1 /*
   2  * Copyright (c) 2003, 2026, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright (c) 2014, Red Hat Inc. All rights reserved.
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This code is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 only, as
   8  * published by the Free Software Foundation.
   9  *
  10  * This code is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  13  * version 2 for more details (a copy is included in the LICENSE file that
  14  * accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License version
  17  * 2 along with this work; if not, write to the Free Software Foundation,
  18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  21  * or visit www.oracle.com if you need additional information or have any
  22  * questions.
  23  *
  24  */
  25 
  26 #include "asm/macroAssembler.inline.hpp"
  27 #include "compiler/disassembler.hpp"
  28 #include "compiler/compilerDefinitions.inline.hpp"
  29 #include "gc/shared/barrierSetAssembler.hpp"
  30 #include "gc/shared/collectedHeap.hpp"
  31 #include "gc/shared/tlab_globals.hpp"
  32 #include "interpreter/interpreter.hpp"
  33 #include "interpreter/interpreterRuntime.hpp"
  34 #include "interpreter/interp_masm.hpp"
  35 #include "interpreter/templateTable.hpp"
  36 #include "memory/universe.hpp"
  37 #include "oops/methodData.hpp"
  38 #include "oops/method.inline.hpp"
  39 #include "oops/objArrayKlass.hpp"
  40 #include "oops/oop.inline.hpp"
  41 #include "oops/resolvedFieldEntry.hpp"
  42 #include "oops/resolvedIndyEntry.hpp"
  43 #include "oops/resolvedMethodEntry.hpp"
  44 #include "prims/jvmtiExport.hpp"
  45 #include "prims/methodHandles.hpp"
  46 #include "runtime/arguments.hpp"
  47 #include "runtime/frame.inline.hpp"
  48 #include "runtime/sharedRuntime.hpp"
  49 #include "runtime/stubRoutines.hpp"
  50 #include "runtime/synchronizer.hpp"
  51 #include "utilities/powerOfTwo.hpp"
  52 
  53 #define __ Disassembler::hook<InterpreterMacroAssembler>(__FILE__, __LINE__, _masm)->
  54 
  55 // Address computation: local variables
  56 
  57 static inline Address iaddress(int n) {
  58   return Address(rlocals, Interpreter::local_offset_in_bytes(n));
  59 }
  60 
  61 static inline Address laddress(int n) {
  62   return iaddress(n + 1);
  63 }
  64 
  65 static inline Address faddress(int n) {
  66   return iaddress(n);
  67 }
  68 
  69 static inline Address daddress(int n) {
  70   return laddress(n);
  71 }
  72 
  73 static inline Address aaddress(int n) {
  74   return iaddress(n);
  75 }
  76 
  77 static inline Address iaddress(Register r) {
  78   return Address(rlocals, r, Address::lsl(3));
  79 }
  80 
  81 static inline Address laddress(Register r, Register scratch,
  82                                InterpreterMacroAssembler* _masm) {
  83   __ lea(scratch, Address(rlocals, r, Address::lsl(3)));
  84   return Address(scratch, Interpreter::local_offset_in_bytes(1));
  85 }
  86 
  87 static inline Address faddress(Register r) {
  88   return iaddress(r);
  89 }
  90 
  91 static inline Address daddress(Register r, Register scratch,
  92                                InterpreterMacroAssembler* _masm) {
  93   return laddress(r, scratch, _masm);
  94 }
  95 
  96 static inline Address aaddress(Register r) {
  97   return iaddress(r);
  98 }
  99 
 100 // At top of Java expression stack which may be different than esp().  It
 101 // isn't for category 1 objects.
 102 static inline Address at_tos   () {
 103   return Address(esp,  Interpreter::expr_offset_in_bytes(0));
 104 }
 105 
 106 static inline Address at_tos_p1() {
 107   return Address(esp,  Interpreter::expr_offset_in_bytes(1));
 108 }
 109 
 110 static inline Address at_tos_p2() {
 111   return Address(esp,  Interpreter::expr_offset_in_bytes(2));
 112 }
 113 
 114 static inline Address at_tos_p3() {
 115   return Address(esp,  Interpreter::expr_offset_in_bytes(3));
 116 }
 117 
 118 static inline Address at_tos_p4() {
 119   return Address(esp,  Interpreter::expr_offset_in_bytes(4));
 120 }
 121 
 122 static inline Address at_tos_p5() {
 123   return Address(esp,  Interpreter::expr_offset_in_bytes(5));
 124 }
 125 
 126 // Condition conversion
 127 static Assembler::Condition j_not(TemplateTable::Condition cc) {
 128   switch (cc) {
 129   case TemplateTable::equal        : return Assembler::NE;
 130   case TemplateTable::not_equal    : return Assembler::EQ;
 131   case TemplateTable::less         : return Assembler::GE;
 132   case TemplateTable::less_equal   : return Assembler::GT;
 133   case TemplateTable::greater      : return Assembler::LE;
 134   case TemplateTable::greater_equal: return Assembler::LT;
 135   }
 136   ShouldNotReachHere();
 137   return Assembler::EQ;
 138 }
 139 
 140 
 141 // Miscellaneous helper routines
 142 // Store an oop (or null) at the Address described by obj.
 143 // If val == noreg this means store a null
 144 static void do_oop_store(InterpreterMacroAssembler* _masm,
 145                          Address dst,
 146                          Register val,
 147                          DecoratorSet decorators) {
 148   assert(val == noreg || val == r0, "parameter is just for looks");
 149   __ store_heap_oop(dst, val, r10, r11, r3, decorators);
 150 }
 151 
 152 static void do_oop_load(InterpreterMacroAssembler* _masm,
 153                         Address src,
 154                         Register dst,
 155                         DecoratorSet decorators) {
 156   __ load_heap_oop(dst, src, r10, r11, decorators);
 157 }
 158 
 159 Address TemplateTable::at_bcp(int offset) {
 160   assert(_desc->uses_bcp(), "inconsistent uses_bcp information");
 161   return Address(rbcp, offset);
 162 }
 163 
 164 void TemplateTable::patch_bytecode(Bytecodes::Code bc, Register bc_reg,
 165                                    Register temp_reg, bool load_bc_into_bc_reg/*=true*/,
 166                                    int byte_no)
 167 {
 168   assert_different_registers(bc_reg, temp_reg);
 169   if (!RewriteBytecodes)  return;
 170   Label L_patch_done;
 171 
 172   switch (bc) {
 173   case Bytecodes::_fast_vputfield:
 174   case Bytecodes::_fast_aputfield:
 175   case Bytecodes::_fast_bputfield:
 176   case Bytecodes::_fast_zputfield:
 177   case Bytecodes::_fast_cputfield:
 178   case Bytecodes::_fast_dputfield:
 179   case Bytecodes::_fast_fputfield:
 180   case Bytecodes::_fast_iputfield:
 181   case Bytecodes::_fast_lputfield:
 182   case Bytecodes::_fast_sputfield:
 183     {
 184       // We skip bytecode quickening for putfield instructions when
 185       // the put_code written to the constant pool cache is zero.
 186       // This is required so that every execution of this instruction
 187       // calls out to InterpreterRuntime::resolve_get_put to do
 188       // additional, required work.
 189       assert(byte_no == f1_byte || byte_no == f2_byte, "byte_no out of range");
 190       assert(load_bc_into_bc_reg, "we use bc_reg as temp");
 191       __ load_field_entry(temp_reg, bc_reg);
 192       if (byte_no == f1_byte) {
 193         __ lea(temp_reg, Address(temp_reg, in_bytes(ResolvedFieldEntry::get_code_offset())));
 194       } else {
 195         __ lea(temp_reg, Address(temp_reg, in_bytes(ResolvedFieldEntry::put_code_offset())));
 196       }
 197       // Load-acquire the bytecode to match store-release in ResolvedFieldEntry::fill_in()
 198       __ ldarb(temp_reg, temp_reg);
 199       __ movw(bc_reg, bc);
 200       __ cbzw(temp_reg, L_patch_done);  // don't patch
 201     }
 202     break;
 203   default:
 204     assert(byte_no == -1, "sanity");
 205     // the pair bytecodes have already done the load.
 206     if (load_bc_into_bc_reg) {
 207       __ movw(bc_reg, bc);
 208     }
 209   }
 210 
 211   if (JvmtiExport::can_post_breakpoint()) {
 212     Label L_fast_patch;
 213     // if a breakpoint is present we can't rewrite the stream directly
 214     __ load_unsigned_byte(temp_reg, at_bcp(0));
 215     __ cmpw(temp_reg, Bytecodes::_breakpoint);
 216     __ br(Assembler::NE, L_fast_patch);
 217     // Let breakpoint table handling rewrite to quicker bytecode
 218     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::set_original_bytecode_at), rmethod, rbcp, bc_reg);
 219     __ b(L_patch_done);
 220     __ bind(L_fast_patch);
 221   }
 222 
 223 #ifdef ASSERT
 224   Label L_okay;
 225   __ load_unsigned_byte(temp_reg, at_bcp(0));
 226   __ cmpw(temp_reg, (int) Bytecodes::java_code(bc));
 227   __ br(Assembler::EQ, L_okay);
 228   __ cmpw(temp_reg, bc_reg);
 229   __ br(Assembler::EQ, L_okay);
 230   __ stop("patching the wrong bytecode");
 231   __ bind(L_okay);
 232 #endif
 233   // Patch bytecode with release store to coordinate with ResolvedFieldEntry loads
 234   // in fast bytecode codelets. load_field_entry has a memory barrier that gains
 235   // the needed ordering, together with control dependency on entering the fast codelet
 236   // itself.
 237   __ lea(temp_reg, at_bcp(0));
 238   __ stlrb(bc_reg, temp_reg);
 239   __ bind(L_patch_done);
 240 }
 241 
 242 
 243 // Individual instructions
 244 
 245 void TemplateTable::nop() {
 246   transition(vtos, vtos);
 247   // nothing to do
 248 }
 249 
 250 void TemplateTable::shouldnotreachhere() {
 251   transition(vtos, vtos);
 252   __ stop("shouldnotreachhere bytecode");
 253 }
 254 
 255 void TemplateTable::aconst_null()
 256 {
 257   transition(vtos, atos);
 258   __ mov(r0, 0);
 259 }
 260 
 261 void TemplateTable::iconst(int value)
 262 {
 263   transition(vtos, itos);
 264   __ mov(r0, value);
 265 }
 266 
 267 void TemplateTable::lconst(int value)
 268 {
 269   __ mov(r0, value);
 270 }
 271 
 272 void TemplateTable::fconst(int value)
 273 {
 274   transition(vtos, ftos);
 275   switch (value) {
 276   case 0:
 277     __ fmovs(v0, 0.0);
 278     break;
 279   case 1:
 280     __ fmovs(v0, 1.0);
 281     break;
 282   case 2:
 283     __ fmovs(v0, 2.0);
 284     break;
 285   default:
 286     ShouldNotReachHere();
 287     break;
 288   }
 289 }
 290 
 291 void TemplateTable::dconst(int value)
 292 {
 293   transition(vtos, dtos);
 294   switch (value) {
 295   case 0:
 296     __ fmovd(v0, 0.0);
 297     break;
 298   case 1:
 299     __ fmovd(v0, 1.0);
 300     break;
 301   case 2:
 302     __ fmovd(v0, 2.0);
 303     break;
 304   default:
 305     ShouldNotReachHere();
 306     break;
 307   }
 308 }
 309 
 310 void TemplateTable::bipush()
 311 {
 312   transition(vtos, itos);
 313   __ load_signed_byte32(r0, at_bcp(1));
 314 }
 315 
 316 void TemplateTable::sipush()
 317 {
 318   transition(vtos, itos);
 319   __ load_unsigned_short(r0, at_bcp(1));
 320   __ revw(r0, r0);
 321   __ asrw(r0, r0, 16);
 322 }
 323 
 324 void TemplateTable::ldc(LdcType type)
 325 {
 326   transition(vtos, vtos);
 327   Label call_ldc, notFloat, notClass, notInt, Done;
 328 
 329   if (is_ldc_wide(type)) {
 330     __ get_unsigned_2_byte_index_at_bcp(r1, 1);
 331   } else {
 332     __ load_unsigned_byte(r1, at_bcp(1));
 333   }
 334   __ get_cpool_and_tags(r2, r0);
 335 
 336   const int base_offset = ConstantPool::header_size() * wordSize;
 337   const int tags_offset = Array<u1>::base_offset_in_bytes();
 338 
 339   // get type
 340   __ add(r3, r1, tags_offset);
 341   __ lea(r3, Address(r0, r3));
 342   __ ldarb(r3, r3);
 343 
 344   // unresolved class - get the resolved class
 345   __ cmp(r3, (u1)JVM_CONSTANT_UnresolvedClass);
 346   __ br(Assembler::EQ, call_ldc);
 347 
 348   // unresolved class in error state - call into runtime to throw the error
 349   // from the first resolution attempt
 350   __ cmp(r3, (u1)JVM_CONSTANT_UnresolvedClassInError);
 351   __ br(Assembler::EQ, call_ldc);
 352 
 353   // resolved class - need to call vm to get java mirror of the class
 354   __ cmp(r3, (u1)JVM_CONSTANT_Class);
 355   __ br(Assembler::NE, notClass);
 356 
 357   __ bind(call_ldc);
 358   __ mov(c_rarg1, is_ldc_wide(type) ? 1 : 0);
 359   call_VM(r0, CAST_FROM_FN_PTR(address, InterpreterRuntime::ldc), c_rarg1);
 360   __ push_ptr(r0);
 361   __ verify_oop(r0);
 362   __ b(Done);
 363 
 364   __ bind(notClass);
 365   __ cmp(r3, (u1)JVM_CONSTANT_Float);
 366   __ br(Assembler::NE, notFloat);
 367   // ftos
 368   __ adds(r1, r2, r1, Assembler::LSL, 3);
 369   __ ldrs(v0, Address(r1, base_offset));
 370   __ push_f();
 371   __ b(Done);
 372 
 373   __ bind(notFloat);
 374 
 375   __ cmp(r3, (u1)JVM_CONSTANT_Integer);
 376   __ br(Assembler::NE, notInt);
 377 
 378   // itos
 379   __ adds(r1, r2, r1, Assembler::LSL, 3);
 380   __ ldrw(r0, Address(r1, base_offset));
 381   __ push_i(r0);
 382   __ b(Done);
 383 
 384   __ bind(notInt);
 385   condy_helper(Done);
 386 
 387   __ bind(Done);
 388 }
 389 
 390 // Fast path for caching oop constants.
 391 void TemplateTable::fast_aldc(LdcType type)
 392 {
 393   transition(vtos, atos);
 394 
 395   Register result = r0;
 396   Register tmp = r1;
 397   Register rarg = r2;
 398 
 399   int index_size = is_ldc_wide(type) ? sizeof(u2) : sizeof(u1);
 400 
 401   Label resolved;
 402 
 403   // We are resolved if the resolved reference cache entry contains a
 404   // non-null object (String, MethodType, etc.)
 405   assert_different_registers(result, tmp);
 406   __ get_cache_index_at_bcp(tmp, 1, index_size);
 407   __ load_resolved_reference_at_index(result, tmp);
 408   __ cbnz(result, resolved);
 409 
 410   address entry = CAST_FROM_FN_PTR(address, InterpreterRuntime::resolve_ldc);
 411 
 412   // first time invocation - must resolve first
 413   __ mov(rarg, (int)bytecode());
 414   __ call_VM(result, entry, rarg);
 415 
 416   __ bind(resolved);
 417 
 418   { // Check for the null sentinel.
 419     // If we just called the VM, it already did the mapping for us,
 420     // but it's harmless to retry.
 421     Label notNull;
 422 
 423     // Stash null_sentinel address to get its value later
 424     __ movptr(rarg, (uintptr_t)Universe::the_null_sentinel_addr());
 425     __ ldr(tmp, Address(rarg));
 426     __ resolve_oop_handle(tmp, r5, rscratch2);
 427     __ cmpoop(result, tmp);
 428     __ br(Assembler::NE, notNull);
 429     __ mov(result, 0);  // null object reference
 430     __ bind(notNull);
 431   }
 432 
 433   if (VerifyOops) {
 434     // Safe to call with 0 result
 435     __ verify_oop(result);
 436   }
 437 }
 438 
 439 void TemplateTable::ldc2_w()
 440 {
 441   transition(vtos, vtos);
 442   Label notDouble, notLong, Done;
 443   __ get_unsigned_2_byte_index_at_bcp(r0, 1);
 444 
 445   __ get_cpool_and_tags(r1, r2);
 446   const int base_offset = ConstantPool::header_size() * wordSize;
 447   const int tags_offset = Array<u1>::base_offset_in_bytes();
 448 
 449   // get type
 450   __ lea(r2, Address(r2, r0, Address::lsl(0)));
 451   __ load_unsigned_byte(r2, Address(r2, tags_offset));
 452   __ cmpw(r2, (int)JVM_CONSTANT_Double);
 453   __ br(Assembler::NE, notDouble);
 454 
 455   // dtos
 456   __ lea (r2, Address(r1, r0, Address::lsl(3)));
 457   __ ldrd(v0, Address(r2, base_offset));
 458   __ push_d();
 459   __ b(Done);
 460 
 461   __ bind(notDouble);
 462   __ cmpw(r2, (int)JVM_CONSTANT_Long);
 463   __ br(Assembler::NE, notLong);
 464 
 465   // ltos
 466   __ lea(r0, Address(r1, r0, Address::lsl(3)));
 467   __ ldr(r0, Address(r0, base_offset));
 468   __ push_l();
 469   __ b(Done);
 470 
 471   __ bind(notLong);
 472   condy_helper(Done);
 473 
 474   __ bind(Done);
 475 }
 476 
 477 void TemplateTable::condy_helper(Label& Done)
 478 {
 479   Register obj = r0;
 480   Register rarg = r1;
 481   Register flags = r2;
 482   Register off = r3;
 483 
 484   address entry = CAST_FROM_FN_PTR(address, InterpreterRuntime::resolve_ldc);
 485 
 486   __ mov(rarg, (int) bytecode());
 487   __ call_VM(obj, entry, rarg);
 488 
 489   __ get_vm_result_metadata(flags, rthread);
 490 
 491   // VMr = obj = base address to find primitive value to push
 492   // VMr2 = flags = (tos, off) using format of CPCE::_flags
 493   __ mov(off, flags);
 494   __ andw(off, off, ConstantPoolCache::field_index_mask);
 495 
 496   const Address field(obj, off);
 497 
 498   // What sort of thing are we loading?
 499   // x86 uses a shift and mask or wings it with a shift plus assert
 500   // the mask is not needed. aarch64 just uses bitfield extract
 501   __ ubfxw(flags, flags, ConstantPoolCache::tos_state_shift,
 502            ConstantPoolCache::tos_state_bits);
 503 
 504   switch (bytecode()) {
 505     case Bytecodes::_ldc:
 506     case Bytecodes::_ldc_w:
 507       {
 508         // tos in (itos, ftos, stos, btos, ctos, ztos)
 509         Label notInt, notFloat, notShort, notByte, notChar, notBool;
 510         __ cmpw(flags, itos);
 511         __ br(Assembler::NE, notInt);
 512         // itos
 513         __ ldrw(r0, field);
 514         __ push(itos);
 515         __ b(Done);
 516 
 517         __ bind(notInt);
 518         __ cmpw(flags, ftos);
 519         __ br(Assembler::NE, notFloat);
 520         // ftos
 521         __ load_float(field);
 522         __ push(ftos);
 523         __ b(Done);
 524 
 525         __ bind(notFloat);
 526         __ cmpw(flags, stos);
 527         __ br(Assembler::NE, notShort);
 528         // stos
 529         __ load_signed_short(r0, field);
 530         __ push(stos);
 531         __ b(Done);
 532 
 533         __ bind(notShort);
 534         __ cmpw(flags, btos);
 535         __ br(Assembler::NE, notByte);
 536         // btos
 537         __ load_signed_byte(r0, field);
 538         __ push(btos);
 539         __ b(Done);
 540 
 541         __ bind(notByte);
 542         __ cmpw(flags, ctos);
 543         __ br(Assembler::NE, notChar);
 544         // ctos
 545         __ load_unsigned_short(r0, field);
 546         __ push(ctos);
 547         __ b(Done);
 548 
 549         __ bind(notChar);
 550         __ cmpw(flags, ztos);
 551         __ br(Assembler::NE, notBool);
 552         // ztos
 553         __ load_signed_byte(r0, field);
 554         __ push(ztos);
 555         __ b(Done);
 556 
 557         __ bind(notBool);
 558         break;
 559       }
 560 
 561     case Bytecodes::_ldc2_w:
 562       {
 563         Label notLong, notDouble;
 564         __ cmpw(flags, ltos);
 565         __ br(Assembler::NE, notLong);
 566         // ltos
 567         __ ldr(r0, field);
 568         __ push(ltos);
 569         __ b(Done);
 570 
 571         __ bind(notLong);
 572         __ cmpw(flags, dtos);
 573         __ br(Assembler::NE, notDouble);
 574         // dtos
 575         __ load_double(field);
 576         __ push(dtos);
 577         __ b(Done);
 578 
 579        __ bind(notDouble);
 580         break;
 581       }
 582 
 583     default:
 584       ShouldNotReachHere();
 585     }
 586 
 587     __ stop("bad ldc/condy");
 588 }
 589 
 590 void TemplateTable::locals_index(Register reg, int offset)
 591 {
 592   __ ldrb(reg, at_bcp(offset));
 593   __ neg(reg, reg);
 594 }
 595 
 596 void TemplateTable::iload() {
 597   iload_internal();
 598 }
 599 
 600 void TemplateTable::nofast_iload() {
 601   iload_internal(may_not_rewrite);
 602 }
 603 
 604 void TemplateTable::iload_internal(RewriteControl rc) {
 605   transition(vtos, itos);
 606   if (RewriteFrequentPairs && rc == may_rewrite) {
 607     Label rewrite, done;
 608     Register bc = r4;
 609 
 610     // get next bytecode
 611     __ load_unsigned_byte(r1, at_bcp(Bytecodes::length_for(Bytecodes::_iload)));
 612 
 613     // if _iload, wait to rewrite to iload2.  We only want to rewrite the
 614     // last two iloads in a pair.  Comparing against fast_iload means that
 615     // the next bytecode is neither an iload or a caload, and therefore
 616     // an iload pair.
 617     __ cmpw(r1, Bytecodes::_iload);
 618     __ br(Assembler::EQ, done);
 619 
 620     // if _fast_iload rewrite to _fast_iload2
 621     __ cmpw(r1, Bytecodes::_fast_iload);
 622     __ movw(bc, Bytecodes::_fast_iload2);
 623     __ br(Assembler::EQ, rewrite);
 624 
 625     // if _caload rewrite to _fast_icaload
 626     __ cmpw(r1, Bytecodes::_caload);
 627     __ movw(bc, Bytecodes::_fast_icaload);
 628     __ br(Assembler::EQ, rewrite);
 629 
 630     // else rewrite to _fast_iload
 631     __ movw(bc, Bytecodes::_fast_iload);
 632 
 633     // rewrite
 634     // bc: new bytecode
 635     __ bind(rewrite);
 636     patch_bytecode(Bytecodes::_iload, bc, r1, false);
 637     __ bind(done);
 638 
 639   }
 640 
 641   // do iload, get the local value into tos
 642   locals_index(r1);
 643   __ ldr(r0, iaddress(r1));
 644 
 645 }
 646 
 647 void TemplateTable::fast_iload2()
 648 {
 649   transition(vtos, itos);
 650   locals_index(r1);
 651   __ ldr(r0, iaddress(r1));
 652   __ push(itos);
 653   locals_index(r1, 3);
 654   __ ldr(r0, iaddress(r1));
 655 }
 656 
 657 void TemplateTable::fast_iload()
 658 {
 659   transition(vtos, itos);
 660   locals_index(r1);
 661   __ ldr(r0, iaddress(r1));
 662 }
 663 
 664 void TemplateTable::lload()
 665 {
 666   transition(vtos, ltos);
 667   __ ldrb(r1, at_bcp(1));
 668   __ sub(r1, rlocals, r1, ext::uxtw, LogBytesPerWord);
 669   __ ldr(r0, Address(r1, Interpreter::local_offset_in_bytes(1)));
 670 }
 671 
 672 void TemplateTable::fload()
 673 {
 674   transition(vtos, ftos);
 675   locals_index(r1);
 676   // n.b. we use ldrd here because this is a 64 bit slot
 677   // this is comparable to the iload case
 678   __ ldrd(v0, faddress(r1));
 679 }
 680 
 681 void TemplateTable::dload()
 682 {
 683   transition(vtos, dtos);
 684   __ ldrb(r1, at_bcp(1));
 685   __ sub(r1, rlocals, r1, ext::uxtw, LogBytesPerWord);
 686   __ ldrd(v0, Address(r1, Interpreter::local_offset_in_bytes(1)));
 687 }
 688 
 689 void TemplateTable::aload()
 690 {
 691   transition(vtos, atos);
 692   locals_index(r1);
 693   __ ldr(r0, iaddress(r1));
 694 }
 695 
 696 void TemplateTable::locals_index_wide(Register reg) {
 697   __ ldrh(reg, at_bcp(2));
 698   __ rev16w(reg, reg);
 699   __ neg(reg, reg);
 700 }
 701 
 702 void TemplateTable::wide_iload() {
 703   transition(vtos, itos);
 704   locals_index_wide(r1);
 705   __ ldr(r0, iaddress(r1));
 706 }
 707 
 708 void TemplateTable::wide_lload()
 709 {
 710   transition(vtos, ltos);
 711   __ ldrh(r1, at_bcp(2));
 712   __ rev16w(r1, r1);
 713   __ sub(r1, rlocals, r1, ext::uxtw, LogBytesPerWord);
 714   __ ldr(r0, Address(r1, Interpreter::local_offset_in_bytes(1)));
 715 }
 716 
 717 void TemplateTable::wide_fload()
 718 {
 719   transition(vtos, ftos);
 720   locals_index_wide(r1);
 721   // n.b. we use ldrd here because this is a 64 bit slot
 722   // this is comparable to the iload case
 723   __ ldrd(v0, faddress(r1));
 724 }
 725 
 726 void TemplateTable::wide_dload()
 727 {
 728   transition(vtos, dtos);
 729   __ ldrh(r1, at_bcp(2));
 730   __ rev16w(r1, r1);
 731   __ sub(r1, rlocals, r1, ext::uxtw, LogBytesPerWord);
 732   __ ldrd(v0, Address(r1, Interpreter::local_offset_in_bytes(1)));
 733 }
 734 
 735 void TemplateTable::wide_aload()
 736 {
 737   transition(vtos, atos);
 738   locals_index_wide(r1);
 739   __ ldr(r0, aaddress(r1));
 740 }
 741 
 742 void TemplateTable::index_check(Register array, Register index)
 743 {
 744   // destroys r1, rscratch1
 745   // sign extend index for use by indexed load
 746   // __ movl2ptr(index, index);
 747   // check index
 748   Register length = rscratch1;
 749   __ ldrw(length, Address(array, arrayOopDesc::length_offset_in_bytes()));
 750   __ cmpw(index, length);
 751   if (index != r1) {
 752     // ??? convention: move aberrant index into r1 for exception message
 753     assert(r1 != array, "different registers");
 754     __ mov(r1, index);
 755   }
 756   Label ok;
 757   __ br(Assembler::LO, ok);
 758   // ??? convention: move array into r3 for exception message
 759    __ mov(r3, array);
 760    __ mov(rscratch1, Interpreter::_throw_ArrayIndexOutOfBoundsException_entry);
 761    __ br(rscratch1);
 762   __ bind(ok);
 763 }
 764 
 765 void TemplateTable::iaload()
 766 {
 767   transition(itos, itos);
 768   __ mov(r1, r0);
 769   __ pop_ptr(r0);
 770   // r0: array
 771   // r1: index
 772   index_check(r0, r1); // leaves index in r1, kills rscratch1
 773   __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_INT) >> 2);
 774   __ access_load_at(T_INT, IN_HEAP | IS_ARRAY, r0, Address(r0, r1, Address::uxtw(2)), noreg, noreg);
 775 }
 776 
 777 void TemplateTable::laload()
 778 {
 779   transition(itos, ltos);
 780   __ mov(r1, r0);
 781   __ pop_ptr(r0);
 782   // r0: array
 783   // r1: index
 784   index_check(r0, r1); // leaves index in r1, kills rscratch1
 785   __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_LONG) >> 3);
 786   __ access_load_at(T_LONG, IN_HEAP | IS_ARRAY, r0, Address(r0, r1, Address::uxtw(3)), noreg, noreg);
 787 }
 788 
 789 void TemplateTable::faload()
 790 {
 791   transition(itos, ftos);
 792   __ mov(r1, r0);
 793   __ pop_ptr(r0);
 794   // r0: array
 795   // r1: index
 796   index_check(r0, r1); // leaves index in r1, kills rscratch1
 797   __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_FLOAT) >> 2);
 798   __ access_load_at(T_FLOAT, IN_HEAP | IS_ARRAY, r0, Address(r0, r1, Address::uxtw(2)), noreg, noreg);
 799 }
 800 
 801 void TemplateTable::daload()
 802 {
 803   transition(itos, dtos);
 804   __ mov(r1, r0);
 805   __ pop_ptr(r0);
 806   // r0: array
 807   // r1: index
 808   index_check(r0, r1); // leaves index in r1, kills rscratch1
 809   __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_DOUBLE) >> 3);
 810   __ access_load_at(T_DOUBLE, IN_HEAP | IS_ARRAY, r0, Address(r0, r1, Address::uxtw(3)), noreg, noreg);
 811 }
 812 
 813 void TemplateTable::aaload()
 814 {
 815   transition(itos, atos);
 816   __ mov(r1, r0);
 817   __ pop_ptr(r0);
 818   // r0: array
 819   // r1: index
 820   index_check(r0, r1); // leaves index in r1, kills rscratch1
 821   __ profile_array_type<ArrayLoadData>(r2, r0, r4);
 822   if (UseArrayFlattening) {
 823     Label is_flat_array, done;
 824 
 825     __ test_flat_array_oop(r0, rscratch1 /*temp*/, is_flat_array);
 826     __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_OBJECT) >> LogBytesPerHeapOop);
 827     do_oop_load(_masm, Address(r0, r1, Address::uxtw(LogBytesPerHeapOop)), r0, IS_ARRAY);
 828 
 829     __ b(done);
 830     __ bind(is_flat_array);
 831     __ call_VM(r0, CAST_FROM_FN_PTR(address, InterpreterRuntime::flat_array_load), r0, r1);
 832     // Ensure the stores to copy the inline field contents are visible
 833     // before any subsequent store that publishes this reference.
 834     __ membar(Assembler::StoreStore);
 835     __ bind(done);
 836   } else {
 837     __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_OBJECT) >> LogBytesPerHeapOop);
 838     do_oop_load(_masm, Address(r0, r1, Address::uxtw(LogBytesPerHeapOop)), r0, IS_ARRAY);
 839   }
 840   __ profile_element_type(r2, r0, r4);
 841 }
 842 
 843 void TemplateTable::baload()
 844 {
 845   transition(itos, itos);
 846   __ mov(r1, r0);
 847   __ pop_ptr(r0);
 848   // r0: array
 849   // r1: index
 850   index_check(r0, r1); // leaves index in r1, kills rscratch1
 851   __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_BYTE) >> 0);
 852   __ access_load_at(T_BYTE, IN_HEAP | IS_ARRAY, r0, Address(r0, r1, Address::uxtw(0)), noreg, noreg);
 853 }
 854 
 855 void TemplateTable::caload()
 856 {
 857   transition(itos, itos);
 858   __ mov(r1, r0);
 859   __ pop_ptr(r0);
 860   // r0: array
 861   // r1: index
 862   index_check(r0, r1); // leaves index in r1, kills rscratch1
 863   __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_CHAR) >> 1);
 864   __ access_load_at(T_CHAR, IN_HEAP | IS_ARRAY, r0, Address(r0, r1, Address::uxtw(1)), noreg, noreg);
 865 }
 866 
 867 // iload followed by caload frequent pair
 868 void TemplateTable::fast_icaload()
 869 {
 870   transition(vtos, itos);
 871   // load index out of locals
 872   locals_index(r2);
 873   __ ldr(r1, iaddress(r2));
 874 
 875   __ pop_ptr(r0);
 876 
 877   // r0: array
 878   // r1: index
 879   index_check(r0, r1); // leaves index in r1, kills rscratch1
 880   __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_CHAR) >> 1);
 881   __ access_load_at(T_CHAR, IN_HEAP | IS_ARRAY, r0, Address(r0, r1, Address::uxtw(1)), noreg, noreg);
 882 }
 883 
 884 void TemplateTable::saload()
 885 {
 886   transition(itos, itos);
 887   __ mov(r1, r0);
 888   __ pop_ptr(r0);
 889   // r0: array
 890   // r1: index
 891   index_check(r0, r1); // leaves index in r1, kills rscratch1
 892   __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_SHORT) >> 1);
 893   __ access_load_at(T_SHORT, IN_HEAP | IS_ARRAY, r0, Address(r0, r1, Address::uxtw(1)), noreg, noreg);
 894 }
 895 
 896 void TemplateTable::iload(int n)
 897 {
 898   transition(vtos, itos);
 899   __ ldr(r0, iaddress(n));
 900 }
 901 
 902 void TemplateTable::lload(int n)
 903 {
 904   transition(vtos, ltos);
 905   __ ldr(r0, laddress(n));
 906 }
 907 
 908 void TemplateTable::fload(int n)
 909 {
 910   transition(vtos, ftos);
 911   __ ldrs(v0, faddress(n));
 912 }
 913 
 914 void TemplateTable::dload(int n)
 915 {
 916   transition(vtos, dtos);
 917   __ ldrd(v0, daddress(n));
 918 }
 919 
 920 void TemplateTable::aload(int n)
 921 {
 922   transition(vtos, atos);
 923   __ ldr(r0, iaddress(n));
 924 }
 925 
 926 void TemplateTable::aload_0() {
 927   aload_0_internal();
 928 }
 929 
 930 void TemplateTable::nofast_aload_0() {
 931   aload_0_internal(may_not_rewrite);
 932 }
 933 
 934 void TemplateTable::aload_0_internal(RewriteControl rc) {
 935   // According to bytecode histograms, the pairs:
 936   //
 937   // _aload_0, _fast_igetfield
 938   // _aload_0, _fast_agetfield
 939   // _aload_0, _fast_fgetfield
 940   //
 941   // occur frequently. If RewriteFrequentPairs is set, the (slow)
 942   // _aload_0 bytecode checks if the next bytecode is either
 943   // _fast_igetfield, _fast_agetfield or _fast_fgetfield and then
 944   // rewrites the current bytecode into a pair bytecode; otherwise it
 945   // rewrites the current bytecode into _fast_aload_0 that doesn't do
 946   // the pair check anymore.
 947   //
 948   // Note: If the next bytecode is _getfield, the rewrite must be
 949   //       delayed, otherwise we may miss an opportunity for a pair.
 950   //
 951   // Also rewrite frequent pairs
 952   //   aload_0, aload_1
 953   //   aload_0, iload_1
 954   // These bytecodes with a small amount of code are most profitable
 955   // to rewrite
 956   if (RewriteFrequentPairs && rc == may_rewrite) {
 957     Label rewrite, done;
 958     const Register bc = r4;
 959 
 960     // get next bytecode
 961     __ load_unsigned_byte(r1, at_bcp(Bytecodes::length_for(Bytecodes::_aload_0)));
 962 
 963     // if _getfield then wait with rewrite
 964     __ cmpw(r1, Bytecodes::Bytecodes::_getfield);
 965     __ br(Assembler::EQ, done);
 966 
 967     // if _igetfield then rewrite to _fast_iaccess_0
 968     assert(Bytecodes::java_code(Bytecodes::_fast_iaccess_0) == Bytecodes::_aload_0, "fix bytecode definition");
 969     __ cmpw(r1, Bytecodes::_fast_igetfield);
 970     __ movw(bc, Bytecodes::_fast_iaccess_0);
 971     __ br(Assembler::EQ, rewrite);
 972 
 973     // if _agetfield then rewrite to _fast_aaccess_0
 974     assert(Bytecodes::java_code(Bytecodes::_fast_aaccess_0) == Bytecodes::_aload_0, "fix bytecode definition");
 975     __ cmpw(r1, Bytecodes::_fast_agetfield);
 976     __ movw(bc, Bytecodes::_fast_aaccess_0);
 977     __ br(Assembler::EQ, rewrite);
 978 
 979     // if _fgetfield then rewrite to _fast_faccess_0
 980     assert(Bytecodes::java_code(Bytecodes::_fast_faccess_0) == Bytecodes::_aload_0, "fix bytecode definition");
 981     __ cmpw(r1, Bytecodes::_fast_fgetfield);
 982     __ movw(bc, Bytecodes::_fast_faccess_0);
 983     __ br(Assembler::EQ, rewrite);
 984 
 985     // else rewrite to _fast_aload0
 986     assert(Bytecodes::java_code(Bytecodes::_fast_aload_0) == Bytecodes::_aload_0, "fix bytecode definition");
 987     __ movw(bc, Bytecodes::Bytecodes::_fast_aload_0);
 988 
 989     // rewrite
 990     // bc: new bytecode
 991     __ bind(rewrite);
 992     patch_bytecode(Bytecodes::_aload_0, bc, r1, false);
 993 
 994     __ bind(done);
 995   }
 996 
 997   // Do actual aload_0 (must do this after patch_bytecode which might call VM and GC might change oop).
 998   aload(0);
 999 }
1000 
1001 void TemplateTable::istore()
1002 {
1003   transition(itos, vtos);
1004   locals_index(r1);
1005   // FIXME: We're being very pernickerty here storing a jint in a
1006   // local with strw, which costs an extra instruction over what we'd
1007   // be able to do with a simple str.  We should just store the whole
1008   // word.
1009   __ lea(rscratch1, iaddress(r1));
1010   __ strw(r0, Address(rscratch1));
1011 }
1012 
1013 void TemplateTable::lstore()
1014 {
1015   transition(ltos, vtos);
1016   locals_index(r1);
1017   __ str(r0, laddress(r1, rscratch1, _masm));
1018 }
1019 
1020 void TemplateTable::fstore() {
1021   transition(ftos, vtos);
1022   locals_index(r1);
1023   __ lea(rscratch1, iaddress(r1));
1024   __ strs(v0, Address(rscratch1));
1025 }
1026 
1027 void TemplateTable::dstore() {
1028   transition(dtos, vtos);
1029   locals_index(r1);
1030   __ strd(v0, daddress(r1, rscratch1, _masm));
1031 }
1032 
1033 void TemplateTable::astore()
1034 {
1035   transition(vtos, vtos);
1036   __ pop_ptr(r0);
1037   locals_index(r1);
1038   __ str(r0, aaddress(r1));
1039 }
1040 
1041 void TemplateTable::wide_istore() {
1042   transition(vtos, vtos);
1043   __ pop_i();
1044   locals_index_wide(r1);
1045   __ lea(rscratch1, iaddress(r1));
1046   __ strw(r0, Address(rscratch1));
1047 }
1048 
1049 void TemplateTable::wide_lstore() {
1050   transition(vtos, vtos);
1051   __ pop_l();
1052   locals_index_wide(r1);
1053   __ str(r0, laddress(r1, rscratch1, _masm));
1054 }
1055 
1056 void TemplateTable::wide_fstore() {
1057   transition(vtos, vtos);
1058   __ pop_f();
1059   locals_index_wide(r1);
1060   __ lea(rscratch1, faddress(r1));
1061   __ strs(v0, rscratch1);
1062 }
1063 
1064 void TemplateTable::wide_dstore() {
1065   transition(vtos, vtos);
1066   __ pop_d();
1067   locals_index_wide(r1);
1068   __ strd(v0, daddress(r1, rscratch1, _masm));
1069 }
1070 
1071 void TemplateTable::wide_astore() {
1072   transition(vtos, vtos);
1073   __ pop_ptr(r0);
1074   locals_index_wide(r1);
1075   __ str(r0, aaddress(r1));
1076 }
1077 
1078 void TemplateTable::iastore() {
1079   transition(itos, vtos);
1080   __ pop_i(r1);
1081   __ pop_ptr(r3);
1082   // r0: value
1083   // r1: index
1084   // r3: array
1085   index_check(r3, r1); // prefer index in r1
1086   __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_INT) >> 2);
1087   __ access_store_at(T_INT, IN_HEAP | IS_ARRAY, Address(r3, r1, Address::uxtw(2)), r0, noreg, noreg, noreg);
1088 }
1089 
1090 void TemplateTable::lastore() {
1091   transition(ltos, vtos);
1092   __ pop_i(r1);
1093   __ pop_ptr(r3);
1094   // r0: value
1095   // r1: index
1096   // r3: array
1097   index_check(r3, r1); // prefer index in r1
1098   __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_LONG) >> 3);
1099   __ access_store_at(T_LONG, IN_HEAP | IS_ARRAY, Address(r3, r1, Address::uxtw(3)), r0, noreg, noreg, noreg);
1100 }
1101 
1102 void TemplateTable::fastore() {
1103   transition(ftos, vtos);
1104   __ pop_i(r1);
1105   __ pop_ptr(r3);
1106   // v0: value
1107   // r1:  index
1108   // r3:  array
1109   index_check(r3, r1); // prefer index in r1
1110   __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_FLOAT) >> 2);
1111   __ access_store_at(T_FLOAT, IN_HEAP | IS_ARRAY, Address(r3, r1, Address::uxtw(2)), noreg /* ftos */, noreg, noreg, noreg);
1112 }
1113 
1114 void TemplateTable::dastore() {
1115   transition(dtos, vtos);
1116   __ pop_i(r1);
1117   __ pop_ptr(r3);
1118   // v0: value
1119   // r1:  index
1120   // r3:  array
1121   index_check(r3, r1); // prefer index in r1
1122   __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_DOUBLE) >> 3);
1123   __ access_store_at(T_DOUBLE, IN_HEAP | IS_ARRAY, Address(r3, r1, Address::uxtw(3)), noreg /* dtos */, noreg, noreg, noreg);
1124 }
1125 
1126 void TemplateTable::aastore() {
1127   Label is_null, is_flat_array, ok_is_subtype, done;
1128   transition(vtos, vtos);
1129   // stack: ..., array, index, value
1130   __ ldr(r0, at_tos());    // value
1131   __ ldr(r2, at_tos_p1()); // index
1132   __ ldr(r3, at_tos_p2()); // array
1133 


1134   index_check(r3, r2);     // kills r1
1135 
1136   __ profile_array_type<ArrayStoreData>(r4, r3, r5);
1137   __ profile_multiple_element_types(r4, r0, r5, r6);
1138 
1139   __ add(r4, r2, arrayOopDesc::base_offset_in_bytes(T_OBJECT) >> LogBytesPerHeapOop);
1140   Address element_address(r3, r4, Address::uxtw(LogBytesPerHeapOop));
1141   // Be careful not to clobber r4 below
1142 
1143   // do array store check - check for null value first
1144   __ cbz(r0, is_null);
1145 
1146   // Move array class to r5
1147   __ load_klass(r5, r3);
1148 
1149   if (UseArrayFlattening) {
1150     __ ldrw(r6, Address(r5, Klass::layout_helper_offset()));
1151     __ test_flat_array_layout(r6, is_flat_array);
1152   }
1153 
1154   // Move subklass into r1
1155   __ load_klass(r1, r0);
1156 
1157   // Move array element superklass into r0
1158   __ ldr(r0, Address(r5, ObjArrayKlass::element_klass_offset()));

1159   // Compress array + index*oopSize + 12 into a single register.  Frees r2.
1160 
1161   // Generate subtype check.  Blows r2, r5
1162   // Superklass in r0.  Subklass in r1.
1163 
1164   // is "r1 <: r0" ? (value subclass <: array element superclass)
1165   __ gen_subtype_check(r1, ok_is_subtype, false);
1166 
1167   // Come here on failure
1168   // object is at TOS
1169   __ b(Interpreter::_throw_ArrayStoreException_entry);
1170 
1171   // Come here on success
1172   __ bind(ok_is_subtype);
1173 
1174   // Get the value we will store
1175   __ ldr(r0, at_tos());
1176   // Now store using the appropriate barrier
1177   // Clobbers: r10, r11, r3
1178   do_oop_store(_masm, element_address, r0, IS_ARRAY);
1179   __ b(done);
1180 
1181   // Have a null in r0, r3=array, r2=index.  Store null at ary[idx]
1182   __ bind(is_null);
1183   if (Arguments::is_valhalla_enabled()) {
1184     Label is_null_into_value_array_npe, store_null;
1185 
1186     if (UseArrayFlattening) {
1187       __ test_flat_array_oop(r3, rscratch1, is_flat_array);
1188     }
1189 
1190     // No way to store null in a null-free array
1191     __ test_null_free_array_oop(r3, rscratch1, is_null_into_value_array_npe);
1192     __ b(store_null);
1193 
1194     __ bind(is_null_into_value_array_npe);
1195     __ b(ExternalAddress(Interpreter::_throw_NullPointerException_entry));
1196 
1197     __ bind(store_null);
1198   }
1199 
1200   // Store a null
1201   // Clobbers: r10, r11, r3
1202   do_oop_store(_masm, element_address, noreg, IS_ARRAY);
1203   __ b(done);
1204 
1205   if (UseArrayFlattening) {
1206      Label is_type_ok;
1207     __ bind(is_flat_array); // Store non-null value to flat
1208 
1209     __ ldr(r0, at_tos());    // value
1210     __ ldr(r3, at_tos_p1()); // index
1211     __ ldr(r2, at_tos_p2()); // array
1212     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::flat_array_store), r0, r2, r3);
1213   }
1214 
1215   // Pop stack arguments
1216   __ bind(done);
1217   __ add(esp, esp, 3 * Interpreter::stackElementSize);
1218 }
1219 
1220 void TemplateTable::bastore()
1221 {
1222   transition(itos, vtos);
1223   __ pop_i(r1);
1224   __ pop_ptr(r3);
1225   // r0: value
1226   // r1: index
1227   // r3: array
1228   index_check(r3, r1); // prefer index in r1
1229 
1230   // Need to check whether array is boolean or byte
1231   // since both types share the bastore bytecode.
1232   __ load_klass(r2, r3);
1233   __ ldrw(r2, Address(r2, Klass::layout_helper_offset()));
1234   int diffbit_index = exact_log2(Klass::layout_helper_boolean_diffbit());
1235   Label L_skip;
1236   __ tbz(r2, diffbit_index, L_skip);
1237   __ andw(r0, r0, 1);  // if it is a T_BOOLEAN array, mask the stored value to 0/1
1238   __ bind(L_skip);
1239 
1240   __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_BYTE) >> 0);
1241   __ access_store_at(T_BYTE, IN_HEAP | IS_ARRAY, Address(r3, r1, Address::uxtw(0)), r0, noreg, noreg, noreg);
1242 }
1243 
1244 void TemplateTable::castore()
1245 {
1246   transition(itos, vtos);
1247   __ pop_i(r1);
1248   __ pop_ptr(r3);
1249   // r0: value
1250   // r1: index
1251   // r3: array
1252   index_check(r3, r1); // prefer index in r1
1253   __ add(r1, r1, arrayOopDesc::base_offset_in_bytes(T_CHAR) >> 1);
1254   __ access_store_at(T_CHAR, IN_HEAP | IS_ARRAY, Address(r3, r1, Address::uxtw(1)), r0, noreg, noreg, noreg);
1255 }
1256 
1257 void TemplateTable::sastore()
1258 {
1259   castore();
1260 }
1261 
1262 void TemplateTable::istore(int n)
1263 {
1264   transition(itos, vtos);
1265   __ str(r0, iaddress(n));
1266 }
1267 
1268 void TemplateTable::lstore(int n)
1269 {
1270   transition(ltos, vtos);
1271   __ str(r0, laddress(n));
1272 }
1273 
1274 void TemplateTable::fstore(int n)
1275 {
1276   transition(ftos, vtos);
1277   __ strs(v0, faddress(n));
1278 }
1279 
1280 void TemplateTable::dstore(int n)
1281 {
1282   transition(dtos, vtos);
1283   __ strd(v0, daddress(n));
1284 }
1285 
1286 void TemplateTable::astore(int n)
1287 {
1288   transition(vtos, vtos);
1289   __ pop_ptr(r0);
1290   __ str(r0, iaddress(n));
1291 }
1292 
1293 void TemplateTable::pop()
1294 {
1295   transition(vtos, vtos);
1296   __ add(esp, esp, Interpreter::stackElementSize);
1297 }
1298 
1299 void TemplateTable::pop2()
1300 {
1301   transition(vtos, vtos);
1302   __ add(esp, esp, 2 * Interpreter::stackElementSize);
1303 }
1304 
1305 void TemplateTable::dup()
1306 {
1307   transition(vtos, vtos);
1308   __ ldr(r0, Address(esp, 0));
1309   __ push(r0);
1310   // stack: ..., a, a
1311 }
1312 
1313 void TemplateTable::dup_x1()
1314 {
1315   transition(vtos, vtos);
1316   // stack: ..., a, b
1317   __ ldr(r0, at_tos());  // load b
1318   __ ldr(r2, at_tos_p1());  // load a
1319   __ str(r0, at_tos_p1());  // store b
1320   __ str(r2, at_tos());  // store a
1321   __ push(r0);                  // push b
1322   // stack: ..., b, a, b
1323 }
1324 
1325 void TemplateTable::dup_x2()
1326 {
1327   transition(vtos, vtos);
1328   // stack: ..., a, b, c
1329   __ ldr(r0, at_tos());  // load c
1330   __ ldr(r2, at_tos_p2());  // load a
1331   __ str(r0, at_tos_p2());  // store c in a
1332   __ push(r0);      // push c
1333   // stack: ..., c, b, c, c
1334   __ ldr(r0, at_tos_p2());  // load b
1335   __ str(r2, at_tos_p2());  // store a in b
1336   // stack: ..., c, a, c, c
1337   __ str(r0, at_tos_p1());  // store b in c
1338   // stack: ..., c, a, b, c
1339 }
1340 
1341 void TemplateTable::dup2()
1342 {
1343   transition(vtos, vtos);
1344   // stack: ..., a, b
1345   __ ldr(r0, at_tos_p1());  // load a
1346   __ push(r0);                  // push a
1347   __ ldr(r0, at_tos_p1());  // load b
1348   __ push(r0);                  // push b
1349   // stack: ..., a, b, a, b
1350 }
1351 
1352 void TemplateTable::dup2_x1()
1353 {
1354   transition(vtos, vtos);
1355   // stack: ..., a, b, c
1356   __ ldr(r2, at_tos());  // load c
1357   __ ldr(r0, at_tos_p1());  // load b
1358   __ push(r0);                  // push b
1359   __ push(r2);                  // push c
1360   // stack: ..., a, b, c, b, c
1361   __ str(r2, at_tos_p3());  // store c in b
1362   // stack: ..., a, c, c, b, c
1363   __ ldr(r2, at_tos_p4());  // load a
1364   __ str(r2, at_tos_p2());  // store a in 2nd c
1365   // stack: ..., a, c, a, b, c
1366   __ str(r0, at_tos_p4());  // store b in a
1367   // stack: ..., b, c, a, b, c
1368 }
1369 
1370 void TemplateTable::dup2_x2()
1371 {
1372   transition(vtos, vtos);
1373   // stack: ..., a, b, c, d
1374   __ ldr(r2, at_tos());  // load d
1375   __ ldr(r0, at_tos_p1());  // load c
1376   __ push(r0)            ;      // push c
1377   __ push(r2);                  // push d
1378   // stack: ..., a, b, c, d, c, d
1379   __ ldr(r0, at_tos_p4());  // load b
1380   __ str(r0, at_tos_p2());  // store b in d
1381   __ str(r2, at_tos_p4());  // store d in b
1382   // stack: ..., a, d, c, b, c, d
1383   __ ldr(r2, at_tos_p5());  // load a
1384   __ ldr(r0, at_tos_p3());  // load c
1385   __ str(r2, at_tos_p3());  // store a in c
1386   __ str(r0, at_tos_p5());  // store c in a
1387   // stack: ..., c, d, a, b, c, d
1388 }
1389 
1390 void TemplateTable::swap()
1391 {
1392   transition(vtos, vtos);
1393   // stack: ..., a, b
1394   __ ldr(r2, at_tos_p1());  // load a
1395   __ ldr(r0, at_tos());  // load b
1396   __ str(r2, at_tos());  // store a in b
1397   __ str(r0, at_tos_p1());  // store b in a
1398   // stack: ..., b, a
1399 }
1400 
1401 void TemplateTable::iop2(Operation op)
1402 {
1403   transition(itos, itos);
1404   // r0 <== r1 op r0
1405   __ pop_i(r1);
1406   switch (op) {
1407   case add  : __ addw(r0, r1, r0); break;
1408   case sub  : __ subw(r0, r1, r0); break;
1409   case mul  : __ mulw(r0, r1, r0); break;
1410   case _and : __ andw(r0, r1, r0); break;
1411   case _or  : __ orrw(r0, r1, r0); break;
1412   case _xor : __ eorw(r0, r1, r0); break;
1413   case shl  : __ lslvw(r0, r1, r0); break;
1414   case shr  : __ asrvw(r0, r1, r0); break;
1415   case ushr : __ lsrvw(r0, r1, r0);break;
1416   default   : ShouldNotReachHere();
1417   }
1418 }
1419 
1420 void TemplateTable::lop2(Operation op)
1421 {
1422   transition(ltos, ltos);
1423   // r0 <== r1 op r0
1424   __ pop_l(r1);
1425   switch (op) {
1426   case add  : __ add(r0, r1, r0); break;
1427   case sub  : __ sub(r0, r1, r0); break;
1428   case mul  : __ mul(r0, r1, r0); break;
1429   case _and : __ andr(r0, r1, r0); break;
1430   case _or  : __ orr(r0, r1, r0); break;
1431   case _xor : __ eor(r0, r1, r0); break;
1432   default   : ShouldNotReachHere();
1433   }
1434 }
1435 
1436 void TemplateTable::idiv()
1437 {
1438   transition(itos, itos);
1439   // explicitly check for div0
1440   Label no_div0;
1441   __ cbnzw(r0, no_div0);
1442   __ mov(rscratch1, Interpreter::_throw_ArithmeticException_entry);
1443   __ br(rscratch1);
1444   __ bind(no_div0);
1445   __ pop_i(r1);
1446   // r0 <== r1 idiv r0
1447   __ corrected_idivl(r0, r1, r0, /* want_remainder */ false);
1448 }
1449 
1450 void TemplateTable::irem()
1451 {
1452   transition(itos, itos);
1453   // explicitly check for div0
1454   Label no_div0;
1455   __ cbnzw(r0, no_div0);
1456   __ mov(rscratch1, Interpreter::_throw_ArithmeticException_entry);
1457   __ br(rscratch1);
1458   __ bind(no_div0);
1459   __ pop_i(r1);
1460   // r0 <== r1 irem r0
1461   __ corrected_idivl(r0, r1, r0, /* want_remainder */ true);
1462 }
1463 
1464 void TemplateTable::lmul()
1465 {
1466   transition(ltos, ltos);
1467   __ pop_l(r1);
1468   __ mul(r0, r0, r1);
1469 }
1470 
1471 void TemplateTable::ldiv()
1472 {
1473   transition(ltos, ltos);
1474   // explicitly check for div0
1475   Label no_div0;
1476   __ cbnz(r0, no_div0);
1477   __ mov(rscratch1, Interpreter::_throw_ArithmeticException_entry);
1478   __ br(rscratch1);
1479   __ bind(no_div0);
1480   __ pop_l(r1);
1481   // r0 <== r1 ldiv r0
1482   __ corrected_idivq(r0, r1, r0, /* want_remainder */ false);
1483 }
1484 
1485 void TemplateTable::lrem()
1486 {
1487   transition(ltos, ltos);
1488   // explicitly check for div0
1489   Label no_div0;
1490   __ cbnz(r0, no_div0);
1491   __ mov(rscratch1, Interpreter::_throw_ArithmeticException_entry);
1492   __ br(rscratch1);
1493   __ bind(no_div0);
1494   __ pop_l(r1);
1495   // r0 <== r1 lrem r0
1496   __ corrected_idivq(r0, r1, r0, /* want_remainder */ true);
1497 }
1498 
1499 void TemplateTable::lshl()
1500 {
1501   transition(itos, ltos);
1502   // shift count is in r0
1503   __ pop_l(r1);
1504   __ lslv(r0, r1, r0);
1505 }
1506 
1507 void TemplateTable::lshr()
1508 {
1509   transition(itos, ltos);
1510   // shift count is in r0
1511   __ pop_l(r1);
1512   __ asrv(r0, r1, r0);
1513 }
1514 
1515 void TemplateTable::lushr()
1516 {
1517   transition(itos, ltos);
1518   // shift count is in r0
1519   __ pop_l(r1);
1520   __ lsrv(r0, r1, r0);
1521 }
1522 
1523 void TemplateTable::fop2(Operation op)
1524 {
1525   transition(ftos, ftos);
1526   switch (op) {
1527   case add:
1528     // n.b. use ldrd because this is a 64 bit slot
1529     __ pop_f(v1);
1530     __ fadds(v0, v1, v0);
1531     break;
1532   case sub:
1533     __ pop_f(v1);
1534     __ fsubs(v0, v1, v0);
1535     break;
1536   case mul:
1537     __ pop_f(v1);
1538     __ fmuls(v0, v1, v0);
1539     break;
1540   case div:
1541     __ pop_f(v1);
1542     __ fdivs(v0, v1, v0);
1543     break;
1544   case rem:
1545     __ fmovs(v1, v0);
1546     __ pop_f(v0);
1547     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::frem));
1548     break;
1549   default:
1550     ShouldNotReachHere();
1551     break;
1552   }
1553 }
1554 
1555 void TemplateTable::dop2(Operation op)
1556 {
1557   transition(dtos, dtos);
1558   switch (op) {
1559   case add:
1560     // n.b. use ldrd because this is a 64 bit slot
1561     __ pop_d(v1);
1562     __ faddd(v0, v1, v0);
1563     break;
1564   case sub:
1565     __ pop_d(v1);
1566     __ fsubd(v0, v1, v0);
1567     break;
1568   case mul:
1569     __ pop_d(v1);
1570     __ fmuld(v0, v1, v0);
1571     break;
1572   case div:
1573     __ pop_d(v1);
1574     __ fdivd(v0, v1, v0);
1575     break;
1576   case rem:
1577     __ fmovd(v1, v0);
1578     __ pop_d(v0);
1579     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::drem));
1580     break;
1581   default:
1582     ShouldNotReachHere();
1583     break;
1584   }
1585 }
1586 
1587 void TemplateTable::ineg()
1588 {
1589   transition(itos, itos);
1590   __ negw(r0, r0);
1591 
1592 }
1593 
1594 void TemplateTable::lneg()
1595 {
1596   transition(ltos, ltos);
1597   __ neg(r0, r0);
1598 }
1599 
1600 void TemplateTable::fneg()
1601 {
1602   transition(ftos, ftos);
1603   __ fnegs(v0, v0);
1604 }
1605 
1606 void TemplateTable::dneg()
1607 {
1608   transition(dtos, dtos);
1609   __ fnegd(v0, v0);
1610 }
1611 
1612 void TemplateTable::iinc()
1613 {
1614   transition(vtos, vtos);
1615   __ load_signed_byte(r1, at_bcp(2)); // get constant
1616   locals_index(r2);
1617   __ ldr(r0, iaddress(r2));
1618   __ addw(r0, r0, r1);
1619   __ str(r0, iaddress(r2));
1620 }
1621 
1622 void TemplateTable::wide_iinc()
1623 {
1624   transition(vtos, vtos);
1625   // __ mov(r1, zr);
1626   __ ldrw(r1, at_bcp(2)); // get constant and index
1627   __ rev16(r1, r1);
1628   __ ubfx(r2, r1, 0, 16);
1629   __ neg(r2, r2);
1630   __ sbfx(r1, r1, 16, 16);
1631   __ ldr(r0, iaddress(r2));
1632   __ addw(r0, r0, r1);
1633   __ str(r0, iaddress(r2));
1634 }
1635 
1636 void TemplateTable::convert()
1637 {
1638   // Checking
1639 #ifdef ASSERT
1640   {
1641     TosState tos_in  = ilgl;
1642     TosState tos_out = ilgl;
1643     switch (bytecode()) {
1644     case Bytecodes::_i2l: // fall through
1645     case Bytecodes::_i2f: // fall through
1646     case Bytecodes::_i2d: // fall through
1647     case Bytecodes::_i2b: // fall through
1648     case Bytecodes::_i2c: // fall through
1649     case Bytecodes::_i2s: tos_in = itos; break;
1650     case Bytecodes::_l2i: // fall through
1651     case Bytecodes::_l2f: // fall through
1652     case Bytecodes::_l2d: tos_in = ltos; break;
1653     case Bytecodes::_f2i: // fall through
1654     case Bytecodes::_f2l: // fall through
1655     case Bytecodes::_f2d: tos_in = ftos; break;
1656     case Bytecodes::_d2i: // fall through
1657     case Bytecodes::_d2l: // fall through
1658     case Bytecodes::_d2f: tos_in = dtos; break;
1659     default             : ShouldNotReachHere();
1660     }
1661     switch (bytecode()) {
1662     case Bytecodes::_l2i: // fall through
1663     case Bytecodes::_f2i: // fall through
1664     case Bytecodes::_d2i: // fall through
1665     case Bytecodes::_i2b: // fall through
1666     case Bytecodes::_i2c: // fall through
1667     case Bytecodes::_i2s: tos_out = itos; break;
1668     case Bytecodes::_i2l: // fall through
1669     case Bytecodes::_f2l: // fall through
1670     case Bytecodes::_d2l: tos_out = ltos; break;
1671     case Bytecodes::_i2f: // fall through
1672     case Bytecodes::_l2f: // fall through
1673     case Bytecodes::_d2f: tos_out = ftos; break;
1674     case Bytecodes::_i2d: // fall through
1675     case Bytecodes::_l2d: // fall through
1676     case Bytecodes::_f2d: tos_out = dtos; break;
1677     default             : ShouldNotReachHere();
1678     }
1679     transition(tos_in, tos_out);
1680   }
1681 #endif // ASSERT
1682   // static const int64_t is_nan = 0x8000000000000000L;
1683 
1684   // Conversion
1685   switch (bytecode()) {
1686   case Bytecodes::_i2l:
1687     __ sxtw(r0, r0);
1688     break;
1689   case Bytecodes::_i2f:
1690     __ scvtfws(v0, r0);
1691     break;
1692   case Bytecodes::_i2d:
1693     __ scvtfwd(v0, r0);
1694     break;
1695   case Bytecodes::_i2b:
1696     __ sxtbw(r0, r0);
1697     break;
1698   case Bytecodes::_i2c:
1699     __ uxthw(r0, r0);
1700     break;
1701   case Bytecodes::_i2s:
1702     __ sxthw(r0, r0);
1703     break;
1704   case Bytecodes::_l2i:
1705     __ uxtw(r0, r0);
1706     break;
1707   case Bytecodes::_l2f:
1708     __ scvtfs(v0, r0);
1709     break;
1710   case Bytecodes::_l2d:
1711     __ scvtfd(v0, r0);
1712     break;
1713   case Bytecodes::_f2i:
1714   {
1715     Label L_Okay;
1716     __ clear_fpsr();
1717     __ fcvtzsw(r0, v0);
1718     __ get_fpsr(r1);
1719     __ cbzw(r1, L_Okay);
1720     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::f2i));
1721     __ bind(L_Okay);
1722   }
1723     break;
1724   case Bytecodes::_f2l:
1725   {
1726     Label L_Okay;
1727     __ clear_fpsr();
1728     __ fcvtzs(r0, v0);
1729     __ get_fpsr(r1);
1730     __ cbzw(r1, L_Okay);
1731     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::f2l));
1732     __ bind(L_Okay);
1733   }
1734     break;
1735   case Bytecodes::_f2d:
1736     __ fcvts(v0, v0);
1737     break;
1738   case Bytecodes::_d2i:
1739   {
1740     Label L_Okay;
1741     __ clear_fpsr();
1742     __ fcvtzdw(r0, v0);
1743     __ get_fpsr(r1);
1744     __ cbzw(r1, L_Okay);
1745     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::d2i));
1746     __ bind(L_Okay);
1747   }
1748     break;
1749   case Bytecodes::_d2l:
1750   {
1751     Label L_Okay;
1752     __ clear_fpsr();
1753     __ fcvtzd(r0, v0);
1754     __ get_fpsr(r1);
1755     __ cbzw(r1, L_Okay);
1756     __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::d2l));
1757     __ bind(L_Okay);
1758   }
1759     break;
1760   case Bytecodes::_d2f:
1761     __ fcvtd(v0, v0);
1762     break;
1763   default:
1764     ShouldNotReachHere();
1765   }
1766 }
1767 
1768 void TemplateTable::lcmp()
1769 {
1770   transition(ltos, itos);
1771   Label done;
1772   __ pop_l(r1);
1773   __ cmp(r1, r0);
1774   __ mov(r0, (uint64_t)-1L);
1775   __ br(Assembler::LT, done);
1776   // __ mov(r0, 1UL);
1777   // __ csel(r0, r0, zr, Assembler::NE);
1778   // and here is a faster way
1779   __ csinc(r0, zr, zr, Assembler::EQ);
1780   __ bind(done);
1781 }
1782 
1783 void TemplateTable::float_cmp(bool is_float, int unordered_result)
1784 {
1785   Label done;
1786   if (is_float) {
1787     // XXX get rid of pop here, use ... reg, mem32
1788     __ pop_f(v1);
1789     __ fcmps(v1, v0);
1790   } else {
1791     // XXX get rid of pop here, use ... reg, mem64
1792     __ pop_d(v1);
1793     __ fcmpd(v1, v0);
1794   }
1795   if (unordered_result < 0) {
1796     // we want -1 for unordered or less than, 0 for equal and 1 for
1797     // greater than.
1798     __ mov(r0, (uint64_t)-1L);
1799     // for FP LT tests less than or unordered
1800     __ br(Assembler::LT, done);
1801     // install 0 for EQ otherwise 1
1802     __ csinc(r0, zr, zr, Assembler::EQ);
1803   } else {
1804     // we want -1 for less than, 0 for equal and 1 for unordered or
1805     // greater than.
1806     __ mov(r0, 1L);
1807     // for FP HI tests greater than or unordered
1808     __ br(Assembler::HI, done);
1809     // install 0 for EQ otherwise ~0
1810     __ csinv(r0, zr, zr, Assembler::EQ);
1811 
1812   }
1813   __ bind(done);
1814 }
1815 
1816 void TemplateTable::branch(bool is_jsr, bool is_wide)
1817 {
1818   __ profile_taken_branch(r0);
1819   const ByteSize be_offset = MethodCounters::backedge_counter_offset() +
1820                              InvocationCounter::counter_offset();
1821   const ByteSize inv_offset = MethodCounters::invocation_counter_offset() +
1822                               InvocationCounter::counter_offset();
1823 
1824   // load branch displacement
1825   if (!is_wide) {
1826     __ ldrh(r2, at_bcp(1));
1827     __ rev16(r2, r2);
1828     // sign extend the 16 bit value in r2
1829     __ sbfm(r2, r2, 0, 15);
1830   } else {
1831     __ ldrw(r2, at_bcp(1));
1832     __ revw(r2, r2);
1833     // sign extend the 32 bit value in r2
1834     __ sbfm(r2, r2, 0, 31);
1835   }
1836 
1837   // Handle all the JSR stuff here, then exit.
1838   // It's much shorter and cleaner than intermingling with the non-JSR
1839   // normal-branch stuff occurring below.
1840 
1841   if (is_jsr) {
1842     // Pre-load the next target bytecode into rscratch1
1843     __ load_unsigned_byte(rscratch1, Address(rbcp, r2));
1844     // compute return address as bci
1845     __ ldr(rscratch2, Address(rmethod, Method::const_offset()));
1846     __ add(rscratch2, rscratch2,
1847            in_bytes(ConstMethod::codes_offset()) - (is_wide ? 5 : 3));
1848     __ sub(r1, rbcp, rscratch2);
1849     __ push_i(r1);
1850     // Adjust the bcp by the 16-bit displacement in r2
1851     __ add(rbcp, rbcp, r2);
1852     __ dispatch_only(vtos, /*generate_poll*/true);
1853     return;
1854   }
1855 
1856   // Normal (non-jsr) branch handling
1857 
1858   // Adjust the bcp by the displacement in r2
1859   __ add(rbcp, rbcp, r2);
1860 
1861   assert(UseLoopCounter || !UseOnStackReplacement,
1862          "on-stack-replacement requires loop counters");
1863   Label backedge_counter_overflow;
1864   Label dispatch;
1865   if (UseLoopCounter) {
1866     // increment backedge counter for backward branches
1867     // r0: MDO
1868     // r2: target offset
1869     __ cmp(r2, zr);
1870     __ br(Assembler::GT, dispatch); // count only if backward branch
1871 
1872     // ECN: FIXME: This code smells
1873     // check if MethodCounters exists
1874     Label has_counters;
1875     __ ldr(rscratch1, Address(rmethod, Method::method_counters_offset()));
1876     __ cbnz(rscratch1, has_counters);
1877     __ push(r0);
1878     __ push(r2);
1879     __ call_VM(noreg, CAST_FROM_FN_PTR(address,
1880             InterpreterRuntime::build_method_counters), rmethod);
1881     __ pop(r2);
1882     __ pop(r0);
1883     __ ldr(rscratch1, Address(rmethod, Method::method_counters_offset()));
1884     __ cbz(rscratch1, dispatch); // No MethodCounters allocated, OutOfMemory
1885     __ bind(has_counters);
1886 
1887     Label no_mdo;
1888     int increment = InvocationCounter::count_increment;
1889     if (ProfileInterpreter) {
1890       // Are we profiling?
1891       __ ldr(r1, Address(rmethod, in_bytes(Method::method_data_offset())));
1892       __ cbz(r1, no_mdo);
1893       // Increment the MDO backedge counter
1894       const Address mdo_backedge_counter(r1, in_bytes(MethodData::backedge_counter_offset()) +
1895                                          in_bytes(InvocationCounter::counter_offset()));
1896       const Address mask(r1, in_bytes(MethodData::backedge_mask_offset()));
1897       __ increment_mask_and_jump(mdo_backedge_counter, increment, mask,
1898                                  r0, rscratch1, false, Assembler::EQ,
1899                                  UseOnStackReplacement ? &backedge_counter_overflow : &dispatch);
1900       __ b(dispatch);
1901     }
1902     __ bind(no_mdo);
1903     // Increment backedge counter in MethodCounters*
1904     __ ldr(rscratch1, Address(rmethod, Method::method_counters_offset()));
1905     const Address mask(rscratch1, in_bytes(MethodCounters::backedge_mask_offset()));
1906     __ increment_mask_and_jump(Address(rscratch1, be_offset), increment, mask,
1907                                r0, rscratch2, false, Assembler::EQ,
1908                                UseOnStackReplacement ? &backedge_counter_overflow : &dispatch);
1909     __ bind(dispatch);
1910   }
1911 
1912   // Pre-load the next target bytecode into rscratch1
1913   __ load_unsigned_byte(rscratch1, Address(rbcp, 0));
1914 
1915   // continue with the bytecode @ target
1916   // rscratch1: target bytecode
1917   // rbcp: target bcp
1918   __ dispatch_only(vtos, /*generate_poll*/true);
1919 
1920   if (UseLoopCounter && UseOnStackReplacement) {
1921     // invocation counter overflow
1922     __ bind(backedge_counter_overflow);
1923     __ neg(r2, r2);
1924     __ add(r2, r2, rbcp);     // branch bcp
1925     // IcoResult frequency_counter_overflow([JavaThread*], address branch_bcp)
1926     __ call_VM(noreg,
1927                CAST_FROM_FN_PTR(address,
1928                                 InterpreterRuntime::frequency_counter_overflow),
1929                r2);
1930     __ load_unsigned_byte(r1, Address(rbcp, 0));  // restore target bytecode
1931 
1932     // r0: osr nmethod (osr ok) or null (osr not possible)
1933     // w1: target bytecode
1934     // r2: scratch
1935     __ cbz(r0, dispatch);     // test result -- no osr if null
1936     // nmethod may have been invalidated (VM may block upon call_VM return)
1937     __ ldrb(r2, Address(r0, nmethod::state_offset()));
1938     if (nmethod::in_use != 0)
1939       __ sub(r2, r2, nmethod::in_use);
1940     __ cbnz(r2, dispatch);
1941 
1942     // We have the address of an on stack replacement routine in r0
1943     // We need to prepare to execute the OSR method. First we must
1944     // migrate the locals and monitors off of the stack.
1945 
1946     __ mov(r19, r0);                             // save the nmethod
1947 
1948     JFR_ONLY(__ enter_jfr_critical_section();)
1949 
1950     call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::OSR_migration_begin));
1951 
1952     // r0 is OSR buffer, move it to expected parameter location
1953     __ mov(j_rarg0, r0);
1954 
1955     // remove activation
1956     // get sender esp
1957     __ ldr(esp,
1958         Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize));
1959     // remove frame anchor
1960     __ leave();
1961 
1962     JFR_ONLY(__ leave_jfr_critical_section();)
1963 
1964     // Ensure compiled code always sees stack at proper alignment
1965     __ andr(sp, esp, -16);
1966 
1967     // and begin the OSR nmethod
1968     __ ldr(rscratch1, Address(r19, nmethod::osr_entry_point_offset()));
1969     __ br(rscratch1);
1970   }
1971 }
1972 
1973 
1974 void TemplateTable::if_0cmp(Condition cc)
1975 {
1976   transition(itos, vtos);
1977   // assume branch is more often taken than not (loops use backward branches)
1978   Label not_taken;
1979   if (cc == equal)
1980     __ cbnzw(r0, not_taken);
1981   else if (cc == not_equal)
1982     __ cbzw(r0, not_taken);
1983   else {
1984     __ andsw(zr, r0, r0);
1985     __ br(j_not(cc), not_taken);
1986   }
1987 
1988   branch(false, false);
1989   __ bind(not_taken);
1990   __ profile_not_taken_branch(r0);
1991 }
1992 
1993 void TemplateTable::if_icmp(Condition cc)
1994 {
1995   transition(itos, vtos);
1996   // assume branch is more often taken than not (loops use backward branches)
1997   Label not_taken;
1998   __ pop_i(r1);
1999   __ cmpw(r1, r0, Assembler::LSL);
2000   __ br(j_not(cc), not_taken);
2001   branch(false, false);
2002   __ bind(not_taken);
2003   __ profile_not_taken_branch(r0);
2004 }
2005 
2006 void TemplateTable::if_nullcmp(Condition cc)
2007 {
2008   transition(atos, vtos);
2009   // assume branch is more often taken than not (loops use backward branches)
2010   Label not_taken;
2011   if (cc == equal)
2012     __ cbnz(r0, not_taken);
2013   else
2014     __ cbz(r0, not_taken);
2015   branch(false, false);
2016   __ bind(not_taken);
2017   __ profile_not_taken_branch(r0);
2018 }
2019 
2020 void TemplateTable::if_acmp(Condition cc) {

2021   transition(atos, vtos);
2022   // assume branch is more often taken than not (loops use backward branches)
2023   Label taken, not_taken;
2024   __ pop_ptr(r1);
2025 
2026   __ profile_acmp(r2, r1, r0, r4);
2027 
2028   Register is_inline_type_mask = rscratch1;
2029   __ mov(is_inline_type_mask, markWord::inline_type_pattern);
2030 
2031   if (Arguments::is_valhalla_enabled()) {
2032     // The substitutability test is only necessary if r1 and r0 are not the same...
2033     __ cmp(r1, r0);
2034     __ br(Assembler::EQ, (cc == equal) ? taken : not_taken);
2035 
2036     // ... neither are null...
2037     __ cbz(r1, (cc == equal) ? not_taken : taken);
2038     __ cbz(r0, (cc == equal) ? not_taken : taken);
2039 
2040     // ...and both are values...
2041     __ ldr(r2, Address(r1, oopDesc::mark_offset_in_bytes()));
2042     __ andr(r2, r2, is_inline_type_mask);
2043     __ ldr(r4, Address(r0, oopDesc::mark_offset_in_bytes()));
2044     __ andr(r4, r4, is_inline_type_mask);
2045     __ andr(r2, r2, r4);
2046     __ cmp(r2,  is_inline_type_mask);
2047     __ br(Assembler::NE, (cc == equal) ? not_taken : taken);
2048 
2049     // ...with the same value klass
2050     __ load_metadata(r2, r1);
2051     __ load_metadata(r4, r0);
2052     __ cmp(r2, r4);
2053     __ br(Assembler::NE, (cc == equal) ? not_taken : taken);
2054 
2055     // Know both are the same type, let's test for substitutability...
2056     if (cc == equal) {
2057       invoke_is_substitutable(r0, r1, taken, not_taken);
2058     } else {
2059       invoke_is_substitutable(r0, r1, not_taken, taken);
2060     }
2061     __ stop("Not reachable");
2062   }
2063 
2064   __ cmpoop(r1, r0);
2065   __ br(j_not(cc), not_taken);
2066   __ bind(taken);
2067   branch(false, false);
2068   __ bind(not_taken);
2069   __ profile_not_taken_branch(r0, true);
2070 }
2071 
2072 void TemplateTable::invoke_is_substitutable(Register aobj, Register bobj,
2073                                             Label& is_subst, Label& not_subst) {
2074 
2075   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::is_substitutable), aobj, bobj);
2076   // Restored... r0 answer, jmp to outcome...
2077   __ cbz(r0, not_subst);
2078   __ b(is_subst);
2079 }
2080 
2081 
2082 void TemplateTable::ret() {
2083   transition(vtos, vtos);
2084   locals_index(r1);
2085   __ ldr(r1, aaddress(r1)); // get return bci, compute return bcp
2086   __ profile_ret(r1, r2);
2087   __ ldr(rbcp, Address(rmethod, Method::const_offset()));
2088   __ lea(rbcp, Address(rbcp, r1));
2089   __ add(rbcp, rbcp, in_bytes(ConstMethod::codes_offset()));
2090   __ dispatch_next(vtos, 0, /*generate_poll*/true);
2091 }
2092 
2093 void TemplateTable::wide_ret() {
2094   transition(vtos, vtos);
2095   locals_index_wide(r1);
2096   __ ldr(r1, aaddress(r1)); // get return bci, compute return bcp
2097   __ profile_ret(r1, r2);
2098   __ ldr(rbcp, Address(rmethod, Method::const_offset()));
2099   __ lea(rbcp, Address(rbcp, r1));
2100   __ add(rbcp, rbcp, in_bytes(ConstMethod::codes_offset()));
2101   __ dispatch_next(vtos, 0, /*generate_poll*/true);
2102 }
2103 
2104 
2105 void TemplateTable::tableswitch() {
2106   Label default_case, continue_execution;
2107   transition(itos, vtos);
2108   // align rbcp
2109   __ lea(r1, at_bcp(BytesPerInt));
2110   __ andr(r1, r1, -BytesPerInt);
2111   // load lo & hi
2112   __ ldrw(r2, Address(r1, BytesPerInt));
2113   __ ldrw(r3, Address(r1, 2 * BytesPerInt));
2114   __ rev32(r2, r2);
2115   __ rev32(r3, r3);
2116   // check against lo & hi
2117   __ cmpw(r0, r2);
2118   __ br(Assembler::LT, default_case);
2119   __ cmpw(r0, r3);
2120   __ br(Assembler::GT, default_case);
2121   // lookup dispatch offset
2122   __ subw(r0, r0, r2);
2123   __ lea(r3, Address(r1, r0, Address::uxtw(2)));
2124   __ ldrw(r3, Address(r3, 3 * BytesPerInt));
2125   __ profile_switch_case(r0, r1, r2);
2126   // continue execution
2127   __ bind(continue_execution);
2128   __ rev32(r3, r3);
2129   __ load_unsigned_byte(rscratch1, Address(rbcp, r3, Address::sxtw(0)));
2130   __ add(rbcp, rbcp, r3, ext::sxtw);
2131   __ dispatch_only(vtos, /*generate_poll*/true);
2132   // handle default
2133   __ bind(default_case);
2134   __ profile_switch_default(r0);
2135   __ ldrw(r3, Address(r1, 0));
2136   __ b(continue_execution);
2137 }
2138 
2139 void TemplateTable::lookupswitch() {
2140   transition(itos, itos);
2141   __ stop("lookupswitch bytecode should have been rewritten");
2142 }
2143 
2144 void TemplateTable::fast_linearswitch() {
2145   transition(itos, vtos);
2146   Label loop_entry, loop, found, continue_execution;
2147   // bswap r0 so we can avoid bswapping the table entries
2148   __ rev32(r0, r0);
2149   // align rbcp
2150   __ lea(r19, at_bcp(BytesPerInt)); // btw: should be able to get rid of
2151                                     // this instruction (change offsets
2152                                     // below)
2153   __ andr(r19, r19, -BytesPerInt);
2154   // set counter
2155   __ ldrw(r1, Address(r19, BytesPerInt));
2156   __ rev32(r1, r1);
2157   __ b(loop_entry);
2158   // table search
2159   __ bind(loop);
2160   __ lea(rscratch1, Address(r19, r1, Address::lsl(3)));
2161   __ ldrw(rscratch1, Address(rscratch1, 2 * BytesPerInt));
2162   __ cmpw(r0, rscratch1);
2163   __ br(Assembler::EQ, found);
2164   __ bind(loop_entry);
2165   __ subs(r1, r1, 1);
2166   __ br(Assembler::PL, loop);
2167   // default case
2168   __ profile_switch_default(r0);
2169   __ ldrw(r3, Address(r19, 0));
2170   __ b(continue_execution);
2171   // entry found -> get offset
2172   __ bind(found);
2173   __ lea(rscratch1, Address(r19, r1, Address::lsl(3)));
2174   __ ldrw(r3, Address(rscratch1, 3 * BytesPerInt));
2175   __ profile_switch_case(r1, r0, r19);
2176   // continue execution
2177   __ bind(continue_execution);
2178   __ rev32(r3, r3);
2179   __ add(rbcp, rbcp, r3, ext::sxtw);
2180   __ ldrb(rscratch1, Address(rbcp, 0));
2181   __ dispatch_only(vtos, /*generate_poll*/true);
2182 }
2183 
2184 void TemplateTable::fast_binaryswitch() {
2185   transition(itos, vtos);
2186   // Implementation using the following core algorithm:
2187   //
2188   // int binary_search(int key, LookupswitchPair* array, int n) {
2189   //   // Binary search according to "Methodik des Programmierens" by
2190   //   // Edsger W. Dijkstra and W.H.J. Feijen, Addison Wesley Germany 1985.
2191   //   int i = 0;
2192   //   int j = n;
2193   //   while (i+1 < j) {
2194   //     // invariant P: 0 <= i < j <= n and (a[i] <= key < a[j] or Q)
2195   //     // with      Q: for all i: 0 <= i < n: key < a[i]
2196   //     // where a stands for the array and assuming that the (inexisting)
2197   //     // element a[n] is infinitely big.
2198   //     int h = (i + j) >> 1;
2199   //     // i < h < j
2200   //     if (key < array[h].fast_match()) {
2201   //       j = h;
2202   //     } else {
2203   //       i = h;
2204   //     }
2205   //   }
2206   //   // R: a[i] <= key < a[i+1] or Q
2207   //   // (i.e., if key is within array, i is the correct index)
2208   //   return i;
2209   // }
2210 
2211   // Register allocation
2212   const Register key   = r0; // already set (tosca)
2213   const Register array = r1;
2214   const Register i     = r2;
2215   const Register j     = r3;
2216   const Register h     = rscratch1;
2217   const Register temp  = rscratch2;
2218 
2219   // Find array start
2220   __ lea(array, at_bcp(3 * BytesPerInt)); // btw: should be able to
2221                                           // get rid of this
2222                                           // instruction (change
2223                                           // offsets below)
2224   __ andr(array, array, -BytesPerInt);
2225 
2226   // Initialize i & j
2227   __ mov(i, 0);                            // i = 0;
2228   __ ldrw(j, Address(array, -BytesPerInt)); // j = length(array);
2229 
2230   // Convert j into native byteordering
2231   __ rev32(j, j);
2232 
2233   // And start
2234   Label entry;
2235   __ b(entry);
2236 
2237   // binary search loop
2238   {
2239     Label loop;
2240     __ bind(loop);
2241     // int h = (i + j) >> 1;
2242     __ addw(h, i, j);                           // h = i + j;
2243     __ lsrw(h, h, 1);                                   // h = (i + j) >> 1;
2244     // if (key < array[h].fast_match()) {
2245     //   j = h;
2246     // } else {
2247     //   i = h;
2248     // }
2249     // Convert array[h].match to native byte-ordering before compare
2250     __ ldr(temp, Address(array, h, Address::lsl(3)));
2251     __ rev32(temp, temp);
2252     __ cmpw(key, temp);
2253     // j = h if (key <  array[h].fast_match())
2254     __ csel(j, h, j, Assembler::LT);
2255     // i = h if (key >= array[h].fast_match())
2256     __ csel(i, h, i, Assembler::GE);
2257     // while (i+1 < j)
2258     __ bind(entry);
2259     __ addw(h, i, 1);          // i+1
2260     __ cmpw(h, j);             // i+1 < j
2261     __ br(Assembler::LT, loop);
2262   }
2263 
2264   // end of binary search, result index is i (must check again!)
2265   Label default_case;
2266   // Convert array[i].match to native byte-ordering before compare
2267   __ ldr(temp, Address(array, i, Address::lsl(3)));
2268   __ rev32(temp, temp);
2269   __ cmpw(key, temp);
2270   __ br(Assembler::NE, default_case);
2271 
2272   // entry found -> j = offset
2273   __ add(j, array, i, ext::uxtx, 3);
2274   __ ldrw(j, Address(j, BytesPerInt));
2275   __ profile_switch_case(i, key, array);
2276   __ rev32(j, j);
2277   __ load_unsigned_byte(rscratch1, Address(rbcp, j, Address::sxtw(0)));
2278   __ lea(rbcp, Address(rbcp, j, Address::sxtw(0)));
2279   __ dispatch_only(vtos, /*generate_poll*/true);
2280 
2281   // default case -> j = default offset
2282   __ bind(default_case);
2283   __ profile_switch_default(i);
2284   __ ldrw(j, Address(array, -2 * BytesPerInt));
2285   __ rev32(j, j);
2286   __ load_unsigned_byte(rscratch1, Address(rbcp, j, Address::sxtw(0)));
2287   __ lea(rbcp, Address(rbcp, j, Address::sxtw(0)));
2288   __ dispatch_only(vtos, /*generate_poll*/true);
2289 }
2290 
2291 
2292 void TemplateTable::_return(TosState state)
2293 {
2294   transition(state, state);
2295   assert(_desc->calls_vm(),
2296          "inconsistent calls_vm information"); // call in remove_activation
2297 
2298   if (_desc->bytecode() == Bytecodes::_return_register_finalizer) {
2299     assert(state == vtos, "only valid state");
2300 
2301     __ ldr(c_rarg1, aaddress(0));
2302     __ load_klass(r3, c_rarg1);
2303     __ ldrb(r3, Address(r3, Klass::misc_flags_offset()));
2304     Label skip_register_finalizer;
2305     __ tbz(r3, exact_log2(KlassFlags::_misc_has_finalizer), skip_register_finalizer);
2306 
2307     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::register_finalizer), c_rarg1);
2308 
2309     __ bind(skip_register_finalizer);
2310   }
2311 
2312   // Issue a StoreStore barrier after all stores but before return
2313   // from any constructor for any class with a final field.  We don't
2314   // know if this is a finalizer, so we always do so.
2315   if (_desc->bytecode() == Bytecodes::_return
2316       || _desc->bytecode() == Bytecodes::_return_register_finalizer)
2317     __ membar(MacroAssembler::StoreStore);
2318 
2319   if (_desc->bytecode() != Bytecodes::_return_register_finalizer) {
2320     Label no_safepoint;
2321     __ ldr(rscratch1, Address(rthread, JavaThread::polling_word_offset()));
2322     __ tbz(rscratch1, log2i_exact(SafepointMechanism::poll_bit()), no_safepoint);
2323     __ push(state);
2324     __ push_cont_fastpath(rthread);
2325     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::at_safepoint));
2326     __ pop_cont_fastpath(rthread);
2327     __ pop(state);
2328     __ bind(no_safepoint);
2329   }
2330 
2331   // Narrow result if state is itos but result type is smaller.
2332   // Need to narrow in the return bytecode rather than in generate_return_entry
2333   // since compiled code callers expect the result to already be narrowed.
2334   if (state == itos) {
2335     __ narrow(r0);
2336   }
2337 
2338   __ remove_activation(state);
2339   __ ret(lr);
2340 }
2341 
2342 // ----------------------------------------------------------------------------
2343 // Volatile variables demand their effects be made known to all CPU's
2344 // in order.  Store buffers on most chips allow reads & writes to
2345 // reorder; the JMM's ReadAfterWrite.java test fails in -Xint mode
2346 // without some kind of memory barrier (i.e., it's not sufficient that
2347 // the interpreter does not reorder volatile references, the hardware
2348 // also must not reorder them).
2349 //
2350 // According to the new Java Memory Model (JMM):
2351 // (1) All volatiles are serialized wrt to each other.  ALSO reads &
2352 //     writes act as acquire & release, so:
2353 // (2) A read cannot let unrelated NON-volatile memory refs that
2354 //     happen after the read float up to before the read.  It's OK for
2355 //     non-volatile memory refs that happen before the volatile read to
2356 //     float down below it.
2357 // (3) Similar a volatile write cannot let unrelated NON-volatile
2358 //     memory refs that happen BEFORE the write float down to after the
2359 //     write.  It's OK for non-volatile memory refs that happen after the
2360 //     volatile write to float up before it.
2361 //
2362 // We only put in barriers around volatile refs (they are expensive),
2363 // not _between_ memory refs (that would require us to track the
2364 // flavor of the previous memory refs).  Requirements (2) and (3)
2365 // require some barriers before volatile stores and after volatile
2366 // loads.  These nearly cover requirement (1) but miss the
2367 // volatile-store-volatile-load case.  This final case is placed after
2368 // volatile-stores although it could just as well go before
2369 // volatile-loads.
2370 
2371 void TemplateTable::resolve_cache_and_index_for_method(int byte_no,
2372                                             Register Rcache,
2373                                             Register index) {
2374   const Register temp = r19;
2375   assert_different_registers(Rcache, index, temp);
2376   assert(byte_no == f1_byte || byte_no == f2_byte, "byte_no out of range");
2377 
2378   Label L_clinit_barrier_slow, L_done;
2379 
2380   Bytecodes::Code code = bytecode();
2381   __ load_method_entry(Rcache, index);
2382   switch(byte_no) {
2383     case f1_byte:
2384       __ lea(temp, Address(Rcache, in_bytes(ResolvedMethodEntry::bytecode1_offset())));
2385       break;
2386     case f2_byte:
2387       __ lea(temp, Address(Rcache, in_bytes(ResolvedMethodEntry::bytecode2_offset())));
2388       break;
2389   }
2390   // Load-acquire the bytecode to match store-release in InterpreterRuntime
2391   __ ldarb(temp, temp);
2392   __ subs(zr, temp, (int) code);  // have we resolved this bytecode?
2393 
2394   // Class initialization barrier for static methods
2395   if (bytecode() == Bytecodes::_invokestatic) {
2396     assert(VM_Version::supports_fast_class_init_checks(), "sanity");
2397     __ br(Assembler::NE, L_clinit_barrier_slow);
2398     __ ldr(temp, Address(Rcache, in_bytes(ResolvedMethodEntry::method_offset())));
2399     __ load_method_holder(temp, temp);
2400     __ clinit_barrier(temp, rscratch1, &L_done, /*L_slow_path*/ nullptr);
2401     __ bind(L_clinit_barrier_slow);
2402   } else {
2403     __ br(Assembler::EQ, L_done);
2404   }
2405 
2406   // resolve first time through
2407   // Class initialization barrier slow path lands here as well.
2408   address entry = CAST_FROM_FN_PTR(address, InterpreterRuntime::resolve_from_cache);
2409   __ mov(temp, (int) code);
2410   __ call_VM_preemptable(noreg, entry, temp);
2411 
2412   // Update registers with resolved info
2413   __ load_method_entry(Rcache, index);
2414   // n.b. unlike x86 Rcache is now rcpool plus the indexed offset
2415   // so all clients ofthis method must be modified accordingly
2416   __ bind(L_done);
2417 }
2418 
2419 void TemplateTable::resolve_cache_and_index_for_field(int byte_no,
2420                                             Register Rcache,
2421                                             Register index) {
2422   const Register temp = r19;
2423   assert_different_registers(Rcache, index, temp);
2424 
2425   Label L_clinit_barrier_slow, L_done;
2426 
2427   Bytecodes::Code code = bytecode();
2428   switch (code) {
2429   case Bytecodes::_nofast_getfield: code = Bytecodes::_getfield; break;
2430   case Bytecodes::_nofast_putfield: code = Bytecodes::_putfield; break;
2431   default: break;
2432   }
2433 
2434   assert(byte_no == f1_byte || byte_no == f2_byte, "byte_no out of range");
2435   __ load_field_entry(Rcache, index);
2436   if (byte_no == f1_byte) {
2437     __ lea(temp, Address(Rcache, in_bytes(ResolvedFieldEntry::get_code_offset())));
2438   } else {
2439     __ lea(temp, Address(Rcache, in_bytes(ResolvedFieldEntry::put_code_offset())));
2440   }
2441   // Load-acquire the bytecode to match store-release in ResolvedFieldEntry::fill_in()
2442   __ ldarb(temp, temp);
2443   __ subs(zr, temp, (int) code);  // have we resolved this bytecode?
2444 
2445   // Class initialization barrier for static fields
2446   if (bytecode() == Bytecodes::_getstatic || bytecode() == Bytecodes::_putstatic) {
2447     assert(VM_Version::supports_fast_class_init_checks(), "sanity");
2448     const Register field_holder = temp;
2449 
2450     __ br(Assembler::NE, L_clinit_barrier_slow);
2451     __ ldr(field_holder, Address(Rcache, in_bytes(ResolvedFieldEntry::field_holder_offset())));
2452     __ clinit_barrier(field_holder, rscratch1, &L_done, /*L_slow_path*/ nullptr);
2453     __ bind(L_clinit_barrier_slow);
2454   } else {
2455     __ br(Assembler::EQ, L_done);
2456   }
2457 
2458   // resolve first time through
2459   // Class initialization barrier slow path lands here as well.
2460   address entry = CAST_FROM_FN_PTR(address, InterpreterRuntime::resolve_from_cache);
2461   __ mov(temp, (int) code);
2462   __ call_VM_preemptable(noreg, entry, temp);
2463 
2464   // Update registers with resolved info
2465   __ load_field_entry(Rcache, index);
2466   __ bind(L_done);
2467 }
2468 
2469 void TemplateTable::load_resolved_field_entry(Register obj,
2470                                               Register cache,
2471                                               Register tos_state,
2472                                               Register offset,
2473                                               Register flags,
2474                                               bool is_static = false) {
2475   assert_different_registers(cache, tos_state, flags, offset);
2476 
2477   // Field offset
2478   __ load_sized_value(offset, Address(cache, in_bytes(ResolvedFieldEntry::field_offset_offset())), sizeof(int), true /*is_signed*/);
2479 
2480   // Flags
2481   __ load_unsigned_byte(flags, Address(cache, in_bytes(ResolvedFieldEntry::flags_offset())));
2482 
2483   // TOS state
2484   if (tos_state != noreg) {
2485     __ load_unsigned_byte(tos_state, Address(cache, in_bytes(ResolvedFieldEntry::type_offset())));
2486   }
2487 
2488   // Klass overwrite register
2489   if (is_static) {
2490     __ ldr(obj, Address(cache, ResolvedFieldEntry::field_holder_offset()));
2491     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
2492     __ ldr(obj, Address(obj, mirror_offset));
2493     __ resolve_oop_handle(obj, r5, rscratch2);
2494   }
2495 }
2496 
2497 void TemplateTable::load_resolved_method_entry_special_or_static(Register cache,
2498                                                                  Register method,
2499                                                                  Register flags) {
2500 
2501   // setup registers
2502   const Register index = flags;
2503   assert_different_registers(method, cache, flags);
2504 
2505   // determine constant pool cache field offsets
2506   resolve_cache_and_index_for_method(f1_byte, cache, index);
2507   __ load_unsigned_byte(flags, Address(cache, in_bytes(ResolvedMethodEntry::flags_offset())));
2508   __ ldr(method, Address(cache, in_bytes(ResolvedMethodEntry::method_offset())));
2509 }
2510 
2511 void TemplateTable::load_resolved_method_entry_handle(Register cache,
2512                                                       Register method,
2513                                                       Register ref_index,
2514                                                       Register flags) {
2515   // setup registers
2516   const Register index = ref_index;
2517   assert_different_registers(method, flags);
2518   assert_different_registers(method, cache, index);
2519 
2520   // determine constant pool cache field offsets
2521   resolve_cache_and_index_for_method(f1_byte, cache, index);
2522   __ load_unsigned_byte(flags, Address(cache, in_bytes(ResolvedMethodEntry::flags_offset())));
2523 
2524   // maybe push appendix to arguments (just before return address)
2525   Label L_no_push;
2526   __ tbz(flags, ResolvedMethodEntry::has_appendix_shift, L_no_push);
2527   // invokehandle uses an index into the resolved references array
2528   __ load_unsigned_short(ref_index, Address(cache, in_bytes(ResolvedMethodEntry::resolved_references_index_offset())));
2529   // Push the appendix as a trailing parameter.
2530   // This must be done before we get the receiver,
2531   // since the parameter_size includes it.
2532   Register appendix = method;
2533   __ load_resolved_reference_at_index(appendix, ref_index);
2534   __ push(appendix);  // push appendix (MethodType, CallSite, etc.)
2535   __ bind(L_no_push);
2536 
2537   __ ldr(method, Address(cache, in_bytes(ResolvedMethodEntry::method_offset())));
2538 }
2539 
2540 void TemplateTable::load_resolved_method_entry_interface(Register cache,
2541                                                          Register klass,
2542                                                          Register method_or_table_index,
2543                                                          Register flags) {
2544   // setup registers
2545   const Register index = method_or_table_index;
2546   assert_different_registers(method_or_table_index, cache, flags);
2547 
2548   // determine constant pool cache field offsets
2549   resolve_cache_and_index_for_method(f1_byte, cache, index);
2550   __ load_unsigned_byte(flags, Address(cache, in_bytes(ResolvedMethodEntry::flags_offset())));
2551 
2552   // Invokeinterface can behave in different ways:
2553   // If calling a method from java.lang.Object, the forced virtual flag is true so the invocation will
2554   // behave like an invokevirtual call. The state of the virtual final flag will determine whether a method or
2555   // vtable index is placed in the register.
2556   // Otherwise, the registers will be populated with the klass and method.
2557 
2558   Label NotVirtual; Label NotVFinal; Label Done;
2559   __ tbz(flags, ResolvedMethodEntry::is_forced_virtual_shift, NotVirtual);
2560   __ tbz(flags, ResolvedMethodEntry::is_vfinal_shift, NotVFinal);
2561   __ ldr(method_or_table_index, Address(cache, in_bytes(ResolvedMethodEntry::method_offset())));
2562   __ b(Done);
2563 
2564   __ bind(NotVFinal);
2565   __ load_unsigned_short(method_or_table_index, Address(cache, in_bytes(ResolvedMethodEntry::table_index_offset())));
2566   __ b(Done);
2567 
2568   __ bind(NotVirtual);
2569   __ ldr(method_or_table_index, Address(cache, in_bytes(ResolvedMethodEntry::method_offset())));
2570   __ ldr(klass, Address(cache, in_bytes(ResolvedMethodEntry::klass_offset())));
2571   __ bind(Done);
2572 }
2573 
2574 void TemplateTable::load_resolved_method_entry_virtual(Register cache,
2575                                                        Register method_or_table_index,
2576                                                        Register flags) {
2577   // setup registers
2578   const Register index = flags;
2579   assert_different_registers(method_or_table_index, cache, flags);
2580 
2581   // determine constant pool cache field offsets
2582   resolve_cache_and_index_for_method(f2_byte, cache, index);
2583   __ load_unsigned_byte(flags, Address(cache, in_bytes(ResolvedMethodEntry::flags_offset())));
2584 
2585   // method_or_table_index can either be an itable index or a method depending on the virtual final flag
2586   Label NotVFinal; Label Done;
2587   __ tbz(flags, ResolvedMethodEntry::is_vfinal_shift, NotVFinal);
2588   __ ldr(method_or_table_index, Address(cache, in_bytes(ResolvedMethodEntry::method_offset())));
2589   __ b(Done);
2590 
2591   __ bind(NotVFinal);
2592   __ load_unsigned_short(method_or_table_index, Address(cache, in_bytes(ResolvedMethodEntry::table_index_offset())));
2593   __ bind(Done);
2594 }
2595 
2596 // The rmethod register is input and overwritten to be the adapter method for the
2597 // indy call. Link Register (lr) is set to the return address for the adapter and
2598 // an appendix may be pushed to the stack. Registers r0-r3 are clobbered
2599 void TemplateTable::load_invokedynamic_entry(Register method) {
2600   // setup registers
2601   const Register appendix = r0;
2602   const Register cache = r2;
2603   const Register index = r3;
2604   assert_different_registers(method, appendix, cache, index, rcpool);
2605 
2606   __ save_bcp();
2607 
2608   Label resolved;
2609 
2610   __ load_resolved_indy_entry(cache, index);
2611   // Load-acquire the adapter method to match store-release in ResolvedIndyEntry::fill_in()
2612   __ lea(method, Address(cache, in_bytes(ResolvedIndyEntry::method_offset())));
2613   __ ldar(method, method);
2614 
2615   // Compare the method to zero
2616   __ cbnz(method, resolved);
2617 
2618   Bytecodes::Code code = bytecode();
2619 
2620   // Call to the interpreter runtime to resolve invokedynamic
2621   address entry = CAST_FROM_FN_PTR(address, InterpreterRuntime::resolve_from_cache);
2622   __ mov(method, code); // this is essentially Bytecodes::_invokedynamic
2623   __ call_VM(noreg, entry, method);
2624   // Update registers with resolved info
2625   __ load_resolved_indy_entry(cache, index);
2626   // Load-acquire the adapter method to match store-release in ResolvedIndyEntry::fill_in()
2627   __ lea(method, Address(cache, in_bytes(ResolvedIndyEntry::method_offset())));
2628   __ ldar(method, method);
2629 
2630 #ifdef ASSERT
2631   __ cbnz(method, resolved);
2632   __ stop("Should be resolved by now");
2633 #endif // ASSERT
2634   __ bind(resolved);
2635 
2636   Label L_no_push;
2637   // Check if there is an appendix
2638   __ load_unsigned_byte(index, Address(cache, in_bytes(ResolvedIndyEntry::flags_offset())));
2639   __ tbz(index, ResolvedIndyEntry::has_appendix_shift, L_no_push);
2640 
2641   // Get appendix
2642   __ load_unsigned_short(index, Address(cache, in_bytes(ResolvedIndyEntry::resolved_references_index_offset())));
2643   // Push the appendix as a trailing parameter
2644   // since the parameter_size includes it.
2645   __ push(method);
2646   __ mov(method, index);
2647   __ load_resolved_reference_at_index(appendix, method);
2648   __ verify_oop(appendix);
2649   __ pop(method);
2650   __ push(appendix);  // push appendix (MethodType, CallSite, etc.)
2651   __ bind(L_no_push);
2652 
2653   // compute return type
2654   __ load_unsigned_byte(index, Address(cache, in_bytes(ResolvedIndyEntry::result_type_offset())));
2655   // load return address
2656   // Return address is loaded into link register(lr) and not pushed to the stack
2657   // like x86
2658   {
2659     const address table_addr = (address) Interpreter::invoke_return_entry_table_for(code);
2660     __ mov(rscratch1, table_addr);
2661     __ ldr(lr, Address(rscratch1, index, Address::lsl(3)));
2662   }
2663 }
2664 
2665 // The registers cache and index expected to be set before call.
2666 // Correct values of the cache and index registers are preserved.
2667 void TemplateTable::jvmti_post_field_access(Register cache, Register index,
2668                                             bool is_static, bool has_tos) {
2669   // do the JVMTI work here to avoid disturbing the register state below
2670   // We use c_rarg registers here because we want to use the register used in
2671   // the call to the VM
2672   if (JvmtiExport::can_post_field_access()) {
2673     // Check to see if a field access watch has been set before we
2674     // take the time to call into the VM.
2675     Label L1;
2676     assert_different_registers(cache, index, r0);
2677     __ lea(rscratch1, ExternalAddress((address) JvmtiExport::get_field_access_count_addr()));
2678     __ ldrw(r0, Address(rscratch1));
2679     __ cbzw(r0, L1);
2680 
2681     __ load_field_entry(c_rarg2, index);
2682 
2683     if (is_static) {
2684       __ mov(c_rarg1, zr); // null object reference
2685     } else {
2686       __ ldr(c_rarg1, at_tos()); // get object pointer without popping it
2687       __ verify_oop(c_rarg1);
2688     }
2689     // c_rarg1: object pointer or null
2690     // c_rarg2: cache entry pointer
2691     __ call_VM(noreg, CAST_FROM_FN_PTR(address,
2692                                        InterpreterRuntime::post_field_access),
2693                c_rarg1, c_rarg2);
2694     __ load_field_entry(cache, index);
2695     __ bind(L1);
2696   }
2697 }
2698 
2699 void TemplateTable::pop_and_check_object(Register r)
2700 {
2701   __ pop_ptr(r);
2702   __ null_check(r);  // for field access must check obj.
2703   __ verify_oop(r);
2704 }
2705 
2706 void TemplateTable::getfield_or_static(int byte_no, bool is_static, RewriteControl rc)
2707 {
2708   const Register cache     = r2;
2709   const Register obj       = r4;
2710   const Register index     = r3;
2711   const Register tos_state = r3;
2712   const Register off       = r19;
2713   const Register flags     = r6;
2714   const Register bc        = r4; // uses same reg as obj, so don't mix them
2715 
2716   resolve_cache_and_index_for_field(byte_no, cache, index);
2717   jvmti_post_field_access(cache, index, is_static, false);
2718 
2719   load_resolved_field_entry(obj, cache, tos_state, off, flags, is_static);
2720 
2721   if (!is_static) {
2722     // obj is on the stack
2723     pop_and_check_object(obj);
2724   }
2725 
2726   // 8179954: We need to make sure that the code generated for
2727   // volatile accesses forms a sequentially-consistent set of
2728   // operations when combined with STLR and LDAR.  Without a leading
2729   // membar it's possible for a simple Dekker test to fail if loads
2730   // use LDR;DMB but stores use STLR.  This can happen if C2 compiles
2731   // the stores in one method and we interpret the loads in another.
2732   if (!CompilerConfig::is_c1_or_interpreter_only_no_jvmci()){
2733     Label notVolatile;
2734     __ tbz(flags, ResolvedFieldEntry::is_volatile_shift, notVolatile);
2735     __ membar(MacroAssembler::AnyAny);
2736     __ bind(notVolatile);
2737   }
2738 
2739   const Address field(obj, off);
2740 
2741   Label Done, notByte, notBool, notInt, notShort, notChar,
2742               notLong, notFloat, notObj, notDouble;
2743 
2744   assert(btos == 0, "change code, btos != 0");
2745   __ cbnz(tos_state, notByte);
2746 
2747   // Don't rewrite getstatic, only getfield
2748   if (is_static) rc = may_not_rewrite;
2749 
2750   // btos
2751   __ access_load_at(T_BYTE, IN_HEAP, r0, field, noreg, noreg);
2752   __ push(btos);
2753   // Rewrite bytecode to be faster
2754   if (rc == may_rewrite) {
2755     patch_bytecode(Bytecodes::_fast_bgetfield, bc, r1);
2756   }
2757   __ b(Done);
2758 
2759   __ bind(notByte);
2760   __ cmp(tos_state, (u1)ztos);
2761   __ br(Assembler::NE, notBool);
2762 
2763   // ztos (same code as btos)
2764   __ access_load_at(T_BOOLEAN, IN_HEAP, r0, field, noreg, noreg);
2765   __ push(ztos);
2766   // Rewrite bytecode to be faster
2767   if (rc == may_rewrite) {
2768     // use btos rewriting, no truncating to t/f bit is needed for getfield.
2769     patch_bytecode(Bytecodes::_fast_bgetfield, bc, r1);
2770   }
2771   __ b(Done);
2772 
2773   __ bind(notBool);
2774   __ cmp(tos_state, (u1)atos);
2775   __ br(Assembler::NE, notObj);
2776   // atos
2777   if (!Arguments::is_valhalla_enabled()) {
2778     do_oop_load(_masm, field, r0, IN_HEAP);
2779     __ push(atos);
2780     if (rc == may_rewrite) {
2781       patch_bytecode(Bytecodes::_fast_agetfield, bc, r1);
2782     }
2783     __ b(Done);
2784   } else { // Valhalla
2785     if (is_static) {
2786       __ load_heap_oop(r0, field, rscratch1, rscratch2);
2787       __ push(atos);
2788       __ b(Done);
2789     } else {
2790       Label is_flat;
2791       __ test_field_is_flat(flags, noreg /* temp */, is_flat);
2792       __ load_heap_oop(r0, field, rscratch1, rscratch2);
2793       __ push(atos);
2794       if (rc == may_rewrite) {
2795         patch_bytecode(Bytecodes::_fast_agetfield, bc, r1);
2796       }
2797       __ b(Done);
2798       __ bind(is_flat);
2799       // field is flat (null-free or nullable with a null-marker)
2800       __ mov(r0, obj);
2801       __ read_flat_field(cache, r0);
2802       __ verify_oop(r0);
2803       __ push(atos);
2804       if (rc == may_rewrite) {
2805         patch_bytecode(Bytecodes::_fast_vgetfield, bc, r1);
2806       }
2807       __ b(Done);
2808     }
2809   }

2810 
2811   __ bind(notObj);
2812   __ cmp(tos_state, (u1)itos);
2813   __ br(Assembler::NE, notInt);
2814   // itos
2815   __ access_load_at(T_INT, IN_HEAP, r0, field, noreg, noreg);
2816   __ push(itos);
2817   // Rewrite bytecode to be faster
2818   if (rc == may_rewrite) {
2819     patch_bytecode(Bytecodes::_fast_igetfield, bc, r1);
2820   }
2821   __ b(Done);
2822 
2823   __ bind(notInt);
2824   __ cmp(tos_state, (u1)ctos);
2825   __ br(Assembler::NE, notChar);
2826   // ctos
2827   __ access_load_at(T_CHAR, IN_HEAP, r0, field, noreg, noreg);
2828   __ push(ctos);
2829   // Rewrite bytecode to be faster
2830   if (rc == may_rewrite) {
2831     patch_bytecode(Bytecodes::_fast_cgetfield, bc, r1);
2832   }
2833   __ b(Done);
2834 
2835   __ bind(notChar);
2836   __ cmp(tos_state, (u1)stos);
2837   __ br(Assembler::NE, notShort);
2838   // stos
2839   __ access_load_at(T_SHORT, IN_HEAP, r0, field, noreg, noreg);
2840   __ push(stos);
2841   // Rewrite bytecode to be faster
2842   if (rc == may_rewrite) {
2843     patch_bytecode(Bytecodes::_fast_sgetfield, bc, r1);
2844   }
2845   __ b(Done);
2846 
2847   __ bind(notShort);
2848   __ cmp(tos_state, (u1)ltos);
2849   __ br(Assembler::NE, notLong);
2850   // ltos
2851   __ access_load_at(T_LONG, IN_HEAP, r0, field, noreg, noreg);
2852   __ push(ltos);
2853   // Rewrite bytecode to be faster
2854   if (rc == may_rewrite) {
2855     patch_bytecode(Bytecodes::_fast_lgetfield, bc, r1);
2856   }
2857   __ b(Done);
2858 
2859   __ bind(notLong);
2860   __ cmp(tos_state, (u1)ftos);
2861   __ br(Assembler::NE, notFloat);
2862   // ftos
2863   __ access_load_at(T_FLOAT, IN_HEAP, noreg /* ftos */, field, noreg, noreg);
2864   __ push(ftos);
2865   // Rewrite bytecode to be faster
2866   if (rc == may_rewrite) {
2867     patch_bytecode(Bytecodes::_fast_fgetfield, bc, r1);
2868   }
2869   __ b(Done);
2870 
2871   __ bind(notFloat);
2872 #ifdef ASSERT
2873   __ cmp(tos_state, (u1)dtos);
2874   __ br(Assembler::NE, notDouble);
2875 #endif
2876   // dtos
2877   __ access_load_at(T_DOUBLE, IN_HEAP, noreg /* ftos */, field, noreg, noreg);
2878   __ push(dtos);
2879   // Rewrite bytecode to be faster
2880   if (rc == may_rewrite) {
2881     patch_bytecode(Bytecodes::_fast_dgetfield, bc, r1);
2882   }
2883 #ifdef ASSERT
2884   __ b(Done);
2885 
2886   __ bind(notDouble);
2887   __ stop("Bad state");
2888 #endif
2889 
2890   __ bind(Done);
2891 
2892   Label notVolatile;
2893   __ tbz(flags, ResolvedFieldEntry::is_volatile_shift, notVolatile);
2894   __ membar(MacroAssembler::LoadLoad | MacroAssembler::LoadStore);
2895   __ bind(notVolatile);
2896 }
2897 
2898 
2899 void TemplateTable::getfield(int byte_no)
2900 {
2901   getfield_or_static(byte_no, false);
2902 }
2903 
2904 void TemplateTable::nofast_getfield(int byte_no) {
2905   getfield_or_static(byte_no, false, may_not_rewrite);
2906 }
2907 
2908 void TemplateTable::getstatic(int byte_no)
2909 {
2910   getfield_or_static(byte_no, true);
2911 }
2912 
2913 // The registers cache and index expected to be set before call.
2914 // The function may destroy various registers, just not the cache and index registers.
2915 void TemplateTable::jvmti_post_field_mod(Register cache, Register index, bool is_static) {
2916   transition(vtos, vtos);
2917 
2918   if (JvmtiExport::can_post_field_modification()) {
2919     // Check to see if a field modification watch has been set before
2920     // we take the time to call into the VM.
2921     Label L1;
2922     assert_different_registers(cache, index, r0);
2923     __ lea(rscratch1, ExternalAddress((address)JvmtiExport::get_field_modification_count_addr()));
2924     __ ldrw(r0, Address(rscratch1));
2925     __ cbz(r0, L1);
2926 
2927     __ mov(c_rarg2, cache);
2928 
2929     if (is_static) {
2930       // Life is simple.  Null out the object pointer.
2931       __ mov(c_rarg1, zr);
2932     } else {
2933       // Life is harder. The stack holds the value on top, followed by
2934       // the object.  We don't know the size of the value, though; it
2935       // could be one or two words depending on its type. As a result,
2936       // we must find the type to determine where the object is.
2937       __ load_unsigned_byte(c_rarg3, Address(c_rarg2, in_bytes(ResolvedFieldEntry::type_offset())));
2938       Label nope2, done, ok;
2939       __ ldr(c_rarg1, at_tos_p1());  // initially assume a one word jvalue
2940       __ cmpw(c_rarg3, ltos);
2941       __ br(Assembler::EQ, ok);
2942       __ cmpw(c_rarg3, dtos);
2943       __ br(Assembler::NE, nope2);
2944       __ bind(ok);
2945       __ ldr(c_rarg1, at_tos_p2()); // ltos (two word jvalue)
2946       __ bind(nope2);
2947     }
2948     // object (tos)
2949     __ mov(c_rarg3, esp);
2950     // c_rarg1: object pointer set up above (null if static)
2951     // c_rarg2: cache entry pointer
2952     // c_rarg3: jvalue object on the stack
2953     __ call_VM(noreg,
2954                CAST_FROM_FN_PTR(address,
2955                                 InterpreterRuntime::post_field_modification),
2956                c_rarg1, c_rarg2, c_rarg3);
2957     __ load_field_entry(cache, index);
2958     __ bind(L1);
2959   }
2960 }
2961 
2962 void TemplateTable::putfield_or_static(int byte_no, bool is_static, RewriteControl rc) {
2963   transition(vtos, vtos);
2964 
2965   const Register cache     = r2;
2966   const Register index     = r3;
2967   const Register tos_state = r3;
2968   const Register obj       = r2;
2969   const Register off       = r19;
2970   const Register flags     = r6;
2971   const Register bc        = r4;
2972 
2973   resolve_cache_and_index_for_field(byte_no, cache, index);
2974   jvmti_post_field_mod(cache, index, is_static);
2975   load_resolved_field_entry(obj, cache, tos_state, off, flags, is_static);
2976 
2977   Label Done;


2978   {
2979     Label notVolatile;
2980     __ tbz(flags, ResolvedFieldEntry::is_volatile_shift, notVolatile);
2981     __ membar(MacroAssembler::StoreStore | MacroAssembler::LoadStore);
2982     __ bind(notVolatile);
2983   }
2984 
2985   // field address
2986   const Address field(obj, off);
2987 
2988   Label notByte, notBool, notInt, notShort, notChar,
2989         notLong, notFloat, notObj, notDouble;
2990 
2991   assert(btos == 0, "change code, btos != 0");
2992   __ cbnz(tos_state, notByte);
2993 
2994   // Don't rewrite putstatic, only putfield
2995   if (is_static) rc = may_not_rewrite;
2996 
2997   // btos
2998   {
2999     __ pop(btos);
3000     if (!is_static) pop_and_check_object(obj);
3001     __ access_store_at(T_BYTE, IN_HEAP, field, r0, noreg, noreg, noreg);
3002     if (rc == may_rewrite) {
3003       patch_bytecode(Bytecodes::_fast_bputfield, bc, r1, true, byte_no);
3004     }
3005     __ b(Done);
3006   }
3007 
3008   __ bind(notByte);
3009   __ cmp(tos_state, (u1)ztos);
3010   __ br(Assembler::NE, notBool);
3011 
3012   // ztos
3013   {
3014     __ pop(ztos);
3015     if (!is_static) pop_and_check_object(obj);
3016     __ access_store_at(T_BOOLEAN, IN_HEAP, field, r0, noreg, noreg, noreg);
3017     if (rc == may_rewrite) {
3018       patch_bytecode(Bytecodes::_fast_zputfield, bc, r1, true, byte_no);
3019     }
3020     __ b(Done);
3021   }
3022 
3023   __ bind(notBool);
3024   __ cmp(tos_state, (u1)atos);
3025   __ br(Assembler::NE, notObj);
3026 
3027   // atos
3028   {
3029     if (!Arguments::is_valhalla_enabled()) {
3030       __ pop(atos);
3031       if (!is_static) pop_and_check_object(obj);
3032       // Store into the field
3033       // Clobbers: r10, r11, r3
3034       do_oop_store(_masm, field, r0, IN_HEAP);
3035       if (rc == may_rewrite) {
3036         patch_bytecode(Bytecodes::_fast_aputfield, bc, r1, true, byte_no);
3037       }
3038       __ b(Done);
3039     } else { // Valhalla
3040       __ pop(atos);
3041       if (is_static) {
3042         Label is_nullable;
3043         __ test_field_is_not_null_free_inline_type(flags, noreg /* temp */, is_nullable);
3044         __ null_check(r0);  // FIXME JDK-8341120
3045         __ bind(is_nullable);
3046         do_oop_store(_masm, field, r0, IN_HEAP);
3047         __ b(Done);
3048       } else {
3049         Label null_free_reference, is_flat, rewrite_inline;
3050         __ test_field_is_flat(flags, noreg /* temp */, is_flat);
3051         __ test_field_is_null_free_inline_type(flags, noreg /* temp */, null_free_reference);
3052         pop_and_check_object(obj);
3053         // Store into the field
3054         // Clobbers: r10, r11, r3
3055         do_oop_store(_masm, field, r0, IN_HEAP);
3056         if (rc == may_rewrite) {
3057           patch_bytecode(Bytecodes::_fast_aputfield, bc, r19, true, byte_no);
3058         }
3059         __ b(Done);
3060         // Implementation of the inline type semantic
3061         __ bind(null_free_reference);
3062         __ null_check(r0);  // FIXME JDK-8341120
3063         pop_and_check_object(obj);
3064         // Store into the field
3065         // Clobbers: r10, r11, r3
3066         do_oop_store(_masm, field, r0, IN_HEAP);
3067         __ b(rewrite_inline);
3068         __ bind(is_flat);
3069         pop_and_check_object(r7);
3070         __ write_flat_field(cache, off, index, flags, r7);
3071         __ bind(rewrite_inline);
3072         if (rc == may_rewrite) {
3073           patch_bytecode(Bytecodes::_fast_vputfield, bc, r19, true, byte_no);
3074         }
3075         __ b(Done);
3076       }
3077     } // Valhalla
3078   }
3079 
3080   __ bind(notObj);
3081   __ cmp(tos_state, (u1)itos);
3082   __ br(Assembler::NE, notInt);
3083 
3084   // itos
3085   {
3086     __ pop(itos);
3087     if (!is_static) pop_and_check_object(obj);
3088     __ access_store_at(T_INT, IN_HEAP, field, r0, noreg, noreg, noreg);
3089     if (rc == may_rewrite) {
3090       patch_bytecode(Bytecodes::_fast_iputfield, bc, r1, true, byte_no);
3091     }
3092     __ b(Done);
3093   }
3094 
3095   __ bind(notInt);
3096   __ cmp(tos_state, (u1)ctos);
3097   __ br(Assembler::NE, notChar);
3098 
3099   // ctos
3100   {
3101     __ pop(ctos);
3102     if (!is_static) pop_and_check_object(obj);
3103     __ access_store_at(T_CHAR, IN_HEAP, field, r0, noreg, noreg, noreg);
3104     if (rc == may_rewrite) {
3105       patch_bytecode(Bytecodes::_fast_cputfield, bc, r1, true, byte_no);
3106     }
3107     __ b(Done);
3108   }
3109 
3110   __ bind(notChar);
3111   __ cmp(tos_state, (u1)stos);
3112   __ br(Assembler::NE, notShort);
3113 
3114   // stos
3115   {
3116     __ pop(stos);
3117     if (!is_static) pop_and_check_object(obj);
3118     __ access_store_at(T_SHORT, IN_HEAP, field, r0, noreg, noreg, noreg);
3119     if (rc == may_rewrite) {
3120       patch_bytecode(Bytecodes::_fast_sputfield, bc, r1, true, byte_no);
3121     }
3122     __ b(Done);
3123   }
3124 
3125   __ bind(notShort);
3126   __ cmp(tos_state, (u1)ltos);
3127   __ br(Assembler::NE, notLong);
3128 
3129   // ltos
3130   {
3131     __ pop(ltos);
3132     if (!is_static) pop_and_check_object(obj);
3133     __ access_store_at(T_LONG, IN_HEAP, field, r0, noreg, noreg, noreg);
3134     if (rc == may_rewrite) {
3135       patch_bytecode(Bytecodes::_fast_lputfield, bc, r1, true, byte_no);
3136     }
3137     __ b(Done);
3138   }
3139 
3140   __ bind(notLong);
3141   __ cmp(tos_state, (u1)ftos);
3142   __ br(Assembler::NE, notFloat);
3143 
3144   // ftos
3145   {
3146     __ pop(ftos);
3147     if (!is_static) pop_and_check_object(obj);
3148     __ access_store_at(T_FLOAT, IN_HEAP, field, noreg /* ftos */, noreg, noreg, noreg);
3149     if (rc == may_rewrite) {
3150       patch_bytecode(Bytecodes::_fast_fputfield, bc, r1, true, byte_no);
3151     }
3152     __ b(Done);
3153   }
3154 
3155   __ bind(notFloat);
3156 #ifdef ASSERT
3157   __ cmp(tos_state, (u1)dtos);
3158   __ br(Assembler::NE, notDouble);
3159 #endif
3160 
3161   // dtos
3162   {
3163     __ pop(dtos);
3164     if (!is_static) pop_and_check_object(obj);
3165     __ access_store_at(T_DOUBLE, IN_HEAP, field, noreg /* dtos */, noreg, noreg, noreg);
3166     if (rc == may_rewrite) {
3167       patch_bytecode(Bytecodes::_fast_dputfield, bc, r1, true, byte_no);
3168     }
3169   }
3170 
3171 #ifdef ASSERT
3172   __ b(Done);
3173 
3174   __ bind(notDouble);
3175   __ stop("Bad state");
3176 #endif
3177 
3178   __ bind(Done);
3179 
3180   {
3181     Label notVolatile;
3182     __ tbz(flags, ResolvedFieldEntry::is_volatile_shift, notVolatile);
3183     __ membar(MacroAssembler::StoreLoad | MacroAssembler::StoreStore);
3184     __ bind(notVolatile);
3185   }
3186 }
3187 
3188 void TemplateTable::putfield(int byte_no)
3189 {
3190   putfield_or_static(byte_no, false);
3191 }
3192 
3193 void TemplateTable::nofast_putfield(int byte_no) {
3194   putfield_or_static(byte_no, false, may_not_rewrite);
3195 }
3196 
3197 void TemplateTable::putstatic(int byte_no) {
3198   putfield_or_static(byte_no, true);
3199 }
3200 
3201 void TemplateTable::jvmti_post_fast_field_mod() {
3202   if (JvmtiExport::can_post_field_modification()) {
3203     // Check to see if a field modification watch has been set before
3204     // we take the time to call into the VM.
3205     Label L2;
3206     __ lea(rscratch1, ExternalAddress((address)JvmtiExport::get_field_modification_count_addr()));
3207     __ ldrw(c_rarg3, Address(rscratch1));
3208     __ cbzw(c_rarg3, L2);
3209     __ pop_ptr(r19);                  // copy the object pointer from tos
3210     __ verify_oop(r19);
3211     __ push_ptr(r19);                 // put the object pointer back on tos
3212     // Save tos values before call_VM() clobbers them. Since we have
3213     // to do it for every data type, we use the saved values as the
3214     // jvalue object.
3215     switch (bytecode()) {          // load values into the jvalue object
3216     case Bytecodes::_fast_vputfield: // fall through
3217     case Bytecodes::_fast_aputfield: __ push_ptr(r0); break;
3218     case Bytecodes::_fast_bputfield: // fall through
3219     case Bytecodes::_fast_zputfield: // fall through
3220     case Bytecodes::_fast_sputfield: // fall through
3221     case Bytecodes::_fast_cputfield: // fall through
3222     case Bytecodes::_fast_iputfield: __ push_i(r0); break;
3223     case Bytecodes::_fast_dputfield: __ push_d(); break;
3224     case Bytecodes::_fast_fputfield: __ push_f(); break;
3225     case Bytecodes::_fast_lputfield: __ push_l(r0); break;
3226 
3227     default:
3228       ShouldNotReachHere();
3229     }
3230     __ mov(c_rarg3, esp);             // points to jvalue on the stack
3231     // access constant pool cache entry
3232     __ load_field_entry(c_rarg2, r0);
3233     __ verify_oop(r19);
3234     // r19: object pointer copied above
3235     // c_rarg2: cache entry pointer
3236     // c_rarg3: jvalue object on the stack
3237     __ call_VM(noreg,
3238                CAST_FROM_FN_PTR(address,
3239                                 InterpreterRuntime::post_field_modification),
3240                r19, c_rarg2, c_rarg3);
3241 
3242     switch (bytecode()) {             // restore tos values
3243     case Bytecodes::_fast_vputfield: // fall through
3244     case Bytecodes::_fast_aputfield: __ pop_ptr(r0); break;
3245     case Bytecodes::_fast_bputfield: // fall through
3246     case Bytecodes::_fast_zputfield: // fall through
3247     case Bytecodes::_fast_sputfield: // fall through
3248     case Bytecodes::_fast_cputfield: // fall through
3249     case Bytecodes::_fast_iputfield: __ pop_i(r0); break;
3250     case Bytecodes::_fast_dputfield: __ pop_d(); break;
3251     case Bytecodes::_fast_fputfield: __ pop_f(); break;
3252     case Bytecodes::_fast_lputfield: __ pop_l(r0); break;
3253     default: break;
3254     }
3255     __ bind(L2);
3256   }
3257 }
3258 
3259 void TemplateTable::fast_storefield(TosState state)
3260 {
3261   transition(state, vtos);
3262 
3263   ByteSize base = ConstantPoolCache::base_offset();
3264 
3265   jvmti_post_fast_field_mod();
3266 
3267   // access constant pool cache
3268   __ load_field_entry(r2, r1);
3269 
3270   // R1: field offset, R2: field holder, R5: flags
3271   load_resolved_field_entry(r2, r2, noreg, r1, r5);
3272   __ verify_field_offset(r1);
3273 
3274   {
3275     Label notVolatile;
3276     __ tbz(r5, ResolvedFieldEntry::is_volatile_shift, notVolatile);
3277     __ membar(MacroAssembler::StoreStore | MacroAssembler::LoadStore);
3278     __ bind(notVolatile);
3279   }
3280 
3281   Label notVolatile;
3282 
3283   // Get object from stack
3284   pop_and_check_object(r2);
3285 
3286   // field address
3287   const Address field(r2, r1);
3288 
3289   // access field
3290   switch (bytecode()) {
3291   case Bytecodes::_fast_vputfield:
3292     {
3293       Label is_flat, done;
3294       __ test_field_is_flat(r5, noreg /* temp */, is_flat);
3295       __ null_check(r0);
3296       do_oop_store(_masm, field, r0, IN_HEAP);
3297       __ b(done);
3298       __ bind(is_flat);
3299       __ load_field_entry(r4, r5);
3300       // Re-shuffle registers because of VM calls calling convention
3301       __ mov(r19, r1);
3302       __ mov(r7, r2);
3303       __ write_flat_field(r4, r19, r6, r8, r7);
3304       __ bind(done);
3305     }
3306     break;
3307   case Bytecodes::_fast_aputfield:
3308     // Clobbers: r10, r11, r3
3309     do_oop_store(_masm, field, r0, IN_HEAP);
3310     break;
3311   case Bytecodes::_fast_lputfield:
3312     __ access_store_at(T_LONG, IN_HEAP, field, r0, noreg, noreg, noreg);
3313     break;
3314   case Bytecodes::_fast_iputfield:
3315     __ access_store_at(T_INT, IN_HEAP, field, r0, noreg, noreg, noreg);
3316     break;
3317   case Bytecodes::_fast_zputfield:
3318     __ access_store_at(T_BOOLEAN, IN_HEAP, field, r0, noreg, noreg, noreg);
3319     break;
3320   case Bytecodes::_fast_bputfield:
3321     __ access_store_at(T_BYTE, IN_HEAP, field, r0, noreg, noreg, noreg);
3322     break;
3323   case Bytecodes::_fast_sputfield:
3324     __ access_store_at(T_SHORT, IN_HEAP, field, r0, noreg, noreg, noreg);
3325     break;
3326   case Bytecodes::_fast_cputfield:
3327     __ access_store_at(T_CHAR, IN_HEAP, field, r0, noreg, noreg, noreg);
3328     break;
3329   case Bytecodes::_fast_fputfield:
3330     __ access_store_at(T_FLOAT, IN_HEAP, field, noreg /* ftos */, noreg, noreg, noreg);
3331     break;
3332   case Bytecodes::_fast_dputfield:
3333     __ access_store_at(T_DOUBLE, IN_HEAP, field, noreg /* dtos */, noreg, noreg, noreg);
3334     break;
3335   default:
3336     ShouldNotReachHere();
3337   }
3338 
3339   {
3340     Label notVolatile;
3341     __ tbz(r5, ResolvedFieldEntry::is_volatile_shift, notVolatile);
3342     __ membar(MacroAssembler::StoreLoad | MacroAssembler::StoreStore);
3343     __ bind(notVolatile);
3344   }
3345 }
3346 
3347 
3348 void TemplateTable::fast_accessfield(TosState state)
3349 {
3350   transition(atos, state);
3351   // Do the JVMTI work here to avoid disturbing the register state below
3352   if (JvmtiExport::can_post_field_access()) {
3353     // Check to see if a field access watch has been set before we
3354     // take the time to call into the VM.
3355     Label L1;
3356     __ lea(rscratch1, ExternalAddress((address) JvmtiExport::get_field_access_count_addr()));
3357     __ ldrw(r2, Address(rscratch1));
3358     __ cbzw(r2, L1);
3359     // access constant pool cache entry
3360     __ load_field_entry(c_rarg2, rscratch2);
3361     __ verify_oop(r0);
3362     __ push_ptr(r0);  // save object pointer before call_VM() clobbers it
3363     __ mov(c_rarg1, r0);
3364     // c_rarg1: object pointer copied above
3365     // c_rarg2: cache entry pointer
3366     __ call_VM(noreg,
3367                CAST_FROM_FN_PTR(address,
3368                                 InterpreterRuntime::post_field_access),
3369                c_rarg1, c_rarg2);
3370     __ pop_ptr(r0); // restore object pointer
3371     __ bind(L1);
3372   }
3373 
3374   // access constant pool cache
3375   __ load_field_entry(r2, r1);
3376 
3377   __ load_sized_value(r1, Address(r2, in_bytes(ResolvedFieldEntry::field_offset_offset())), sizeof(int), true /*is_signed*/);
3378   __ verify_field_offset(r1);
3379 
3380   __ load_unsigned_byte(r3, Address(r2, in_bytes(ResolvedFieldEntry::flags_offset())));
3381 
3382   // r0: object
3383   __ verify_oop(r0);
3384   __ null_check(r0);
3385   const Address field(r0, r1);
3386 
3387   // 8179954: We need to make sure that the code generated for
3388   // volatile accesses forms a sequentially-consistent set of
3389   // operations when combined with STLR and LDAR.  Without a leading
3390   // membar it's possible for a simple Dekker test to fail if loads
3391   // use LDR;DMB but stores use STLR.  This can happen if C2 compiles
3392   // the stores in one method and we interpret the loads in another.
3393   if (!CompilerConfig::is_c1_or_interpreter_only_no_jvmci()) {
3394     Label notVolatile;
3395     __ tbz(r3, ResolvedFieldEntry::is_volatile_shift, notVolatile);
3396     __ membar(MacroAssembler::AnyAny);
3397     __ bind(notVolatile);
3398   }
3399 
3400   // access field
3401   switch (bytecode()) {
3402   case Bytecodes::_fast_vgetfield:
3403     {
3404       // field is flat
3405       __ read_flat_field(r2, r0);
3406       __ verify_oop(r0);
3407     }
3408     break;
3409   case Bytecodes::_fast_agetfield:
3410     do_oop_load(_masm, field, r0, IN_HEAP);
3411     __ verify_oop(r0);
3412     break;
3413   case Bytecodes::_fast_lgetfield:
3414     __ access_load_at(T_LONG, IN_HEAP, r0, field, noreg, noreg);
3415     break;
3416   case Bytecodes::_fast_igetfield:
3417     __ access_load_at(T_INT, IN_HEAP, r0, field, noreg, noreg);
3418     break;
3419   case Bytecodes::_fast_bgetfield:
3420     __ access_load_at(T_BYTE, IN_HEAP, r0, field, noreg, noreg);
3421     break;
3422   case Bytecodes::_fast_sgetfield:
3423     __ access_load_at(T_SHORT, IN_HEAP, r0, field, noreg, noreg);
3424     break;
3425   case Bytecodes::_fast_cgetfield:
3426     __ access_load_at(T_CHAR, IN_HEAP, r0, field, noreg, noreg);
3427     break;
3428   case Bytecodes::_fast_fgetfield:
3429     __ access_load_at(T_FLOAT, IN_HEAP, noreg /* ftos */, field, noreg, noreg);
3430     break;
3431   case Bytecodes::_fast_dgetfield:
3432     __ access_load_at(T_DOUBLE, IN_HEAP, noreg /* dtos */, field, noreg, noreg);
3433     break;
3434   default:
3435     ShouldNotReachHere();
3436   }
3437   {
3438     Label notVolatile;
3439     __ tbz(r3, ResolvedFieldEntry::is_volatile_shift, notVolatile);
3440     __ membar(MacroAssembler::LoadLoad | MacroAssembler::LoadStore);
3441     __ bind(notVolatile);
3442   }
3443 }
3444 
3445 void TemplateTable::fast_xaccess(TosState state)
3446 {
3447   transition(vtos, state);
3448 
3449   // get receiver
3450   __ ldr(r0, aaddress(0));
3451   // access constant pool cache
3452   __ load_field_entry(r2, r3, 2);
3453 
3454   __ load_sized_value(r1, Address(r2, in_bytes(ResolvedFieldEntry::field_offset_offset())), sizeof(int), true /*is_signed*/);
3455   __ verify_field_offset(r1);
3456 
3457   // 8179954: We need to make sure that the code generated for
3458   // volatile accesses forms a sequentially-consistent set of
3459   // operations when combined with STLR and LDAR.  Without a leading
3460   // membar it's possible for a simple Dekker test to fail if loads
3461   // use LDR;DMB but stores use STLR.  This can happen if C2 compiles
3462   // the stores in one method and we interpret the loads in another.
3463   if (!CompilerConfig::is_c1_or_interpreter_only_no_jvmci()) {
3464     Label notVolatile;
3465     __ load_unsigned_byte(r3, Address(r2, in_bytes(ResolvedFieldEntry::flags_offset())));
3466     __ tbz(r3, ResolvedFieldEntry::is_volatile_shift, notVolatile);
3467     __ membar(MacroAssembler::AnyAny);
3468     __ bind(notVolatile);
3469   }
3470 
3471   // make sure exception is reported in correct bcp range (getfield is
3472   // next instruction)
3473   __ increment(rbcp);
3474   __ null_check(r0);
3475   switch (state) {
3476   case itos:
3477     __ access_load_at(T_INT, IN_HEAP, r0, Address(r0, r1, Address::lsl(0)), noreg, noreg);
3478     break;
3479   case atos:
3480     do_oop_load(_masm, Address(r0, r1, Address::lsl(0)), r0, IN_HEAP);
3481     __ verify_oop(r0);
3482     break;
3483   case ftos:
3484     __ access_load_at(T_FLOAT, IN_HEAP, noreg /* ftos */, Address(r0, r1, Address::lsl(0)), noreg, noreg);
3485     break;
3486   default:
3487     ShouldNotReachHere();
3488   }
3489 
3490   {
3491     Label notVolatile;
3492     __ load_unsigned_byte(r3, Address(r2, in_bytes(ResolvedFieldEntry::flags_offset())));
3493     __ tbz(r3, ResolvedFieldEntry::is_volatile_shift, notVolatile);
3494     __ membar(MacroAssembler::LoadLoad | MacroAssembler::LoadStore);
3495     __ bind(notVolatile);
3496   }
3497 
3498   __ decrement(rbcp);
3499 }
3500 
3501 
3502 
3503 //-----------------------------------------------------------------------------
3504 // Calls
3505 
3506 void TemplateTable::prepare_invoke(Register cache, Register recv) {
3507 
3508   Bytecodes::Code code = bytecode();
3509   const bool load_receiver       = (code != Bytecodes::_invokestatic) && (code != Bytecodes::_invokedynamic);
3510 
3511   // save 'interpreter return address'
3512   __ save_bcp();
3513 
3514   // Load TOS state for later
3515   __ load_unsigned_byte(rscratch2, Address(cache, in_bytes(ResolvedMethodEntry::type_offset())));
3516 
3517   // load receiver if needed (note: no return address pushed yet)
3518   if (load_receiver) {
3519     __ load_unsigned_short(recv, Address(cache, in_bytes(ResolvedMethodEntry::num_parameters_offset())));
3520     __ add(rscratch1, esp, recv, ext::uxtx, 3);
3521     __ ldr(recv, Address(rscratch1, -Interpreter::expr_offset_in_bytes(1)));
3522     __ verify_oop(recv);
3523   }
3524 
3525   // load return address
3526   {
3527     const address table_addr = (address) Interpreter::invoke_return_entry_table_for(code);
3528     __ mov(rscratch1, table_addr);
3529     __ ldr(lr, Address(rscratch1, rscratch2, Address::lsl(3)));
3530   }
3531 }
3532 
3533 
3534 void TemplateTable::invokevirtual_helper(Register index,
3535                                          Register recv,
3536                                          Register flags)
3537 {
3538   // Uses temporary registers r0, r3
3539   assert_different_registers(index, recv, r0, r3);
3540   // Test for an invoke of a final method
3541   Label notFinal;
3542   __ tbz(flags, ResolvedMethodEntry::is_vfinal_shift, notFinal);
3543 
3544   const Register method = index;  // method must be rmethod
3545   assert(method == rmethod,
3546          "Method must be rmethod for interpreter calling convention");
3547 
3548   // do the call - the index is actually the method to call
3549   // that is, f2 is a vtable index if !is_vfinal, else f2 is a Method*
3550 
3551   // It's final, need a null check here!
3552   __ null_check(recv);
3553 
3554   // profile this call
3555   __ profile_final_call(r0);
3556   __ profile_arguments_type(r0, method, r4, true);
3557 
3558   __ jump_from_interpreted(method, r0);
3559 
3560   __ bind(notFinal);
3561 
3562   // get receiver klass
3563   __ load_klass(r0, recv);
3564 
3565   // profile this call
3566   __ profile_virtual_call(r0, rlocals);
3567 
3568   // get target Method & entry point
3569   __ lookup_virtual_method(r0, index, method);
3570   __ profile_arguments_type(r3, method, r4, true);
3571   // FIXME -- this looks completely redundant. is it?
3572   // __ ldr(r3, Address(method, Method::interpreter_entry_offset()));
3573   __ jump_from_interpreted(method, r3);
3574 }
3575 
3576 void TemplateTable::invokevirtual(int byte_no)
3577 {
3578   transition(vtos, vtos);
3579   assert(byte_no == f2_byte, "use this argument");
3580 
3581   load_resolved_method_entry_virtual(r2,      // ResolvedMethodEntry*
3582                                      rmethod, // Method* or itable index
3583                                      r3);     // flags
3584   prepare_invoke(r2, r2); // recv
3585 
3586   // rmethod: index (actually a Method*)
3587   // r2: receiver
3588   // r3: flags
3589 
3590   invokevirtual_helper(rmethod, r2, r3);
3591 }
3592 
3593 void TemplateTable::invokespecial(int byte_no)
3594 {
3595   transition(vtos, vtos);
3596   assert(byte_no == f1_byte, "use this argument");
3597 
3598   load_resolved_method_entry_special_or_static(r2,      // ResolvedMethodEntry*
3599                                                rmethod, // Method*
3600                                                r3);     // flags
3601   prepare_invoke(r2, r2);  // get receiver also for null check
3602   __ verify_oop(r2);
3603   __ null_check(r2);
3604   // do the call
3605   __ profile_call(r0);
3606   __ profile_arguments_type(r0, rmethod, rbcp, false);
3607   __ jump_from_interpreted(rmethod, r0);
3608 }
3609 
3610 void TemplateTable::invokestatic(int byte_no)
3611 {
3612   transition(vtos, vtos);
3613   assert(byte_no == f1_byte, "use this argument");
3614 
3615   load_resolved_method_entry_special_or_static(r2,      // ResolvedMethodEntry*
3616                                                rmethod, // Method*
3617                                                r3);     // flags
3618   prepare_invoke(r2, r2);  // get receiver also for null check
3619 
3620   // do the call
3621   __ profile_call(r0);
3622   __ profile_arguments_type(r0, rmethod, r4, false);
3623   __ jump_from_interpreted(rmethod, r0);
3624 }
3625 
3626 void TemplateTable::fast_invokevfinal(int byte_no)
3627 {
3628   __ call_Unimplemented();
3629 }
3630 
3631 void TemplateTable::invokeinterface(int byte_no) {
3632   transition(vtos, vtos);
3633   assert(byte_no == f1_byte, "use this argument");
3634 
3635   load_resolved_method_entry_interface(r2,      // ResolvedMethodEntry*
3636                                        r0,      // Klass*
3637                                        rmethod, // Method* or itable/vtable index
3638                                        r3);     // flags
3639   prepare_invoke(r2, r2); // receiver
3640 
3641   // r0: interface klass (from f1)
3642   // rmethod: method (from f2)
3643   // r2: receiver
3644   // r3: flags
3645 
3646   // First check for Object case, then private interface method,
3647   // then regular interface method.
3648 
3649   // Special case of invokeinterface called for virtual method of
3650   // java.lang.Object.  See cpCache.cpp for details.
3651   Label notObjectMethod;
3652   __ tbz(r3, ResolvedMethodEntry::is_forced_virtual_shift, notObjectMethod);
3653 
3654   invokevirtual_helper(rmethod, r2, r3);
3655   __ bind(notObjectMethod);
3656 
3657   Label no_such_interface;
3658 
3659   // Check for private method invocation - indicated by vfinal
3660   Label notVFinal;
3661   __ tbz(r3, ResolvedMethodEntry::is_vfinal_shift, notVFinal);
3662 
3663   // Get receiver klass into r3
3664   __ load_klass(r3, r2);
3665 
3666   Label subtype;
3667   __ check_klass_subtype(r3, r0, r4, subtype);
3668   // If we get here the typecheck failed
3669   __ b(no_such_interface);
3670   __ bind(subtype);
3671 
3672   __ profile_final_call(r0);
3673   __ profile_arguments_type(r0, rmethod, r4, true);
3674   __ jump_from_interpreted(rmethod, r0);
3675 
3676   __ bind(notVFinal);
3677 
3678   // Get receiver klass into r3
3679   __ load_klass(r3, r2);
3680 
3681   Label no_such_method;
3682 
3683   // Preserve method for throw_AbstractMethodErrorVerbose.
3684   __ mov(r16, rmethod);
3685   // Receiver subtype check against REFC.
3686   // Superklass in r0. Subklass in r3. Blows rscratch2, r13
3687   __ lookup_interface_method(// inputs: rec. class, interface, itable index
3688                              r3, r0, noreg,
3689                              // outputs: scan temp. reg, scan temp. reg
3690                              rscratch2, r13,
3691                              no_such_interface,
3692                              /*return_method=*/false);
3693 
3694   // profile this call
3695   __ profile_virtual_call(r3, r13);
3696 
3697   // Get declaring interface class from method, and itable index
3698 
3699   __ load_method_holder(r0, rmethod);
3700   __ ldrw(rmethod, Address(rmethod, Method::itable_index_offset()));
3701   __ subw(rmethod, rmethod, Method::itable_index_max);
3702   __ negw(rmethod, rmethod);
3703 
3704   // Preserve recvKlass for throw_AbstractMethodErrorVerbose.
3705   __ mov(rlocals, r3);
3706   __ lookup_interface_method(// inputs: rec. class, interface, itable index
3707                              rlocals, r0, rmethod,
3708                              // outputs: method, scan temp. reg
3709                              rmethod, r13,
3710                              no_such_interface);
3711 
3712   // rmethod,: Method to call
3713   // r2: receiver
3714   // Check for abstract method error
3715   // Note: This should be done more efficiently via a throw_abstract_method_error
3716   //       interpreter entry point and a conditional jump to it in case of a null
3717   //       method.
3718   __ cbz(rmethod, no_such_method);
3719 
3720   __ profile_arguments_type(r3, rmethod, r13, true);
3721 
3722   // do the call
3723   // r2: receiver
3724   // rmethod,: Method
3725   __ jump_from_interpreted(rmethod, r3);
3726   __ should_not_reach_here();
3727 
3728   // exception handling code follows...
3729   // note: must restore interpreter registers to canonical
3730   //       state for exception handling to work correctly!
3731 
3732   __ bind(no_such_method);
3733   // throw exception
3734   __ restore_bcp();      // bcp must be correct for exception handler   (was destroyed)
3735   __ restore_locals();   // make sure locals pointer is correct as well (was destroyed)
3736   // Pass arguments for generating a verbose error message.
3737   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodErrorVerbose), r3, r16);
3738   // the call_VM checks for exception, so we should never return here.
3739   __ should_not_reach_here();
3740 
3741   __ bind(no_such_interface);
3742   // throw exception
3743   __ restore_bcp();      // bcp must be correct for exception handler   (was destroyed)
3744   __ restore_locals();   // make sure locals pointer is correct as well (was destroyed)
3745   // Pass arguments for generating a verbose error message.
3746   __ call_VM(noreg, CAST_FROM_FN_PTR(address,
3747                    InterpreterRuntime::throw_IncompatibleClassChangeErrorVerbose), r3, r0);
3748   // the call_VM checks for exception, so we should never return here.
3749   __ should_not_reach_here();
3750   return;
3751 }
3752 
3753 void TemplateTable::invokehandle(int byte_no) {
3754   transition(vtos, vtos);
3755   assert(byte_no == f1_byte, "use this argument");
3756 
3757   load_resolved_method_entry_handle(r2,      // ResolvedMethodEntry*
3758                                     rmethod, // Method*
3759                                     r0,      // Resolved reference
3760                                     r3);     // flags
3761   prepare_invoke(r2, r2);
3762 
3763   __ verify_method_ptr(r2);
3764   __ verify_oop(r2);
3765   __ null_check(r2);
3766 
3767   // FIXME: profile the LambdaForm also
3768 
3769   // r13 is safe to use here as a scratch reg because it is about to
3770   // be clobbered by jump_from_interpreted().
3771   __ profile_final_call(r13);
3772   __ profile_arguments_type(r13, rmethod, r4, true);
3773 
3774   __ jump_from_interpreted(rmethod, r0);
3775 }
3776 
3777 void TemplateTable::invokedynamic(int byte_no) {
3778   transition(vtos, vtos);
3779   assert(byte_no == f1_byte, "use this argument");
3780 
3781   load_invokedynamic_entry(rmethod);
3782 
3783   // r0: CallSite object (from cpool->resolved_references[])
3784   // rmethod: MH.linkToCallSite method
3785 
3786   // Note:  r0_callsite is already pushed
3787 
3788   // %%% should make a type profile for any invokedynamic that takes a ref argument
3789   // profile this call
3790   __ profile_call(rbcp);
3791   __ profile_arguments_type(r3, rmethod, r13, false);
3792 
3793   __ verify_oop(r0);
3794 
3795   __ jump_from_interpreted(rmethod, r0);
3796 }
3797 
3798 
3799 //-----------------------------------------------------------------------------
3800 // Allocation
3801 
3802 void TemplateTable::_new() {
3803   transition(vtos, atos);
3804 
3805   __ get_unsigned_2_byte_index_at_bcp(r3, 1);
3806   Label slow_case;
3807   Label done;
3808   Label initialize_header;
3809 
3810   __ get_cpool_and_tags(r4, r0);
3811   // Make sure the class we're about to instantiate has been resolved.
3812   // This is done before loading InstanceKlass to be consistent with the order
3813   // how Constant Pool is updated (see ConstantPool::klass_at_put)
3814   const int tags_offset = Array<u1>::base_offset_in_bytes();
3815   __ lea(rscratch1, Address(r0, r3, Address::lsl(0)));
3816   __ lea(rscratch1, Address(rscratch1, tags_offset));
3817   __ ldarb(rscratch1, rscratch1);
3818   __ cmp(rscratch1, (u1)JVM_CONSTANT_Class);
3819   __ br(Assembler::NE, slow_case);
3820 
3821   // get InstanceKlass
3822   __ load_resolved_klass_at_offset(r4, r3, r4, rscratch1);
3823 
3824   // make sure klass is initialized
3825   assert(VM_Version::supports_fast_class_init_checks(), "Optimization requires support for fast class initialization checks");
3826   __ clinit_barrier(r4, rscratch1, nullptr /*L_fast_path*/, &slow_case);
3827 
3828   // get instance_size in InstanceKlass (scaled to a count of bytes)
3829   __ ldrw(r3,
3830           Address(r4,
3831                   Klass::layout_helper_offset()));
3832   // test to see if it is malformed in some way
3833   __ tbnz(r3, exact_log2(Klass::_lh_instance_slow_path_bit), slow_case);
3834 
3835   // Allocate the instance:
3836   //  If TLAB is enabled:
3837   //    Try to allocate in the TLAB.
3838   //    If fails, go to the slow path.
3839   //    Initialize the allocation.
3840   //    Exit.
3841   //
3842   //  Go to slow path.
3843 
3844   if (UseTLAB) {
3845     __ tlab_allocate(r0, r3, 0, noreg, r1, slow_case);
3846 
3847     if (ZeroTLAB) {
3848       // the fields have been already cleared
3849       __ b(initialize_header);
3850     }
3851 
3852     // The object is initialized before the header.  If the object size is
3853     // zero, go directly to the header initialization.
3854     int header_size = oopDesc::header_size() * HeapWordSize;
3855     assert(is_aligned(header_size, BytesPerLong), "oop header size must be 8-byte-aligned");
3856     __ sub(r3, r3, header_size);
3857     __ cbz(r3, initialize_header);
3858 
3859   #ifdef ASSERT
3860     // make sure instance_size was multiple of 8
3861     Label L;
3862     __ tst(r3, 7);
3863     __ br(Assembler::EQ, L);
3864     __ stop("object size is not multiple of 8 - adjust this code");
3865     __ bind(L);
3866     // must be > 0, no extra check needed here
3867   #endif
3868 
3869     // Initialize object fields
3870     {
3871       __ add(r2, r0, header_size);
3872       Label loop;
3873       __ bind(loop);
3874       __ str(zr, Address(__ post(r2, BytesPerLong)));
3875       __ sub(r3, r3, BytesPerLong);
3876       __ cbnz(r3, loop);
3877     }
3878 
3879     // initialize object header only.
3880     __ bind(initialize_header);
3881     if (UseCompactObjectHeaders || Arguments::is_valhalla_enabled()) {
3882       __ ldr(rscratch1, Address(r4, Klass::prototype_header_offset()));
3883       __ str(rscratch1, Address(r0, oopDesc::mark_offset_in_bytes()));
3884     } else {
3885       __ mov(rscratch1, (intptr_t)markWord::prototype().value());
3886       __ str(rscratch1, Address(r0, oopDesc::mark_offset_in_bytes()));
3887     }
3888     if (!UseCompactObjectHeaders) {
3889       __ store_klass_gap(r0, zr);  // zero klass gap for compressed oops
3890       __ store_klass(r0, r4);      // store klass last
3891     }
3892 
3893     if (DTraceAllocProbes) {
3894       // Trigger dtrace event for fastpath
3895       __ push(atos); // save the return value
3896       __ call_VM_leaf(
3897            CAST_FROM_FN_PTR(address, static_cast<int (*)(oopDesc*)>(SharedRuntime::dtrace_object_alloc)), r0);
3898       __ pop(atos); // restore the return value
3899 
3900     }
3901     __ b(done);
3902   }
3903 
3904   // slow case
3905   __ bind(slow_case);
3906   __ get_constant_pool(c_rarg1);
3907   __ get_unsigned_2_byte_index_at_bcp(c_rarg2, 1);
3908   __ call_VM_preemptable(r0, CAST_FROM_FN_PTR(address, InterpreterRuntime::_new), c_rarg1, c_rarg2);
3909   __ verify_oop(r0);
3910 
3911   // continue
3912   __ bind(done);
3913   // Must prevent reordering of stores for object initialization with stores that publish the new object.
3914   __ membar(Assembler::StoreStore);
3915 }
3916 
3917 void TemplateTable::newarray() {
3918   transition(itos, atos);
3919   __ load_unsigned_byte(c_rarg1, at_bcp(1));
3920   __ mov(c_rarg2, r0);
3921   call_VM(r0, CAST_FROM_FN_PTR(address, InterpreterRuntime::newarray),
3922           c_rarg1, c_rarg2);
3923   // Must prevent reordering of stores for object initialization with stores that publish the new object.
3924   __ membar(Assembler::StoreStore);
3925 }
3926 
3927 void TemplateTable::anewarray() {
3928   transition(itos, atos);
3929   __ get_unsigned_2_byte_index_at_bcp(c_rarg2, 1);
3930   __ get_constant_pool(c_rarg1);
3931   __ mov(c_rarg3, r0);
3932   call_VM(r0, CAST_FROM_FN_PTR(address, InterpreterRuntime::anewarray),
3933           c_rarg1, c_rarg2, c_rarg3);
3934   // Must prevent reordering of stores for object initialization with stores that publish the new object.
3935   __ membar(Assembler::StoreStore);
3936 }
3937 
3938 void TemplateTable::arraylength() {
3939   transition(atos, itos);
3940   __ ldrw(r0, Address(r0, arrayOopDesc::length_offset_in_bytes()));
3941 }
3942 
3943 void TemplateTable::checkcast()
3944 {
3945   transition(atos, atos);
3946   Label done, is_null, ok_is_subtype, quicked, resolved;
3947   __ cbz(r0, is_null);
3948 
3949   // Get cpool & tags index
3950   __ get_cpool_and_tags(r2, r3); // r2=cpool, r3=tags array
3951   __ get_unsigned_2_byte_index_at_bcp(r19, 1); // r19=index
3952   // See if bytecode has already been quicked
3953   __ add(rscratch1, r3, Array<u1>::base_offset_in_bytes());
3954   __ lea(r1, Address(rscratch1, r19));
3955   __ ldarb(r1, r1);
3956   __ cmp(r1, (u1)JVM_CONSTANT_Class);
3957   __ br(Assembler::EQ, quicked);
3958 
3959   __ push(atos); // save receiver for result, and for GC
3960   call_VM(r0, CAST_FROM_FN_PTR(address, InterpreterRuntime::quicken_io_cc));
3961   __ get_vm_result_metadata(r0, rthread);
3962   __ pop(r3); // restore receiver
3963   __ b(resolved);
3964 
3965   // Get superklass in r0 and subklass in r3
3966   __ bind(quicked);
3967   __ mov(r3, r0); // Save object in r3; r0 needed for subtype check
3968   __ load_resolved_klass_at_offset(r2, r19, r0, rscratch1); // r0 = klass
3969 
3970   __ bind(resolved);
3971   __ load_klass(r19, r3);
3972 
3973   // Generate subtype check.  Blows r2, r5.  Object in r3.
3974   // Superklass in r0.  Subklass in r19.
3975   __ gen_subtype_check(r19, ok_is_subtype);
3976 
3977   // Come here on failure
3978   __ push(r3);
3979   // object is at TOS
3980   __ b(Interpreter::_throw_ClassCastException_entry);
3981 
3982   // Come here on success
3983   __ bind(ok_is_subtype);
3984   __ mov(r0, r3); // Restore object in r3
3985 
3986   __ b(done);
3987   __ bind(is_null);
3988 
3989   // Collect counts on whether this test sees nulls a lot or not.
3990   if (ProfileInterpreter) {


3991     __ profile_null_seen(r2);


3992   }
3993   __ bind(done);
3994 }
3995 
3996 void TemplateTable::instanceof() {
3997   transition(atos, itos);
3998   Label done, is_null, ok_is_subtype, quicked, resolved;
3999   __ cbz(r0, is_null);
4000 
4001   // Get cpool & tags index
4002   __ get_cpool_and_tags(r2, r3); // r2=cpool, r3=tags array
4003   __ get_unsigned_2_byte_index_at_bcp(r19, 1); // r19=index
4004   // See if bytecode has already been quicked
4005   __ add(rscratch1, r3, Array<u1>::base_offset_in_bytes());
4006   __ lea(r1, Address(rscratch1, r19));
4007   __ ldarb(r1, r1);
4008   __ cmp(r1, (u1)JVM_CONSTANT_Class);
4009   __ br(Assembler::EQ, quicked);
4010 
4011   __ push(atos); // save receiver for result, and for GC
4012   call_VM(r0, CAST_FROM_FN_PTR(address, InterpreterRuntime::quicken_io_cc));
4013   __ get_vm_result_metadata(r0, rthread);
4014   __ pop(r3); // restore receiver
4015   __ verify_oop(r3);
4016   __ load_klass(r3, r3);
4017   __ b(resolved);
4018 
4019   // Get superklass in r0 and subklass in r3
4020   __ bind(quicked);
4021   __ load_klass(r3, r0);
4022   __ load_resolved_klass_at_offset(r2, r19, r0, rscratch1);
4023 
4024   __ bind(resolved);
4025 
4026   // Generate subtype check.  Blows r2, r5
4027   // Superklass in r0.  Subklass in r3.
4028   __ gen_subtype_check(r3, ok_is_subtype);
4029 
4030   // Come here on failure
4031   __ mov(r0, 0);
4032   __ b(done);
4033   // Come here on success
4034   __ bind(ok_is_subtype);
4035   __ mov(r0, 1);
4036 
4037   // Collect counts on whether this test sees nulls a lot or not.
4038   if (ProfileInterpreter) {
4039     __ b(done);
4040     __ bind(is_null);
4041     __ profile_null_seen(r2);
4042   } else {
4043     __ bind(is_null);   // same as 'done'
4044   }
4045   __ bind(done);
4046   // r0 = 0: obj == nullptr or  obj is not an instanceof the specified klass
4047   // r0 = 1: obj != nullptr and obj is     an instanceof the specified klass
4048 }
4049 
4050 //-----------------------------------------------------------------------------
4051 // Breakpoints
4052 void TemplateTable::_breakpoint() {
4053   // Note: We get here even if we are single stepping..
4054   // jbug inists on setting breakpoints at every bytecode
4055   // even if we are in single step mode.
4056 
4057   transition(vtos, vtos);
4058 
4059   // get the unpatched byte code
4060   __ get_method(c_rarg1);
4061   __ call_VM(noreg,
4062              CAST_FROM_FN_PTR(address,
4063                               InterpreterRuntime::get_original_bytecode_at),
4064              c_rarg1, rbcp);
4065   __ mov(r19, r0);
4066 
4067   // post the breakpoint event
4068   __ call_VM(noreg,
4069              CAST_FROM_FN_PTR(address, InterpreterRuntime::_breakpoint),
4070              rmethod, rbcp);
4071 
4072   // complete the execution of original bytecode
4073   __ mov(rscratch1, r19);
4074   __ dispatch_only_normal(vtos);
4075 }
4076 
4077 //-----------------------------------------------------------------------------
4078 // Exceptions
4079 
4080 void TemplateTable::athrow() {
4081   transition(atos, vtos);
4082   __ null_check(r0);
4083   __ b(Interpreter::throw_exception_entry());
4084 }
4085 
4086 //-----------------------------------------------------------------------------
4087 // Synchronization
4088 //
4089 // Note: monitorenter & exit are symmetric routines; which is reflected
4090 //       in the assembly code structure as well
4091 //
4092 // Stack layout:
4093 //
4094 // [expressions  ] <--- esp               = expression stack top
4095 // ..
4096 // [expressions  ]
4097 // [monitor entry] <--- monitor block top = expression stack bot
4098 // ..
4099 // [monitor entry]
4100 // [frame data   ] <--- monitor block bot
4101 // ...
4102 // [saved rfp    ] <--- rfp
4103 void TemplateTable::monitorenter()
4104 {
4105   transition(atos, vtos);
4106 
4107   // check for null object
4108   __ null_check(r0);
4109 
4110   Label is_inline_type;
4111   __ ldr(rscratch1, Address(r0, oopDesc::mark_offset_in_bytes()));
4112   __ test_markword_is_inline_type(rscratch1, is_inline_type);
4113 
4114   const Address monitor_block_top(
4115         rfp, frame::interpreter_frame_monitor_block_top_offset * wordSize);
4116   const Address monitor_block_bot(
4117         rfp, frame::interpreter_frame_initial_sp_offset * wordSize);
4118   const int entry_size = frame::interpreter_frame_monitor_size_in_bytes();
4119 
4120   Label allocated;
4121 
4122   // initialize entry pointer
4123   __ mov(c_rarg1, zr); // points to free slot or null
4124 
4125   // find a free slot in the monitor block (result in c_rarg1)
4126   {
4127     Label entry, loop, exit;
4128     __ ldr(c_rarg3, monitor_block_top); // derelativize pointer
4129     __ lea(c_rarg3, Address(rfp, c_rarg3, Address::lsl(Interpreter::logStackElementSize)));
4130     // c_rarg3 points to current entry, starting with top-most entry
4131 
4132     __ lea(c_rarg2, monitor_block_bot); // points to word before bottom
4133 
4134     __ b(entry);
4135 
4136     __ bind(loop);
4137     // check if current entry is used
4138     // if not used then remember entry in c_rarg1
4139     __ ldr(rscratch1, Address(c_rarg3, BasicObjectLock::obj_offset()));
4140     __ cmp(zr, rscratch1);
4141     __ csel(c_rarg1, c_rarg3, c_rarg1, Assembler::EQ);
4142     // check if current entry is for same object
4143     __ cmp(r0, rscratch1);
4144     // if same object then stop searching
4145     __ br(Assembler::EQ, exit);
4146     // otherwise advance to next entry
4147     __ add(c_rarg3, c_rarg3, entry_size);
4148     __ bind(entry);
4149     // check if bottom reached
4150     __ cmp(c_rarg3, c_rarg2);
4151     // if not at bottom then check this entry
4152     __ br(Assembler::NE, loop);
4153     __ bind(exit);
4154   }
4155 
4156   __ cbnz(c_rarg1, allocated); // check if a slot has been found and
4157                             // if found, continue with that on
4158 
4159   // allocate one if there's no free slot
4160   {
4161     Label entry, loop;
4162     // 1. compute new pointers            // rsp: old expression stack top
4163 
4164     __ check_extended_sp();
4165     __ sub(sp, sp, entry_size);           // make room for the monitor
4166     __ sub(rscratch1, sp, rfp);
4167     __ asr(rscratch1, rscratch1, Interpreter::logStackElementSize);
4168     __ str(rscratch1, Address(rfp, frame::interpreter_frame_extended_sp_offset * wordSize));
4169 
4170     __ ldr(c_rarg1, monitor_block_bot);   // derelativize pointer
4171     __ lea(c_rarg1, Address(rfp, c_rarg1, Address::lsl(Interpreter::logStackElementSize)));
4172     // c_rarg1 points to the old expression stack bottom
4173 
4174     __ sub(esp, esp, entry_size);         // move expression stack top
4175     __ sub(c_rarg1, c_rarg1, entry_size); // move expression stack bottom
4176     __ mov(c_rarg3, esp);                 // set start value for copy loop
4177     __ sub(rscratch1, c_rarg1, rfp);      // relativize pointer
4178     __ asr(rscratch1, rscratch1, Interpreter::logStackElementSize);
4179     __ str(rscratch1, monitor_block_bot);  // set new monitor block bottom
4180 
4181     __ b(entry);
4182     // 2. move expression stack contents
4183     __ bind(loop);
4184     __ ldr(c_rarg2, Address(c_rarg3, entry_size)); // load expression stack
4185                                                    // word from old location
4186     __ str(c_rarg2, Address(c_rarg3, 0));          // and store it at new location
4187     __ add(c_rarg3, c_rarg3, wordSize);            // advance to next word
4188     __ bind(entry);
4189     __ cmp(c_rarg3, c_rarg1);        // check if bottom reached
4190     __ br(Assembler::NE, loop);      // if not at bottom then
4191                                      // copy next word
4192   }
4193 
4194   // call run-time routine
4195   // c_rarg1: points to monitor entry
4196   __ bind(allocated);
4197 
4198   // Increment bcp to point to the next bytecode, so exception
4199   // handling for async. exceptions work correctly.
4200   // The object has already been popped from the stack, so the
4201   // expression stack looks correct.
4202   __ increment(rbcp);
4203 
4204   // store object
4205   __ str(r0, Address(c_rarg1, BasicObjectLock::obj_offset()));
4206   __ lock_object(c_rarg1);
4207 
4208   // check to make sure this monitor doesn't cause stack overflow after locking
4209   __ save_bcp();  // in case of exception
4210   __ generate_stack_overflow_check(0);
4211 
4212   // The bcp has already been incremented. Just need to dispatch to
4213   // next instruction.
4214   __ dispatch_next(vtos);
4215 
4216   __ bind(is_inline_type);
4217   __ call_VM(noreg, CAST_FROM_FN_PTR(address,
4218                     InterpreterRuntime::throw_identity_exception), r0);
4219   __ should_not_reach_here();
4220 }
4221 
4222 
4223 void TemplateTable::monitorexit()
4224 {
4225   transition(atos, vtos);
4226 
4227   // check for null object
4228   __ null_check(r0);
4229 
4230   const int is_inline_type_mask = markWord::inline_type_pattern;
4231   Label has_identity;
4232   __ ldr(rscratch1, Address(r0, oopDesc::mark_offset_in_bytes()));
4233   __ mov(rscratch2, is_inline_type_mask);
4234   __ andr(rscratch1, rscratch1, rscratch2);
4235   __ cmp(rscratch1, rscratch2);
4236   __ br(Assembler::NE, has_identity);
4237   __ call_VM(noreg, CAST_FROM_FN_PTR(address,
4238                      InterpreterRuntime::throw_illegal_monitor_state_exception));
4239   __ should_not_reach_here();
4240   __ bind(has_identity);
4241 
4242   const Address monitor_block_top(
4243         rfp, frame::interpreter_frame_monitor_block_top_offset * wordSize);
4244   const Address monitor_block_bot(
4245         rfp, frame::interpreter_frame_initial_sp_offset * wordSize);
4246   const int entry_size = frame::interpreter_frame_monitor_size_in_bytes();
4247 
4248   Label found;
4249 
4250   // find matching slot
4251   {
4252     Label entry, loop;
4253     __ ldr(c_rarg1, monitor_block_top); // derelativize pointer
4254     __ lea(c_rarg1, Address(rfp, c_rarg1, Address::lsl(Interpreter::logStackElementSize)));
4255     // c_rarg1 points to current entry, starting with top-most entry
4256 
4257     __ lea(c_rarg2, monitor_block_bot); // points to word before bottom
4258                                         // of monitor block
4259     __ b(entry);
4260 
4261     __ bind(loop);
4262     // check if current entry is for same object
4263     __ ldr(rscratch1, Address(c_rarg1, BasicObjectLock::obj_offset()));
4264     __ cmp(r0, rscratch1);
4265     // if same object then stop searching
4266     __ br(Assembler::EQ, found);
4267     // otherwise advance to next entry
4268     __ add(c_rarg1, c_rarg1, entry_size);
4269     __ bind(entry);
4270     // check if bottom reached
4271     __ cmp(c_rarg1, c_rarg2);
4272     // if not at bottom then check this entry
4273     __ br(Assembler::NE, loop);
4274   }
4275 
4276   // error handling. Unlocking was not block-structured
4277   __ call_VM(noreg, CAST_FROM_FN_PTR(address,
4278                    InterpreterRuntime::throw_illegal_monitor_state_exception));
4279   __ should_not_reach_here();
4280 
4281   // call run-time routine
4282   __ bind(found);
4283   __ push_ptr(r0); // make sure object is on stack (contract with oopMaps)
4284   __ unlock_object(c_rarg1);
4285   __ pop_ptr(r0); // discard object
4286 }
4287 
4288 
4289 // Wide instructions
4290 void TemplateTable::wide()
4291 {
4292   __ load_unsigned_byte(r19, at_bcp(1));
4293   __ mov(rscratch1, (address)Interpreter::_wentry_point);
4294   __ ldr(rscratch1, Address(rscratch1, r19, Address::uxtw(3)));
4295   __ br(rscratch1);
4296 }
4297 
4298 
4299 // Multi arrays
4300 void TemplateTable::multianewarray() {
4301   transition(vtos, atos);
4302   __ load_unsigned_byte(r0, at_bcp(3)); // get number of dimensions
4303   // last dim is on top of stack; we want address of first one:
4304   // first_addr = last_addr + (ndims - 1) * wordSize
4305   __ lea(c_rarg1, Address(esp, r0, Address::uxtw(3)));
4306   __ sub(c_rarg1, c_rarg1, wordSize);
4307   call_VM(r0,
4308           CAST_FROM_FN_PTR(address, InterpreterRuntime::multianewarray),
4309           c_rarg1);
4310   __ load_unsigned_byte(r1, at_bcp(3));
4311   __ lea(esp, Address(esp, r1, Address::uxtw(3)));
4312 }
--- EOF ---