1 /*
   2  * Copyright (c) 2007, 2013, 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 "oops/arrayOop.hpp"
  33 #include "oops/methodData.hpp"
  34 #include "oops/method.hpp"
  35 #include "oops/oop.inline.hpp"
  36 #include "prims/jvmtiExport.hpp"
  37 #include "prims/jvmtiThreadState.hpp"
  38 #include "runtime/arguments.hpp"
  39 #include "runtime/deoptimization.hpp"
  40 #include "runtime/frame.inline.hpp"
  41 #include "runtime/interfaceSupport.hpp"
  42 #include "runtime/sharedRuntime.hpp"
  43 #include "runtime/stubRoutines.hpp"
  44 #include "runtime/synchronizer.hpp"
  45 #include "runtime/timer.hpp"
  46 #include "runtime/vframeArray.hpp"
  47 #include "utilities/debug.hpp"
  48 #include "utilities/macros.hpp"
  49 #ifdef SHARK
  50 #include "shark/shark_globals.hpp"
  51 #endif
  52 
  53 #ifdef CC_INTERP
  54 
  55 // Routine exists to make tracebacks look decent in debugger
  56 // while "shadow" interpreter frames are on stack. It is also
  57 // used to distinguish interpreter frames.
  58 
  59 extern "C" void RecursiveInterpreterActivation(interpreterState istate) {
  60   ShouldNotReachHere();
  61 }
  62 
  63 bool CppInterpreter::contains(address pc) {
  64   return ( _code->contains(pc) ||
  65          ( pc == (CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset)));
  66 }
  67 
  68 #define STATE(field_name) Lstate, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
  69 #define __ _masm->
  70 
  71 Label frame_manager_entry;
  72 Label fast_accessor_slow_entry_path;  // fast accessor methods need to be able to jmp to unsynchronized
  73                                       // c++ interpreter entry point this holds that entry point label.
  74 
  75 static address unctrap_frame_manager_entry  = NULL;
  76 
  77 static address interpreter_return_address  = NULL;
  78 static address deopt_frame_manager_return_atos  = NULL;
  79 static address deopt_frame_manager_return_btos  = NULL;
  80 static address deopt_frame_manager_return_itos  = NULL;
  81 static address deopt_frame_manager_return_ltos  = NULL;
  82 static address deopt_frame_manager_return_ftos  = NULL;
  83 static address deopt_frame_manager_return_dtos  = NULL;
  84 static address deopt_frame_manager_return_vtos  = NULL;
  85 
  86 const Register prevState = G1_scratch;
  87 
  88 void InterpreterGenerator::save_native_result(void) {
  89   // result potentially in O0/O1: save it across calls
  90   __ stf(FloatRegisterImpl::D, F0, STATE(_native_fresult));
  91 #ifdef _LP64
  92   __ stx(O0, STATE(_native_lresult));
  93 #else
  94   __ std(O0, STATE(_native_lresult));
  95 #endif
  96 }
  97 
  98 void InterpreterGenerator::restore_native_result(void) {
  99 
 100   // Restore any method result value
 101   __ ldf(FloatRegisterImpl::D, STATE(_native_fresult), F0);
 102 #ifdef _LP64
 103   __ ldx(STATE(_native_lresult), O0);
 104 #else
 105   __ ldd(STATE(_native_lresult), O0);
 106 #endif
 107 }
 108 
 109 // A result handler converts/unboxes a native call result into
 110 // a java interpreter/compiler result. The current frame is an
 111 // interpreter frame. The activation frame unwind code must be
 112 // consistent with that of TemplateTable::_return(...). In the
 113 // case of native methods, the caller's SP was not modified.
 114 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) {
 115   address entry = __ pc();
 116   Register Itos_i  = Otos_i ->after_save();
 117   Register Itos_l  = Otos_l ->after_save();
 118   Register Itos_l1 = Otos_l1->after_save();
 119   Register Itos_l2 = Otos_l2->after_save();
 120   switch (type) {
 121     case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
 122     case T_CHAR   : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i);   break; // cannot use and3, 0xFFFF too big as immediate value!
 123     case T_BYTE   : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i);   break;
 124     case T_SHORT  : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i);   break;
 125     case T_LONG   :
 126 #ifndef _LP64
 127                     __ mov(O1, Itos_l2);  // move other half of long
 128 #endif              // ifdef or no ifdef, fall through to the T_INT case
 129     case T_INT    : __ mov(O0, Itos_i);                         break;
 130     case T_VOID   : /* nothing to do */                         break;
 131     case T_FLOAT  : assert(F0 == Ftos_f, "fix this code" );     break;
 132     case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" );     break;
 133     case T_OBJECT :
 134       __ ld_ptr(STATE(_oop_temp), Itos_i);
 135       __ verify_oop(Itos_i);
 136       break;
 137     default       : ShouldNotReachHere();
 138   }
 139   __ ret();                           // return from interpreter activation
 140   __ delayed()->restore(I5_savedSP, G0, SP);  // remove interpreter frame
 141   NOT_PRODUCT(__ emit_int32(0);)       // marker for disassembly
 142   return entry;
 143 }
 144 
 145 // tosca based result to c++ interpreter stack based result.
 146 // Result goes to address in L1_scratch
 147 
 148 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) {
 149   // A result is in the native abi result register from a native method call.
 150   // We need to return this result to the interpreter by pushing the result on the interpreter's
 151   // stack. This is relatively simple the destination is in L1_scratch
 152   // i.e. L1_scratch is the first free element on the stack. If we "push" a return value we must
 153   // adjust L1_scratch
 154   address entry = __ pc();
 155   switch (type) {
 156     case T_BOOLEAN:
 157       // !0 => true; 0 => false
 158       __ subcc(G0, O0, G0);
 159       __ addc(G0, 0, O0);
 160       __ st(O0, L1_scratch, 0);
 161       __ sub(L1_scratch, wordSize, L1_scratch);
 162       break;
 163 
 164     // cannot use and3, 0xFFFF too big as immediate value!
 165     case T_CHAR   :
 166       __ sll(O0, 16, O0);
 167       __ srl(O0, 16, O0);
 168       __ st(O0, L1_scratch, 0);
 169       __ sub(L1_scratch, wordSize, L1_scratch);
 170       break;
 171 
 172     case T_BYTE   :
 173       __ sll(O0, 24, O0);
 174       __ sra(O0, 24, O0);
 175       __ st(O0, L1_scratch, 0);
 176       __ sub(L1_scratch, wordSize, L1_scratch);
 177       break;
 178 
 179     case T_SHORT  :
 180       __ sll(O0, 16, O0);
 181       __ sra(O0, 16, O0);
 182       __ st(O0, L1_scratch, 0);
 183       __ sub(L1_scratch, wordSize, L1_scratch);
 184       break;
 185     case T_LONG   :
 186 #ifndef _LP64
 187 #if defined(COMPILER2)
 188   // All return values are where we want them, except for Longs.  C2 returns
 189   // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
 190   // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
 191   // build even if we are returning from interpreted we just do a little
 192   // stupid shuffing.
 193   // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
 194   // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
 195   // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
 196       __ stx(G1, L1_scratch, -wordSize);
 197 #else
 198       // native result is in O0, O1
 199       __ st(O1, L1_scratch, 0);                      // Low order
 200       __ st(O0, L1_scratch, -wordSize);              // High order
 201 #endif /* COMPILER2 */
 202 #else
 203       __ stx(O0, L1_scratch, -wordSize);
 204 #endif
 205       __ sub(L1_scratch, 2*wordSize, L1_scratch);
 206       break;
 207 
 208     case T_INT    :
 209       __ st(O0, L1_scratch, 0);
 210       __ sub(L1_scratch, wordSize, L1_scratch);
 211       break;
 212 
 213     case T_VOID   : /* nothing to do */
 214       break;
 215 
 216     case T_FLOAT  :
 217       __ stf(FloatRegisterImpl::S, F0, L1_scratch, 0);
 218       __ sub(L1_scratch, wordSize, L1_scratch);
 219       break;
 220 
 221     case T_DOUBLE :
 222       // Every stack slot is aligned on 64 bit, However is this
 223       // the correct stack slot on 64bit?? QQQ
 224       __ stf(FloatRegisterImpl::D, F0, L1_scratch, -wordSize);
 225       __ sub(L1_scratch, 2*wordSize, L1_scratch);
 226       break;
 227     case T_OBJECT :
 228       __ verify_oop(O0);
 229       __ st_ptr(O0, L1_scratch, 0);
 230       __ sub(L1_scratch, wordSize, L1_scratch);
 231       break;
 232     default       : ShouldNotReachHere();
 233   }
 234   __ retl();                          // return from interpreter activation
 235   __ delayed()->nop();                // schedule this better
 236   NOT_PRODUCT(__ emit_int32(0);)       // marker for disassembly
 237   return entry;
 238 }
 239 
 240 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) {
 241   // A result is in the java expression stack of the interpreted method that has just
 242   // returned. Place this result on the java expression stack of the caller.
 243   //
 244   // The current interpreter activation in Lstate is for the method just returning its
 245   // result. So we know that the result of this method is on the top of the current
 246   // execution stack (which is pre-pushed) and will be return to the top of the caller
 247   // stack. The top of the callers stack is the bottom of the locals of the current
 248   // activation.
 249   // Because of the way activation are managed by the frame manager the value of esp is
 250   // below both the stack top of the current activation and naturally the stack top
 251   // of the calling activation. This enable this routine to leave the return address
 252   // to the frame manager on the stack and do a vanilla return.
 253   //
 254   // On entry: O0 - points to source (callee stack top)
 255   //           O1 - points to destination (caller stack top [i.e. free location])
 256   // destroys O2, O3
 257   //
 258 
 259   address entry = __ pc();
 260   switch (type) {
 261     case T_VOID:  break;
 262       break;
 263     case T_FLOAT  :
 264     case T_BOOLEAN:
 265     case T_CHAR   :
 266     case T_BYTE   :
 267     case T_SHORT  :
 268     case T_INT    :
 269       // 1 word result
 270       __ ld(O0, 0, O2);
 271       __ st(O2, O1, 0);
 272       __ sub(O1, wordSize, O1);
 273       break;
 274     case T_DOUBLE  :
 275     case T_LONG    :
 276       // return top two words on current expression stack to caller's expression stack
 277       // The caller's expression stack is adjacent to the current frame manager's intepretState
 278       // except we allocated one extra word for this intepretState so we won't overwrite it
 279       // when we return a two word result.
 280 #ifdef _LP64
 281       __ ld_ptr(O0, 0, O2);
 282       __ st_ptr(O2, O1, -wordSize);
 283 #else
 284       __ ld(O0, 0, O2);
 285       __ ld(O0, wordSize, O3);
 286       __ st(O3, O1, 0);
 287       __ st(O2, O1, -wordSize);
 288 #endif
 289       __ sub(O1, 2*wordSize, O1);
 290       break;
 291     case T_OBJECT :
 292       __ ld_ptr(O0, 0, O2);
 293       __ verify_oop(O2);                                               // verify it
 294       __ st_ptr(O2, O1, 0);
 295       __ sub(O1, wordSize, O1);
 296       break;
 297     default       : ShouldNotReachHere();
 298   }
 299   __ retl();
 300   __ delayed()->nop(); // QQ schedule this better
 301   return entry;
 302 }
 303 
 304 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) {
 305   // A result is in the java expression stack of the interpreted method that has just
 306   // returned. Place this result in the native abi that the caller expects.
 307   // We are in a new frame registers we set must be in caller (i.e. callstub) frame.
 308   //
 309   // Similar to generate_stack_to_stack_converter above. Called at a similar time from the
 310   // frame manager execept in this situation the caller is native code (c1/c2/call_stub)
 311   // and so rather than return result onto caller's java expression stack we return the
 312   // result in the expected location based on the native abi.
 313   // On entry: O0 - source (stack top)
 314   // On exit result in expected output register
 315   // QQQ schedule this better
 316 
 317   address entry = __ pc();
 318   switch (type) {
 319     case T_VOID:  break;
 320       break;
 321     case T_FLOAT  :
 322       __ ldf(FloatRegisterImpl::S, O0, 0, F0);
 323       break;
 324     case T_BOOLEAN:
 325     case T_CHAR   :
 326     case T_BYTE   :
 327     case T_SHORT  :
 328     case T_INT    :
 329       // 1 word result
 330       __ ld(O0, 0, O0->after_save());
 331       break;
 332     case T_DOUBLE  :
 333       __ ldf(FloatRegisterImpl::D, O0, 0, F0);
 334       break;
 335     case T_LONG    :
 336       // return top two words on current expression stack to caller's expression stack
 337       // The caller's expression stack is adjacent to the current frame manager's interpretState
 338       // except we allocated one extra word for this intepretState so we won't overwrite it
 339       // when we return a two word result.
 340 #ifdef _LP64
 341       __ ld_ptr(O0, 0, O0->after_save());
 342 #else
 343       __ ld(O0, wordSize, O1->after_save());
 344       __ ld(O0, 0, O0->after_save());
 345 #endif
 346 #if defined(COMPILER2) && !defined(_LP64)
 347       // C2 expects long results in G1 we can't tell if we're returning to interpreted
 348       // or compiled so just be safe use G1 and O0/O1
 349 
 350       // Shift bits into high (msb) of G1
 351       __ sllx(Otos_l1->after_save(), 32, G1);
 352       // Zero extend low bits
 353       __ srl (Otos_l2->after_save(), 0, Otos_l2->after_save());
 354       __ or3 (Otos_l2->after_save(), G1, G1);
 355 #endif /* COMPILER2 */
 356       break;
 357     case T_OBJECT :
 358       __ ld_ptr(O0, 0, O0->after_save());
 359       __ verify_oop(O0->after_save());                                               // verify it
 360       break;
 361     default       : ShouldNotReachHere();
 362   }
 363   __ retl();
 364   __ delayed()->nop();
 365   return entry;
 366 }
 367 
 368 address CppInterpreter::return_entry(TosState state, int length, Bytecodes::Code code) {
 369   // make it look good in the debugger
 370   return CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset;
 371 }
 372 
 373 address CppInterpreter::deopt_entry(TosState state, int length) {
 374   address ret = NULL;
 375   if (length != 0) {
 376     switch (state) {
 377       case atos: ret = deopt_frame_manager_return_atos; break;
 378       case btos: ret = deopt_frame_manager_return_btos; break;
 379       case ctos:
 380       case stos:
 381       case itos: ret = deopt_frame_manager_return_itos; break;
 382       case ltos: ret = deopt_frame_manager_return_ltos; break;
 383       case ftos: ret = deopt_frame_manager_return_ftos; break;
 384       case dtos: ret = deopt_frame_manager_return_dtos; break;
 385       case vtos: ret = deopt_frame_manager_return_vtos; break;
 386     }
 387   } else {
 388     ret = unctrap_frame_manager_entry;  // re-execute the bytecode ( e.g. uncommon trap)
 389   }
 390   assert(ret != NULL, "Not initialized");
 391   return ret;
 392 }
 393 
 394 //
 395 // Helpers for commoning out cases in the various type of method entries.
 396 //
 397 
 398 // increment invocation count & check for overflow
 399 //
 400 // Note: checking for negative value instead of overflow
 401 //       so we have a 'sticky' overflow test
 402 //
 403 // Lmethod: method
 404 // ??: invocation counter
 405 //
 406 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
 407   Label done;
 408   const Register Rcounters = G3_scratch;
 409 
 410   __ ld_ptr(STATE(_method), G5_method);
 411   __ get_method_counters(G5_method, Rcounters, done);
 412 
 413   // Update standard invocation counters
 414   __ increment_invocation_counter(Rcounters, O0, G4_scratch);
 415   if (ProfileInterpreter) {
 416     Address interpreter_invocation_counter(Rcounters,
 417             in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
 418     __ ld(interpreter_invocation_counter, G4_scratch);
 419     __ inc(G4_scratch);
 420     __ st(G4_scratch, interpreter_invocation_counter);
 421   }
 422 
 423   AddressLiteral invocation_limit((address)&InvocationCounter::InterpreterInvocationLimit);
 424   __ load_contents(invocation_limit, G3_scratch);
 425   __ cmp(O0, G3_scratch);
 426   __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
 427   __ delayed()->nop();
 428   __ bind(done);
 429 }
 430 
 431 address InterpreterGenerator::generate_empty_entry(void) {
 432 
 433   // A method that does nothing but return...
 434 
 435   address entry = __ pc();
 436   Label slow_path;
 437 
 438   // do nothing for empty methods (do not even increment invocation counter)
 439   if ( UseFastEmptyMethods) {
 440     // If we need a safepoint check, generate full interpreter entry.
 441     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
 442     __ load_contents(sync_state, G3_scratch);
 443     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 444     __ br(Assembler::notEqual, false, Assembler::pn, frame_manager_entry);
 445     __ delayed()->nop();
 446 
 447     // Code: _return
 448     __ retl();
 449     __ delayed()->mov(O5_savedSP, SP);
 450     return entry;
 451   }
 452   return NULL;
 453 }
 454 
 455 // Call an accessor method (assuming it is resolved, otherwise drop into
 456 // vanilla (slow path) entry
 457 
 458 // Generates code to elide accessor methods
 459 // Uses G3_scratch and G1_scratch as scratch
 460 address InterpreterGenerator::generate_accessor_entry(void) {
 461 
 462   // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
 463   // parameter size = 1
 464   // Note: We can only use this code if the getfield has been resolved
 465   //       and if we don't have a null-pointer exception => check for
 466   //       these conditions first and use slow path if necessary.
 467   address entry = __ pc();
 468   Label slow_path;
 469 
 470   if ( UseFastAccessorMethods) {
 471     // Check if we need to reach a safepoint and generate full interpreter
 472     // frame if so.
 473     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
 474     __ load_contents(sync_state, G3_scratch);
 475     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 476     __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
 477     __ delayed()->nop();
 478 
 479     // Check if local 0 != NULL
 480     __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
 481     __ tst(Otos_i);  // check if local 0 == NULL and go the slow path
 482     __ brx(Assembler::zero, false, Assembler::pn, slow_path);
 483     __ delayed()->nop();
 484 
 485 
 486     // read first instruction word and extract bytecode @ 1 and index @ 2
 487     // get first 4 bytes of the bytecodes (big endian!)
 488     __ ld_ptr(Address(G5_method, in_bytes(Method::const_offset())), G1_scratch);
 489     __ ld(Address(G1_scratch, in_bytes(ConstMethod::codes_offset())), G1_scratch);
 490 
 491     // move index @ 2 far left then to the right most two bytes.
 492     __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
 493     __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
 494                       ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
 495 
 496     // get constant pool cache
 497     __ ld_ptr(G5_method, in_bytes(Method::const_offset()), G3_scratch);
 498     __ ld_ptr(G3_scratch, in_bytes(ConstMethod::constants_offset()), G3_scratch);
 499     __ ld_ptr(G3_scratch, ConstantPool::cache_offset_in_bytes(), G3_scratch);
 500 
 501     // get specific constant pool cache entry
 502     __ add(G3_scratch, G1_scratch, G3_scratch);
 503 
 504     // Check the constant Pool cache entry to see if it has been resolved.
 505     // If not, need the slow path.
 506     ByteSize cp_base_offset = ConstantPoolCache::base_offset();
 507     __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::indices_offset()), G1_scratch);
 508     __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
 509     __ and3(G1_scratch, 0xFF, G1_scratch);
 510     __ cmp(G1_scratch, Bytecodes::_getfield);
 511     __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
 512     __ delayed()->nop();
 513 
 514     // Get the type and return field offset from the constant pool cache
 515     __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset()), G1_scratch);
 516     __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset()), G3_scratch);
 517 
 518     Label xreturn_path;
 519     // Need to differentiate between igetfield, agetfield, bgetfield etc.
 520     // because they are different sizes.
 521     // Get the type from the constant pool cache
 522     __ srl(G1_scratch, ConstantPoolCacheEntry::tos_state_shift, G1_scratch);
 523     // Make sure we don't need to mask G1_scratch after the above shift
 524     ConstantPoolCacheEntry::verify_tos_state_shift();
 525     __ cmp(G1_scratch, atos );
 526     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 527     __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
 528     __ cmp(G1_scratch, itos);
 529     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 530     __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
 531     __ cmp(G1_scratch, stos);
 532     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 533     __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
 534     __ cmp(G1_scratch, ctos);
 535     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 536     __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
 537 #ifdef ASSERT
 538     __ cmp(G1_scratch, btos);
 539     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 540     __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
 541     __ should_not_reach_here();
 542 #endif
 543     __ ldsb(Otos_i, G3_scratch, Otos_i);
 544     __ bind(xreturn_path);
 545 
 546     // _ireturn/_areturn
 547     __ retl();                      // return from leaf routine
 548     __ delayed()->mov(O5_savedSP, SP);
 549 
 550     // Generate regular method entry
 551     __ bind(slow_path);
 552     __ ba(fast_accessor_slow_entry_path);
 553     __ delayed()->nop();
 554     return entry;
 555   }
 556   return NULL;
 557 }
 558 
 559 address InterpreterGenerator::generate_Reference_get_entry(void) {
 560 #if INCLUDE_ALL_GCS
 561   if (UseG1GC) {
 562     // We need to generate have a routine that generates code to:
 563     //   * load the value in the referent field
 564     //   * passes that value to the pre-barrier.
 565     //
 566     // In the case of G1 this will record the value of the
 567     // referent in an SATB buffer if marking is active.
 568     // This will cause concurrent marking to mark the referent
 569     // field as live.
 570     Unimplemented();
 571   }
 572 #endif // INCLUDE_ALL_GCS
 573 
 574   // If G1 is not enabled then attempt to go through the accessor entry point
 575   // Reference.get is an accessor
 576   return generate_accessor_entry();
 577 }
 578 
 579 //
 580 // Interpreter stub for calling a native method. (C++ interpreter)
 581 // This sets up a somewhat different looking stack for calling the native method
 582 // than the typical interpreter frame setup.
 583 //
 584 
 585 address InterpreterGenerator::generate_native_entry(bool synchronized) {
 586   address entry = __ pc();
 587 
 588   // the following temporary registers are used during frame creation
 589   const Register Gtmp1 = G3_scratch ;
 590   const Register Gtmp2 = G1_scratch;
 591   const Register RconstMethod = Gtmp1;
 592   const Address constMethod(G5_method, in_bytes(Method::const_offset()));
 593   const Address size_of_parameters(RconstMethod, in_bytes(ConstMethod::size_of_parameters_offset()));
 594 
 595   bool inc_counter  = UseCompiler || CountCompiledCalls;
 596 
 597   // make sure registers are different!
 598   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
 599 
 600   const Address access_flags      (G5_method, in_bytes(Method::access_flags_offset()));
 601 
 602   Label Lentry;
 603   __ bind(Lentry);
 604 
 605   const Register Glocals_size = G3;
 606   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
 607 
 608   // make sure method is native & not abstract
 609   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
 610 #ifdef ASSERT
 611   __ ld(access_flags, Gtmp1);
 612   {
 613     Label L;
 614     __ btst(JVM_ACC_NATIVE, Gtmp1);
 615     __ br(Assembler::notZero, false, Assembler::pt, L);
 616     __ delayed()->nop();
 617     __ stop("tried to execute non-native method as native");
 618     __ bind(L);
 619   }
 620   { Label L;
 621     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
 622     __ br(Assembler::zero, false, Assembler::pt, L);
 623     __ delayed()->nop();
 624     __ stop("tried to execute abstract method as non-abstract");
 625     __ bind(L);
 626   }
 627 #endif // ASSERT
 628 
 629   __ ld_ptr(constMethod, RconstMethod);
 630   __ lduh(size_of_parameters, Gtmp1);
 631   __ sll(Gtmp1, LogBytesPerWord, Gtmp2);       // parameter size in bytes
 632   __ add(Gargs, Gtmp2, Gargs);                 // points to first local + BytesPerWord
 633   // NEW
 634   __ add(Gargs, -wordSize, Gargs);             // points to first local[0]
 635   // generate the code to allocate the interpreter stack frame
 636   // NEW FRAME ALLOCATED HERE
 637   // save callers original sp
 638   // __ mov(SP, I5_savedSP->after_restore());
 639 
 640   generate_compute_interpreter_state(Lstate, G0, true);
 641 
 642   // At this point Lstate points to new interpreter state
 643   //
 644 
 645   const Address do_not_unlock_if_synchronized(G2_thread,
 646       in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
 647   // Since at this point in the method invocation the exception handler
 648   // would try to exit the monitor of synchronized methods which hasn't
 649   // been entered yet, we set the thread local variable
 650   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
 651   // runtime, exception handling i.e. unlock_if_synchronized_method will
 652   // check this thread local flag.
 653   // This flag has two effects, one is to force an unwind in the topmost
 654   // interpreter frame and not perform an unlock while doing so.
 655 
 656   __ movbool(true, G3_scratch);
 657   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
 658 
 659 
 660   // increment invocation counter and check for overflow
 661   //
 662   // Note: checking for negative value instead of overflow
 663   //       so we have a 'sticky' overflow test (may be of
 664   //       importance as soon as we have true MT/MP)
 665   Label invocation_counter_overflow;
 666   if (inc_counter) {
 667     generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
 668   }
 669   Label Lcontinue;
 670   __ bind(Lcontinue);
 671 
 672   bang_stack_shadow_pages(true);
 673   // reset the _do_not_unlock_if_synchronized flag
 674   __ stbool(G0, do_not_unlock_if_synchronized);
 675 
 676   // check for synchronized methods
 677   // Must happen AFTER invocation_counter check, so method is not locked
 678   // if counter overflows.
 679 
 680   if (synchronized) {
 681     lock_method();
 682     // Don't see how G2_thread is preserved here...
 683     // __ verify_thread(); QQQ destroys L0,L1 can't use
 684   } else {
 685 #ifdef ASSERT
 686     { Label ok;
 687       __ ld_ptr(STATE(_method), G5_method);
 688       __ ld(access_flags, O0);
 689       __ btst(JVM_ACC_SYNCHRONIZED, O0);
 690       __ br( Assembler::zero, false, Assembler::pt, ok);
 691       __ delayed()->nop();
 692       __ stop("method needs synchronization");
 693       __ bind(ok);
 694     }
 695 #endif // ASSERT
 696   }
 697 
 698   // start execution
 699 
 700 //   __ verify_thread(); kills L1,L2 can't  use at the moment
 701 
 702   // jvmti/jvmpi support
 703   __ notify_method_entry();
 704 
 705   // native call
 706 
 707   // (note that O0 is never an oop--at most it is a handle)
 708   // It is important not to smash any handles created by this call,
 709   // until any oop handle in O0 is dereferenced.
 710 
 711   // (note that the space for outgoing params is preallocated)
 712 
 713   // get signature handler
 714 
 715   Label pending_exception_present;
 716 
 717   { Label L;
 718     __ ld_ptr(STATE(_method), G5_method);
 719     __ ld_ptr(Address(G5_method, in_bytes(Method::signature_handler_offset())), G3_scratch);
 720     __ tst(G3_scratch);
 721     __ brx(Assembler::notZero, false, Assembler::pt, L);
 722     __ delayed()->nop();
 723     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), G5_method, false);
 724     __ ld_ptr(STATE(_method), G5_method);
 725 
 726     Address exception_addr(G2_thread, in_bytes(Thread::pending_exception_offset()));
 727     __ ld_ptr(exception_addr, G3_scratch);
 728     __ br_notnull_short(G3_scratch, Assembler::pn, pending_exception_present);
 729     __ ld_ptr(Address(G5_method, in_bytes(Method::signature_handler_offset())), G3_scratch);
 730     __ bind(L);
 731   }
 732 
 733   // Push a new frame so that the args will really be stored in
 734   // Copy a few locals across so the new frame has the variables
 735   // we need but these values will be dead at the jni call and
 736   // therefore not gc volatile like the values in the current
 737   // frame (Lstate in particular)
 738 
 739   // Flush the state pointer to the register save area
 740   // Which is the only register we need for a stack walk.
 741   __ st_ptr(Lstate, SP, (Lstate->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
 742 
 743   __ mov(Lstate, O1);         // Need to pass the state pointer across the frame
 744 
 745   // Calculate current frame size
 746   __ sub(SP, FP, O3);         // Calculate negative of current frame size
 747   __ save(SP, O3, SP);        // Allocate an identical sized frame
 748 
 749   __ mov(I1, Lstate);          // In the "natural" register.
 750 
 751   // Note I7 has leftover trash. Slow signature handler will fill it in
 752   // should we get there. Normal jni call will set reasonable last_Java_pc
 753   // below (and fix I7 so the stack trace doesn't have a meaningless frame
 754   // in it).
 755 
 756 
 757   // call signature handler
 758   __ ld_ptr(STATE(_method), Lmethod);
 759   __ ld_ptr(STATE(_locals), Llocals);
 760 
 761   __ callr(G3_scratch, 0);
 762   __ delayed()->nop();
 763   __ ld_ptr(STATE(_thread), G2_thread);        // restore thread (shouldn't be needed)
 764 
 765   { Label not_static;
 766 
 767     __ ld_ptr(STATE(_method), G5_method);
 768     __ ld(access_flags, O0);
 769     __ btst(JVM_ACC_STATIC, O0);
 770     __ br( Assembler::zero, false, Assembler::pt, not_static);
 771     __ delayed()->
 772       // get native function entry point(O0 is a good temp until the very end)
 773        ld_ptr(Address(G5_method, in_bytes(Method::native_function_offset())), O0);
 774     // for static methods insert the mirror argument
 775     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
 776 
 777     __ ld_ptr(Address(G5_method, in_bytes(Method:: const_offset())), O1);
 778     __ ld_ptr(Address(O1, in_bytes(ConstMethod::constants_offset())), O1);
 779     __ ld_ptr(Address(O1, ConstantPool::pool_holder_offset_in_bytes()), O1);
 780     __ ld_ptr(O1, mirror_offset, O1);
 781     // where the mirror handle body is allocated:
 782 #ifdef ASSERT
 783     if (!PrintSignatureHandlers)  // do not dirty the output with this
 784     { Label L;
 785       __ tst(O1);
 786       __ brx(Assembler::notZero, false, Assembler::pt, L);
 787       __ delayed()->nop();
 788       __ stop("mirror is missing");
 789       __ bind(L);
 790     }
 791 #endif // ASSERT
 792     __ st_ptr(O1, STATE(_oop_temp));
 793     __ add(STATE(_oop_temp), O1);            // this is really an LEA not an add
 794     __ bind(not_static);
 795   }
 796 
 797   // At this point, arguments have been copied off of stack into
 798   // their JNI positions, which are O1..O5 and SP[68..].
 799   // Oops are boxed in-place on the stack, with handles copied to arguments.
 800   // The result handler is in Lscratch.  O0 will shortly hold the JNIEnv*.
 801 
 802 #ifdef ASSERT
 803   { Label L;
 804     __ tst(O0);
 805     __ brx(Assembler::notZero, false, Assembler::pt, L);
 806     __ delayed()->nop();
 807     __ stop("native entry point is missing");
 808     __ bind(L);
 809   }
 810 #endif // ASSERT
 811 
 812   //
 813   // setup the java frame anchor
 814   //
 815   // The scavenge function only needs to know that the PC of this frame is
 816   // in the interpreter method entry code, it doesn't need to know the exact
 817   // PC and hence we can use O7 which points to the return address from the
 818   // previous call in the code stream (signature handler function)
 819   //
 820   // The other trick is we set last_Java_sp to FP instead of the usual SP because
 821   // we have pushed the extra frame in order to protect the volatile register(s)
 822   // in that frame when we return from the jni call
 823   //
 824 
 825 
 826   __ set_last_Java_frame(FP, O7);
 827   __ mov(O7, I7);  // make dummy interpreter frame look like one above,
 828                    // not meaningless information that'll confuse me.
 829 
 830   // flush the windows now. We don't care about the current (protection) frame
 831   // only the outer frames
 832 
 833   __ flushw();
 834 
 835   // mark windows as flushed
 836   Address flags(G2_thread,
 837                 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
 838   __ set(JavaFrameAnchor::flushed, G3_scratch);
 839   __ st(G3_scratch, flags);
 840 
 841   // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
 842 
 843   Address thread_state(G2_thread, in_bytes(JavaThread::thread_state_offset()));
 844 #ifdef ASSERT
 845   { Label L;
 846     __ ld(thread_state, G3_scratch);
 847     __ cmp(G3_scratch, _thread_in_Java);
 848     __ br(Assembler::equal, false, Assembler::pt, L);
 849     __ delayed()->nop();
 850     __ stop("Wrong thread state in native stub");
 851     __ bind(L);
 852   }
 853 #endif // ASSERT
 854   __ set(_thread_in_native, G3_scratch);
 855   __ st(G3_scratch, thread_state);
 856 
 857   // Call the jni method, using the delay slot to set the JNIEnv* argument.
 858   __ callr(O0, 0);
 859   __ delayed()->
 860      add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
 861   __ ld_ptr(STATE(_thread), G2_thread);  // restore thread
 862 
 863   // must we block?
 864 
 865   // Block, if necessary, before resuming in _thread_in_Java state.
 866   // In order for GC to work, don't clear the last_Java_sp until after blocking.
 867   { Label no_block;
 868     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
 869 
 870     // Switch thread to "native transition" state before reading the synchronization state.
 871     // This additional state is necessary because reading and testing the synchronization
 872     // state is not atomic w.r.t. GC, as this scenario demonstrates:
 873     //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
 874     //     VM thread changes sync state to synchronizing and suspends threads for GC.
 875     //     Thread A is resumed to finish this native method, but doesn't block here since it
 876     //     didn't see any synchronization is progress, and escapes.
 877     __ set(_thread_in_native_trans, G3_scratch);
 878     __ st(G3_scratch, thread_state);
 879     if(os::is_MP()) {
 880       // Write serialization page so VM thread can do a pseudo remote membar.
 881       // We use the current thread pointer to calculate a thread specific
 882       // offset to write to within the page. This minimizes bus traffic
 883       // due to cache line collision.
 884       __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
 885     }
 886     __ load_contents(sync_state, G3_scratch);
 887     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 888 
 889 
 890     Label L;
 891     Address suspend_state(G2_thread, in_bytes(JavaThread::suspend_flags_offset()));
 892     __ br(Assembler::notEqual, false, Assembler::pn, L);
 893     __ delayed()->
 894       ld(suspend_state, G3_scratch);
 895     __ cmp(G3_scratch, 0);
 896     __ br(Assembler::equal, false, Assembler::pt, no_block);
 897     __ delayed()->nop();
 898     __ bind(L);
 899 
 900     // Block.  Save any potential method result value before the operation and
 901     // use a leaf call to leave the last_Java_frame setup undisturbed.
 902     save_native_result();
 903     __ call_VM_leaf(noreg,
 904                     CAST_FROM_FN_PTR(address, JavaThread::check_safepoint_and_suspend_for_native_trans),
 905                     G2_thread);
 906     __ ld_ptr(STATE(_thread), G2_thread);  // restore thread
 907     // Restore any method result value
 908     restore_native_result();
 909     __ bind(no_block);
 910   }
 911 
 912   // Clear the frame anchor now
 913 
 914   __ reset_last_Java_frame();
 915 
 916   // Move the result handler address
 917   __ mov(Lscratch, G3_scratch);
 918   // return possible result to the outer frame
 919 #ifndef __LP64
 920   __ mov(O0, I0);
 921   __ restore(O1, G0, O1);
 922 #else
 923   __ restore(O0, G0, O0);
 924 #endif /* __LP64 */
 925 
 926   // Move result handler to expected register
 927   __ mov(G3_scratch, Lscratch);
 928 
 929 
 930   // thread state is thread_in_native_trans. Any safepoint blocking has
 931   // happened in the trampoline we are ready to switch to thread_in_Java.
 932 
 933   __ set(_thread_in_Java, G3_scratch);
 934   __ st(G3_scratch, thread_state);
 935 
 936   // If we have an oop result store it where it will be safe for any further gc
 937   // until we return now that we've released the handle it might be protected by
 938 
 939   {
 940     Label no_oop, store_result;
 941 
 942     __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
 943     __ cmp(G3_scratch, Lscratch);
 944     __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
 945     __ delayed()->nop();
 946     __ addcc(G0, O0, O0);
 947     __ brx(Assembler::notZero, true, Assembler::pt, store_result);     // if result is not NULL:
 948     __ delayed()->ld_ptr(O0, 0, O0);                                   // unbox it
 949     __ mov(G0, O0);
 950 
 951     __ bind(store_result);
 952     // Store it where gc will look for it and result handler expects it.
 953     __ st_ptr(O0, STATE(_oop_temp));
 954 
 955     __ bind(no_oop);
 956 
 957   }
 958 
 959   // reset handle block
 960   __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch);
 961   __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
 962 
 963 
 964   // handle exceptions (exception handling will handle unlocking!)
 965   { Label L;
 966     Address exception_addr (G2_thread, in_bytes(Thread::pending_exception_offset()));
 967 
 968     __ ld_ptr(exception_addr, Gtemp);
 969     __ tst(Gtemp);
 970     __ brx(Assembler::equal, false, Assembler::pt, L);
 971     __ delayed()->nop();
 972     __ bind(pending_exception_present);
 973     // With c++ interpreter we just leave it pending caller will do the correct thing. However...
 974     // Like x86 we ignore the result of the native call and leave the method locked. This
 975     // seems wrong to leave things locked.
 976 
 977     __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
 978     __ delayed()->restore(I5_savedSP, G0, SP);  // remove interpreter frame
 979 
 980     __ bind(L);
 981   }
 982 
 983   // jvmdi/jvmpi support (preserves thread register)
 984   __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
 985 
 986   if (synchronized) {
 987     // save and restore any potential method result value around the unlocking operation
 988     save_native_result();
 989 
 990     const int entry_size            = frame::interpreter_frame_monitor_size() * wordSize;
 991     // Get the initial monitor we allocated
 992     __ sub(Lstate, entry_size, O1);                        // initial monitor
 993     __ unlock_object(O1);
 994     restore_native_result();
 995   }
 996 
 997 #if defined(COMPILER2) && !defined(_LP64)
 998 
 999   // C2 expects long results in G1 we can't tell if we're returning to interpreted
1000   // or compiled so just be safe.
1001 
1002   __ sllx(O0, 32, G1);          // Shift bits into high G1
1003   __ srl (O1, 0, O1);           // Zero extend O1
1004   __ or3 (O1, G1, G1);          // OR 64 bits into G1
1005 
1006 #endif /* COMPILER2 && !_LP64 */
1007 
1008 #ifdef ASSERT
1009   {
1010     Label ok;
1011     __ cmp(I5_savedSP, FP);
1012     __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
1013     __ delayed()->nop();
1014     __ stop("bad I5_savedSP value");
1015     __ should_not_reach_here();
1016     __ bind(ok);
1017   }
1018 #endif
1019   // Calls result handler which POPS FRAME
1020   if (TraceJumps) {
1021     // Move target to register that is recordable
1022     __ mov(Lscratch, G3_scratch);
1023     __ JMP(G3_scratch, 0);
1024   } else {
1025     __ jmp(Lscratch, 0);
1026   }
1027   __ delayed()->nop();
1028 
1029   if (inc_counter) {
1030     // handle invocation counter overflow
1031     __ bind(invocation_counter_overflow);
1032     generate_counter_overflow(Lcontinue);
1033   }
1034 
1035 
1036   return entry;
1037 }
1038 
1039 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state,
1040                                                               const Register prev_state,
1041                                                               bool native) {
1042 
1043   // On entry
1044   // G5_method - caller's method
1045   // Gargs - points to initial parameters (i.e. locals[0])
1046   // G2_thread - valid? (C1 only??)
1047   // "prev_state" - contains any previous frame manager state which we must save a link
1048   //
1049   // On return
1050   // "state" is a pointer to the newly allocated  state object. We must allocate and initialize
1051   // a new interpretState object and the method expression stack.
1052 
1053   assert_different_registers(state, prev_state);
1054   assert_different_registers(prev_state, G3_scratch);
1055   const Register Gtmp = G3_scratch;
1056   const Address constMethod       (G5_method, in_bytes(Method::const_offset()));
1057   const Address access_flags      (G5_method, in_bytes(Method::access_flags_offset()));
1058 
1059   // slop factor is two extra slots on the expression stack so that
1060   // we always have room to store a result when returning from a call without parameters
1061   // that returns a result.
1062 
1063   const int slop_factor = 2*wordSize;
1064 
1065   const int fixed_size = ((sizeof(BytecodeInterpreter) + slop_factor) >> LogBytesPerWord) + // what is the slop factor?
1066                          Method::extra_stack_entries() + // extra stack for jsr 292
1067                          frame::memory_parameter_word_sp_offset +  // register save area + param window
1068                          (native ?  frame::interpreter_frame_extra_outgoing_argument_words : 0); // JNI, class
1069 
1070   // XXX G5_method valid
1071 
1072   // Now compute new frame size
1073 
1074   if (native) {
1075     const Register RconstMethod = Gtmp;
1076     const Address size_of_parameters(RconstMethod, in_bytes(ConstMethod::size_of_parameters_offset()));
1077     __ ld_ptr(constMethod, RconstMethod);
1078     __ lduh( size_of_parameters, Gtmp );
1079     __ calc_mem_param_words(Gtmp, Gtmp);     // space for native call parameters passed on the stack in words
1080   } else {
1081     // Full size expression stack
1082     __ ld_ptr(constMethod, Gtmp);
1083     __ lduh(Gtmp, in_bytes(ConstMethod::max_stack_offset()), Gtmp);
1084   }
1085   __ add(Gtmp, fixed_size, Gtmp);           // plus the fixed portion
1086 
1087   __ neg(Gtmp);                               // negative space for stack/parameters in words
1088   __ and3(Gtmp, -WordsPerLong, Gtmp);        // make multiple of 2 (SP must be 2-word aligned)
1089   __ sll(Gtmp, LogBytesPerWord, Gtmp);       // negative space for frame in bytes
1090 
1091   // Need to do stack size check here before we fault on large frames
1092 
1093   Label stack_ok;
1094 
1095   const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages :
1096                                                                               (StackRedPages+StackYellowPages);
1097 
1098 
1099   __ ld_ptr(G2_thread, in_bytes(Thread::stack_base_offset()), O0);
1100   __ ld_ptr(G2_thread, in_bytes(Thread::stack_size_offset()), O1);
1101   // compute stack bottom
1102   __ sub(O0, O1, O0);
1103 
1104   // Avoid touching the guard pages
1105   // Also a fudge for frame size of BytecodeInterpreter::run
1106   // It varies from 1k->4k depending on build type
1107   const int fudge = 6 * K;
1108 
1109   __ set(fudge + (max_pages * os::vm_page_size()), O1);
1110 
1111   __ add(O0, O1, O0);
1112   __ sub(O0, Gtmp, O0);
1113   __ cmp(SP, O0);
1114   __ brx(Assembler::greaterUnsigned, false, Assembler::pt, stack_ok);
1115   __ delayed()->nop();
1116 
1117      // throw exception return address becomes throwing pc
1118 
1119   __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
1120   __ stop("never reached");
1121 
1122   __ bind(stack_ok);
1123 
1124   __ save(SP, Gtmp, SP);                      // setup new frame and register window
1125 
1126   // New window I7 call_stub or previous activation
1127   // O6 - register save area, BytecodeInterpreter just below it, args/locals just above that
1128   //
1129   __ sub(FP, sizeof(BytecodeInterpreter), state);        // Point to new Interpreter state
1130   __ add(state, STACK_BIAS, state );         // Account for 64bit bias
1131 
1132 #define XXX_STATE(field_name) state, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
1133 
1134   // Initialize a new Interpreter state
1135   // orig_sp - caller's original sp
1136   // G2_thread - thread
1137   // Gargs - &locals[0] (unbiased?)
1138   // G5_method - method
1139   // SP (biased) - accounts for full size java stack, BytecodeInterpreter object, register save area, and register parameter save window
1140 
1141 
1142   __ set(0xdead0004, O1);
1143 
1144 
1145   __ st_ptr(Gargs, XXX_STATE(_locals));
1146   __ st_ptr(G0, XXX_STATE(_oop_temp));
1147 
1148   __ st_ptr(state, XXX_STATE(_self_link));                // point to self
1149   __ st_ptr(prev_state->after_save(), XXX_STATE(_prev_link)); // Chain interpreter states
1150   __ st_ptr(G2_thread, XXX_STATE(_thread));               // Store javathread
1151 
1152   if (native) {
1153     __ st_ptr(G0, XXX_STATE(_bcp));
1154   } else {
1155     __ ld_ptr(G5_method, in_bytes(Method::const_offset()), O2); // get ConstMethod*
1156     __ add(O2, in_bytes(ConstMethod::codes_offset()), O2);        // get bcp
1157     __ st_ptr(O2, XXX_STATE(_bcp));
1158   }
1159 
1160   __ st_ptr(G0, XXX_STATE(_mdx));
1161   __ st_ptr(G5_method, XXX_STATE(_method));
1162 
1163   __ set((int) BytecodeInterpreter::method_entry, O1);
1164   __ st(O1, XXX_STATE(_msg));
1165 
1166   __ ld_ptr(constMethod, O3);
1167   __ ld_ptr(O3, in_bytes(ConstMethod::constants_offset()), O3);
1168   __ ld_ptr(O3, ConstantPool::cache_offset_in_bytes(), O2);
1169   __ st_ptr(O2, XXX_STATE(_constants));
1170 
1171   __ st_ptr(G0, XXX_STATE(_result._to_call._callee));
1172 
1173   // Monitor base is just start of BytecodeInterpreter object;
1174   __ mov(state, O2);
1175   __ st_ptr(O2, XXX_STATE(_monitor_base));
1176 
1177   // Do we need a monitor for synchonized method?
1178   {
1179     __ ld(access_flags, O1);
1180     Label done;
1181     Label got_obj;
1182     __ btst(JVM_ACC_SYNCHRONIZED, O1);
1183     __ br( Assembler::zero, false, Assembler::pt, done);
1184 
1185     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
1186     __ delayed()->btst(JVM_ACC_STATIC, O1);
1187     __ ld_ptr(XXX_STATE(_locals), O1);
1188     __ br( Assembler::zero, true, Assembler::pt, got_obj);
1189     __ delayed()->ld_ptr(O1, 0, O1);                  // get receiver for not-static case
1190     __ ld_ptr(constMethod, O1);
1191     __ ld_ptr( O1, in_bytes(ConstMethod::constants_offset()), O1);
1192     __ ld_ptr( O1, ConstantPool::pool_holder_offset_in_bytes(), O1);
1193     // lock the mirror, not the Klass*
1194     __ ld_ptr( O1, mirror_offset, O1);
1195 
1196     __ bind(got_obj);
1197 
1198   #ifdef ASSERT
1199     __ tst(O1);
1200     __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
1201   #endif // ASSERT
1202 
1203     const int entry_size            = frame::interpreter_frame_monitor_size() * wordSize;
1204     __ sub(SP, entry_size, SP);                         // account for initial monitor
1205     __ sub(O2, entry_size, O2);                        // initial monitor
1206     __ st_ptr(O1, O2, BasicObjectLock::obj_offset_in_bytes()); // and allocate it for interpreter use
1207     __ bind(done);
1208   }
1209 
1210   // Remember initial frame bottom
1211 
1212   __ st_ptr(SP, XXX_STATE(_frame_bottom));
1213 
1214   __ st_ptr(O2, XXX_STATE(_stack_base));
1215 
1216   __ sub(O2, wordSize, O2);                    // prepush
1217   __ st_ptr(O2, XXX_STATE(_stack));                // PREPUSH
1218 
1219   // Full size expression stack
1220   __ ld_ptr(constMethod, O3);
1221   __ lduh(O3, in_bytes(ConstMethod::max_stack_offset()), O3);
1222   __ inc(O3, Method::extra_stack_entries());
1223   __ sll(O3, LogBytesPerWord, O3);
1224   __ sub(O2, O3, O3);
1225 //  __ sub(O3, wordSize, O3);                    // so prepush doesn't look out of bounds
1226   __ st_ptr(O3, XXX_STATE(_stack_limit));
1227 
1228   if (!native) {
1229     //
1230     // Code to initialize locals
1231     //
1232     Register init_value = noreg;    // will be G0 if we must clear locals
1233     // Now zero locals
1234     if (true /* zerolocals */ || ClearInterpreterLocals) {
1235       // explicitly initialize locals
1236       init_value = G0;
1237     } else {
1238     #ifdef ASSERT
1239       // initialize locals to a garbage pattern for better debugging
1240       init_value = O3;
1241       __ set( 0x0F0F0F0F, init_value );
1242     #endif // ASSERT
1243     }
1244     if (init_value != noreg) {
1245       Label clear_loop;
1246       const Register RconstMethod = O1;
1247       const Address size_of_parameters(RconstMethod, in_bytes(ConstMethod::size_of_parameters_offset()));
1248       const Address size_of_locals    (RconstMethod, in_bytes(ConstMethod::size_of_locals_offset()));
1249 
1250       // NOTE: If you change the frame layout, this code will need to
1251       // be updated!
1252       __ ld_ptr( constMethod, RconstMethod );
1253       __ lduh( size_of_locals, O2 );
1254       __ lduh( size_of_parameters, O1 );
1255       __ sll( O2, LogBytesPerWord, O2);
1256       __ sll( O1, LogBytesPerWord, O1 );
1257       __ ld_ptr(XXX_STATE(_locals), L2_scratch);
1258       __ sub( L2_scratch, O2, O2 );
1259       __ sub( L2_scratch, O1, O1 );
1260 
1261       __ bind( clear_loop );
1262       __ inc( O2, wordSize );
1263 
1264       __ cmp( O2, O1 );
1265       __ br( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1266       __ delayed()->st_ptr( init_value, O2, 0 );
1267     }
1268   }
1269 }
1270 // Find preallocated  monitor and lock method (C++ interpreter)
1271 //
1272 void InterpreterGenerator::lock_method(void) {
1273 // Lock the current method.
1274 // Destroys registers L2_scratch, L3_scratch, O0
1275 //
1276 // Find everything relative to Lstate
1277 
1278 #ifdef ASSERT
1279   __ ld_ptr(STATE(_method), L2_scratch);
1280   __ ld(L2_scratch, in_bytes(Method::access_flags_offset()), O0);
1281 
1282  { Label ok;
1283    __ btst(JVM_ACC_SYNCHRONIZED, O0);
1284    __ br( Assembler::notZero, false, Assembler::pt, ok);
1285    __ delayed()->nop();
1286    __ stop("method doesn't need synchronization");
1287    __ bind(ok);
1288   }
1289 #endif // ASSERT
1290 
1291   // monitor is already allocated at stack base
1292   // and the lockee is already present
1293   __ ld_ptr(STATE(_stack_base), L2_scratch);
1294   __ ld_ptr(L2_scratch, BasicObjectLock::obj_offset_in_bytes(), O0);   // get object
1295   __ lock_object(L2_scratch, O0);
1296 
1297 }
1298 
1299 //  Generate code for handling resuming a deopted method
1300 void CppInterpreterGenerator::generate_deopt_handling() {
1301 
1302   Label return_from_deopt_common;
1303 
1304   // deopt needs to jump to here to enter the interpreter (return a result)
1305   deopt_frame_manager_return_atos  = __ pc();
1306 
1307   // O0/O1 live
1308   __ ba(return_from_deopt_common);
1309   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_OBJECT), L3_scratch);    // Result stub address array index
1310 
1311 
1312   // deopt needs to jump to here to enter the interpreter (return a result)
1313   deopt_frame_manager_return_btos  = __ pc();
1314 
1315   // O0/O1 live
1316   __ ba(return_from_deopt_common);
1317   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_BOOLEAN), L3_scratch);    // Result stub address array index
1318 
1319   // deopt needs to jump to here to enter the interpreter (return a result)
1320   deopt_frame_manager_return_itos  = __ pc();
1321 
1322   // O0/O1 live
1323   __ ba(return_from_deopt_common);
1324   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_INT), L3_scratch);    // Result stub address array index
1325 
1326   // deopt needs to jump to here to enter the interpreter (return a result)
1327 
1328   deopt_frame_manager_return_ltos  = __ pc();
1329 #if !defined(_LP64) && defined(COMPILER2)
1330   // All return values are where we want them, except for Longs.  C2 returns
1331   // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
1332   // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
1333   // build even if we are returning from interpreted we just do a little
1334   // stupid shuffing.
1335   // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
1336   // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
1337   // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
1338 
1339   __ srl (G1, 0,O1);
1340   __ srlx(G1,32,O0);
1341 #endif /* !_LP64 && COMPILER2 */
1342   // O0/O1 live
1343   __ ba(return_from_deopt_common);
1344   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_LONG), L3_scratch);    // Result stub address array index
1345 
1346   // deopt needs to jump to here to enter the interpreter (return a result)
1347 
1348   deopt_frame_manager_return_ftos  = __ pc();
1349   // O0/O1 live
1350   __ ba(return_from_deopt_common);
1351   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_FLOAT), L3_scratch);    // Result stub address array index
1352 
1353   // deopt needs to jump to here to enter the interpreter (return a result)
1354   deopt_frame_manager_return_dtos  = __ pc();
1355 
1356   // O0/O1 live
1357   __ ba(return_from_deopt_common);
1358   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_DOUBLE), L3_scratch);    // Result stub address array index
1359 
1360   // deopt needs to jump to here to enter the interpreter (return a result)
1361   deopt_frame_manager_return_vtos  = __ pc();
1362 
1363   // O0/O1 live
1364   __ set(AbstractInterpreter::BasicType_as_index(T_VOID), L3_scratch);
1365 
1366   // Deopt return common
1367   // an index is present that lets us move any possible result being
1368   // return to the interpreter's stack
1369   //
1370   __ bind(return_from_deopt_common);
1371 
1372   // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1373   // stack is in the state that the  calling convention left it.
1374   // Copy the result from native abi result and place it on java expression stack.
1375 
1376   // Current interpreter state is present in Lstate
1377 
1378   // Get current pre-pushed top of interpreter stack
1379   // Any result (if any) is in native abi
1380   // result type index is in L3_scratch
1381 
1382   __ ld_ptr(STATE(_stack), L1_scratch);                                          // get top of java expr stack
1383 
1384   __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1385   __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1386   __ ld_ptr(L4_scratch, L3_scratch, Lscratch);                                       // get typed result converter address
1387   __ jmpl(Lscratch, G0, O7);                                         // and convert it
1388   __ delayed()->nop();
1389 
1390   // L1_scratch points to top of stack (prepushed)
1391   __ st_ptr(L1_scratch, STATE(_stack));
1392 }
1393 
1394 // Generate the code to handle a more_monitors message from the c++ interpreter
1395 void CppInterpreterGenerator::generate_more_monitors() {
1396 
1397   Label entry, loop;
1398   const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1399   // 1. compute new pointers                                // esp: old expression stack top
1400   __ delayed()->ld_ptr(STATE(_stack_base), L4_scratch);            // current expression stack bottom
1401   __ sub(L4_scratch, entry_size, L4_scratch);
1402   __ st_ptr(L4_scratch, STATE(_stack_base));
1403 
1404   __ sub(SP, entry_size, SP);                  // Grow stack
1405   __ st_ptr(SP, STATE(_frame_bottom));
1406 
1407   __ ld_ptr(STATE(_stack_limit), L2_scratch);
1408   __ sub(L2_scratch, entry_size, L2_scratch);
1409   __ st_ptr(L2_scratch, STATE(_stack_limit));
1410 
1411   __ ld_ptr(STATE(_stack), L1_scratch);                // Get current stack top
1412   __ sub(L1_scratch, entry_size, L1_scratch);
1413   __ st_ptr(L1_scratch, STATE(_stack));
1414   __ ba(entry);
1415   __ delayed()->add(L1_scratch, wordSize, L1_scratch);        // first real entry (undo prepush)
1416 
1417   // 2. move expression stack
1418 
1419   __ bind(loop);
1420   __ st_ptr(L3_scratch, Address(L1_scratch, 0));
1421   __ add(L1_scratch, wordSize, L1_scratch);
1422   __ bind(entry);
1423   __ cmp(L1_scratch, L4_scratch);
1424   __ br(Assembler::notEqual, false, Assembler::pt, loop);
1425   __ delayed()->ld_ptr(L1_scratch, entry_size, L3_scratch);
1426 
1427   // now zero the slot so we can find it.
1428   __ st_ptr(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes());
1429 
1430 }
1431 
1432 // Initial entry to C++ interpreter from the call_stub.
1433 // This entry point is called the frame manager since it handles the generation
1434 // of interpreter activation frames via requests directly from the vm (via call_stub)
1435 // and via requests from the interpreter. The requests from the call_stub happen
1436 // directly thru the entry point. Requests from the interpreter happen via returning
1437 // from the interpreter and examining the message the interpreter has returned to
1438 // the frame manager. The frame manager can take the following requests:
1439 
1440 // NO_REQUEST - error, should never happen.
1441 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and
1442 //                 allocate a new monitor.
1443 // CALL_METHOD - setup a new activation to call a new method. Very similar to what
1444 //               happens during entry during the entry via the call stub.
1445 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub.
1446 //
1447 // Arguments:
1448 //
1449 // ebx: Method*
1450 // ecx: receiver - unused (retrieved from stack as needed)
1451 // esi: previous frame manager state (NULL from the call_stub/c1/c2)
1452 //
1453 //
1454 // Stack layout at entry
1455 //
1456 // [ return address     ] <--- esp
1457 // [ parameter n        ]
1458 //   ...
1459 // [ parameter 1        ]
1460 // [ expression stack   ]
1461 //
1462 //
1463 // We are free to blow any registers we like because the call_stub which brought us here
1464 // initially has preserved the callee save registers already.
1465 //
1466 //
1467 
1468 static address interpreter_frame_manager = NULL;
1469 
1470 #ifdef ASSERT
1471   #define VALIDATE_STATE(scratch, marker)                         \
1472   {                                                               \
1473     Label skip;                                                   \
1474     __ ld_ptr(STATE(_self_link), scratch);                        \
1475     __ cmp(Lstate, scratch);                                      \
1476     __ brx(Assembler::equal, false, Assembler::pt, skip);         \
1477     __ delayed()->nop();                                          \
1478     __ breakpoint_trap();                                         \
1479     __ emit_int32(marker);                                         \
1480     __ bind(skip);                                                \
1481   }
1482 #else
1483   #define VALIDATE_STATE(scratch, marker)
1484 #endif /* ASSERT */
1485 
1486 void CppInterpreterGenerator::adjust_callers_stack(Register args) {
1487 //
1488 // Adjust caller's stack so that all the locals can be contiguous with
1489 // the parameters.
1490 // Worries about stack overflow make this a pain.
1491 //
1492 // Destroys args, G3_scratch, G3_scratch
1493 // In/Out O5_savedSP (sender's original SP)
1494 //
1495 //  assert_different_registers(state, prev_state);
1496   const Register Gtmp = G3_scratch;
1497   const Register RconstMethod = G3_scratch;
1498   const Register tmp = O2;
1499   const Address constMethod(G5_method, in_bytes(Method::const_offset()));
1500   const Address size_of_parameters(RconstMethod, in_bytes(ConstMethod::size_of_parameters_offset()));
1501   const Address size_of_locals    (RconstMethod, in_bytes(ConstMethod::size_of_locals_offset()));
1502 
1503   __ ld_ptr(constMethod, RconstMethod);
1504   __ lduh(size_of_parameters, tmp);
1505   __ sll(tmp, LogBytesPerWord, Gargs);       // parameter size in bytes
1506   __ add(args, Gargs, Gargs);                // points to first local + BytesPerWord
1507   // NEW
1508   __ add(Gargs, -wordSize, Gargs);             // points to first local[0]
1509   // determine extra space for non-argument locals & adjust caller's SP
1510   // Gtmp1: parameter size in words
1511   __ lduh(size_of_locals, Gtmp);
1512   __ compute_extra_locals_size_in_bytes(tmp, Gtmp, Gtmp);
1513 
1514 #if 1
1515   // c2i adapters place the final interpreter argument in the register save area for O0/I0
1516   // the call_stub will place the final interpreter argument at
1517   // frame::memory_parameter_word_sp_offset. This is mostly not noticable for either asm
1518   // or c++ interpreter. However with the c++ interpreter when we do a recursive call
1519   // and try to make it look good in the debugger we will store the argument to
1520   // RecursiveInterpreterActivation in the register argument save area. Without allocating
1521   // extra space for the compiler this will overwrite locals in the local array of the
1522   // interpreter.
1523   // QQQ still needed with frameless adapters???
1524 
1525   const int c2i_adjust_words = frame::memory_parameter_word_sp_offset - frame::callee_register_argument_save_area_sp_offset;
1526 
1527   __ add(Gtmp, c2i_adjust_words*wordSize, Gtmp);
1528 #endif // 1
1529 
1530 
1531   __ sub(SP, Gtmp, SP);                      // just caller's frame for the additional space we need.
1532 }
1533 
1534 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1535 
1536   // G5_method: Method*
1537   // G2_thread: thread (unused)
1538   // Gargs:   bottom of args (sender_sp)
1539   // O5: sender's sp
1540 
1541   // A single frame manager is plenty as we don't specialize for synchronized. We could and
1542   // the code is pretty much ready. Would need to change the test below and for good measure
1543   // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized
1544   // routines. Not clear this is worth it yet.
1545 
1546   if (interpreter_frame_manager) {
1547     return interpreter_frame_manager;
1548   }
1549 
1550   __ bind(frame_manager_entry);
1551 
1552   // the following temporary registers are used during frame creation
1553   const Register Gtmp1 = G3_scratch;
1554   // const Register Lmirror = L1;     // native mirror (native calls only)
1555 
1556   const Address constMethod       (G5_method, in_bytes(Method::const_offset()));
1557   const Address access_flags      (G5_method, in_bytes(Method::access_flags_offset()));
1558 
1559   address entry_point = __ pc();
1560   __ mov(G0, prevState);                                                 // no current activation
1561 
1562 
1563   Label re_dispatch;
1564 
1565   __ bind(re_dispatch);
1566 
1567   // Interpreter needs to have locals completely contiguous. In order to do that
1568   // We must adjust the caller's stack pointer for any locals beyond just the
1569   // parameters
1570   adjust_callers_stack(Gargs);
1571 
1572   // O5_savedSP still contains sender's sp
1573 
1574   // NEW FRAME
1575 
1576   generate_compute_interpreter_state(Lstate, prevState, false);
1577 
1578   // At this point a new interpreter frame and state object are created and initialized
1579   // Lstate has the pointer to the new activation
1580   // Any stack banging or limit check should already be done.
1581 
1582   Label call_interpreter;
1583 
1584   __ bind(call_interpreter);
1585 
1586 
1587 #if 1
1588   __ set(0xdead002, Lmirror);
1589   __ set(0xdead002, L2_scratch);
1590   __ set(0xdead003, L3_scratch);
1591   __ set(0xdead004, L4_scratch);
1592   __ set(0xdead005, Lscratch);
1593   __ set(0xdead006, Lscratch2);
1594   __ set(0xdead007, L7_scratch);
1595 
1596   __ set(0xdeaf002, O2);
1597   __ set(0xdeaf003, O3);
1598   __ set(0xdeaf004, O4);
1599   __ set(0xdeaf005, O5);
1600 #endif
1601 
1602   // Call interpreter (stack bang complete) enter here if message is
1603   // set and we know stack size is valid
1604 
1605   Label call_interpreter_2;
1606 
1607   __ bind(call_interpreter_2);
1608 
1609 #ifdef ASSERT
1610   {
1611     Label skip;
1612     __ ld_ptr(STATE(_frame_bottom), G3_scratch);
1613     __ cmp(G3_scratch, SP);
1614     __ brx(Assembler::equal, false, Assembler::pt, skip);
1615     __ delayed()->nop();
1616     __ stop("SP not restored to frame bottom");
1617     __ bind(skip);
1618   }
1619 #endif
1620 
1621   VALIDATE_STATE(G3_scratch, 4);
1622   __ set_last_Java_frame(SP, noreg);
1623   __ mov(Lstate, O0);                 // (arg) pointer to current state
1624 
1625   __ call(CAST_FROM_FN_PTR(address,
1626                            JvmtiExport::can_post_interpreter_events() ?
1627                                                                   BytecodeInterpreter::runWithChecks
1628                                                                 : BytecodeInterpreter::run),
1629          relocInfo::runtime_call_type);
1630 
1631   __ delayed()->nop();
1632 
1633   __ ld_ptr(STATE(_thread), G2_thread);
1634   __ reset_last_Java_frame();
1635 
1636   // examine msg from interpreter to determine next action
1637   __ ld_ptr(STATE(_thread), G2_thread);                                  // restore G2_thread
1638 
1639   __ ld(STATE(_msg), L1_scratch);                                       // Get new message
1640 
1641   Label call_method;
1642   Label return_from_interpreted_method;
1643   Label throw_exception;
1644   Label do_OSR;
1645   Label bad_msg;
1646   Label resume_interpreter;
1647 
1648   __ cmp(L1_scratch, (int)BytecodeInterpreter::call_method);
1649   __ br(Assembler::equal, false, Assembler::pt, call_method);
1650   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::return_from_method);
1651   __ br(Assembler::equal, false, Assembler::pt, return_from_interpreted_method);
1652   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::throwing_exception);
1653   __ br(Assembler::equal, false, Assembler::pt, throw_exception);
1654   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::do_osr);
1655   __ br(Assembler::equal, false, Assembler::pt, do_OSR);
1656   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::more_monitors);
1657   __ br(Assembler::notEqual, false, Assembler::pt, bad_msg);
1658 
1659   // Allocate more monitor space, shuffle expression stack....
1660 
1661   generate_more_monitors();
1662 
1663   // new monitor slot allocated, resume the interpreter.
1664 
1665   __ set((int)BytecodeInterpreter::got_monitors, L1_scratch);
1666   VALIDATE_STATE(G3_scratch, 5);
1667   __ ba(call_interpreter);
1668   __ delayed()->st(L1_scratch, STATE(_msg));
1669 
1670   // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode)
1671   unctrap_frame_manager_entry  = __ pc();
1672 
1673   // QQQ what message do we send
1674 
1675   __ ba(call_interpreter);
1676   __ delayed()->ld_ptr(STATE(_frame_bottom), SP);                  // restore to full stack frame
1677 
1678   //=============================================================================
1679   // Returning from a compiled method into a deopted method. The bytecode at the
1680   // bcp has completed. The result of the bytecode is in the native abi (the tosca
1681   // for the template based interpreter). Any stack space that was used by the
1682   // bytecode that has completed has been removed (e.g. parameters for an invoke)
1683   // so all that we have to do is place any pending result on the expression stack
1684   // and resume execution on the next bytecode.
1685 
1686   generate_deopt_handling();
1687 
1688   // ready to resume the interpreter
1689 
1690   __ set((int)BytecodeInterpreter::deopt_resume, L1_scratch);
1691   __ ba(call_interpreter);
1692   __ delayed()->st(L1_scratch, STATE(_msg));
1693 
1694   // Current frame has caught an exception we need to dispatch to the
1695   // handler. We can get here because a native interpreter frame caught
1696   // an exception in which case there is no handler and we must rethrow
1697   // If it is a vanilla interpreted frame the we simply drop into the
1698   // interpreter and let it do the lookup.
1699 
1700   Interpreter::_rethrow_exception_entry = __ pc();
1701 
1702   Label return_with_exception;
1703   Label unwind_and_forward;
1704 
1705   // O0: exception
1706   // O7: throwing pc
1707 
1708   // We want exception in the thread no matter what we ultimately decide about frame type.
1709 
1710   Address exception_addr (G2_thread, in_bytes(Thread::pending_exception_offset()));
1711   __ verify_thread();
1712   __ st_ptr(O0, exception_addr);
1713 
1714   // get the Method*
1715   __ ld_ptr(STATE(_method), G5_method);
1716 
1717   // if this current frame vanilla or native?
1718 
1719   __ ld(access_flags, Gtmp1);
1720   __ btst(JVM_ACC_NATIVE, Gtmp1);
1721   __ br(Assembler::zero, false, Assembler::pt, return_with_exception);  // vanilla interpreted frame handle directly
1722   __ delayed()->nop();
1723 
1724   // We drop thru to unwind a native interpreted frame with a pending exception
1725   // We jump here for the initial interpreter frame with exception pending
1726   // We unwind the current acivation and forward it to our caller.
1727 
1728   __ bind(unwind_and_forward);
1729 
1730   // Unwind frame and jump to forward exception. unwinding will place throwing pc in O7
1731   // as expected by forward_exception.
1732 
1733   __ restore(FP, G0, SP);                  // unwind interpreter state frame
1734   __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
1735   __ delayed()->mov(I5_savedSP->after_restore(), SP);
1736 
1737   // Return point from a call which returns a result in the native abi
1738   // (c1/c2/jni-native). This result must be processed onto the java
1739   // expression stack.
1740   //
1741   // A pending exception may be present in which case there is no result present
1742 
1743   address return_from_native_method = __ pc();
1744 
1745   VALIDATE_STATE(G3_scratch, 6);
1746 
1747   // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1748   // stack is in the state that the  calling convention left it.
1749   // Copy the result from native abi result and place it on java expression stack.
1750 
1751   // Current interpreter state is present in Lstate
1752 
1753   // Exception pending?
1754 
1755   __ ld_ptr(STATE(_frame_bottom), SP);                             // restore to full stack frame
1756   __ ld_ptr(exception_addr, Lscratch);                                         // get any pending exception
1757   __ tst(Lscratch);                                                            // exception pending?
1758   __ brx(Assembler::notZero, false, Assembler::pt, return_with_exception);
1759   __ delayed()->nop();
1760 
1761   // Process the native abi result to java expression stack
1762 
1763   __ ld_ptr(STATE(_result._to_call._callee), L4_scratch);                        // called method
1764   __ ld_ptr(STATE(_stack), L1_scratch);                                          // get top of java expr stack
1765   // get parameter size
1766   __ ld_ptr(L4_scratch, in_bytes(Method::const_offset()), L2_scratch);
1767   __ lduh(L2_scratch, in_bytes(ConstMethod::size_of_parameters_offset()), L2_scratch);
1768   __ sll(L2_scratch, LogBytesPerWord, L2_scratch     );                           // parameter size in bytes
1769   __ add(L1_scratch, L2_scratch, L1_scratch);                                      // stack destination for result
1770   __ ld(L4_scratch, in_bytes(Method::result_index_offset()), L3_scratch); // called method result type index
1771 
1772   // tosca is really just native abi
1773   __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1774   __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1775   __ ld_ptr(L4_scratch, L3_scratch, Lscratch);                                       // get typed result converter address
1776   __ jmpl(Lscratch, G0, O7);                                                   // and convert it
1777   __ delayed()->nop();
1778 
1779   // L1_scratch points to top of stack (prepushed)
1780 
1781   __ ba(resume_interpreter);
1782   __ delayed()->mov(L1_scratch, O1);
1783 
1784   // An exception is being caught on return to a vanilla interpreter frame.
1785   // Empty the stack and resume interpreter
1786 
1787   __ bind(return_with_exception);
1788 
1789   __ ld_ptr(STATE(_frame_bottom), SP);                             // restore to full stack frame
1790   __ ld_ptr(STATE(_stack_base), O1);                               // empty java expression stack
1791   __ ba(resume_interpreter);
1792   __ delayed()->sub(O1, wordSize, O1);                             // account for prepush
1793 
1794   // Return from interpreted method we return result appropriate to the caller (i.e. "recursive"
1795   // interpreter call, or native) and unwind this interpreter activation.
1796   // All monitors should be unlocked.
1797 
1798   __ bind(return_from_interpreted_method);
1799 
1800   VALIDATE_STATE(G3_scratch, 7);
1801 
1802   Label return_to_initial_caller;
1803 
1804   // Interpreted result is on the top of the completed activation expression stack.
1805   // We must return it to the top of the callers stack if caller was interpreted
1806   // otherwise we convert to native abi result and return to call_stub/c1/c2
1807   // The caller's expression stack was truncated by the call however the current activation
1808   // has enough stuff on the stack that we have usable space there no matter what. The
1809   // other thing that makes it easy is that the top of the caller's stack is stored in STATE(_locals)
1810   // for the current activation
1811 
1812   __ ld_ptr(STATE(_prev_link), L1_scratch);
1813   __ ld_ptr(STATE(_method), L2_scratch);                               // get method just executed
1814   __ ld(L2_scratch, in_bytes(Method::result_index_offset()), L2_scratch);
1815   __ tst(L1_scratch);
1816   __ brx(Assembler::zero, false, Assembler::pt, return_to_initial_caller);
1817   __ delayed()->sll(L2_scratch, LogBytesPerWord, L2_scratch);
1818 
1819   // Copy result to callers java stack
1820 
1821   __ set((intptr_t)CppInterpreter::_stack_to_stack, L4_scratch);
1822   __ ld_ptr(L4_scratch, L2_scratch, Lscratch);                          // get typed result converter address
1823   __ ld_ptr(STATE(_stack), O0);                                       // current top (prepushed)
1824   __ ld_ptr(STATE(_locals), O1);                                      // stack destination
1825 
1826   // O0 - will be source, O1 - will be destination (preserved)
1827   __ jmpl(Lscratch, G0, O7);                                          // and convert it
1828   __ delayed()->add(O0, wordSize, O0);                                // get source (top of current expr stack)
1829 
1830   // O1 == &locals[0]
1831 
1832   // Result is now on caller's stack. Just unwind current activation and resume
1833 
1834   Label unwind_recursive_activation;
1835 
1836 
1837   __ bind(unwind_recursive_activation);
1838 
1839   // O1 == &locals[0] (really callers stacktop) for activation now returning
1840   // returning to interpreter method from "recursive" interpreter call
1841   // result converter left O1 pointing to top of the( prepushed) java stack for method we are returning
1842   // to. Now all we must do is unwind the state from the completed call
1843 
1844   // Must restore stack
1845   VALIDATE_STATE(G3_scratch, 8);
1846 
1847   // Return to interpreter method after a method call (interpreted/native/c1/c2) has completed.
1848   // Result if any is already on the caller's stack. All we must do now is remove the now dead
1849   // frame and tell interpreter to resume.
1850 
1851 
1852   __ mov(O1, I1);                                                     // pass back new stack top across activation
1853   // POP FRAME HERE ==================================
1854   __ restore(FP, G0, SP);                                             // unwind interpreter state frame
1855   __ ld_ptr(STATE(_frame_bottom), SP);                                // restore to full stack frame
1856 
1857 
1858   // Resume the interpreter. The current frame contains the current interpreter
1859   // state object.
1860   //
1861   // O1 == new java stack pointer
1862 
1863   __ bind(resume_interpreter);
1864   VALIDATE_STATE(G3_scratch, 10);
1865 
1866   // A frame we have already used before so no need to bang stack so use call_interpreter_2 entry
1867 
1868   __ set((int)BytecodeInterpreter::method_resume, L1_scratch);
1869   __ st(L1_scratch, STATE(_msg));
1870   __ ba(call_interpreter_2);
1871   __ delayed()->st_ptr(O1, STATE(_stack));
1872 
1873 
1874   // Fast accessor methods share this entry point.
1875   // This works because frame manager is in the same codelet
1876   // This can either be an entry via call_stub/c1/c2 or a recursive interpreter call
1877   // we need to do a little register fixup here once we distinguish the two of them
1878   if (UseFastAccessorMethods && !synchronized) {
1879   // Call stub_return address still in O7
1880     __ bind(fast_accessor_slow_entry_path);
1881     __ set((intptr_t)return_from_native_method - 8, Gtmp1);
1882     __ cmp(Gtmp1, O7);                                                // returning to interpreter?
1883     __ brx(Assembler::equal, true, Assembler::pt, re_dispatch);       // yep
1884     __ delayed()->nop();
1885     __ ba(re_dispatch);
1886     __ delayed()->mov(G0, prevState);                                 // initial entry
1887 
1888   }
1889 
1890   // interpreter returning to native code (call_stub/c1/c2)
1891   // convert result and unwind initial activation
1892   // L2_scratch - scaled result type index
1893 
1894   __ bind(return_to_initial_caller);
1895 
1896   __ set((intptr_t)CppInterpreter::_stack_to_native_abi, L4_scratch);
1897   __ ld_ptr(L4_scratch, L2_scratch, Lscratch);                           // get typed result converter address
1898   __ ld_ptr(STATE(_stack), O0);                                        // current top (prepushed)
1899   __ jmpl(Lscratch, G0, O7);                                           // and convert it
1900   __ delayed()->add(O0, wordSize, O0);                                 // get source (top of current expr stack)
1901 
1902   Label unwind_initial_activation;
1903   __ bind(unwind_initial_activation);
1904 
1905   // RETURN TO CALL_STUB/C1/C2 code (result if any in I0..I1/(F0/..F1)
1906   // we can return here with an exception that wasn't handled by interpreted code
1907   // how does c1/c2 see it on return?
1908 
1909   // compute resulting sp before/after args popped depending upon calling convention
1910   // __ ld_ptr(STATE(_saved_sp), Gtmp1);
1911   //
1912   // POP FRAME HERE ==================================
1913   __ restore(FP, G0, SP);
1914   __ retl();
1915   __ delayed()->mov(I5_savedSP->after_restore(), SP);
1916 
1917   // OSR request, unwind the current frame and transfer to the OSR entry
1918   // and enter OSR nmethod
1919 
1920   __ bind(do_OSR);
1921   Label remove_initial_frame;
1922   __ ld_ptr(STATE(_prev_link), L1_scratch);
1923   __ ld_ptr(STATE(_result._osr._osr_buf), G1_scratch);
1924 
1925   // We are going to pop this frame. Is there another interpreter frame underneath
1926   // it or is it callstub/compiled?
1927 
1928   __ tst(L1_scratch);
1929   __ brx(Assembler::zero, false, Assembler::pt, remove_initial_frame);
1930   __ delayed()->ld_ptr(STATE(_result._osr._osr_entry), G3_scratch);
1931 
1932   // Frame underneath is an interpreter frame simply unwind
1933   // POP FRAME HERE ==================================
1934   __ restore(FP, G0, SP);                                             // unwind interpreter state frame
1935   __ mov(I5_savedSP->after_restore(), SP);
1936 
1937   // Since we are now calling native need to change our "return address" from the
1938   // dummy RecursiveInterpreterActivation to a return from native
1939 
1940   __ set((intptr_t)return_from_native_method - 8, O7);
1941 
1942   __ jmpl(G3_scratch, G0, G0);
1943   __ delayed()->mov(G1_scratch, O0);
1944 
1945   __ bind(remove_initial_frame);
1946 
1947   // POP FRAME HERE ==================================
1948   __ restore(FP, G0, SP);
1949   __ mov(I5_savedSP->after_restore(), SP);
1950   __ jmpl(G3_scratch, G0, G0);
1951   __ delayed()->mov(G1_scratch, O0);
1952 
1953   // Call a new method. All we do is (temporarily) trim the expression stack
1954   // push a return address to bring us back to here and leap to the new entry.
1955   // At this point we have a topmost frame that was allocated by the frame manager
1956   // which contains the current method interpreted state. We trim this frame
1957   // of excess java expression stack entries and then recurse.
1958 
1959   __ bind(call_method);
1960 
1961   // stack points to next free location and not top element on expression stack
1962   // method expects sp to be pointing to topmost element
1963 
1964   __ ld_ptr(STATE(_thread), G2_thread);
1965   __ ld_ptr(STATE(_result._to_call._callee), G5_method);
1966 
1967 
1968   // SP already takes in to account the 2 extra words we use for slop
1969   // when we call a "static long no_params()" method. So if
1970   // we trim back sp by the amount of unused java expression stack
1971   // there will be automagically the 2 extra words we need.
1972   // We also have to worry about keeping SP aligned.
1973 
1974   __ ld_ptr(STATE(_stack), Gargs);
1975   __ ld_ptr(STATE(_stack_limit), L1_scratch);
1976 
1977   // compute the unused java stack size
1978   __ sub(Gargs, L1_scratch, L2_scratch);                       // compute unused space
1979 
1980   // Round down the unused space to that stack is always 16-byte aligned
1981   // by making the unused space a multiple of the size of two longs.
1982 
1983   __ and3(L2_scratch, -2*BytesPerLong, L2_scratch);
1984 
1985   // Now trim the stack
1986   __ add(SP, L2_scratch, SP);
1987 
1988 
1989   // Now point to the final argument (account for prepush)
1990   __ add(Gargs, wordSize, Gargs);
1991 #ifdef ASSERT
1992   // Make sure we have space for the window
1993   __ sub(Gargs, SP, L1_scratch);
1994   __ cmp(L1_scratch, 16*wordSize);
1995   {
1996     Label skip;
1997     __ brx(Assembler::greaterEqual, false, Assembler::pt, skip);
1998     __ delayed()->nop();
1999     __ stop("killed stack");
2000     __ bind(skip);
2001   }
2002 #endif // ASSERT
2003 
2004   // Create a new frame where we can store values that make it look like the interpreter
2005   // really recursed.
2006 
2007   // prepare to recurse or call specialized entry
2008 
2009   // First link the registers we need
2010 
2011   // make the pc look good in debugger
2012   __ set(CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation), O7);
2013   // argument too
2014   __ mov(Lstate, I0);
2015 
2016   // Record our sending SP
2017   __ mov(SP, O5_savedSP);
2018 
2019   __ ld_ptr(STATE(_result._to_call._callee_entry_point), L2_scratch);
2020   __ set((intptr_t) entry_point, L1_scratch);
2021   __ cmp(L1_scratch, L2_scratch);
2022   __ brx(Assembler::equal, false, Assembler::pt, re_dispatch);
2023   __ delayed()->mov(Lstate, prevState);                                // link activations
2024 
2025   // method uses specialized entry, push a return so we look like call stub setup
2026   // this path will handle fact that result is returned in registers and not
2027   // on the java stack.
2028 
2029   __ set((intptr_t)return_from_native_method - 8, O7);
2030   __ jmpl(L2_scratch, G0, G0);                               // Do specialized entry
2031   __ delayed()->nop();
2032 
2033   //
2034   // Bad Message from interpreter
2035   //
2036   __ bind(bad_msg);
2037   __ stop("Bad message from interpreter");
2038 
2039   // Interpreted method "returned" with an exception pass it on...
2040   // Pass result, unwind activation and continue/return to interpreter/call_stub
2041   // We handle result (if any) differently based on return to interpreter or call_stub
2042 
2043   __ bind(throw_exception);
2044   __ ld_ptr(STATE(_prev_link), L1_scratch);
2045   __ tst(L1_scratch);
2046   __ brx(Assembler::zero, false, Assembler::pt, unwind_and_forward);
2047   __ delayed()->nop();
2048 
2049   __ ld_ptr(STATE(_locals), O1); // get result of popping callee's args
2050   __ ba(unwind_recursive_activation);
2051   __ delayed()->nop();
2052 
2053   interpreter_frame_manager = entry_point;
2054   return entry_point;
2055 }
2056 
2057 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2058  : CppInterpreterGenerator(code) {
2059    generate_all(); // down here so it can be "virtual"
2060 }
2061 
2062 
2063 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
2064 
2065   // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
2066   // expression stack, the callee will have callee_extra_locals (so we can account for
2067   // frame extension) and monitor_size for monitors. Basically we need to calculate
2068   // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
2069   //
2070   //
2071   // The big complicating thing here is that we must ensure that the stack stays properly
2072   // aligned. This would be even uglier if monitor size wasn't modulo what the stack
2073   // needs to be aligned for). We are given that the sp (fp) is already aligned by
2074   // the caller so we must ensure that it is properly aligned for our callee.
2075   //
2076   // Ths c++ interpreter always makes sure that we have a enough extra space on the
2077   // stack at all times to deal with the "stack long no_params()" method issue. This
2078   // is "slop_factor" here.
2079   const int slop_factor = 2;
2080 
2081   const int fixed_size = sizeof(BytecodeInterpreter)/wordSize +           // interpreter state object
2082                          frame::memory_parameter_word_sp_offset;   // register save area + param window
2083   return (round_to(max_stack +
2084                    slop_factor +
2085                    fixed_size +
2086                    monitor_size +
2087                    (callee_extra_locals * Interpreter::stackElementWords), WordsPerLong));
2088 
2089 }
2090 
2091 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
2092 
2093   // See call_stub code
2094   int call_stub_size  = round_to(7 + frame::memory_parameter_word_sp_offset,
2095                                  WordsPerLong);    // 7 + register save area
2096 
2097   // Save space for one monitor to get into the interpreted method in case
2098   // the method is synchronized
2099   int monitor_size    = method->is_synchronized() ?
2100                                 1*frame::interpreter_frame_monitor_size() : 0;
2101   return size_activation_helper(method->max_locals(), method->max_stack(),
2102                                  monitor_size) + call_stub_size;
2103 }
2104 
2105 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
2106                                            frame* caller,
2107                                            frame* current,
2108                                            Method* method,
2109                                            intptr_t* locals,
2110                                            intptr_t* stack,
2111                                            intptr_t* stack_base,
2112                                            intptr_t* monitor_base,
2113                                            intptr_t* frame_bottom,
2114                                            bool is_top_frame
2115                                            )
2116 {
2117   // What about any vtable?
2118   //
2119   to_fill->_thread = JavaThread::current();
2120   // This gets filled in later but make it something recognizable for now
2121   to_fill->_bcp = method->code_base();
2122   to_fill->_locals = locals;
2123   to_fill->_constants = method->constants()->cache();
2124   to_fill->_method = method;
2125   to_fill->_mdx = NULL;
2126   to_fill->_stack = stack;
2127   if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) {
2128     to_fill->_msg = deopt_resume2;
2129   } else {
2130     to_fill->_msg = method_resume;
2131   }
2132   to_fill->_result._to_call._bcp_advance = 0;
2133   to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2134   to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2135   to_fill->_prev_link = NULL;
2136 
2137   // Fill in the registers for the frame
2138 
2139   // Need to install _sender_sp. Actually not too hard in C++!
2140   // When the skeletal frames are layed out we fill in a value
2141   // for _sender_sp. That value is only correct for the oldest
2142   // skeletal frame constructed (because there is only a single
2143   // entry for "caller_adjustment". While the skeletal frames
2144   // exist that is good enough. We correct that calculation
2145   // here and get all the frames correct.
2146 
2147   // to_fill->_sender_sp = locals - (method->size_of_parameters() - 1);
2148 
2149   *current->register_addr(Lstate) = (intptr_t) to_fill;
2150   // skeletal already places a useful value here and this doesn't account
2151   // for alignment so don't bother.
2152   // *current->register_addr(I5_savedSP) =     (intptr_t) locals - (method->size_of_parameters() - 1);
2153 
2154   if (caller->is_interpreted_frame()) {
2155     interpreterState prev  = caller->get_interpreterState();
2156     to_fill->_prev_link = prev;
2157     // Make the prev callee look proper
2158     prev->_result._to_call._callee = method;
2159     if (*prev->_bcp == Bytecodes::_invokeinterface) {
2160       prev->_result._to_call._bcp_advance = 5;
2161     } else {
2162       prev->_result._to_call._bcp_advance = 3;
2163     }
2164   }
2165   to_fill->_oop_temp = NULL;
2166   to_fill->_stack_base = stack_base;
2167   // Need +1 here because stack_base points to the word just above the first expr stack entry
2168   // and stack_limit is supposed to point to the word just below the last expr stack entry.
2169   // See generate_compute_interpreter_state.
2170   to_fill->_stack_limit = stack_base - (method->max_stack() + 1);
2171   to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2172 
2173   // sparc specific
2174   to_fill->_frame_bottom = frame_bottom;
2175   to_fill->_self_link = to_fill;
2176 #ifdef ASSERT
2177   to_fill->_native_fresult = 123456.789;
2178   to_fill->_native_lresult = CONST64(0xdeadcafedeafcafe);
2179 #endif
2180 }
2181 
2182 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, address last_Java_pc, intptr_t* last_Java_fp) {
2183   istate->_last_Java_pc = (intptr_t*) last_Java_pc;
2184 }
2185 
2186 
2187 int AbstractInterpreter::layout_activation(Method* method,
2188                                            int tempcount, // Number of slots on java expression stack in use
2189                                            int popframe_extra_args,
2190                                            int moncount,  // Number of active monitors
2191                                            int caller_actual_parameters,
2192                                            int callee_param_size,
2193                                            int callee_locals_size,
2194                                            frame* caller,
2195                                            frame* interpreter_frame,
2196                                            bool is_top_frame,
2197                                            bool is_bottom_frame) {
2198 
2199   assert(popframe_extra_args == 0, "NEED TO FIX");
2200   // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state()
2201   // does as far as allocating an interpreter frame.
2202   // If interpreter_frame!=NULL, set up the method, locals, and monitors.
2203   // The frame interpreter_frame, if not NULL, is guaranteed to be the right size,
2204   // as determined by a previous call to this method.
2205   // It is also guaranteed to be walkable even though it is in a skeletal state
2206   // NOTE: return size is in words not bytes
2207   // NOTE: tempcount is the current size of the java expression stack. For top most
2208   //       frames we will allocate a full sized expression stack and not the curback
2209   //       version that non-top frames have.
2210 
2211   // Calculate the amount our frame will be adjust by the callee. For top frame
2212   // this is zero.
2213 
2214   // NOTE: ia64 seems to do this wrong (or at least backwards) in that it
2215   // calculates the extra locals based on itself. Not what the callee does
2216   // to it. So it ignores last_frame_adjust value. Seems suspicious as far
2217   // as getting sender_sp correct.
2218 
2219   int extra_locals_size = callee_locals_size - callee_param_size;
2220   int monitor_size = (sizeof(BasicObjectLock) * moncount) / wordSize;
2221   int full_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2222   int short_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2223   int frame_words = is_top_frame ? full_frame_words : short_frame_words;
2224 
2225 
2226   /*
2227     if we actually have a frame to layout we must now fill in all the pieces. This means both
2228     the interpreterState and the registers.
2229   */
2230   if (interpreter_frame != NULL) {
2231 
2232     // MUCHO HACK
2233 
2234     intptr_t* frame_bottom = interpreter_frame->sp() - (full_frame_words - frame_words);
2235     // 'interpreter_frame->sp()' is unbiased while 'frame_bottom' must be a biased value in 64bit mode.
2236     assert(((intptr_t)frame_bottom & 0xf) == 0, "SP biased in layout_activation");
2237     frame_bottom = (intptr_t*)((intptr_t)frame_bottom - STACK_BIAS);
2238 
2239     /* Now fillin the interpreterState object */
2240 
2241     interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() -  sizeof(BytecodeInterpreter));
2242 
2243 
2244     intptr_t* locals;
2245 
2246     // Calculate the postion of locals[0]. This is painful because of
2247     // stack alignment (same as ia64). The problem is that we can
2248     // not compute the location of locals from fp(). fp() will account
2249     // for the extra locals but it also accounts for aligning the stack
2250     // and we can't determine if the locals[0] was misaligned but max_locals
2251     // was enough to have the
2252     // calculate postion of locals. fp already accounts for extra locals.
2253     // +2 for the static long no_params() issue.
2254 
2255     if (caller->is_interpreted_frame()) {
2256       // locals must agree with the caller because it will be used to set the
2257       // caller's tos when we return.
2258       interpreterState prev  = caller->get_interpreterState();
2259       // stack() is prepushed.
2260       locals = prev->stack() + method->size_of_parameters();
2261     } else {
2262       // Lay out locals block in the caller adjacent to the register window save area.
2263       //
2264       // Compiled frames do not allocate a varargs area which is why this if
2265       // statement is needed.
2266       //
2267       intptr_t* fp = interpreter_frame->fp();
2268       int local_words = method->max_locals() * Interpreter::stackElementWords;
2269 
2270       if (caller->is_compiled_frame()) {
2271         locals = fp + frame::register_save_words + local_words - 1;
2272       } else {
2273         locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
2274       }
2275 
2276     }
2277     // END MUCHO HACK
2278 
2279     intptr_t* monitor_base = (intptr_t*) cur_state;
2280     intptr_t* stack_base =  monitor_base - monitor_size;
2281     /* +1 because stack is always prepushed */
2282     intptr_t* stack = stack_base - (tempcount + 1);
2283 
2284 
2285     BytecodeInterpreter::layout_interpreterState(cur_state,
2286                                           caller,
2287                                           interpreter_frame,
2288                                           method,
2289                                           locals,
2290                                           stack,
2291                                           stack_base,
2292                                           monitor_base,
2293                                           frame_bottom,
2294                                           is_top_frame);
2295 
2296     BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp());
2297 
2298   }
2299   return frame_words;
2300 }
2301 
2302 #endif // CC_INTERP