1 /*
   2  * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/assembler.hpp"
  27 #include "asm/assembler.inline.hpp"
  28 #include "compiler/disassembler.hpp"
  29 #include "gc_interface/collectedHeap.inline.hpp"
  30 #include "interpreter/interpreter.hpp"
  31 #include "memory/cardTableModRefBS.hpp"
  32 #include "memory/resourceArea.hpp"
  33 #include "memory/universe.hpp"
  34 #include "prims/methodHandles.hpp"
  35 #include "runtime/biasedLocking.hpp"
  36 #include "runtime/interfaceSupport.hpp"
  37 #include "runtime/objectMonitor.hpp"
  38 #include "runtime/os.hpp"
  39 #include "runtime/sharedRuntime.hpp"
  40 #include "runtime/stubRoutines.hpp"
  41 #include "utilities/macros.hpp"
  42 #if INCLUDE_ALL_GCS
  43 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
  44 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
  45 #include "gc_implementation/g1/heapRegion.hpp"
  46 #endif // INCLUDE_ALL_GCS
  47 
  48 #ifdef PRODUCT
  49 #define BLOCK_COMMENT(str) /* nothing */
  50 #define STOP(error) stop(error)
  51 #else
  52 #define BLOCK_COMMENT(str) block_comment(str)
  53 #define STOP(error) block_comment(error); stop(error)
  54 #endif
  55 
  56 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
  57 
  58 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
  59 
  60 #ifdef ASSERT
  61 bool AbstractAssembler::pd_check_instruction_mark() { return true; }
  62 #endif
  63 
  64 static Assembler::Condition reverse[] = {
  65     Assembler::noOverflow     /* overflow      = 0x0 */ ,
  66     Assembler::overflow       /* noOverflow    = 0x1 */ ,
  67     Assembler::aboveEqual     /* carrySet      = 0x2, below         = 0x2 */ ,
  68     Assembler::below          /* aboveEqual    = 0x3, carryClear    = 0x3 */ ,
  69     Assembler::notZero        /* zero          = 0x4, equal         = 0x4 */ ,
  70     Assembler::zero           /* notZero       = 0x5, notEqual      = 0x5 */ ,
  71     Assembler::above          /* belowEqual    = 0x6 */ ,
  72     Assembler::belowEqual     /* above         = 0x7 */ ,
  73     Assembler::positive       /* negative      = 0x8 */ ,
  74     Assembler::negative       /* positive      = 0x9 */ ,
  75     Assembler::noParity       /* parity        = 0xa */ ,
  76     Assembler::parity         /* noParity      = 0xb */ ,
  77     Assembler::greaterEqual   /* less          = 0xc */ ,
  78     Assembler::less           /* greaterEqual  = 0xd */ ,
  79     Assembler::greater        /* lessEqual     = 0xe */ ,
  80     Assembler::lessEqual      /* greater       = 0xf, */
  81 
  82 };
  83 
  84 
  85 // Implementation of MacroAssembler
  86 
  87 // First all the versions that have distinct versions depending on 32/64 bit
  88 // Unless the difference is trivial (1 line or so).
  89 
  90 #ifndef _LP64
  91 
  92 // 32bit versions
  93 
  94 Address MacroAssembler::as_Address(AddressLiteral adr) {
  95   return Address(adr.target(), adr.rspec());
  96 }
  97 
  98 Address MacroAssembler::as_Address(ArrayAddress adr) {
  99   return Address::make_array(adr);
 100 }
 101 
 102 void MacroAssembler::call_VM_leaf_base(address entry_point,
 103                                        int number_of_arguments) {
 104   call(RuntimeAddress(entry_point));
 105   increment(rsp, number_of_arguments * wordSize);
 106 }
 107 
 108 void MacroAssembler::cmpklass(Address src1, Metadata* obj) {
 109   cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate());
 110 }
 111 
 112 void MacroAssembler::cmpklass(Register src1, Metadata* obj) {
 113   cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate());
 114 }
 115 
 116 void MacroAssembler::cmpoop(Address src1, jobject obj) {
 117   cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate());
 118 }
 119 
 120 void MacroAssembler::cmpoop(Register src1, jobject obj) {
 121   cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate());
 122 }
 123 
 124 void MacroAssembler::extend_sign(Register hi, Register lo) {
 125   // According to Intel Doc. AP-526, "Integer Divide", p.18.
 126   if (VM_Version::is_P6() && hi == rdx && lo == rax) {
 127     cdql();
 128   } else {
 129     movl(hi, lo);
 130     sarl(hi, 31);
 131   }
 132 }
 133 
 134 void MacroAssembler::jC2(Register tmp, Label& L) {
 135   // set parity bit if FPU flag C2 is set (via rax)
 136   save_rax(tmp);
 137   fwait(); fnstsw_ax();
 138   sahf();
 139   restore_rax(tmp);
 140   // branch
 141   jcc(Assembler::parity, L);
 142 }
 143 
 144 void MacroAssembler::jnC2(Register tmp, Label& L) {
 145   // set parity bit if FPU flag C2 is set (via rax)
 146   save_rax(tmp);
 147   fwait(); fnstsw_ax();
 148   sahf();
 149   restore_rax(tmp);
 150   // branch
 151   jcc(Assembler::noParity, L);
 152 }
 153 
 154 // 32bit can do a case table jump in one instruction but we no longer allow the base
 155 // to be installed in the Address class
 156 void MacroAssembler::jump(ArrayAddress entry) {
 157   jmp(as_Address(entry));
 158 }
 159 
 160 // Note: y_lo will be destroyed
 161 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) {
 162   // Long compare for Java (semantics as described in JVM spec.)
 163   Label high, low, done;
 164 
 165   cmpl(x_hi, y_hi);
 166   jcc(Assembler::less, low);
 167   jcc(Assembler::greater, high);
 168   // x_hi is the return register
 169   xorl(x_hi, x_hi);
 170   cmpl(x_lo, y_lo);
 171   jcc(Assembler::below, low);
 172   jcc(Assembler::equal, done);
 173 
 174   bind(high);
 175   xorl(x_hi, x_hi);
 176   increment(x_hi);
 177   jmp(done);
 178 
 179   bind(low);
 180   xorl(x_hi, x_hi);
 181   decrementl(x_hi);
 182 
 183   bind(done);
 184 }
 185 
 186 void MacroAssembler::lea(Register dst, AddressLiteral src) {
 187     mov_literal32(dst, (int32_t)src.target(), src.rspec());
 188 }
 189 
 190 void MacroAssembler::lea(Address dst, AddressLiteral adr) {
 191   // leal(dst, as_Address(adr));
 192   // see note in movl as to why we must use a move
 193   mov_literal32(dst, (int32_t) adr.target(), adr.rspec());
 194 }
 195 
 196 void MacroAssembler::leave() {
 197   mov(rsp, rbp);
 198   pop(rbp);
 199 }
 200 
 201 void MacroAssembler::lmul(int x_rsp_offset, int y_rsp_offset) {
 202   // Multiplication of two Java long values stored on the stack
 203   // as illustrated below. Result is in rdx:rax.
 204   //
 205   // rsp ---> [  ??  ] \               \
 206   //            ....    | y_rsp_offset  |
 207   //          [ y_lo ] /  (in bytes)    | x_rsp_offset
 208   //          [ y_hi ]                  | (in bytes)
 209   //            ....                    |
 210   //          [ x_lo ]                 /
 211   //          [ x_hi ]
 212   //            ....
 213   //
 214   // Basic idea: lo(result) = lo(x_lo * y_lo)
 215   //             hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi)
 216   Address x_hi(rsp, x_rsp_offset + wordSize); Address x_lo(rsp, x_rsp_offset);
 217   Address y_hi(rsp, y_rsp_offset + wordSize); Address y_lo(rsp, y_rsp_offset);
 218   Label quick;
 219   // load x_hi, y_hi and check if quick
 220   // multiplication is possible
 221   movl(rbx, x_hi);
 222   movl(rcx, y_hi);
 223   movl(rax, rbx);
 224   orl(rbx, rcx);                                 // rbx, = 0 <=> x_hi = 0 and y_hi = 0
 225   jcc(Assembler::zero, quick);                   // if rbx, = 0 do quick multiply
 226   // do full multiplication
 227   // 1st step
 228   mull(y_lo);                                    // x_hi * y_lo
 229   movl(rbx, rax);                                // save lo(x_hi * y_lo) in rbx,
 230   // 2nd step
 231   movl(rax, x_lo);
 232   mull(rcx);                                     // x_lo * y_hi
 233   addl(rbx, rax);                                // add lo(x_lo * y_hi) to rbx,
 234   // 3rd step
 235   bind(quick);                                   // note: rbx, = 0 if quick multiply!
 236   movl(rax, x_lo);
 237   mull(y_lo);                                    // x_lo * y_lo
 238   addl(rdx, rbx);                                // correct hi(x_lo * y_lo)
 239 }
 240 
 241 void MacroAssembler::lneg(Register hi, Register lo) {
 242   negl(lo);
 243   adcl(hi, 0);
 244   negl(hi);
 245 }
 246 
 247 void MacroAssembler::lshl(Register hi, Register lo) {
 248   // Java shift left long support (semantics as described in JVM spec., p.305)
 249   // (basic idea for shift counts s >= n: x << s == (x << n) << (s - n))
 250   // shift value is in rcx !
 251   assert(hi != rcx, "must not use rcx");
 252   assert(lo != rcx, "must not use rcx");
 253   const Register s = rcx;                        // shift count
 254   const int      n = BitsPerWord;
 255   Label L;
 256   andl(s, 0x3f);                                 // s := s & 0x3f (s < 0x40)
 257   cmpl(s, n);                                    // if (s < n)
 258   jcc(Assembler::less, L);                       // else (s >= n)
 259   movl(hi, lo);                                  // x := x << n
 260   xorl(lo, lo);
 261   // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n!
 262   bind(L);                                       // s (mod n) < n
 263   shldl(hi, lo);                                 // x := x << s
 264   shll(lo);
 265 }
 266 
 267 
 268 void MacroAssembler::lshr(Register hi, Register lo, bool sign_extension) {
 269   // Java shift right long support (semantics as described in JVM spec., p.306 & p.310)
 270   // (basic idea for shift counts s >= n: x >> s == (x >> n) >> (s - n))
 271   assert(hi != rcx, "must not use rcx");
 272   assert(lo != rcx, "must not use rcx");
 273   const Register s = rcx;                        // shift count
 274   const int      n = BitsPerWord;
 275   Label L;
 276   andl(s, 0x3f);                                 // s := s & 0x3f (s < 0x40)
 277   cmpl(s, n);                                    // if (s < n)
 278   jcc(Assembler::less, L);                       // else (s >= n)
 279   movl(lo, hi);                                  // x := x >> n
 280   if (sign_extension) sarl(hi, 31);
 281   else                xorl(hi, hi);
 282   // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n!
 283   bind(L);                                       // s (mod n) < n
 284   shrdl(lo, hi);                                 // x := x >> s
 285   if (sign_extension) sarl(hi);
 286   else                shrl(hi);
 287 }
 288 
 289 void MacroAssembler::movoop(Register dst, jobject obj) {
 290   mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate());
 291 }
 292 
 293 void MacroAssembler::movoop(Address dst, jobject obj) {
 294   mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate());
 295 }
 296 
 297 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) {
 298   mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate());
 299 }
 300 
 301 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) {
 302   mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate());
 303 }
 304 
 305 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) {
 306   // scratch register is not used,
 307   // it is defined to match parameters of 64-bit version of this method.
 308   if (src.is_lval()) {
 309     mov_literal32(dst, (intptr_t)src.target(), src.rspec());
 310   } else {
 311     movl(dst, as_Address(src));
 312   }
 313 }
 314 
 315 void MacroAssembler::movptr(ArrayAddress dst, Register src) {
 316   movl(as_Address(dst), src);
 317 }
 318 
 319 void MacroAssembler::movptr(Register dst, ArrayAddress src) {
 320   movl(dst, as_Address(src));
 321 }
 322 
 323 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
 324 void MacroAssembler::movptr(Address dst, intptr_t src) {
 325   movl(dst, src);
 326 }
 327 
 328 
 329 void MacroAssembler::pop_callee_saved_registers() {
 330   pop(rcx);
 331   pop(rdx);
 332   pop(rdi);
 333   pop(rsi);
 334 }
 335 
 336 void MacroAssembler::pop_fTOS() {
 337   fld_d(Address(rsp, 0));
 338   addl(rsp, 2 * wordSize);
 339 }
 340 
 341 void MacroAssembler::push_callee_saved_registers() {
 342   push(rsi);
 343   push(rdi);
 344   push(rdx);
 345   push(rcx);
 346 }
 347 
 348 void MacroAssembler::push_fTOS() {
 349   subl(rsp, 2 * wordSize);
 350   fstp_d(Address(rsp, 0));
 351 }
 352 
 353 
 354 void MacroAssembler::pushoop(jobject obj) {
 355   push_literal32((int32_t)obj, oop_Relocation::spec_for_immediate());
 356 }
 357 
 358 void MacroAssembler::pushklass(Metadata* obj) {
 359   push_literal32((int32_t)obj, metadata_Relocation::spec_for_immediate());
 360 }
 361 
 362 void MacroAssembler::pushptr(AddressLiteral src) {
 363   if (src.is_lval()) {
 364     push_literal32((int32_t)src.target(), src.rspec());
 365   } else {
 366     pushl(as_Address(src));
 367   }
 368 }
 369 
 370 void MacroAssembler::set_word_if_not_zero(Register dst) {
 371   xorl(dst, dst);
 372   set_byte_if_not_zero(dst);
 373 }
 374 
 375 static void pass_arg0(MacroAssembler* masm, Register arg) {
 376   masm->push(arg);
 377 }
 378 
 379 static void pass_arg1(MacroAssembler* masm, Register arg) {
 380   masm->push(arg);
 381 }
 382 
 383 static void pass_arg2(MacroAssembler* masm, Register arg) {
 384   masm->push(arg);
 385 }
 386 
 387 static void pass_arg3(MacroAssembler* masm, Register arg) {
 388   masm->push(arg);
 389 }
 390 
 391 #ifndef PRODUCT
 392 extern "C" void findpc(intptr_t x);
 393 #endif
 394 
 395 void MacroAssembler::debug32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip, char* msg) {
 396   // In order to get locks to work, we need to fake a in_VM state
 397   JavaThread* thread = JavaThread::current();
 398   JavaThreadState saved_state = thread->thread_state();
 399   thread->set_thread_state(_thread_in_vm);
 400   if (ShowMessageBoxOnError) {
 401     JavaThread* thread = JavaThread::current();
 402     JavaThreadState saved_state = thread->thread_state();
 403     thread->set_thread_state(_thread_in_vm);
 404     if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
 405       ttyLocker ttyl;
 406       BytecodeCounter::print();
 407     }
 408     // To see where a verify_oop failed, get $ebx+40/X for this frame.
 409     // This is the value of eip which points to where verify_oop will return.
 410     if (os::message_box(msg, "Execution stopped, print registers?")) {
 411       print_state32(rdi, rsi, rbp, rsp, rbx, rdx, rcx, rax, eip);
 412       BREAKPOINT;
 413     }
 414   } else {
 415     ttyLocker ttyl;
 416     ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", msg);
 417   }
 418   // Don't assert holding the ttyLock
 419     assert(false, err_msg("DEBUG MESSAGE: %s", msg));
 420   ThreadStateTransition::transition(thread, _thread_in_vm, saved_state);
 421 }
 422 
 423 void MacroAssembler::print_state32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip) {
 424   ttyLocker ttyl;
 425   FlagSetting fs(Debugging, true);
 426   tty->print_cr("eip = 0x%08x", eip);
 427 #ifndef PRODUCT
 428   if ((WizardMode || Verbose) && PrintMiscellaneous) {
 429     tty->cr();
 430     findpc(eip);
 431     tty->cr();
 432   }
 433 #endif
 434 #define PRINT_REG(rax) \
 435   { tty->print("%s = ", #rax); os::print_location(tty, rax); }
 436   PRINT_REG(rax);
 437   PRINT_REG(rbx);
 438   PRINT_REG(rcx);
 439   PRINT_REG(rdx);
 440   PRINT_REG(rdi);
 441   PRINT_REG(rsi);
 442   PRINT_REG(rbp);
 443   PRINT_REG(rsp);
 444 #undef PRINT_REG
 445   // Print some words near top of staack.
 446   int* dump_sp = (int*) rsp;
 447   for (int col1 = 0; col1 < 8; col1++) {
 448     tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
 449     os::print_location(tty, *dump_sp++);
 450   }
 451   for (int row = 0; row < 16; row++) {
 452     tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
 453     for (int col = 0; col < 8; col++) {
 454       tty->print(" 0x%08x", *dump_sp++);
 455     }
 456     tty->cr();
 457   }
 458   // Print some instructions around pc:
 459   Disassembler::decode((address)eip-64, (address)eip);
 460   tty->print_cr("--------");
 461   Disassembler::decode((address)eip, (address)eip+32);
 462 }
 463 
 464 void MacroAssembler::stop(const char* msg) {
 465   ExternalAddress message((address)msg);
 466   // push address of message
 467   pushptr(message.addr());
 468   { Label L; call(L, relocInfo::none); bind(L); }     // push eip
 469   pusha();                                            // push registers
 470   call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug32)));
 471   hlt();
 472 }
 473 
 474 void MacroAssembler::warn(const char* msg) {
 475   push_CPU_state();
 476 
 477   ExternalAddress message((address) msg);
 478   // push address of message
 479   pushptr(message.addr());
 480 
 481   call(RuntimeAddress(CAST_FROM_FN_PTR(address, warning)));
 482   addl(rsp, wordSize);       // discard argument
 483   pop_CPU_state();
 484 }
 485 
 486 void MacroAssembler::print_state() {
 487   { Label L; call(L, relocInfo::none); bind(L); }     // push eip
 488   pusha();                                            // push registers
 489 
 490   push_CPU_state();
 491   call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::print_state32)));
 492   pop_CPU_state();
 493 
 494   popa();
 495   addl(rsp, wordSize);
 496 }
 497 
 498 #else // _LP64
 499 
 500 // 64 bit versions
 501 
 502 Address MacroAssembler::as_Address(AddressLiteral adr) {
 503   // amd64 always does this as a pc-rel
 504   // we can be absolute or disp based on the instruction type
 505   // jmp/call are displacements others are absolute
 506   assert(!adr.is_lval(), "must be rval");
 507   assert(reachable(adr), "must be");
 508   return Address((int32_t)(intptr_t)(adr.target() - pc()), adr.target(), adr.reloc());
 509 
 510 }
 511 
 512 Address MacroAssembler::as_Address(ArrayAddress adr) {
 513   AddressLiteral base = adr.base();
 514   lea(rscratch1, base);
 515   Address index = adr.index();
 516   assert(index._disp == 0, "must not have disp"); // maybe it can?
 517   Address array(rscratch1, index._index, index._scale, index._disp);
 518   return array;
 519 }
 520 
 521 void MacroAssembler::call_VM_leaf_base(address entry_point, int num_args) {
 522   Label L, E;
 523 
 524 #ifdef _WIN64
 525   // Windows always allocates space for it's register args
 526   assert(num_args <= 4, "only register arguments supported");
 527   subq(rsp,  frame::arg_reg_save_area_bytes);
 528 #endif
 529 
 530   // Align stack if necessary
 531   testl(rsp, 15);
 532   jcc(Assembler::zero, L);
 533 
 534   subq(rsp, 8);
 535   {
 536     call(RuntimeAddress(entry_point));
 537   }
 538   addq(rsp, 8);
 539   jmp(E);
 540 
 541   bind(L);
 542   {
 543     call(RuntimeAddress(entry_point));
 544   }
 545 
 546   bind(E);
 547 
 548 #ifdef _WIN64
 549   // restore stack pointer
 550   addq(rsp, frame::arg_reg_save_area_bytes);
 551 #endif
 552 
 553 }
 554 
 555 void MacroAssembler::cmp64(Register src1, AddressLiteral src2) {
 556   assert(!src2.is_lval(), "should use cmpptr");
 557 
 558   if (reachable(src2)) {
 559     cmpq(src1, as_Address(src2));
 560   } else {
 561     lea(rscratch1, src2);
 562     Assembler::cmpq(src1, Address(rscratch1, 0));
 563   }
 564 }
 565 
 566 int MacroAssembler::corrected_idivq(Register reg) {
 567   // Full implementation of Java ldiv and lrem; checks for special
 568   // case as described in JVM spec., p.243 & p.271.  The function
 569   // returns the (pc) offset of the idivl instruction - may be needed
 570   // for implicit exceptions.
 571   //
 572   //         normal case                           special case
 573   //
 574   // input : rax: dividend                         min_long
 575   //         reg: divisor   (may not be eax/edx)   -1
 576   //
 577   // output: rax: quotient  (= rax idiv reg)       min_long
 578   //         rdx: remainder (= rax irem reg)       0
 579   assert(reg != rax && reg != rdx, "reg cannot be rax or rdx register");
 580   static const int64_t min_long = 0x8000000000000000;
 581   Label normal_case, special_case;
 582 
 583   // check for special case
 584   cmp64(rax, ExternalAddress((address) &min_long));
 585   jcc(Assembler::notEqual, normal_case);
 586   xorl(rdx, rdx); // prepare rdx for possible special case (where
 587                   // remainder = 0)
 588   cmpq(reg, -1);
 589   jcc(Assembler::equal, special_case);
 590 
 591   // handle normal case
 592   bind(normal_case);
 593   cdqq();
 594   int idivq_offset = offset();
 595   idivq(reg);
 596 
 597   // normal and special case exit
 598   bind(special_case);
 599 
 600   return idivq_offset;
 601 }
 602 
 603 void MacroAssembler::decrementq(Register reg, int value) {
 604   if (value == min_jint) { subq(reg, value); return; }
 605   if (value <  0) { incrementq(reg, -value); return; }
 606   if (value == 0) {                        ; return; }
 607   if (value == 1 && UseIncDec) { decq(reg) ; return; }
 608   /* else */      { subq(reg, value)       ; return; }
 609 }
 610 
 611 void MacroAssembler::decrementq(Address dst, int value) {
 612   if (value == min_jint) { subq(dst, value); return; }
 613   if (value <  0) { incrementq(dst, -value); return; }
 614   if (value == 0) {                        ; return; }
 615   if (value == 1 && UseIncDec) { decq(dst) ; return; }
 616   /* else */      { subq(dst, value)       ; return; }
 617 }
 618 
 619 void MacroAssembler::incrementq(AddressLiteral dst) {
 620   if (reachable(dst)) {
 621     incrementq(as_Address(dst));
 622   } else {
 623     lea(rscratch1, dst);
 624     incrementq(Address(rscratch1, 0));
 625   }
 626 }
 627 
 628 void MacroAssembler::incrementq(Register reg, int value) {
 629   if (value == min_jint) { addq(reg, value); return; }
 630   if (value <  0) { decrementq(reg, -value); return; }
 631   if (value == 0) {                        ; return; }
 632   if (value == 1 && UseIncDec) { incq(reg) ; return; }
 633   /* else */      { addq(reg, value)       ; return; }
 634 }
 635 
 636 void MacroAssembler::incrementq(Address dst, int value) {
 637   if (value == min_jint) { addq(dst, value); return; }
 638   if (value <  0) { decrementq(dst, -value); return; }
 639   if (value == 0) {                        ; return; }
 640   if (value == 1 && UseIncDec) { incq(dst) ; return; }
 641   /* else */      { addq(dst, value)       ; return; }
 642 }
 643 
 644 // 32bit can do a case table jump in one instruction but we no longer allow the base
 645 // to be installed in the Address class
 646 void MacroAssembler::jump(ArrayAddress entry) {
 647   lea(rscratch1, entry.base());
 648   Address dispatch = entry.index();
 649   assert(dispatch._base == noreg, "must be");
 650   dispatch._base = rscratch1;
 651   jmp(dispatch);
 652 }
 653 
 654 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) {
 655   ShouldNotReachHere(); // 64bit doesn't use two regs
 656   cmpq(x_lo, y_lo);
 657 }
 658 
 659 void MacroAssembler::lea(Register dst, AddressLiteral src) {
 660     mov_literal64(dst, (intptr_t)src.target(), src.rspec());
 661 }
 662 
 663 void MacroAssembler::lea(Address dst, AddressLiteral adr) {
 664   mov_literal64(rscratch1, (intptr_t)adr.target(), adr.rspec());
 665   movptr(dst, rscratch1);
 666 }
 667 
 668 void MacroAssembler::leave() {
 669   // %%% is this really better? Why not on 32bit too?
 670   emit_int8((unsigned char)0xC9); // LEAVE
 671 }
 672 
 673 void MacroAssembler::lneg(Register hi, Register lo) {
 674   ShouldNotReachHere(); // 64bit doesn't use two regs
 675   negq(lo);
 676 }
 677 
 678 void MacroAssembler::movoop(Register dst, jobject obj) {
 679   mov_literal64(dst, (intptr_t)obj, oop_Relocation::spec_for_immediate());
 680 }
 681 
 682 void MacroAssembler::movoop(Address dst, jobject obj) {
 683   mov_literal64(rscratch1, (intptr_t)obj, oop_Relocation::spec_for_immediate());
 684   movq(dst, rscratch1);
 685 }
 686 
 687 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) {
 688   mov_literal64(dst, (intptr_t)obj, metadata_Relocation::spec_for_immediate());
 689 }
 690 
 691 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) {
 692   mov_literal64(rscratch1, (intptr_t)obj, metadata_Relocation::spec_for_immediate());
 693   movq(dst, rscratch1);
 694 }
 695 
 696 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) {
 697   if (src.is_lval()) {
 698     mov_literal64(dst, (intptr_t)src.target(), src.rspec());
 699   } else {
 700     if (reachable(src)) {
 701       movq(dst, as_Address(src));
 702     } else {
 703       lea(scratch, src);
 704       movq(dst, Address(scratch, 0));
 705     }
 706   }
 707 }
 708 
 709 void MacroAssembler::movptr(ArrayAddress dst, Register src) {
 710   movq(as_Address(dst), src);
 711 }
 712 
 713 void MacroAssembler::movptr(Register dst, ArrayAddress src) {
 714   movq(dst, as_Address(src));
 715 }
 716 
 717 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
 718 void MacroAssembler::movptr(Address dst, intptr_t src) {
 719   mov64(rscratch1, src);
 720   movq(dst, rscratch1);
 721 }
 722 
 723 // These are mostly for initializing NULL
 724 void MacroAssembler::movptr(Address dst, int32_t src) {
 725   movslq(dst, src);
 726 }
 727 
 728 void MacroAssembler::movptr(Register dst, int32_t src) {
 729   mov64(dst, (intptr_t)src);
 730 }
 731 
 732 void MacroAssembler::pushoop(jobject obj) {
 733   movoop(rscratch1, obj);
 734   push(rscratch1);
 735 }
 736 
 737 void MacroAssembler::pushklass(Metadata* obj) {
 738   mov_metadata(rscratch1, obj);
 739   push(rscratch1);
 740 }
 741 
 742 void MacroAssembler::pushptr(AddressLiteral src) {
 743   lea(rscratch1, src);
 744   if (src.is_lval()) {
 745     push(rscratch1);
 746   } else {
 747     pushq(Address(rscratch1, 0));
 748   }
 749 }
 750 
 751 void MacroAssembler::reset_last_Java_frame(bool clear_fp,
 752                                            bool clear_pc) {
 753   // we must set sp to zero to clear frame
 754   movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
 755   // must clear fp, so that compiled frames are not confused; it is
 756   // possible that we need it only for debugging
 757   if (clear_fp) {
 758     movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
 759   }
 760 
 761   if (clear_pc) {
 762     movptr(Address(r15_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
 763   }
 764 }
 765 
 766 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
 767                                          Register last_java_fp,
 768                                          address  last_java_pc) {
 769   // determine last_java_sp register
 770   if (!last_java_sp->is_valid()) {
 771     last_java_sp = rsp;
 772   }
 773 
 774   // last_java_fp is optional
 775   if (last_java_fp->is_valid()) {
 776     movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()),
 777            last_java_fp);
 778   }
 779 
 780   // last_java_pc is optional
 781   if (last_java_pc != NULL) {
 782     Address java_pc(r15_thread,
 783                     JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset());
 784     lea(rscratch1, InternalAddress(last_java_pc));
 785     movptr(java_pc, rscratch1);
 786   }
 787 
 788   movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), last_java_sp);
 789 }
 790 
 791 static void pass_arg0(MacroAssembler* masm, Register arg) {
 792   if (c_rarg0 != arg ) {
 793     masm->mov(c_rarg0, arg);
 794   }
 795 }
 796 
 797 static void pass_arg1(MacroAssembler* masm, Register arg) {
 798   if (c_rarg1 != arg ) {
 799     masm->mov(c_rarg1, arg);
 800   }
 801 }
 802 
 803 static void pass_arg2(MacroAssembler* masm, Register arg) {
 804   if (c_rarg2 != arg ) {
 805     masm->mov(c_rarg2, arg);
 806   }
 807 }
 808 
 809 static void pass_arg3(MacroAssembler* masm, Register arg) {
 810   if (c_rarg3 != arg ) {
 811     masm->mov(c_rarg3, arg);
 812   }
 813 }
 814 
 815 void MacroAssembler::stop(const char* msg) {
 816   address rip = pc();
 817   pusha(); // get regs on stack
 818   lea(c_rarg0, ExternalAddress((address) msg));
 819   lea(c_rarg1, InternalAddress(rip));
 820   movq(c_rarg2, rsp); // pass pointer to regs array
 821   andq(rsp, -16); // align stack as required by ABI
 822   call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
 823   hlt();
 824 }
 825 
 826 void MacroAssembler::warn(const char* msg) {
 827   push(rbp);
 828   movq(rbp, rsp);
 829   andq(rsp, -16);     // align stack as required by push_CPU_state and call
 830   push_CPU_state();   // keeps alignment at 16 bytes
 831   lea(c_rarg0, ExternalAddress((address) msg));
 832   call_VM_leaf(CAST_FROM_FN_PTR(address, warning), c_rarg0);
 833   pop_CPU_state();
 834   mov(rsp, rbp);
 835   pop(rbp);
 836 }
 837 
 838 void MacroAssembler::print_state() {
 839   address rip = pc();
 840   pusha();            // get regs on stack
 841   push(rbp);
 842   movq(rbp, rsp);
 843   andq(rsp, -16);     // align stack as required by push_CPU_state and call
 844   push_CPU_state();   // keeps alignment at 16 bytes
 845 
 846   lea(c_rarg0, InternalAddress(rip));
 847   lea(c_rarg1, Address(rbp, wordSize)); // pass pointer to regs array
 848   call_VM_leaf(CAST_FROM_FN_PTR(address, MacroAssembler::print_state64), c_rarg0, c_rarg1);
 849 
 850   pop_CPU_state();
 851   mov(rsp, rbp);
 852   pop(rbp);
 853   popa();
 854 }
 855 
 856 #ifndef PRODUCT
 857 extern "C" void findpc(intptr_t x);
 858 #endif
 859 
 860 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[]) {
 861   // In order to get locks to work, we need to fake a in_VM state
 862   if (ShowMessageBoxOnError) {
 863     JavaThread* thread = JavaThread::current();
 864     JavaThreadState saved_state = thread->thread_state();
 865     thread->set_thread_state(_thread_in_vm);
 866 #ifndef PRODUCT
 867     if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
 868       ttyLocker ttyl;
 869       BytecodeCounter::print();
 870     }
 871 #endif
 872     // To see where a verify_oop failed, get $ebx+40/X for this frame.
 873     // XXX correct this offset for amd64
 874     // This is the value of eip which points to where verify_oop will return.
 875     if (os::message_box(msg, "Execution stopped, print registers?")) {
 876       print_state64(pc, regs);
 877       BREAKPOINT;
 878       assert(false, "start up GDB");
 879     }
 880     ThreadStateTransition::transition(thread, _thread_in_vm, saved_state);
 881   } else {
 882     ttyLocker ttyl;
 883     ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n",
 884                     msg);
 885     assert(false, err_msg("DEBUG MESSAGE: %s", msg));
 886   }
 887 }
 888 
 889 void MacroAssembler::print_state64(int64_t pc, int64_t regs[]) {
 890   ttyLocker ttyl;
 891   FlagSetting fs(Debugging, true);
 892   tty->print_cr("rip = 0x%016lx", pc);
 893 #ifndef PRODUCT
 894   tty->cr();
 895   findpc(pc);
 896   tty->cr();
 897 #endif
 898 #define PRINT_REG(rax, value) \
 899   { tty->print("%s = ", #rax); os::print_location(tty, value); }
 900   PRINT_REG(rax, regs[15]);
 901   PRINT_REG(rbx, regs[12]);
 902   PRINT_REG(rcx, regs[14]);
 903   PRINT_REG(rdx, regs[13]);
 904   PRINT_REG(rdi, regs[8]);
 905   PRINT_REG(rsi, regs[9]);
 906   PRINT_REG(rbp, regs[10]);
 907   PRINT_REG(rsp, regs[11]);
 908   PRINT_REG(r8 , regs[7]);
 909   PRINT_REG(r9 , regs[6]);
 910   PRINT_REG(r10, regs[5]);
 911   PRINT_REG(r11, regs[4]);
 912   PRINT_REG(r12, regs[3]);
 913   PRINT_REG(r13, regs[2]);
 914   PRINT_REG(r14, regs[1]);
 915   PRINT_REG(r15, regs[0]);
 916 #undef PRINT_REG
 917   // Print some words near top of staack.
 918   int64_t* rsp = (int64_t*) regs[11];
 919   int64_t* dump_sp = rsp;
 920   for (int col1 = 0; col1 < 8; col1++) {
 921     tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (int64_t)dump_sp);
 922     os::print_location(tty, *dump_sp++);
 923   }
 924   for (int row = 0; row < 25; row++) {
 925     tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (int64_t)dump_sp);
 926     for (int col = 0; col < 4; col++) {
 927       tty->print(" 0x%016lx", *dump_sp++);
 928     }
 929     tty->cr();
 930   }
 931   // Print some instructions around pc:
 932   Disassembler::decode((address)pc-64, (address)pc);
 933   tty->print_cr("--------");
 934   Disassembler::decode((address)pc, (address)pc+32);
 935 }
 936 
 937 #endif // _LP64
 938 
 939 // Now versions that are common to 32/64 bit
 940 
 941 void MacroAssembler::addptr(Register dst, int32_t imm32) {
 942   LP64_ONLY(addq(dst, imm32)) NOT_LP64(addl(dst, imm32));
 943 }
 944 
 945 void MacroAssembler::addptr(Register dst, Register src) {
 946   LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src));
 947 }
 948 
 949 void MacroAssembler::addptr(Address dst, Register src) {
 950   LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src));
 951 }
 952 
 953 void MacroAssembler::addsd(XMMRegister dst, AddressLiteral src) {
 954   if (reachable(src)) {
 955     Assembler::addsd(dst, as_Address(src));
 956   } else {
 957     lea(rscratch1, src);
 958     Assembler::addsd(dst, Address(rscratch1, 0));
 959   }
 960 }
 961 
 962 void MacroAssembler::addss(XMMRegister dst, AddressLiteral src) {
 963   if (reachable(src)) {
 964     addss(dst, as_Address(src));
 965   } else {
 966     lea(rscratch1, src);
 967     addss(dst, Address(rscratch1, 0));
 968   }
 969 }
 970 
 971 void MacroAssembler::align(int modulus) {
 972   if (offset() % modulus != 0) {
 973     nop(modulus - (offset() % modulus));
 974   }
 975 }
 976 
 977 void MacroAssembler::andpd(XMMRegister dst, AddressLiteral src) {
 978   // Used in sign-masking with aligned address.
 979   assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
 980   if (reachable(src)) {
 981     Assembler::andpd(dst, as_Address(src));
 982   } else {
 983     lea(rscratch1, src);
 984     Assembler::andpd(dst, Address(rscratch1, 0));
 985   }
 986 }
 987 
 988 void MacroAssembler::andps(XMMRegister dst, AddressLiteral src) {
 989   // Used in sign-masking with aligned address.
 990   assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
 991   if (reachable(src)) {
 992     Assembler::andps(dst, as_Address(src));
 993   } else {
 994     lea(rscratch1, src);
 995     Assembler::andps(dst, Address(rscratch1, 0));
 996   }
 997 }
 998 
 999 void MacroAssembler::andptr(Register dst, int32_t imm32) {
1000   LP64_ONLY(andq(dst, imm32)) NOT_LP64(andl(dst, imm32));
1001 }
1002 
1003 void MacroAssembler::atomic_incl(Address counter_addr) {
1004   if (os::is_MP())
1005     lock();
1006   incrementl(counter_addr);
1007 }
1008 
1009 void MacroAssembler::atomic_incl(AddressLiteral counter_addr, Register scr) {
1010   if (reachable(counter_addr)) {
1011     atomic_incl(as_Address(counter_addr));
1012   } else {
1013     lea(scr, counter_addr);
1014     atomic_incl(Address(scr, 0));
1015   }
1016 }
1017 
1018 #ifdef _LP64
1019 void MacroAssembler::atomic_incq(Address counter_addr) {
1020   if (os::is_MP())
1021     lock();
1022   incrementq(counter_addr);
1023 }
1024 
1025 void MacroAssembler::atomic_incq(AddressLiteral counter_addr, Register scr) {
1026   if (reachable(counter_addr)) {
1027     atomic_incq(as_Address(counter_addr));
1028   } else {
1029     lea(scr, counter_addr);
1030     atomic_incq(Address(scr, 0));
1031   }
1032 }
1033 #endif
1034 
1035 // Writes to stack successive pages until offset reached to check for
1036 // stack overflow + shadow pages.  This clobbers tmp.
1037 void MacroAssembler::bang_stack_size(Register size, Register tmp) {
1038   movptr(tmp, rsp);
1039   // Bang stack for total size given plus shadow page size.
1040   // Bang one page at a time because large size can bang beyond yellow and
1041   // red zones.
1042   Label loop;
1043   bind(loop);
1044   movl(Address(tmp, (-os::vm_page_size())), size );
1045   subptr(tmp, os::vm_page_size());
1046   subl(size, os::vm_page_size());
1047   jcc(Assembler::greater, loop);
1048 
1049   // Bang down shadow pages too.
1050   // At this point, (tmp-0) is the last address touched, so don't
1051   // touch it again.  (It was touched as (tmp-pagesize) but then tmp
1052   // was post-decremented.)  Skip this address by starting at i=1, and
1053   // touch a few more pages below.  N.B.  It is important to touch all
1054   // the way down to and including i=StackShadowPages.
1055   for (int i = 1; i < StackShadowPages; i++) {
1056     // this could be any sized move but this is can be a debugging crumb
1057     // so the bigger the better.
1058     movptr(Address(tmp, (-i*os::vm_page_size())), size );
1059   }
1060 }
1061 
1062 int MacroAssembler::biased_locking_enter(Register lock_reg,
1063                                          Register obj_reg,
1064                                          Register swap_reg,
1065                                          Register tmp_reg,
1066                                          bool swap_reg_contains_mark,
1067                                          Label& done,
1068                                          Label* slow_case,
1069                                          BiasedLockingCounters* counters) {
1070   assert(UseBiasedLocking, "why call this otherwise?");
1071   assert(swap_reg == rax, "swap_reg must be rax for cmpxchgq");
1072   LP64_ONLY( assert(tmp_reg != noreg, "tmp_reg must be supplied"); )
1073   bool need_tmp_reg = false;
1074   if (tmp_reg == noreg) {
1075     need_tmp_reg = true;
1076     tmp_reg = lock_reg;
1077     assert_different_registers(lock_reg, obj_reg, swap_reg);
1078   } else {
1079     assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg);
1080   }
1081   assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout");
1082   Address mark_addr      (obj_reg, oopDesc::mark_offset_in_bytes());
1083   Address saved_mark_addr(lock_reg, 0);
1084 
1085   if (PrintBiasedLockingStatistics && counters == NULL) {
1086     counters = BiasedLocking::counters();
1087   }
1088   // Biased locking
1089   // See whether the lock is currently biased toward our thread and
1090   // whether the epoch is still valid
1091   // Note that the runtime guarantees sufficient alignment of JavaThread
1092   // pointers to allow age to be placed into low bits
1093   // First check to see whether biasing is even enabled for this object
1094   Label cas_label;
1095   int null_check_offset = -1;
1096   if (!swap_reg_contains_mark) {
1097     null_check_offset = offset();
1098     movptr(swap_reg, mark_addr);
1099   }
1100   if (need_tmp_reg) {
1101     push(tmp_reg);
1102   }
1103   movptr(tmp_reg, swap_reg);
1104   andptr(tmp_reg, markOopDesc::biased_lock_mask_in_place);
1105   cmpptr(tmp_reg, markOopDesc::biased_lock_pattern);
1106   if (need_tmp_reg) {
1107     pop(tmp_reg);
1108   }
1109   jcc(Assembler::notEqual, cas_label);
1110   // The bias pattern is present in the object's header. Need to check
1111   // whether the bias owner and the epoch are both still current.
1112 #ifndef _LP64
1113   // Note that because there is no current thread register on x86_32 we
1114   // need to store off the mark word we read out of the object to
1115   // avoid reloading it and needing to recheck invariants below. This
1116   // store is unfortunate but it makes the overall code shorter and
1117   // simpler.
1118   movptr(saved_mark_addr, swap_reg);
1119 #endif
1120   if (need_tmp_reg) {
1121     push(tmp_reg);
1122   }
1123   if (swap_reg_contains_mark) {
1124     null_check_offset = offset();
1125   }
1126   load_prototype_header(tmp_reg, obj_reg);
1127 #ifdef _LP64
1128   orptr(tmp_reg, r15_thread);
1129   xorptr(tmp_reg, swap_reg);
1130   Register header_reg = tmp_reg;
1131 #else
1132   xorptr(tmp_reg, swap_reg);
1133   get_thread(swap_reg);
1134   xorptr(swap_reg, tmp_reg);
1135   Register header_reg = swap_reg;
1136 #endif
1137   andptr(header_reg, ~((int) markOopDesc::age_mask_in_place));
1138   if (need_tmp_reg) {
1139     pop(tmp_reg);
1140   }
1141   if (counters != NULL) {
1142     cond_inc32(Assembler::zero,
1143                ExternalAddress((address) counters->biased_lock_entry_count_addr()));
1144   }
1145   jcc(Assembler::equal, done);
1146 
1147   Label try_revoke_bias;
1148   Label try_rebias;
1149 
1150   // At this point we know that the header has the bias pattern and
1151   // that we are not the bias owner in the current epoch. We need to
1152   // figure out more details about the state of the header in order to
1153   // know what operations can be legally performed on the object's
1154   // header.
1155 
1156   // If the low three bits in the xor result aren't clear, that means
1157   // the prototype header is no longer biased and we have to revoke
1158   // the bias on this object.
1159   testptr(header_reg, markOopDesc::biased_lock_mask_in_place);
1160   jccb(Assembler::notZero, try_revoke_bias);
1161 
1162   // Biasing is still enabled for this data type. See whether the
1163   // epoch of the current bias is still valid, meaning that the epoch
1164   // bits of the mark word are equal to the epoch bits of the
1165   // prototype header. (Note that the prototype header's epoch bits
1166   // only change at a safepoint.) If not, attempt to rebias the object
1167   // toward the current thread. Note that we must be absolutely sure
1168   // that the current epoch is invalid in order to do this because
1169   // otherwise the manipulations it performs on the mark word are
1170   // illegal.
1171   testptr(header_reg, markOopDesc::epoch_mask_in_place);
1172   jccb(Assembler::notZero, try_rebias);
1173 
1174   // The epoch of the current bias is still valid but we know nothing
1175   // about the owner; it might be set or it might be clear. Try to
1176   // acquire the bias of the object using an atomic operation. If this
1177   // fails we will go in to the runtime to revoke the object's bias.
1178   // Note that we first construct the presumed unbiased header so we
1179   // don't accidentally blow away another thread's valid bias.
1180   NOT_LP64( movptr(swap_reg, saved_mark_addr); )
1181   andptr(swap_reg,
1182          markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place);
1183   if (need_tmp_reg) {
1184     push(tmp_reg);
1185   }
1186 #ifdef _LP64
1187   movptr(tmp_reg, swap_reg);
1188   orptr(tmp_reg, r15_thread);
1189 #else
1190   get_thread(tmp_reg);
1191   orptr(tmp_reg, swap_reg);
1192 #endif
1193   if (os::is_MP()) {
1194     lock();
1195   }
1196   cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1197   if (need_tmp_reg) {
1198     pop(tmp_reg);
1199   }
1200   // If the biasing toward our thread failed, this means that
1201   // another thread succeeded in biasing it toward itself and we
1202   // need to revoke that bias. The revocation will occur in the
1203   // interpreter runtime in the slow case.
1204   if (counters != NULL) {
1205     cond_inc32(Assembler::zero,
1206                ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr()));
1207   }
1208   if (slow_case != NULL) {
1209     jcc(Assembler::notZero, *slow_case);
1210   }
1211   jmp(done);
1212 
1213   bind(try_rebias);
1214   // At this point we know the epoch has expired, meaning that the
1215   // current "bias owner", if any, is actually invalid. Under these
1216   // circumstances _only_, we are allowed to use the current header's
1217   // value as the comparison value when doing the cas to acquire the
1218   // bias in the current epoch. In other words, we allow transfer of
1219   // the bias from one thread to another directly in this situation.
1220   //
1221   // FIXME: due to a lack of registers we currently blow away the age
1222   // bits in this situation. Should attempt to preserve them.
1223   if (need_tmp_reg) {
1224     push(tmp_reg);
1225   }
1226   load_prototype_header(tmp_reg, obj_reg);
1227 #ifdef _LP64
1228   orptr(tmp_reg, r15_thread);
1229 #else
1230   get_thread(swap_reg);
1231   orptr(tmp_reg, swap_reg);
1232   movptr(swap_reg, saved_mark_addr);
1233 #endif
1234   if (os::is_MP()) {
1235     lock();
1236   }
1237   cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1238   if (need_tmp_reg) {
1239     pop(tmp_reg);
1240   }
1241   // If the biasing toward our thread failed, then another thread
1242   // succeeded in biasing it toward itself and we need to revoke that
1243   // bias. The revocation will occur in the runtime in the slow case.
1244   if (counters != NULL) {
1245     cond_inc32(Assembler::zero,
1246                ExternalAddress((address) counters->rebiased_lock_entry_count_addr()));
1247   }
1248   if (slow_case != NULL) {
1249     jcc(Assembler::notZero, *slow_case);
1250   }
1251   jmp(done);
1252 
1253   bind(try_revoke_bias);
1254   // The prototype mark in the klass doesn't have the bias bit set any
1255   // more, indicating that objects of this data type are not supposed
1256   // to be biased any more. We are going to try to reset the mark of
1257   // this object to the prototype value and fall through to the
1258   // CAS-based locking scheme. Note that if our CAS fails, it means
1259   // that another thread raced us for the privilege of revoking the
1260   // bias of this particular object, so it's okay to continue in the
1261   // normal locking code.
1262   //
1263   // FIXME: due to a lack of registers we currently blow away the age
1264   // bits in this situation. Should attempt to preserve them.
1265   NOT_LP64( movptr(swap_reg, saved_mark_addr); )
1266   if (need_tmp_reg) {
1267     push(tmp_reg);
1268   }
1269   load_prototype_header(tmp_reg, obj_reg);
1270   if (os::is_MP()) {
1271     lock();
1272   }
1273   cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1274   if (need_tmp_reg) {
1275     pop(tmp_reg);
1276   }
1277   // Fall through to the normal CAS-based lock, because no matter what
1278   // the result of the above CAS, some thread must have succeeded in
1279   // removing the bias bit from the object's header.
1280   if (counters != NULL) {
1281     cond_inc32(Assembler::zero,
1282                ExternalAddress((address) counters->revoked_lock_entry_count_addr()));
1283   }
1284 
1285   bind(cas_label);
1286 
1287   return null_check_offset;
1288 }
1289 
1290 void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) {
1291   assert(UseBiasedLocking, "why call this otherwise?");
1292 
1293   // Check for biased locking unlock case, which is a no-op
1294   // Note: we do not have to check the thread ID for two reasons.
1295   // First, the interpreter checks for IllegalMonitorStateException at
1296   // a higher level. Second, if the bias was revoked while we held the
1297   // lock, the object could not be rebiased toward another thread, so
1298   // the bias bit would be clear.
1299   movptr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
1300   andptr(temp_reg, markOopDesc::biased_lock_mask_in_place);
1301   cmpptr(temp_reg, markOopDesc::biased_lock_pattern);
1302   jcc(Assembler::equal, done);
1303 }
1304 
1305 #ifdef COMPILER2
1306 
1307 #if INCLUDE_RTM_OPT
1308 
1309 // Update rtm_counters based on abort status
1310 // input: abort_status
1311 //        rtm_counters (RTMLockingCounters*)
1312 // flags are killed
1313 void MacroAssembler::rtm_counters_update(Register abort_status, Register rtm_counters) {
1314 
1315   atomic_incptr(Address(rtm_counters, RTMLockingCounters::abort_count_offset()));
1316   if (PrintPreciseRTMLockingStatistics) {
1317     for (int i = 0; i < RTMLockingCounters::ABORT_STATUS_LIMIT; i++) {
1318       Label check_abort;
1319       testl(abort_status, (1<<i));
1320       jccb(Assembler::equal, check_abort);
1321       atomic_incptr(Address(rtm_counters, RTMLockingCounters::abortX_count_offset() + (i * sizeof(uintx))));
1322       bind(check_abort);
1323     }
1324   }
1325 }
1326 
1327 // Branch if (random & (count-1) != 0), count is 2^n
1328 // tmp, scr and flags are killed
1329 void MacroAssembler::branch_on_random_using_rdtsc(Register tmp, Register scr, int count, Label& brLabel) {
1330   assert(tmp == rax, "");
1331   assert(scr == rdx, "");
1332   rdtsc(); // modifies EDX:EAX
1333   andptr(tmp, count-1);
1334   jccb(Assembler::notZero, brLabel);
1335 }
1336 
1337 // Perform abort ratio calculation, set no_rtm bit if high ratio
1338 // input:  rtm_counters_Reg (RTMLockingCounters* address)
1339 // tmpReg, rtm_counters_Reg and flags are killed
1340 void MacroAssembler::rtm_abort_ratio_calculation(Register tmpReg,
1341                                                  Register rtm_counters_Reg,
1342                                                  RTMLockingCounters* rtm_counters,
1343                                                  Metadata* method_data) {
1344   Label L_done, L_check_always_rtm1, L_check_always_rtm2;
1345 
1346   if (RTMLockingCalculationDelay > 0) {
1347     // Delay calculation
1348     movptr(tmpReg, ExternalAddress((address) RTMLockingCounters::rtm_calculation_flag_addr()), tmpReg);
1349     testptr(tmpReg, tmpReg);
1350     jccb(Assembler::equal, L_done);
1351   }
1352   // Abort ratio calculation only if abort_count > RTMAbortThreshold
1353   //   Aborted transactions = abort_count * 100
1354   //   All transactions = total_count *  RTMTotalCountIncrRate
1355   //   Set no_rtm bit if (Aborted transactions >= All transactions * RTMAbortRatio)
1356 
1357   movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::abort_count_offset()));
1358   cmpptr(tmpReg, RTMAbortThreshold);
1359   jccb(Assembler::below, L_check_always_rtm2);
1360   imulptr(tmpReg, tmpReg, 100);
1361 
1362   Register scrReg = rtm_counters_Reg;
1363   movptr(scrReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset()));
1364   imulptr(scrReg, scrReg, RTMTotalCountIncrRate);
1365   imulptr(scrReg, scrReg, RTMAbortRatio);
1366   cmpptr(tmpReg, scrReg);
1367   jccb(Assembler::below, L_check_always_rtm1);
1368   if (method_data != NULL) {
1369     // set rtm_state to "no rtm" in MDO
1370     mov_metadata(tmpReg, method_data);
1371     if (os::is_MP()) {
1372       lock();
1373     }
1374     orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), NoRTM);
1375   }
1376   jmpb(L_done);
1377   bind(L_check_always_rtm1);
1378   // Reload RTMLockingCounters* address
1379   lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters));
1380   bind(L_check_always_rtm2);
1381   movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset()));
1382   cmpptr(tmpReg, RTMLockingThreshold / RTMTotalCountIncrRate);
1383   jccb(Assembler::below, L_done);
1384   if (method_data != NULL) {
1385     // set rtm_state to "always rtm" in MDO
1386     mov_metadata(tmpReg, method_data);
1387     if (os::is_MP()) {
1388       lock();
1389     }
1390     orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), UseRTM);
1391   }
1392   bind(L_done);
1393 }
1394 
1395 // Update counters and perform abort ratio calculation
1396 // input:  abort_status_Reg
1397 // rtm_counters_Reg, flags are killed
1398 void MacroAssembler::rtm_profiling(Register abort_status_Reg,
1399                                    Register rtm_counters_Reg,
1400                                    RTMLockingCounters* rtm_counters,
1401                                    Metadata* method_data,
1402                                    bool profile_rtm) {
1403 
1404   assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1405   // update rtm counters based on rax value at abort
1406   // reads abort_status_Reg, updates flags
1407   lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters));
1408   rtm_counters_update(abort_status_Reg, rtm_counters_Reg);
1409   if (profile_rtm) {
1410     // Save abort status because abort_status_Reg is used by following code.
1411     if (RTMRetryCount > 0) {
1412       push(abort_status_Reg);
1413     }
1414     assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1415     rtm_abort_ratio_calculation(abort_status_Reg, rtm_counters_Reg, rtm_counters, method_data);
1416     // restore abort status
1417     if (RTMRetryCount > 0) {
1418       pop(abort_status_Reg);
1419     }
1420   }
1421 }
1422 
1423 // Retry on abort if abort's status is 0x6: can retry (0x2) | memory conflict (0x4)
1424 // inputs: retry_count_Reg
1425 //       : abort_status_Reg
1426 // output: retry_count_Reg decremented by 1
1427 // flags are killed
1428 void MacroAssembler::rtm_retry_lock_on_abort(Register retry_count_Reg, Register abort_status_Reg, Label& retryLabel) {
1429   Label doneRetry;
1430   assert(abort_status_Reg == rax, "");
1431   // The abort reason bits are in eax (see all states in rtmLocking.hpp)
1432   // 0x6 = conflict on which we can retry (0x2) | memory conflict (0x4)
1433   // if reason is in 0x6 and retry count != 0 then retry
1434   andptr(abort_status_Reg, 0x6);
1435   jccb(Assembler::zero, doneRetry);
1436   testl(retry_count_Reg, retry_count_Reg);
1437   jccb(Assembler::zero, doneRetry);
1438   pause();
1439   decrementl(retry_count_Reg);
1440   jmp(retryLabel);
1441   bind(doneRetry);
1442 }
1443 
1444 // Spin and retry if lock is busy,
1445 // inputs: box_Reg (monitor address)
1446 //       : retry_count_Reg
1447 // output: retry_count_Reg decremented by 1
1448 //       : clear z flag if retry count exceeded
1449 // tmp_Reg, scr_Reg, flags are killed
1450 void MacroAssembler::rtm_retry_lock_on_busy(Register retry_count_Reg, Register box_Reg,
1451                                             Register tmp_Reg, Register scr_Reg, Label& retryLabel) {
1452   Label SpinLoop, SpinExit, doneRetry;
1453   int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1454 
1455   testl(retry_count_Reg, retry_count_Reg);
1456   jccb(Assembler::zero, doneRetry);
1457   decrementl(retry_count_Reg);
1458   movptr(scr_Reg, RTMSpinLoopCount);
1459 
1460   bind(SpinLoop);
1461   pause();
1462   decrementl(scr_Reg);
1463   jccb(Assembler::lessEqual, SpinExit);
1464   movptr(tmp_Reg, Address(box_Reg, owner_offset));
1465   testptr(tmp_Reg, tmp_Reg);
1466   jccb(Assembler::notZero, SpinLoop);
1467 
1468   bind(SpinExit);
1469   jmp(retryLabel);
1470   bind(doneRetry);
1471   incrementl(retry_count_Reg); // clear z flag
1472 }
1473 
1474 // Use RTM for normal stack locks
1475 // Input: objReg (object to lock)
1476 void MacroAssembler::rtm_stack_locking(Register objReg, Register tmpReg, Register scrReg,
1477                                        Register retry_on_abort_count_Reg,
1478                                        RTMLockingCounters* stack_rtm_counters,
1479                                        Metadata* method_data, bool profile_rtm,
1480                                        Label& DONE_LABEL, Label& IsInflated) {
1481   assert(UseRTMForStackLocks, "why call this otherwise?");
1482   assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking");
1483   assert(tmpReg == rax, "");
1484   assert(scrReg == rdx, "");
1485   Label L_rtm_retry, L_decrement_retry, L_on_abort;
1486 
1487   if (RTMRetryCount > 0) {
1488     movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort
1489     bind(L_rtm_retry);
1490   }
1491   movptr(tmpReg, Address(objReg, 0));
1492   testptr(tmpReg, markOopDesc::monitor_value);  // inflated vs stack-locked|neutral|biased
1493   jcc(Assembler::notZero, IsInflated);
1494 
1495   if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1496     Label L_noincrement;
1497     if (RTMTotalCountIncrRate > 1) {
1498       // tmpReg, scrReg and flags are killed
1499       branch_on_random_using_rdtsc(tmpReg, scrReg, (int)RTMTotalCountIncrRate, L_noincrement);
1500     }
1501     assert(stack_rtm_counters != NULL, "should not be NULL when profiling RTM");
1502     atomic_incptr(ExternalAddress((address)stack_rtm_counters->total_count_addr()), scrReg);
1503     bind(L_noincrement);
1504   }
1505   xbegin(L_on_abort);
1506   movptr(tmpReg, Address(objReg, 0));       // fetch markword
1507   andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits
1508   cmpptr(tmpReg, markOopDesc::unlocked_value);            // bits = 001 unlocked
1509   jcc(Assembler::equal, DONE_LABEL);        // all done if unlocked
1510 
1511   Register abort_status_Reg = tmpReg; // status of abort is stored in RAX
1512   if (UseRTMXendForLockBusy) {
1513     xend();
1514     movptr(abort_status_Reg, 0x2);   // Set the abort status to 2 (so we can retry)
1515     jmp(L_decrement_retry);
1516   }
1517   else {
1518     xabort(0);
1519   }
1520   bind(L_on_abort);
1521   if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1522     rtm_profiling(abort_status_Reg, scrReg, stack_rtm_counters, method_data, profile_rtm);
1523   }
1524   bind(L_decrement_retry);
1525   if (RTMRetryCount > 0) {
1526     // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4)
1527     rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry);
1528   }
1529 }
1530 
1531 // Use RTM for inflating locks
1532 // inputs: objReg (object to lock)
1533 //         boxReg (on-stack box address (displaced header location) - KILLED)
1534 //         tmpReg (ObjectMonitor address + markOopDesc::monitor_value)
1535 void MacroAssembler::rtm_inflated_locking(Register objReg, Register boxReg, Register tmpReg,
1536                                           Register scrReg, Register retry_on_busy_count_Reg,
1537                                           Register retry_on_abort_count_Reg,
1538                                           RTMLockingCounters* rtm_counters,
1539                                           Metadata* method_data, bool profile_rtm,
1540                                           Label& DONE_LABEL) {
1541   assert(UseRTMLocking, "why call this otherwise?");
1542   assert(tmpReg == rax, "");
1543   assert(scrReg == rdx, "");
1544   Label L_rtm_retry, L_decrement_retry, L_on_abort;
1545   int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1546 
1547   // Without cast to int32_t a movptr will destroy r10 which is typically obj
1548   movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1549   movptr(boxReg, tmpReg); // Save ObjectMonitor address
1550 
1551   if (RTMRetryCount > 0) {
1552     movl(retry_on_busy_count_Reg, RTMRetryCount);  // Retry on lock busy
1553     movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort
1554     bind(L_rtm_retry);
1555   }
1556   if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1557     Label L_noincrement;
1558     if (RTMTotalCountIncrRate > 1) {
1559       // tmpReg, scrReg and flags are killed
1560       branch_on_random_using_rdtsc(tmpReg, scrReg, (int)RTMTotalCountIncrRate, L_noincrement);
1561     }
1562     assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1563     atomic_incptr(ExternalAddress((address)rtm_counters->total_count_addr()), scrReg);
1564     bind(L_noincrement);
1565   }
1566   xbegin(L_on_abort);
1567   movptr(tmpReg, Address(objReg, 0));
1568   movptr(tmpReg, Address(tmpReg, owner_offset));
1569   testptr(tmpReg, tmpReg);
1570   jcc(Assembler::zero, DONE_LABEL);
1571   if (UseRTMXendForLockBusy) {
1572     xend();
1573     jmp(L_decrement_retry);
1574   }
1575   else {
1576     xabort(0);
1577   }
1578   bind(L_on_abort);
1579   Register abort_status_Reg = tmpReg; // status of abort is stored in RAX
1580   if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1581     rtm_profiling(abort_status_Reg, scrReg, rtm_counters, method_data, profile_rtm);
1582   }
1583   if (RTMRetryCount > 0) {
1584     // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4)
1585     rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry);
1586   }
1587 
1588   movptr(tmpReg, Address(boxReg, owner_offset)) ;
1589   testptr(tmpReg, tmpReg) ;
1590   jccb(Assembler::notZero, L_decrement_retry) ;
1591 
1592   // Appears unlocked - try to swing _owner from null to non-null.
1593   // Invariant: tmpReg == 0.  tmpReg is EAX which is the implicit cmpxchg comparand.
1594 #ifdef _LP64
1595   Register threadReg = r15_thread;
1596 #else
1597   get_thread(scrReg);
1598   Register threadReg = scrReg;
1599 #endif
1600   if (os::is_MP()) {
1601     lock();
1602   }
1603   cmpxchgptr(threadReg, Address(boxReg, owner_offset)); // Updates tmpReg
1604 
1605   if (RTMRetryCount > 0) {
1606     // success done else retry
1607     jccb(Assembler::equal, DONE_LABEL) ;
1608     bind(L_decrement_retry);
1609     // Spin and retry if lock is busy.
1610     rtm_retry_lock_on_busy(retry_on_busy_count_Reg, boxReg, tmpReg, scrReg, L_rtm_retry);
1611   }
1612   else {
1613     bind(L_decrement_retry);
1614   }
1615 }
1616 
1617 #endif //  INCLUDE_RTM_OPT
1618 
1619 // Fast_Lock and Fast_Unlock used by C2
1620 
1621 // Because the transitions from emitted code to the runtime
1622 // monitorenter/exit helper stubs are so slow it's critical that
1623 // we inline both the stack-locking fast-path and the inflated fast path.
1624 //
1625 // See also: cmpFastLock and cmpFastUnlock.
1626 //
1627 // What follows is a specialized inline transliteration of the code
1628 // in slow_enter() and slow_exit().  If we're concerned about I$ bloat
1629 // another option would be to emit TrySlowEnter and TrySlowExit methods
1630 // at startup-time.  These methods would accept arguments as
1631 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure
1632 // indications in the icc.ZFlag.  Fast_Lock and Fast_Unlock would simply
1633 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit.
1634 // In practice, however, the # of lock sites is bounded and is usually small.
1635 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer
1636 // if the processor uses simple bimodal branch predictors keyed by EIP
1637 // Since the helper routines would be called from multiple synchronization
1638 // sites.
1639 //
1640 // An even better approach would be write "MonitorEnter()" and "MonitorExit()"
1641 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites
1642 // to those specialized methods.  That'd give us a mostly platform-independent
1643 // implementation that the JITs could optimize and inline at their pleasure.
1644 // Done correctly, the only time we'd need to cross to native could would be
1645 // to park() or unpark() threads.  We'd also need a few more unsafe operators
1646 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and
1647 // (b) explicit barriers or fence operations.
1648 //
1649 // TODO:
1650 //
1651 // *  Arrange for C2 to pass "Self" into Fast_Lock and Fast_Unlock in one of the registers (scr).
1652 //    This avoids manifesting the Self pointer in the Fast_Lock and Fast_Unlock terminals.
1653 //    Given TLAB allocation, Self is usually manifested in a register, so passing it into
1654 //    the lock operators would typically be faster than reifying Self.
1655 //
1656 // *  Ideally I'd define the primitives as:
1657 //       fast_lock   (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED.
1658 //       fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED
1659 //    Unfortunately ADLC bugs prevent us from expressing the ideal form.
1660 //    Instead, we're stuck with a rather awkward and brittle register assignments below.
1661 //    Furthermore the register assignments are overconstrained, possibly resulting in
1662 //    sub-optimal code near the synchronization site.
1663 //
1664 // *  Eliminate the sp-proximity tests and just use "== Self" tests instead.
1665 //    Alternately, use a better sp-proximity test.
1666 //
1667 // *  Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value.
1668 //    Either one is sufficient to uniquely identify a thread.
1669 //    TODO: eliminate use of sp in _owner and use get_thread(tr) instead.
1670 //
1671 // *  Intrinsify notify() and notifyAll() for the common cases where the
1672 //    object is locked by the calling thread but the waitlist is empty.
1673 //    avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll().
1674 //
1675 // *  use jccb and jmpb instead of jcc and jmp to improve code density.
1676 //    But beware of excessive branch density on AMD Opterons.
1677 //
1678 // *  Both Fast_Lock and Fast_Unlock set the ICC.ZF to indicate success
1679 //    or failure of the fast-path.  If the fast-path fails then we pass
1680 //    control to the slow-path, typically in C.  In Fast_Lock and
1681 //    Fast_Unlock we often branch to DONE_LABEL, just to find that C2
1682 //    will emit a conditional branch immediately after the node.
1683 //    So we have branches to branches and lots of ICC.ZF games.
1684 //    Instead, it might be better to have C2 pass a "FailureLabel"
1685 //    into Fast_Lock and Fast_Unlock.  In the case of success, control
1686 //    will drop through the node.  ICC.ZF is undefined at exit.
1687 //    In the case of failure, the node will branch directly to the
1688 //    FailureLabel
1689 
1690 
1691 // obj: object to lock
1692 // box: on-stack box address (displaced header location) - KILLED
1693 // rax,: tmp -- KILLED
1694 // scr: tmp -- KILLED
1695 void MacroAssembler::fast_lock(Register objReg, Register boxReg, Register tmpReg,
1696                                Register scrReg, Register cx1Reg, Register cx2Reg,
1697                                BiasedLockingCounters* counters,
1698                                RTMLockingCounters* rtm_counters,
1699                                RTMLockingCounters* stack_rtm_counters,
1700                                Metadata* method_data,
1701                                bool use_rtm, bool profile_rtm) {
1702   // Ensure the register assignents are disjoint
1703   assert(tmpReg == rax, "");
1704 
1705   if (use_rtm) {
1706     assert_different_registers(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg);
1707   } else {
1708     assert(cx1Reg == noreg, "");
1709     assert(cx2Reg == noreg, "");
1710     assert_different_registers(objReg, boxReg, tmpReg, scrReg);
1711   }
1712 
1713   if (counters != NULL) {
1714     atomic_incl(ExternalAddress((address)counters->total_entry_count_addr()), scrReg);
1715   }
1716   if (EmitSync & 1) {
1717       // set box->dhw = markOopDesc::unused_mark()
1718       // Force all sync thru slow-path: slow_enter() and slow_exit()
1719       movptr (Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1720       cmpptr (rsp, (int32_t)NULL_WORD);
1721   } else {
1722     // Possible cases that we'll encounter in fast_lock
1723     // ------------------------------------------------
1724     // * Inflated
1725     //    -- unlocked
1726     //    -- Locked
1727     //       = by self
1728     //       = by other
1729     // * biased
1730     //    -- by Self
1731     //    -- by other
1732     // * neutral
1733     // * stack-locked
1734     //    -- by self
1735     //       = sp-proximity test hits
1736     //       = sp-proximity test generates false-negative
1737     //    -- by other
1738     //
1739 
1740     Label IsInflated, DONE_LABEL;
1741 
1742     // it's stack-locked, biased or neutral
1743     // TODO: optimize away redundant LDs of obj->mark and improve the markword triage
1744     // order to reduce the number of conditional branches in the most common cases.
1745     // Beware -- there's a subtle invariant that fetch of the markword
1746     // at [FETCH], below, will never observe a biased encoding (*101b).
1747     // If this invariant is not held we risk exclusion (safety) failure.
1748     if (UseBiasedLocking && !UseOptoBiasInlining) {
1749       biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, counters);
1750     }
1751 
1752 #if INCLUDE_RTM_OPT
1753     if (UseRTMForStackLocks && use_rtm) {
1754       rtm_stack_locking(objReg, tmpReg, scrReg, cx2Reg,
1755                         stack_rtm_counters, method_data, profile_rtm,
1756                         DONE_LABEL, IsInflated);
1757     }
1758 #endif // INCLUDE_RTM_OPT
1759 
1760     movptr(tmpReg, Address(objReg, 0));          // [FETCH]
1761     testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased
1762     jccb(Assembler::notZero, IsInflated);
1763 
1764     // Attempt stack-locking ...
1765     orptr (tmpReg, markOopDesc::unlocked_value);
1766     movptr(Address(boxReg, 0), tmpReg);          // Anticipate successful CAS
1767     if (os::is_MP()) {
1768       lock();
1769     }
1770     cmpxchgptr(boxReg, Address(objReg, 0));      // Updates tmpReg
1771     if (counters != NULL) {
1772       cond_inc32(Assembler::equal,
1773                  ExternalAddress((address)counters->fast_path_entry_count_addr()));
1774     }
1775     jcc(Assembler::equal, DONE_LABEL);           // Success
1776 
1777     // Recursive locking.
1778     // The object is stack-locked: markword contains stack pointer to BasicLock.
1779     // Locked by current thread if difference with current SP is less than one page.
1780     subptr(tmpReg, rsp);
1781     // Next instruction set ZFlag == 1 (Success) if difference is less then one page.
1782     andptr(tmpReg, (int32_t) (NOT_LP64(0xFFFFF003) LP64_ONLY(7 - os::vm_page_size())) );
1783     movptr(Address(boxReg, 0), tmpReg);
1784     if (counters != NULL) {
1785       cond_inc32(Assembler::equal,
1786                  ExternalAddress((address)counters->fast_path_entry_count_addr()));
1787     }
1788     jmp(DONE_LABEL);
1789 
1790     bind(IsInflated);
1791     // The object is inflated. tmpReg contains pointer to ObjectMonitor* + markOopDesc::monitor_value
1792 
1793 #if INCLUDE_RTM_OPT
1794     // Use the same RTM locking code in 32- and 64-bit VM.
1795     if (use_rtm) {
1796       rtm_inflated_locking(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg,
1797                            rtm_counters, method_data, profile_rtm, DONE_LABEL);
1798     } else {
1799 #endif // INCLUDE_RTM_OPT
1800 
1801 #ifndef _LP64
1802     // The object is inflated.
1803 
1804     // boxReg refers to the on-stack BasicLock in the current frame.
1805     // We'd like to write:
1806     //   set box->_displaced_header = markOopDesc::unused_mark().  Any non-0 value suffices.
1807     // This is convenient but results a ST-before-CAS penalty.  The following CAS suffers
1808     // additional latency as we have another ST in the store buffer that must drain.
1809 
1810     if (EmitSync & 8192) {
1811        movptr(Address(boxReg, 0), 3);            // results in ST-before-CAS penalty
1812        get_thread (scrReg);
1813        movptr(boxReg, tmpReg);                    // consider: LEA box, [tmp-2]
1814        movptr(tmpReg, NULL_WORD);                 // consider: xor vs mov
1815        if (os::is_MP()) {
1816          lock();
1817        }
1818        cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1819     } else
1820     if ((EmitSync & 128) == 0) {                      // avoid ST-before-CAS
1821        movptr(scrReg, boxReg);
1822        movptr(boxReg, tmpReg);                   // consider: LEA box, [tmp-2]
1823 
1824        // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
1825        if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
1826           // prefetchw [eax + Offset(_owner)-2]
1827           prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1828        }
1829 
1830        if ((EmitSync & 64) == 0) {
1831          // Optimistic form: consider XORL tmpReg,tmpReg
1832          movptr(tmpReg, NULL_WORD);
1833        } else {
1834          // Can suffer RTS->RTO upgrades on shared or cold $ lines
1835          // Test-And-CAS instead of CAS
1836          movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));   // rax, = m->_owner
1837          testptr(tmpReg, tmpReg);                   // Locked ?
1838          jccb  (Assembler::notZero, DONE_LABEL);
1839        }
1840 
1841        // Appears unlocked - try to swing _owner from null to non-null.
1842        // Ideally, I'd manifest "Self" with get_thread and then attempt
1843        // to CAS the register containing Self into m->Owner.
1844        // But we don't have enough registers, so instead we can either try to CAS
1845        // rsp or the address of the box (in scr) into &m->owner.  If the CAS succeeds
1846        // we later store "Self" into m->Owner.  Transiently storing a stack address
1847        // (rsp or the address of the box) into  m->owner is harmless.
1848        // Invariant: tmpReg == 0.  tmpReg is EAX which is the implicit cmpxchg comparand.
1849        if (os::is_MP()) {
1850          lock();
1851        }
1852        cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1853        movptr(Address(scrReg, 0), 3);          // box->_displaced_header = 3
1854        jccb  (Assembler::notZero, DONE_LABEL);
1855        get_thread (scrReg);                    // beware: clobbers ICCs
1856        movptr(Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), scrReg);
1857        xorptr(boxReg, boxReg);                 // set icc.ZFlag = 1 to indicate success
1858 
1859        // If the CAS fails we can either retry or pass control to the slow-path.
1860        // We use the latter tactic.
1861        // Pass the CAS result in the icc.ZFlag into DONE_LABEL
1862        // If the CAS was successful ...
1863        //   Self has acquired the lock
1864        //   Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
1865        // Intentional fall-through into DONE_LABEL ...
1866     } else {
1867        movptr(Address(boxReg, 0), intptr_t(markOopDesc::unused_mark()));  // results in ST-before-CAS penalty
1868        movptr(boxReg, tmpReg);
1869 
1870        // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
1871        if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
1872           // prefetchw [eax + Offset(_owner)-2]
1873           prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1874        }
1875 
1876        if ((EmitSync & 64) == 0) {
1877          // Optimistic form
1878          xorptr  (tmpReg, tmpReg);
1879        } else {
1880          // Can suffer RTS->RTO upgrades on shared or cold $ lines
1881          movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));   // rax, = m->_owner
1882          testptr(tmpReg, tmpReg);                   // Locked ?
1883          jccb  (Assembler::notZero, DONE_LABEL);
1884        }
1885 
1886        // Appears unlocked - try to swing _owner from null to non-null.
1887        // Use either "Self" (in scr) or rsp as thread identity in _owner.
1888        // Invariant: tmpReg == 0.  tmpReg is EAX which is the implicit cmpxchg comparand.
1889        get_thread (scrReg);
1890        if (os::is_MP()) {
1891          lock();
1892        }
1893        cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1894 
1895        // If the CAS fails we can either retry or pass control to the slow-path.
1896        // We use the latter tactic.
1897        // Pass the CAS result in the icc.ZFlag into DONE_LABEL
1898        // If the CAS was successful ...
1899        //   Self has acquired the lock
1900        //   Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
1901        // Intentional fall-through into DONE_LABEL ...
1902     }
1903 #else // _LP64
1904     // It's inflated
1905     movq(scrReg, tmpReg);
1906     xorq(tmpReg, tmpReg);
1907 
1908     if (os::is_MP()) {
1909       lock();
1910     }
1911     cmpxchgptr(r15_thread, Address(scrReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1912     // Unconditionally set box->_displaced_header = markOopDesc::unused_mark().
1913     // Without cast to int32_t movptr will destroy r10 which is typically obj.
1914     movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1915     // Intentional fall-through into DONE_LABEL ...
1916     // Propagate ICC.ZF from CAS above into DONE_LABEL.
1917 #endif // _LP64
1918 #if INCLUDE_RTM_OPT
1919     } // use_rtm()
1920 #endif
1921     // DONE_LABEL is a hot target - we'd really like to place it at the
1922     // start of cache line by padding with NOPs.
1923     // See the AMD and Intel software optimization manuals for the
1924     // most efficient "long" NOP encodings.
1925     // Unfortunately none of our alignment mechanisms suffice.
1926     bind(DONE_LABEL);
1927 
1928     // At DONE_LABEL the icc ZFlag is set as follows ...
1929     // Fast_Unlock uses the same protocol.
1930     // ZFlag == 1 -> Success
1931     // ZFlag == 0 -> Failure - force control through the slow-path
1932   }
1933 }
1934 
1935 // obj: object to unlock
1936 // box: box address (displaced header location), killed.  Must be EAX.
1937 // tmp: killed, cannot be obj nor box.
1938 //
1939 // Some commentary on balanced locking:
1940 //
1941 // Fast_Lock and Fast_Unlock are emitted only for provably balanced lock sites.
1942 // Methods that don't have provably balanced locking are forced to run in the
1943 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock.
1944 // The interpreter provides two properties:
1945 // I1:  At return-time the interpreter automatically and quietly unlocks any
1946 //      objects acquired the current activation (frame).  Recall that the
1947 //      interpreter maintains an on-stack list of locks currently held by
1948 //      a frame.
1949 // I2:  If a method attempts to unlock an object that is not held by the
1950 //      the frame the interpreter throws IMSX.
1951 //
1952 // Lets say A(), which has provably balanced locking, acquires O and then calls B().
1953 // B() doesn't have provably balanced locking so it runs in the interpreter.
1954 // Control returns to A() and A() unlocks O.  By I1 and I2, above, we know that O
1955 // is still locked by A().
1956 //
1957 // The only other source of unbalanced locking would be JNI.  The "Java Native Interface:
1958 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter
1959 // should not be unlocked by "normal" java-level locking and vice-versa.  The specification
1960 // doesn't specify what will occur if a program engages in such mixed-mode locking, however.
1961 
1962 void MacroAssembler::fast_unlock(Register objReg, Register boxReg, Register tmpReg, bool use_rtm) {
1963   assert(boxReg == rax, "");
1964   assert_different_registers(objReg, boxReg, tmpReg);
1965 
1966   if (EmitSync & 4) {
1967     // Disable - inhibit all inlining.  Force control through the slow-path
1968     cmpptr (rsp, 0);
1969   } else
1970   if (EmitSync & 8) {
1971     Label DONE_LABEL;
1972     if (UseBiasedLocking) {
1973        biased_locking_exit(objReg, tmpReg, DONE_LABEL);
1974     }
1975     // Classic stack-locking code ...
1976     // Check whether the displaced header is 0
1977     //(=> recursive unlock)
1978     movptr(tmpReg, Address(boxReg, 0));
1979     testptr(tmpReg, tmpReg);
1980     jccb(Assembler::zero, DONE_LABEL);
1981     // If not recursive lock, reset the header to displaced header
1982     if (os::is_MP()) {
1983       lock();
1984     }
1985     cmpxchgptr(tmpReg, Address(objReg, 0));   // Uses RAX which is box
1986     bind(DONE_LABEL);
1987   } else {
1988     Label DONE_LABEL, Stacked, CheckSucc;
1989 
1990     // Critically, the biased locking test must have precedence over
1991     // and appear before the (box->dhw == 0) recursive stack-lock test.
1992     if (UseBiasedLocking && !UseOptoBiasInlining) {
1993        biased_locking_exit(objReg, tmpReg, DONE_LABEL);
1994     }
1995 
1996 #if INCLUDE_RTM_OPT
1997     if (UseRTMForStackLocks && use_rtm) {
1998       assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking");
1999       Label L_regular_unlock;
2000       movptr(tmpReg, Address(objReg, 0));           // fetch markword
2001       andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits
2002       cmpptr(tmpReg, markOopDesc::unlocked_value);            // bits = 001 unlocked
2003       jccb(Assembler::notEqual, L_regular_unlock);  // if !HLE RegularLock
2004       xend();                                       // otherwise end...
2005       jmp(DONE_LABEL);                              // ... and we're done
2006       bind(L_regular_unlock);
2007     }
2008 #endif
2009 
2010     cmpptr(Address(boxReg, 0), (int32_t)NULL_WORD); // Examine the displaced header
2011     jcc   (Assembler::zero, DONE_LABEL);            // 0 indicates recursive stack-lock
2012     movptr(tmpReg, Address(objReg, 0));             // Examine the object's markword
2013     testptr(tmpReg, markOopDesc::monitor_value);    // Inflated?
2014     jccb  (Assembler::zero, Stacked);
2015 
2016     // It's inflated.
2017 #if INCLUDE_RTM_OPT
2018     if (use_rtm) {
2019       Label L_regular_inflated_unlock;
2020       int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
2021       movptr(boxReg, Address(tmpReg, owner_offset));
2022       testptr(boxReg, boxReg);
2023       jccb(Assembler::notZero, L_regular_inflated_unlock);
2024       xend();
2025       jmpb(DONE_LABEL);
2026       bind(L_regular_inflated_unlock);
2027     }
2028 #endif
2029 
2030     // Despite our balanced locking property we still check that m->_owner == Self
2031     // as java routines or native JNI code called by this thread might
2032     // have released the lock.
2033     // Refer to the comments in synchronizer.cpp for how we might encode extra
2034     // state in _succ so we can avoid fetching EntryList|cxq.
2035     //
2036     // I'd like to add more cases in fast_lock() and fast_unlock() --
2037     // such as recursive enter and exit -- but we have to be wary of
2038     // I$ bloat, T$ effects and BP$ effects.
2039     //
2040     // If there's no contention try a 1-0 exit.  That is, exit without
2041     // a costly MEMBAR or CAS.  See synchronizer.cpp for details on how
2042     // we detect and recover from the race that the 1-0 exit admits.
2043     //
2044     // Conceptually Fast_Unlock() must execute a STST|LDST "release" barrier
2045     // before it STs null into _owner, releasing the lock.  Updates
2046     // to data protected by the critical section must be visible before
2047     // we drop the lock (and thus before any other thread could acquire
2048     // the lock and observe the fields protected by the lock).
2049     // IA32's memory-model is SPO, so STs are ordered with respect to
2050     // each other and there's no need for an explicit barrier (fence).
2051     // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
2052 #ifndef _LP64
2053     get_thread (boxReg);
2054     if ((EmitSync & 4096) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
2055       // prefetchw [ebx + Offset(_owner)-2]
2056       prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2057     }
2058 
2059     // Note that we could employ various encoding schemes to reduce
2060     // the number of loads below (currently 4) to just 2 or 3.
2061     // Refer to the comments in synchronizer.cpp.
2062     // In practice the chain of fetches doesn't seem to impact performance, however.
2063     if ((EmitSync & 65536) == 0 && (EmitSync & 256)) {
2064        // Attempt to reduce branch density - AMD's branch predictor.
2065        xorptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2066        orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2067        orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2068        orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2069        jccb  (Assembler::notZero, DONE_LABEL);
2070        movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2071        jmpb  (DONE_LABEL);
2072     } else {
2073        xorptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2074        orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2075        jccb  (Assembler::notZero, DONE_LABEL);
2076        movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2077        orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2078        jccb  (Assembler::notZero, CheckSucc);
2079        movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2080        jmpb  (DONE_LABEL);
2081     }
2082 
2083     // The Following code fragment (EmitSync & 65536) improves the performance of
2084     // contended applications and contended synchronization microbenchmarks.
2085     // Unfortunately the emission of the code - even though not executed - causes regressions
2086     // in scimark and jetstream, evidently because of $ effects.  Replacing the code
2087     // with an equal number of never-executed NOPs results in the same regression.
2088     // We leave it off by default.
2089 
2090     if ((EmitSync & 65536) != 0) {
2091        Label LSuccess, LGoSlowPath ;
2092 
2093        bind  (CheckSucc);
2094 
2095        // Optional pre-test ... it's safe to elide this
2096        if ((EmitSync & 16) == 0) {
2097           cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2098           jccb  (Assembler::zero, LGoSlowPath);
2099        }
2100 
2101        // We have a classic Dekker-style idiom:
2102        //    ST m->_owner = 0 ; MEMBAR; LD m->_succ
2103        // There are a number of ways to implement the barrier:
2104        // (1) lock:andl &m->_owner, 0
2105        //     is fast, but mask doesn't currently support the "ANDL M,IMM32" form.
2106        //     LOCK: ANDL [ebx+Offset(_Owner)-2], 0
2107        //     Encodes as 81 31 OFF32 IMM32 or 83 63 OFF8 IMM8
2108        // (2) If supported, an explicit MFENCE is appealing.
2109        //     In older IA32 processors MFENCE is slower than lock:add or xchg
2110        //     particularly if the write-buffer is full as might be the case if
2111        //     if stores closely precede the fence or fence-equivalent instruction.
2112        //     In more modern implementations MFENCE appears faster, however.
2113        // (3) In lieu of an explicit fence, use lock:addl to the top-of-stack
2114        //     The $lines underlying the top-of-stack should be in M-state.
2115        //     The locked add instruction is serializing, of course.
2116        // (4) Use xchg, which is serializing
2117        //     mov boxReg, 0; xchgl boxReg, [tmpReg + Offset(_owner)-2] also works
2118        // (5) ST m->_owner = 0 and then execute lock:orl &m->_succ, 0.
2119        //     The integer condition codes will tell us if succ was 0.
2120        //     Since _succ and _owner should reside in the same $line and
2121        //     we just stored into _owner, it's likely that the $line
2122        //     remains in M-state for the lock:orl.
2123        //
2124        // We currently use (3), although it's likely that switching to (2)
2125        // is correct for the future.
2126 
2127        movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2128        if (os::is_MP()) {
2129           if (VM_Version::supports_sse2() && 1 == FenceInstruction) {
2130             mfence();
2131           } else {
2132             lock (); addptr(Address(rsp, 0), 0);
2133           }
2134        }
2135        // Ratify _succ remains non-null
2136        cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), 0);
2137        jccb  (Assembler::notZero, LSuccess);
2138 
2139        xorptr(boxReg, boxReg);                  // box is really EAX
2140        if (os::is_MP()) { lock(); }
2141        cmpxchgptr(rsp, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2142        jccb  (Assembler::notEqual, LSuccess);
2143        // Since we're low on registers we installed rsp as a placeholding in _owner.
2144        // Now install Self over rsp.  This is safe as we're transitioning from
2145        // non-null to non=null
2146        get_thread (boxReg);
2147        movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), boxReg);
2148        // Intentional fall-through into LGoSlowPath ...
2149 
2150        bind  (LGoSlowPath);
2151        orptr(boxReg, 1);                      // set ICC.ZF=0 to indicate failure
2152        jmpb  (DONE_LABEL);
2153 
2154        bind  (LSuccess);
2155        xorptr(boxReg, boxReg);                 // set ICC.ZF=1 to indicate success
2156        jmpb  (DONE_LABEL);
2157     }
2158 
2159     bind (Stacked);
2160     // It's not inflated and it's not recursively stack-locked and it's not biased.
2161     // It must be stack-locked.
2162     // Try to reset the header to displaced header.
2163     // The "box" value on the stack is stable, so we can reload
2164     // and be assured we observe the same value as above.
2165     movptr(tmpReg, Address(boxReg, 0));
2166     if (os::is_MP()) {
2167       lock();
2168     }
2169     cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses RAX which is box
2170     // Intention fall-thru into DONE_LABEL
2171 
2172     // DONE_LABEL is a hot target - we'd really like to place it at the
2173     // start of cache line by padding with NOPs.
2174     // See the AMD and Intel software optimization manuals for the
2175     // most efficient "long" NOP encodings.
2176     // Unfortunately none of our alignment mechanisms suffice.
2177     if ((EmitSync & 65536) == 0) {
2178        bind (CheckSucc);
2179     }
2180 #else // _LP64
2181     // It's inflated
2182     movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2183     xorptr(boxReg, r15_thread);
2184     orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2185     jccb  (Assembler::notZero, DONE_LABEL);
2186     movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2187     orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2188     jccb  (Assembler::notZero, CheckSucc);
2189     movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD);
2190     jmpb  (DONE_LABEL);
2191 
2192     if ((EmitSync & 65536) == 0) {
2193       Label LSuccess, LGoSlowPath ;
2194       bind  (CheckSucc);
2195       cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2196       jccb  (Assembler::zero, LGoSlowPath);
2197 
2198       // I'd much rather use lock:andl m->_owner, 0 as it's faster than the
2199       // the explicit ST;MEMBAR combination, but masm doesn't currently support
2200       // "ANDQ M,IMM".  Don't use MFENCE here.  lock:add to TOS, xchg, etc
2201       // are all faster when the write buffer is populated.
2202       movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD);
2203       if (os::is_MP()) {
2204          lock (); addl (Address(rsp, 0), 0);
2205       }
2206       cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2207       jccb  (Assembler::notZero, LSuccess);
2208 
2209       movptr (boxReg, (int32_t)NULL_WORD);                   // box is really EAX
2210       if (os::is_MP()) { lock(); }
2211       cmpxchgptr(r15_thread, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2212       jccb  (Assembler::notEqual, LSuccess);
2213       // Intentional fall-through into slow-path
2214 
2215       bind  (LGoSlowPath);
2216       orl   (boxReg, 1);                      // set ICC.ZF=0 to indicate failure
2217       jmpb  (DONE_LABEL);
2218 
2219       bind  (LSuccess);
2220       testl (boxReg, 0);                      // set ICC.ZF=1 to indicate success
2221       jmpb  (DONE_LABEL);
2222     }
2223 
2224     bind  (Stacked);
2225     movptr(tmpReg, Address (boxReg, 0));      // re-fetch
2226     if (os::is_MP()) { lock(); }
2227     cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses RAX which is box
2228 
2229     if (EmitSync & 65536) {
2230        bind (CheckSucc);
2231     }
2232 #endif
2233     bind(DONE_LABEL);
2234     // Avoid branch to branch on AMD processors
2235     if (EmitSync & 32768) {
2236        nop();
2237     }
2238   }
2239 }
2240 #endif // COMPILER2
2241 
2242 void MacroAssembler::c2bool(Register x) {
2243   // implements x == 0 ? 0 : 1
2244   // note: must only look at least-significant byte of x
2245   //       since C-style booleans are stored in one byte
2246   //       only! (was bug)
2247   andl(x, 0xFF);
2248   setb(Assembler::notZero, x);
2249 }
2250 
2251 // Wouldn't need if AddressLiteral version had new name
2252 void MacroAssembler::call(Label& L, relocInfo::relocType rtype) {
2253   Assembler::call(L, rtype);
2254 }
2255 
2256 void MacroAssembler::call(Register entry) {
2257   Assembler::call(entry);
2258 }
2259 
2260 void MacroAssembler::call(AddressLiteral entry) {
2261   if (reachable(entry)) {
2262     Assembler::call_literal(entry.target(), entry.rspec());
2263   } else {
2264     lea(rscratch1, entry);
2265     Assembler::call(rscratch1);
2266   }
2267 }
2268 
2269 void MacroAssembler::ic_call(address entry) {
2270   RelocationHolder rh = virtual_call_Relocation::spec(pc());
2271   movptr(rax, (intptr_t)Universe::non_oop_word());
2272   call(AddressLiteral(entry, rh));
2273 }
2274 
2275 // Implementation of call_VM versions
2276 
2277 void MacroAssembler::call_VM(Register oop_result,
2278                              address entry_point,
2279                              bool check_exceptions) {
2280   Label C, E;
2281   call(C, relocInfo::none);
2282   jmp(E);
2283 
2284   bind(C);
2285   call_VM_helper(oop_result, entry_point, 0, check_exceptions);
2286   ret(0);
2287 
2288   bind(E);
2289 }
2290 
2291 void MacroAssembler::call_VM(Register oop_result,
2292                              address entry_point,
2293                              Register arg_1,
2294                              bool check_exceptions) {
2295   Label C, E;
2296   call(C, relocInfo::none);
2297   jmp(E);
2298 
2299   bind(C);
2300   pass_arg1(this, arg_1);
2301   call_VM_helper(oop_result, entry_point, 1, check_exceptions);
2302   ret(0);
2303 
2304   bind(E);
2305 }
2306 
2307 void MacroAssembler::call_VM(Register oop_result,
2308                              address entry_point,
2309                              Register arg_1,
2310                              Register arg_2,
2311                              bool check_exceptions) {
2312   Label C, E;
2313   call(C, relocInfo::none);
2314   jmp(E);
2315 
2316   bind(C);
2317 
2318   LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2319 
2320   pass_arg2(this, arg_2);
2321   pass_arg1(this, arg_1);
2322   call_VM_helper(oop_result, entry_point, 2, check_exceptions);
2323   ret(0);
2324 
2325   bind(E);
2326 }
2327 
2328 void MacroAssembler::call_VM(Register oop_result,
2329                              address entry_point,
2330                              Register arg_1,
2331                              Register arg_2,
2332                              Register arg_3,
2333                              bool check_exceptions) {
2334   Label C, E;
2335   call(C, relocInfo::none);
2336   jmp(E);
2337 
2338   bind(C);
2339 
2340   LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2341   LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2342   pass_arg3(this, arg_3);
2343 
2344   LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2345   pass_arg2(this, arg_2);
2346 
2347   pass_arg1(this, arg_1);
2348   call_VM_helper(oop_result, entry_point, 3, check_exceptions);
2349   ret(0);
2350 
2351   bind(E);
2352 }
2353 
2354 void MacroAssembler::call_VM(Register oop_result,
2355                              Register last_java_sp,
2356                              address entry_point,
2357                              int number_of_arguments,
2358                              bool check_exceptions) {
2359   Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg);
2360   call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
2361 }
2362 
2363 void MacroAssembler::call_VM(Register oop_result,
2364                              Register last_java_sp,
2365                              address entry_point,
2366                              Register arg_1,
2367                              bool check_exceptions) {
2368   pass_arg1(this, arg_1);
2369   call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
2370 }
2371 
2372 void MacroAssembler::call_VM(Register oop_result,
2373                              Register last_java_sp,
2374                              address entry_point,
2375                              Register arg_1,
2376                              Register arg_2,
2377                              bool check_exceptions) {
2378 
2379   LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2380   pass_arg2(this, arg_2);
2381   pass_arg1(this, arg_1);
2382   call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
2383 }
2384 
2385 void MacroAssembler::call_VM(Register oop_result,
2386                              Register last_java_sp,
2387                              address entry_point,
2388                              Register arg_1,
2389                              Register arg_2,
2390                              Register arg_3,
2391                              bool check_exceptions) {
2392   LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2393   LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2394   pass_arg3(this, arg_3);
2395   LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2396   pass_arg2(this, arg_2);
2397   pass_arg1(this, arg_1);
2398   call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
2399 }
2400 
2401 void MacroAssembler::super_call_VM(Register oop_result,
2402                                    Register last_java_sp,
2403                                    address entry_point,
2404                                    int number_of_arguments,
2405                                    bool check_exceptions) {
2406   Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg);
2407   MacroAssembler::call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
2408 }
2409 
2410 void MacroAssembler::super_call_VM(Register oop_result,
2411                                    Register last_java_sp,
2412                                    address entry_point,
2413                                    Register arg_1,
2414                                    bool check_exceptions) {
2415   pass_arg1(this, arg_1);
2416   super_call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
2417 }
2418 
2419 void MacroAssembler::super_call_VM(Register oop_result,
2420                                    Register last_java_sp,
2421                                    address entry_point,
2422                                    Register arg_1,
2423                                    Register arg_2,
2424                                    bool check_exceptions) {
2425 
2426   LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2427   pass_arg2(this, arg_2);
2428   pass_arg1(this, arg_1);
2429   super_call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
2430 }
2431 
2432 void MacroAssembler::super_call_VM(Register oop_result,
2433                                    Register last_java_sp,
2434                                    address entry_point,
2435                                    Register arg_1,
2436                                    Register arg_2,
2437                                    Register arg_3,
2438                                    bool check_exceptions) {
2439   LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2440   LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2441   pass_arg3(this, arg_3);
2442   LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2443   pass_arg2(this, arg_2);
2444   pass_arg1(this, arg_1);
2445   super_call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
2446 }
2447 
2448 void MacroAssembler::call_VM_base(Register oop_result,
2449                                   Register java_thread,
2450                                   Register last_java_sp,
2451                                   address  entry_point,
2452                                   int      number_of_arguments,
2453                                   bool     check_exceptions) {
2454   // determine java_thread register
2455   if (!java_thread->is_valid()) {
2456 #ifdef _LP64
2457     java_thread = r15_thread;
2458 #else
2459     java_thread = rdi;
2460     get_thread(java_thread);
2461 #endif // LP64
2462   }
2463   // determine last_java_sp register
2464   if (!last_java_sp->is_valid()) {
2465     last_java_sp = rsp;
2466   }
2467   // debugging support
2468   assert(number_of_arguments >= 0   , "cannot have negative number of arguments");
2469   LP64_ONLY(assert(java_thread == r15_thread, "unexpected register"));
2470 #ifdef ASSERT
2471   // TraceBytecodes does not use r12 but saves it over the call, so don't verify
2472   // r12 is the heapbase.
2473   LP64_ONLY(if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");)
2474 #endif // ASSERT
2475 
2476   assert(java_thread != oop_result  , "cannot use the same register for java_thread & oop_result");
2477   assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp");
2478 
2479   // push java thread (becomes first argument of C function)
2480 
2481   NOT_LP64(push(java_thread); number_of_arguments++);
2482   LP64_ONLY(mov(c_rarg0, r15_thread));
2483 
2484   // set last Java frame before call
2485   assert(last_java_sp != rbp, "can't use ebp/rbp");
2486 
2487   // Only interpreter should have to set fp
2488   set_last_Java_frame(java_thread, last_java_sp, rbp, NULL);
2489 
2490   // do the call, remove parameters
2491   MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments);
2492 
2493   // restore the thread (cannot use the pushed argument since arguments
2494   // may be overwritten by C code generated by an optimizing compiler);
2495   // however can use the register value directly if it is callee saved.
2496   if (LP64_ONLY(true ||) java_thread == rdi || java_thread == rsi) {
2497     // rdi & rsi (also r15) are callee saved -> nothing to do
2498 #ifdef ASSERT
2499     guarantee(java_thread != rax, "change this code");
2500     push(rax);
2501     { Label L;
2502       get_thread(rax);
2503       cmpptr(java_thread, rax);
2504       jcc(Assembler::equal, L);
2505       STOP("MacroAssembler::call_VM_base: rdi not callee saved?");
2506       bind(L);
2507     }
2508     pop(rax);
2509 #endif
2510   } else {
2511     get_thread(java_thread);
2512   }
2513   // reset last Java frame
2514   // Only interpreter should have to clear fp
2515   reset_last_Java_frame(java_thread, true, false);
2516 
2517 #ifndef CC_INTERP
2518    // C++ interp handles this in the interpreter
2519   check_and_handle_popframe(java_thread);
2520   check_and_handle_earlyret(java_thread);
2521 #endif /* CC_INTERP */
2522 
2523   if (check_exceptions) {
2524     // check for pending exceptions (java_thread is set upon return)
2525     cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t) NULL_WORD);
2526 #ifndef _LP64
2527     jump_cc(Assembler::notEqual,
2528             RuntimeAddress(StubRoutines::forward_exception_entry()));
2529 #else
2530     // This used to conditionally jump to forward_exception however it is
2531     // possible if we relocate that the branch will not reach. So we must jump
2532     // around so we can always reach
2533 
2534     Label ok;
2535     jcc(Assembler::equal, ok);
2536     jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2537     bind(ok);
2538 #endif // LP64
2539   }
2540 
2541   // get oop result if there is one and reset the value in the thread
2542   if (oop_result->is_valid()) {
2543     get_vm_result(oop_result, java_thread);
2544   }
2545 }
2546 
2547 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) {
2548 
2549   // Calculate the value for last_Java_sp
2550   // somewhat subtle. call_VM does an intermediate call
2551   // which places a return address on the stack just under the
2552   // stack pointer as the user finsihed with it. This allows
2553   // use to retrieve last_Java_pc from last_Java_sp[-1].
2554   // On 32bit we then have to push additional args on the stack to accomplish
2555   // the actual requested call. On 64bit call_VM only can use register args
2556   // so the only extra space is the return address that call_VM created.
2557   // This hopefully explains the calculations here.
2558 
2559 #ifdef _LP64
2560   // We've pushed one address, correct last_Java_sp
2561   lea(rax, Address(rsp, wordSize));
2562 #else
2563   lea(rax, Address(rsp, (1 + number_of_arguments) * wordSize));
2564 #endif // LP64
2565 
2566   call_VM_base(oop_result, noreg, rax, entry_point, number_of_arguments, check_exceptions);
2567 
2568 }
2569 
2570 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) {
2571   call_VM_leaf_base(entry_point, number_of_arguments);
2572 }
2573 
2574 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) {
2575   pass_arg0(this, arg_0);
2576   call_VM_leaf(entry_point, 1);
2577 }
2578 
2579 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2580 
2581   LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2582   pass_arg1(this, arg_1);
2583   pass_arg0(this, arg_0);
2584   call_VM_leaf(entry_point, 2);
2585 }
2586 
2587 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2588   LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2589   LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2590   pass_arg2(this, arg_2);
2591   LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2592   pass_arg1(this, arg_1);
2593   pass_arg0(this, arg_0);
2594   call_VM_leaf(entry_point, 3);
2595 }
2596 
2597 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) {
2598   pass_arg0(this, arg_0);
2599   MacroAssembler::call_VM_leaf_base(entry_point, 1);
2600 }
2601 
2602 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2603 
2604   LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2605   pass_arg1(this, arg_1);
2606   pass_arg0(this, arg_0);
2607   MacroAssembler::call_VM_leaf_base(entry_point, 2);
2608 }
2609 
2610 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2611   LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2612   LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2613   pass_arg2(this, arg_2);
2614   LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2615   pass_arg1(this, arg_1);
2616   pass_arg0(this, arg_0);
2617   MacroAssembler::call_VM_leaf_base(entry_point, 3);
2618 }
2619 
2620 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) {
2621   LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg"));
2622   LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2623   LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2624   pass_arg3(this, arg_3);
2625   LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2626   LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2627   pass_arg2(this, arg_2);
2628   LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2629   pass_arg1(this, arg_1);
2630   pass_arg0(this, arg_0);
2631   MacroAssembler::call_VM_leaf_base(entry_point, 4);
2632 }
2633 
2634 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) {
2635   movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset()));
2636   movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD);
2637   verify_oop(oop_result, "broken oop in call_VM_base");
2638 }
2639 
2640 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) {
2641   movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset()));
2642   movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD);
2643 }
2644 
2645 void MacroAssembler::check_and_handle_earlyret(Register java_thread) {
2646 }
2647 
2648 void MacroAssembler::check_and_handle_popframe(Register java_thread) {
2649 }
2650 
2651 void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) {
2652   if (reachable(src1)) {
2653     cmpl(as_Address(src1), imm);
2654   } else {
2655     lea(rscratch1, src1);
2656     cmpl(Address(rscratch1, 0), imm);
2657   }
2658 }
2659 
2660 void MacroAssembler::cmp32(Register src1, AddressLiteral src2) {
2661   assert(!src2.is_lval(), "use cmpptr");
2662   if (reachable(src2)) {
2663     cmpl(src1, as_Address(src2));
2664   } else {
2665     lea(rscratch1, src2);
2666     cmpl(src1, Address(rscratch1, 0));
2667   }
2668 }
2669 
2670 void MacroAssembler::cmp32(Register src1, int32_t imm) {
2671   Assembler::cmpl(src1, imm);
2672 }
2673 
2674 void MacroAssembler::cmp32(Register src1, Address src2) {
2675   Assembler::cmpl(src1, src2);
2676 }
2677 
2678 void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2679   ucomisd(opr1, opr2);
2680 
2681   Label L;
2682   if (unordered_is_less) {
2683     movl(dst, -1);
2684     jcc(Assembler::parity, L);
2685     jcc(Assembler::below , L);
2686     movl(dst, 0);
2687     jcc(Assembler::equal , L);
2688     increment(dst);
2689   } else { // unordered is greater
2690     movl(dst, 1);
2691     jcc(Assembler::parity, L);
2692     jcc(Assembler::above , L);
2693     movl(dst, 0);
2694     jcc(Assembler::equal , L);
2695     decrementl(dst);
2696   }
2697   bind(L);
2698 }
2699 
2700 void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2701   ucomiss(opr1, opr2);
2702 
2703   Label L;
2704   if (unordered_is_less) {
2705     movl(dst, -1);
2706     jcc(Assembler::parity, L);
2707     jcc(Assembler::below , L);
2708     movl(dst, 0);
2709     jcc(Assembler::equal , L);
2710     increment(dst);
2711   } else { // unordered is greater
2712     movl(dst, 1);
2713     jcc(Assembler::parity, L);
2714     jcc(Assembler::above , L);
2715     movl(dst, 0);
2716     jcc(Assembler::equal , L);
2717     decrementl(dst);
2718   }
2719   bind(L);
2720 }
2721 
2722 
2723 void MacroAssembler::cmp8(AddressLiteral src1, int imm) {
2724   if (reachable(src1)) {
2725     cmpb(as_Address(src1), imm);
2726   } else {
2727     lea(rscratch1, src1);
2728     cmpb(Address(rscratch1, 0), imm);
2729   }
2730 }
2731 
2732 void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) {
2733 #ifdef _LP64
2734   if (src2.is_lval()) {
2735     movptr(rscratch1, src2);
2736     Assembler::cmpq(src1, rscratch1);
2737   } else if (reachable(src2)) {
2738     cmpq(src1, as_Address(src2));
2739   } else {
2740     lea(rscratch1, src2);
2741     Assembler::cmpq(src1, Address(rscratch1, 0));
2742   }
2743 #else
2744   if (src2.is_lval()) {
2745     cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2746   } else {
2747     cmpl(src1, as_Address(src2));
2748   }
2749 #endif // _LP64
2750 }
2751 
2752 void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) {
2753   assert(src2.is_lval(), "not a mem-mem compare");
2754 #ifdef _LP64
2755   // moves src2's literal address
2756   movptr(rscratch1, src2);
2757   Assembler::cmpq(src1, rscratch1);
2758 #else
2759   cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2760 #endif // _LP64
2761 }
2762 
2763 void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) {
2764   if (reachable(adr)) {
2765     if (os::is_MP())
2766       lock();
2767     cmpxchgptr(reg, as_Address(adr));
2768   } else {
2769     lea(rscratch1, adr);
2770     if (os::is_MP())
2771       lock();
2772     cmpxchgptr(reg, Address(rscratch1, 0));
2773   }
2774 }
2775 
2776 void MacroAssembler::cmpxchgptr(Register reg, Address adr) {
2777   LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr));
2778 }
2779 
2780 void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) {
2781   if (reachable(src)) {
2782     Assembler::comisd(dst, as_Address(src));
2783   } else {
2784     lea(rscratch1, src);
2785     Assembler::comisd(dst, Address(rscratch1, 0));
2786   }
2787 }
2788 
2789 void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) {
2790   if (reachable(src)) {
2791     Assembler::comiss(dst, as_Address(src));
2792   } else {
2793     lea(rscratch1, src);
2794     Assembler::comiss(dst, Address(rscratch1, 0));
2795   }
2796 }
2797 
2798 
2799 void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) {
2800   Condition negated_cond = negate_condition(cond);
2801   Label L;
2802   jcc(negated_cond, L);
2803   pushf(); // Preserve flags
2804   atomic_incl(counter_addr);
2805   popf();
2806   bind(L);
2807 }
2808 
2809 int MacroAssembler::corrected_idivl(Register reg) {
2810   // Full implementation of Java idiv and irem; checks for
2811   // special case as described in JVM spec., p.243 & p.271.
2812   // The function returns the (pc) offset of the idivl
2813   // instruction - may be needed for implicit exceptions.
2814   //
2815   //         normal case                           special case
2816   //
2817   // input : rax,: dividend                         min_int
2818   //         reg: divisor   (may not be rax,/rdx)   -1
2819   //
2820   // output: rax,: quotient  (= rax, idiv reg)       min_int
2821   //         rdx: remainder (= rax, irem reg)       0
2822   assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register");
2823   const int min_int = 0x80000000;
2824   Label normal_case, special_case;
2825 
2826   // check for special case
2827   cmpl(rax, min_int);
2828   jcc(Assembler::notEqual, normal_case);
2829   xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0)
2830   cmpl(reg, -1);
2831   jcc(Assembler::equal, special_case);
2832 
2833   // handle normal case
2834   bind(normal_case);
2835   cdql();
2836   int idivl_offset = offset();
2837   idivl(reg);
2838 
2839   // normal and special case exit
2840   bind(special_case);
2841 
2842   return idivl_offset;
2843 }
2844 
2845 
2846 
2847 void MacroAssembler::decrementl(Register reg, int value) {
2848   if (value == min_jint) {subl(reg, value) ; return; }
2849   if (value <  0) { incrementl(reg, -value); return; }
2850   if (value == 0) {                        ; return; }
2851   if (value == 1 && UseIncDec) { decl(reg) ; return; }
2852   /* else */      { subl(reg, value)       ; return; }
2853 }
2854 
2855 void MacroAssembler::decrementl(Address dst, int value) {
2856   if (value == min_jint) {subl(dst, value) ; return; }
2857   if (value <  0) { incrementl(dst, -value); return; }
2858   if (value == 0) {                        ; return; }
2859   if (value == 1 && UseIncDec) { decl(dst) ; return; }
2860   /* else */      { subl(dst, value)       ; return; }
2861 }
2862 
2863 void MacroAssembler::division_with_shift (Register reg, int shift_value) {
2864   assert (shift_value > 0, "illegal shift value");
2865   Label _is_positive;
2866   testl (reg, reg);
2867   jcc (Assembler::positive, _is_positive);
2868   int offset = (1 << shift_value) - 1 ;
2869 
2870   if (offset == 1) {
2871     incrementl(reg);
2872   } else {
2873     addl(reg, offset);
2874   }
2875 
2876   bind (_is_positive);
2877   sarl(reg, shift_value);
2878 }
2879 
2880 void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) {
2881   if (reachable(src)) {
2882     Assembler::divsd(dst, as_Address(src));
2883   } else {
2884     lea(rscratch1, src);
2885     Assembler::divsd(dst, Address(rscratch1, 0));
2886   }
2887 }
2888 
2889 void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) {
2890   if (reachable(src)) {
2891     Assembler::divss(dst, as_Address(src));
2892   } else {
2893     lea(rscratch1, src);
2894     Assembler::divss(dst, Address(rscratch1, 0));
2895   }
2896 }
2897 
2898 // !defined(COMPILER2) is because of stupid core builds
2899 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2)
2900 void MacroAssembler::empty_FPU_stack() {
2901   if (VM_Version::supports_mmx()) {
2902     emms();
2903   } else {
2904     for (int i = 8; i-- > 0; ) ffree(i);
2905   }
2906 }
2907 #endif // !LP64 || C1 || !C2
2908 
2909 
2910 // Defines obj, preserves var_size_in_bytes
2911 void MacroAssembler::eden_allocate(Register obj,
2912                                    Register var_size_in_bytes,
2913                                    int con_size_in_bytes,
2914                                    Register t1,
2915                                    Label& slow_case) {
2916   assert(obj == rax, "obj must be in rax, for cmpxchg");
2917   assert_different_registers(obj, var_size_in_bytes, t1);
2918   if (!Universe::heap()->supports_inline_contig_alloc()) {
2919     jmp(slow_case);
2920   } else {
2921     Register end = t1;
2922     Label retry;
2923     bind(retry);
2924     ExternalAddress heap_top((address) Universe::heap()->top_addr());
2925     movptr(obj, heap_top);
2926     if (var_size_in_bytes == noreg) {
2927       lea(end, Address(obj, con_size_in_bytes));
2928     } else {
2929       lea(end, Address(obj, var_size_in_bytes, Address::times_1));
2930     }
2931     // if end < obj then we wrapped around => object too long => slow case
2932     cmpptr(end, obj);
2933     jcc(Assembler::below, slow_case);
2934     cmpptr(end, ExternalAddress((address) Universe::heap()->end_addr()));
2935     jcc(Assembler::above, slow_case);
2936     // Compare obj with the top addr, and if still equal, store the new top addr in
2937     // end at the address of the top addr pointer. Sets ZF if was equal, and clears
2938     // it otherwise. Use lock prefix for atomicity on MPs.
2939     locked_cmpxchgptr(end, heap_top);
2940     jcc(Assembler::notEqual, retry);
2941   }
2942 }
2943 
2944 void MacroAssembler::enter() {
2945   push(rbp);
2946   mov(rbp, rsp);
2947 }
2948 
2949 // A 5 byte nop that is safe for patching (see patch_verified_entry)
2950 void MacroAssembler::fat_nop() {
2951   if (UseAddressNop) {
2952     addr_nop_5();
2953   } else {
2954     emit_int8(0x26); // es:
2955     emit_int8(0x2e); // cs:
2956     emit_int8(0x64); // fs:
2957     emit_int8(0x65); // gs:
2958     emit_int8((unsigned char)0x90);
2959   }
2960 }
2961 
2962 void MacroAssembler::fcmp(Register tmp) {
2963   fcmp(tmp, 1, true, true);
2964 }
2965 
2966 void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) {
2967   assert(!pop_right || pop_left, "usage error");
2968   if (VM_Version::supports_cmov()) {
2969     assert(tmp == noreg, "unneeded temp");
2970     if (pop_left) {
2971       fucomip(index);
2972     } else {
2973       fucomi(index);
2974     }
2975     if (pop_right) {
2976       fpop();
2977     }
2978   } else {
2979     assert(tmp != noreg, "need temp");
2980     if (pop_left) {
2981       if (pop_right) {
2982         fcompp();
2983       } else {
2984         fcomp(index);
2985       }
2986     } else {
2987       fcom(index);
2988     }
2989     // convert FPU condition into eflags condition via rax,
2990     save_rax(tmp);
2991     fwait(); fnstsw_ax();
2992     sahf();
2993     restore_rax(tmp);
2994   }
2995   // condition codes set as follows:
2996   //
2997   // CF (corresponds to C0) if x < y
2998   // PF (corresponds to C2) if unordered
2999   // ZF (corresponds to C3) if x = y
3000 }
3001 
3002 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) {
3003   fcmp2int(dst, unordered_is_less, 1, true, true);
3004 }
3005 
3006 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) {
3007   fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right);
3008   Label L;
3009   if (unordered_is_less) {
3010     movl(dst, -1);
3011     jcc(Assembler::parity, L);
3012     jcc(Assembler::below , L);
3013     movl(dst, 0);
3014     jcc(Assembler::equal , L);
3015     increment(dst);
3016   } else { // unordered is greater
3017     movl(dst, 1);
3018     jcc(Assembler::parity, L);
3019     jcc(Assembler::above , L);
3020     movl(dst, 0);
3021     jcc(Assembler::equal , L);
3022     decrementl(dst);
3023   }
3024   bind(L);
3025 }
3026 
3027 void MacroAssembler::fld_d(AddressLiteral src) {
3028   fld_d(as_Address(src));
3029 }
3030 
3031 void MacroAssembler::fld_s(AddressLiteral src) {
3032   fld_s(as_Address(src));
3033 }
3034 
3035 void MacroAssembler::fld_x(AddressLiteral src) {
3036   Assembler::fld_x(as_Address(src));
3037 }
3038 
3039 void MacroAssembler::fldcw(AddressLiteral src) {
3040   Assembler::fldcw(as_Address(src));
3041 }
3042 
3043 void MacroAssembler::pow_exp_core_encoding() {
3044   // kills rax, rcx, rdx
3045   subptr(rsp,sizeof(jdouble));
3046   // computes 2^X. Stack: X ...
3047   // f2xm1 computes 2^X-1 but only operates on -1<=X<=1. Get int(X) and
3048   // keep it on the thread's stack to compute 2^int(X) later
3049   // then compute 2^(X-int(X)) as (2^(X-int(X)-1+1)
3050   // final result is obtained with: 2^X = 2^int(X) * 2^(X-int(X))
3051   fld_s(0);                 // Stack: X X ...
3052   frndint();                // Stack: int(X) X ...
3053   fsuba(1);                 // Stack: int(X) X-int(X) ...
3054   fistp_s(Address(rsp,0));  // move int(X) as integer to thread's stack. Stack: X-int(X) ...
3055   f2xm1();                  // Stack: 2^(X-int(X))-1 ...
3056   fld1();                   // Stack: 1 2^(X-int(X))-1 ...
3057   faddp(1);                 // Stack: 2^(X-int(X))
3058   // computes 2^(int(X)): add exponent bias (1023) to int(X), then
3059   // shift int(X)+1023 to exponent position.
3060   // Exponent is limited to 11 bits if int(X)+1023 does not fit in 11
3061   // bits, set result to NaN. 0x000 and 0x7FF are reserved exponent
3062   // values so detect them and set result to NaN.
3063   movl(rax,Address(rsp,0));
3064   movl(rcx, -2048); // 11 bit mask and valid NaN binary encoding
3065   addl(rax, 1023);
3066   movl(rdx,rax);
3067   shll(rax,20);
3068   // Check that 0 < int(X)+1023 < 2047. Otherwise set rax to NaN.
3069   addl(rdx,1);
3070   // Check that 1 < int(X)+1023+1 < 2048
3071   // in 3 steps:
3072   // 1- (int(X)+1023+1)&-2048 == 0 => 0 <= int(X)+1023+1 < 2048
3073   // 2- (int(X)+1023+1)&-2048 != 0
3074   // 3- (int(X)+1023+1)&-2048 != 1
3075   // Do 2- first because addl just updated the flags.
3076   cmov32(Assembler::equal,rax,rcx);
3077   cmpl(rdx,1);
3078   cmov32(Assembler::equal,rax,rcx);
3079   testl(rdx,rcx);
3080   cmov32(Assembler::notEqual,rax,rcx);
3081   movl(Address(rsp,4),rax);
3082   movl(Address(rsp,0),0);
3083   fmul_d(Address(rsp,0));   // Stack: 2^X ...
3084   addptr(rsp,sizeof(jdouble));
3085 }
3086 
3087 void MacroAssembler::increase_precision() {
3088   subptr(rsp, BytesPerWord);
3089   fnstcw(Address(rsp, 0));
3090   movl(rax, Address(rsp, 0));
3091   orl(rax, 0x300);
3092   push(rax);
3093   fldcw(Address(rsp, 0));
3094   pop(rax);
3095 }
3096 
3097 void MacroAssembler::restore_precision() {
3098   fldcw(Address(rsp, 0));
3099   addptr(rsp, BytesPerWord);
3100 }
3101 
3102 void MacroAssembler::fast_pow() {
3103   // computes X^Y = 2^(Y * log2(X))
3104   // if fast computation is not possible, result is NaN. Requires
3105   // fallback from user of this macro.
3106   // increase precision for intermediate steps of the computation
3107   BLOCK_COMMENT("fast_pow {");
3108   increase_precision();
3109   fyl2x();                 // Stack: (Y*log2(X)) ...
3110   pow_exp_core_encoding(); // Stack: exp(X) ...
3111   restore_precision();
3112   BLOCK_COMMENT("} fast_pow");
3113 }
3114 
3115 void MacroAssembler::fast_exp() {
3116   // computes exp(X) = 2^(X * log2(e))
3117   // if fast computation is not possible, result is NaN. Requires
3118   // fallback from user of this macro.
3119   // increase precision for intermediate steps of the computation
3120   increase_precision();
3121   fldl2e();                // Stack: log2(e) X ...
3122   fmulp(1);                // Stack: (X*log2(e)) ...
3123   pow_exp_core_encoding(); // Stack: exp(X) ...
3124   restore_precision();
3125 }
3126 
3127 void MacroAssembler::pow_or_exp(bool is_exp, int num_fpu_regs_in_use) {
3128   // kills rax, rcx, rdx
3129   // pow and exp needs 2 extra registers on the fpu stack.
3130   Label slow_case, done;
3131   Register tmp = noreg;
3132   if (!VM_Version::supports_cmov()) {
3133     // fcmp needs a temporary so preserve rdx,
3134     tmp = rdx;
3135   }
3136   Register tmp2 = rax;
3137   Register tmp3 = rcx;
3138 
3139   if (is_exp) {
3140     // Stack: X
3141     fld_s(0);                   // duplicate argument for runtime call. Stack: X X
3142     fast_exp();                 // Stack: exp(X) X
3143     fcmp(tmp, 0, false, false); // Stack: exp(X) X
3144     // exp(X) not equal to itself: exp(X) is NaN go to slow case.
3145     jcc(Assembler::parity, slow_case);
3146     // get rid of duplicate argument. Stack: exp(X)
3147     if (num_fpu_regs_in_use > 0) {
3148       fxch();
3149       fpop();
3150     } else {
3151       ffree(1);
3152     }
3153     jmp(done);
3154   } else {
3155     // Stack: X Y
3156     Label x_negative, y_not_2;
3157 
3158     static double two = 2.0;
3159     ExternalAddress two_addr((address)&two);
3160 
3161     // constant maybe too far on 64 bit
3162     lea(tmp2, two_addr);
3163     fld_d(Address(tmp2, 0));    // Stack: 2 X Y
3164     fcmp(tmp, 2, true, false);  // Stack: X Y
3165     jcc(Assembler::parity, y_not_2);
3166     jcc(Assembler::notEqual, y_not_2);
3167 
3168     fxch(); fpop();             // Stack: X
3169     fmul(0);                    // Stack: X*X
3170 
3171     jmp(done);
3172 
3173     bind(y_not_2);
3174 
3175     fldz();                     // Stack: 0 X Y
3176     fcmp(tmp, 1, true, false);  // Stack: X Y
3177     jcc(Assembler::above, x_negative);
3178 
3179     // X >= 0
3180 
3181     fld_s(1);                   // duplicate arguments for runtime call. Stack: Y X Y
3182     fld_s(1);                   // Stack: X Y X Y
3183     fast_pow();                 // Stack: X^Y X Y
3184     fcmp(tmp, 0, false, false); // Stack: X^Y X Y
3185     // X^Y not equal to itself: X^Y is NaN go to slow case.
3186     jcc(Assembler::parity, slow_case);
3187     // get rid of duplicate arguments. Stack: X^Y
3188     if (num_fpu_regs_in_use > 0) {
3189       fxch(); fpop();
3190       fxch(); fpop();
3191     } else {
3192       ffree(2);
3193       ffree(1);
3194     }
3195     jmp(done);
3196 
3197     // X <= 0
3198     bind(x_negative);
3199 
3200     fld_s(1);                   // Stack: Y X Y
3201     frndint();                  // Stack: int(Y) X Y
3202     fcmp(tmp, 2, false, false); // Stack: int(Y) X Y
3203     jcc(Assembler::notEqual, slow_case);
3204 
3205     subptr(rsp, 8);
3206 
3207     // For X^Y, when X < 0, Y has to be an integer and the final
3208     // result depends on whether it's odd or even. We just checked
3209     // that int(Y) == Y.  We move int(Y) to gp registers as a 64 bit
3210     // integer to test its parity. If int(Y) is huge and doesn't fit
3211     // in the 64 bit integer range, the integer indefinite value will
3212     // end up in the gp registers. Huge numbers are all even, the
3213     // integer indefinite number is even so it's fine.
3214 
3215 #ifdef ASSERT
3216     // Let's check we don't end up with an integer indefinite number
3217     // when not expected. First test for huge numbers: check whether
3218     // int(Y)+1 == int(Y) which is true for very large numbers and
3219     // those are all even. A 64 bit integer is guaranteed to not
3220     // overflow for numbers where y+1 != y (when precision is set to
3221     // double precision).
3222     Label y_not_huge;
3223 
3224     fld1();                     // Stack: 1 int(Y) X Y
3225     fadd(1);                    // Stack: 1+int(Y) int(Y) X Y
3226 
3227 #ifdef _LP64
3228     // trip to memory to force the precision down from double extended
3229     // precision
3230     fstp_d(Address(rsp, 0));
3231     fld_d(Address(rsp, 0));
3232 #endif
3233 
3234     fcmp(tmp, 1, true, false);  // Stack: int(Y) X Y
3235 #endif
3236 
3237     // move int(Y) as 64 bit integer to thread's stack
3238     fistp_d(Address(rsp,0));    // Stack: X Y
3239 
3240 #ifdef ASSERT
3241     jcc(Assembler::notEqual, y_not_huge);
3242 
3243     // Y is huge so we know it's even. It may not fit in a 64 bit
3244     // integer and we don't want the debug code below to see the
3245     // integer indefinite value so overwrite int(Y) on the thread's
3246     // stack with 0.
3247     movl(Address(rsp, 0), 0);
3248     movl(Address(rsp, 4), 0);
3249 
3250     bind(y_not_huge);
3251 #endif
3252 
3253     fld_s(1);                   // duplicate arguments for runtime call. Stack: Y X Y
3254     fld_s(1);                   // Stack: X Y X Y
3255     fabs();                     // Stack: abs(X) Y X Y
3256     fast_pow();                 // Stack: abs(X)^Y X Y
3257     fcmp(tmp, 0, false, false); // Stack: abs(X)^Y X Y
3258     // abs(X)^Y not equal to itself: abs(X)^Y is NaN go to slow case.
3259 
3260     pop(tmp2);
3261     NOT_LP64(pop(tmp3));
3262     jcc(Assembler::parity, slow_case);
3263 
3264 #ifdef ASSERT
3265     // Check that int(Y) is not integer indefinite value (int
3266     // overflow). Shouldn't happen because for values that would
3267     // overflow, 1+int(Y)==Y which was tested earlier.
3268 #ifndef _LP64
3269     {
3270       Label integer;
3271       testl(tmp2, tmp2);
3272       jcc(Assembler::notZero, integer);
3273       cmpl(tmp3, 0x80000000);
3274       jcc(Assembler::notZero, integer);
3275       STOP("integer indefinite value shouldn't be seen here");
3276       bind(integer);
3277     }
3278 #else
3279     {
3280       Label integer;
3281       mov(tmp3, tmp2); // preserve tmp2 for parity check below
3282       shlq(tmp3, 1);
3283       jcc(Assembler::carryClear, integer);
3284       jcc(Assembler::notZero, integer);
3285       STOP("integer indefinite value shouldn't be seen here");
3286       bind(integer);
3287     }
3288 #endif
3289 #endif
3290 
3291     // get rid of duplicate arguments. Stack: X^Y
3292     if (num_fpu_regs_in_use > 0) {
3293       fxch(); fpop();
3294       fxch(); fpop();
3295     } else {
3296       ffree(2);
3297       ffree(1);
3298     }
3299 
3300     testl(tmp2, 1);
3301     jcc(Assembler::zero, done); // X <= 0, Y even: X^Y = abs(X)^Y
3302     // X <= 0, Y even: X^Y = -abs(X)^Y
3303 
3304     fchs();                     // Stack: -abs(X)^Y Y
3305     jmp(done);
3306   }
3307 
3308   // slow case: runtime call
3309   bind(slow_case);
3310 
3311   fpop();                       // pop incorrect result or int(Y)
3312 
3313   fp_runtime_fallback(is_exp ? CAST_FROM_FN_PTR(address, SharedRuntime::dexp) : CAST_FROM_FN_PTR(address, SharedRuntime::dpow),
3314                       is_exp ? 1 : 2, num_fpu_regs_in_use);
3315 
3316   // Come here with result in F-TOS
3317   bind(done);
3318 }
3319 
3320 void MacroAssembler::fpop() {
3321   ffree();
3322   fincstp();
3323 }
3324 
3325 void MacroAssembler::fremr(Register tmp) {
3326   save_rax(tmp);
3327   { Label L;
3328     bind(L);
3329     fprem();
3330     fwait(); fnstsw_ax();
3331 #ifdef _LP64
3332     testl(rax, 0x400);
3333     jcc(Assembler::notEqual, L);
3334 #else
3335     sahf();
3336     jcc(Assembler::parity, L);
3337 #endif // _LP64
3338   }
3339   restore_rax(tmp);
3340   // Result is in ST0.
3341   // Note: fxch & fpop to get rid of ST1
3342   // (otherwise FPU stack could overflow eventually)
3343   fxch(1);
3344   fpop();
3345 }
3346 
3347 
3348 void MacroAssembler::incrementl(AddressLiteral dst) {
3349   if (reachable(dst)) {
3350     incrementl(as_Address(dst));
3351   } else {
3352     lea(rscratch1, dst);
3353     incrementl(Address(rscratch1, 0));
3354   }
3355 }
3356 
3357 void MacroAssembler::incrementl(ArrayAddress dst) {
3358   incrementl(as_Address(dst));
3359 }
3360 
3361 void MacroAssembler::incrementl(Register reg, int value) {
3362   if (value == min_jint) {addl(reg, value) ; return; }
3363   if (value <  0) { decrementl(reg, -value); return; }
3364   if (value == 0) {                        ; return; }
3365   if (value == 1 && UseIncDec) { incl(reg) ; return; }
3366   /* else */      { addl(reg, value)       ; return; }
3367 }
3368 
3369 void MacroAssembler::incrementl(Address dst, int value) {
3370   if (value == min_jint) {addl(dst, value) ; return; }
3371   if (value <  0) { decrementl(dst, -value); return; }
3372   if (value == 0) {                        ; return; }
3373   if (value == 1 && UseIncDec) { incl(dst) ; return; }
3374   /* else */      { addl(dst, value)       ; return; }
3375 }
3376 
3377 void MacroAssembler::jump(AddressLiteral dst) {
3378   if (reachable(dst)) {
3379     jmp_literal(dst.target(), dst.rspec());
3380   } else {
3381     lea(rscratch1, dst);
3382     jmp(rscratch1);
3383   }
3384 }
3385 
3386 void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) {
3387   if (reachable(dst)) {
3388     InstructionMark im(this);
3389     relocate(dst.reloc());
3390     const int short_size = 2;
3391     const int long_size = 6;
3392     int offs = (intptr_t)dst.target() - ((intptr_t)pc());
3393     if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) {
3394       // 0111 tttn #8-bit disp
3395       emit_int8(0x70 | cc);
3396       emit_int8((offs - short_size) & 0xFF);
3397     } else {
3398       // 0000 1111 1000 tttn #32-bit disp
3399       emit_int8(0x0F);
3400       emit_int8((unsigned char)(0x80 | cc));
3401       emit_int32(offs - long_size);
3402     }
3403   } else {
3404 #ifdef ASSERT
3405     warning("reversing conditional branch");
3406 #endif /* ASSERT */
3407     Label skip;
3408     jccb(reverse[cc], skip);
3409     lea(rscratch1, dst);
3410     Assembler::jmp(rscratch1);
3411     bind(skip);
3412   }
3413 }
3414 
3415 void MacroAssembler::ldmxcsr(AddressLiteral src) {
3416   if (reachable(src)) {
3417     Assembler::ldmxcsr(as_Address(src));
3418   } else {
3419     lea(rscratch1, src);
3420     Assembler::ldmxcsr(Address(rscratch1, 0));
3421   }
3422 }
3423 
3424 int MacroAssembler::load_signed_byte(Register dst, Address src) {
3425   int off;
3426   if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3427     off = offset();
3428     movsbl(dst, src); // movsxb
3429   } else {
3430     off = load_unsigned_byte(dst, src);
3431     shll(dst, 24);
3432     sarl(dst, 24);
3433   }
3434   return off;
3435 }
3436 
3437 // Note: load_signed_short used to be called load_signed_word.
3438 // Although the 'w' in x86 opcodes refers to the term "word" in the assembler
3439 // manual, which means 16 bits, that usage is found nowhere in HotSpot code.
3440 // The term "word" in HotSpot means a 32- or 64-bit machine word.
3441 int MacroAssembler::load_signed_short(Register dst, Address src) {
3442   int off;
3443   if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3444     // This is dubious to me since it seems safe to do a signed 16 => 64 bit
3445     // version but this is what 64bit has always done. This seems to imply
3446     // that users are only using 32bits worth.
3447     off = offset();
3448     movswl(dst, src); // movsxw
3449   } else {
3450     off = load_unsigned_short(dst, src);
3451     shll(dst, 16);
3452     sarl(dst, 16);
3453   }
3454   return off;
3455 }
3456 
3457 int MacroAssembler::load_unsigned_byte(Register dst, Address src) {
3458   // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3459   // and "3.9 Partial Register Penalties", p. 22).
3460   int off;
3461   if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) {
3462     off = offset();
3463     movzbl(dst, src); // movzxb
3464   } else {
3465     xorl(dst, dst);
3466     off = offset();
3467     movb(dst, src);
3468   }
3469   return off;
3470 }
3471 
3472 // Note: load_unsigned_short used to be called load_unsigned_word.
3473 int MacroAssembler::load_unsigned_short(Register dst, Address src) {
3474   // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3475   // and "3.9 Partial Register Penalties", p. 22).
3476   int off;
3477   if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) {
3478     off = offset();
3479     movzwl(dst, src); // movzxw
3480   } else {
3481     xorl(dst, dst);
3482     off = offset();
3483     movw(dst, src);
3484   }
3485   return off;
3486 }
3487 
3488 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) {
3489   switch (size_in_bytes) {
3490 #ifndef _LP64
3491   case  8:
3492     assert(dst2 != noreg, "second dest register required");
3493     movl(dst,  src);
3494     movl(dst2, src.plus_disp(BytesPerInt));
3495     break;
3496 #else
3497   case  8:  movq(dst, src); break;
3498 #endif
3499   case  4:  movl(dst, src); break;
3500   case  2:  is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break;
3501   case  1:  is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break;
3502   default:  ShouldNotReachHere();
3503   }
3504 }
3505 
3506 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) {
3507   switch (size_in_bytes) {
3508 #ifndef _LP64
3509   case  8:
3510     assert(src2 != noreg, "second source register required");
3511     movl(dst,                        src);
3512     movl(dst.plus_disp(BytesPerInt), src2);
3513     break;
3514 #else
3515   case  8:  movq(dst, src); break;
3516 #endif
3517   case  4:  movl(dst, src); break;
3518   case  2:  movw(dst, src); break;
3519   case  1:  movb(dst, src); break;
3520   default:  ShouldNotReachHere();
3521   }
3522 }
3523 
3524 void MacroAssembler::mov32(AddressLiteral dst, Register src) {
3525   if (reachable(dst)) {
3526     movl(as_Address(dst), src);
3527   } else {
3528     lea(rscratch1, dst);
3529     movl(Address(rscratch1, 0), src);
3530   }
3531 }
3532 
3533 void MacroAssembler::mov32(Register dst, AddressLiteral src) {
3534   if (reachable(src)) {
3535     movl(dst, as_Address(src));
3536   } else {
3537     lea(rscratch1, src);
3538     movl(dst, Address(rscratch1, 0));
3539   }
3540 }
3541 
3542 // C++ bool manipulation
3543 
3544 void MacroAssembler::movbool(Register dst, Address src) {
3545   if(sizeof(bool) == 1)
3546     movb(dst, src);
3547   else if(sizeof(bool) == 2)
3548     movw(dst, src);
3549   else if(sizeof(bool) == 4)
3550     movl(dst, src);
3551   else
3552     // unsupported
3553     ShouldNotReachHere();
3554 }
3555 
3556 void MacroAssembler::movbool(Address dst, bool boolconst) {
3557   if(sizeof(bool) == 1)
3558     movb(dst, (int) boolconst);
3559   else if(sizeof(bool) == 2)
3560     movw(dst, (int) boolconst);
3561   else if(sizeof(bool) == 4)
3562     movl(dst, (int) boolconst);
3563   else
3564     // unsupported
3565     ShouldNotReachHere();
3566 }
3567 
3568 void MacroAssembler::movbool(Address dst, Register src) {
3569   if(sizeof(bool) == 1)
3570     movb(dst, src);
3571   else if(sizeof(bool) == 2)
3572     movw(dst, src);
3573   else if(sizeof(bool) == 4)
3574     movl(dst, src);
3575   else
3576     // unsupported
3577     ShouldNotReachHere();
3578 }
3579 
3580 void MacroAssembler::movbyte(ArrayAddress dst, int src) {
3581   movb(as_Address(dst), src);
3582 }
3583 
3584 void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) {
3585   if (reachable(src)) {
3586     movdl(dst, as_Address(src));
3587   } else {
3588     lea(rscratch1, src);
3589     movdl(dst, Address(rscratch1, 0));
3590   }
3591 }
3592 
3593 void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) {
3594   if (reachable(src)) {
3595     movq(dst, as_Address(src));
3596   } else {
3597     lea(rscratch1, src);
3598     movq(dst, Address(rscratch1, 0));
3599   }
3600 }
3601 
3602 void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) {
3603   if (reachable(src)) {
3604     if (UseXmmLoadAndClearUpper) {
3605       movsd (dst, as_Address(src));
3606     } else {
3607       movlpd(dst, as_Address(src));
3608     }
3609   } else {
3610     lea(rscratch1, src);
3611     if (UseXmmLoadAndClearUpper) {
3612       movsd (dst, Address(rscratch1, 0));
3613     } else {
3614       movlpd(dst, Address(rscratch1, 0));
3615     }
3616   }
3617 }
3618 
3619 void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) {
3620   if (reachable(src)) {
3621     movss(dst, as_Address(src));
3622   } else {
3623     lea(rscratch1, src);
3624     movss(dst, Address(rscratch1, 0));
3625   }
3626 }
3627 
3628 void MacroAssembler::movptr(Register dst, Register src) {
3629   LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3630 }
3631 
3632 void MacroAssembler::movptr(Register dst, Address src) {
3633   LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3634 }
3635 
3636 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
3637 void MacroAssembler::movptr(Register dst, intptr_t src) {
3638   LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src));
3639 }
3640 
3641 void MacroAssembler::movptr(Address dst, Register src) {
3642   LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3643 }
3644 
3645 void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src) {
3646   if (reachable(src)) {
3647     Assembler::movdqu(dst, as_Address(src));
3648   } else {
3649     lea(rscratch1, src);
3650     Assembler::movdqu(dst, Address(rscratch1, 0));
3651   }
3652 }
3653 
3654 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) {
3655   if (reachable(src)) {
3656     Assembler::movdqa(dst, as_Address(src));
3657   } else {
3658     lea(rscratch1, src);
3659     Assembler::movdqa(dst, Address(rscratch1, 0));
3660   }
3661 }
3662 
3663 void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) {
3664   if (reachable(src)) {
3665     Assembler::movsd(dst, as_Address(src));
3666   } else {
3667     lea(rscratch1, src);
3668     Assembler::movsd(dst, Address(rscratch1, 0));
3669   }
3670 }
3671 
3672 void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) {
3673   if (reachable(src)) {
3674     Assembler::movss(dst, as_Address(src));
3675   } else {
3676     lea(rscratch1, src);
3677     Assembler::movss(dst, Address(rscratch1, 0));
3678   }
3679 }
3680 
3681 void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) {
3682   if (reachable(src)) {
3683     Assembler::mulsd(dst, as_Address(src));
3684   } else {
3685     lea(rscratch1, src);
3686     Assembler::mulsd(dst, Address(rscratch1, 0));
3687   }
3688 }
3689 
3690 void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) {
3691   if (reachable(src)) {
3692     Assembler::mulss(dst, as_Address(src));
3693   } else {
3694     lea(rscratch1, src);
3695     Assembler::mulss(dst, Address(rscratch1, 0));
3696   }
3697 }
3698 
3699 void MacroAssembler::null_check(Register reg, int offset) {
3700   if (needs_explicit_null_check(offset)) {
3701     // provoke OS NULL exception if reg = NULL by
3702     // accessing M[reg] w/o changing any (non-CC) registers
3703     // NOTE: cmpl is plenty here to provoke a segv
3704     cmpptr(rax, Address(reg, 0));
3705     // Note: should probably use testl(rax, Address(reg, 0));
3706     //       may be shorter code (however, this version of
3707     //       testl needs to be implemented first)
3708   } else {
3709     // nothing to do, (later) access of M[reg + offset]
3710     // will provoke OS NULL exception if reg = NULL
3711   }
3712 }
3713 
3714 void MacroAssembler::os_breakpoint() {
3715   // instead of directly emitting a breakpoint, call os:breakpoint for better debugability
3716   // (e.g., MSVC can't call ps() otherwise)
3717   call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint)));
3718 }
3719 
3720 void MacroAssembler::pop_CPU_state() {
3721   pop_FPU_state();
3722   pop_IU_state();
3723 }
3724 
3725 void MacroAssembler::pop_FPU_state() {
3726   NOT_LP64(frstor(Address(rsp, 0));)
3727   LP64_ONLY(fxrstor(Address(rsp, 0));)
3728   addptr(rsp, FPUStateSizeInWords * wordSize);
3729 }
3730 
3731 void MacroAssembler::pop_IU_state() {
3732   popa();
3733   LP64_ONLY(addq(rsp, 8));
3734   popf();
3735 }
3736 
3737 // Save Integer and Float state
3738 // Warning: Stack must be 16 byte aligned (64bit)
3739 void MacroAssembler::push_CPU_state() {
3740   push_IU_state();
3741   push_FPU_state();
3742 }
3743 
3744 void MacroAssembler::push_FPU_state() {
3745   subptr(rsp, FPUStateSizeInWords * wordSize);
3746 #ifndef _LP64
3747   fnsave(Address(rsp, 0));
3748   fwait();
3749 #else
3750   fxsave(Address(rsp, 0));
3751 #endif // LP64
3752 }
3753 
3754 void MacroAssembler::push_IU_state() {
3755   // Push flags first because pusha kills them
3756   pushf();
3757   // Make sure rsp stays 16-byte aligned
3758   LP64_ONLY(subq(rsp, 8));
3759   pusha();
3760 }
3761 
3762 void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp, bool clear_pc) {
3763   // determine java_thread register
3764   if (!java_thread->is_valid()) {
3765     java_thread = rdi;
3766     get_thread(java_thread);
3767   }
3768   // we must set sp to zero to clear frame
3769   movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
3770   if (clear_fp) {
3771     movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
3772   }
3773 
3774   if (clear_pc)
3775     movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
3776 
3777 }
3778 
3779 void MacroAssembler::restore_rax(Register tmp) {
3780   if (tmp == noreg) pop(rax);
3781   else if (tmp != rax) mov(rax, tmp);
3782 }
3783 
3784 void MacroAssembler::round_to(Register reg, int modulus) {
3785   addptr(reg, modulus - 1);
3786   andptr(reg, -modulus);
3787 }
3788 
3789 void MacroAssembler::save_rax(Register tmp) {
3790   if (tmp == noreg) push(rax);
3791   else if (tmp != rax) mov(tmp, rax);
3792 }
3793 
3794 // Write serialization page so VM thread can do a pseudo remote membar.
3795 // We use the current thread pointer to calculate a thread specific
3796 // offset to write to within the page. This minimizes bus traffic
3797 // due to cache line collision.
3798 void MacroAssembler::serialize_memory(Register thread, Register tmp) {
3799   movl(tmp, thread);
3800   shrl(tmp, os::get_serialize_page_shift_count());
3801   andl(tmp, (os::vm_page_size() - sizeof(int)));
3802 
3803   Address index(noreg, tmp, Address::times_1);
3804   ExternalAddress page(os::get_memory_serialize_page());
3805 
3806   // Size of store must match masking code above
3807   movl(as_Address(ArrayAddress(page, index)), tmp);
3808 }
3809 
3810 // Calls to C land
3811 //
3812 // When entering C land, the rbp, & rsp of the last Java frame have to be recorded
3813 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp
3814 // has to be reset to 0. This is required to allow proper stack traversal.
3815 void MacroAssembler::set_last_Java_frame(Register java_thread,
3816                                          Register last_java_sp,
3817                                          Register last_java_fp,
3818                                          address  last_java_pc) {
3819   // determine java_thread register
3820   if (!java_thread->is_valid()) {
3821     java_thread = rdi;
3822     get_thread(java_thread);
3823   }
3824   // determine last_java_sp register
3825   if (!last_java_sp->is_valid()) {
3826     last_java_sp = rsp;
3827   }
3828 
3829   // last_java_fp is optional
3830 
3831   if (last_java_fp->is_valid()) {
3832     movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp);
3833   }
3834 
3835   // last_java_pc is optional
3836 
3837   if (last_java_pc != NULL) {
3838     lea(Address(java_thread,
3839                  JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()),
3840         InternalAddress(last_java_pc));
3841 
3842   }
3843   movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp);
3844 }
3845 
3846 void MacroAssembler::shlptr(Register dst, int imm8) {
3847   LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8));
3848 }
3849 
3850 void MacroAssembler::shrptr(Register dst, int imm8) {
3851   LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8));
3852 }
3853 
3854 void MacroAssembler::sign_extend_byte(Register reg) {
3855   if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) {
3856     movsbl(reg, reg); // movsxb
3857   } else {
3858     shll(reg, 24);
3859     sarl(reg, 24);
3860   }
3861 }
3862 
3863 void MacroAssembler::sign_extend_short(Register reg) {
3864   if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3865     movswl(reg, reg); // movsxw
3866   } else {
3867     shll(reg, 16);
3868     sarl(reg, 16);
3869   }
3870 }
3871 
3872 void MacroAssembler::testl(Register dst, AddressLiteral src) {
3873   assert(reachable(src), "Address should be reachable");
3874   testl(dst, as_Address(src));
3875 }
3876 
3877 void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) {
3878   if (reachable(src)) {
3879     Assembler::sqrtsd(dst, as_Address(src));
3880   } else {
3881     lea(rscratch1, src);
3882     Assembler::sqrtsd(dst, Address(rscratch1, 0));
3883   }
3884 }
3885 
3886 void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) {
3887   if (reachable(src)) {
3888     Assembler::sqrtss(dst, as_Address(src));
3889   } else {
3890     lea(rscratch1, src);
3891     Assembler::sqrtss(dst, Address(rscratch1, 0));
3892   }
3893 }
3894 
3895 void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) {
3896   if (reachable(src)) {
3897     Assembler::subsd(dst, as_Address(src));
3898   } else {
3899     lea(rscratch1, src);
3900     Assembler::subsd(dst, Address(rscratch1, 0));
3901   }
3902 }
3903 
3904 void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) {
3905   if (reachable(src)) {
3906     Assembler::subss(dst, as_Address(src));
3907   } else {
3908     lea(rscratch1, src);
3909     Assembler::subss(dst, Address(rscratch1, 0));
3910   }
3911 }
3912 
3913 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) {
3914   if (reachable(src)) {
3915     Assembler::ucomisd(dst, as_Address(src));
3916   } else {
3917     lea(rscratch1, src);
3918     Assembler::ucomisd(dst, Address(rscratch1, 0));
3919   }
3920 }
3921 
3922 void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) {
3923   if (reachable(src)) {
3924     Assembler::ucomiss(dst, as_Address(src));
3925   } else {
3926     lea(rscratch1, src);
3927     Assembler::ucomiss(dst, Address(rscratch1, 0));
3928   }
3929 }
3930 
3931 void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src) {
3932   // Used in sign-bit flipping with aligned address.
3933   assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
3934   if (reachable(src)) {
3935     Assembler::xorpd(dst, as_Address(src));
3936   } else {
3937     lea(rscratch1, src);
3938     Assembler::xorpd(dst, Address(rscratch1, 0));
3939   }
3940 }
3941 
3942 void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src) {
3943   // Used in sign-bit flipping with aligned address.
3944   assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
3945   if (reachable(src)) {
3946     Assembler::xorps(dst, as_Address(src));
3947   } else {
3948     lea(rscratch1, src);
3949     Assembler::xorps(dst, Address(rscratch1, 0));
3950   }
3951 }
3952 
3953 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) {
3954   // Used in sign-bit flipping with aligned address.
3955   bool aligned_adr = (((intptr_t)src.target() & 15) == 0);
3956   assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes");
3957   if (reachable(src)) {
3958     Assembler::pshufb(dst, as_Address(src));
3959   } else {
3960     lea(rscratch1, src);
3961     Assembler::pshufb(dst, Address(rscratch1, 0));
3962   }
3963 }
3964 
3965 // AVX 3-operands instructions
3966 
3967 void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3968   if (reachable(src)) {
3969     vaddsd(dst, nds, as_Address(src));
3970   } else {
3971     lea(rscratch1, src);
3972     vaddsd(dst, nds, Address(rscratch1, 0));
3973   }
3974 }
3975 
3976 void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3977   if (reachable(src)) {
3978     vaddss(dst, nds, as_Address(src));
3979   } else {
3980     lea(rscratch1, src);
3981     vaddss(dst, nds, Address(rscratch1, 0));
3982   }
3983 }
3984 
3985 void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256) {
3986   if (reachable(src)) {
3987     vandpd(dst, nds, as_Address(src), vector256);
3988   } else {
3989     lea(rscratch1, src);
3990     vandpd(dst, nds, Address(rscratch1, 0), vector256);
3991   }
3992 }
3993 
3994 void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256) {
3995   if (reachable(src)) {
3996     vandps(dst, nds, as_Address(src), vector256);
3997   } else {
3998     lea(rscratch1, src);
3999     vandps(dst, nds, Address(rscratch1, 0), vector256);
4000   }
4001 }
4002 
4003 void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4004   if (reachable(src)) {
4005     vdivsd(dst, nds, as_Address(src));
4006   } else {
4007     lea(rscratch1, src);
4008     vdivsd(dst, nds, Address(rscratch1, 0));
4009   }
4010 }
4011 
4012 void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4013   if (reachable(src)) {
4014     vdivss(dst, nds, as_Address(src));
4015   } else {
4016     lea(rscratch1, src);
4017     vdivss(dst, nds, Address(rscratch1, 0));
4018   }
4019 }
4020 
4021 void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4022   if (reachable(src)) {
4023     vmulsd(dst, nds, as_Address(src));
4024   } else {
4025     lea(rscratch1, src);
4026     vmulsd(dst, nds, Address(rscratch1, 0));
4027   }
4028 }
4029 
4030 void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4031   if (reachable(src)) {
4032     vmulss(dst, nds, as_Address(src));
4033   } else {
4034     lea(rscratch1, src);
4035     vmulss(dst, nds, Address(rscratch1, 0));
4036   }
4037 }
4038 
4039 void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4040   if (reachable(src)) {
4041     vsubsd(dst, nds, as_Address(src));
4042   } else {
4043     lea(rscratch1, src);
4044     vsubsd(dst, nds, Address(rscratch1, 0));
4045   }
4046 }
4047 
4048 void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4049   if (reachable(src)) {
4050     vsubss(dst, nds, as_Address(src));
4051   } else {
4052     lea(rscratch1, src);
4053     vsubss(dst, nds, Address(rscratch1, 0));
4054   }
4055 }
4056 
4057 void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256) {
4058   if (reachable(src)) {
4059     vxorpd(dst, nds, as_Address(src), vector256);
4060   } else {
4061     lea(rscratch1, src);
4062     vxorpd(dst, nds, Address(rscratch1, 0), vector256);
4063   }
4064 }
4065 
4066 void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256) {
4067   if (reachable(src)) {
4068     vxorps(dst, nds, as_Address(src), vector256);
4069   } else {
4070     lea(rscratch1, src);
4071     vxorps(dst, nds, Address(rscratch1, 0), vector256);
4072   }
4073 }
4074 
4075 
4076 //////////////////////////////////////////////////////////////////////////////////
4077 #if INCLUDE_ALL_GCS
4078 
4079 void MacroAssembler::g1_write_barrier_pre(Register obj,
4080                                           Register pre_val,
4081                                           Register thread,
4082                                           Register tmp,
4083                                           bool tosca_live,
4084                                           bool expand_call) {
4085 
4086   // If expand_call is true then we expand the call_VM_leaf macro
4087   // directly to skip generating the check by
4088   // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
4089 
4090 #ifdef _LP64
4091   assert(thread == r15_thread, "must be");
4092 #endif // _LP64
4093 
4094   Label done;
4095   Label runtime;
4096 
4097   assert(pre_val != noreg, "check this code");
4098 
4099   if (obj != noreg) {
4100     assert_different_registers(obj, pre_val, tmp);
4101     assert(pre_val != rax, "check this code");
4102   }
4103 
4104   Address in_progress(thread, in_bytes(JavaThread::satb_mark_queue_offset() +
4105                                        PtrQueue::byte_offset_of_active()));
4106   Address index(thread, in_bytes(JavaThread::satb_mark_queue_offset() +
4107                                        PtrQueue::byte_offset_of_index()));
4108   Address buffer(thread, in_bytes(JavaThread::satb_mark_queue_offset() +
4109                                        PtrQueue::byte_offset_of_buf()));
4110 
4111 
4112   // Is marking active?
4113   if (in_bytes(PtrQueue::byte_width_of_active()) == 4) {
4114     cmpl(in_progress, 0);
4115   } else {
4116     assert(in_bytes(PtrQueue::byte_width_of_active()) == 1, "Assumption");
4117     cmpb(in_progress, 0);
4118   }
4119   jcc(Assembler::equal, done);
4120 
4121   // Do we need to load the previous value?
4122   if (obj != noreg) {
4123     load_heap_oop(pre_val, Address(obj, 0));
4124   }
4125 
4126   // Is the previous value null?
4127   cmpptr(pre_val, (int32_t) NULL_WORD);
4128   jcc(Assembler::equal, done);
4129 
4130   // Can we store original value in the thread's buffer?
4131   // Is index == 0?
4132   // (The index field is typed as size_t.)
4133 
4134   movptr(tmp, index);                   // tmp := *index_adr
4135   cmpptr(tmp, 0);                       // tmp == 0?
4136   jcc(Assembler::equal, runtime);       // If yes, goto runtime
4137 
4138   subptr(tmp, wordSize);                // tmp := tmp - wordSize
4139   movptr(index, tmp);                   // *index_adr := tmp
4140   addptr(tmp, buffer);                  // tmp := tmp + *buffer_adr
4141 
4142   // Record the previous value
4143   movptr(Address(tmp, 0), pre_val);
4144   jmp(done);
4145 
4146   bind(runtime);
4147   // save the live input values
4148   if(tosca_live) push(rax);
4149 
4150   if (obj != noreg && obj != rax)
4151     push(obj);
4152 
4153   if (pre_val != rax)
4154     push(pre_val);
4155 
4156   // Calling the runtime using the regular call_VM_leaf mechanism generates
4157   // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
4158   // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL.
4159   //
4160   // If we care generating the pre-barrier without a frame (e.g. in the
4161   // intrinsified Reference.get() routine) then ebp might be pointing to
4162   // the caller frame and so this check will most likely fail at runtime.
4163   //
4164   // Expanding the call directly bypasses the generation of the check.
4165   // So when we do not have have a full interpreter frame on the stack
4166   // expand_call should be passed true.
4167 
4168   NOT_LP64( push(thread); )
4169 
4170   if (expand_call) {
4171     LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); )
4172     pass_arg1(this, thread);
4173     pass_arg0(this, pre_val);
4174     MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), 2);
4175   } else {
4176     call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), pre_val, thread);
4177   }
4178 
4179   NOT_LP64( pop(thread); )
4180 
4181   // save the live input values
4182   if (pre_val != rax)
4183     pop(pre_val);
4184 
4185   if (obj != noreg && obj != rax)
4186     pop(obj);
4187 
4188   if(tosca_live) pop(rax);
4189 
4190   bind(done);
4191 }
4192 
4193 void MacroAssembler::g1_write_barrier_post(Register store_addr,
4194                                            Register new_val,
4195                                            Register thread,
4196                                            Register tmp,
4197                                            Register tmp2) {
4198 #ifdef _LP64
4199   assert(thread == r15_thread, "must be");
4200 #endif // _LP64
4201 
4202   Address queue_index(thread, in_bytes(JavaThread::dirty_card_queue_offset() +
4203                                        PtrQueue::byte_offset_of_index()));
4204   Address buffer(thread, in_bytes(JavaThread::dirty_card_queue_offset() +
4205                                        PtrQueue::byte_offset_of_buf()));
4206 
4207   CardTableModRefBS* ct =
4208     barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set());
4209   assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
4210 
4211   Label done;
4212   Label runtime;
4213 
4214   // Does store cross heap regions?
4215 
4216   movptr(tmp, store_addr);
4217   xorptr(tmp, new_val);
4218   shrptr(tmp, HeapRegion::LogOfHRGrainBytes);
4219   jcc(Assembler::equal, done);
4220 
4221   // crosses regions, storing NULL?
4222 
4223   cmpptr(new_val, (int32_t) NULL_WORD);
4224   jcc(Assembler::equal, done);
4225 
4226   // storing region crossing non-NULL, is card already dirty?
4227 
4228   const Register card_addr = tmp;
4229   const Register cardtable = tmp2;
4230 
4231   movptr(card_addr, store_addr);
4232   shrptr(card_addr, CardTableModRefBS::card_shift);
4233   // Do not use ExternalAddress to load 'byte_map_base', since 'byte_map_base' is NOT
4234   // a valid address and therefore is not properly handled by the relocation code.
4235   movptr(cardtable, (intptr_t)ct->byte_map_base);
4236   addptr(card_addr, cardtable);
4237 
4238   cmpb(Address(card_addr, 0), (int)G1SATBCardTableModRefBS::g1_young_card_val());
4239   jcc(Assembler::equal, done);
4240 
4241   membar(Assembler::Membar_mask_bits(Assembler::StoreLoad));
4242   cmpb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val());
4243   jcc(Assembler::equal, done);
4244 
4245 
4246   // storing a region crossing, non-NULL oop, card is clean.
4247   // dirty card and log.
4248 
4249   movb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val());
4250 
4251   cmpl(queue_index, 0);
4252   jcc(Assembler::equal, runtime);
4253   subl(queue_index, wordSize);
4254   movptr(tmp2, buffer);
4255 #ifdef _LP64
4256   movslq(rscratch1, queue_index);
4257   addq(tmp2, rscratch1);
4258   movq(Address(tmp2, 0), card_addr);
4259 #else
4260   addl(tmp2, queue_index);
4261   movl(Address(tmp2, 0), card_addr);
4262 #endif
4263   jmp(done);
4264 
4265   bind(runtime);
4266   // save the live input values
4267   push(store_addr);
4268   push(new_val);
4269 #ifdef _LP64
4270   call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, r15_thread);
4271 #else
4272   push(thread);
4273   call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, thread);
4274   pop(thread);
4275 #endif
4276   pop(new_val);
4277   pop(store_addr);
4278 
4279   bind(done);
4280 }
4281 
4282 #endif // INCLUDE_ALL_GCS
4283 //////////////////////////////////////////////////////////////////////////////////
4284 
4285 
4286 void MacroAssembler::store_check(Register obj) {
4287   // Does a store check for the oop in register obj. The content of
4288   // register obj is destroyed afterwards.
4289   store_check_part_1(obj);
4290   store_check_part_2(obj);
4291 }
4292 
4293 void MacroAssembler::store_check(Register obj, Address dst) {
4294   store_check(obj);
4295 }
4296 
4297 
4298 // split the store check operation so that other instructions can be scheduled inbetween
4299 void MacroAssembler::store_check_part_1(Register obj) {
4300   BarrierSet* bs = Universe::heap()->barrier_set();
4301   assert(bs->kind() == BarrierSet::CardTableModRef, "Wrong barrier set kind");
4302   shrptr(obj, CardTableModRefBS::card_shift);
4303 }
4304 
4305 void MacroAssembler::store_check_part_2(Register obj) {
4306   BarrierSet* bs = Universe::heap()->barrier_set();
4307   assert(bs->kind() == BarrierSet::CardTableModRef, "Wrong barrier set kind");
4308   CardTableModRefBS* ct = barrier_set_cast<CardTableModRefBS>(bs);
4309   assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
4310 
4311   // The calculation for byte_map_base is as follows:
4312   // byte_map_base = _byte_map - (uintptr_t(low_bound) >> card_shift);
4313   // So this essentially converts an address to a displacement and it will
4314   // never need to be relocated. On 64bit however the value may be too
4315   // large for a 32bit displacement.
4316   intptr_t disp = (intptr_t) ct->byte_map_base;
4317   if (is_simm32(disp)) {
4318     Address cardtable(noreg, obj, Address::times_1, disp);
4319     movb(cardtable, 0);
4320   } else {
4321     // By doing it as an ExternalAddress 'disp' could be converted to a rip-relative
4322     // displacement and done in a single instruction given favorable mapping and a
4323     // smarter version of as_Address. However, 'ExternalAddress' generates a relocation
4324     // entry and that entry is not properly handled by the relocation code.
4325     AddressLiteral cardtable((address)ct->byte_map_base, relocInfo::none);
4326     Address index(noreg, obj, Address::times_1);
4327     movb(as_Address(ArrayAddress(cardtable, index)), 0);
4328   }
4329 }
4330 
4331 void MacroAssembler::subptr(Register dst, int32_t imm32) {
4332   LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32));
4333 }
4334 
4335 // Force generation of a 4 byte immediate value even if it fits into 8bit
4336 void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) {
4337   LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32));
4338 }
4339 
4340 void MacroAssembler::subptr(Register dst, Register src) {
4341   LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src));
4342 }
4343 
4344 // C++ bool manipulation
4345 void MacroAssembler::testbool(Register dst) {
4346   if(sizeof(bool) == 1)
4347     testb(dst, 0xff);
4348   else if(sizeof(bool) == 2) {
4349     // testw implementation needed for two byte bools
4350     ShouldNotReachHere();
4351   } else if(sizeof(bool) == 4)
4352     testl(dst, dst);
4353   else
4354     // unsupported
4355     ShouldNotReachHere();
4356 }
4357 
4358 void MacroAssembler::testptr(Register dst, Register src) {
4359   LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src));
4360 }
4361 
4362 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes.
4363 void MacroAssembler::tlab_allocate(Register obj,
4364                                    Register var_size_in_bytes,
4365                                    int con_size_in_bytes,
4366                                    Register t1,
4367                                    Register t2,
4368                                    Label& slow_case) {
4369   assert_different_registers(obj, t1, t2);
4370   assert_different_registers(obj, var_size_in_bytes, t1);
4371   Register end = t2;
4372   Register thread = NOT_LP64(t1) LP64_ONLY(r15_thread);
4373 
4374   verify_tlab();
4375 
4376   NOT_LP64(get_thread(thread));
4377 
4378   movptr(obj, Address(thread, JavaThread::tlab_top_offset()));
4379   if (var_size_in_bytes == noreg) {
4380     lea(end, Address(obj, con_size_in_bytes));
4381   } else {
4382     lea(end, Address(obj, var_size_in_bytes, Address::times_1));
4383   }
4384   cmpptr(end, Address(thread, JavaThread::tlab_end_offset()));
4385   jcc(Assembler::above, slow_case);
4386 
4387   // update the tlab top pointer
4388   movptr(Address(thread, JavaThread::tlab_top_offset()), end);
4389 
4390   // recover var_size_in_bytes if necessary
4391   if (var_size_in_bytes == end) {
4392     subptr(var_size_in_bytes, obj);
4393   }
4394   verify_tlab();
4395 }
4396 
4397 // Preserves rbx, and rdx.
4398 Register MacroAssembler::tlab_refill(Label& retry,
4399                                      Label& try_eden,
4400                                      Label& slow_case) {
4401   Register top = rax;
4402   Register t1  = rcx;
4403   Register t2  = rsi;
4404   Register thread_reg = NOT_LP64(rdi) LP64_ONLY(r15_thread);
4405   assert_different_registers(top, thread_reg, t1, t2, /* preserve: */ rbx, rdx);
4406   Label do_refill, discard_tlab;
4407 
4408   if (!Universe::heap()->supports_inline_contig_alloc()) {
4409     // No allocation in the shared eden.
4410     jmp(slow_case);
4411   }
4412 
4413   NOT_LP64(get_thread(thread_reg));
4414 
4415   movptr(top, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
4416   movptr(t1,  Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())));
4417 
4418   // calculate amount of free space
4419   subptr(t1, top);
4420   shrptr(t1, LogHeapWordSize);
4421 
4422   // Retain tlab and allocate object in shared space if
4423   // the amount free in the tlab is too large to discard.
4424   cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset())));
4425   jcc(Assembler::lessEqual, discard_tlab);
4426 
4427   // Retain
4428   // %%% yuck as movptr...
4429   movptr(t2, (int32_t) ThreadLocalAllocBuffer::refill_waste_limit_increment());
4430   addptr(Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset())), t2);
4431   if (TLABStats) {
4432     // increment number of slow_allocations
4433     addl(Address(thread_reg, in_bytes(JavaThread::tlab_slow_allocations_offset())), 1);
4434   }
4435   jmp(try_eden);
4436 
4437   bind(discard_tlab);
4438   if (TLABStats) {
4439     // increment number of refills
4440     addl(Address(thread_reg, in_bytes(JavaThread::tlab_number_of_refills_offset())), 1);
4441     // accumulate wastage -- t1 is amount free in tlab
4442     addl(Address(thread_reg, in_bytes(JavaThread::tlab_fast_refill_waste_offset())), t1);
4443   }
4444 
4445   // if tlab is currently allocated (top or end != null) then
4446   // fill [top, end + alignment_reserve) with array object
4447   testptr(top, top);
4448   jcc(Assembler::zero, do_refill);
4449 
4450   // set up the mark word
4451   movptr(Address(top, oopDesc::mark_offset_in_bytes()), (intptr_t)markOopDesc::prototype()->copy_set_hash(0x2));
4452   // set the length to the remaining space
4453   subptr(t1, typeArrayOopDesc::header_size(T_INT));
4454   addptr(t1, (int32_t)ThreadLocalAllocBuffer::alignment_reserve());
4455   shlptr(t1, log2_intptr(HeapWordSize/sizeof(jint)));
4456   movl(Address(top, arrayOopDesc::length_offset_in_bytes()), t1);
4457   // set klass to intArrayKlass
4458   // dubious reloc why not an oop reloc?
4459   movptr(t1, ExternalAddress((address)Universe::intArrayKlassObj_addr()));
4460   // store klass last.  concurrent gcs assumes klass length is valid if
4461   // klass field is not null.
4462   store_klass(top, t1);
4463 
4464   movptr(t1, top);
4465   subptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())));
4466   incr_allocated_bytes(thread_reg, t1, 0);
4467 
4468   // refill the tlab with an eden allocation
4469   bind(do_refill);
4470   movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset())));
4471   shlptr(t1, LogHeapWordSize);
4472   // allocate new tlab, address returned in top
4473   eden_allocate(top, t1, 0, t2, slow_case);
4474 
4475   // Check that t1 was preserved in eden_allocate.
4476 #ifdef ASSERT
4477   if (UseTLAB) {
4478     Label ok;
4479     Register tsize = rsi;
4480     assert_different_registers(tsize, thread_reg, t1);
4481     push(tsize);
4482     movptr(tsize, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset())));
4483     shlptr(tsize, LogHeapWordSize);
4484     cmpptr(t1, tsize);
4485     jcc(Assembler::equal, ok);
4486     STOP("assert(t1 != tlab size)");
4487     should_not_reach_here();
4488 
4489     bind(ok);
4490     pop(tsize);
4491   }
4492 #endif
4493   movptr(Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())), top);
4494   movptr(Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())), top);
4495   addptr(top, t1);
4496   subptr(top, (int32_t)ThreadLocalAllocBuffer::alignment_reserve_in_bytes());
4497   movptr(Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())), top);
4498   verify_tlab();
4499   jmp(retry);
4500 
4501   return thread_reg; // for use by caller
4502 }
4503 
4504 void MacroAssembler::incr_allocated_bytes(Register thread,
4505                                           Register var_size_in_bytes,
4506                                           int con_size_in_bytes,
4507                                           Register t1) {
4508   if (!thread->is_valid()) {
4509 #ifdef _LP64
4510     thread = r15_thread;
4511 #else
4512     assert(t1->is_valid(), "need temp reg");
4513     thread = t1;
4514     get_thread(thread);
4515 #endif
4516   }
4517 
4518 #ifdef _LP64
4519   if (var_size_in_bytes->is_valid()) {
4520     addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes);
4521   } else {
4522     addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes);
4523   }
4524 #else
4525   if (var_size_in_bytes->is_valid()) {
4526     addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes);
4527   } else {
4528     addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes);
4529   }
4530   adcl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())+4), 0);
4531 #endif
4532 }
4533 
4534 void MacroAssembler::fp_runtime_fallback(address runtime_entry, int nb_args, int num_fpu_regs_in_use) {
4535   pusha();
4536 
4537   // if we are coming from c1, xmm registers may be live
4538   int off = 0;
4539   if (UseSSE == 1)  {
4540     subptr(rsp, sizeof(jdouble)*8);
4541     movflt(Address(rsp,off++*sizeof(jdouble)),xmm0);
4542     movflt(Address(rsp,off++*sizeof(jdouble)),xmm1);
4543     movflt(Address(rsp,off++*sizeof(jdouble)),xmm2);
4544     movflt(Address(rsp,off++*sizeof(jdouble)),xmm3);
4545     movflt(Address(rsp,off++*sizeof(jdouble)),xmm4);
4546     movflt(Address(rsp,off++*sizeof(jdouble)),xmm5);
4547     movflt(Address(rsp,off++*sizeof(jdouble)),xmm6);
4548     movflt(Address(rsp,off++*sizeof(jdouble)),xmm7);
4549   } else if (UseSSE >= 2)  {
4550 #ifdef COMPILER2
4551     if (MaxVectorSize > 16) {
4552       assert(UseAVX > 0, "256bit vectors are supported only with AVX");
4553       // Save upper half of YMM registes
4554       subptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8));
4555       vextractf128h(Address(rsp,  0),xmm0);
4556       vextractf128h(Address(rsp, 16),xmm1);
4557       vextractf128h(Address(rsp, 32),xmm2);
4558       vextractf128h(Address(rsp, 48),xmm3);
4559       vextractf128h(Address(rsp, 64),xmm4);
4560       vextractf128h(Address(rsp, 80),xmm5);
4561       vextractf128h(Address(rsp, 96),xmm6);
4562       vextractf128h(Address(rsp,112),xmm7);
4563 #ifdef _LP64
4564       vextractf128h(Address(rsp,128),xmm8);
4565       vextractf128h(Address(rsp,144),xmm9);
4566       vextractf128h(Address(rsp,160),xmm10);
4567       vextractf128h(Address(rsp,176),xmm11);
4568       vextractf128h(Address(rsp,192),xmm12);
4569       vextractf128h(Address(rsp,208),xmm13);
4570       vextractf128h(Address(rsp,224),xmm14);
4571       vextractf128h(Address(rsp,240),xmm15);
4572 #endif
4573     }
4574 #endif
4575     // Save whole 128bit (16 bytes) XMM regiters
4576     subptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8));
4577     movdqu(Address(rsp,off++*16),xmm0);
4578     movdqu(Address(rsp,off++*16),xmm1);
4579     movdqu(Address(rsp,off++*16),xmm2);
4580     movdqu(Address(rsp,off++*16),xmm3);
4581     movdqu(Address(rsp,off++*16),xmm4);
4582     movdqu(Address(rsp,off++*16),xmm5);
4583     movdqu(Address(rsp,off++*16),xmm6);
4584     movdqu(Address(rsp,off++*16),xmm7);
4585 #ifdef _LP64
4586     movdqu(Address(rsp,off++*16),xmm8);
4587     movdqu(Address(rsp,off++*16),xmm9);
4588     movdqu(Address(rsp,off++*16),xmm10);
4589     movdqu(Address(rsp,off++*16),xmm11);
4590     movdqu(Address(rsp,off++*16),xmm12);
4591     movdqu(Address(rsp,off++*16),xmm13);
4592     movdqu(Address(rsp,off++*16),xmm14);
4593     movdqu(Address(rsp,off++*16),xmm15);
4594 #endif
4595   }
4596 
4597   // Preserve registers across runtime call
4598   int incoming_argument_and_return_value_offset = -1;
4599   if (num_fpu_regs_in_use > 1) {
4600     // Must preserve all other FPU regs (could alternatively convert
4601     // SharedRuntime::dsin, dcos etc. into assembly routines known not to trash
4602     // FPU state, but can not trust C compiler)
4603     NEEDS_CLEANUP;
4604     // NOTE that in this case we also push the incoming argument(s) to
4605     // the stack and restore it later; we also use this stack slot to
4606     // hold the return value from dsin, dcos etc.
4607     for (int i = 0; i < num_fpu_regs_in_use; i++) {
4608       subptr(rsp, sizeof(jdouble));
4609       fstp_d(Address(rsp, 0));
4610     }
4611     incoming_argument_and_return_value_offset = sizeof(jdouble)*(num_fpu_regs_in_use-1);
4612     for (int i = nb_args-1; i >= 0; i--) {
4613       fld_d(Address(rsp, incoming_argument_and_return_value_offset-i*sizeof(jdouble)));
4614     }
4615   }
4616 
4617   subptr(rsp, nb_args*sizeof(jdouble));
4618   for (int i = 0; i < nb_args; i++) {
4619     fstp_d(Address(rsp, i*sizeof(jdouble)));
4620   }
4621 
4622 #ifdef _LP64
4623   if (nb_args > 0) {
4624     movdbl(xmm0, Address(rsp, 0));
4625   }
4626   if (nb_args > 1) {
4627     movdbl(xmm1, Address(rsp, sizeof(jdouble)));
4628   }
4629   assert(nb_args <= 2, "unsupported number of args");
4630 #endif // _LP64
4631 
4632   // NOTE: we must not use call_VM_leaf here because that requires a
4633   // complete interpreter frame in debug mode -- same bug as 4387334
4634   // MacroAssembler::call_VM_leaf_base is perfectly safe and will
4635   // do proper 64bit abi
4636 
4637   NEEDS_CLEANUP;
4638   // Need to add stack banging before this runtime call if it needs to
4639   // be taken; however, there is no generic stack banging routine at
4640   // the MacroAssembler level
4641 
4642   MacroAssembler::call_VM_leaf_base(runtime_entry, 0);
4643 
4644 #ifdef _LP64
4645   movsd(Address(rsp, 0), xmm0);
4646   fld_d(Address(rsp, 0));
4647 #endif // _LP64
4648   addptr(rsp, sizeof(jdouble) * nb_args);
4649   if (num_fpu_regs_in_use > 1) {
4650     // Must save return value to stack and then restore entire FPU
4651     // stack except incoming arguments
4652     fstp_d(Address(rsp, incoming_argument_and_return_value_offset));
4653     for (int i = 0; i < num_fpu_regs_in_use - nb_args; i++) {
4654       fld_d(Address(rsp, 0));
4655       addptr(rsp, sizeof(jdouble));
4656     }
4657     fld_d(Address(rsp, (nb_args-1)*sizeof(jdouble)));
4658     addptr(rsp, sizeof(jdouble) * nb_args);
4659   }
4660 
4661   off = 0;
4662   if (UseSSE == 1)  {
4663     movflt(xmm0, Address(rsp,off++*sizeof(jdouble)));
4664     movflt(xmm1, Address(rsp,off++*sizeof(jdouble)));
4665     movflt(xmm2, Address(rsp,off++*sizeof(jdouble)));
4666     movflt(xmm3, Address(rsp,off++*sizeof(jdouble)));
4667     movflt(xmm4, Address(rsp,off++*sizeof(jdouble)));
4668     movflt(xmm5, Address(rsp,off++*sizeof(jdouble)));
4669     movflt(xmm6, Address(rsp,off++*sizeof(jdouble)));
4670     movflt(xmm7, Address(rsp,off++*sizeof(jdouble)));
4671     addptr(rsp, sizeof(jdouble)*8);
4672   } else if (UseSSE >= 2)  {
4673     // Restore whole 128bit (16 bytes) XMM regiters
4674     movdqu(xmm0, Address(rsp,off++*16));
4675     movdqu(xmm1, Address(rsp,off++*16));
4676     movdqu(xmm2, Address(rsp,off++*16));
4677     movdqu(xmm3, Address(rsp,off++*16));
4678     movdqu(xmm4, Address(rsp,off++*16));
4679     movdqu(xmm5, Address(rsp,off++*16));
4680     movdqu(xmm6, Address(rsp,off++*16));
4681     movdqu(xmm7, Address(rsp,off++*16));
4682 #ifdef _LP64
4683     movdqu(xmm8, Address(rsp,off++*16));
4684     movdqu(xmm9, Address(rsp,off++*16));
4685     movdqu(xmm10, Address(rsp,off++*16));
4686     movdqu(xmm11, Address(rsp,off++*16));
4687     movdqu(xmm12, Address(rsp,off++*16));
4688     movdqu(xmm13, Address(rsp,off++*16));
4689     movdqu(xmm14, Address(rsp,off++*16));
4690     movdqu(xmm15, Address(rsp,off++*16));
4691 #endif
4692     addptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8));
4693 #ifdef COMPILER2
4694     if (MaxVectorSize > 16) {
4695       // Restore upper half of YMM registes.
4696       vinsertf128h(xmm0, Address(rsp,  0));
4697       vinsertf128h(xmm1, Address(rsp, 16));
4698       vinsertf128h(xmm2, Address(rsp, 32));
4699       vinsertf128h(xmm3, Address(rsp, 48));
4700       vinsertf128h(xmm4, Address(rsp, 64));
4701       vinsertf128h(xmm5, Address(rsp, 80));
4702       vinsertf128h(xmm6, Address(rsp, 96));
4703       vinsertf128h(xmm7, Address(rsp,112));
4704 #ifdef _LP64
4705       vinsertf128h(xmm8, Address(rsp,128));
4706       vinsertf128h(xmm9, Address(rsp,144));
4707       vinsertf128h(xmm10, Address(rsp,160));
4708       vinsertf128h(xmm11, Address(rsp,176));
4709       vinsertf128h(xmm12, Address(rsp,192));
4710       vinsertf128h(xmm13, Address(rsp,208));
4711       vinsertf128h(xmm14, Address(rsp,224));
4712       vinsertf128h(xmm15, Address(rsp,240));
4713 #endif
4714       addptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8));
4715     }
4716 #endif
4717   }
4718   popa();
4719 }
4720 
4721 static const double     pi_4 =  0.7853981633974483;
4722 
4723 void MacroAssembler::trigfunc(char trig, int num_fpu_regs_in_use) {
4724   // A hand-coded argument reduction for values in fabs(pi/4, pi/2)
4725   // was attempted in this code; unfortunately it appears that the
4726   // switch to 80-bit precision and back causes this to be
4727   // unprofitable compared with simply performing a runtime call if
4728   // the argument is out of the (-pi/4, pi/4) range.
4729 
4730   Register tmp = noreg;
4731   if (!VM_Version::supports_cmov()) {
4732     // fcmp needs a temporary so preserve rbx,
4733     tmp = rbx;
4734     push(tmp);
4735   }
4736 
4737   Label slow_case, done;
4738 
4739   ExternalAddress pi4_adr = (address)&pi_4;
4740   if (reachable(pi4_adr)) {
4741     // x ?<= pi/4
4742     fld_d(pi4_adr);
4743     fld_s(1);                // Stack:  X  PI/4  X
4744     fabs();                  // Stack: |X| PI/4  X
4745     fcmp(tmp);
4746     jcc(Assembler::above, slow_case);
4747 
4748     // fastest case: -pi/4 <= x <= pi/4
4749     switch(trig) {
4750     case 's':
4751       fsin();
4752       break;
4753     case 'c':
4754       fcos();
4755       break;
4756     case 't':
4757       ftan();
4758       break;
4759     default:
4760       assert(false, "bad intrinsic");
4761       break;
4762     }
4763     jmp(done);
4764   }
4765 
4766   // slow case: runtime call
4767   bind(slow_case);
4768 
4769   switch(trig) {
4770   case 's':
4771     {
4772       fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1, num_fpu_regs_in_use);
4773     }
4774     break;
4775   case 'c':
4776     {
4777       fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1, num_fpu_regs_in_use);
4778     }
4779     break;
4780   case 't':
4781     {
4782       fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1, num_fpu_regs_in_use);
4783     }
4784     break;
4785   default:
4786     assert(false, "bad intrinsic");
4787     break;
4788   }
4789 
4790   // Come here with result in F-TOS
4791   bind(done);
4792 
4793   if (tmp != noreg) {
4794     pop(tmp);
4795   }
4796 }
4797 
4798 
4799 // Look up the method for a megamorphic invokeinterface call.
4800 // The target method is determined by <intf_klass, itable_index>.
4801 // The receiver klass is in recv_klass.
4802 // On success, the result will be in method_result, and execution falls through.
4803 // On failure, execution transfers to the given label.
4804 void MacroAssembler::lookup_interface_method(Register recv_klass,
4805                                              Register intf_klass,
4806                                              RegisterOrConstant itable_index,
4807                                              Register method_result,
4808                                              Register scan_temp,
4809                                              Label& L_no_such_interface) {
4810   assert_different_registers(recv_klass, intf_klass, method_result, scan_temp);
4811   assert(itable_index.is_constant() || itable_index.as_register() == method_result,
4812          "caller must use same register for non-constant itable index as for method");
4813 
4814   // Compute start of first itableOffsetEntry (which is at the end of the vtable)
4815   int vtable_base = InstanceKlass::vtable_start_offset() * wordSize;
4816   int itentry_off = itableMethodEntry::method_offset_in_bytes();
4817   int scan_step   = itableOffsetEntry::size() * wordSize;
4818   int vte_size    = vtableEntry::size() * wordSize;
4819   Address::ScaleFactor times_vte_scale = Address::times_ptr;
4820   assert(vte_size == wordSize, "else adjust times_vte_scale");
4821 
4822   movl(scan_temp, Address(recv_klass, InstanceKlass::vtable_length_offset() * wordSize));
4823 
4824   // %%% Could store the aligned, prescaled offset in the klassoop.
4825   lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base));
4826   if (HeapWordsPerLong > 1) {
4827     // Round up to align_object_offset boundary
4828     // see code for InstanceKlass::start_of_itable!
4829     round_to(scan_temp, BytesPerLong);
4830   }
4831 
4832   // Adjust recv_klass by scaled itable_index, so we can free itable_index.
4833   assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
4834   lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off));
4835 
4836   // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) {
4837   //   if (scan->interface() == intf) {
4838   //     result = (klass + scan->offset() + itable_index);
4839   //   }
4840   // }
4841   Label search, found_method;
4842 
4843   for (int peel = 1; peel >= 0; peel--) {
4844     movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes()));
4845     cmpptr(intf_klass, method_result);
4846 
4847     if (peel) {
4848       jccb(Assembler::equal, found_method);
4849     } else {
4850       jccb(Assembler::notEqual, search);
4851       // (invert the test to fall through to found_method...)
4852     }
4853 
4854     if (!peel)  break;
4855 
4856     bind(search);
4857 
4858     // Check that the previous entry is non-null.  A null entry means that
4859     // the receiver class doesn't implement the interface, and wasn't the
4860     // same as when the caller was compiled.
4861     testptr(method_result, method_result);
4862     jcc(Assembler::zero, L_no_such_interface);
4863     addptr(scan_temp, scan_step);
4864   }
4865 
4866   bind(found_method);
4867 
4868   // Got a hit.
4869   movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes()));
4870   movptr(method_result, Address(recv_klass, scan_temp, Address::times_1));
4871 }
4872 
4873 
4874 // virtual method calling
4875 void MacroAssembler::lookup_virtual_method(Register recv_klass,
4876                                            RegisterOrConstant vtable_index,
4877                                            Register method_result) {
4878   const int base = InstanceKlass::vtable_start_offset() * wordSize;
4879   assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below");
4880   Address vtable_entry_addr(recv_klass,
4881                             vtable_index, Address::times_ptr,
4882                             base + vtableEntry::method_offset_in_bytes());
4883   movptr(method_result, vtable_entry_addr);
4884 }
4885 
4886 
4887 void MacroAssembler::check_klass_subtype(Register sub_klass,
4888                            Register super_klass,
4889                            Register temp_reg,
4890                            Label& L_success) {
4891   Label L_failure;
4892   check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg,        &L_success, &L_failure, NULL);
4893   check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL);
4894   bind(L_failure);
4895 }
4896 
4897 
4898 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
4899                                                    Register super_klass,
4900                                                    Register temp_reg,
4901                                                    Label* L_success,
4902                                                    Label* L_failure,
4903                                                    Label* L_slow_path,
4904                                         RegisterOrConstant super_check_offset) {
4905   assert_different_registers(sub_klass, super_klass, temp_reg);
4906   bool must_load_sco = (super_check_offset.constant_or_zero() == -1);
4907   if (super_check_offset.is_register()) {
4908     assert_different_registers(sub_klass, super_klass,
4909                                super_check_offset.as_register());
4910   } else if (must_load_sco) {
4911     assert(temp_reg != noreg, "supply either a temp or a register offset");
4912   }
4913 
4914   Label L_fallthrough;
4915   int label_nulls = 0;
4916   if (L_success == NULL)   { L_success   = &L_fallthrough; label_nulls++; }
4917   if (L_failure == NULL)   { L_failure   = &L_fallthrough; label_nulls++; }
4918   if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; }
4919   assert(label_nulls <= 1, "at most one NULL in the batch");
4920 
4921   int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
4922   int sco_offset = in_bytes(Klass::super_check_offset_offset());
4923   Address super_check_offset_addr(super_klass, sco_offset);
4924 
4925   // Hacked jcc, which "knows" that L_fallthrough, at least, is in
4926   // range of a jccb.  If this routine grows larger, reconsider at
4927   // least some of these.
4928 #define local_jcc(assembler_cond, label)                                \
4929   if (&(label) == &L_fallthrough)  jccb(assembler_cond, label);         \
4930   else                             jcc( assembler_cond, label) /*omit semi*/
4931 
4932   // Hacked jmp, which may only be used just before L_fallthrough.
4933 #define final_jmp(label)                                                \
4934   if (&(label) == &L_fallthrough) { /*do nothing*/ }                    \
4935   else                            jmp(label)                /*omit semi*/
4936 
4937   // If the pointers are equal, we are done (e.g., String[] elements).
4938   // This self-check enables sharing of secondary supertype arrays among
4939   // non-primary types such as array-of-interface.  Otherwise, each such
4940   // type would need its own customized SSA.
4941   // We move this check to the front of the fast path because many
4942   // type checks are in fact trivially successful in this manner,
4943   // so we get a nicely predicted branch right at the start of the check.
4944   cmpptr(sub_klass, super_klass);
4945   local_jcc(Assembler::equal, *L_success);
4946 
4947   // Check the supertype display:
4948   if (must_load_sco) {
4949     // Positive movl does right thing on LP64.
4950     movl(temp_reg, super_check_offset_addr);
4951     super_check_offset = RegisterOrConstant(temp_reg);
4952   }
4953   Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0);
4954   cmpptr(super_klass, super_check_addr); // load displayed supertype
4955 
4956   // This check has worked decisively for primary supers.
4957   // Secondary supers are sought in the super_cache ('super_cache_addr').
4958   // (Secondary supers are interfaces and very deeply nested subtypes.)
4959   // This works in the same check above because of a tricky aliasing
4960   // between the super_cache and the primary super display elements.
4961   // (The 'super_check_addr' can address either, as the case requires.)
4962   // Note that the cache is updated below if it does not help us find
4963   // what we need immediately.
4964   // So if it was a primary super, we can just fail immediately.
4965   // Otherwise, it's the slow path for us (no success at this point).
4966 
4967   if (super_check_offset.is_register()) {
4968     local_jcc(Assembler::equal, *L_success);
4969     cmpl(super_check_offset.as_register(), sc_offset);
4970     if (L_failure == &L_fallthrough) {
4971       local_jcc(Assembler::equal, *L_slow_path);
4972     } else {
4973       local_jcc(Assembler::notEqual, *L_failure);
4974       final_jmp(*L_slow_path);
4975     }
4976   } else if (super_check_offset.as_constant() == sc_offset) {
4977     // Need a slow path; fast failure is impossible.
4978     if (L_slow_path == &L_fallthrough) {
4979       local_jcc(Assembler::equal, *L_success);
4980     } else {
4981       local_jcc(Assembler::notEqual, *L_slow_path);
4982       final_jmp(*L_success);
4983     }
4984   } else {
4985     // No slow path; it's a fast decision.
4986     if (L_failure == &L_fallthrough) {
4987       local_jcc(Assembler::equal, *L_success);
4988     } else {
4989       local_jcc(Assembler::notEqual, *L_failure);
4990       final_jmp(*L_success);
4991     }
4992   }
4993 
4994   bind(L_fallthrough);
4995 
4996 #undef local_jcc
4997 #undef final_jmp
4998 }
4999 
5000 
5001 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
5002                                                    Register super_klass,
5003                                                    Register temp_reg,
5004                                                    Register temp2_reg,
5005                                                    Label* L_success,
5006                                                    Label* L_failure,
5007                                                    bool set_cond_codes) {
5008   assert_different_registers(sub_klass, super_klass, temp_reg);
5009   if (temp2_reg != noreg)
5010     assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg);
5011 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg)
5012 
5013   Label L_fallthrough;
5014   int label_nulls = 0;
5015   if (L_success == NULL)   { L_success   = &L_fallthrough; label_nulls++; }
5016   if (L_failure == NULL)   { L_failure   = &L_fallthrough; label_nulls++; }
5017   assert(label_nulls <= 1, "at most one NULL in the batch");
5018 
5019   // a couple of useful fields in sub_klass:
5020   int ss_offset = in_bytes(Klass::secondary_supers_offset());
5021   int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
5022   Address secondary_supers_addr(sub_klass, ss_offset);
5023   Address super_cache_addr(     sub_klass, sc_offset);
5024 
5025   // Do a linear scan of the secondary super-klass chain.
5026   // This code is rarely used, so simplicity is a virtue here.
5027   // The repne_scan instruction uses fixed registers, which we must spill.
5028   // Don't worry too much about pre-existing connections with the input regs.
5029 
5030   assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super)
5031   assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter)
5032 
5033   // Get super_klass value into rax (even if it was in rdi or rcx).
5034   bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false;
5035   if (super_klass != rax || UseCompressedOops) {
5036     if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; }
5037     mov(rax, super_klass);
5038   }
5039   if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; }
5040   if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; }
5041 
5042 #ifndef PRODUCT
5043   int* pst_counter = &SharedRuntime::_partial_subtype_ctr;
5044   ExternalAddress pst_counter_addr((address) pst_counter);
5045   NOT_LP64(  incrementl(pst_counter_addr) );
5046   LP64_ONLY( lea(rcx, pst_counter_addr) );
5047   LP64_ONLY( incrementl(Address(rcx, 0)) );
5048 #endif //PRODUCT
5049 
5050   // We will consult the secondary-super array.
5051   movptr(rdi, secondary_supers_addr);
5052   // Load the array length.  (Positive movl does right thing on LP64.)
5053   movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes()));
5054   // Skip to start of data.
5055   addptr(rdi, Array<Klass*>::base_offset_in_bytes());
5056 
5057   // Scan RCX words at [RDI] for an occurrence of RAX.
5058   // Set NZ/Z based on last compare.
5059   // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does
5060   // not change flags (only scas instruction which is repeated sets flags).
5061   // Set Z = 0 (not equal) before 'repne' to indicate that class was not found.
5062 
5063     testptr(rax,rax); // Set Z = 0
5064     repne_scan();
5065 
5066   // Unspill the temp. registers:
5067   if (pushed_rdi)  pop(rdi);
5068   if (pushed_rcx)  pop(rcx);
5069   if (pushed_rax)  pop(rax);
5070 
5071   if (set_cond_codes) {
5072     // Special hack for the AD files:  rdi is guaranteed non-zero.
5073     assert(!pushed_rdi, "rdi must be left non-NULL");
5074     // Also, the condition codes are properly set Z/NZ on succeed/failure.
5075   }
5076 
5077   if (L_failure == &L_fallthrough)
5078         jccb(Assembler::notEqual, *L_failure);
5079   else  jcc(Assembler::notEqual, *L_failure);
5080 
5081   // Success.  Cache the super we found and proceed in triumph.
5082   movptr(super_cache_addr, super_klass);
5083 
5084   if (L_success != &L_fallthrough) {
5085     jmp(*L_success);
5086   }
5087 
5088 #undef IS_A_TEMP
5089 
5090   bind(L_fallthrough);
5091 }
5092 
5093 
5094 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) {
5095   if (VM_Version::supports_cmov()) {
5096     cmovl(cc, dst, src);
5097   } else {
5098     Label L;
5099     jccb(negate_condition(cc), L);
5100     movl(dst, src);
5101     bind(L);
5102   }
5103 }
5104 
5105 void MacroAssembler::cmov32(Condition cc, Register dst, Register src) {
5106   if (VM_Version::supports_cmov()) {
5107     cmovl(cc, dst, src);
5108   } else {
5109     Label L;
5110     jccb(negate_condition(cc), L);
5111     movl(dst, src);
5112     bind(L);
5113   }
5114 }
5115 
5116 void MacroAssembler::verify_oop(Register reg, const char* s) {
5117   if (!VerifyOops) return;
5118 
5119   // Pass register number to verify_oop_subroutine
5120   const char* b = NULL;
5121   {
5122     ResourceMark rm;
5123     stringStream ss;
5124     ss.print("verify_oop: %s: %s", reg->name(), s);
5125     b = code_string(ss.as_string());
5126   }
5127   BLOCK_COMMENT("verify_oop {");
5128 #ifdef _LP64
5129   push(rscratch1);                    // save r10, trashed by movptr()
5130 #endif
5131   push(rax);                          // save rax,
5132   push(reg);                          // pass register argument
5133   ExternalAddress buffer((address) b);
5134   // avoid using pushptr, as it modifies scratch registers
5135   // and our contract is not to modify anything
5136   movptr(rax, buffer.addr());
5137   push(rax);
5138   // call indirectly to solve generation ordering problem
5139   movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
5140   call(rax);
5141   // Caller pops the arguments (oop, message) and restores rax, r10
5142   BLOCK_COMMENT("} verify_oop");
5143 }
5144 
5145 
5146 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr,
5147                                                       Register tmp,
5148                                                       int offset) {
5149   intptr_t value = *delayed_value_addr;
5150   if (value != 0)
5151     return RegisterOrConstant(value + offset);
5152 
5153   // load indirectly to solve generation ordering problem
5154   movptr(tmp, ExternalAddress((address) delayed_value_addr));
5155 
5156 #ifdef ASSERT
5157   { Label L;
5158     testptr(tmp, tmp);
5159     if (WizardMode) {
5160       const char* buf = NULL;
5161       {
5162         ResourceMark rm;
5163         stringStream ss;
5164         ss.print("DelayedValue="INTPTR_FORMAT, delayed_value_addr[1]);
5165         buf = code_string(ss.as_string());
5166       }
5167       jcc(Assembler::notZero, L);
5168       STOP(buf);
5169     } else {
5170       jccb(Assembler::notZero, L);
5171       hlt();
5172     }
5173     bind(L);
5174   }
5175 #endif
5176 
5177   if (offset != 0)
5178     addptr(tmp, offset);
5179 
5180   return RegisterOrConstant(tmp);
5181 }
5182 
5183 
5184 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot,
5185                                          int extra_slot_offset) {
5186   // cf. TemplateTable::prepare_invoke(), if (load_receiver).
5187   int stackElementSize = Interpreter::stackElementSize;
5188   int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0);
5189 #ifdef ASSERT
5190   int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1);
5191   assert(offset1 - offset == stackElementSize, "correct arithmetic");
5192 #endif
5193   Register             scale_reg    = noreg;
5194   Address::ScaleFactor scale_factor = Address::no_scale;
5195   if (arg_slot.is_constant()) {
5196     offset += arg_slot.as_constant() * stackElementSize;
5197   } else {
5198     scale_reg    = arg_slot.as_register();
5199     scale_factor = Address::times(stackElementSize);
5200   }
5201   offset += wordSize;           // return PC is on stack
5202   return Address(rsp, scale_reg, scale_factor, offset);
5203 }
5204 
5205 
5206 void MacroAssembler::verify_oop_addr(Address addr, const char* s) {
5207   if (!VerifyOops) return;
5208 
5209   // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord);
5210   // Pass register number to verify_oop_subroutine
5211   const char* b = NULL;
5212   {
5213     ResourceMark rm;
5214     stringStream ss;
5215     ss.print("verify_oop_addr: %s", s);
5216     b = code_string(ss.as_string());
5217   }
5218 #ifdef _LP64
5219   push(rscratch1);                    // save r10, trashed by movptr()
5220 #endif
5221   push(rax);                          // save rax,
5222   // addr may contain rsp so we will have to adjust it based on the push
5223   // we just did (and on 64 bit we do two pushes)
5224   // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which
5225   // stores rax into addr which is backwards of what was intended.
5226   if (addr.uses(rsp)) {
5227     lea(rax, addr);
5228     pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord));
5229   } else {
5230     pushptr(addr);
5231   }
5232 
5233   ExternalAddress buffer((address) b);
5234   // pass msg argument
5235   // avoid using pushptr, as it modifies scratch registers
5236   // and our contract is not to modify anything
5237   movptr(rax, buffer.addr());
5238   push(rax);
5239 
5240   // call indirectly to solve generation ordering problem
5241   movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
5242   call(rax);
5243   // Caller pops the arguments (addr, message) and restores rax, r10.
5244 }
5245 
5246 void MacroAssembler::verify_tlab() {
5247 #ifdef ASSERT
5248   if (UseTLAB && VerifyOops) {
5249     Label next, ok;
5250     Register t1 = rsi;
5251     Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread);
5252 
5253     push(t1);
5254     NOT_LP64(push(thread_reg));
5255     NOT_LP64(get_thread(thread_reg));
5256 
5257     movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
5258     cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())));
5259     jcc(Assembler::aboveEqual, next);
5260     STOP("assert(top >= start)");
5261     should_not_reach_here();
5262 
5263     bind(next);
5264     movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())));
5265     cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
5266     jcc(Assembler::aboveEqual, ok);
5267     STOP("assert(top <= end)");
5268     should_not_reach_here();
5269 
5270     bind(ok);
5271     NOT_LP64(pop(thread_reg));
5272     pop(t1);
5273   }
5274 #endif
5275 }
5276 
5277 class ControlWord {
5278  public:
5279   int32_t _value;
5280 
5281   int  rounding_control() const        { return  (_value >> 10) & 3      ; }
5282   int  precision_control() const       { return  (_value >>  8) & 3      ; }
5283   bool precision() const               { return ((_value >>  5) & 1) != 0; }
5284   bool underflow() const               { return ((_value >>  4) & 1) != 0; }
5285   bool overflow() const                { return ((_value >>  3) & 1) != 0; }
5286   bool zero_divide() const             { return ((_value >>  2) & 1) != 0; }
5287   bool denormalized() const            { return ((_value >>  1) & 1) != 0; }
5288   bool invalid() const                 { return ((_value >>  0) & 1) != 0; }
5289 
5290   void print() const {
5291     // rounding control
5292     const char* rc;
5293     switch (rounding_control()) {
5294       case 0: rc = "round near"; break;
5295       case 1: rc = "round down"; break;
5296       case 2: rc = "round up  "; break;
5297       case 3: rc = "chop      "; break;
5298     };
5299     // precision control
5300     const char* pc;
5301     switch (precision_control()) {
5302       case 0: pc = "24 bits "; break;
5303       case 1: pc = "reserved"; break;
5304       case 2: pc = "53 bits "; break;
5305       case 3: pc = "64 bits "; break;
5306     };
5307     // flags
5308     char f[9];
5309     f[0] = ' ';
5310     f[1] = ' ';
5311     f[2] = (precision   ()) ? 'P' : 'p';
5312     f[3] = (underflow   ()) ? 'U' : 'u';
5313     f[4] = (overflow    ()) ? 'O' : 'o';
5314     f[5] = (zero_divide ()) ? 'Z' : 'z';
5315     f[6] = (denormalized()) ? 'D' : 'd';
5316     f[7] = (invalid     ()) ? 'I' : 'i';
5317     f[8] = '\x0';
5318     // output
5319     printf("%04x  masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc);
5320   }
5321 
5322 };
5323 
5324 class StatusWord {
5325  public:
5326   int32_t _value;
5327 
5328   bool busy() const                    { return ((_value >> 15) & 1) != 0; }
5329   bool C3() const                      { return ((_value >> 14) & 1) != 0; }
5330   bool C2() const                      { return ((_value >> 10) & 1) != 0; }
5331   bool C1() const                      { return ((_value >>  9) & 1) != 0; }
5332   bool C0() const                      { return ((_value >>  8) & 1) != 0; }
5333   int  top() const                     { return  (_value >> 11) & 7      ; }
5334   bool error_status() const            { return ((_value >>  7) & 1) != 0; }
5335   bool stack_fault() const             { return ((_value >>  6) & 1) != 0; }
5336   bool precision() const               { return ((_value >>  5) & 1) != 0; }
5337   bool underflow() const               { return ((_value >>  4) & 1) != 0; }
5338   bool overflow() const                { return ((_value >>  3) & 1) != 0; }
5339   bool zero_divide() const             { return ((_value >>  2) & 1) != 0; }
5340   bool denormalized() const            { return ((_value >>  1) & 1) != 0; }
5341   bool invalid() const                 { return ((_value >>  0) & 1) != 0; }
5342 
5343   void print() const {
5344     // condition codes
5345     char c[5];
5346     c[0] = (C3()) ? '3' : '-';
5347     c[1] = (C2()) ? '2' : '-';
5348     c[2] = (C1()) ? '1' : '-';
5349     c[3] = (C0()) ? '0' : '-';
5350     c[4] = '\x0';
5351     // flags
5352     char f[9];
5353     f[0] = (error_status()) ? 'E' : '-';
5354     f[1] = (stack_fault ()) ? 'S' : '-';
5355     f[2] = (precision   ()) ? 'P' : '-';
5356     f[3] = (underflow   ()) ? 'U' : '-';
5357     f[4] = (overflow    ()) ? 'O' : '-';
5358     f[5] = (zero_divide ()) ? 'Z' : '-';
5359     f[6] = (denormalized()) ? 'D' : '-';
5360     f[7] = (invalid     ()) ? 'I' : '-';
5361     f[8] = '\x0';
5362     // output
5363     printf("%04x  flags = %s, cc =  %s, top = %d", _value & 0xFFFF, f, c, top());
5364   }
5365 
5366 };
5367 
5368 class TagWord {
5369  public:
5370   int32_t _value;
5371 
5372   int tag_at(int i) const              { return (_value >> (i*2)) & 3; }
5373 
5374   void print() const {
5375     printf("%04x", _value & 0xFFFF);
5376   }
5377 
5378 };
5379 
5380 class FPU_Register {
5381  public:
5382   int32_t _m0;
5383   int32_t _m1;
5384   int16_t _ex;
5385 
5386   bool is_indefinite() const           {
5387     return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0;
5388   }
5389 
5390   void print() const {
5391     char  sign = (_ex < 0) ? '-' : '+';
5392     const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : "   ";
5393     printf("%c%04hx.%08x%08x  %s", sign, _ex, _m1, _m0, kind);
5394   };
5395 
5396 };
5397 
5398 class FPU_State {
5399  public:
5400   enum {
5401     register_size       = 10,
5402     number_of_registers =  8,
5403     register_mask       =  7
5404   };
5405 
5406   ControlWord  _control_word;
5407   StatusWord   _status_word;
5408   TagWord      _tag_word;
5409   int32_t      _error_offset;
5410   int32_t      _error_selector;
5411   int32_t      _data_offset;
5412   int32_t      _data_selector;
5413   int8_t       _register[register_size * number_of_registers];
5414 
5415   int tag_for_st(int i) const          { return _tag_word.tag_at((_status_word.top() + i) & register_mask); }
5416   FPU_Register* st(int i) const        { return (FPU_Register*)&_register[register_size * i]; }
5417 
5418   const char* tag_as_string(int tag) const {
5419     switch (tag) {
5420       case 0: return "valid";
5421       case 1: return "zero";
5422       case 2: return "special";
5423       case 3: return "empty";
5424     }
5425     ShouldNotReachHere();
5426     return NULL;
5427   }
5428 
5429   void print() const {
5430     // print computation registers
5431     { int t = _status_word.top();
5432       for (int i = 0; i < number_of_registers; i++) {
5433         int j = (i - t) & register_mask;
5434         printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j);
5435         st(j)->print();
5436         printf(" %s\n", tag_as_string(_tag_word.tag_at(i)));
5437       }
5438     }
5439     printf("\n");
5440     // print control registers
5441     printf("ctrl = "); _control_word.print(); printf("\n");
5442     printf("stat = "); _status_word .print(); printf("\n");
5443     printf("tags = "); _tag_word    .print(); printf("\n");
5444   }
5445 
5446 };
5447 
5448 class Flag_Register {
5449  public:
5450   int32_t _value;
5451 
5452   bool overflow() const                { return ((_value >> 11) & 1) != 0; }
5453   bool direction() const               { return ((_value >> 10) & 1) != 0; }
5454   bool sign() const                    { return ((_value >>  7) & 1) != 0; }
5455   bool zero() const                    { return ((_value >>  6) & 1) != 0; }
5456   bool auxiliary_carry() const         { return ((_value >>  4) & 1) != 0; }
5457   bool parity() const                  { return ((_value >>  2) & 1) != 0; }
5458   bool carry() const                   { return ((_value >>  0) & 1) != 0; }
5459 
5460   void print() const {
5461     // flags
5462     char f[8];
5463     f[0] = (overflow       ()) ? 'O' : '-';
5464     f[1] = (direction      ()) ? 'D' : '-';
5465     f[2] = (sign           ()) ? 'S' : '-';
5466     f[3] = (zero           ()) ? 'Z' : '-';
5467     f[4] = (auxiliary_carry()) ? 'A' : '-';
5468     f[5] = (parity         ()) ? 'P' : '-';
5469     f[6] = (carry          ()) ? 'C' : '-';
5470     f[7] = '\x0';
5471     // output
5472     printf("%08x  flags = %s", _value, f);
5473   }
5474 
5475 };
5476 
5477 class IU_Register {
5478  public:
5479   int32_t _value;
5480 
5481   void print() const {
5482     printf("%08x  %11d", _value, _value);
5483   }
5484 
5485 };
5486 
5487 class IU_State {
5488  public:
5489   Flag_Register _eflags;
5490   IU_Register   _rdi;
5491   IU_Register   _rsi;
5492   IU_Register   _rbp;
5493   IU_Register   _rsp;
5494   IU_Register   _rbx;
5495   IU_Register   _rdx;
5496   IU_Register   _rcx;
5497   IU_Register   _rax;
5498 
5499   void print() const {
5500     // computation registers
5501     printf("rax,  = "); _rax.print(); printf("\n");
5502     printf("rbx,  = "); _rbx.print(); printf("\n");
5503     printf("rcx  = "); _rcx.print(); printf("\n");
5504     printf("rdx  = "); _rdx.print(); printf("\n");
5505     printf("rdi  = "); _rdi.print(); printf("\n");
5506     printf("rsi  = "); _rsi.print(); printf("\n");
5507     printf("rbp,  = "); _rbp.print(); printf("\n");
5508     printf("rsp  = "); _rsp.print(); printf("\n");
5509     printf("\n");
5510     // control registers
5511     printf("flgs = "); _eflags.print(); printf("\n");
5512   }
5513 };
5514 
5515 
5516 class CPU_State {
5517  public:
5518   FPU_State _fpu_state;
5519   IU_State  _iu_state;
5520 
5521   void print() const {
5522     printf("--------------------------------------------------\n");
5523     _iu_state .print();
5524     printf("\n");
5525     _fpu_state.print();
5526     printf("--------------------------------------------------\n");
5527   }
5528 
5529 };
5530 
5531 
5532 static void _print_CPU_state(CPU_State* state) {
5533   state->print();
5534 };
5535 
5536 
5537 void MacroAssembler::print_CPU_state() {
5538   push_CPU_state();
5539   push(rsp);                // pass CPU state
5540   call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state)));
5541   addptr(rsp, wordSize);       // discard argument
5542   pop_CPU_state();
5543 }
5544 
5545 
5546 static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) {
5547   static int counter = 0;
5548   FPU_State* fs = &state->_fpu_state;
5549   counter++;
5550   // For leaf calls, only verify that the top few elements remain empty.
5551   // We only need 1 empty at the top for C2 code.
5552   if( stack_depth < 0 ) {
5553     if( fs->tag_for_st(7) != 3 ) {
5554       printf("FPR7 not empty\n");
5555       state->print();
5556       assert(false, "error");
5557       return false;
5558     }
5559     return true;                // All other stack states do not matter
5560   }
5561 
5562   assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std,
5563          "bad FPU control word");
5564 
5565   // compute stack depth
5566   int i = 0;
5567   while (i < FPU_State::number_of_registers && fs->tag_for_st(i)  < 3) i++;
5568   int d = i;
5569   while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++;
5570   // verify findings
5571   if (i != FPU_State::number_of_registers) {
5572     // stack not contiguous
5573     printf("%s: stack not contiguous at ST%d\n", s, i);
5574     state->print();
5575     assert(false, "error");
5576     return false;
5577   }
5578   // check if computed stack depth corresponds to expected stack depth
5579   if (stack_depth < 0) {
5580     // expected stack depth is -stack_depth or less
5581     if (d > -stack_depth) {
5582       // too many elements on the stack
5583       printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d);
5584       state->print();
5585       assert(false, "error");
5586       return false;
5587     }
5588   } else {
5589     // expected stack depth is stack_depth
5590     if (d != stack_depth) {
5591       // wrong stack depth
5592       printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d);
5593       state->print();
5594       assert(false, "error");
5595       return false;
5596     }
5597   }
5598   // everything is cool
5599   return true;
5600 }
5601 
5602 
5603 void MacroAssembler::verify_FPU(int stack_depth, const char* s) {
5604   if (!VerifyFPU) return;
5605   push_CPU_state();
5606   push(rsp);                // pass CPU state
5607   ExternalAddress msg((address) s);
5608   // pass message string s
5609   pushptr(msg.addr());
5610   push(stack_depth);        // pass stack depth
5611   call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU)));
5612   addptr(rsp, 3 * wordSize);   // discard arguments
5613   // check for error
5614   { Label L;
5615     testl(rax, rax);
5616     jcc(Assembler::notZero, L);
5617     int3();                  // break if error condition
5618     bind(L);
5619   }
5620   pop_CPU_state();
5621 }
5622 
5623 void MacroAssembler::restore_cpu_control_state_after_jni() {
5624   // Either restore the MXCSR register after returning from the JNI Call
5625   // or verify that it wasn't changed (with -Xcheck:jni flag).
5626   if (VM_Version::supports_sse()) {
5627     if (RestoreMXCSROnJNICalls) {
5628       ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std()));
5629     } else if (CheckJNICalls) {
5630       call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry()));
5631     }
5632   }
5633   if (VM_Version::supports_avx()) {
5634     // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty.
5635     vzeroupper();
5636   }
5637 
5638 #ifndef _LP64
5639   // Either restore the x87 floating pointer control word after returning
5640   // from the JNI call or verify that it wasn't changed.
5641   if (CheckJNICalls) {
5642     call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry()));
5643   }
5644 #endif // _LP64
5645 }
5646 
5647 
5648 void MacroAssembler::load_klass(Register dst, Register src) {
5649 #ifdef _LP64
5650   if (UseCompressedClassPointers) {
5651     movl(dst, Address(src, oopDesc::klass_offset_in_bytes()));
5652     decode_klass_not_null(dst);
5653   } else
5654 #endif
5655     movptr(dst, Address(src, oopDesc::klass_offset_in_bytes()));
5656 }
5657 
5658 void MacroAssembler::load_prototype_header(Register dst, Register src) {
5659   load_klass(dst, src);
5660   movptr(dst, Address(dst, Klass::prototype_header_offset()));
5661 }
5662 
5663 void MacroAssembler::store_klass(Register dst, Register src) {
5664 #ifdef _LP64
5665   if (UseCompressedClassPointers) {
5666     encode_klass_not_null(src);
5667     movl(Address(dst, oopDesc::klass_offset_in_bytes()), src);
5668   } else
5669 #endif
5670     movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src);
5671 }
5672 
5673 void MacroAssembler::load_heap_oop(Register dst, Address src) {
5674 #ifdef _LP64
5675   // FIXME: Must change all places where we try to load the klass.
5676   if (UseCompressedOops) {
5677     movl(dst, src);
5678     decode_heap_oop(dst);
5679   } else
5680 #endif
5681     movptr(dst, src);
5682 }
5683 
5684 // Doesn't do verfication, generates fixed size code
5685 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src) {
5686 #ifdef _LP64
5687   if (UseCompressedOops) {
5688     movl(dst, src);
5689     decode_heap_oop_not_null(dst);
5690   } else
5691 #endif
5692     movptr(dst, src);
5693 }
5694 
5695 void MacroAssembler::store_heap_oop(Address dst, Register src) {
5696 #ifdef _LP64
5697   if (UseCompressedOops) {
5698     assert(!dst.uses(src), "not enough registers");
5699     encode_heap_oop(src);
5700     movl(dst, src);
5701   } else
5702 #endif
5703     movptr(dst, src);
5704 }
5705 
5706 void MacroAssembler::cmp_heap_oop(Register src1, Address src2, Register tmp) {
5707   assert_different_registers(src1, tmp);
5708 #ifdef _LP64
5709   if (UseCompressedOops) {
5710     bool did_push = false;
5711     if (tmp == noreg) {
5712       tmp = rax;
5713       push(tmp);
5714       did_push = true;
5715       assert(!src2.uses(rsp), "can't push");
5716     }
5717     load_heap_oop(tmp, src2);
5718     cmpptr(src1, tmp);
5719     if (did_push)  pop(tmp);
5720   } else
5721 #endif
5722     cmpptr(src1, src2);
5723 }
5724 
5725 // Used for storing NULLs.
5726 void MacroAssembler::store_heap_oop_null(Address dst) {
5727 #ifdef _LP64
5728   if (UseCompressedOops) {
5729     movl(dst, (int32_t)NULL_WORD);
5730   } else {
5731     movslq(dst, (int32_t)NULL_WORD);
5732   }
5733 #else
5734   movl(dst, (int32_t)NULL_WORD);
5735 #endif
5736 }
5737 
5738 #ifdef _LP64
5739 void MacroAssembler::store_klass_gap(Register dst, Register src) {
5740   if (UseCompressedClassPointers) {
5741     // Store to klass gap in destination
5742     movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src);
5743   }
5744 }
5745 
5746 #ifdef ASSERT
5747 void MacroAssembler::verify_heapbase(const char* msg) {
5748   assert (UseCompressedOops, "should be compressed");
5749   assert (Universe::heap() != NULL, "java heap should be initialized");
5750   if (CheckCompressedOops) {
5751     Label ok;
5752     push(rscratch1); // cmpptr trashes rscratch1
5753     cmpptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr()));
5754     jcc(Assembler::equal, ok);
5755     STOP(msg);
5756     bind(ok);
5757     pop(rscratch1);
5758   }
5759 }
5760 #endif
5761 
5762 // Algorithm must match oop.inline.hpp encode_heap_oop.
5763 void MacroAssembler::encode_heap_oop(Register r) {
5764 #ifdef ASSERT
5765   verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?");
5766 #endif
5767   verify_oop(r, "broken oop in encode_heap_oop");
5768   if (Universe::narrow_oop_base() == NULL) {
5769     if (Universe::narrow_oop_shift() != 0) {
5770       assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5771       shrq(r, LogMinObjAlignmentInBytes);
5772     }
5773     return;
5774   }
5775   testq(r, r);
5776   cmovq(Assembler::equal, r, r12_heapbase);
5777   subq(r, r12_heapbase);
5778   shrq(r, LogMinObjAlignmentInBytes);
5779 }
5780 
5781 void MacroAssembler::encode_heap_oop_not_null(Register r) {
5782 #ifdef ASSERT
5783   verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?");
5784   if (CheckCompressedOops) {
5785     Label ok;
5786     testq(r, r);
5787     jcc(Assembler::notEqual, ok);
5788     STOP("null oop passed to encode_heap_oop_not_null");
5789     bind(ok);
5790   }
5791 #endif
5792   verify_oop(r, "broken oop in encode_heap_oop_not_null");
5793   if (Universe::narrow_oop_base() != NULL) {
5794     subq(r, r12_heapbase);
5795   }
5796   if (Universe::narrow_oop_shift() != 0) {
5797     assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5798     shrq(r, LogMinObjAlignmentInBytes);
5799   }
5800 }
5801 
5802 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) {
5803 #ifdef ASSERT
5804   verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?");
5805   if (CheckCompressedOops) {
5806     Label ok;
5807     testq(src, src);
5808     jcc(Assembler::notEqual, ok);
5809     STOP("null oop passed to encode_heap_oop_not_null2");
5810     bind(ok);
5811   }
5812 #endif
5813   verify_oop(src, "broken oop in encode_heap_oop_not_null2");
5814   if (dst != src) {
5815     movq(dst, src);
5816   }
5817   if (Universe::narrow_oop_base() != NULL) {
5818     subq(dst, r12_heapbase);
5819   }
5820   if (Universe::narrow_oop_shift() != 0) {
5821     assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5822     shrq(dst, LogMinObjAlignmentInBytes);
5823   }
5824 }
5825 
5826 void  MacroAssembler::decode_heap_oop(Register r) {
5827 #ifdef ASSERT
5828   verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?");
5829 #endif
5830   if (Universe::narrow_oop_base() == NULL) {
5831     if (Universe::narrow_oop_shift() != 0) {
5832       assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5833       shlq(r, LogMinObjAlignmentInBytes);
5834     }
5835   } else {
5836     Label done;
5837     shlq(r, LogMinObjAlignmentInBytes);
5838     jccb(Assembler::equal, done);
5839     addq(r, r12_heapbase);
5840     bind(done);
5841   }
5842   verify_oop(r, "broken oop in decode_heap_oop");
5843 }
5844 
5845 void  MacroAssembler::decode_heap_oop_not_null(Register r) {
5846   // Note: it will change flags
5847   assert (UseCompressedOops, "should only be used for compressed headers");
5848   assert (Universe::heap() != NULL, "java heap should be initialized");
5849   // Cannot assert, unverified entry point counts instructions (see .ad file)
5850   // vtableStubs also counts instructions in pd_code_size_limit.
5851   // Also do not verify_oop as this is called by verify_oop.
5852   if (Universe::narrow_oop_shift() != 0) {
5853     assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5854     shlq(r, LogMinObjAlignmentInBytes);
5855     if (Universe::narrow_oop_base() != NULL) {
5856       addq(r, r12_heapbase);
5857     }
5858   } else {
5859     assert (Universe::narrow_oop_base() == NULL, "sanity");
5860   }
5861 }
5862 
5863 void  MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) {
5864   // Note: it will change flags
5865   assert (UseCompressedOops, "should only be used for compressed headers");
5866   assert (Universe::heap() != NULL, "java heap should be initialized");
5867   // Cannot assert, unverified entry point counts instructions (see .ad file)
5868   // vtableStubs also counts instructions in pd_code_size_limit.
5869   // Also do not verify_oop as this is called by verify_oop.
5870   if (Universe::narrow_oop_shift() != 0) {
5871     assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5872     if (LogMinObjAlignmentInBytes == Address::times_8) {
5873       leaq(dst, Address(r12_heapbase, src, Address::times_8, 0));
5874     } else {
5875       if (dst != src) {
5876         movq(dst, src);
5877       }
5878       shlq(dst, LogMinObjAlignmentInBytes);
5879       if (Universe::narrow_oop_base() != NULL) {
5880         addq(dst, r12_heapbase);
5881       }
5882     }
5883   } else {
5884     assert (Universe::narrow_oop_base() == NULL, "sanity");
5885     if (dst != src) {
5886       movq(dst, src);
5887     }
5888   }
5889 }
5890 
5891 void MacroAssembler::encode_klass_not_null(Register r) {
5892   if (Universe::narrow_klass_base() != NULL) {
5893     // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
5894     assert(r != r12_heapbase, "Encoding a klass in r12");
5895     mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base());
5896     subq(r, r12_heapbase);
5897   }
5898   if (Universe::narrow_klass_shift() != 0) {
5899     assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
5900     shrq(r, LogKlassAlignmentInBytes);
5901   }
5902   if (Universe::narrow_klass_base() != NULL) {
5903     reinit_heapbase();
5904   }
5905 }
5906 
5907 void MacroAssembler::encode_klass_not_null(Register dst, Register src) {
5908   if (dst == src) {
5909     encode_klass_not_null(src);
5910   } else {
5911     if (Universe::narrow_klass_base() != NULL) {
5912       mov64(dst, (int64_t)Universe::narrow_klass_base());
5913       negq(dst);
5914       addq(dst, src);
5915     } else {
5916       movptr(dst, src);
5917     }
5918     if (Universe::narrow_klass_shift() != 0) {
5919       assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
5920       shrq(dst, LogKlassAlignmentInBytes);
5921     }
5922   }
5923 }
5924 
5925 // Function instr_size_for_decode_klass_not_null() counts the instructions
5926 // generated by decode_klass_not_null(register r) and reinit_heapbase(),
5927 // when (Universe::heap() != NULL).  Hence, if the instructions they
5928 // generate change, then this method needs to be updated.
5929 int MacroAssembler::instr_size_for_decode_klass_not_null() {
5930   assert (UseCompressedClassPointers, "only for compressed klass ptrs");
5931   if (Universe::narrow_klass_base() != NULL) {
5932     // mov64 + addq + shlq? + mov64  (for reinit_heapbase()).
5933     return (Universe::narrow_klass_shift() == 0 ? 20 : 24);
5934   } else {
5935     // longest load decode klass function, mov64, leaq
5936     return 16;
5937   }
5938 }
5939 
5940 // !!! If the instructions that get generated here change then function
5941 // instr_size_for_decode_klass_not_null() needs to get updated.
5942 void  MacroAssembler::decode_klass_not_null(Register r) {
5943   // Note: it will change flags
5944   assert (UseCompressedClassPointers, "should only be used for compressed headers");
5945   assert(r != r12_heapbase, "Decoding a klass in r12");
5946   // Cannot assert, unverified entry point counts instructions (see .ad file)
5947   // vtableStubs also counts instructions in pd_code_size_limit.
5948   // Also do not verify_oop as this is called by verify_oop.
5949   if (Universe::narrow_klass_shift() != 0) {
5950     assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
5951     shlq(r, LogKlassAlignmentInBytes);
5952   }
5953   // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
5954   if (Universe::narrow_klass_base() != NULL) {
5955     mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base());
5956     addq(r, r12_heapbase);
5957     reinit_heapbase();
5958   }
5959 }
5960 
5961 void  MacroAssembler::decode_klass_not_null(Register dst, Register src) {
5962   // Note: it will change flags
5963   assert (UseCompressedClassPointers, "should only be used for compressed headers");
5964   if (dst == src) {
5965     decode_klass_not_null(dst);
5966   } else {
5967     // Cannot assert, unverified entry point counts instructions (see .ad file)
5968     // vtableStubs also counts instructions in pd_code_size_limit.
5969     // Also do not verify_oop as this is called by verify_oop.
5970     mov64(dst, (int64_t)Universe::narrow_klass_base());
5971     if (Universe::narrow_klass_shift() != 0) {
5972       assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
5973       assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?");
5974       leaq(dst, Address(dst, src, Address::times_8, 0));
5975     } else {
5976       addq(dst, src);
5977     }
5978   }
5979 }
5980 
5981 void  MacroAssembler::set_narrow_oop(Register dst, jobject obj) {
5982   assert (UseCompressedOops, "should only be used for compressed headers");
5983   assert (Universe::heap() != NULL, "java heap should be initialized");
5984   assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5985   int oop_index = oop_recorder()->find_index(obj);
5986   RelocationHolder rspec = oop_Relocation::spec(oop_index);
5987   mov_narrow_oop(dst, oop_index, rspec);
5988 }
5989 
5990 void  MacroAssembler::set_narrow_oop(Address dst, jobject obj) {
5991   assert (UseCompressedOops, "should only be used for compressed headers");
5992   assert (Universe::heap() != NULL, "java heap should be initialized");
5993   assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5994   int oop_index = oop_recorder()->find_index(obj);
5995   RelocationHolder rspec = oop_Relocation::spec(oop_index);
5996   mov_narrow_oop(dst, oop_index, rspec);
5997 }
5998 
5999 void  MacroAssembler::set_narrow_klass(Register dst, Klass* k) {
6000   assert (UseCompressedClassPointers, "should only be used for compressed headers");
6001   assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6002   int klass_index = oop_recorder()->find_index(k);
6003   RelocationHolder rspec = metadata_Relocation::spec(klass_index);
6004   mov_narrow_oop(dst, Klass::encode_klass(k), rspec);
6005 }
6006 
6007 void  MacroAssembler::set_narrow_klass(Address dst, Klass* k) {
6008   assert (UseCompressedClassPointers, "should only be used for compressed headers");
6009   assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6010   int klass_index = oop_recorder()->find_index(k);
6011   RelocationHolder rspec = metadata_Relocation::spec(klass_index);
6012   mov_narrow_oop(dst, Klass::encode_klass(k), rspec);
6013 }
6014 
6015 void  MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) {
6016   assert (UseCompressedOops, "should only be used for compressed headers");
6017   assert (Universe::heap() != NULL, "java heap should be initialized");
6018   assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6019   int oop_index = oop_recorder()->find_index(obj);
6020   RelocationHolder rspec = oop_Relocation::spec(oop_index);
6021   Assembler::cmp_narrow_oop(dst, oop_index, rspec);
6022 }
6023 
6024 void  MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) {
6025   assert (UseCompressedOops, "should only be used for compressed headers");
6026   assert (Universe::heap() != NULL, "java heap should be initialized");
6027   assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6028   int oop_index = oop_recorder()->find_index(obj);
6029   RelocationHolder rspec = oop_Relocation::spec(oop_index);
6030   Assembler::cmp_narrow_oop(dst, oop_index, rspec);
6031 }
6032 
6033 void  MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) {
6034   assert (UseCompressedClassPointers, "should only be used for compressed headers");
6035   assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6036   int klass_index = oop_recorder()->find_index(k);
6037   RelocationHolder rspec = metadata_Relocation::spec(klass_index);
6038   Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec);
6039 }
6040 
6041 void  MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) {
6042   assert (UseCompressedClassPointers, "should only be used for compressed headers");
6043   assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6044   int klass_index = oop_recorder()->find_index(k);
6045   RelocationHolder rspec = metadata_Relocation::spec(klass_index);
6046   Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec);
6047 }
6048 
6049 void MacroAssembler::reinit_heapbase() {
6050   if (UseCompressedOops || UseCompressedClassPointers) {
6051     if (Universe::heap() != NULL) {
6052       if (Universe::narrow_oop_base() == NULL) {
6053         MacroAssembler::xorptr(r12_heapbase, r12_heapbase);
6054       } else {
6055         mov64(r12_heapbase, (int64_t)Universe::narrow_ptrs_base());
6056       }
6057     } else {
6058       movptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr()));
6059     }
6060   }
6061 }
6062 
6063 #endif // _LP64
6064 
6065 
6066 // C2 compiled method's prolog code.
6067 void MacroAssembler::verified_entry(int framesize, int stack_bang_size, bool fp_mode_24b) {
6068 
6069   // WARNING: Initial instruction MUST be 5 bytes or longer so that
6070   // NativeJump::patch_verified_entry will be able to patch out the entry
6071   // code safely. The push to verify stack depth is ok at 5 bytes,
6072   // the frame allocation can be either 3 or 6 bytes. So if we don't do
6073   // stack bang then we must use the 6 byte frame allocation even if
6074   // we have no frame. :-(
6075   assert(stack_bang_size >= framesize || stack_bang_size <= 0, "stack bang size incorrect");
6076 
6077   assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
6078   // Remove word for return addr
6079   framesize -= wordSize;
6080   stack_bang_size -= wordSize;
6081 
6082   // Calls to C2R adapters often do not accept exceptional returns.
6083   // We require that their callers must bang for them.  But be careful, because
6084   // some VM calls (such as call site linkage) can use several kilobytes of
6085   // stack.  But the stack safety zone should account for that.
6086   // See bugs 4446381, 4468289, 4497237.
6087   if (stack_bang_size > 0) {
6088     generate_stack_overflow_check(stack_bang_size);
6089 
6090     // We always push rbp, so that on return to interpreter rbp, will be
6091     // restored correctly and we can correct the stack.
6092     push(rbp);
6093     // Remove word for ebp
6094     framesize -= wordSize;
6095 
6096     // Create frame
6097     if (framesize) {
6098       subptr(rsp, framesize);
6099     }
6100   } else {
6101     // Create frame (force generation of a 4 byte immediate value)
6102     subptr_imm32(rsp, framesize);
6103 
6104     // Save RBP register now.
6105     framesize -= wordSize;
6106     movptr(Address(rsp, framesize), rbp);
6107   }
6108 
6109   if (VerifyStackAtCalls) { // Majik cookie to verify stack depth
6110     framesize -= wordSize;
6111     movptr(Address(rsp, framesize), (int32_t)0xbadb100d);
6112   }
6113 
6114 #ifndef _LP64
6115   // If method sets FPU control word do it now
6116   if (fp_mode_24b) {
6117     fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
6118   }
6119   if (UseSSE >= 2 && VerifyFPU) {
6120     verify_FPU(0, "FPU stack must be clean on entry");
6121   }
6122 #endif
6123 
6124 #ifdef ASSERT
6125   if (VerifyStackAtCalls) {
6126     Label L;
6127     push(rax);
6128     mov(rax, rsp);
6129     andptr(rax, StackAlignmentInBytes-1);
6130     cmpptr(rax, StackAlignmentInBytes-wordSize);
6131     pop(rax);
6132     jcc(Assembler::equal, L);
6133     STOP("Stack is not properly aligned!");
6134     bind(L);
6135   }
6136 #endif
6137 
6138 }
6139 
6140 void MacroAssembler::clear_mem(Register base, Register cnt, Register tmp) {
6141   // cnt - number of qwords (8-byte words).
6142   // base - start address, qword aligned.
6143   assert(base==rdi, "base register must be edi for rep stos");
6144   assert(tmp==rax,   "tmp register must be eax for rep stos");
6145   assert(cnt==rcx,   "cnt register must be ecx for rep stos");
6146 
6147   xorptr(tmp, tmp);
6148   if (UseFastStosb) {
6149     shlptr(cnt,3); // convert to number of bytes
6150     rep_stosb();
6151   } else {
6152     NOT_LP64(shlptr(cnt,1);) // convert to number of dwords for 32-bit VM
6153     rep_stos();
6154   }
6155 }
6156 
6157 // IndexOf for constant substrings with size >= 8 chars
6158 // which don't need to be loaded through stack.
6159 void MacroAssembler::string_indexofC8(Register str1, Register str2,
6160                                       Register cnt1, Register cnt2,
6161                                       int int_cnt2,  Register result,
6162                                       XMMRegister vec, Register tmp) {
6163   ShortBranchVerifier sbv(this);
6164   assert(UseSSE42Intrinsics, "SSE4.2 is required");
6165 
6166   // This method uses pcmpestri instruction with bound registers
6167   //   inputs:
6168   //     xmm - substring
6169   //     rax - substring length (elements count)
6170   //     mem - scanned string
6171   //     rdx - string length (elements count)
6172   //     0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
6173   //   outputs:
6174   //     rcx - matched index in string
6175   assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
6176 
6177   Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR,
6178         RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR,
6179         MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE;
6180 
6181   // Note, inline_string_indexOf() generates checks:
6182   // if (substr.count > string.count) return -1;
6183   // if (substr.count == 0) return 0;
6184   assert(int_cnt2 >= 8, "this code isused only for cnt2 >= 8 chars");
6185 
6186   // Load substring.
6187   movdqu(vec, Address(str2, 0));
6188   movl(cnt2, int_cnt2);
6189   movptr(result, str1); // string addr
6190 
6191   if (int_cnt2 > 8) {
6192     jmpb(SCAN_TO_SUBSTR);
6193 
6194     // Reload substr for rescan, this code
6195     // is executed only for large substrings (> 8 chars)
6196     bind(RELOAD_SUBSTR);
6197     movdqu(vec, Address(str2, 0));
6198     negptr(cnt2); // Jumped here with negative cnt2, convert to positive
6199 
6200     bind(RELOAD_STR);
6201     // We came here after the beginning of the substring was
6202     // matched but the rest of it was not so we need to search
6203     // again. Start from the next element after the previous match.
6204 
6205     // cnt2 is number of substring reminding elements and
6206     // cnt1 is number of string reminding elements when cmp failed.
6207     // Restored cnt1 = cnt1 - cnt2 + int_cnt2
6208     subl(cnt1, cnt2);
6209     addl(cnt1, int_cnt2);
6210     movl(cnt2, int_cnt2); // Now restore cnt2
6211 
6212     decrementl(cnt1);     // Shift to next element
6213     cmpl(cnt1, cnt2);
6214     jccb(Assembler::negative, RET_NOT_FOUND);  // Left less then substring
6215 
6216     addptr(result, 2);
6217 
6218   } // (int_cnt2 > 8)
6219 
6220   // Scan string for start of substr in 16-byte vectors
6221   bind(SCAN_TO_SUBSTR);
6222   pcmpestri(vec, Address(result, 0), 0x0d);
6223   jccb(Assembler::below, FOUND_CANDIDATE);   // CF == 1
6224   subl(cnt1, 8);
6225   jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
6226   cmpl(cnt1, cnt2);
6227   jccb(Assembler::negative, RET_NOT_FOUND);  // Left less then substring
6228   addptr(result, 16);
6229   jmpb(SCAN_TO_SUBSTR);
6230 
6231   // Found a potential substr
6232   bind(FOUND_CANDIDATE);
6233   // Matched whole vector if first element matched (tmp(rcx) == 0).
6234   if (int_cnt2 == 8) {
6235     jccb(Assembler::overflow, RET_FOUND);    // OF == 1
6236   } else { // int_cnt2 > 8
6237     jccb(Assembler::overflow, FOUND_SUBSTR);
6238   }
6239   // After pcmpestri tmp(rcx) contains matched element index
6240   // Compute start addr of substr
6241   lea(result, Address(result, tmp, Address::times_2));
6242 
6243   // Make sure string is still long enough
6244   subl(cnt1, tmp);
6245   cmpl(cnt1, cnt2);
6246   if (int_cnt2 == 8) {
6247     jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
6248   } else { // int_cnt2 > 8
6249     jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD);
6250   }
6251   // Left less then substring.
6252 
6253   bind(RET_NOT_FOUND);
6254   movl(result, -1);
6255   jmpb(EXIT);
6256 
6257   if (int_cnt2 > 8) {
6258     // This code is optimized for the case when whole substring
6259     // is matched if its head is matched.
6260     bind(MATCH_SUBSTR_HEAD);
6261     pcmpestri(vec, Address(result, 0), 0x0d);
6262     // Reload only string if does not match
6263     jccb(Assembler::noOverflow, RELOAD_STR); // OF == 0
6264 
6265     Label CONT_SCAN_SUBSTR;
6266     // Compare the rest of substring (> 8 chars).
6267     bind(FOUND_SUBSTR);
6268     // First 8 chars are already matched.
6269     negptr(cnt2);
6270     addptr(cnt2, 8);
6271 
6272     bind(SCAN_SUBSTR);
6273     subl(cnt1, 8);
6274     cmpl(cnt2, -8); // Do not read beyond substring
6275     jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR);
6276     // Back-up strings to avoid reading beyond substring:
6277     // cnt1 = cnt1 - cnt2 + 8
6278     addl(cnt1, cnt2); // cnt2 is negative
6279     addl(cnt1, 8);
6280     movl(cnt2, 8); negptr(cnt2);
6281     bind(CONT_SCAN_SUBSTR);
6282     if (int_cnt2 < (int)G) {
6283       movdqu(vec, Address(str2, cnt2, Address::times_2, int_cnt2*2));
6284       pcmpestri(vec, Address(result, cnt2, Address::times_2, int_cnt2*2), 0x0d);
6285     } else {
6286       // calculate index in register to avoid integer overflow (int_cnt2*2)
6287       movl(tmp, int_cnt2);
6288       addptr(tmp, cnt2);
6289       movdqu(vec, Address(str2, tmp, Address::times_2, 0));
6290       pcmpestri(vec, Address(result, tmp, Address::times_2, 0), 0x0d);
6291     }
6292     // Need to reload strings pointers if not matched whole vector
6293     jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
6294     addptr(cnt2, 8);
6295     jcc(Assembler::negative, SCAN_SUBSTR);
6296     // Fall through if found full substring
6297 
6298   } // (int_cnt2 > 8)
6299 
6300   bind(RET_FOUND);
6301   // Found result if we matched full small substring.
6302   // Compute substr offset
6303   subptr(result, str1);
6304   shrl(result, 1); // index
6305   bind(EXIT);
6306 
6307 } // string_indexofC8
6308 
6309 // Small strings are loaded through stack if they cross page boundary.
6310 void MacroAssembler::string_indexof(Register str1, Register str2,
6311                                     Register cnt1, Register cnt2,
6312                                     int int_cnt2,  Register result,
6313                                     XMMRegister vec, Register tmp) {
6314   ShortBranchVerifier sbv(this);
6315   assert(UseSSE42Intrinsics, "SSE4.2 is required");
6316   //
6317   // int_cnt2 is length of small (< 8 chars) constant substring
6318   // or (-1) for non constant substring in which case its length
6319   // is in cnt2 register.
6320   //
6321   // Note, inline_string_indexOf() generates checks:
6322   // if (substr.count > string.count) return -1;
6323   // if (substr.count == 0) return 0;
6324   //
6325   assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < 8), "should be != 0");
6326 
6327   // This method uses pcmpestri instruction with bound registers
6328   //   inputs:
6329   //     xmm - substring
6330   //     rax - substring length (elements count)
6331   //     mem - scanned string
6332   //     rdx - string length (elements count)
6333   //     0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
6334   //   outputs:
6335   //     rcx - matched index in string
6336   assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
6337 
6338   Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR,
6339         RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR,
6340         FOUND_CANDIDATE;
6341 
6342   { //========================================================
6343     // We don't know where these strings are located
6344     // and we can't read beyond them. Load them through stack.
6345     Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR;
6346 
6347     movptr(tmp, rsp); // save old SP
6348 
6349     if (int_cnt2 > 0) {     // small (< 8 chars) constant substring
6350       if (int_cnt2 == 1) {  // One char
6351         load_unsigned_short(result, Address(str2, 0));
6352         movdl(vec, result); // move 32 bits
6353       } else if (int_cnt2 == 2) { // Two chars
6354         movdl(vec, Address(str2, 0)); // move 32 bits
6355       } else if (int_cnt2 == 4) { // Four chars
6356         movq(vec, Address(str2, 0));  // move 64 bits
6357       } else { // cnt2 = { 3, 5, 6, 7 }
6358         // Array header size is 12 bytes in 32-bit VM
6359         // + 6 bytes for 3 chars == 18 bytes,
6360         // enough space to load vec and shift.
6361         assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity");
6362         movdqu(vec, Address(str2, (int_cnt2*2)-16));
6363         psrldq(vec, 16-(int_cnt2*2));
6364       }
6365     } else { // not constant substring
6366       cmpl(cnt2, 8);
6367       jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough
6368 
6369       // We can read beyond string if srt+16 does not cross page boundary
6370       // since heaps are aligned and mapped by pages.
6371       assert(os::vm_page_size() < (int)G, "default page should be small");
6372       movl(result, str2); // We need only low 32 bits
6373       andl(result, (os::vm_page_size()-1));
6374       cmpl(result, (os::vm_page_size()-16));
6375       jccb(Assembler::belowEqual, CHECK_STR);
6376 
6377       // Move small strings to stack to allow load 16 bytes into vec.
6378       subptr(rsp, 16);
6379       int stk_offset = wordSize-2;
6380       push(cnt2);
6381 
6382       bind(COPY_SUBSTR);
6383       load_unsigned_short(result, Address(str2, cnt2, Address::times_2, -2));
6384       movw(Address(rsp, cnt2, Address::times_2, stk_offset), result);
6385       decrement(cnt2);
6386       jccb(Assembler::notZero, COPY_SUBSTR);
6387 
6388       pop(cnt2);
6389       movptr(str2, rsp);  // New substring address
6390     } // non constant
6391 
6392     bind(CHECK_STR);
6393     cmpl(cnt1, 8);
6394     jccb(Assembler::aboveEqual, BIG_STRINGS);
6395 
6396     // Check cross page boundary.
6397     movl(result, str1); // We need only low 32 bits
6398     andl(result, (os::vm_page_size()-1));
6399     cmpl(result, (os::vm_page_size()-16));
6400     jccb(Assembler::belowEqual, BIG_STRINGS);
6401 
6402     subptr(rsp, 16);
6403     int stk_offset = -2;
6404     if (int_cnt2 < 0) { // not constant
6405       push(cnt2);
6406       stk_offset += wordSize;
6407     }
6408     movl(cnt2, cnt1);
6409 
6410     bind(COPY_STR);
6411     load_unsigned_short(result, Address(str1, cnt2, Address::times_2, -2));
6412     movw(Address(rsp, cnt2, Address::times_2, stk_offset), result);
6413     decrement(cnt2);
6414     jccb(Assembler::notZero, COPY_STR);
6415 
6416     if (int_cnt2 < 0) { // not constant
6417       pop(cnt2);
6418     }
6419     movptr(str1, rsp);  // New string address
6420 
6421     bind(BIG_STRINGS);
6422     // Load substring.
6423     if (int_cnt2 < 0) { // -1
6424       movdqu(vec, Address(str2, 0));
6425       push(cnt2);       // substr count
6426       push(str2);       // substr addr
6427       push(str1);       // string addr
6428     } else {
6429       // Small (< 8 chars) constant substrings are loaded already.
6430       movl(cnt2, int_cnt2);
6431     }
6432     push(tmp);  // original SP
6433 
6434   } // Finished loading
6435 
6436   //========================================================
6437   // Start search
6438   //
6439 
6440   movptr(result, str1); // string addr
6441 
6442   if (int_cnt2  < 0) {  // Only for non constant substring
6443     jmpb(SCAN_TO_SUBSTR);
6444 
6445     // SP saved at sp+0
6446     // String saved at sp+1*wordSize
6447     // Substr saved at sp+2*wordSize
6448     // Substr count saved at sp+3*wordSize
6449 
6450     // Reload substr for rescan, this code
6451     // is executed only for large substrings (> 8 chars)
6452     bind(RELOAD_SUBSTR);
6453     movptr(str2, Address(rsp, 2*wordSize));
6454     movl(cnt2, Address(rsp, 3*wordSize));
6455     movdqu(vec, Address(str2, 0));
6456     // We came here after the beginning of the substring was
6457     // matched but the rest of it was not so we need to search
6458     // again. Start from the next element after the previous match.
6459     subptr(str1, result); // Restore counter
6460     shrl(str1, 1);
6461     addl(cnt1, str1);
6462     decrementl(cnt1);   // Shift to next element
6463     cmpl(cnt1, cnt2);
6464     jccb(Assembler::negative, RET_NOT_FOUND);  // Left less then substring
6465 
6466     addptr(result, 2);
6467   } // non constant
6468 
6469   // Scan string for start of substr in 16-byte vectors
6470   bind(SCAN_TO_SUBSTR);
6471   assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
6472   pcmpestri(vec, Address(result, 0), 0x0d);
6473   jccb(Assembler::below, FOUND_CANDIDATE);   // CF == 1
6474   subl(cnt1, 8);
6475   jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
6476   cmpl(cnt1, cnt2);
6477   jccb(Assembler::negative, RET_NOT_FOUND);  // Left less then substring
6478   addptr(result, 16);
6479 
6480   bind(ADJUST_STR);
6481   cmpl(cnt1, 8); // Do not read beyond string
6482   jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
6483   // Back-up string to avoid reading beyond string.
6484   lea(result, Address(result, cnt1, Address::times_2, -16));
6485   movl(cnt1, 8);
6486   jmpb(SCAN_TO_SUBSTR);
6487 
6488   // Found a potential substr
6489   bind(FOUND_CANDIDATE);
6490   // After pcmpestri tmp(rcx) contains matched element index
6491 
6492   // Make sure string is still long enough
6493   subl(cnt1, tmp);
6494   cmpl(cnt1, cnt2);
6495   jccb(Assembler::greaterEqual, FOUND_SUBSTR);
6496   // Left less then substring.
6497 
6498   bind(RET_NOT_FOUND);
6499   movl(result, -1);
6500   jmpb(CLEANUP);
6501 
6502   bind(FOUND_SUBSTR);
6503   // Compute start addr of substr
6504   lea(result, Address(result, tmp, Address::times_2));
6505 
6506   if (int_cnt2 > 0) { // Constant substring
6507     // Repeat search for small substring (< 8 chars)
6508     // from new point without reloading substring.
6509     // Have to check that we don't read beyond string.
6510     cmpl(tmp, 8-int_cnt2);
6511     jccb(Assembler::greater, ADJUST_STR);
6512     // Fall through if matched whole substring.
6513   } else { // non constant
6514     assert(int_cnt2 == -1, "should be != 0");
6515 
6516     addl(tmp, cnt2);
6517     // Found result if we matched whole substring.
6518     cmpl(tmp, 8);
6519     jccb(Assembler::lessEqual, RET_FOUND);
6520 
6521     // Repeat search for small substring (<= 8 chars)
6522     // from new point 'str1' without reloading substring.
6523     cmpl(cnt2, 8);
6524     // Have to check that we don't read beyond string.
6525     jccb(Assembler::lessEqual, ADJUST_STR);
6526 
6527     Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG;
6528     // Compare the rest of substring (> 8 chars).
6529     movptr(str1, result);
6530 
6531     cmpl(tmp, cnt2);
6532     // First 8 chars are already matched.
6533     jccb(Assembler::equal, CHECK_NEXT);
6534 
6535     bind(SCAN_SUBSTR);
6536     pcmpestri(vec, Address(str1, 0), 0x0d);
6537     // Need to reload strings pointers if not matched whole vector
6538     jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
6539 
6540     bind(CHECK_NEXT);
6541     subl(cnt2, 8);
6542     jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring
6543     addptr(str1, 16);
6544     addptr(str2, 16);
6545     subl(cnt1, 8);
6546     cmpl(cnt2, 8); // Do not read beyond substring
6547     jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR);
6548     // Back-up strings to avoid reading beyond substring.
6549     lea(str2, Address(str2, cnt2, Address::times_2, -16));
6550     lea(str1, Address(str1, cnt2, Address::times_2, -16));
6551     subl(cnt1, cnt2);
6552     movl(cnt2, 8);
6553     addl(cnt1, 8);
6554     bind(CONT_SCAN_SUBSTR);
6555     movdqu(vec, Address(str2, 0));
6556     jmpb(SCAN_SUBSTR);
6557 
6558     bind(RET_FOUND_LONG);
6559     movptr(str1, Address(rsp, wordSize));
6560   } // non constant
6561 
6562   bind(RET_FOUND);
6563   // Compute substr offset
6564   subptr(result, str1);
6565   shrl(result, 1); // index
6566 
6567   bind(CLEANUP);
6568   pop(rsp); // restore SP
6569 
6570 } // string_indexof
6571 
6572 // Compare strings.
6573 void MacroAssembler::string_compare(Register str1, Register str2,
6574                                     Register cnt1, Register cnt2, Register result,
6575                                     XMMRegister vec1) {
6576   ShortBranchVerifier sbv(this);
6577   Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL;
6578 
6579   // Compute the minimum of the string lengths and the
6580   // difference of the string lengths (stack).
6581   // Do the conditional move stuff
6582   movl(result, cnt1);
6583   subl(cnt1, cnt2);
6584   push(cnt1);
6585   cmov32(Assembler::lessEqual, cnt2, result);
6586 
6587   // Is the minimum length zero?
6588   testl(cnt2, cnt2);
6589   jcc(Assembler::zero, LENGTH_DIFF_LABEL);
6590 
6591   // Compare first characters
6592   load_unsigned_short(result, Address(str1, 0));
6593   load_unsigned_short(cnt1, Address(str2, 0));
6594   subl(result, cnt1);
6595   jcc(Assembler::notZero,  POP_LABEL);
6596   cmpl(cnt2, 1);
6597   jcc(Assembler::equal, LENGTH_DIFF_LABEL);
6598 
6599   // Check if the strings start at the same location.
6600   cmpptr(str1, str2);
6601   jcc(Assembler::equal, LENGTH_DIFF_LABEL);
6602 
6603   Address::ScaleFactor scale = Address::times_2;
6604   int stride = 8;
6605 
6606   if (UseAVX >= 2 && UseSSE42Intrinsics) {
6607     Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR;
6608     Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR;
6609     Label COMPARE_TAIL_LONG;
6610     int pcmpmask = 0x19;
6611 
6612     // Setup to compare 16-chars (32-bytes) vectors,
6613     // start from first character again because it has aligned address.
6614     int stride2 = 16;
6615     int adr_stride  = stride  << scale;
6616 
6617     assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
6618     // rax and rdx are used by pcmpestri as elements counters
6619     movl(result, cnt2);
6620     andl(cnt2, ~(stride2-1));   // cnt2 holds the vector count
6621     jcc(Assembler::zero, COMPARE_TAIL_LONG);
6622 
6623     // fast path : compare first 2 8-char vectors.
6624     bind(COMPARE_16_CHARS);
6625     movdqu(vec1, Address(str1, 0));
6626     pcmpestri(vec1, Address(str2, 0), pcmpmask);
6627     jccb(Assembler::below, COMPARE_INDEX_CHAR);
6628 
6629     movdqu(vec1, Address(str1, adr_stride));
6630     pcmpestri(vec1, Address(str2, adr_stride), pcmpmask);
6631     jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS);
6632     addl(cnt1, stride);
6633 
6634     // Compare the characters at index in cnt1
6635     bind(COMPARE_INDEX_CHAR); //cnt1 has the offset of the mismatching character
6636     load_unsigned_short(result, Address(str1, cnt1, scale));
6637     load_unsigned_short(cnt2, Address(str2, cnt1, scale));
6638     subl(result, cnt2);
6639     jmp(POP_LABEL);
6640 
6641     // Setup the registers to start vector comparison loop
6642     bind(COMPARE_WIDE_VECTORS);
6643     lea(str1, Address(str1, result, scale));
6644     lea(str2, Address(str2, result, scale));
6645     subl(result, stride2);
6646     subl(cnt2, stride2);
6647     jccb(Assembler::zero, COMPARE_WIDE_TAIL);
6648     negptr(result);
6649 
6650     //  In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest)
6651     bind(COMPARE_WIDE_VECTORS_LOOP);
6652     vmovdqu(vec1, Address(str1, result, scale));
6653     vpxor(vec1, Address(str2, result, scale));
6654     vptest(vec1, vec1);
6655     jccb(Assembler::notZero, VECTOR_NOT_EQUAL);
6656     addptr(result, stride2);
6657     subl(cnt2, stride2);
6658     jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP);
6659     // clean upper bits of YMM registers
6660     vpxor(vec1, vec1);
6661 
6662     // compare wide vectors tail
6663     bind(COMPARE_WIDE_TAIL);
6664     testptr(result, result);
6665     jccb(Assembler::zero, LENGTH_DIFF_LABEL);
6666 
6667     movl(result, stride2);
6668     movl(cnt2, result);
6669     negptr(result);
6670     jmpb(COMPARE_WIDE_VECTORS_LOOP);
6671 
6672     // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors.
6673     bind(VECTOR_NOT_EQUAL);
6674     // clean upper bits of YMM registers
6675     vpxor(vec1, vec1);
6676     lea(str1, Address(str1, result, scale));
6677     lea(str2, Address(str2, result, scale));
6678     jmp(COMPARE_16_CHARS);
6679 
6680     // Compare tail chars, length between 1 to 15 chars
6681     bind(COMPARE_TAIL_LONG);
6682     movl(cnt2, result);
6683     cmpl(cnt2, stride);
6684     jccb(Assembler::less, COMPARE_SMALL_STR);
6685 
6686     movdqu(vec1, Address(str1, 0));
6687     pcmpestri(vec1, Address(str2, 0), pcmpmask);
6688     jcc(Assembler::below, COMPARE_INDEX_CHAR);
6689     subptr(cnt2, stride);
6690     jccb(Assembler::zero, LENGTH_DIFF_LABEL);
6691     lea(str1, Address(str1, result, scale));
6692     lea(str2, Address(str2, result, scale));
6693     negptr(cnt2);
6694     jmpb(WHILE_HEAD_LABEL);
6695 
6696     bind(COMPARE_SMALL_STR);
6697   } else if (UseSSE42Intrinsics) {
6698     Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL;
6699     int pcmpmask = 0x19;
6700     // Setup to compare 8-char (16-byte) vectors,
6701     // start from first character again because it has aligned address.
6702     movl(result, cnt2);
6703     andl(cnt2, ~(stride - 1));   // cnt2 holds the vector count
6704     jccb(Assembler::zero, COMPARE_TAIL);
6705 
6706     lea(str1, Address(str1, result, scale));
6707     lea(str2, Address(str2, result, scale));
6708     negptr(result);
6709 
6710     // pcmpestri
6711     //   inputs:
6712     //     vec1- substring
6713     //     rax - negative string length (elements count)
6714     //     mem - scanned string
6715     //     rdx - string length (elements count)
6716     //     pcmpmask - cmp mode: 11000 (string compare with negated result)
6717     //               + 00 (unsigned bytes) or  + 01 (unsigned shorts)
6718     //   outputs:
6719     //     rcx - first mismatched element index
6720     assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
6721 
6722     bind(COMPARE_WIDE_VECTORS);
6723     movdqu(vec1, Address(str1, result, scale));
6724     pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
6725     // After pcmpestri cnt1(rcx) contains mismatched element index
6726 
6727     jccb(Assembler::below, VECTOR_NOT_EQUAL);  // CF==1
6728     addptr(result, stride);
6729     subptr(cnt2, stride);
6730     jccb(Assembler::notZero, COMPARE_WIDE_VECTORS);
6731 
6732     // compare wide vectors tail
6733     testptr(result, result);
6734     jccb(Assembler::zero, LENGTH_DIFF_LABEL);
6735 
6736     movl(cnt2, stride);
6737     movl(result, stride);
6738     negptr(result);
6739     movdqu(vec1, Address(str1, result, scale));
6740     pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
6741     jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL);
6742 
6743     // Mismatched characters in the vectors
6744     bind(VECTOR_NOT_EQUAL);
6745     addptr(cnt1, result);
6746     load_unsigned_short(result, Address(str1, cnt1, scale));
6747     load_unsigned_short(cnt2, Address(str2, cnt1, scale));
6748     subl(result, cnt2);
6749     jmpb(POP_LABEL);
6750 
6751     bind(COMPARE_TAIL); // limit is zero
6752     movl(cnt2, result);
6753     // Fallthru to tail compare
6754   }
6755   // Shift str2 and str1 to the end of the arrays, negate min
6756   lea(str1, Address(str1, cnt2, scale));
6757   lea(str2, Address(str2, cnt2, scale));
6758   decrementl(cnt2);  // first character was compared already
6759   negptr(cnt2);
6760 
6761   // Compare the rest of the elements
6762   bind(WHILE_HEAD_LABEL);
6763   load_unsigned_short(result, Address(str1, cnt2, scale, 0));
6764   load_unsigned_short(cnt1, Address(str2, cnt2, scale, 0));
6765   subl(result, cnt1);
6766   jccb(Assembler::notZero, POP_LABEL);
6767   increment(cnt2);
6768   jccb(Assembler::notZero, WHILE_HEAD_LABEL);
6769 
6770   // Strings are equal up to min length.  Return the length difference.
6771   bind(LENGTH_DIFF_LABEL);
6772   pop(result);
6773   jmpb(DONE_LABEL);
6774 
6775   // Discard the stored length difference
6776   bind(POP_LABEL);
6777   pop(cnt1);
6778 
6779   // That's it
6780   bind(DONE_LABEL);
6781 }
6782 
6783 // Compare char[] arrays aligned to 4 bytes or substrings.
6784 void MacroAssembler::char_arrays_equals(bool is_array_equ, Register ary1, Register ary2,
6785                                         Register limit, Register result, Register chr,
6786                                         XMMRegister vec1, XMMRegister vec2) {
6787   ShortBranchVerifier sbv(this);
6788   Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR;
6789 
6790   int length_offset  = arrayOopDesc::length_offset_in_bytes();
6791   int base_offset    = arrayOopDesc::base_offset_in_bytes(T_CHAR);
6792 
6793   // Check the input args
6794   cmpptr(ary1, ary2);
6795   jcc(Assembler::equal, TRUE_LABEL);
6796 
6797   if (is_array_equ) {
6798     // Need additional checks for arrays_equals.
6799     testptr(ary1, ary1);
6800     jcc(Assembler::zero, FALSE_LABEL);
6801     testptr(ary2, ary2);
6802     jcc(Assembler::zero, FALSE_LABEL);
6803 
6804     // Check the lengths
6805     movl(limit, Address(ary1, length_offset));
6806     cmpl(limit, Address(ary2, length_offset));
6807     jcc(Assembler::notEqual, FALSE_LABEL);
6808   }
6809 
6810   // count == 0
6811   testl(limit, limit);
6812   jcc(Assembler::zero, TRUE_LABEL);
6813 
6814   if (is_array_equ) {
6815     // Load array address
6816     lea(ary1, Address(ary1, base_offset));
6817     lea(ary2, Address(ary2, base_offset));
6818   }
6819 
6820   shll(limit, 1);      // byte count != 0
6821   movl(result, limit); // copy
6822 
6823   if (UseAVX >= 2) {
6824     // With AVX2, use 32-byte vector compare
6825     Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
6826 
6827     // Compare 32-byte vectors
6828     andl(result, 0x0000001e);  //   tail count (in bytes)
6829     andl(limit, 0xffffffe0);   // vector count (in bytes)
6830     jccb(Assembler::zero, COMPARE_TAIL);
6831 
6832     lea(ary1, Address(ary1, limit, Address::times_1));
6833     lea(ary2, Address(ary2, limit, Address::times_1));
6834     negptr(limit);
6835 
6836     bind(COMPARE_WIDE_VECTORS);
6837     vmovdqu(vec1, Address(ary1, limit, Address::times_1));
6838     vmovdqu(vec2, Address(ary2, limit, Address::times_1));
6839     vpxor(vec1, vec2);
6840 
6841     vptest(vec1, vec1);
6842     jccb(Assembler::notZero, FALSE_LABEL);
6843     addptr(limit, 32);
6844     jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
6845 
6846     testl(result, result);
6847     jccb(Assembler::zero, TRUE_LABEL);
6848 
6849     vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
6850     vmovdqu(vec2, Address(ary2, result, Address::times_1, -32));
6851     vpxor(vec1, vec2);
6852 
6853     vptest(vec1, vec1);
6854     jccb(Assembler::notZero, FALSE_LABEL);
6855     jmpb(TRUE_LABEL);
6856 
6857     bind(COMPARE_TAIL); // limit is zero
6858     movl(limit, result);
6859     // Fallthru to tail compare
6860   } else if (UseSSE42Intrinsics) {
6861     // With SSE4.2, use double quad vector compare
6862     Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
6863 
6864     // Compare 16-byte vectors
6865     andl(result, 0x0000000e);  //   tail count (in bytes)
6866     andl(limit, 0xfffffff0);   // vector count (in bytes)
6867     jccb(Assembler::zero, COMPARE_TAIL);
6868 
6869     lea(ary1, Address(ary1, limit, Address::times_1));
6870     lea(ary2, Address(ary2, limit, Address::times_1));
6871     negptr(limit);
6872 
6873     bind(COMPARE_WIDE_VECTORS);
6874     movdqu(vec1, Address(ary1, limit, Address::times_1));
6875     movdqu(vec2, Address(ary2, limit, Address::times_1));
6876     pxor(vec1, vec2);
6877 
6878     ptest(vec1, vec1);
6879     jccb(Assembler::notZero, FALSE_LABEL);
6880     addptr(limit, 16);
6881     jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
6882 
6883     testl(result, result);
6884     jccb(Assembler::zero, TRUE_LABEL);
6885 
6886     movdqu(vec1, Address(ary1, result, Address::times_1, -16));
6887     movdqu(vec2, Address(ary2, result, Address::times_1, -16));
6888     pxor(vec1, vec2);
6889 
6890     ptest(vec1, vec1);
6891     jccb(Assembler::notZero, FALSE_LABEL);
6892     jmpb(TRUE_LABEL);
6893 
6894     bind(COMPARE_TAIL); // limit is zero
6895     movl(limit, result);
6896     // Fallthru to tail compare
6897   }
6898 
6899   // Compare 4-byte vectors
6900   andl(limit, 0xfffffffc); // vector count (in bytes)
6901   jccb(Assembler::zero, COMPARE_CHAR);
6902 
6903   lea(ary1, Address(ary1, limit, Address::times_1));
6904   lea(ary2, Address(ary2, limit, Address::times_1));
6905   negptr(limit);
6906 
6907   bind(COMPARE_VECTORS);
6908   movl(chr, Address(ary1, limit, Address::times_1));
6909   cmpl(chr, Address(ary2, limit, Address::times_1));
6910   jccb(Assembler::notEqual, FALSE_LABEL);
6911   addptr(limit, 4);
6912   jcc(Assembler::notZero, COMPARE_VECTORS);
6913 
6914   // Compare trailing char (final 2 bytes), if any
6915   bind(COMPARE_CHAR);
6916   testl(result, 0x2);   // tail  char
6917   jccb(Assembler::zero, TRUE_LABEL);
6918   load_unsigned_short(chr, Address(ary1, 0));
6919   load_unsigned_short(limit, Address(ary2, 0));
6920   cmpl(chr, limit);
6921   jccb(Assembler::notEqual, FALSE_LABEL);
6922 
6923   bind(TRUE_LABEL);
6924   movl(result, 1);   // return true
6925   jmpb(DONE);
6926 
6927   bind(FALSE_LABEL);
6928   xorl(result, result); // return false
6929 
6930   // That's it
6931   bind(DONE);
6932   if (UseAVX >= 2) {
6933     // clean upper bits of YMM registers
6934     vpxor(vec1, vec1);
6935     vpxor(vec2, vec2);
6936   }
6937 }
6938 
6939 void MacroAssembler::generate_fill(BasicType t, bool aligned,
6940                                    Register to, Register value, Register count,
6941                                    Register rtmp, XMMRegister xtmp) {
6942   ShortBranchVerifier sbv(this);
6943   assert_different_registers(to, value, count, rtmp);
6944   Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte;
6945   Label L_fill_2_bytes, L_fill_4_bytes;
6946 
6947   int shift = -1;
6948   switch (t) {
6949     case T_BYTE:
6950       shift = 2;
6951       break;
6952     case T_SHORT:
6953       shift = 1;
6954       break;
6955     case T_INT:
6956       shift = 0;
6957       break;
6958     default: ShouldNotReachHere();
6959   }
6960 
6961   if (t == T_BYTE) {
6962     andl(value, 0xff);
6963     movl(rtmp, value);
6964     shll(rtmp, 8);
6965     orl(value, rtmp);
6966   }
6967   if (t == T_SHORT) {
6968     andl(value, 0xffff);
6969   }
6970   if (t == T_BYTE || t == T_SHORT) {
6971     movl(rtmp, value);
6972     shll(rtmp, 16);
6973     orl(value, rtmp);
6974   }
6975 
6976   cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element
6977   jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp
6978   if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) {
6979     // align source address at 4 bytes address boundary
6980     if (t == T_BYTE) {
6981       // One byte misalignment happens only for byte arrays
6982       testptr(to, 1);
6983       jccb(Assembler::zero, L_skip_align1);
6984       movb(Address(to, 0), value);
6985       increment(to);
6986       decrement(count);
6987       BIND(L_skip_align1);
6988     }
6989     // Two bytes misalignment happens only for byte and short (char) arrays
6990     testptr(to, 2);
6991     jccb(Assembler::zero, L_skip_align2);
6992     movw(Address(to, 0), value);
6993     addptr(to, 2);
6994     subl(count, 1<<(shift-1));
6995     BIND(L_skip_align2);
6996   }
6997   if (UseSSE < 2) {
6998     Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
6999     // Fill 32-byte chunks
7000     subl(count, 8 << shift);
7001     jcc(Assembler::less, L_check_fill_8_bytes);
7002     align(16);
7003 
7004     BIND(L_fill_32_bytes_loop);
7005 
7006     for (int i = 0; i < 32; i += 4) {
7007       movl(Address(to, i), value);
7008     }
7009 
7010     addptr(to, 32);
7011     subl(count, 8 << shift);
7012     jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
7013     BIND(L_check_fill_8_bytes);
7014     addl(count, 8 << shift);
7015     jccb(Assembler::zero, L_exit);
7016     jmpb(L_fill_8_bytes);
7017 
7018     //
7019     // length is too short, just fill qwords
7020     //
7021     BIND(L_fill_8_bytes_loop);
7022     movl(Address(to, 0), value);
7023     movl(Address(to, 4), value);
7024     addptr(to, 8);
7025     BIND(L_fill_8_bytes);
7026     subl(count, 1 << (shift + 1));
7027     jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
7028     // fall through to fill 4 bytes
7029   } else {
7030     Label L_fill_32_bytes;
7031     if (!UseUnalignedLoadStores) {
7032       // align to 8 bytes, we know we are 4 byte aligned to start
7033       testptr(to, 4);
7034       jccb(Assembler::zero, L_fill_32_bytes);
7035       movl(Address(to, 0), value);
7036       addptr(to, 4);
7037       subl(count, 1<<shift);
7038     }
7039     BIND(L_fill_32_bytes);
7040     {
7041       assert( UseSSE >= 2, "supported cpu only" );
7042       Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
7043       movdl(xtmp, value);
7044       if (UseAVX >= 2 && UseUnalignedLoadStores) {
7045         // Fill 64-byte chunks
7046         Label L_fill_64_bytes_loop, L_check_fill_32_bytes;
7047         vpbroadcastd(xtmp, xtmp);
7048 
7049         subl(count, 16 << shift);
7050         jcc(Assembler::less, L_check_fill_32_bytes);
7051         align(16);
7052 
7053         BIND(L_fill_64_bytes_loop);
7054         vmovdqu(Address(to, 0), xtmp);
7055         vmovdqu(Address(to, 32), xtmp);
7056         addptr(to, 64);
7057         subl(count, 16 << shift);
7058         jcc(Assembler::greaterEqual, L_fill_64_bytes_loop);
7059 
7060         BIND(L_check_fill_32_bytes);
7061         addl(count, 8 << shift);
7062         jccb(Assembler::less, L_check_fill_8_bytes);
7063         vmovdqu(Address(to, 0), xtmp);
7064         addptr(to, 32);
7065         subl(count, 8 << shift);
7066 
7067         BIND(L_check_fill_8_bytes);
7068         // clean upper bits of YMM registers
7069         movdl(xtmp, value);
7070         pshufd(xtmp, xtmp, 0);
7071       } else {
7072         // Fill 32-byte chunks
7073         pshufd(xtmp, xtmp, 0);
7074 
7075         subl(count, 8 << shift);
7076         jcc(Assembler::less, L_check_fill_8_bytes);
7077         align(16);
7078 
7079         BIND(L_fill_32_bytes_loop);
7080 
7081         if (UseUnalignedLoadStores) {
7082           movdqu(Address(to, 0), xtmp);
7083           movdqu(Address(to, 16), xtmp);
7084         } else {
7085           movq(Address(to, 0), xtmp);
7086           movq(Address(to, 8), xtmp);
7087           movq(Address(to, 16), xtmp);
7088           movq(Address(to, 24), xtmp);
7089         }
7090 
7091         addptr(to, 32);
7092         subl(count, 8 << shift);
7093         jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
7094 
7095         BIND(L_check_fill_8_bytes);
7096       }
7097       addl(count, 8 << shift);
7098       jccb(Assembler::zero, L_exit);
7099       jmpb(L_fill_8_bytes);
7100 
7101       //
7102       // length is too short, just fill qwords
7103       //
7104       BIND(L_fill_8_bytes_loop);
7105       movq(Address(to, 0), xtmp);
7106       addptr(to, 8);
7107       BIND(L_fill_8_bytes);
7108       subl(count, 1 << (shift + 1));
7109       jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
7110     }
7111   }
7112   // fill trailing 4 bytes
7113   BIND(L_fill_4_bytes);
7114   testl(count, 1<<shift);
7115   jccb(Assembler::zero, L_fill_2_bytes);
7116   movl(Address(to, 0), value);
7117   if (t == T_BYTE || t == T_SHORT) {
7118     addptr(to, 4);
7119     BIND(L_fill_2_bytes);
7120     // fill trailing 2 bytes
7121     testl(count, 1<<(shift-1));
7122     jccb(Assembler::zero, L_fill_byte);
7123     movw(Address(to, 0), value);
7124     if (t == T_BYTE) {
7125       addptr(to, 2);
7126       BIND(L_fill_byte);
7127       // fill trailing byte
7128       testl(count, 1);
7129       jccb(Assembler::zero, L_exit);
7130       movb(Address(to, 0), value);
7131     } else {
7132       BIND(L_fill_byte);
7133     }
7134   } else {
7135     BIND(L_fill_2_bytes);
7136   }
7137   BIND(L_exit);
7138 }
7139 
7140 // encode char[] to byte[] in ISO_8859_1
7141 void MacroAssembler::encode_iso_array(Register src, Register dst, Register len,
7142                                       XMMRegister tmp1Reg, XMMRegister tmp2Reg,
7143                                       XMMRegister tmp3Reg, XMMRegister tmp4Reg,
7144                                       Register tmp5, Register result) {
7145   // rsi: src
7146   // rdi: dst
7147   // rdx: len
7148   // rcx: tmp5
7149   // rax: result
7150   ShortBranchVerifier sbv(this);
7151   assert_different_registers(src, dst, len, tmp5, result);
7152   Label L_done, L_copy_1_char, L_copy_1_char_exit;
7153 
7154   // set result
7155   xorl(result, result);
7156   // check for zero length
7157   testl(len, len);
7158   jcc(Assembler::zero, L_done);
7159   movl(result, len);
7160 
7161   // Setup pointers
7162   lea(src, Address(src, len, Address::times_2)); // char[]
7163   lea(dst, Address(dst, len, Address::times_1)); // byte[]
7164   negptr(len);
7165 
7166   if (UseSSE42Intrinsics || UseAVX >= 2) {
7167     Label L_chars_8_check, L_copy_8_chars, L_copy_8_chars_exit;
7168     Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit;
7169 
7170     if (UseAVX >= 2) {
7171       Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit;
7172       movl(tmp5, 0xff00ff00);   // create mask to test for Unicode chars in vector
7173       movdl(tmp1Reg, tmp5);
7174       vpbroadcastd(tmp1Reg, tmp1Reg);
7175       jmpb(L_chars_32_check);
7176 
7177       bind(L_copy_32_chars);
7178       vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64));
7179       vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32));
7180       vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector256 */ true);
7181       vptest(tmp2Reg, tmp1Reg);       // check for Unicode chars in  vector
7182       jccb(Assembler::notZero, L_copy_32_chars_exit);
7183       vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector256 */ true);
7184       vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector256 */ true);
7185       vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg);
7186 
7187       bind(L_chars_32_check);
7188       addptr(len, 32);
7189       jccb(Assembler::lessEqual, L_copy_32_chars);
7190 
7191       bind(L_copy_32_chars_exit);
7192       subptr(len, 16);
7193       jccb(Assembler::greater, L_copy_16_chars_exit);
7194 
7195     } else if (UseSSE42Intrinsics) {
7196       movl(tmp5, 0xff00ff00);   // create mask to test for Unicode chars in vector
7197       movdl(tmp1Reg, tmp5);
7198       pshufd(tmp1Reg, tmp1Reg, 0);
7199       jmpb(L_chars_16_check);
7200     }
7201 
7202     bind(L_copy_16_chars);
7203     if (UseAVX >= 2) {
7204       vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32));
7205       vptest(tmp2Reg, tmp1Reg);
7206       jccb(Assembler::notZero, L_copy_16_chars_exit);
7207       vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector256 */ true);
7208       vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector256 */ true);
7209     } else {
7210       if (UseAVX > 0) {
7211         movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
7212         movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
7213         vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector256 */ false);
7214       } else {
7215         movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
7216         por(tmp2Reg, tmp3Reg);
7217         movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
7218         por(tmp2Reg, tmp4Reg);
7219       }
7220       ptest(tmp2Reg, tmp1Reg);       // check for Unicode chars in  vector
7221       jccb(Assembler::notZero, L_copy_16_chars_exit);
7222       packuswb(tmp3Reg, tmp4Reg);
7223     }
7224     movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg);
7225 
7226     bind(L_chars_16_check);
7227     addptr(len, 16);
7228     jccb(Assembler::lessEqual, L_copy_16_chars);
7229 
7230     bind(L_copy_16_chars_exit);
7231     if (UseAVX >= 2) {
7232       // clean upper bits of YMM registers
7233       vpxor(tmp2Reg, tmp2Reg);
7234       vpxor(tmp3Reg, tmp3Reg);
7235       vpxor(tmp4Reg, tmp4Reg);
7236       movdl(tmp1Reg, tmp5);
7237       pshufd(tmp1Reg, tmp1Reg, 0);
7238     }
7239     subptr(len, 8);
7240     jccb(Assembler::greater, L_copy_8_chars_exit);
7241 
7242     bind(L_copy_8_chars);
7243     movdqu(tmp3Reg, Address(src, len, Address::times_2, -16));
7244     ptest(tmp3Reg, tmp1Reg);
7245     jccb(Assembler::notZero, L_copy_8_chars_exit);
7246     packuswb(tmp3Reg, tmp1Reg);
7247     movq(Address(dst, len, Address::times_1, -8), tmp3Reg);
7248     addptr(len, 8);
7249     jccb(Assembler::lessEqual, L_copy_8_chars);
7250 
7251     bind(L_copy_8_chars_exit);
7252     subptr(len, 8);
7253     jccb(Assembler::zero, L_done);
7254   }
7255 
7256   bind(L_copy_1_char);
7257   load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0));
7258   testl(tmp5, 0xff00);      // check if Unicode char
7259   jccb(Assembler::notZero, L_copy_1_char_exit);
7260   movb(Address(dst, len, Address::times_1, 0), tmp5);
7261   addptr(len, 1);
7262   jccb(Assembler::less, L_copy_1_char);
7263 
7264   bind(L_copy_1_char_exit);
7265   addptr(result, len); // len is negative count of not processed elements
7266   bind(L_done);
7267 }
7268 
7269 #ifdef _LP64
7270 /**
7271  * Helper for multiply_to_len().
7272  */
7273 void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo, Register src1, Register src2) {
7274   addq(dest_lo, src1);
7275   adcq(dest_hi, 0);
7276   addq(dest_lo, src2);
7277   adcq(dest_hi, 0);
7278 }
7279 
7280 /**
7281  * Multiply 64 bit by 64 bit first loop.
7282  */
7283 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart,
7284                                            Register y, Register y_idx, Register z,
7285                                            Register carry, Register product,
7286                                            Register idx, Register kdx) {
7287   //
7288   //  jlong carry, x[], y[], z[];
7289   //  for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
7290   //    huge_128 product = y[idx] * x[xstart] + carry;
7291   //    z[kdx] = (jlong)product;
7292   //    carry  = (jlong)(product >>> 64);
7293   //  }
7294   //  z[xstart] = carry;
7295   //
7296 
7297   Label L_first_loop, L_first_loop_exit;
7298   Label L_one_x, L_one_y, L_multiply;
7299 
7300   decrementl(xstart);
7301   jcc(Assembler::negative, L_one_x);
7302 
7303   movq(x_xstart, Address(x, xstart, Address::times_4,  0));
7304   rorq(x_xstart, 32); // convert big-endian to little-endian
7305 
7306   bind(L_first_loop);
7307   decrementl(idx);
7308   jcc(Assembler::negative, L_first_loop_exit);
7309   decrementl(idx);
7310   jcc(Assembler::negative, L_one_y);
7311   movq(y_idx, Address(y, idx, Address::times_4,  0));
7312   rorq(y_idx, 32); // convert big-endian to little-endian
7313   bind(L_multiply);
7314   movq(product, x_xstart);
7315   mulq(y_idx); // product(rax) * y_idx -> rdx:rax
7316   addq(product, carry);
7317   adcq(rdx, 0);
7318   subl(kdx, 2);
7319   movl(Address(z, kdx, Address::times_4,  4), product);
7320   shrq(product, 32);
7321   movl(Address(z, kdx, Address::times_4,  0), product);
7322   movq(carry, rdx);
7323   jmp(L_first_loop);
7324 
7325   bind(L_one_y);
7326   movl(y_idx, Address(y,  0));
7327   jmp(L_multiply);
7328 
7329   bind(L_one_x);
7330   movl(x_xstart, Address(x,  0));
7331   jmp(L_first_loop);
7332 
7333   bind(L_first_loop_exit);
7334 }
7335 
7336 /**
7337  * Multiply 64 bit by 64 bit and add 128 bit.
7338  */
7339 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y, Register z,
7340                                             Register yz_idx, Register idx,
7341                                             Register carry, Register product, int offset) {
7342   //     huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry;
7343   //     z[kdx] = (jlong)product;
7344 
7345   movq(yz_idx, Address(y, idx, Address::times_4,  offset));
7346   rorq(yz_idx, 32); // convert big-endian to little-endian
7347   movq(product, x_xstart);
7348   mulq(yz_idx);     // product(rax) * yz_idx -> rdx:product(rax)
7349   movq(yz_idx, Address(z, idx, Address::times_4,  offset));
7350   rorq(yz_idx, 32); // convert big-endian to little-endian
7351 
7352   add2_with_carry(rdx, product, carry, yz_idx);
7353 
7354   movl(Address(z, idx, Address::times_4,  offset+4), product);
7355   shrq(product, 32);
7356   movl(Address(z, idx, Address::times_4,  offset), product);
7357 
7358 }
7359 
7360 /**
7361  * Multiply 128 bit by 128 bit. Unrolled inner loop.
7362  */
7363 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart, Register y, Register z,
7364                                              Register yz_idx, Register idx, Register jdx,
7365                                              Register carry, Register product,
7366                                              Register carry2) {
7367   //   jlong carry, x[], y[], z[];
7368   //   int kdx = ystart+1;
7369   //   for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
7370   //     huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry;
7371   //     z[kdx+idx+1] = (jlong)product;
7372   //     jlong carry2  = (jlong)(product >>> 64);
7373   //     product = (y[idx] * x_xstart) + z[kdx+idx] + carry2;
7374   //     z[kdx+idx] = (jlong)product;
7375   //     carry  = (jlong)(product >>> 64);
7376   //   }
7377   //   idx += 2;
7378   //   if (idx > 0) {
7379   //     product = (y[idx] * x_xstart) + z[kdx+idx] + carry;
7380   //     z[kdx+idx] = (jlong)product;
7381   //     carry  = (jlong)(product >>> 64);
7382   //   }
7383   //
7384 
7385   Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
7386 
7387   movl(jdx, idx);
7388   andl(jdx, 0xFFFFFFFC);
7389   shrl(jdx, 2);
7390 
7391   bind(L_third_loop);
7392   subl(jdx, 1);
7393   jcc(Assembler::negative, L_third_loop_exit);
7394   subl(idx, 4);
7395 
7396   multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8);
7397   movq(carry2, rdx);
7398 
7399   multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0);
7400   movq(carry, rdx);
7401   jmp(L_third_loop);
7402 
7403   bind (L_third_loop_exit);
7404 
7405   andl (idx, 0x3);
7406   jcc(Assembler::zero, L_post_third_loop_done);
7407 
7408   Label L_check_1;
7409   subl(idx, 2);
7410   jcc(Assembler::negative, L_check_1);
7411 
7412   multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0);
7413   movq(carry, rdx);
7414 
7415   bind (L_check_1);
7416   addl (idx, 0x2);
7417   andl (idx, 0x1);
7418   subl(idx, 1);
7419   jcc(Assembler::negative, L_post_third_loop_done);
7420 
7421   movl(yz_idx, Address(y, idx, Address::times_4,  0));
7422   movq(product, x_xstart);
7423   mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
7424   movl(yz_idx, Address(z, idx, Address::times_4,  0));
7425 
7426   add2_with_carry(rdx, product, yz_idx, carry);
7427 
7428   movl(Address(z, idx, Address::times_4,  0), product);
7429   shrq(product, 32);
7430 
7431   shlq(rdx, 32);
7432   orq(product, rdx);
7433   movq(carry, product);
7434 
7435   bind(L_post_third_loop_done);
7436 }
7437 
7438 /**
7439  * Multiply 128 bit by 128 bit using BMI2. Unrolled inner loop.
7440  *
7441  */
7442 void MacroAssembler::multiply_128_x_128_bmi2_loop(Register y, Register z,
7443                                                   Register carry, Register carry2,
7444                                                   Register idx, Register jdx,
7445                                                   Register yz_idx1, Register yz_idx2,
7446                                                   Register tmp, Register tmp3, Register tmp4) {
7447   assert(UseBMI2Instructions, "should be used only when BMI2 is available");
7448 
7449   //   jlong carry, x[], y[], z[];
7450   //   int kdx = ystart+1;
7451   //   for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
7452   //     huge_128 tmp3 = (y[idx+1] * rdx) + z[kdx+idx+1] + carry;
7453   //     jlong carry2  = (jlong)(tmp3 >>> 64);
7454   //     huge_128 tmp4 = (y[idx]   * rdx) + z[kdx+idx] + carry2;
7455   //     carry  = (jlong)(tmp4 >>> 64);
7456   //     z[kdx+idx+1] = (jlong)tmp3;
7457   //     z[kdx+idx] = (jlong)tmp4;
7458   //   }
7459   //   idx += 2;
7460   //   if (idx > 0) {
7461   //     yz_idx1 = (y[idx] * rdx) + z[kdx+idx] + carry;
7462   //     z[kdx+idx] = (jlong)yz_idx1;
7463   //     carry  = (jlong)(yz_idx1 >>> 64);
7464   //   }
7465   //
7466 
7467   Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
7468 
7469   movl(jdx, idx);
7470   andl(jdx, 0xFFFFFFFC);
7471   shrl(jdx, 2);
7472 
7473   bind(L_third_loop);
7474   subl(jdx, 1);
7475   jcc(Assembler::negative, L_third_loop_exit);
7476   subl(idx, 4);
7477 
7478   movq(yz_idx1,  Address(y, idx, Address::times_4,  8));
7479   rorxq(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian
7480   movq(yz_idx2, Address(y, idx, Address::times_4,  0));
7481   rorxq(yz_idx2, yz_idx2, 32);
7482 
7483   mulxq(tmp4, tmp3, yz_idx1);  //  yz_idx1 * rdx -> tmp4:tmp3
7484   mulxq(carry2, tmp, yz_idx2); //  yz_idx2 * rdx -> carry2:tmp
7485 
7486   movq(yz_idx1,  Address(z, idx, Address::times_4,  8));
7487   rorxq(yz_idx1, yz_idx1, 32);
7488   movq(yz_idx2, Address(z, idx, Address::times_4,  0));
7489   rorxq(yz_idx2, yz_idx2, 32);
7490 
7491   if (VM_Version::supports_adx()) {
7492     adcxq(tmp3, carry);
7493     adoxq(tmp3, yz_idx1);
7494 
7495     adcxq(tmp4, tmp);
7496     adoxq(tmp4, yz_idx2);
7497 
7498     movl(carry, 0); // does not affect flags
7499     adcxq(carry2, carry);
7500     adoxq(carry2, carry);
7501   } else {
7502     add2_with_carry(tmp4, tmp3, carry, yz_idx1);
7503     add2_with_carry(carry2, tmp4, tmp, yz_idx2);
7504   }
7505   movq(carry, carry2);
7506 
7507   movl(Address(z, idx, Address::times_4, 12), tmp3);
7508   shrq(tmp3, 32);
7509   movl(Address(z, idx, Address::times_4,  8), tmp3);
7510 
7511   movl(Address(z, idx, Address::times_4,  4), tmp4);
7512   shrq(tmp4, 32);
7513   movl(Address(z, idx, Address::times_4,  0), tmp4);
7514 
7515   jmp(L_third_loop);
7516 
7517   bind (L_third_loop_exit);
7518 
7519   andl (idx, 0x3);
7520   jcc(Assembler::zero, L_post_third_loop_done);
7521 
7522   Label L_check_1;
7523   subl(idx, 2);
7524   jcc(Assembler::negative, L_check_1);
7525 
7526   movq(yz_idx1, Address(y, idx, Address::times_4,  0));
7527   rorxq(yz_idx1, yz_idx1, 32);
7528   mulxq(tmp4, tmp3, yz_idx1); //  yz_idx1 * rdx -> tmp4:tmp3
7529   movq(yz_idx2, Address(z, idx, Address::times_4,  0));
7530   rorxq(yz_idx2, yz_idx2, 32);
7531 
7532   add2_with_carry(tmp4, tmp3, carry, yz_idx2);
7533 
7534   movl(Address(z, idx, Address::times_4,  4), tmp3);
7535   shrq(tmp3, 32);
7536   movl(Address(z, idx, Address::times_4,  0), tmp3);
7537   movq(carry, tmp4);
7538 
7539   bind (L_check_1);
7540   addl (idx, 0x2);
7541   andl (idx, 0x1);
7542   subl(idx, 1);
7543   jcc(Assembler::negative, L_post_third_loop_done);
7544   movl(tmp4, Address(y, idx, Address::times_4,  0));
7545   mulxq(carry2, tmp3, tmp4);  //  tmp4 * rdx -> carry2:tmp3
7546   movl(tmp4, Address(z, idx, Address::times_4,  0));
7547 
7548   add2_with_carry(carry2, tmp3, tmp4, carry);
7549 
7550   movl(Address(z, idx, Address::times_4,  0), tmp3);
7551   shrq(tmp3, 32);
7552 
7553   shlq(carry2, 32);
7554   orq(tmp3, carry2);
7555   movq(carry, tmp3);
7556 
7557   bind(L_post_third_loop_done);
7558 }
7559 
7560 /**
7561  * Code for BigInteger::multiplyToLen() instrinsic.
7562  *
7563  * rdi: x
7564  * rax: xlen
7565  * rsi: y
7566  * rcx: ylen
7567  * r8:  z
7568  * r11: zlen
7569  * r12: tmp1
7570  * r13: tmp2
7571  * r14: tmp3
7572  * r15: tmp4
7573  * rbx: tmp5
7574  *
7575  */
7576 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, Register z, Register zlen,
7577                                      Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5) {
7578   ShortBranchVerifier sbv(this);
7579   assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx);
7580 
7581   push(tmp1);
7582   push(tmp2);
7583   push(tmp3);
7584   push(tmp4);
7585   push(tmp5);
7586 
7587   push(xlen);
7588   push(zlen);
7589 
7590   const Register idx = tmp1;
7591   const Register kdx = tmp2;
7592   const Register xstart = tmp3;
7593 
7594   const Register y_idx = tmp4;
7595   const Register carry = tmp5;
7596   const Register product  = xlen;
7597   const Register x_xstart = zlen;  // reuse register
7598 
7599   // First Loop.
7600   //
7601   //  final static long LONG_MASK = 0xffffffffL;
7602   //  int xstart = xlen - 1;
7603   //  int ystart = ylen - 1;
7604   //  long carry = 0;
7605   //  for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
7606   //    long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry;
7607   //    z[kdx] = (int)product;
7608   //    carry = product >>> 32;
7609   //  }
7610   //  z[xstart] = (int)carry;
7611   //
7612 
7613   movl(idx, ylen);      // idx = ylen;
7614   movl(kdx, zlen);      // kdx = xlen+ylen;
7615   xorq(carry, carry);   // carry = 0;
7616 
7617   Label L_done;
7618 
7619   movl(xstart, xlen);
7620   decrementl(xstart);
7621   jcc(Assembler::negative, L_done);
7622 
7623   multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx);
7624 
7625   Label L_second_loop;
7626   testl(kdx, kdx);
7627   jcc(Assembler::zero, L_second_loop);
7628 
7629   Label L_carry;
7630   subl(kdx, 1);
7631   jcc(Assembler::zero, L_carry);
7632 
7633   movl(Address(z, kdx, Address::times_4,  0), carry);
7634   shrq(carry, 32);
7635   subl(kdx, 1);
7636 
7637   bind(L_carry);
7638   movl(Address(z, kdx, Address::times_4,  0), carry);
7639 
7640   // Second and third (nested) loops.
7641   //
7642   // for (int i = xstart-1; i >= 0; i--) { // Second loop
7643   //   carry = 0;
7644   //   for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop
7645   //     long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) +
7646   //                    (z[k] & LONG_MASK) + carry;
7647   //     z[k] = (int)product;
7648   //     carry = product >>> 32;
7649   //   }
7650   //   z[i] = (int)carry;
7651   // }
7652   //
7653   // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx
7654 
7655   const Register jdx = tmp1;
7656 
7657   bind(L_second_loop);
7658   xorl(carry, carry);    // carry = 0;
7659   movl(jdx, ylen);       // j = ystart+1
7660 
7661   subl(xstart, 1);       // i = xstart-1;
7662   jcc(Assembler::negative, L_done);
7663 
7664   push (z);
7665 
7666   Label L_last_x;
7667   lea(z, Address(z, xstart, Address::times_4, 4)); // z = z + k - j
7668   subl(xstart, 1);       // i = xstart-1;
7669   jcc(Assembler::negative, L_last_x);
7670 
7671   if (UseBMI2Instructions) {
7672     movq(rdx,  Address(x, xstart, Address::times_4,  0));
7673     rorxq(rdx, rdx, 32); // convert big-endian to little-endian
7674   } else {
7675     movq(x_xstart, Address(x, xstart, Address::times_4,  0));
7676     rorq(x_xstart, 32);  // convert big-endian to little-endian
7677   }
7678 
7679   Label L_third_loop_prologue;
7680   bind(L_third_loop_prologue);
7681 
7682   push (x);
7683   push (xstart);
7684   push (ylen);
7685 
7686 
7687   if (UseBMI2Instructions) {
7688     multiply_128_x_128_bmi2_loop(y, z, carry, x, jdx, ylen, product, tmp2, x_xstart, tmp3, tmp4);
7689   } else { // !UseBMI2Instructions
7690     multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x);
7691   }
7692 
7693   pop(ylen);
7694   pop(xlen);
7695   pop(x);
7696   pop(z);
7697 
7698   movl(tmp3, xlen);
7699   addl(tmp3, 1);
7700   movl(Address(z, tmp3, Address::times_4,  0), carry);
7701   subl(tmp3, 1);
7702   jccb(Assembler::negative, L_done);
7703 
7704   shrq(carry, 32);
7705   movl(Address(z, tmp3, Address::times_4,  0), carry);
7706   jmp(L_second_loop);
7707 
7708   // Next infrequent code is moved outside loops.
7709   bind(L_last_x);
7710   if (UseBMI2Instructions) {
7711     movl(rdx, Address(x,  0));
7712   } else {
7713     movl(x_xstart, Address(x,  0));
7714   }
7715   jmp(L_third_loop_prologue);
7716 
7717   bind(L_done);
7718 
7719   pop(zlen);
7720   pop(xlen);
7721 
7722   pop(tmp5);
7723   pop(tmp4);
7724   pop(tmp3);
7725   pop(tmp2);
7726   pop(tmp1);
7727 }
7728 #endif
7729 
7730 /**
7731  * Emits code to update CRC-32 with a byte value according to constants in table
7732  *
7733  * @param [in,out]crc   Register containing the crc.
7734  * @param [in]val       Register containing the byte to fold into the CRC.
7735  * @param [in]table     Register containing the table of crc constants.
7736  *
7737  * uint32_t crc;
7738  * val = crc_table[(val ^ crc) & 0xFF];
7739  * crc = val ^ (crc >> 8);
7740  *
7741  */
7742 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) {
7743   xorl(val, crc);
7744   andl(val, 0xFF);
7745   shrl(crc, 8); // unsigned shift
7746   xorl(crc, Address(table, val, Address::times_4, 0));
7747 }
7748 
7749 /**
7750  * Fold 128-bit data chunk
7751  */
7752 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) {
7753   if (UseAVX > 0) {
7754     vpclmulhdq(xtmp, xK, xcrc); // [123:64]
7755     vpclmulldq(xcrc, xK, xcrc); // [63:0]
7756     vpxor(xcrc, xcrc, Address(buf, offset), false /* vector256 */);
7757     pxor(xcrc, xtmp);
7758   } else {
7759     movdqa(xtmp, xcrc);
7760     pclmulhdq(xtmp, xK);   // [123:64]
7761     pclmulldq(xcrc, xK);   // [63:0]
7762     pxor(xcrc, xtmp);
7763     movdqu(xtmp, Address(buf, offset));
7764     pxor(xcrc, xtmp);
7765   }
7766 }
7767 
7768 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) {
7769   if (UseAVX > 0) {
7770     vpclmulhdq(xtmp, xK, xcrc);
7771     vpclmulldq(xcrc, xK, xcrc);
7772     pxor(xcrc, xbuf);
7773     pxor(xcrc, xtmp);
7774   } else {
7775     movdqa(xtmp, xcrc);
7776     pclmulhdq(xtmp, xK);
7777     pclmulldq(xcrc, xK);
7778     pxor(xcrc, xbuf);
7779     pxor(xcrc, xtmp);
7780   }
7781 }
7782 
7783 /**
7784  * 8-bit folds to compute 32-bit CRC
7785  *
7786  * uint64_t xcrc;
7787  * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8);
7788  */
7789 void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) {
7790   movdl(tmp, xcrc);
7791   andl(tmp, 0xFF);
7792   movdl(xtmp, Address(table, tmp, Address::times_4, 0));
7793   psrldq(xcrc, 1); // unsigned shift one byte
7794   pxor(xcrc, xtmp);
7795 }
7796 
7797 /**
7798  * uint32_t crc;
7799  * timesXtoThe32[crc & 0xFF] ^ (crc >> 8);
7800  */
7801 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) {
7802   movl(tmp, crc);
7803   andl(tmp, 0xFF);
7804   shrl(crc, 8);
7805   xorl(crc, Address(table, tmp, Address::times_4, 0));
7806 }
7807 
7808 /**
7809  * @param crc   register containing existing CRC (32-bit)
7810  * @param buf   register pointing to input byte buffer (byte*)
7811  * @param len   register containing number of bytes
7812  * @param table register that will contain address of CRC table
7813  * @param tmp   scratch register
7814  */
7815 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) {
7816   assert_different_registers(crc, buf, len, table, tmp, rax);
7817 
7818   Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned;
7819   Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop;
7820 
7821   lea(table, ExternalAddress(StubRoutines::crc_table_addr()));
7822   notl(crc); // ~crc
7823   cmpl(len, 16);
7824   jcc(Assembler::less, L_tail);
7825 
7826   // Align buffer to 16 bytes
7827   movl(tmp, buf);
7828   andl(tmp, 0xF);
7829   jccb(Assembler::zero, L_aligned);
7830   subl(tmp,  16);
7831   addl(len, tmp);
7832 
7833   align(4);
7834   BIND(L_align_loop);
7835   movsbl(rax, Address(buf, 0)); // load byte with sign extension
7836   update_byte_crc32(crc, rax, table);
7837   increment(buf);
7838   incrementl(tmp);
7839   jccb(Assembler::less, L_align_loop);
7840 
7841   BIND(L_aligned);
7842   movl(tmp, len); // save
7843   shrl(len, 4);
7844   jcc(Assembler::zero, L_tail_restore);
7845 
7846   // Fold crc into first bytes of vector
7847   movdqa(xmm1, Address(buf, 0));
7848   movdl(rax, xmm1);
7849   xorl(crc, rax);
7850   pinsrd(xmm1, crc, 0);
7851   addptr(buf, 16);
7852   subl(len, 4); // len > 0
7853   jcc(Assembler::less, L_fold_tail);
7854 
7855   movdqa(xmm2, Address(buf,  0));
7856   movdqa(xmm3, Address(buf, 16));
7857   movdqa(xmm4, Address(buf, 32));
7858   addptr(buf, 48);
7859   subl(len, 3);
7860   jcc(Assembler::lessEqual, L_fold_512b);
7861 
7862   // Fold total 512 bits of polynomial on each iteration,
7863   // 128 bits per each of 4 parallel streams.
7864   movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32));
7865 
7866   align(32);
7867   BIND(L_fold_512b_loop);
7868   fold_128bit_crc32(xmm1, xmm0, xmm5, buf,  0);
7869   fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16);
7870   fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32);
7871   fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48);
7872   addptr(buf, 64);
7873   subl(len, 4);
7874   jcc(Assembler::greater, L_fold_512b_loop);
7875 
7876   // Fold 512 bits to 128 bits.
7877   BIND(L_fold_512b);
7878   movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
7879   fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2);
7880   fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3);
7881   fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4);
7882 
7883   // Fold the rest of 128 bits data chunks
7884   BIND(L_fold_tail);
7885   addl(len, 3);
7886   jccb(Assembler::lessEqual, L_fold_128b);
7887   movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
7888 
7889   BIND(L_fold_tail_loop);
7890   fold_128bit_crc32(xmm1, xmm0, xmm5, buf,  0);
7891   addptr(buf, 16);
7892   decrementl(len);
7893   jccb(Assembler::greater, L_fold_tail_loop);
7894 
7895   // Fold 128 bits in xmm1 down into 32 bits in crc register.
7896   BIND(L_fold_128b);
7897   movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr()));
7898   if (UseAVX > 0) {
7899     vpclmulqdq(xmm2, xmm0, xmm1, 0x1);
7900     vpand(xmm3, xmm0, xmm2, false /* vector256 */);
7901     vpclmulqdq(xmm0, xmm0, xmm3, 0x1);
7902   } else {
7903     movdqa(xmm2, xmm0);
7904     pclmulqdq(xmm2, xmm1, 0x1);
7905     movdqa(xmm3, xmm0);
7906     pand(xmm3, xmm2);
7907     pclmulqdq(xmm0, xmm3, 0x1);
7908   }
7909   psrldq(xmm1, 8);
7910   psrldq(xmm2, 4);
7911   pxor(xmm0, xmm1);
7912   pxor(xmm0, xmm2);
7913 
7914   // 8 8-bit folds to compute 32-bit CRC.
7915   for (int j = 0; j < 4; j++) {
7916     fold_8bit_crc32(xmm0, table, xmm1, rax);
7917   }
7918   movdl(crc, xmm0); // mov 32 bits to general register
7919   for (int j = 0; j < 4; j++) {
7920     fold_8bit_crc32(crc, table, rax);
7921   }
7922 
7923   BIND(L_tail_restore);
7924   movl(len, tmp); // restore
7925   BIND(L_tail);
7926   andl(len, 0xf);
7927   jccb(Assembler::zero, L_exit);
7928 
7929   // Fold the rest of bytes
7930   align(4);
7931   BIND(L_tail_loop);
7932   movsbl(rax, Address(buf, 0)); // load byte with sign extension
7933   update_byte_crc32(crc, rax, table);
7934   increment(buf);
7935   decrementl(len);
7936   jccb(Assembler::greater, L_tail_loop);
7937 
7938   BIND(L_exit);
7939   notl(crc); // ~c
7940 }
7941 
7942 #undef BIND
7943 #undef BLOCK_COMMENT
7944 
7945 
7946 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) {
7947   switch (cond) {
7948     // Note some conditions are synonyms for others
7949     case Assembler::zero:         return Assembler::notZero;
7950     case Assembler::notZero:      return Assembler::zero;
7951     case Assembler::less:         return Assembler::greaterEqual;
7952     case Assembler::lessEqual:    return Assembler::greater;
7953     case Assembler::greater:      return Assembler::lessEqual;
7954     case Assembler::greaterEqual: return Assembler::less;
7955     case Assembler::below:        return Assembler::aboveEqual;
7956     case Assembler::belowEqual:   return Assembler::above;
7957     case Assembler::above:        return Assembler::belowEqual;
7958     case Assembler::aboveEqual:   return Assembler::below;
7959     case Assembler::overflow:     return Assembler::noOverflow;
7960     case Assembler::noOverflow:   return Assembler::overflow;
7961     case Assembler::negative:     return Assembler::positive;
7962     case Assembler::positive:     return Assembler::negative;
7963     case Assembler::parity:       return Assembler::noParity;
7964     case Assembler::noParity:     return Assembler::parity;
7965   }
7966   ShouldNotReachHere(); return Assembler::overflow;
7967 }
7968 
7969 SkipIfEqual::SkipIfEqual(
7970     MacroAssembler* masm, const bool* flag_addr, bool value) {
7971   _masm = masm;
7972   _masm->cmp8(ExternalAddress((address)flag_addr), value);
7973   _masm->jcc(Assembler::equal, _label);
7974 }
7975 
7976 SkipIfEqual::~SkipIfEqual() {
7977   _masm->bind(_label);
7978 }