1 /*
   2  * Copyright (c) 1997, 2020, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/macroAssembler.inline.hpp"
  27 #include "gc/shared/barrierSetAssembler.hpp"
  28 #include "interpreter/interpreter.hpp"
  29 #include "interpreter/interpreterRuntime.hpp"
  30 #include "interpreter/interp_masm.hpp"
  31 #include "interpreter/templateTable.hpp"
  32 #include "memory/universe.hpp"
  33 #include "oops/methodData.hpp"
  34 #include "oops/objArrayKlass.hpp"
  35 #include "oops/oop.inline.hpp"
  36 #include "prims/methodHandles.hpp"
  37 #include "runtime/frame.inline.hpp"
  38 #include "runtime/safepointMechanism.hpp"
  39 #include "runtime/sharedRuntime.hpp"
  40 #include "runtime/stubRoutines.hpp"
  41 #include "runtime/synchronizer.hpp"
  42 #include "utilities/macros.hpp"
  43 
  44 #define __ _masm->
  45 
  46 // Misc helpers
  47 
  48 // Do an oop store like *(base + index + offset) = val
  49 // index can be noreg,
  50 static void do_oop_store(InterpreterMacroAssembler* _masm,
  51                          Register base,
  52                          Register index,
  53                          int offset,
  54                          Register val,
  55                          Register tmp,
  56                          DecoratorSet decorators = 0) {
  57   assert(tmp != val && tmp != base && tmp != index, "register collision");
  58   assert(index == noreg || offset == 0, "only one offset");
  59   if (index == noreg) {
  60     __ store_heap_oop(val, base, offset, tmp, decorators);
  61   } else {
  62     __ store_heap_oop(val, base, index, tmp, decorators);
  63   }
  64 }
  65 
  66 // Do an oop load like val = *(base + index + offset)
  67 // index can be noreg.
  68 static void do_oop_load(InterpreterMacroAssembler* _masm,
  69                         Register base,
  70                         Register index,
  71                         int offset,
  72                         Register dst,
  73                         Register tmp,
  74                         DecoratorSet decorators = 0) {
  75   assert(tmp != dst && tmp != base && tmp != index, "register collision");
  76   assert(index == noreg || offset == 0, "only one offset");
  77   if (index == noreg) {
  78     __ load_heap_oop(base, offset, dst, tmp, decorators);
  79   } else {
  80     __ load_heap_oop(base, index, dst, tmp, decorators);
  81   }
  82 }
  83 
  84 
  85 //----------------------------------------------------------------------------------------------------
  86 // Platform-dependent initialization
  87 
  88 void TemplateTable::pd_initialize() {
  89   // (none)
  90 }
  91 
  92 
  93 //----------------------------------------------------------------------------------------------------
  94 // Condition conversion
  95 Assembler::Condition ccNot(TemplateTable::Condition cc) {
  96   switch (cc) {
  97     case TemplateTable::equal        : return Assembler::notEqual;
  98     case TemplateTable::not_equal    : return Assembler::equal;
  99     case TemplateTable::less         : return Assembler::greaterEqual;
 100     case TemplateTable::less_equal   : return Assembler::greater;
 101     case TemplateTable::greater      : return Assembler::lessEqual;
 102     case TemplateTable::greater_equal: return Assembler::less;
 103   }
 104   ShouldNotReachHere();
 105   return Assembler::zero;
 106 }
 107 
 108 //----------------------------------------------------------------------------------------------------
 109 // Miscelaneous helper routines
 110 
 111 
 112 Address TemplateTable::at_bcp(int offset) {
 113   assert(_desc->uses_bcp(), "inconsistent uses_bcp information");
 114   return Address(Lbcp, offset);
 115 }
 116 
 117 
 118 void TemplateTable::patch_bytecode(Bytecodes::Code bc, Register bc_reg,
 119                                    Register temp_reg, bool load_bc_into_bc_reg/*=true*/,
 120                                    int byte_no) {
 121   // With sharing on, may need to test Method* flag.
 122   if (!RewriteBytecodes)  return;
 123   Label L_patch_done;
 124 
 125   switch (bc) {
 126   case Bytecodes::_fast_aputfield:
 127   case Bytecodes::_fast_bputfield:
 128   case Bytecodes::_fast_zputfield:
 129   case Bytecodes::_fast_cputfield:
 130   case Bytecodes::_fast_dputfield:
 131   case Bytecodes::_fast_fputfield:
 132   case Bytecodes::_fast_iputfield:
 133   case Bytecodes::_fast_lputfield:
 134   case Bytecodes::_fast_sputfield:
 135     {
 136       // We skip bytecode quickening for putfield instructions when
 137       // the put_code written to the constant pool cache is zero.
 138       // This is required so that every execution of this instruction
 139       // calls out to InterpreterRuntime::resolve_get_put to do
 140       // additional, required work.
 141       assert(byte_no == f1_byte || byte_no == f2_byte, "byte_no out of range");
 142       assert(load_bc_into_bc_reg, "we use bc_reg as temp");
 143       __ get_cache_and_index_and_bytecode_at_bcp(bc_reg, temp_reg, temp_reg, byte_no, 1);
 144       __ set(bc, bc_reg);
 145       __ cmp_and_br_short(temp_reg, 0, Assembler::equal, Assembler::pn, L_patch_done);  // don't patch
 146     }
 147     break;
 148   default:
 149     assert(byte_no == -1, "sanity");
 150     if (load_bc_into_bc_reg) {
 151       __ set(bc, bc_reg);
 152     }
 153   }
 154 
 155   if (JvmtiExport::can_post_breakpoint()) {
 156     Label L_fast_patch;
 157     __ ldub(at_bcp(0), temp_reg);
 158     __ cmp_and_br_short(temp_reg, Bytecodes::_breakpoint, Assembler::notEqual, Assembler::pt, L_fast_patch);
 159     // perform the quickening, slowly, in the bowels of the breakpoint table
 160     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::set_original_bytecode_at), Lmethod, Lbcp, bc_reg);
 161     __ ba_short(L_patch_done);
 162     __ bind(L_fast_patch);
 163   }
 164 
 165 #ifdef ASSERT
 166   Bytecodes::Code orig_bytecode =  Bytecodes::java_code(bc);
 167   Label L_okay;
 168   __ ldub(at_bcp(0), temp_reg);
 169   __ cmp(temp_reg, orig_bytecode);
 170   __ br(Assembler::equal, false, Assembler::pt, L_okay);
 171   __ delayed()->cmp(temp_reg, bc_reg);
 172   __ br(Assembler::equal, false, Assembler::pt, L_okay);
 173   __ delayed()->nop();
 174   __ stop("patching the wrong bytecode");
 175   __ bind(L_okay);
 176 #endif
 177 
 178   // patch bytecode
 179   __ stb(bc_reg, at_bcp(0));
 180   __ bind(L_patch_done);
 181 }
 182 
 183 //----------------------------------------------------------------------------------------------------
 184 // Individual instructions
 185 
 186 void TemplateTable::nop() {
 187   transition(vtos, vtos);
 188   // nothing to do
 189 }
 190 
 191 void TemplateTable::shouldnotreachhere() {
 192   transition(vtos, vtos);
 193   __ stop("shouldnotreachhere bytecode");
 194 }
 195 
 196 void TemplateTable::aconst_null() {
 197   transition(vtos, atos);
 198   __ clr(Otos_i);
 199 }
 200 
 201 
 202 void TemplateTable::iconst(int value) {
 203   transition(vtos, itos);
 204   __ set(value, Otos_i);
 205 }
 206 
 207 
 208 void TemplateTable::lconst(int value) {
 209   transition(vtos, ltos);
 210   assert(value >= 0, "check this code");
 211   __ set(value, Otos_l);
 212 }
 213 
 214 
 215 void TemplateTable::fconst(int value) {
 216   transition(vtos, ftos);
 217   static float zero = 0.0, one = 1.0, two = 2.0;
 218   float* p;
 219   switch( value ) {
 220    default: ShouldNotReachHere();
 221    case 0:  p = &zero;  break;
 222    case 1:  p = &one;   break;
 223    case 2:  p = &two;   break;
 224   }
 225   AddressLiteral a(p);
 226   __ sethi(a, G3_scratch);
 227   __ ldf(FloatRegisterImpl::S, G3_scratch, a.low10(), Ftos_f);
 228 }
 229 
 230 
 231 void TemplateTable::dconst(int value) {
 232   transition(vtos, dtos);
 233   static double zero = 0.0, one = 1.0;
 234   double* p;
 235   switch( value ) {
 236    default: ShouldNotReachHere();
 237    case 0:  p = &zero;  break;
 238    case 1:  p = &one;   break;
 239   }
 240   AddressLiteral a(p);
 241   __ sethi(a, G3_scratch);
 242   __ ldf(FloatRegisterImpl::D, G3_scratch, a.low10(), Ftos_d);
 243 }
 244 
 245 
 246 // %%%%% Should factore most snippet templates across platforms
 247 
 248 void TemplateTable::bipush() {
 249   transition(vtos, itos);
 250   __ ldsb( at_bcp(1), Otos_i );
 251 }
 252 
 253 void TemplateTable::sipush() {
 254   transition(vtos, itos);
 255   __ get_2_byte_integer_at_bcp(1, G3_scratch, Otos_i, InterpreterMacroAssembler::Signed);
 256 }
 257 
 258 void TemplateTable::ldc(bool wide) {
 259   transition(vtos, vtos);
 260   Label call_ldc, notInt, isString, notString, notClass, notFloat, exit;
 261 
 262   if (wide) {
 263     __ get_2_byte_integer_at_bcp(1, G3_scratch, O1, InterpreterMacroAssembler::Unsigned);
 264   } else {
 265     __ ldub(Lbcp, 1, O1);
 266   }
 267   __ get_cpool_and_tags(O0, O2);
 268 
 269   const int base_offset = ConstantPool::header_size() * wordSize;
 270   const int tags_offset = Array<u1>::base_offset_in_bytes();
 271 
 272   // get type from tags
 273   __ add(O2, tags_offset, O2);
 274   __ ldub(O2, O1, O2);
 275 
 276   // unresolved class? If so, must resolve
 277   __ cmp_and_brx_short(O2, JVM_CONSTANT_UnresolvedClass, Assembler::equal, Assembler::pt, call_ldc);
 278 
 279   // unresolved class in error state
 280   __ cmp_and_brx_short(O2, JVM_CONSTANT_UnresolvedClassInError, Assembler::equal, Assembler::pn, call_ldc);
 281 
 282   __ cmp(O2, JVM_CONSTANT_Class);      // need to call vm to get java mirror of the class
 283   __ brx(Assembler::notEqual, true, Assembler::pt, notClass);
 284   __ delayed()->add(O0, base_offset, O0);
 285 
 286   __ bind(call_ldc);
 287   __ set(wide, O1);
 288   call_VM(Otos_i, CAST_FROM_FN_PTR(address, InterpreterRuntime::ldc), O1);
 289   __ push(atos);
 290   __ ba(exit);
 291   __ delayed()->nop();
 292 
 293   __ bind(notClass);
 294  // __ add(O0, base_offset, O0);
 295   __ sll(O1, LogBytesPerWord, O1);
 296   __ cmp(O2, JVM_CONSTANT_Integer);
 297   __ brx(Assembler::notEqual, true, Assembler::pt, notInt);
 298   __ delayed()->cmp(O2, JVM_CONSTANT_String);
 299   __ ld(O0, O1, Otos_i);
 300   __ push(itos);
 301   __ ba(exit);
 302   __ delayed()->nop();
 303 
 304   __ bind(notInt);
 305  // __ cmp(O2, JVM_CONSTANT_String);
 306   __ brx(Assembler::notEqual, true, Assembler::pt, notString);
 307   __ delayed()->cmp(O2, JVM_CONSTANT_Float);
 308   __ bind(isString);
 309   __ stop("string should be rewritten to fast_aldc");
 310   __ ba(exit);
 311   __ delayed()->nop();
 312 
 313   __ bind(notString);
 314  //__ cmp(O2, JVM_CONSTANT_Float);
 315   __ brx(Assembler::notEqual, true, Assembler::pt, notFloat);
 316   __ delayed()->nop();
 317   __ ldf(FloatRegisterImpl::S, O0, O1, Ftos_f);
 318   __ push(ftos);
 319   __ ba(exit);
 320   __ delayed()->nop();
 321 
 322   // assume the tag is for condy; if not, the VM runtime will tell us
 323   __ bind(notFloat);
 324   condy_helper(exit);
 325 
 326   __ bind(exit);
 327 }
 328 
 329 // Fast path for caching oop constants.
 330 // %%% We should use this to handle Class and String constants also.
 331 // %%% It will simplify the ldc/primitive path considerably.
 332 void TemplateTable::fast_aldc(bool wide) {
 333   transition(vtos, atos);
 334 
 335   int index_size = wide ? sizeof(u2) : sizeof(u1);
 336   Label resolved;
 337 
 338   // We are resolved if the resolved reference cache entry contains a
 339   // non-null object (CallSite, etc.)
 340   assert_different_registers(Otos_i, G3_scratch);
 341   __ get_cache_index_at_bcp(Otos_i, G3_scratch, 1, index_size);  // load index => G3_scratch
 342   __ load_resolved_reference_at_index(Otos_i, G3_scratch, Lscratch);
 343   __ tst(Otos_i);
 344   __ br(Assembler::notEqual, false, Assembler::pt, resolved);
 345   __ delayed()->set((int)bytecode(), O1);
 346 
 347   address entry = CAST_FROM_FN_PTR(address, InterpreterRuntime::resolve_ldc);
 348 
 349   // first time invocation - must resolve first
 350   __ call_VM(Otos_i, entry, O1);
 351   __ bind(resolved);
 352 
 353   { // Check for the null sentinel.
 354     // If we just called the VM, it already did the mapping for us,
 355     // but it's harmless to retry.
 356     Label notNull;
 357     __ set(ExternalAddress((address)Universe::the_null_sentinel_addr()), G3_scratch);
 358     __ ld_ptr(G3_scratch, 0, G3_scratch);
 359     __ cmp(G3_scratch, Otos_i);
 360     __ br(Assembler::notEqual, true, Assembler::pt, notNull);
 361     __ delayed()->nop();
 362     __ clr(Otos_i);  // NULL object reference
 363     __ bind(notNull);
 364   }
 365 
 366   // Safe to call with 0 result
 367   __ verify_oop(Otos_i);
 368 }
 369 
 370 void TemplateTable::ldc2_w() {
 371   transition(vtos, vtos);
 372   Label notDouble, notLong, exit;
 373 
 374   __ get_2_byte_integer_at_bcp(1, G3_scratch, O1, InterpreterMacroAssembler::Unsigned);
 375   __ get_cpool_and_tags(O0, O2);
 376 
 377   const int base_offset = ConstantPool::header_size() * wordSize;
 378   const int tags_offset = Array<u1>::base_offset_in_bytes();
 379   // get type from tags
 380   __ add(O2, tags_offset, O2);
 381   __ ldub(O2, O1, O2);
 382 
 383   __ sll(O1, LogBytesPerWord, O1);
 384   __ add(O0, O1, G3_scratch);
 385 
 386   __ cmp_and_brx_short(O2, JVM_CONSTANT_Double, Assembler::notEqual, Assembler::pt, notDouble);
 387   // A double can be placed at word-aligned locations in the constant pool.
 388   // Check out Conversions.java for an example.
 389   // Also ConstantPool::header_size() is 20, which makes it very difficult
 390   // to double-align double on the constant pool.  SG, 11/7/97
 391   __ ldf(FloatRegisterImpl::D, G3_scratch, base_offset, Ftos_d);
 392   __ push(dtos);
 393   __ ba_short(exit);
 394 
 395   __ bind(notDouble);
 396   __ cmp_and_brx_short(O2, JVM_CONSTANT_Long, Assembler::notEqual, Assembler::pt, notLong);
 397   __ ldx(G3_scratch, base_offset, Otos_l);
 398   __ push(ltos);
 399   __ ba_short(exit);
 400 
 401   __ bind(notLong);
 402   condy_helper(exit);
 403 
 404   __ bind(exit);
 405 }
 406 
 407 void TemplateTable::condy_helper(Label& exit) {
 408   Register Robj = Otos_i;
 409   Register Roffset = G4_scratch;
 410   Register Rflags = G1_scratch;
 411 
 412   address entry = CAST_FROM_FN_PTR(address, InterpreterRuntime::resolve_ldc);
 413 
 414   __ set((int)bytecode(), O1);
 415   __ call_VM(Robj, entry, O1);
 416 
 417   // Get vm_result_2 has flags = (tos, off) using format CPCE::_flags
 418   __ get_vm_result_2(G3_scratch);
 419 
 420   // Get offset
 421   __ set((int)ConstantPoolCacheEntry::field_index_mask, Roffset);
 422   __ and3(G3_scratch, Roffset, Roffset);
 423 
 424   // compute type
 425   __ srl(G3_scratch, ConstantPoolCacheEntry::tos_state_shift, Rflags);
 426   // Make sure we don't need to mask Rflags after the above shift
 427   ConstantPoolCacheEntry::verify_tos_state_shift();
 428 
 429   switch (bytecode()) {
 430   case Bytecodes::_ldc:
 431   case Bytecodes::_ldc_w:
 432     {
 433       // tos in (itos, ftos, stos, btos, ctos, ztos)
 434       Label notInt, notFloat, notShort, notByte, notChar, notBool;
 435       __ cmp(Rflags, itos);
 436       __ br(Assembler::notEqual, false, Assembler::pt, notInt);
 437       __ delayed()->cmp(Rflags, ftos);
 438       // itos
 439       __ ld(Robj, Roffset, Otos_i);
 440       __ push(itos);
 441       __ ba_short(exit);
 442 
 443       __ bind(notInt);
 444       __ br(Assembler::notEqual, false, Assembler::pt, notFloat);
 445       __ delayed()->cmp(Rflags, stos);
 446       // ftos
 447       __ ldf(FloatRegisterImpl::S, Robj, Roffset, Ftos_f);
 448       __ push(ftos);
 449       __ ba_short(exit);
 450 
 451       __ bind(notFloat);
 452       __ br(Assembler::notEqual, false, Assembler::pt, notShort);
 453       __ delayed()->cmp(Rflags, btos);
 454       // stos
 455       __ ldsh(Robj, Roffset, Otos_i);
 456       __ push(itos);
 457       __ ba_short(exit);
 458 
 459       __ bind(notShort);
 460       __ br(Assembler::notEqual, false, Assembler::pt, notByte);
 461       __ delayed()->cmp(Rflags, ctos);
 462       // btos
 463       __ ldsb(Robj, Roffset, Otos_i);
 464       __ push(itos);
 465       __ ba_short(exit);
 466 
 467       __ bind(notByte);
 468       __ br(Assembler::notEqual, false, Assembler::pt, notChar);
 469       __ delayed()->cmp(Rflags, ztos);
 470       // ctos
 471       __ lduh(Robj, Roffset, Otos_i);
 472       __ push(itos);
 473       __ ba_short(exit);
 474 
 475       __ bind(notChar);
 476       __ br(Assembler::notEqual, false, Assembler::pt, notBool);
 477       __ delayed()->nop();
 478       // ztos
 479       __ ldsb(Robj, Roffset, Otos_i);
 480       __ push(itos);
 481       __ ba_short(exit);
 482 
 483       __ bind(notBool);
 484       break;
 485     }
 486 
 487   case Bytecodes::_ldc2_w:
 488     {
 489       Label notLong, notDouble;
 490       __ cmp(Rflags, ltos);
 491       __ br(Assembler::notEqual, false, Assembler::pt, notLong);
 492       __ delayed()->cmp(Rflags, dtos);
 493       // ltos
 494       // load must be atomic
 495       __ ld_long(Robj, Roffset, Otos_l);
 496       __ push(ltos);
 497       __ ba_short(exit);
 498 
 499       __ bind(notLong);
 500       __ br(Assembler::notEqual, false, Assembler::pt, notDouble);
 501       __ delayed()->nop();
 502       // dtos
 503       __ ldf(FloatRegisterImpl::D, Robj, Roffset, Ftos_d);
 504       __ push(dtos);
 505       __ ba_short(exit);
 506 
 507       __ bind(notDouble);
 508       break;
 509     }
 510 
 511   default:
 512     ShouldNotReachHere();
 513   }
 514 
 515   __ stop("bad ldc/condy");
 516 
 517   __ bind(exit);
 518 }
 519 
 520 void TemplateTable::locals_index(Register reg, int offset) {
 521   __ ldub( at_bcp(offset), reg );
 522 }
 523 
 524 void TemplateTable::locals_index_wide(Register reg) {
 525   // offset is 2, not 1, because Lbcp points to wide prefix code
 526   __ get_2_byte_integer_at_bcp(2, G4_scratch, reg, InterpreterMacroAssembler::Unsigned);
 527 }
 528 
 529 void TemplateTable::iload() {
 530   iload_internal();
 531 }
 532 
 533 void TemplateTable::nofast_iload() {
 534   iload_internal(may_not_rewrite);
 535 }
 536 
 537 void TemplateTable::iload_internal(RewriteControl rc) {
 538   transition(vtos, itos);
 539   // Rewrite iload,iload  pair into fast_iload2
 540   //         iload,caload pair into fast_icaload
 541   if (RewriteFrequentPairs && rc == may_rewrite) {
 542     Label rewrite, done;
 543 
 544     // get next byte
 545     __ ldub(at_bcp(Bytecodes::length_for(Bytecodes::_iload)), G3_scratch);
 546 
 547     // if _iload, wait to rewrite to iload2.  We only want to rewrite the
 548     // last two iloads in a pair.  Comparing against fast_iload means that
 549     // the next bytecode is neither an iload or a caload, and therefore
 550     // an iload pair.
 551     __ cmp_and_br_short(G3_scratch, (int)Bytecodes::_iload, Assembler::equal, Assembler::pn, done);
 552 
 553     __ cmp(G3_scratch, (int)Bytecodes::_fast_iload);
 554     __ br(Assembler::equal, false, Assembler::pn, rewrite);
 555     __ delayed()->set(Bytecodes::_fast_iload2, G4_scratch);
 556 
 557     __ cmp(G3_scratch, (int)Bytecodes::_caload);
 558     __ br(Assembler::equal, false, Assembler::pn, rewrite);
 559     __ delayed()->set(Bytecodes::_fast_icaload, G4_scratch);
 560 
 561     __ set(Bytecodes::_fast_iload, G4_scratch);  // don't check again
 562     // rewrite
 563     // G4_scratch: fast bytecode
 564     __ bind(rewrite);
 565     patch_bytecode(Bytecodes::_iload, G4_scratch, G3_scratch, false);
 566     __ bind(done);
 567   }
 568 
 569   // Get the local value into tos
 570   locals_index(G3_scratch);
 571   __ access_local_int( G3_scratch, Otos_i );
 572 }
 573 
 574 void TemplateTable::fast_iload2() {
 575   transition(vtos, itos);
 576   locals_index(G3_scratch);
 577   __ access_local_int( G3_scratch, Otos_i );
 578   __ push_i();
 579   locals_index(G3_scratch, 3);  // get next bytecode's local index.
 580   __ access_local_int( G3_scratch, Otos_i );
 581 }
 582 
 583 void TemplateTable::fast_iload() {
 584   transition(vtos, itos);
 585   locals_index(G3_scratch);
 586   __ access_local_int( G3_scratch, Otos_i );
 587 }
 588 
 589 void TemplateTable::lload() {
 590   transition(vtos, ltos);
 591   locals_index(G3_scratch);
 592   __ access_local_long( G3_scratch, Otos_l );
 593 }
 594 
 595 
 596 void TemplateTable::fload() {
 597   transition(vtos, ftos);
 598   locals_index(G3_scratch);
 599   __ access_local_float( G3_scratch, Ftos_f );
 600 }
 601 
 602 
 603 void TemplateTable::dload() {
 604   transition(vtos, dtos);
 605   locals_index(G3_scratch);
 606   __ access_local_double( G3_scratch, Ftos_d );
 607 }
 608 
 609 
 610 void TemplateTable::aload() {
 611   transition(vtos, atos);
 612   locals_index(G3_scratch);
 613   __ access_local_ptr( G3_scratch, Otos_i);
 614 }
 615 
 616 
 617 void TemplateTable::wide_iload() {
 618   transition(vtos, itos);
 619   locals_index_wide(G3_scratch);
 620   __ access_local_int( G3_scratch, Otos_i );
 621 }
 622 
 623 
 624 void TemplateTable::wide_lload() {
 625   transition(vtos, ltos);
 626   locals_index_wide(G3_scratch);
 627   __ access_local_long( G3_scratch, Otos_l );
 628 }
 629 
 630 
 631 void TemplateTable::wide_fload() {
 632   transition(vtos, ftos);
 633   locals_index_wide(G3_scratch);
 634   __ access_local_float( G3_scratch, Ftos_f );
 635 }
 636 
 637 
 638 void TemplateTable::wide_dload() {
 639   transition(vtos, dtos);
 640   locals_index_wide(G3_scratch);
 641   __ access_local_double( G3_scratch, Ftos_d );
 642 }
 643 
 644 
 645 void TemplateTable::wide_aload() {
 646   transition(vtos, atos);
 647   locals_index_wide(G3_scratch);
 648   __ access_local_ptr( G3_scratch, Otos_i );
 649   __ verify_oop(Otos_i);
 650 }
 651 
 652 
 653 void TemplateTable::iaload() {
 654   transition(itos, itos);
 655   // Otos_i: index
 656   // tos: array
 657   __ index_check(O2, Otos_i, LogBytesPerInt, G3_scratch, O3);
 658   __ ld(O3, arrayOopDesc::base_offset_in_bytes(T_INT), Otos_i);
 659 }
 660 
 661 
 662 void TemplateTable::laload() {
 663   transition(itos, ltos);
 664   // Otos_i: index
 665   // O2: array
 666   __ index_check(O2, Otos_i, LogBytesPerLong, G3_scratch, O3);
 667   __ ld_long(O3, arrayOopDesc::base_offset_in_bytes(T_LONG), Otos_l);
 668 }
 669 
 670 
 671 void TemplateTable::faload() {
 672   transition(itos, ftos);
 673   // Otos_i: index
 674   // O2: array
 675   __ index_check(O2, Otos_i, LogBytesPerInt, G3_scratch, O3);
 676   __ ldf(FloatRegisterImpl::S, O3, arrayOopDesc::base_offset_in_bytes(T_FLOAT), Ftos_f);
 677 }
 678 
 679 
 680 void TemplateTable::daload() {
 681   transition(itos, dtos);
 682   // Otos_i: index
 683   // O2: array
 684   __ index_check(O2, Otos_i, LogBytesPerLong, G3_scratch, O3);
 685   __ ldf(FloatRegisterImpl::D, O3, arrayOopDesc::base_offset_in_bytes(T_DOUBLE), Ftos_d);
 686 }
 687 
 688 
 689 void TemplateTable::aaload() {
 690   transition(itos, atos);
 691   // Otos_i: index
 692   // tos: array
 693   __ index_check(O2, Otos_i, UseCompressedOops ? 2 : LogBytesPerWord, G3_scratch, O3);
 694   do_oop_load(_masm,
 695               O3,
 696               noreg,
 697               arrayOopDesc::base_offset_in_bytes(T_OBJECT),
 698               Otos_i,
 699               G3_scratch,
 700               IS_ARRAY);
 701   __ verify_oop(Otos_i);
 702 }
 703 
 704 
 705 void TemplateTable::baload() {
 706   transition(itos, itos);
 707   // Otos_i: index
 708   // tos: array
 709   __ index_check(O2, Otos_i, 0, G3_scratch, O3);
 710   __ ldsb(O3, arrayOopDesc::base_offset_in_bytes(T_BYTE), Otos_i);
 711 }
 712 
 713 
 714 void TemplateTable::caload() {
 715   transition(itos, itos);
 716   // Otos_i: index
 717   // tos: array
 718   __ index_check(O2, Otos_i, LogBytesPerShort, G3_scratch, O3);
 719   __ lduh(O3, arrayOopDesc::base_offset_in_bytes(T_CHAR), Otos_i);
 720 }
 721 
 722 void TemplateTable::fast_icaload() {
 723   transition(vtos, itos);
 724   // Otos_i: index
 725   // tos: array
 726   locals_index(G3_scratch);
 727   __ access_local_int( G3_scratch, Otos_i );
 728   __ index_check(O2, Otos_i, LogBytesPerShort, G3_scratch, O3);
 729   __ lduh(O3, arrayOopDesc::base_offset_in_bytes(T_CHAR), Otos_i);
 730 }
 731 
 732 
 733 void TemplateTable::saload() {
 734   transition(itos, itos);
 735   // Otos_i: index
 736   // tos: array
 737   __ index_check(O2, Otos_i, LogBytesPerShort, G3_scratch, O3);
 738   __ ldsh(O3, arrayOopDesc::base_offset_in_bytes(T_SHORT), Otos_i);
 739 }
 740 
 741 
 742 void TemplateTable::iload(int n) {
 743   transition(vtos, itos);
 744   __ ld( Llocals, Interpreter::local_offset_in_bytes(n), Otos_i );
 745 }
 746 
 747 
 748 void TemplateTable::lload(int n) {
 749   transition(vtos, ltos);
 750   assert(n+1 < Argument::n_register_parameters, "would need more code");
 751   __ load_unaligned_long(Llocals, Interpreter::local_offset_in_bytes(n+1), Otos_l);
 752 }
 753 
 754 
 755 void TemplateTable::fload(int n) {
 756   transition(vtos, ftos);
 757   assert(n < Argument::n_register_parameters, "would need more code");
 758   __ ldf( FloatRegisterImpl::S, Llocals, Interpreter::local_offset_in_bytes(n),     Ftos_f );
 759 }
 760 
 761 
 762 void TemplateTable::dload(int n) {
 763   transition(vtos, dtos);
 764   FloatRegister dst = Ftos_d;
 765   __ load_unaligned_double(Llocals, Interpreter::local_offset_in_bytes(n+1), dst);
 766 }
 767 
 768 
 769 void TemplateTable::aload(int n) {
 770   transition(vtos, atos);
 771   __ ld_ptr( Llocals, Interpreter::local_offset_in_bytes(n), Otos_i );
 772 }
 773 
 774 void TemplateTable::aload_0() {
 775   aload_0_internal();
 776 }
 777 
 778 void TemplateTable::nofast_aload_0() {
 779   aload_0_internal(may_not_rewrite);
 780 }
 781 
 782 void TemplateTable::aload_0_internal(RewriteControl rc) {
 783   transition(vtos, atos);
 784 
 785   // According to bytecode histograms, the pairs:
 786   //
 787   // _aload_0, _fast_igetfield (itos)
 788   // _aload_0, _fast_agetfield (atos)
 789   // _aload_0, _fast_fgetfield (ftos)
 790   //
 791   // occur frequently. If RewriteFrequentPairs is set, the (slow) _aload_0
 792   // bytecode checks the next bytecode and then rewrites the current
 793   // bytecode into a pair bytecode; otherwise it rewrites the current
 794   // bytecode into _fast_aload_0 that doesn't do the pair check anymore.
 795   //
 796   if (RewriteFrequentPairs && rc == may_rewrite) {
 797     Label rewrite, done;
 798 
 799     // get next byte
 800     __ ldub(at_bcp(Bytecodes::length_for(Bytecodes::_aload_0)), G3_scratch);
 801 
 802     // if _getfield then wait with rewrite
 803     __ cmp_and_br_short(G3_scratch, (int)Bytecodes::_getfield, Assembler::equal, Assembler::pn, done);
 804 
 805     // if _igetfield then rewrite to _fast_iaccess_0
 806     assert(Bytecodes::java_code(Bytecodes::_fast_iaccess_0) == Bytecodes::_aload_0, "adjust fast bytecode def");
 807     __ cmp(G3_scratch, (int)Bytecodes::_fast_igetfield);
 808     __ br(Assembler::equal, false, Assembler::pn, rewrite);
 809     __ delayed()->set(Bytecodes::_fast_iaccess_0, G4_scratch);
 810 
 811     // if _agetfield then rewrite to _fast_aaccess_0
 812     assert(Bytecodes::java_code(Bytecodes::_fast_aaccess_0) == Bytecodes::_aload_0, "adjust fast bytecode def");
 813     __ cmp(G3_scratch, (int)Bytecodes::_fast_agetfield);
 814     __ br(Assembler::equal, false, Assembler::pn, rewrite);
 815     __ delayed()->set(Bytecodes::_fast_aaccess_0, G4_scratch);
 816 
 817     // if _fgetfield then rewrite to _fast_faccess_0
 818     assert(Bytecodes::java_code(Bytecodes::_fast_faccess_0) == Bytecodes::_aload_0, "adjust fast bytecode def");
 819     __ cmp(G3_scratch, (int)Bytecodes::_fast_fgetfield);
 820     __ br(Assembler::equal, false, Assembler::pn, rewrite);
 821     __ delayed()->set(Bytecodes::_fast_faccess_0, G4_scratch);
 822 
 823     // else rewrite to _fast_aload0
 824     assert(Bytecodes::java_code(Bytecodes::_fast_aload_0) == Bytecodes::_aload_0, "adjust fast bytecode def");
 825     __ set(Bytecodes::_fast_aload_0, G4_scratch);
 826 
 827     // rewrite
 828     // G4_scratch: fast bytecode
 829     __ bind(rewrite);
 830     patch_bytecode(Bytecodes::_aload_0, G4_scratch, G3_scratch, false);
 831     __ bind(done);
 832   }
 833 
 834   // Do actual aload_0 (must do this after patch_bytecode which might call VM and GC might change oop).
 835   aload(0);
 836 }
 837 
 838 void TemplateTable::istore() {
 839   transition(itos, vtos);
 840   locals_index(G3_scratch);
 841   __ store_local_int( G3_scratch, Otos_i );
 842 }
 843 
 844 
 845 void TemplateTable::lstore() {
 846   transition(ltos, vtos);
 847   locals_index(G3_scratch);
 848   __ store_local_long( G3_scratch, Otos_l );
 849 }
 850 
 851 
 852 void TemplateTable::fstore() {
 853   transition(ftos, vtos);
 854   locals_index(G3_scratch);
 855   __ store_local_float( G3_scratch, Ftos_f );
 856 }
 857 
 858 
 859 void TemplateTable::dstore() {
 860   transition(dtos, vtos);
 861   locals_index(G3_scratch);
 862   __ store_local_double( G3_scratch, Ftos_d );
 863 }
 864 
 865 
 866 void TemplateTable::astore() {
 867   transition(vtos, vtos);
 868   __ load_ptr(0, Otos_i);
 869   __ inc(Lesp, Interpreter::stackElementSize);
 870   __ verify_oop_or_return_address(Otos_i, G3_scratch);
 871   locals_index(G3_scratch);
 872   __ store_local_ptr(G3_scratch, Otos_i);
 873 }
 874 
 875 
 876 void TemplateTable::wide_istore() {
 877   transition(vtos, vtos);
 878   __ pop_i();
 879   locals_index_wide(G3_scratch);
 880   __ store_local_int( G3_scratch, Otos_i );
 881 }
 882 
 883 
 884 void TemplateTable::wide_lstore() {
 885   transition(vtos, vtos);
 886   __ pop_l();
 887   locals_index_wide(G3_scratch);
 888   __ store_local_long( G3_scratch, Otos_l );
 889 }
 890 
 891 
 892 void TemplateTable::wide_fstore() {
 893   transition(vtos, vtos);
 894   __ pop_f();
 895   locals_index_wide(G3_scratch);
 896   __ store_local_float( G3_scratch, Ftos_f );
 897 }
 898 
 899 
 900 void TemplateTable::wide_dstore() {
 901   transition(vtos, vtos);
 902   __ pop_d();
 903   locals_index_wide(G3_scratch);
 904   __ store_local_double( G3_scratch, Ftos_d );
 905 }
 906 
 907 
 908 void TemplateTable::wide_astore() {
 909   transition(vtos, vtos);
 910   __ load_ptr(0, Otos_i);
 911   __ inc(Lesp, Interpreter::stackElementSize);
 912   __ verify_oop_or_return_address(Otos_i, G3_scratch);
 913   locals_index_wide(G3_scratch);
 914   __ store_local_ptr(G3_scratch, Otos_i);
 915 }
 916 
 917 
 918 void TemplateTable::iastore() {
 919   transition(itos, vtos);
 920   __ pop_i(O2); // index
 921   // Otos_i: val
 922   // O3: array
 923   __ index_check(O3, O2, LogBytesPerInt, G3_scratch, O2);
 924   __ st(Otos_i, O2, arrayOopDesc::base_offset_in_bytes(T_INT));
 925 }
 926 
 927 
 928 void TemplateTable::lastore() {
 929   transition(ltos, vtos);
 930   __ pop_i(O2); // index
 931   // Otos_l: val
 932   // O3: array
 933   __ index_check(O3, O2, LogBytesPerLong, G3_scratch, O2);
 934   __ st_long(Otos_l, O2, arrayOopDesc::base_offset_in_bytes(T_LONG));
 935 }
 936 
 937 
 938 void TemplateTable::fastore() {
 939   transition(ftos, vtos);
 940   __ pop_i(O2); // index
 941   // Ftos_f: val
 942   // O3: array
 943   __ index_check(O3, O2, LogBytesPerInt, G3_scratch, O2);
 944   __ stf(FloatRegisterImpl::S, Ftos_f, O2, arrayOopDesc::base_offset_in_bytes(T_FLOAT));
 945 }
 946 
 947 
 948 void TemplateTable::dastore() {
 949   transition(dtos, vtos);
 950   __ pop_i(O2); // index
 951   // Fos_d: val
 952   // O3: array
 953   __ index_check(O3, O2, LogBytesPerLong, G3_scratch, O2);
 954   __ stf(FloatRegisterImpl::D, Ftos_d, O2, arrayOopDesc::base_offset_in_bytes(T_DOUBLE));
 955 }
 956 
 957 
 958 void TemplateTable::aastore() {
 959   Label store_ok, is_null, done;
 960   transition(vtos, vtos);
 961   __ ld_ptr(Lesp, Interpreter::expr_offset_in_bytes(0), Otos_i);
 962   __ ld(Lesp, Interpreter::expr_offset_in_bytes(1), O2);         // get index
 963   __ ld_ptr(Lesp, Interpreter::expr_offset_in_bytes(2), O3);     // get array
 964   // Otos_i: val
 965   // O2: index
 966   // O3: array
 967   __ verify_oop(Otos_i);
 968   __ index_check_without_pop(O3, O2, UseCompressedOops ? 2 : LogBytesPerWord, G3_scratch, O1);
 969 
 970   // do array store check - check for NULL value first
 971   __ br_null_short( Otos_i, Assembler::pn, is_null );
 972 
 973   __ load_klass(O3, O4); // get array klass
 974   __ load_klass(Otos_i, O5); // get value klass
 975 
 976   // do fast instanceof cache test
 977 
 978   __ ld_ptr(O4,     in_bytes(ObjArrayKlass::element_klass_offset()),  O4);
 979 
 980   assert(Otos_i == O0, "just checking");
 981 
 982   // Otos_i:    value
 983   // O1:        addr - offset
 984   // O2:        index
 985   // O3:        array
 986   // O4:        array element klass
 987   // O5:        value klass
 988 
 989   // Address element(O1, 0, arrayOopDesc::base_offset_in_bytes(T_OBJECT));
 990 
 991   // Generate a fast subtype check.  Branch to store_ok if no
 992   // failure.  Throw if failure.
 993   __ gen_subtype_check( O5, O4, G3_scratch, G4_scratch, G1_scratch, store_ok );
 994 
 995   // Not a subtype; so must throw exception
 996   __ throw_if_not_x( Assembler::never, Interpreter::_throw_ArrayStoreException_entry, G3_scratch );
 997 
 998   // Store is OK.
 999   __ bind(store_ok);
1000   do_oop_store(_masm, O1, noreg, arrayOopDesc::base_offset_in_bytes(T_OBJECT), Otos_i, G3_scratch, IS_ARRAY);
1001 
1002   __ ba(done);
1003   __ delayed()->inc(Lesp, 3* Interpreter::stackElementSize); // adj sp (pops array, index and value)
1004 
1005   __ bind(is_null);
1006   do_oop_store(_masm, O1, noreg, arrayOopDesc::base_offset_in_bytes(T_OBJECT), G0, G4_scratch, IS_ARRAY);
1007 
1008   __ profile_null_seen(G3_scratch);
1009   __ inc(Lesp, 3* Interpreter::stackElementSize);     // adj sp (pops array, index and value)
1010   __ bind(done);
1011 }
1012 
1013 
1014 void TemplateTable::bastore() {
1015   transition(itos, vtos);
1016   __ pop_i(O2); // index
1017   // Otos_i: val
1018   // O2: index
1019   // O3: array
1020   __ index_check(O3, O2, 0, G3_scratch, O2);
1021   // Need to check whether array is boolean or byte
1022   // since both types share the bastore bytecode.
1023   __ load_klass(O3, G4_scratch);
1024   __ ld(G4_scratch, in_bytes(Klass::layout_helper_offset()), G4_scratch);
1025   __ set(Klass::layout_helper_boolean_diffbit(), G3_scratch);
1026   __ andcc(G3_scratch, G4_scratch, G0);
1027   Label L_skip;
1028   __ br(Assembler::zero, false, Assembler::pn, L_skip);
1029   __ delayed()->nop();
1030   __ and3(Otos_i, 1, Otos_i);  // if it is a T_BOOLEAN array, mask the stored value to 0/1
1031   __ bind(L_skip);
1032   __ stb(Otos_i, O2, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1033 }
1034 
1035 
1036 void TemplateTable::castore() {
1037   transition(itos, vtos);
1038   __ pop_i(O2); // index
1039   // Otos_i: val
1040   // O3: array
1041   __ index_check(O3, O2, LogBytesPerShort, G3_scratch, O2);
1042   __ sth(Otos_i, O2, arrayOopDesc::base_offset_in_bytes(T_CHAR));
1043 }
1044 
1045 
1046 void TemplateTable::sastore() {
1047   // %%%%% Factor across platform
1048   castore();
1049 }
1050 
1051 
1052 void TemplateTable::istore(int n) {
1053   transition(itos, vtos);
1054   __ st(Otos_i, Llocals, Interpreter::local_offset_in_bytes(n));
1055 }
1056 
1057 
1058 void TemplateTable::lstore(int n) {
1059   transition(ltos, vtos);
1060   assert(n+1 < Argument::n_register_parameters, "only handle register cases");
1061   __ store_unaligned_long(Otos_l, Llocals, Interpreter::local_offset_in_bytes(n+1));
1062 
1063 }
1064 
1065 
1066 void TemplateTable::fstore(int n) {
1067   transition(ftos, vtos);
1068   assert(n < Argument::n_register_parameters, "only handle register cases");
1069   __ stf(FloatRegisterImpl::S, Ftos_f, Llocals, Interpreter::local_offset_in_bytes(n));
1070 }
1071 
1072 
1073 void TemplateTable::dstore(int n) {
1074   transition(dtos, vtos);
1075   FloatRegister src = Ftos_d;
1076   __ store_unaligned_double(src, Llocals, Interpreter::local_offset_in_bytes(n+1));
1077 }
1078 
1079 
1080 void TemplateTable::astore(int n) {
1081   transition(vtos, vtos);
1082   __ load_ptr(0, Otos_i);
1083   __ inc(Lesp, Interpreter::stackElementSize);
1084   __ verify_oop_or_return_address(Otos_i, G3_scratch);
1085   __ store_local_ptr(n, Otos_i);
1086 }
1087 
1088 
1089 void TemplateTable::pop() {
1090   transition(vtos, vtos);
1091   __ inc(Lesp, Interpreter::stackElementSize);
1092 }
1093 
1094 
1095 void TemplateTable::pop2() {
1096   transition(vtos, vtos);
1097   __ inc(Lesp, 2 * Interpreter::stackElementSize);
1098 }
1099 
1100 
1101 void TemplateTable::dup() {
1102   transition(vtos, vtos);
1103   // stack: ..., a
1104   // load a and tag
1105   __ load_ptr(0, Otos_i);
1106   __ push_ptr(Otos_i);
1107   // stack: ..., a, a
1108 }
1109 
1110 
1111 void TemplateTable::dup_x1() {
1112   transition(vtos, vtos);
1113   // stack: ..., a, b
1114   __ load_ptr( 1, G3_scratch);  // get a
1115   __ load_ptr( 0, Otos_l1);     // get b
1116   __ store_ptr(1, Otos_l1);     // put b
1117   __ store_ptr(0, G3_scratch);  // put a - like swap
1118   __ push_ptr(Otos_l1);         // push b
1119   // stack: ..., b, a, b
1120 }
1121 
1122 
1123 void TemplateTable::dup_x2() {
1124   transition(vtos, vtos);
1125   // stack: ..., a, b, c
1126   // get c and push on stack, reuse registers
1127   __ load_ptr( 0, G3_scratch);  // get c
1128   __ push_ptr(G3_scratch);      // push c with tag
1129   // stack: ..., a, b, c, c  (c in reg)  (Lesp - 4)
1130   // (stack offsets n+1 now)
1131   __ load_ptr( 3, Otos_l1);     // get a
1132   __ store_ptr(3, G3_scratch);  // put c at 3
1133   // stack: ..., c, b, c, c  (a in reg)
1134   __ load_ptr( 2, G3_scratch);  // get b
1135   __ store_ptr(2, Otos_l1);     // put a at 2
1136   // stack: ..., c, a, c, c  (b in reg)
1137   __ store_ptr(1, G3_scratch);  // put b at 1
1138   // stack: ..., c, a, b, c
1139 }
1140 
1141 
1142 void TemplateTable::dup2() {
1143   transition(vtos, vtos);
1144   __ load_ptr(1, G3_scratch);  // get a
1145   __ load_ptr(0, Otos_l1);     // get b
1146   __ push_ptr(G3_scratch);     // push a
1147   __ push_ptr(Otos_l1);        // push b
1148   // stack: ..., a, b, a, b
1149 }
1150 
1151 
1152 void TemplateTable::dup2_x1() {
1153   transition(vtos, vtos);
1154   // stack: ..., a, b, c
1155   __ load_ptr( 1, Lscratch);    // get b
1156   __ load_ptr( 2, Otos_l1);     // get a
1157   __ store_ptr(2, Lscratch);    // put b at a
1158   // stack: ..., b, b, c
1159   __ load_ptr( 0, G3_scratch);  // get c
1160   __ store_ptr(1, G3_scratch);  // put c at b
1161   // stack: ..., b, c, c
1162   __ store_ptr(0, Otos_l1);     // put a at c
1163   // stack: ..., b, c, a
1164   __ push_ptr(Lscratch);        // push b
1165   __ push_ptr(G3_scratch);      // push c
1166   // stack: ..., b, c, a, b, c
1167 }
1168 
1169 
1170 // The spec says that these types can be a mixture of category 1 (1 word)
1171 // types and/or category 2 types (long and doubles)
1172 void TemplateTable::dup2_x2() {
1173   transition(vtos, vtos);
1174   // stack: ..., a, b, c, d
1175   __ load_ptr( 1, Lscratch);    // get c
1176   __ load_ptr( 3, Otos_l1);     // get a
1177   __ store_ptr(3, Lscratch);    // put c at 3
1178   __ store_ptr(1, Otos_l1);     // put a at 1
1179   // stack: ..., c, b, a, d
1180   __ load_ptr( 2, G3_scratch);  // get b
1181   __ load_ptr( 0, Otos_l1);     // get d
1182   __ store_ptr(0, G3_scratch);  // put b at 0
1183   __ store_ptr(2, Otos_l1);     // put d at 2
1184   // stack: ..., c, d, a, b
1185   __ push_ptr(Lscratch);        // push c
1186   __ push_ptr(Otos_l1);         // push d
1187   // stack: ..., c, d, a, b, c, d
1188 }
1189 
1190 
1191 void TemplateTable::swap() {
1192   transition(vtos, vtos);
1193   // stack: ..., a, b
1194   __ load_ptr( 1, G3_scratch);  // get a
1195   __ load_ptr( 0, Otos_l1);     // get b
1196   __ store_ptr(0, G3_scratch);  // put b
1197   __ store_ptr(1, Otos_l1);     // put a
1198   // stack: ..., b, a
1199 }
1200 
1201 
1202 void TemplateTable::iop2(Operation op) {
1203   transition(itos, itos);
1204   __ pop_i(O1);
1205   switch (op) {
1206    case  add:  __  add(O1, Otos_i, Otos_i);  break;
1207    case  sub:  __  sub(O1, Otos_i, Otos_i);  break;
1208      // %%%%% Mul may not exist: better to call .mul?
1209    case  mul:  __ smul(O1, Otos_i, Otos_i);  break;
1210    case _and:  __ and3(O1, Otos_i, Otos_i);  break;
1211    case  _or:  __  or3(O1, Otos_i, Otos_i);  break;
1212    case _xor:  __ xor3(O1, Otos_i, Otos_i);  break;
1213    case  shl:  __  sll(O1, Otos_i, Otos_i);  break;
1214    case  shr:  __  sra(O1, Otos_i, Otos_i);  break;
1215    case ushr:  __  srl(O1, Otos_i, Otos_i);  break;
1216    default: ShouldNotReachHere();
1217   }
1218 }
1219 
1220 
1221 void TemplateTable::lop2(Operation op) {
1222   transition(ltos, ltos);
1223   __ pop_l(O2);
1224   switch (op) {
1225    case  add:  __  add(O2, Otos_l, Otos_l);  break;
1226    case  sub:  __  sub(O2, Otos_l, Otos_l);  break;
1227    case _and:  __ and3(O2, Otos_l, Otos_l);  break;
1228    case  _or:  __  or3(O2, Otos_l, Otos_l);  break;
1229    case _xor:  __ xor3(O2, Otos_l, Otos_l);  break;
1230    default: ShouldNotReachHere();
1231   }
1232 }
1233 
1234 
1235 void TemplateTable::idiv() {
1236   // %%%%% Later: ForSPARC/V7 call .sdiv library routine,
1237   // %%%%% Use ldsw...sdivx on pure V9 ABI. 64 bit safe.
1238 
1239   transition(itos, itos);
1240   __ pop_i(O1); // get 1st op
1241 
1242   // Y contains upper 32 bits of result, set it to 0 or all ones
1243   __ wry(G0);
1244   __ mov(~0, G3_scratch);
1245 
1246   __ tst(O1);
1247      Label neg;
1248   __ br(Assembler::negative, true, Assembler::pn, neg);
1249   __ delayed()->wry(G3_scratch);
1250   __ bind(neg);
1251 
1252      Label ok;
1253   __ tst(Otos_i);
1254   __ throw_if_not_icc( Assembler::notZero, Interpreter::_throw_ArithmeticException_entry, G3_scratch );
1255 
1256   const int min_int = 0x80000000;
1257   Label regular;
1258   __ cmp(Otos_i, -1);
1259   __ br(Assembler::notEqual, false, Assembler::pt, regular);
1260   // Don't put set in delay slot
1261   // Set will turn into multiple instructions in 64 bit mode
1262   __ delayed()->nop();
1263   __ set(min_int, G4_scratch);
1264   Label done;
1265   __ cmp(O1, G4_scratch);
1266   __ br(Assembler::equal, true, Assembler::pt, done);
1267   __ delayed()->mov(O1, Otos_i);   // (mov only executed if branch taken)
1268 
1269   __ bind(regular);
1270   __ sdiv(O1, Otos_i, Otos_i); // note: irem uses O1 after this instruction!
1271   __ bind(done);
1272 }
1273 
1274 
1275 void TemplateTable::irem() {
1276   transition(itos, itos);
1277   __ mov(Otos_i, O2); // save divisor
1278   idiv();                               // %%%% Hack: exploits fact that idiv leaves dividend in O1
1279   __ smul(Otos_i, O2, Otos_i);
1280   __ sub(O1, Otos_i, Otos_i);
1281 }
1282 
1283 
1284 void TemplateTable::lmul() {
1285   transition(ltos, ltos);
1286   __ pop_l(O2);
1287   __ mulx(Otos_l, O2, Otos_l);
1288 
1289 }
1290 
1291 
1292 void TemplateTable::ldiv() {
1293   transition(ltos, ltos);
1294 
1295   // check for zero
1296   __ pop_l(O2);
1297   __ tst(Otos_l);
1298   __ throw_if_not_xcc( Assembler::notZero, Interpreter::_throw_ArithmeticException_entry, G3_scratch);
1299   __ sdivx(O2, Otos_l, Otos_l);
1300 }
1301 
1302 
1303 void TemplateTable::lrem() {
1304   transition(ltos, ltos);
1305 
1306   // check for zero
1307   __ pop_l(O2);
1308   __ tst(Otos_l);
1309   __ throw_if_not_xcc( Assembler::notZero, Interpreter::_throw_ArithmeticException_entry, G3_scratch);
1310   __ sdivx(O2, Otos_l, Otos_l2);
1311   __ mulx (Otos_l2, Otos_l, Otos_l2);
1312   __ sub  (O2, Otos_l2, Otos_l);
1313 }
1314 
1315 
1316 void TemplateTable::lshl() {
1317   transition(itos, ltos); // %%%% could optimize, fill delay slot or opt for ultra
1318 
1319   __ pop_l(O2);                          // shift value in O2, O3
1320   __ sllx(O2, Otos_i, Otos_l);
1321 }
1322 
1323 
1324 void TemplateTable::lshr() {
1325   transition(itos, ltos); // %%%% see lshl comment
1326 
1327   __ pop_l(O2);                          // shift value in O2, O3
1328   __ srax(O2, Otos_i, Otos_l);
1329 }
1330 
1331 
1332 
1333 void TemplateTable::lushr() {
1334   transition(itos, ltos); // %%%% see lshl comment
1335 
1336   __ pop_l(O2);                          // shift value in O2, O3
1337   __ srlx(O2, Otos_i, Otos_l);
1338 }
1339 
1340 
1341 void TemplateTable::fop2(Operation op) {
1342   transition(ftos, ftos);
1343   switch (op) {
1344    case  add:  __  pop_f(F4); __ fadd(FloatRegisterImpl::S, F4, Ftos_f, Ftos_f);  break;
1345    case  sub:  __  pop_f(F4); __ fsub(FloatRegisterImpl::S, F4, Ftos_f, Ftos_f);  break;
1346    case  mul:  __  pop_f(F4); __ fmul(FloatRegisterImpl::S, F4, Ftos_f, Ftos_f);  break;
1347    case  div:  __  pop_f(F4); __ fdiv(FloatRegisterImpl::S, F4, Ftos_f, Ftos_f);  break;
1348    case  rem:
1349      assert(Ftos_f == F0, "just checking");
1350      // LP64 calling conventions use F1, F3 for passing 2 floats
1351      __ pop_f(F1);
1352      __ fmov(FloatRegisterImpl::S, Ftos_f, F3);
1353      __ call_VM_leaf(Lscratch, CAST_FROM_FN_PTR(address, SharedRuntime::frem));
1354      assert( Ftos_f == F0, "fix this code" );
1355      break;
1356 
1357    default: ShouldNotReachHere();
1358   }
1359 }
1360 
1361 
1362 void TemplateTable::dop2(Operation op) {
1363   transition(dtos, dtos);
1364   switch (op) {
1365    case  add:  __  pop_d(F4); __ fadd(FloatRegisterImpl::D, F4, Ftos_d, Ftos_d);  break;
1366    case  sub:  __  pop_d(F4); __ fsub(FloatRegisterImpl::D, F4, Ftos_d, Ftos_d);  break;
1367    case  mul:  __  pop_d(F4); __ fmul(FloatRegisterImpl::D, F4, Ftos_d, Ftos_d);  break;
1368    case  div:  __  pop_d(F4); __ fdiv(FloatRegisterImpl::D, F4, Ftos_d, Ftos_d);  break;
1369    case  rem:
1370      // Pass arguments in D0, D2
1371      __ fmov(FloatRegisterImpl::D, Ftos_f, F2 );
1372      __ pop_d( F0 );
1373      __ call_VM_leaf(Lscratch, CAST_FROM_FN_PTR(address, SharedRuntime::drem));
1374      assert( Ftos_d == F0, "fix this code" );
1375      break;
1376 
1377    default: ShouldNotReachHere();
1378   }
1379 }
1380 
1381 
1382 void TemplateTable::ineg() {
1383   transition(itos, itos);
1384   __ neg(Otos_i);
1385 }
1386 
1387 
1388 void TemplateTable::lneg() {
1389   transition(ltos, ltos);
1390   __ sub(G0, Otos_l, Otos_l);
1391 }
1392 
1393 
1394 void TemplateTable::fneg() {
1395   transition(ftos, ftos);
1396   __ fneg(FloatRegisterImpl::S, Ftos_f, Ftos_f);
1397 }
1398 
1399 
1400 void TemplateTable::dneg() {
1401   transition(dtos, dtos);
1402   __ fneg(FloatRegisterImpl::D, Ftos_f, Ftos_f);
1403 }
1404 
1405 
1406 void TemplateTable::iinc() {
1407   transition(vtos, vtos);
1408   locals_index(G3_scratch);
1409   __ ldsb(Lbcp, 2, O2);  // load constant
1410   __ access_local_int(G3_scratch, Otos_i);
1411   __ add(Otos_i, O2, Otos_i);
1412   __ st(Otos_i, G3_scratch, 0);    // access_local_int puts E.A. in G3_scratch
1413 }
1414 
1415 
1416 void TemplateTable::wide_iinc() {
1417   transition(vtos, vtos);
1418   locals_index_wide(G3_scratch);
1419   __ get_2_byte_integer_at_bcp( 4,  O2, O3, InterpreterMacroAssembler::Signed);
1420   __ access_local_int(G3_scratch, Otos_i);
1421   __ add(Otos_i, O3, Otos_i);
1422   __ st(Otos_i, G3_scratch, 0);    // access_local_int puts E.A. in G3_scratch
1423 }
1424 
1425 
1426 void TemplateTable::convert() {
1427 // %%%%% Factor this first part accross platforms
1428   #ifdef ASSERT
1429     TosState tos_in  = ilgl;
1430     TosState tos_out = ilgl;
1431     switch (bytecode()) {
1432       case Bytecodes::_i2l: // fall through
1433       case Bytecodes::_i2f: // fall through
1434       case Bytecodes::_i2d: // fall through
1435       case Bytecodes::_i2b: // fall through
1436       case Bytecodes::_i2c: // fall through
1437       case Bytecodes::_i2s: tos_in = itos; break;
1438       case Bytecodes::_l2i: // fall through
1439       case Bytecodes::_l2f: // fall through
1440       case Bytecodes::_l2d: tos_in = ltos; break;
1441       case Bytecodes::_f2i: // fall through
1442       case Bytecodes::_f2l: // fall through
1443       case Bytecodes::_f2d: tos_in = ftos; break;
1444       case Bytecodes::_d2i: // fall through
1445       case Bytecodes::_d2l: // fall through
1446       case Bytecodes::_d2f: tos_in = dtos; break;
1447       default             : ShouldNotReachHere();
1448     }
1449     switch (bytecode()) {
1450       case Bytecodes::_l2i: // fall through
1451       case Bytecodes::_f2i: // fall through
1452       case Bytecodes::_d2i: // fall through
1453       case Bytecodes::_i2b: // fall through
1454       case Bytecodes::_i2c: // fall through
1455       case Bytecodes::_i2s: tos_out = itos; break;
1456       case Bytecodes::_i2l: // fall through
1457       case Bytecodes::_f2l: // fall through
1458       case Bytecodes::_d2l: tos_out = ltos; break;
1459       case Bytecodes::_i2f: // fall through
1460       case Bytecodes::_l2f: // fall through
1461       case Bytecodes::_d2f: tos_out = ftos; break;
1462       case Bytecodes::_i2d: // fall through
1463       case Bytecodes::_l2d: // fall through
1464       case Bytecodes::_f2d: tos_out = dtos; break;
1465       default             : ShouldNotReachHere();
1466     }
1467     transition(tos_in, tos_out);
1468   #endif
1469 
1470 
1471   // Conversion
1472   Label done;
1473   switch (bytecode()) {
1474    case Bytecodes::_i2l:
1475     // Sign extend the 32 bits
1476     __ sra ( Otos_i, 0, Otos_l );
1477     break;
1478 
1479    case Bytecodes::_i2f:
1480     __ st(Otos_i, __ d_tmp );
1481     __ ldf(FloatRegisterImpl::S,  __ d_tmp, F0);
1482     __ fitof(FloatRegisterImpl::S, F0, Ftos_f);
1483     break;
1484 
1485    case Bytecodes::_i2d:
1486     __ st(Otos_i, __ d_tmp);
1487     __ ldf(FloatRegisterImpl::S,  __ d_tmp, F0);
1488     __ fitof(FloatRegisterImpl::D, F0, Ftos_f);
1489     break;
1490 
1491    case Bytecodes::_i2b:
1492     __ sll(Otos_i, 24, Otos_i);
1493     __ sra(Otos_i, 24, Otos_i);
1494     break;
1495 
1496    case Bytecodes::_i2c:
1497     __ sll(Otos_i, 16, Otos_i);
1498     __ srl(Otos_i, 16, Otos_i);
1499     break;
1500 
1501    case Bytecodes::_i2s:
1502     __ sll(Otos_i, 16, Otos_i);
1503     __ sra(Otos_i, 16, Otos_i);
1504     break;
1505 
1506    case Bytecodes::_l2i:
1507     // Sign-extend into the high 32 bits
1508     __ sra(Otos_l, 0, Otos_i);
1509     break;
1510 
1511    case Bytecodes::_l2f:
1512    case Bytecodes::_l2d:
1513     __ st_long(Otos_l, __ d_tmp);
1514     __ ldf(FloatRegisterImpl::D, __ d_tmp, Ftos_d);
1515 
1516     if (bytecode() == Bytecodes::_l2f) {
1517       __ fxtof(FloatRegisterImpl::S, Ftos_d, Ftos_f);
1518     } else {
1519       __ fxtof(FloatRegisterImpl::D, Ftos_d, Ftos_d);
1520     }
1521     break;
1522 
1523   case Bytecodes::_f2i:  {
1524       Label isNaN;
1525       // result must be 0 if value is NaN; test by comparing value to itself
1526       __ fcmp(FloatRegisterImpl::S, Assembler::fcc0, Ftos_f, Ftos_f);
1527       __ fb(Assembler::f_unordered, true, Assembler::pn, isNaN);
1528       __ delayed()->clr(Otos_i);                                     // NaN
1529       __ ftoi(FloatRegisterImpl::S, Ftos_f, F30);
1530       __ stf(FloatRegisterImpl::S, F30, __ d_tmp);
1531       __ ld(__ d_tmp, Otos_i);
1532       __ bind(isNaN);
1533     }
1534     break;
1535 
1536    case Bytecodes::_f2l:
1537     // must uncache tos
1538     __ push_f();
1539     __ pop_f(F1);
1540     __ call_VM_leaf(Lscratch, CAST_FROM_FN_PTR(address, SharedRuntime::f2l));
1541     break;
1542 
1543    case Bytecodes::_f2d:
1544     __ ftof( FloatRegisterImpl::S, FloatRegisterImpl::D, Ftos_f, Ftos_f);
1545     break;
1546 
1547    case Bytecodes::_d2i:
1548    case Bytecodes::_d2l:
1549     // must uncache tos
1550     __ push_d();
1551     // LP64 calling conventions pass first double arg in D0
1552     __ pop_d( Ftos_d );
1553     __ call_VM_leaf(Lscratch,
1554         bytecode() == Bytecodes::_d2i
1555           ? CAST_FROM_FN_PTR(address, SharedRuntime::d2i)
1556           : CAST_FROM_FN_PTR(address, SharedRuntime::d2l));
1557     break;
1558 
1559     case Bytecodes::_d2f:
1560       __ ftof( FloatRegisterImpl::D, FloatRegisterImpl::S, Ftos_d, Ftos_f);
1561     break;
1562 
1563     default: ShouldNotReachHere();
1564   }
1565   __ bind(done);
1566 }
1567 
1568 
1569 void TemplateTable::lcmp() {
1570   transition(ltos, itos);
1571 
1572   __ pop_l(O1); // pop off value 1, value 2 is in O0
1573   __ lcmp( O1, Otos_l, Otos_i );
1574 }
1575 
1576 
1577 void TemplateTable::float_cmp(bool is_float, int unordered_result) {
1578 
1579   if (is_float) __ pop_f(F2);
1580   else          __ pop_d(F2);
1581 
1582   assert(Ftos_f == F0  &&  Ftos_d == F0,  "alias checking:");
1583 
1584   __ float_cmp( is_float, unordered_result, F2, F0, Otos_i );
1585 }
1586 
1587 void TemplateTable::branch(bool is_jsr, bool is_wide) {
1588   // Note: on SPARC, we use InterpreterMacroAssembler::if_cmp also.
1589   __ verify_thread();
1590 
1591   const Register O2_bumped_count = O2;
1592   __ profile_taken_branch(G3_scratch, O2_bumped_count);
1593 
1594   // get (wide) offset to O1_disp
1595   const Register O1_disp = O1;
1596   if (is_wide)  __ get_4_byte_integer_at_bcp( 1,  G4_scratch, O1_disp,                                    InterpreterMacroAssembler::set_CC);
1597   else          __ get_2_byte_integer_at_bcp( 1,  G4_scratch, O1_disp, InterpreterMacroAssembler::Signed, InterpreterMacroAssembler::set_CC);
1598 
1599   // Handle all the JSR stuff here, then exit.
1600   // It's much shorter and cleaner than intermingling with the
1601   // non-JSR normal-branch stuff occurring below.
1602   if( is_jsr ) {
1603     // compute return address as bci in Otos_i
1604     __ ld_ptr(Lmethod, Method::const_offset(), G3_scratch);
1605     __ sub(Lbcp, G3_scratch, G3_scratch);
1606     __ sub(G3_scratch, in_bytes(ConstMethod::codes_offset()) - (is_wide ? 5 : 3), Otos_i);
1607 
1608     // Bump Lbcp to target of JSR
1609     __ add(Lbcp, O1_disp, Lbcp);
1610     // Push returnAddress for "ret" on stack
1611     __ push_ptr(Otos_i);
1612     // And away we go!
1613     __ dispatch_next(vtos, 0, true);
1614     return;
1615   }
1616 
1617   // Normal (non-jsr) branch handling
1618 
1619   // Save the current Lbcp
1620   const Register l_cur_bcp = Lscratch;
1621   __ mov( Lbcp, l_cur_bcp );
1622 
1623   bool increment_invocation_counter_for_backward_branches = UseCompiler && UseLoopCounter;
1624   if ( increment_invocation_counter_for_backward_branches ) {
1625     Label Lforward;
1626     // check branch direction
1627     __ br( Assembler::positive, false,  Assembler::pn, Lforward );
1628     // Bump bytecode pointer by displacement (take the branch)
1629     __ delayed()->add( O1_disp, Lbcp, Lbcp );     // add to bc addr
1630 
1631     const Register G3_method_counters = G3_scratch;
1632     __ get_method_counters(Lmethod, G3_method_counters, Lforward);
1633 
1634     if (TieredCompilation) {
1635       Label Lno_mdo, Loverflow;
1636       int increment = InvocationCounter::count_increment;
1637       if (ProfileInterpreter) {
1638         // If no method data exists, go to profile_continue.
1639         __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
1640         __ br_null_short(G4_scratch, Assembler::pn, Lno_mdo);
1641 
1642         // Increment backedge counter in the MDO
1643         Address mdo_backedge_counter(G4_scratch, in_bytes(MethodData::backedge_counter_offset()) +
1644                                                  in_bytes(InvocationCounter::counter_offset()));
1645         Address mask(G4_scratch, in_bytes(MethodData::backedge_mask_offset()));
1646         __ increment_mask_and_jump(mdo_backedge_counter, increment, mask, G3_scratch, O0,
1647                                    (UseOnStackReplacement ? Assembler::notZero : Assembler::always), &Lforward);
1648         __ ba_short(Loverflow);
1649       }
1650 
1651       // If there's no MDO, increment counter in MethodCounters*
1652       __ bind(Lno_mdo);
1653       Address backedge_counter(G3_method_counters,
1654               in_bytes(MethodCounters::backedge_counter_offset()) +
1655               in_bytes(InvocationCounter::counter_offset()));
1656       Address mask(G3_method_counters, in_bytes(MethodCounters::backedge_mask_offset()));
1657       __ increment_mask_and_jump(backedge_counter, increment, mask, G4_scratch, O0,
1658                                  (UseOnStackReplacement ? Assembler::notZero : Assembler::always), &Lforward);
1659       __ bind(Loverflow);
1660 
1661       // notify point for loop, pass branch bytecode
1662       __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), l_cur_bcp);
1663 
1664       // Was an OSR adapter generated?
1665       // O0 = osr nmethod
1666       __ br_null_short(O0, Assembler::pn, Lforward);
1667 
1668       // Has the nmethod been invalidated already?
1669       __ ldub(O0, nmethod::state_offset(), O2);
1670       __ cmp_and_br_short(O2, nmethod::in_use, Assembler::notEqual, Assembler::pn, Lforward);
1671 
1672       // migrate the interpreter frame off of the stack
1673 
1674       __ mov(G2_thread, L7);
1675       // save nmethod
1676       __ mov(O0, L6);
1677       __ set_last_Java_frame(SP, noreg);
1678       __ call_VM_leaf(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::OSR_migration_begin), L7);
1679       __ reset_last_Java_frame();
1680       __ mov(L7, G2_thread);
1681 
1682       // move OSR nmethod to I1
1683       __ mov(L6, I1);
1684 
1685       // OSR buffer to I0
1686       __ mov(O0, I0);
1687 
1688       // remove the interpreter frame
1689       __ restore(I5_savedSP, 0, SP);
1690 
1691       // Jump to the osr code.
1692       __ ld_ptr(O1, nmethod::osr_entry_point_offset(), O2);
1693       __ jmp(O2, G0);
1694       __ delayed()->nop();
1695 
1696     } else { // not TieredCompilation
1697       // Update Backedge branch separately from invocations
1698       const Register G4_invoke_ctr = G4;
1699       __ increment_backedge_counter(G3_method_counters, G4_invoke_ctr, G1_scratch);
1700       if (ProfileInterpreter) {
1701         __ test_invocation_counter_for_mdp(G4_invoke_ctr, G3_method_counters, G1_scratch, Lforward);
1702         if (UseOnStackReplacement) {
1703 
1704           __ test_backedge_count_for_osr(O2_bumped_count, G3_method_counters, l_cur_bcp, G1_scratch);
1705         }
1706       } else {
1707         if (UseOnStackReplacement) {
1708           __ test_backedge_count_for_osr(G4_invoke_ctr, G3_method_counters, l_cur_bcp, G1_scratch);
1709         }
1710       }
1711     }
1712 
1713     __ bind(Lforward);
1714   } else
1715     // Bump bytecode pointer by displacement (take the branch)
1716     __ add( O1_disp, Lbcp, Lbcp );// add to bc addr
1717 
1718   // continue with bytecode @ target
1719   // %%%%% Like Intel, could speed things up by moving bytecode fetch to code above,
1720   // %%%%% and changing dispatch_next to dispatch_only
1721   __ dispatch_next(vtos, 0, true);
1722 }
1723 
1724 
1725 // Note Condition in argument is TemplateTable::Condition
1726 // arg scope is within class scope
1727 
1728 void TemplateTable::if_0cmp(Condition cc) {
1729   // no pointers, integer only!
1730   transition(itos, vtos);
1731   // assume branch is more often taken than not (loops use backward branches)
1732   __ cmp( Otos_i, 0);
1733   __ if_cmp(ccNot(cc), false);
1734 }
1735 
1736 
1737 void TemplateTable::if_icmp(Condition cc) {
1738   transition(itos, vtos);
1739   __ pop_i(O1);
1740   __ cmp(O1, Otos_i);
1741   __ if_cmp(ccNot(cc), false);
1742 }
1743 
1744 
1745 void TemplateTable::if_nullcmp(Condition cc) {
1746   transition(atos, vtos);
1747   __ tst(Otos_i);
1748   __ if_cmp(ccNot(cc), true);
1749 }
1750 
1751 
1752 void TemplateTable::if_acmp(Condition cc) {
1753   transition(atos, vtos);
1754   __ pop_ptr(O1);
1755   __ verify_oop(O1);
1756   __ verify_oop(Otos_i);
1757   __ cmp(O1, Otos_i);
1758   __ if_cmp(ccNot(cc), true);
1759 }
1760 
1761 
1762 
1763 void TemplateTable::ret() {
1764   transition(vtos, vtos);
1765   locals_index(G3_scratch);
1766   __ access_local_returnAddress(G3_scratch, Otos_i);
1767   // Otos_i contains the bci, compute the bcp from that
1768 
1769 #ifdef ASSERT
1770   // jsr result was labeled as an 'itos' not an 'atos' because we cannot GC
1771   // the result.  The return address (really a BCI) was stored with an
1772   // 'astore' because JVM specs claim it's a pointer-sized thing.  Hence in
1773   // the 64-bit build the 32-bit BCI is actually in the low bits of a 64-bit
1774   // loaded value.
1775   { Label zzz ;
1776      __ set (65536, G3_scratch) ;
1777      __ cmp (Otos_i, G3_scratch) ;
1778      __ bp( Assembler::lessEqualUnsigned, false, Assembler::xcc, Assembler::pn, zzz);
1779      __ delayed()->nop();
1780      __ stop("BCI is in the wrong register half?");
1781      __ bind (zzz) ;
1782   }
1783 #endif
1784 
1785   __ profile_ret(vtos, Otos_i, G4_scratch);
1786 
1787   __ ld_ptr(Lmethod, Method::const_offset(), G3_scratch);
1788   __ add(G3_scratch, Otos_i, G3_scratch);
1789   __ add(G3_scratch, in_bytes(ConstMethod::codes_offset()), Lbcp);
1790   __ dispatch_next(vtos, 0, true);
1791 }
1792 
1793 
1794 void TemplateTable::wide_ret() {
1795   transition(vtos, vtos);
1796   locals_index_wide(G3_scratch);
1797   __ access_local_returnAddress(G3_scratch, Otos_i);
1798   // Otos_i contains the bci, compute the bcp from that
1799 
1800   __ profile_ret(vtos, Otos_i, G4_scratch);
1801 
1802   __ ld_ptr(Lmethod, Method::const_offset(), G3_scratch);
1803   __ add(G3_scratch, Otos_i, G3_scratch);
1804   __ add(G3_scratch, in_bytes(ConstMethod::codes_offset()), Lbcp);
1805   __ dispatch_next(vtos, 0, true);
1806 }
1807 
1808 
1809 void TemplateTable::tableswitch() {
1810   transition(itos, vtos);
1811   Label default_case, continue_execution;
1812 
1813   // align bcp
1814   __ add(Lbcp, BytesPerInt, O1);
1815   __ and3(O1, -BytesPerInt, O1);
1816   // load lo, hi
1817   __ ld(O1, 1 * BytesPerInt, O2);       // Low Byte
1818   __ ld(O1, 2 * BytesPerInt, O3);       // High Byte
1819   // Sign extend the 32 bits
1820   __ sra ( Otos_i, 0, Otos_i );
1821 
1822   // check against lo & hi
1823   __ cmp( Otos_i, O2);
1824   __ br( Assembler::less, false, Assembler::pn, default_case);
1825   __ delayed()->cmp( Otos_i, O3 );
1826   __ br( Assembler::greater, false, Assembler::pn, default_case);
1827   // lookup dispatch offset
1828   __ delayed()->sub(Otos_i, O2, O2);
1829   __ profile_switch_case(O2, O3, G3_scratch, G4_scratch);
1830   __ sll(O2, LogBytesPerInt, O2);
1831   __ add(O2, 3 * BytesPerInt, O2);
1832   __ ba(continue_execution);
1833   __ delayed()->ld(O1, O2, O2);
1834   // handle default
1835   __ bind(default_case);
1836   __ profile_switch_default(O3);
1837   __ ld(O1, 0, O2); // get default offset
1838   // continue execution
1839   __ bind(continue_execution);
1840   __ add(Lbcp, O2, Lbcp);
1841   __ dispatch_next(vtos, 0, true);
1842 }
1843 
1844 
1845 void TemplateTable::lookupswitch() {
1846   transition(itos, itos);
1847   __ stop("lookupswitch bytecode should have been rewritten");
1848 }
1849 
1850 void TemplateTable::fast_linearswitch() {
1851   transition(itos, vtos);
1852     Label loop_entry, loop, found, continue_execution;
1853   // align bcp
1854   __ add(Lbcp, BytesPerInt, O1);
1855   __ and3(O1, -BytesPerInt, O1);
1856  // set counter
1857   __ ld(O1, BytesPerInt, O2);
1858   __ sll(O2, LogBytesPerInt + 1, O2); // in word-pairs
1859   __ add(O1, 2 * BytesPerInt, O3); // set first pair addr
1860   __ ba(loop_entry);
1861   __ delayed()->add(O3, O2, O2); // counter now points past last pair
1862 
1863   // table search
1864   __ bind(loop);
1865   __ cmp(O4, Otos_i);
1866   __ br(Assembler::equal, true, Assembler::pn, found);
1867   __ delayed()->ld(O3, BytesPerInt, O4); // offset -> O4
1868   __ inc(O3, 2 * BytesPerInt);
1869 
1870   __ bind(loop_entry);
1871   __ cmp(O2, O3);
1872   __ brx(Assembler::greaterUnsigned, true, Assembler::pt, loop);
1873   __ delayed()->ld(O3, 0, O4);
1874 
1875   // default case
1876   __ ld(O1, 0, O4); // get default offset
1877   if (ProfileInterpreter) {
1878     __ profile_switch_default(O3);
1879     __ ba_short(continue_execution);
1880   }
1881 
1882   // entry found -> get offset
1883   __ bind(found);
1884   if (ProfileInterpreter) {
1885     __ sub(O3, O1, O3);
1886     __ sub(O3, 2*BytesPerInt, O3);
1887     __ srl(O3, LogBytesPerInt + 1, O3); // in word-pairs
1888     __ profile_switch_case(O3, O1, O2, G3_scratch);
1889 
1890     __ bind(continue_execution);
1891   }
1892   __ add(Lbcp, O4, Lbcp);
1893   __ dispatch_next(vtos, 0, true);
1894 }
1895 
1896 
1897 void TemplateTable::fast_binaryswitch() {
1898   transition(itos, vtos);
1899   // Implementation using the following core algorithm: (copied from Intel)
1900   //
1901   // int binary_search(int key, LookupswitchPair* array, int n) {
1902   //   // Binary search according to "Methodik des Programmierens" by
1903   //   // Edsger W. Dijkstra and W.H.J. Feijen, Addison Wesley Germany 1985.
1904   //   int i = 0;
1905   //   int j = n;
1906   //   while (i+1 < j) {
1907   //     // invariant P: 0 <= i < j <= n and (a[i] <= key < a[j] or Q)
1908   //     // with      Q: for all i: 0 <= i < n: key < a[i]
1909   //     // where a stands for the array and assuming that the (inexisting)
1910   //     // element a[n] is infinitely big.
1911   //     int h = (i + j) >> 1;
1912   //     // i < h < j
1913   //     if (key < array[h].fast_match()) {
1914   //       j = h;
1915   //     } else {
1916   //       i = h;
1917   //     }
1918   //   }
1919   //   // R: a[i] <= key < a[i+1] or Q
1920   //   // (i.e., if key is within array, i is the correct index)
1921   //   return i;
1922   // }
1923 
1924   // register allocation
1925   assert(Otos_i == O0, "alias checking");
1926   const Register Rkey     = Otos_i;                    // already set (tosca)
1927   const Register Rarray   = O1;
1928   const Register Ri       = O2;
1929   const Register Rj       = O3;
1930   const Register Rh       = O4;
1931   const Register Rscratch = O5;
1932 
1933   const int log_entry_size = 3;
1934   const int entry_size = 1 << log_entry_size;
1935 
1936   Label found;
1937   // Find Array start
1938   __ add(Lbcp, 3 * BytesPerInt, Rarray);
1939   __ and3(Rarray, -BytesPerInt, Rarray);
1940   // initialize i & j (in delay slot)
1941   __ clr( Ri );
1942 
1943   // and start
1944   Label entry;
1945   __ ba(entry);
1946   __ delayed()->ld( Rarray, -BytesPerInt, Rj);
1947   // (Rj is already in the native byte-ordering.)
1948 
1949   // binary search loop
1950   { Label loop;
1951     __ bind( loop );
1952     // int h = (i + j) >> 1;
1953     __ sra( Rh, 1, Rh );
1954     // if (key < array[h].fast_match()) {
1955     //   j = h;
1956     // } else {
1957     //   i = h;
1958     // }
1959     __ sll( Rh, log_entry_size, Rscratch );
1960     __ ld( Rarray, Rscratch, Rscratch );
1961     // (Rscratch is already in the native byte-ordering.)
1962     __ cmp( Rkey, Rscratch );
1963     __ movcc( Assembler::less,         false, Assembler::icc, Rh, Rj );  // j = h if (key <  array[h].fast_match())
1964     __ movcc( Assembler::greaterEqual, false, Assembler::icc, Rh, Ri );  // i = h if (key >= array[h].fast_match())
1965 
1966     // while (i+1 < j)
1967     __ bind( entry );
1968     __ add( Ri, 1, Rscratch );
1969     __ cmp(Rscratch, Rj);
1970     __ br( Assembler::less, true, Assembler::pt, loop );
1971     __ delayed()->add( Ri, Rj, Rh ); // start h = i + j  >> 1;
1972   }
1973 
1974   // end of binary search, result index is i (must check again!)
1975   Label default_case;
1976   Label continue_execution;
1977   if (ProfileInterpreter) {
1978     __ mov( Ri, Rh );              // Save index in i for profiling
1979   }
1980   __ sll( Ri, log_entry_size, Ri );
1981   __ ld( Rarray, Ri, Rscratch );
1982   // (Rscratch is already in the native byte-ordering.)
1983   __ cmp( Rkey, Rscratch );
1984   __ br( Assembler::notEqual, true, Assembler::pn, default_case );
1985   __ delayed()->ld( Rarray, -2 * BytesPerInt, Rj ); // load default offset -> j
1986 
1987   // entry found -> j = offset
1988   __ inc( Ri, BytesPerInt );
1989   __ profile_switch_case(Rh, Rj, Rscratch, Rkey);
1990   __ ld( Rarray, Ri, Rj );
1991   // (Rj is already in the native byte-ordering.)
1992 
1993   if (ProfileInterpreter) {
1994     __ ba_short(continue_execution);
1995   }
1996 
1997   __ bind(default_case); // fall through (if not profiling)
1998   __ profile_switch_default(Ri);
1999 
2000   __ bind(continue_execution);
2001   __ add( Lbcp, Rj, Lbcp );
2002   __ dispatch_next(vtos, 0, true);
2003 }
2004 
2005 
2006 void TemplateTable::_return(TosState state) {
2007   transition(state, state);
2008   assert(_desc->calls_vm(), "inconsistent calls_vm information");
2009 
2010   if (_desc->bytecode() == Bytecodes::_return_register_finalizer) {
2011     assert(state == vtos, "only valid state");
2012     __ mov(G0, G3_scratch);
2013     __ access_local_ptr(G3_scratch, Otos_i);
2014     __ load_klass(Otos_i, O2);
2015     __ set(JVM_ACC_HAS_FINALIZER, G3);
2016     __ ld(O2, in_bytes(Klass::access_flags_offset()), O2);
2017     __ andcc(G3, O2, G0);
2018     Label skip_register_finalizer;
2019     __ br(Assembler::zero, false, Assembler::pn, skip_register_finalizer);
2020     __ delayed()->nop();
2021 
2022     // Call out to do finalizer registration
2023     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::register_finalizer), Otos_i);
2024 
2025     __ bind(skip_register_finalizer);
2026   }
2027 
2028   if (_desc->bytecode() != Bytecodes::_return_register_finalizer) {
2029     Label no_safepoint;
2030     __ ldx(Address(G2_thread, Thread::polling_page_offset()), G3_scratch, 0);
2031     __ btst(SafepointMechanism::poll_bit(), G3_scratch);
2032     __ br(Assembler::zero, false, Assembler::pt, no_safepoint);
2033     __ delayed()->nop();
2034     __ push(state);
2035     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::at_safepoint));
2036     __ pop(state);
2037     __ bind(no_safepoint);
2038   }
2039 
2040   // Narrow result if state is itos but result type is smaller.
2041   // Need to narrow in the return bytecode rather than in generate_return_entry
2042   // since compiled code callers expect the result to already be narrowed.
2043   if (state == itos) {
2044     __ narrow(Otos_i);
2045   }
2046   __ remove_activation(state, /* throw_monitor_exception */ true);
2047 
2048   // The caller's SP was adjusted upon method entry to accomodate
2049   // the callee's non-argument locals. Undo that adjustment.
2050   __ ret();                             // return to caller
2051   __ delayed()->restore(I5_savedSP, G0, SP);
2052 }
2053 
2054 
2055 // ----------------------------------------------------------------------------
2056 // Volatile variables demand their effects be made known to all CPU's in
2057 // order.  Store buffers on most chips allow reads & writes to reorder; the
2058 // JMM's ReadAfterWrite.java test fails in -Xint mode without some kind of
2059 // memory barrier (i.e., it's not sufficient that the interpreter does not
2060 // reorder volatile references, the hardware also must not reorder them).
2061 //
2062 // According to the new Java Memory Model (JMM):
2063 // (1) All volatiles are serialized wrt to each other.
2064 // ALSO reads & writes act as aquire & release, so:
2065 // (2) A read cannot let unrelated NON-volatile memory refs that happen after
2066 // the read float up to before the read.  It's OK for non-volatile memory refs
2067 // that happen before the volatile read to float down below it.
2068 // (3) Similar a volatile write cannot let unrelated NON-volatile memory refs
2069 // that happen BEFORE the write float down to after the write.  It's OK for
2070 // non-volatile memory refs that happen after the volatile write to float up
2071 // before it.
2072 //
2073 // We only put in barriers around volatile refs (they are expensive), not
2074 // _between_ memory refs (that would require us to track the flavor of the
2075 // previous memory refs).  Requirements (2) and (3) require some barriers
2076 // before volatile stores and after volatile loads.  These nearly cover
2077 // requirement (1) but miss the volatile-store-volatile-load case.  This final
2078 // case is placed after volatile-stores although it could just as well go
2079 // before volatile-loads.
2080 void TemplateTable::volatile_barrier(Assembler::Membar_mask_bits order_constraint) {
2081   // Helper function to insert a is-volatile test and memory barrier
2082   // All current sparc implementations run in TSO, needing only StoreLoad
2083   if ((order_constraint & Assembler::StoreLoad) == 0) return;
2084   __ membar( order_constraint );
2085 }
2086 
2087 // ----------------------------------------------------------------------------
2088 void TemplateTable::resolve_cache_and_index(int byte_no,
2089                                             Register Rcache,
2090                                             Register index,
2091                                             size_t index_size) {
2092   // Depends on cpCacheOop layout!
2093 
2094   Label resolved;
2095   Bytecodes::Code code = bytecode();
2096   switch (code) {
2097   case Bytecodes::_nofast_getfield: code = Bytecodes::_getfield; break;
2098   case Bytecodes::_nofast_putfield: code = Bytecodes::_putfield; break;
2099   }
2100 
2101   assert(byte_no == f1_byte || byte_no == f2_byte, "byte_no out of range");
2102   __ get_cache_and_index_and_bytecode_at_bcp(Rcache, index, Lbyte_code, byte_no, 1, index_size);
2103   __ cmp(Lbyte_code, code);  // have we resolved this bytecode?
2104   __ br(Assembler::equal, false, Assembler::pt, resolved);
2105   __ delayed()->set(code, O1);
2106 
2107   address entry = CAST_FROM_FN_PTR(address, InterpreterRuntime::resolve_from_cache);
2108   // first time invocation - must resolve first
2109   __ call_VM(noreg, entry, O1);
2110   // Update registers with resolved info
2111   __ get_cache_and_index_at_bcp(Rcache, index, 1, index_size);
2112   __ bind(resolved);
2113 }
2114 
2115 void TemplateTable::load_invoke_cp_cache_entry(int byte_no,
2116                                                Register method,
2117                                                Register itable_index,
2118                                                Register flags,
2119                                                bool is_invokevirtual,
2120                                                bool is_invokevfinal,
2121                                                bool is_invokedynamic) {
2122   // Uses both G3_scratch and G4_scratch
2123   Register cache = G3_scratch;
2124   Register index = G4_scratch;
2125   assert_different_registers(cache, method, itable_index);
2126 
2127   // determine constant pool cache field offsets
2128   assert(is_invokevirtual == (byte_no == f2_byte), "is_invokevirtual flag redundant");
2129   const int method_offset = in_bytes(
2130       ConstantPoolCache::base_offset() +
2131       ((byte_no == f2_byte)
2132        ? ConstantPoolCacheEntry::f2_offset()
2133        : ConstantPoolCacheEntry::f1_offset()
2134       )
2135     );
2136   const int flags_offset = in_bytes(ConstantPoolCache::base_offset() +
2137                                     ConstantPoolCacheEntry::flags_offset());
2138   // access constant pool cache fields
2139   const int index_offset = in_bytes(ConstantPoolCache::base_offset() +
2140                                     ConstantPoolCacheEntry::f2_offset());
2141 
2142   if (is_invokevfinal) {
2143     __ get_cache_and_index_at_bcp(cache, index, 1);
2144     __ ld_ptr(Address(cache, method_offset), method);
2145   } else {
2146     size_t index_size = (is_invokedynamic ? sizeof(u4) : sizeof(u2));
2147     resolve_cache_and_index(byte_no, cache, index, index_size);
2148     __ ld_ptr(Address(cache, method_offset), method);
2149   }
2150 
2151   if (itable_index != noreg) {
2152     // pick up itable or appendix index from f2 also:
2153     __ ld_ptr(Address(cache, index_offset), itable_index);
2154   }
2155   __ ld_ptr(Address(cache, flags_offset), flags);
2156 }
2157 
2158 // The Rcache register must be set before call
2159 void TemplateTable::load_field_cp_cache_entry(Register Robj,
2160                                               Register Rcache,
2161                                               Register index,
2162                                               Register Roffset,
2163                                               Register Rflags,
2164                                               bool is_static) {
2165   assert_different_registers(Rcache, Rflags, Roffset, Lscratch);
2166 
2167   ByteSize cp_base_offset = ConstantPoolCache::base_offset();
2168 
2169   __ ld_ptr(Rcache, cp_base_offset + ConstantPoolCacheEntry::flags_offset(), Rflags);
2170   __ ld_ptr(Rcache, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), Roffset);
2171   if (is_static) {
2172     __ ld_ptr(Rcache, cp_base_offset + ConstantPoolCacheEntry::f1_offset(), Robj);
2173     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
2174     __ ld_ptr( Robj, mirror_offset, Robj);
2175     __ resolve_oop_handle(Robj, Lscratch);
2176   }
2177 }
2178 
2179 // The registers Rcache and index expected to be set before call.
2180 // Correct values of the Rcache and index registers are preserved.
2181 void TemplateTable::jvmti_post_field_access(Register Rcache,
2182                                             Register index,
2183                                             bool is_static,
2184                                             bool has_tos) {
2185   ByteSize cp_base_offset = ConstantPoolCache::base_offset();
2186 
2187   if (JvmtiExport::can_post_field_access()) {
2188     // Check to see if a field access watch has been set before we take
2189     // the time to call into the VM.
2190     Label Label1;
2191     assert_different_registers(Rcache, index, G1_scratch);
2192     AddressLiteral get_field_access_count_addr(JvmtiExport::get_field_access_count_addr());
2193     __ load_contents(get_field_access_count_addr, G1_scratch);
2194     __ cmp_and_br_short(G1_scratch, 0, Assembler::equal, Assembler::pt, Label1);
2195 
2196     __ add(Rcache, in_bytes(cp_base_offset), Rcache);
2197 
2198     if (is_static) {
2199       __ clr(Otos_i);
2200     } else {
2201       if (has_tos) {
2202       // save object pointer before call_VM() clobbers it
2203         __ push_ptr(Otos_i);  // put object on tos where GC wants it.
2204       } else {
2205         // Load top of stack (do not pop the value off the stack);
2206         __ ld_ptr(Lesp, Interpreter::expr_offset_in_bytes(0), Otos_i);
2207       }
2208       __ verify_oop(Otos_i);
2209     }
2210     // Otos_i: object pointer or NULL if static
2211     // Rcache: cache entry pointer
2212     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::post_field_access),
2213                Otos_i, Rcache);
2214     if (!is_static && has_tos) {
2215       __ pop_ptr(Otos_i);  // restore object pointer
2216       __ verify_oop(Otos_i);
2217     }
2218     __ get_cache_and_index_at_bcp(Rcache, index, 1);
2219     __ bind(Label1);
2220   }
2221 }
2222 
2223 void TemplateTable::getfield_or_static(int byte_no, bool is_static, RewriteControl rc) {
2224   transition(vtos, vtos);
2225 
2226   Register Rcache = G3_scratch;
2227   Register index  = G4_scratch;
2228   Register Rclass = Rcache;
2229   Register Roffset= G4_scratch;
2230   Register Rflags = G1_scratch;
2231   ByteSize cp_base_offset = ConstantPoolCache::base_offset();
2232 
2233   resolve_cache_and_index(byte_no, Rcache, index, sizeof(u2));
2234   jvmti_post_field_access(Rcache, index, is_static, false);
2235   load_field_cp_cache_entry(Rclass, Rcache, index, Roffset, Rflags, is_static);
2236 
2237   if (!is_static) {
2238     pop_and_check_object(Rclass);
2239   } else {
2240     __ verify_oop(Rclass);
2241   }
2242 
2243   Label exit;
2244 
2245   Assembler::Membar_mask_bits membar_bits =
2246     Assembler::Membar_mask_bits(Assembler::LoadLoad | Assembler::LoadStore);
2247 
2248   if (__ membar_has_effect(membar_bits)) {
2249     // Get volatile flag
2250     __ set((1 << ConstantPoolCacheEntry::is_volatile_shift), Lscratch);
2251     __ and3(Rflags, Lscratch, Lscratch);
2252   }
2253 
2254   Label checkVolatile;
2255 
2256   // compute field type
2257   Label notByte, notBool, notInt, notShort, notChar, notLong, notFloat, notObj;
2258   __ srl(Rflags, ConstantPoolCacheEntry::tos_state_shift, Rflags);
2259   // Make sure we don't need to mask Rflags after the above shift
2260   ConstantPoolCacheEntry::verify_tos_state_shift();
2261 
2262   // Check atos before itos for getstatic, more likely (in Queens at least)
2263   __ cmp(Rflags, atos);
2264   __ br(Assembler::notEqual, false, Assembler::pt, notObj);
2265   __ delayed() ->cmp(Rflags, itos);
2266 
2267   // atos
2268   do_oop_load(_masm, Rclass, Roffset, 0, Otos_i, noreg);
2269   __ verify_oop(Otos_i);
2270   __ push(atos);
2271   if (!is_static && rc == may_rewrite) {
2272     patch_bytecode(Bytecodes::_fast_agetfield, G3_scratch, G4_scratch);
2273   }
2274   __ ba(checkVolatile);
2275   __ delayed()->tst(Lscratch);
2276 
2277   __ bind(notObj);
2278 
2279   // cmp(Rflags, itos);
2280   __ br(Assembler::notEqual, false, Assembler::pt, notInt);
2281   __ delayed() ->cmp(Rflags, ltos);
2282 
2283   // itos
2284   __ ld(Rclass, Roffset, Otos_i);
2285   __ push(itos);
2286   if (!is_static && rc == may_rewrite) {
2287     patch_bytecode(Bytecodes::_fast_igetfield, G3_scratch, G4_scratch);
2288   }
2289   __ ba(checkVolatile);
2290   __ delayed()->tst(Lscratch);
2291 
2292   __ bind(notInt);
2293 
2294   // cmp(Rflags, ltos);
2295   __ br(Assembler::notEqual, false, Assembler::pt, notLong);
2296   __ delayed() ->cmp(Rflags, btos);
2297 
2298   // ltos
2299   // load must be atomic
2300   __ ld_long(Rclass, Roffset, Otos_l);
2301   __ push(ltos);
2302   if (!is_static && rc == may_rewrite) {
2303     patch_bytecode(Bytecodes::_fast_lgetfield, G3_scratch, G4_scratch);
2304   }
2305   __ ba(checkVolatile);
2306   __ delayed()->tst(Lscratch);
2307 
2308   __ bind(notLong);
2309 
2310   // cmp(Rflags, btos);
2311   __ br(Assembler::notEqual, false, Assembler::pt, notByte);
2312   __ delayed() ->cmp(Rflags, ztos);
2313 
2314   // btos
2315   __ ldsb(Rclass, Roffset, Otos_i);
2316   __ push(itos);
2317   if (!is_static && rc == may_rewrite) {
2318     patch_bytecode(Bytecodes::_fast_bgetfield, G3_scratch, G4_scratch);
2319   }
2320   __ ba(checkVolatile);
2321   __ delayed()->tst(Lscratch);
2322 
2323   __ bind(notByte);
2324 
2325   // cmp(Rflags, ztos);
2326   __ br(Assembler::notEqual, false, Assembler::pt, notBool);
2327   __ delayed() ->cmp(Rflags, ctos);
2328 
2329   // ztos
2330   __ ldsb(Rclass, Roffset, Otos_i);
2331   __ push(itos);
2332   if (!is_static && rc == may_rewrite) {
2333     // use btos rewriting, no truncating to t/f bit is needed for getfield.
2334     patch_bytecode(Bytecodes::_fast_bgetfield, G3_scratch, G4_scratch);
2335   }
2336   __ ba(checkVolatile);
2337   __ delayed()->tst(Lscratch);
2338 
2339   __ bind(notBool);
2340 
2341   // cmp(Rflags, ctos);
2342   __ br(Assembler::notEqual, false, Assembler::pt, notChar);
2343   __ delayed() ->cmp(Rflags, stos);
2344 
2345   // ctos
2346   __ lduh(Rclass, Roffset, Otos_i);
2347   __ push(itos);
2348   if (!is_static && rc == may_rewrite) {
2349     patch_bytecode(Bytecodes::_fast_cgetfield, G3_scratch, G4_scratch);
2350   }
2351   __ ba(checkVolatile);
2352   __ delayed()->tst(Lscratch);
2353 
2354   __ bind(notChar);
2355 
2356   // cmp(Rflags, stos);
2357   __ br(Assembler::notEqual, false, Assembler::pt, notShort);
2358   __ delayed() ->cmp(Rflags, ftos);
2359 
2360   // stos
2361   __ ldsh(Rclass, Roffset, Otos_i);
2362   __ push(itos);
2363   if (!is_static && rc == may_rewrite) {
2364     patch_bytecode(Bytecodes::_fast_sgetfield, G3_scratch, G4_scratch);
2365   }
2366   __ ba(checkVolatile);
2367   __ delayed()->tst(Lscratch);
2368 
2369   __ bind(notShort);
2370 
2371 
2372   // cmp(Rflags, ftos);
2373   __ br(Assembler::notEqual, false, Assembler::pt, notFloat);
2374   __ delayed() ->tst(Lscratch);
2375 
2376   // ftos
2377   __ ldf(FloatRegisterImpl::S, Rclass, Roffset, Ftos_f);
2378   __ push(ftos);
2379   if (!is_static && rc == may_rewrite) {
2380     patch_bytecode(Bytecodes::_fast_fgetfield, G3_scratch, G4_scratch);
2381   }
2382   __ ba(checkVolatile);
2383   __ delayed()->tst(Lscratch);
2384 
2385   __ bind(notFloat);
2386 
2387 
2388   // dtos
2389   __ ldf(FloatRegisterImpl::D, Rclass, Roffset, Ftos_d);
2390   __ push(dtos);
2391   if (!is_static && rc == may_rewrite) {
2392     patch_bytecode(Bytecodes::_fast_dgetfield, G3_scratch, G4_scratch);
2393   }
2394 
2395   __ bind(checkVolatile);
2396   if (__ membar_has_effect(membar_bits)) {
2397     // __ tst(Lscratch); executed in delay slot
2398     __ br(Assembler::zero, false, Assembler::pt, exit);
2399     __ delayed()->nop();
2400     volatile_barrier(membar_bits);
2401   }
2402 
2403   __ bind(exit);
2404 }
2405 
2406 void TemplateTable::getfield(int byte_no) {
2407   getfield_or_static(byte_no, false);
2408 }
2409 
2410 void TemplateTable::nofast_getfield(int byte_no) {
2411   getfield_or_static(byte_no, false, may_not_rewrite);
2412 }
2413 
2414 void TemplateTable::getstatic(int byte_no) {
2415   getfield_or_static(byte_no, true);
2416 }
2417 
2418 void TemplateTable::fast_accessfield(TosState state) {
2419   transition(atos, state);
2420   Register Rcache  = G3_scratch;
2421   Register index   = G4_scratch;
2422   Register Roffset = G4_scratch;
2423   Register Rflags  = Rcache;
2424   ByteSize cp_base_offset = ConstantPoolCache::base_offset();
2425 
2426   __ get_cache_and_index_at_bcp(Rcache, index, 1);
2427   jvmti_post_field_access(Rcache, index, /*is_static*/false, /*has_tos*/true);
2428 
2429   __ ld_ptr(Rcache, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), Roffset);
2430 
2431   __ null_check(Otos_i);
2432   __ verify_oop(Otos_i);
2433 
2434   Label exit;
2435 
2436   Assembler::Membar_mask_bits membar_bits =
2437     Assembler::Membar_mask_bits(Assembler::LoadLoad | Assembler::LoadStore);
2438   if (__ membar_has_effect(membar_bits)) {
2439     // Get volatile flag
2440     __ ld_ptr(Rcache, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), Rflags);
2441     __ set((1 << ConstantPoolCacheEntry::is_volatile_shift), Lscratch);
2442   }
2443 
2444   switch (bytecode()) {
2445     case Bytecodes::_fast_bgetfield:
2446       __ ldsb(Otos_i, Roffset, Otos_i);
2447       break;
2448     case Bytecodes::_fast_cgetfield:
2449       __ lduh(Otos_i, Roffset, Otos_i);
2450       break;
2451     case Bytecodes::_fast_sgetfield:
2452       __ ldsh(Otos_i, Roffset, Otos_i);
2453       break;
2454     case Bytecodes::_fast_igetfield:
2455       __ ld(Otos_i, Roffset, Otos_i);
2456       break;
2457     case Bytecodes::_fast_lgetfield:
2458       __ ld_long(Otos_i, Roffset, Otos_l);
2459       break;
2460     case Bytecodes::_fast_fgetfield:
2461       __ ldf(FloatRegisterImpl::S, Otos_i, Roffset, Ftos_f);
2462       break;
2463     case Bytecodes::_fast_dgetfield:
2464       __ ldf(FloatRegisterImpl::D, Otos_i, Roffset, Ftos_d);
2465       break;
2466     case Bytecodes::_fast_agetfield:
2467       do_oop_load(_masm, Otos_i, Roffset, 0, Otos_i, noreg);
2468       break;
2469     default:
2470       ShouldNotReachHere();
2471   }
2472 
2473   if (__ membar_has_effect(membar_bits)) {
2474     __ btst(Lscratch, Rflags);
2475     __ br(Assembler::zero, false, Assembler::pt, exit);
2476     __ delayed()->nop();
2477     volatile_barrier(membar_bits);
2478     __ bind(exit);
2479   }
2480 
2481   if (state == atos) {
2482     __ verify_oop(Otos_i);    // does not blow flags!
2483   }
2484 }
2485 
2486 void TemplateTable::jvmti_post_fast_field_mod() {
2487   if (JvmtiExport::can_post_field_modification()) {
2488     // Check to see if a field modification watch has been set before we take
2489     // the time to call into the VM.
2490     Label done;
2491     AddressLiteral get_field_modification_count_addr(JvmtiExport::get_field_modification_count_addr());
2492     __ load_contents(get_field_modification_count_addr, G4_scratch);
2493     __ cmp_and_br_short(G4_scratch, 0, Assembler::equal, Assembler::pt, done);
2494     __ pop_ptr(G4_scratch);     // copy the object pointer from tos
2495     __ verify_oop(G4_scratch);
2496     __ push_ptr(G4_scratch);    // put the object pointer back on tos
2497     __ get_cache_entry_pointer_at_bcp(G1_scratch, G3_scratch, 1);
2498     // Save tos values before call_VM() clobbers them. Since we have
2499     // to do it for every data type, we use the saved values as the
2500     // jvalue object.
2501     switch (bytecode()) {  // save tos values before call_VM() clobbers them
2502     case Bytecodes::_fast_aputfield: __ push_ptr(Otos_i); break;
2503     case Bytecodes::_fast_bputfield: // fall through
2504     case Bytecodes::_fast_zputfield: // fall through
2505     case Bytecodes::_fast_sputfield: // fall through
2506     case Bytecodes::_fast_cputfield: // fall through
2507     case Bytecodes::_fast_iputfield: __ push_i(Otos_i); break;
2508     case Bytecodes::_fast_dputfield: __ push_d(Ftos_d); break;
2509     case Bytecodes::_fast_fputfield: __ push_f(Ftos_f); break;
2510     // get words in right order for use as jvalue object
2511     case Bytecodes::_fast_lputfield: __ push_l(Otos_l); break;
2512     }
2513     // setup pointer to jvalue object
2514     __ mov(Lesp, G3_scratch);  __ inc(G3_scratch, wordSize);
2515     // G4_scratch:  object pointer
2516     // G1_scratch: cache entry pointer
2517     // G3_scratch: jvalue object on the stack
2518     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::post_field_modification), G4_scratch, G1_scratch, G3_scratch);
2519     switch (bytecode()) {             // restore tos values
2520     case Bytecodes::_fast_aputfield: __ pop_ptr(Otos_i); break;
2521     case Bytecodes::_fast_bputfield: // fall through
2522     case Bytecodes::_fast_zputfield: // fall through
2523     case Bytecodes::_fast_sputfield: // fall through
2524     case Bytecodes::_fast_cputfield: // fall through
2525     case Bytecodes::_fast_iputfield: __ pop_i(Otos_i); break;
2526     case Bytecodes::_fast_dputfield: __ pop_d(Ftos_d); break;
2527     case Bytecodes::_fast_fputfield: __ pop_f(Ftos_f); break;
2528     case Bytecodes::_fast_lputfield: __ pop_l(Otos_l); break;
2529     }
2530     __ bind(done);
2531   }
2532 }
2533 
2534 // The registers Rcache and index expected to be set before call.
2535 // The function may destroy various registers, just not the Rcache and index registers.
2536 void TemplateTable::jvmti_post_field_mod(Register Rcache, Register index, bool is_static) {
2537   ByteSize cp_base_offset = ConstantPoolCache::base_offset();
2538 
2539   if (JvmtiExport::can_post_field_modification()) {
2540     // Check to see if a field modification watch has been set before we take
2541     // the time to call into the VM.
2542     Label Label1;
2543     assert_different_registers(Rcache, index, G1_scratch);
2544     AddressLiteral get_field_modification_count_addr(JvmtiExport::get_field_modification_count_addr());
2545     __ load_contents(get_field_modification_count_addr, G1_scratch);
2546     __ cmp_and_br_short(G1_scratch, 0, Assembler::zero, Assembler::pt, Label1);
2547 
2548     // The Rcache and index registers have been already set.
2549     // This allows to eliminate this call but the Rcache and index
2550     // registers must be correspondingly used after this line.
2551     __ get_cache_and_index_at_bcp(G1_scratch, G4_scratch, 1);
2552 
2553     __ add(G1_scratch, in_bytes(cp_base_offset), G3_scratch);
2554     if (is_static) {
2555       // Life is simple.  Null out the object pointer.
2556       __ clr(G4_scratch);
2557     } else {
2558       Register Rflags = G1_scratch;
2559       // Life is harder. The stack holds the value on top, followed by the
2560       // object.  We don't know the size of the value, though; it could be
2561       // one or two words depending on its type. As a result, we must find
2562       // the type to determine where the object is.
2563 
2564       Label two_word, valsizeknown;
2565       __ ld_ptr(G1_scratch, cp_base_offset + ConstantPoolCacheEntry::flags_offset(), Rflags);
2566       __ mov(Lesp, G4_scratch);
2567       __ srl(Rflags, ConstantPoolCacheEntry::tos_state_shift, Rflags);
2568       // Make sure we don't need to mask Rflags after the above shift
2569       ConstantPoolCacheEntry::verify_tos_state_shift();
2570       __ cmp(Rflags, ltos);
2571       __ br(Assembler::equal, false, Assembler::pt, two_word);
2572       __ delayed()->cmp(Rflags, dtos);
2573       __ br(Assembler::equal, false, Assembler::pt, two_word);
2574       __ delayed()->nop();
2575       __ inc(G4_scratch, Interpreter::expr_offset_in_bytes(1));
2576       __ ba_short(valsizeknown);
2577       __ bind(two_word);
2578 
2579       __ inc(G4_scratch, Interpreter::expr_offset_in_bytes(2));
2580 
2581       __ bind(valsizeknown);
2582       // setup object pointer
2583       __ ld_ptr(G4_scratch, 0, G4_scratch);
2584       __ verify_oop(G4_scratch);
2585     }
2586     // setup pointer to jvalue object
2587     __ mov(Lesp, G1_scratch);  __ inc(G1_scratch, wordSize);
2588     // G4_scratch:  object pointer or NULL if static
2589     // G3_scratch: cache entry pointer
2590     // G1_scratch: jvalue object on the stack
2591     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::post_field_modification),
2592                G4_scratch, G3_scratch, G1_scratch);
2593     __ get_cache_and_index_at_bcp(Rcache, index, 1);
2594     __ bind(Label1);
2595   }
2596 }
2597 
2598 void TemplateTable::pop_and_check_object(Register r) {
2599   __ pop_ptr(r);
2600   __ null_check(r);  // for field access must check obj.
2601   __ verify_oop(r);
2602 }
2603 
2604 void TemplateTable::putfield_or_static(int byte_no, bool is_static, RewriteControl rc) {
2605   transition(vtos, vtos);
2606   Register Rcache = G3_scratch;
2607   Register index  = G4_scratch;
2608   Register Rclass = Rcache;
2609   Register Roffset= G4_scratch;
2610   Register Rflags = G1_scratch;
2611   ByteSize cp_base_offset = ConstantPoolCache::base_offset();
2612 
2613   resolve_cache_and_index(byte_no, Rcache, index, sizeof(u2));
2614   jvmti_post_field_mod(Rcache, index, is_static);
2615   load_field_cp_cache_entry(Rclass, Rcache, index, Roffset, Rflags, is_static);
2616 
2617   Assembler::Membar_mask_bits read_bits =
2618     Assembler::Membar_mask_bits(Assembler::LoadStore | Assembler::StoreStore);
2619   Assembler::Membar_mask_bits write_bits = Assembler::StoreLoad;
2620 
2621   Label notVolatile, checkVolatile, exit;
2622   if (__ membar_has_effect(read_bits) || __ membar_has_effect(write_bits)) {
2623     __ set((1 << ConstantPoolCacheEntry::is_volatile_shift), Lscratch);
2624     __ and3(Rflags, Lscratch, Lscratch);
2625 
2626     if (__ membar_has_effect(read_bits)) {
2627       __ cmp_and_br_short(Lscratch, 0, Assembler::equal, Assembler::pt, notVolatile);
2628       volatile_barrier(read_bits);
2629       __ bind(notVolatile);
2630     }
2631   }
2632 
2633   __ srl(Rflags, ConstantPoolCacheEntry::tos_state_shift, Rflags);
2634   // Make sure we don't need to mask Rflags after the above shift
2635   ConstantPoolCacheEntry::verify_tos_state_shift();
2636 
2637   // compute field type
2638   Label notInt, notShort, notChar, notObj, notByte, notBool, notLong, notFloat;
2639 
2640   if (is_static) {
2641     // putstatic with object type most likely, check that first
2642     __ cmp(Rflags, atos);
2643     __ br(Assembler::notEqual, false, Assembler::pt, notObj);
2644     __ delayed()->cmp(Rflags, itos);
2645 
2646     // atos
2647     {
2648       __ pop_ptr();
2649       __ verify_oop(Otos_i);
2650       do_oop_store(_masm, Rclass, Roffset, 0, Otos_i, G1_scratch);
2651       __ ba(checkVolatile);
2652       __ delayed()->tst(Lscratch);
2653     }
2654 
2655     __ bind(notObj);
2656     // cmp(Rflags, itos);
2657     __ br(Assembler::notEqual, false, Assembler::pt, notInt);
2658     __ delayed()->cmp(Rflags, btos);
2659 
2660     // itos
2661     {
2662       __ pop_i();
2663       __ st(Otos_i, Rclass, Roffset);
2664       __ ba(checkVolatile);
2665       __ delayed()->tst(Lscratch);
2666     }
2667 
2668     __ bind(notInt);
2669   } else {
2670     // putfield with int type most likely, check that first
2671     __ cmp(Rflags, itos);
2672     __ br(Assembler::notEqual, false, Assembler::pt, notInt);
2673     __ delayed()->cmp(Rflags, atos);
2674 
2675     // itos
2676     {
2677       __ pop_i();
2678       pop_and_check_object(Rclass);
2679       __ st(Otos_i, Rclass, Roffset);
2680       if (rc == may_rewrite) patch_bytecode(Bytecodes::_fast_iputfield, G3_scratch, G4_scratch, true, byte_no);
2681       __ ba(checkVolatile);
2682       __ delayed()->tst(Lscratch);
2683     }
2684 
2685     __ bind(notInt);
2686     // cmp(Rflags, atos);
2687     __ br(Assembler::notEqual, false, Assembler::pt, notObj);
2688     __ delayed()->cmp(Rflags, btos);
2689 
2690     // atos
2691     {
2692       __ pop_ptr();
2693       pop_and_check_object(Rclass);
2694       __ verify_oop(Otos_i);
2695       do_oop_store(_masm, Rclass, Roffset, 0, Otos_i, G1_scratch);
2696       if (rc == may_rewrite) patch_bytecode(Bytecodes::_fast_aputfield, G3_scratch, G4_scratch, true, byte_no);
2697       __ ba(checkVolatile);
2698       __ delayed()->tst(Lscratch);
2699     }
2700 
2701     __ bind(notObj);
2702   }
2703 
2704   // cmp(Rflags, btos);
2705   __ br(Assembler::notEqual, false, Assembler::pt, notByte);
2706   __ delayed()->cmp(Rflags, ztos);
2707 
2708   // btos
2709   {
2710     __ pop_i();
2711     if (!is_static) pop_and_check_object(Rclass);
2712     __ stb(Otos_i, Rclass, Roffset);
2713     if (!is_static && rc == may_rewrite) {
2714       patch_bytecode(Bytecodes::_fast_bputfield, G3_scratch, G4_scratch, true, byte_no);
2715     }
2716     __ ba(checkVolatile);
2717     __ delayed()->tst(Lscratch);
2718   }
2719 
2720   __ bind(notByte);
2721 
2722   // cmp(Rflags, btos);
2723   __ br(Assembler::notEqual, false, Assembler::pt, notBool);
2724   __ delayed()->cmp(Rflags, ltos);
2725 
2726   // ztos
2727   {
2728     __ pop_i();
2729     if (!is_static) pop_and_check_object(Rclass);
2730     __ and3(Otos_i, 1, Otos_i);
2731     __ stb(Otos_i, Rclass, Roffset);
2732     if (!is_static && rc == may_rewrite) {
2733       patch_bytecode(Bytecodes::_fast_zputfield, G3_scratch, G4_scratch, true, byte_no);
2734     }
2735     __ ba(checkVolatile);
2736     __ delayed()->tst(Lscratch);
2737   }
2738 
2739   __ bind(notBool);
2740   // cmp(Rflags, ltos);
2741   __ br(Assembler::notEqual, false, Assembler::pt, notLong);
2742   __ delayed()->cmp(Rflags, ctos);
2743 
2744   // ltos
2745   {
2746     __ pop_l();
2747     if (!is_static) pop_and_check_object(Rclass);
2748     __ st_long(Otos_l, Rclass, Roffset);
2749     if (!is_static && rc == may_rewrite) {
2750       patch_bytecode(Bytecodes::_fast_lputfield, G3_scratch, G4_scratch, true, byte_no);
2751     }
2752     __ ba(checkVolatile);
2753     __ delayed()->tst(Lscratch);
2754   }
2755 
2756   __ bind(notLong);
2757   // cmp(Rflags, ctos);
2758   __ br(Assembler::notEqual, false, Assembler::pt, notChar);
2759   __ delayed()->cmp(Rflags, stos);
2760 
2761   // ctos (char)
2762   {
2763     __ pop_i();
2764     if (!is_static) pop_and_check_object(Rclass);
2765     __ sth(Otos_i, Rclass, Roffset);
2766     if (!is_static && rc == may_rewrite) {
2767       patch_bytecode(Bytecodes::_fast_cputfield, G3_scratch, G4_scratch, true, byte_no);
2768     }
2769     __ ba(checkVolatile);
2770     __ delayed()->tst(Lscratch);
2771   }
2772 
2773   __ bind(notChar);
2774   // cmp(Rflags, stos);
2775   __ br(Assembler::notEqual, false, Assembler::pt, notShort);
2776   __ delayed()->cmp(Rflags, ftos);
2777 
2778   // stos (short)
2779   {
2780     __ pop_i();
2781     if (!is_static) pop_and_check_object(Rclass);
2782     __ sth(Otos_i, Rclass, Roffset);
2783     if (!is_static && rc == may_rewrite) {
2784       patch_bytecode(Bytecodes::_fast_sputfield, G3_scratch, G4_scratch, true, byte_no);
2785     }
2786     __ ba(checkVolatile);
2787     __ delayed()->tst(Lscratch);
2788   }
2789 
2790   __ bind(notShort);
2791   // cmp(Rflags, ftos);
2792   __ br(Assembler::notZero, false, Assembler::pt, notFloat);
2793   __ delayed()->nop();
2794 
2795   // ftos
2796   {
2797     __ pop_f();
2798     if (!is_static) pop_and_check_object(Rclass);
2799     __ stf(FloatRegisterImpl::S, Ftos_f, Rclass, Roffset);
2800     if (!is_static && rc == may_rewrite) {
2801       patch_bytecode(Bytecodes::_fast_fputfield, G3_scratch, G4_scratch, true, byte_no);
2802     }
2803     __ ba(checkVolatile);
2804     __ delayed()->tst(Lscratch);
2805   }
2806 
2807   __ bind(notFloat);
2808 
2809   // dtos
2810   {
2811     __ pop_d();
2812     if (!is_static) pop_and_check_object(Rclass);
2813     __ stf(FloatRegisterImpl::D, Ftos_d, Rclass, Roffset);
2814     if (!is_static && rc == may_rewrite) {
2815       patch_bytecode(Bytecodes::_fast_dputfield, G3_scratch, G4_scratch, true, byte_no);
2816     }
2817   }
2818 
2819   __ bind(checkVolatile);
2820   __ tst(Lscratch);
2821 
2822   if (__ membar_has_effect(write_bits)) {
2823     // __ tst(Lscratch); in delay slot
2824     __ br(Assembler::zero, false, Assembler::pt, exit);
2825     __ delayed()->nop();
2826     volatile_barrier(Assembler::StoreLoad);
2827     __ bind(exit);
2828   }
2829 }
2830 
2831 void TemplateTable::fast_storefield(TosState state) {
2832   transition(state, vtos);
2833   Register Rcache = G3_scratch;
2834   Register Rclass = Rcache;
2835   Register Roffset= G4_scratch;
2836   Register Rflags = G1_scratch;
2837   ByteSize cp_base_offset = ConstantPoolCache::base_offset();
2838 
2839   jvmti_post_fast_field_mod();
2840 
2841   __ get_cache_and_index_at_bcp(Rcache, G4_scratch, 1);
2842 
2843   Assembler::Membar_mask_bits read_bits =
2844     Assembler::Membar_mask_bits(Assembler::LoadStore | Assembler::StoreStore);
2845   Assembler::Membar_mask_bits write_bits = Assembler::StoreLoad;
2846 
2847   Label notVolatile, checkVolatile, exit;
2848   if (__ membar_has_effect(read_bits) || __ membar_has_effect(write_bits)) {
2849     __ ld_ptr(Rcache, cp_base_offset + ConstantPoolCacheEntry::flags_offset(), Rflags);
2850     __ set((1 << ConstantPoolCacheEntry::is_volatile_shift), Lscratch);
2851     __ and3(Rflags, Lscratch, Lscratch);
2852     if (__ membar_has_effect(read_bits)) {
2853       __ cmp_and_br_short(Lscratch, 0, Assembler::equal, Assembler::pt, notVolatile);
2854       volatile_barrier(read_bits);
2855       __ bind(notVolatile);
2856     }
2857   }
2858 
2859   __ ld_ptr(Rcache, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), Roffset);
2860   pop_and_check_object(Rclass);
2861 
2862   switch (bytecode()) {
2863     case Bytecodes::_fast_zputfield: __ and3(Otos_i, 1, Otos_i);  // fall through to bputfield
2864     case Bytecodes::_fast_bputfield: __ stb(Otos_i, Rclass, Roffset); break;
2865     case Bytecodes::_fast_cputfield: /* fall through */
2866     case Bytecodes::_fast_sputfield: __ sth(Otos_i, Rclass, Roffset); break;
2867     case Bytecodes::_fast_iputfield: __ st(Otos_i, Rclass, Roffset);  break;
2868     case Bytecodes::_fast_lputfield: __ st_long(Otos_l, Rclass, Roffset); break;
2869     case Bytecodes::_fast_fputfield:
2870       __ stf(FloatRegisterImpl::S, Ftos_f, Rclass, Roffset);
2871       break;
2872     case Bytecodes::_fast_dputfield:
2873       __ stf(FloatRegisterImpl::D, Ftos_d, Rclass, Roffset);
2874       break;
2875     case Bytecodes::_fast_aputfield:
2876       do_oop_store(_masm, Rclass, Roffset, 0, Otos_i, G1_scratch);
2877       break;
2878     default:
2879       ShouldNotReachHere();
2880   }
2881 
2882   if (__ membar_has_effect(write_bits)) {
2883     __ cmp_and_br_short(Lscratch, 0, Assembler::equal, Assembler::pt, exit);
2884     volatile_barrier(Assembler::StoreLoad);
2885     __ bind(exit);
2886   }
2887 }
2888 
2889 void TemplateTable::putfield(int byte_no) {
2890   putfield_or_static(byte_no, false);
2891 }
2892 
2893 void TemplateTable::nofast_putfield(int byte_no) {
2894   putfield_or_static(byte_no, false, may_not_rewrite);
2895 }
2896 
2897 void TemplateTable::putstatic(int byte_no) {
2898   putfield_or_static(byte_no, true);
2899 }
2900 
2901 void TemplateTable::fast_xaccess(TosState state) {
2902   transition(vtos, state);
2903   Register Rcache = G3_scratch;
2904   Register Roffset = G4_scratch;
2905   Register Rflags  = G4_scratch;
2906   Register Rreceiver = Lscratch;
2907 
2908   __ ld_ptr(Llocals, 0, Rreceiver);
2909 
2910   // access constant pool cache  (is resolved)
2911   __ get_cache_and_index_at_bcp(Rcache, G4_scratch, 2);
2912   __ ld_ptr(Rcache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::f2_offset(), Roffset);
2913   __ add(Lbcp, 1, Lbcp);       // needed to report exception at the correct bcp
2914 
2915   __ verify_oop(Rreceiver);
2916   __ null_check(Rreceiver);
2917   if (state == atos) {
2918     do_oop_load(_masm, Rreceiver, Roffset, 0, Otos_i, noreg);
2919   } else if (state == itos) {
2920     __ ld (Rreceiver, Roffset, Otos_i) ;
2921   } else if (state == ftos) {
2922     __ ldf(FloatRegisterImpl::S, Rreceiver, Roffset, Ftos_f);
2923   } else {
2924     ShouldNotReachHere();
2925   }
2926 
2927   Assembler::Membar_mask_bits membar_bits =
2928     Assembler::Membar_mask_bits(Assembler::LoadLoad | Assembler::LoadStore);
2929   if (__ membar_has_effect(membar_bits)) {
2930 
2931     // Get is_volatile value in Rflags and check if membar is needed
2932     __ ld_ptr(Rcache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset(), Rflags);
2933 
2934     // Test volatile
2935     Label notVolatile;
2936     __ set((1 << ConstantPoolCacheEntry::is_volatile_shift), Lscratch);
2937     __ btst(Rflags, Lscratch);
2938     __ br(Assembler::zero, false, Assembler::pt, notVolatile);
2939     __ delayed()->nop();
2940     volatile_barrier(membar_bits);
2941     __ bind(notVolatile);
2942   }
2943 
2944   __ interp_verify_oop(Otos_i, state, __FILE__, __LINE__);
2945   __ sub(Lbcp, 1, Lbcp);
2946 }
2947 
2948 //----------------------------------------------------------------------------------------------------
2949 // Calls
2950 
2951 void TemplateTable::count_calls(Register method, Register temp) {
2952   // implemented elsewhere
2953   ShouldNotReachHere();
2954 }
2955 
2956 void TemplateTable::prepare_invoke(int byte_no,
2957                                    Register method,  // linked method (or i-klass)
2958                                    Register ra,      // return address
2959                                    Register index,   // itable index, MethodType, etc.
2960                                    Register recv,    // if caller wants to see it
2961                                    Register flags    // if caller wants to test it
2962                                    ) {
2963   // determine flags
2964   const Bytecodes::Code code = bytecode();
2965   const bool is_invokeinterface  = code == Bytecodes::_invokeinterface;
2966   const bool is_invokedynamic    = code == Bytecodes::_invokedynamic;
2967   const bool is_invokehandle     = code == Bytecodes::_invokehandle;
2968   const bool is_invokevirtual    = code == Bytecodes::_invokevirtual;
2969   const bool is_invokespecial    = code == Bytecodes::_invokespecial;
2970   const bool load_receiver       = (recv != noreg);
2971   assert(load_receiver == (code != Bytecodes::_invokestatic && code != Bytecodes::_invokedynamic), "");
2972   assert(recv  == noreg || recv  == O0, "");
2973   assert(flags == noreg || flags == O1, "");
2974 
2975   // setup registers & access constant pool cache
2976   if (recv  == noreg)  recv  = O0;
2977   if (flags == noreg)  flags = O1;
2978   const Register temp = O2;
2979   assert_different_registers(method, ra, index, recv, flags, temp);
2980 
2981   load_invoke_cp_cache_entry(byte_no, method, index, flags, is_invokevirtual, false, is_invokedynamic);
2982 
2983   __ mov(SP, O5_savedSP);  // record SP that we wanted the callee to restore
2984 
2985   // maybe push appendix to arguments
2986   if (is_invokedynamic || is_invokehandle) {
2987     Label L_no_push;
2988     __ set((1 << ConstantPoolCacheEntry::has_appendix_shift), temp);
2989     __ btst(flags, temp);
2990     __ br(Assembler::zero, false, Assembler::pt, L_no_push);
2991     __ delayed()->nop();
2992     // Push the appendix as a trailing parameter.
2993     // This must be done before we get the receiver,
2994     // since the parameter_size includes it.
2995     __ load_resolved_reference_at_index(temp, index, /*tmp*/recv);
2996     __ verify_oop(temp);
2997     __ push_ptr(temp);  // push appendix (MethodType, CallSite, etc.)
2998     __ bind(L_no_push);
2999   }
3000 
3001   // load receiver if needed (after appendix is pushed so parameter size is correct)
3002   if (load_receiver) {
3003     __ and3(flags, ConstantPoolCacheEntry::parameter_size_mask, temp);  // get parameter size
3004     __ load_receiver(temp, recv);  //  __ argument_address uses Gargs but we need Lesp
3005     __ verify_oop(recv);
3006   }
3007 
3008   // compute return type
3009   __ srl(flags, ConstantPoolCacheEntry::tos_state_shift, ra);
3010   // Make sure we don't need to mask flags after the above shift
3011   ConstantPoolCacheEntry::verify_tos_state_shift();
3012   // load return address
3013   {
3014     const address table_addr = (address) Interpreter::invoke_return_entry_table_for(code);
3015     AddressLiteral table(table_addr);
3016     __ set(table, temp);
3017     __ sll(ra, LogBytesPerWord, ra);
3018     __ ld_ptr(Address(temp, ra), ra);
3019   }
3020 }
3021 
3022 
3023 void TemplateTable::generate_vtable_call(Register Rrecv, Register Rindex, Register Rret) {
3024   Register Rtemp = G4_scratch;
3025   Register Rcall = Rindex;
3026   assert_different_registers(Rcall, G5_method, Gargs, Rret);
3027 
3028   // get target Method* & entry point
3029   __ lookup_virtual_method(Rrecv, Rindex, G5_method);
3030   __ profile_arguments_type(G5_method, Rcall, Gargs, true);
3031   __ call_from_interpreter(Rcall, Gargs, Rret);
3032 }
3033 
3034 void TemplateTable::invokevirtual(int byte_no) {
3035   transition(vtos, vtos);
3036   assert(byte_no == f2_byte, "use this argument");
3037 
3038   Register Rscratch = G3_scratch;
3039   Register Rtemp    = G4_scratch;
3040   Register Rret     = Lscratch;
3041   Register O0_recv  = O0;
3042   Label notFinal;
3043 
3044   load_invoke_cp_cache_entry(byte_no, G5_method, noreg, Rret, true, false, false);
3045   __ mov(SP, O5_savedSP); // record SP that we wanted the callee to restore
3046 
3047   // Check for vfinal
3048   __ set((1 << ConstantPoolCacheEntry::is_vfinal_shift), G4_scratch);
3049   __ btst(Rret, G4_scratch);
3050   __ br(Assembler::zero, false, Assembler::pt, notFinal);
3051   __ delayed()->and3(Rret, 0xFF, G4_scratch);      // gets number of parameters
3052 
3053   if (RewriteBytecodes && !UseSharedSpaces && !DumpSharedSpaces) {
3054     patch_bytecode(Bytecodes::_fast_invokevfinal, Rscratch, Rtemp);
3055   }
3056 
3057   invokevfinal_helper(Rscratch, Rret);
3058 
3059   __ bind(notFinal);
3060 
3061   __ mov(G5_method, Rscratch);  // better scratch register
3062   __ load_receiver(G4_scratch, O0_recv);  // gets receiverOop
3063   // receiver is in O0_recv
3064   __ verify_oop(O0_recv);
3065 
3066   // get return address
3067   AddressLiteral table(Interpreter::invoke_return_entry_table());
3068   __ set(table, Rtemp);
3069   __ srl(Rret, ConstantPoolCacheEntry::tos_state_shift, Rret);          // get return type
3070   // Make sure we don't need to mask Rret after the above shift
3071   ConstantPoolCacheEntry::verify_tos_state_shift();
3072   __ sll(Rret,  LogBytesPerWord, Rret);
3073   __ ld_ptr(Rtemp, Rret, Rret);         // get return address
3074 
3075   // get receiver klass
3076   __ null_check(O0_recv, oopDesc::klass_offset_in_bytes());
3077   __ load_klass(O0_recv, O0_recv);
3078   __ verify_klass_ptr(O0_recv);
3079 
3080   __ profile_virtual_call(O0_recv, O4);
3081 
3082   generate_vtable_call(O0_recv, Rscratch, Rret);
3083 }
3084 
3085 void TemplateTable::fast_invokevfinal(int byte_no) {
3086   transition(vtos, vtos);
3087   assert(byte_no == f2_byte, "use this argument");
3088 
3089   load_invoke_cp_cache_entry(byte_no, G5_method, noreg, Lscratch, true,
3090                              /*is_invokevfinal*/true, false);
3091   __ mov(SP, O5_savedSP); // record SP that we wanted the callee to restore
3092   invokevfinal_helper(G3_scratch, Lscratch);
3093 }
3094 
3095 void TemplateTable::invokevfinal_helper(Register Rscratch, Register Rret) {
3096   Register Rtemp = G4_scratch;
3097 
3098   // Load receiver from stack slot
3099   __ ld_ptr(G5_method, in_bytes(Method::const_offset()), G4_scratch);
3100   __ lduh(G4_scratch, in_bytes(ConstMethod::size_of_parameters_offset()), G4_scratch);
3101   __ load_receiver(G4_scratch, O0);
3102 
3103   // receiver NULL check
3104   __ null_check(O0);
3105 
3106   __ profile_final_call(O4);
3107   __ profile_arguments_type(G5_method, Rscratch, Gargs, true);
3108 
3109   // get return address
3110   AddressLiteral table(Interpreter::invoke_return_entry_table());
3111   __ set(table, Rtemp);
3112   __ srl(Rret, ConstantPoolCacheEntry::tos_state_shift, Rret);          // get return type
3113   // Make sure we don't need to mask Rret after the above shift
3114   ConstantPoolCacheEntry::verify_tos_state_shift();
3115   __ sll(Rret,  LogBytesPerWord, Rret);
3116   __ ld_ptr(Rtemp, Rret, Rret);         // get return address
3117 
3118 
3119   // do the call
3120   __ call_from_interpreter(Rscratch, Gargs, Rret);
3121 }
3122 
3123 
3124 void TemplateTable::invokespecial(int byte_no) {
3125   transition(vtos, vtos);
3126   assert(byte_no == f1_byte, "use this argument");
3127 
3128   const Register Rret     = Lscratch;
3129   const Register O0_recv  = O0;
3130   const Register Rscratch = G3_scratch;
3131 
3132   prepare_invoke(byte_no, G5_method, Rret, noreg, O0_recv);  // get receiver also for null check
3133   __ null_check(O0_recv);
3134 
3135   // do the call
3136   __ profile_call(O4);
3137   __ profile_arguments_type(G5_method, Rscratch, Gargs, false);
3138   __ call_from_interpreter(Rscratch, Gargs, Rret);
3139 }
3140 
3141 
3142 void TemplateTable::invokestatic(int byte_no) {
3143   transition(vtos, vtos);
3144   assert(byte_no == f1_byte, "use this argument");
3145 
3146   const Register Rret     = Lscratch;
3147   const Register Rscratch = G3_scratch;
3148 
3149   prepare_invoke(byte_no, G5_method, Rret);  // get f1 Method*
3150 
3151   // do the call
3152   __ profile_call(O4);
3153   __ profile_arguments_type(G5_method, Rscratch, Gargs, false);
3154   __ call_from_interpreter(Rscratch, Gargs, Rret);
3155 }
3156 
3157 void TemplateTable::invokeinterface_object_method(Register RKlass,
3158                                                   Register Rcall,
3159                                                   Register Rret,
3160                                                   Register Rflags) {
3161   Register Rscratch = G4_scratch;
3162   Register Rindex = Lscratch;
3163 
3164   assert_different_registers(Rscratch, Rindex, Rret);
3165 
3166   Label notFinal;
3167 
3168   // Check for vfinal
3169   __ set((1 << ConstantPoolCacheEntry::is_vfinal_shift), Rscratch);
3170   __ btst(Rflags, Rscratch);
3171   __ br(Assembler::zero, false, Assembler::pt, notFinal);
3172   __ delayed()->nop();
3173 
3174   __ profile_final_call(O4);
3175 
3176   // do the call - the index (f2) contains the Method*
3177   assert_different_registers(G5_method, Gargs, Rcall);
3178   __ mov(Rindex, G5_method);
3179   __ profile_arguments_type(G5_method, Rcall, Gargs, true);
3180   __ call_from_interpreter(Rcall, Gargs, Rret);
3181   __ bind(notFinal);
3182 
3183   __ profile_virtual_call(RKlass, O4);
3184   generate_vtable_call(RKlass, Rindex, Rret);
3185 }
3186 
3187 
3188 void TemplateTable::invokeinterface(int byte_no) {
3189   transition(vtos, vtos);
3190   assert(byte_no == f1_byte, "use this argument");
3191 
3192   const Register Rinterface  = G1_scratch;
3193   const Register Rmethod     = Lscratch;
3194   const Register Rret        = G3_scratch;
3195   const Register O0_recv     = O0;
3196   const Register O1_flags    = O1;
3197   const Register O2_Klass    = O2;
3198   const Register Rscratch    = G4_scratch;
3199   assert_different_registers(Rscratch, G5_method);
3200 
3201   prepare_invoke(byte_no, Rinterface, Rret, Rmethod, O0_recv, O1_flags);
3202 
3203   // First check for Object case, then private interface method,
3204   // then regular interface method.
3205 
3206   // get receiver klass - this is also a null check
3207   __ null_check(O0_recv, oopDesc::klass_offset_in_bytes());
3208   __ load_klass(O0_recv, O2_Klass);
3209 
3210   // Special case of invokeinterface called for virtual method of
3211   // java.lang.Object.  See cpCache.cpp for details.
3212   Label notObjectMethod;
3213   __ set((1 << ConstantPoolCacheEntry::is_forced_virtual_shift), Rscratch);
3214   __ btst(O1_flags, Rscratch);
3215   __ br(Assembler::zero, false, Assembler::pt, notObjectMethod);
3216   __ delayed()->nop();
3217 
3218   invokeinterface_object_method(O2_Klass, Rinterface, Rret, O1_flags);
3219 
3220   __ bind(notObjectMethod);
3221 
3222   Label L_no_such_interface;
3223 
3224   // Check for private method invocation - indicated by vfinal
3225   Label notVFinal;
3226   {
3227     __ set((1 << ConstantPoolCacheEntry::is_vfinal_shift), Rscratch);
3228     __ btst(O1_flags, Rscratch);
3229     __ br(Assembler::zero, false, Assembler::pt, notVFinal);
3230     __ delayed()->nop();
3231 
3232     Label subtype;
3233     Register Rtemp = O1_flags;
3234     __ check_klass_subtype(O2_Klass, Rinterface, Rscratch, Rtemp, subtype);
3235     // If we get here the typecheck failed
3236     __ ba(L_no_such_interface);
3237     __ delayed()->nop();
3238     __ bind(subtype);
3239 
3240     // do the call
3241     Register Rcall = Rinterface;
3242     __ mov(Rmethod, G5_method);
3243     assert_different_registers(Rcall, G5_method, Gargs, Rret);
3244 
3245     __ profile_arguments_type(G5_method, Rcall, Gargs, true);
3246     __ profile_final_call(Rscratch);
3247     __ call_from_interpreter(Rcall, Gargs, Rret);
3248   }
3249   __ bind(notVFinal);
3250 
3251   Register Rtemp = O1_flags;
3252 
3253   // Receiver subtype check against REFC.
3254   __ lookup_interface_method(// inputs: rec. class, interface, itable index
3255                              O2_Klass, Rinterface, noreg,
3256                              // outputs: temp reg1, temp reg2, temp reg3
3257                              G5_method, Rscratch, Rtemp,
3258                              L_no_such_interface,
3259                              /*return_method=*/false);
3260 
3261   __ profile_virtual_call(O2_Klass, O4);
3262 
3263   //
3264   // find entry point to call
3265   //
3266 
3267   // Get declaring interface class from method
3268   __ ld_ptr(Rmethod, Method::const_offset(), Rinterface);
3269   __ ld_ptr(Rinterface, ConstMethod::constants_offset(), Rinterface);
3270   __ ld_ptr(Rinterface, ConstantPool::pool_holder_offset_in_bytes(), Rinterface);
3271 
3272   // Get itable index from method
3273   const Register Rindex = G5_method;
3274   __ ld(Rmethod, Method::itable_index_offset(), Rindex);
3275   __ sub(Rindex, Method::itable_index_max, Rindex);
3276   __ neg(Rindex);
3277 
3278   // Preserve O2_Klass for throw_AbstractMethodErrorVerbose
3279   __ mov(O2_Klass, O4);
3280   __ lookup_interface_method(// inputs: rec. class, interface, itable index
3281                              O4, Rinterface, Rindex,
3282                              // outputs: method, scan temp reg, temp reg
3283                              G5_method, Rscratch, Rtemp,
3284                              L_no_such_interface);
3285 
3286   // Check for abstract method error.
3287   {
3288     Label ok;
3289     __ br_notnull_short(G5_method, Assembler::pt, ok);
3290     // Pass arguments for generating a verbose error message.
3291     call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodErrorVerbose),
3292             O2_Klass, Rmethod);
3293     __ should_not_reach_here();
3294     __ bind(ok);
3295   }
3296 
3297   Register Rcall = Rinterface;
3298   assert_different_registers(Rcall, G5_method, Gargs, Rret);
3299 
3300   __ profile_arguments_type(G5_method, Rcall, Gargs, true);
3301   __ call_from_interpreter(Rcall, Gargs, Rret);
3302 
3303   __ bind(L_no_such_interface);
3304   // Pass arguments for generating a verbose error message.
3305   call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_IncompatibleClassChangeErrorVerbose),
3306           O2_Klass, Rinterface);
3307   __ should_not_reach_here();
3308 }
3309 
3310 void TemplateTable::invokehandle(int byte_no) {
3311   transition(vtos, vtos);
3312   assert(byte_no == f1_byte, "use this argument");
3313 
3314   const Register Rret       = Lscratch;
3315   const Register G4_mtype   = G4_scratch;
3316   const Register O0_recv    = O0;
3317   const Register Rscratch   = G3_scratch;
3318 
3319   prepare_invoke(byte_no, G5_method, Rret, G4_mtype, O0_recv);
3320   __ null_check(O0_recv);
3321 
3322   // G4: MethodType object (from cpool->resolved_references[f1], if necessary)
3323   // G5: MH.invokeExact_MT method (from f2)
3324 
3325   // Note:  G4_mtype is already pushed (if necessary) by prepare_invoke
3326 
3327   // do the call
3328   __ verify_oop(G4_mtype);
3329   __ profile_final_call(O4);  // FIXME: profile the LambdaForm also
3330   __ profile_arguments_type(G5_method, Rscratch, Gargs, true);
3331   __ call_from_interpreter(Rscratch, Gargs, Rret);
3332 }
3333 
3334 
3335 void TemplateTable::invokedynamic(int byte_no) {
3336   transition(vtos, vtos);
3337   assert(byte_no == f1_byte, "use this argument");
3338 
3339   const Register Rret        = Lscratch;
3340   const Register G4_callsite = G4_scratch;
3341   const Register Rscratch    = G3_scratch;
3342 
3343   prepare_invoke(byte_no, G5_method, Rret, G4_callsite);
3344 
3345   // G4: CallSite object (from cpool->resolved_references[f1])
3346   // G5: MH.linkToCallSite method (from f2)
3347 
3348   // Note:  G4_callsite is already pushed by prepare_invoke
3349 
3350   // %%% should make a type profile for any invokedynamic that takes a ref argument
3351   // profile this call
3352   __ profile_call(O4);
3353 
3354   // do the call
3355   __ verify_oop(G4_callsite);
3356   __ profile_arguments_type(G5_method, Rscratch, Gargs, false);
3357   __ call_from_interpreter(Rscratch, Gargs, Rret);
3358 }
3359 
3360 
3361 //----------------------------------------------------------------------------------------------------
3362 // Allocation
3363 
3364 void TemplateTable::_new() {
3365   transition(vtos, atos);
3366 
3367   Label slow_case;
3368   Label done;
3369   Label initialize_header;
3370   Label initialize_object;  // including clearing the fields
3371 
3372   Register RallocatedObject = Otos_i;
3373   Register RinstanceKlass = O1;
3374   Register Roffset = O3;
3375   Register Rscratch = O4;
3376 
3377   __ get_2_byte_integer_at_bcp(1, Rscratch, Roffset, InterpreterMacroAssembler::Unsigned);
3378   __ get_cpool_and_tags(Rscratch, G3_scratch);
3379   // make sure the class we're about to instantiate has been resolved
3380   // This is done before loading InstanceKlass to be consistent with the order
3381   // how Constant Pool is updated (see ConstantPool::klass_at_put)
3382   __ add(G3_scratch, Array<u1>::base_offset_in_bytes(), G3_scratch);
3383   __ ldub(G3_scratch, Roffset, G3_scratch);
3384   __ cmp(G3_scratch, JVM_CONSTANT_Class);
3385   __ br(Assembler::notEqual, false, Assembler::pn, slow_case);
3386   __ delayed()->sll(Roffset, LogBytesPerWord, Roffset);
3387   // get InstanceKlass
3388   __ load_resolved_klass_at_offset(Rscratch, Roffset, RinstanceKlass);
3389 
3390   // make sure klass is fully initialized:
3391   __ ldub(RinstanceKlass, in_bytes(InstanceKlass::init_state_offset()), G3_scratch);
3392   __ cmp(G3_scratch, InstanceKlass::fully_initialized);
3393   __ br(Assembler::notEqual, false, Assembler::pn, slow_case);
3394   __ delayed()->ld(RinstanceKlass, in_bytes(Klass::layout_helper_offset()), Roffset);
3395 
3396   // get instance_size in InstanceKlass (already aligned)
3397   //__ ld(RinstanceKlass, in_bytes(Klass::layout_helper_offset()), Roffset);
3398 
3399   // make sure klass does not have has_finalizer, or is abstract, or interface or java/lang/Class
3400   __ btst(Klass::_lh_instance_slow_path_bit, Roffset);
3401   __ br(Assembler::notZero, false, Assembler::pn, slow_case);
3402   __ delayed()->nop();
3403 
3404   // Allocate the instance:
3405   //  If TLAB is enabled:
3406   //    Try to allocate in the TLAB.
3407   //    If fails, go to the slow path.
3408   //  Else If inline contiguous allocations are enabled:
3409   //    Try to allocate in eden.
3410   //    If fails due to heap end, go to slow path.
3411   //
3412   //  If TLAB is enabled OR inline contiguous is enabled:
3413   //    Initialize the allocation.
3414   //    Exit.
3415   //
3416   //  Go to slow path.
3417 
3418   const bool allow_shared_alloc =
3419     Universe::heap()->supports_inline_contig_alloc();
3420 
3421   if(UseTLAB) {
3422     Register RoldTopValue = RallocatedObject;
3423     Register RtlabWasteLimitValue = G3_scratch;
3424     Register RnewTopValue = G1_scratch;
3425     Register RendValue = Rscratch;
3426     Register RfreeValue = RnewTopValue;
3427 
3428     // check if we can allocate in the TLAB
3429     __ ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), RoldTopValue); // sets up RalocatedObject
3430     __ ld_ptr(G2_thread, in_bytes(JavaThread::tlab_end_offset()), RendValue);
3431     __ add(RoldTopValue, Roffset, RnewTopValue);
3432 
3433     // if there is enough space, we do not CAS and do not clear
3434     __ cmp(RnewTopValue, RendValue);
3435     if(ZeroTLAB) {
3436       // the fields have already been cleared
3437       __ brx(Assembler::lessEqualUnsigned, true, Assembler::pt, initialize_header);
3438     } else {
3439       // initialize both the header and fields
3440       __ brx(Assembler::lessEqualUnsigned, true, Assembler::pt, initialize_object);
3441     }
3442     __ delayed()->st_ptr(RnewTopValue, G2_thread, in_bytes(JavaThread::tlab_top_offset()));
3443 
3444     // Allocation does not fit in the TLAB.
3445     __ ba_short(slow_case);
3446   } else {
3447     // Allocation in the shared Eden
3448     if (allow_shared_alloc) {
3449       Register RoldTopValue = G1_scratch;
3450       Register RtopAddr = G3_scratch;
3451       Register RnewTopValue = RallocatedObject;
3452       Register RendValue = Rscratch;
3453 
3454       __ set((intptr_t)Universe::heap()->top_addr(), RtopAddr);
3455 
3456       Label retry;
3457       __ bind(retry);
3458       __ set((intptr_t)Universe::heap()->end_addr(), RendValue);
3459       __ ld_ptr(RendValue, 0, RendValue);
3460       __ ld_ptr(RtopAddr, 0, RoldTopValue);
3461       __ add(RoldTopValue, Roffset, RnewTopValue);
3462 
3463       // RnewTopValue contains the top address after the new object
3464       // has been allocated.
3465       __ cmp_and_brx_short(RnewTopValue, RendValue, Assembler::greaterUnsigned, Assembler::pn, slow_case);
3466 
3467       __ cas_ptr(RtopAddr, RoldTopValue, RnewTopValue);
3468 
3469       // if someone beat us on the allocation, try again, otherwise continue
3470       __ cmp_and_brx_short(RoldTopValue, RnewTopValue, Assembler::notEqual, Assembler::pn, retry);
3471 
3472       // bump total bytes allocated by this thread
3473       // RoldTopValue and RtopAddr are dead, so can use G1 and G3
3474       __ incr_allocated_bytes(Roffset, G1_scratch, G3_scratch);
3475     }
3476   }
3477 
3478   // If UseTLAB or allow_shared_alloc are true, the object is created above and
3479   // there is an initialize need. Otherwise, skip and go to the slow path.
3480   if (UseTLAB || allow_shared_alloc) {
3481     // clear object fields
3482     __ bind(initialize_object);
3483     __ deccc(Roffset, sizeof(oopDesc));
3484     __ br(Assembler::zero, false, Assembler::pt, initialize_header);
3485     __ delayed()->add(RallocatedObject, sizeof(oopDesc), G3_scratch);
3486 
3487     // initialize remaining object fields
3488     if (UseBlockZeroing) {
3489       // Use BIS for zeroing
3490       __ bis_zeroing(G3_scratch, Roffset, G1_scratch, initialize_header);
3491     } else {
3492       Label loop;
3493       __ subcc(Roffset, wordSize, Roffset);
3494       __ bind(loop);
3495       //__ subcc(Roffset, wordSize, Roffset);      // executed above loop or in delay slot
3496       __ st_ptr(G0, G3_scratch, Roffset);
3497       __ br(Assembler::notEqual, false, Assembler::pt, loop);
3498       __ delayed()->subcc(Roffset, wordSize, Roffset);
3499     }
3500     __ ba_short(initialize_header);
3501   }
3502 
3503   // slow case
3504   __ bind(slow_case);
3505   __ get_2_byte_integer_at_bcp(1, G3_scratch, O2, InterpreterMacroAssembler::Unsigned);
3506   __ get_constant_pool(O1);
3507 
3508   call_VM(Otos_i, CAST_FROM_FN_PTR(address, InterpreterRuntime::_new), O1, O2);
3509 
3510   __ ba_short(done);
3511 
3512   // Initialize the header: mark, klass
3513   __ bind(initialize_header);
3514 
3515   if (UseBiasedLocking) {
3516     __ ld_ptr(RinstanceKlass, in_bytes(Klass::prototype_header_offset()), G4_scratch);
3517   } else {
3518     __ set((intptr_t)markWord::prototype().value(), G4_scratch);
3519   }
3520   __ st_ptr(G4_scratch, RallocatedObject, oopDesc::mark_offset_in_bytes());       // mark
3521   __ store_klass_gap(G0, RallocatedObject);         // klass gap if compressed
3522   __ store_klass(RinstanceKlass, RallocatedObject); // klass (last for cms)
3523 
3524   {
3525     SkipIfEqual skip_if(
3526       _masm, G4_scratch, &DTraceAllocProbes, Assembler::zero);
3527     // Trigger dtrace event
3528     __ push(atos);
3529     __ call_VM_leaf(noreg,
3530        CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc), O0);
3531     __ pop(atos);
3532   }
3533 
3534   // continue
3535   __ bind(done);
3536 }
3537 
3538 
3539 
3540 void TemplateTable::newarray() {
3541   transition(itos, atos);
3542   __ ldub(Lbcp, 1, O1);
3543      call_VM(Otos_i, CAST_FROM_FN_PTR(address, InterpreterRuntime::newarray), O1, Otos_i);
3544 }
3545 
3546 
3547 void TemplateTable::anewarray() {
3548   transition(itos, atos);
3549   __ get_constant_pool(O1);
3550   __ get_2_byte_integer_at_bcp(1, G4_scratch, O2, InterpreterMacroAssembler::Unsigned);
3551      call_VM(Otos_i, CAST_FROM_FN_PTR(address, InterpreterRuntime::anewarray), O1, O2, Otos_i);
3552 }
3553 
3554 
3555 void TemplateTable::arraylength() {
3556   transition(atos, itos);
3557   Label ok;
3558   __ verify_oop(Otos_i);
3559   __ tst(Otos_i);
3560   __ throw_if_not_1_x( Assembler::notZero, ok );
3561   __ delayed()->ld(Otos_i, arrayOopDesc::length_offset_in_bytes(), Otos_i);
3562   __ throw_if_not_2( Interpreter::_throw_NullPointerException_entry, G3_scratch, ok);
3563 }
3564 
3565 
3566 void TemplateTable::checkcast() {
3567   transition(atos, atos);
3568   Label done, is_null, quicked, cast_ok, resolved;
3569   Register Roffset = G1_scratch;
3570   Register RobjKlass = O5;
3571   Register RspecifiedKlass = O4;
3572 
3573   // Check for casting a NULL
3574   __ br_null(Otos_i, false, Assembler::pn, is_null);
3575   __ delayed()->nop();
3576 
3577   // Get value klass in RobjKlass
3578   __ load_klass(Otos_i, RobjKlass); // get value klass
3579 
3580   // Get constant pool tag
3581   __ get_2_byte_integer_at_bcp(1, Lscratch, Roffset, InterpreterMacroAssembler::Unsigned);
3582 
3583   // See if the checkcast has been quickened
3584   __ get_cpool_and_tags(Lscratch, G3_scratch);
3585   __ add(G3_scratch, Array<u1>::base_offset_in_bytes(), G3_scratch);
3586   __ ldub(G3_scratch, Roffset, G3_scratch);
3587   __ cmp(G3_scratch, JVM_CONSTANT_Class);
3588   __ br(Assembler::equal, true, Assembler::pt, quicked);
3589   __ delayed()->sll(Roffset, LogBytesPerWord, Roffset);
3590 
3591   __ push_ptr(); // save receiver for result, and for GC
3592   call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::quicken_io_cc) );
3593   __ get_vm_result_2(RspecifiedKlass);
3594   __ pop_ptr(Otos_i, G3_scratch); // restore receiver
3595 
3596   __ ba_short(resolved);
3597 
3598   // Extract target class from constant pool
3599   __ bind(quicked);
3600   __ load_resolved_klass_at_offset(Lscratch, Roffset, RspecifiedKlass);
3601 
3602 
3603   __ bind(resolved);
3604   __ load_klass(Otos_i, RobjKlass); // get value klass
3605 
3606   // Generate a fast subtype check.  Branch to cast_ok if no
3607   // failure.  Throw exception if failure.
3608   __ gen_subtype_check( RobjKlass, RspecifiedKlass, G3_scratch, G4_scratch, G1_scratch, cast_ok );
3609 
3610   // Not a subtype; so must throw exception
3611   __ throw_if_not_x( Assembler::never, Interpreter::_throw_ClassCastException_entry, G3_scratch );
3612 
3613   __ bind(cast_ok);
3614 
3615   if (ProfileInterpreter) {
3616     __ ba_short(done);
3617   }
3618   __ bind(is_null);
3619   __ profile_null_seen(G3_scratch);
3620   __ bind(done);
3621 }
3622 
3623 
3624 void TemplateTable::instanceof() {
3625   Label done, is_null, quicked, resolved;
3626   transition(atos, itos);
3627   Register Roffset = G1_scratch;
3628   Register RobjKlass = O5;
3629   Register RspecifiedKlass = O4;
3630 
3631   // Check for casting a NULL
3632   __ br_null(Otos_i, false, Assembler::pt, is_null);
3633   __ delayed()->nop();
3634 
3635   // Get value klass in RobjKlass
3636   __ load_klass(Otos_i, RobjKlass); // get value klass
3637 
3638   // Get constant pool tag
3639   __ get_2_byte_integer_at_bcp(1, Lscratch, Roffset, InterpreterMacroAssembler::Unsigned);
3640 
3641   // See if the checkcast has been quickened
3642   __ get_cpool_and_tags(Lscratch, G3_scratch);
3643   __ add(G3_scratch, Array<u1>::base_offset_in_bytes(), G3_scratch);
3644   __ ldub(G3_scratch, Roffset, G3_scratch);
3645   __ cmp(G3_scratch, JVM_CONSTANT_Class);
3646   __ br(Assembler::equal, true, Assembler::pt, quicked);
3647   __ delayed()->sll(Roffset, LogBytesPerWord, Roffset);
3648 
3649   __ push_ptr(); // save receiver for result, and for GC
3650   call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::quicken_io_cc) );
3651   __ get_vm_result_2(RspecifiedKlass);
3652   __ pop_ptr(Otos_i, G3_scratch); // restore receiver
3653 
3654   __ ba_short(resolved);
3655 
3656   // Extract target class from constant pool
3657   __ bind(quicked);
3658   __ get_constant_pool(Lscratch);
3659   __ load_resolved_klass_at_offset(Lscratch, Roffset, RspecifiedKlass);
3660 
3661   __ bind(resolved);
3662   __ load_klass(Otos_i, RobjKlass); // get value klass
3663 
3664   // Generate a fast subtype check.  Branch to cast_ok if no
3665   // failure.  Return 0 if failure.
3666   __ or3(G0, 1, Otos_i);      // set result assuming quick tests succeed
3667   __ gen_subtype_check( RobjKlass, RspecifiedKlass, G3_scratch, G4_scratch, G1_scratch, done );
3668   // Not a subtype; return 0;
3669   __ clr( Otos_i );
3670 
3671   if (ProfileInterpreter) {
3672     __ ba_short(done);
3673   }
3674   __ bind(is_null);
3675   __ profile_null_seen(G3_scratch);
3676   __ bind(done);
3677 }
3678 
3679 void TemplateTable::_breakpoint() {
3680 
3681    // Note: We get here even if we are single stepping..
3682    // jbug insists on setting breakpoints at every bytecode
3683    // even if we are in single step mode.
3684 
3685    transition(vtos, vtos);
3686    // get the unpatched byte code
3687    __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::get_original_bytecode_at), Lmethod, Lbcp);
3688    __ mov(O0, Lbyte_code);
3689 
3690    // post the breakpoint event
3691    __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::_breakpoint), Lmethod, Lbcp);
3692 
3693    // complete the execution of original bytecode
3694    __ dispatch_normal(vtos);
3695 }
3696 
3697 
3698 //----------------------------------------------------------------------------------------------------
3699 // Exceptions
3700 
3701 void TemplateTable::athrow() {
3702   transition(atos, vtos);
3703 
3704   // This works because exception is cached in Otos_i which is same as O0,
3705   // which is same as what throw_exception_entry_expects
3706   assert(Otos_i == Oexception, "see explanation above");
3707 
3708   __ verify_oop(Otos_i);
3709   __ null_check(Otos_i);
3710   __ throw_if_not_x(Assembler::never, Interpreter::throw_exception_entry(), G3_scratch);
3711 }
3712 
3713 
3714 //----------------------------------------------------------------------------------------------------
3715 // Synchronization
3716 
3717 
3718 // See frame_sparc.hpp for monitor block layout.
3719 // Monitor elements are dynamically allocated by growing stack as needed.
3720 
3721 void TemplateTable::monitorenter() {
3722   transition(atos, vtos);
3723   __ verify_oop(Otos_i);
3724   // Try to acquire a lock on the object
3725   // Repeat until succeeded (i.e., until
3726   // monitorenter returns true).
3727 
3728   {   Label ok;
3729     __ tst(Otos_i);
3730     __ throw_if_not_1_x( Assembler::notZero,  ok);
3731     __ delayed()->mov(Otos_i, Lscratch); // save obj
3732     __ throw_if_not_2( Interpreter::_throw_NullPointerException_entry, G3_scratch, ok);
3733   }
3734 
3735   assert(O0 == Otos_i, "Be sure where the object to lock is");
3736 
3737   // find a free slot in the monitor block
3738 
3739 
3740   // initialize entry pointer
3741   __ clr(O1); // points to free slot or NULL
3742 
3743   {
3744     Label entry, loop, exit;
3745     __ add( __ top_most_monitor(), O2 ); // last one to check
3746     __ ba( entry );
3747     __ delayed()->mov( Lmonitors, O3 ); // first one to check
3748 
3749 
3750     __ bind( loop );
3751 
3752     __ verify_oop(O4);          // verify each monitor's oop
3753     __ tst(O4); // is this entry unused?
3754     __ movcc( Assembler::zero, false, Assembler::ptr_cc, O3, O1);
3755 
3756     __ cmp(O4, O0); // check if current entry is for same object
3757     __ brx( Assembler::equal, false, Assembler::pn, exit );
3758     __ delayed()->inc( O3, frame::interpreter_frame_monitor_size() * wordSize ); // check next one
3759 
3760     __ bind( entry );
3761 
3762     __ cmp( O3, O2 );
3763     __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, loop );
3764     __ delayed()->ld_ptr(O3, BasicObjectLock::obj_offset_in_bytes(), O4);
3765 
3766     __ bind( exit );
3767   }
3768 
3769   { Label allocated;
3770 
3771     // found free slot?
3772     __ br_notnull_short(O1, Assembler::pn, allocated);
3773 
3774     __ add_monitor_to_stack( false, O2, O3 );
3775     __ mov(Lmonitors, O1);
3776 
3777     __ bind(allocated);
3778   }
3779 
3780   // Increment bcp to point to the next bytecode, so exception handling for async. exceptions work correctly.
3781   // The object has already been poped from the stack, so the expression stack looks correct.
3782   __ inc(Lbcp);
3783 
3784   __ st_ptr(O0, O1, BasicObjectLock::obj_offset_in_bytes()); // store object
3785   __ lock_object(O1, O0);
3786 
3787   // check if there's enough space on the stack for the monitors after locking
3788   __ generate_stack_overflow_check(0);
3789 
3790   // The bcp has already been incremented. Just need to dispatch to next instruction.
3791   __ dispatch_next(vtos);
3792 }
3793 
3794 
3795 void TemplateTable::monitorexit() {
3796   transition(atos, vtos);
3797   __ verify_oop(Otos_i);
3798   __ tst(Otos_i);
3799   __ throw_if_not_x( Assembler::notZero, Interpreter::_throw_NullPointerException_entry, G3_scratch );
3800 
3801   assert(O0 == Otos_i, "just checking");
3802 
3803   { Label entry, loop, found;
3804     __ add( __ top_most_monitor(), O2 ); // last one to check
3805     __ ba(entry);
3806     // use Lscratch to hold monitor elem to check, start with most recent monitor,
3807     // By using a local it survives the call to the C routine.
3808     __ delayed()->mov( Lmonitors, Lscratch );
3809 
3810     __ bind( loop );
3811 
3812     __ verify_oop(O4);          // verify each monitor's oop
3813     __ cmp(O4, O0); // check if current entry is for desired object
3814     __ brx( Assembler::equal, true, Assembler::pt, found );
3815     __ delayed()->mov(Lscratch, O1); // pass found entry as argument to monitorexit
3816 
3817     __ inc( Lscratch, frame::interpreter_frame_monitor_size() * wordSize ); // advance to next
3818 
3819     __ bind( entry );
3820 
3821     __ cmp( Lscratch, O2 );
3822     __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, loop );
3823     __ delayed()->ld_ptr(Lscratch, BasicObjectLock::obj_offset_in_bytes(), O4);
3824 
3825     call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_illegal_monitor_state_exception));
3826     __ should_not_reach_here();
3827 
3828     __ bind(found);
3829   }
3830   __ unlock_object(O1);
3831 }
3832 
3833 
3834 //----------------------------------------------------------------------------------------------------
3835 // Wide instructions
3836 
3837 void TemplateTable::wide() {
3838   transition(vtos, vtos);
3839   __ ldub(Lbcp, 1, G3_scratch);// get next bc
3840   __ sll(G3_scratch, LogBytesPerWord, G3_scratch);
3841   AddressLiteral ep(Interpreter::_wentry_point);
3842   __ set(ep, G4_scratch);
3843   __ ld_ptr(G4_scratch, G3_scratch, G3_scratch);
3844   __ jmp(G3_scratch, G0);
3845   __ delayed()->nop();
3846   // Note: the Lbcp increment step is part of the individual wide bytecode implementations
3847 }
3848 
3849 
3850 //----------------------------------------------------------------------------------------------------
3851 // Multi arrays
3852 
3853 void TemplateTable::multianewarray() {
3854   transition(vtos, atos);
3855      // put ndims * wordSize into Lscratch
3856   __ ldub( Lbcp,     3,               Lscratch);
3857   __ sll(  Lscratch, Interpreter::logStackElementSize, Lscratch);
3858      // Lesp points past last_dim, so set to O1 to first_dim address
3859   __ add(  Lesp,     Lscratch,        O1);
3860      call_VM(Otos_i, CAST_FROM_FN_PTR(address, InterpreterRuntime::multianewarray), O1);
3861   __ add(  Lesp,     Lscratch,        Lesp); // pop all dimensions off the stack
3862 }