1 /*
   2  * Copyright (c) 2007, 2011, 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/methodDataOop.hpp"
  34 #include "oops/methodOop.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 #ifdef SHARK
  49 #include "shark/shark_globals.hpp"
  50 #endif
  51 
  52 #ifdef CC_INTERP
  53 
  54 // Routine exists to make tracebacks look decent in debugger
  55 // while "shadow" interpreter frames are on stack. It is also
  56 // used to distinguish interpreter frames.
  57 
  58 extern "C" void RecursiveInterpreterActivation(interpreterState istate) {
  59   ShouldNotReachHere();
  60 }
  61 
  62 bool CppInterpreter::contains(address pc) {
  63   return ( _code->contains(pc) ||
  64          ( pc == (CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset)));
  65 }
  66 
  67 #define STATE(field_name) Lstate, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
  68 #define __ _masm->
  69 
  70 Label frame_manager_entry;
  71 Label fast_accessor_slow_entry_path;  // fast accessor methods need to be able to jmp to unsynchronized
  72                                       // c++ interpreter entry point this holds that entry point label.
  73 
  74 static address unctrap_frame_manager_entry  = NULL;
  75 
  76 static address interpreter_return_address  = NULL;
  77 static address deopt_frame_manager_return_atos  = NULL;
  78 static address deopt_frame_manager_return_btos  = NULL;
  79 static address deopt_frame_manager_return_itos  = NULL;
  80 static address deopt_frame_manager_return_ltos  = NULL;
  81 static address deopt_frame_manager_return_ftos  = NULL;
  82 static address deopt_frame_manager_return_dtos  = NULL;
  83 static address deopt_frame_manager_return_vtos  = NULL;
  84 
  85 const Register prevState = G1_scratch;
  86 
  87 void InterpreterGenerator::save_native_result(void) {
  88   // result potentially in O0/O1: save it across calls
  89   __ stf(FloatRegisterImpl::D, F0, STATE(_native_fresult));
  90 #ifdef _LP64
  91   __ stx(O0, STATE(_native_lresult));
  92 #else
  93   __ std(O0, STATE(_native_lresult));
  94 #endif
  95 }
  96 
  97 void InterpreterGenerator::restore_native_result(void) {
  98 
  99   // Restore any method result value
 100   __ ldf(FloatRegisterImpl::D, STATE(_native_fresult), F0);
 101 #ifdef _LP64
 102   __ ldx(STATE(_native_lresult), O0);
 103 #else
 104   __ ldd(STATE(_native_lresult), O0);
 105 #endif
 106 }
 107 
 108 // A result handler converts/unboxes a native call result into
 109 // a java interpreter/compiler result. The current frame is an
 110 // interpreter frame. The activation frame unwind code must be
 111 // consistent with that of TemplateTable::_return(...). In the
 112 // case of native methods, the caller's SP was not modified.
 113 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) {
 114   address entry = __ pc();
 115   Register Itos_i  = Otos_i ->after_save();
 116   Register Itos_l  = Otos_l ->after_save();
 117   Register Itos_l1 = Otos_l1->after_save();
 118   Register Itos_l2 = Otos_l2->after_save();
 119   switch (type) {
 120     case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
 121     case T_CHAR   : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i);   break; // cannot use and3, 0xFFFF too big as immediate value!
 122     case T_BYTE   : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i);   break;
 123     case T_SHORT  : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i);   break;
 124     case T_LONG   :
 125 #ifndef _LP64
 126                     __ mov(O1, Itos_l2);  // move other half of long
 127 #endif              // ifdef or no ifdef, fall through to the T_INT case
 128     case T_INT    : __ mov(O0, Itos_i);                         break;
 129     case T_VOID   : /* nothing to do */                         break;
 130     case T_FLOAT  : assert(F0 == Ftos_f, "fix this code" );     break;
 131     case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" );     break;
 132     case T_OBJECT :
 133       __ ld_ptr(STATE(_oop_temp), Itos_i);
 134       __ verify_oop(Itos_i);
 135       break;
 136     default       : ShouldNotReachHere();
 137   }
 138   __ ret();                           // return from interpreter activation
 139   __ delayed()->restore(I5_savedSP, G0, SP);  // remove interpreter frame
 140   NOT_PRODUCT(__ emit_long(0);)       // marker for disassembly
 141   return entry;
 142 }
 143 
 144 // tosca based result to c++ interpreter stack based result.
 145 // Result goes to address in L1_scratch
 146 
 147 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) {
 148   // A result is in the native abi result register from a native method call.
 149   // We need to return this result to the interpreter by pushing the result on the interpreter's
 150   // stack. This is relatively simple the destination is in L1_scratch
 151   // i.e. L1_scratch is the first free element on the stack. If we "push" a return value we must
 152   // adjust L1_scratch
 153   address entry = __ pc();
 154   switch (type) {
 155     case T_BOOLEAN:
 156       // !0 => true; 0 => false
 157       __ subcc(G0, O0, G0);
 158       __ addc(G0, 0, O0);
 159       __ st(O0, L1_scratch, 0);
 160       __ sub(L1_scratch, wordSize, L1_scratch);
 161       break;
 162 
 163     // cannot use and3, 0xFFFF too big as immediate value!
 164     case T_CHAR   :
 165       __ sll(O0, 16, O0);
 166       __ srl(O0, 16, O0);
 167       __ st(O0, L1_scratch, 0);
 168       __ sub(L1_scratch, wordSize, L1_scratch);
 169       break;
 170 
 171     case T_BYTE   :
 172       __ sll(O0, 24, O0);
 173       __ sra(O0, 24, O0);
 174       __ st(O0, L1_scratch, 0);
 175       __ sub(L1_scratch, wordSize, L1_scratch);
 176       break;
 177 
 178     case T_SHORT  :
 179       __ sll(O0, 16, O0);
 180       __ sra(O0, 16, O0);
 181       __ st(O0, L1_scratch, 0);
 182       __ sub(L1_scratch, wordSize, L1_scratch);
 183       break;
 184     case T_LONG   :
 185 #ifndef _LP64
 186 #if defined(COMPILER2)
 187   // All return values are where we want them, except for Longs.  C2 returns
 188   // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
 189   // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
 190   // build even if we are returning from interpreted we just do a little
 191   // stupid shuffing.
 192   // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
 193   // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
 194   // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
 195       __ stx(G1, L1_scratch, -wordSize);
 196 #else
 197       // native result is in O0, O1
 198       __ st(O1, L1_scratch, 0);                      // Low order
 199       __ st(O0, L1_scratch, -wordSize);              // High order
 200 #endif /* COMPILER2 */
 201 #else
 202       __ stx(O0, L1_scratch, -wordSize);
 203 #endif
 204       __ sub(L1_scratch, 2*wordSize, L1_scratch);
 205       break;
 206 
 207     case T_INT    :
 208       __ st(O0, L1_scratch, 0);
 209       __ sub(L1_scratch, wordSize, L1_scratch);
 210       break;
 211 
 212     case T_VOID   : /* nothing to do */
 213       break;
 214 
 215     case T_FLOAT  :
 216       __ stf(FloatRegisterImpl::S, F0, L1_scratch, 0);
 217       __ sub(L1_scratch, wordSize, L1_scratch);
 218       break;
 219 
 220     case T_DOUBLE :
 221       // Every stack slot is aligned on 64 bit, However is this
 222       // the correct stack slot on 64bit?? QQQ
 223       __ stf(FloatRegisterImpl::D, F0, L1_scratch, -wordSize);
 224       __ sub(L1_scratch, 2*wordSize, L1_scratch);
 225       break;
 226     case T_OBJECT :
 227       __ verify_oop(O0);
 228       __ st_ptr(O0, L1_scratch, 0);
 229       __ sub(L1_scratch, wordSize, L1_scratch);
 230       break;
 231     default       : ShouldNotReachHere();
 232   }
 233   __ retl();                          // return from interpreter activation
 234   __ delayed()->nop();                // schedule this better
 235   NOT_PRODUCT(__ emit_long(0);)       // marker for disassembly
 236   return entry;
 237 }
 238 
 239 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) {
 240   // A result is in the java expression stack of the interpreted method that has just
 241   // returned. Place this result on the java expression stack of the caller.
 242   //
 243   // The current interpreter activation in Lstate is for the method just returning its
 244   // result. So we know that the result of this method is on the top of the current
 245   // execution stack (which is pre-pushed) and will be return to the top of the caller
 246   // stack. The top of the callers stack is the bottom of the locals of the current
 247   // activation.
 248   // Because of the way activation are managed by the frame manager the value of esp is
 249   // below both the stack top of the current activation and naturally the stack top
 250   // of the calling activation. This enable this routine to leave the return address
 251   // to the frame manager on the stack and do a vanilla return.
 252   //
 253   // On entry: O0 - points to source (callee stack top)
 254   //           O1 - points to destination (caller stack top [i.e. free location])
 255   // destroys O2, O3
 256   //
 257 
 258   address entry = __ pc();
 259   switch (type) {
 260     case T_VOID:  break;
 261       break;
 262     case T_FLOAT  :
 263     case T_BOOLEAN:
 264     case T_CHAR   :
 265     case T_BYTE   :
 266     case T_SHORT  :
 267     case T_INT    :
 268       // 1 word result
 269       __ ld(O0, 0, O2);
 270       __ st(O2, O1, 0);
 271       __ sub(O1, wordSize, O1);
 272       break;
 273     case T_DOUBLE  :
 274     case T_LONG    :
 275       // return top two words on current expression stack to caller's expression stack
 276       // The caller's expression stack is adjacent to the current frame manager's intepretState
 277       // except we allocated one extra word for this intepretState so we won't overwrite it
 278       // when we return a two word result.
 279 #ifdef _LP64
 280       __ ld_ptr(O0, 0, O2);
 281       __ st_ptr(O2, O1, -wordSize);
 282 #else
 283       __ ld(O0, 0, O2);
 284       __ ld(O0, wordSize, O3);
 285       __ st(O3, O1, 0);
 286       __ st(O2, O1, -wordSize);
 287 #endif
 288       __ sub(O1, 2*wordSize, O1);
 289       break;
 290     case T_OBJECT :
 291       __ ld_ptr(O0, 0, O2);
 292       __ verify_oop(O2);                                               // verify it
 293       __ st_ptr(O2, O1, 0);
 294       __ sub(O1, wordSize, O1);
 295       break;
 296     default       : ShouldNotReachHere();
 297   }
 298   __ retl();
 299   __ delayed()->nop(); // QQ schedule this better
 300   return entry;
 301 }
 302 
 303 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) {
 304   // A result is in the java expression stack of the interpreted method that has just
 305   // returned. Place this result in the native abi that the caller expects.
 306   // We are in a new frame registers we set must be in caller (i.e. callstub) frame.
 307   //
 308   // Similar to generate_stack_to_stack_converter above. Called at a similar time from the
 309   // frame manager execept in this situation the caller is native code (c1/c2/call_stub)
 310   // and so rather than return result onto caller's java expression stack we return the
 311   // result in the expected location based on the native abi.
 312   // On entry: O0 - source (stack top)
 313   // On exit result in expected output register
 314   // QQQ schedule this better
 315 
 316   address entry = __ pc();
 317   switch (type) {
 318     case T_VOID:  break;
 319       break;
 320     case T_FLOAT  :
 321       __ ldf(FloatRegisterImpl::S, O0, 0, F0);
 322       break;
 323     case T_BOOLEAN:
 324     case T_CHAR   :
 325     case T_BYTE   :
 326     case T_SHORT  :
 327     case T_INT    :
 328       // 1 word result
 329       __ ld(O0, 0, O0->after_save());
 330       break;
 331     case T_DOUBLE  :
 332       __ ldf(FloatRegisterImpl::D, O0, 0, F0);
 333       break;
 334     case T_LONG    :
 335       // return top two words on current expression stack to caller's expression stack
 336       // The caller's expression stack is adjacent to the current frame manager's interpretState
 337       // except we allocated one extra word for this intepretState so we won't overwrite it
 338       // when we return a two word result.
 339 #ifdef _LP64
 340       __ ld_ptr(O0, 0, O0->after_save());
 341 #else
 342       __ ld(O0, wordSize, O1->after_save());
 343       __ ld(O0, 0, O0->after_save());
 344 #endif
 345 #if defined(COMPILER2) && !defined(_LP64)
 346       // C2 expects long results in G1 we can't tell if we're returning to interpreted
 347       // or compiled so just be safe use G1 and O0/O1
 348 
 349       // Shift bits into high (msb) of G1
 350       __ sllx(Otos_l1->after_save(), 32, G1);
 351       // Zero extend low bits
 352       __ srl (Otos_l2->after_save(), 0, Otos_l2->after_save());
 353       __ or3 (Otos_l2->after_save(), G1, G1);
 354 #endif /* COMPILER2 */
 355       break;
 356     case T_OBJECT :
 357       __ ld_ptr(O0, 0, O0->after_save());
 358       __ verify_oop(O0->after_save());                                               // verify it
 359       break;
 360     default       : ShouldNotReachHere();
 361   }
 362   __ retl();
 363   __ delayed()->nop();
 364   return entry;
 365 }
 366 
 367 address CppInterpreter::return_entry(TosState state, int length) {
 368   // make it look good in the debugger
 369   return CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset;
 370 }
 371 
 372 address CppInterpreter::deopt_entry(TosState state, int length) {
 373   address ret = NULL;
 374   if (length != 0) {
 375     switch (state) {
 376       case atos: ret = deopt_frame_manager_return_atos; break;
 377       case btos: ret = deopt_frame_manager_return_btos; break;
 378       case ctos:
 379       case stos:
 380       case itos: ret = deopt_frame_manager_return_itos; break;
 381       case ltos: ret = deopt_frame_manager_return_ltos; break;
 382       case ftos: ret = deopt_frame_manager_return_ftos; break;
 383       case dtos: ret = deopt_frame_manager_return_dtos; break;
 384       case vtos: ret = deopt_frame_manager_return_vtos; break;
 385     }
 386   } else {
 387     ret = unctrap_frame_manager_entry;  // re-execute the bytecode ( e.g. uncommon trap)
 388   }
 389   assert(ret != NULL, "Not initialized");
 390   return ret;
 391 }
 392 
 393 //
 394 // Helpers for commoning out cases in the various type of method entries.
 395 //
 396 
 397 // increment invocation count & check for overflow
 398 //
 399 // Note: checking for negative value instead of overflow
 400 //       so we have a 'sticky' overflow test
 401 //
 402 // Lmethod: method
 403 // ??: invocation counter
 404 //
 405 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
 406   // Update standard invocation counters
 407   __ increment_invocation_counter(O0, G3_scratch);
 408   if (ProfileInterpreter) {  // %%% Merge this into methodDataOop
 409     __ ld_ptr(STATE(_method), G3_scratch);
 410     Address interpreter_invocation_counter(G3_scratch, 0, in_bytes(methodOopDesc::interpreter_invocation_counter_offset()));
 411     __ ld(interpreter_invocation_counter, G3_scratch);
 412     __ inc(G3_scratch);
 413     __ st(G3_scratch, interpreter_invocation_counter);
 414   }
 415 
 416   Address invocation_limit(G3_scratch, (address)&InvocationCounter::InterpreterInvocationLimit);
 417   __ sethi(invocation_limit);
 418   __ ld(invocation_limit, G3_scratch);
 419   __ cmp(O0, G3_scratch);
 420   __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
 421   __ delayed()->nop();
 422 
 423 }
 424 
 425 address InterpreterGenerator::generate_empty_entry(void) {
 426 
 427   // A method that does nothing but return...
 428 
 429   address entry = __ pc();
 430   Label slow_path;
 431 
 432   __ verify_oop(G5_method);
 433 
 434   // do nothing for empty methods (do not even increment invocation counter)
 435   if ( UseFastEmptyMethods) {
 436     // If we need a safepoint check, generate full interpreter entry.
 437     Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
 438     __ load_contents(sync_state, G3_scratch);
 439     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 440     __ br(Assembler::notEqual, false, Assembler::pn, frame_manager_entry);
 441     __ delayed()->nop();
 442 
 443     // Code: _return
 444     __ retl();
 445     __ delayed()->mov(O5_savedSP, SP);
 446     return entry;
 447   }
 448   return NULL;
 449 }
 450 
 451 // Call an accessor method (assuming it is resolved, otherwise drop into
 452 // vanilla (slow path) entry
 453 
 454 // Generates code to elide accessor methods
 455 // Uses G3_scratch and G1_scratch as scratch
 456 address InterpreterGenerator::generate_accessor_entry(void) {
 457 
 458   // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
 459   // parameter size = 1
 460   // Note: We can only use this code if the getfield has been resolved
 461   //       and if we don't have a null-pointer exception => check for
 462   //       these conditions first and use slow path if necessary.
 463   address entry = __ pc();
 464   Label slow_path;
 465 
 466   if ( UseFastAccessorMethods) {
 467     // Check if we need to reach a safepoint and generate full interpreter
 468     // frame if so.
 469     Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
 470     __ load_contents(sync_state, G3_scratch);
 471     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 472     __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
 473     __ delayed()->nop();
 474 
 475     // Check if local 0 != NULL
 476     __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
 477     __ tst(Otos_i);  // check if local 0 == NULL and go the slow path
 478     __ brx(Assembler::zero, false, Assembler::pn, slow_path);
 479     __ delayed()->nop();
 480 
 481 
 482     // read first instruction word and extract bytecode @ 1 and index @ 2
 483     // get first 4 bytes of the bytecodes (big endian!)
 484     __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::const_offset())), G1_scratch);
 485     __ ld(Address(G1_scratch, 0, in_bytes(constMethodOopDesc::codes_offset())), G1_scratch);
 486 
 487     // move index @ 2 far left then to the right most two bytes.
 488     __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
 489     __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
 490                       ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
 491 
 492     // get constant pool cache
 493     __ ld_ptr(G5_method, in_bytes(methodOopDesc::constants_offset()), G3_scratch);
 494     __ ld_ptr(G3_scratch, constantPoolOopDesc::cache_offset_in_bytes(), G3_scratch);
 495 
 496     // get specific constant pool cache entry
 497     __ add(G3_scratch, G1_scratch, G3_scratch);
 498 
 499     // Check the constant Pool cache entry to see if it has been resolved.
 500     // If not, need the slow path.
 501     ByteSize cp_base_offset = constantPoolCacheOopDesc::base_offset();
 502     __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::indices_offset()), G1_scratch);
 503     __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
 504     __ and3(G1_scratch, 0xFF, G1_scratch);
 505     __ cmp(G1_scratch, Bytecodes::_getfield);
 506     __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
 507     __ delayed()->nop();
 508 
 509     // Get the type and return field offset from the constant pool cache
 510     __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset()), G1_scratch);
 511     __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset()), G3_scratch);
 512 
 513     Label xreturn_path;
 514     // Need to differentiate between igetfield, agetfield, bgetfield etc.
 515     // because they are different sizes.
 516     // Get the type from the constant pool cache
 517     __ srl(G1_scratch, ConstantPoolCacheEntry::tosBits, G1_scratch);
 518     // Make sure we don't need to mask G1_scratch for tosBits after the above shift
 519     ConstantPoolCacheEntry::verify_tosBits();
 520     __ cmp(G1_scratch, atos );
 521     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 522     __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
 523     __ cmp(G1_scratch, itos);
 524     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 525     __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
 526     __ cmp(G1_scratch, stos);
 527     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 528     __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
 529     __ cmp(G1_scratch, ctos);
 530     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 531     __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
 532 #ifdef ASSERT
 533     __ cmp(G1_scratch, btos);
 534     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 535     __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
 536     __ should_not_reach_here();
 537 #endif
 538     __ ldsb(Otos_i, G3_scratch, Otos_i);
 539     __ bind(xreturn_path);
 540 
 541     // _ireturn/_areturn
 542     __ retl();                      // return from leaf routine
 543     __ delayed()->mov(O5_savedSP, SP);
 544 
 545     // Generate regular method entry
 546     __ bind(slow_path);
 547     __ ba(false, fast_accessor_slow_entry_path);
 548     __ delayed()->nop();
 549     return entry;
 550   }
 551   return NULL;
 552 }
 553 
 554 address InterpreterGenerator::generate_Reference_get_entry(void) {
 555 #ifndef SERIALGC
 556   if (UseG1GC) {
 557     // We need to generate have a routine that generates code to:
 558     //   * load the value in the referent field
 559     //   * passes that value to the pre-barrier.
 560     //
 561     // In the case of G1 this will record the value of the
 562     // referent in an SATB buffer if marking is active.
 563     // This will cause concurrent marking to mark the referent
 564     // field as live.
 565     Unimplemented();
 566   }
 567 #endif // SERIALGC
 568 
 569   // If G1 is not enabled then attempt to go through the accessor entry point
 570   // Reference.get is an accessor
 571   return generate_accessor_entry();
 572 }
 573 
 574 //
 575 // Interpreter stub for calling a native method. (C++ interpreter)
 576 // This sets up a somewhat different looking stack for calling the native method
 577 // than the typical interpreter frame setup.
 578 //
 579 
 580 address InterpreterGenerator::generate_native_entry(bool synchronized) {
 581   address entry = __ pc();
 582 
 583   // the following temporary registers are used during frame creation
 584   const Register Gtmp1 = G3_scratch ;
 585   const Register Gtmp2 = G1_scratch;
 586   const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
 587 
 588   bool inc_counter  = UseCompiler || CountCompiledCalls;
 589 
 590   // make sure registers are different!
 591   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
 592 
 593   const Address access_flags      (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
 594 
 595   Label Lentry;
 596   __ bind(Lentry);
 597 
 598   __ verify_oop(G5_method);
 599 
 600   const Register Glocals_size = G3;
 601   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
 602 
 603   // make sure method is native & not abstract
 604   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
 605 #ifdef ASSERT
 606   __ ld(access_flags, Gtmp1);
 607   {
 608     Label L;
 609     __ btst(JVM_ACC_NATIVE, Gtmp1);
 610     __ br(Assembler::notZero, false, Assembler::pt, L);
 611     __ delayed()->nop();
 612     __ stop("tried to execute non-native method as native");
 613     __ bind(L);
 614   }
 615   { Label L;
 616     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
 617     __ br(Assembler::zero, false, Assembler::pt, L);
 618     __ delayed()->nop();
 619     __ stop("tried to execute abstract method as non-abstract");
 620     __ bind(L);
 621   }
 622 #endif // ASSERT
 623 
 624   __ lduh(size_of_parameters, Gtmp1);
 625   __ sll(Gtmp1, LogBytesPerWord, Gtmp2);       // parameter size in bytes
 626   __ add(Gargs, Gtmp2, Gargs);                 // points to first local + BytesPerWord
 627   // NEW
 628   __ add(Gargs, -wordSize, Gargs);             // points to first local[0]
 629   // generate the code to allocate the interpreter stack frame
 630   // NEW FRAME ALLOCATED HERE
 631   // save callers original sp
 632   // __ mov(SP, I5_savedSP->after_restore());
 633 
 634   generate_compute_interpreter_state(Lstate, G0, true);
 635 
 636   // At this point Lstate points to new interpreter state
 637   //
 638 
 639   const Address do_not_unlock_if_synchronized(G2_thread, 0,
 640       in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
 641   // Since at this point in the method invocation the exception handler
 642   // would try to exit the monitor of synchronized methods which hasn't
 643   // been entered yet, we set the thread local variable
 644   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
 645   // runtime, exception handling i.e. unlock_if_synchronized_method will
 646   // check this thread local flag.
 647   // This flag has two effects, one is to force an unwind in the topmost
 648   // interpreter frame and not perform an unlock while doing so.
 649 
 650   __ movbool(true, G3_scratch);
 651   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
 652 
 653 
 654   // increment invocation counter and check for overflow
 655   //
 656   // Note: checking for negative value instead of overflow
 657   //       so we have a 'sticky' overflow test (may be of
 658   //       importance as soon as we have true MT/MP)
 659   Label invocation_counter_overflow;
 660   if (inc_counter) {
 661     generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
 662   }
 663   Label Lcontinue;
 664   __ bind(Lcontinue);
 665 
 666   bang_stack_shadow_pages(true);
 667   // reset the _do_not_unlock_if_synchronized flag
 668   __ stbool(G0, do_not_unlock_if_synchronized);
 669 
 670   // check for synchronized methods
 671   // Must happen AFTER invocation_counter check, so method is not locked
 672   // if counter overflows.
 673 
 674   if (synchronized) {
 675     lock_method();
 676     // Don't see how G2_thread is preserved here...
 677     // __ verify_thread(); QQQ destroys L0,L1 can't use
 678   } else {
 679 #ifdef ASSERT
 680     { Label ok;
 681       __ ld_ptr(STATE(_method), G5_method);
 682       __ ld(access_flags, O0);
 683       __ btst(JVM_ACC_SYNCHRONIZED, O0);
 684       __ br( Assembler::zero, false, Assembler::pt, ok);
 685       __ delayed()->nop();
 686       __ stop("method needs synchronization");
 687       __ bind(ok);
 688     }
 689 #endif // ASSERT
 690   }
 691 
 692   // start execution
 693 
 694 //   __ verify_thread(); kills L1,L2 can't  use at the moment
 695 
 696   // jvmti/jvmpi support
 697   __ notify_method_entry();
 698 
 699   // native call
 700 
 701   // (note that O0 is never an oop--at most it is a handle)
 702   // It is important not to smash any handles created by this call,
 703   // until any oop handle in O0 is dereferenced.
 704 
 705   // (note that the space for outgoing params is preallocated)
 706 
 707   // get signature handler
 708 
 709   Label pending_exception_present;
 710 
 711   { Label L;
 712     __ ld_ptr(STATE(_method), G5_method);
 713     __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::signature_handler_offset())), G3_scratch);
 714     __ tst(G3_scratch);
 715     __ brx(Assembler::notZero, false, Assembler::pt, L);
 716     __ delayed()->nop();
 717     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), G5_method, false);
 718     __ ld_ptr(STATE(_method), G5_method);
 719 
 720     Address exception_addr(G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
 721     __ ld_ptr(exception_addr, G3_scratch);
 722     __ br_notnull(G3_scratch, false, Assembler::pn, pending_exception_present);
 723     __ delayed()->nop();
 724     __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::signature_handler_offset())), G3_scratch);
 725     __ bind(L);
 726   }
 727 
 728   // Push a new frame so that the args will really be stored in
 729   // Copy a few locals across so the new frame has the variables
 730   // we need but these values will be dead at the jni call and
 731   // therefore not gc volatile like the values in the current
 732   // frame (Lstate in particular)
 733 
 734   // Flush the state pointer to the register save area
 735   // Which is the only register we need for a stack walk.
 736   __ st_ptr(Lstate, SP, (Lstate->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
 737 
 738   __ mov(Lstate, O1);         // Need to pass the state pointer across the frame
 739 
 740   // Calculate current frame size
 741   __ sub(SP, FP, O3);         // Calculate negative of current frame size
 742   __ save(SP, O3, SP);        // Allocate an identical sized frame
 743 
 744   __ mov(I1, Lstate);          // In the "natural" register.
 745 
 746   // Note I7 has leftover trash. Slow signature handler will fill it in
 747   // should we get there. Normal jni call will set reasonable last_Java_pc
 748   // below (and fix I7 so the stack trace doesn't have a meaningless frame
 749   // in it).
 750 
 751 
 752   // call signature handler
 753   __ ld_ptr(STATE(_method), Lmethod);
 754   __ ld_ptr(STATE(_locals), Llocals);
 755 
 756   __ callr(G3_scratch, 0);
 757   __ delayed()->nop();
 758   __ ld_ptr(STATE(_thread), G2_thread);        // restore thread (shouldn't be needed)
 759 
 760   { Label not_static;
 761 
 762     __ ld_ptr(STATE(_method), G5_method);
 763     __ ld(access_flags, O0);
 764     __ btst(JVM_ACC_STATIC, O0);
 765     __ br( Assembler::zero, false, Assembler::pt, not_static);
 766     __ delayed()->
 767       // get native function entry point(O0 is a good temp until the very end)
 768        ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::native_function_offset())), O0);
 769     // for static methods insert the mirror argument
 770     const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
 771 
 772     __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc:: constants_offset())), O1);
 773     __ ld_ptr(Address(O1, 0, constantPoolOopDesc::pool_holder_offset_in_bytes()), O1);
 774     __ ld_ptr(O1, mirror_offset, O1);
 775     // where the mirror handle body is allocated:
 776 #ifdef ASSERT
 777     if (!PrintSignatureHandlers)  // do not dirty the output with this
 778     { Label L;
 779       __ tst(O1);
 780       __ brx(Assembler::notZero, false, Assembler::pt, L);
 781       __ delayed()->nop();
 782       __ stop("mirror is missing");
 783       __ bind(L);
 784     }
 785 #endif // ASSERT
 786     __ st_ptr(O1, STATE(_oop_temp));
 787     __ add(STATE(_oop_temp), O1);            // this is really an LEA not an add
 788     __ bind(not_static);
 789   }
 790 
 791   // At this point, arguments have been copied off of stack into
 792   // their JNI positions, which are O1..O5 and SP[68..].
 793   // Oops are boxed in-place on the stack, with handles copied to arguments.
 794   // The result handler is in Lscratch.  O0 will shortly hold the JNIEnv*.
 795 
 796 #ifdef ASSERT
 797   { Label L;
 798     __ tst(O0);
 799     __ brx(Assembler::notZero, false, Assembler::pt, L);
 800     __ delayed()->nop();
 801     __ stop("native entry point is missing");
 802     __ bind(L);
 803   }
 804 #endif // ASSERT
 805 
 806   //
 807   // setup the java frame anchor
 808   //
 809   // The scavenge function only needs to know that the PC of this frame is
 810   // in the interpreter method entry code, it doesn't need to know the exact
 811   // PC and hence we can use O7 which points to the return address from the
 812   // previous call in the code stream (signature handler function)
 813   //
 814   // The other trick is we set last_Java_sp to FP instead of the usual SP because
 815   // we have pushed the extra frame in order to protect the volatile register(s)
 816   // in that frame when we return from the jni call
 817   //
 818 
 819 
 820   __ set_last_Java_frame(FP, O7);
 821   __ mov(O7, I7);  // make dummy interpreter frame look like one above,
 822                    // not meaningless information that'll confuse me.
 823 
 824   // flush the windows now. We don't care about the current (protection) frame
 825   // only the outer frames
 826 
 827   __ flush_windows();
 828 
 829   // mark windows as flushed
 830   Address flags(G2_thread,
 831                 0,
 832                 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
 833   __ set(JavaFrameAnchor::flushed, G3_scratch);
 834   __ st(G3_scratch, flags);
 835 
 836   // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
 837 
 838   Address thread_state(G2_thread, 0, in_bytes(JavaThread::thread_state_offset()));
 839 #ifdef ASSERT
 840   { Label L;
 841     __ ld(thread_state, G3_scratch);
 842     __ cmp(G3_scratch, _thread_in_Java);
 843     __ br(Assembler::equal, false, Assembler::pt, L);
 844     __ delayed()->nop();
 845     __ stop("Wrong thread state in native stub");
 846     __ bind(L);
 847   }
 848 #endif // ASSERT
 849   __ set(_thread_in_native, G3_scratch);
 850   __ st(G3_scratch, thread_state);
 851 
 852   // Call the jni method, using the delay slot to set the JNIEnv* argument.
 853   __ callr(O0, 0);
 854   __ delayed()->
 855      add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
 856   __ ld_ptr(STATE(_thread), G2_thread);  // restore thread
 857 
 858   // must we block?
 859 
 860   // Block, if necessary, before resuming in _thread_in_Java state.
 861   // In order for GC to work, don't clear the last_Java_sp until after blocking.
 862   { Label no_block;
 863     Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
 864 
 865     // Switch thread to "native transition" state before reading the synchronization state.
 866     // This additional state is necessary because reading and testing the synchronization
 867     // state is not atomic w.r.t. GC, as this scenario demonstrates:
 868     //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
 869     //     VM thread changes sync state to synchronizing and suspends threads for GC.
 870     //     Thread A is resumed to finish this native method, but doesn't block here since it
 871     //     didn't see any synchronization is progress, and escapes.
 872     __ set(_thread_in_native_trans, G3_scratch);
 873     __ st(G3_scratch, thread_state);
 874     if(os::is_MP()) {
 875       // Write serialization page so VM thread can do a pseudo remote membar.
 876       // We use the current thread pointer to calculate a thread specific
 877       // offset to write to within the page. This minimizes bus traffic
 878       // due to cache line collision.
 879       __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
 880     }
 881     __ load_contents(sync_state, G3_scratch);
 882     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 883 
 884 
 885     Label L;
 886     Address suspend_state(G2_thread, 0, in_bytes(JavaThread::suspend_flags_offset()));
 887     __ br(Assembler::notEqual, false, Assembler::pn, L);
 888     __ delayed()->
 889       ld(suspend_state, G3_scratch);
 890     __ cmp(G3_scratch, 0);
 891     __ br(Assembler::equal, false, Assembler::pt, no_block);
 892     __ delayed()->nop();
 893     __ bind(L);
 894 
 895     // Block.  Save any potential method result value before the operation and
 896     // use a leaf call to leave the last_Java_frame setup undisturbed.
 897     save_native_result();
 898     __ call_VM_leaf(noreg,
 899                     CAST_FROM_FN_PTR(address, JavaThread::check_safepoint_and_suspend_for_native_trans),
 900                     G2_thread);
 901     __ ld_ptr(STATE(_thread), G2_thread);  // restore thread
 902     // Restore any method result value
 903     restore_native_result();
 904     __ bind(no_block);
 905   }
 906 
 907   // Clear the frame anchor now
 908 
 909   __ reset_last_Java_frame();
 910 
 911   // Move the result handler address
 912   __ mov(Lscratch, G3_scratch);
 913   // return possible result to the outer frame
 914 #ifndef __LP64
 915   __ mov(O0, I0);
 916   __ restore(O1, G0, O1);
 917 #else
 918   __ restore(O0, G0, O0);
 919 #endif /* __LP64 */
 920 
 921   // Move result handler to expected register
 922   __ mov(G3_scratch, Lscratch);
 923 
 924 
 925   // thread state is thread_in_native_trans. Any safepoint blocking has
 926   // happened in the trampoline we are ready to switch to thread_in_Java.
 927 
 928   __ set(_thread_in_Java, G3_scratch);
 929   __ st(G3_scratch, thread_state);
 930 
 931   // If we have an oop result store it where it will be safe for any further gc
 932   // until we return now that we've released the handle it might be protected by
 933 
 934   {
 935     Label no_oop, store_result;
 936 
 937     __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
 938     __ cmp(G3_scratch, Lscratch);
 939     __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
 940     __ delayed()->nop();
 941     __ addcc(G0, O0, O0);
 942     __ brx(Assembler::notZero, true, Assembler::pt, store_result);     // if result is not NULL:
 943     __ delayed()->ld_ptr(O0, 0, O0);                                   // unbox it
 944     __ mov(G0, O0);
 945 
 946     __ bind(store_result);
 947     // Store it where gc will look for it and result handler expects it.
 948     __ st_ptr(O0, STATE(_oop_temp));
 949 
 950     __ bind(no_oop);
 951 
 952   }
 953 
 954   // reset handle block
 955   __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch);
 956   __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
 957 
 958 
 959   // handle exceptions (exception handling will handle unlocking!)
 960   { Label L;
 961     Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
 962 
 963     __ ld_ptr(exception_addr, Gtemp);
 964     __ tst(Gtemp);
 965     __ brx(Assembler::equal, false, Assembler::pt, L);
 966     __ delayed()->nop();
 967     __ bind(pending_exception_present);
 968     // With c++ interpreter we just leave it pending caller will do the correct thing. However...
 969     // Like x86 we ignore the result of the native call and leave the method locked. This
 970     // seems wrong to leave things locked.
 971 
 972     __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
 973     __ delayed()->restore(I5_savedSP, G0, SP);  // remove interpreter frame
 974 
 975     __ bind(L);
 976   }
 977 
 978   // jvmdi/jvmpi support (preserves thread register)
 979   __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
 980 
 981   if (synchronized) {
 982     // save and restore any potential method result value around the unlocking operation
 983     save_native_result();
 984 
 985     const int entry_size            = frame::interpreter_frame_monitor_size() * wordSize;
 986     // Get the initial monitor we allocated
 987     __ sub(Lstate, entry_size, O1);                        // initial monitor
 988     __ unlock_object(O1);
 989     restore_native_result();
 990   }
 991 
 992 #if defined(COMPILER2) && !defined(_LP64)
 993 
 994   // C2 expects long results in G1 we can't tell if we're returning to interpreted
 995   // or compiled so just be safe.
 996 
 997   __ sllx(O0, 32, G1);          // Shift bits into high G1
 998   __ srl (O1, 0, O1);           // Zero extend O1
 999   __ or3 (O1, G1, G1);          // OR 64 bits into G1
1000 
1001 #endif /* COMPILER2 && !_LP64 */
1002 
1003 #ifdef ASSERT
1004   {
1005     Label ok;
1006     __ cmp(I5_savedSP, FP);
1007     __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
1008     __ delayed()->nop();
1009     __ stop("bad I5_savedSP value");
1010     __ should_not_reach_here();
1011     __ bind(ok);
1012   }
1013 #endif
1014   // Calls result handler which POPS FRAME
1015   if (TraceJumps) {
1016     // Move target to register that is recordable
1017     __ mov(Lscratch, G3_scratch);
1018     __ JMP(G3_scratch, 0);
1019   } else {
1020     __ jmp(Lscratch, 0);
1021   }
1022   __ delayed()->nop();
1023 
1024   if (inc_counter) {
1025     // handle invocation counter overflow
1026     __ bind(invocation_counter_overflow);
1027     generate_counter_overflow(Lcontinue);
1028   }
1029 
1030 
1031   return entry;
1032 }
1033 
1034 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state,
1035                                                               const Register prev_state,
1036                                                               bool native) {
1037 
1038   // On entry
1039   // G5_method - caller's method
1040   // Gargs - points to initial parameters (i.e. locals[0])
1041   // G2_thread - valid? (C1 only??)
1042   // "prev_state" - contains any previous frame manager state which we must save a link
1043   //
1044   // On return
1045   // "state" is a pointer to the newly allocated  state object. We must allocate and initialize
1046   // a new interpretState object and the method expression stack.
1047 
1048   assert_different_registers(state, prev_state);
1049   assert_different_registers(prev_state, G3_scratch);
1050   const Register Gtmp = G3_scratch;
1051   const Address constants         (G5_method, 0, in_bytes(methodOopDesc::constants_offset()));
1052   const Address access_flags      (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
1053   const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1054   const Address max_stack         (G5_method, 0, in_bytes(methodOopDesc::max_stack_offset()));
1055   const Address size_of_locals    (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1056 
1057   // slop factor is two extra slots on the expression stack so that
1058   // we always have room to store a result when returning from a call without parameters
1059   // that returns a result.
1060 
1061   const int slop_factor = 2*wordSize;
1062 
1063   const int fixed_size = ((sizeof(BytecodeInterpreter) + slop_factor) >> LogBytesPerWord) + // what is the slop factor?
1064                          //6815692//methodOopDesc::extra_stack_words() +  // extra push slots for MH adapters
1065                          frame::memory_parameter_word_sp_offset +  // register save area + param window
1066                          (native ?  frame::interpreter_frame_extra_outgoing_argument_words : 0); // JNI, class
1067 
1068   // XXX G5_method valid
1069 
1070   // Now compute new frame size
1071 
1072   if (native) {
1073     __ lduh( size_of_parameters, Gtmp );
1074     __ calc_mem_param_words(Gtmp, Gtmp);     // space for native call parameters passed on the stack in words
1075   } else {
1076     __ lduh(max_stack, Gtmp);                // Full size expression stack
1077   }
1078   __ add(Gtmp, fixed_size, Gtmp);           // plus the fixed portion
1079 
1080   __ neg(Gtmp);                               // negative space for stack/parameters in words
1081   __ and3(Gtmp, -WordsPerLong, Gtmp);        // make multiple of 2 (SP must be 2-word aligned)
1082   __ sll(Gtmp, LogBytesPerWord, Gtmp);       // negative space for frame in bytes
1083 
1084   // Need to do stack size check here before we fault on large frames
1085 
1086   Label stack_ok;
1087 
1088   const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages :
1089                                                                               (StackRedPages+StackYellowPages);
1090 
1091 
1092   __ ld_ptr(G2_thread, in_bytes(Thread::stack_base_offset()), O0);
1093   __ ld_ptr(G2_thread, in_bytes(Thread::stack_size_offset()), O1);
1094   // compute stack bottom
1095   __ sub(O0, O1, O0);
1096 
1097   // Avoid touching the guard pages
1098   // Also a fudge for frame size of BytecodeInterpreter::run
1099   // It varies from 1k->4k depending on build type
1100   const int fudge = 6 * K;
1101 
1102   __ set(fudge + (max_pages * os::vm_page_size()), O1);
1103 
1104   __ add(O0, O1, O0);
1105   __ sub(O0, Gtmp, O0);
1106   __ cmp(SP, O0);
1107   __ brx(Assembler::greaterUnsigned, false, Assembler::pt, stack_ok);
1108   __ delayed()->nop();
1109 
1110      // throw exception return address becomes throwing pc
1111 
1112   __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
1113   __ stop("never reached");
1114 
1115   __ bind(stack_ok);
1116 
1117   __ save(SP, Gtmp, SP);                      // setup new frame and register window
1118 
1119   // New window I7 call_stub or previous activation
1120   // O6 - register save area, BytecodeInterpreter just below it, args/locals just above that
1121   //
1122   __ sub(FP, sizeof(BytecodeInterpreter), state);        // Point to new Interpreter state
1123   __ add(state, STACK_BIAS, state );         // Account for 64bit bias
1124 
1125 #define XXX_STATE(field_name) state, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
1126 
1127   // Initialize a new Interpreter state
1128   // orig_sp - caller's original sp
1129   // G2_thread - thread
1130   // Gargs - &locals[0] (unbiased?)
1131   // G5_method - method
1132   // SP (biased) - accounts for full size java stack, BytecodeInterpreter object, register save area, and register parameter save window
1133 
1134 
1135   __ set(0xdead0004, O1);
1136 
1137 
1138   __ st_ptr(Gargs, XXX_STATE(_locals));
1139   __ st_ptr(G0, XXX_STATE(_oop_temp));
1140 
1141   __ st_ptr(state, XXX_STATE(_self_link));                // point to self
1142   __ st_ptr(prev_state->after_save(), XXX_STATE(_prev_link)); // Chain interpreter states
1143   __ st_ptr(G2_thread, XXX_STATE(_thread));               // Store javathread
1144 
1145   if (native) {
1146     __ st_ptr(G0, XXX_STATE(_bcp));
1147   } else {
1148     __ ld_ptr(G5_method, in_bytes(methodOopDesc::const_offset()), O2); // get constMethodOop
1149     __ add(O2, in_bytes(constMethodOopDesc::codes_offset()), O2);        // get bcp
1150     __ st_ptr(O2, XXX_STATE(_bcp));
1151   }
1152 
1153   __ st_ptr(G0, XXX_STATE(_mdx));
1154   __ st_ptr(G5_method, XXX_STATE(_method));
1155 
1156   __ set((int) BytecodeInterpreter::method_entry, O1);
1157   __ st(O1, XXX_STATE(_msg));
1158 
1159   __ ld_ptr(constants, O3);
1160   __ ld_ptr(O3, constantPoolOopDesc::cache_offset_in_bytes(), O2);
1161   __ st_ptr(O2, XXX_STATE(_constants));
1162 
1163   __ st_ptr(G0, XXX_STATE(_result._to_call._callee));
1164 
1165   // Monitor base is just start of BytecodeInterpreter object;
1166   __ mov(state, O2);
1167   __ st_ptr(O2, XXX_STATE(_monitor_base));
1168 
1169   // Do we need a monitor for synchonized method?
1170   {
1171     __ ld(access_flags, O1);
1172     Label done;
1173     Label got_obj;
1174     __ btst(JVM_ACC_SYNCHRONIZED, O1);
1175     __ br( Assembler::zero, false, Assembler::pt, done);
1176 
1177     const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
1178     __ delayed()->btst(JVM_ACC_STATIC, O1);
1179     __ ld_ptr(XXX_STATE(_locals), O1);
1180     __ br( Assembler::zero, true, Assembler::pt, got_obj);
1181     __ delayed()->ld_ptr(O1, 0, O1);                  // get receiver for not-static case
1182     __ ld_ptr(constants, O1);
1183     __ ld_ptr( O1, constantPoolOopDesc::pool_holder_offset_in_bytes(), O1);
1184     // lock the mirror, not the klassOop
1185     __ ld_ptr( O1, mirror_offset, O1);
1186 
1187     __ bind(got_obj);
1188 
1189   #ifdef ASSERT
1190     __ tst(O1);
1191     __ breakpoint_trap(Assembler::zero);
1192   #endif // ASSERT
1193 
1194     const int entry_size            = frame::interpreter_frame_monitor_size() * wordSize;
1195     __ sub(SP, entry_size, SP);                         // account for initial monitor
1196     __ sub(O2, entry_size, O2);                        // initial monitor
1197     __ st_ptr(O1, O2, BasicObjectLock::obj_offset_in_bytes()); // and allocate it for interpreter use
1198     __ bind(done);
1199   }
1200 
1201   // Remember initial frame bottom
1202 
1203   __ st_ptr(SP, XXX_STATE(_frame_bottom));
1204 
1205   __ st_ptr(O2, XXX_STATE(_stack_base));
1206 
1207   __ sub(O2, wordSize, O2);                    // prepush
1208   __ st_ptr(O2, XXX_STATE(_stack));                // PREPUSH
1209 
1210   __ lduh(max_stack, O3);                      // Full size expression stack
1211   guarantee(!EnableInvokeDynamic, "no support yet for java.lang.invoke.MethodHandle"); //6815692
1212   //6815692//if (EnableInvokeDynamic)
1213   //6815692//  __ inc(O3, methodOopDesc::extra_stack_entries());
1214   __ sll(O3, LogBytesPerWord, O3);
1215   __ sub(O2, O3, O3);
1216 //  __ sub(O3, wordSize, O3);                    // so prepush doesn't look out of bounds
1217   __ st_ptr(O3, XXX_STATE(_stack_limit));
1218 
1219   if (!native) {
1220     //
1221     // Code to initialize locals
1222     //
1223     Register init_value = noreg;    // will be G0 if we must clear locals
1224     // Now zero locals
1225     if (true /* zerolocals */ || ClearInterpreterLocals) {
1226       // explicitly initialize locals
1227       init_value = G0;
1228     } else {
1229     #ifdef ASSERT
1230       // initialize locals to a garbage pattern for better debugging
1231       init_value = O3;
1232       __ set( 0x0F0F0F0F, init_value );
1233     #endif // ASSERT
1234     }
1235     if (init_value != noreg) {
1236       Label clear_loop;
1237 
1238       // NOTE: If you change the frame layout, this code will need to
1239       // be updated!
1240       __ lduh( size_of_locals, O2 );
1241       __ lduh( size_of_parameters, O1 );
1242       __ sll( O2, LogBytesPerWord, O2);
1243       __ sll( O1, LogBytesPerWord, O1 );
1244       __ ld_ptr(XXX_STATE(_locals), L2_scratch);
1245       __ sub( L2_scratch, O2, O2 );
1246       __ sub( L2_scratch, O1, O1 );
1247 
1248       __ bind( clear_loop );
1249       __ inc( O2, wordSize );
1250 
1251       __ cmp( O2, O1 );
1252       __ br( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1253       __ delayed()->st_ptr( init_value, O2, 0 );
1254     }
1255   }
1256 }
1257 // Find preallocated  monitor and lock method (C++ interpreter)
1258 //
1259 void InterpreterGenerator::lock_method(void) {
1260 // Lock the current method.
1261 // Destroys registers L2_scratch, L3_scratch, O0
1262 //
1263 // Find everything relative to Lstate
1264 
1265 #ifdef ASSERT
1266   __ ld_ptr(STATE(_method), L2_scratch);
1267   __ ld(L2_scratch, in_bytes(methodOopDesc::access_flags_offset()), O0);
1268 
1269  { Label ok;
1270    __ btst(JVM_ACC_SYNCHRONIZED, O0);
1271    __ br( Assembler::notZero, false, Assembler::pt, ok);
1272    __ delayed()->nop();
1273    __ stop("method doesn't need synchronization");
1274    __ bind(ok);
1275   }
1276 #endif // ASSERT
1277 
1278   // monitor is already allocated at stack base
1279   // and the lockee is already present
1280   __ ld_ptr(STATE(_stack_base), L2_scratch);
1281   __ ld_ptr(L2_scratch, BasicObjectLock::obj_offset_in_bytes(), O0);   // get object
1282   __ lock_object(L2_scratch, O0);
1283 
1284 }
1285 
1286 //  Generate code for handling resuming a deopted method
1287 void CppInterpreterGenerator::generate_deopt_handling() {
1288 
1289   Label return_from_deopt_common;
1290 
1291   // deopt needs to jump to here to enter the interpreter (return a result)
1292   deopt_frame_manager_return_atos  = __ pc();
1293 
1294   // O0/O1 live
1295   __ ba(false, return_from_deopt_common);
1296   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_OBJECT), L3_scratch);    // Result stub address array index
1297 
1298 
1299   // deopt needs to jump to here to enter the interpreter (return a result)
1300   deopt_frame_manager_return_btos  = __ pc();
1301 
1302   // O0/O1 live
1303   __ ba(false, return_from_deopt_common);
1304   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_BOOLEAN), L3_scratch);    // Result stub address array index
1305 
1306   // deopt needs to jump to here to enter the interpreter (return a result)
1307   deopt_frame_manager_return_itos  = __ pc();
1308 
1309   // O0/O1 live
1310   __ ba(false, return_from_deopt_common);
1311   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_INT), L3_scratch);    // Result stub address array index
1312 
1313   // deopt needs to jump to here to enter the interpreter (return a result)
1314 
1315   deopt_frame_manager_return_ltos  = __ pc();
1316 #if !defined(_LP64) && defined(COMPILER2)
1317   // All return values are where we want them, except for Longs.  C2 returns
1318   // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
1319   // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
1320   // build even if we are returning from interpreted we just do a little
1321   // stupid shuffing.
1322   // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
1323   // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
1324   // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
1325 
1326   __ srl (G1, 0,O1);
1327   __ srlx(G1,32,O0);
1328 #endif /* !_LP64 && COMPILER2 */
1329   // O0/O1 live
1330   __ ba(false, return_from_deopt_common);
1331   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_LONG), L3_scratch);    // Result stub address array index
1332 
1333   // deopt needs to jump to here to enter the interpreter (return a result)
1334 
1335   deopt_frame_manager_return_ftos  = __ pc();
1336   // O0/O1 live
1337   __ ba(false, return_from_deopt_common);
1338   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_FLOAT), L3_scratch);    // Result stub address array index
1339 
1340   // deopt needs to jump to here to enter the interpreter (return a result)
1341   deopt_frame_manager_return_dtos  = __ pc();
1342 
1343   // O0/O1 live
1344   __ ba(false, return_from_deopt_common);
1345   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_DOUBLE), L3_scratch);    // Result stub address array index
1346 
1347   // deopt needs to jump to here to enter the interpreter (return a result)
1348   deopt_frame_manager_return_vtos  = __ pc();
1349 
1350   // O0/O1 live
1351   __ set(AbstractInterpreter::BasicType_as_index(T_VOID), L3_scratch);
1352 
1353   // Deopt return common
1354   // an index is present that lets us move any possible result being
1355   // return to the interpreter's stack
1356   //
1357   __ bind(return_from_deopt_common);
1358 
1359   // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1360   // stack is in the state that the  calling convention left it.
1361   // Copy the result from native abi result and place it on java expression stack.
1362 
1363   // Current interpreter state is present in Lstate
1364 
1365   // Get current pre-pushed top of interpreter stack
1366   // Any result (if any) is in native abi
1367   // result type index is in L3_scratch
1368 
1369   __ ld_ptr(STATE(_stack), L1_scratch);                                          // get top of java expr stack
1370 
1371   __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1372   __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1373   __ ld_ptr(L4_scratch, L3_scratch, Lscratch);                                       // get typed result converter address
1374   __ jmpl(Lscratch, G0, O7);                                         // and convert it
1375   __ delayed()->nop();
1376 
1377   // L1_scratch points to top of stack (prepushed)
1378   __ st_ptr(L1_scratch, STATE(_stack));
1379 }
1380 
1381 // Generate the code to handle a more_monitors message from the c++ interpreter
1382 void CppInterpreterGenerator::generate_more_monitors() {
1383 
1384   Label entry, loop;
1385   const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1386   // 1. compute new pointers                                // esp: old expression stack top
1387   __ delayed()->ld_ptr(STATE(_stack_base), L4_scratch);            // current expression stack bottom
1388   __ sub(L4_scratch, entry_size, L4_scratch);
1389   __ st_ptr(L4_scratch, STATE(_stack_base));
1390 
1391   __ sub(SP, entry_size, SP);                  // Grow stack
1392   __ st_ptr(SP, STATE(_frame_bottom));
1393 
1394   __ ld_ptr(STATE(_stack_limit), L2_scratch);
1395   __ sub(L2_scratch, entry_size, L2_scratch);
1396   __ st_ptr(L2_scratch, STATE(_stack_limit));
1397 
1398   __ ld_ptr(STATE(_stack), L1_scratch);                // Get current stack top
1399   __ sub(L1_scratch, entry_size, L1_scratch);
1400   __ st_ptr(L1_scratch, STATE(_stack));
1401   __ ba(false, entry);
1402   __ delayed()->add(L1_scratch, wordSize, L1_scratch);        // first real entry (undo prepush)
1403 
1404   // 2. move expression stack
1405 
1406   __ bind(loop);
1407   __ st_ptr(L3_scratch, Address(L1_scratch, 0));
1408   __ add(L1_scratch, wordSize, L1_scratch);
1409   __ bind(entry);
1410   __ cmp(L1_scratch, L4_scratch);
1411   __ br(Assembler::notEqual, false, Assembler::pt, loop);
1412   __ delayed()->ld_ptr(L1_scratch, entry_size, L3_scratch);
1413 
1414   // now zero the slot so we can find it.
1415   __ st_ptr(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes());
1416 
1417 }
1418 
1419 // Initial entry to C++ interpreter from the call_stub.
1420 // This entry point is called the frame manager since it handles the generation
1421 // of interpreter activation frames via requests directly from the vm (via call_stub)
1422 // and via requests from the interpreter. The requests from the call_stub happen
1423 // directly thru the entry point. Requests from the interpreter happen via returning
1424 // from the interpreter and examining the message the interpreter has returned to
1425 // the frame manager. The frame manager can take the following requests:
1426 
1427 // NO_REQUEST - error, should never happen.
1428 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and
1429 //                 allocate a new monitor.
1430 // CALL_METHOD - setup a new activation to call a new method. Very similar to what
1431 //               happens during entry during the entry via the call stub.
1432 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub.
1433 //
1434 // Arguments:
1435 //
1436 // ebx: methodOop
1437 // ecx: receiver - unused (retrieved from stack as needed)
1438 // esi: previous frame manager state (NULL from the call_stub/c1/c2)
1439 //
1440 //
1441 // Stack layout at entry
1442 //
1443 // [ return address     ] <--- esp
1444 // [ parameter n        ]
1445 //   ...
1446 // [ parameter 1        ]
1447 // [ expression stack   ]
1448 //
1449 //
1450 // We are free to blow any registers we like because the call_stub which brought us here
1451 // initially has preserved the callee save registers already.
1452 //
1453 //
1454 
1455 static address interpreter_frame_manager = NULL;
1456 
1457 #ifdef ASSERT
1458   #define VALIDATE_STATE(scratch, marker)                         \
1459   {                                                               \
1460     Label skip;                                                   \
1461     __ ld_ptr(STATE(_self_link), scratch);                        \
1462     __ cmp(Lstate, scratch);                                      \
1463     __ brx(Assembler::equal, false, Assembler::pt, skip);         \
1464     __ delayed()->nop();                                          \
1465     __ breakpoint_trap();                                         \
1466     __ emit_long(marker);                                         \
1467     __ bind(skip);                                                \
1468   }
1469 #else
1470   #define VALIDATE_STATE(scratch, marker)
1471 #endif /* ASSERT */
1472 
1473 void CppInterpreterGenerator::adjust_callers_stack(Register args) {
1474 //
1475 // Adjust caller's stack so that all the locals can be contiguous with
1476 // the parameters.
1477 // Worries about stack overflow make this a pain.
1478 //
1479 // Destroys args, G3_scratch, G3_scratch
1480 // In/Out O5_savedSP (sender's original SP)
1481 //
1482 //  assert_different_registers(state, prev_state);
1483   const Register Gtmp = G3_scratch;
1484   const Register tmp = O2;
1485   const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1486   const Address size_of_locals    (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1487 
1488   __ lduh(size_of_parameters, tmp);
1489   __ sll(tmp, LogBytesPerWord, Gtmp);       // parameter size in bytes
1490   __ add(args, Gtmp, Gargs);                // points to first local + BytesPerWord
1491   // NEW
1492   __ add(Gargs, -wordSize, Gargs);             // points to first local[0]
1493   // determine extra space for non-argument locals & adjust caller's SP
1494   // Gtmp1: parameter size in words
1495   __ lduh(size_of_locals, Gtmp);
1496   __ compute_extra_locals_size_in_bytes(tmp, Gtmp, Gtmp);
1497 
1498 #if 1
1499   // c2i adapters place the final interpreter argument in the register save area for O0/I0
1500   // the call_stub will place the final interpreter argument at
1501   // frame::memory_parameter_word_sp_offset. This is mostly not noticable for either asm
1502   // or c++ interpreter. However with the c++ interpreter when we do a recursive call
1503   // and try to make it look good in the debugger we will store the argument to
1504   // RecursiveInterpreterActivation in the register argument save area. Without allocating
1505   // extra space for the compiler this will overwrite locals in the local array of the
1506   // interpreter.
1507   // QQQ still needed with frameless adapters???
1508 
1509   const int c2i_adjust_words = frame::memory_parameter_word_sp_offset - frame::callee_register_argument_save_area_sp_offset;
1510 
1511   __ add(Gtmp, c2i_adjust_words*wordSize, Gtmp);
1512 #endif // 1
1513 
1514 
1515   __ sub(SP, Gtmp, SP);                      // just caller's frame for the additional space we need.
1516 }
1517 
1518 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1519 
1520   // G5_method: methodOop
1521   // G2_thread: thread (unused)
1522   // Gargs:   bottom of args (sender_sp)
1523   // O5: sender's sp
1524 
1525   // A single frame manager is plenty as we don't specialize for synchronized. We could and
1526   // the code is pretty much ready. Would need to change the test below and for good measure
1527   // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized
1528   // routines. Not clear this is worth it yet.
1529 
1530   if (interpreter_frame_manager) {
1531     return interpreter_frame_manager;
1532   }
1533 
1534   __ bind(frame_manager_entry);
1535 
1536   // the following temporary registers are used during frame creation
1537   const Register Gtmp1 = G3_scratch;
1538   // const Register Lmirror = L1;     // native mirror (native calls only)
1539 
1540   const Address constants         (G5_method, 0, in_bytes(methodOopDesc::constants_offset()));
1541   const Address access_flags      (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
1542   const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1543   const Address max_stack         (G5_method, 0, in_bytes(methodOopDesc::max_stack_offset()));
1544   const Address size_of_locals    (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1545 
1546   address entry_point = __ pc();
1547   __ mov(G0, prevState);                                                 // no current activation
1548 
1549 
1550   Label re_dispatch;
1551 
1552   __ bind(re_dispatch);
1553 
1554   // Interpreter needs to have locals completely contiguous. In order to do that
1555   // We must adjust the caller's stack pointer for any locals beyond just the
1556   // parameters
1557   adjust_callers_stack(Gargs);
1558 
1559   // O5_savedSP still contains sender's sp
1560 
1561   // NEW FRAME
1562 
1563   generate_compute_interpreter_state(Lstate, prevState, false);
1564 
1565   // At this point a new interpreter frame and state object are created and initialized
1566   // Lstate has the pointer to the new activation
1567   // Any stack banging or limit check should already be done.
1568 
1569   Label call_interpreter;
1570 
1571   __ bind(call_interpreter);
1572 
1573 
1574 #if 1
1575   __ set(0xdead002, Lmirror);
1576   __ set(0xdead002, L2_scratch);
1577   __ set(0xdead003, L3_scratch);
1578   __ set(0xdead004, L4_scratch);
1579   __ set(0xdead005, Lscratch);
1580   __ set(0xdead006, Lscratch2);
1581   __ set(0xdead007, L7_scratch);
1582 
1583   __ set(0xdeaf002, O2);
1584   __ set(0xdeaf003, O3);
1585   __ set(0xdeaf004, O4);
1586   __ set(0xdeaf005, O5);
1587 #endif
1588 
1589   // Call interpreter (stack bang complete) enter here if message is
1590   // set and we know stack size is valid
1591 
1592   Label call_interpreter_2;
1593 
1594   __ bind(call_interpreter_2);
1595 
1596 #ifdef ASSERT
1597   {
1598     Label skip;
1599     __ ld_ptr(STATE(_frame_bottom), G3_scratch);
1600     __ cmp(G3_scratch, SP);
1601     __ brx(Assembler::equal, false, Assembler::pt, skip);
1602     __ delayed()->nop();
1603     __ stop("SP not restored to frame bottom");
1604     __ bind(skip);
1605   }
1606 #endif
1607 
1608   VALIDATE_STATE(G3_scratch, 4);
1609   __ set_last_Java_frame(SP, noreg);
1610   __ mov(Lstate, O0);                 // (arg) pointer to current state
1611 
1612   __ call(CAST_FROM_FN_PTR(address,
1613                            JvmtiExport::can_post_interpreter_events() ?
1614                                                                   BytecodeInterpreter::runWithChecks
1615                                                                 : BytecodeInterpreter::run),
1616          relocInfo::runtime_call_type);
1617 
1618   __ delayed()->nop();
1619 
1620   __ ld_ptr(STATE(_thread), G2_thread);
1621   __ reset_last_Java_frame();
1622 
1623   // examine msg from interpreter to determine next action
1624   __ ld_ptr(STATE(_thread), G2_thread);                                  // restore G2_thread
1625 
1626   __ ld(STATE(_msg), L1_scratch);                                       // Get new message
1627 
1628   Label call_method;
1629   Label return_from_interpreted_method;
1630   Label throw_exception;
1631   Label do_OSR;
1632   Label bad_msg;
1633   Label resume_interpreter;
1634 
1635   __ cmp(L1_scratch, (int)BytecodeInterpreter::call_method);
1636   __ br(Assembler::equal, false, Assembler::pt, call_method);
1637   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::return_from_method);
1638   __ br(Assembler::equal, false, Assembler::pt, return_from_interpreted_method);
1639   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::throwing_exception);
1640   __ br(Assembler::equal, false, Assembler::pt, throw_exception);
1641   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::do_osr);
1642   __ br(Assembler::equal, false, Assembler::pt, do_OSR);
1643   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::more_monitors);
1644   __ br(Assembler::notEqual, false, Assembler::pt, bad_msg);
1645 
1646   // Allocate more monitor space, shuffle expression stack....
1647 
1648   generate_more_monitors();
1649 
1650   // new monitor slot allocated, resume the interpreter.
1651 
1652   __ set((int)BytecodeInterpreter::got_monitors, L1_scratch);
1653   VALIDATE_STATE(G3_scratch, 5);
1654   __ ba(false, call_interpreter);
1655   __ delayed()->st(L1_scratch, STATE(_msg));
1656 
1657   // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode)
1658   unctrap_frame_manager_entry  = __ pc();
1659 
1660   // QQQ what message do we send
1661 
1662   __ ba(false, call_interpreter);
1663   __ delayed()->ld_ptr(STATE(_frame_bottom), SP);                  // restore to full stack frame
1664 
1665   //=============================================================================
1666   // Returning from a compiled method into a deopted method. The bytecode at the
1667   // bcp has completed. The result of the bytecode is in the native abi (the tosca
1668   // for the template based interpreter). Any stack space that was used by the
1669   // bytecode that has completed has been removed (e.g. parameters for an invoke)
1670   // so all that we have to do is place any pending result on the expression stack
1671   // and resume execution on the next bytecode.
1672 
1673   generate_deopt_handling();
1674 
1675   // ready to resume the interpreter
1676 
1677   __ set((int)BytecodeInterpreter::deopt_resume, L1_scratch);
1678   __ ba(false, call_interpreter);
1679   __ delayed()->st(L1_scratch, STATE(_msg));
1680 
1681   // Current frame has caught an exception we need to dispatch to the
1682   // handler. We can get here because a native interpreter frame caught
1683   // an exception in which case there is no handler and we must rethrow
1684   // If it is a vanilla interpreted frame the we simply drop into the
1685   // interpreter and let it do the lookup.
1686 
1687   Interpreter::_rethrow_exception_entry = __ pc();
1688 
1689   Label return_with_exception;
1690   Label unwind_and_forward;
1691 
1692   // O0: exception
1693   // O7: throwing pc
1694 
1695   // We want exception in the thread no matter what we ultimately decide about frame type.
1696 
1697   Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
1698   __ verify_thread();
1699   __ st_ptr(O0, exception_addr);
1700 
1701   // get the methodOop
1702   __ ld_ptr(STATE(_method), G5_method);
1703 
1704   // if this current frame vanilla or native?
1705 
1706   __ ld(access_flags, Gtmp1);
1707   __ btst(JVM_ACC_NATIVE, Gtmp1);
1708   __ br(Assembler::zero, false, Assembler::pt, return_with_exception);  // vanilla interpreted frame handle directly
1709   __ delayed()->nop();
1710 
1711   // We drop thru to unwind a native interpreted frame with a pending exception
1712   // We jump here for the initial interpreter frame with exception pending
1713   // We unwind the current acivation and forward it to our caller.
1714 
1715   __ bind(unwind_and_forward);
1716 
1717   // Unwind frame and jump to forward exception. unwinding will place throwing pc in O7
1718   // as expected by forward_exception.
1719 
1720   __ restore(FP, G0, SP);                  // unwind interpreter state frame
1721   __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
1722   __ delayed()->mov(I5_savedSP->after_restore(), SP);
1723 
1724   // Return point from a call which returns a result in the native abi
1725   // (c1/c2/jni-native). This result must be processed onto the java
1726   // expression stack.
1727   //
1728   // A pending exception may be present in which case there is no result present
1729 
1730   address return_from_native_method = __ pc();
1731 
1732   VALIDATE_STATE(G3_scratch, 6);
1733 
1734   // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1735   // stack is in the state that the  calling convention left it.
1736   // Copy the result from native abi result and place it on java expression stack.
1737 
1738   // Current interpreter state is present in Lstate
1739 
1740   // Exception pending?
1741 
1742   __ ld_ptr(STATE(_frame_bottom), SP);                             // restore to full stack frame
1743   __ ld_ptr(exception_addr, Lscratch);                                         // get any pending exception
1744   __ tst(Lscratch);                                                            // exception pending?
1745   __ brx(Assembler::notZero, false, Assembler::pt, return_with_exception);
1746   __ delayed()->nop();
1747 
1748   // Process the native abi result to java expression stack
1749 
1750   __ ld_ptr(STATE(_result._to_call._callee), L4_scratch);                        // called method
1751   __ ld_ptr(STATE(_stack), L1_scratch);                                          // get top of java expr stack
1752   __ lduh(L4_scratch, in_bytes(methodOopDesc::size_of_parameters_offset()), L2_scratch); // get parameter size
1753   __ sll(L2_scratch, LogBytesPerWord, L2_scratch     );                           // parameter size in bytes
1754   __ add(L1_scratch, L2_scratch, L1_scratch);                                      // stack destination for result
1755   __ ld(L4_scratch, in_bytes(methodOopDesc::result_index_offset()), L3_scratch); // called method result type index
1756 
1757   // tosca is really just native abi
1758   __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1759   __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1760   __ ld_ptr(L4_scratch, L3_scratch, Lscratch);                                       // get typed result converter address
1761   __ jmpl(Lscratch, G0, O7);                                                   // and convert it
1762   __ delayed()->nop();
1763 
1764   // L1_scratch points to top of stack (prepushed)
1765 
1766   __ ba(false, resume_interpreter);
1767   __ delayed()->mov(L1_scratch, O1);
1768 
1769   // An exception is being caught on return to a vanilla interpreter frame.
1770   // Empty the stack and resume interpreter
1771 
1772   __ bind(return_with_exception);
1773 
1774   __ ld_ptr(STATE(_frame_bottom), SP);                             // restore to full stack frame
1775   __ ld_ptr(STATE(_stack_base), O1);                               // empty java expression stack
1776   __ ba(false, resume_interpreter);
1777   __ delayed()->sub(O1, wordSize, O1);                             // account for prepush
1778 
1779   // Return from interpreted method we return result appropriate to the caller (i.e. "recursive"
1780   // interpreter call, or native) and unwind this interpreter activation.
1781   // All monitors should be unlocked.
1782 
1783   __ bind(return_from_interpreted_method);
1784 
1785   VALIDATE_STATE(G3_scratch, 7);
1786 
1787   Label return_to_initial_caller;
1788 
1789   // Interpreted result is on the top of the completed activation expression stack.
1790   // We must return it to the top of the callers stack if caller was interpreted
1791   // otherwise we convert to native abi result and return to call_stub/c1/c2
1792   // The caller's expression stack was truncated by the call however the current activation
1793   // has enough stuff on the stack that we have usable space there no matter what. The
1794   // other thing that makes it easy is that the top of the caller's stack is stored in STATE(_locals)
1795   // for the current activation
1796 
1797   __ ld_ptr(STATE(_prev_link), L1_scratch);
1798   __ ld_ptr(STATE(_method), L2_scratch);                               // get method just executed
1799   __ ld(L2_scratch, in_bytes(methodOopDesc::result_index_offset()), L2_scratch);
1800   __ tst(L1_scratch);
1801   __ brx(Assembler::zero, false, Assembler::pt, return_to_initial_caller);
1802   __ delayed()->sll(L2_scratch, LogBytesPerWord, L2_scratch);
1803 
1804   // Copy result to callers java stack
1805 
1806   __ set((intptr_t)CppInterpreter::_stack_to_stack, L4_scratch);
1807   __ ld_ptr(L4_scratch, L2_scratch, Lscratch);                          // get typed result converter address
1808   __ ld_ptr(STATE(_stack), O0);                                       // current top (prepushed)
1809   __ ld_ptr(STATE(_locals), O1);                                      // stack destination
1810 
1811   // O0 - will be source, O1 - will be destination (preserved)
1812   __ jmpl(Lscratch, G0, O7);                                          // and convert it
1813   __ delayed()->add(O0, wordSize, O0);                                // get source (top of current expr stack)
1814 
1815   // O1 == &locals[0]
1816 
1817   // Result is now on caller's stack. Just unwind current activation and resume
1818 
1819   Label unwind_recursive_activation;
1820 
1821 
1822   __ bind(unwind_recursive_activation);
1823 
1824   // O1 == &locals[0] (really callers stacktop) for activation now returning
1825   // returning to interpreter method from "recursive" interpreter call
1826   // result converter left O1 pointing to top of the( prepushed) java stack for method we are returning
1827   // to. Now all we must do is unwind the state from the completed call
1828 
1829   // Must restore stack
1830   VALIDATE_STATE(G3_scratch, 8);
1831 
1832   // Return to interpreter method after a method call (interpreted/native/c1/c2) has completed.
1833   // Result if any is already on the caller's stack. All we must do now is remove the now dead
1834   // frame and tell interpreter to resume.
1835 
1836 
1837   __ mov(O1, I1);                                                     // pass back new stack top across activation
1838   // POP FRAME HERE ==================================
1839   __ restore(FP, G0, SP);                                             // unwind interpreter state frame
1840   __ ld_ptr(STATE(_frame_bottom), SP);                                // restore to full stack frame
1841 
1842 
1843   // Resume the interpreter. The current frame contains the current interpreter
1844   // state object.
1845   //
1846   // O1 == new java stack pointer
1847 
1848   __ bind(resume_interpreter);
1849   VALIDATE_STATE(G3_scratch, 10);
1850 
1851   // A frame we have already used before so no need to bang stack so use call_interpreter_2 entry
1852 
1853   __ set((int)BytecodeInterpreter::method_resume, L1_scratch);
1854   __ st(L1_scratch, STATE(_msg));
1855   __ ba(false, call_interpreter_2);
1856   __ delayed()->st_ptr(O1, STATE(_stack));
1857 
1858 
1859   // Fast accessor methods share this entry point.
1860   // This works because frame manager is in the same codelet
1861   // This can either be an entry via call_stub/c1/c2 or a recursive interpreter call
1862   // we need to do a little register fixup here once we distinguish the two of them
1863   if (UseFastAccessorMethods && !synchronized) {
1864   // Call stub_return address still in O7
1865     __ bind(fast_accessor_slow_entry_path);
1866     __ set((intptr_t)return_from_native_method - 8, Gtmp1);
1867     __ cmp(Gtmp1, O7);                                                // returning to interpreter?
1868     __ brx(Assembler::equal, true, Assembler::pt, re_dispatch);       // yep
1869     __ delayed()->nop();
1870     __ ba(false, re_dispatch);
1871     __ delayed()->mov(G0, prevState);                                   // initial entry
1872 
1873   }
1874 
1875   // interpreter returning to native code (call_stub/c1/c2)
1876   // convert result and unwind initial activation
1877   // L2_scratch - scaled result type index
1878 
1879   __ bind(return_to_initial_caller);
1880 
1881   __ set((intptr_t)CppInterpreter::_stack_to_native_abi, L4_scratch);
1882   __ ld_ptr(L4_scratch, L2_scratch, Lscratch);                           // get typed result converter address
1883   __ ld_ptr(STATE(_stack), O0);                                        // current top (prepushed)
1884   __ jmpl(Lscratch, G0, O7);                                           // and convert it
1885   __ delayed()->add(O0, wordSize, O0);                                 // get source (top of current expr stack)
1886 
1887   Label unwind_initial_activation;
1888   __ bind(unwind_initial_activation);
1889 
1890   // RETURN TO CALL_STUB/C1/C2 code (result if any in I0..I1/(F0/..F1)
1891   // we can return here with an exception that wasn't handled by interpreted code
1892   // how does c1/c2 see it on return?
1893 
1894   // compute resulting sp before/after args popped depending upon calling convention
1895   // __ ld_ptr(STATE(_saved_sp), Gtmp1);
1896   //
1897   // POP FRAME HERE ==================================
1898   __ restore(FP, G0, SP);
1899   __ retl();
1900   __ delayed()->mov(I5_savedSP->after_restore(), SP);
1901 
1902   // OSR request, unwind the current frame and transfer to the OSR entry
1903   // and enter OSR nmethod
1904 
1905   __ bind(do_OSR);
1906   Label remove_initial_frame;
1907   __ ld_ptr(STATE(_prev_link), L1_scratch);
1908   __ ld_ptr(STATE(_result._osr._osr_buf), G1_scratch);
1909 
1910   // We are going to pop this frame. Is there another interpreter frame underneath
1911   // it or is it callstub/compiled?
1912 
1913   __ tst(L1_scratch);
1914   __ brx(Assembler::zero, false, Assembler::pt, remove_initial_frame);
1915   __ delayed()->ld_ptr(STATE(_result._osr._osr_entry), G3_scratch);
1916 
1917   // Frame underneath is an interpreter frame simply unwind
1918   // POP FRAME HERE ==================================
1919   __ restore(FP, G0, SP);                                             // unwind interpreter state frame
1920   __ mov(I5_savedSP->after_restore(), SP);
1921 
1922   // Since we are now calling native need to change our "return address" from the
1923   // dummy RecursiveInterpreterActivation to a return from native
1924 
1925   __ set((intptr_t)return_from_native_method - 8, O7);
1926 
1927   __ jmpl(G3_scratch, G0, G0);
1928   __ delayed()->mov(G1_scratch, O0);
1929 
1930   __ bind(remove_initial_frame);
1931 
1932   // POP FRAME HERE ==================================
1933   __ restore(FP, G0, SP);
1934   __ mov(I5_savedSP->after_restore(), SP);
1935   __ jmpl(G3_scratch, G0, G0);
1936   __ delayed()->mov(G1_scratch, O0);
1937 
1938   // Call a new method. All we do is (temporarily) trim the expression stack
1939   // push a return address to bring us back to here and leap to the new entry.
1940   // At this point we have a topmost frame that was allocated by the frame manager
1941   // which contains the current method interpreted state. We trim this frame
1942   // of excess java expression stack entries and then recurse.
1943 
1944   __ bind(call_method);
1945 
1946   // stack points to next free location and not top element on expression stack
1947   // method expects sp to be pointing to topmost element
1948 
1949   __ ld_ptr(STATE(_thread), G2_thread);
1950   __ ld_ptr(STATE(_result._to_call._callee), G5_method);
1951 
1952 
1953   // SP already takes in to account the 2 extra words we use for slop
1954   // when we call a "static long no_params()" method. So if
1955   // we trim back sp by the amount of unused java expression stack
1956   // there will be automagically the 2 extra words we need.
1957   // We also have to worry about keeping SP aligned.
1958 
1959   __ ld_ptr(STATE(_stack), Gargs);
1960   __ ld_ptr(STATE(_stack_limit), L1_scratch);
1961 
1962   // compute the unused java stack size
1963   __ sub(Gargs, L1_scratch, L2_scratch);                       // compute unused space
1964 
1965   // Round down the unused space to that stack is always 16-byte aligned
1966   // by making the unused space a multiple of the size of two longs.
1967 
1968   __ and3(L2_scratch, -2*BytesPerLong, L2_scratch);
1969 
1970   // Now trim the stack
1971   __ add(SP, L2_scratch, SP);
1972 
1973 
1974   // Now point to the final argument (account for prepush)
1975   __ add(Gargs, wordSize, Gargs);
1976 #ifdef ASSERT
1977   // Make sure we have space for the window
1978   __ sub(Gargs, SP, L1_scratch);
1979   __ cmp(L1_scratch, 16*wordSize);
1980   {
1981     Label skip;
1982     __ brx(Assembler::greaterEqual, false, Assembler::pt, skip);
1983     __ delayed()->nop();
1984     __ stop("killed stack");
1985     __ bind(skip);
1986   }
1987 #endif // ASSERT
1988 
1989   // Create a new frame where we can store values that make it look like the interpreter
1990   // really recursed.
1991 
1992   // prepare to recurse or call specialized entry
1993 
1994   // First link the registers we need
1995 
1996   // make the pc look good in debugger
1997   __ set(CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation), O7);
1998   // argument too
1999   __ mov(Lstate, I0);
2000 
2001   // Record our sending SP
2002   __ mov(SP, O5_savedSP);
2003 
2004   __ ld_ptr(STATE(_result._to_call._callee_entry_point), L2_scratch);
2005   __ set((intptr_t) entry_point, L1_scratch);
2006   __ cmp(L1_scratch, L2_scratch);
2007   __ brx(Assembler::equal, false, Assembler::pt, re_dispatch);
2008   __ delayed()->mov(Lstate, prevState);                                // link activations
2009 
2010   // method uses specialized entry, push a return so we look like call stub setup
2011   // this path will handle fact that result is returned in registers and not
2012   // on the java stack.
2013 
2014   __ set((intptr_t)return_from_native_method - 8, O7);
2015   __ jmpl(L2_scratch, G0, G0);                               // Do specialized entry
2016   __ delayed()->nop();
2017 
2018   //
2019   // Bad Message from interpreter
2020   //
2021   __ bind(bad_msg);
2022   __ stop("Bad message from interpreter");
2023 
2024   // Interpreted method "returned" with an exception pass it on...
2025   // Pass result, unwind activation and continue/return to interpreter/call_stub
2026   // We handle result (if any) differently based on return to interpreter or call_stub
2027 
2028   __ bind(throw_exception);
2029   __ ld_ptr(STATE(_prev_link), L1_scratch);
2030   __ tst(L1_scratch);
2031   __ brx(Assembler::zero, false, Assembler::pt, unwind_and_forward);
2032   __ delayed()->nop();
2033 
2034   __ ld_ptr(STATE(_locals), O1);                                   // get result of popping callee's args
2035   __ ba(false, unwind_recursive_activation);
2036   __ delayed()->nop();
2037 
2038   interpreter_frame_manager = entry_point;
2039   return entry_point;
2040 }
2041 
2042 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2043  : CppInterpreterGenerator(code) {
2044    generate_all(); // down here so it can be "virtual"
2045 }
2046 
2047 
2048 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
2049 
2050   // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
2051   // expression stack, the callee will have callee_extra_locals (so we can account for
2052   // frame extension) and monitor_size for monitors. Basically we need to calculate
2053   // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
2054   //
2055   //
2056   // The big complicating thing here is that we must ensure that the stack stays properly
2057   // aligned. This would be even uglier if monitor size wasn't modulo what the stack
2058   // needs to be aligned for). We are given that the sp (fp) is already aligned by
2059   // the caller so we must ensure that it is properly aligned for our callee.
2060   //
2061   // Ths c++ interpreter always makes sure that we have a enough extra space on the
2062   // stack at all times to deal with the "stack long no_params()" method issue. This
2063   // is "slop_factor" here.
2064   const int slop_factor = 2;
2065 
2066   const int fixed_size = sizeof(BytecodeInterpreter)/wordSize +           // interpreter state object
2067                          frame::memory_parameter_word_sp_offset;   // register save area + param window
2068   const int extra_stack = 0; //6815692//methodOopDesc::extra_stack_entries();
2069   return (round_to(max_stack +
2070                    extra_stack +
2071                    slop_factor +
2072                    fixed_size +
2073                    monitor_size +
2074                    (callee_extra_locals * Interpreter::stackElementWords()), WordsPerLong));
2075 
2076 }
2077 
2078 int AbstractInterpreter::size_top_interpreter_activation(methodOop method) {
2079 
2080   // See call_stub code
2081   int call_stub_size  = round_to(7 + frame::memory_parameter_word_sp_offset,
2082                                  WordsPerLong);    // 7 + register save area
2083 
2084   // Save space for one monitor to get into the interpreted method in case
2085   // the method is synchronized
2086   int monitor_size    = method->is_synchronized() ?
2087                                 1*frame::interpreter_frame_monitor_size() : 0;
2088   return size_activation_helper(method->max_locals(), method->max_stack(),
2089                                  monitor_size) + call_stub_size;
2090 }
2091 
2092 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
2093                                            frame* caller,
2094                                            frame* current,
2095                                            methodOop method,
2096                                            intptr_t* locals,
2097                                            intptr_t* stack,
2098                                            intptr_t* stack_base,
2099                                            intptr_t* monitor_base,
2100                                            intptr_t* frame_bottom,
2101                                            bool is_top_frame
2102                                            )
2103 {
2104   // What about any vtable?
2105   //
2106   to_fill->_thread = JavaThread::current();
2107   // This gets filled in later but make it something recognizable for now
2108   to_fill->_bcp = method->code_base();
2109   to_fill->_locals = locals;
2110   to_fill->_constants = method->constants()->cache();
2111   to_fill->_method = method;
2112   to_fill->_mdx = NULL;
2113   to_fill->_stack = stack;
2114   if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) {
2115     to_fill->_msg = deopt_resume2;
2116   } else {
2117     to_fill->_msg = method_resume;
2118   }
2119   to_fill->_result._to_call._bcp_advance = 0;
2120   to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2121   to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2122   to_fill->_prev_link = NULL;
2123 
2124   // Fill in the registers for the frame
2125 
2126   // Need to install _sender_sp. Actually not too hard in C++!
2127   // When the skeletal frames are layed out we fill in a value
2128   // for _sender_sp. That value is only correct for the oldest
2129   // skeletal frame constructed (because there is only a single
2130   // entry for "caller_adjustment". While the skeletal frames
2131   // exist that is good enough. We correct that calculation
2132   // here and get all the frames correct.
2133 
2134   // to_fill->_sender_sp = locals - (method->size_of_parameters() - 1);
2135 
2136   *current->register_addr(Lstate) = (intptr_t) to_fill;
2137   // skeletal already places a useful value here and this doesn't account
2138   // for alignment so don't bother.
2139   // *current->register_addr(I5_savedSP) =     (intptr_t) locals - (method->size_of_parameters() - 1);
2140 
2141   if (caller->is_interpreted_frame()) {
2142     interpreterState prev  = caller->get_interpreterState();
2143     to_fill->_prev_link = prev;
2144     // Make the prev callee look proper
2145     prev->_result._to_call._callee = method;
2146     if (*prev->_bcp == Bytecodes::_invokeinterface) {
2147       prev->_result._to_call._bcp_advance = 5;
2148     } else {
2149       prev->_result._to_call._bcp_advance = 3;
2150     }
2151   }
2152   to_fill->_oop_temp = NULL;
2153   to_fill->_stack_base = stack_base;
2154   // Need +1 here because stack_base points to the word just above the first expr stack entry
2155   // and stack_limit is supposed to point to the word just below the last expr stack entry.
2156   // See generate_compute_interpreter_state.
2157   int extra_stack = 0; //6815692//methodOopDesc::extra_stack_entries();
2158   to_fill->_stack_limit = stack_base - (method->max_stack() + 1 + extra_stack);
2159   to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2160 
2161   // sparc specific
2162   to_fill->_frame_bottom = frame_bottom;
2163   to_fill->_self_link = to_fill;
2164 #ifdef ASSERT
2165   to_fill->_native_fresult = 123456.789;
2166   to_fill->_native_lresult = CONST64(0xdeadcafedeafcafe);
2167 #endif
2168 }
2169 
2170 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, address last_Java_pc, intptr_t* last_Java_fp) {
2171   istate->_last_Java_pc = (intptr_t*) last_Java_pc;
2172 }
2173 
2174 
2175 int AbstractInterpreter::layout_activation(methodOop method,
2176                                            int tempcount, // Number of slots on java expression stack in use
2177                                            int popframe_extra_args,
2178                                            int moncount,  // Number of active monitors
2179                                            int caller_actual_parameters,
2180                                            int callee_param_size,
2181                                            int callee_locals_size,
2182                                            frame* caller,
2183                                            frame* interpreter_frame,
2184                                            bool is_top_frame) {
2185 
2186   assert(popframe_extra_args == 0, "NEED TO FIX");
2187   // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state()
2188   // does as far as allocating an interpreter frame.
2189   // If interpreter_frame!=NULL, set up the method, locals, and monitors.
2190   // The frame interpreter_frame, if not NULL, is guaranteed to be the right size,
2191   // as determined by a previous call to this method.
2192   // It is also guaranteed to be walkable even though it is in a skeletal state
2193   // NOTE: return size is in words not bytes
2194   // NOTE: tempcount is the current size of the java expression stack. For top most
2195   //       frames we will allocate a full sized expression stack and not the curback
2196   //       version that non-top frames have.
2197 
2198   // Calculate the amount our frame will be adjust by the callee. For top frame
2199   // this is zero.
2200 
2201   // NOTE: ia64 seems to do this wrong (or at least backwards) in that it
2202   // calculates the extra locals based on itself. Not what the callee does
2203   // to it. So it ignores last_frame_adjust value. Seems suspicious as far
2204   // as getting sender_sp correct.
2205 
2206   int extra_locals_size = callee_locals_size - callee_param_size;
2207   int monitor_size = (sizeof(BasicObjectLock) * moncount) / wordSize;
2208   int full_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2209   int short_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2210   int frame_words = is_top_frame ? full_frame_words : short_frame_words;
2211 
2212 
2213   /*
2214     if we actually have a frame to layout we must now fill in all the pieces. This means both
2215     the interpreterState and the registers.
2216   */
2217   if (interpreter_frame != NULL) {
2218 
2219     // MUCHO HACK
2220 
2221     intptr_t* frame_bottom = interpreter_frame->sp() - (full_frame_words - frame_words);
2222     // 'interpreter_frame->sp()' is unbiased while 'frame_bottom' must be a biased value in 64bit mode.
2223     assert(((intptr_t)frame_bottom & 0xf) == 0, "SP biased in layout_activation");
2224     frame_bottom = (intptr_t*)((intptr_t)frame_bottom - STACK_BIAS);
2225 
2226     /* Now fillin the interpreterState object */
2227 
2228     interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() -  sizeof(BytecodeInterpreter));
2229 
2230 
2231     intptr_t* locals;
2232 
2233     // Calculate the postion of locals[0]. This is painful because of
2234     // stack alignment (same as ia64). The problem is that we can
2235     // not compute the location of locals from fp(). fp() will account
2236     // for the extra locals but it also accounts for aligning the stack
2237     // and we can't determine if the locals[0] was misaligned but max_locals
2238     // was enough to have the
2239     // calculate postion of locals. fp already accounts for extra locals.
2240     // +2 for the static long no_params() issue.
2241 
2242     if (caller->is_interpreted_frame()) {
2243       // locals must agree with the caller because it will be used to set the
2244       // caller's tos when we return.
2245       interpreterState prev  = caller->get_interpreterState();
2246       // stack() is prepushed.
2247       locals = prev->stack() + method->size_of_parameters();
2248     } else {
2249       // Lay out locals block in the caller adjacent to the register window save area.
2250       //
2251       // Compiled frames do not allocate a varargs area which is why this if
2252       // statement is needed.
2253       //
2254       intptr_t* fp = interpreter_frame->fp();
2255       int local_words = method->max_locals() * Interpreter::stackElementWords();
2256 
2257       if (caller->is_compiled_frame()) {
2258         locals = fp + frame::register_save_words + local_words - 1;
2259       } else {
2260         locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
2261       }
2262 
2263     }
2264     // END MUCHO HACK
2265 
2266     intptr_t* monitor_base = (intptr_t*) cur_state;
2267     intptr_t* stack_base =  monitor_base - monitor_size;
2268     /* +1 because stack is always prepushed */
2269     intptr_t* stack = stack_base - (tempcount + 1);
2270 
2271 
2272     BytecodeInterpreter::layout_interpreterState(cur_state,
2273                                           caller,
2274                                           interpreter_frame,
2275                                           method,
2276                                           locals,
2277                                           stack,
2278                                           stack_base,
2279                                           monitor_base,
2280                                           frame_bottom,
2281                                           is_top_frame);
2282 
2283     BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp());
2284 
2285   }
2286   return frame_words;
2287 }
2288 
2289 #endif // CC_INTERP