1 /*
   2  * Copyright (c) 2007, 2014, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/assembler.hpp"
  27 #include "interpreter/bytecodeHistogram.hpp"
  28 #include "interpreter/cppInterpreter.hpp"
  29 #include "interpreter/interpreter.hpp"
  30 #include "interpreter/interpreterGenerator.hpp"
  31 #include "interpreter/interpreterRuntime.hpp"
  32 #include "interpreter/interp_masm.hpp"
  33 #include "oops/arrayOop.hpp"
  34 #include "oops/methodData.hpp"
  35 #include "oops/method.hpp"
  36 #include "oops/oop.inline.hpp"
  37 #include "prims/jvmtiExport.hpp"
  38 #include "prims/jvmtiThreadState.hpp"
  39 #include "runtime/arguments.hpp"
  40 #include "runtime/deoptimization.hpp"
  41 #include "runtime/frame.inline.hpp"
  42 #include "runtime/interfaceSupport.hpp"
  43 #include "runtime/sharedRuntime.hpp"
  44 #include "runtime/stubRoutines.hpp"
  45 #include "runtime/synchronizer.hpp"
  46 #include "runtime/timer.hpp"
  47 #include "runtime/vframeArray.hpp"
  48 #include "utilities/debug.hpp"
  49 #include "utilities/macros.hpp"
  50 #ifdef SHARK
  51 #include "shark/shark_globals.hpp"
  52 #endif
  53 
  54 #ifdef CC_INTERP
  55 
  56 // Routine exists to make tracebacks look decent in debugger
  57 // while "shadow" interpreter frames are on stack. It is also
  58 // used to distinguish interpreter frames.
  59 
  60 extern "C" void RecursiveInterpreterActivation(interpreterState istate) {
  61   ShouldNotReachHere();
  62 }
  63 
  64 bool CppInterpreter::contains(address pc) {
  65   return ( _code->contains(pc) ||
  66          ( pc == (CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset)));
  67 }
  68 
  69 #define STATE(field_name) Lstate, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
  70 #define __ _masm->
  71 
  72 Label frame_manager_entry; // c++ interpreter entry point this holds that entry point label.
  73 
  74 static address unctrap_frame_manager_entry  = NULL;
  75 
  76 static address interpreter_return_address  = NULL;
  77 static address deopt_frame_manager_return_atos  = NULL;
  78 static address deopt_frame_manager_return_btos  = NULL;
  79 static address deopt_frame_manager_return_itos  = NULL;
  80 static address deopt_frame_manager_return_ltos  = NULL;
  81 static address deopt_frame_manager_return_ftos  = NULL;
  82 static address deopt_frame_manager_return_dtos  = NULL;
  83 static address deopt_frame_manager_return_vtos  = NULL;
  84 
  85 const Register prevState = G1_scratch;
  86 
  87 void InterpreterGenerator::save_native_result(void) {
  88   // result potentially in O0/O1: save it across calls
  89   __ stf(FloatRegisterImpl::D, F0, STATE(_native_fresult));
  90 #ifdef _LP64
  91   __ stx(O0, STATE(_native_lresult));
  92 #else
  93   __ std(O0, STATE(_native_lresult));
  94 #endif
  95 }
  96 
  97 void InterpreterGenerator::restore_native_result(void) {
  98 
  99   // Restore any method result value
 100   __ ldf(FloatRegisterImpl::D, STATE(_native_fresult), F0);
 101 #ifdef _LP64
 102   __ ldx(STATE(_native_lresult), O0);
 103 #else
 104   __ ldd(STATE(_native_lresult), O0);
 105 #endif
 106 }
 107 
 108 // A result handler converts/unboxes a native call result into
 109 // a java interpreter/compiler result. The current frame is an
 110 // interpreter frame. The activation frame unwind code must be
 111 // consistent with that of TemplateTable::_return(...). In the
 112 // case of native methods, the caller's SP was not modified.
 113 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) {
 114   address entry = __ pc();
 115   Register Itos_i  = Otos_i ->after_save();
 116   Register Itos_l  = Otos_l ->after_save();
 117   Register Itos_l1 = Otos_l1->after_save();
 118   Register Itos_l2 = Otos_l2->after_save();
 119   switch (type) {
 120     case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
 121     case T_CHAR   : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i);   break; // cannot use and3, 0xFFFF too big as immediate value!
 122     case T_BYTE   : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i);   break;
 123     case T_SHORT  : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i);   break;
 124     case T_LONG   :
 125 #ifndef _LP64
 126                     __ mov(O1, Itos_l2);  // move other half of long
 127 #endif              // ifdef or no ifdef, fall through to the T_INT case
 128     case T_INT    : __ mov(O0, Itos_i);                         break;
 129     case T_VOID   : /* nothing to do */                         break;
 130     case T_FLOAT  : assert(F0 == Ftos_f, "fix this code" );     break;
 131     case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" );     break;
 132     case T_OBJECT :
 133       __ ld_ptr(STATE(_oop_temp), Itos_i);
 134       __ verify_oop(Itos_i);
 135       break;
 136     default       : ShouldNotReachHere();
 137   }
 138   __ ret();                           // return from interpreter activation
 139   __ delayed()->restore(I5_savedSP, G0, SP);  // remove interpreter frame
 140   NOT_PRODUCT(__ emit_int32(0);)       // marker for disassembly
 141   return entry;
 142 }
 143 
 144 // tosca based result to c++ interpreter stack based result.
 145 // Result goes to address in L1_scratch
 146 
 147 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) {
 148   // A result is in the native abi result register from a native method call.
 149   // We need to return this result to the interpreter by pushing the result on the interpreter's
 150   // stack. This is relatively simple the destination is in L1_scratch
 151   // i.e. L1_scratch is the first free element on the stack. If we "push" a return value we must
 152   // adjust L1_scratch
 153   address entry = __ pc();
 154   switch (type) {
 155     case T_BOOLEAN:
 156       // !0 => true; 0 => false
 157       __ subcc(G0, O0, G0);
 158       __ addc(G0, 0, O0);
 159       __ st(O0, L1_scratch, 0);
 160       __ sub(L1_scratch, wordSize, L1_scratch);
 161       break;
 162 
 163     // cannot use and3, 0xFFFF too big as immediate value!
 164     case T_CHAR   :
 165       __ sll(O0, 16, O0);
 166       __ srl(O0, 16, O0);
 167       __ st(O0, L1_scratch, 0);
 168       __ sub(L1_scratch, wordSize, L1_scratch);
 169       break;
 170 
 171     case T_BYTE   :
 172       __ sll(O0, 24, O0);
 173       __ sra(O0, 24, O0);
 174       __ st(O0, L1_scratch, 0);
 175       __ sub(L1_scratch, wordSize, L1_scratch);
 176       break;
 177 
 178     case T_SHORT  :
 179       __ sll(O0, 16, O0);
 180       __ sra(O0, 16, O0);
 181       __ st(O0, L1_scratch, 0);
 182       __ sub(L1_scratch, wordSize, L1_scratch);
 183       break;
 184     case T_LONG   :
 185 #ifndef _LP64
 186 #if defined(COMPILER2)
 187   // All return values are where we want them, except for Longs.  C2 returns
 188   // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
 189   // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
 190   // build even if we are returning from interpreted we just do a little
 191   // stupid shuffing.
 192   // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
 193   // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
 194   // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
 195       __ stx(G1, L1_scratch, -wordSize);
 196 #else
 197       // native result is in O0, O1
 198       __ st(O1, L1_scratch, 0);                      // Low order
 199       __ st(O0, L1_scratch, -wordSize);              // High order
 200 #endif /* COMPILER2 */
 201 #else
 202       __ stx(O0, L1_scratch, -wordSize);
 203 #endif
 204       __ sub(L1_scratch, 2*wordSize, L1_scratch);
 205       break;
 206 
 207     case T_INT    :
 208       __ st(O0, L1_scratch, 0);
 209       __ sub(L1_scratch, wordSize, L1_scratch);
 210       break;
 211 
 212     case T_VOID   : /* nothing to do */
 213       break;
 214 
 215     case T_FLOAT  :
 216       __ stf(FloatRegisterImpl::S, F0, L1_scratch, 0);
 217       __ sub(L1_scratch, wordSize, L1_scratch);
 218       break;
 219 
 220     case T_DOUBLE :
 221       // Every stack slot is aligned on 64 bit, However is this
 222       // the correct stack slot on 64bit?? QQQ
 223       __ stf(FloatRegisterImpl::D, F0, L1_scratch, -wordSize);
 224       __ sub(L1_scratch, 2*wordSize, L1_scratch);
 225       break;
 226     case T_OBJECT :
 227       __ verify_oop(O0);
 228       __ st_ptr(O0, L1_scratch, 0);
 229       __ sub(L1_scratch, wordSize, L1_scratch);
 230       break;
 231     default       : ShouldNotReachHere();
 232   }
 233   __ retl();                          // return from interpreter activation
 234   __ delayed()->nop();                // schedule this better
 235   NOT_PRODUCT(__ emit_int32(0);)       // marker for disassembly
 236   return entry;
 237 }
 238 
 239 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) {
 240   // A result is in the java expression stack of the interpreted method that has just
 241   // returned. Place this result on the java expression stack of the caller.
 242   //
 243   // The current interpreter activation in Lstate is for the method just returning its
 244   // result. So we know that the result of this method is on the top of the current
 245   // execution stack (which is pre-pushed) and will be return to the top of the caller
 246   // stack. The top of the callers stack is the bottom of the locals of the current
 247   // activation.
 248   // Because of the way activation are managed by the frame manager the value of esp is
 249   // below both the stack top of the current activation and naturally the stack top
 250   // of the calling activation. This enable this routine to leave the return address
 251   // to the frame manager on the stack and do a vanilla return.
 252   //
 253   // On entry: O0 - points to source (callee stack top)
 254   //           O1 - points to destination (caller stack top [i.e. free location])
 255   // destroys O2, O3
 256   //
 257 
 258   address entry = __ pc();
 259   switch (type) {
 260     case T_VOID:  break;
 261       break;
 262     case T_FLOAT  :
 263     case T_BOOLEAN:
 264     case T_CHAR   :
 265     case T_BYTE   :
 266     case T_SHORT  :
 267     case T_INT    :
 268       // 1 word result
 269       __ ld(O0, 0, O2);
 270       __ st(O2, O1, 0);
 271       __ sub(O1, wordSize, O1);
 272       break;
 273     case T_DOUBLE  :
 274     case T_LONG    :
 275       // return top two words on current expression stack to caller's expression stack
 276       // The caller's expression stack is adjacent to the current frame manager's intepretState
 277       // except we allocated one extra word for this intepretState so we won't overwrite it
 278       // when we return a two word result.
 279 #ifdef _LP64
 280       __ ld_ptr(O0, 0, O2);
 281       __ st_ptr(O2, O1, -wordSize);
 282 #else
 283       __ ld(O0, 0, O2);
 284       __ ld(O0, wordSize, O3);
 285       __ st(O3, O1, 0);
 286       __ st(O2, O1, -wordSize);
 287 #endif
 288       __ sub(O1, 2*wordSize, O1);
 289       break;
 290     case T_OBJECT :
 291       __ ld_ptr(O0, 0, O2);
 292       __ verify_oop(O2);                                               // verify it
 293       __ st_ptr(O2, O1, 0);
 294       __ sub(O1, wordSize, O1);
 295       break;
 296     default       : ShouldNotReachHere();
 297   }
 298   __ retl();
 299   __ delayed()->nop(); // QQ schedule this better
 300   return entry;
 301 }
 302 
 303 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) {
 304   // A result is in the java expression stack of the interpreted method that has just
 305   // returned. Place this result in the native abi that the caller expects.
 306   // We are in a new frame registers we set must be in caller (i.e. callstub) frame.
 307   //
 308   // Similar to generate_stack_to_stack_converter above. Called at a similar time from the
 309   // frame manager execept in this situation the caller is native code (c1/c2/call_stub)
 310   // and so rather than return result onto caller's java expression stack we return the
 311   // result in the expected location based on the native abi.
 312   // On entry: O0 - source (stack top)
 313   // On exit result in expected output register
 314   // QQQ schedule this better
 315 
 316   address entry = __ pc();
 317   switch (type) {
 318     case T_VOID:  break;
 319       break;
 320     case T_FLOAT  :
 321       __ ldf(FloatRegisterImpl::S, O0, 0, F0);
 322       break;
 323     case T_BOOLEAN:
 324     case T_CHAR   :
 325     case T_BYTE   :
 326     case T_SHORT  :
 327     case T_INT    :
 328       // 1 word result
 329       __ ld(O0, 0, O0->after_save());
 330       break;
 331     case T_DOUBLE  :
 332       __ ldf(FloatRegisterImpl::D, O0, 0, F0);
 333       break;
 334     case T_LONG    :
 335       // return top two words on current expression stack to caller's expression stack
 336       // The caller's expression stack is adjacent to the current frame manager's interpretState
 337       // except we allocated one extra word for this intepretState so we won't overwrite it
 338       // when we return a two word result.
 339 #ifdef _LP64
 340       __ ld_ptr(O0, 0, O0->after_save());
 341 #else
 342       __ ld(O0, wordSize, O1->after_save());
 343       __ ld(O0, 0, O0->after_save());
 344 #endif
 345 #if defined(COMPILER2) && !defined(_LP64)
 346       // C2 expects long results in G1 we can't tell if we're returning to interpreted
 347       // or compiled so just be safe use G1 and O0/O1
 348 
 349       // Shift bits into high (msb) of G1
 350       __ sllx(Otos_l1->after_save(), 32, G1);
 351       // Zero extend low bits
 352       __ srl (Otos_l2->after_save(), 0, Otos_l2->after_save());
 353       __ or3 (Otos_l2->after_save(), G1, G1);
 354 #endif /* COMPILER2 */
 355       break;
 356     case T_OBJECT :
 357       __ ld_ptr(O0, 0, O0->after_save());
 358       __ verify_oop(O0->after_save());                                               // verify it
 359       break;
 360     default       : ShouldNotReachHere();
 361   }
 362   __ retl();
 363   __ delayed()->nop();
 364   return entry;
 365 }
 366 
 367 address CppInterpreter::return_entry(TosState state, int length, Bytecodes::Code code) {
 368   // make it look good in the debugger
 369   return CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset;
 370 }
 371 
 372 address CppInterpreter::deopt_entry(TosState state, int length) {
 373   address ret = NULL;
 374   if (length != 0) {
 375     switch (state) {
 376       case atos: ret = deopt_frame_manager_return_atos; break;
 377       case btos: ret = deopt_frame_manager_return_btos; break;
 378       case ctos:
 379       case stos:
 380       case itos: ret = deopt_frame_manager_return_itos; break;
 381       case ltos: ret = deopt_frame_manager_return_ltos; break;
 382       case ftos: ret = deopt_frame_manager_return_ftos; break;
 383       case dtos: ret = deopt_frame_manager_return_dtos; break;
 384       case vtos: ret = deopt_frame_manager_return_vtos; break;
 385     }
 386   } else {
 387     ret = unctrap_frame_manager_entry;  // re-execute the bytecode ( e.g. uncommon trap)
 388   }
 389   assert(ret != NULL, "Not initialized");
 390   return ret;
 391 }
 392 
 393 //
 394 // Helpers for commoning out cases in the various type of method entries.
 395 //
 396 
 397 // increment invocation count & check for overflow
 398 //
 399 // Note: checking for negative value instead of overflow
 400 //       so we have a 'sticky' overflow test
 401 //
 402 // Lmethod: method
 403 // ??: invocation counter
 404 //
 405 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
 406   Label done;
 407   const Register Rcounters = G3_scratch;
 408 
 409   __ ld_ptr(STATE(_method), G5_method);
 410   __ get_method_counters(G5_method, Rcounters, done);
 411 
 412   // Update standard invocation counters
 413   __ increment_invocation_counter(Rcounters, O0, G4_scratch);
 414   if (ProfileInterpreter) {
 415     Address interpreter_invocation_counter(Rcounters,
 416             in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
 417     __ ld(interpreter_invocation_counter, G4_scratch);
 418     __ inc(G4_scratch);
 419     __ st(G4_scratch, interpreter_invocation_counter);
 420   }
 421 
 422   AddressLiteral invocation_limit((address)&InvocationCounter::InterpreterInvocationLimit);
 423   __ load_contents(invocation_limit, G3_scratch);
 424   __ cmp(O0, G3_scratch);
 425   __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
 426   __ delayed()->nop();
 427   __ bind(done);
 428 }
 429 
 430 address InterpreterGenerator::generate_empty_entry(void) {
 431 
 432   // A method that does nothing but return...
 433 
 434   address entry = __ pc();
 435   Label slow_path;
 436 
 437   // do nothing for empty methods (do not even increment invocation counter)
 438   if ( UseFastEmptyMethods) {
 439     // If we need a safepoint check, generate full interpreter entry.
 440     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
 441     __ load_contents(sync_state, G3_scratch);
 442     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 443     __ br(Assembler::notEqual, false, Assembler::pn, frame_manager_entry);
 444     __ delayed()->nop();
 445 
 446     // Code: _return
 447     __ retl();
 448     __ delayed()->mov(O5_savedSP, SP);
 449     return entry;
 450   }
 451   return NULL;
 452 }
 453 
 454 address InterpreterGenerator::generate_Reference_get_entry(void) {
 455 #if INCLUDE_ALL_GCS
 456   if (UseG1GC) {
 457     // We need to generate have a routine that generates code to:
 458     //   * load the value in the referent field
 459     //   * passes that value to the pre-barrier.
 460     //
 461     // In the case of G1 this will record the value of the
 462     // referent in an SATB buffer if marking is active.
 463     // This will cause concurrent marking to mark the referent
 464     // field as live.
 465     Unimplemented();
 466   }
 467 #endif // INCLUDE_ALL_GCS
 468 
 469   // If G1 is not enabled then attempt to go through the accessor entry point
 470   // Reference.get is an accessor
 471   return NULL;
 472 }
 473 
 474 //
 475 // Interpreter stub for calling a native method. (C++ interpreter)
 476 // This sets up a somewhat different looking stack for calling the native method
 477 // than the typical interpreter frame setup.
 478 //
 479 
 480 address InterpreterGenerator::generate_native_entry(bool synchronized) {
 481   address entry = __ pc();
 482 
 483   // the following temporary registers are used during frame creation
 484   const Register Gtmp1 = G3_scratch ;
 485   const Register Gtmp2 = G1_scratch;
 486   const Register RconstMethod = Gtmp1;
 487   const Address constMethod(G5_method, in_bytes(Method::const_offset()));
 488   const Address size_of_parameters(RconstMethod, in_bytes(ConstMethod::size_of_parameters_offset()));
 489 
 490   bool inc_counter  = UseCompiler || CountCompiledCalls;
 491 
 492   // make sure registers are different!
 493   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
 494 
 495   const Address access_flags      (G5_method, in_bytes(Method::access_flags_offset()));
 496 
 497   Label Lentry;
 498   __ bind(Lentry);
 499 
 500   const Register Glocals_size = G3;
 501   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
 502 
 503   // make sure method is native & not abstract
 504   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
 505 #ifdef ASSERT
 506   __ ld(access_flags, Gtmp1);
 507   {
 508     Label L;
 509     __ btst(JVM_ACC_NATIVE, Gtmp1);
 510     __ br(Assembler::notZero, false, Assembler::pt, L);
 511     __ delayed()->nop();
 512     __ stop("tried to execute non-native method as native");
 513     __ bind(L);
 514   }
 515   { Label L;
 516     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
 517     __ br(Assembler::zero, false, Assembler::pt, L);
 518     __ delayed()->nop();
 519     __ stop("tried to execute abstract method as non-abstract");
 520     __ bind(L);
 521   }
 522 #endif // ASSERT
 523 
 524   __ ld_ptr(constMethod, RconstMethod);
 525   __ lduh(size_of_parameters, Gtmp1);
 526   __ sll(Gtmp1, LogBytesPerWord, Gtmp2);       // parameter size in bytes
 527   __ add(Gargs, Gtmp2, Gargs);                 // points to first local + BytesPerWord
 528   // NEW
 529   __ add(Gargs, -wordSize, Gargs);             // points to first local[0]
 530   // generate the code to allocate the interpreter stack frame
 531   // NEW FRAME ALLOCATED HERE
 532   // save callers original sp
 533   // __ mov(SP, I5_savedSP->after_restore());
 534 
 535   generate_compute_interpreter_state(Lstate, G0, true);
 536 
 537   // At this point Lstate points to new interpreter state
 538   //
 539 
 540   const Address do_not_unlock_if_synchronized(G2_thread,
 541       in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
 542   // Since at this point in the method invocation the exception handler
 543   // would try to exit the monitor of synchronized methods which hasn't
 544   // been entered yet, we set the thread local variable
 545   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
 546   // runtime, exception handling i.e. unlock_if_synchronized_method will
 547   // check this thread local flag.
 548   // This flag has two effects, one is to force an unwind in the topmost
 549   // interpreter frame and not perform an unlock while doing so.
 550 
 551   __ movbool(true, G3_scratch);
 552   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
 553 
 554 
 555   // increment invocation counter and check for overflow
 556   //
 557   // Note: checking for negative value instead of overflow
 558   //       so we have a 'sticky' overflow test (may be of
 559   //       importance as soon as we have true MT/MP)
 560   Label invocation_counter_overflow;
 561   if (inc_counter) {
 562     generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
 563   }
 564   Label Lcontinue;
 565   __ bind(Lcontinue);
 566 
 567   bang_stack_shadow_pages(true);
 568   // reset the _do_not_unlock_if_synchronized flag
 569   __ stbool(G0, do_not_unlock_if_synchronized);
 570 
 571   // check for synchronized methods
 572   // Must happen AFTER invocation_counter check, so method is not locked
 573   // if counter overflows.
 574 
 575   if (synchronized) {
 576     lock_method();
 577     // Don't see how G2_thread is preserved here...
 578     // __ verify_thread(); QQQ destroys L0,L1 can't use
 579   } else {
 580 #ifdef ASSERT
 581     { Label ok;
 582       __ ld_ptr(STATE(_method), G5_method);
 583       __ ld(access_flags, O0);
 584       __ btst(JVM_ACC_SYNCHRONIZED, O0);
 585       __ br( Assembler::zero, false, Assembler::pt, ok);
 586       __ delayed()->nop();
 587       __ stop("method needs synchronization");
 588       __ bind(ok);
 589     }
 590 #endif // ASSERT
 591   }
 592 
 593   // start execution
 594 
 595 //   __ verify_thread(); kills L1,L2 can't  use at the moment
 596 
 597   // jvmti/jvmpi support
 598   __ notify_method_entry();
 599 
 600   // native call
 601 
 602   // (note that O0 is never an oop--at most it is a handle)
 603   // It is important not to smash any handles created by this call,
 604   // until any oop handle in O0 is dereferenced.
 605 
 606   // (note that the space for outgoing params is preallocated)
 607 
 608   // get signature handler
 609 
 610   Label pending_exception_present;
 611 
 612   { Label L;
 613     __ ld_ptr(STATE(_method), G5_method);
 614     __ ld_ptr(Address(G5_method, in_bytes(Method::signature_handler_offset())), G3_scratch);
 615     __ tst(G3_scratch);
 616     __ brx(Assembler::notZero, false, Assembler::pt, L);
 617     __ delayed()->nop();
 618     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), G5_method, false);
 619     __ ld_ptr(STATE(_method), G5_method);
 620 
 621     Address exception_addr(G2_thread, in_bytes(Thread::pending_exception_offset()));
 622     __ ld_ptr(exception_addr, G3_scratch);
 623     __ br_notnull_short(G3_scratch, Assembler::pn, pending_exception_present);
 624     __ ld_ptr(Address(G5_method, in_bytes(Method::signature_handler_offset())), G3_scratch);
 625     __ bind(L);
 626   }
 627 
 628   // Push a new frame so that the args will really be stored in
 629   // Copy a few locals across so the new frame has the variables
 630   // we need but these values will be dead at the jni call and
 631   // therefore not gc volatile like the values in the current
 632   // frame (Lstate in particular)
 633 
 634   // Flush the state pointer to the register save area
 635   // Which is the only register we need for a stack walk.
 636   __ st_ptr(Lstate, SP, (Lstate->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
 637 
 638   __ mov(Lstate, O1);         // Need to pass the state pointer across the frame
 639 
 640   // Calculate current frame size
 641   __ sub(SP, FP, O3);         // Calculate negative of current frame size
 642   __ save(SP, O3, SP);        // Allocate an identical sized frame
 643 
 644   __ mov(I1, Lstate);          // In the "natural" register.
 645 
 646   // Note I7 has leftover trash. Slow signature handler will fill it in
 647   // should we get there. Normal jni call will set reasonable last_Java_pc
 648   // below (and fix I7 so the stack trace doesn't have a meaningless frame
 649   // in it).
 650 
 651 
 652   // call signature handler
 653   __ ld_ptr(STATE(_method), Lmethod);
 654   __ ld_ptr(STATE(_locals), Llocals);
 655 
 656   __ callr(G3_scratch, 0);
 657   __ delayed()->nop();
 658   __ ld_ptr(STATE(_thread), G2_thread);        // restore thread (shouldn't be needed)
 659 
 660   { Label not_static;
 661 
 662     __ ld_ptr(STATE(_method), G5_method);
 663     __ ld(access_flags, O0);
 664     __ btst(JVM_ACC_STATIC, O0);
 665     __ br( Assembler::zero, false, Assembler::pt, not_static);
 666     __ delayed()->
 667       // get native function entry point(O0 is a good temp until the very end)
 668        ld_ptr(Address(G5_method, in_bytes(Method::native_function_offset())), O0);
 669     // for static methods insert the mirror argument
 670     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
 671 
 672     __ ld_ptr(Address(G5_method, in_bytes(Method:: const_offset())), O1);
 673     __ ld_ptr(Address(O1, in_bytes(ConstMethod::constants_offset())), O1);
 674     __ ld_ptr(Address(O1, ConstantPool::pool_holder_offset_in_bytes()), O1);
 675     __ ld_ptr(O1, mirror_offset, O1);
 676     // where the mirror handle body is allocated:
 677 #ifdef ASSERT
 678     if (!PrintSignatureHandlers)  // do not dirty the output with this
 679     { Label L;
 680       __ tst(O1);
 681       __ brx(Assembler::notZero, false, Assembler::pt, L);
 682       __ delayed()->nop();
 683       __ stop("mirror is missing");
 684       __ bind(L);
 685     }
 686 #endif // ASSERT
 687     __ st_ptr(O1, STATE(_oop_temp));
 688     __ add(STATE(_oop_temp), O1);            // this is really an LEA not an add
 689     __ bind(not_static);
 690   }
 691 
 692   // At this point, arguments have been copied off of stack into
 693   // their JNI positions, which are O1..O5 and SP[68..].
 694   // Oops are boxed in-place on the stack, with handles copied to arguments.
 695   // The result handler is in Lscratch.  O0 will shortly hold the JNIEnv*.
 696 
 697 #ifdef ASSERT
 698   { Label L;
 699     __ tst(O0);
 700     __ brx(Assembler::notZero, false, Assembler::pt, L);
 701     __ delayed()->nop();
 702     __ stop("native entry point is missing");
 703     __ bind(L);
 704   }
 705 #endif // ASSERT
 706 
 707   //
 708   // setup the java frame anchor
 709   //
 710   // The scavenge function only needs to know that the PC of this frame is
 711   // in the interpreter method entry code, it doesn't need to know the exact
 712   // PC and hence we can use O7 which points to the return address from the
 713   // previous call in the code stream (signature handler function)
 714   //
 715   // The other trick is we set last_Java_sp to FP instead of the usual SP because
 716   // we have pushed the extra frame in order to protect the volatile register(s)
 717   // in that frame when we return from the jni call
 718   //
 719 
 720 
 721   __ set_last_Java_frame(FP, O7);
 722   __ mov(O7, I7);  // make dummy interpreter frame look like one above,
 723                    // not meaningless information that'll confuse me.
 724 
 725   // flush the windows now. We don't care about the current (protection) frame
 726   // only the outer frames
 727 
 728   __ flushw();
 729 
 730   // mark windows as flushed
 731   Address flags(G2_thread,
 732                 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
 733   __ set(JavaFrameAnchor::flushed, G3_scratch);
 734   __ st(G3_scratch, flags);
 735 
 736   // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
 737 
 738   Address thread_state(G2_thread, in_bytes(JavaThread::thread_state_offset()));
 739 #ifdef ASSERT
 740   { Label L;
 741     __ ld(thread_state, G3_scratch);
 742     __ cmp(G3_scratch, _thread_in_Java);
 743     __ br(Assembler::equal, false, Assembler::pt, L);
 744     __ delayed()->nop();
 745     __ stop("Wrong thread state in native stub");
 746     __ bind(L);
 747   }
 748 #endif // ASSERT
 749   __ set(_thread_in_native, G3_scratch);
 750   __ st(G3_scratch, thread_state);
 751 
 752   // Call the jni method, using the delay slot to set the JNIEnv* argument.
 753   __ callr(O0, 0);
 754   __ delayed()->
 755      add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
 756   __ ld_ptr(STATE(_thread), G2_thread);  // restore thread
 757 
 758   // must we block?
 759 
 760   // Block, if necessary, before resuming in _thread_in_Java state.
 761   // In order for GC to work, don't clear the last_Java_sp until after blocking.
 762   { Label no_block;
 763     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
 764 
 765     // Switch thread to "native transition" state before reading the synchronization state.
 766     // This additional state is necessary because reading and testing the synchronization
 767     // state is not atomic w.r.t. GC, as this scenario demonstrates:
 768     //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
 769     //     VM thread changes sync state to synchronizing and suspends threads for GC.
 770     //     Thread A is resumed to finish this native method, but doesn't block here since it
 771     //     didn't see any synchronization is progress, and escapes.
 772     __ set(_thread_in_native_trans, G3_scratch);
 773     __ st(G3_scratch, thread_state);
 774     if(os::is_MP()) {
 775       // Write serialization page so VM thread can do a pseudo remote membar.
 776       // We use the current thread pointer to calculate a thread specific
 777       // offset to write to within the page. This minimizes bus traffic
 778       // due to cache line collision.
 779       __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
 780     }
 781     __ load_contents(sync_state, G3_scratch);
 782     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 783 
 784 
 785     Label L;
 786     Address suspend_state(G2_thread, in_bytes(JavaThread::suspend_flags_offset()));
 787     __ br(Assembler::notEqual, false, Assembler::pn, L);
 788     __ delayed()->
 789       ld(suspend_state, G3_scratch);
 790     __ cmp(G3_scratch, 0);
 791     __ br(Assembler::equal, false, Assembler::pt, no_block);
 792     __ delayed()->nop();
 793     __ bind(L);
 794 
 795     // Block.  Save any potential method result value before the operation and
 796     // use a leaf call to leave the last_Java_frame setup undisturbed.
 797     save_native_result();
 798     __ call_VM_leaf(noreg,
 799                     CAST_FROM_FN_PTR(address, JavaThread::check_safepoint_and_suspend_for_native_trans),
 800                     G2_thread);
 801     __ ld_ptr(STATE(_thread), G2_thread);  // restore thread
 802     // Restore any method result value
 803     restore_native_result();
 804     __ bind(no_block);
 805   }
 806 
 807   // Clear the frame anchor now
 808 
 809   __ reset_last_Java_frame();
 810 
 811   // Move the result handler address
 812   __ mov(Lscratch, G3_scratch);
 813   // return possible result to the outer frame
 814 #ifndef __LP64
 815   __ mov(O0, I0);
 816   __ restore(O1, G0, O1);
 817 #else
 818   __ restore(O0, G0, O0);
 819 #endif /* __LP64 */
 820 
 821   // Move result handler to expected register
 822   __ mov(G3_scratch, Lscratch);
 823 
 824 
 825   // thread state is thread_in_native_trans. Any safepoint blocking has
 826   // happened in the trampoline we are ready to switch to thread_in_Java.
 827 
 828   __ set(_thread_in_Java, G3_scratch);
 829   __ st(G3_scratch, thread_state);
 830 
 831   // If we have an oop result store it where it will be safe for any further gc
 832   // until we return now that we've released the handle it might be protected by
 833 
 834   {
 835     Label no_oop, store_result;
 836 
 837     __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
 838     __ cmp(G3_scratch, Lscratch);
 839     __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
 840     __ delayed()->nop();
 841     __ addcc(G0, O0, O0);
 842     __ brx(Assembler::notZero, true, Assembler::pt, store_result);     // if result is not NULL:
 843     __ delayed()->ld_ptr(O0, 0, O0);                                   // unbox it
 844     __ mov(G0, O0);
 845 
 846     __ bind(store_result);
 847     // Store it where gc will look for it and result handler expects it.
 848     __ st_ptr(O0, STATE(_oop_temp));
 849 
 850     __ bind(no_oop);
 851 
 852   }
 853 
 854   // reset handle block
 855   __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch);
 856   __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
 857 
 858 
 859   // handle exceptions (exception handling will handle unlocking!)
 860   { Label L;
 861     Address exception_addr (G2_thread, in_bytes(Thread::pending_exception_offset()));
 862 
 863     __ ld_ptr(exception_addr, Gtemp);
 864     __ tst(Gtemp);
 865     __ brx(Assembler::equal, false, Assembler::pt, L);
 866     __ delayed()->nop();
 867     __ bind(pending_exception_present);
 868     // With c++ interpreter we just leave it pending caller will do the correct thing. However...
 869     // Like x86 we ignore the result of the native call and leave the method locked. This
 870     // seems wrong to leave things locked.
 871 
 872     __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
 873     __ delayed()->restore(I5_savedSP, G0, SP);  // remove interpreter frame
 874 
 875     __ bind(L);
 876   }
 877 
 878   // jvmdi/jvmpi support (preserves thread register)
 879   __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
 880 
 881   if (synchronized) {
 882     // save and restore any potential method result value around the unlocking operation
 883     save_native_result();
 884 
 885     const int entry_size            = frame::interpreter_frame_monitor_size() * wordSize;
 886     // Get the initial monitor we allocated
 887     __ sub(Lstate, entry_size, O1);                        // initial monitor
 888     __ unlock_object(O1);
 889     restore_native_result();
 890   }
 891 
 892 #if defined(COMPILER2) && !defined(_LP64)
 893 
 894   // C2 expects long results in G1 we can't tell if we're returning to interpreted
 895   // or compiled so just be safe.
 896 
 897   __ sllx(O0, 32, G1);          // Shift bits into high G1
 898   __ srl (O1, 0, O1);           // Zero extend O1
 899   __ or3 (O1, G1, G1);          // OR 64 bits into G1
 900 
 901 #endif /* COMPILER2 && !_LP64 */
 902 
 903 #ifdef ASSERT
 904   {
 905     Label ok;
 906     __ cmp(I5_savedSP, FP);
 907     __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
 908     __ delayed()->nop();
 909     __ stop("bad I5_savedSP value");
 910     __ should_not_reach_here();
 911     __ bind(ok);
 912   }
 913 #endif
 914   // Calls result handler which POPS FRAME
 915   if (TraceJumps) {
 916     // Move target to register that is recordable
 917     __ mov(Lscratch, G3_scratch);
 918     __ JMP(G3_scratch, 0);
 919   } else {
 920     __ jmp(Lscratch, 0);
 921   }
 922   __ delayed()->nop();
 923 
 924   if (inc_counter) {
 925     // handle invocation counter overflow
 926     __ bind(invocation_counter_overflow);
 927     generate_counter_overflow(Lcontinue);
 928   }
 929 
 930 
 931   return entry;
 932 }
 933 
 934 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state,
 935                                                               const Register prev_state,
 936                                                               bool native) {
 937 
 938   // On entry
 939   // G5_method - caller's method
 940   // Gargs - points to initial parameters (i.e. locals[0])
 941   // G2_thread - valid? (C1 only??)
 942   // "prev_state" - contains any previous frame manager state which we must save a link
 943   //
 944   // On return
 945   // "state" is a pointer to the newly allocated  state object. We must allocate and initialize
 946   // a new interpretState object and the method expression stack.
 947 
 948   assert_different_registers(state, prev_state);
 949   assert_different_registers(prev_state, G3_scratch);
 950   const Register Gtmp = G3_scratch;
 951   const Address constMethod       (G5_method, in_bytes(Method::const_offset()));
 952   const Address access_flags      (G5_method, in_bytes(Method::access_flags_offset()));
 953 
 954   // slop factor is two extra slots on the expression stack so that
 955   // we always have room to store a result when returning from a call without parameters
 956   // that returns a result.
 957 
 958   const int slop_factor = 2*wordSize;
 959 
 960   const int fixed_size = ((sizeof(BytecodeInterpreter) + slop_factor) >> LogBytesPerWord) + // what is the slop factor?
 961                          Method::extra_stack_entries() + // extra stack for jsr 292
 962                          frame::memory_parameter_word_sp_offset +  // register save area + param window
 963                          (native ?  frame::interpreter_frame_extra_outgoing_argument_words : 0); // JNI, class
 964 
 965   // XXX G5_method valid
 966 
 967   // Now compute new frame size
 968 
 969   if (native) {
 970     const Register RconstMethod = Gtmp;
 971     const Address size_of_parameters(RconstMethod, in_bytes(ConstMethod::size_of_parameters_offset()));
 972     __ ld_ptr(constMethod, RconstMethod);
 973     __ lduh( size_of_parameters, Gtmp );
 974     __ calc_mem_param_words(Gtmp, Gtmp);     // space for native call parameters passed on the stack in words
 975   } else {
 976     // Full size expression stack
 977     __ ld_ptr(constMethod, Gtmp);
 978     __ lduh(Gtmp, in_bytes(ConstMethod::max_stack_offset()), Gtmp);
 979   }
 980   __ add(Gtmp, fixed_size, Gtmp);           // plus the fixed portion
 981 
 982   __ neg(Gtmp);                               // negative space for stack/parameters in words
 983   __ and3(Gtmp, -WordsPerLong, Gtmp);        // make multiple of 2 (SP must be 2-word aligned)
 984   __ sll(Gtmp, LogBytesPerWord, Gtmp);       // negative space for frame in bytes
 985 
 986   // Need to do stack size check here before we fault on large frames
 987 
 988   Label stack_ok;
 989 
 990   const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages :
 991                                                                               (StackRedPages+StackYellowPages);
 992 
 993 
 994   __ ld_ptr(G2_thread, in_bytes(Thread::stack_base_offset()), O0);
 995   __ ld_ptr(G2_thread, in_bytes(Thread::stack_size_offset()), O1);
 996   // compute stack bottom
 997   __ sub(O0, O1, O0);
 998 
 999   // Avoid touching the guard pages
1000   // Also a fudge for frame size of BytecodeInterpreter::run
1001   // It varies from 1k->4k depending on build type
1002   const int fudge = 6 * K;
1003 
1004   __ set(fudge + (max_pages * os::vm_page_size()), O1);
1005 
1006   __ add(O0, O1, O0);
1007   __ sub(O0, Gtmp, O0);
1008   __ cmp(SP, O0);
1009   __ brx(Assembler::greaterUnsigned, false, Assembler::pt, stack_ok);
1010   __ delayed()->nop();
1011 
1012      // throw exception return address becomes throwing pc
1013 
1014   __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
1015   __ stop("never reached");
1016 
1017   __ bind(stack_ok);
1018 
1019   __ save(SP, Gtmp, SP);                      // setup new frame and register window
1020 
1021   // New window I7 call_stub or previous activation
1022   // O6 - register save area, BytecodeInterpreter just below it, args/locals just above that
1023   //
1024   __ sub(FP, sizeof(BytecodeInterpreter), state);        // Point to new Interpreter state
1025   __ add(state, STACK_BIAS, state );         // Account for 64bit bias
1026 
1027 #define XXX_STATE(field_name) state, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
1028 
1029   // Initialize a new Interpreter state
1030   // orig_sp - caller's original sp
1031   // G2_thread - thread
1032   // Gargs - &locals[0] (unbiased?)
1033   // G5_method - method
1034   // SP (biased) - accounts for full size java stack, BytecodeInterpreter object, register save area, and register parameter save window
1035 
1036 
1037   __ set(0xdead0004, O1);
1038 
1039 
1040   __ st_ptr(Gargs, XXX_STATE(_locals));
1041   __ st_ptr(G0, XXX_STATE(_oop_temp));
1042 
1043   __ st_ptr(state, XXX_STATE(_self_link));                // point to self
1044   __ st_ptr(prev_state->after_save(), XXX_STATE(_prev_link)); // Chain interpreter states
1045   __ st_ptr(G2_thread, XXX_STATE(_thread));               // Store javathread
1046 
1047   if (native) {
1048     __ st_ptr(G0, XXX_STATE(_bcp));
1049   } else {
1050     __ ld_ptr(G5_method, in_bytes(Method::const_offset()), O2); // get ConstMethod*
1051     __ add(O2, in_bytes(ConstMethod::codes_offset()), O2);        // get bcp
1052     __ st_ptr(O2, XXX_STATE(_bcp));
1053   }
1054 
1055   __ st_ptr(G0, XXX_STATE(_mdx));
1056   __ st_ptr(G5_method, XXX_STATE(_method));
1057 
1058   __ set((int) BytecodeInterpreter::method_entry, O1);
1059   __ st(O1, XXX_STATE(_msg));
1060 
1061   __ ld_ptr(constMethod, O3);
1062   __ ld_ptr(O3, in_bytes(ConstMethod::constants_offset()), O3);
1063   __ ld_ptr(O3, ConstantPool::cache_offset_in_bytes(), O2);
1064   __ st_ptr(O2, XXX_STATE(_constants));
1065 
1066   __ st_ptr(G0, XXX_STATE(_result._to_call._callee));
1067 
1068   // Monitor base is just start of BytecodeInterpreter object;
1069   __ mov(state, O2);
1070   __ st_ptr(O2, XXX_STATE(_monitor_base));
1071 
1072   // Do we need a monitor for synchonized method?
1073   {
1074     __ ld(access_flags, O1);
1075     Label done;
1076     Label got_obj;
1077     __ btst(JVM_ACC_SYNCHRONIZED, O1);
1078     __ br( Assembler::zero, false, Assembler::pt, done);
1079 
1080     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
1081     __ delayed()->btst(JVM_ACC_STATIC, O1);
1082     __ ld_ptr(XXX_STATE(_locals), O1);
1083     __ br( Assembler::zero, true, Assembler::pt, got_obj);
1084     __ delayed()->ld_ptr(O1, 0, O1);                  // get receiver for not-static case
1085     __ ld_ptr(constMethod, O1);
1086     __ ld_ptr( O1, in_bytes(ConstMethod::constants_offset()), O1);
1087     __ ld_ptr( O1, ConstantPool::pool_holder_offset_in_bytes(), O1);
1088     // lock the mirror, not the Klass*
1089     __ ld_ptr( O1, mirror_offset, O1);
1090 
1091     __ bind(got_obj);
1092 
1093   #ifdef ASSERT
1094     __ tst(O1);
1095     __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
1096   #endif // ASSERT
1097 
1098     const int entry_size            = frame::interpreter_frame_monitor_size() * wordSize;
1099     __ sub(SP, entry_size, SP);                         // account for initial monitor
1100     __ sub(O2, entry_size, O2);                        // initial monitor
1101     __ st_ptr(O1, O2, BasicObjectLock::obj_offset_in_bytes()); // and allocate it for interpreter use
1102     __ bind(done);
1103   }
1104 
1105   // Remember initial frame bottom
1106 
1107   __ st_ptr(SP, XXX_STATE(_frame_bottom));
1108 
1109   __ st_ptr(O2, XXX_STATE(_stack_base));
1110 
1111   __ sub(O2, wordSize, O2);                    // prepush
1112   __ st_ptr(O2, XXX_STATE(_stack));                // PREPUSH
1113 
1114   // Full size expression stack
1115   __ ld_ptr(constMethod, O3);
1116   __ lduh(O3, in_bytes(ConstMethod::max_stack_offset()), O3);
1117   __ inc(O3, Method::extra_stack_entries());
1118   __ sll(O3, LogBytesPerWord, O3);
1119   __ sub(O2, O3, O3);
1120 //  __ sub(O3, wordSize, O3);                    // so prepush doesn't look out of bounds
1121   __ st_ptr(O3, XXX_STATE(_stack_limit));
1122 
1123   if (!native) {
1124     //
1125     // Code to initialize locals
1126     //
1127     Register init_value = noreg;    // will be G0 if we must clear locals
1128     // Now zero locals
1129     if (true /* zerolocals */ || ClearInterpreterLocals) {
1130       // explicitly initialize locals
1131       init_value = G0;
1132     } else {
1133     #ifdef ASSERT
1134       // initialize locals to a garbage pattern for better debugging
1135       init_value = O3;
1136       __ set( 0x0F0F0F0F, init_value );
1137     #endif // ASSERT
1138     }
1139     if (init_value != noreg) {
1140       Label clear_loop;
1141       const Register RconstMethod = O1;
1142       const Address size_of_parameters(RconstMethod, in_bytes(ConstMethod::size_of_parameters_offset()));
1143       const Address size_of_locals    (RconstMethod, in_bytes(ConstMethod::size_of_locals_offset()));
1144 
1145       // NOTE: If you change the frame layout, this code will need to
1146       // be updated!
1147       __ ld_ptr( constMethod, RconstMethod );
1148       __ lduh( size_of_locals, O2 );
1149       __ lduh( size_of_parameters, O1 );
1150       __ sll( O2, LogBytesPerWord, O2);
1151       __ sll( O1, LogBytesPerWord, O1 );
1152       __ ld_ptr(XXX_STATE(_locals), L2_scratch);
1153       __ sub( L2_scratch, O2, O2 );
1154       __ sub( L2_scratch, O1, O1 );
1155 
1156       __ bind( clear_loop );
1157       __ inc( O2, wordSize );
1158 
1159       __ cmp( O2, O1 );
1160       __ br( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1161       __ delayed()->st_ptr( init_value, O2, 0 );
1162     }
1163   }
1164 }
1165 // Find preallocated  monitor and lock method (C++ interpreter)
1166 //
1167 void InterpreterGenerator::lock_method(void) {
1168 // Lock the current method.
1169 // Destroys registers L2_scratch, L3_scratch, O0
1170 //
1171 // Find everything relative to Lstate
1172 
1173 #ifdef ASSERT
1174   __ ld_ptr(STATE(_method), L2_scratch);
1175   __ ld(L2_scratch, in_bytes(Method::access_flags_offset()), O0);
1176 
1177  { Label ok;
1178    __ btst(JVM_ACC_SYNCHRONIZED, O0);
1179    __ br( Assembler::notZero, false, Assembler::pt, ok);
1180    __ delayed()->nop();
1181    __ stop("method doesn't need synchronization");
1182    __ bind(ok);
1183   }
1184 #endif // ASSERT
1185 
1186   // monitor is already allocated at stack base
1187   // and the lockee is already present
1188   __ ld_ptr(STATE(_stack_base), L2_scratch);
1189   __ ld_ptr(L2_scratch, BasicObjectLock::obj_offset_in_bytes(), O0);   // get object
1190   __ lock_object(L2_scratch, O0);
1191 
1192 }
1193 
1194 //  Generate code for handling resuming a deopted method
1195 void CppInterpreterGenerator::generate_deopt_handling() {
1196 
1197   Label return_from_deopt_common;
1198 
1199   // deopt needs to jump to here to enter the interpreter (return a result)
1200   deopt_frame_manager_return_atos  = __ pc();
1201 
1202   // O0/O1 live
1203   __ ba(return_from_deopt_common);
1204   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_OBJECT), L3_scratch);    // Result stub address array index
1205 
1206 
1207   // deopt needs to jump to here to enter the interpreter (return a result)
1208   deopt_frame_manager_return_btos  = __ pc();
1209 
1210   // O0/O1 live
1211   __ ba(return_from_deopt_common);
1212   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_BOOLEAN), L3_scratch);    // Result stub address array index
1213 
1214   // deopt needs to jump to here to enter the interpreter (return a result)
1215   deopt_frame_manager_return_itos  = __ pc();
1216 
1217   // O0/O1 live
1218   __ ba(return_from_deopt_common);
1219   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_INT), L3_scratch);    // Result stub address array index
1220 
1221   // deopt needs to jump to here to enter the interpreter (return a result)
1222 
1223   deopt_frame_manager_return_ltos  = __ pc();
1224 #if !defined(_LP64) && defined(COMPILER2)
1225   // All return values are where we want them, except for Longs.  C2 returns
1226   // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
1227   // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
1228   // build even if we are returning from interpreted we just do a little
1229   // stupid shuffing.
1230   // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
1231   // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
1232   // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
1233 
1234   __ srl (G1, 0,O1);
1235   __ srlx(G1,32,O0);
1236 #endif /* !_LP64 && COMPILER2 */
1237   // O0/O1 live
1238   __ ba(return_from_deopt_common);
1239   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_LONG), L3_scratch);    // Result stub address array index
1240 
1241   // deopt needs to jump to here to enter the interpreter (return a result)
1242 
1243   deopt_frame_manager_return_ftos  = __ pc();
1244   // O0/O1 live
1245   __ ba(return_from_deopt_common);
1246   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_FLOAT), L3_scratch);    // Result stub address array index
1247 
1248   // deopt needs to jump to here to enter the interpreter (return a result)
1249   deopt_frame_manager_return_dtos  = __ pc();
1250 
1251   // O0/O1 live
1252   __ ba(return_from_deopt_common);
1253   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_DOUBLE), L3_scratch);    // Result stub address array index
1254 
1255   // deopt needs to jump to here to enter the interpreter (return a result)
1256   deopt_frame_manager_return_vtos  = __ pc();
1257 
1258   // O0/O1 live
1259   __ set(AbstractInterpreter::BasicType_as_index(T_VOID), L3_scratch);
1260 
1261   // Deopt return common
1262   // an index is present that lets us move any possible result being
1263   // return to the interpreter's stack
1264   //
1265   __ bind(return_from_deopt_common);
1266 
1267   // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1268   // stack is in the state that the  calling convention left it.
1269   // Copy the result from native abi result and place it on java expression stack.
1270 
1271   // Current interpreter state is present in Lstate
1272 
1273   // Get current pre-pushed top of interpreter stack
1274   // Any result (if any) is in native abi
1275   // result type index is in L3_scratch
1276 
1277   __ ld_ptr(STATE(_stack), L1_scratch);                                          // get top of java expr stack
1278 
1279   __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1280   __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1281   __ ld_ptr(L4_scratch, L3_scratch, Lscratch);                                       // get typed result converter address
1282   __ jmpl(Lscratch, G0, O7);                                         // and convert it
1283   __ delayed()->nop();
1284 
1285   // L1_scratch points to top of stack (prepushed)
1286   __ st_ptr(L1_scratch, STATE(_stack));
1287 }
1288 
1289 // Generate the code to handle a more_monitors message from the c++ interpreter
1290 void CppInterpreterGenerator::generate_more_monitors() {
1291 
1292   Label entry, loop;
1293   const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1294   // 1. compute new pointers                                // esp: old expression stack top
1295   __ delayed()->ld_ptr(STATE(_stack_base), L4_scratch);            // current expression stack bottom
1296   __ sub(L4_scratch, entry_size, L4_scratch);
1297   __ st_ptr(L4_scratch, STATE(_stack_base));
1298 
1299   __ sub(SP, entry_size, SP);                  // Grow stack
1300   __ st_ptr(SP, STATE(_frame_bottom));
1301 
1302   __ ld_ptr(STATE(_stack_limit), L2_scratch);
1303   __ sub(L2_scratch, entry_size, L2_scratch);
1304   __ st_ptr(L2_scratch, STATE(_stack_limit));
1305 
1306   __ ld_ptr(STATE(_stack), L1_scratch);                // Get current stack top
1307   __ sub(L1_scratch, entry_size, L1_scratch);
1308   __ st_ptr(L1_scratch, STATE(_stack));
1309   __ ba(entry);
1310   __ delayed()->add(L1_scratch, wordSize, L1_scratch);        // first real entry (undo prepush)
1311 
1312   // 2. move expression stack
1313 
1314   __ bind(loop);
1315   __ st_ptr(L3_scratch, Address(L1_scratch, 0));
1316   __ add(L1_scratch, wordSize, L1_scratch);
1317   __ bind(entry);
1318   __ cmp(L1_scratch, L4_scratch);
1319   __ br(Assembler::notEqual, false, Assembler::pt, loop);
1320   __ delayed()->ld_ptr(L1_scratch, entry_size, L3_scratch);
1321 
1322   // now zero the slot so we can find it.
1323   __ st_ptr(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes());
1324 
1325 }
1326 
1327 // Initial entry to C++ interpreter from the call_stub.
1328 // This entry point is called the frame manager since it handles the generation
1329 // of interpreter activation frames via requests directly from the vm (via call_stub)
1330 // and via requests from the interpreter. The requests from the call_stub happen
1331 // directly thru the entry point. Requests from the interpreter happen via returning
1332 // from the interpreter and examining the message the interpreter has returned to
1333 // the frame manager. The frame manager can take the following requests:
1334 
1335 // NO_REQUEST - error, should never happen.
1336 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and
1337 //                 allocate a new monitor.
1338 // CALL_METHOD - setup a new activation to call a new method. Very similar to what
1339 //               happens during entry during the entry via the call stub.
1340 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub.
1341 //
1342 // Arguments:
1343 //
1344 // ebx: Method*
1345 // ecx: receiver - unused (retrieved from stack as needed)
1346 // esi: previous frame manager state (NULL from the call_stub/c1/c2)
1347 //
1348 //
1349 // Stack layout at entry
1350 //
1351 // [ return address     ] <--- esp
1352 // [ parameter n        ]
1353 //   ...
1354 // [ parameter 1        ]
1355 // [ expression stack   ]
1356 //
1357 //
1358 // We are free to blow any registers we like because the call_stub which brought us here
1359 // initially has preserved the callee save registers already.
1360 //
1361 //
1362 
1363 static address interpreter_frame_manager = NULL;
1364 
1365 #ifdef ASSERT
1366   #define VALIDATE_STATE(scratch, marker)                         \
1367   {                                                               \
1368     Label skip;                                                   \
1369     __ ld_ptr(STATE(_self_link), scratch);                        \
1370     __ cmp(Lstate, scratch);                                      \
1371     __ brx(Assembler::equal, false, Assembler::pt, skip);         \
1372     __ delayed()->nop();                                          \
1373     __ breakpoint_trap();                                         \
1374     __ emit_int32(marker);                                         \
1375     __ bind(skip);                                                \
1376   }
1377 #else
1378   #define VALIDATE_STATE(scratch, marker)
1379 #endif /* ASSERT */
1380 
1381 void CppInterpreterGenerator::adjust_callers_stack(Register args) {
1382 //
1383 // Adjust caller's stack so that all the locals can be contiguous with
1384 // the parameters.
1385 // Worries about stack overflow make this a pain.
1386 //
1387 // Destroys args, G3_scratch, G3_scratch
1388 // In/Out O5_savedSP (sender's original SP)
1389 //
1390 //  assert_different_registers(state, prev_state);
1391   const Register Gtmp = G3_scratch;
1392   const Register RconstMethod = G3_scratch;
1393   const Register tmp = O2;
1394   const Address constMethod(G5_method, in_bytes(Method::const_offset()));
1395   const Address size_of_parameters(RconstMethod, in_bytes(ConstMethod::size_of_parameters_offset()));
1396   const Address size_of_locals    (RconstMethod, in_bytes(ConstMethod::size_of_locals_offset()));
1397 
1398   __ ld_ptr(constMethod, RconstMethod);
1399   __ lduh(size_of_parameters, tmp);
1400   __ sll(tmp, LogBytesPerWord, Gargs);       // parameter size in bytes
1401   __ add(args, Gargs, Gargs);                // points to first local + BytesPerWord
1402   // NEW
1403   __ add(Gargs, -wordSize, Gargs);             // points to first local[0]
1404   // determine extra space for non-argument locals & adjust caller's SP
1405   // Gtmp1: parameter size in words
1406   __ lduh(size_of_locals, Gtmp);
1407   __ compute_extra_locals_size_in_bytes(tmp, Gtmp, Gtmp);
1408 
1409 #if 1
1410   // c2i adapters place the final interpreter argument in the register save area for O0/I0
1411   // the call_stub will place the final interpreter argument at
1412   // frame::memory_parameter_word_sp_offset. This is mostly not noticable for either asm
1413   // or c++ interpreter. However with the c++ interpreter when we do a recursive call
1414   // and try to make it look good in the debugger we will store the argument to
1415   // RecursiveInterpreterActivation in the register argument save area. Without allocating
1416   // extra space for the compiler this will overwrite locals in the local array of the
1417   // interpreter.
1418   // QQQ still needed with frameless adapters???
1419 
1420   const int c2i_adjust_words = frame::memory_parameter_word_sp_offset - frame::callee_register_argument_save_area_sp_offset;
1421 
1422   __ add(Gtmp, c2i_adjust_words*wordSize, Gtmp);
1423 #endif // 1
1424 
1425 
1426   __ sub(SP, Gtmp, SP);                      // just caller's frame for the additional space we need.
1427 }
1428 
1429 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1430 
1431   // G5_method: Method*
1432   // G2_thread: thread (unused)
1433   // Gargs:   bottom of args (sender_sp)
1434   // O5: sender's sp
1435 
1436   // A single frame manager is plenty as we don't specialize for synchronized. We could and
1437   // the code is pretty much ready. Would need to change the test below and for good measure
1438   // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized
1439   // routines. Not clear this is worth it yet.
1440 
1441   if (interpreter_frame_manager) {
1442     return interpreter_frame_manager;
1443   }
1444 
1445   __ bind(frame_manager_entry);
1446 
1447   // the following temporary registers are used during frame creation
1448   const Register Gtmp1 = G3_scratch;
1449   // const Register Lmirror = L1;     // native mirror (native calls only)
1450 
1451   const Address constMethod       (G5_method, in_bytes(Method::const_offset()));
1452   const Address access_flags      (G5_method, in_bytes(Method::access_flags_offset()));
1453 
1454   address entry_point = __ pc();
1455   __ mov(G0, prevState);                                                 // no current activation
1456 
1457 
1458   Label re_dispatch;
1459 
1460   __ bind(re_dispatch);
1461 
1462   // Interpreter needs to have locals completely contiguous. In order to do that
1463   // We must adjust the caller's stack pointer for any locals beyond just the
1464   // parameters
1465   adjust_callers_stack(Gargs);
1466 
1467   // O5_savedSP still contains sender's sp
1468 
1469   // NEW FRAME
1470 
1471   generate_compute_interpreter_state(Lstate, prevState, false);
1472 
1473   // At this point a new interpreter frame and state object are created and initialized
1474   // Lstate has the pointer to the new activation
1475   // Any stack banging or limit check should already be done.
1476 
1477   Label call_interpreter;
1478 
1479   __ bind(call_interpreter);
1480 
1481 
1482 #if 1
1483   __ set(0xdead002, Lmirror);
1484   __ set(0xdead002, L2_scratch);
1485   __ set(0xdead003, L3_scratch);
1486   __ set(0xdead004, L4_scratch);
1487   __ set(0xdead005, Lscratch);
1488   __ set(0xdead006, Lscratch2);
1489   __ set(0xdead007, L7_scratch);
1490 
1491   __ set(0xdeaf002, O2);
1492   __ set(0xdeaf003, O3);
1493   __ set(0xdeaf004, O4);
1494   __ set(0xdeaf005, O5);
1495 #endif
1496 
1497   // Call interpreter (stack bang complete) enter here if message is
1498   // set and we know stack size is valid
1499 
1500   Label call_interpreter_2;
1501 
1502   __ bind(call_interpreter_2);
1503 
1504 #ifdef ASSERT
1505   {
1506     Label skip;
1507     __ ld_ptr(STATE(_frame_bottom), G3_scratch);
1508     __ cmp(G3_scratch, SP);
1509     __ brx(Assembler::equal, false, Assembler::pt, skip);
1510     __ delayed()->nop();
1511     __ stop("SP not restored to frame bottom");
1512     __ bind(skip);
1513   }
1514 #endif
1515 
1516   VALIDATE_STATE(G3_scratch, 4);
1517   __ set_last_Java_frame(SP, noreg);
1518   __ mov(Lstate, O0);                 // (arg) pointer to current state
1519 
1520   __ call(CAST_FROM_FN_PTR(address,
1521                            JvmtiExport::can_post_interpreter_events() ?
1522                                                                   BytecodeInterpreter::runWithChecks
1523                                                                 : BytecodeInterpreter::run),
1524          relocInfo::runtime_call_type);
1525 
1526   __ delayed()->nop();
1527 
1528   __ ld_ptr(STATE(_thread), G2_thread);
1529   __ reset_last_Java_frame();
1530 
1531   // examine msg from interpreter to determine next action
1532   __ ld_ptr(STATE(_thread), G2_thread);                                  // restore G2_thread
1533 
1534   __ ld(STATE(_msg), L1_scratch);                                       // Get new message
1535 
1536   Label call_method;
1537   Label return_from_interpreted_method;
1538   Label throw_exception;
1539   Label do_OSR;
1540   Label bad_msg;
1541   Label resume_interpreter;
1542 
1543   __ cmp(L1_scratch, (int)BytecodeInterpreter::call_method);
1544   __ br(Assembler::equal, false, Assembler::pt, call_method);
1545   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::return_from_method);
1546   __ br(Assembler::equal, false, Assembler::pt, return_from_interpreted_method);
1547   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::throwing_exception);
1548   __ br(Assembler::equal, false, Assembler::pt, throw_exception);
1549   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::do_osr);
1550   __ br(Assembler::equal, false, Assembler::pt, do_OSR);
1551   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::more_monitors);
1552   __ br(Assembler::notEqual, false, Assembler::pt, bad_msg);
1553 
1554   // Allocate more monitor space, shuffle expression stack....
1555 
1556   generate_more_monitors();
1557 
1558   // new monitor slot allocated, resume the interpreter.
1559 
1560   __ set((int)BytecodeInterpreter::got_monitors, L1_scratch);
1561   VALIDATE_STATE(G3_scratch, 5);
1562   __ ba(call_interpreter);
1563   __ delayed()->st(L1_scratch, STATE(_msg));
1564 
1565   // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode)
1566   unctrap_frame_manager_entry  = __ pc();
1567 
1568   // QQQ what message do we send
1569 
1570   __ ba(call_interpreter);
1571   __ delayed()->ld_ptr(STATE(_frame_bottom), SP);                  // restore to full stack frame
1572 
1573   //=============================================================================
1574   // Returning from a compiled method into a deopted method. The bytecode at the
1575   // bcp has completed. The result of the bytecode is in the native abi (the tosca
1576   // for the template based interpreter). Any stack space that was used by the
1577   // bytecode that has completed has been removed (e.g. parameters for an invoke)
1578   // so all that we have to do is place any pending result on the expression stack
1579   // and resume execution on the next bytecode.
1580 
1581   generate_deopt_handling();
1582 
1583   // ready to resume the interpreter
1584 
1585   __ set((int)BytecodeInterpreter::deopt_resume, L1_scratch);
1586   __ ba(call_interpreter);
1587   __ delayed()->st(L1_scratch, STATE(_msg));
1588 
1589   // Current frame has caught an exception we need to dispatch to the
1590   // handler. We can get here because a native interpreter frame caught
1591   // an exception in which case there is no handler and we must rethrow
1592   // If it is a vanilla interpreted frame the we simply drop into the
1593   // interpreter and let it do the lookup.
1594 
1595   Interpreter::_rethrow_exception_entry = __ pc();
1596 
1597   Label return_with_exception;
1598   Label unwind_and_forward;
1599 
1600   // O0: exception
1601   // O7: throwing pc
1602 
1603   // We want exception in the thread no matter what we ultimately decide about frame type.
1604 
1605   Address exception_addr (G2_thread, in_bytes(Thread::pending_exception_offset()));
1606   __ verify_thread();
1607   __ st_ptr(O0, exception_addr);
1608 
1609   // get the Method*
1610   __ ld_ptr(STATE(_method), G5_method);
1611 
1612   // if this current frame vanilla or native?
1613 
1614   __ ld(access_flags, Gtmp1);
1615   __ btst(JVM_ACC_NATIVE, Gtmp1);
1616   __ br(Assembler::zero, false, Assembler::pt, return_with_exception);  // vanilla interpreted frame handle directly
1617   __ delayed()->nop();
1618 
1619   // We drop thru to unwind a native interpreted frame with a pending exception
1620   // We jump here for the initial interpreter frame with exception pending
1621   // We unwind the current acivation and forward it to our caller.
1622 
1623   __ bind(unwind_and_forward);
1624 
1625   // Unwind frame and jump to forward exception. unwinding will place throwing pc in O7
1626   // as expected by forward_exception.
1627 
1628   __ restore(FP, G0, SP);                  // unwind interpreter state frame
1629   __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
1630   __ delayed()->mov(I5_savedSP->after_restore(), SP);
1631 
1632   // Return point from a call which returns a result in the native abi
1633   // (c1/c2/jni-native). This result must be processed onto the java
1634   // expression stack.
1635   //
1636   // A pending exception may be present in which case there is no result present
1637 
1638   address return_from_native_method = __ pc();
1639 
1640   VALIDATE_STATE(G3_scratch, 6);
1641 
1642   // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1643   // stack is in the state that the  calling convention left it.
1644   // Copy the result from native abi result and place it on java expression stack.
1645 
1646   // Current interpreter state is present in Lstate
1647 
1648   // Exception pending?
1649 
1650   __ ld_ptr(STATE(_frame_bottom), SP);                             // restore to full stack frame
1651   __ ld_ptr(exception_addr, Lscratch);                                         // get any pending exception
1652   __ tst(Lscratch);                                                            // exception pending?
1653   __ brx(Assembler::notZero, false, Assembler::pt, return_with_exception);
1654   __ delayed()->nop();
1655 
1656   // Process the native abi result to java expression stack
1657 
1658   __ ld_ptr(STATE(_result._to_call._callee), L4_scratch);                        // called method
1659   __ ld_ptr(STATE(_stack), L1_scratch);                                          // get top of java expr stack
1660   // get parameter size
1661   __ ld_ptr(L4_scratch, in_bytes(Method::const_offset()), L2_scratch);
1662   __ lduh(L2_scratch, in_bytes(ConstMethod::size_of_parameters_offset()), L2_scratch);
1663   __ sll(L2_scratch, LogBytesPerWord, L2_scratch     );                           // parameter size in bytes
1664   __ add(L1_scratch, L2_scratch, L1_scratch);                                      // stack destination for result
1665   __ ld(L4_scratch, in_bytes(Method::result_index_offset()), L3_scratch); // called method result type index
1666 
1667   // tosca is really just native abi
1668   __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1669   __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1670   __ ld_ptr(L4_scratch, L3_scratch, Lscratch);                                       // get typed result converter address
1671   __ jmpl(Lscratch, G0, O7);                                                   // and convert it
1672   __ delayed()->nop();
1673 
1674   // L1_scratch points to top of stack (prepushed)
1675 
1676   __ ba(resume_interpreter);
1677   __ delayed()->mov(L1_scratch, O1);
1678 
1679   // An exception is being caught on return to a vanilla interpreter frame.
1680   // Empty the stack and resume interpreter
1681 
1682   __ bind(return_with_exception);
1683 
1684   __ ld_ptr(STATE(_frame_bottom), SP);                             // restore to full stack frame
1685   __ ld_ptr(STATE(_stack_base), O1);                               // empty java expression stack
1686   __ ba(resume_interpreter);
1687   __ delayed()->sub(O1, wordSize, O1);                             // account for prepush
1688 
1689   // Return from interpreted method we return result appropriate to the caller (i.e. "recursive"
1690   // interpreter call, or native) and unwind this interpreter activation.
1691   // All monitors should be unlocked.
1692 
1693   __ bind(return_from_interpreted_method);
1694 
1695   VALIDATE_STATE(G3_scratch, 7);
1696 
1697   Label return_to_initial_caller;
1698 
1699   // Interpreted result is on the top of the completed activation expression stack.
1700   // We must return it to the top of the callers stack if caller was interpreted
1701   // otherwise we convert to native abi result and return to call_stub/c1/c2
1702   // The caller's expression stack was truncated by the call however the current activation
1703   // has enough stuff on the stack that we have usable space there no matter what. The
1704   // other thing that makes it easy is that the top of the caller's stack is stored in STATE(_locals)
1705   // for the current activation
1706 
1707   __ ld_ptr(STATE(_prev_link), L1_scratch);
1708   __ ld_ptr(STATE(_method), L2_scratch);                               // get method just executed
1709   __ ld(L2_scratch, in_bytes(Method::result_index_offset()), L2_scratch);
1710   __ tst(L1_scratch);
1711   __ brx(Assembler::zero, false, Assembler::pt, return_to_initial_caller);
1712   __ delayed()->sll(L2_scratch, LogBytesPerWord, L2_scratch);
1713 
1714   // Copy result to callers java stack
1715 
1716   __ set((intptr_t)CppInterpreter::_stack_to_stack, L4_scratch);
1717   __ ld_ptr(L4_scratch, L2_scratch, Lscratch);                          // get typed result converter address
1718   __ ld_ptr(STATE(_stack), O0);                                       // current top (prepushed)
1719   __ ld_ptr(STATE(_locals), O1);                                      // stack destination
1720 
1721   // O0 - will be source, O1 - will be destination (preserved)
1722   __ jmpl(Lscratch, G0, O7);                                          // and convert it
1723   __ delayed()->add(O0, wordSize, O0);                                // get source (top of current expr stack)
1724 
1725   // O1 == &locals[0]
1726 
1727   // Result is now on caller's stack. Just unwind current activation and resume
1728 
1729   Label unwind_recursive_activation;
1730 
1731 
1732   __ bind(unwind_recursive_activation);
1733 
1734   // O1 == &locals[0] (really callers stacktop) for activation now returning
1735   // returning to interpreter method from "recursive" interpreter call
1736   // result converter left O1 pointing to top of the( prepushed) java stack for method we are returning
1737   // to. Now all we must do is unwind the state from the completed call
1738 
1739   // Must restore stack
1740   VALIDATE_STATE(G3_scratch, 8);
1741 
1742   // Return to interpreter method after a method call (interpreted/native/c1/c2) has completed.
1743   // Result if any is already on the caller's stack. All we must do now is remove the now dead
1744   // frame and tell interpreter to resume.
1745 
1746 
1747   __ mov(O1, I1);                                                     // pass back new stack top across activation
1748   // POP FRAME HERE ==================================
1749   __ restore(FP, G0, SP);                                             // unwind interpreter state frame
1750   __ ld_ptr(STATE(_frame_bottom), SP);                                // restore to full stack frame
1751 
1752 
1753   // Resume the interpreter. The current frame contains the current interpreter
1754   // state object.
1755   //
1756   // O1 == new java stack pointer
1757 
1758   __ bind(resume_interpreter);
1759   VALIDATE_STATE(G3_scratch, 10);
1760 
1761   // A frame we have already used before so no need to bang stack so use call_interpreter_2 entry
1762 
1763   __ set((int)BytecodeInterpreter::method_resume, L1_scratch);
1764   __ st(L1_scratch, STATE(_msg));
1765   __ ba(call_interpreter_2);
1766   __ delayed()->st_ptr(O1, STATE(_stack));
1767 
1768   // interpreter returning to native code (call_stub/c1/c2)
1769   // convert result and unwind initial activation
1770   // L2_scratch - scaled result type index
1771 
1772   __ bind(return_to_initial_caller);
1773 
1774   __ set((intptr_t)CppInterpreter::_stack_to_native_abi, L4_scratch);
1775   __ ld_ptr(L4_scratch, L2_scratch, Lscratch);                           // get typed result converter address
1776   __ ld_ptr(STATE(_stack), O0);                                        // current top (prepushed)
1777   __ jmpl(Lscratch, G0, O7);                                           // and convert it
1778   __ delayed()->add(O0, wordSize, O0);                                 // get source (top of current expr stack)
1779 
1780   Label unwind_initial_activation;
1781   __ bind(unwind_initial_activation);
1782 
1783   // RETURN TO CALL_STUB/C1/C2 code (result if any in I0..I1/(F0/..F1)
1784   // we can return here with an exception that wasn't handled by interpreted code
1785   // how does c1/c2 see it on return?
1786 
1787   // compute resulting sp before/after args popped depending upon calling convention
1788   // __ ld_ptr(STATE(_saved_sp), Gtmp1);
1789   //
1790   // POP FRAME HERE ==================================
1791   __ restore(FP, G0, SP);
1792   __ retl();
1793   __ delayed()->mov(I5_savedSP->after_restore(), SP);
1794 
1795   // OSR request, unwind the current frame and transfer to the OSR entry
1796   // and enter OSR nmethod
1797 
1798   __ bind(do_OSR);
1799   Label remove_initial_frame;
1800   __ ld_ptr(STATE(_prev_link), L1_scratch);
1801   __ ld_ptr(STATE(_result._osr._osr_buf), G1_scratch);
1802 
1803   // We are going to pop this frame. Is there another interpreter frame underneath
1804   // it or is it callstub/compiled?
1805 
1806   __ tst(L1_scratch);
1807   __ brx(Assembler::zero, false, Assembler::pt, remove_initial_frame);
1808   __ delayed()->ld_ptr(STATE(_result._osr._osr_entry), G3_scratch);
1809 
1810   // Frame underneath is an interpreter frame simply unwind
1811   // POP FRAME HERE ==================================
1812   __ restore(FP, G0, SP);                                             // unwind interpreter state frame
1813   __ mov(I5_savedSP->after_restore(), SP);
1814 
1815   // Since we are now calling native need to change our "return address" from the
1816   // dummy RecursiveInterpreterActivation to a return from native
1817 
1818   __ set((intptr_t)return_from_native_method - 8, O7);
1819 
1820   __ jmpl(G3_scratch, G0, G0);
1821   __ delayed()->mov(G1_scratch, O0);
1822 
1823   __ bind(remove_initial_frame);
1824 
1825   // POP FRAME HERE ==================================
1826   __ restore(FP, G0, SP);
1827   __ mov(I5_savedSP->after_restore(), SP);
1828   __ jmpl(G3_scratch, G0, G0);
1829   __ delayed()->mov(G1_scratch, O0);
1830 
1831   // Call a new method. All we do is (temporarily) trim the expression stack
1832   // push a return address to bring us back to here and leap to the new entry.
1833   // At this point we have a topmost frame that was allocated by the frame manager
1834   // which contains the current method interpreted state. We trim this frame
1835   // of excess java expression stack entries and then recurse.
1836 
1837   __ bind(call_method);
1838 
1839   // stack points to next free location and not top element on expression stack
1840   // method expects sp to be pointing to topmost element
1841 
1842   __ ld_ptr(STATE(_thread), G2_thread);
1843   __ ld_ptr(STATE(_result._to_call._callee), G5_method);
1844 
1845 
1846   // SP already takes in to account the 2 extra words we use for slop
1847   // when we call a "static long no_params()" method. So if
1848   // we trim back sp by the amount of unused java expression stack
1849   // there will be automagically the 2 extra words we need.
1850   // We also have to worry about keeping SP aligned.
1851 
1852   __ ld_ptr(STATE(_stack), Gargs);
1853   __ ld_ptr(STATE(_stack_limit), L1_scratch);
1854 
1855   // compute the unused java stack size
1856   __ sub(Gargs, L1_scratch, L2_scratch);                       // compute unused space
1857 
1858   // Round down the unused space to that stack is always 16-byte aligned
1859   // by making the unused space a multiple of the size of two longs.
1860 
1861   __ and3(L2_scratch, -2*BytesPerLong, L2_scratch);
1862 
1863   // Now trim the stack
1864   __ add(SP, L2_scratch, SP);
1865 
1866 
1867   // Now point to the final argument (account for prepush)
1868   __ add(Gargs, wordSize, Gargs);
1869 #ifdef ASSERT
1870   // Make sure we have space for the window
1871   __ sub(Gargs, SP, L1_scratch);
1872   __ cmp(L1_scratch, 16*wordSize);
1873   {
1874     Label skip;
1875     __ brx(Assembler::greaterEqual, false, Assembler::pt, skip);
1876     __ delayed()->nop();
1877     __ stop("killed stack");
1878     __ bind(skip);
1879   }
1880 #endif // ASSERT
1881 
1882   // Create a new frame where we can store values that make it look like the interpreter
1883   // really recursed.
1884 
1885   // prepare to recurse or call specialized entry
1886 
1887   // First link the registers we need
1888 
1889   // make the pc look good in debugger
1890   __ set(CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation), O7);
1891   // argument too
1892   __ mov(Lstate, I0);
1893 
1894   // Record our sending SP
1895   __ mov(SP, O5_savedSP);
1896 
1897   __ ld_ptr(STATE(_result._to_call._callee_entry_point), L2_scratch);
1898   __ set((intptr_t) entry_point, L1_scratch);
1899   __ cmp(L1_scratch, L2_scratch);
1900   __ brx(Assembler::equal, false, Assembler::pt, re_dispatch);
1901   __ delayed()->mov(Lstate, prevState);                                // link activations
1902 
1903   // method uses specialized entry, push a return so we look like call stub setup
1904   // this path will handle fact that result is returned in registers and not
1905   // on the java stack.
1906 
1907   __ set((intptr_t)return_from_native_method - 8, O7);
1908   __ jmpl(L2_scratch, G0, G0);                               // Do specialized entry
1909   __ delayed()->nop();
1910 
1911   //
1912   // Bad Message from interpreter
1913   //
1914   __ bind(bad_msg);
1915   __ stop("Bad message from interpreter");
1916 
1917   // Interpreted method "returned" with an exception pass it on...
1918   // Pass result, unwind activation and continue/return to interpreter/call_stub
1919   // We handle result (if any) differently based on return to interpreter or call_stub
1920 
1921   __ bind(throw_exception);
1922   __ ld_ptr(STATE(_prev_link), L1_scratch);
1923   __ tst(L1_scratch);
1924   __ brx(Assembler::zero, false, Assembler::pt, unwind_and_forward);
1925   __ delayed()->nop();
1926 
1927   __ ld_ptr(STATE(_locals), O1); // get result of popping callee's args
1928   __ ba(unwind_recursive_activation);
1929   __ delayed()->nop();
1930 
1931   interpreter_frame_manager = entry_point;
1932   return entry_point;
1933 }
1934 
1935 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
1936  : CppInterpreterGenerator(code) {
1937    generate_all(); // down here so it can be "virtual"
1938 }
1939 
1940 
1941 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
1942 
1943   // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
1944   // expression stack, the callee will have callee_extra_locals (so we can account for
1945   // frame extension) and monitor_size for monitors. Basically we need to calculate
1946   // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
1947   //
1948   //
1949   // The big complicating thing here is that we must ensure that the stack stays properly
1950   // aligned. This would be even uglier if monitor size wasn't modulo what the stack
1951   // needs to be aligned for). We are given that the sp (fp) is already aligned by
1952   // the caller so we must ensure that it is properly aligned for our callee.
1953   //
1954   // Ths c++ interpreter always makes sure that we have a enough extra space on the
1955   // stack at all times to deal with the "stack long no_params()" method issue. This
1956   // is "slop_factor" here.
1957   const int slop_factor = 2;
1958 
1959   const int fixed_size = sizeof(BytecodeInterpreter)/wordSize +           // interpreter state object
1960                          frame::memory_parameter_word_sp_offset;   // register save area + param window
1961   return (round_to(max_stack +
1962                    slop_factor +
1963                    fixed_size +
1964                    monitor_size +
1965                    (callee_extra_locals * Interpreter::stackElementWords), WordsPerLong));
1966 
1967 }
1968 
1969 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
1970 
1971   // See call_stub code
1972   int call_stub_size  = round_to(7 + frame::memory_parameter_word_sp_offset,
1973                                  WordsPerLong);    // 7 + register save area
1974 
1975   // Save space for one monitor to get into the interpreted method in case
1976   // the method is synchronized
1977   int monitor_size    = method->is_synchronized() ?
1978                                 1*frame::interpreter_frame_monitor_size() : 0;
1979   return size_activation_helper(method->max_locals(), method->max_stack(),
1980                                 monitor_size) + call_stub_size;
1981 }
1982 
1983 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
1984                                            frame* caller,
1985                                            frame* current,
1986                                            Method* method,
1987                                            intptr_t* locals,
1988                                            intptr_t* stack,
1989                                            intptr_t* stack_base,
1990                                            intptr_t* monitor_base,
1991                                            intptr_t* frame_bottom,
1992                                            bool is_top_frame
1993                                            )
1994 {
1995   // What about any vtable?
1996   //
1997   to_fill->_thread = JavaThread::current();
1998   // This gets filled in later but make it something recognizable for now
1999   to_fill->_bcp = method->code_base();
2000   to_fill->_locals = locals;
2001   to_fill->_constants = method->constants()->cache();
2002   to_fill->_method = method;
2003   to_fill->_mdx = NULL;
2004   to_fill->_stack = stack;
2005   if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) {
2006     to_fill->_msg = deopt_resume2;
2007   } else {
2008     to_fill->_msg = method_resume;
2009   }
2010   to_fill->_result._to_call._bcp_advance = 0;
2011   to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2012   to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2013   to_fill->_prev_link = NULL;
2014 
2015   // Fill in the registers for the frame
2016 
2017   // Need to install _sender_sp. Actually not too hard in C++!
2018   // When the skeletal frames are layed out we fill in a value
2019   // for _sender_sp. That value is only correct for the oldest
2020   // skeletal frame constructed (because there is only a single
2021   // entry for "caller_adjustment". While the skeletal frames
2022   // exist that is good enough. We correct that calculation
2023   // here and get all the frames correct.
2024 
2025   // to_fill->_sender_sp = locals - (method->size_of_parameters() - 1);
2026 
2027   *current->register_addr(Lstate) = (intptr_t) to_fill;
2028   // skeletal already places a useful value here and this doesn't account
2029   // for alignment so don't bother.
2030   // *current->register_addr(I5_savedSP) =     (intptr_t) locals - (method->size_of_parameters() - 1);
2031 
2032   if (caller->is_interpreted_frame()) {
2033     interpreterState prev  = caller->get_interpreterState();
2034     to_fill->_prev_link = prev;
2035     // Make the prev callee look proper
2036     prev->_result._to_call._callee = method;
2037     if (*prev->_bcp == Bytecodes::_invokeinterface) {
2038       prev->_result._to_call._bcp_advance = 5;
2039     } else {
2040       prev->_result._to_call._bcp_advance = 3;
2041     }
2042   }
2043   to_fill->_oop_temp = NULL;
2044   to_fill->_stack_base = stack_base;
2045   // Need +1 here because stack_base points to the word just above the first expr stack entry
2046   // and stack_limit is supposed to point to the word just below the last expr stack entry.
2047   // See generate_compute_interpreter_state.
2048   to_fill->_stack_limit = stack_base - (method->max_stack() + 1);
2049   to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2050 
2051   // sparc specific
2052   to_fill->_frame_bottom = frame_bottom;
2053   to_fill->_self_link = to_fill;
2054 #ifdef ASSERT
2055   to_fill->_native_fresult = 123456.789;
2056   to_fill->_native_lresult = CONST64(0xdeadcafedeafcafe);
2057 #endif
2058 }
2059 
2060 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, address last_Java_pc, intptr_t* last_Java_fp) {
2061   istate->_last_Java_pc = (intptr_t*) last_Java_pc;
2062 }
2063 
2064 static int frame_size_helper(int max_stack,
2065                              int moncount,
2066                              int callee_param_size,
2067                              int callee_locals_size,
2068                              bool is_top_frame,
2069                              int& monitor_size,
2070                              int& full_frame_words) {
2071   int extra_locals_size = callee_locals_size - callee_param_size;
2072   monitor_size = (sizeof(BasicObjectLock) * moncount) / wordSize;
2073   full_frame_words = size_activation_helper(extra_locals_size, max_stack, monitor_size);
2074   int short_frame_words = size_activation_helper(extra_locals_size, max_stack, monitor_size);
2075   int frame_words = is_top_frame ? full_frame_words : short_frame_words;
2076 
2077   return frame_words;
2078 }
2079 
2080 int AbstractInterpreter::size_activation(int max_stack,
2081                                          int tempcount,
2082                                          int extra_args,
2083                                          int moncount,
2084                                          int callee_param_size,
2085                                          int callee_locals_size,
2086                                          bool is_top_frame) {
2087   assert(extra_args == 0, "NEED TO FIX");
2088   // NOTE: return size is in words not bytes
2089   // Calculate the amount our frame will be adjust by the callee. For top frame
2090   // this is zero.
2091 
2092   // NOTE: ia64 seems to do this wrong (or at least backwards) in that it
2093   // calculates the extra locals based on itself. Not what the callee does
2094   // to it. So it ignores last_frame_adjust value. Seems suspicious as far
2095   // as getting sender_sp correct.
2096 
2097   int unused_monitor_size = 0;
2098   int unused_full_frame_words = 0;
2099   return frame_size_helper(max_stack, moncount, callee_param_size, callee_locals_size, is_top_frame,
2100                            unused_monitor_size, unused_full_frame_words);
2101 }
2102 void AbstractInterpreter::layout_activation(Method* method,
2103                                             int tempcount, // Number of slots on java expression stack in use
2104                                             int popframe_extra_args,
2105                                             int moncount,  // Number of active monitors
2106                                             int caller_actual_parameters,
2107                                             int callee_param_size,
2108                                             int callee_locals_size,
2109                                             frame* caller,
2110                                             frame* interpreter_frame,
2111                                             bool is_top_frame,
2112                                             bool is_bottom_frame) {
2113   assert(popframe_extra_args == 0, "NEED TO FIX");
2114   // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state()
2115   // does as far as allocating an interpreter frame.
2116   // Set up the method, locals, and monitors.
2117   // The frame interpreter_frame is guaranteed to be the right size,
2118   // as determined by a previous call to the size_activation() method.
2119   // It is also guaranteed to be walkable even though it is in a skeletal state
2120   // NOTE: tempcount is the current size of the java expression stack. For top most
2121   //       frames we will allocate a full sized expression stack and not the curback
2122   //       version that non-top frames have.
2123 
2124   int monitor_size = 0;
2125   int full_frame_words = 0;
2126   int frame_words = frame_size_helper(method->max_stack(), moncount, callee_param_size, callee_locals_size,
2127                                       is_top_frame, monitor_size, full_frame_words);
2128 
2129   /*
2130     We must now fill in all the pieces of the frame. This means both
2131     the interpreterState and the registers.
2132   */
2133 
2134   // MUCHO HACK
2135 
2136   intptr_t* frame_bottom = interpreter_frame->sp() - (full_frame_words - frame_words);
2137   // 'interpreter_frame->sp()' is unbiased while 'frame_bottom' must be a biased value in 64bit mode.
2138   assert(((intptr_t)frame_bottom & 0xf) == 0, "SP biased in layout_activation");
2139   frame_bottom = (intptr_t*)((intptr_t)frame_bottom - STACK_BIAS);
2140 
2141   /* Now fillin the interpreterState object */
2142 
2143   interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() -  sizeof(BytecodeInterpreter));
2144 
2145 
2146   intptr_t* locals;
2147 
2148   // Calculate the postion of locals[0]. This is painful because of
2149   // stack alignment (same as ia64). The problem is that we can
2150   // not compute the location of locals from fp(). fp() will account
2151   // for the extra locals but it also accounts for aligning the stack
2152   // and we can't determine if the locals[0] was misaligned but max_locals
2153   // was enough to have the
2154   // calculate postion of locals. fp already accounts for extra locals.
2155   // +2 for the static long no_params() issue.
2156 
2157   if (caller->is_interpreted_frame()) {
2158     // locals must agree with the caller because it will be used to set the
2159     // caller's tos when we return.
2160     interpreterState prev  = caller->get_interpreterState();
2161     // stack() is prepushed.
2162     locals = prev->stack() + method->size_of_parameters();
2163   } else {
2164     // Lay out locals block in the caller adjacent to the register window save area.
2165     //
2166     // Compiled frames do not allocate a varargs area which is why this if
2167     // statement is needed.
2168     //
2169     intptr_t* fp = interpreter_frame->fp();
2170     int local_words = method->max_locals() * Interpreter::stackElementWords;
2171 
2172     if (caller->is_compiled_frame()) {
2173       locals = fp + frame::register_save_words + local_words - 1;
2174     } else {
2175       locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
2176     }
2177 
2178   }
2179   // END MUCHO HACK
2180 
2181   intptr_t* monitor_base = (intptr_t*) cur_state;
2182   intptr_t* stack_base =  monitor_base - monitor_size;
2183   /* +1 because stack is always prepushed */
2184   intptr_t* stack = stack_base - (tempcount + 1);
2185 
2186 
2187   BytecodeInterpreter::layout_interpreterState(cur_state,
2188                                                caller,
2189                                                interpreter_frame,
2190                                                method,
2191                                                locals,
2192                                                stack,
2193                                                stack_base,
2194                                                monitor_base,
2195                                                frame_bottom,
2196                                                is_top_frame);
2197 
2198   BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp());
2199 }
2200 
2201 #endif // CC_INTERP