in the current epoch. We need to 1177 // figure out more details about the state of the header in order to 1178 // know what operations can be legally performed on the object's 1179 // header. 1180 1181 // If the low three bits in the xor result aren't clear, that means 1182 // the prototype header is no longer biased and we have to revoke 1183 // the bias on this object. 1184 testptr(header_reg, markOopDesc::biased_lock_mask_in_place); 1185 jccb(Assembler::notZero, try_revoke_bias); 1186 1187 // Biasing is still enabled for this data type. See whether the 1188 // epoch of the current bias is still valid, meaning that the epoch 1189 // bits of the mark word are equal to the epoch bits of the 1190 // prototype header. (Note that the prototype header's epoch bits 1191 // only change at a safepoint.) If not, attempt to rebias the object 1192 // toward the current thread. Note that we must be absolutely sure 1193 // that the current epoch is invalid in order to do this because 1194 // otherwise the manipulations it performs on the mark word are 1195 // illegal. 1196 testptr(header_reg, markOopDesc::epoch_mask_in_place); 1197 jccb(Assembler::notZero, try_rebias); 1198 1199 // The epoch of the current bias is still valid but we know nothing 1200 // about the owner; it might be set or it might be clear. Try to 1201 // acquire the bias of the object using an atomic operation. If this 1202 // fails we will go in to the runtime to revoke the object's bias. 1203 // Note that we first construct the presumed unbiased header so we 1204 // don't accidentally blow away another thread's valid bias. 1205 NOT_LP64( movptr(swap_reg, saved_mark_addr); ) 1206 andptr(swap_reg, 1207 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place); 1208 #ifdef _LP64 1209 movptr(tmp_reg, swap_reg); 1210 orptr(tmp_reg, r15_thread); 1211 #else 1212 get_thread(tmp_reg); 1213 orptr(tmp_reg, swap_reg); 1214 #endif 1215 lock(); 1216 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1217 // If the biasing toward our thread failed, this means that 1218 // another thread succeeded in biasing it toward itself and we 1219 // need to revoke that bias. The revocation will occur in the 1220 // interpreter runtime in the slow case. 1221 if (counters != NULL) { 1222 cond_inc32(Assembler::zero, 1223 ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr())); 1224 } 1225 if (slow_case != NULL) { 1226 jcc(Assembler::notZero, *slow_case); 1227 } 1228 jmp(done); 1229 1230 bind(try_rebias); 1231 // At this point we know the epoch has expired, meaning that the 1232 // current "bias owner", if any, is actually invalid. Under these 1233 // circumstances _only_, we are allowed to use the current header's 1234 // value as the comparison value when doing the cas to acquire the 1235 // bias in the current epoch. In other words, we allow transfer of 1236 // the bias from one thread to another directly in this situation. 1237 // 1238 // FIXME: due to a lack of registers we currently blow away the age 1239 // bits in this situation. Should attempt to preserve them. 1240 load_prototype_header(tmp_reg, obj_reg); 1241 #ifdef _LP64 1242 orptr(tmp_reg, r15_thread); 1243 #else 1244 get_thread(swap_reg); 1245 orptr(tmp_reg, swap_reg); 1246 movptr(swap_reg, saved_mark_addr); 1247 #endif 1248 lock(); 1249 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1250 // If the biasing toward our thread failed, then another thread 1251 // succeeded in biasing it toward itself and we need to revoke that 1252 // bias. The revocation will occur in the runtime in the slow case. 1253 if (counters != NULL) { 1254 cond_inc32(Assembler::zero, 1255 ExternalAddress((address) counters->rebiased_lock_entry_count_addr())); 1256 } 1257 if (slow_case != NULL) { 1258 jcc(Assembler::notZero, *slow_case); 1259 } 1260 jmp(done); 1261 1262 bind(try_revoke_bias); 1263 // The prototype mark in the klass doesn't have the bias bit set any 1264 // more, indicating that objects of this data type are not supposed 1265 // to be biased any more. We are going to try to reset the mark of 1266 // this object to the prototype value and fall through to the 1267 // CAS-based locking scheme. Note that if our CAS fails, it means 1268 // that another thread raced us for the privilege of revoking the 1269 // bias of this particular object, so it's okay to continue in the 1270 // normal locking code. 1271 // 1272 // FIXME: due to a lack of registers we currently blow away the age 1273 // bits in this situation. Should attempt to preserve them. 1274 NOT_LP64( movptr(swap_reg, saved_mark_addr); ) 1275 load_prototype_header(tmp_reg, obj_reg); 1276 lock(); 1277 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg 1278 // Fall through to the normal CAS-based lock, because no matter what 1279 // the result of the above CAS, some thread must have succeeded in 1280 // removing the bias bit from the object's header. 1281 if (counters != NULL) { 1282 cond_inc32(Assembler::zero, 1283 ExternalAddress((address) counters->revoked_lock_entry_count_addr())); 1284 } 1285 1286 bind(cas_label); 1287 1288 return null_check_offset; 1289 } 1290 1291 void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) { 1292 assert(UseBiasedLocking, "why call this otherwise?"); 1293 1294 // Check for biased locking unlock case, which is a no-op 1295 // Note: we do not have to check the thread ID for two reasons. 1296 // First, the interpreter checks for IllegalMonitorStateException at 1297 // a higher level. Second, if the bias was revoked while we held the 1298 // lock, the object could not be rebiased toward another thread, so 1299 // the bias bit would be clear. 1300 movptr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 1301 andptr(temp_reg, markOopDesc::biased_lock_mask_in_place); 1302 cmpptr(temp_reg, markOopDesc::biased_lock_pattern); 1303 jcc(Assembler::equal, done); 1304 } 1305 1306 #ifdef COMPILER2 1307 1308 #if INCLUDE_RTM_OPT 1309 1310 // Update rtm_counters based on abort status 1311 // input: abort_status 1312 // rtm_counters (RTMLockingCounters*) 1313 // flags are killed 1314 void MacroAssembler::rtm_counters_update(Register abort_status, Register rtm_counters) { 1315 1316 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abort_count_offset())); 1317 if (PrintPreciseRTMLockingStatistics) { 1318 for (int i = 0; i < RTMLockingCounters::ABORT_STATUS_LIMIT; i++) { 1319 Label check_abort; 1320 testl(abort_status, (1<<i)); 1321 jccb(Assembler::equal, check_abort); 1322 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abortX_count_offset() + (i * sizeof(uintx)))); 1323 bind(check_abort); 1324 } 1325 } 1326 } 1327 1328 // Branch if (random & (count-1) != 0), count is 2^n 1329 // tmp, scr and flags are killed 1330 void MacroAssembler::branch_on_random_using_rdtsc(Register tmp, Register scr, int count, Label& brLabel) { 1331 assert(tmp == rax, ""); 1332 assert(scr == rdx, ""); 1333 rdtsc(); // modifies EDX:EAX 1334 andptr(tmp, count-1); 1335 jccb(Assembler::notZero, brLabel); 1336 } 1337 1338 // Perform abort ratio calculation, set no_rtm bit if high ratio 1339 // input: rtm_counters_Reg (RTMLockingCounters* address) 1340 // tmpReg, rtm_counters_Reg and flags are killed 1341 void MacroAssembler::rtm_abort_ratio_calculation(Register tmpReg, 1342 Register rtm_counters_Reg, 1343 RTMLockingCounters* rtm_counters, 1344 Metadata* method_data) { 1345 Label L_done, L_check_always_rtm1, L_check_always_rtm2; 1346 1347 if (RTMLockingCalculationDelay > 0) { 1348 // Delay calculation 1349 movptr(tmpReg, ExternalAddress((address) RTMLockingCounters::rtm_calculation_flag_addr()), tmpReg); 1350 testptr(tmpReg, tmpReg); 1351 jccb(Assembler::equal, L_done); 1352 } 1353 // Abort ratio calculation only if abort_count > RTMAbortThreshold 1354 // Aborted transactions = abort_count * 100 1355 // All transactions = total_count * RTMTotalCountIncrRate 1356 // Set no_rtm bit if (Aborted transactions >= All transactions * RTMAbortRatio) 1357 1358 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::abort_count_offset())); 1359 cmpptr(tmpReg, RTMAbortThreshold); 1360 jccb(Assembler::below, L_check_always_rtm2); 1361 imulptr(tmpReg, tmpReg, 100); 1362 1363 Register scrReg = rtm_counters_Reg; 1364 movptr(scrReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset())); 1365 imulptr(scrReg, scrReg, RTMTotalCountIncrRate); 1366 imulptr(scrReg, scrReg, RTMAbortRatio); 1367 cmpptr(tmpReg, scrReg); 1368 jccb(Assembler::below, L_check_always_rtm1); 1369 if (method_data != NULL) { 1370 // set rtm_state to "no rtm" in MDO 1371 mov_metadata(tmpReg, method_data); 1372 lock(); 1373 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), NoRTM); 1374 } 1375 jmpb(L_done); 1376 bind(L_check_always_rtm1); 1377 // Reload RTMLockingCounters* address 1378 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters)); 1379 bind(L_check_always_rtm2); 1380 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset())); 1381 cmpptr(tmpReg, RTMLockingThreshold / RTMTotalCountIncrRate); 1382 jccb(Assembler::below, L_done); 1383 if (method_data != NULL) { 1384 // set rtm_state to "always rtm" in MDO 1385 mov_metadata(tmpReg, method_data); 1386 lock(); 1387 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), UseRTM); 1388 } 1389 bind(L_done); 1390 } 1391 1392 // Update counters and perform abort ratio calculation 1393 // input: abort_status_Reg 1394 // rtm_counters_Reg, flags are killed 1395 void MacroAssembler::rtm_profiling(Register abort_status_Reg, 1396 Register rtm_counters_Reg, 1397 RTMLockingCounters* rtm_counters, 1398 Metadata* method_data, 1399 bool profile_rtm) { 1400 1401 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1402 // update rtm counters based on rax value at abort 1403 // reads abort_status_Reg, updates flags 1404 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters)); 1405 rtm_counters_update(abort_status_Reg, rtm_counters_Reg); 1406 if (profile_rtm) { 1407 // Save abort status because abort_status_Reg is used by following code. 1408 if (RTMRetryCount > 0) { 1409 push(abort_status_Reg); 1410 } 1411 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1412 rtm_abort_ratio_calculation(abort_status_Reg, rtm_counters_Reg, rtm_counters, method_data); 1413 // restore abort status 1414 if (RTMRetryCount > 0) { 1415 pop(abort_status_Reg); 1416 } 1417 } 1418 } 1419 1420 // Retry on abort if abort's status is 0x6: can retry (0x2) | memory conflict (0x4) 1421 // inputs: retry_count_Reg 1422 // : abort_status_Reg 1423 // output: retry_count_Reg decremented by 1 1424 // flags are killed 1425 void MacroAssembler::rtm_retry_lock_on_abort(Register retry_count_Reg, Register abort_status_Reg, Label& retryLabel) { 1426 Label doneRetry; 1427 assert(abort_status_Reg == rax, ""); 1428 // The abort reason bits are in eax (see all states in rtmLocking.hpp) 1429 // 0x6 = conflict on which we can retry (0x2) | memory conflict (0x4) 1430 // if reason is in 0x6 and retry count != 0 then retry 1431 andptr(abort_status_Reg, 0x6); 1432 jccb(Assembler::zero, doneRetry); 1433 testl(retry_count_Reg, retry_count_Reg); 1434 jccb(Assembler::zero, doneRetry); 1435 pause(); 1436 decrementl(retry_count_Reg); 1437 jmp(retryLabel); 1438 bind(doneRetry); 1439 } 1440 1441 // Spin and retry if lock is busy, 1442 // inputs: box_Reg (monitor address) 1443 // : retry_count_Reg 1444 // output: retry_count_Reg decremented by 1 1445 // : clear z flag if retry count exceeded 1446 // tmp_Reg, scr_Reg, flags are killed 1447 void MacroAssembler::rtm_retry_lock_on_busy(Register retry_count_Reg, Register box_Reg, 1448 Register tmp_Reg, Register scr_Reg, Label& retryLabel) { 1449 Label SpinLoop, SpinExit, doneRetry; 1450 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1451 1452 testl(retry_count_Reg, retry_count_Reg); 1453 jccb(Assembler::zero, doneRetry); 1454 decrementl(retry_count_Reg); 1455 movptr(scr_Reg, RTMSpinLoopCount); 1456 1457 bind(SpinLoop); 1458 pause(); 1459 decrementl(scr_Reg); 1460 jccb(Assembler::lessEqual, SpinExit); 1461 movptr(tmp_Reg, Address(box_Reg, owner_offset)); 1462 testptr(tmp_Reg, tmp_Reg); 1463 jccb(Assembler::notZero, SpinLoop); 1464 1465 bind(SpinExit); 1466 jmp(retryLabel); 1467 bind(doneRetry); 1468 incrementl(retry_count_Reg); // clear z flag 1469 } 1470 1471 // Use RTM for normal stack locks 1472 // Input: objReg (object to lock) 1473 void MacroAssembler::rtm_stack_locking(Register objReg, Register tmpReg, Register scrReg, 1474 Register retry_on_abort_count_Reg, 1475 RTMLockingCounters* stack_rtm_counters, 1476 Metadata* method_data, bool profile_rtm, 1477 Label& DONE_LABEL, Label& IsInflated) { 1478 assert(UseRTMForStackLocks, "why call this otherwise?"); 1479 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking"); 1480 assert(tmpReg == rax, ""); 1481 assert(scrReg == rdx, ""); 1482 Label L_rtm_retry, L_decrement_retry, L_on_abort; 1483 1484 if (RTMRetryCount > 0) { 1485 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort 1486 bind(L_rtm_retry); 1487 } 1488 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); 1489 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased 1490 jcc(Assembler::notZero, IsInflated); 1491 1492 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1493 Label L_noincrement; 1494 if (RTMTotalCountIncrRate > 1) { 1495 // tmpReg, scrReg and flags are killed 1496 branch_on_random_using_rdtsc(tmpReg, scrReg, RTMTotalCountIncrRate, L_noincrement); 1497 } 1498 assert(stack_rtm_counters != NULL, "should not be NULL when profiling RTM"); 1499 atomic_incptr(ExternalAddress((address)stack_rtm_counters->total_count_addr()), scrReg); 1500 bind(L_noincrement); 1501 } 1502 xbegin(L_on_abort); 1503 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // fetch markword 1504 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits 1505 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked 1506 jcc(Assembler::equal, DONE_LABEL); // all done if unlocked 1507 1508 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX 1509 if (UseRTMXendForLockBusy) { 1510 xend(); 1511 movptr(abort_status_Reg, 0x2); // Set the abort status to 2 (so we can retry) 1512 jmp(L_decrement_retry); 1513 } 1514 else { 1515 xabort(0); 1516 } 1517 bind(L_on_abort); 1518 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1519 rtm_profiling(abort_status_Reg, scrReg, stack_rtm_counters, method_data, profile_rtm); 1520 } 1521 bind(L_decrement_retry); 1522 if (RTMRetryCount > 0) { 1523 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4) 1524 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry); 1525 } 1526 } 1527 1528 // Use RTM for inflating locks 1529 // inputs: objReg (object to lock) 1530 // boxReg (on-stack box address (displaced header location) - KILLED) 1531 // tmpReg (ObjectMonitor address + markOopDesc::monitor_value) 1532 void MacroAssembler::rtm_inflated_locking(Register objReg, Register boxReg, Register tmpReg, 1533 Register scrReg, Register retry_on_busy_count_Reg, 1534 Register retry_on_abort_count_Reg, 1535 RTMLockingCounters* rtm_counters, 1536 Metadata* method_data, bool profile_rtm, 1537 Label& DONE_LABEL) { 1538 assert(UseRTMLocking, "why call this otherwise?"); 1539 assert(tmpReg == rax, ""); 1540 assert(scrReg == rdx, ""); 1541 Label L_rtm_retry, L_decrement_retry, L_on_abort; 1542 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1543 1544 // Without cast to int32_t a movptr will destroy r10 which is typically obj 1545 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark())); 1546 movptr(boxReg, tmpReg); // Save ObjectMonitor address 1547 1548 if (RTMRetryCount > 0) { 1549 movl(retry_on_busy_count_Reg, RTMRetryCount); // Retry on lock busy 1550 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort 1551 bind(L_rtm_retry); 1552 } 1553 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1554 Label L_noincrement; 1555 if (RTMTotalCountIncrRate > 1) { 1556 // tmpReg, scrReg and flags are killed 1557 branch_on_random_using_rdtsc(tmpReg, scrReg, RTMTotalCountIncrRate, L_noincrement); 1558 } 1559 assert(rtm_counters != NULL, "should not be NULL when profiling RTM"); 1560 atomic_incptr(ExternalAddress((address)rtm_counters->total_count_addr()), scrReg); 1561 bind(L_noincrement); 1562 } 1563 xbegin(L_on_abort); 1564 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); 1565 movptr(tmpReg, Address(tmpReg, owner_offset)); 1566 testptr(tmpReg, tmpReg); 1567 jcc(Assembler::zero, DONE_LABEL); 1568 if (UseRTMXendForLockBusy) { 1569 xend(); 1570 jmp(L_decrement_retry); 1571 } 1572 else { 1573 xabort(0); 1574 } 1575 bind(L_on_abort); 1576 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX 1577 if (PrintPreciseRTMLockingStatistics || profile_rtm) { 1578 rtm_profiling(abort_status_Reg, scrReg, rtm_counters, method_data, profile_rtm); 1579 } 1580 if (RTMRetryCount > 0) { 1581 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4) 1582 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry); 1583 } 1584 1585 movptr(tmpReg, Address(boxReg, owner_offset)) ; 1586 testptr(tmpReg, tmpReg) ; 1587 jccb(Assembler::notZero, L_decrement_retry) ; 1588 1589 // Appears unlocked - try to swing _owner from null to non-null. 1590 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1591 #ifdef _LP64 1592 Register threadReg = r15_thread; 1593 #else 1594 get_thread(scrReg); 1595 Register threadReg = scrReg; 1596 #endif 1597 lock(); 1598 cmpxchgptr(threadReg, Address(boxReg, owner_offset)); // Updates tmpReg 1599 1600 if (RTMRetryCount > 0) { 1601 // success done else retry 1602 jccb(Assembler::equal, DONE_LABEL) ; 1603 bind(L_decrement_retry); 1604 // Spin and retry if lock is busy. 1605 rtm_retry_lock_on_busy(retry_on_busy_count_Reg, boxReg, tmpReg, scrReg, L_rtm_retry); 1606 } 1607 else { 1608 bind(L_decrement_retry); 1609 } 1610 } 1611 1612 #endif // INCLUDE_RTM_OPT 1613 1614 // Fast_Lock and Fast_Unlock used by C2 1615 1616 // Because the transitions from emitted code to the runtime 1617 // monitorenter/exit helper stubs are so slow it's critical that 1618 // we inline both the stack-locking fast-path and the inflated fast path. 1619 // 1620 // See also: cmpFastLock and cmpFastUnlock. 1621 // 1622 // What follows is a specialized inline transliteration of the code 1623 // in slow_enter() and slow_exit(). If we're concerned about I$ bloat 1624 // another option would be to emit TrySlowEnter and TrySlowExit methods 1625 // at startup-time. These methods would accept arguments as 1626 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure 1627 // indications in the icc.ZFlag. Fast_Lock and Fast_Unlock would simply 1628 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit. 1629 // In practice, however, the # of lock sites is bounded and is usually small. 1630 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer 1631 // if the processor uses simple bimodal branch predictors keyed by EIP 1632 // Since the helper routines would be called from multiple synchronization 1633 // sites. 1634 // 1635 // An even better approach would be write "MonitorEnter()" and "MonitorExit()" 1636 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites 1637 // to those specialized methods. That'd give us a mostly platform-independent 1638 // implementation that the JITs could optimize and inline at their pleasure. 1639 // Done correctly, the only time we'd need to cross to native could would be 1640 // to park() or unpark() threads. We'd also need a few more unsafe operators 1641 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and 1642 // (b) explicit barriers or fence operations. 1643 // 1644 // TODO: 1645 // 1646 // * Arrange for C2 to pass "Self" into Fast_Lock and Fast_Unlock in one of the registers (scr). 1647 // This avoids manifesting the Self pointer in the Fast_Lock and Fast_Unlock terminals. 1648 // Given TLAB allocation, Self is usually manifested in a register, so passing it into 1649 // the lock operators would typically be faster than reifying Self. 1650 // 1651 // * Ideally I'd define the primitives as: 1652 // fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED. 1653 // fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED 1654 // Unfortunately ADLC bugs prevent us from expressing the ideal form. 1655 // Instead, we're stuck with a rather awkward and brittle register assignments below. 1656 // Furthermore the register assignments are overconstrained, possibly resulting in 1657 // sub-optimal code near the synchronization site. 1658 // 1659 // * Eliminate the sp-proximity tests and just use "== Self" tests instead. 1660 // Alternately, use a better sp-proximity test. 1661 // 1662 // * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value. 1663 // Either one is sufficient to uniquely identify a thread. 1664 // TODO: eliminate use of sp in _owner and use get_thread(tr) instead. 1665 // 1666 // * Intrinsify notify() and notifyAll() for the common cases where the 1667 // object is locked by the calling thread but the waitlist is empty. 1668 // avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll(). 1669 // 1670 // * use jccb and jmpb instead of jcc and jmp to improve code density. 1671 // But beware of excessive branch density on AMD Opterons. 1672 // 1673 // * Both Fast_Lock and Fast_Unlock set the ICC.ZF to indicate success 1674 // or failure of the fast-path. If the fast-path fails then we pass 1675 // control to the slow-path, typically in C. In Fast_Lock and 1676 // Fast_Unlock we often branch to DONE_LABEL, just to find that C2 1677 // will emit a conditional branch immediately after the node. 1678 // So we have branches to branches and lots of ICC.ZF games. 1679 // Instead, it might be better to have C2 pass a "FailureLabel" 1680 // into Fast_Lock and Fast_Unlock. In the case of success, control 1681 // will drop through the node. ICC.ZF is undefined at exit. 1682 // In the case of failure, the node will branch directly to the 1683 // FailureLabel 1684 1685 1686 // obj: object to lock 1687 // box: on-stack box address (displaced header location) - KILLED 1688 // rax,: tmp -- KILLED 1689 // scr: tmp -- KILLED 1690 void MacroAssembler::fast_lock(Register objReg, Register boxReg, Register tmpReg, 1691 Register scrReg, Register cx1Reg, Register cx2Reg, 1692 BiasedLockingCounters* counters, 1693 RTMLockingCounters* rtm_counters, 1694 RTMLockingCounters* stack_rtm_counters, 1695 Metadata* method_data, 1696 bool use_rtm, bool profile_rtm) { 1697 // Ensure the register assignments are disjoint 1698 assert(tmpReg == rax, ""); 1699 1700 if (use_rtm) { 1701 assert_different_registers(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg); 1702 } else { 1703 assert(cx1Reg == noreg, ""); 1704 assert(cx2Reg == noreg, ""); 1705 assert_different_registers(objReg, boxReg, tmpReg, scrReg); 1706 } 1707 1708 if (counters != NULL) { 1709 atomic_incl(ExternalAddress((address)counters->total_entry_count_addr()), scrReg); 1710 } 1711 1712 // Possible cases that we'll encounter in fast_lock 1713 // ------------------------------------------------ 1714 // * Inflated 1715 // -- unlocked 1716 // -- Locked 1717 // = by self 1718 // = by other 1719 // * biased 1720 // -- by Self 1721 // -- by other 1722 // * neutral 1723 // * stack-locked 1724 // -- by self 1725 // = sp-proximity test hits 1726 // = sp-proximity test generates false-negative 1727 // -- by other 1728 // 1729 1730 Label IsInflated, DONE_LABEL; 1731 1732 // it's stack-locked, biased or neutral 1733 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage 1734 // order to reduce the number of conditional branches in the most common cases. 1735 // Beware -- there's a subtle invariant that fetch of the markword 1736 // at [FETCH], below, will never observe a biased encoding (*101b). 1737 // If this invariant is not held we risk exclusion (safety) failure. 1738 if (UseBiasedLocking && !UseOptoBiasInlining) { 1739 biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, counters); 1740 } 1741 1742 #if INCLUDE_RTM_OPT 1743 if (UseRTMForStackLocks && use_rtm) { 1744 rtm_stack_locking(objReg, tmpReg, scrReg, cx2Reg, 1745 stack_rtm_counters, method_data, profile_rtm, 1746 DONE_LABEL, IsInflated); 1747 } 1748 #endif // INCLUDE_RTM_OPT 1749 1750 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // [FETCH] 1751 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased 1752 jccb(Assembler::notZero, IsInflated); 1753 1754 // Attempt stack-locking ... 1755 orptr (tmpReg, markOopDesc::unlocked_value); 1756 movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS 1757 lock(); 1758 cmpxchgptr(boxReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Updates tmpReg 1759 if (counters != NULL) { 1760 cond_inc32(Assembler::equal, 1761 ExternalAddress((address)counters->fast_path_entry_count_addr())); 1762 } 1763 jcc(Assembler::equal, DONE_LABEL); // Success 1764 1765 // Recursive locking. 1766 // The object is stack-locked: markword contains stack pointer to BasicLock. 1767 // Locked by current thread if difference with current SP is less than one page. 1768 subptr(tmpReg, rsp); 1769 // Next instruction set ZFlag == 1 (Success) if difference is less then one page. 1770 andptr(tmpReg, (int32_t) (NOT_LP64(0xFFFFF003) LP64_ONLY(7 - os::vm_page_size())) ); 1771 movptr(Address(boxReg, 0), tmpReg); 1772 if (counters != NULL) { 1773 cond_inc32(Assembler::equal, 1774 ExternalAddress((address)counters->fast_path_entry_count_addr())); 1775 } 1776 jmp(DONE_LABEL); 1777 1778 bind(IsInflated); 1779 // The object is inflated. tmpReg contains pointer to ObjectMonitor* + markOopDesc::monitor_value 1780 1781 #if INCLUDE_RTM_OPT 1782 // Use the same RTM locking code in 32- and 64-bit VM. 1783 if (use_rtm) { 1784 rtm_inflated_locking(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg, 1785 rtm_counters, method_data, profile_rtm, DONE_LABEL); 1786 } else { 1787 #endif // INCLUDE_RTM_OPT 1788 1789 #ifndef _LP64 1790 // The object is inflated. 1791 1792 // boxReg refers to the on-stack BasicLock in the current frame. 1793 // We'd like to write: 1794 // set box->_displaced_header = markOopDesc::unused_mark(). Any non-0 value suffices. 1795 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers 1796 // additional latency as we have another ST in the store buffer that must drain. 1797 1798 // avoid ST-before-CAS 1799 // register juggle because we need tmpReg for cmpxchgptr below 1800 movptr(scrReg, boxReg); 1801 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2] 1802 1803 // Optimistic form: consider XORL tmpReg,tmpReg 1804 movptr(tmpReg, NULL_WORD); 1805 1806 // Appears unlocked - try to swing _owner from null to non-null. 1807 // Ideally, I'd manifest "Self" with get_thread and then attempt 1808 // to CAS the register containing Self into m->Owner. 1809 // But we don't have enough registers, so instead we can either try to CAS 1810 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds 1811 // we later store "Self" into m->Owner. Transiently storing a stack address 1812 // (rsp or the address of the box) into m->owner is harmless. 1813 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand. 1814 lock(); 1815 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1816 movptr(Address(scrReg, 0), 3); // box->_displaced_header = 3 1817 // If we weren't able to swing _owner from NULL to the BasicLock 1818 // then take the slow path. 1819 jccb (Assembler::notZero, DONE_LABEL); 1820 // update _owner from BasicLock to thread 1821 get_thread (scrReg); // beware: clobbers ICCs 1822 movptr(Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), scrReg); 1823 xorptr(boxReg, boxReg); // set icc.ZFlag = 1 to indicate success 1824 1825 // If the CAS fails we can either retry or pass control to the slow-path. 1826 // We use the latter tactic. 1827 // Pass the CAS result in the icc.ZFlag into DONE_LABEL 1828 // If the CAS was successful ... 1829 // Self has acquired the lock 1830 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it. 1831 // Intentional fall-through into DONE_LABEL ... 1832 #else // _LP64 1833 // It's inflated 1834 movq(scrReg, tmpReg); 1835 xorq(tmpReg, tmpReg); 1836 1837 lock(); 1838 cmpxchgptr(r15_thread, Address(scrReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 1839 // Unconditionally set box->_displaced_header = markOopDesc::unused_mark(). 1840 // Without cast to int32_t movptr will destroy r10 which is typically obj. 1841 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark())); 1842 // Intentional fall-through into DONE_LABEL ... 1843 // Propagate ICC.ZF from CAS above into DONE_LABEL. 1844 #endif // _LP64 1845 #if INCLUDE_RTM_OPT 1846 } // use_rtm() 1847 #endif 1848 // DONE_LABEL is a hot target - we'd really like to place it at the 1849 // start of cache line by padding with NOPs. 1850 // See the AMD and Intel software optimization manuals for the 1851 // most efficient "long" NOP encodings. 1852 // Unfortunately none of our alignment mechanisms suffice. 1853 bind(DONE_LABEL); 1854 1855 // At DONE_LABEL the icc ZFlag is set as follows ... 1856 // Fast_Unlock uses the same protocol. 1857 // ZFlag == 1 -> Success 1858 // ZFlag == 0 -> Failure - force control through the slow-path 1859 } 1860 1861 // obj: object to unlock 1862 // box: box address (displaced header location), killed. Must be EAX. 1863 // tmp: killed, cannot be obj nor box. 1864 // 1865 // Some commentary on balanced locking: 1866 // 1867 // Fast_Lock and Fast_Unlock are emitted only for provably balanced lock sites. 1868 // Methods that don't have provably balanced locking are forced to run in the 1869 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock. 1870 // The interpreter provides two properties: 1871 // I1: At return-time the interpreter automatically and quietly unlocks any 1872 // objects acquired the current activation (frame). Recall that the 1873 // interpreter maintains an on-stack list of locks currently held by 1874 // a frame. 1875 // I2: If a method attempts to unlock an object that is not held by the 1876 // the frame the interpreter throws IMSX. 1877 // 1878 // Lets say A(), which has provably balanced locking, acquires O and then calls B(). 1879 // B() doesn't have provably balanced locking so it runs in the interpreter. 1880 // Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O 1881 // is still locked by A(). 1882 // 1883 // The only other source of unbalanced locking would be JNI. The "Java Native Interface: 1884 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter 1885 // should not be unlocked by "normal" java-level locking and vice-versa. The specification 1886 // doesn't specify what will occur if a program engages in such mixed-mode locking, however. 1887 // Arguably given that the spec legislates the JNI case as undefined our implementation 1888 // could reasonably *avoid* checking owner in Fast_Unlock(). 1889 // In the interest of performance we elide m->Owner==Self check in unlock. 1890 // A perfectly viable alternative is to elide the owner check except when 1891 // Xcheck:jni is enabled. 1892 1893 void MacroAssembler::fast_unlock(Register objReg, Register boxReg, Register tmpReg, bool use_rtm) { 1894 assert(boxReg == rax, ""); 1895 assert_different_registers(objReg, boxReg, tmpReg); 1896 1897 Label DONE_LABEL, Stacked, CheckSucc; 1898 1899 // Critically, the biased locking test must have precedence over 1900 // and appear before the (box->dhw == 0) recursive stack-lock test. 1901 if (UseBiasedLocking && !UseOptoBiasInlining) { 1902 biased_locking_exit(objReg, tmpReg, DONE_LABEL); 1903 } 1904 1905 #if INCLUDE_RTM_OPT 1906 if (UseRTMForStackLocks && use_rtm) { 1907 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking"); 1908 Label L_regular_unlock; 1909 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // fetch markword 1910 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits 1911 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked 1912 jccb(Assembler::notEqual, L_regular_unlock); // if !HLE RegularLock 1913 xend(); // otherwise end... 1914 jmp(DONE_LABEL); // ... and we're done 1915 bind(L_regular_unlock); 1916 } 1917 #endif 1918 1919 cmpptr(Address(boxReg, 0), (int32_t)NULL_WORD); // Examine the displaced header 1920 jcc (Assembler::zero, DONE_LABEL); // 0 indicates recursive stack-lock 1921 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Examine the object's markword 1922 testptr(tmpReg, markOopDesc::monitor_value); // Inflated? 1923 jccb (Assembler::zero, Stacked); 1924 1925 // It's inflated. 1926 #if INCLUDE_RTM_OPT 1927 if (use_rtm) { 1928 Label L_regular_inflated_unlock; 1929 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner); 1930 movptr(boxReg, Address(tmpReg, owner_offset)); 1931 testptr(boxReg, boxReg); 1932 jccb(Assembler::notZero, L_regular_inflated_unlock); 1933 xend(); 1934 jmpb(DONE_LABEL); 1935 bind(L_regular_inflated_unlock); 1936 } 1937 #endif 1938 1939 // Despite our balanced locking property we still check that m->_owner == Self 1940 // as java routines or native JNI code called by this thread might 1941 // have released the lock. 1942 // Refer to the comments in synchronizer.cpp for how we might encode extra 1943 // state in _succ so we can avoid fetching EntryList|cxq. 1944 // 1945 // I'd like to add more cases in fast_lock() and fast_unlock() -- 1946 // such as recursive enter and exit -- but we have to be wary of 1947 // I$ bloat, T$ effects and BP$ effects. 1948 // 1949 // If there's no contention try a 1-0 exit. That is, exit without 1950 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how 1951 // we detect and recover from the race that the 1-0 exit admits. 1952 // 1953 // Conceptually Fast_Unlock() must execute a STST|LDST "release" barrier 1954 // before it STs null into _owner, releasing the lock. Updates 1955 // to data protected by the critical section must be visible before 1956 // we drop the lock (and thus before any other thread could acquire 1957 // the lock and observe the fields protected by the lock). 1958 // IA32's memory-model is SPO, so STs are ordered with respect to 1959 // each other and there's no need for an explicit barrier (fence). 1960 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 1961 #ifndef _LP64 1962 get_thread (boxReg); 1963 1964 // Note that we could employ various encoding schemes to reduce 1965 // the number of loads below (currently 4) to just 2 or 3. 1966 // Refer to the comments in synchronizer.cpp. 1967 // In practice the chain of fetches doesn't seem to impact performance, however. 1968 xorptr(boxReg, boxReg); 1969 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 1970 jccb (Assembler::notZero, DONE_LABEL); 1971 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 1972 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 1973 jccb (Assembler::notZero, CheckSucc); 1974 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD); 1975 jmpb (DONE_LABEL); 1976 1977 bind (Stacked); 1978 // It's not inflated and it's not recursively stack-locked and it's not biased. 1979 // It must be stack-locked. 1980 // Try to reset the header to displaced header. 1981 // The "box" value on the stack is stable, so we can reload 1982 // and be assured we observe the same value as above. 1983 movptr(tmpReg, Address(boxReg, 0)); 1984 lock(); 1985 cmpxchgptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Uses RAX which is box 1986 // Intention fall-thru into DONE_LABEL 1987 1988 // DONE_LABEL is a hot target - we'd really like to place it at the 1989 // start of cache line by padding with NOPs. 1990 // See the AMD and Intel software optimization manuals for the 1991 // most efficient "long" NOP encodings. 1992 // Unfortunately none of our alignment mechanisms suffice. 1993 bind (CheckSucc); 1994 #else // _LP64 1995 // It's inflated 1996 xorptr(boxReg, boxReg); 1997 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions))); 1998 jccb (Assembler::notZero, DONE_LABEL); 1999 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq))); 2000 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList))); 2001 jccb (Assembler::notZero, CheckSucc); 2002 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD); 2003 jmpb (DONE_LABEL); 2004 2005 // Try to avoid passing control into the slow_path ... 2006 Label LSuccess, LGoSlowPath ; 2007 bind (CheckSucc); 2008 2009 // The following optional optimization can be elided if necessary 2010 // Effectively: if (succ == null) goto SlowPath 2011 // The code reduces the window for a race, however, 2012 // and thus benefits performance. 2013 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2014 jccb (Assembler::zero, LGoSlowPath); 2015 2016 xorptr(boxReg, boxReg); 2017 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD); 2018 2019 // Memory barrier/fence 2020 // Dekker pivot point -- fulcrum : ST Owner; MEMBAR; LD Succ 2021 // Instead of MFENCE we use a dummy locked add of 0 to the top-of-stack. 2022 // This is faster on Nehalem and AMD Shanghai/Barcelona. 2023 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences 2024 // We might also restructure (ST Owner=0;barrier;LD _Succ) to 2025 // (mov box,0; xchgq box, &m->Owner; LD _succ) . 2026 lock(); addl(Address(rsp, 0), 0); 2027 2028 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD); 2029 jccb (Assembler::notZero, LSuccess); 2030 2031 // Rare inopportune interleaving - race. 2032 // The successor vanished in the small window above. 2033 // The lock is contended -- (cxq|EntryList) != null -- and there's no apparent successor. 2034 // We need to ensure progress and succession. 2035 // Try to reacquire the lock. 2036 // If that fails then the new owner is responsible for succession and this 2037 // thread needs to take no further action and can exit via the fast path (success). 2038 // If the re-acquire succeeds then pass control into the slow path. 2039 // As implemented, this latter mode is horrible because we generated more 2040 // coherence traffic on the lock *and* artifically extended the critical section 2041 // length while by virtue of passing control into the slow path. 2042 2043 // box is really RAX -- the following CMPXCHG depends on that binding 2044 // cmpxchg R,[M] is equivalent to rax = CAS(M,rax,R) 2045 lock(); 2046 cmpxchgptr(r15_thread, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); 2047 // There's no successor so we tried to regrab the lock. 2048 // If that didn't work, then another thread grabbed the 2049 // lock so we're done (and exit was a success). 2050 jccb (Assembler::notEqual, LSuccess); 2051 // Intentional fall-through into slow-path 2052 2053 bind (LGoSlowPath); 2054 orl (boxReg, 1); // set ICC.ZF=0 to indicate failure 2055 jmpb (DONE_LABEL); 2056 2057 bind (LSuccess); 2058 testl (boxReg, 0); // set ICC.ZF=1 to indicate success 2059 jmpb (DONE_LABEL); 2060 2061 bind (Stacked); 2062 movptr(tmpReg, Address (boxReg, 0)); // re-fetch 2063 lock(); 2064 cmpxchgptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Uses RAX which is box 2065 2066 #endif 2067 bind(DONE_LABEL); 2068 } 2069 #endif // COMPILER2 2070 2071 void MacroAssembler::c2bool(Register x) { 2072 // implements x == 0 ? 0 : 1 2073 // note: must only look at least-significant byte of x 2074 // since C-style booleans are stored in one byte 2075 // only! (was bug) 2076 andl(x, 0xFF); 2077 setb(Assembler::notZero, x); 2078 } 2079 2080 // Wouldn't need if AddressLiteral version had new name 2081 void MacroAssembler::call(Label& L, relocInfo::relocType rtype) { 2082 Assembler::call(L, rtype); 2083 } 2084 2085 void MacroAssembler::call(Register entry) { 2086 Assembler::call(entry); 2087 } 2088 2089 void MacroAssembler::call(AddressLiteral entry) { 2090 if (reachable(entry)) { 2091 Assembler::call_literal(entry.target(), entry.rspec()); 2092 } else { 2093 lea(rscratch1, entry); 2094 Assembler::call(rscratch1); 2095 } 2096 } 2097 2098 void MacroAssembler::ic_call(address entry, jint method_index) { 2099 RelocationHolder rh = virtual_call_Relocation::spec(pc(), method_index); 2100 movptr(rax, (intptr_t)Universe::non_oop_word()); 2101 call(AddressLiteral(entry, rh)); 2102 } 2103 2104 // Implementation of call_VM versions 2105 2106 void MacroAssembler::call_VM(Register oop_result, 2107 address entry_point, 2108 bool check_exceptions) { 2109 Label C, E; 2110 call(C, relocInfo::none); 2111 jmp(E); 2112 2113 bind(C); 2114 call_VM_helper(oop_result, entry_point, 0, check_exceptions); 2115 ret(0); 2116 2117 bind(E); 2118 } 2119 2120 void MacroAssembler::call_VM(Register oop_result, 2121 address entry_point, 2122 Register arg_1, 2123 bool check_exceptions) { 2124 Label C, E; 2125 call(C, relocInfo::none); 2126 jmp(E); 2127 2128 bind(C); 2129 pass_arg1(this, arg_1); 2130 call_VM_helper(oop_result, entry_point, 1, check_exceptions); 2131 ret(0); 2132 2133 bind(E); 2134 } 2135 2136 void MacroAssembler::call_VM(Register oop_result, 2137 address entry_point, 2138 Register arg_1, 2139 Register arg_2, 2140 bool check_exceptions) { 2141 Label C, E; 2142 call(C, relocInfo::none); 2143 jmp(E); 2144 2145 bind(C); 2146 2147 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2148 2149 pass_arg2(this, arg_2); 2150 pass_arg1(this, arg_1); 2151 call_VM_helper(oop_result, entry_point, 2, check_exceptions); 2152 ret(0); 2153 2154 bind(E); 2155 } 2156 2157 void MacroAssembler::call_VM(Register oop_result, 2158 address entry_point, 2159 Register arg_1, 2160 Register arg_2, 2161 Register arg_3, 2162 bool check_exceptions) { 2163 Label C, E; 2164 call(C, relocInfo::none); 2165 jmp(E); 2166 2167 bind(C); 2168 2169 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2170 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2171 pass_arg3(this, arg_3); 2172 2173 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2174 pass_arg2(this, arg_2); 2175 2176 pass_arg1(this, arg_1); 2177 call_VM_helper(oop_result, entry_point, 3, check_exceptions); 2178 ret(0); 2179 2180 bind(E); 2181 } 2182 2183 void MacroAssembler::call_VM(Register oop_result, 2184 Register last_java_sp, 2185 address entry_point, 2186 int number_of_arguments, 2187 bool check_exceptions) { 2188 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg); 2189 call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions); 2190 } 2191 2192 void MacroAssembler::call_VM(Register oop_result, 2193 Register last_java_sp, 2194 address entry_point, 2195 Register arg_1, 2196 bool check_exceptions) { 2197 pass_arg1(this, arg_1); 2198 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 2199 } 2200 2201 void MacroAssembler::call_VM(Register oop_result, 2202 Register last_java_sp, 2203 address entry_point, 2204 Register arg_1, 2205 Register arg_2, 2206 bool check_exceptions) { 2207 2208 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2209 pass_arg2(this, arg_2); 2210 pass_arg1(this, arg_1); 2211 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 2212 } 2213 2214 void MacroAssembler::call_VM(Register oop_result, 2215 Register last_java_sp, 2216 address entry_point, 2217 Register arg_1, 2218 Register arg_2, 2219 Register arg_3, 2220 bool check_exceptions) { 2221 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2222 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2223 pass_arg3(this, arg_3); 2224 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2225 pass_arg2(this, arg_2); 2226 pass_arg1(this, arg_1); 2227 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 2228 } 2229 2230 void MacroAssembler::super_call_VM(Register oop_result, 2231 Register last_java_sp, 2232 address entry_point, 2233 int number_of_arguments, 2234 bool check_exceptions) { 2235 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg); 2236 MacroAssembler::call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions); 2237 } 2238 2239 void MacroAssembler::super_call_VM(Register oop_result, 2240 Register last_java_sp, 2241 address entry_point, 2242 Register arg_1, 2243 bool check_exceptions) { 2244 pass_arg1(this, arg_1); 2245 super_call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 2246 } 2247 2248 void MacroAssembler::super_call_VM(Register oop_result, 2249 Register last_java_sp, 2250 address entry_point, 2251 Register arg_1, 2252 Register arg_2, 2253 bool check_exceptions) { 2254 2255 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2256 pass_arg2(this, arg_2); 2257 pass_arg1(this, arg_1); 2258 super_call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 2259 } 2260 2261 void MacroAssembler::super_call_VM(Register oop_result, 2262 Register last_java_sp, 2263 address entry_point, 2264 Register arg_1, 2265 Register arg_2, 2266 Register arg_3, 2267 bool check_exceptions) { 2268 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2269 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2270 pass_arg3(this, arg_3); 2271 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2272 pass_arg2(this, arg_2); 2273 pass_arg1(this, arg_1); 2274 super_call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 2275 } 2276 2277 void MacroAssembler::call_VM_base(Register oop_result, 2278 Register java_thread, 2279 Register last_java_sp, 2280 address entry_point, 2281 int number_of_arguments, 2282 bool check_exceptions) { 2283 // determine java_thread register 2284 if (!java_thread->is_valid()) { 2285 #ifdef _LP64 2286 java_thread = r15_thread; 2287 #else 2288 java_thread = rdi; 2289 get_thread(java_thread); 2290 #endif // LP64 2291 } 2292 // determine last_java_sp register 2293 if (!last_java_sp->is_valid()) { 2294 last_java_sp = rsp; 2295 } 2296 // debugging support 2297 assert(number_of_arguments >= 0 , "cannot have negative number of arguments"); 2298 LP64_ONLY(assert(java_thread == r15_thread, "unexpected register")); 2299 #ifdef ASSERT 2300 // TraceBytecodes does not use r12 but saves it over the call, so don't verify 2301 // r12 is the heapbase. 2302 LP64_ONLY(if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");) 2303 #endif // ASSERT 2304 2305 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result"); 2306 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp"); 2307 2308 // push java thread (becomes first argument of C function) 2309 2310 NOT_LP64(push(java_thread); number_of_arguments++); 2311 LP64_ONLY(mov(c_rarg0, r15_thread)); 2312 2313 // set last Java frame before call 2314 assert(last_java_sp != rbp, "can't use ebp/rbp"); 2315 2316 // Only interpreter should have to set fp 2317 set_last_Java_frame(java_thread, last_java_sp, rbp, NULL); 2318 2319 // do the call, remove parameters 2320 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments); 2321 2322 // restore the thread (cannot use the pushed argument since arguments 2323 // may be overwritten by C code generated by an optimizing compiler); 2324 // however can use the register value directly if it is callee saved. 2325 if (LP64_ONLY(true ||) java_thread == rdi || java_thread == rsi) { 2326 // rdi & rsi (also r15) are callee saved -> nothing to do 2327 #ifdef ASSERT 2328 guarantee(java_thread != rax, "change this code"); 2329 push(rax); 2330 { Label L; 2331 get_thread(rax); 2332 cmpptr(java_thread, rax); 2333 jcc(Assembler::equal, L); 2334 STOP("MacroAssembler::call_VM_base: rdi not callee saved?"); 2335 bind(L); 2336 } 2337 pop(rax); 2338 #endif 2339 } else { 2340 get_thread(java_thread); 2341 } 2342 // reset last Java frame 2343 // Only interpreter should have to clear fp 2344 reset_last_Java_frame(java_thread, true); 2345 2346 // C++ interp handles this in the interpreter 2347 check_and_handle_popframe(java_thread); 2348 check_and_handle_earlyret(java_thread); 2349 2350 if (check_exceptions) { 2351 // check for pending exceptions (java_thread is set upon return) 2352 cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t) NULL_WORD); 2353 #ifndef _LP64 2354 jump_cc(Assembler::notEqual, 2355 RuntimeAddress(StubRoutines::forward_exception_entry())); 2356 #else 2357 // This used to conditionally jump to forward_exception however it is 2358 // possible if we relocate that the branch will not reach. So we must jump 2359 // around so we can always reach 2360 2361 Label ok; 2362 jcc(Assembler::equal, ok); 2363 jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2364 bind(ok); 2365 #endif // LP64 2366 } 2367 2368 // get oop result if there is one and reset the value in the thread 2369 if (oop_result->is_valid()) { 2370 get_vm_result(oop_result, java_thread); 2371 } 2372 } 2373 2374 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) { 2375 2376 // Calculate the value for last_Java_sp 2377 // somewhat subtle. call_VM does an intermediate call 2378 // which places a return address on the stack just under the 2379 // stack pointer as the user finsihed with it. This allows 2380 // use to retrieve last_Java_pc from last_Java_sp[-1]. 2381 // On 32bit we then have to push additional args on the stack to accomplish 2382 // the actual requested call. On 64bit call_VM only can use register args 2383 // so the only extra space is the return address that call_VM created. 2384 // This hopefully explains the calculations here. 2385 2386 #ifdef _LP64 2387 // We've pushed one address, correct last_Java_sp 2388 lea(rax, Address(rsp, wordSize)); 2389 #else 2390 lea(rax, Address(rsp, (1 + number_of_arguments) * wordSize)); 2391 #endif // LP64 2392 2393 call_VM_base(oop_result, noreg, rax, entry_point, number_of_arguments, check_exceptions); 2394 2395 } 2396 2397 // Use this method when MacroAssembler version of call_VM_leaf_base() should be called from Interpreter. 2398 void MacroAssembler::call_VM_leaf0(address entry_point) { 2399 MacroAssembler::call_VM_leaf_base(entry_point, 0); 2400 } 2401 2402 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) { 2403 call_VM_leaf_base(entry_point, number_of_arguments); 2404 } 2405 2406 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) { 2407 pass_arg0(this, arg_0); 2408 call_VM_leaf(entry_point, 1); 2409 } 2410 2411 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { 2412 2413 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2414 pass_arg1(this, arg_1); 2415 pass_arg0(this, arg_0); 2416 call_VM_leaf(entry_point, 2); 2417 } 2418 2419 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { 2420 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2421 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2422 pass_arg2(this, arg_2); 2423 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2424 pass_arg1(this, arg_1); 2425 pass_arg0(this, arg_0); 2426 call_VM_leaf(entry_point, 3); 2427 } 2428 2429 void MacroAssembler::super_call_VM_leaf(address entry_point) { 2430 MacroAssembler::call_VM_leaf_base(entry_point, 1); 2431 } 2432 2433 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) { 2434 pass_arg0(this, arg_0); 2435 MacroAssembler::call_VM_leaf_base(entry_point, 1); 2436 } 2437 2438 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { 2439 2440 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2441 pass_arg1(this, arg_1); 2442 pass_arg0(this, arg_0); 2443 MacroAssembler::call_VM_leaf_base(entry_point, 2); 2444 } 2445 2446 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { 2447 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2448 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2449 pass_arg2(this, arg_2); 2450 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2451 pass_arg1(this, arg_1); 2452 pass_arg0(this, arg_0); 2453 MacroAssembler::call_VM_leaf_base(entry_point, 3); 2454 } 2455 2456 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) { 2457 LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg")); 2458 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); 2459 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); 2460 pass_arg3(this, arg_3); 2461 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); 2462 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); 2463 pass_arg2(this, arg_2); 2464 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); 2465 pass_arg1(this, arg_1); 2466 pass_arg0(this, arg_0); 2467 MacroAssembler::call_VM_leaf_base(entry_point, 4); 2468 } 2469 2470 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) { 2471 movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset())); 2472 movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD); 2473 verify_oop(oop_result, "broken oop in call_VM_base"); 2474 } 2475 2476 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) { 2477 movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset())); 2478 movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD); 2479 } 2480 2481 void MacroAssembler::check_and_handle_earlyret(Register java_thread) { 2482 } 2483 2484 void MacroAssembler::check_and_handle_popframe(Register java_thread) { 2485 } 2486 2487 void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) { 2488 if (reachable(src1)) { 2489 cmpl(as_Address(src1), imm); 2490 } else { 2491 lea(rscratch1, src1); 2492 cmpl(Address(rscratch1, 0), imm); 2493 } 2494 } 2495 2496 void MacroAssembler::cmp32(Register src1, AddressLiteral src2) { 2497 assert(!src2.is_lval(), "use cmpptr"); 2498 if (reachable(src2)) { 2499 cmpl(src1, as_Address(src2)); 2500 } else { 2501 lea(rscratch1, src2); 2502 cmpl(src1, Address(rscratch1, 0)); 2503 } 2504 } 2505 2506 void MacroAssembler::cmp32(Register src1, int32_t imm) { 2507 Assembler::cmpl(src1, imm); 2508 } 2509 2510 void MacroAssembler::cmp32(Register src1, Address src2) { 2511 Assembler::cmpl(src1, src2); 2512 } 2513 2514 void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) { 2515 ucomisd(opr1, opr2); 2516 2517 Label L; 2518 if (unordered_is_less) { 2519 movl(dst, -1); 2520 jcc(Assembler::parity, L); 2521 jcc(Assembler::below , L); 2522 movl(dst, 0); 2523 jcc(Assembler::equal , L); 2524 increment(dst); 2525 } else { // unordered is greater 2526 movl(dst, 1); 2527 jcc(Assembler::parity, L); 2528 jcc(Assembler::above , L); 2529 movl(dst, 0); 2530 jcc(Assembler::equal , L); 2531 decrementl(dst); 2532 } 2533 bind(L); 2534 } 2535 2536 void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) { 2537 ucomiss(opr1, opr2); 2538 2539 Label L; 2540 if (unordered_is_less) { 2541 movl(dst, -1); 2542 jcc(Assembler::parity, L); 2543 jcc(Assembler::below , L); 2544 movl(dst, 0); 2545 jcc(Assembler::equal , L); 2546 increment(dst); 2547 } else { // unordered is greater 2548 movl(dst, 1); 2549 jcc(Assembler::parity, L); 2550 jcc(Assembler::above , L); 2551 movl(dst, 0); 2552 jcc(Assembler::equal , L); 2553 decrementl(dst); 2554 } 2555 bind(L); 2556 } 2557 2558 2559 void MacroAssembler::cmp8(AddressLiteral src1, int imm) { 2560 if (reachable(src1)) { 2561 cmpb(as_Address(src1), imm); 2562 } else { 2563 lea(rscratch1, src1); 2564 cmpb(Address(rscratch1, 0), imm); 2565 } 2566 } 2567 2568 void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) { 2569 #ifdef _LP64 2570 if (src2.is_lval()) { 2571 movptr(rscratch1, src2); 2572 Assembler::cmpq(src1, rscratch1); 2573 } else if (reachable(src2)) { 2574 cmpq(src1, as_Address(src2)); 2575 } else { 2576 lea(rscratch1, src2); 2577 Assembler::cmpq(src1, Address(rscratch1, 0)); 2578 } 2579 #else 2580 if (src2.is_lval()) { 2581 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec()); 2582 } else { 2583 cmpl(src1, as_Address(src2)); 2584 } 2585 #endif // _LP64 2586 } 2587 2588 void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) { 2589 assert(src2.is_lval(), "not a mem-mem compare"); 2590 #ifdef _LP64 2591 // moves src2's literal address 2592 movptr(rscratch1, src2); 2593 Assembler::cmpq(src1, rscratch1); 2594 #else 2595 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec()); 2596 #endif // _LP64 2597 } 2598 2599 void MacroAssembler::cmpoop(Register src1, Register src2) { 2600 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 2601 bs->obj_equals(this, src1, src2); 2602 } 2603 2604 void MacroAssembler::cmpoop(Register src1, Address src2) { 2605 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 2606 bs->obj_equals(this, src1, src2); 2607 } 2608 2609 #ifdef _LP64 2610 void MacroAssembler::cmpoop(Register src1, jobject src2) { 2611 movoop(rscratch1, src2); 2612 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 2613 bs->obj_equals(this, src1, rscratch1); 2614 } 2615 #endif 2616 2617 void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) { 2618 if (reachable(adr)) { 2619 lock(); 2620 cmpxchgptr(reg, as_Address(adr)); 2621 } else { 2622 lea(rscratch1, adr); 2623 lock(); 2624 cmpxchgptr(reg, Address(rscratch1, 0)); 2625 } 2626 } 2627 2628 void MacroAssembler::cmpxchgptr(Register reg, Address adr) { 2629 LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr)); 2630 } 2631 2632 void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) { 2633 if (reachable(src)) { 2634 Assembler::comisd(dst, as_Address(src)); 2635 } else { 2636 lea(rscratch1, src); 2637 Assembler::comisd(dst, Address(rscratch1, 0)); 2638 } 2639 } 2640 2641 void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) { 2642 if (reachable(src)) { 2643 Assembler::comiss(dst, as_Address(src)); 2644 } else { 2645 lea(rscratch1, src); 2646 Assembler::comiss(dst, Address(rscratch1, 0)); 2647 } 2648 } 2649 2650 2651 void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) { 2652 Condition negated_cond = negate_condition(cond); 2653 Label L; 2654 jcc(negated_cond, L); 2655 pushf(); // Preserve flags 2656 atomic_incl(counter_addr); 2657 popf(); 2658 bind(L); 2659 } 2660 2661 int MacroAssembler::corrected_idivl(Register reg) { 2662 // Full implementation of Java idiv and irem; checks for 2663 // special case as described in JVM spec., p.243 & p.271. 2664 // The function returns the (pc) offset of the idivl 2665 // instruction - may be needed for implicit exceptions. 2666 // 2667 // normal case special case 2668 // 2669 // input : rax,: dividend min_int 2670 // reg: divisor (may not be rax,/rdx) -1 2671 // 2672 // output: rax,: quotient (= rax, idiv reg) min_int 2673 // rdx: remainder (= rax, irem reg) 0 2674 assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register"); 2675 const int min_int = 0x80000000; 2676 Label normal_case, special_case; 2677 2678 // check for special case 2679 cmpl(rax, min_int); 2680 jcc(Assembler::notEqual, normal_case); 2681 xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0) 2682 cmpl(reg, -1); 2683 jcc(Assembler::equal, special_case); 2684 2685 // handle normal case 2686 bind(normal_case); 2687 cdql(); 2688 int idivl_offset = offset(); 2689 idivl(reg); 2690 2691 // normal and special case exit 2692 bind(special_case); 2693 2694 return idivl_offset; 2695 } 2696 2697 2698 2699 void MacroAssembler::decrementl(Register reg, int value) { 2700 if (value == min_jint) {subl(reg, value) ; return; } 2701 if (value < 0) { incrementl(reg, -value); return; } 2702 if (value == 0) { ; return; } 2703 if (value == 1 && UseIncDec) { decl(reg) ; return; } 2704 /* else */ { subl(reg, value) ; return; } 2705 } 2706 2707 void MacroAssembler::decrementl(Address dst, int value) { 2708 if (value == min_jint) {subl(dst, value) ; return; } 2709 if (value < 0) { incrementl(dst, -value); return; } 2710 if (value == 0) { ; return; } 2711 if (value == 1 && UseIncDec) { decl(dst) ; return; } 2712 /* else */ { subl(dst, value) ; return; } 2713 } 2714 2715 void MacroAssembler::division_with_shift (Register reg, int shift_value) { 2716 assert (shift_value > 0, "illegal shift value"); 2717 Label _is_positive; 2718 testl (reg, reg); 2719 jcc (Assembler::positive, _is_positive); 2720 int offset = (1 << shift_value) - 1 ; 2721 2722 if (offset == 1) { 2723 incrementl(reg); 2724 } else { 2725 addl(reg, offset); 2726 } 2727 2728 bind (_is_positive); 2729 sarl(reg, shift_value); 2730 } 2731 2732 void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) { 2733 if (reachable(src)) { 2734 Assembler::divsd(dst, as_Address(src)); 2735 } else { 2736 lea(rscratch1, src); 2737 Assembler::divsd(dst, Address(rscratch1, 0)); 2738 } 2739 } 2740 2741 void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) { 2742 if (reachable(src)) { 2743 Assembler::divss(dst, as_Address(src)); 2744 } else { 2745 lea(rscratch1, src); 2746 Assembler::divss(dst, Address(rscratch1, 0)); 2747 } 2748 } 2749 2750 // !defined(COMPILER2) is because of stupid core builds 2751 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2) || INCLUDE_JVMCI 2752 void MacroAssembler::empty_FPU_stack() { 2753 if (VM_Version::supports_mmx()) { 2754 emms(); 2755 } else { 2756 for (int i = 8; i-- > 0; ) ffree(i); 2757 } 2758 } 2759 #endif // !LP64 || C1 || !C2 || INCLUDE_JVMCI 2760 2761 2762 void MacroAssembler::enter() { 2763 push(rbp); 2764 mov(rbp, rsp); 2765 } 2766 2767 // A 5 byte nop that is safe for patching (see patch_verified_entry) 2768 void MacroAssembler::fat_nop() { 2769 if (UseAddressNop) { 2770 addr_nop_5(); 2771 } else { 2772 emit_int8(0x26); // es: 2773 emit_int8(0x2e); // cs: 2774 emit_int8(0x64); // fs: 2775 emit_int8(0x65); // gs: 2776 emit_int8((unsigned char)0x90); 2777 } 2778 } 2779 2780 void MacroAssembler::fcmp(Register tmp) { 2781 fcmp(tmp, 1, true, true); 2782 } 2783 2784 void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) { 2785 assert(!pop_right || pop_left, "usage error"); 2786 if (VM_Version::supports_cmov()) { 2787 assert(tmp == noreg, "unneeded temp"); 2788 if (pop_left) { 2789 fucomip(index); 2790 } else { 2791 fucomi(index); 2792 } 2793 if (pop_right) { 2794 fpop(); 2795 } 2796 } else { 2797 assert(tmp != noreg, "need temp"); 2798 if (pop_left) { 2799 if (pop_right) { 2800 fcompp(); 2801 } else { 2802 fcomp(index); 2803 } 2804 } else { 2805 fcom(index); 2806 } 2807 // convert FPU condition into eflags condition via rax, 2808 save_rax(tmp); 2809 fwait(); fnstsw_ax(); 2810 sahf(); 2811 restore_rax(tmp); 2812 } 2813 // condition codes set as follows: 2814 // 2815 // CF (corresponds to C0) if x < y 2816 // PF (corresponds to C2) if unordered 2817 // ZF (corresponds to C3) if x = y 2818 } 2819 2820 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) { 2821 fcmp2int(dst, unordered_is_less, 1, true, true); 2822 } 2823 2824 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) { 2825 fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right); 2826 Label L; 2827 if (unordered_is_less) { 2828 movl(dst, -1); 2829 jcc(Assembler::parity, L); 2830 jcc(Assembler::below , L); 2831 movl(dst, 0); 2832 jcc(Assembler::equal , L); 2833 increment(dst); 2834 } else { // unordered is greater 2835 movl(dst, 1); 2836 jcc(Assembler::parity, L); 2837 jcc(Assembler::above , L); 2838 movl(dst, 0); 2839 jcc(Assembler::equal , L); 2840 decrementl(dst); 2841 } 2842 bind(L); 2843 } 2844 2845 void MacroAssembler::fld_d(AddressLiteral src) { 2846 fld_d(as_Address(src)); 2847 } 2848 2849 void MacroAssembler::fld_s(AddressLiteral src) { 2850 fld_s(as_Address(src)); 2851 } 2852 2853 void MacroAssembler::fld_x(AddressLiteral src) { 2854 Assembler::fld_x(as_Address(src)); 2855 } 2856 2857 void MacroAssembler::fldcw(AddressLiteral src) { 2858 Assembler::fldcw(as_Address(src)); 2859 } 2860 2861 void MacroAssembler::mulpd(XMMRegister dst, AddressLiteral src) { 2862 if (reachable(src)) { 2863 Assembler::mulpd(dst, as_Address(src)); 2864 } else { 2865 lea(rscratch1, src); 2866 Assembler::mulpd(dst, Address(rscratch1, 0)); 2867 } 2868 } 2869 2870 void MacroAssembler::increase_precision() { 2871 subptr(rsp, BytesPerWord); 2872 fnstcw(Address(rsp, 0)); 2873 movl(rax, Address(rsp, 0)); 2874 orl(rax, 0x300); 2875 push(rax); 2876 fldcw(Address(rsp, 0)); 2877 pop(rax); 2878 } 2879 2880 void MacroAssembler::restore_precision() { 2881 fldcw(Address(rsp, 0)); 2882 addptr(rsp, BytesPerWord); 2883 } 2884 2885 void MacroAssembler::fpop() { 2886 ffree(); 2887 fincstp(); 2888 } 2889 2890 void MacroAssembler::load_float(Address src) { 2891 if (UseSSE >= 1) { 2892 movflt(xmm0, src); 2893 } else { 2894 LP64_ONLY(ShouldNotReachHere()); 2895 NOT_LP64(fld_s(src)); 2896 } 2897 } 2898 2899 void MacroAssembler::store_float(Address dst) { 2900 if (UseSSE >= 1) { 2901 movflt(dst, xmm0); 2902 } else { 2903 LP64_ONLY(ShouldNotReachHere()); 2904 NOT_LP64(fstp_s(dst)); 2905 } 2906 } 2907 2908 void MacroAssembler::load_double(Address src) { 2909 if (UseSSE >= 2) { 2910 movdbl(xmm0, src); 2911 } else { 2912 LP64_ONLY(ShouldNotReachHere()); 2913 NOT_LP64(fld_d(src)); 2914 } 2915 } 2916 2917 void MacroAssembler::store_double(Address dst) { 2918 if (UseSSE >= 2) { 2919 movdbl(dst, xmm0); 2920 } else { 2921 LP64_ONLY(ShouldNotReachHere()); 2922 NOT_LP64(fstp_d(dst)); 2923 } 2924 } 2925 2926 void MacroAssembler::fremr(Register tmp) { 2927 save_rax(tmp); 2928 { Label L; 2929 bind(L); 2930 fprem(); 2931 fwait(); fnstsw_ax(); 2932 #ifdef _LP64 2933 testl(rax, 0x400); 2934 jcc(Assembler::notEqual, L); 2935 #else 2936 sahf(); 2937 jcc(Assembler::parity, L); 2938 #endif // _LP64 2939 } 2940 restore_rax(tmp); 2941 // Result is in ST0. 2942 // Note: fxch & fpop to get rid of ST1 2943 // (otherwise FPU stack could overflow eventually) 2944 fxch(1); 2945 fpop(); 2946 } 2947 2948 // dst = c = a * b + c 2949 void MacroAssembler::fmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) { 2950 Assembler::vfmadd231sd(c, a, b); 2951 if (dst != c) { 2952 movdbl(dst, c); 2953 } 2954 } 2955 2956 // dst = c = a * b + c 2957 void MacroAssembler::fmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) { 2958 Assembler::vfmadd231ss(c, a, b); 2959 if (dst != c) { 2960 movflt(dst, c); 2961 } 2962 } 2963 2964 // dst = c = a * b + c 2965 void MacroAssembler::vfmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c, int vector_len) { 2966 Assembler::vfmadd231pd(c, a, b, vector_len); 2967 if (dst != c) { 2968 vmovdqu(dst, c); 2969 } 2970 } 2971 2972 // dst = c = a * b + c 2973 void MacroAssembler::vfmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c, int vector_len) { 2974 Assembler::vfmadd231ps(c, a, b, vector_len); 2975 if (dst != c) { 2976 vmovdqu(dst, c); 2977 } 2978 } 2979 2980 // dst = c = a * b + c 2981 void MacroAssembler::vfmad(XMMRegister dst, XMMRegister a, Address b, XMMRegister c, int vector_len) { 2982 Assembler::vfmadd231pd(c, a, b, vector_len); 2983 if (dst != c) { 2984 vmovdqu(dst, c); 2985 } 2986 } 2987 2988 // dst = c = a * b + c 2989 void MacroAssembler::vfmaf(XMMRegister dst, XMMRegister a, Address b, XMMRegister c, int vector_len) { 2990 Assembler::vfmadd231ps(c, a, b, vector_len); 2991 if (dst != c) { 2992 vmovdqu(dst, c); 2993 } 2994 } 2995 2996 void MacroAssembler::incrementl(AddressLiteral dst) { 2997 if (reachable(dst)) { 2998 incrementl(as_Address(dst)); 2999 } else { 3000 lea(rscratch1, dst); 3001 incrementl(Address(rscratch1, 0)); 3002 } 3003 } 3004 3005 void MacroAssembler::incrementl(ArrayAddress dst) { 3006 incrementl(as_Address(dst)); 3007 } 3008 3009 void MacroAssembler::incrementl(Register reg, int value) { 3010 if (value == min_jint) {addl(reg, value) ; return; } 3011 if (value < 0) { decrementl(reg, -value); return; } 3012 if (value == 0) { ; return; } 3013 if (value == 1 && UseIncDec) { incl(reg) ; return; } 3014 /* else */ { addl(reg, value) ; return; } 3015 } 3016 3017 void MacroAssembler::incrementl(Address dst, int value) { 3018 if (value == min_jint) {addl(dst, value) ; return; } 3019 if (value < 0) { decrementl(dst, -value); return; } 3020 if (value == 0) { ; return; } 3021 if (value == 1 && UseIncDec) { incl(dst) ; return; } 3022 /* else */ { addl(dst, value) ; return; } 3023 } 3024 3025 void MacroAssembler::jump(AddressLiteral dst) { 3026 if (reachable(dst)) { 3027 jmp_literal(dst.target(), dst.rspec()); 3028 } else { 3029 lea(rscratch1, dst); 3030 jmp(rscratch1); 3031 } 3032 } 3033 3034 void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) { 3035 if (reachable(dst)) { 3036 InstructionMark im(this); 3037 relocate(dst.reloc()); 3038 const int short_size = 2; 3039 const int long_size = 6; 3040 int offs = (intptr_t)dst.target() - ((intptr_t)pc()); 3041 if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) { 3042 // 0111 tttn #8-bit disp 3043 emit_int8(0x70 | cc); 3044 emit_int8((offs - short_size) & 0xFF); 3045 } else { 3046 // 0000 1111 1000 tttn #32-bit disp 3047 emit_int8(0x0F); 3048 emit_int8((unsigned char)(0x80 | cc)); 3049 emit_int32(offs - long_size); 3050 } 3051 } else { 3052 #ifdef ASSERT 3053 warning("reversing conditional branch"); 3054 #endif /* ASSERT */ 3055 Label skip; 3056 jccb(reverse[cc], skip); 3057 lea(rscratch1, dst); 3058 Assembler::jmp(rscratch1); 3059 bind(skip); 3060 } 3061 } 3062 3063 void MacroAssembler::ldmxcsr(AddressLiteral src) { 3064 if (reachable(src)) { 3065 Assembler::ldmxcsr(as_Address(src)); 3066 } else { 3067 lea(rscratch1, src); 3068 Assembler::ldmxcsr(Address(rscratch1, 0)); 3069 } 3070 } 3071 3072 int MacroAssembler::load_signed_byte(Register dst, Address src) { 3073 int off; 3074 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3075 off = offset(); 3076 movsbl(dst, src); // movsxb 3077 } else { 3078 off = load_unsigned_byte(dst, src); 3079 shll(dst, 24); 3080 sarl(dst, 24); 3081 } 3082 return off; 3083 } 3084 3085 // Note: load_signed_short used to be called load_signed_word. 3086 // Although the 'w' in x86 opcodes refers to the term "word" in the assembler 3087 // manual, which means 16 bits, that usage is found nowhere in HotSpot code. 3088 // The term "word" in HotSpot means a 32- or 64-bit machine word. 3089 int MacroAssembler::load_signed_short(Register dst, Address src) { 3090 int off; 3091 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3092 // This is dubious to me since it seems safe to do a signed 16 => 64 bit 3093 // version but this is what 64bit has always done. This seems to imply 3094 // that users are only using 32bits worth. 3095 off = offset(); 3096 movswl(dst, src); // movsxw 3097 } else { 3098 off = load_unsigned_short(dst, src); 3099 shll(dst, 16); 3100 sarl(dst, 16); 3101 } 3102 return off; 3103 } 3104 3105 int MacroAssembler::load_unsigned_byte(Register dst, Address src) { 3106 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16, 3107 // and "3.9 Partial Register Penalties", p. 22). 3108 int off; 3109 if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) { 3110 off = offset(); 3111 movzbl(dst, src); // movzxb 3112 } else { 3113 xorl(dst, dst); 3114 off = offset(); 3115 movb(dst, src); 3116 } 3117 return off; 3118 } 3119 3120 // Note: load_unsigned_short used to be called load_unsigned_word. 3121 int MacroAssembler::load_unsigned_short(Register dst, Address src) { 3122 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16, 3123 // and "3.9 Partial Register Penalties", p. 22). 3124 int off; 3125 if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) { 3126 off = offset(); 3127 movzwl(dst, src); // movzxw 3128 } else { 3129 xorl(dst, dst); 3130 off = offset(); 3131 movw(dst, src); 3132 } 3133 return off; 3134 } 3135 3136 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) { 3137 switch (size_in_bytes) { 3138 #ifndef _LP64 3139 case 8: 3140 assert(dst2 != noreg, "second dest register required"); 3141 movl(dst, src); 3142 movl(dst2, src.plus_disp(BytesPerInt)); 3143 break; 3144 #else 3145 case 8: movq(dst, src); break; 3146 #endif 3147 case 4: movl(dst, src); break; 3148 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break; 3149 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break; 3150 default: ShouldNotReachHere(); 3151 } 3152 } 3153 3154 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) { 3155 switch (size_in_bytes) { 3156 #ifndef _LP64 3157 case 8: 3158 assert(src2 != noreg, "second source register required"); 3159 movl(dst, src); 3160 movl(dst.plus_disp(BytesPerInt), src2); 3161 break; 3162 #else 3163 case 8: movq(dst, src); break; 3164 #endif 3165 case 4: movl(dst, src); break; 3166 case 2: movw(dst, src); break; 3167 case 1: movb(dst, src); break; 3168 default: ShouldNotReachHere(); 3169 } 3170 } 3171 3172 void MacroAssembler::mov32(AddressLiteral dst, Register src) { 3173 if (reachable(dst)) { 3174 movl(as_Address(dst), src); 3175 } else { 3176 lea(rscratch1, dst); 3177 movl(Address(rscratch1, 0), src); 3178 } 3179 } 3180 3181 void MacroAssembler::mov32(Register dst, AddressLiteral src) { 3182 if (reachable(src)) { 3183 movl(dst, as_Address(src)); 3184 } else { 3185 lea(rscratch1, src); 3186 movl(dst, Address(rscratch1, 0)); 3187 } 3188 } 3189 3190 // C++ bool manipulation 3191 3192 void MacroAssembler::movbool(Register dst, Address src) { 3193 if(sizeof(bool) == 1) 3194 movb(dst, src); 3195 else if(sizeof(bool) == 2) 3196 movw(dst, src); 3197 else if(sizeof(bool) == 4) 3198 movl(dst, src); 3199 else 3200 // unsupported 3201 ShouldNotReachHere(); 3202 } 3203 3204 void MacroAssembler::movbool(Address dst, bool boolconst) { 3205 if(sizeof(bool) == 1) 3206 movb(dst, (int) boolconst); 3207 else if(sizeof(bool) == 2) 3208 movw(dst, (int) boolconst); 3209 else if(sizeof(bool) == 4) 3210 movl(dst, (int) boolconst); 3211 else 3212 // unsupported 3213 ShouldNotReachHere(); 3214 } 3215 3216 void MacroAssembler::movbool(Address dst, Register src) { 3217 if(sizeof(bool) == 1) 3218 movb(dst, src); 3219 else if(sizeof(bool) == 2) 3220 movw(dst, src); 3221 else if(sizeof(bool) == 4) 3222 movl(dst, src); 3223 else 3224 // unsupported 3225 ShouldNotReachHere(); 3226 } 3227 3228 void MacroAssembler::movbyte(ArrayAddress dst, int src) { 3229 movb(as_Address(dst), src); 3230 } 3231 3232 void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) { 3233 if (reachable(src)) { 3234 movdl(dst, as_Address(src)); 3235 } else { 3236 lea(rscratch1, src); 3237 movdl(dst, Address(rscratch1, 0)); 3238 } 3239 } 3240 3241 void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) { 3242 if (reachable(src)) { 3243 movq(dst, as_Address(src)); 3244 } else { 3245 lea(rscratch1, src); 3246 movq(dst, Address(rscratch1, 0)); 3247 } 3248 } 3249 3250 #ifdef COMPILER2 3251 void MacroAssembler::setvectmask(Register dst, Register src) { 3252 guarantee(PostLoopMultiversioning, "must be"); 3253 Assembler::movl(dst, 1); 3254 Assembler::shlxl(dst, dst, src); 3255 Assembler::decl(dst); 3256 Assembler::kmovdl(k1, dst); 3257 Assembler::movl(dst, src); 3258 } 3259 3260 void MacroAssembler::restorevectmask() { 3261 guarantee(PostLoopMultiversioning, "must be"); 3262 Assembler::knotwl(k1, k0); 3263 } 3264 #endif // COMPILER2 3265 3266 void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) { 3267 if (reachable(src)) { 3268 if (UseXmmLoadAndClearUpper) { 3269 movsd (dst, as_Address(src)); 3270 } else { 3271 movlpd(dst, as_Address(src)); 3272 } 3273 } else { 3274 lea(rscratch1, src); 3275 if (UseXmmLoadAndClearUpper) { 3276 movsd (dst, Address(rscratch1, 0)); 3277 } else { 3278 movlpd(dst, Address(rscratch1, 0)); 3279 } 3280 } 3281 } 3282 3283 void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) { 3284 if (reachable(src)) { 3285 movss(dst, as_Address(src)); 3286 } else { 3287 lea(rscratch1, src); 3288 movss(dst, Address(rscratch1, 0)); 3289 } 3290 } 3291 3292 void MacroAssembler::movptr(Register dst, Register src) { 3293 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3294 } 3295 3296 void MacroAssembler::movptr(Register dst, Address src) { 3297 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3298 } 3299 3300 // src should NEVER be a real pointer. Use AddressLiteral for true pointers 3301 void MacroAssembler::movptr(Register dst, intptr_t src) { 3302 LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src)); 3303 } 3304 3305 void MacroAssembler::movptr(Address dst, Register src) { 3306 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); 3307 } 3308 3309 void MacroAssembler::movdqu(Address dst, XMMRegister src) { 3310 assert(((src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15"); 3311 Assembler::movdqu(dst, src); 3312 } 3313 3314 void MacroAssembler::movdqu(XMMRegister dst, Address src) { 3315 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15"); 3316 Assembler::movdqu(dst, src); 3317 } 3318 3319 void MacroAssembler::movdqu(XMMRegister dst, XMMRegister src) { 3320 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15"); 3321 Assembler::movdqu(dst, src); 3322 } 3323 3324 void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src, Register scratchReg) { 3325 if (reachable(src)) { 3326 movdqu(dst, as_Address(src)); 3327 } else { 3328 lea(scratchReg, src); 3329 movdqu(dst, Address(scratchReg, 0)); 3330 } 3331 } 3332 3333 void MacroAssembler::vmovdqu(Address dst, XMMRegister src) { 3334 assert(((src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15"); 3335 Assembler::vmovdqu(dst, src); 3336 } 3337 3338 void MacroAssembler::vmovdqu(XMMRegister dst, Address src) { 3339 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15"); 3340 Assembler::vmovdqu(dst, src); 3341 } 3342 3343 void MacroAssembler::vmovdqu(XMMRegister dst, XMMRegister src) { 3344 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15"); 3345 Assembler::vmovdqu(dst, src); 3346 } 3347 3348 void MacroAssembler::vmovdqu(XMMRegister dst, AddressLiteral src) { 3349 if (reachable(src)) { 3350 vmovdqu(dst, as_Address(src)); 3351 } 3352 else { 3353 lea(rscratch1, src); 3354 vmovdqu(dst, Address(rscratch1, 0)); 3355 } 3356 } 3357 3358 void MacroAssembler::evmovdquq(XMMRegister dst, AddressLiteral src, int vector_len, Register rscratch) { 3359 if (reachable(src)) { 3360 Assembler::evmovdquq(dst, as_Address(src), vector_len); 3361 } else { 3362 lea(rscratch, src); 3363 Assembler::evmovdquq(dst, Address(rscratch, 0), vector_len); 3364 } 3365 } 3366 3367 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) { 3368 if (reachable(src)) { 3369 Assembler::movdqa(dst, as_Address(src)); 3370 } else { 3371 lea(rscratch1, src); 3372 Assembler::movdqa(dst, Address(rscratch1, 0)); 3373 } 3374 } 3375 3376 void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) { 3377 if (reachable(src)) { 3378 Assembler::movsd(dst, as_Address(src)); 3379 } else { 3380 lea(rscratch1, src); 3381 Assembler::movsd(dst, Address(rscratch1, 0)); 3382 } 3383 } 3384 3385 void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) { 3386 if (reachable(src)) { 3387 Assembler::movss(dst, as_Address(src)); 3388 } else { 3389 lea(rscratch1, src); 3390 Assembler::movss(dst, Address(rscratch1, 0)); 3391 } 3392 } 3393 3394 void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) { 3395 if (reachable(src)) { 3396 Assembler::mulsd(dst, as_Address(src)); 3397 } else { 3398 lea(rscratch1, src); 3399 Assembler::mulsd(dst, Address(rscratch1, 0)); 3400 } 3401 } 3402 3403 void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) { 3404 if (reachable(src)) { 3405 Assembler::mulss(dst, as_Address(src)); 3406 } else { 3407 lea(rscratch1, src); 3408 Assembler::mulss(dst, Address(rscratch1, 0)); 3409 } 3410 } 3411 3412 void MacroAssembler::null_check(Register reg, int offset) { 3413 if (needs_explicit_null_check(offset)) { 3414 // provoke OS NULL exception if reg = NULL by 3415 // accessing M[reg] w/o changing any (non-CC) registers 3416 // NOTE: cmpl is plenty here to provoke a segv 3417 cmpptr(rax, Address(reg, 0)); 3418 // Note: should probably use testl(rax, Address(reg, 0)); 3419 // may be shorter code (however, this version of 3420 // testl needs to be implemented first) 3421 } else { 3422 // nothing to do, (later) access of M[reg + offset] 3423 // will provoke OS NULL exception if reg = NULL 3424 } 3425 } 3426 3427 void MacroAssembler::test_klass_is_value(Register klass, Register temp_reg, Label& is_value) { 3428 movl(temp_reg, Address(klass, Klass::access_flags_offset())); 3429 testl(temp_reg, JVM_ACC_VALUE); 3430 jcc(Assembler::notZero, is_value); 3431 } 3432 3433 void MacroAssembler::test_field_is_flattenable(Register flags, Register temp_reg, Label& is_flattenable) { 3434 movl(temp_reg, flags); 3435 shrl(temp_reg, ConstantPoolCacheEntry::is_flattenable_field_shift); 3436 andl(temp_reg, 0x1); 3437 testl(temp_reg, temp_reg); 3438 jcc(Assembler::notZero, is_flattenable); 3439 } 3440 3441 void MacroAssembler::test_field_is_not_flattenable(Register flags, Register temp_reg, Label& notFlattenable) { 3442 movl(temp_reg, flags); 3443 shrl(temp_reg, ConstantPoolCacheEntry::is_flattenable_field_shift); 3444 andl(temp_reg, 0x1); 3445 testl(temp_reg, temp_reg); 3446 jcc(Assembler::zero, notFlattenable); 3447 } 3448 3449 void MacroAssembler::test_field_is_flattened(Register flags, Register temp_reg, Label& is_flattened) { 3450 movl(temp_reg, flags); 3451 shrl(temp_reg, ConstantPoolCacheEntry::is_flattened_field_shift); 3452 andl(temp_reg, 0x1); 3453 testl(temp_reg, temp_reg); 3454 jcc(Assembler::notZero, is_flattened); 3455 } 3456 3457 void MacroAssembler::test_flattened_array_oop(Register oop, Register temp_reg, 3458 Label&is_flattened_array) { 3459 load_storage_props(temp_reg, oop); 3460 testb(temp_reg, ArrayStorageProperties::flattened_value); 3461 jcc(Assembler::notZero, is_flattened_array); 3462 } 3463 3464 void MacroAssembler::test_null_free_array_oop(Register oop, Register temp_reg, Label&is_null_free_array) { 3465 load_storage_props(temp_reg, oop); 3466 testb(temp_reg, ArrayStorageProperties::null_free_value); 3467 jcc(Assembler::notZero, is_null_free_array); 3468 } 3469 3470 void MacroAssembler::os_breakpoint() { 3471 // instead of directly emitting a breakpoint, call os:breakpoint for better debugability 3472 // (e.g., MSVC can't call ps() otherwise) 3473 call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint))); 3474 } 3475 3476 void MacroAssembler::unimplemented(const char* what) { 3477 const char* buf = NULL; 3478 { 3479 ResourceMark rm; 3480 stringStream ss; 3481 ss.print("unimplemented: %s", what); 3482 buf = code_string(ss.as_string()); 3483 } 3484 stop(buf); 3485 } 3486 3487 #ifdef _LP64 3488 #define XSTATE_BV 0x200 3489 #endif 3490 3491 void MacroAssembler::pop_CPU_state() { 3492 pop_FPU_state(); 3493 pop_IU_state(); 3494 } 3495 3496 void MacroAssembler::pop_FPU_state() { 3497 #ifndef _LP64 3498 frstor(Address(rsp, 0)); 3499 #else 3500 fxrstor(Address(rsp, 0)); 3501 #endif 3502 addptr(rsp, FPUStateSizeInWords * wordSize); 3503 } 3504 3505 void MacroAssembler::pop_IU_state() { 3506 popa(); 3507 LP64_ONLY(addq(rsp, 8)); 3508 popf(); 3509 } 3510 3511 // Save Integer and Float state 3512 // Warning: Stack must be 16 byte aligned (64bit) 3513 void MacroAssembler::push_CPU_state() { 3514 push_IU_state(); 3515 push_FPU_state(); 3516 } 3517 3518 void MacroAssembler::push_FPU_state() { 3519 subptr(rsp, FPUStateSizeInWords * wordSize); 3520 #ifndef _LP64 3521 fnsave(Address(rsp, 0)); 3522 fwait(); 3523 #else 3524 fxsave(Address(rsp, 0)); 3525 #endif // LP64 3526 } 3527 3528 void MacroAssembler::push_IU_state() { 3529 // Push flags first because pusha kills them 3530 pushf(); 3531 // Make sure rsp stays 16-byte aligned 3532 LP64_ONLY(subq(rsp, 8)); 3533 pusha(); 3534 } 3535 3536 void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp) { // determine java_thread register 3537 if (!java_thread->is_valid()) { 3538 java_thread = rdi; 3539 get_thread(java_thread); 3540 } 3541 // we must set sp to zero to clear frame 3542 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD); 3543 if (clear_fp) { 3544 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD); 3545 } 3546 3547 // Always clear the pc because it could have been set by make_walkable() 3548 movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD); 3549 3550 vzeroupper(); 3551 } 3552 3553 void MacroAssembler::restore_rax(Register tmp) { 3554 if (tmp == noreg) pop(rax); 3555 else if (tmp != rax) mov(rax, tmp); 3556 } 3557 3558 void MacroAssembler::round_to(Register reg, int modulus) { 3559 addptr(reg, modulus - 1); 3560 andptr(reg, -modulus); 3561 } 3562 3563 void MacroAssembler::save_rax(Register tmp) { 3564 if (tmp == noreg) push(rax); 3565 else if (tmp != rax) mov(tmp, rax); 3566 } 3567 3568 void MacroAssembler::safepoint_poll(Label& slow_path, Register thread_reg, Register temp_reg) { 3569 if (SafepointMechanism::uses_thread_local_poll()) { 3570 #ifdef _LP64 3571 assert(thread_reg == r15_thread, "should be"); 3572 #else 3573 if (thread_reg == noreg) { 3574 thread_reg = temp_reg; 3575 get_thread(thread_reg); 3576 } 3577 #endif 3578 testb(Address(thread_reg, Thread::polling_page_offset()), SafepointMechanism::poll_bit()); 3579 jcc(Assembler::notZero, slow_path); // handshake bit set implies poll 3580 } else { 3581 cmp32(ExternalAddress(SafepointSynchronize::address_of_state()), 3582 SafepointSynchronize::_not_synchronized); 3583 jcc(Assembler::notEqual, slow_path); 3584 } 3585 } 3586 3587 // Calls to C land 3588 // 3589 // When entering C land, the rbp, & rsp of the last Java frame have to be recorded 3590 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp 3591 // has to be reset to 0. This is required to allow proper stack traversal. 3592 void MacroAssembler::set_last_Java_frame(Register java_thread, 3593 Register last_java_sp, 3594 Register last_java_fp, 3595 address last_java_pc) { 3596 vzeroupper(); 3597 // determine java_thread register 3598 if (!java_thread->is_valid()) { 3599 java_thread = rdi; 3600 get_thread(java_thread); 3601 } 3602 // determine last_java_sp register 3603 if (!last_java_sp->is_valid()) { 3604 last_java_sp = rsp; 3605 } 3606 3607 // last_java_fp is optional 3608 3609 if (last_java_fp->is_valid()) { 3610 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp); 3611 } 3612 3613 // last_java_pc is optional 3614 3615 if (last_java_pc != NULL) { 3616 lea(Address(java_thread, 3617 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()), 3618 InternalAddress(last_java_pc)); 3619 3620 } 3621 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp); 3622 } 3623 3624 void MacroAssembler::shlptr(Register dst, int imm8) { 3625 LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8)); 3626 } 3627 3628 void MacroAssembler::shrptr(Register dst, int imm8) { 3629 LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8)); 3630 } 3631 3632 void MacroAssembler::sign_extend_byte(Register reg) { 3633 if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) { 3634 movsbl(reg, reg); // movsxb 3635 } else { 3636 shll(reg, 24); 3637 sarl(reg, 24); 3638 } 3639 } 3640 3641 void MacroAssembler::sign_extend_short(Register reg) { 3642 if (LP64_ONLY(true ||) VM_Version::is_P6()) { 3643 movswl(reg, reg); // movsxw 3644 } else { 3645 shll(reg, 16); 3646 sarl(reg, 16); 3647 } 3648 } 3649 3650 void MacroAssembler::testl(Register dst, AddressLiteral src) { 3651 assert(reachable(src), "Address should be reachable"); 3652 testl(dst, as_Address(src)); 3653 } 3654 3655 void MacroAssembler::pcmpeqb(XMMRegister dst, XMMRegister src) { 3656 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3657 Assembler::pcmpeqb(dst, src); 3658 } 3659 3660 void MacroAssembler::pcmpeqw(XMMRegister dst, XMMRegister src) { 3661 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3662 Assembler::pcmpeqw(dst, src); 3663 } 3664 3665 void MacroAssembler::pcmpestri(XMMRegister dst, Address src, int imm8) { 3666 assert((dst->encoding() < 16),"XMM register should be 0-15"); 3667 Assembler::pcmpestri(dst, src, imm8); 3668 } 3669 3670 void MacroAssembler::pcmpestri(XMMRegister dst, XMMRegister src, int imm8) { 3671 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15"); 3672 Assembler::pcmpestri(dst, src, imm8); 3673 } 3674 3675 void MacroAssembler::pmovzxbw(XMMRegister dst, XMMRegister src) { 3676 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3677 Assembler::pmovzxbw(dst, src); 3678 } 3679 3680 void MacroAssembler::pmovzxbw(XMMRegister dst, Address src) { 3681 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3682 Assembler::pmovzxbw(dst, src); 3683 } 3684 3685 void MacroAssembler::pmovmskb(Register dst, XMMRegister src) { 3686 assert((src->encoding() < 16),"XMM register should be 0-15"); 3687 Assembler::pmovmskb(dst, src); 3688 } 3689 3690 void MacroAssembler::ptest(XMMRegister dst, XMMRegister src) { 3691 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15"); 3692 Assembler::ptest(dst, src); 3693 } 3694 3695 void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) { 3696 if (reachable(src)) { 3697 Assembler::sqrtsd(dst, as_Address(src)); 3698 } else { 3699 lea(rscratch1, src); 3700 Assembler::sqrtsd(dst, Address(rscratch1, 0)); 3701 } 3702 } 3703 3704 void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) { 3705 if (reachable(src)) { 3706 Assembler::sqrtss(dst, as_Address(src)); 3707 } else { 3708 lea(rscratch1, src); 3709 Assembler::sqrtss(dst, Address(rscratch1, 0)); 3710 } 3711 } 3712 3713 void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) { 3714 if (reachable(src)) { 3715 Assembler::subsd(dst, as_Address(src)); 3716 } else { 3717 lea(rscratch1, src); 3718 Assembler::subsd(dst, Address(rscratch1, 0)); 3719 } 3720 } 3721 3722 void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) { 3723 if (reachable(src)) { 3724 Assembler::subss(dst, as_Address(src)); 3725 } else { 3726 lea(rscratch1, src); 3727 Assembler::subss(dst, Address(rscratch1, 0)); 3728 } 3729 } 3730 3731 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) { 3732 if (reachable(src)) { 3733 Assembler::ucomisd(dst, as_Address(src)); 3734 } else { 3735 lea(rscratch1, src); 3736 Assembler::ucomisd(dst, Address(rscratch1, 0)); 3737 } 3738 } 3739 3740 void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) { 3741 if (reachable(src)) { 3742 Assembler::ucomiss(dst, as_Address(src)); 3743 } else { 3744 lea(rscratch1, src); 3745 Assembler::ucomiss(dst, Address(rscratch1, 0)); 3746 } 3747 } 3748 3749 void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src) { 3750 // Used in sign-bit flipping with aligned address. 3751 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 3752 if (reachable(src)) { 3753 Assembler::xorpd(dst, as_Address(src)); 3754 } else { 3755 lea(rscratch1, src); 3756 Assembler::xorpd(dst, Address(rscratch1, 0)); 3757 } 3758 } 3759 3760 void MacroAssembler::xorpd(XMMRegister dst, XMMRegister src) { 3761 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) { 3762 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit); 3763 } 3764 else { 3765 Assembler::xorpd(dst, src); 3766 } 3767 } 3768 3769 void MacroAssembler::xorps(XMMRegister dst, XMMRegister src) { 3770 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) { 3771 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit); 3772 } else { 3773 Assembler::xorps(dst, src); 3774 } 3775 } 3776 3777 void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src) { 3778 // Used in sign-bit flipping with aligned address. 3779 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); 3780 if (reachable(src)) { 3781 Assembler::xorps(dst, as_Address(src)); 3782 } else { 3783 lea(rscratch1, src); 3784 Assembler::xorps(dst, Address(rscratch1, 0)); 3785 } 3786 } 3787 3788 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) { 3789 // Used in sign-bit flipping with aligned address. 3790 bool aligned_adr = (((intptr_t)src.target() & 15) == 0); 3791 assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes"); 3792 if (reachable(src)) { 3793 Assembler::pshufb(dst, as_Address(src)); 3794 } else { 3795 lea(rscratch1, src); 3796 Assembler::pshufb(dst, Address(rscratch1, 0)); 3797 } 3798 } 3799 3800 // AVX 3-operands instructions 3801 3802 void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 3803 if (reachable(src)) { 3804 vaddsd(dst, nds, as_Address(src)); 3805 } else { 3806 lea(rscratch1, src); 3807 vaddsd(dst, nds, Address(rscratch1, 0)); 3808 } 3809 } 3810 3811 void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 3812 if (reachable(src)) { 3813 vaddss(dst, nds, as_Address(src)); 3814 } else { 3815 lea(rscratch1, src); 3816 vaddss(dst, nds, Address(rscratch1, 0)); 3817 } 3818 } 3819 3820 void MacroAssembler::vabsss(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) { 3821 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15"); 3822 vandps(dst, nds, negate_field, vector_len); 3823 } 3824 3825 void MacroAssembler::vabssd(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) { 3826 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15"); 3827 vandpd(dst, nds, negate_field, vector_len); 3828 } 3829 3830 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 3831 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3832 Assembler::vpaddb(dst, nds, src, vector_len); 3833 } 3834 3835 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 3836 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3837 Assembler::vpaddb(dst, nds, src, vector_len); 3838 } 3839 3840 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 3841 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3842 Assembler::vpaddw(dst, nds, src, vector_len); 3843 } 3844 3845 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 3846 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3847 Assembler::vpaddw(dst, nds, src, vector_len); 3848 } 3849 3850 void MacroAssembler::vpand(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 3851 if (reachable(src)) { 3852 Assembler::vpand(dst, nds, as_Address(src), vector_len); 3853 } else { 3854 lea(rscratch1, src); 3855 Assembler::vpand(dst, nds, Address(rscratch1, 0), vector_len); 3856 } 3857 } 3858 3859 void MacroAssembler::vpbroadcastw(XMMRegister dst, XMMRegister src, int vector_len) { 3860 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3861 Assembler::vpbroadcastw(dst, src, vector_len); 3862 } 3863 3864 void MacroAssembler::vpcmpeqb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 3865 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3866 Assembler::vpcmpeqb(dst, nds, src, vector_len); 3867 } 3868 3869 void MacroAssembler::vpcmpeqw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 3870 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3871 Assembler::vpcmpeqw(dst, nds, src, vector_len); 3872 } 3873 3874 void MacroAssembler::vpmovzxbw(XMMRegister dst, Address src, int vector_len) { 3875 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3876 Assembler::vpmovzxbw(dst, src, vector_len); 3877 } 3878 3879 void MacroAssembler::vpmovmskb(Register dst, XMMRegister src) { 3880 assert((src->encoding() < 16),"XMM register should be 0-15"); 3881 Assembler::vpmovmskb(dst, src); 3882 } 3883 3884 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 3885 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3886 Assembler::vpmullw(dst, nds, src, vector_len); 3887 } 3888 3889 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 3890 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3891 Assembler::vpmullw(dst, nds, src, vector_len); 3892 } 3893 3894 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 3895 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3896 Assembler::vpsubb(dst, nds, src, vector_len); 3897 } 3898 3899 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 3900 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3901 Assembler::vpsubb(dst, nds, src, vector_len); 3902 } 3903 3904 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) { 3905 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3906 Assembler::vpsubw(dst, nds, src, vector_len); 3907 } 3908 3909 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) { 3910 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3911 Assembler::vpsubw(dst, nds, src, vector_len); 3912 } 3913 3914 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 3915 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3916 Assembler::vpsraw(dst, nds, shift, vector_len); 3917 } 3918 3919 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 3920 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3921 Assembler::vpsraw(dst, nds, shift, vector_len); 3922 } 3923 3924 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 3925 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3926 Assembler::vpsrlw(dst, nds, shift, vector_len); 3927 } 3928 3929 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 3930 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3931 Assembler::vpsrlw(dst, nds, shift, vector_len); 3932 } 3933 3934 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) { 3935 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3936 Assembler::vpsllw(dst, nds, shift, vector_len); 3937 } 3938 3939 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) { 3940 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3941 Assembler::vpsllw(dst, nds, shift, vector_len); 3942 } 3943 3944 void MacroAssembler::vptest(XMMRegister dst, XMMRegister src) { 3945 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15"); 3946 Assembler::vptest(dst, src); 3947 } 3948 3949 void MacroAssembler::punpcklbw(XMMRegister dst, XMMRegister src) { 3950 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3951 Assembler::punpcklbw(dst, src); 3952 } 3953 3954 void MacroAssembler::pshufd(XMMRegister dst, Address src, int mode) { 3955 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15"); 3956 Assembler::pshufd(dst, src, mode); 3957 } 3958 3959 void MacroAssembler::pshuflw(XMMRegister dst, XMMRegister src, int mode) { 3960 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15"); 3961 Assembler::pshuflw(dst, src, mode); 3962 } 3963 3964 void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 3965 if (reachable(src)) { 3966 vandpd(dst, nds, as_Address(src), vector_len); 3967 } else { 3968 lea(rscratch1, src); 3969 vandpd(dst, nds, Address(rscratch1, 0), vector_len); 3970 } 3971 } 3972 3973 void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 3974 if (reachable(src)) { 3975 vandps(dst, nds, as_Address(src), vector_len); 3976 } else { 3977 lea(rscratch1, src); 3978 vandps(dst, nds, Address(rscratch1, 0), vector_len); 3979 } 3980 } 3981 3982 void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 3983 if (reachable(src)) { 3984 vdivsd(dst, nds, as_Address(src)); 3985 } else { 3986 lea(rscratch1, src); 3987 vdivsd(dst, nds, Address(rscratch1, 0)); 3988 } 3989 } 3990 3991 void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 3992 if (reachable(src)) { 3993 vdivss(dst, nds, as_Address(src)); 3994 } else { 3995 lea(rscratch1, src); 3996 vdivss(dst, nds, Address(rscratch1, 0)); 3997 } 3998 } 3999 4000 void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4001 if (reachable(src)) { 4002 vmulsd(dst, nds, as_Address(src)); 4003 } else { 4004 lea(rscratch1, src); 4005 vmulsd(dst, nds, Address(rscratch1, 0)); 4006 } 4007 } 4008 4009 void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4010 if (reachable(src)) { 4011 vmulss(dst, nds, as_Address(src)); 4012 } else { 4013 lea(rscratch1, src); 4014 vmulss(dst, nds, Address(rscratch1, 0)); 4015 } 4016 } 4017 4018 void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4019 if (reachable(src)) { 4020 vsubsd(dst, nds, as_Address(src)); 4021 } else { 4022 lea(rscratch1, src); 4023 vsubsd(dst, nds, Address(rscratch1, 0)); 4024 } 4025 } 4026 4027 void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4028 if (reachable(src)) { 4029 vsubss(dst, nds, as_Address(src)); 4030 } else { 4031 lea(rscratch1, src); 4032 vsubss(dst, nds, Address(rscratch1, 0)); 4033 } 4034 } 4035 4036 void MacroAssembler::vnegatess(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4037 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15"); 4038 vxorps(dst, nds, src, Assembler::AVX_128bit); 4039 } 4040 4041 void MacroAssembler::vnegatesd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { 4042 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15"); 4043 vxorpd(dst, nds, src, Assembler::AVX_128bit); 4044 } 4045 4046 void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 4047 if (reachable(src)) { 4048 vxorpd(dst, nds, as_Address(src), vector_len); 4049 } else { 4050 lea(rscratch1, src); 4051 vxorpd(dst, nds, Address(rscratch1, 0), vector_len); 4052 } 4053 } 4054 4055 void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) { 4056 if (reachable(src)) { 4057 vxorps(dst, nds, as_Address(src), vector_len); 4058 } else { 4059 lea(rscratch1, src); 4060 vxorps(dst, nds, Address(rscratch1, 0), vector_len); 4061 } 4062 } 4063 4064 void MacroAssembler::clear_jweak_tag(Register possibly_jweak) { 4065 const int32_t inverted_jweak_mask = ~static_cast<int32_t>(JNIHandles::weak_tag_mask); 4066 STATIC_ASSERT(inverted_jweak_mask == -2); // otherwise check this code 4067 // The inverted mask is sign-extended 4068 andptr(possibly_jweak, inverted_jweak_mask); 4069 } 4070 4071 void MacroAssembler::resolve_jobject(Register value, 4072 Register thread, 4073 Register tmp) { 4074 assert_different_registers(value, thread, tmp); 4075 Label done, not_weak; 4076 testptr(value, value); 4077 jcc(Assembler::zero, done); // Use NULL as-is. 4078 testptr(value, JNIHandles::weak_tag_mask); // Test for jweak tag. 4079 jcc(Assembler::zero, not_weak); 4080 // Resolve jweak. 4081 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF, 4082 value, Address(value, -JNIHandles::weak_tag_value), tmp, thread); 4083 verify_oop(value); 4084 jmp(done); 4085 bind(not_weak); 4086 // Resolve (untagged) jobject. 4087 access_load_at(T_OBJECT, IN_NATIVE, value, Address(value, 0), tmp, thread); 4088 verify_oop(value); 4089 bind(done); 4090 } 4091 4092 void MacroAssembler::subptr(Register dst, int32_t imm32) { 4093 LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32)); 4094 } 4095 4096 // Force generation of a 4 byte immediate value even if it fits into 8bit 4097 void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) { 4098 LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32)); 4099 } 4100 4101 void MacroAssembler::subptr(Register dst, Register src) { 4102 LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src)); 4103 } 4104 4105 // C++ bool manipulation 4106 void MacroAssembler::testbool(Register dst) { 4107 if(sizeof(bool) == 1) 4108 testb(dst, 0xff); 4109 else if(sizeof(bool) == 2) { 4110 // testw implementation needed for two byte bools 4111 ShouldNotReachHere(); 4112 } else if(sizeof(bool) == 4) 4113 testl(dst, dst); 4114 else 4115 // unsupported 4116 ShouldNotReachHere(); 4117 } 4118 4119 void MacroAssembler::testptr(Register dst, Register src) { 4120 LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src)); 4121 } 4122 4123 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes. 4124 void MacroAssembler::tlab_allocate(Register thread, Register obj, 4125 Register var_size_in_bytes, 4126 int con_size_in_bytes, 4127 Register t1, 4128 Register t2, 4129 Label& slow_case) { 4130 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 4131 bs->tlab_allocate(this, thread, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case); 4132 } 4133 4134 // Defines obj, preserves var_size_in_bytes 4135 void MacroAssembler::eden_allocate(Register thread, Register obj, 4136 Register var_size_in_bytes, 4137 int con_size_in_bytes, 4138 Register t1, 4139 Label& slow_case) { 4140 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 4141 bs->eden_allocate(this, thread, obj, var_size_in_bytes, con_size_in_bytes, t1, slow_case); 4142 } 4143 4144 // Preserves the contents of address, destroys the contents length_in_bytes and temp. 4145 void MacroAssembler::zero_memory(Register address, Register length_in_bytes, int offset_in_bytes, Register temp) { 4146 assert(address != length_in_bytes && address != temp && temp != length_in_bytes, "registers must be different"); 4147 assert((offset_in_bytes & (BytesPerWord - 1)) == 0, "offset must be a multiple of BytesPerWord"); 4148 Label done; 4149 4150 testptr(length_in_bytes, length_in_bytes); 4151 jcc(Assembler::zero, done); 4152 4153 // initialize topmost word, divide index by 2, check if odd and test if zero 4154 // note: for the remaining code to work, index must be a multiple of BytesPerWord 4155 #ifdef ASSERT 4156 { 4157 Label L; 4158 testptr(length_in_bytes, BytesPerWord - 1); 4159 jcc(Assembler::zero, L); 4160 stop("length must be a multiple of BytesPerWord"); 4161 bind(L); 4162 } 4163 #endif 4164 Register index = length_in_bytes; 4165 xorptr(temp, temp); // use _zero reg to clear memory (shorter code) 4166 if (UseIncDec) { 4167 shrptr(index, 3); // divide by 8/16 and set carry flag if bit 2 was set 4168 } else { 4169 shrptr(index, 2); // use 2 instructions to avoid partial flag stall 4170 shrptr(index, 1); 4171 } 4172 #ifndef _LP64 4173 // index could have not been a multiple of 8 (i.e., bit 2 was set) 4174 { 4175 Label even; 4176 // note: if index was a multiple of 8, then it cannot 4177 // be 0 now otherwise it must have been 0 before 4178 // => if it is even, we don't need to check for 0 again 4179 jcc(Assembler::carryClear, even); 4180 // clear topmost word (no jump would be needed if conditional assignment worked here) 4181 movptr(Address(address, index, Address::times_8, offset_in_bytes - 0*BytesPerWord), temp); 4182 // index could be 0 now, must check again 4183 jcc(Assembler::zero, done); 4184 bind(even); 4185 } 4186 #endif // !_LP64 4187 // initialize remaining object fields: index is a multiple of 2 now 4188 { 4189 Label loop; 4190 bind(loop); 4191 movptr(Address(address, index, Address::times_8, offset_in_bytes - 1*BytesPerWord), temp); 4192 NOT_LP64(movptr(Address(address, index, Address::times_8, offset_in_bytes - 2*BytesPerWord), temp);) 4193 decrement(index); 4194 jcc(Assembler::notZero, loop); 4195 } 4196 4197 bind(done); 4198 } 4199 4200 // Look up the method for a megamorphic invokeinterface call. 4201 // The target method is determined by <intf_klass, itable_index>. 4202 // The receiver klass is in recv_klass. 4203 // On success, the result will be in method_result, and execution falls through. 4204 // On failure, execution transfers to the given label. 4205 void MacroAssembler::lookup_interface_method(Register recv_klass, 4206 Register intf_klass, 4207 RegisterOrConstant itable_index, 4208 Register method_result, 4209 Register scan_temp, 4210 Label& L_no_such_interface, 4211 bool return_method) { 4212 assert_different_registers(recv_klass, intf_klass, scan_temp); 4213 assert_different_registers(method_result, intf_klass, scan_temp); 4214 assert(recv_klass != method_result || !return_method, 4215 "recv_klass can be destroyed when method isn't needed"); 4216 4217 assert(itable_index.is_constant() || itable_index.as_register() == method_result, 4218 "caller must use same register for non-constant itable index as for method"); 4219 4220 // Compute start of first itableOffsetEntry (which is at the end of the vtable) 4221 int vtable_base = in_bytes(Klass::vtable_start_offset()); 4222 int itentry_off = itableMethodEntry::method_offset_in_bytes(); 4223 int scan_step = itableOffsetEntry::size() * wordSize; 4224 int vte_size = vtableEntry::size_in_bytes(); 4225 Address::ScaleFactor times_vte_scale = Address::times_ptr; 4226 assert(vte_size == wordSize, "else adjust times_vte_scale"); 4227 4228 movl(scan_temp, Address(recv_klass, Klass::vtable_length_offset())); 4229 4230 // %%% Could store the aligned, prescaled offset in the klassoop. 4231 lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base)); 4232 4233 if (return_method) { 4234 // Adjust recv_klass by scaled itable_index, so we can free itable_index. 4235 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below"); 4236 lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off)); 4237 } 4238 4239 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) { 4240 // if (scan->interface() == intf) { 4241 // result = (klass + scan->offset() + itable_index); 4242 // } 4243 // } 4244 Label search, found_method; 4245 4246 for (int peel = 1; peel >= 0; peel--) { 4247 movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes())); 4248 cmpptr(intf_klass, method_result); 4249 4250 if (peel) { 4251 jccb(Assembler::equal, found_method); 4252 } else { 4253 jccb(Assembler::notEqual, search); 4254 // (invert the test to fall through to found_method...) 4255 } 4256 4257 if (!peel) break; 4258 4259 bind(search); 4260 4261 // Check that the previous entry is non-null. A null entry means that 4262 // the receiver class doesn't implement the interface, and wasn't the 4263 // same as when the caller was compiled. 4264 testptr(method_result, method_result); 4265 jcc(Assembler::zero, L_no_such_interface); 4266 addptr(scan_temp, scan_step); 4267 } 4268 4269 bind(found_method); 4270 4271 if (return_method) { 4272 // Got a hit. 4273 movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes())); 4274 movptr(method_result, Address(recv_klass, scan_temp, Address::times_1)); 4275 } 4276 } 4277 4278 4279 // virtual method calling 4280 void MacroAssembler::lookup_virtual_method(Register recv_klass, 4281 RegisterOrConstant vtable_index, 4282 Register method_result) { 4283 const int base = in_bytes(Klass::vtable_start_offset()); 4284 assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below"); 4285 Address vtable_entry_addr(recv_klass, 4286 vtable_index, Address::times_ptr, 4287 base + vtableEntry::method_offset_in_bytes()); 4288 movptr(method_result, vtable_entry_addr); 4289 } 4290 4291 4292 void MacroAssembler::check_klass_subtype(Register sub_klass, 4293 Register super_klass, 4294 Register temp_reg, 4295 Label& L_success) { 4296 Label L_failure; 4297 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL); 4298 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL); 4299 bind(L_failure); 4300 } 4301 4302 4303 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass, 4304 Register super_klass, 4305 Register temp_reg, 4306 Label* L_success, 4307 Label* L_failure, 4308 Label* L_slow_path, 4309 RegisterOrConstant super_check_offset) { 4310 assert_different_registers(sub_klass, super_klass, temp_reg); 4311 bool must_load_sco = (super_check_offset.constant_or_zero() == -1); 4312 if (super_check_offset.is_register()) { 4313 assert_different_registers(sub_klass, super_klass, 4314 super_check_offset.as_register()); 4315 } else if (must_load_sco) { 4316 assert(temp_reg != noreg, "supply either a temp or a register offset"); 4317 } 4318 4319 Label L_fallthrough; 4320 int label_nulls = 0; 4321 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 4322 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 4323 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; } 4324 assert(label_nulls <= 1, "at most one NULL in the batch"); 4325 4326 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 4327 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 4328 Address super_check_offset_addr(super_klass, sco_offset); 4329 4330 // Hacked jcc, which "knows" that L_fallthrough, at least, is in 4331 // range of a jccb. If this routine grows larger, reconsider at 4332 // least some of these. 4333 #define local_jcc(assembler_cond, label) \ 4334 if (&(label) == &L_fallthrough) jccb(assembler_cond, label); \ 4335 else jcc( assembler_cond, label) /*omit semi*/ 4336 4337 // Hacked jmp, which may only be used just before L_fallthrough. 4338 #define final_jmp(label) \ 4339 if (&(label) == &L_fallthrough) { /*do nothing*/ } \ 4340 else jmp(label) /*omit semi*/ 4341 4342 // If the pointers are equal, we are done (e.g., String[] elements). 4343 // This self-check enables sharing of secondary supertype arrays among 4344 // non-primary types such as array-of-interface. Otherwise, each such 4345 // type would need its own customized SSA. 4346 // We move this check to the front of the fast path because many 4347 // type checks are in fact trivially successful in this manner, 4348 // so we get a nicely predicted branch right at the start of the check. 4349 cmpptr(sub_klass, super_klass); 4350 local_jcc(Assembler::equal, *L_success); 4351 4352 // Check the supertype display: 4353 if (must_load_sco) { 4354 // Positive movl does right thing on LP64. 4355 movl(temp_reg, super_check_offset_addr); 4356 super_check_offset = RegisterOrConstant(temp_reg); 4357 } 4358 Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0); 4359 cmpptr(super_klass, super_check_addr); // load displayed supertype 4360 4361 // This check has worked decisively for primary supers. 4362 // Secondary supers are sought in the super_cache ('super_cache_addr'). 4363 // (Secondary supers are interfaces and very deeply nested subtypes.) 4364 // This works in the same check above because of a tricky aliasing 4365 // between the super_cache and the primary super display elements. 4366 // (The 'super_check_addr' can address either, as the case requires.) 4367 // Note that the cache is updated below if it does not help us find 4368 // what we need immediately. 4369 // So if it was a primary super, we can just fail immediately. 4370 // Otherwise, it's the slow path for us (no success at this point). 4371 4372 if (super_check_offset.is_register()) { 4373 local_jcc(Assembler::equal, *L_success); 4374 cmpl(super_check_offset.as_register(), sc_offset); 4375 if (L_failure == &L_fallthrough) { 4376 local_jcc(Assembler::equal, *L_slow_path); 4377 } else { 4378 local_jcc(Assembler::notEqual, *L_failure); 4379 final_jmp(*L_slow_path); 4380 } 4381 } else if (super_check_offset.as_constant() == sc_offset) { 4382 // Need a slow path; fast failure is impossible. 4383 if (L_slow_path == &L_fallthrough) { 4384 local_jcc(Assembler::equal, *L_success); 4385 } else { 4386 local_jcc(Assembler::notEqual, *L_slow_path); 4387 final_jmp(*L_success); 4388 } 4389 } else { 4390 // No slow path; it's a fast decision. 4391 if (L_failure == &L_fallthrough) { 4392 local_jcc(Assembler::equal, *L_success); 4393 } else { 4394 local_jcc(Assembler::notEqual, *L_failure); 4395 final_jmp(*L_success); 4396 } 4397 } 4398 4399 bind(L_fallthrough); 4400 4401 #undef local_jcc 4402 #undef final_jmp 4403 } 4404 4405 4406 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass, 4407 Register super_klass, 4408 Register temp_reg, 4409 Register temp2_reg, 4410 Label* L_success, 4411 Label* L_failure, 4412 bool set_cond_codes) { 4413 assert_different_registers(sub_klass, super_klass, temp_reg); 4414 if (temp2_reg != noreg) 4415 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg); 4416 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg) 4417 4418 Label L_fallthrough; 4419 int label_nulls = 0; 4420 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } 4421 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } 4422 assert(label_nulls <= 1, "at most one NULL in the batch"); 4423 4424 // a couple of useful fields in sub_klass: 4425 int ss_offset = in_bytes(Klass::secondary_supers_offset()); 4426 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 4427 Address secondary_supers_addr(sub_klass, ss_offset); 4428 Address super_cache_addr( sub_klass, sc_offset); 4429 4430 // Do a linear scan of the secondary super-klass chain. 4431 // This code is rarely used, so simplicity is a virtue here. 4432 // The repne_scan instruction uses fixed registers, which we must spill. 4433 // Don't worry too much about pre-existing connections with the input regs. 4434 4435 assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super) 4436 assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter) 4437 4438 // Get super_klass value into rax (even if it was in rdi or rcx). 4439 bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false; 4440 if (super_klass != rax || UseCompressedOops) { 4441 if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; } 4442 mov(rax, super_klass); 4443 } 4444 if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; } 4445 if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; } 4446 4447 #ifndef PRODUCT 4448 int* pst_counter = &SharedRuntime::_partial_subtype_ctr; 4449 ExternalAddress pst_counter_addr((address) pst_counter); 4450 NOT_LP64( incrementl(pst_counter_addr) ); 4451 LP64_ONLY( lea(rcx, pst_counter_addr) ); 4452 LP64_ONLY( incrementl(Address(rcx, 0)) ); 4453 #endif //PRODUCT 4454 4455 // We will consult the secondary-super array. 4456 movptr(rdi, secondary_supers_addr); 4457 // Load the array length. (Positive movl does right thing on LP64.) 4458 movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes())); 4459 // Skip to start of data. 4460 addptr(rdi, Array<Klass*>::base_offset_in_bytes()); 4461 4462 // Scan RCX words at [RDI] for an occurrence of RAX. 4463 // Set NZ/Z based on last compare. 4464 // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does 4465 // not change flags (only scas instruction which is repeated sets flags). 4466 // Set Z = 0 (not equal) before 'repne' to indicate that class was not found. 4467 4468 testptr(rax,rax); // Set Z = 0 4469 repne_scan(); 4470 4471 // Unspill the temp. registers: 4472 if (pushed_rdi) pop(rdi); 4473 if (pushed_rcx) pop(rcx); 4474 if (pushed_rax) pop(rax); 4475 4476 if (set_cond_codes) { 4477 // Special hack for the AD files: rdi is guaranteed non-zero. 4478 assert(!pushed_rdi, "rdi must be left non-NULL"); 4479 // Also, the condition codes are properly set Z/NZ on succeed/failure. 4480 } 4481 4482 if (L_failure == &L_fallthrough) 4483 jccb(Assembler::notEqual, *L_failure); 4484 else jcc(Assembler::notEqual, *L_failure); 4485 4486 // Success. Cache the super we found and proceed in triumph. 4487 movptr(super_cache_addr, super_klass); 4488 4489 if (L_success != &L_fallthrough) { 4490 jmp(*L_success); 4491 } 4492 4493 #undef IS_A_TEMP 4494 4495 bind(L_fallthrough); 4496 } 4497 4498 4499 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) { 4500 if (VM_Version::supports_cmov()) { 4501 cmovl(cc, dst, src); 4502 } else { 4503 Label L; 4504 jccb(negate_condition(cc), L); 4505 movl(dst, src); 4506 bind(L); 4507 } 4508 } 4509 4510 void MacroAssembler::cmov32(Condition cc, Register dst, Register src) { 4511 if (VM_Version::supports_cmov()) { 4512 cmovl(cc, dst, src); 4513 } else { 4514 Label L; 4515 jccb(negate_condition(cc), L); 4516 movl(dst, src); 4517 bind(L); 4518 } 4519 } 4520 4521 void MacroAssembler::verify_oop(Register reg, const char* s) { 4522 if (!VerifyOops || VerifyAdapterSharing) { 4523 // Below address of the code string confuses VerifyAdapterSharing 4524 // because it may differ between otherwise equivalent adapters. 4525 return; 4526 } 4527 4528 // Pass register number to verify_oop_subroutine 4529 const char* b = NULL; 4530 { 4531 ResourceMark rm; 4532 stringStream ss; 4533 ss.print("verify_oop: %s: %s", reg->name(), s); 4534 b = code_string(ss.as_string()); 4535 } 4536 BLOCK_COMMENT("verify_oop {"); 4537 #ifdef _LP64 4538 push(rscratch1); // save r10, trashed by movptr() 4539 #endif 4540 push(rax); // save rax, 4541 push(reg); // pass register argument 4542 ExternalAddress buffer((address) b); 4543 // avoid using pushptr, as it modifies scratch registers 4544 // and our contract is not to modify anything 4545 movptr(rax, buffer.addr()); 4546 push(rax); 4547 // call indirectly to solve generation ordering problem 4548 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); 4549 call(rax); 4550 // Caller pops the arguments (oop, message) and restores rax, r10 4551 BLOCK_COMMENT("} verify_oop"); 4552 } 4553 4554 4555 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr, 4556 Register tmp, 4557 int offset) { 4558 intptr_t value = *delayed_value_addr; 4559 if (value != 0) 4560 return RegisterOrConstant(value + offset); 4561 4562 // load indirectly to solve generation ordering problem 4563 movptr(tmp, ExternalAddress((address) delayed_value_addr)); 4564 4565 #ifdef ASSERT 4566 { Label L; 4567 testptr(tmp, tmp); 4568 if (WizardMode) { 4569 const char* buf = NULL; 4570 { 4571 ResourceMark rm; 4572 stringStream ss; 4573 ss.print("DelayedValue=" INTPTR_FORMAT, delayed_value_addr[1]); 4574 buf = code_string(ss.as_string()); 4575 } 4576 jcc(Assembler::notZero, L); 4577 STOP(buf); 4578 } else { 4579 jccb(Assembler::notZero, L); 4580 hlt(); 4581 } 4582 bind(L); 4583 } 4584 #endif 4585 4586 if (offset != 0) 4587 addptr(tmp, offset); 4588 4589 return RegisterOrConstant(tmp); 4590 } 4591 4592 4593 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot, 4594 int extra_slot_offset) { 4595 // cf. TemplateTable::prepare_invoke(), if (load_receiver). 4596 int stackElementSize = Interpreter::stackElementSize; 4597 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0); 4598 #ifdef ASSERT 4599 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1); 4600 assert(offset1 - offset == stackElementSize, "correct arithmetic"); 4601 #endif 4602 Register scale_reg = noreg; 4603 Address::ScaleFactor scale_factor = Address::no_scale; 4604 if (arg_slot.is_constant()) { 4605 offset += arg_slot.as_constant() * stackElementSize; 4606 } else { 4607 scale_reg = arg_slot.as_register(); 4608 scale_factor = Address::times(stackElementSize); 4609 } 4610 offset += wordSize; // return PC is on stack 4611 return Address(rsp, scale_reg, scale_factor, offset); 4612 } 4613 4614 4615 void MacroAssembler::verify_oop_addr(Address addr, const char* s) { 4616 if (!VerifyOops || VerifyAdapterSharing) { 4617 // Below address of the code string confuses VerifyAdapterSharing 4618 // because it may differ between otherwise equivalent adapters. 4619 return; 4620 } 4621 4622 // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord); 4623 // Pass register number to verify_oop_subroutine 4624 const char* b = NULL; 4625 { 4626 ResourceMark rm; 4627 stringStream ss; 4628 ss.print("verify_oop_addr: %s", s); 4629 b = code_string(ss.as_string()); 4630 } 4631 #ifdef _LP64 4632 push(rscratch1); // save r10, trashed by movptr() 4633 #endif 4634 push(rax); // save rax, 4635 // addr may contain rsp so we will have to adjust it based on the push 4636 // we just did (and on 64 bit we do two pushes) 4637 // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which 4638 // stores rax into addr which is backwards of what was intended. 4639 if (addr.uses(rsp)) { 4640 lea(rax, addr); 4641 pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord)); 4642 } else { 4643 pushptr(addr); 4644 } 4645 4646 ExternalAddress buffer((address) b); 4647 // pass msg argument 4648 // avoid using pushptr, as it modifies scratch registers 4649 // and our contract is not to modify anything 4650 movptr(rax, buffer.addr()); 4651 push(rax); 4652 4653 // call indirectly to solve generation ordering problem 4654 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); 4655 call(rax); 4656 // Caller pops the arguments (addr, message) and restores rax, r10. 4657 } 4658 4659 void MacroAssembler::verify_tlab() { 4660 #ifdef ASSERT 4661 if (UseTLAB && VerifyOops) { 4662 Label next, ok; 4663 Register t1 = rsi; 4664 Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread); 4665 4666 push(t1); 4667 NOT_LP64(push(thread_reg)); 4668 NOT_LP64(get_thread(thread_reg)); 4669 4670 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 4671 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset()))); 4672 jcc(Assembler::aboveEqual, next); 4673 STOP("assert(top >= start)"); 4674 should_not_reach_here(); 4675 4676 bind(next); 4677 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset()))); 4678 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); 4679 jcc(Assembler::aboveEqual, ok); 4680 STOP("assert(top <= end)"); 4681 should_not_reach_here(); 4682 4683 bind(ok); 4684 NOT_LP64(pop(thread_reg)); 4685 pop(t1); 4686 } 4687 #endif 4688 } 4689 4690 class ControlWord { 4691 public: 4692 int32_t _value; 4693 4694 int rounding_control() const { return (_value >> 10) & 3 ; } 4695 int precision_control() const { return (_value >> 8) & 3 ; } 4696 bool precision() const { return ((_value >> 5) & 1) != 0; } 4697 bool underflow() const { return ((_value >> 4) & 1) != 0; } 4698 bool overflow() const { return ((_value >> 3) & 1) != 0; } 4699 bool zero_divide() const { return ((_value >> 2) & 1) != 0; } 4700 bool denormalized() const { return ((_value >> 1) & 1) != 0; } 4701 bool invalid() const { return ((_value >> 0) & 1) != 0; } 4702 4703 void print() const { 4704 // rounding control 4705 const char* rc; 4706 switch (rounding_control()) { 4707 case 0: rc = "round near"; break; 4708 case 1: rc = "round down"; break; 4709 case 2: rc = "round up "; break; 4710 case 3: rc = "chop "; break; 4711 }; 4712 // precision control 4713 const char* pc; 4714 switch (precision_control()) { 4715 case 0: pc = "24 bits "; break; 4716 case 1: pc = "reserved"; break; 4717 case 2: pc = "53 bits "; break; 4718 case 3: pc = "64 bits "; break; 4719 }; 4720 // flags 4721 char f[9]; 4722 f[0] = ' '; 4723 f[1] = ' '; 4724 f[2] = (precision ()) ? 'P' : 'p'; 4725 f[3] = (underflow ()) ? 'U' : 'u'; 4726 f[4] = (overflow ()) ? 'O' : 'o'; 4727 f[5] = (zero_divide ()) ? 'Z' : 'z'; 4728 f[6] = (denormalized()) ? 'D' : 'd'; 4729 f[7] = (invalid ()) ? 'I' : 'i'; 4730 f[8] = '\x0'; 4731 // output 4732 printf("%04x masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc); 4733 } 4734 4735 }; 4736 4737 class StatusWord { 4738 public: 4739 int32_t _value; 4740 4741 bool busy() const { return ((_value >> 15) & 1) != 0; } 4742 bool C3() const { return ((_value >> 14) & 1) != 0; } 4743 bool C2() const { return ((_value >> 10) & 1) != 0; } 4744 bool C1() const { return ((_value >> 9) & 1) != 0; } 4745 bool C0() const { return ((_value >> 8) & 1) != 0; } 4746 int top() const { return (_value >> 11) & 7 ; } 4747 bool error_status() const { return ((_value >> 7) & 1) != 0; } 4748 bool stack_fault() const { return ((_value >> 6) & 1) != 0; } 4749 bool precision() const { return ((_value >> 5) & 1) != 0; } 4750 bool underflow() const { return ((_value >> 4) & 1) != 0; } 4751 bool overflow() const { return ((_value >> 3) & 1) != 0; } 4752 bool zero_divide() const { return ((_value >> 2) & 1) != 0; } 4753 bool denormalized() const { return ((_value >> 1) & 1) != 0; } 4754 bool invalid() const { return ((_value >> 0) & 1) != 0; } 4755 4756 void print() const { 4757 // condition codes 4758 char c[5]; 4759 c[0] = (C3()) ? '3' : '-'; 4760 c[1] = (C2()) ? '2' : '-'; 4761 c[2] = (C1()) ? '1' : '-'; 4762 c[3] = (C0()) ? '0' : '-'; 4763 c[4] = '\x0'; 4764 // flags 4765 char f[9]; 4766 f[0] = (error_status()) ? 'E' : '-'; 4767 f[1] = (stack_fault ()) ? 'S' : '-'; 4768 f[2] = (precision ()) ? 'P' : '-'; 4769 f[3] = (underflow ()) ? 'U' : '-'; 4770 f[4] = (overflow ()) ? 'O' : '-'; 4771 f[5] = (zero_divide ()) ? 'Z' : '-'; 4772 f[6] = (denormalized()) ? 'D' : '-'; 4773 f[7] = (invalid ()) ? 'I' : '-'; 4774 f[8] = '\x0'; 4775 // output 4776 printf("%04x flags = %s, cc = %s, top = %d", _value & 0xFFFF, f, c, top()); 4777 } 4778 4779 }; 4780 4781 class TagWord { 4782 public: 4783 int32_t _value; 4784 4785 int tag_at(int i) const { return (_value >> (i*2)) & 3; } 4786 4787 void print() const { 4788 printf("%04x", _value & 0xFFFF); 4789 } 4790 4791 }; 4792 4793 class FPU_Register { 4794 public: 4795 int32_t _m0; 4796 int32_t _m1; 4797 int16_t _ex; 4798 4799 bool is_indefinite() const { 4800 return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0; 4801 } 4802 4803 void print() const { 4804 char sign = (_ex < 0) ? '-' : '+'; 4805 const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : " "; 4806 printf("%c%04hx.%08x%08x %s", sign, _ex, _m1, _m0, kind); 4807 }; 4808 4809 }; 4810 4811 class FPU_State { 4812 public: 4813 enum { 4814 register_size = 10, 4815 number_of_registers = 8, 4816 register_mask = 7 4817 }; 4818 4819 ControlWord _control_word; 4820 StatusWord _status_word; 4821 TagWord _tag_word; 4822 int32_t _error_offset; 4823 int32_t _error_selector; 4824 int32_t _data_offset; 4825 int32_t _data_selector; 4826 int8_t _register[register_size * number_of_registers]; 4827 4828 int tag_for_st(int i) const { return _tag_word.tag_at((_status_word.top() + i) & register_mask); } 4829 FPU_Register* st(int i) const { return (FPU_Register*)&_register[register_size * i]; } 4830 4831 const char* tag_as_string(int tag) const { 4832 switch (tag) { 4833 case 0: return "valid"; 4834 case 1: return "zero"; 4835 case 2: return "special"; 4836 case 3: return "empty"; 4837 } 4838 ShouldNotReachHere(); 4839 return NULL; 4840 } 4841 4842 void print() const { 4843 // print computation registers 4844 { int t = _status_word.top(); 4845 for (int i = 0; i < number_of_registers; i++) { 4846 int j = (i - t) & register_mask; 4847 printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j); 4848 st(j)->print(); 4849 printf(" %s\n", tag_as_string(_tag_word.tag_at(i))); 4850 } 4851 } 4852 printf("\n"); 4853 // print control registers 4854 printf("ctrl = "); _control_word.print(); printf("\n"); 4855 printf("stat = "); _status_word .print(); printf("\n"); 4856 printf("tags = "); _tag_word .print(); printf("\n"); 4857 } 4858 4859 }; 4860 4861 class Flag_Register { 4862 public: 4863 int32_t _value; 4864 4865 bool overflow() const { return ((_value >> 11) & 1) != 0; } 4866 bool direction() const { return ((_value >> 10) & 1) != 0; } 4867 bool sign() const { return ((_value >> 7) & 1) != 0; } 4868 bool zero() const { return ((_value >> 6) & 1) != 0; } 4869 bool auxiliary_carry() const { return ((_value >> 4) & 1) != 0; } 4870 bool parity() const { return ((_value >> 2) & 1) != 0; } 4871 bool carry() const { return ((_value >> 0) & 1) != 0; } 4872 4873 void print() const { 4874 // flags 4875 char f[8]; 4876 f[0] = (overflow ()) ? 'O' : '-'; 4877 f[1] = (direction ()) ? 'D' : '-'; 4878 f[2] = (sign ()) ? 'S' : '-'; 4879 f[3] = (zero ()) ? 'Z' : '-'; 4880 f[4] = (auxiliary_carry()) ? 'A' : '-'; 4881 f[5] = (parity ()) ? 'P' : '-'; 4882 f[6] = (carry ()) ? 'C' : '-'; 4883 f[7] = '\x0'; 4884 // output 4885 printf("%08x flags = %s", _value, f); 4886 } 4887 4888 }; 4889 4890 class IU_Register { 4891 public: 4892 int32_t _value; 4893 4894 void print() const { 4895 printf("%08x %11d", _value, _value); 4896 } 4897 4898 }; 4899 4900 class IU_State { 4901 public: 4902 Flag_Register _eflags; 4903 IU_Register _rdi; 4904 IU_Register _rsi; 4905 IU_Register _rbp; 4906 IU_Register _rsp; 4907 IU_Register _rbx; 4908 IU_Register _rdx; 4909 IU_Register _rcx; 4910 IU_Register _rax; 4911 4912 void print() const { 4913 // computation registers 4914 printf("rax, = "); _rax.print(); printf("\n"); 4915 printf("rbx, = "); _rbx.print(); printf("\n"); 4916 printf("rcx = "); _rcx.print(); printf("\n"); 4917 printf("rdx = "); _rdx.print(); printf("\n"); 4918 printf("rdi = "); _rdi.print(); printf("\n"); 4919 printf("rsi = "); _rsi.print(); printf("\n"); 4920 printf("rbp, = "); _rbp.print(); printf("\n"); 4921 printf("rsp = "); _rsp.print(); printf("\n"); 4922 printf("\n"); 4923 // control registers 4924 printf("flgs = "); _eflags.print(); printf("\n"); 4925 } 4926 }; 4927 4928 4929 class CPU_State { 4930 public: 4931 FPU_State _fpu_state; 4932 IU_State _iu_state; 4933 4934 void print() const { 4935 printf("--------------------------------------------------\n"); 4936 _iu_state .print(); 4937 printf("\n"); 4938 _fpu_state.print(); 4939 printf("--------------------------------------------------\n"); 4940 } 4941 4942 }; 4943 4944 4945 static void _print_CPU_state(CPU_State* state) { 4946 state->print(); 4947 }; 4948 4949 4950 void MacroAssembler::print_CPU_state() { 4951 push_CPU_state(); 4952 push(rsp); // pass CPU state 4953 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state))); 4954 addptr(rsp, wordSize); // discard argument 4955 pop_CPU_state(); 4956 } 4957 4958 4959 static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) { 4960 static int counter = 0; 4961 FPU_State* fs = &state->_fpu_state; 4962 counter++; 4963 // For leaf calls, only verify that the top few elements remain empty. 4964 // We only need 1 empty at the top for C2 code. 4965 if( stack_depth < 0 ) { 4966 if( fs->tag_for_st(7) != 3 ) { 4967 printf("FPR7 not empty\n"); 4968 state->print(); 4969 assert(false, "error"); 4970 return false; 4971 } 4972 return true; // All other stack states do not matter 4973 } 4974 4975 assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std, 4976 "bad FPU control word"); 4977 4978 // compute stack depth 4979 int i = 0; 4980 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) < 3) i++; 4981 int d = i; 4982 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++; 4983 // verify findings 4984 if (i != FPU_State::number_of_registers) { 4985 // stack not contiguous 4986 printf("%s: stack not contiguous at ST%d\n", s, i); 4987 state->print(); 4988 assert(false, "error"); 4989 return false; 4990 } 4991 // check if computed stack depth corresponds to expected stack depth 4992 if (stack_depth < 0) { 4993 // expected stack depth is -stack_depth or less 4994 if (d > -stack_depth) { 4995 // too many elements on the stack 4996 printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d); 4997 state->print(); 4998 assert(false, "error"); 4999 return false; 5000 } 5001 } else { 5002 // expected stack depth is stack_depth 5003 if (d != stack_depth) { 5004 // wrong stack depth 5005 printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d); 5006 state->print(); 5007 assert(false, "error"); 5008 return false; 5009 } 5010 } 5011 // everything is cool 5012 return true; 5013 } 5014 5015 5016 void MacroAssembler::verify_FPU(int stack_depth, const char* s) { 5017 if (!VerifyFPU) return; 5018 push_CPU_state(); 5019 push(rsp); // pass CPU state 5020 ExternalAddress msg((address) s); 5021 // pass message string s 5022 pushptr(msg.addr()); 5023 push(stack_depth); // pass stack depth 5024 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU))); 5025 addptr(rsp, 3 * wordSize); // discard arguments 5026 // check for error 5027 { Label L; 5028 testl(rax, rax); 5029 jcc(Assembler::notZero, L); 5030 int3(); // break if error condition 5031 bind(L); 5032 } 5033 pop_CPU_state(); 5034 } 5035 5036 void MacroAssembler::restore_cpu_control_state_after_jni() { 5037 // Either restore the MXCSR register after returning from the JNI Call 5038 // or verify that it wasn't changed (with -Xcheck:jni flag). 5039 if (VM_Version::supports_sse()) { 5040 if (RestoreMXCSROnJNICalls) { 5041 ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std())); 5042 } else if (CheckJNICalls) { 5043 call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry())); 5044 } 5045 } 5046 // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty. 5047 vzeroupper(); 5048 // Reset k1 to 0xffff. 5049 5050 #ifdef COMPILER2 5051 if (PostLoopMultiversioning && VM_Version::supports_evex()) { 5052 push(rcx); 5053 movl(rcx, 0xffff); 5054 kmovwl(k1, rcx); 5055 pop(rcx); 5056 } 5057 #endif // COMPILER2 5058 5059 #ifndef _LP64 5060 // Either restore the x87 floating pointer control word after returning 5061 // from the JNI call or verify that it wasn't changed. 5062 if (CheckJNICalls) { 5063 call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry())); 5064 } 5065 #endif // _LP64 5066 } 5067 5068 // ((OopHandle)result).resolve(); 5069 void MacroAssembler::resolve_oop_handle(Register result, Register tmp) { 5070 assert_different_registers(result, tmp); 5071 5072 // Only 64 bit platforms support GCs that require a tmp register 5073 // Only IN_HEAP loads require a thread_tmp register 5074 // OopHandle::resolve is an indirection like jobject. 5075 access_load_at(T_OBJECT, IN_NATIVE, 5076 result, Address(result, 0), tmp, /*tmp_thread*/noreg); 5077 } 5078 5079 void MacroAssembler::load_mirror(Register mirror, Register method, Register tmp) { 5080 // get mirror 5081 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 5082 movptr(mirror, Address(method, Method::const_offset())); 5083 movptr(mirror, Address(mirror, ConstMethod::constants_offset())); 5084 movptr(mirror, Address(mirror, ConstantPool::pool_holder_offset_in_bytes())); 5085 movptr(mirror, Address(mirror, mirror_offset)); 5086 resolve_oop_handle(mirror, tmp); 5087 } 5088 5089 void MacroAssembler::load_metadata(Register dst, Register src) { 5090 if (UseCompressedClassPointers) { 5091 movl(dst, Address(src, oopDesc::klass_offset_in_bytes())); 5092 } else { 5093 movptr(dst, Address(src, oopDesc::klass_offset_in_bytes())); 5094 } 5095 } 5096 5097 void MacroAssembler::load_storage_props(Register dst, Register src) { 5098 load_metadata(dst, src); 5099 if (UseCompressedClassPointers) { 5100 shrl(dst, oopDesc::narrow_storage_props_shift); 5101 } else { 5102 shrq(dst, oopDesc::wide_storage_props_shift); 5103 } 5104 } 5105 5106 void MacroAssembler::load_klass(Register dst, Register src) { 5107 load_metadata(dst, src); 5108 #ifdef _LP64 5109 if (UseCompressedClassPointers) { 5110 decode_klass_not_null(dst); 5111 } else 5112 #endif 5113 { 5114 #ifdef _LP64 5115 shlq(dst, oopDesc::storage_props_nof_bits); 5116 shrq(dst, oopDesc::storage_props_nof_bits); 5117 #else 5118 andl(dst, oopDesc::wide_klass_mask()); 5119 #endif 5120 } 5121 } 5122 5123 void MacroAssembler::load_prototype_header(Register dst, Register src) { 5124 load_klass(dst, src); 5125 movptr(dst, Address(dst, Klass::prototype_header_offset())); 5126 } 5127 5128 void MacroAssembler::store_klass(Register dst, Register src) { 5129 #ifdef _LP64 5130 if (UseCompressedClassPointers) { 5131 encode_klass_not_null(src); 5132 movl(Address(dst, oopDesc::klass_offset_in_bytes()), src); 5133 } else 5134 #endif 5135 movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src); 5136 } 5137 5138 void MacroAssembler::access_load_at(BasicType type, DecoratorSet decorators, Register dst, Address src, 5139 Register tmp1, Register thread_tmp) { 5140 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 5141 decorators = AccessInternal::decorator_fixup(decorators); 5142 bool as_raw = (decorators & AS_RAW) != 0; 5143 if (as_raw) { 5144 bs->BarrierSetAssembler::load_at(this, decorators, type, dst, src, tmp1, thread_tmp); 5145 } else { 5146 bs->load_at(this, decorators, type, dst, src, tmp1, thread_tmp); 5147 } 5148 } 5149 5150 void MacroAssembler::access_store_at(BasicType type, DecoratorSet decorators, Address dst, Register src, 5151 Register tmp1, Register tmp2, Register tmp3) { 5152 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 5153 decorators = AccessInternal::decorator_fixup(decorators); 5154 bool as_raw = (decorators & AS_RAW) != 0; 5155 if (as_raw) { 5156 bs->BarrierSetAssembler::store_at(this, decorators, type, dst, src, tmp1, tmp2, tmp3); 5157 } else { 5158 bs->store_at(this, decorators, type, dst, src, tmp1, tmp2, tmp3); 5159 } 5160 } 5161 5162 void MacroAssembler::resolve(DecoratorSet decorators, Register obj) { 5163 // Use stronger ACCESS_WRITE|ACCESS_READ by default. 5164 if ((decorators & (ACCESS_READ | ACCESS_WRITE)) == 0) { 5165 decorators |= ACCESS_READ | ACCESS_WRITE; 5166 } 5167 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 5168 return bs->resolve(this, decorators, obj); 5169 } 5170 5171 void MacroAssembler::load_heap_oop(Register dst, Address src, Register tmp1, 5172 Register thread_tmp, DecoratorSet decorators) { 5173 access_load_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, thread_tmp); 5174 } 5175 5176 // Doesn't do verfication, generates fixed size code 5177 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src, Register tmp1, 5178 Register thread_tmp, DecoratorSet decorators) { 5179 access_load_at(T_OBJECT, IN_HEAP | IS_NOT_NULL | decorators, dst, src, tmp1, thread_tmp); 5180 } 5181 5182 void MacroAssembler::store_heap_oop(Address dst, Register src, Register tmp1, 5183 Register tmp2, Register tmp3, DecoratorSet decorators) { 5184 access_store_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, tmp2, tmp3); 5185 } 5186 5187 // Used for storing NULLs. 5188 void MacroAssembler::store_heap_oop_null(Address dst) { 5189 access_store_at(T_OBJECT, IN_HEAP, dst, noreg, noreg, noreg, noreg); 5190 } 5191 5192 #ifdef _LP64 5193 void MacroAssembler::store_klass_gap(Register dst, Register src) { 5194 if (UseCompressedClassPointers) { 5195 // Store to klass gap in destination 5196 movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src); 5197 } 5198 } 5199 5200 #ifdef ASSERT 5201 void MacroAssembler::verify_heapbase(const char* msg) { 5202 assert (UseCompressedOops, "should be compressed"); 5203 assert (Universe::heap() != NULL, "java heap should be initialized"); 5204 if (CheckCompressedOops) { 5205 Label ok; 5206 push(rscratch1); // cmpptr trashes rscratch1 5207 cmpptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr())); 5208 jcc(Assembler::equal, ok); 5209 STOP(msg); 5210 bind(ok); 5211 pop(rscratch1); 5212 } 5213 } 5214 #endif 5215 5216 // Algorithm must match oop.inline.hpp encode_heap_oop. 5217 void MacroAssembler::encode_heap_oop(Register r) { 5218 #ifdef ASSERT 5219 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?"); 5220 #endif 5221 verify_oop(r, "broken oop in encode_heap_oop"); 5222 if (Universe::narrow_oop_base() == NULL) { 5223 if (Universe::narrow_oop_shift() != 0) { 5224 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 5225 shrq(r, LogMinObjAlignmentInBytes); 5226 } 5227 return; 5228 } 5229 testq(r, r); 5230 cmovq(Assembler::equal, r, r12_heapbase); 5231 subq(r, r12_heapbase); 5232 shrq(r, LogMinObjAlignmentInBytes); 5233 } 5234 5235 void MacroAssembler::encode_heap_oop_not_null(Register r) { 5236 #ifdef ASSERT 5237 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?"); 5238 if (CheckCompressedOops) { 5239 Label ok; 5240 testq(r, r); 5241 jcc(Assembler::notEqual, ok); 5242 STOP("null oop passed to encode_heap_oop_not_null"); 5243 bind(ok); 5244 } 5245 #endif 5246 verify_oop(r, "broken oop in encode_heap_oop_not_null"); 5247 if (Universe::narrow_oop_base() != NULL) { 5248 subq(r, r12_heapbase); 5249 } 5250 if (Universe::narrow_oop_shift() != 0) { 5251 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 5252 shrq(r, LogMinObjAlignmentInBytes); 5253 } 5254 } 5255 5256 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) { 5257 #ifdef ASSERT 5258 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?"); 5259 if (CheckCompressedOops) { 5260 Label ok; 5261 testq(src, src); 5262 jcc(Assembler::notEqual, ok); 5263 STOP("null oop passed to encode_heap_oop_not_null2"); 5264 bind(ok); 5265 } 5266 #endif 5267 verify_oop(src, "broken oop in encode_heap_oop_not_null2"); 5268 if (dst != src) { 5269 movq(dst, src); 5270 } 5271 if (Universe::narrow_oop_base() != NULL) { 5272 subq(dst, r12_heapbase); 5273 } 5274 if (Universe::narrow_oop_shift() != 0) { 5275 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 5276 shrq(dst, LogMinObjAlignmentInBytes); 5277 } 5278 } 5279 5280 void MacroAssembler::decode_heap_oop(Register r) { 5281 #ifdef ASSERT 5282 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?"); 5283 #endif 5284 if (Universe::narrow_oop_base() == NULL) { 5285 if (Universe::narrow_oop_shift() != 0) { 5286 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 5287 shlq(r, LogMinObjAlignmentInBytes); 5288 } 5289 } else { 5290 Label done; 5291 shlq(r, LogMinObjAlignmentInBytes); 5292 jccb(Assembler::equal, done); 5293 addq(r, r12_heapbase); 5294 bind(done); 5295 } 5296 verify_oop(r, "broken oop in decode_heap_oop"); 5297 } 5298 5299 void MacroAssembler::decode_heap_oop_not_null(Register r) { 5300 // Note: it will change flags 5301 assert (UseCompressedOops, "should only be used for compressed headers"); 5302 assert (Universe::heap() != NULL, "java heap should be initialized"); 5303 // Cannot assert, unverified entry point counts instructions (see .ad file) 5304 // vtableStubs also counts instructions in pd_code_size_limit. 5305 // Also do not verify_oop as this is called by verify_oop. 5306 if (Universe::narrow_oop_shift() != 0) { 5307 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 5308 shlq(r, LogMinObjAlignmentInBytes); 5309 if (Universe::narrow_oop_base() != NULL) { 5310 addq(r, r12_heapbase); 5311 } 5312 } else { 5313 assert (Universe::narrow_oop_base() == NULL, "sanity"); 5314 } 5315 } 5316 5317 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) { 5318 // Note: it will change flags 5319 assert (UseCompressedOops, "should only be used for compressed headers"); 5320 assert (Universe::heap() != NULL, "java heap should be initialized"); 5321 // Cannot assert, unverified entry point counts instructions (see .ad file) 5322 // vtableStubs also counts instructions in pd_code_size_limit. 5323 // Also do not verify_oop as this is called by verify_oop. 5324 if (Universe::narrow_oop_shift() != 0) { 5325 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); 5326 if (LogMinObjAlignmentInBytes == Address::times_8) { 5327 leaq(dst, Address(r12_heapbase, src, Address::times_8, 0)); 5328 } else { 5329 if (dst != src) { 5330 movq(dst, src); 5331 } 5332 shlq(dst, LogMinObjAlignmentInBytes); 5333 if (Universe::narrow_oop_base() != NULL) { 5334 addq(dst, r12_heapbase); 5335 } 5336 } 5337 } else { 5338 assert (Universe::narrow_oop_base() == NULL, "sanity"); 5339 if (dst != src) { 5340 movq(dst, src); 5341 } 5342 } 5343 } 5344 5345 void MacroAssembler::encode_klass_not_null(Register r) { 5346 if (Universe::narrow_klass_base() != NULL) { 5347 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base. 5348 assert(r != r12_heapbase, "Encoding a klass in r12"); 5349 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base()); 5350 subq(r, r12_heapbase); 5351 } 5352 if (Universe::narrow_klass_shift() != 0) { 5353 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 5354 shrq(r, LogKlassAlignmentInBytes); 5355 } 5356 if (Universe::narrow_klass_base() != NULL) { 5357 reinit_heapbase(); 5358 } 5359 } 5360 5361 void MacroAssembler::encode_klass_not_null(Register dst, Register src) { 5362 if (dst == src) { 5363 encode_klass_not_null(src); 5364 } else { 5365 if (Universe::narrow_klass_base() != NULL) { 5366 mov64(dst, (int64_t)Universe::narrow_klass_base()); 5367 negq(dst); 5368 addq(dst, src); 5369 } else { 5370 movptr(dst, src); 5371 } 5372 if (Universe::narrow_klass_shift() != 0) { 5373 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 5374 shrq(dst, LogKlassAlignmentInBytes); 5375 } 5376 } 5377 } 5378 5379 // Function instr_size_for_decode_klass_not_null() counts the instructions 5380 // generated by decode_klass_not_null(register r) and reinit_heapbase(), 5381 // when (Universe::heap() != NULL). Hence, if the instructions they 5382 // generate change, then this method needs to be updated. 5383 int MacroAssembler::instr_size_for_decode_klass_not_null() { 5384 assert (UseCompressedClassPointers, "only for compressed klass ptrs"); 5385 if (Universe::narrow_klass_base() != NULL) { 5386 // mov64 + addq + shlq? + mov64 (for reinit_heapbase()). 5387 return (Universe::narrow_klass_shift() == 0 ? 20 : 24); 5388 } else { 5389 // longest load decode klass function, mov64, leaq 5390 return 16; 5391 } 5392 } 5393 5394 // !!! If the instructions that get generated here change then function 5395 // instr_size_for_decode_klass_not_null() needs to get updated. 5396 void MacroAssembler::decode_klass_not_null(Register r) { 5397 // Note: it will change flags 5398 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 5399 assert(r != r12_heapbase, "Decoding a klass in r12"); 5400 andl(r, oopDesc::compressed_klass_mask()); 5401 // Cannot assert, unverified entry point counts instructions (see .ad file) 5402 // vtableStubs also counts instructions in pd_code_size_limit. 5403 // Also do not verify_oop as this is called by verify_oop. 5404 if (Universe::narrow_klass_shift() != 0) { 5405 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 5406 shlq(r, LogKlassAlignmentInBytes); 5407 } 5408 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base. 5409 if (Universe::narrow_klass_base() != NULL) { 5410 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base()); 5411 addq(r, r12_heapbase); 5412 reinit_heapbase(); 5413 } 5414 } 5415 5416 void MacroAssembler::decode_klass_not_null(Register dst, Register src) { 5417 // Note: it will change flags 5418 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 5419 if (dst == src) { 5420 decode_klass_not_null(dst); 5421 } else { 5422 // Cannot assert, unverified entry point counts instructions (see .ad file) 5423 // vtableStubs also counts instructions in pd_code_size_limit. 5424 // Also do not verify_oop as this is called by verify_oop. 5425 andl(src, oopDesc::compressed_klass_mask()); 5426 mov64(dst, (int64_t)Universe::narrow_klass_base()); 5427 if (Universe::narrow_klass_shift() != 0) { 5428 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); 5429 assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?"); 5430 leaq(dst, Address(dst, src, Address::times_8, 0)); 5431 } else { 5432 addq(dst, src); 5433 } 5434 } 5435 } 5436 5437 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) { 5438 assert (UseCompressedOops, "should only be used for compressed headers"); 5439 assert (Universe::heap() != NULL, "java heap should be initialized"); 5440 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5441 int oop_index = oop_recorder()->find_index(obj); 5442 RelocationHolder rspec = oop_Relocation::spec(oop_index); 5443 mov_narrow_oop(dst, oop_index, rspec); 5444 } 5445 5446 void MacroAssembler::set_narrow_oop(Address dst, jobject obj) { 5447 assert (UseCompressedOops, "should only be used for compressed headers"); 5448 assert (Universe::heap() != NULL, "java heap should be initialized"); 5449 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5450 int oop_index = oop_recorder()->find_index(obj); 5451 RelocationHolder rspec = oop_Relocation::spec(oop_index); 5452 mov_narrow_oop(dst, oop_index, rspec); 5453 } 5454 5455 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) { 5456 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 5457 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5458 int klass_index = oop_recorder()->find_index(k); 5459 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 5460 mov_narrow_oop(dst, Klass::encode_klass(k), rspec); 5461 } 5462 5463 void MacroAssembler::set_narrow_klass(Address dst, Klass* k) { 5464 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 5465 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5466 int klass_index = oop_recorder()->find_index(k); 5467 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 5468 mov_narrow_oop(dst, Klass::encode_klass(k), rspec); 5469 } 5470 5471 void MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) { 5472 assert (UseCompressedOops, "should only be used for compressed headers"); 5473 assert (Universe::heap() != NULL, "java heap should be initialized"); 5474 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5475 int oop_index = oop_recorder()->find_index(obj); 5476 RelocationHolder rspec = oop_Relocation::spec(oop_index); 5477 Assembler::cmp_narrow_oop(dst, oop_index, rspec); 5478 } 5479 5480 void MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) { 5481 assert (UseCompressedOops, "should only be used for compressed headers"); 5482 assert (Universe::heap() != NULL, "java heap should be initialized"); 5483 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5484 int oop_index = oop_recorder()->find_index(obj); 5485 RelocationHolder rspec = oop_Relocation::spec(oop_index); 5486 Assembler::cmp_narrow_oop(dst, oop_index, rspec); 5487 } 5488 5489 void MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) { 5490 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 5491 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5492 int klass_index = oop_recorder()->find_index(k); 5493 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 5494 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec); 5495 } 5496 5497 void MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) { 5498 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 5499 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); 5500 int klass_index = oop_recorder()->find_index(k); 5501 RelocationHolder rspec = metadata_Relocation::spec(klass_index); 5502 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec); 5503 } 5504 5505 void MacroAssembler::reinit_heapbase() { 5506 if (UseCompressedOops || UseCompressedClassPointers) { 5507 if (Universe::heap() != NULL) { 5508 if (Universe::narrow_oop_base() == NULL) { 5509 MacroAssembler::xorptr(r12_heapbase, r12_heapbase); 5510 } else { 5511 mov64(r12_heapbase, (int64_t)Universe::narrow_ptrs_base()); 5512 } 5513 } else { 5514 movptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr())); 5515 } 5516 } 5517 } 5518 5519 #endif // _LP64 5520 5521 // C2 compiled method's prolog code. 5522 void MacroAssembler::verified_entry(Compile* C, int sp_inc) { 5523 int framesize = C->frame_size_in_bytes(); 5524 int bangsize = C->bang_size_in_bytes(); 5525 bool fp_mode_24b = C->in_24_bit_fp_mode(); 5526 int stack_bang_size = C->need_stack_bang(bangsize) ? bangsize : 0; 5527 bool is_stub = C->stub_function() != NULL; 5528 5529 // WARNING: Initial instruction MUST be 5 bytes or longer so that 5530 // NativeJump::patch_verified_entry will be able to patch out the entry 5531 // code safely. The push to verify stack depth is ok at 5 bytes, 5532 // the frame allocation can be either 3 or 6 bytes. So if we don't do 5533 // stack bang then we must use the 6 byte frame allocation even if 5534 // we have no frame. :-( 5535 assert(stack_bang_size >= framesize || stack_bang_size <= 0, "stack bang size incorrect"); 5536 5537 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned"); 5538 // Remove word for return addr 5539 framesize -= wordSize; 5540 stack_bang_size -= wordSize; 5541 5542 // Calls to C2R adapters often do not accept exceptional returns. 5543 // We require that their callers must bang for them. But be careful, because 5544 // some VM calls (such as call site linkage) can use several kilobytes of 5545 // stack. But the stack safety zone should account for that. 5546 // See bugs 4446381, 4468289, 4497237. 5547 if (stack_bang_size > 0) { 5548 generate_stack_overflow_check(stack_bang_size); 5549 5550 // We always push rbp, so that on return to interpreter rbp, will be 5551 // restored correctly and we can correct the stack. 5552 push(rbp); 5553 // Save caller's stack pointer into RBP if the frame pointer is preserved. 5554 if (PreserveFramePointer) { 5555 mov(rbp, rsp); 5556 } 5557 // Remove word for ebp 5558 framesize -= wordSize; 5559 5560 // Create frame 5561 if (framesize) { 5562 subptr(rsp, framesize); 5563 } 5564 } else { 5565 // Create frame (force generation of a 4 byte immediate value) 5566 subptr_imm32(rsp, framesize); 5567 5568 // Save RBP register now. 5569 framesize -= wordSize; 5570 movptr(Address(rsp, framesize), rbp); 5571 // Save caller's stack pointer into RBP if the frame pointer is preserved. 5572 if (PreserveFramePointer) { 5573 movptr(rbp, rsp); 5574 if (framesize > 0) { 5575 addptr(rbp, framesize); 5576 } 5577 } 5578 } 5579 5580 if (C->needs_stack_repair()) { 5581 // Save stack increment (also account for fixed framesize and rbp) 5582 assert((sp_inc & (StackAlignmentInBytes-1)) == 0, "stack increment not aligned"); 5583 movptr(Address(rsp, C->sp_inc_offset()), sp_inc + framesize + wordSize); 5584 } 5585 5586 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth 5587 framesize -= wordSize; 5588 movptr(Address(rsp, framesize), (int32_t)0xbadb100d); 5589 } 5590 5591 #ifndef _LP64 5592 // If method sets FPU control word do it now 5593 if (fp_mode_24b) { 5594 fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24())); 5595 } 5596 if (UseSSE >= 2 && VerifyFPU) { 5597 verify_FPU(0, "FPU stack must be clean on entry"); 5598 } 5599 #endif 5600 5601 #ifdef ASSERT 5602 if (VerifyStackAtCalls) { 5603 Label L; 5604 push(rax); 5605 mov(rax, rsp); 5606 andptr(rax, StackAlignmentInBytes-1); 5607 cmpptr(rax, StackAlignmentInBytes-wordSize); 5608 pop(rax); 5609 jcc(Assembler::equal, L); 5610 STOP("Stack is not properly aligned!"); 5611 bind(L); 5612 } 5613 #endif 5614 5615 if (!is_stub) { 5616 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 5617 bs->nmethod_entry_barrier(this); 5618 } 5619 } 5620 5621 // clear memory of size 'cnt' qwords, starting at 'base' using XMM/YMM registers 5622 void MacroAssembler::xmm_clear_mem(Register base, Register cnt, Register val, XMMRegister xtmp) { 5623 // cnt - number of qwords (8-byte words). 5624 // base - start address, qword aligned. 5625 Label L_zero_64_bytes, L_loop, L_sloop, L_tail, L_end; 5626 movdq(xtmp, val); 5627 if (UseAVX >= 2) { 5628 punpcklqdq(xtmp, xtmp); 5629 vinserti128_high(xtmp, xtmp); 5630 } else { 5631 punpcklqdq(xtmp, xtmp); 5632 } 5633 jmp(L_zero_64_bytes); 5634 5635 BIND(L_loop); 5636 if (UseAVX >= 2) { 5637 vmovdqu(Address(base, 0), xtmp); 5638 vmovdqu(Address(base, 32), xtmp); 5639 } else { 5640 movdqu(Address(base, 0), xtmp); 5641 movdqu(Address(base, 16), xtmp); 5642 movdqu(Address(base, 32), xtmp); 5643 movdqu(Address(base, 48), xtmp); 5644 } 5645 addptr(base, 64); 5646 5647 BIND(L_zero_64_bytes); 5648 subptr(cnt, 8); 5649 jccb(Assembler::greaterEqual, L_loop); 5650 addptr(cnt, 4); 5651 jccb(Assembler::less, L_tail); 5652 // Copy trailing 32 bytes 5653 if (UseAVX >= 2) { 5654 vmovdqu(Address(base, 0), xtmp); 5655 } else { 5656 movdqu(Address(base, 0), xtmp); 5657 movdqu(Address(base, 16), xtmp); 5658 } 5659 addptr(base, 32); 5660 subptr(cnt, 4); 5661 5662 BIND(L_tail); 5663 addptr(cnt, 4); 5664 jccb(Assembler::lessEqual, L_end); 5665 decrement(cnt); 5666 5667 BIND(L_sloop); 5668 movq(Address(base, 0), xtmp); 5669 addptr(base, 8); 5670 decrement(cnt); 5671 jccb(Assembler::greaterEqual, L_sloop); 5672 BIND(L_end); 5673 } 5674 5675 void MacroAssembler::store_value_type_fields_to_buf(ciValueKlass* vk) { 5676 #ifndef _LP64 5677 super_call_VM_leaf(StubRoutines::store_value_type_fields_to_buf()); 5678 #else 5679 // A value type might be returned. If fields are in registers we 5680 // need to allocate a value type instance and initialize it with 5681 // the value of the fields. 5682 Label skip, slow_case; 5683 // We only need a new buffered value if a new one is not returned 5684 testptr(rax, 1); 5685 jcc(Assembler::zero, skip); 5686 5687 // Try to allocate a new buffered value (from the heap) 5688 if (UseTLAB) { 5689 // FIXME -- for smaller code, the inline allocation (and the slow case) should be moved inside the pack handler. 5690 if (vk != NULL) { 5691 // Called from C1, where the return type is statically known. 5692 movptr(rbx, (intptr_t)vk->get_ValueKlass()); 5693 jint lh = vk->layout_helper(); 5694 assert(lh != Klass::_lh_neutral_value, "inline class in return type must have been resolved"); 5695 movl(r14, lh); 5696 } else { 5697 // Call from interpreter. RAX contains ((the ValueKlass* of the return type) | 0x01) 5698 mov(rbx, rax); 5699 andptr(rbx, -2); 5700 movl(r14, Address(rbx, Klass::layout_helper_offset())); 5701 } 5702 5703 movptr(r13, Address(r15_thread, in_bytes(JavaThread::tlab_top_offset()))); 5704 lea(r14, Address(r13, r14, Address::times_1)); 5705 cmpptr(r14, Address(r15_thread, in_bytes(JavaThread::tlab_end_offset()))); 5706 jcc(Assembler::above, slow_case); 5707 movptr(Address(r15_thread, in_bytes(JavaThread::tlab_top_offset())), r14); 5708 movptr(Address(r13, oopDesc::mark_offset_in_bytes()), (intptr_t)markOopDesc::always_locked_prototype()); 5709 5710 xorl(rax, rax); // use zero reg to clear memory (shorter code) 5711 store_klass_gap(r13, rax); // zero klass gap for compressed oops 5712 5713 if (vk == NULL) { 5714 // store_klass corrupts rbx, so save it in rax for later use (interpreter case only). 5715 mov(rax, rbx); 5716 } 5717 store_klass(r13, rbx); // klass 5718 5719 // We have our new buffered value, initialize its fields with a 5720 // value class specific handler 5721 if (vk != NULL) { 5722 // FIXME -- do the packing in-line to avoid the runtime call 5723 mov(rax, r13); 5724 call(RuntimeAddress(vk->pack_handler())); 5725 } else { 5726 movptr(rbx, Address(rax, InstanceKlass::adr_valueklass_fixed_block_offset())); 5727 movptr(rbx, Address(rbx, ValueKlass::pack_handler_offset())); 5728 mov(rax, r13); 5729 call(rbx); 5730 } 5731 jmp(skip); 5732 } 5733 5734 bind(slow_case); 5735 // We failed to allocate a new value, fall back to a runtime 5736 // call. Some oop field may be live in some registers but we can't 5737 // tell. That runtime call will take care of preserving them 5738 // across a GC if there's one. 5739 super_call_VM_leaf(StubRoutines::store_value_type_fields_to_buf()); 5740 bind(skip); 5741 #endif 5742 } 5743 5744 5745 // Move a value between registers/stack slots and update the reg_state 5746 bool MacroAssembler::move_helper(VMReg from, VMReg to, BasicType bt, RegState reg_state[], int ret_off, int extra_stack_offset) { 5747 if (reg_state[to->value()] == reg_written) { 5748 return true; // Already written 5749 } 5750 if (from != to && bt != T_VOID) { 5751 if (reg_state[to->value()] == reg_readonly) { 5752 return false; // Not yet writable 5753 } 5754 if (from->is_reg()) { 5755 if (to->is_reg()) { 5756 if (from->is_XMMRegister()) { 5757 if (bt == T_DOUBLE) { 5758 movdbl(to->as_XMMRegister(), from->as_XMMRegister()); 5759 } else { 5760 assert(bt == T_FLOAT, "must be float"); 5761 movflt(to->as_XMMRegister(), from->as_XMMRegister()); 5762 } 5763 } else { 5764 movq(to->as_Register(), from->as_Register()); 5765 } 5766 } else { 5767 int st_off = to->reg2stack() * VMRegImpl::stack_slot_size + extra_stack_offset; 5768 assert(st_off != ret_off, "overwriting return address at %d", st_off); 5769 Address to_addr = Address(rsp, st_off); 5770 if (from->is_XMMRegister()) { 5771 if (bt == T_DOUBLE) { 5772 movdbl(to_addr, from->as_XMMRegister()); 5773 } else { 5774 assert(bt == T_FLOAT, "must be float"); 5775 movflt(to_addr, from->as_XMMRegister()); 5776 } 5777 } else { 5778 movq(to_addr, from->as_Register()); 5779 } 5780 } 5781 } else { 5782 Address from_addr = Address(rsp, from->reg2stack() * VMRegImpl::stack_slot_size + extra_stack_offset); 5783 if (to->is_reg()) { 5784 if (to->is_XMMRegister()) { 5785 if (bt == T_DOUBLE) { 5786 movdbl(to->as_XMMRegister(), from_addr); 5787 } else { 5788 assert(bt == T_FLOAT, "must be float"); 5789 movflt(to->as_XMMRegister(), from_addr); 5790 } 5791 } else { 5792 movq(to->as_Register(), from_addr); 5793 } 5794 } else { 5795 int st_off = to->reg2stack() * VMRegImpl::stack_slot_size + extra_stack_offset; 5796 assert(st_off != ret_off, "overwriting return address at %d", st_off); 5797 movq(r13, from_addr); 5798 movq(Address(rsp, st_off), r13); 5799 } 5800 } 5801 } 5802 // Update register states 5803 reg_state[from->value()] = reg_writable; 5804 reg_state[to->value()] = reg_written; 5805 return true; 5806 } 5807 5808 // Read all fields from a value type oop and store the values in registers/stack slots 5809 bool MacroAssembler::unpack_value_helper(const GrowableArray<SigEntry>* sig, int& sig_index, VMReg from, VMRegPair* regs_to, 5810 int& to_index, RegState reg_state[], int ret_off, int extra_stack_offset) { 5811 Register fromReg = from->is_reg() ? from->as_Register() : noreg; 5812 assert(sig->at(sig_index)._bt == T_VOID, "should be at end delimiter"); 5813 5814 int vt = 1; 5815 bool done = true; 5816 bool mark_done = true; 5817 do { 5818 sig_index--; 5819 BasicType bt = sig->at(sig_index)._bt; 5820 if (bt == T_VALUETYPE) { 5821 vt--; 5822 } else if (bt == T_VOID && 5823 sig->at(sig_index-1)._bt != T_LONG && 5824 sig->at(sig_index-1)._bt != T_DOUBLE) { 5825 vt++; 5826 } else if (SigEntry::is_reserved_entry(sig, sig_index)) { 5827 to_index--; // Ignore this 5828 } else { 5829 assert(to_index >= 0, "invalid to_index"); 5830 VMRegPair pair_to = regs_to[to_index--]; 5831 VMReg to = pair_to.first(); 5832 5833 if (bt == T_VOID) continue; 5834 5835 int idx = (int)to->value(); 5836 if (reg_state[idx] == reg_readonly) { 5837 if (idx != from->value()) { 5838 mark_done = false; 5839 } 5840 done = false; 5841 continue; 5842 } else if (reg_state[idx] == reg_written) { 5843 continue; 5844 } else { 5845 assert(reg_state[idx] == reg_writable, "must be writable"); 5846 reg_state[idx] = reg_written; 5847 } 5848 5849 if (fromReg == noreg) { 5850 int st_off = from->reg2stack() * VMRegImpl::stack_slot_size + extra_stack_offset; 5851 movq(r10, Address(rsp, st_off)); 5852 fromReg = r10; 5853 } 5854 5855 int off = sig->at(sig_index)._offset; 5856 assert(off > 0, "offset in object should be positive"); 5857 bool is_oop = (bt == T_OBJECT || bt == T_ARRAY); 5858 5859 Address fromAddr = Address(fromReg, off); 5860 bool is_signed = (bt != T_CHAR) && (bt != T_BOOLEAN); 5861 if (!to->is_XMMRegister()) { 5862 Register dst = to->is_stack() ? r13 : to->as_Register(); 5863 if (is_oop) { 5864 load_heap_oop(dst, fromAddr); 5865 } else { 5866 load_sized_value(dst, fromAddr, type2aelembytes(bt), is_signed); 5867 } 5868 if (to->is_stack()) { 5869 int st_off = to->reg2stack() * VMRegImpl::stack_slot_size + extra_stack_offset; 5870 assert(st_off != ret_off, "overwriting return address at %d", st_off); 5871 movq(Address(rsp, st_off), dst); 5872 } 5873 } else { 5874 if (bt == T_DOUBLE) { 5875 movdbl(to->as_XMMRegister(), fromAddr); 5876 } else { 5877 assert(bt == T_FLOAT, "must be float"); 5878 movflt(to->as_XMMRegister(), fromAddr); 5879 } 5880 } 5881 } 5882 } while (vt != 0); 5883 if (mark_done && reg_state[from->value()] != reg_written) { 5884 // This is okay because no one else will write to that slot 5885 reg_state[from->value()] = reg_writable; 5886 } 5887 return done; 5888 } 5889 5890 class ScalarizedValueArgsStream : public StackObj { 5891 const GrowableArray<SigEntry>* _sig_cc; 5892 int _sig_cc_index; 5893 const VMRegPair* _regs_cc; 5894 int _regs_cc_count; 5895 int _regs_cc_index; 5896 int _vt; 5897 DEBUG_ONLY(bool _finished); 5898 public: 5899 ScalarizedValueArgsStream(const GrowableArray<SigEntry>* sig_cc, int sig_cc_index, VMRegPair* regs_cc, int regs_cc_count, int regs_cc_index) : 5900 _sig_cc(sig_cc), _sig_cc_index(sig_cc_index), _regs_cc(regs_cc), _regs_cc_count(regs_cc_count), _regs_cc_index(regs_cc_index) { 5901 assert(_sig_cc->at(_sig_cc_index)._bt == T_VALUETYPE, "should be at end delimiter"); 5902 _vt = 1; 5903 DEBUG_ONLY(_finished = false); 5904 } 5905 5906 bool next(VMRegPair& pair, BasicType& bt) { 5907 assert(!_finished, "sanity"); 5908 do { 5909 _sig_cc_index++; 5910 bt = _sig_cc->at(_sig_cc_index)._bt; 5911 if (bt == T_VALUETYPE) { 5912 _vt++; 5913 } else if (bt == T_VOID && 5914 _sig_cc->at(_sig_cc_index-1)._bt != T_LONG && 5915 _sig_cc->at(_sig_cc_index-1)._bt != T_DOUBLE) { 5916 _vt--; 5917 } else if (SigEntry::is_reserved_entry(_sig_cc, _sig_cc_index)) { 5918 _regs_cc_index++; 5919 } else { 5920 assert(_regs_cc_index < _regs_cc_count, "must be"); 5921 pair = _regs_cc[_regs_cc_index++]; 5922 VMReg r1 = pair.first(); 5923 VMReg r2 = pair.second(); 5924 5925 if (!r1->is_valid()) { 5926 assert(!r2->is_valid(), "must be invalid"); 5927 } else { 5928 return true; 5929 } 5930 } 5931 } while (_vt != 0); 5932 5933 DEBUG_ONLY(_finished = true); 5934 return false; 5935 } 5936 5937 int sig_cc_index() {return _sig_cc_index;} 5938 int regs_cc_index() {return _regs_cc_index;} 5939 }; 5940 5941 static void skip_unpacked_fields(const GrowableArray<SigEntry>* sig, int& sig_index, VMRegPair* regs_from, int regs_from_count, int& from_index) { 5942 ScalarizedValueArgsStream stream(sig, sig_index, regs_from, regs_from_count, from_index); 5943 VMRegPair from_pair; 5944 BasicType bt; 5945 while (stream.next(from_pair, bt)) {} 5946 sig_index = stream.sig_cc_index(); 5947 from_index = stream.regs_cc_index(); 5948 } 5949 5950 static bool is_reg_in_unpacked_fields(const GrowableArray<SigEntry>* sig, int sig_index, VMReg to, VMRegPair* regs_from, int regs_from_count, int from_index) { 5951 ScalarizedValueArgsStream stream(sig, sig_index, regs_from, regs_from_count, from_index); 5952 VMRegPair from_pair; 5953 BasicType bt; 5954 while (stream.next(from_pair, bt)) { 5955 if (from_pair.first() == to) { 5956 return true; 5957 } 5958 } 5959 5960 return false; 5961 } 5962 5963 // Pack fields back into a value type oop 5964 bool MacroAssembler::pack_value_helper(const GrowableArray<SigEntry>* sig, int& sig_index, int vtarg_index, 5965 VMReg to, VMRegPair* regs_from, int regs_from_count, int& from_index, RegState reg_state[], 5966 int ret_off, int extra_stack_offset) { 5967 assert(sig->at(sig_index)._bt == T_VALUETYPE, "should be at end delimiter"); 5968 assert(to->is_valid(), "must be"); 5969 5970 if (reg_state[to->value()] == reg_written) { 5971 skip_unpacked_fields(sig, sig_index, regs_from, regs_from_count, from_index); 5972 return true; // Already written 5973 } 5974 5975 Register val_array = rax; 5976 Register val_obj_tmp = r11; 5977 Register from_reg_tmp = r10; 5978 Register tmp1 = r14; 5979 Register tmp2 = r13; 5980 Register tmp3 = rbx; 5981 Register val_obj = to->is_stack() ? val_obj_tmp : to->as_Register(); 5982 5983 if (reg_state[to->value()] == reg_readonly) { 5984 if (!is_reg_in_unpacked_fields(sig, sig_index, to, regs_from, regs_from_count, from_index)) { 5985 skip_unpacked_fields(sig, sig_index, regs_from, regs_from_count, from_index); 5986 return false; // Not yet writable 5987 } 5988 val_obj = val_obj_tmp; 5989 } 5990 5991 int index = arrayOopDesc::base_offset_in_bytes(T_OBJECT) + vtarg_index * type2aelembytes(T_VALUETYPE); 5992 load_heap_oop(val_obj, Address(val_array, index)); 5993 5994 ScalarizedValueArgsStream stream(sig, sig_index, regs_from, regs_from_count, from_index); 5995 VMRegPair from_pair; 5996 BasicType bt; 5997 while (stream.next(from_pair, bt)) { 5998 int off = sig->at(stream.sig_cc_index())._offset; 5999 assert(off > 0, "offset in object should be positive"); 6000 bool is_oop = (bt == T_OBJECT || bt == T_ARRAY); 6001 size_t size_in_bytes = is_java_primitive(bt) ? type2aelembytes(bt) : wordSize; 6002 6003 VMReg from_r1 = from_pair.first(); 6004 VMReg from_r2 = from_pair.second(); 6005 6006 // Pack the scalarized field into the value object. 6007 Address dst(val_obj, off); 6008 if (!from_r1->is_XMMRegister()) { 6009 Register from_reg; 6010 6011 if (from_r1->is_stack()) { 6012 from_reg = from_reg_tmp; 6013 int ld_off = from_r1->reg2stack() * VMRegImpl::stack_slot_size + extra_stack_offset; 6014 load_sized_value(from_reg, Address(rsp, ld_off), size_in_bytes, /* is_signed */ false); 6015 } else { 6016 from_reg = from_r1->as_Register(); 6017 } 6018 6019 if (is_oop) { 6020 DecoratorSet decorators = IN_HEAP | ACCESS_WRITE; 6021 store_heap_oop(dst, from_reg, tmp1, tmp2, tmp3, decorators); 6022 } else { 6023 store_sized_value(dst, from_reg, size_in_bytes); 6024 } 6025 } else { 6026 if (from_r2->is_valid()) { 6027 movdbl(dst, from_r1->as_XMMRegister()); 6028 } else { 6029 movflt(dst, from_r1->as_XMMRegister()); 6030 } 6031 } 6032 reg_state[from_r1->value()] = reg_writable; 6033 } 6034 sig_index = stream.sig_cc_index(); 6035 from_index = stream.regs_cc_index(); 6036 6037 assert(reg_state[to->value()] == reg_writable, "must have already been read"); 6038 bool success = move_helper(val_obj->as_VMReg(), to, T_OBJECT, reg_state, ret_off, extra_stack_offset); 6039 assert(success, "to register must be writeable"); 6040 6041 return true; 6042 } 6043 6044 // Unpack all value type arguments passed as oops 6045 void MacroAssembler::unpack_value_args(Compile* C, bool receiver_only) { 6046 assert(C->has_scalarized_args(), "value type argument scalarization is disabled"); 6047 Method* method = C->method()->get_Method(); 6048 const GrowableArray<SigEntry>* sig_cc = method->adapter()->get_sig_cc(); 6049 assert(sig_cc != NULL, "must have scalarized signature"); 6050 6051 // Get unscalarized calling convention 6052 BasicType* sig_bt = NEW_RESOURCE_ARRAY(BasicType, sig_cc->length()); // FIXME - may underflow if we support values with no fields! 6053 int args_passed = 0; 6054 if (!method->is_static()) { 6055 sig_bt[args_passed++] = T_OBJECT; 6056 } 6057 if (!receiver_only) { 6058 for (SignatureStream ss(method->signature()); !ss.at_return_type(); ss.next()) { 6059 BasicType bt = ss.type(); 6060 sig_bt[args_passed++] = bt; 6061 if (type2size[bt] == 2) { 6062 sig_bt[args_passed++] = T_VOID; 6063 } 6064 } 6065 } else { 6066 // Only unpack the receiver, all other arguments are already scalarized 6067 InstanceKlass* holder = method->method_holder(); 6068 int rec_len = holder->is_value() ? ValueKlass::cast(holder)->extended_sig()->length() : 1; 6069 // Copy scalarized signature but skip receiver, value type delimiters and reserved entries 6070 for (int i = 0; i < sig_cc->length(); i++) { 6071 if (!SigEntry::is_reserved_entry(sig_cc, i)) { 6072 if (SigEntry::skip_value_delimiters(sig_cc, i) && rec_len <= 0) { 6073 sig_bt[args_passed++] = sig_cc->at(i)._bt; 6074 } 6075 rec_len--; 6076 } 6077 } 6078 } 6079 VMRegPair* regs = NEW_RESOURCE_ARRAY(VMRegPair, args_passed); 6080 int args_on_stack = SharedRuntime::java_calling_convention(sig_bt, regs, args_passed, false); 6081 6082 // Get scalarized calling convention 6083 int args_passed_cc = SigEntry::fill_sig_bt(sig_cc, sig_bt); 6084 VMRegPair* regs_cc = NEW_RESOURCE_ARRAY(VMRegPair, sig_cc->length()); 6085 int args_on_stack_cc = SharedRuntime::java_calling_convention(sig_bt, regs_cc, args_passed_cc, false); 6086 6087 int extra_stack_offset = wordSize; // stack has the returned address 6088 int sp_inc = shuffle_value_args(false, receiver_only, extra_stack_offset, sig_bt, sig_cc, 6089 args_passed, args_on_stack, regs, 6090 args_passed_cc, args_on_stack_cc, regs_cc); 6091 // Emit code for verified entry and save increment for stack repair on return 6092 verified_entry(C, sp_inc); 6093 } 6094 6095 int MacroAssembler::shuffle_value_args(bool is_packing, bool receiver_only, int extra_stack_offset, 6096 BasicType* sig_bt, const GrowableArray<SigEntry>* sig_cc, 6097 int args_passed, int args_on_stack, VMRegPair* regs, // from 6098 int args_passed_to, int args_on_stack_to, VMRegPair* regs_to) { // to 6099 // Check if we need to extend the stack for unpacking 6100 int sp_inc = (args_on_stack_to - args_on_stack) * VMRegImpl::stack_slot_size; 6101 if (sp_inc > 0) { 6102 // Save the return address, adjust the stack (make sure it is properly 6103 // 16-byte aligned) and copy the return address to the new top of the stack. 6104 pop(r13); 6105 sp_inc = align_up(sp_inc, StackAlignmentInBytes); 6106 subptr(rsp, sp_inc); 6107 push(r13); 6108 } else { 6109 // The scalarized calling convention needs less stack space than the unscalarized one. 6110 // No need to extend the stack, the caller will take care of these adjustments. 6111 sp_inc = 0; 6112 } 6113 6114 int ret_off; // make sure we don't overwrite the return address 6115 if (is_packing) { 6116 // For C1 code, the VVEP doesn't have reserved slots, so we store the returned address at 6117 // rsp[0] during shuffling. 6118 ret_off = 0; 6119 } else { 6120 // C2 code ensures that sp_inc is a reserved slot. 6121 ret_off = sp_inc; 6122 } 6123 6124 // Initialize register/stack slot states (make all writable) 6125 int max_stack = MAX2(args_on_stack + sp_inc/VMRegImpl::stack_slot_size, args_on_stack_to); 6126 int max_reg = VMRegImpl::stack2reg(max_stack)->value(); 6127 RegState* reg_state = NEW_RESOURCE_ARRAY(RegState, max_reg); 6128 for (int i = 0; i < max_reg; ++i) { 6129 reg_state[i] = reg_writable; 6130 } 6131 // Set all source registers/stack slots to readonly to prevent accidental overwriting 6132 for (int i = 0; i < args_passed; ++i) { 6133 VMReg reg = regs[i].first(); 6134 if (!reg->is_valid()) continue; 6135 if (reg->is_stack()) { 6136 // Update source stack location by adding stack increment 6137 reg = VMRegImpl::stack2reg(reg->reg2stack() + sp_inc/VMRegImpl::stack_slot_size); 6138 regs[i] = reg; 6139 } 6140 assert(reg->value() >= 0 && reg->value() < max_reg, "reg value out of bounds"); 6141 reg_state[reg->value()] = reg_readonly; 6142 } 6143 6144 // Emit code for packing/unpacking value type arguments 6145 // We try multiple times and eventually start spilling to resolve (circular) dependencies 6146 bool done = false; 6147 for (int i = 0; i < 2*args_passed_to && !done; ++i) { 6148 done = true; 6149 bool spill = (i > args_passed_to); // Start spilling? 6150 // Iterate over all arguments (when unpacking, do in reverse) 6151 int step = is_packing ? 1 : -1; 6152 int from_index = is_packing ? 0 : args_passed - 1; 6153 int to_index = is_packing ? 0 : args_passed_to - 1; 6154 int sig_index = is_packing ? 0 : sig_cc->length() - 1; 6155 int sig_index_end = is_packing ? sig_cc->length() : -1; 6156 int vtarg_index = 0; 6157 for (; sig_index != sig_index_end; sig_index += step) { 6158 assert(0 <= sig_index && sig_index < sig_cc->length(), "index out of bounds"); 6159 if (SigEntry::is_reserved_entry(sig_cc, sig_index)) { 6160 if (is_packing) { 6161 if (i == 0) { 6162 // The reserved entries are not used by the packed args, so make them writable 6163 VMReg from_reg = regs[from_index].first(); 6164 if (from_reg->is_valid()) { 6165 assert(from_reg->is_stack(), "reserved entries must be stack"); 6166 reg_state[from_reg->value()] = reg_writable; 6167 } 6168 } 6169 from_index += step; 6170 } else { 6171 to_index += step; 6172 } 6173 } else { 6174 assert(0 <= from_index && from_index < args_passed, "index out of bounds"); 6175 assert(0 <= to_index && to_index < args_passed_to, "index out of bounds"); 6176 if (spill) { 6177 // This call returns true IFF we should keep trying to spill in this round. 6178 spill = shuffle_value_args_spill(is_packing, sig_cc, sig_index, regs, from_index, args_passed, 6179 reg_state, ret_off, extra_stack_offset); 6180 } 6181 BasicType bt = sig_cc->at(sig_index)._bt; 6182 if (SigEntry::skip_value_delimiters(sig_cc, sig_index)) { 6183 VMReg from_reg = regs[from_index].first(); 6184 done &= move_helper(from_reg, regs_to[to_index].first(), bt, reg_state, ret_off, extra_stack_offset); 6185 to_index += step; 6186 } else if (is_packing || !receiver_only || (from_index == 0 && bt == T_VOID)) { 6187 if (is_packing) { 6188 VMReg reg_to = regs_to[to_index].first(); 6189 done &= pack_value_helper(sig_cc, sig_index, vtarg_index, reg_to, regs, args_passed, from_index, 6190 reg_state, ret_off, extra_stack_offset); 6191 vtarg_index ++; 6192 to_index ++; 6193 continue; // from_index already adjusted 6194 } else { 6195 VMReg from_reg = regs[from_index].first(); 6196 done &= unpack_value_helper(sig_cc, sig_index, from_reg, regs_to, to_index, reg_state, ret_off, extra_stack_offset); 6197 } 6198 } else { 6199 continue; 6200 } 6201 from_index += step; 6202 } 6203 } 6204 } 6205 guarantee(done, "Could not resolve circular dependency when shuffling value type arguments"); 6206 return sp_inc; 6207 } 6208 6209 bool MacroAssembler::shuffle_value_args_spill(bool is_packing, const GrowableArray<SigEntry>* sig_cc, int sig_cc_index, 6210 VMRegPair* regs_from, int from_index, int regs_from_count, 6211 RegState* reg_state, int ret_off, int extra_stack_offset) { 6212 VMReg reg; 6213 6214 if (!is_packing || SigEntry::skip_value_delimiters(sig_cc, sig_cc_index)) { 6215 reg = regs_from[from_index].first(); 6216 if (!reg->is_valid() || reg_state[reg->value()] != reg_readonly) { 6217 // Spilling this won't break circles 6218 return true; 6219 } 6220 } else { 6221 ScalarizedValueArgsStream stream(sig_cc, sig_cc_index, regs_from, regs_from_count, from_index); 6222 VMRegPair from_pair; 6223 BasicType bt; 6224 bool found = false; 6225 while (stream.next(from_pair, bt)) { 6226 reg = from_pair.first(); 6227 assert(reg->is_valid(), "must be"); 6228 if (reg_state[reg->value()] == reg_readonly) { 6229 found = true; 6230 break; 6231 } 6232 } 6233 if (!found) { 6234 // Spilling fields in this value arg won't break circles 6235 return true; 6236 } 6237 } 6238 6239 // Spill argument to be able to write the source and resolve circular dependencies 6240 VMReg spill_reg = reg->is_XMMRegister() ? xmm8->as_VMReg() : r14->as_VMReg(); 6241 if (reg_state[spill_reg->value()] == reg_readonly) { 6242 // We have already spilled (in previous round). The spilled register should be consumed by this round. 6243 } else { 6244 bool res = move_helper(reg, spill_reg, T_DOUBLE, reg_state, ret_off, extra_stack_offset); 6245 assert(res, "Spilling should not fail"); 6246 // Set spill_reg as new source and update state 6247 reg = spill_reg; 6248 regs_from[from_index].set1(reg); 6249 reg_state[reg->value()] = reg_readonly; 6250 } 6251 6252 return false; // Do not spill again in this round 6253 } 6254 6255 // Restores the stack on return 6256 void MacroAssembler::restore_stack(Compile* C) { 6257 int framesize = C->frame_size_in_bytes(); 6258 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned"); 6259 // Remove word for return addr already pushed and RBP 6260 framesize -= 2*wordSize; 6261 6262 if (C->needs_stack_repair()) { 6263 // Restore rbp and repair rsp by adding the stack increment 6264 movq(rbp, Address(rsp, framesize)); 6265 addq(rsp, Address(rsp, C->sp_inc_offset())); 6266 } else { 6267 if (framesize > 0) { 6268 addq(rsp, framesize); 6269 } 6270 pop(rbp); 6271 } 6272 } 6273 6274 void MacroAssembler::clear_mem(Register base, Register cnt, Register val, XMMRegister xtmp, bool is_large, bool word_copy_only) { 6275 // cnt - number of qwords (8-byte words). 6276 // base - start address, qword aligned. 6277 // is_large - if optimizers know cnt is larger than InitArrayShortSize 6278 assert(base==rdi, "base register must be edi for rep stos"); 6279 assert(val==rax, "tmp register must be eax for rep stos"); 6280 assert(cnt==rcx, "cnt register must be ecx for rep stos"); 6281 assert(InitArrayShortSize % BytesPerLong == 0, 6282 "InitArrayShortSize should be the multiple of BytesPerLong"); 6283 6284 Label DONE; 6285 6286 if (!is_large) { 6287 Label LOOP, LONG; 6288 cmpptr(cnt, InitArrayShortSize/BytesPerLong); 6289 jccb(Assembler::greater, LONG); 6290 6291 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM 6292 6293 decrement(cnt); 6294 jccb(Assembler::negative, DONE); // Zero length 6295 6296 // Use individual pointer-sized stores for small counts: 6297 BIND(LOOP); 6298 movptr(Address(base, cnt, Address::times_ptr), val); 6299 decrement(cnt); 6300 jccb(Assembler::greaterEqual, LOOP); 6301 jmpb(DONE); 6302 6303 BIND(LONG); 6304 } 6305 6306 // Use longer rep-prefixed ops for non-small counts: 6307 if (UseFastStosb && !word_copy_only) { 6308 shlptr(cnt, 3); // convert to number of bytes 6309 rep_stosb(); 6310 } else if (UseXMMForObjInit) { 6311 xmm_clear_mem(base, cnt, val, xtmp); 6312 } else { 6313 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM 6314 rep_stos(); 6315 } 6316 6317 BIND(DONE); 6318 } 6319 6320 #ifdef COMPILER2 6321 6322 // IndexOf for constant substrings with size >= 8 chars 6323 // which don't need to be loaded through stack. 6324 void MacroAssembler::string_indexofC8(Register str1, Register str2, 6325 Register cnt1, Register cnt2, 6326 int int_cnt2, Register result, 6327 XMMRegister vec, Register tmp, 6328 int ae) { 6329 ShortBranchVerifier sbv(this); 6330 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 6331 assert(ae != StrIntrinsicNode::LU, "Invalid encoding"); 6332 6333 // This method uses the pcmpestri instruction with bound registers 6334 // inputs: 6335 // xmm - substring 6336 // rax - substring length (elements count) 6337 // mem - scanned string 6338 // rdx - string length (elements count) 6339 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts) 6340 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes) 6341 // outputs: 6342 // rcx - matched index in string 6343 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 6344 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts 6345 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8 6346 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2; 6347 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1; 6348 6349 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, 6350 RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR, 6351 MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE; 6352 6353 // Note, inline_string_indexOf() generates checks: 6354 // if (substr.count > string.count) return -1; 6355 // if (substr.count == 0) return 0; 6356 assert(int_cnt2 >= stride, "this code is used only for cnt2 >= 8 chars"); 6357 6358 // Load substring. 6359 if (ae == StrIntrinsicNode::UL) { 6360 pmovzxbw(vec, Address(str2, 0)); 6361 } else { 6362 movdqu(vec, Address(str2, 0)); 6363 } 6364 movl(cnt2, int_cnt2); 6365 movptr(result, str1); // string addr 6366 6367 if (int_cnt2 > stride) { 6368 jmpb(SCAN_TO_SUBSTR); 6369 6370 // Reload substr for rescan, this code 6371 // is executed only for large substrings (> 8 chars) 6372 bind(RELOAD_SUBSTR); 6373 if (ae == StrIntrinsicNode::UL) { 6374 pmovzxbw(vec, Address(str2, 0)); 6375 } else { 6376 movdqu(vec, Address(str2, 0)); 6377 } 6378 negptr(cnt2); // Jumped here with negative cnt2, convert to positive 6379 6380 bind(RELOAD_STR); 6381 // We came here after the beginning of the substring was 6382 // matched but the rest of it was not so we need to search 6383 // again. Start from the next element after the previous match. 6384 6385 // cnt2 is number of substring reminding elements and 6386 // cnt1 is number of string reminding elements when cmp failed. 6387 // Restored cnt1 = cnt1 - cnt2 + int_cnt2 6388 subl(cnt1, cnt2); 6389 addl(cnt1, int_cnt2); 6390 movl(cnt2, int_cnt2); // Now restore cnt2 6391 6392 decrementl(cnt1); // Shift to next element 6393 cmpl(cnt1, cnt2); 6394 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring 6395 6396 addptr(result, (1<<scale1)); 6397 6398 } // (int_cnt2 > 8) 6399 6400 // Scan string for start of substr in 16-byte vectors 6401 bind(SCAN_TO_SUBSTR); 6402 pcmpestri(vec, Address(result, 0), mode); 6403 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1 6404 subl(cnt1, stride); 6405 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string 6406 cmpl(cnt1, cnt2); 6407 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 6408 addptr(result, 16); 6409 jmpb(SCAN_TO_SUBSTR); 6410 6411 // Found a potential substr 6412 bind(FOUND_CANDIDATE); 6413 // Matched whole vector if first element matched (tmp(rcx) == 0). 6414 if (int_cnt2 == stride) { 6415 jccb(Assembler::overflow, RET_FOUND); // OF == 1 6416 } else { // int_cnt2 > 8 6417 jccb(Assembler::overflow, FOUND_SUBSTR); 6418 } 6419 // After pcmpestri tmp(rcx) contains matched element index 6420 // Compute start addr of substr 6421 lea(result, Address(result, tmp, scale1)); 6422 6423 // Make sure string is still long enough 6424 subl(cnt1, tmp); 6425 cmpl(cnt1, cnt2); 6426 if (int_cnt2 == stride) { 6427 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR); 6428 } else { // int_cnt2 > 8 6429 jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD); 6430 } 6431 // Left less then substring. 6432 6433 bind(RET_NOT_FOUND); 6434 movl(result, -1); 6435 jmp(EXIT); 6436 6437 if (int_cnt2 > stride) { 6438 // This code is optimized for the case when whole substring 6439 // is matched if its head is matched. 6440 bind(MATCH_SUBSTR_HEAD); 6441 pcmpestri(vec, Address(result, 0), mode); 6442 // Reload only string if does not match 6443 jcc(Assembler::noOverflow, RELOAD_STR); // OF == 0 6444 6445 Label CONT_SCAN_SUBSTR; 6446 // Compare the rest of substring (> 8 chars). 6447 bind(FOUND_SUBSTR); 6448 // First 8 chars are already matched. 6449 negptr(cnt2); 6450 addptr(cnt2, stride); 6451 6452 bind(SCAN_SUBSTR); 6453 subl(cnt1, stride); 6454 cmpl(cnt2, -stride); // Do not read beyond substring 6455 jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR); 6456 // Back-up strings to avoid reading beyond substring: 6457 // cnt1 = cnt1 - cnt2 + 8 6458 addl(cnt1, cnt2); // cnt2 is negative 6459 addl(cnt1, stride); 6460 movl(cnt2, stride); negptr(cnt2); 6461 bind(CONT_SCAN_SUBSTR); 6462 if (int_cnt2 < (int)G) { 6463 int tail_off1 = int_cnt2<<scale1; 6464 int tail_off2 = int_cnt2<<scale2; 6465 if (ae == StrIntrinsicNode::UL) { 6466 pmovzxbw(vec, Address(str2, cnt2, scale2, tail_off2)); 6467 } else { 6468 movdqu(vec, Address(str2, cnt2, scale2, tail_off2)); 6469 } 6470 pcmpestri(vec, Address(result, cnt2, scale1, tail_off1), mode); 6471 } else { 6472 // calculate index in register to avoid integer overflow (int_cnt2*2) 6473 movl(tmp, int_cnt2); 6474 addptr(tmp, cnt2); 6475 if (ae == StrIntrinsicNode::UL) { 6476 pmovzxbw(vec, Address(str2, tmp, scale2, 0)); 6477 } else { 6478 movdqu(vec, Address(str2, tmp, scale2, 0)); 6479 } 6480 pcmpestri(vec, Address(result, tmp, scale1, 0), mode); 6481 } 6482 // Need to reload strings pointers if not matched whole vector 6483 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0 6484 addptr(cnt2, stride); 6485 jcc(Assembler::negative, SCAN_SUBSTR); 6486 // Fall through if found full substring 6487 6488 } // (int_cnt2 > 8) 6489 6490 bind(RET_FOUND); 6491 // Found result if we matched full small substring. 6492 // Compute substr offset 6493 subptr(result, str1); 6494 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 6495 shrl(result, 1); // index 6496 } 6497 bind(EXIT); 6498 6499 } // string_indexofC8 6500 6501 // Small strings are loaded through stack if they cross page boundary. 6502 void MacroAssembler::string_indexof(Register str1, Register str2, 6503 Register cnt1, Register cnt2, 6504 int int_cnt2, Register result, 6505 XMMRegister vec, Register tmp, 6506 int ae) { 6507 ShortBranchVerifier sbv(this); 6508 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 6509 assert(ae != StrIntrinsicNode::LU, "Invalid encoding"); 6510 6511 // 6512 // int_cnt2 is length of small (< 8 chars) constant substring 6513 // or (-1) for non constant substring in which case its length 6514 // is in cnt2 register. 6515 // 6516 // Note, inline_string_indexOf() generates checks: 6517 // if (substr.count > string.count) return -1; 6518 // if (substr.count == 0) return 0; 6519 // 6520 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8 6521 assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < stride), "should be != 0"); 6522 // This method uses the pcmpestri instruction with bound registers 6523 // inputs: 6524 // xmm - substring 6525 // rax - substring length (elements count) 6526 // mem - scanned string 6527 // rdx - string length (elements count) 6528 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts) 6529 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes) 6530 // outputs: 6531 // rcx - matched index in string 6532 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 6533 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts 6534 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2; 6535 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1; 6536 6537 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR, 6538 RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR, 6539 FOUND_CANDIDATE; 6540 6541 { //======================================================== 6542 // We don't know where these strings are located 6543 // and we can't read beyond them. Load them through stack. 6544 Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR; 6545 6546 movptr(tmp, rsp); // save old SP 6547 6548 if (int_cnt2 > 0) { // small (< 8 chars) constant substring 6549 if (int_cnt2 == (1>>scale2)) { // One byte 6550 assert((ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL), "Only possible for latin1 encoding"); 6551 load_unsigned_byte(result, Address(str2, 0)); 6552 movdl(vec, result); // move 32 bits 6553 } else if (ae == StrIntrinsicNode::LL && int_cnt2 == 3) { // Three bytes 6554 // Not enough header space in 32-bit VM: 12+3 = 15. 6555 movl(result, Address(str2, -1)); 6556 shrl(result, 8); 6557 movdl(vec, result); // move 32 bits 6558 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (2>>scale2)) { // One char 6559 load_unsigned_short(result, Address(str2, 0)); 6560 movdl(vec, result); // move 32 bits 6561 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (4>>scale2)) { // Two chars 6562 movdl(vec, Address(str2, 0)); // move 32 bits 6563 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (8>>scale2)) { // Four chars 6564 movq(vec, Address(str2, 0)); // move 64 bits 6565 } else { // cnt2 = { 3, 5, 6, 7 } || (ae == StrIntrinsicNode::UL && cnt2 ={2, ..., 7}) 6566 // Array header size is 12 bytes in 32-bit VM 6567 // + 6 bytes for 3 chars == 18 bytes, 6568 // enough space to load vec and shift. 6569 assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity"); 6570 if (ae == StrIntrinsicNode::UL) { 6571 int tail_off = int_cnt2-8; 6572 pmovzxbw(vec, Address(str2, tail_off)); 6573 psrldq(vec, -2*tail_off); 6574 } 6575 else { 6576 int tail_off = int_cnt2*(1<<scale2); 6577 movdqu(vec, Address(str2, tail_off-16)); 6578 psrldq(vec, 16-tail_off); 6579 } 6580 } 6581 } else { // not constant substring 6582 cmpl(cnt2, stride); 6583 jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough 6584 6585 // We can read beyond string if srt+16 does not cross page boundary 6586 // since heaps are aligned and mapped by pages. 6587 assert(os::vm_page_size() < (int)G, "default page should be small"); 6588 movl(result, str2); // We need only low 32 bits 6589 andl(result, (os::vm_page_size()-1)); 6590 cmpl(result, (os::vm_page_size()-16)); 6591 jccb(Assembler::belowEqual, CHECK_STR); 6592 6593 // Move small strings to stack to allow load 16 bytes into vec. 6594 subptr(rsp, 16); 6595 int stk_offset = wordSize-(1<<scale2); 6596 push(cnt2); 6597 6598 bind(COPY_SUBSTR); 6599 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL) { 6600 load_unsigned_byte(result, Address(str2, cnt2, scale2, -1)); 6601 movb(Address(rsp, cnt2, scale2, stk_offset), result); 6602 } else if (ae == StrIntrinsicNode::UU) { 6603 load_unsigned_short(result, Address(str2, cnt2, scale2, -2)); 6604 movw(Address(rsp, cnt2, scale2, stk_offset), result); 6605 } 6606 decrement(cnt2); 6607 jccb(Assembler::notZero, COPY_SUBSTR); 6608 6609 pop(cnt2); 6610 movptr(str2, rsp); // New substring address 6611 } // non constant 6612 6613 bind(CHECK_STR); 6614 cmpl(cnt1, stride); 6615 jccb(Assembler::aboveEqual, BIG_STRINGS); 6616 6617 // Check cross page boundary. 6618 movl(result, str1); // We need only low 32 bits 6619 andl(result, (os::vm_page_size()-1)); 6620 cmpl(result, (os::vm_page_size()-16)); 6621 jccb(Assembler::belowEqual, BIG_STRINGS); 6622 6623 subptr(rsp, 16); 6624 int stk_offset = -(1<<scale1); 6625 if (int_cnt2 < 0) { // not constant 6626 push(cnt2); 6627 stk_offset += wordSize; 6628 } 6629 movl(cnt2, cnt1); 6630 6631 bind(COPY_STR); 6632 if (ae == StrIntrinsicNode::LL) { 6633 load_unsigned_byte(result, Address(str1, cnt2, scale1, -1)); 6634 movb(Address(rsp, cnt2, scale1, stk_offset), result); 6635 } else { 6636 load_unsigned_short(result, Address(str1, cnt2, scale1, -2)); 6637 movw(Address(rsp, cnt2, scale1, stk_offset), result); 6638 } 6639 decrement(cnt2); 6640 jccb(Assembler::notZero, COPY_STR); 6641 6642 if (int_cnt2 < 0) { // not constant 6643 pop(cnt2); 6644 } 6645 movptr(str1, rsp); // New string address 6646 6647 bind(BIG_STRINGS); 6648 // Load substring. 6649 if (int_cnt2 < 0) { // -1 6650 if (ae == StrIntrinsicNode::UL) { 6651 pmovzxbw(vec, Address(str2, 0)); 6652 } else { 6653 movdqu(vec, Address(str2, 0)); 6654 } 6655 push(cnt2); // substr count 6656 push(str2); // substr addr 6657 push(str1); // string addr 6658 } else { 6659 // Small (< 8 chars) constant substrings are loaded already. 6660 movl(cnt2, int_cnt2); 6661 } 6662 push(tmp); // original SP 6663 6664 } // Finished loading 6665 6666 //======================================================== 6667 // Start search 6668 // 6669 6670 movptr(result, str1); // string addr 6671 6672 if (int_cnt2 < 0) { // Only for non constant substring 6673 jmpb(SCAN_TO_SUBSTR); 6674 6675 // SP saved at sp+0 6676 // String saved at sp+1*wordSize 6677 // Substr saved at sp+2*wordSize 6678 // Substr count saved at sp+3*wordSize 6679 6680 // Reload substr for rescan, this code 6681 // is executed only for large substrings (> 8 chars) 6682 bind(RELOAD_SUBSTR); 6683 movptr(str2, Address(rsp, 2*wordSize)); 6684 movl(cnt2, Address(rsp, 3*wordSize)); 6685 if (ae == StrIntrinsicNode::UL) { 6686 pmovzxbw(vec, Address(str2, 0)); 6687 } else { 6688 movdqu(vec, Address(str2, 0)); 6689 } 6690 // We came here after the beginning of the substring was 6691 // matched but the rest of it was not so we need to search 6692 // again. Start from the next element after the previous match. 6693 subptr(str1, result); // Restore counter 6694 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 6695 shrl(str1, 1); 6696 } 6697 addl(cnt1, str1); 6698 decrementl(cnt1); // Shift to next element 6699 cmpl(cnt1, cnt2); 6700 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring 6701 6702 addptr(result, (1<<scale1)); 6703 } // non constant 6704 6705 // Scan string for start of substr in 16-byte vectors 6706 bind(SCAN_TO_SUBSTR); 6707 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); 6708 pcmpestri(vec, Address(result, 0), mode); 6709 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1 6710 subl(cnt1, stride); 6711 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string 6712 cmpl(cnt1, cnt2); 6713 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring 6714 addptr(result, 16); 6715 6716 bind(ADJUST_STR); 6717 cmpl(cnt1, stride); // Do not read beyond string 6718 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR); 6719 // Back-up string to avoid reading beyond string. 6720 lea(result, Address(result, cnt1, scale1, -16)); 6721 movl(cnt1, stride); 6722 jmpb(SCAN_TO_SUBSTR); 6723 6724 // Found a potential substr 6725 bind(FOUND_CANDIDATE); 6726 // After pcmpestri tmp(rcx) contains matched element index 6727 6728 // Make sure string is still long enough 6729 subl(cnt1, tmp); 6730 cmpl(cnt1, cnt2); 6731 jccb(Assembler::greaterEqual, FOUND_SUBSTR); 6732 // Left less then substring. 6733 6734 bind(RET_NOT_FOUND); 6735 movl(result, -1); 6736 jmp(CLEANUP); 6737 6738 bind(FOUND_SUBSTR); 6739 // Compute start addr of substr 6740 lea(result, Address(result, tmp, scale1)); 6741 if (int_cnt2 > 0) { // Constant substring 6742 // Repeat search for small substring (< 8 chars) 6743 // from new point without reloading substring. 6744 // Have to check that we don't read beyond string. 6745 cmpl(tmp, stride-int_cnt2); 6746 jccb(Assembler::greater, ADJUST_STR); 6747 // Fall through if matched whole substring. 6748 } else { // non constant 6749 assert(int_cnt2 == -1, "should be != 0"); 6750 6751 addl(tmp, cnt2); 6752 // Found result if we matched whole substring. 6753 cmpl(tmp, stride); 6754 jcc(Assembler::lessEqual, RET_FOUND); 6755 6756 // Repeat search for small substring (<= 8 chars) 6757 // from new point 'str1' without reloading substring. 6758 cmpl(cnt2, stride); 6759 // Have to check that we don't read beyond string. 6760 jccb(Assembler::lessEqual, ADJUST_STR); 6761 6762 Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG; 6763 // Compare the rest of substring (> 8 chars). 6764 movptr(str1, result); 6765 6766 cmpl(tmp, cnt2); 6767 // First 8 chars are already matched. 6768 jccb(Assembler::equal, CHECK_NEXT); 6769 6770 bind(SCAN_SUBSTR); 6771 pcmpestri(vec, Address(str1, 0), mode); 6772 // Need to reload strings pointers if not matched whole vector 6773 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0 6774 6775 bind(CHECK_NEXT); 6776 subl(cnt2, stride); 6777 jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring 6778 addptr(str1, 16); 6779 if (ae == StrIntrinsicNode::UL) { 6780 addptr(str2, 8); 6781 } else { 6782 addptr(str2, 16); 6783 } 6784 subl(cnt1, stride); 6785 cmpl(cnt2, stride); // Do not read beyond substring 6786 jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR); 6787 // Back-up strings to avoid reading beyond substring. 6788 6789 if (ae == StrIntrinsicNode::UL) { 6790 lea(str2, Address(str2, cnt2, scale2, -8)); 6791 lea(str1, Address(str1, cnt2, scale1, -16)); 6792 } else { 6793 lea(str2, Address(str2, cnt2, scale2, -16)); 6794 lea(str1, Address(str1, cnt2, scale1, -16)); 6795 } 6796 subl(cnt1, cnt2); 6797 movl(cnt2, stride); 6798 addl(cnt1, stride); 6799 bind(CONT_SCAN_SUBSTR); 6800 if (ae == StrIntrinsicNode::UL) { 6801 pmovzxbw(vec, Address(str2, 0)); 6802 } else { 6803 movdqu(vec, Address(str2, 0)); 6804 } 6805 jmp(SCAN_SUBSTR); 6806 6807 bind(RET_FOUND_LONG); 6808 movptr(str1, Address(rsp, wordSize)); 6809 } // non constant 6810 6811 bind(RET_FOUND); 6812 // Compute substr offset 6813 subptr(result, str1); 6814 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 6815 shrl(result, 1); // index 6816 } 6817 bind(CLEANUP); 6818 pop(rsp); // restore SP 6819 6820 } // string_indexof 6821 6822 void MacroAssembler::string_indexof_char(Register str1, Register cnt1, Register ch, Register result, 6823 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp) { 6824 ShortBranchVerifier sbv(this); 6825 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required"); 6826 6827 int stride = 8; 6828 6829 Label FOUND_CHAR, SCAN_TO_CHAR, SCAN_TO_CHAR_LOOP, 6830 SCAN_TO_8_CHAR, SCAN_TO_8_CHAR_LOOP, SCAN_TO_16_CHAR_LOOP, 6831 RET_NOT_FOUND, SCAN_TO_8_CHAR_INIT, 6832 FOUND_SEQ_CHAR, DONE_LABEL; 6833 6834 movptr(result, str1); 6835 if (UseAVX >= 2) { 6836 cmpl(cnt1, stride); 6837 jcc(Assembler::less, SCAN_TO_CHAR_LOOP); 6838 cmpl(cnt1, 2*stride); 6839 jcc(Assembler::less, SCAN_TO_8_CHAR_INIT); 6840 movdl(vec1, ch); 6841 vpbroadcastw(vec1, vec1, Assembler::AVX_256bit); 6842 vpxor(vec2, vec2); 6843 movl(tmp, cnt1); 6844 andl(tmp, 0xFFFFFFF0); //vector count (in chars) 6845 andl(cnt1,0x0000000F); //tail count (in chars) 6846 6847 bind(SCAN_TO_16_CHAR_LOOP); 6848 vmovdqu(vec3, Address(result, 0)); 6849 vpcmpeqw(vec3, vec3, vec1, 1); 6850 vptest(vec2, vec3); 6851 jcc(Assembler::carryClear, FOUND_CHAR); 6852 addptr(result, 32); 6853 subl(tmp, 2*stride); 6854 jcc(Assembler::notZero, SCAN_TO_16_CHAR_LOOP); 6855 jmp(SCAN_TO_8_CHAR); 6856 bind(SCAN_TO_8_CHAR_INIT); 6857 movdl(vec1, ch); 6858 pshuflw(vec1, vec1, 0x00); 6859 pshufd(vec1, vec1, 0); 6860 pxor(vec2, vec2); 6861 } 6862 bind(SCAN_TO_8_CHAR); 6863 cmpl(cnt1, stride); 6864 if (UseAVX >= 2) { 6865 jcc(Assembler::less, SCAN_TO_CHAR); 6866 } else { 6867 jcc(Assembler::less, SCAN_TO_CHAR_LOOP); 6868 movdl(vec1, ch); 6869 pshuflw(vec1, vec1, 0x00); 6870 pshufd(vec1, vec1, 0); 6871 pxor(vec2, vec2); 6872 } 6873 movl(tmp, cnt1); 6874 andl(tmp, 0xFFFFFFF8); //vector count (in chars) 6875 andl(cnt1,0x00000007); //tail count (in chars) 6876 6877 bind(SCAN_TO_8_CHAR_LOOP); 6878 movdqu(vec3, Address(result, 0)); 6879 pcmpeqw(vec3, vec1); 6880 ptest(vec2, vec3); 6881 jcc(Assembler::carryClear, FOUND_CHAR); 6882 addptr(result, 16); 6883 subl(tmp, stride); 6884 jcc(Assembler::notZero, SCAN_TO_8_CHAR_LOOP); 6885 bind(SCAN_TO_CHAR); 6886 testl(cnt1, cnt1); 6887 jcc(Assembler::zero, RET_NOT_FOUND); 6888 bind(SCAN_TO_CHAR_LOOP); 6889 load_unsigned_short(tmp, Address(result, 0)); 6890 cmpl(ch, tmp); 6891 jccb(Assembler::equal, FOUND_SEQ_CHAR); 6892 addptr(result, 2); 6893 subl(cnt1, 1); 6894 jccb(Assembler::zero, RET_NOT_FOUND); 6895 jmp(SCAN_TO_CHAR_LOOP); 6896 6897 bind(RET_NOT_FOUND); 6898 movl(result, -1); 6899 jmpb(DONE_LABEL); 6900 6901 bind(FOUND_CHAR); 6902 if (UseAVX >= 2) { 6903 vpmovmskb(tmp, vec3); 6904 } else { 6905 pmovmskb(tmp, vec3); 6906 } 6907 bsfl(ch, tmp); 6908 addl(result, ch); 6909 6910 bind(FOUND_SEQ_CHAR); 6911 subptr(result, str1); 6912 shrl(result, 1); 6913 6914 bind(DONE_LABEL); 6915 } // string_indexof_char 6916 6917 // helper function for string_compare 6918 void MacroAssembler::load_next_elements(Register elem1, Register elem2, Register str1, Register str2, 6919 Address::ScaleFactor scale, Address::ScaleFactor scale1, 6920 Address::ScaleFactor scale2, Register index, int ae) { 6921 if (ae == StrIntrinsicNode::LL) { 6922 load_unsigned_byte(elem1, Address(str1, index, scale, 0)); 6923 load_unsigned_byte(elem2, Address(str2, index, scale, 0)); 6924 } else if (ae == StrIntrinsicNode::UU) { 6925 load_unsigned_short(elem1, Address(str1, index, scale, 0)); 6926 load_unsigned_short(elem2, Address(str2, index, scale, 0)); 6927 } else { 6928 load_unsigned_byte(elem1, Address(str1, index, scale1, 0)); 6929 load_unsigned_short(elem2, Address(str2, index, scale2, 0)); 6930 } 6931 } 6932 6933 // Compare strings, used for char[] and byte[]. 6934 void MacroAssembler::string_compare(Register str1, Register str2, 6935 Register cnt1, Register cnt2, Register result, 6936 XMMRegister vec1, int ae) { 6937 ShortBranchVerifier sbv(this); 6938 Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL; 6939 Label COMPARE_WIDE_VECTORS_LOOP_FAILED; // used only _LP64 && AVX3 6940 int stride, stride2, adr_stride, adr_stride1, adr_stride2; 6941 int stride2x2 = 0x40; 6942 Address::ScaleFactor scale = Address::no_scale; 6943 Address::ScaleFactor scale1 = Address::no_scale; 6944 Address::ScaleFactor scale2 = Address::no_scale; 6945 6946 if (ae != StrIntrinsicNode::LL) { 6947 stride2x2 = 0x20; 6948 } 6949 6950 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) { 6951 shrl(cnt2, 1); 6952 } 6953 // Compute the minimum of the string lengths and the 6954 // difference of the string lengths (stack). 6955 // Do the conditional move stuff 6956 movl(result, cnt1); 6957 subl(cnt1, cnt2); 6958 push(cnt1); 6959 cmov32(Assembler::lessEqual, cnt2, result); // cnt2 = min(cnt1, cnt2) 6960 6961 // Is the minimum length zero? 6962 testl(cnt2, cnt2); 6963 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 6964 if (ae == StrIntrinsicNode::LL) { 6965 // Load first bytes 6966 load_unsigned_byte(result, Address(str1, 0)); // result = str1[0] 6967 load_unsigned_byte(cnt1, Address(str2, 0)); // cnt1 = str2[0] 6968 } else if (ae == StrIntrinsicNode::UU) { 6969 // Load first characters 6970 load_unsigned_short(result, Address(str1, 0)); 6971 load_unsigned_short(cnt1, Address(str2, 0)); 6972 } else { 6973 load_unsigned_byte(result, Address(str1, 0)); 6974 load_unsigned_short(cnt1, Address(str2, 0)); 6975 } 6976 subl(result, cnt1); 6977 jcc(Assembler::notZero, POP_LABEL); 6978 6979 if (ae == StrIntrinsicNode::UU) { 6980 // Divide length by 2 to get number of chars 6981 shrl(cnt2, 1); 6982 } 6983 cmpl(cnt2, 1); 6984 jcc(Assembler::equal, LENGTH_DIFF_LABEL); 6985 6986 // Check if the strings start at the same location and setup scale and stride 6987 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 6988 cmpptr(str1, str2); 6989 jcc(Assembler::equal, LENGTH_DIFF_LABEL); 6990 if (ae == StrIntrinsicNode::LL) { 6991 scale = Address::times_1; 6992 stride = 16; 6993 } else { 6994 scale = Address::times_2; 6995 stride = 8; 6996 } 6997 } else { 6998 scale1 = Address::times_1; 6999 scale2 = Address::times_2; 7000 // scale not used 7001 stride = 8; 7002 } 7003 7004 if (UseAVX >= 2 && UseSSE42Intrinsics) { 7005 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR; 7006 Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR; 7007 Label COMPARE_WIDE_VECTORS_LOOP_AVX2; 7008 Label COMPARE_TAIL_LONG; 7009 Label COMPARE_WIDE_VECTORS_LOOP_AVX3; // used only _LP64 && AVX3 7010 7011 int pcmpmask = 0x19; 7012 if (ae == StrIntrinsicNode::LL) { 7013 pcmpmask &= ~0x01; 7014 } 7015 7016 // Setup to compare 16-chars (32-bytes) vectors, 7017 // start from first character again because it has aligned address. 7018 if (ae == StrIntrinsicNode::LL) { 7019 stride2 = 32; 7020 } else { 7021 stride2 = 16; 7022 } 7023 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7024 adr_stride = stride << scale; 7025 } else { 7026 adr_stride1 = 8; //stride << scale1; 7027 adr_stride2 = 16; //stride << scale2; 7028 } 7029 7030 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri"); 7031 // rax and rdx are used by pcmpestri as elements counters 7032 movl(result, cnt2); 7033 andl(cnt2, ~(stride2-1)); // cnt2 holds the vector count 7034 jcc(Assembler::zero, COMPARE_TAIL_LONG); 7035 7036 // fast path : compare first 2 8-char vectors. 7037 bind(COMPARE_16_CHARS); 7038 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7039 movdqu(vec1, Address(str1, 0)); 7040 } else { 7041 pmovzxbw(vec1, Address(str1, 0)); 7042 } 7043 pcmpestri(vec1, Address(str2, 0), pcmpmask); 7044 jccb(Assembler::below, COMPARE_INDEX_CHAR); 7045 7046 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7047 movdqu(vec1, Address(str1, adr_stride)); 7048 pcmpestri(vec1, Address(str2, adr_stride), pcmpmask); 7049 } else { 7050 pmovzxbw(vec1, Address(str1, adr_stride1)); 7051 pcmpestri(vec1, Address(str2, adr_stride2), pcmpmask); 7052 } 7053 jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS); 7054 addl(cnt1, stride); 7055 7056 // Compare the characters at index in cnt1 7057 bind(COMPARE_INDEX_CHAR); // cnt1 has the offset of the mismatching character 7058 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae); 7059 subl(result, cnt2); 7060 jmp(POP_LABEL); 7061 7062 // Setup the registers to start vector comparison loop 7063 bind(COMPARE_WIDE_VECTORS); 7064 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7065 lea(str1, Address(str1, result, scale)); 7066 lea(str2, Address(str2, result, scale)); 7067 } else { 7068 lea(str1, Address(str1, result, scale1)); 7069 lea(str2, Address(str2, result, scale2)); 7070 } 7071 subl(result, stride2); 7072 subl(cnt2, stride2); 7073 jcc(Assembler::zero, COMPARE_WIDE_TAIL); 7074 negptr(result); 7075 7076 // In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest) 7077 bind(COMPARE_WIDE_VECTORS_LOOP); 7078 7079 #ifdef _LP64 7080 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop 7081 cmpl(cnt2, stride2x2); 7082 jccb(Assembler::below, COMPARE_WIDE_VECTORS_LOOP_AVX2); 7083 testl(cnt2, stride2x2-1); // cnt2 holds the vector count 7084 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX2); // means we cannot subtract by 0x40 7085 7086 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop 7087 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7088 evmovdquq(vec1, Address(str1, result, scale), Assembler::AVX_512bit); 7089 evpcmpeqb(k7, vec1, Address(str2, result, scale), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0 7090 } else { 7091 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_512bit); 7092 evpcmpeqb(k7, vec1, Address(str2, result, scale2), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0 7093 } 7094 kortestql(k7, k7); 7095 jcc(Assembler::aboveEqual, COMPARE_WIDE_VECTORS_LOOP_FAILED); // miscompare 7096 addptr(result, stride2x2); // update since we already compared at this addr 7097 subl(cnt2, stride2x2); // and sub the size too 7098 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX3); 7099 7100 vpxor(vec1, vec1); 7101 jmpb(COMPARE_WIDE_TAIL); 7102 }//if (VM_Version::supports_avx512vlbw()) 7103 #endif // _LP64 7104 7105 7106 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2); 7107 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7108 vmovdqu(vec1, Address(str1, result, scale)); 7109 vpxor(vec1, Address(str2, result, scale)); 7110 } else { 7111 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_256bit); 7112 vpxor(vec1, Address(str2, result, scale2)); 7113 } 7114 vptest(vec1, vec1); 7115 jcc(Assembler::notZero, VECTOR_NOT_EQUAL); 7116 addptr(result, stride2); 7117 subl(cnt2, stride2); 7118 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP); 7119 // clean upper bits of YMM registers 7120 vpxor(vec1, vec1); 7121 7122 // compare wide vectors tail 7123 bind(COMPARE_WIDE_TAIL); 7124 testptr(result, result); 7125 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 7126 7127 movl(result, stride2); 7128 movl(cnt2, result); 7129 negptr(result); 7130 jmp(COMPARE_WIDE_VECTORS_LOOP_AVX2); 7131 7132 // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors. 7133 bind(VECTOR_NOT_EQUAL); 7134 // clean upper bits of YMM registers 7135 vpxor(vec1, vec1); 7136 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7137 lea(str1, Address(str1, result, scale)); 7138 lea(str2, Address(str2, result, scale)); 7139 } else { 7140 lea(str1, Address(str1, result, scale1)); 7141 lea(str2, Address(str2, result, scale2)); 7142 } 7143 jmp(COMPARE_16_CHARS); 7144 7145 // Compare tail chars, length between 1 to 15 chars 7146 bind(COMPARE_TAIL_LONG); 7147 movl(cnt2, result); 7148 cmpl(cnt2, stride); 7149 jcc(Assembler::less, COMPARE_SMALL_STR); 7150 7151 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7152 movdqu(vec1, Address(str1, 0)); 7153 } else { 7154 pmovzxbw(vec1, Address(str1, 0)); 7155 } 7156 pcmpestri(vec1, Address(str2, 0), pcmpmask); 7157 jcc(Assembler::below, COMPARE_INDEX_CHAR); 7158 subptr(cnt2, stride); 7159 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 7160 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7161 lea(str1, Address(str1, result, scale)); 7162 lea(str2, Address(str2, result, scale)); 7163 } else { 7164 lea(str1, Address(str1, result, scale1)); 7165 lea(str2, Address(str2, result, scale2)); 7166 } 7167 negptr(cnt2); 7168 jmpb(WHILE_HEAD_LABEL); 7169 7170 bind(COMPARE_SMALL_STR); 7171 } else if (UseSSE42Intrinsics) { 7172 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL; 7173 int pcmpmask = 0x19; 7174 // Setup to compare 8-char (16-byte) vectors, 7175 // start from first character again because it has aligned address. 7176 movl(result, cnt2); 7177 andl(cnt2, ~(stride - 1)); // cnt2 holds the vector count 7178 if (ae == StrIntrinsicNode::LL) { 7179 pcmpmask &= ~0x01; 7180 } 7181 jcc(Assembler::zero, COMPARE_TAIL); 7182 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7183 lea(str1, Address(str1, result, scale)); 7184 lea(str2, Address(str2, result, scale)); 7185 } else { 7186 lea(str1, Address(str1, result, scale1)); 7187 lea(str2, Address(str2, result, scale2)); 7188 } 7189 negptr(result); 7190 7191 // pcmpestri 7192 // inputs: 7193 // vec1- substring 7194 // rax - negative string length (elements count) 7195 // mem - scanned string 7196 // rdx - string length (elements count) 7197 // pcmpmask - cmp mode: 11000 (string compare with negated result) 7198 // + 00 (unsigned bytes) or + 01 (unsigned shorts) 7199 // outputs: 7200 // rcx - first mismatched element index 7201 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri"); 7202 7203 bind(COMPARE_WIDE_VECTORS); 7204 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7205 movdqu(vec1, Address(str1, result, scale)); 7206 pcmpestri(vec1, Address(str2, result, scale), pcmpmask); 7207 } else { 7208 pmovzxbw(vec1, Address(str1, result, scale1)); 7209 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask); 7210 } 7211 // After pcmpestri cnt1(rcx) contains mismatched element index 7212 7213 jccb(Assembler::below, VECTOR_NOT_EQUAL); // CF==1 7214 addptr(result, stride); 7215 subptr(cnt2, stride); 7216 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS); 7217 7218 // compare wide vectors tail 7219 testptr(result, result); 7220 jcc(Assembler::zero, LENGTH_DIFF_LABEL); 7221 7222 movl(cnt2, stride); 7223 movl(result, stride); 7224 negptr(result); 7225 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7226 movdqu(vec1, Address(str1, result, scale)); 7227 pcmpestri(vec1, Address(str2, result, scale), pcmpmask); 7228 } else { 7229 pmovzxbw(vec1, Address(str1, result, scale1)); 7230 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask); 7231 } 7232 jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL); 7233 7234 // Mismatched characters in the vectors 7235 bind(VECTOR_NOT_EQUAL); 7236 addptr(cnt1, result); 7237 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae); 7238 subl(result, cnt2); 7239 jmpb(POP_LABEL); 7240 7241 bind(COMPARE_TAIL); // limit is zero 7242 movl(cnt2, result); 7243 // Fallthru to tail compare 7244 } 7245 // Shift str2 and str1 to the end of the arrays, negate min 7246 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) { 7247 lea(str1, Address(str1, cnt2, scale)); 7248 lea(str2, Address(str2, cnt2, scale)); 7249 } else { 7250 lea(str1, Address(str1, cnt2, scale1)); 7251 lea(str2, Address(str2, cnt2, scale2)); 7252 } 7253 decrementl(cnt2); // first character was compared already 7254 negptr(cnt2); 7255 7256 // Compare the rest of the elements 7257 bind(WHILE_HEAD_LABEL); 7258 load_next_elements(result, cnt1, str1, str2, scale, scale1, scale2, cnt2, ae); 7259 subl(result, cnt1); 7260 jccb(Assembler::notZero, POP_LABEL); 7261 increment(cnt2); 7262 jccb(Assembler::notZero, WHILE_HEAD_LABEL); 7263 7264 // Strings are equal up to min length. Return the length difference. 7265 bind(LENGTH_DIFF_LABEL); 7266 pop(result); 7267 if (ae == StrIntrinsicNode::UU) { 7268 // Divide diff by 2 to get number of chars 7269 sarl(result, 1); 7270 } 7271 jmpb(DONE_LABEL); 7272 7273 #ifdef _LP64 7274 if (VM_Version::supports_avx512vlbw()) { 7275 7276 bind(COMPARE_WIDE_VECTORS_LOOP_FAILED); 7277 7278 kmovql(cnt1, k7); 7279 notq(cnt1); 7280 bsfq(cnt2, cnt1); 7281 if (ae != StrIntrinsicNode::LL) { 7282 // Divide diff by 2 to get number of chars 7283 sarl(cnt2, 1); 7284 } 7285 addq(result, cnt2); 7286 if (ae == StrIntrinsicNode::LL) { 7287 load_unsigned_byte(cnt1, Address(str2, result)); 7288 load_unsigned_byte(result, Address(str1, result)); 7289 } else if (ae == StrIntrinsicNode::UU) { 7290 load_unsigned_short(cnt1, Address(str2, result, scale)); 7291 load_unsigned_short(result, Address(str1, result, scale)); 7292 } else { 7293 load_unsigned_short(cnt1, Address(str2, result, scale2)); 7294 load_unsigned_byte(result, Address(str1, result, scale1)); 7295 } 7296 subl(result, cnt1); 7297 jmpb(POP_LABEL); 7298 }//if (VM_Version::supports_avx512vlbw()) 7299 #endif // _LP64 7300 7301 // Discard the stored length difference 7302 bind(POP_LABEL); 7303 pop(cnt1); 7304 7305 // That's it 7306 bind(DONE_LABEL); 7307 if(ae == StrIntrinsicNode::UL) { 7308 negl(result); 7309 } 7310 7311 } 7312 7313 // Search for Non-ASCII character (Negative byte value) in a byte array, 7314 // return true if it has any and false otherwise. 7315 // ..\jdk\src\java.base\share\classes\java\lang\StringCoding.java 7316 // @HotSpotIntrinsicCandidate 7317 // private static boolean hasNegatives(byte[] ba, int off, int len) { 7318 // for (int i = off; i < off + len; i++) { 7319 // if (ba[i] < 0) { 7320 // return true; 7321 // } 7322 // } 7323 // return false; 7324 // } 7325 void MacroAssembler::has_negatives(Register ary1, Register len, 7326 Register result, Register tmp1, 7327 XMMRegister vec1, XMMRegister vec2) { 7328 // rsi: byte array 7329 // rcx: len 7330 // rax: result 7331 ShortBranchVerifier sbv(this); 7332 assert_different_registers(ary1, len, result, tmp1); 7333 assert_different_registers(vec1, vec2); 7334 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_CHAR, COMPARE_VECTORS, COMPARE_BYTE; 7335 7336 // len == 0 7337 testl(len, len); 7338 jcc(Assembler::zero, FALSE_LABEL); 7339 7340 if ((UseAVX > 2) && // AVX512 7341 VM_Version::supports_avx512vlbw() && 7342 VM_Version::supports_bmi2()) { 7343 7344 Label test_64_loop, test_tail; 7345 Register tmp3_aliased = len; 7346 7347 movl(tmp1, len); 7348 vpxor(vec2, vec2, vec2, Assembler::AVX_512bit); 7349 7350 andl(tmp1, 64 - 1); // tail count (in chars) 0x3F 7351 andl(len, ~(64 - 1)); // vector count (in chars) 7352 jccb(Assembler::zero, test_tail); 7353 7354 lea(ary1, Address(ary1, len, Address::times_1)); 7355 negptr(len); 7356 7357 bind(test_64_loop); 7358 // Check whether our 64 elements of size byte contain negatives 7359 evpcmpgtb(k2, vec2, Address(ary1, len, Address::times_1), Assembler::AVX_512bit); 7360 kortestql(k2, k2); 7361 jcc(Assembler::notZero, TRUE_LABEL); 7362 7363 addptr(len, 64); 7364 jccb(Assembler::notZero, test_64_loop); 7365 7366 7367 bind(test_tail); 7368 // bail out when there is nothing to be done 7369 testl(tmp1, -1); 7370 jcc(Assembler::zero, FALSE_LABEL); 7371 7372 // ~(~0 << len) applied up to two times (for 32-bit scenario) 7373 #ifdef _LP64 7374 mov64(tmp3_aliased, 0xFFFFFFFFFFFFFFFF); 7375 shlxq(tmp3_aliased, tmp3_aliased, tmp1); 7376 notq(tmp3_aliased); 7377 kmovql(k3, tmp3_aliased); 7378 #else 7379 Label k_init; 7380 jmp(k_init); 7381 7382 // We could not read 64-bits from a general purpose register thus we move 7383 // data required to compose 64 1's to the instruction stream 7384 // We emit 64 byte wide series of elements from 0..63 which later on would 7385 // be used as a compare targets with tail count contained in tmp1 register. 7386 // Result would be a k register having tmp1 consecutive number or 1 7387 // counting from least significant bit. 7388 address tmp = pc(); 7389 emit_int64(0x0706050403020100); 7390 emit_int64(0x0F0E0D0C0B0A0908); 7391 emit_int64(0x1716151413121110); 7392 emit_int64(0x1F1E1D1C1B1A1918); 7393 emit_int64(0x2726252423222120); 7394 emit_int64(0x2F2E2D2C2B2A2928); 7395 emit_int64(0x3736353433323130); 7396 emit_int64(0x3F3E3D3C3B3A3938); 7397 7398 bind(k_init); 7399 lea(len, InternalAddress(tmp)); 7400 // create mask to test for negative byte inside a vector 7401 evpbroadcastb(vec1, tmp1, Assembler::AVX_512bit); 7402 evpcmpgtb(k3, vec1, Address(len, 0), Assembler::AVX_512bit); 7403 7404 #endif 7405 evpcmpgtb(k2, k3, vec2, Address(ary1, 0), Assembler::AVX_512bit); 7406 ktestq(k2, k3); 7407 jcc(Assembler::notZero, TRUE_LABEL); 7408 7409 jmp(FALSE_LABEL); 7410 } else { 7411 movl(result, len); // copy 7412 7413 if (UseAVX == 2 && UseSSE >= 2) { 7414 // With AVX2, use 32-byte vector compare 7415 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 7416 7417 // Compare 32-byte vectors 7418 andl(result, 0x0000001f); // tail count (in bytes) 7419 andl(len, 0xffffffe0); // vector count (in bytes) 7420 jccb(Assembler::zero, COMPARE_TAIL); 7421 7422 lea(ary1, Address(ary1, len, Address::times_1)); 7423 negptr(len); 7424 7425 movl(tmp1, 0x80808080); // create mask to test for Unicode chars in vector 7426 movdl(vec2, tmp1); 7427 vpbroadcastd(vec2, vec2, Assembler::AVX_256bit); 7428 7429 bind(COMPARE_WIDE_VECTORS); 7430 vmovdqu(vec1, Address(ary1, len, Address::times_1)); 7431 vptest(vec1, vec2); 7432 jccb(Assembler::notZero, TRUE_LABEL); 7433 addptr(len, 32); 7434 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 7435 7436 testl(result, result); 7437 jccb(Assembler::zero, FALSE_LABEL); 7438 7439 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32)); 7440 vptest(vec1, vec2); 7441 jccb(Assembler::notZero, TRUE_LABEL); 7442 jmpb(FALSE_LABEL); 7443 7444 bind(COMPARE_TAIL); // len is zero 7445 movl(len, result); 7446 // Fallthru to tail compare 7447 } else if (UseSSE42Intrinsics) { 7448 // With SSE4.2, use double quad vector compare 7449 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 7450 7451 // Compare 16-byte vectors 7452 andl(result, 0x0000000f); // tail count (in bytes) 7453 andl(len, 0xfffffff0); // vector count (in bytes) 7454 jcc(Assembler::zero, COMPARE_TAIL); 7455 7456 lea(ary1, Address(ary1, len, Address::times_1)); 7457 negptr(len); 7458 7459 movl(tmp1, 0x80808080); 7460 movdl(vec2, tmp1); 7461 pshufd(vec2, vec2, 0); 7462 7463 bind(COMPARE_WIDE_VECTORS); 7464 movdqu(vec1, Address(ary1, len, Address::times_1)); 7465 ptest(vec1, vec2); 7466 jcc(Assembler::notZero, TRUE_LABEL); 7467 addptr(len, 16); 7468 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 7469 7470 testl(result, result); 7471 jcc(Assembler::zero, FALSE_LABEL); 7472 7473 movdqu(vec1, Address(ary1, result, Address::times_1, -16)); 7474 ptest(vec1, vec2); 7475 jccb(Assembler::notZero, TRUE_LABEL); 7476 jmpb(FALSE_LABEL); 7477 7478 bind(COMPARE_TAIL); // len is zero 7479 movl(len, result); 7480 // Fallthru to tail compare 7481 } 7482 } 7483 // Compare 4-byte vectors 7484 andl(len, 0xfffffffc); // vector count (in bytes) 7485 jccb(Assembler::zero, COMPARE_CHAR); 7486 7487 lea(ary1, Address(ary1, len, Address::times_1)); 7488 negptr(len); 7489 7490 bind(COMPARE_VECTORS); 7491 movl(tmp1, Address(ary1, len, Address::times_1)); 7492 andl(tmp1, 0x80808080); 7493 jccb(Assembler::notZero, TRUE_LABEL); 7494 addptr(len, 4); 7495 jcc(Assembler::notZero, COMPARE_VECTORS); 7496 7497 // Compare trailing char (final 2 bytes), if any 7498 bind(COMPARE_CHAR); 7499 testl(result, 0x2); // tail char 7500 jccb(Assembler::zero, COMPARE_BYTE); 7501 load_unsigned_short(tmp1, Address(ary1, 0)); 7502 andl(tmp1, 0x00008080); 7503 jccb(Assembler::notZero, TRUE_LABEL); 7504 subptr(result, 2); 7505 lea(ary1, Address(ary1, 2)); 7506 7507 bind(COMPARE_BYTE); 7508 testl(result, 0x1); // tail byte 7509 jccb(Assembler::zero, FALSE_LABEL); 7510 load_unsigned_byte(tmp1, Address(ary1, 0)); 7511 andl(tmp1, 0x00000080); 7512 jccb(Assembler::notEqual, TRUE_LABEL); 7513 jmpb(FALSE_LABEL); 7514 7515 bind(TRUE_LABEL); 7516 movl(result, 1); // return true 7517 jmpb(DONE); 7518 7519 bind(FALSE_LABEL); 7520 xorl(result, result); // return false 7521 7522 // That's it 7523 bind(DONE); 7524 if (UseAVX >= 2 && UseSSE >= 2) { 7525 // clean upper bits of YMM registers 7526 vpxor(vec1, vec1); 7527 vpxor(vec2, vec2); 7528 } 7529 } 7530 // Compare char[] or byte[] arrays aligned to 4 bytes or substrings. 7531 void MacroAssembler::arrays_equals(bool is_array_equ, Register ary1, Register ary2, 7532 Register limit, Register result, Register chr, 7533 XMMRegister vec1, XMMRegister vec2, bool is_char) { 7534 ShortBranchVerifier sbv(this); 7535 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR, COMPARE_BYTE; 7536 7537 int length_offset = arrayOopDesc::length_offset_in_bytes(); 7538 int base_offset = arrayOopDesc::base_offset_in_bytes(is_char ? T_CHAR : T_BYTE); 7539 7540 if (is_array_equ) { 7541 // Check the input args 7542 cmpoop(ary1, ary2); 7543 jcc(Assembler::equal, TRUE_LABEL); 7544 7545 // Need additional checks for arrays_equals. 7546 testptr(ary1, ary1); 7547 jcc(Assembler::zero, FALSE_LABEL); 7548 testptr(ary2, ary2); 7549 jcc(Assembler::zero, FALSE_LABEL); 7550 7551 // Check the lengths 7552 movl(limit, Address(ary1, length_offset)); 7553 cmpl(limit, Address(ary2, length_offset)); 7554 jcc(Assembler::notEqual, FALSE_LABEL); 7555 } 7556 7557 // count == 0 7558 testl(limit, limit); 7559 jcc(Assembler::zero, TRUE_LABEL); 7560 7561 if (is_array_equ) { 7562 // Load array address 7563 lea(ary1, Address(ary1, base_offset)); 7564 lea(ary2, Address(ary2, base_offset)); 7565 } 7566 7567 if (is_array_equ && is_char) { 7568 // arrays_equals when used for char[]. 7569 shll(limit, 1); // byte count != 0 7570 } 7571 movl(result, limit); // copy 7572 7573 if (UseAVX >= 2) { 7574 // With AVX2, use 32-byte vector compare 7575 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 7576 7577 // Compare 32-byte vectors 7578 andl(result, 0x0000001f); // tail count (in bytes) 7579 andl(limit, 0xffffffe0); // vector count (in bytes) 7580 jcc(Assembler::zero, COMPARE_TAIL); 7581 7582 lea(ary1, Address(ary1, limit, Address::times_1)); 7583 lea(ary2, Address(ary2, limit, Address::times_1)); 7584 negptr(limit); 7585 7586 bind(COMPARE_WIDE_VECTORS); 7587 7588 #ifdef _LP64 7589 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop 7590 Label COMPARE_WIDE_VECTORS_LOOP_AVX2, COMPARE_WIDE_VECTORS_LOOP_AVX3; 7591 7592 cmpl(limit, -64); 7593 jccb(Assembler::greater, COMPARE_WIDE_VECTORS_LOOP_AVX2); 7594 7595 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop 7596 7597 evmovdquq(vec1, Address(ary1, limit, Address::times_1), Assembler::AVX_512bit); 7598 evpcmpeqb(k7, vec1, Address(ary2, limit, Address::times_1), Assembler::AVX_512bit); 7599 kortestql(k7, k7); 7600 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare 7601 addptr(limit, 64); // update since we already compared at this addr 7602 cmpl(limit, -64); 7603 jccb(Assembler::lessEqual, COMPARE_WIDE_VECTORS_LOOP_AVX3); 7604 7605 // At this point we may still need to compare -limit+result bytes. 7606 // We could execute the next two instruction and just continue via non-wide path: 7607 // cmpl(limit, 0); 7608 // jcc(Assembler::equal, COMPARE_TAIL); // true 7609 // But since we stopped at the points ary{1,2}+limit which are 7610 // not farther than 64 bytes from the ends of arrays ary{1,2}+result 7611 // (|limit| <= 32 and result < 32), 7612 // we may just compare the last 64 bytes. 7613 // 7614 addptr(result, -64); // it is safe, bc we just came from this area 7615 evmovdquq(vec1, Address(ary1, result, Address::times_1), Assembler::AVX_512bit); 7616 evpcmpeqb(k7, vec1, Address(ary2, result, Address::times_1), Assembler::AVX_512bit); 7617 kortestql(k7, k7); 7618 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare 7619 7620 jmp(TRUE_LABEL); 7621 7622 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2); 7623 7624 }//if (VM_Version::supports_avx512vlbw()) 7625 #endif //_LP64 7626 7627 vmovdqu(vec1, Address(ary1, limit, Address::times_1)); 7628 vmovdqu(vec2, Address(ary2, limit, Address::times_1)); 7629 vpxor(vec1, vec2); 7630 7631 vptest(vec1, vec1); 7632 jcc(Assembler::notZero, FALSE_LABEL); 7633 addptr(limit, 32); 7634 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 7635 7636 testl(result, result); 7637 jcc(Assembler::zero, TRUE_LABEL); 7638 7639 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32)); 7640 vmovdqu(vec2, Address(ary2, result, Address::times_1, -32)); 7641 vpxor(vec1, vec2); 7642 7643 vptest(vec1, vec1); 7644 jccb(Assembler::notZero, FALSE_LABEL); 7645 jmpb(TRUE_LABEL); 7646 7647 bind(COMPARE_TAIL); // limit is zero 7648 movl(limit, result); 7649 // Fallthru to tail compare 7650 } else if (UseSSE42Intrinsics) { 7651 // With SSE4.2, use double quad vector compare 7652 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; 7653 7654 // Compare 16-byte vectors 7655 andl(result, 0x0000000f); // tail count (in bytes) 7656 andl(limit, 0xfffffff0); // vector count (in bytes) 7657 jcc(Assembler::zero, COMPARE_TAIL); 7658 7659 lea(ary1, Address(ary1, limit, Address::times_1)); 7660 lea(ary2, Address(ary2, limit, Address::times_1)); 7661 negptr(limit); 7662 7663 bind(COMPARE_WIDE_VECTORS); 7664 movdqu(vec1, Address(ary1, limit, Address::times_1)); 7665 movdqu(vec2, Address(ary2, limit, Address::times_1)); 7666 pxor(vec1, vec2); 7667 7668 ptest(vec1, vec1); 7669 jcc(Assembler::notZero, FALSE_LABEL); 7670 addptr(limit, 16); 7671 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); 7672 7673 testl(result, result); 7674 jcc(Assembler::zero, TRUE_LABEL); 7675 7676 movdqu(vec1, Address(ary1, result, Address::times_1, -16)); 7677 movdqu(vec2, Address(ary2, result, Address::times_1, -16)); 7678 pxor(vec1, vec2); 7679 7680 ptest(vec1, vec1); 7681 jccb(Assembler::notZero, FALSE_LABEL); 7682 jmpb(TRUE_LABEL); 7683 7684 bind(COMPARE_TAIL); // limit is zero 7685 movl(limit, result); 7686 // Fallthru to tail compare 7687 } 7688 7689 // Compare 4-byte vectors 7690 andl(limit, 0xfffffffc); // vector count (in bytes) 7691 jccb(Assembler::zero, COMPARE_CHAR); 7692 7693 lea(ary1, Address(ary1, limit, Address::times_1)); 7694 lea(ary2, Address(ary2, limit, Address::times_1)); 7695 negptr(limit); 7696 7697 bind(COMPARE_VECTORS); 7698 movl(chr, Address(ary1, limit, Address::times_1)); 7699 cmpl(chr, Address(ary2, limit, Address::times_1)); 7700 jccb(Assembler::notEqual, FALSE_LABEL); 7701 addptr(limit, 4); 7702 jcc(Assembler::notZero, COMPARE_VECTORS); 7703 7704 // Compare trailing char (final 2 bytes), if any 7705 bind(COMPARE_CHAR); 7706 testl(result, 0x2); // tail char 7707 jccb(Assembler::zero, COMPARE_BYTE); 7708 load_unsigned_short(chr, Address(ary1, 0)); 7709 load_unsigned_short(limit, Address(ary2, 0)); 7710 cmpl(chr, limit); 7711 jccb(Assembler::notEqual, FALSE_LABEL); 7712 7713 if (is_array_equ && is_char) { 7714 bind(COMPARE_BYTE); 7715 } else { 7716 lea(ary1, Address(ary1, 2)); 7717 lea(ary2, Address(ary2, 2)); 7718 7719 bind(COMPARE_BYTE); 7720 testl(result, 0x1); // tail byte 7721 jccb(Assembler::zero, TRUE_LABEL); 7722 load_unsigned_byte(chr, Address(ary1, 0)); 7723 load_unsigned_byte(limit, Address(ary2, 0)); 7724 cmpl(chr, limit); 7725 jccb(Assembler::notEqual, FALSE_LABEL); 7726 } 7727 bind(TRUE_LABEL); 7728 movl(result, 1); // return true 7729 jmpb(DONE); 7730 7731 bind(FALSE_LABEL); 7732 xorl(result, result); // return false 7733 7734 // That's it 7735 bind(DONE); 7736 if (UseAVX >= 2) { 7737 // clean upper bits of YMM registers 7738 vpxor(vec1, vec1); 7739 vpxor(vec2, vec2); 7740 } 7741 } 7742 7743 #endif 7744 7745 void MacroAssembler::generate_fill(BasicType t, bool aligned, 7746 Register to, Register value, Register count, 7747 Register rtmp, XMMRegister xtmp) { 7748 ShortBranchVerifier sbv(this); 7749 assert_different_registers(to, value, count, rtmp); 7750 Label L_exit; 7751 Label L_fill_2_bytes, L_fill_4_bytes; 7752 7753 int shift = -1; 7754 switch (t) { 7755 case T_BYTE: 7756 shift = 2; 7757 break; 7758 case T_SHORT: 7759 shift = 1; 7760 break; 7761 case T_INT: 7762 shift = 0; 7763 break; 7764 default: ShouldNotReachHere(); 7765 } 7766 7767 if (t == T_BYTE) { 7768 andl(value, 0xff); 7769 movl(rtmp, value); 7770 shll(rtmp, 8); 7771 orl(value, rtmp); 7772 } 7773 if (t == T_SHORT) { 7774 andl(value, 0xffff); 7775 } 7776 if (t == T_BYTE || t == T_SHORT) { 7777 movl(rtmp, value); 7778 shll(rtmp, 16); 7779 orl(value, rtmp); 7780 } 7781 7782 cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element 7783 jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp 7784 if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) { 7785 Label L_skip_align2; 7786 // align source address at 4 bytes address boundary 7787 if (t == T_BYTE) { 7788 Label L_skip_align1; 7789 // One byte misalignment happens only for byte arrays 7790 testptr(to, 1); 7791 jccb(Assembler::zero, L_skip_align1); 7792 movb(Address(to, 0), value); 7793 increment(to); 7794 decrement(count); 7795 BIND(L_skip_align1); 7796 } 7797 // Two bytes misalignment happens only for byte and short (char) arrays 7798 testptr(to, 2); 7799 jccb(Assembler::zero, L_skip_align2); 7800 movw(Address(to, 0), value); 7801 addptr(to, 2); 7802 subl(count, 1<<(shift-1)); 7803 BIND(L_skip_align2); 7804 } 7805 if (UseSSE < 2) { 7806 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes; 7807 // Fill 32-byte chunks 7808 subl(count, 8 << shift); 7809 jcc(Assembler::less, L_check_fill_8_bytes); 7810 align(16); 7811 7812 BIND(L_fill_32_bytes_loop); 7813 7814 for (int i = 0; i < 32; i += 4) { 7815 movl(Address(to, i), value); 7816 } 7817 7818 addptr(to, 32); 7819 subl(count, 8 << shift); 7820 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop); 7821 BIND(L_check_fill_8_bytes); 7822 addl(count, 8 << shift); 7823 jccb(Assembler::zero, L_exit); 7824 jmpb(L_fill_8_bytes); 7825 7826 // 7827 // length is too short, just fill qwords 7828 // 7829 BIND(L_fill_8_bytes_loop); 7830 movl(Address(to, 0), value); 7831 movl(Address(to, 4), value); 7832 addptr(to, 8); 7833 BIND(L_fill_8_bytes); 7834 subl(count, 1 << (shift + 1)); 7835 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop); 7836 // fall through to fill 4 bytes 7837 } else { 7838 Label L_fill_32_bytes; 7839 if (!UseUnalignedLoadStores) { 7840 // align to 8 bytes, we know we are 4 byte aligned to start 7841 testptr(to, 4); 7842 jccb(Assembler::zero, L_fill_32_bytes); 7843 movl(Address(to, 0), value); 7844 addptr(to, 4); 7845 subl(count, 1<<shift); 7846 } 7847 BIND(L_fill_32_bytes); 7848 { 7849 assert( UseSSE >= 2, "supported cpu only" ); 7850 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes; 7851 movdl(xtmp, value); 7852 if (UseAVX > 2 && UseUnalignedLoadStores) { 7853 // Fill 64-byte chunks 7854 Label L_fill_64_bytes_loop, L_check_fill_32_bytes; 7855 vpbroadcastd(xtmp, xtmp, Assembler::AVX_512bit); 7856 7857 subl(count, 16 << shift); 7858 jcc(Assembler::less, L_check_fill_32_bytes); 7859 align(16); 7860 7861 BIND(L_fill_64_bytes_loop); 7862 evmovdqul(Address(to, 0), xtmp, Assembler::AVX_512bit); 7863 addptr(to, 64); 7864 subl(count, 16 << shift); 7865 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop); 7866 7867 BIND(L_check_fill_32_bytes); 7868 addl(count, 8 << shift); 7869 jccb(Assembler::less, L_check_fill_8_bytes); 7870 vmovdqu(Address(to, 0), xtmp); 7871 addptr(to, 32); 7872 subl(count, 8 << shift); 7873 7874 BIND(L_check_fill_8_bytes); 7875 } else if (UseAVX == 2 && UseUnalignedLoadStores) { 7876 // Fill 64-byte chunks 7877 Label L_fill_64_bytes_loop, L_check_fill_32_bytes; 7878 vpbroadcastd(xtmp, xtmp, Assembler::AVX_256bit); 7879 7880 subl(count, 16 << shift); 7881 jcc(Assembler::less, L_check_fill_32_bytes); 7882 align(16); 7883 7884 BIND(L_fill_64_bytes_loop); 7885 vmovdqu(Address(to, 0), xtmp); 7886 vmovdqu(Address(to, 32), xtmp); 7887 addptr(to, 64); 7888 subl(count, 16 << shift); 7889 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop); 7890 7891 BIND(L_check_fill_32_bytes); 7892 addl(count, 8 << shift); 7893 jccb(Assembler::less, L_check_fill_8_bytes); 7894 vmovdqu(Address(to, 0), xtmp); 7895 addptr(to, 32); 7896 subl(count, 8 << shift); 7897 7898 BIND(L_check_fill_8_bytes); 7899 // clean upper bits of YMM registers 7900 movdl(xtmp, value); 7901 pshufd(xtmp, xtmp, 0); 7902 } else { 7903 // Fill 32-byte chunks 7904 pshufd(xtmp, xtmp, 0); 7905 7906 subl(count, 8 << shift); 7907 jcc(Assembler::less, L_check_fill_8_bytes); 7908 align(16); 7909 7910 BIND(L_fill_32_bytes_loop); 7911 7912 if (UseUnalignedLoadStores) { 7913 movdqu(Address(to, 0), xtmp); 7914 movdqu(Address(to, 16), xtmp); 7915 } else { 7916 movq(Address(to, 0), xtmp); 7917 movq(Address(to, 8), xtmp); 7918 movq(Address(to, 16), xtmp); 7919 movq(Address(to, 24), xtmp); 7920 } 7921 7922 addptr(to, 32); 7923 subl(count, 8 << shift); 7924 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop); 7925 7926 BIND(L_check_fill_8_bytes); 7927 } 7928 addl(count, 8 << shift); 7929 jccb(Assembler::zero, L_exit); 7930 jmpb(L_fill_8_bytes); 7931 7932 // 7933 // length is too short, just fill qwords 7934 // 7935 BIND(L_fill_8_bytes_loop); 7936 movq(Address(to, 0), xtmp); 7937 addptr(to, 8); 7938 BIND(L_fill_8_bytes); 7939 subl(count, 1 << (shift + 1)); 7940 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop); 7941 } 7942 } 7943 // fill trailing 4 bytes 7944 BIND(L_fill_4_bytes); 7945 testl(count, 1<<shift); 7946 jccb(Assembler::zero, L_fill_2_bytes); 7947 movl(Address(to, 0), value); 7948 if (t == T_BYTE || t == T_SHORT) { 7949 Label L_fill_byte; 7950 addptr(to, 4); 7951 BIND(L_fill_2_bytes); 7952 // fill trailing 2 bytes 7953 testl(count, 1<<(shift-1)); 7954 jccb(Assembler::zero, L_fill_byte); 7955 movw(Address(to, 0), value); 7956 if (t == T_BYTE) { 7957 addptr(to, 2); 7958 BIND(L_fill_byte); 7959 // fill trailing byte 7960 testl(count, 1); 7961 jccb(Assembler::zero, L_exit); 7962 movb(Address(to, 0), value); 7963 } else { 7964 BIND(L_fill_byte); 7965 } 7966 } else { 7967 BIND(L_fill_2_bytes); 7968 } 7969 BIND(L_exit); 7970 } 7971 7972 // encode char[] to byte[] in ISO_8859_1 7973 //@HotSpotIntrinsicCandidate 7974 //private static int implEncodeISOArray(byte[] sa, int sp, 7975 //byte[] da, int dp, int len) { 7976 // int i = 0; 7977 // for (; i < len; i++) { 7978 // char c = StringUTF16.getChar(sa, sp++); 7979 // if (c > '\u00FF') 7980 // break; 7981 // da[dp++] = (byte)c; 7982 // } 7983 // return i; 7984 //} 7985 void MacroAssembler::encode_iso_array(Register src, Register dst, Register len, 7986 XMMRegister tmp1Reg, XMMRegister tmp2Reg, 7987 XMMRegister tmp3Reg, XMMRegister tmp4Reg, 7988 Register tmp5, Register result) { 7989 7990 // rsi: src 7991 // rdi: dst 7992 // rdx: len 7993 // rcx: tmp5 7994 // rax: result 7995 ShortBranchVerifier sbv(this); 7996 assert_different_registers(src, dst, len, tmp5, result); 7997 Label L_done, L_copy_1_char, L_copy_1_char_exit; 7998 7999 // set result 8000 xorl(result, result); 8001 // check for zero length 8002 testl(len, len); 8003 jcc(Assembler::zero, L_done); 8004 8005 movl(result, len); 8006 8007 // Setup pointers 8008 lea(src, Address(src, len, Address::times_2)); // char[] 8009 lea(dst, Address(dst, len, Address::times_1)); // byte[] 8010 negptr(len); 8011 8012 if (UseSSE42Intrinsics || UseAVX >= 2) { 8013 Label L_copy_8_chars, L_copy_8_chars_exit; 8014 Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit; 8015 8016 if (UseAVX >= 2) { 8017 Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit; 8018 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector 8019 movdl(tmp1Reg, tmp5); 8020 vpbroadcastd(tmp1Reg, tmp1Reg, Assembler::AVX_256bit); 8021 jmp(L_chars_32_check); 8022 8023 bind(L_copy_32_chars); 8024 vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64)); 8025 vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32)); 8026 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1); 8027 vptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 8028 jccb(Assembler::notZero, L_copy_32_chars_exit); 8029 vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1); 8030 vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector_len */ 1); 8031 vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg); 8032 8033 bind(L_chars_32_check); 8034 addptr(len, 32); 8035 jcc(Assembler::lessEqual, L_copy_32_chars); 8036 8037 bind(L_copy_32_chars_exit); 8038 subptr(len, 16); 8039 jccb(Assembler::greater, L_copy_16_chars_exit); 8040 8041 } else if (UseSSE42Intrinsics) { 8042 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector 8043 movdl(tmp1Reg, tmp5); 8044 pshufd(tmp1Reg, tmp1Reg, 0); 8045 jmpb(L_chars_16_check); 8046 } 8047 8048 bind(L_copy_16_chars); 8049 if (UseAVX >= 2) { 8050 vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32)); 8051 vptest(tmp2Reg, tmp1Reg); 8052 jcc(Assembler::notZero, L_copy_16_chars_exit); 8053 vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector_len */ 1); 8054 vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector_len */ 1); 8055 } else { 8056 if (UseAVX > 0) { 8057 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32)); 8058 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16)); 8059 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 0); 8060 } else { 8061 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32)); 8062 por(tmp2Reg, tmp3Reg); 8063 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16)); 8064 por(tmp2Reg, tmp4Reg); 8065 } 8066 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 8067 jccb(Assembler::notZero, L_copy_16_chars_exit); 8068 packuswb(tmp3Reg, tmp4Reg); 8069 } 8070 movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg); 8071 8072 bind(L_chars_16_check); 8073 addptr(len, 16); 8074 jcc(Assembler::lessEqual, L_copy_16_chars); 8075 8076 bind(L_copy_16_chars_exit); 8077 if (UseAVX >= 2) { 8078 // clean upper bits of YMM registers 8079 vpxor(tmp2Reg, tmp2Reg); 8080 vpxor(tmp3Reg, tmp3Reg); 8081 vpxor(tmp4Reg, tmp4Reg); 8082 movdl(tmp1Reg, tmp5); 8083 pshufd(tmp1Reg, tmp1Reg, 0); 8084 } 8085 subptr(len, 8); 8086 jccb(Assembler::greater, L_copy_8_chars_exit); 8087 8088 bind(L_copy_8_chars); 8089 movdqu(tmp3Reg, Address(src, len, Address::times_2, -16)); 8090 ptest(tmp3Reg, tmp1Reg); 8091 jccb(Assembler::notZero, L_copy_8_chars_exit); 8092 packuswb(tmp3Reg, tmp1Reg); 8093 movq(Address(dst, len, Address::times_1, -8), tmp3Reg); 8094 addptr(len, 8); 8095 jccb(Assembler::lessEqual, L_copy_8_chars); 8096 8097 bind(L_copy_8_chars_exit); 8098 subptr(len, 8); 8099 jccb(Assembler::zero, L_done); 8100 } 8101 8102 bind(L_copy_1_char); 8103 load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0)); 8104 testl(tmp5, 0xff00); // check if Unicode char 8105 jccb(Assembler::notZero, L_copy_1_char_exit); 8106 movb(Address(dst, len, Address::times_1, 0), tmp5); 8107 addptr(len, 1); 8108 jccb(Assembler::less, L_copy_1_char); 8109 8110 bind(L_copy_1_char_exit); 8111 addptr(result, len); // len is negative count of not processed elements 8112 8113 bind(L_done); 8114 } 8115 8116 #ifdef _LP64 8117 /** 8118 * Helper for multiply_to_len(). 8119 */ 8120 void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo, Register src1, Register src2) { 8121 addq(dest_lo, src1); 8122 adcq(dest_hi, 0); 8123 addq(dest_lo, src2); 8124 adcq(dest_hi, 0); 8125 } 8126 8127 /** 8128 * Multiply 64 bit by 64 bit first loop. 8129 */ 8130 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart, 8131 Register y, Register y_idx, Register z, 8132 Register carry, Register product, 8133 Register idx, Register kdx) { 8134 // 8135 // jlong carry, x[], y[], z[]; 8136 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) { 8137 // huge_128 product = y[idx] * x[xstart] + carry; 8138 // z[kdx] = (jlong)product; 8139 // carry = (jlong)(product >>> 64); 8140 // } 8141 // z[xstart] = carry; 8142 // 8143 8144 Label L_first_loop, L_first_loop_exit; 8145 Label L_one_x, L_one_y, L_multiply; 8146 8147 decrementl(xstart); 8148 jcc(Assembler::negative, L_one_x); 8149 8150 movq(x_xstart, Address(x, xstart, Address::times_4, 0)); 8151 rorq(x_xstart, 32); // convert big-endian to little-endian 8152 8153 bind(L_first_loop); 8154 decrementl(idx); 8155 jcc(Assembler::negative, L_first_loop_exit); 8156 decrementl(idx); 8157 jcc(Assembler::negative, L_one_y); 8158 movq(y_idx, Address(y, idx, Address::times_4, 0)); 8159 rorq(y_idx, 32); // convert big-endian to little-endian 8160 bind(L_multiply); 8161 movq(product, x_xstart); 8162 mulq(y_idx); // product(rax) * y_idx -> rdx:rax 8163 addq(product, carry); 8164 adcq(rdx, 0); 8165 subl(kdx, 2); 8166 movl(Address(z, kdx, Address::times_4, 4), product); 8167 shrq(product, 32); 8168 movl(Address(z, kdx, Address::times_4, 0), product); 8169 movq(carry, rdx); 8170 jmp(L_first_loop); 8171 8172 bind(L_one_y); 8173 movl(y_idx, Address(y, 0)); 8174 jmp(L_multiply); 8175 8176 bind(L_one_x); 8177 movl(x_xstart, Address(x, 0)); 8178 jmp(L_first_loop); 8179 8180 bind(L_first_loop_exit); 8181 } 8182 8183 /** 8184 * Multiply 64 bit by 64 bit and add 128 bit. 8185 */ 8186 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y, Register z, 8187 Register yz_idx, Register idx, 8188 Register carry, Register product, int offset) { 8189 // huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry; 8190 // z[kdx] = (jlong)product; 8191 8192 movq(yz_idx, Address(y, idx, Address::times_4, offset)); 8193 rorq(yz_idx, 32); // convert big-endian to little-endian 8194 movq(product, x_xstart); 8195 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax) 8196 movq(yz_idx, Address(z, idx, Address::times_4, offset)); 8197 rorq(yz_idx, 32); // convert big-endian to little-endian 8198 8199 add2_with_carry(rdx, product, carry, yz_idx); 8200 8201 movl(Address(z, idx, Address::times_4, offset+4), product); 8202 shrq(product, 32); 8203 movl(Address(z, idx, Address::times_4, offset), product); 8204 8205 } 8206 8207 /** 8208 * Multiply 128 bit by 128 bit. Unrolled inner loop. 8209 */ 8210 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart, Register y, Register z, 8211 Register yz_idx, Register idx, Register jdx, 8212 Register carry, Register product, 8213 Register carry2) { 8214 // jlong carry, x[], y[], z[]; 8215 // int kdx = ystart+1; 8216 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop 8217 // huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry; 8218 // z[kdx+idx+1] = (jlong)product; 8219 // jlong carry2 = (jlong)(product >>> 64); 8220 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry2; 8221 // z[kdx+idx] = (jlong)product; 8222 // carry = (jlong)(product >>> 64); 8223 // } 8224 // idx += 2; 8225 // if (idx > 0) { 8226 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry; 8227 // z[kdx+idx] = (jlong)product; 8228 // carry = (jlong)(product >>> 64); 8229 // } 8230 // 8231 8232 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done; 8233 8234 movl(jdx, idx); 8235 andl(jdx, 0xFFFFFFFC); 8236 shrl(jdx, 2); 8237 8238 bind(L_third_loop); 8239 subl(jdx, 1); 8240 jcc(Assembler::negative, L_third_loop_exit); 8241 subl(idx, 4); 8242 8243 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8); 8244 movq(carry2, rdx); 8245 8246 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0); 8247 movq(carry, rdx); 8248 jmp(L_third_loop); 8249 8250 bind (L_third_loop_exit); 8251 8252 andl (idx, 0x3); 8253 jcc(Assembler::zero, L_post_third_loop_done); 8254 8255 Label L_check_1; 8256 subl(idx, 2); 8257 jcc(Assembler::negative, L_check_1); 8258 8259 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0); 8260 movq(carry, rdx); 8261 8262 bind (L_check_1); 8263 addl (idx, 0x2); 8264 andl (idx, 0x1); 8265 subl(idx, 1); 8266 jcc(Assembler::negative, L_post_third_loop_done); 8267 8268 movl(yz_idx, Address(y, idx, Address::times_4, 0)); 8269 movq(product, x_xstart); 8270 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax) 8271 movl(yz_idx, Address(z, idx, Address::times_4, 0)); 8272 8273 add2_with_carry(rdx, product, yz_idx, carry); 8274 8275 movl(Address(z, idx, Address::times_4, 0), product); 8276 shrq(product, 32); 8277 8278 shlq(rdx, 32); 8279 orq(product, rdx); 8280 movq(carry, product); 8281 8282 bind(L_post_third_loop_done); 8283 } 8284 8285 /** 8286 * Multiply 128 bit by 128 bit using BMI2. Unrolled inner loop. 8287 * 8288 */ 8289 void MacroAssembler::multiply_128_x_128_bmi2_loop(Register y, Register z, 8290 Register carry, Register carry2, 8291 Register idx, Register jdx, 8292 Register yz_idx1, Register yz_idx2, 8293 Register tmp, Register tmp3, Register tmp4) { 8294 assert(UseBMI2Instructions, "should be used only when BMI2 is available"); 8295 8296 // jlong carry, x[], y[], z[]; 8297 // int kdx = ystart+1; 8298 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop 8299 // huge_128 tmp3 = (y[idx+1] * rdx) + z[kdx+idx+1] + carry; 8300 // jlong carry2 = (jlong)(tmp3 >>> 64); 8301 // huge_128 tmp4 = (y[idx] * rdx) + z[kdx+idx] + carry2; 8302 // carry = (jlong)(tmp4 >>> 64); 8303 // z[kdx+idx+1] = (jlong)tmp3; 8304 // z[kdx+idx] = (jlong)tmp4; 8305 // } 8306 // idx += 2; 8307 // if (idx > 0) { 8308 // yz_idx1 = (y[idx] * rdx) + z[kdx+idx] + carry; 8309 // z[kdx+idx] = (jlong)yz_idx1; 8310 // carry = (jlong)(yz_idx1 >>> 64); 8311 // } 8312 // 8313 8314 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done; 8315 8316 movl(jdx, idx); 8317 andl(jdx, 0xFFFFFFFC); 8318 shrl(jdx, 2); 8319 8320 bind(L_third_loop); 8321 subl(jdx, 1); 8322 jcc(Assembler::negative, L_third_loop_exit); 8323 subl(idx, 4); 8324 8325 movq(yz_idx1, Address(y, idx, Address::times_4, 8)); 8326 rorxq(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian 8327 movq(yz_idx2, Address(y, idx, Address::times_4, 0)); 8328 rorxq(yz_idx2, yz_idx2, 32); 8329 8330 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3 8331 mulxq(carry2, tmp, yz_idx2); // yz_idx2 * rdx -> carry2:tmp 8332 8333 movq(yz_idx1, Address(z, idx, Address::times_4, 8)); 8334 rorxq(yz_idx1, yz_idx1, 32); 8335 movq(yz_idx2, Address(z, idx, Address::times_4, 0)); 8336 rorxq(yz_idx2, yz_idx2, 32); 8337 8338 if (VM_Version::supports_adx()) { 8339 adcxq(tmp3, carry); 8340 adoxq(tmp3, yz_idx1); 8341 8342 adcxq(tmp4, tmp); 8343 adoxq(tmp4, yz_idx2); 8344 8345 movl(carry, 0); // does not affect flags 8346 adcxq(carry2, carry); 8347 adoxq(carry2, carry); 8348 } else { 8349 add2_with_carry(tmp4, tmp3, carry, yz_idx1); 8350 add2_with_carry(carry2, tmp4, tmp, yz_idx2); 8351 } 8352 movq(carry, carry2); 8353 8354 movl(Address(z, idx, Address::times_4, 12), tmp3); 8355 shrq(tmp3, 32); 8356 movl(Address(z, idx, Address::times_4, 8), tmp3); 8357 8358 movl(Address(z, idx, Address::times_4, 4), tmp4); 8359 shrq(tmp4, 32); 8360 movl(Address(z, idx, Address::times_4, 0), tmp4); 8361 8362 jmp(L_third_loop); 8363 8364 bind (L_third_loop_exit); 8365 8366 andl (idx, 0x3); 8367 jcc(Assembler::zero, L_post_third_loop_done); 8368 8369 Label L_check_1; 8370 subl(idx, 2); 8371 jcc(Assembler::negative, L_check_1); 8372 8373 movq(yz_idx1, Address(y, idx, Address::times_4, 0)); 8374 rorxq(yz_idx1, yz_idx1, 32); 8375 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3 8376 movq(yz_idx2, Address(z, idx, Address::times_4, 0)); 8377 rorxq(yz_idx2, yz_idx2, 32); 8378 8379 add2_with_carry(tmp4, tmp3, carry, yz_idx2); 8380 8381 movl(Address(z, idx, Address::times_4, 4), tmp3); 8382 shrq(tmp3, 32); 8383 movl(Address(z, idx, Address::times_4, 0), tmp3); 8384 movq(carry, tmp4); 8385 8386 bind (L_check_1); 8387 addl (idx, 0x2); 8388 andl (idx, 0x1); 8389 subl(idx, 1); 8390 jcc(Assembler::negative, L_post_third_loop_done); 8391 movl(tmp4, Address(y, idx, Address::times_4, 0)); 8392 mulxq(carry2, tmp3, tmp4); // tmp4 * rdx -> carry2:tmp3 8393 movl(tmp4, Address(z, idx, Address::times_4, 0)); 8394 8395 add2_with_carry(carry2, tmp3, tmp4, carry); 8396 8397 movl(Address(z, idx, Address::times_4, 0), tmp3); 8398 shrq(tmp3, 32); 8399 8400 shlq(carry2, 32); 8401 orq(tmp3, carry2); 8402 movq(carry, tmp3); 8403 8404 bind(L_post_third_loop_done); 8405 } 8406 8407 /** 8408 * Code for BigInteger::multiplyToLen() instrinsic. 8409 * 8410 * rdi: x 8411 * rax: xlen 8412 * rsi: y 8413 * rcx: ylen 8414 * r8: z 8415 * r11: zlen 8416 * r12: tmp1 8417 * r13: tmp2 8418 * r14: tmp3 8419 * r15: tmp4 8420 * rbx: tmp5 8421 * 8422 */ 8423 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, Register z, Register zlen, 8424 Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5) { 8425 ShortBranchVerifier sbv(this); 8426 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx); 8427 8428 push(tmp1); 8429 push(tmp2); 8430 push(tmp3); 8431 push(tmp4); 8432 push(tmp5); 8433 8434 push(xlen); 8435 push(zlen); 8436 8437 const Register idx = tmp1; 8438 const Register kdx = tmp2; 8439 const Register xstart = tmp3; 8440 8441 const Register y_idx = tmp4; 8442 const Register carry = tmp5; 8443 const Register product = xlen; 8444 const Register x_xstart = zlen; // reuse register 8445 8446 // First Loop. 8447 // 8448 // final static long LONG_MASK = 0xffffffffL; 8449 // int xstart = xlen - 1; 8450 // int ystart = ylen - 1; 8451 // long carry = 0; 8452 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) { 8453 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry; 8454 // z[kdx] = (int)product; 8455 // carry = product >>> 32; 8456 // } 8457 // z[xstart] = (int)carry; 8458 // 8459 8460 movl(idx, ylen); // idx = ylen; 8461 movl(kdx, zlen); // kdx = xlen+ylen; 8462 xorq(carry, carry); // carry = 0; 8463 8464 Label L_done; 8465 8466 movl(xstart, xlen); 8467 decrementl(xstart); 8468 jcc(Assembler::negative, L_done); 8469 8470 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx); 8471 8472 Label L_second_loop; 8473 testl(kdx, kdx); 8474 jcc(Assembler::zero, L_second_loop); 8475 8476 Label L_carry; 8477 subl(kdx, 1); 8478 jcc(Assembler::zero, L_carry); 8479 8480 movl(Address(z, kdx, Address::times_4, 0), carry); 8481 shrq(carry, 32); 8482 subl(kdx, 1); 8483 8484 bind(L_carry); 8485 movl(Address(z, kdx, Address::times_4, 0), carry); 8486 8487 // Second and third (nested) loops. 8488 // 8489 // for (int i = xstart-1; i >= 0; i--) { // Second loop 8490 // carry = 0; 8491 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop 8492 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) + 8493 // (z[k] & LONG_MASK) + carry; 8494 // z[k] = (int)product; 8495 // carry = product >>> 32; 8496 // } 8497 // z[i] = (int)carry; 8498 // } 8499 // 8500 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx 8501 8502 const Register jdx = tmp1; 8503 8504 bind(L_second_loop); 8505 xorl(carry, carry); // carry = 0; 8506 movl(jdx, ylen); // j = ystart+1 8507 8508 subl(xstart, 1); // i = xstart-1; 8509 jcc(Assembler::negative, L_done); 8510 8511 push (z); 8512 8513 Label L_last_x; 8514 lea(z, Address(z, xstart, Address::times_4, 4)); // z = z + k - j 8515 subl(xstart, 1); // i = xstart-1; 8516 jcc(Assembler::negative, L_last_x); 8517 8518 if (UseBMI2Instructions) { 8519 movq(rdx, Address(x, xstart, Address::times_4, 0)); 8520 rorxq(rdx, rdx, 32); // convert big-endian to little-endian 8521 } else { 8522 movq(x_xstart, Address(x, xstart, Address::times_4, 0)); 8523 rorq(x_xstart, 32); // convert big-endian to little-endian 8524 } 8525 8526 Label L_third_loop_prologue; 8527 bind(L_third_loop_prologue); 8528 8529 push (x); 8530 push (xstart); 8531 push (ylen); 8532 8533 8534 if (UseBMI2Instructions) { 8535 multiply_128_x_128_bmi2_loop(y, z, carry, x, jdx, ylen, product, tmp2, x_xstart, tmp3, tmp4); 8536 } else { // !UseBMI2Instructions 8537 multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x); 8538 } 8539 8540 pop(ylen); 8541 pop(xlen); 8542 pop(x); 8543 pop(z); 8544 8545 movl(tmp3, xlen); 8546 addl(tmp3, 1); 8547 movl(Address(z, tmp3, Address::times_4, 0), carry); 8548 subl(tmp3, 1); 8549 jccb(Assembler::negative, L_done); 8550 8551 shrq(carry, 32); 8552 movl(Address(z, tmp3, Address::times_4, 0), carry); 8553 jmp(L_second_loop); 8554 8555 // Next infrequent code is moved outside loops. 8556 bind(L_last_x); 8557 if (UseBMI2Instructions) { 8558 movl(rdx, Address(x, 0)); 8559 } else { 8560 movl(x_xstart, Address(x, 0)); 8561 } 8562 jmp(L_third_loop_prologue); 8563 8564 bind(L_done); 8565 8566 pop(zlen); 8567 pop(xlen); 8568 8569 pop(tmp5); 8570 pop(tmp4); 8571 pop(tmp3); 8572 pop(tmp2); 8573 pop(tmp1); 8574 } 8575 8576 void MacroAssembler::vectorized_mismatch(Register obja, Register objb, Register length, Register log2_array_indxscale, 8577 Register result, Register tmp1, Register tmp2, XMMRegister rymm0, XMMRegister rymm1, XMMRegister rymm2){ 8578 assert(UseSSE42Intrinsics, "SSE4.2 must be enabled."); 8579 Label VECTOR16_LOOP, VECTOR8_LOOP, VECTOR4_LOOP; 8580 Label VECTOR8_TAIL, VECTOR4_TAIL; 8581 Label VECTOR32_NOT_EQUAL, VECTOR16_NOT_EQUAL, VECTOR8_NOT_EQUAL, VECTOR4_NOT_EQUAL; 8582 Label SAME_TILL_END, DONE; 8583 Label BYTES_LOOP, BYTES_TAIL, BYTES_NOT_EQUAL; 8584 8585 //scale is in rcx in both Win64 and Unix 8586 ShortBranchVerifier sbv(this); 8587 8588 shlq(length); 8589 xorq(result, result); 8590 8591 if ((UseAVX > 2) && 8592 VM_Version::supports_avx512vlbw()) { 8593 Label VECTOR64_LOOP, VECTOR64_NOT_EQUAL, VECTOR32_TAIL; 8594 8595 cmpq(length, 64); 8596 jcc(Assembler::less, VECTOR32_TAIL); 8597 movq(tmp1, length); 8598 andq(tmp1, 0x3F); // tail count 8599 andq(length, ~(0x3F)); //vector count 8600 8601 bind(VECTOR64_LOOP); 8602 // AVX512 code to compare 64 byte vectors. 8603 evmovdqub(rymm0, Address(obja, result), Assembler::AVX_512bit); 8604 evpcmpeqb(k7, rymm0, Address(objb, result), Assembler::AVX_512bit); 8605 kortestql(k7, k7); 8606 jcc(Assembler::aboveEqual, VECTOR64_NOT_EQUAL); // mismatch 8607 addq(result, 64); 8608 subq(length, 64); 8609 jccb(Assembler::notZero, VECTOR64_LOOP); 8610 8611 //bind(VECTOR64_TAIL); 8612 testq(tmp1, tmp1); 8613 jcc(Assembler::zero, SAME_TILL_END); 8614 8615 //bind(VECTOR64_TAIL); 8616 // AVX512 code to compare upto 63 byte vectors. 8617 mov64(tmp2, 0xFFFFFFFFFFFFFFFF); 8618 shlxq(tmp2, tmp2, tmp1); 8619 notq(tmp2); 8620 kmovql(k3, tmp2); 8621 8622 evmovdqub(rymm0, k3, Address(obja, result), Assembler::AVX_512bit); 8623 evpcmpeqb(k7, k3, rymm0, Address(objb, result), Assembler::AVX_512bit); 8624 8625 ktestql(k7, k3); 8626 jcc(Assembler::below, SAME_TILL_END); // not mismatch 8627 8628 bind(VECTOR64_NOT_EQUAL); 8629 kmovql(tmp1, k7); 8630 notq(tmp1); 8631 tzcntq(tmp1, tmp1); 8632 addq(result, tmp1); 8633 shrq(result); 8634 jmp(DONE); 8635 bind(VECTOR32_TAIL); 8636 } 8637 8638 cmpq(length, 8); 8639 jcc(Assembler::equal, VECTOR8_LOOP); 8640 jcc(Assembler::less, VECTOR4_TAIL); 8641 8642 if (UseAVX >= 2) { 8643 Label VECTOR16_TAIL, VECTOR32_LOOP; 8644 8645 cmpq(length, 16); 8646 jcc(Assembler::equal, VECTOR16_LOOP); 8647 jcc(Assembler::less, VECTOR8_LOOP); 8648 8649 cmpq(length, 32); 8650 jccb(Assembler::less, VECTOR16_TAIL); 8651 8652 subq(length, 32); 8653 bind(VECTOR32_LOOP); 8654 vmovdqu(rymm0, Address(obja, result)); 8655 vmovdqu(rymm1, Address(objb, result)); 8656 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_256bit); 8657 vptest(rymm2, rymm2); 8658 jcc(Assembler::notZero, VECTOR32_NOT_EQUAL);//mismatch found 8659 addq(result, 32); 8660 subq(length, 32); 8661 jcc(Assembler::greaterEqual, VECTOR32_LOOP); 8662 addq(length, 32); 8663 jcc(Assembler::equal, SAME_TILL_END); 8664 //falling through if less than 32 bytes left //close the branch here. 8665 8666 bind(VECTOR16_TAIL); 8667 cmpq(length, 16); 8668 jccb(Assembler::less, VECTOR8_TAIL); 8669 bind(VECTOR16_LOOP); 8670 movdqu(rymm0, Address(obja, result)); 8671 movdqu(rymm1, Address(objb, result)); 8672 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_128bit); 8673 ptest(rymm2, rymm2); 8674 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found 8675 addq(result, 16); 8676 subq(length, 16); 8677 jcc(Assembler::equal, SAME_TILL_END); 8678 //falling through if less than 16 bytes left 8679 } else {//regular intrinsics 8680 8681 cmpq(length, 16); 8682 jccb(Assembler::less, VECTOR8_TAIL); 8683 8684 subq(length, 16); 8685 bind(VECTOR16_LOOP); 8686 movdqu(rymm0, Address(obja, result)); 8687 movdqu(rymm1, Address(objb, result)); 8688 pxor(rymm0, rymm1); 8689 ptest(rymm0, rymm0); 8690 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found 8691 addq(result, 16); 8692 subq(length, 16); 8693 jccb(Assembler::greaterEqual, VECTOR16_LOOP); 8694 addq(length, 16); 8695 jcc(Assembler::equal, SAME_TILL_END); 8696 //falling through if less than 16 bytes left 8697 } 8698 8699 bind(VECTOR8_TAIL); 8700 cmpq(length, 8); 8701 jccb(Assembler::less, VECTOR4_TAIL); 8702 bind(VECTOR8_LOOP); 8703 movq(tmp1, Address(obja, result)); 8704 movq(tmp2, Address(objb, result)); 8705 xorq(tmp1, tmp2); 8706 testq(tmp1, tmp1); 8707 jcc(Assembler::notZero, VECTOR8_NOT_EQUAL);//mismatch found 8708 addq(result, 8); 8709 subq(length, 8); 8710 jcc(Assembler::equal, SAME_TILL_END); 8711 //falling through if less than 8 bytes left 8712 8713 bind(VECTOR4_TAIL); 8714 cmpq(length, 4); 8715 jccb(Assembler::less, BYTES_TAIL); 8716 bind(VECTOR4_LOOP); 8717 movl(tmp1, Address(obja, result)); 8718 xorl(tmp1, Address(objb, result)); 8719 testl(tmp1, tmp1); 8720 jcc(Assembler::notZero, VECTOR4_NOT_EQUAL);//mismatch found 8721 addq(result, 4); 8722 subq(length, 4); 8723 jcc(Assembler::equal, SAME_TILL_END); 8724 //falling through if less than 4 bytes left 8725 8726 bind(BYTES_TAIL); 8727 bind(BYTES_LOOP); 8728 load_unsigned_byte(tmp1, Address(obja, result)); 8729 load_unsigned_byte(tmp2, Address(objb, result)); 8730 xorl(tmp1, tmp2); 8731 testl(tmp1, tmp1); 8732 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 8733 decq(length); 8734 jcc(Assembler::zero, SAME_TILL_END); 8735 incq(result); 8736 load_unsigned_byte(tmp1, Address(obja, result)); 8737 load_unsigned_byte(tmp2, Address(objb, result)); 8738 xorl(tmp1, tmp2); 8739 testl(tmp1, tmp1); 8740 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 8741 decq(length); 8742 jcc(Assembler::zero, SAME_TILL_END); 8743 incq(result); 8744 load_unsigned_byte(tmp1, Address(obja, result)); 8745 load_unsigned_byte(tmp2, Address(objb, result)); 8746 xorl(tmp1, tmp2); 8747 testl(tmp1, tmp1); 8748 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found 8749 jmp(SAME_TILL_END); 8750 8751 if (UseAVX >= 2) { 8752 bind(VECTOR32_NOT_EQUAL); 8753 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_256bit); 8754 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_256bit); 8755 vpxor(rymm0, rymm0, rymm2, Assembler::AVX_256bit); 8756 vpmovmskb(tmp1, rymm0); 8757 bsfq(tmp1, tmp1); 8758 addq(result, tmp1); 8759 shrq(result); 8760 jmp(DONE); 8761 } 8762 8763 bind(VECTOR16_NOT_EQUAL); 8764 if (UseAVX >= 2) { 8765 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_128bit); 8766 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_128bit); 8767 pxor(rymm0, rymm2); 8768 } else { 8769 pcmpeqb(rymm2, rymm2); 8770 pxor(rymm0, rymm1); 8771 pcmpeqb(rymm0, rymm1); 8772 pxor(rymm0, rymm2); 8773 } 8774 pmovmskb(tmp1, rymm0); 8775 bsfq(tmp1, tmp1); 8776 addq(result, tmp1); 8777 shrq(result); 8778 jmpb(DONE); 8779 8780 bind(VECTOR8_NOT_EQUAL); 8781 bind(VECTOR4_NOT_EQUAL); 8782 bsfq(tmp1, tmp1); 8783 shrq(tmp1, 3); 8784 addq(result, tmp1); 8785 bind(BYTES_NOT_EQUAL); 8786 shrq(result); 8787 jmpb(DONE); 8788 8789 bind(SAME_TILL_END); 8790 mov64(result, -1); 8791 8792 bind(DONE); 8793 } 8794 8795 //Helper functions for square_to_len() 8796 8797 /** 8798 * Store the squares of x[], right shifted one bit (divided by 2) into z[] 8799 * Preserves x and z and modifies rest of the registers. 8800 */ 8801 void MacroAssembler::square_rshift(Register x, Register xlen, Register z, Register tmp1, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 8802 // Perform square and right shift by 1 8803 // Handle odd xlen case first, then for even xlen do the following 8804 // jlong carry = 0; 8805 // for (int j=0, i=0; j < xlen; j+=2, i+=4) { 8806 // huge_128 product = x[j:j+1] * x[j:j+1]; 8807 // z[i:i+1] = (carry << 63) | (jlong)(product >>> 65); 8808 // z[i+2:i+3] = (jlong)(product >>> 1); 8809 // carry = (jlong)product; 8810 // } 8811 8812 xorq(tmp5, tmp5); // carry 8813 xorq(rdxReg, rdxReg); 8814 xorl(tmp1, tmp1); // index for x 8815 xorl(tmp4, tmp4); // index for z 8816 8817 Label L_first_loop, L_first_loop_exit; 8818 8819 testl(xlen, 1); 8820 jccb(Assembler::zero, L_first_loop); //jump if xlen is even 8821 8822 // Square and right shift by 1 the odd element using 32 bit multiply 8823 movl(raxReg, Address(x, tmp1, Address::times_4, 0)); 8824 imulq(raxReg, raxReg); 8825 shrq(raxReg, 1); 8826 adcq(tmp5, 0); 8827 movq(Address(z, tmp4, Address::times_4, 0), raxReg); 8828 incrementl(tmp1); 8829 addl(tmp4, 2); 8830 8831 // Square and right shift by 1 the rest using 64 bit multiply 8832 bind(L_first_loop); 8833 cmpptr(tmp1, xlen); 8834 jccb(Assembler::equal, L_first_loop_exit); 8835 8836 // Square 8837 movq(raxReg, Address(x, tmp1, Address::times_4, 0)); 8838 rorq(raxReg, 32); // convert big-endian to little-endian 8839 mulq(raxReg); // 64-bit multiply rax * rax -> rdx:rax 8840 8841 // Right shift by 1 and save carry 8842 shrq(tmp5, 1); // rdx:rax:tmp5 = (tmp5:rdx:rax) >>> 1 8843 rcrq(rdxReg, 1); 8844 rcrq(raxReg, 1); 8845 adcq(tmp5, 0); 8846 8847 // Store result in z 8848 movq(Address(z, tmp4, Address::times_4, 0), rdxReg); 8849 movq(Address(z, tmp4, Address::times_4, 8), raxReg); 8850 8851 // Update indices for x and z 8852 addl(tmp1, 2); 8853 addl(tmp4, 4); 8854 jmp(L_first_loop); 8855 8856 bind(L_first_loop_exit); 8857 } 8858 8859 8860 /** 8861 * Perform the following multiply add operation using BMI2 instructions 8862 * carry:sum = sum + op1*op2 + carry 8863 * op2 should be in rdx 8864 * op2 is preserved, all other registers are modified 8865 */ 8866 void MacroAssembler::multiply_add_64_bmi2(Register sum, Register op1, Register op2, Register carry, Register tmp2) { 8867 // assert op2 is rdx 8868 mulxq(tmp2, op1, op1); // op1 * op2 -> tmp2:op1 8869 addq(sum, carry); 8870 adcq(tmp2, 0); 8871 addq(sum, op1); 8872 adcq(tmp2, 0); 8873 movq(carry, tmp2); 8874 } 8875 8876 /** 8877 * Perform the following multiply add operation: 8878 * carry:sum = sum + op1*op2 + carry 8879 * Preserves op1, op2 and modifies rest of registers 8880 */ 8881 void MacroAssembler::multiply_add_64(Register sum, Register op1, Register op2, Register carry, Register rdxReg, Register raxReg) { 8882 // rdx:rax = op1 * op2 8883 movq(raxReg, op2); 8884 mulq(op1); 8885 8886 // rdx:rax = sum + carry + rdx:rax 8887 addq(sum, carry); 8888 adcq(rdxReg, 0); 8889 addq(sum, raxReg); 8890 adcq(rdxReg, 0); 8891 8892 // carry:sum = rdx:sum 8893 movq(carry, rdxReg); 8894 } 8895 8896 /** 8897 * Add 64 bit long carry into z[] with carry propogation. 8898 * Preserves z and carry register values and modifies rest of registers. 8899 * 8900 */ 8901 void MacroAssembler::add_one_64(Register z, Register zlen, Register carry, Register tmp1) { 8902 Label L_fourth_loop, L_fourth_loop_exit; 8903 8904 movl(tmp1, 1); 8905 subl(zlen, 2); 8906 addq(Address(z, zlen, Address::times_4, 0), carry); 8907 8908 bind(L_fourth_loop); 8909 jccb(Assembler::carryClear, L_fourth_loop_exit); 8910 subl(zlen, 2); 8911 jccb(Assembler::negative, L_fourth_loop_exit); 8912 addq(Address(z, zlen, Address::times_4, 0), tmp1); 8913 jmp(L_fourth_loop); 8914 bind(L_fourth_loop_exit); 8915 } 8916 8917 /** 8918 * Shift z[] left by 1 bit. 8919 * Preserves x, len, z and zlen registers and modifies rest of the registers. 8920 * 8921 */ 8922 void MacroAssembler::lshift_by_1(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4) { 8923 8924 Label L_fifth_loop, L_fifth_loop_exit; 8925 8926 // Fifth loop 8927 // Perform primitiveLeftShift(z, zlen, 1) 8928 8929 const Register prev_carry = tmp1; 8930 const Register new_carry = tmp4; 8931 const Register value = tmp2; 8932 const Register zidx = tmp3; 8933 8934 // int zidx, carry; 8935 // long value; 8936 // carry = 0; 8937 // for (zidx = zlen-2; zidx >=0; zidx -= 2) { 8938 // (carry:value) = (z[i] << 1) | carry ; 8939 // z[i] = value; 8940 // } 8941 8942 movl(zidx, zlen); 8943 xorl(prev_carry, prev_carry); // clear carry flag and prev_carry register 8944 8945 bind(L_fifth_loop); 8946 decl(zidx); // Use decl to preserve carry flag 8947 decl(zidx); 8948 jccb(Assembler::negative, L_fifth_loop_exit); 8949 8950 if (UseBMI2Instructions) { 8951 movq(value, Address(z, zidx, Address::times_4, 0)); 8952 rclq(value, 1); 8953 rorxq(value, value, 32); 8954 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form 8955 } 8956 else { 8957 // clear new_carry 8958 xorl(new_carry, new_carry); 8959 8960 // Shift z[i] by 1, or in previous carry and save new carry 8961 movq(value, Address(z, zidx, Address::times_4, 0)); 8962 shlq(value, 1); 8963 adcl(new_carry, 0); 8964 8965 orq(value, prev_carry); 8966 rorq(value, 0x20); 8967 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form 8968 8969 // Set previous carry = new carry 8970 movl(prev_carry, new_carry); 8971 } 8972 jmp(L_fifth_loop); 8973 8974 bind(L_fifth_loop_exit); 8975 } 8976 8977 8978 /** 8979 * Code for BigInteger::squareToLen() intrinsic 8980 * 8981 * rdi: x 8982 * rsi: len 8983 * r8: z 8984 * rcx: zlen 8985 * r12: tmp1 8986 * r13: tmp2 8987 * r14: tmp3 8988 * r15: tmp4 8989 * rbx: tmp5 8990 * 8991 */ 8992 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) { 8993 8994 Label L_second_loop, L_second_loop_exit, L_third_loop, L_third_loop_exit, L_last_x, L_multiply; 8995 push(tmp1); 8996 push(tmp2); 8997 push(tmp3); 8998 push(tmp4); 8999 push(tmp5); 9000 9001 // First loop 9002 // Store the squares, right shifted one bit (i.e., divided by 2). 9003 square_rshift(x, len, z, tmp1, tmp3, tmp4, tmp5, rdxReg, raxReg); 9004 9005 // Add in off-diagonal sums. 9006 // 9007 // Second, third (nested) and fourth loops. 9008 // zlen +=2; 9009 // for (int xidx=len-2,zidx=zlen-4; xidx > 0; xidx-=2,zidx-=4) { 9010 // carry = 0; 9011 // long op2 = x[xidx:xidx+1]; 9012 // for (int j=xidx-2,k=zidx; j >= 0; j-=2) { 9013 // k -= 2; 9014 // long op1 = x[j:j+1]; 9015 // long sum = z[k:k+1]; 9016 // carry:sum = multiply_add_64(sum, op1, op2, carry, tmp_regs); 9017 // z[k:k+1] = sum; 9018 // } 9019 // add_one_64(z, k, carry, tmp_regs); 9020 // } 9021 9022 const Register carry = tmp5; 9023 const Register sum = tmp3; 9024 const Register op1 = tmp4; 9025 Register op2 = tmp2; 9026 9027 push(zlen); 9028 push(len); 9029 addl(zlen,2); 9030 bind(L_second_loop); 9031 xorq(carry, carry); 9032 subl(zlen, 4); 9033 subl(len, 2); 9034 push(zlen); 9035 push(len); 9036 cmpl(len, 0); 9037 jccb(Assembler::lessEqual, L_second_loop_exit); 9038 9039 // Multiply an array by one 64 bit long. 9040 if (UseBMI2Instructions) { 9041 op2 = rdxReg; 9042 movq(op2, Address(x, len, Address::times_4, 0)); 9043 rorxq(op2, op2, 32); 9044 } 9045 else { 9046 movq(op2, Address(x, len, Address::times_4, 0)); 9047 rorq(op2, 32); 9048 } 9049 9050 bind(L_third_loop); 9051 decrementl(len); 9052 jccb(Assembler::negative, L_third_loop_exit); 9053 decrementl(len); 9054 jccb(Assembler::negative, L_last_x); 9055 9056 movq(op1, Address(x, len, Address::times_4, 0)); 9057 rorq(op1, 32); 9058 9059 bind(L_multiply); 9060 subl(zlen, 2); 9061 movq(sum, Address(z, zlen, Address::times_4, 0)); 9062 9063 // Multiply 64 bit by 64 bit and add 64 bits lower half and upper 64 bits as carry. 9064 if (UseBMI2Instructions) { 9065 multiply_add_64_bmi2(sum, op1, op2, carry, tmp2); 9066 } 9067 else { 9068 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9069 } 9070 9071 movq(Address(z, zlen, Address::times_4, 0), sum); 9072 9073 jmp(L_third_loop); 9074 bind(L_third_loop_exit); 9075 9076 // Fourth loop 9077 // Add 64 bit long carry into z with carry propogation. 9078 // Uses offsetted zlen. 9079 add_one_64(z, zlen, carry, tmp1); 9080 9081 pop(len); 9082 pop(zlen); 9083 jmp(L_second_loop); 9084 9085 // Next infrequent code is moved outside loops. 9086 bind(L_last_x); 9087 movl(op1, Address(x, 0)); 9088 jmp(L_multiply); 9089 9090 bind(L_second_loop_exit); 9091 pop(len); 9092 pop(zlen); 9093 pop(len); 9094 pop(zlen); 9095 9096 // Fifth loop 9097 // Shift z left 1 bit. 9098 lshift_by_1(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4); 9099 9100 // z[zlen-1] |= x[len-1] & 1; 9101 movl(tmp3, Address(x, len, Address::times_4, -4)); 9102 andl(tmp3, 1); 9103 orl(Address(z, zlen, Address::times_4, -4), tmp3); 9104 9105 pop(tmp5); 9106 pop(tmp4); 9107 pop(tmp3); 9108 pop(tmp2); 9109 pop(tmp1); 9110 } 9111 9112 /** 9113 * Helper function for mul_add() 9114 * Multiply the in[] by int k and add to out[] starting at offset offs using 9115 * 128 bit by 32 bit multiply and return the carry in tmp5. 9116 * Only quad int aligned length of in[] is operated on in this function. 9117 * k is in rdxReg for BMI2Instructions, for others it is in tmp2. 9118 * This function preserves out, in and k registers. 9119 * len and offset point to the appropriate index in "in" & "out" correspondingly 9120 * tmp5 has the carry. 9121 * other registers are temporary and are modified. 9122 * 9123 */ 9124 void MacroAssembler::mul_add_128_x_32_loop(Register out, Register in, 9125 Register offset, Register len, Register tmp1, Register tmp2, Register tmp3, 9126 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 9127 9128 Label L_first_loop, L_first_loop_exit; 9129 9130 movl(tmp1, len); 9131 shrl(tmp1, 2); 9132 9133 bind(L_first_loop); 9134 subl(tmp1, 1); 9135 jccb(Assembler::negative, L_first_loop_exit); 9136 9137 subl(len, 4); 9138 subl(offset, 4); 9139 9140 Register op2 = tmp2; 9141 const Register sum = tmp3; 9142 const Register op1 = tmp4; 9143 const Register carry = tmp5; 9144 9145 if (UseBMI2Instructions) { 9146 op2 = rdxReg; 9147 } 9148 9149 movq(op1, Address(in, len, Address::times_4, 8)); 9150 rorq(op1, 32); 9151 movq(sum, Address(out, offset, Address::times_4, 8)); 9152 rorq(sum, 32); 9153 if (UseBMI2Instructions) { 9154 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 9155 } 9156 else { 9157 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9158 } 9159 // Store back in big endian from little endian 9160 rorq(sum, 0x20); 9161 movq(Address(out, offset, Address::times_4, 8), sum); 9162 9163 movq(op1, Address(in, len, Address::times_4, 0)); 9164 rorq(op1, 32); 9165 movq(sum, Address(out, offset, Address::times_4, 0)); 9166 rorq(sum, 32); 9167 if (UseBMI2Instructions) { 9168 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 9169 } 9170 else { 9171 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9172 } 9173 // Store back in big endian from little endian 9174 rorq(sum, 0x20); 9175 movq(Address(out, offset, Address::times_4, 0), sum); 9176 9177 jmp(L_first_loop); 9178 bind(L_first_loop_exit); 9179 } 9180 9181 /** 9182 * Code for BigInteger::mulAdd() intrinsic 9183 * 9184 * rdi: out 9185 * rsi: in 9186 * r11: offs (out.length - offset) 9187 * rcx: len 9188 * r8: k 9189 * r12: tmp1 9190 * r13: tmp2 9191 * r14: tmp3 9192 * r15: tmp4 9193 * rbx: tmp5 9194 * Multiply the in[] by word k and add to out[], return the carry in rax 9195 */ 9196 void MacroAssembler::mul_add(Register out, Register in, Register offs, 9197 Register len, Register k, Register tmp1, Register tmp2, Register tmp3, 9198 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) { 9199 9200 Label L_carry, L_last_in, L_done; 9201 9202 // carry = 0; 9203 // for (int j=len-1; j >= 0; j--) { 9204 // long product = (in[j] & LONG_MASK) * kLong + 9205 // (out[offs] & LONG_MASK) + carry; 9206 // out[offs--] = (int)product; 9207 // carry = product >>> 32; 9208 // } 9209 // 9210 push(tmp1); 9211 push(tmp2); 9212 push(tmp3); 9213 push(tmp4); 9214 push(tmp5); 9215 9216 Register op2 = tmp2; 9217 const Register sum = tmp3; 9218 const Register op1 = tmp4; 9219 const Register carry = tmp5; 9220 9221 if (UseBMI2Instructions) { 9222 op2 = rdxReg; 9223 movl(op2, k); 9224 } 9225 else { 9226 movl(op2, k); 9227 } 9228 9229 xorq(carry, carry); 9230 9231 //First loop 9232 9233 //Multiply in[] by k in a 4 way unrolled loop using 128 bit by 32 bit multiply 9234 //The carry is in tmp5 9235 mul_add_128_x_32_loop(out, in, offs, len, tmp1, tmp2, tmp3, tmp4, tmp5, rdxReg, raxReg); 9236 9237 //Multiply the trailing in[] entry using 64 bit by 32 bit, if any 9238 decrementl(len); 9239 jccb(Assembler::negative, L_carry); 9240 decrementl(len); 9241 jccb(Assembler::negative, L_last_in); 9242 9243 movq(op1, Address(in, len, Address::times_4, 0)); 9244 rorq(op1, 32); 9245 9246 subl(offs, 2); 9247 movq(sum, Address(out, offs, Address::times_4, 0)); 9248 rorq(sum, 32); 9249 9250 if (UseBMI2Instructions) { 9251 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg); 9252 } 9253 else { 9254 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg); 9255 } 9256 9257 // Store back in big endian from little endian 9258 rorq(sum, 0x20); 9259 movq(Address(out, offs, Address::times_4, 0), sum); 9260 9261 testl(len, len); 9262 jccb(Assembler::zero, L_carry); 9263 9264 //Multiply the last in[] entry, if any 9265 bind(L_last_in); 9266 movl(op1, Address(in, 0)); 9267 movl(sum, Address(out, offs, Address::times_4, -4)); 9268 9269 movl(raxReg, k); 9270 mull(op1); //tmp4 * eax -> edx:eax 9271 addl(sum, carry); 9272 adcl(rdxReg, 0); 9273 addl(sum, raxReg); 9274 adcl(rdxReg, 0); 9275 movl(carry, rdxReg); 9276 9277 movl(Address(out, offs, Address::times_4, -4), sum); 9278 9279 bind(L_carry); 9280 //return tmp5/carry as carry in rax 9281 movl(rax, carry); 9282 9283 bind(L_done); 9284 pop(tmp5); 9285 pop(tmp4); 9286 pop(tmp3); 9287 pop(tmp2); 9288 pop(tmp1); 9289 } 9290 #endif 9291 9292 /** 9293 * Emits code to update CRC-32 with a byte value according to constants in table 9294 * 9295 * @param [in,out]crc Register containing the crc. 9296 * @param [in]val Register containing the byte to fold into the CRC. 9297 * @param [in]table Register containing the table of crc constants. 9298 * 9299 * uint32_t crc; 9300 * val = crc_table[(val ^ crc) & 0xFF]; 9301 * crc = val ^ (crc >> 8); 9302 * 9303 */ 9304 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) { 9305 xorl(val, crc); 9306 andl(val, 0xFF); 9307 shrl(crc, 8); // unsigned shift 9308 xorl(crc, Address(table, val, Address::times_4, 0)); 9309 } 9310 9311 /** 9312 * Fold four 128-bit data chunks 9313 */ 9314 void MacroAssembler::fold_128bit_crc32_avx512(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) { 9315 evpclmulhdq(xtmp, xK, xcrc, Assembler::AVX_512bit); // [123:64] 9316 evpclmulldq(xcrc, xK, xcrc, Assembler::AVX_512bit); // [63:0] 9317 evpxorq(xcrc, xcrc, Address(buf, offset), Assembler::AVX_512bit /* vector_len */); 9318 evpxorq(xcrc, xcrc, xtmp, Assembler::AVX_512bit /* vector_len */); 9319 } 9320 9321 /** 9322 * Fold 128-bit data chunk 9323 */ 9324 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) { 9325 if (UseAVX > 0) { 9326 vpclmulhdq(xtmp, xK, xcrc); // [123:64] 9327 vpclmulldq(xcrc, xK, xcrc); // [63:0] 9328 vpxor(xcrc, xcrc, Address(buf, offset), 0 /* vector_len */); 9329 pxor(xcrc, xtmp); 9330 } else { 9331 movdqa(xtmp, xcrc); 9332 pclmulhdq(xtmp, xK); // [123:64] 9333 pclmulldq(xcrc, xK); // [63:0] 9334 pxor(xcrc, xtmp); 9335 movdqu(xtmp, Address(buf, offset)); 9336 pxor(xcrc, xtmp); 9337 } 9338 } 9339 9340 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) { 9341 if (UseAVX > 0) { 9342 vpclmulhdq(xtmp, xK, xcrc); 9343 vpclmulldq(xcrc, xK, xcrc); 9344 pxor(xcrc, xbuf); 9345 pxor(xcrc, xtmp); 9346 } else { 9347 movdqa(xtmp, xcrc); 9348 pclmulhdq(xtmp, xK); 9349 pclmulldq(xcrc, xK); 9350 pxor(xcrc, xbuf); 9351 pxor(xcrc, xtmp); 9352 } 9353 } 9354 9355 /** 9356 * 8-bit folds to compute 32-bit CRC 9357 * 9358 * uint64_t xcrc; 9359 * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8); 9360 */ 9361 void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) { 9362 movdl(tmp, xcrc); 9363 andl(tmp, 0xFF); 9364 movdl(xtmp, Address(table, tmp, Address::times_4, 0)); 9365 psrldq(xcrc, 1); // unsigned shift one byte 9366 pxor(xcrc, xtmp); 9367 } 9368 9369 /** 9370 * uint32_t crc; 9371 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8); 9372 */ 9373 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) { 9374 movl(tmp, crc); 9375 andl(tmp, 0xFF); 9376 shrl(crc, 8); 9377 xorl(crc, Address(table, tmp, Address::times_4, 0)); 9378 } 9379 9380 /** 9381 * @param crc register containing existing CRC (32-bit) 9382 * @param buf register pointing to input byte buffer (byte*) 9383 * @param len register containing number of bytes 9384 * @param table register that will contain address of CRC table 9385 * @param tmp scratch register 9386 */ 9387 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) { 9388 assert_different_registers(crc, buf, len, table, tmp, rax); 9389 9390 Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned; 9391 Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop; 9392 9393 // For EVEX with VL and BW, provide a standard mask, VL = 128 will guide the merge 9394 // context for the registers used, where all instructions below are using 128-bit mode 9395 // On EVEX without VL and BW, these instructions will all be AVX. 9396 lea(table, ExternalAddress(StubRoutines::crc_table_addr())); 9397 notl(crc); // ~crc 9398 cmpl(len, 16); 9399 jcc(Assembler::less, L_tail); 9400 9401 // Align buffer to 16 bytes 9402 movl(tmp, buf); 9403 andl(tmp, 0xF); 9404 jccb(Assembler::zero, L_aligned); 9405 subl(tmp, 16); 9406 addl(len, tmp); 9407 9408 align(4); 9409 BIND(L_align_loop); 9410 movsbl(rax, Address(buf, 0)); // load byte with sign extension 9411 update_byte_crc32(crc, rax, table); 9412 increment(buf); 9413 incrementl(tmp); 9414 jccb(Assembler::less, L_align_loop); 9415 9416 BIND(L_aligned); 9417 movl(tmp, len); // save 9418 shrl(len, 4); 9419 jcc(Assembler::zero, L_tail_restore); 9420 9421 // Fold total 512 bits of polynomial on each iteration 9422 if (VM_Version::supports_vpclmulqdq()) { 9423 Label Parallel_loop, L_No_Parallel; 9424 9425 cmpl(len, 8); 9426 jccb(Assembler::less, L_No_Parallel); 9427 9428 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32)); 9429 evmovdquq(xmm1, Address(buf, 0), Assembler::AVX_512bit); 9430 movdl(xmm5, crc); 9431 evpxorq(xmm1, xmm1, xmm5, Assembler::AVX_512bit); 9432 addptr(buf, 64); 9433 subl(len, 7); 9434 evshufi64x2(xmm0, xmm0, xmm0, 0x00, Assembler::AVX_512bit); //propagate the mask from 128 bits to 512 bits 9435 9436 BIND(Parallel_loop); 9437 fold_128bit_crc32_avx512(xmm1, xmm0, xmm5, buf, 0); 9438 addptr(buf, 64); 9439 subl(len, 4); 9440 jcc(Assembler::greater, Parallel_loop); 9441 9442 vextracti64x2(xmm2, xmm1, 0x01); 9443 vextracti64x2(xmm3, xmm1, 0x02); 9444 vextracti64x2(xmm4, xmm1, 0x03); 9445 jmp(L_fold_512b); 9446 9447 BIND(L_No_Parallel); 9448 } 9449 // Fold crc into first bytes of vector 9450 movdqa(xmm1, Address(buf, 0)); 9451 movdl(rax, xmm1); 9452 xorl(crc, rax); 9453 if (VM_Version::supports_sse4_1()) { 9454 pinsrd(xmm1, crc, 0); 9455 } else { 9456 pinsrw(xmm1, crc, 0); 9457 shrl(crc, 16); 9458 pinsrw(xmm1, crc, 1); 9459 } 9460 addptr(buf, 16); 9461 subl(len, 4); // len > 0 9462 jcc(Assembler::less, L_fold_tail); 9463 9464 movdqa(xmm2, Address(buf, 0)); 9465 movdqa(xmm3, Address(buf, 16)); 9466 movdqa(xmm4, Address(buf, 32)); 9467 addptr(buf, 48); 9468 subl(len, 3); 9469 jcc(Assembler::lessEqual, L_fold_512b); 9470 9471 // Fold total 512 bits of polynomial on each iteration, 9472 // 128 bits per each of 4 parallel streams. 9473 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32)); 9474 9475 align(32); 9476 BIND(L_fold_512b_loop); 9477 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0); 9478 fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16); 9479 fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32); 9480 fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48); 9481 addptr(buf, 64); 9482 subl(len, 4); 9483 jcc(Assembler::greater, L_fold_512b_loop); 9484 9485 // Fold 512 bits to 128 bits. 9486 BIND(L_fold_512b); 9487 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16)); 9488 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2); 9489 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3); 9490 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4); 9491 9492 // Fold the rest of 128 bits data chunks 9493 BIND(L_fold_tail); 9494 addl(len, 3); 9495 jccb(Assembler::lessEqual, L_fold_128b); 9496 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16)); 9497 9498 BIND(L_fold_tail_loop); 9499 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0); 9500 addptr(buf, 16); 9501 decrementl(len); 9502 jccb(Assembler::greater, L_fold_tail_loop); 9503 9504 // Fold 128 bits in xmm1 down into 32 bits in crc register. 9505 BIND(L_fold_128b); 9506 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr())); 9507 if (UseAVX > 0) { 9508 vpclmulqdq(xmm2, xmm0, xmm1, 0x1); 9509 vpand(xmm3, xmm0, xmm2, 0 /* vector_len */); 9510 vpclmulqdq(xmm0, xmm0, xmm3, 0x1); 9511 } else { 9512 movdqa(xmm2, xmm0); 9513 pclmulqdq(xmm2, xmm1, 0x1); 9514 movdqa(xmm3, xmm0); 9515 pand(xmm3, xmm2); 9516 pclmulqdq(xmm0, xmm3, 0x1); 9517 } 9518 psrldq(xmm1, 8); 9519 psrldq(xmm2, 4); 9520 pxor(xmm0, xmm1); 9521 pxor(xmm0, xmm2); 9522 9523 // 8 8-bit folds to compute 32-bit CRC. 9524 for (int j = 0; j < 4; j++) { 9525 fold_8bit_crc32(xmm0, table, xmm1, rax); 9526 } 9527 movdl(crc, xmm0); // mov 32 bits to general register 9528 for (int j = 0; j < 4; j++) { 9529 fold_8bit_crc32(crc, table, rax); 9530 } 9531 9532 BIND(L_tail_restore); 9533 movl(len, tmp); // restore 9534 BIND(L_tail); 9535 andl(len, 0xf); 9536 jccb(Assembler::zero, L_exit); 9537 9538 // Fold the rest of bytes 9539 align(4); 9540 BIND(L_tail_loop); 9541 movsbl(rax, Address(buf, 0)); // load byte with sign extension 9542 update_byte_crc32(crc, rax, table); 9543 increment(buf); 9544 decrementl(len); 9545 jccb(Assembler::greater, L_tail_loop); 9546 9547 BIND(L_exit); 9548 notl(crc); // ~c 9549 } 9550 9551 #ifdef _LP64 9552 // S. Gueron / Information Processing Letters 112 (2012) 184 9553 // Algorithm 4: Computing carry-less multiplication using a precomputed lookup table. 9554 // Input: A 32 bit value B = [byte3, byte2, byte1, byte0]. 9555 // Output: the 64-bit carry-less product of B * CONST 9556 void MacroAssembler::crc32c_ipl_alg4(Register in, uint32_t n, 9557 Register tmp1, Register tmp2, Register tmp3) { 9558 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr())); 9559 if (n > 0) { 9560 addq(tmp3, n * 256 * 8); 9561 } 9562 // Q1 = TABLEExt[n][B & 0xFF]; 9563 movl(tmp1, in); 9564 andl(tmp1, 0x000000FF); 9565 shll(tmp1, 3); 9566 addq(tmp1, tmp3); 9567 movq(tmp1, Address(tmp1, 0)); 9568 9569 // Q2 = TABLEExt[n][B >> 8 & 0xFF]; 9570 movl(tmp2, in); 9571 shrl(tmp2, 8); 9572 andl(tmp2, 0x000000FF); 9573 shll(tmp2, 3); 9574 addq(tmp2, tmp3); 9575 movq(tmp2, Address(tmp2, 0)); 9576 9577 shlq(tmp2, 8); 9578 xorq(tmp1, tmp2); 9579 9580 // Q3 = TABLEExt[n][B >> 16 & 0xFF]; 9581 movl(tmp2, in); 9582 shrl(tmp2, 16); 9583 andl(tmp2, 0x000000FF); 9584 shll(tmp2, 3); 9585 addq(tmp2, tmp3); 9586 movq(tmp2, Address(tmp2, 0)); 9587 9588 shlq(tmp2, 16); 9589 xorq(tmp1, tmp2); 9590 9591 // Q4 = TABLEExt[n][B >> 24 & 0xFF]; 9592 shrl(in, 24); 9593 andl(in, 0x000000FF); 9594 shll(in, 3); 9595 addq(in, tmp3); 9596 movq(in, Address(in, 0)); 9597 9598 shlq(in, 24); 9599 xorq(in, tmp1); 9600 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24; 9601 } 9602 9603 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1, 9604 Register in_out, 9605 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported, 9606 XMMRegister w_xtmp2, 9607 Register tmp1, 9608 Register n_tmp2, Register n_tmp3) { 9609 if (is_pclmulqdq_supported) { 9610 movdl(w_xtmp1, in_out); // modified blindly 9611 9612 movl(tmp1, const_or_pre_comp_const_index); 9613 movdl(w_xtmp2, tmp1); 9614 pclmulqdq(w_xtmp1, w_xtmp2, 0); 9615 9616 movdq(in_out, w_xtmp1); 9617 } else { 9618 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3); 9619 } 9620 } 9621 9622 // Recombination Alternative 2: No bit-reflections 9623 // T1 = (CRC_A * U1) << 1 9624 // T2 = (CRC_B * U2) << 1 9625 // C1 = T1 >> 32 9626 // C2 = T2 >> 32 9627 // T1 = T1 & 0xFFFFFFFF 9628 // T2 = T2 & 0xFFFFFFFF 9629 // T1 = CRC32(0, T1) 9630 // T2 = CRC32(0, T2) 9631 // C1 = C1 ^ T1 9632 // C2 = C2 ^ T2 9633 // CRC = C1 ^ C2 ^ CRC_C 9634 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, 9635 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 9636 Register tmp1, Register tmp2, 9637 Register n_tmp3) { 9638 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 9639 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 9640 shlq(in_out, 1); 9641 movl(tmp1, in_out); 9642 shrq(in_out, 32); 9643 xorl(tmp2, tmp2); 9644 crc32(tmp2, tmp1, 4); 9645 xorl(in_out, tmp2); // we don't care about upper 32 bit contents here 9646 shlq(in1, 1); 9647 movl(tmp1, in1); 9648 shrq(in1, 32); 9649 xorl(tmp2, tmp2); 9650 crc32(tmp2, tmp1, 4); 9651 xorl(in1, tmp2); 9652 xorl(in_out, in1); 9653 xorl(in_out, in2); 9654 } 9655 9656 // Set N to predefined value 9657 // Subtract from a lenght of a buffer 9658 // execute in a loop: 9659 // CRC_A = 0xFFFFFFFF, CRC_B = 0, CRC_C = 0 9660 // for i = 1 to N do 9661 // CRC_A = CRC32(CRC_A, A[i]) 9662 // CRC_B = CRC32(CRC_B, B[i]) 9663 // CRC_C = CRC32(CRC_C, C[i]) 9664 // end for 9665 // Recombine 9666 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, 9667 Register in_out1, Register in_out2, Register in_out3, 9668 Register tmp1, Register tmp2, Register tmp3, 9669 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 9670 Register tmp4, Register tmp5, 9671 Register n_tmp6) { 9672 Label L_processPartitions; 9673 Label L_processPartition; 9674 Label L_exit; 9675 9676 bind(L_processPartitions); 9677 cmpl(in_out1, 3 * size); 9678 jcc(Assembler::less, L_exit); 9679 xorl(tmp1, tmp1); 9680 xorl(tmp2, tmp2); 9681 movq(tmp3, in_out2); 9682 addq(tmp3, size); 9683 9684 bind(L_processPartition); 9685 crc32(in_out3, Address(in_out2, 0), 8); 9686 crc32(tmp1, Address(in_out2, size), 8); 9687 crc32(tmp2, Address(in_out2, size * 2), 8); 9688 addq(in_out2, 8); 9689 cmpq(in_out2, tmp3); 9690 jcc(Assembler::less, L_processPartition); 9691 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2, 9692 w_xtmp1, w_xtmp2, w_xtmp3, 9693 tmp4, tmp5, 9694 n_tmp6); 9695 addq(in_out2, 2 * size); 9696 subl(in_out1, 3 * size); 9697 jmp(L_processPartitions); 9698 9699 bind(L_exit); 9700 } 9701 #else 9702 void MacroAssembler::crc32c_ipl_alg4(Register in_out, uint32_t n, 9703 Register tmp1, Register tmp2, Register tmp3, 9704 XMMRegister xtmp1, XMMRegister xtmp2) { 9705 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr())); 9706 if (n > 0) { 9707 addl(tmp3, n * 256 * 8); 9708 } 9709 // Q1 = TABLEExt[n][B & 0xFF]; 9710 movl(tmp1, in_out); 9711 andl(tmp1, 0x000000FF); 9712 shll(tmp1, 3); 9713 addl(tmp1, tmp3); 9714 movq(xtmp1, Address(tmp1, 0)); 9715 9716 // Q2 = TABLEExt[n][B >> 8 & 0xFF]; 9717 movl(tmp2, in_out); 9718 shrl(tmp2, 8); 9719 andl(tmp2, 0x000000FF); 9720 shll(tmp2, 3); 9721 addl(tmp2, tmp3); 9722 movq(xtmp2, Address(tmp2, 0)); 9723 9724 psllq(xtmp2, 8); 9725 pxor(xtmp1, xtmp2); 9726 9727 // Q3 = TABLEExt[n][B >> 16 & 0xFF]; 9728 movl(tmp2, in_out); 9729 shrl(tmp2, 16); 9730 andl(tmp2, 0x000000FF); 9731 shll(tmp2, 3); 9732 addl(tmp2, tmp3); 9733 movq(xtmp2, Address(tmp2, 0)); 9734 9735 psllq(xtmp2, 16); 9736 pxor(xtmp1, xtmp2); 9737 9738 // Q4 = TABLEExt[n][B >> 24 & 0xFF]; 9739 shrl(in_out, 24); 9740 andl(in_out, 0x000000FF); 9741 shll(in_out, 3); 9742 addl(in_out, tmp3); 9743 movq(xtmp2, Address(in_out, 0)); 9744 9745 psllq(xtmp2, 24); 9746 pxor(xtmp1, xtmp2); // Result in CXMM 9747 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24; 9748 } 9749 9750 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1, 9751 Register in_out, 9752 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported, 9753 XMMRegister w_xtmp2, 9754 Register tmp1, 9755 Register n_tmp2, Register n_tmp3) { 9756 if (is_pclmulqdq_supported) { 9757 movdl(w_xtmp1, in_out); 9758 9759 movl(tmp1, const_or_pre_comp_const_index); 9760 movdl(w_xtmp2, tmp1); 9761 pclmulqdq(w_xtmp1, w_xtmp2, 0); 9762 // Keep result in XMM since GPR is 32 bit in length 9763 } else { 9764 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3, w_xtmp1, w_xtmp2); 9765 } 9766 } 9767 9768 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, 9769 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 9770 Register tmp1, Register tmp2, 9771 Register n_tmp3) { 9772 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 9773 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3); 9774 9775 psllq(w_xtmp1, 1); 9776 movdl(tmp1, w_xtmp1); 9777 psrlq(w_xtmp1, 32); 9778 movdl(in_out, w_xtmp1); 9779 9780 xorl(tmp2, tmp2); 9781 crc32(tmp2, tmp1, 4); 9782 xorl(in_out, tmp2); 9783 9784 psllq(w_xtmp2, 1); 9785 movdl(tmp1, w_xtmp2); 9786 psrlq(w_xtmp2, 32); 9787 movdl(in1, w_xtmp2); 9788 9789 xorl(tmp2, tmp2); 9790 crc32(tmp2, tmp1, 4); 9791 xorl(in1, tmp2); 9792 xorl(in_out, in1); 9793 xorl(in_out, in2); 9794 } 9795 9796 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, 9797 Register in_out1, Register in_out2, Register in_out3, 9798 Register tmp1, Register tmp2, Register tmp3, 9799 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 9800 Register tmp4, Register tmp5, 9801 Register n_tmp6) { 9802 Label L_processPartitions; 9803 Label L_processPartition; 9804 Label L_exit; 9805 9806 bind(L_processPartitions); 9807 cmpl(in_out1, 3 * size); 9808 jcc(Assembler::less, L_exit); 9809 xorl(tmp1, tmp1); 9810 xorl(tmp2, tmp2); 9811 movl(tmp3, in_out2); 9812 addl(tmp3, size); 9813 9814 bind(L_processPartition); 9815 crc32(in_out3, Address(in_out2, 0), 4); 9816 crc32(tmp1, Address(in_out2, size), 4); 9817 crc32(tmp2, Address(in_out2, size*2), 4); 9818 crc32(in_out3, Address(in_out2, 0+4), 4); 9819 crc32(tmp1, Address(in_out2, size+4), 4); 9820 crc32(tmp2, Address(in_out2, size*2+4), 4); 9821 addl(in_out2, 8); 9822 cmpl(in_out2, tmp3); 9823 jcc(Assembler::less, L_processPartition); 9824 9825 push(tmp3); 9826 push(in_out1); 9827 push(in_out2); 9828 tmp4 = tmp3; 9829 tmp5 = in_out1; 9830 n_tmp6 = in_out2; 9831 9832 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2, 9833 w_xtmp1, w_xtmp2, w_xtmp3, 9834 tmp4, tmp5, 9835 n_tmp6); 9836 9837 pop(in_out2); 9838 pop(in_out1); 9839 pop(tmp3); 9840 9841 addl(in_out2, 2 * size); 9842 subl(in_out1, 3 * size); 9843 jmp(L_processPartitions); 9844 9845 bind(L_exit); 9846 } 9847 #endif //LP64 9848 9849 #ifdef _LP64 9850 // Algorithm 2: Pipelined usage of the CRC32 instruction. 9851 // Input: A buffer I of L bytes. 9852 // Output: the CRC32C value of the buffer. 9853 // Notations: 9854 // Write L = 24N + r, with N = floor (L/24). 9855 // r = L mod 24 (0 <= r < 24). 9856 // Consider I as the concatenation of A|B|C|R, where A, B, C, each, 9857 // N quadwords, and R consists of r bytes. 9858 // A[j] = I [8j+7:8j], j= 0, 1, ..., N-1 9859 // B[j] = I [N + 8j+7:N + 8j], j= 0, 1, ..., N-1 9860 // C[j] = I [2N + 8j+7:2N + 8j], j= 0, 1, ..., N-1 9861 // if r > 0 R[j] = I [3N +j], j= 0, 1, ...,r-1 9862 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2, 9863 Register tmp1, Register tmp2, Register tmp3, 9864 Register tmp4, Register tmp5, Register tmp6, 9865 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 9866 bool is_pclmulqdq_supported) { 9867 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS]; 9868 Label L_wordByWord; 9869 Label L_byteByByteProlog; 9870 Label L_byteByByte; 9871 Label L_exit; 9872 9873 if (is_pclmulqdq_supported ) { 9874 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr; 9875 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr+1); 9876 9877 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2); 9878 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3); 9879 9880 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4); 9881 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5); 9882 assert((CRC32C_NUM_PRECOMPUTED_CONSTANTS - 1 ) == 5, "Checking whether you declared all of the constants based on the number of \"chunks\""); 9883 } else { 9884 const_or_pre_comp_const_index[0] = 1; 9885 const_or_pre_comp_const_index[1] = 0; 9886 9887 const_or_pre_comp_const_index[2] = 3; 9888 const_or_pre_comp_const_index[3] = 2; 9889 9890 const_or_pre_comp_const_index[4] = 5; 9891 const_or_pre_comp_const_index[5] = 4; 9892 } 9893 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported, 9894 in2, in1, in_out, 9895 tmp1, tmp2, tmp3, 9896 w_xtmp1, w_xtmp2, w_xtmp3, 9897 tmp4, tmp5, 9898 tmp6); 9899 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported, 9900 in2, in1, in_out, 9901 tmp1, tmp2, tmp3, 9902 w_xtmp1, w_xtmp2, w_xtmp3, 9903 tmp4, tmp5, 9904 tmp6); 9905 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported, 9906 in2, in1, in_out, 9907 tmp1, tmp2, tmp3, 9908 w_xtmp1, w_xtmp2, w_xtmp3, 9909 tmp4, tmp5, 9910 tmp6); 9911 movl(tmp1, in2); 9912 andl(tmp1, 0x00000007); 9913 negl(tmp1); 9914 addl(tmp1, in2); 9915 addq(tmp1, in1); 9916 9917 BIND(L_wordByWord); 9918 cmpq(in1, tmp1); 9919 jcc(Assembler::greaterEqual, L_byteByByteProlog); 9920 crc32(in_out, Address(in1, 0), 4); 9921 addq(in1, 4); 9922 jmp(L_wordByWord); 9923 9924 BIND(L_byteByByteProlog); 9925 andl(in2, 0x00000007); 9926 movl(tmp2, 1); 9927 9928 BIND(L_byteByByte); 9929 cmpl(tmp2, in2); 9930 jccb(Assembler::greater, L_exit); 9931 crc32(in_out, Address(in1, 0), 1); 9932 incq(in1); 9933 incl(tmp2); 9934 jmp(L_byteByByte); 9935 9936 BIND(L_exit); 9937 } 9938 #else 9939 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2, 9940 Register tmp1, Register tmp2, Register tmp3, 9941 Register tmp4, Register tmp5, Register tmp6, 9942 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3, 9943 bool is_pclmulqdq_supported) { 9944 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS]; 9945 Label L_wordByWord; 9946 Label L_byteByByteProlog; 9947 Label L_byteByByte; 9948 Label L_exit; 9949 9950 if (is_pclmulqdq_supported) { 9951 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr; 9952 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 1); 9953 9954 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2); 9955 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3); 9956 9957 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4); 9958 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5); 9959 } else { 9960 const_or_pre_comp_const_index[0] = 1; 9961 const_or_pre_comp_const_index[1] = 0; 9962 9963 const_or_pre_comp_const_index[2] = 3; 9964 const_or_pre_comp_const_index[3] = 2; 9965 9966 const_or_pre_comp_const_index[4] = 5; 9967 const_or_pre_comp_const_index[5] = 4; 9968 } 9969 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported, 9970 in2, in1, in_out, 9971 tmp1, tmp2, tmp3, 9972 w_xtmp1, w_xtmp2, w_xtmp3, 9973 tmp4, tmp5, 9974 tmp6); 9975 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported, 9976 in2, in1, in_out, 9977 tmp1, tmp2, tmp3, 9978 w_xtmp1, w_xtmp2, w_xtmp3, 9979 tmp4, tmp5, 9980 tmp6); 9981 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported, 9982 in2, in1, in_out, 9983 tmp1, tmp2, tmp3, 9984 w_xtmp1, w_xtmp2, w_xtmp3, 9985 tmp4, tmp5, 9986 tmp6); 9987 movl(tmp1, in2); 9988 andl(tmp1, 0x00000007); 9989 negl(tmp1); 9990 addl(tmp1, in2); 9991 addl(tmp1, in1); 9992 9993 BIND(L_wordByWord); 9994 cmpl(in1, tmp1); 9995 jcc(Assembler::greaterEqual, L_byteByByteProlog); 9996 crc32(in_out, Address(in1,0), 4); 9997 addl(in1, 4); 9998 jmp(L_wordByWord); 9999 10000 BIND(L_byteByByteProlog); 10001 andl(in2, 0x00000007); 10002 movl(tmp2, 1); 10003 10004 BIND(L_byteByByte); 10005 cmpl(tmp2, in2); 10006 jccb(Assembler::greater, L_exit); 10007 movb(tmp1, Address(in1, 0)); 10008 crc32(in_out, tmp1, 1); 10009 incl(in1); 10010 incl(tmp2); 10011 jmp(L_byteByByte); 10012 10013 BIND(L_exit); 10014 } 10015 #endif // LP64 10016 #undef BIND 10017 #undef BLOCK_COMMENT 10018 10019 // Compress char[] array to byte[]. 10020 // ..\jdk\src\java.base\share\classes\java\lang\StringUTF16.java 10021 // @HotSpotIntrinsicCandidate 10022 // private static int compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) { 10023 // for (int i = 0; i < len; i++) { 10024 // int c = src[srcOff++]; 10025 // if (c >>> 8 != 0) { 10026 // return 0; 10027 // } 10028 // dst[dstOff++] = (byte)c; 10029 // } 10030 // return len; 10031 // } 10032 void MacroAssembler::char_array_compress(Register src, Register dst, Register len, 10033 XMMRegister tmp1Reg, XMMRegister tmp2Reg, 10034 XMMRegister tmp3Reg, XMMRegister tmp4Reg, 10035 Register tmp5, Register result) { 10036 Label copy_chars_loop, return_length, return_zero, done; 10037 10038 // rsi: src 10039 // rdi: dst 10040 // rdx: len 10041 // rcx: tmp5 10042 // rax: result 10043 10044 // rsi holds start addr of source char[] to be compressed 10045 // rdi holds start addr of destination byte[] 10046 // rdx holds length 10047 10048 assert(len != result, ""); 10049 10050 // save length for return 10051 push(len); 10052 10053 if ((UseAVX > 2) && // AVX512 10054 VM_Version::supports_avx512vlbw() && 10055 VM_Version::supports_bmi2()) { 10056 10057 Label copy_32_loop, copy_loop_tail, below_threshold; 10058 10059 // alignment 10060 Label post_alignment; 10061 10062 // if length of the string is less than 16, handle it in an old fashioned way 10063 testl(len, -32); 10064 jcc(Assembler::zero, below_threshold); 10065 10066 // First check whether a character is compressable ( <= 0xFF). 10067 // Create mask to test for Unicode chars inside zmm vector 10068 movl(result, 0x00FF); 10069 evpbroadcastw(tmp2Reg, result, Assembler::AVX_512bit); 10070 10071 testl(len, -64); 10072 jcc(Assembler::zero, post_alignment); 10073 10074 movl(tmp5, dst); 10075 andl(tmp5, (32 - 1)); 10076 negl(tmp5); 10077 andl(tmp5, (32 - 1)); 10078 10079 // bail out when there is nothing to be done 10080 testl(tmp5, 0xFFFFFFFF); 10081 jcc(Assembler::zero, post_alignment); 10082 10083 // ~(~0 << len), where len is the # of remaining elements to process 10084 movl(result, 0xFFFFFFFF); 10085 shlxl(result, result, tmp5); 10086 notl(result); 10087 kmovdl(k3, result); 10088 10089 evmovdquw(tmp1Reg, k3, Address(src, 0), Assembler::AVX_512bit); 10090 evpcmpuw(k2, k3, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit); 10091 ktestd(k2, k3); 10092 jcc(Assembler::carryClear, return_zero); 10093 10094 evpmovwb(Address(dst, 0), k3, tmp1Reg, Assembler::AVX_512bit); 10095 10096 addptr(src, tmp5); 10097 addptr(src, tmp5); 10098 addptr(dst, tmp5); 10099 subl(len, tmp5); 10100 10101 bind(post_alignment); 10102 // end of alignment 10103 10104 movl(tmp5, len); 10105 andl(tmp5, (32 - 1)); // tail count (in chars) 10106 andl(len, ~(32 - 1)); // vector count (in chars) 10107 jcc(Assembler::zero, copy_loop_tail); 10108 10109 lea(src, Address(src, len, Address::times_2)); 10110 lea(dst, Address(dst, len, Address::times_1)); 10111 negptr(len); 10112 10113 bind(copy_32_loop); 10114 evmovdquw(tmp1Reg, Address(src, len, Address::times_2), Assembler::AVX_512bit); 10115 evpcmpuw(k2, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit); 10116 kortestdl(k2, k2); 10117 jcc(Assembler::carryClear, return_zero); 10118 10119 // All elements in current processed chunk are valid candidates for 10120 // compression. Write a truncated byte elements to the memory. 10121 evpmovwb(Address(dst, len, Address::times_1), tmp1Reg, Assembler::AVX_512bit); 10122 addptr(len, 32); 10123 jcc(Assembler::notZero, copy_32_loop); 10124 10125 bind(copy_loop_tail); 10126 // bail out when there is nothing to be done 10127 testl(tmp5, 0xFFFFFFFF); 10128 jcc(Assembler::zero, return_length); 10129 10130 movl(len, tmp5); 10131 10132 // ~(~0 << len), where len is the # of remaining elements to process 10133 movl(result, 0xFFFFFFFF); 10134 shlxl(result, result, len); 10135 notl(result); 10136 10137 kmovdl(k3, result); 10138 10139 evmovdquw(tmp1Reg, k3, Address(src, 0), Assembler::AVX_512bit); 10140 evpcmpuw(k2, k3, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit); 10141 ktestd(k2, k3); 10142 jcc(Assembler::carryClear, return_zero); 10143 10144 evpmovwb(Address(dst, 0), k3, tmp1Reg, Assembler::AVX_512bit); 10145 jmp(return_length); 10146 10147 bind(below_threshold); 10148 } 10149 10150 if (UseSSE42Intrinsics) { 10151 Label copy_32_loop, copy_16, copy_tail; 10152 10153 movl(result, len); 10154 10155 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vectors 10156 10157 // vectored compression 10158 andl(len, 0xfffffff0); // vector count (in chars) 10159 andl(result, 0x0000000f); // tail count (in chars) 10160 testl(len, len); 10161 jcc(Assembler::zero, copy_16); 10162 10163 // compress 16 chars per iter 10164 movdl(tmp1Reg, tmp5); 10165 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg 10166 pxor(tmp4Reg, tmp4Reg); 10167 10168 lea(src, Address(src, len, Address::times_2)); 10169 lea(dst, Address(dst, len, Address::times_1)); 10170 negptr(len); 10171 10172 bind(copy_32_loop); 10173 movdqu(tmp2Reg, Address(src, len, Address::times_2)); // load 1st 8 characters 10174 por(tmp4Reg, tmp2Reg); 10175 movdqu(tmp3Reg, Address(src, len, Address::times_2, 16)); // load next 8 characters 10176 por(tmp4Reg, tmp3Reg); 10177 ptest(tmp4Reg, tmp1Reg); // check for Unicode chars in next vector 10178 jcc(Assembler::notZero, return_zero); 10179 packuswb(tmp2Reg, tmp3Reg); // only ASCII chars; compress each to 1 byte 10180 movdqu(Address(dst, len, Address::times_1), tmp2Reg); 10181 addptr(len, 16); 10182 jcc(Assembler::notZero, copy_32_loop); 10183 10184 // compress next vector of 8 chars (if any) 10185 bind(copy_16); 10186 movl(len, result); 10187 andl(len, 0xfffffff8); // vector count (in chars) 10188 andl(result, 0x00000007); // tail count (in chars) 10189 testl(len, len); 10190 jccb(Assembler::zero, copy_tail); 10191 10192 movdl(tmp1Reg, tmp5); 10193 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg 10194 pxor(tmp3Reg, tmp3Reg); 10195 10196 movdqu(tmp2Reg, Address(src, 0)); 10197 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector 10198 jccb(Assembler::notZero, return_zero); 10199 packuswb(tmp2Reg, tmp3Reg); // only LATIN1 chars; compress each to 1 byte 10200 movq(Address(dst, 0), tmp2Reg); 10201 addptr(src, 16); 10202 addptr(dst, 8); 10203 10204 bind(copy_tail); 10205 movl(len, result); 10206 } 10207 // compress 1 char per iter 10208 testl(len, len); 10209 jccb(Assembler::zero, return_length); 10210 lea(src, Address(src, len, Address::times_2)); 10211 lea(dst, Address(dst, len, Address::times_1)); 10212 negptr(len); 10213 10214 bind(copy_chars_loop); 10215 load_unsigned_short(result, Address(src, len, Address::times_2)); 10216 testl(result, 0xff00); // check if Unicode char 10217 jccb(Assembler::notZero, return_zero); 10218 movb(Address(dst, len, Address::times_1), result); // ASCII char; compress to 1 byte 10219 increment(len); 10220 jcc(Assembler::notZero, copy_chars_loop); 10221 10222 // if compression succeeded, return length 10223 bind(return_length); 10224 pop(result); 10225 jmpb(done); 10226 10227 // if compression failed, return 0 10228 bind(return_zero); 10229 xorl(result, result); 10230 addptr(rsp, wordSize); 10231 10232 bind(done); 10233 } 10234 10235 // Inflate byte[] array to char[]. 10236 // ..\jdk\src\java.base\share\classes\java\lang\StringLatin1.java 10237 // @HotSpotIntrinsicCandidate 10238 // private static void inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) { 10239 // for (int i = 0; i < len; i++) { 10240 // dst[dstOff++] = (char)(src[srcOff++] & 0xff); 10241 // } 10242 // } 10243 void MacroAssembler::byte_array_inflate(Register src, Register dst, Register len, 10244 XMMRegister tmp1, Register tmp2) { 10245 Label copy_chars_loop, done, below_threshold; 10246 // rsi: src 10247 // rdi: dst 10248 // rdx: len 10249 // rcx: tmp2 10250 10251 // rsi holds start addr of source byte[] to be inflated 10252 // rdi holds start addr of destination char[] 10253 // rdx holds length 10254 assert_different_registers(src, dst, len, tmp2); 10255 10256 if ((UseAVX > 2) && // AVX512 10257 VM_Version::supports_avx512vlbw() && 10258 VM_Version::supports_bmi2()) { 10259 10260 Label copy_32_loop, copy_tail; 10261 Register tmp3_aliased = len; 10262 10263 // if length of the string is less than 16, handle it in an old fashioned way 10264 testl(len, -16); 10265 jcc(Assembler::zero, below_threshold); 10266 10267 // In order to use only one arithmetic operation for the main loop we use 10268 // this pre-calculation 10269 movl(tmp2, len); 10270 andl(tmp2, (32 - 1)); // tail count (in chars), 32 element wide loop 10271 andl(len, -32); // vector count 10272 jccb(Assembler::zero, copy_tail); 10273 10274 lea(src, Address(src, len, Address::times_1)); 10275 lea(dst, Address(dst, len, Address::times_2)); 10276 negptr(len); 10277 10278 10279 // inflate 32 chars per iter 10280 bind(copy_32_loop); 10281 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_512bit); 10282 evmovdquw(Address(dst, len, Address::times_2), tmp1, Assembler::AVX_512bit); 10283 addptr(len, 32); 10284 jcc(Assembler::notZero, copy_32_loop); 10285 10286 bind(copy_tail); 10287 // bail out when there is nothing to be done 10288 testl(tmp2, -1); // we don't destroy the contents of tmp2 here 10289 jcc(Assembler::zero, done); 10290 10291 // ~(~0 << length), where length is the # of remaining elements to process 10292 movl(tmp3_aliased, -1); 10293 shlxl(tmp3_aliased, tmp3_aliased, tmp2); 10294 notl(tmp3_aliased); 10295 kmovdl(k2, tmp3_aliased); 10296 evpmovzxbw(tmp1, k2, Address(src, 0), Assembler::AVX_512bit); 10297 evmovdquw(Address(dst, 0), k2, tmp1, Assembler::AVX_512bit); 10298 10299 jmp(done); 10300 } 10301 if (UseSSE42Intrinsics) { 10302 Label copy_16_loop, copy_8_loop, copy_bytes, copy_new_tail, copy_tail; 10303 10304 movl(tmp2, len); 10305 10306 if (UseAVX > 1) { 10307 andl(tmp2, (16 - 1)); 10308 andl(len, -16); 10309 jccb(Assembler::zero, copy_new_tail); 10310 } else { 10311 andl(tmp2, 0x00000007); // tail count (in chars) 10312 andl(len, 0xfffffff8); // vector count (in chars) 10313 jccb(Assembler::zero, copy_tail); 10314 } 10315 10316 // vectored inflation 10317 lea(src, Address(src, len, Address::times_1)); 10318 lea(dst, Address(dst, len, Address::times_2)); 10319 negptr(len); 10320 10321 if (UseAVX > 1) { 10322 bind(copy_16_loop); 10323 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_256bit); 10324 vmovdqu(Address(dst, len, Address::times_2), tmp1); 10325 addptr(len, 16); 10326 jcc(Assembler::notZero, copy_16_loop); 10327 10328 bind(below_threshold); 10329 bind(copy_new_tail); 10330 if ((UseAVX > 2) && 10331 VM_Version::supports_avx512vlbw() && 10332 VM_Version::supports_bmi2()) { 10333 movl(tmp2, len); 10334 } else { 10335 movl(len, tmp2); 10336 } 10337 andl(tmp2, 0x00000007); 10338 andl(len, 0xFFFFFFF8); 10339 jccb(Assembler::zero, copy_tail); 10340 10341 pmovzxbw(tmp1, Address(src, 0)); 10342 movdqu(Address(dst, 0), tmp1); 10343 addptr(src, 8); 10344 addptr(dst, 2 * 8); 10345 10346 jmp(copy_tail, true); 10347 } 10348 10349 // inflate 8 chars per iter 10350 bind(copy_8_loop); 10351 pmovzxbw(tmp1, Address(src, len, Address::times_1)); // unpack to 8 words 10352 movdqu(Address(dst, len, Address::times_2), tmp1); 10353 addptr(len, 8); 10354 jcc(Assembler::notZero, copy_8_loop); 10355 10356 bind(copy_tail); 10357 movl(len, tmp2); 10358 10359 cmpl(len, 4); 10360 jccb(Assembler::less, copy_bytes); 10361 10362 movdl(tmp1, Address(src, 0)); // load 4 byte chars 10363 pmovzxbw(tmp1, tmp1); 10364 movq(Address(dst, 0), tmp1); 10365 subptr(len, 4); 10366 addptr(src, 4); 10367 addptr(dst, 8); 10368 10369 bind(copy_bytes); 10370 } else { 10371 bind(below_threshold); 10372 } 10373 10374 testl(len, len); 10375 jccb(Assembler::zero, done); 10376 lea(src, Address(src, len, Address::times_1)); 10377 lea(dst, Address(dst, len, Address::times_2)); 10378 negptr(len); 10379 10380 // inflate 1 char per iter 10381 bind(copy_chars_loop); 10382 load_unsigned_byte(tmp2, Address(src, len, Address::times_1)); // load byte char 10383 movw(Address(dst, len, Address::times_2), tmp2); // inflate byte char to word 10384 increment(len); 10385 jcc(Assembler::notZero, copy_chars_loop); 10386 10387 bind(done); 10388 } 10389 10390 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) { 10391 switch (cond) { 10392 // Note some conditions are synonyms for others 10393 case Assembler::zero: return Assembler::notZero; 10394 case Assembler::notZero: return Assembler::zero; 10395 case Assembler::less: return Assembler::greaterEqual; 10396 case Assembler::lessEqual: return Assembler::greater; 10397 case Assembler::greater: return Assembler::lessEqual; 10398 case Assembler::greaterEqual: return Assembler::less; 10399 case Assembler::below: return Assembler::aboveEqual; 10400 case Assembler::belowEqual: return Assembler::above; 10401 case Assembler::above: return Assembler::belowEqual; 10402 case Assembler::aboveEqual: return Assembler::below; 10403 case Assembler::overflow: return Assembler::noOverflow; 10404 case Assembler::noOverflow: return Assembler::overflow; 10405 case Assembler::negative: return Assembler::positive; 10406 case Assembler::positive: return Assembler::negative; 10407 case Assembler::parity: return Assembler::noParity; 10408 case Assembler::noParity: return Assembler::parity; 10409 } 10410 ShouldNotReachHere(); return Assembler::overflow; 10411 } 10412 10413 SkipIfEqual::SkipIfEqual( 10414 MacroAssembler* masm, const bool* flag_addr, bool value) { 10415 _masm = masm; 10416 _masm->cmp8(ExternalAddress((address)flag_addr), value); 10417 _masm->jcc(Assembler::equal, _label); 10418 } 10419 10420 SkipIfEqual::~SkipIfEqual() { 10421 _masm->bind(_label); 10422 } 10423 10424 // 32-bit Windows has its own fast-path implementation 10425 // of get_thread 10426 #if !defined(WIN32) || defined(_LP64) 10427 10428 // This is simply a call to Thread::current() 10429 void MacroAssembler::get_thread(Register thread) { 10430 if (thread != rax) { 10431 push(rax); 10432 } 10433 LP64_ONLY(push(rdi);) 10434 LP64_ONLY(push(rsi);) 10435 push(rdx); 10436 push(rcx); 10437 #ifdef _LP64 10438 push(r8); 10439 push(r9); 10440 push(r10); 10441 push(r11); 10442 #endif 10443 10444 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, Thread::current), 0); 10445 10446 #ifdef _LP64 10447 pop(r11); 10448 pop(r10); 10449 pop(r9); 10450 pop(r8); 10451 #endif 10452 pop(rcx); 10453 pop(rdx); 10454 LP64_ONLY(pop(rsi);) 10455 LP64_ONLY(pop(rdi);) 10456 if (thread != rax) { 10457 mov(thread, rax); 10458 pop(rax); 10459 } 10460 } 10461 10462 #endif