1 /*
   2  * Copyright (c) 2007, 2012, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/assembler.hpp"
  27 #include "interpreter/bytecodeHistogram.hpp"
  28 #include "interpreter/cppInterpreter.hpp"
  29 #include "interpreter/interpreter.hpp"
  30 #include "interpreter/interpreterGenerator.hpp"
  31 #include "interpreter/interpreterRuntime.hpp"
  32 #include "oops/arrayOop.hpp"
  33 #include "oops/methodData.hpp"
  34 #include "oops/method.hpp"
  35 #include "oops/oop.inline.hpp"
  36 #include "prims/jvmtiExport.hpp"
  37 #include "prims/jvmtiThreadState.hpp"
  38 #include "runtime/arguments.hpp"
  39 #include "runtime/deoptimization.hpp"
  40 #include "runtime/frame.inline.hpp"
  41 #include "runtime/interfaceSupport.hpp"
  42 #include "runtime/sharedRuntime.hpp"
  43 #include "runtime/stubRoutines.hpp"
  44 #include "runtime/synchronizer.hpp"
  45 #include "runtime/timer.hpp"
  46 #include "runtime/vframeArray.hpp"
  47 #include "utilities/debug.hpp"
  48 #include "utilities/macros.hpp"
  49 #ifdef SHARK
  50 #include "shark/shark_globals.hpp"
  51 #endif
  52 
  53 #ifdef CC_INTERP
  54 
  55 // Routine exists to make tracebacks look decent in debugger
  56 // while "shadow" interpreter frames are on stack. It is also
  57 // used to distinguish interpreter frames.
  58 
  59 extern "C" void RecursiveInterpreterActivation(interpreterState istate) {
  60   ShouldNotReachHere();
  61 }
  62 
  63 bool CppInterpreter::contains(address pc) {
  64   return ( _code->contains(pc) ||
  65          ( pc == (CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset)));
  66 }
  67 
  68 #define STATE(field_name) Lstate, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
  69 #define __ _masm->
  70 
  71 Label frame_manager_entry;
  72 Label fast_accessor_slow_entry_path;  // fast accessor methods need to be able to jmp to unsynchronized
  73                                       // c++ interpreter entry point this holds that entry point label.
  74 
  75 static address unctrap_frame_manager_entry  = NULL;
  76 
  77 static address interpreter_return_address  = NULL;
  78 static address deopt_frame_manager_return_atos  = NULL;
  79 static address deopt_frame_manager_return_btos  = NULL;
  80 static address deopt_frame_manager_return_itos  = NULL;
  81 static address deopt_frame_manager_return_ltos  = NULL;
  82 static address deopt_frame_manager_return_ftos  = NULL;
  83 static address deopt_frame_manager_return_dtos  = NULL;
  84 static address deopt_frame_manager_return_vtos  = NULL;
  85 
  86 const Register prevState = G1_scratch;
  87 
  88 void InterpreterGenerator::save_native_result(void) {
  89   // result potentially in O0/O1: save it across calls
  90   __ stf(FloatRegisterImpl::D, F0, STATE(_native_fresult));
  91 #ifdef _LP64
  92   __ stx(O0, STATE(_native_lresult));
  93 #else
  94   __ std(O0, STATE(_native_lresult));
  95 #endif
  96 }
  97 
  98 void InterpreterGenerator::restore_native_result(void) {
  99 
 100   // Restore any method result value
 101   __ ldf(FloatRegisterImpl::D, STATE(_native_fresult), F0);
 102 #ifdef _LP64
 103   __ ldx(STATE(_native_lresult), O0);
 104 #else
 105   __ ldd(STATE(_native_lresult), O0);
 106 #endif
 107 }
 108 
 109 // A result handler converts/unboxes a native call result into
 110 // a java interpreter/compiler result. The current frame is an
 111 // interpreter frame. The activation frame unwind code must be
 112 // consistent with that of TemplateTable::_return(...). In the
 113 // case of native methods, the caller's SP was not modified.
 114 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) {
 115   address entry = __ pc();
 116   Register Itos_i  = Otos_i ->after_save();
 117   Register Itos_l  = Otos_l ->after_save();
 118   Register Itos_l1 = Otos_l1->after_save();
 119   Register Itos_l2 = Otos_l2->after_save();
 120   switch (type) {
 121     case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
 122     case T_CHAR   : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i);   break; // cannot use and3, 0xFFFF too big as immediate value!
 123     case T_BYTE   : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i);   break;
 124     case T_SHORT  : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i);   break;
 125     case T_LONG   :
 126 #ifndef _LP64
 127                     __ mov(O1, Itos_l2);  // move other half of long
 128 #endif              // ifdef or no ifdef, fall through to the T_INT case
 129     case T_INT    : __ mov(O0, Itos_i);                         break;
 130     case T_VOID   : /* nothing to do */                         break;
 131     case T_FLOAT  : assert(F0 == Ftos_f, "fix this code" );     break;
 132     case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" );     break;
 133     case T_OBJECT :
 134       __ ld_ptr(STATE(_oop_temp), Itos_i);
 135       __ verify_oop(Itos_i);
 136       break;
 137     default       : ShouldNotReachHere();
 138   }
 139   __ ret();                           // return from interpreter activation
 140   __ delayed()->restore(I5_savedSP, G0, SP);  // remove interpreter frame
 141   NOT_PRODUCT(__ emit_int32(0);)       // marker for disassembly
 142   return entry;
 143 }
 144 
 145 // tosca based result to c++ interpreter stack based result.
 146 // Result goes to address in L1_scratch
 147 
 148 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) {
 149   // A result is in the native abi result register from a native method call.
 150   // We need to return this result to the interpreter by pushing the result on the interpreter's
 151   // stack. This is relatively simple the destination is in L1_scratch
 152   // i.e. L1_scratch is the first free element on the stack. If we "push" a return value we must
 153   // adjust L1_scratch
 154   address entry = __ pc();
 155   switch (type) {
 156     case T_BOOLEAN:
 157       // !0 => true; 0 => false
 158       __ subcc(G0, O0, G0);
 159       __ addc(G0, 0, O0);
 160       __ st(O0, L1_scratch, 0);
 161       __ sub(L1_scratch, wordSize, L1_scratch);
 162       break;
 163 
 164     // cannot use and3, 0xFFFF too big as immediate value!
 165     case T_CHAR   :
 166       __ sll(O0, 16, O0);
 167       __ srl(O0, 16, O0);
 168       __ st(O0, L1_scratch, 0);
 169       __ sub(L1_scratch, wordSize, L1_scratch);
 170       break;
 171 
 172     case T_BYTE   :
 173       __ sll(O0, 24, O0);
 174       __ sra(O0, 24, O0);
 175       __ st(O0, L1_scratch, 0);
 176       __ sub(L1_scratch, wordSize, L1_scratch);
 177       break;
 178 
 179     case T_SHORT  :
 180       __ sll(O0, 16, O0);
 181       __ sra(O0, 16, O0);
 182       __ st(O0, L1_scratch, 0);
 183       __ sub(L1_scratch, wordSize, L1_scratch);
 184       break;
 185     case T_LONG   :
 186 #ifndef _LP64
 187 #if defined(COMPILER2)
 188   // All return values are where we want them, except for Longs.  C2 returns
 189   // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
 190   // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
 191   // build even if we are returning from interpreted we just do a little
 192   // stupid shuffing.
 193   // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
 194   // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
 195   // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
 196       __ stx(G1, L1_scratch, -wordSize);
 197 #else
 198       // native result is in O0, O1
 199       __ st(O1, L1_scratch, 0);                      // Low order
 200       __ st(O0, L1_scratch, -wordSize);              // High order
 201 #endif /* COMPILER2 */
 202 #else
 203       __ stx(O0, L1_scratch, -wordSize);
 204 #endif
 205       __ sub(L1_scratch, 2*wordSize, L1_scratch);
 206       break;
 207 
 208     case T_INT    :
 209       __ st(O0, L1_scratch, 0);
 210       __ sub(L1_scratch, wordSize, L1_scratch);
 211       break;
 212 
 213     case T_VOID   : /* nothing to do */
 214       break;
 215 
 216     case T_FLOAT  :
 217       __ stf(FloatRegisterImpl::S, F0, L1_scratch, 0);
 218       __ sub(L1_scratch, wordSize, L1_scratch);
 219       break;
 220 
 221     case T_DOUBLE :
 222       // Every stack slot is aligned on 64 bit, However is this
 223       // the correct stack slot on 64bit?? QQQ
 224       __ stf(FloatRegisterImpl::D, F0, L1_scratch, -wordSize);
 225       __ sub(L1_scratch, 2*wordSize, L1_scratch);
 226       break;
 227     case T_OBJECT :
 228       __ verify_oop(O0);
 229       __ st_ptr(O0, L1_scratch, 0);
 230       __ sub(L1_scratch, wordSize, L1_scratch);
 231       break;
 232     default       : ShouldNotReachHere();
 233   }
 234   __ retl();                          // return from interpreter activation
 235   __ delayed()->nop();                // schedule this better
 236   NOT_PRODUCT(__ emit_int32(0);)       // marker for disassembly
 237   return entry;
 238 }
 239 
 240 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) {
 241   // A result is in the java expression stack of the interpreted method that has just
 242   // returned. Place this result on the java expression stack of the caller.
 243   //
 244   // The current interpreter activation in Lstate is for the method just returning its
 245   // result. So we know that the result of this method is on the top of the current
 246   // execution stack (which is pre-pushed) and will be return to the top of the caller
 247   // stack. The top of the callers stack is the bottom of the locals of the current
 248   // activation.
 249   // Because of the way activation are managed by the frame manager the value of esp is
 250   // below both the stack top of the current activation and naturally the stack top
 251   // of the calling activation. This enable this routine to leave the return address
 252   // to the frame manager on the stack and do a vanilla return.
 253   //
 254   // On entry: O0 - points to source (callee stack top)
 255   //           O1 - points to destination (caller stack top [i.e. free location])
 256   // destroys O2, O3
 257   //
 258 
 259   address entry = __ pc();
 260   switch (type) {
 261     case T_VOID:  break;
 262       break;
 263     case T_FLOAT  :
 264     case T_BOOLEAN:
 265     case T_CHAR   :
 266     case T_BYTE   :
 267     case T_SHORT  :
 268     case T_INT    :
 269       // 1 word result
 270       __ ld(O0, 0, O2);
 271       __ st(O2, O1, 0);
 272       __ sub(O1, wordSize, O1);
 273       break;
 274     case T_DOUBLE  :
 275     case T_LONG    :
 276       // return top two words on current expression stack to caller's expression stack
 277       // The caller's expression stack is adjacent to the current frame manager's intepretState
 278       // except we allocated one extra word for this intepretState so we won't overwrite it
 279       // when we return a two word result.
 280 #ifdef _LP64
 281       __ ld_ptr(O0, 0, O2);
 282       __ st_ptr(O2, O1, -wordSize);
 283 #else
 284       __ ld(O0, 0, O2);
 285       __ ld(O0, wordSize, O3);
 286       __ st(O3, O1, 0);
 287       __ st(O2, O1, -wordSize);
 288 #endif
 289       __ sub(O1, 2*wordSize, O1);
 290       break;
 291     case T_OBJECT :
 292       __ ld_ptr(O0, 0, O2);
 293       __ verify_oop(O2);                                               // verify it
 294       __ st_ptr(O2, O1, 0);
 295       __ sub(O1, wordSize, O1);
 296       break;
 297     default       : ShouldNotReachHere();
 298   }
 299   __ retl();
 300   __ delayed()->nop(); // QQ schedule this better
 301   return entry;
 302 }
 303 
 304 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) {
 305   // A result is in the java expression stack of the interpreted method that has just
 306   // returned. Place this result in the native abi that the caller expects.
 307   // We are in a new frame registers we set must be in caller (i.e. callstub) frame.
 308   //
 309   // Similar to generate_stack_to_stack_converter above. Called at a similar time from the
 310   // frame manager execept in this situation the caller is native code (c1/c2/call_stub)
 311   // and so rather than return result onto caller's java expression stack we return the
 312   // result in the expected location based on the native abi.
 313   // On entry: O0 - source (stack top)
 314   // On exit result in expected output register
 315   // QQQ schedule this better
 316 
 317   address entry = __ pc();
 318   switch (type) {
 319     case T_VOID:  break;
 320       break;
 321     case T_FLOAT  :
 322       __ ldf(FloatRegisterImpl::S, O0, 0, F0);
 323       break;
 324     case T_BOOLEAN:
 325     case T_CHAR   :
 326     case T_BYTE   :
 327     case T_SHORT  :
 328     case T_INT    :
 329       // 1 word result
 330       __ ld(O0, 0, O0->after_save());
 331       break;
 332     case T_DOUBLE  :
 333       __ ldf(FloatRegisterImpl::D, O0, 0, F0);
 334       break;
 335     case T_LONG    :
 336       // return top two words on current expression stack to caller's expression stack
 337       // The caller's expression stack is adjacent to the current frame manager's interpretState
 338       // except we allocated one extra word for this intepretState so we won't overwrite it
 339       // when we return a two word result.
 340 #ifdef _LP64
 341       __ ld_ptr(O0, 0, O0->after_save());
 342 #else
 343       __ ld(O0, wordSize, O1->after_save());
 344       __ ld(O0, 0, O0->after_save());
 345 #endif
 346 #if defined(COMPILER2) && !defined(_LP64)
 347       // C2 expects long results in G1 we can't tell if we're returning to interpreted
 348       // or compiled so just be safe use G1 and O0/O1
 349 
 350       // Shift bits into high (msb) of G1
 351       __ sllx(Otos_l1->after_save(), 32, G1);
 352       // Zero extend low bits
 353       __ srl (Otos_l2->after_save(), 0, Otos_l2->after_save());
 354       __ or3 (Otos_l2->after_save(), G1, G1);
 355 #endif /* COMPILER2 */
 356       break;
 357     case T_OBJECT :
 358       __ ld_ptr(O0, 0, O0->after_save());
 359       __ verify_oop(O0->after_save());                                               // verify it
 360       break;
 361     default       : ShouldNotReachHere();
 362   }
 363   __ retl();
 364   __ delayed()->nop();
 365   return entry;
 366 }
 367 
 368 address CppInterpreter::return_entry(TosState state, int length) {
 369   // make it look good in the debugger
 370   return CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset;
 371 }
 372 
 373 address CppInterpreter::deopt_entry(TosState state, int length) {
 374   address ret = NULL;
 375   if (length != 0) {
 376     switch (state) {
 377       case atos: ret = deopt_frame_manager_return_atos; break;
 378       case btos: ret = deopt_frame_manager_return_btos; break;
 379       case ctos:
 380       case stos:
 381       case itos: ret = deopt_frame_manager_return_itos; break;
 382       case ltos: ret = deopt_frame_manager_return_ltos; break;
 383       case ftos: ret = deopt_frame_manager_return_ftos; break;
 384       case dtos: ret = deopt_frame_manager_return_dtos; break;
 385       case vtos: ret = deopt_frame_manager_return_vtos; break;
 386     }
 387   } else {
 388     ret = unctrap_frame_manager_entry;  // re-execute the bytecode ( e.g. uncommon trap)
 389   }
 390   assert(ret != NULL, "Not initialized");
 391   return ret;
 392 }
 393 
 394 //
 395 // Helpers for commoning out cases in the various type of method entries.
 396 //
 397 
 398 // increment invocation count & check for overflow
 399 //
 400 // Note: checking for negative value instead of overflow
 401 //       so we have a 'sticky' overflow test
 402 //
 403 // Lmethod: method
 404 // ??: invocation counter
 405 //
 406 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
 407   // Update standard invocation counters
 408   __ increment_invocation_counter(O0, G3_scratch);
 409   if (ProfileInterpreter) {  // %%% Merge this into MethodData*
 410     __ ld_ptr(STATE(_method), G3_scratch);
 411     Address interpreter_invocation_counter(G3_scratch, 0, in_bytes(Method::interpreter_invocation_counter_offset()));
 412     __ ld(interpreter_invocation_counter, G3_scratch);
 413     __ inc(G3_scratch);
 414     __ st(G3_scratch, interpreter_invocation_counter);
 415   }
 416 
 417   Address invocation_limit(G3_scratch, (address)&InvocationCounter::InterpreterInvocationLimit);
 418   __ sethi(invocation_limit);
 419   __ ld(invocation_limit, G3_scratch);
 420   __ cmp(O0, G3_scratch);
 421   __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
 422   __ delayed()->nop();
 423 
 424 }
 425 
 426 address InterpreterGenerator::generate_empty_entry(void) {
 427 
 428   // A method that does nothing but return...
 429 
 430   address entry = __ pc();
 431   Label slow_path;
 432 
 433   // do nothing for empty methods (do not even increment invocation counter)
 434   if ( UseFastEmptyMethods) {
 435     // If we need a safepoint check, generate full interpreter entry.
 436     Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
 437     __ load_contents(sync_state, G3_scratch);
 438     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 439     __ br(Assembler::notEqual, false, Assembler::pn, frame_manager_entry);
 440     __ delayed()->nop();
 441 
 442     // Code: _return
 443     __ retl();
 444     __ delayed()->mov(O5_savedSP, SP);
 445     return entry;
 446   }
 447   return NULL;
 448 }
 449 
 450 // Call an accessor method (assuming it is resolved, otherwise drop into
 451 // vanilla (slow path) entry
 452 
 453 // Generates code to elide accessor methods
 454 // Uses G3_scratch and G1_scratch as scratch
 455 address InterpreterGenerator::generate_accessor_entry(void) {
 456 
 457   // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
 458   // parameter size = 1
 459   // Note: We can only use this code if the getfield has been resolved
 460   //       and if we don't have a null-pointer exception => check for
 461   //       these conditions first and use slow path if necessary.
 462   address entry = __ pc();
 463   Label slow_path;
 464 
 465   if ( UseFastAccessorMethods) {
 466     // Check if we need to reach a safepoint and generate full interpreter
 467     // frame if so.
 468     Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
 469     __ load_contents(sync_state, G3_scratch);
 470     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 471     __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
 472     __ delayed()->nop();
 473 
 474     // Check if local 0 != NULL
 475     __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
 476     __ tst(Otos_i);  // check if local 0 == NULL and go the slow path
 477     __ brx(Assembler::zero, false, Assembler::pn, slow_path);
 478     __ delayed()->nop();
 479 
 480 
 481     // read first instruction word and extract bytecode @ 1 and index @ 2
 482     // get first 4 bytes of the bytecodes (big endian!)
 483     __ ld_ptr(Address(G5_method, 0, in_bytes(Method::const_offset())), G1_scratch);
 484     __ ld(Address(G1_scratch, 0, in_bytes(ConstMethod::codes_offset())), G1_scratch);
 485 
 486     // move index @ 2 far left then to the right most two bytes.
 487     __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
 488     __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
 489                       ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
 490 
 491     // get constant pool cache
 492     __ ld_ptr(G5_method, in_bytes(Method::const_offset()), G3_scratch);
 493     __ ld_ptr(G3_scratch, in_bytes(ConstMethod::constants_offset()), G3_scratch);
 494     __ ld_ptr(G3_scratch, ConstantPool::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 = ConstantPoolCache::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::tos_state_shift, G1_scratch);
 518     // Make sure we don't need to mask G1_scratch after the above shift
 519     ConstantPoolCacheEntry::verify_tos_state_shift();
 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(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 #if INCLUDE_ALL_GCS
 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 // INCLUDE_ALL_GCS
 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 Register RconstMethod = Gtmp1;
 587   const Address constMethod(G5_method, 0, in_bytes(Method::const_offset()));
 588   const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
 589 
 590   bool inc_counter  = UseCompiler || CountCompiledCalls;
 591 
 592   // make sure registers are different!
 593   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
 594 
 595   const Address access_flags      (G5_method, 0, in_bytes(Method::access_flags_offset()));
 596 
 597   Label Lentry;
 598   __ bind(Lentry);
 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   __ ld_ptr(constMethod, RconstMethod);
 625   __ lduh(size_of_parameters, Gtmp1);
 626   __ sll(Gtmp1, LogBytesPerWord, Gtmp2);       // parameter size in bytes
 627   __ add(Gargs, Gtmp2, Gargs);                 // points to first local + BytesPerWord
 628   // NEW
 629   __ add(Gargs, -wordSize, Gargs);             // points to first local[0]
 630   // generate the code to allocate the interpreter stack frame
 631   // NEW FRAME ALLOCATED HERE
 632   // save callers original sp
 633   // __ mov(SP, I5_savedSP->after_restore());
 634 
 635   generate_compute_interpreter_state(Lstate, G0, true);
 636 
 637   // At this point Lstate points to new interpreter state
 638   //
 639 
 640   const Address do_not_unlock_if_synchronized(G2_thread, 0,
 641       in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
 642   // Since at this point in the method invocation the exception handler
 643   // would try to exit the monitor of synchronized methods which hasn't
 644   // been entered yet, we set the thread local variable
 645   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
 646   // runtime, exception handling i.e. unlock_if_synchronized_method will
 647   // check this thread local flag.
 648   // This flag has two effects, one is to force an unwind in the topmost
 649   // interpreter frame and not perform an unlock while doing so.
 650 
 651   __ movbool(true, G3_scratch);
 652   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
 653 
 654 
 655   // increment invocation counter and check for overflow
 656   //
 657   // Note: checking for negative value instead of overflow
 658   //       so we have a 'sticky' overflow test (may be of
 659   //       importance as soon as we have true MT/MP)
 660   Label invocation_counter_overflow;
 661   if (inc_counter) {
 662     generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
 663   }
 664   Label Lcontinue;
 665   __ bind(Lcontinue);
 666 
 667   bang_stack_shadow_pages(true);
 668   // reset the _do_not_unlock_if_synchronized flag
 669   __ stbool(G0, do_not_unlock_if_synchronized);
 670 
 671   // check for synchronized methods
 672   // Must happen AFTER invocation_counter check, so method is not locked
 673   // if counter overflows.
 674 
 675   if (synchronized) {
 676     lock_method();
 677     // Don't see how G2_thread is preserved here...
 678     // __ verify_thread(); QQQ destroys L0,L1 can't use
 679   } else {
 680 #ifdef ASSERT
 681     { Label ok;
 682       __ ld_ptr(STATE(_method), G5_method);
 683       __ ld(access_flags, O0);
 684       __ btst(JVM_ACC_SYNCHRONIZED, O0);
 685       __ br( Assembler::zero, false, Assembler::pt, ok);
 686       __ delayed()->nop();
 687       __ stop("method needs synchronization");
 688       __ bind(ok);
 689     }
 690 #endif // ASSERT
 691   }
 692 
 693   // start execution
 694 
 695 //   __ verify_thread(); kills L1,L2 can't  use at the moment
 696 
 697   // jvmti/jvmpi support
 698   __ notify_method_entry();
 699 
 700   // native call
 701 
 702   // (note that O0 is never an oop--at most it is a handle)
 703   // It is important not to smash any handles created by this call,
 704   // until any oop handle in O0 is dereferenced.
 705 
 706   // (note that the space for outgoing params is preallocated)
 707 
 708   // get signature handler
 709 
 710   Label pending_exception_present;
 711 
 712   { Label L;
 713     __ ld_ptr(STATE(_method), G5_method);
 714     __ ld_ptr(Address(G5_method, 0, in_bytes(Method::signature_handler_offset())), G3_scratch);
 715     __ tst(G3_scratch);
 716     __ brx(Assembler::notZero, false, Assembler::pt, L);
 717     __ delayed()->nop();
 718     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), G5_method, false);
 719     __ ld_ptr(STATE(_method), G5_method);
 720 
 721     Address exception_addr(G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
 722     __ ld_ptr(exception_addr, G3_scratch);
 723     __ br_notnull_short(G3_scratch, Assembler::pn, pending_exception_present);
 724     __ ld_ptr(Address(G5_method, 0, in_bytes(Method::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(Method::native_function_offset())), O0);
 769     // for static methods insert the mirror argument
 770     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
 771 
 772     __ ld_ptr(Address(G5_method, 0, in_bytes(Method:: const_offset())), O1);
 773     __ ld_ptr(Address(O1, 0, in_bytes(ConstMethod::constants_offset())), O1);
 774     __ ld_ptr(Address(O1, 0, ConstantPool::pool_holder_offset_in_bytes()), O1);
 775     __ ld_ptr(O1, mirror_offset, O1);
 776     // where the mirror handle body is allocated:
 777 #ifdef ASSERT
 778     if (!PrintSignatureHandlers)  // do not dirty the output with this
 779     { Label L;
 780       __ tst(O1);
 781       __ brx(Assembler::notZero, false, Assembler::pt, L);
 782       __ delayed()->nop();
 783       __ stop("mirror is missing");
 784       __ bind(L);
 785     }
 786 #endif // ASSERT
 787     __ st_ptr(O1, STATE(_oop_temp));
 788     __ add(STATE(_oop_temp), O1);            // this is really an LEA not an add
 789     __ bind(not_static);
 790   }
 791 
 792   // At this point, arguments have been copied off of stack into
 793   // their JNI positions, which are O1..O5 and SP[68..].
 794   // Oops are boxed in-place on the stack, with handles copied to arguments.
 795   // The result handler is in Lscratch.  O0 will shortly hold the JNIEnv*.
 796 
 797 #ifdef ASSERT
 798   { Label L;
 799     __ tst(O0);
 800     __ brx(Assembler::notZero, false, Assembler::pt, L);
 801     __ delayed()->nop();
 802     __ stop("native entry point is missing");
 803     __ bind(L);
 804   }
 805 #endif // ASSERT
 806 
 807   //
 808   // setup the java frame anchor
 809   //
 810   // The scavenge function only needs to know that the PC of this frame is
 811   // in the interpreter method entry code, it doesn't need to know the exact
 812   // PC and hence we can use O7 which points to the return address from the
 813   // previous call in the code stream (signature handler function)
 814   //
 815   // The other trick is we set last_Java_sp to FP instead of the usual SP because
 816   // we have pushed the extra frame in order to protect the volatile register(s)
 817   // in that frame when we return from the jni call
 818   //
 819 
 820 
 821   __ set_last_Java_frame(FP, O7);
 822   __ mov(O7, I7);  // make dummy interpreter frame look like one above,
 823                    // not meaningless information that'll confuse me.
 824 
 825   // flush the windows now. We don't care about the current (protection) frame
 826   // only the outer frames
 827 
 828   __ flush_windows();
 829 
 830   // mark windows as flushed
 831   Address flags(G2_thread,
 832                 0,
 833                 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
 834   __ set(JavaFrameAnchor::flushed, G3_scratch);
 835   __ st(G3_scratch, flags);
 836 
 837   // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
 838 
 839   Address thread_state(G2_thread, 0, in_bytes(JavaThread::thread_state_offset()));
 840 #ifdef ASSERT
 841   { Label L;
 842     __ ld(thread_state, G3_scratch);
 843     __ cmp(G3_scratch, _thread_in_Java);
 844     __ br(Assembler::equal, false, Assembler::pt, L);
 845     __ delayed()->nop();
 846     __ stop("Wrong thread state in native stub");
 847     __ bind(L);
 848   }
 849 #endif // ASSERT
 850   __ set(_thread_in_native, G3_scratch);
 851   __ st(G3_scratch, thread_state);
 852 
 853   // Call the jni method, using the delay slot to set the JNIEnv* argument.
 854   __ callr(O0, 0);
 855   __ delayed()->
 856      add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
 857   __ ld_ptr(STATE(_thread), G2_thread);  // restore thread
 858 
 859   // must we block?
 860 
 861   // Block, if necessary, before resuming in _thread_in_Java state.
 862   // In order for GC to work, don't clear the last_Java_sp until after blocking.
 863   { Label no_block;
 864     Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
 865 
 866     // Switch thread to "native transition" state before reading the synchronization state.
 867     // This additional state is necessary because reading and testing the synchronization
 868     // state is not atomic w.r.t. GC, as this scenario demonstrates:
 869     //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
 870     //     VM thread changes sync state to synchronizing and suspends threads for GC.
 871     //     Thread A is resumed to finish this native method, but doesn't block here since it
 872     //     didn't see any synchronization is progress, and escapes.
 873     __ set(_thread_in_native_trans, G3_scratch);
 874     __ st(G3_scratch, thread_state);
 875     if(os::is_MP()) {
 876       // Write serialization page so VM thread can do a pseudo remote membar.
 877       // We use the current thread pointer to calculate a thread specific
 878       // offset to write to within the page. This minimizes bus traffic
 879       // due to cache line collision.
 880       __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
 881     }
 882     __ load_contents(sync_state, G3_scratch);
 883     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 884 
 885 
 886     Label L;
 887     Address suspend_state(G2_thread, 0, in_bytes(JavaThread::suspend_flags_offset()));
 888     __ br(Assembler::notEqual, false, Assembler::pn, L);
 889     __ delayed()->
 890       ld(suspend_state, G3_scratch);
 891     __ cmp(G3_scratch, 0);
 892     __ br(Assembler::equal, false, Assembler::pt, no_block);
 893     __ delayed()->nop();
 894     __ bind(L);
 895 
 896     // Block.  Save any potential method result value before the operation and
 897     // use a leaf call to leave the last_Java_frame setup undisturbed.
 898     save_native_result();
 899     __ call_VM_leaf(noreg,
 900                     CAST_FROM_FN_PTR(address, JavaThread::check_safepoint_and_suspend_for_native_trans),
 901                     G2_thread);
 902     __ ld_ptr(STATE(_thread), G2_thread);  // restore thread
 903     // Restore any method result value
 904     restore_native_result();
 905     __ bind(no_block);
 906   }
 907 
 908   // Clear the frame anchor now
 909 
 910   __ reset_last_Java_frame();
 911 
 912   // Move the result handler address
 913   __ mov(Lscratch, G3_scratch);
 914   // return possible result to the outer frame
 915 #ifndef __LP64
 916   __ mov(O0, I0);
 917   __ restore(O1, G0, O1);
 918 #else
 919   __ restore(O0, G0, O0);
 920 #endif /* __LP64 */
 921 
 922   // Move result handler to expected register
 923   __ mov(G3_scratch, Lscratch);
 924 
 925 
 926   // thread state is thread_in_native_trans. Any safepoint blocking has
 927   // happened in the trampoline we are ready to switch to thread_in_Java.
 928 
 929   __ set(_thread_in_Java, G3_scratch);
 930   __ st(G3_scratch, thread_state);
 931 
 932   // If we have an oop result store it where it will be safe for any further gc
 933   // until we return now that we've released the handle it might be protected by
 934 
 935   {
 936     Label no_oop, store_result;
 937 
 938     __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
 939     __ cmp(G3_scratch, Lscratch);
 940     __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
 941     __ delayed()->nop();
 942     __ addcc(G0, O0, O0);
 943     __ brx(Assembler::notZero, true, Assembler::pt, store_result);     // if result is not NULL:
 944     __ delayed()->ld_ptr(O0, 0, O0);                                   // unbox it
 945     __ mov(G0, O0);
 946 
 947     __ bind(store_result);
 948     // Store it where gc will look for it and result handler expects it.
 949     __ st_ptr(O0, STATE(_oop_temp));
 950 
 951     __ bind(no_oop);
 952 
 953   }
 954 
 955   // reset handle block
 956   __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch);
 957   __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
 958 
 959 
 960   // handle exceptions (exception handling will handle unlocking!)
 961   { Label L;
 962     Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
 963 
 964     __ ld_ptr(exception_addr, Gtemp);
 965     __ tst(Gtemp);
 966     __ brx(Assembler::equal, false, Assembler::pt, L);
 967     __ delayed()->nop();
 968     __ bind(pending_exception_present);
 969     // With c++ interpreter we just leave it pending caller will do the correct thing. However...
 970     // Like x86 we ignore the result of the native call and leave the method locked. This
 971     // seems wrong to leave things locked.
 972 
 973     __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
 974     __ delayed()->restore(I5_savedSP, G0, SP);  // remove interpreter frame
 975 
 976     __ bind(L);
 977   }
 978 
 979   // jvmdi/jvmpi support (preserves thread register)
 980   __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
 981 
 982   if (synchronized) {
 983     // save and restore any potential method result value around the unlocking operation
 984     save_native_result();
 985 
 986     const int entry_size            = frame::interpreter_frame_monitor_size() * wordSize;
 987     // Get the initial monitor we allocated
 988     __ sub(Lstate, entry_size, O1);                        // initial monitor
 989     __ unlock_object(O1);
 990     restore_native_result();
 991   }
 992 
 993 #if defined(COMPILER2) && !defined(_LP64)
 994 
 995   // C2 expects long results in G1 we can't tell if we're returning to interpreted
 996   // or compiled so just be safe.
 997 
 998   __ sllx(O0, 32, G1);          // Shift bits into high G1
 999   __ srl (O1, 0, O1);           // Zero extend O1
1000   __ or3 (O1, G1, G1);          // OR 64 bits into G1
1001 
1002 #endif /* COMPILER2 && !_LP64 */
1003 
1004 #ifdef ASSERT
1005   {
1006     Label ok;
1007     __ cmp(I5_savedSP, FP);
1008     __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
1009     __ delayed()->nop();
1010     __ stop("bad I5_savedSP value");
1011     __ should_not_reach_here();
1012     __ bind(ok);
1013   }
1014 #endif
1015   // Calls result handler which POPS FRAME
1016   if (TraceJumps) {
1017     // Move target to register that is recordable
1018     __ mov(Lscratch, G3_scratch);
1019     __ JMP(G3_scratch, 0);
1020   } else {
1021     __ jmp(Lscratch, 0);
1022   }
1023   __ delayed()->nop();
1024 
1025   if (inc_counter) {
1026     // handle invocation counter overflow
1027     __ bind(invocation_counter_overflow);
1028     generate_counter_overflow(Lcontinue);
1029   }
1030 
1031 
1032   return entry;
1033 }
1034 
1035 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state,
1036                                                               const Register prev_state,
1037                                                               bool native) {
1038 
1039   // On entry
1040   // G5_method - caller's method
1041   // Gargs - points to initial parameters (i.e. locals[0])
1042   // G2_thread - valid? (C1 only??)
1043   // "prev_state" - contains any previous frame manager state which we must save a link
1044   //
1045   // On return
1046   // "state" is a pointer to the newly allocated  state object. We must allocate and initialize
1047   // a new interpretState object and the method expression stack.
1048 
1049   assert_different_registers(state, prev_state);
1050   assert_different_registers(prev_state, G3_scratch);
1051   const Register Gtmp = G3_scratch;
1052   const Address constMethod       (G5_method, 0, in_bytes(Method::const_offset()));
1053   const Address access_flags      (G5_method, 0, in_bytes(Method::access_flags_offset()));
1054 
1055   // slop factor is two extra slots on the expression stack so that
1056   // we always have room to store a result when returning from a call without parameters
1057   // that returns a result.
1058 
1059   const int slop_factor = 2*wordSize;
1060 
1061   const int fixed_size = ((sizeof(BytecodeInterpreter) + slop_factor) >> LogBytesPerWord) + // what is the slop factor?
1062                          frame::memory_parameter_word_sp_offset +  // register save area + param window
1063                          (native ?  frame::interpreter_frame_extra_outgoing_argument_words : 0); // JNI, class
1064 
1065   // XXX G5_method valid
1066 
1067   // Now compute new frame size
1068 
1069   if (native) {
1070     const Register RconstMethod = Gtmp;
1071     const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
1072     __ ld_ptr(constMethod, RconstMethod);
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     // Full size expression stack
1077     __ ld_ptr(constMethod, Gtmp);
1078     __ lduh(Gtmp, in_bytes(ConstMethod::max_stack_offset()), Gtmp);
1079   }
1080   __ add(Gtmp, fixed_size, Gtmp);           // plus the fixed portion
1081 
1082   __ neg(Gtmp);                               // negative space for stack/parameters in words
1083   __ and3(Gtmp, -WordsPerLong, Gtmp);        // make multiple of 2 (SP must be 2-word aligned)
1084   __ sll(Gtmp, LogBytesPerWord, Gtmp);       // negative space for frame in bytes
1085 
1086   // Need to do stack size check here before we fault on large frames
1087 
1088   Label stack_ok;
1089 
1090   const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages :
1091                                                                               (StackRedPages+StackYellowPages);
1092 
1093 
1094   __ ld_ptr(G2_thread, in_bytes(Thread::stack_base_offset()), O0);
1095   __ ld_ptr(G2_thread, in_bytes(Thread::stack_size_offset()), O1);
1096   // compute stack bottom
1097   __ sub(O0, O1, O0);
1098 
1099   // Avoid touching the guard pages
1100   // Also a fudge for frame size of BytecodeInterpreter::run
1101   // It varies from 1k->4k depending on build type
1102   const int fudge = 6 * K;
1103 
1104   __ set(fudge + (max_pages * os::vm_page_size()), O1);
1105 
1106   __ add(O0, O1, O0);
1107   __ sub(O0, Gtmp, O0);
1108   __ cmp(SP, O0);
1109   __ brx(Assembler::greaterUnsigned, false, Assembler::pt, stack_ok);
1110   __ delayed()->nop();
1111 
1112      // throw exception return address becomes throwing pc
1113 
1114   __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
1115   __ stop("never reached");
1116 
1117   __ bind(stack_ok);
1118 
1119   __ save(SP, Gtmp, SP);                      // setup new frame and register window
1120 
1121   // New window I7 call_stub or previous activation
1122   // O6 - register save area, BytecodeInterpreter just below it, args/locals just above that
1123   //
1124   __ sub(FP, sizeof(BytecodeInterpreter), state);        // Point to new Interpreter state
1125   __ add(state, STACK_BIAS, state );         // Account for 64bit bias
1126 
1127 #define XXX_STATE(field_name) state, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
1128 
1129   // Initialize a new Interpreter state
1130   // orig_sp - caller's original sp
1131   // G2_thread - thread
1132   // Gargs - &locals[0] (unbiased?)
1133   // G5_method - method
1134   // SP (biased) - accounts for full size java stack, BytecodeInterpreter object, register save area, and register parameter save window
1135 
1136 
1137   __ set(0xdead0004, O1);
1138 
1139 
1140   __ st_ptr(Gargs, XXX_STATE(_locals));
1141   __ st_ptr(G0, XXX_STATE(_oop_temp));
1142 
1143   __ st_ptr(state, XXX_STATE(_self_link));                // point to self
1144   __ st_ptr(prev_state->after_save(), XXX_STATE(_prev_link)); // Chain interpreter states
1145   __ st_ptr(G2_thread, XXX_STATE(_thread));               // Store javathread
1146 
1147   if (native) {
1148     __ st_ptr(G0, XXX_STATE(_bcp));
1149   } else {
1150     __ ld_ptr(G5_method, in_bytes(Method::const_offset()), O2); // get ConstMethod*
1151     __ add(O2, in_bytes(ConstMethod::codes_offset()), O2);        // get bcp
1152     __ st_ptr(O2, XXX_STATE(_bcp));
1153   }
1154 
1155   __ st_ptr(G0, XXX_STATE(_mdx));
1156   __ st_ptr(G5_method, XXX_STATE(_method));
1157 
1158   __ set((int) BytecodeInterpreter::method_entry, O1);
1159   __ st(O1, XXX_STATE(_msg));
1160 
1161   __ ld_ptr(constMethod, O3);
1162   __ ld_ptr(O3, in_bytes(ConstMethod::constants_offset()), O3);
1163   __ ld_ptr(O3, ConstantPool::cache_offset_in_bytes(), O2);
1164   __ st_ptr(O2, XXX_STATE(_constants));
1165 
1166   __ st_ptr(G0, XXX_STATE(_result._to_call._callee));
1167 
1168   // Monitor base is just start of BytecodeInterpreter object;
1169   __ mov(state, O2);
1170   __ st_ptr(O2, XXX_STATE(_monitor_base));
1171 
1172   // Do we need a monitor for synchonized method?
1173   {
1174     __ ld(access_flags, O1);
1175     Label done;
1176     Label got_obj;
1177     __ btst(JVM_ACC_SYNCHRONIZED, O1);
1178     __ br( Assembler::zero, false, Assembler::pt, done);
1179 
1180     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
1181     __ delayed()->btst(JVM_ACC_STATIC, O1);
1182     __ ld_ptr(XXX_STATE(_locals), O1);
1183     __ br( Assembler::zero, true, Assembler::pt, got_obj);
1184     __ delayed()->ld_ptr(O1, 0, O1);                  // get receiver for not-static case
1185     __ ld_ptr(constMethod, O1);
1186     __ ld_ptr( O1, in_bytes(ConstMethod::constants_offset()), O1);
1187     __ ld_ptr( O1, ConstantPool::pool_holder_offset_in_bytes(), O1);
1188     // lock the mirror, not the Klass*
1189     __ ld_ptr( O1, mirror_offset, O1);
1190 
1191     __ bind(got_obj);
1192 
1193   #ifdef ASSERT
1194     __ tst(O1);
1195     __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
1196   #endif // ASSERT
1197 
1198     const int entry_size            = frame::interpreter_frame_monitor_size() * wordSize;
1199     __ sub(SP, entry_size, SP);                         // account for initial monitor
1200     __ sub(O2, entry_size, O2);                        // initial monitor
1201     __ st_ptr(O1, O2, BasicObjectLock::obj_offset_in_bytes()); // and allocate it for interpreter use
1202     __ bind(done);
1203   }
1204 
1205   // Remember initial frame bottom
1206 
1207   __ st_ptr(SP, XXX_STATE(_frame_bottom));
1208 
1209   __ st_ptr(O2, XXX_STATE(_stack_base));
1210 
1211   __ sub(O2, wordSize, O2);                    // prepush
1212   __ st_ptr(O2, XXX_STATE(_stack));                // PREPUSH
1213 
1214   // Full size expression stack
1215   __ ld_ptr(constMethod, O3);
1216   __ lduh(O3, in_bytes(ConstMethod::max_stack_offset()), O3);
1217   guarantee(!EnableInvokeDynamic, "no support yet for java.lang.invoke.MethodHandle"); //6815692
1218   //6815692//if (EnableInvokeDynamic)
1219   //6815692//  __ inc(O3, Method::extra_stack_entries());
1220   __ sll(O3, LogBytesPerWord, O3);
1221   __ sub(O2, O3, O3);
1222 //  __ sub(O3, wordSize, O3);                    // so prepush doesn't look out of bounds
1223   __ st_ptr(O3, XXX_STATE(_stack_limit));
1224 
1225   if (!native) {
1226     //
1227     // Code to initialize locals
1228     //
1229     Register init_value = noreg;    // will be G0 if we must clear locals
1230     // Now zero locals
1231     if (true /* zerolocals */ || ClearInterpreterLocals) {
1232       // explicitly initialize locals
1233       init_value = G0;
1234     } else {
1235     #ifdef ASSERT
1236       // initialize locals to a garbage pattern for better debugging
1237       init_value = O3;
1238       __ set( 0x0F0F0F0F, init_value );
1239     #endif // ASSERT
1240     }
1241     if (init_value != noreg) {
1242       Label clear_loop;
1243       const Register RconstMethod = O1;
1244       const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
1245       const Address size_of_locals    (RconstMethod, 0, in_bytes(ConstMethod::size_of_locals_offset()));
1246 
1247       // NOTE: If you change the frame layout, this code will need to
1248       // be updated!
1249       __ ld_ptr( constMethod, RconstMethod );
1250       __ lduh( size_of_locals, O2 );
1251       __ lduh( size_of_parameters, O1 );
1252       __ sll( O2, LogBytesPerWord, O2);
1253       __ sll( O1, LogBytesPerWord, O1 );
1254       __ ld_ptr(XXX_STATE(_locals), L2_scratch);
1255       __ sub( L2_scratch, O2, O2 );
1256       __ sub( L2_scratch, O1, O1 );
1257 
1258       __ bind( clear_loop );
1259       __ inc( O2, wordSize );
1260 
1261       __ cmp( O2, O1 );
1262       __ br( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1263       __ delayed()->st_ptr( init_value, O2, 0 );
1264     }
1265   }
1266 }
1267 // Find preallocated  monitor and lock method (C++ interpreter)
1268 //
1269 void InterpreterGenerator::lock_method(void) {
1270 // Lock the current method.
1271 // Destroys registers L2_scratch, L3_scratch, O0
1272 //
1273 // Find everything relative to Lstate
1274 
1275 #ifdef ASSERT
1276   __ ld_ptr(STATE(_method), L2_scratch);
1277   __ ld(L2_scratch, in_bytes(Method::access_flags_offset()), O0);
1278 
1279  { Label ok;
1280    __ btst(JVM_ACC_SYNCHRONIZED, O0);
1281    __ br( Assembler::notZero, false, Assembler::pt, ok);
1282    __ delayed()->nop();
1283    __ stop("method doesn't need synchronization");
1284    __ bind(ok);
1285   }
1286 #endif // ASSERT
1287 
1288   // monitor is already allocated at stack base
1289   // and the lockee is already present
1290   __ ld_ptr(STATE(_stack_base), L2_scratch);
1291   __ ld_ptr(L2_scratch, BasicObjectLock::obj_offset_in_bytes(), O0);   // get object
1292   __ lock_object(L2_scratch, O0);
1293 
1294 }
1295 
1296 //  Generate code for handling resuming a deopted method
1297 void CppInterpreterGenerator::generate_deopt_handling() {
1298 
1299   Label return_from_deopt_common;
1300 
1301   // deopt needs to jump to here to enter the interpreter (return a result)
1302   deopt_frame_manager_return_atos  = __ pc();
1303 
1304   // O0/O1 live
1305   __ ba(return_from_deopt_common);
1306   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_OBJECT), L3_scratch);    // Result stub address array index
1307 
1308 
1309   // deopt needs to jump to here to enter the interpreter (return a result)
1310   deopt_frame_manager_return_btos  = __ pc();
1311 
1312   // O0/O1 live
1313   __ ba(return_from_deopt_common);
1314   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_BOOLEAN), L3_scratch);    // Result stub address array index
1315 
1316   // deopt needs to jump to here to enter the interpreter (return a result)
1317   deopt_frame_manager_return_itos  = __ pc();
1318 
1319   // O0/O1 live
1320   __ ba(return_from_deopt_common);
1321   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_INT), L3_scratch);    // Result stub address array index
1322 
1323   // deopt needs to jump to here to enter the interpreter (return a result)
1324 
1325   deopt_frame_manager_return_ltos  = __ pc();
1326 #if !defined(_LP64) && defined(COMPILER2)
1327   // All return values are where we want them, except for Longs.  C2 returns
1328   // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
1329   // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
1330   // build even if we are returning from interpreted we just do a little
1331   // stupid shuffing.
1332   // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
1333   // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
1334   // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
1335 
1336   __ srl (G1, 0,O1);
1337   __ srlx(G1,32,O0);
1338 #endif /* !_LP64 && COMPILER2 */
1339   // O0/O1 live
1340   __ ba(return_from_deopt_common);
1341   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_LONG), L3_scratch);    // Result stub address array index
1342 
1343   // deopt needs to jump to here to enter the interpreter (return a result)
1344 
1345   deopt_frame_manager_return_ftos  = __ pc();
1346   // O0/O1 live
1347   __ ba(return_from_deopt_common);
1348   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_FLOAT), L3_scratch);    // Result stub address array index
1349 
1350   // deopt needs to jump to here to enter the interpreter (return a result)
1351   deopt_frame_manager_return_dtos  = __ pc();
1352 
1353   // O0/O1 live
1354   __ ba(return_from_deopt_common);
1355   __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_DOUBLE), L3_scratch);    // Result stub address array index
1356 
1357   // deopt needs to jump to here to enter the interpreter (return a result)
1358   deopt_frame_manager_return_vtos  = __ pc();
1359 
1360   // O0/O1 live
1361   __ set(AbstractInterpreter::BasicType_as_index(T_VOID), L3_scratch);
1362 
1363   // Deopt return common
1364   // an index is present that lets us move any possible result being
1365   // return to the interpreter's stack
1366   //
1367   __ bind(return_from_deopt_common);
1368 
1369   // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1370   // stack is in the state that the  calling convention left it.
1371   // Copy the result from native abi result and place it on java expression stack.
1372 
1373   // Current interpreter state is present in Lstate
1374 
1375   // Get current pre-pushed top of interpreter stack
1376   // Any result (if any) is in native abi
1377   // result type index is in L3_scratch
1378 
1379   __ ld_ptr(STATE(_stack), L1_scratch);                                          // get top of java expr stack
1380 
1381   __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1382   __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1383   __ ld_ptr(L4_scratch, L3_scratch, Lscratch);                                       // get typed result converter address
1384   __ jmpl(Lscratch, G0, O7);                                         // and convert it
1385   __ delayed()->nop();
1386 
1387   // L1_scratch points to top of stack (prepushed)
1388   __ st_ptr(L1_scratch, STATE(_stack));
1389 }
1390 
1391 // Generate the code to handle a more_monitors message from the c++ interpreter
1392 void CppInterpreterGenerator::generate_more_monitors() {
1393 
1394   Label entry, loop;
1395   const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1396   // 1. compute new pointers                                // esp: old expression stack top
1397   __ delayed()->ld_ptr(STATE(_stack_base), L4_scratch);            // current expression stack bottom
1398   __ sub(L4_scratch, entry_size, L4_scratch);
1399   __ st_ptr(L4_scratch, STATE(_stack_base));
1400 
1401   __ sub(SP, entry_size, SP);                  // Grow stack
1402   __ st_ptr(SP, STATE(_frame_bottom));
1403 
1404   __ ld_ptr(STATE(_stack_limit), L2_scratch);
1405   __ sub(L2_scratch, entry_size, L2_scratch);
1406   __ st_ptr(L2_scratch, STATE(_stack_limit));
1407 
1408   __ ld_ptr(STATE(_stack), L1_scratch);                // Get current stack top
1409   __ sub(L1_scratch, entry_size, L1_scratch);
1410   __ st_ptr(L1_scratch, STATE(_stack));
1411   __ ba(entry);
1412   __ delayed()->add(L1_scratch, wordSize, L1_scratch);        // first real entry (undo prepush)
1413 
1414   // 2. move expression stack
1415 
1416   __ bind(loop);
1417   __ st_ptr(L3_scratch, Address(L1_scratch, 0));
1418   __ add(L1_scratch, wordSize, L1_scratch);
1419   __ bind(entry);
1420   __ cmp(L1_scratch, L4_scratch);
1421   __ br(Assembler::notEqual, false, Assembler::pt, loop);
1422   __ delayed()->ld_ptr(L1_scratch, entry_size, L3_scratch);
1423 
1424   // now zero the slot so we can find it.
1425   __ st_ptr(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes());
1426 
1427 }
1428 
1429 // Initial entry to C++ interpreter from the call_stub.
1430 // This entry point is called the frame manager since it handles the generation
1431 // of interpreter activation frames via requests directly from the vm (via call_stub)
1432 // and via requests from the interpreter. The requests from the call_stub happen
1433 // directly thru the entry point. Requests from the interpreter happen via returning
1434 // from the interpreter and examining the message the interpreter has returned to
1435 // the frame manager. The frame manager can take the following requests:
1436 
1437 // NO_REQUEST - error, should never happen.
1438 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and
1439 //                 allocate a new monitor.
1440 // CALL_METHOD - setup a new activation to call a new method. Very similar to what
1441 //               happens during entry during the entry via the call stub.
1442 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub.
1443 //
1444 // Arguments:
1445 //
1446 // ebx: Method*
1447 // ecx: receiver - unused (retrieved from stack as needed)
1448 // esi: previous frame manager state (NULL from the call_stub/c1/c2)
1449 //
1450 //
1451 // Stack layout at entry
1452 //
1453 // [ return address     ] <--- esp
1454 // [ parameter n        ]
1455 //   ...
1456 // [ parameter 1        ]
1457 // [ expression stack   ]
1458 //
1459 //
1460 // We are free to blow any registers we like because the call_stub which brought us here
1461 // initially has preserved the callee save registers already.
1462 //
1463 //
1464 
1465 static address interpreter_frame_manager = NULL;
1466 
1467 #ifdef ASSERT
1468   #define VALIDATE_STATE(scratch, marker)                         \
1469   {                                                               \
1470     Label skip;                                                   \
1471     __ ld_ptr(STATE(_self_link), scratch);                        \
1472     __ cmp(Lstate, scratch);                                      \
1473     __ brx(Assembler::equal, false, Assembler::pt, skip);         \
1474     __ delayed()->nop();                                          \
1475     __ breakpoint_trap();                                         \
1476     __ emit_int32(marker);                                         \
1477     __ bind(skip);                                                \
1478   }
1479 #else
1480   #define VALIDATE_STATE(scratch, marker)
1481 #endif /* ASSERT */
1482 
1483 void CppInterpreterGenerator::adjust_callers_stack(Register args) {
1484 //
1485 // Adjust caller's stack so that all the locals can be contiguous with
1486 // the parameters.
1487 // Worries about stack overflow make this a pain.
1488 //
1489 // Destroys args, G3_scratch, G3_scratch
1490 // In/Out O5_savedSP (sender's original SP)
1491 //
1492 //  assert_different_registers(state, prev_state);
1493   const Register Gtmp = G3_scratch;
1494   const RconstMethod = G3_scratch;
1495   const Register tmp = O2;
1496   const Address constMethod(G5_method, 0, in_bytes(Method::const_offset()));
1497   const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
1498   const Address size_of_locals    (RconstMethod, 0, in_bytes(ConstMethod::size_of_locals_offset()));
1499 
1500   __ ld_ptr(constMethod, RconstMethod);
1501   __ lduh(size_of_parameters, tmp);
1502   __ sll(tmp, LogBytesPerWord, Gargs);       // parameter size in bytes
1503   __ add(args, Gargs, Gargs);                // points to first local + BytesPerWord
1504   // NEW
1505   __ add(Gargs, -wordSize, Gargs);             // points to first local[0]
1506   // determine extra space for non-argument locals & adjust caller's SP
1507   // Gtmp1: parameter size in words
1508   __ lduh(size_of_locals, Gtmp);
1509   __ compute_extra_locals_size_in_bytes(tmp, Gtmp, Gtmp);
1510 
1511 #if 1
1512   // c2i adapters place the final interpreter argument in the register save area for O0/I0
1513   // the call_stub will place the final interpreter argument at
1514   // frame::memory_parameter_word_sp_offset. This is mostly not noticable for either asm
1515   // or c++ interpreter. However with the c++ interpreter when we do a recursive call
1516   // and try to make it look good in the debugger we will store the argument to
1517   // RecursiveInterpreterActivation in the register argument save area. Without allocating
1518   // extra space for the compiler this will overwrite locals in the local array of the
1519   // interpreter.
1520   // QQQ still needed with frameless adapters???
1521 
1522   const int c2i_adjust_words = frame::memory_parameter_word_sp_offset - frame::callee_register_argument_save_area_sp_offset;
1523 
1524   __ add(Gtmp, c2i_adjust_words*wordSize, Gtmp);
1525 #endif // 1
1526 
1527 
1528   __ sub(SP, Gtmp, SP);                      // just caller's frame for the additional space we need.
1529 }
1530 
1531 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1532 
1533   // G5_method: Method*
1534   // G2_thread: thread (unused)
1535   // Gargs:   bottom of args (sender_sp)
1536   // O5: sender's sp
1537 
1538   // A single frame manager is plenty as we don't specialize for synchronized. We could and
1539   // the code is pretty much ready. Would need to change the test below and for good measure
1540   // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized
1541   // routines. Not clear this is worth it yet.
1542 
1543   if (interpreter_frame_manager) {
1544     return interpreter_frame_manager;
1545   }
1546 
1547   __ bind(frame_manager_entry);
1548 
1549   // the following temporary registers are used during frame creation
1550   const Register Gtmp1 = G3_scratch;
1551   // const Register Lmirror = L1;     // native mirror (native calls only)
1552 
1553   const Address constMethod       (G5_method, 0, in_bytes(Method::const_offset()));
1554   const Address access_flags      (G5_method, 0, in_bytes(Method::access_flags_offset()));
1555 
1556   address entry_point = __ pc();
1557   __ mov(G0, prevState);                                                 // no current activation
1558 
1559 
1560   Label re_dispatch;
1561 
1562   __ bind(re_dispatch);
1563 
1564   // Interpreter needs to have locals completely contiguous. In order to do that
1565   // We must adjust the caller's stack pointer for any locals beyond just the
1566   // parameters
1567   adjust_callers_stack(Gargs);
1568 
1569   // O5_savedSP still contains sender's sp
1570 
1571   // NEW FRAME
1572 
1573   generate_compute_interpreter_state(Lstate, prevState, false);
1574 
1575   // At this point a new interpreter frame and state object are created and initialized
1576   // Lstate has the pointer to the new activation
1577   // Any stack banging or limit check should already be done.
1578 
1579   Label call_interpreter;
1580 
1581   __ bind(call_interpreter);
1582 
1583 
1584 #if 1
1585   __ set(0xdead002, Lmirror);
1586   __ set(0xdead002, L2_scratch);
1587   __ set(0xdead003, L3_scratch);
1588   __ set(0xdead004, L4_scratch);
1589   __ set(0xdead005, Lscratch);
1590   __ set(0xdead006, Lscratch2);
1591   __ set(0xdead007, L7_scratch);
1592 
1593   __ set(0xdeaf002, O2);
1594   __ set(0xdeaf003, O3);
1595   __ set(0xdeaf004, O4);
1596   __ set(0xdeaf005, O5);
1597 #endif
1598 
1599   // Call interpreter (stack bang complete) enter here if message is
1600   // set and we know stack size is valid
1601 
1602   Label call_interpreter_2;
1603 
1604   __ bind(call_interpreter_2);
1605 
1606 #ifdef ASSERT
1607   {
1608     Label skip;
1609     __ ld_ptr(STATE(_frame_bottom), G3_scratch);
1610     __ cmp(G3_scratch, SP);
1611     __ brx(Assembler::equal, false, Assembler::pt, skip);
1612     __ delayed()->nop();
1613     __ stop("SP not restored to frame bottom");
1614     __ bind(skip);
1615   }
1616 #endif
1617 
1618   VALIDATE_STATE(G3_scratch, 4);
1619   __ set_last_Java_frame(SP, noreg);
1620   __ mov(Lstate, O0);                 // (arg) pointer to current state
1621 
1622   __ call(CAST_FROM_FN_PTR(address,
1623                            JvmtiExport::can_post_interpreter_events() ?
1624                                                                   BytecodeInterpreter::runWithChecks
1625                                                                 : BytecodeInterpreter::run),
1626          relocInfo::runtime_call_type);
1627 
1628   __ delayed()->nop();
1629 
1630   __ ld_ptr(STATE(_thread), G2_thread);
1631   __ reset_last_Java_frame();
1632 
1633   // examine msg from interpreter to determine next action
1634   __ ld_ptr(STATE(_thread), G2_thread);                                  // restore G2_thread
1635 
1636   __ ld(STATE(_msg), L1_scratch);                                       // Get new message
1637 
1638   Label call_method;
1639   Label return_from_interpreted_method;
1640   Label throw_exception;
1641   Label do_OSR;
1642   Label bad_msg;
1643   Label resume_interpreter;
1644 
1645   __ cmp(L1_scratch, (int)BytecodeInterpreter::call_method);
1646   __ br(Assembler::equal, false, Assembler::pt, call_method);
1647   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::return_from_method);
1648   __ br(Assembler::equal, false, Assembler::pt, return_from_interpreted_method);
1649   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::throwing_exception);
1650   __ br(Assembler::equal, false, Assembler::pt, throw_exception);
1651   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::do_osr);
1652   __ br(Assembler::equal, false, Assembler::pt, do_OSR);
1653   __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::more_monitors);
1654   __ br(Assembler::notEqual, false, Assembler::pt, bad_msg);
1655 
1656   // Allocate more monitor space, shuffle expression stack....
1657 
1658   generate_more_monitors();
1659 
1660   // new monitor slot allocated, resume the interpreter.
1661 
1662   __ set((int)BytecodeInterpreter::got_monitors, L1_scratch);
1663   VALIDATE_STATE(G3_scratch, 5);
1664   __ ba(call_interpreter);
1665   __ delayed()->st(L1_scratch, STATE(_msg));
1666 
1667   // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode)
1668   unctrap_frame_manager_entry  = __ pc();
1669 
1670   // QQQ what message do we send
1671 
1672   __ ba(call_interpreter);
1673   __ delayed()->ld_ptr(STATE(_frame_bottom), SP);                  // restore to full stack frame
1674 
1675   //=============================================================================
1676   // Returning from a compiled method into a deopted method. The bytecode at the
1677   // bcp has completed. The result of the bytecode is in the native abi (the tosca
1678   // for the template based interpreter). Any stack space that was used by the
1679   // bytecode that has completed has been removed (e.g. parameters for an invoke)
1680   // so all that we have to do is place any pending result on the expression stack
1681   // and resume execution on the next bytecode.
1682 
1683   generate_deopt_handling();
1684 
1685   // ready to resume the interpreter
1686 
1687   __ set((int)BytecodeInterpreter::deopt_resume, L1_scratch);
1688   __ ba(call_interpreter);
1689   __ delayed()->st(L1_scratch, STATE(_msg));
1690 
1691   // Current frame has caught an exception we need to dispatch to the
1692   // handler. We can get here because a native interpreter frame caught
1693   // an exception in which case there is no handler and we must rethrow
1694   // If it is a vanilla interpreted frame the we simply drop into the
1695   // interpreter and let it do the lookup.
1696 
1697   Interpreter::_rethrow_exception_entry = __ pc();
1698 
1699   Label return_with_exception;
1700   Label unwind_and_forward;
1701 
1702   // O0: exception
1703   // O7: throwing pc
1704 
1705   // We want exception in the thread no matter what we ultimately decide about frame type.
1706 
1707   Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
1708   __ verify_thread();
1709   __ st_ptr(O0, exception_addr);
1710 
1711   // get the Method*
1712   __ ld_ptr(STATE(_method), G5_method);
1713 
1714   // if this current frame vanilla or native?
1715 
1716   __ ld(access_flags, Gtmp1);
1717   __ btst(JVM_ACC_NATIVE, Gtmp1);
1718   __ br(Assembler::zero, false, Assembler::pt, return_with_exception);  // vanilla interpreted frame handle directly
1719   __ delayed()->nop();
1720 
1721   // We drop thru to unwind a native interpreted frame with a pending exception
1722   // We jump here for the initial interpreter frame with exception pending
1723   // We unwind the current acivation and forward it to our caller.
1724 
1725   __ bind(unwind_and_forward);
1726 
1727   // Unwind frame and jump to forward exception. unwinding will place throwing pc in O7
1728   // as expected by forward_exception.
1729 
1730   __ restore(FP, G0, SP);                  // unwind interpreter state frame
1731   __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
1732   __ delayed()->mov(I5_savedSP->after_restore(), SP);
1733 
1734   // Return point from a call which returns a result in the native abi
1735   // (c1/c2/jni-native). This result must be processed onto the java
1736   // expression stack.
1737   //
1738   // A pending exception may be present in which case there is no result present
1739 
1740   address return_from_native_method = __ pc();
1741 
1742   VALIDATE_STATE(G3_scratch, 6);
1743 
1744   // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1745   // stack is in the state that the  calling convention left it.
1746   // Copy the result from native abi result and place it on java expression stack.
1747 
1748   // Current interpreter state is present in Lstate
1749 
1750   // Exception pending?
1751 
1752   __ ld_ptr(STATE(_frame_bottom), SP);                             // restore to full stack frame
1753   __ ld_ptr(exception_addr, Lscratch);                                         // get any pending exception
1754   __ tst(Lscratch);                                                            // exception pending?
1755   __ brx(Assembler::notZero, false, Assembler::pt, return_with_exception);
1756   __ delayed()->nop();
1757 
1758   // Process the native abi result to java expression stack
1759 
1760   __ ld_ptr(STATE(_result._to_call._callee), L4_scratch);                        // called method
1761   __ ld_ptr(STATE(_stack), L1_scratch);                                          // get top of java expr stack
1762   // get parameter size
1763   __ ld_ptr(L4_scratch, in_bytes(Method::const_offset()), L2_scratch);
1764   __ lduh(L2_scratch, in_bytes(ConstMethod::size_of_parameters_offset()), L2_scratch);
1765   __ sll(L2_scratch, LogBytesPerWord, L2_scratch     );                           // parameter size in bytes
1766   __ add(L1_scratch, L2_scratch, L1_scratch);                                      // stack destination for result
1767   __ ld(L4_scratch, in_bytes(Method::result_index_offset()), L3_scratch); // called method result type index
1768 
1769   // tosca is really just native abi
1770   __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1771   __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1772   __ ld_ptr(L4_scratch, L3_scratch, Lscratch);                                       // get typed result converter address
1773   __ jmpl(Lscratch, G0, O7);                                                   // and convert it
1774   __ delayed()->nop();
1775 
1776   // L1_scratch points to top of stack (prepushed)
1777 
1778   __ ba(resume_interpreter);
1779   __ delayed()->mov(L1_scratch, O1);
1780 
1781   // An exception is being caught on return to a vanilla interpreter frame.
1782   // Empty the stack and resume interpreter
1783 
1784   __ bind(return_with_exception);
1785 
1786   __ ld_ptr(STATE(_frame_bottom), SP);                             // restore to full stack frame
1787   __ ld_ptr(STATE(_stack_base), O1);                               // empty java expression stack
1788   __ ba(resume_interpreter);
1789   __ delayed()->sub(O1, wordSize, O1);                             // account for prepush
1790 
1791   // Return from interpreted method we return result appropriate to the caller (i.e. "recursive"
1792   // interpreter call, or native) and unwind this interpreter activation.
1793   // All monitors should be unlocked.
1794 
1795   __ bind(return_from_interpreted_method);
1796 
1797   VALIDATE_STATE(G3_scratch, 7);
1798 
1799   Label return_to_initial_caller;
1800 
1801   // Interpreted result is on the top of the completed activation expression stack.
1802   // We must return it to the top of the callers stack if caller was interpreted
1803   // otherwise we convert to native abi result and return to call_stub/c1/c2
1804   // The caller's expression stack was truncated by the call however the current activation
1805   // has enough stuff on the stack that we have usable space there no matter what. The
1806   // other thing that makes it easy is that the top of the caller's stack is stored in STATE(_locals)
1807   // for the current activation
1808 
1809   __ ld_ptr(STATE(_prev_link), L1_scratch);
1810   __ ld_ptr(STATE(_method), L2_scratch);                               // get method just executed
1811   __ ld(L2_scratch, in_bytes(Method::result_index_offset()), L2_scratch);
1812   __ tst(L1_scratch);
1813   __ brx(Assembler::zero, false, Assembler::pt, return_to_initial_caller);
1814   __ delayed()->sll(L2_scratch, LogBytesPerWord, L2_scratch);
1815 
1816   // Copy result to callers java stack
1817 
1818   __ set((intptr_t)CppInterpreter::_stack_to_stack, L4_scratch);
1819   __ ld_ptr(L4_scratch, L2_scratch, Lscratch);                          // get typed result converter address
1820   __ ld_ptr(STATE(_stack), O0);                                       // current top (prepushed)
1821   __ ld_ptr(STATE(_locals), O1);                                      // stack destination
1822 
1823   // O0 - will be source, O1 - will be destination (preserved)
1824   __ jmpl(Lscratch, G0, O7);                                          // and convert it
1825   __ delayed()->add(O0, wordSize, O0);                                // get source (top of current expr stack)
1826 
1827   // O1 == &locals[0]
1828 
1829   // Result is now on caller's stack. Just unwind current activation and resume
1830 
1831   Label unwind_recursive_activation;
1832 
1833 
1834   __ bind(unwind_recursive_activation);
1835 
1836   // O1 == &locals[0] (really callers stacktop) for activation now returning
1837   // returning to interpreter method from "recursive" interpreter call
1838   // result converter left O1 pointing to top of the( prepushed) java stack for method we are returning
1839   // to. Now all we must do is unwind the state from the completed call
1840 
1841   // Must restore stack
1842   VALIDATE_STATE(G3_scratch, 8);
1843 
1844   // Return to interpreter method after a method call (interpreted/native/c1/c2) has completed.
1845   // Result if any is already on the caller's stack. All we must do now is remove the now dead
1846   // frame and tell interpreter to resume.
1847 
1848 
1849   __ mov(O1, I1);                                                     // pass back new stack top across activation
1850   // POP FRAME HERE ==================================
1851   __ restore(FP, G0, SP);                                             // unwind interpreter state frame
1852   __ ld_ptr(STATE(_frame_bottom), SP);                                // restore to full stack frame
1853 
1854 
1855   // Resume the interpreter. The current frame contains the current interpreter
1856   // state object.
1857   //
1858   // O1 == new java stack pointer
1859 
1860   __ bind(resume_interpreter);
1861   VALIDATE_STATE(G3_scratch, 10);
1862 
1863   // A frame we have already used before so no need to bang stack so use call_interpreter_2 entry
1864 
1865   __ set((int)BytecodeInterpreter::method_resume, L1_scratch);
1866   __ st(L1_scratch, STATE(_msg));
1867   __ ba(call_interpreter_2);
1868   __ delayed()->st_ptr(O1, STATE(_stack));
1869 
1870 
1871   // Fast accessor methods share this entry point.
1872   // This works because frame manager is in the same codelet
1873   // This can either be an entry via call_stub/c1/c2 or a recursive interpreter call
1874   // we need to do a little register fixup here once we distinguish the two of them
1875   if (UseFastAccessorMethods && !synchronized) {
1876   // Call stub_return address still in O7
1877     __ bind(fast_accessor_slow_entry_path);
1878     __ set((intptr_t)return_from_native_method - 8, Gtmp1);
1879     __ cmp(Gtmp1, O7);                                                // returning to interpreter?
1880     __ brx(Assembler::equal, true, Assembler::pt, re_dispatch);       // yep
1881     __ delayed()->nop();
1882     __ ba(re_dispatch);
1883     __ delayed()->mov(G0, prevState);                                 // initial entry
1884 
1885   }
1886 
1887   // interpreter returning to native code (call_stub/c1/c2)
1888   // convert result and unwind initial activation
1889   // L2_scratch - scaled result type index
1890 
1891   __ bind(return_to_initial_caller);
1892 
1893   __ set((intptr_t)CppInterpreter::_stack_to_native_abi, L4_scratch);
1894   __ ld_ptr(L4_scratch, L2_scratch, Lscratch);                           // get typed result converter address
1895   __ ld_ptr(STATE(_stack), O0);                                        // current top (prepushed)
1896   __ jmpl(Lscratch, G0, O7);                                           // and convert it
1897   __ delayed()->add(O0, wordSize, O0);                                 // get source (top of current expr stack)
1898 
1899   Label unwind_initial_activation;
1900   __ bind(unwind_initial_activation);
1901 
1902   // RETURN TO CALL_STUB/C1/C2 code (result if any in I0..I1/(F0/..F1)
1903   // we can return here with an exception that wasn't handled by interpreted code
1904   // how does c1/c2 see it on return?
1905 
1906   // compute resulting sp before/after args popped depending upon calling convention
1907   // __ ld_ptr(STATE(_saved_sp), Gtmp1);
1908   //
1909   // POP FRAME HERE ==================================
1910   __ restore(FP, G0, SP);
1911   __ retl();
1912   __ delayed()->mov(I5_savedSP->after_restore(), SP);
1913 
1914   // OSR request, unwind the current frame and transfer to the OSR entry
1915   // and enter OSR nmethod
1916 
1917   __ bind(do_OSR);
1918   Label remove_initial_frame;
1919   __ ld_ptr(STATE(_prev_link), L1_scratch);
1920   __ ld_ptr(STATE(_result._osr._osr_buf), G1_scratch);
1921 
1922   // We are going to pop this frame. Is there another interpreter frame underneath
1923   // it or is it callstub/compiled?
1924 
1925   __ tst(L1_scratch);
1926   __ brx(Assembler::zero, false, Assembler::pt, remove_initial_frame);
1927   __ delayed()->ld_ptr(STATE(_result._osr._osr_entry), G3_scratch);
1928 
1929   // Frame underneath is an interpreter frame simply unwind
1930   // POP FRAME HERE ==================================
1931   __ restore(FP, G0, SP);                                             // unwind interpreter state frame
1932   __ mov(I5_savedSP->after_restore(), SP);
1933 
1934   // Since we are now calling native need to change our "return address" from the
1935   // dummy RecursiveInterpreterActivation to a return from native
1936 
1937   __ set((intptr_t)return_from_native_method - 8, O7);
1938 
1939   __ jmpl(G3_scratch, G0, G0);
1940   __ delayed()->mov(G1_scratch, O0);
1941 
1942   __ bind(remove_initial_frame);
1943 
1944   // POP FRAME HERE ==================================
1945   __ restore(FP, G0, SP);
1946   __ mov(I5_savedSP->after_restore(), SP);
1947   __ jmpl(G3_scratch, G0, G0);
1948   __ delayed()->mov(G1_scratch, O0);
1949 
1950   // Call a new method. All we do is (temporarily) trim the expression stack
1951   // push a return address to bring us back to here and leap to the new entry.
1952   // At this point we have a topmost frame that was allocated by the frame manager
1953   // which contains the current method interpreted state. We trim this frame
1954   // of excess java expression stack entries and then recurse.
1955 
1956   __ bind(call_method);
1957 
1958   // stack points to next free location and not top element on expression stack
1959   // method expects sp to be pointing to topmost element
1960 
1961   __ ld_ptr(STATE(_thread), G2_thread);
1962   __ ld_ptr(STATE(_result._to_call._callee), G5_method);
1963 
1964 
1965   // SP already takes in to account the 2 extra words we use for slop
1966   // when we call a "static long no_params()" method. So if
1967   // we trim back sp by the amount of unused java expression stack
1968   // there will be automagically the 2 extra words we need.
1969   // We also have to worry about keeping SP aligned.
1970 
1971   __ ld_ptr(STATE(_stack), Gargs);
1972   __ ld_ptr(STATE(_stack_limit), L1_scratch);
1973 
1974   // compute the unused java stack size
1975   __ sub(Gargs, L1_scratch, L2_scratch);                       // compute unused space
1976 
1977   // Round down the unused space to that stack is always 16-byte aligned
1978   // by making the unused space a multiple of the size of two longs.
1979 
1980   __ and3(L2_scratch, -2*BytesPerLong, L2_scratch);
1981 
1982   // Now trim the stack
1983   __ add(SP, L2_scratch, SP);
1984 
1985 
1986   // Now point to the final argument (account for prepush)
1987   __ add(Gargs, wordSize, Gargs);
1988 #ifdef ASSERT
1989   // Make sure we have space for the window
1990   __ sub(Gargs, SP, L1_scratch);
1991   __ cmp(L1_scratch, 16*wordSize);
1992   {
1993     Label skip;
1994     __ brx(Assembler::greaterEqual, false, Assembler::pt, skip);
1995     __ delayed()->nop();
1996     __ stop("killed stack");
1997     __ bind(skip);
1998   }
1999 #endif // ASSERT
2000 
2001   // Create a new frame where we can store values that make it look like the interpreter
2002   // really recursed.
2003 
2004   // prepare to recurse or call specialized entry
2005 
2006   // First link the registers we need
2007 
2008   // make the pc look good in debugger
2009   __ set(CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation), O7);
2010   // argument too
2011   __ mov(Lstate, I0);
2012 
2013   // Record our sending SP
2014   __ mov(SP, O5_savedSP);
2015 
2016   __ ld_ptr(STATE(_result._to_call._callee_entry_point), L2_scratch);
2017   __ set((intptr_t) entry_point, L1_scratch);
2018   __ cmp(L1_scratch, L2_scratch);
2019   __ brx(Assembler::equal, false, Assembler::pt, re_dispatch);
2020   __ delayed()->mov(Lstate, prevState);                                // link activations
2021 
2022   // method uses specialized entry, push a return so we look like call stub setup
2023   // this path will handle fact that result is returned in registers and not
2024   // on the java stack.
2025 
2026   __ set((intptr_t)return_from_native_method - 8, O7);
2027   __ jmpl(L2_scratch, G0, G0);                               // Do specialized entry
2028   __ delayed()->nop();
2029 
2030   //
2031   // Bad Message from interpreter
2032   //
2033   __ bind(bad_msg);
2034   __ stop("Bad message from interpreter");
2035 
2036   // Interpreted method "returned" with an exception pass it on...
2037   // Pass result, unwind activation and continue/return to interpreter/call_stub
2038   // We handle result (if any) differently based on return to interpreter or call_stub
2039 
2040   __ bind(throw_exception);
2041   __ ld_ptr(STATE(_prev_link), L1_scratch);
2042   __ tst(L1_scratch);
2043   __ brx(Assembler::zero, false, Assembler::pt, unwind_and_forward);
2044   __ delayed()->nop();
2045 
2046   __ ld_ptr(STATE(_locals), O1); // get result of popping callee's args
2047   __ ba(unwind_recursive_activation);
2048   __ delayed()->nop();
2049 
2050   interpreter_frame_manager = entry_point;
2051   return entry_point;
2052 }
2053 
2054 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2055  : CppInterpreterGenerator(code) {
2056    generate_all(); // down here so it can be "virtual"
2057 }
2058 
2059 
2060 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
2061 
2062   // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
2063   // expression stack, the callee will have callee_extra_locals (so we can account for
2064   // frame extension) and monitor_size for monitors. Basically we need to calculate
2065   // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
2066   //
2067   //
2068   // The big complicating thing here is that we must ensure that the stack stays properly
2069   // aligned. This would be even uglier if monitor size wasn't modulo what the stack
2070   // needs to be aligned for). We are given that the sp (fp) is already aligned by
2071   // the caller so we must ensure that it is properly aligned for our callee.
2072   //
2073   // Ths c++ interpreter always makes sure that we have a enough extra space on the
2074   // stack at all times to deal with the "stack long no_params()" method issue. This
2075   // is "slop_factor" here.
2076   const int slop_factor = 2;
2077 
2078   const int fixed_size = sizeof(BytecodeInterpreter)/wordSize +           // interpreter state object
2079                          frame::memory_parameter_word_sp_offset;   // register save area + param window
2080   return (round_to(max_stack +
2081                    slop_factor +
2082                    fixed_size +
2083                    monitor_size +
2084                    (callee_extra_locals * Interpreter::stackElementWords), WordsPerLong));
2085 
2086 }
2087 
2088 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
2089 
2090   // See call_stub code
2091   int call_stub_size  = round_to(7 + frame::memory_parameter_word_sp_offset,
2092                                  WordsPerLong);    // 7 + register save area
2093 
2094   // Save space for one monitor to get into the interpreted method in case
2095   // the method is synchronized
2096   int monitor_size    = method->is_synchronized() ?
2097                                 1*frame::interpreter_frame_monitor_size() : 0;
2098   return size_activation_helper(method->max_locals(), method->max_stack(),
2099                                  monitor_size) + call_stub_size;
2100 }
2101 
2102 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
2103                                            frame* caller,
2104                                            frame* current,
2105                                            Method* method,
2106                                            intptr_t* locals,
2107                                            intptr_t* stack,
2108                                            intptr_t* stack_base,
2109                                            intptr_t* monitor_base,
2110                                            intptr_t* frame_bottom,
2111                                            bool is_top_frame
2112                                            )
2113 {
2114   // What about any vtable?
2115   //
2116   to_fill->_thread = JavaThread::current();
2117   // This gets filled in later but make it something recognizable for now
2118   to_fill->_bcp = method->code_base();
2119   to_fill->_locals = locals;
2120   to_fill->_constants = method->constants()->cache();
2121   to_fill->_method = method;
2122   to_fill->_mdx = NULL;
2123   to_fill->_stack = stack;
2124   if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) {
2125     to_fill->_msg = deopt_resume2;
2126   } else {
2127     to_fill->_msg = method_resume;
2128   }
2129   to_fill->_result._to_call._bcp_advance = 0;
2130   to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2131   to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2132   to_fill->_prev_link = NULL;
2133 
2134   // Fill in the registers for the frame
2135 
2136   // Need to install _sender_sp. Actually not too hard in C++!
2137   // When the skeletal frames are layed out we fill in a value
2138   // for _sender_sp. That value is only correct for the oldest
2139   // skeletal frame constructed (because there is only a single
2140   // entry for "caller_adjustment". While the skeletal frames
2141   // exist that is good enough. We correct that calculation
2142   // here and get all the frames correct.
2143 
2144   // to_fill->_sender_sp = locals - (method->size_of_parameters() - 1);
2145 
2146   *current->register_addr(Lstate) = (intptr_t) to_fill;
2147   // skeletal already places a useful value here and this doesn't account
2148   // for alignment so don't bother.
2149   // *current->register_addr(I5_savedSP) =     (intptr_t) locals - (method->size_of_parameters() - 1);
2150 
2151   if (caller->is_interpreted_frame()) {
2152     interpreterState prev  = caller->get_interpreterState();
2153     to_fill->_prev_link = prev;
2154     // Make the prev callee look proper
2155     prev->_result._to_call._callee = method;
2156     if (*prev->_bcp == Bytecodes::_invokeinterface) {
2157       prev->_result._to_call._bcp_advance = 5;
2158     } else {
2159       prev->_result._to_call._bcp_advance = 3;
2160     }
2161   }
2162   to_fill->_oop_temp = NULL;
2163   to_fill->_stack_base = stack_base;
2164   // Need +1 here because stack_base points to the word just above the first expr stack entry
2165   // and stack_limit is supposed to point to the word just below the last expr stack entry.
2166   // See generate_compute_interpreter_state.
2167   to_fill->_stack_limit = stack_base - (method->max_stack() + 1);
2168   to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2169 
2170   // sparc specific
2171   to_fill->_frame_bottom = frame_bottom;
2172   to_fill->_self_link = to_fill;
2173 #ifdef ASSERT
2174   to_fill->_native_fresult = 123456.789;
2175   to_fill->_native_lresult = CONST64(0xdeadcafedeafcafe);
2176 #endif
2177 }
2178 
2179 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, address last_Java_pc, intptr_t* last_Java_fp) {
2180   istate->_last_Java_pc = (intptr_t*) last_Java_pc;
2181 }
2182 
2183 
2184 int AbstractInterpreter::layout_activation(Method* method,
2185                                            int tempcount, // Number of slots on java expression stack in use
2186                                            int popframe_extra_args,
2187                                            int moncount,  // Number of active monitors
2188                                            int caller_actual_parameters,
2189                                            int callee_param_size,
2190                                            int callee_locals_size,
2191                                            frame* caller,
2192                                            frame* interpreter_frame,
2193                                            bool is_top_frame,
2194                                            bool is_bottom_frame) {
2195 
2196   assert(popframe_extra_args == 0, "NEED TO FIX");
2197   // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state()
2198   // does as far as allocating an interpreter frame.
2199   // If interpreter_frame!=NULL, set up the method, locals, and monitors.
2200   // The frame interpreter_frame, if not NULL, is guaranteed to be the right size,
2201   // as determined by a previous call to this method.
2202   // It is also guaranteed to be walkable even though it is in a skeletal state
2203   // NOTE: return size is in words not bytes
2204   // NOTE: tempcount is the current size of the java expression stack. For top most
2205   //       frames we will allocate a full sized expression stack and not the curback
2206   //       version that non-top frames have.
2207 
2208   // Calculate the amount our frame will be adjust by the callee. For top frame
2209   // this is zero.
2210 
2211   // NOTE: ia64 seems to do this wrong (or at least backwards) in that it
2212   // calculates the extra locals based on itself. Not what the callee does
2213   // to it. So it ignores last_frame_adjust value. Seems suspicious as far
2214   // as getting sender_sp correct.
2215 
2216   int extra_locals_size = callee_locals_size - callee_param_size;
2217   int monitor_size = (sizeof(BasicObjectLock) * moncount) / wordSize;
2218   int full_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2219   int short_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2220   int frame_words = is_top_frame ? full_frame_words : short_frame_words;
2221 
2222 
2223   /*
2224     if we actually have a frame to layout we must now fill in all the pieces. This means both
2225     the interpreterState and the registers.
2226   */
2227   if (interpreter_frame != NULL) {
2228 
2229     // MUCHO HACK
2230 
2231     intptr_t* frame_bottom = interpreter_frame->sp() - (full_frame_words - frame_words);
2232     // 'interpreter_frame->sp()' is unbiased while 'frame_bottom' must be a biased value in 64bit mode.
2233     assert(((intptr_t)frame_bottom & 0xf) == 0, "SP biased in layout_activation");
2234     frame_bottom = (intptr_t*)((intptr_t)frame_bottom - STACK_BIAS);
2235 
2236     /* Now fillin the interpreterState object */
2237 
2238     interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() -  sizeof(BytecodeInterpreter));
2239 
2240 
2241     intptr_t* locals;
2242 
2243     // Calculate the postion of locals[0]. This is painful because of
2244     // stack alignment (same as ia64). The problem is that we can
2245     // not compute the location of locals from fp(). fp() will account
2246     // for the extra locals but it also accounts for aligning the stack
2247     // and we can't determine if the locals[0] was misaligned but max_locals
2248     // was enough to have the
2249     // calculate postion of locals. fp already accounts for extra locals.
2250     // +2 for the static long no_params() issue.
2251 
2252     if (caller->is_interpreted_frame()) {
2253       // locals must agree with the caller because it will be used to set the
2254       // caller's tos when we return.
2255       interpreterState prev  = caller->get_interpreterState();
2256       // stack() is prepushed.
2257       locals = prev->stack() + method->size_of_parameters();
2258     } else {
2259       // Lay out locals block in the caller adjacent to the register window save area.
2260       //
2261       // Compiled frames do not allocate a varargs area which is why this if
2262       // statement is needed.
2263       //
2264       intptr_t* fp = interpreter_frame->fp();
2265       int local_words = method->max_locals() * Interpreter::stackElementWords;
2266 
2267       if (caller->is_compiled_frame()) {
2268         locals = fp + frame::register_save_words + local_words - 1;
2269       } else {
2270         locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
2271       }
2272 
2273     }
2274     // END MUCHO HACK
2275 
2276     intptr_t* monitor_base = (intptr_t*) cur_state;
2277     intptr_t* stack_base =  monitor_base - monitor_size;
2278     /* +1 because stack is always prepushed */
2279     intptr_t* stack = stack_base - (tempcount + 1);
2280 
2281 
2282     BytecodeInterpreter::layout_interpreterState(cur_state,
2283                                           caller,
2284                                           interpreter_frame,
2285                                           method,
2286                                           locals,
2287                                           stack,
2288                                           stack_base,
2289                                           monitor_base,
2290                                           frame_bottom,
2291                                           is_top_frame);
2292 
2293     BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp());
2294 
2295   }
2296   return frame_words;
2297 }
2298 
2299 #endif // CC_INTERP