1 /*
   2  * Copyright (c) 1997, 2011, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/assembler.hpp"
  27 #include "interpreter/bytecodeHistogram.hpp"
  28 #include "interpreter/interpreter.hpp"
  29 #include "interpreter/interpreterGenerator.hpp"
  30 #include "interpreter/interpreterRuntime.hpp"
  31 #include "interpreter/templateTable.hpp"
  32 #include "oops/arrayOop.hpp"
  33 #include "oops/methodDataOop.hpp"
  34 #include "oops/methodOop.hpp"
  35 #include "oops/oop.inline.hpp"
  36 #include "prims/jvmtiExport.hpp"
  37 #include "prims/jvmtiThreadState.hpp"
  38 #include "runtime/arguments.hpp"
  39 #include "runtime/deoptimization.hpp"
  40 #include "runtime/frame.inline.hpp"
  41 #include "runtime/sharedRuntime.hpp"
  42 #include "runtime/stubRoutines.hpp"
  43 #include "runtime/synchronizer.hpp"
  44 #include "runtime/timer.hpp"
  45 #include "runtime/vframeArray.hpp"
  46 #include "utilities/debug.hpp"
  47 
  48 #ifndef CC_INTERP
  49 #ifndef FAST_DISPATCH
  50 #define FAST_DISPATCH 1
  51 #endif
  52 #undef FAST_DISPATCH
  53 
  54 
  55 // Generation of Interpreter
  56 //
  57 // The InterpreterGenerator generates the interpreter into Interpreter::_code.
  58 
  59 
  60 #define __ _masm->
  61 
  62 
  63 //----------------------------------------------------------------------------------------------------
  64 
  65 
  66 void InterpreterGenerator::save_native_result(void) {
  67   // result potentially in O0/O1: save it across calls
  68   const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
  69 
  70   // result potentially in F0/F1: save it across calls
  71   const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
  72 
  73   // save and restore any potential method result value around the unlocking operation
  74   __ stf(FloatRegisterImpl::D, F0, d_tmp);
  75 #ifdef _LP64
  76   __ stx(O0, l_tmp);
  77 #else
  78   __ std(O0, l_tmp);
  79 #endif
  80 }
  81 
  82 void InterpreterGenerator::restore_native_result(void) {
  83   const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
  84   const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
  85 
  86   // Restore any method result value
  87   __ ldf(FloatRegisterImpl::D, d_tmp, F0);
  88 #ifdef _LP64
  89   __ ldx(l_tmp, O0);
  90 #else
  91   __ ldd(l_tmp, O0);
  92 #endif
  93 }
  94 
  95 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
  96   assert(!pass_oop || message == NULL, "either oop or message but not both");
  97   address entry = __ pc();
  98   // expression stack must be empty before entering the VM if an exception happened
  99   __ empty_expression_stack();
 100   // load exception object
 101   __ set((intptr_t)name, G3_scratch);
 102   if (pass_oop) {
 103     __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
 104   } else {
 105     __ set((intptr_t)message, G4_scratch);
 106     __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
 107   }
 108   // throw exception
 109   assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
 110   AddressLiteral thrower(Interpreter::throw_exception_entry());
 111   __ jump_to(thrower, G3_scratch);
 112   __ delayed()->nop();
 113   return entry;
 114 }
 115 
 116 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
 117   address entry = __ pc();
 118   // expression stack must be empty before entering the VM if an exception
 119   // happened
 120   __ empty_expression_stack();
 121   // load exception object
 122   __ call_VM(Oexception,
 123              CAST_FROM_FN_PTR(address,
 124                               InterpreterRuntime::throw_ClassCastException),
 125              Otos_i);
 126   __ should_not_reach_here();
 127   return entry;
 128 }
 129 
 130 
 131 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
 132   address entry = __ pc();
 133   // expression stack must be empty before entering the VM if an exception happened
 134   __ empty_expression_stack();
 135   // convention: expect aberrant index in register G3_scratch, then shuffle the
 136   // index to G4_scratch for the VM call
 137   __ mov(G3_scratch, G4_scratch);
 138   __ set((intptr_t)name, G3_scratch);
 139   __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
 140   __ should_not_reach_here();
 141   return entry;
 142 }
 143 
 144 
 145 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
 146   address entry = __ pc();
 147   // expression stack must be empty before entering the VM if an exception happened
 148   __ empty_expression_stack();
 149   __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
 150   __ should_not_reach_here();
 151   return entry;
 152 }
 153 
 154 
 155 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step) {
 156   TosState incoming_state = state;
 157 
 158   Label cont;
 159   address compiled_entry = __ pc();
 160 
 161   address entry = __ pc();
 162 #if !defined(_LP64) && defined(COMPILER2)
 163   // All return values are where we want them, except for Longs.  C2 returns
 164   // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
 165   // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
 166   // build even if we are returning from interpreted we just do a little
 167   // stupid shuffing.
 168   // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
 169   // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
 170   // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
 171 
 172   if (incoming_state == ltos) {
 173     __ srl (G1,  0, O1);
 174     __ srlx(G1, 32, O0);
 175   }
 176 #endif // !_LP64 && COMPILER2
 177 
 178   __ bind(cont);
 179 
 180   // The callee returns with the stack possibly adjusted by adapter transition
 181   // We remove that possible adjustment here.
 182   // All interpreter local registers are untouched. Any result is passed back
 183   // in the O0/O1 or float registers. Before continuing, the arguments must be
 184   // popped from the java expression stack; i.e., Lesp must be adjusted.
 185 
 186   __ mov(Llast_SP, SP);   // Remove any adapter added stack space.
 187 
 188   Label L_got_cache, L_giant_index;
 189   const Register cache = G3_scratch;
 190   const Register size  = G1_scratch;
 191   if (EnableInvokeDynamic) {
 192     __ ldub(Address(Lbcp, 0), G1_scratch);  // Load current bytecode.
 193     __ cmp_and_br_short(G1_scratch, Bytecodes::_invokedynamic, Assembler::equal, Assembler::pn, L_giant_index);
 194   }
 195   __ get_cache_and_index_at_bcp(cache, G1_scratch, 1);
 196   __ bind(L_got_cache);
 197   __ ld_ptr(cache, constantPoolCacheOopDesc::base_offset() +
 198                    ConstantPoolCacheEntry::flags_offset(), size);
 199   __ and3(size, 0xFF, size);                   // argument size in words
 200   __ sll(size, Interpreter::logStackElementSize, size); // each argument size in bytes
 201   __ add(Lesp, size, Lesp);                    // pop arguments
 202   __ dispatch_next(state, step);
 203 
 204   // out of the main line of code...
 205   if (EnableInvokeDynamic) {
 206     __ bind(L_giant_index);
 207     __ get_cache_and_index_at_bcp(cache, G1_scratch, 1, sizeof(u4));
 208     __ ba_short(L_got_cache);
 209   }
 210 
 211   return entry;
 212 }
 213 
 214 
 215 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
 216   address entry = __ pc();
 217   __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
 218   { Label L;
 219     Address exception_addr(G2_thread, Thread::pending_exception_offset());
 220     __ ld_ptr(exception_addr, Gtemp);  // Load pending exception.
 221     __ br_null_short(Gtemp, Assembler::pt, L);
 222     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
 223     __ should_not_reach_here();
 224     __ bind(L);
 225   }
 226   __ dispatch_next(state, step);
 227   return entry;
 228 }
 229 
 230 // A result handler converts/unboxes a native call result into
 231 // a java interpreter/compiler result. The current frame is an
 232 // interpreter frame. The activation frame unwind code must be
 233 // consistent with that of TemplateTable::_return(...). In the
 234 // case of native methods, the caller's SP was not modified.
 235 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
 236   address entry = __ pc();
 237   Register Itos_i  = Otos_i ->after_save();
 238   Register Itos_l  = Otos_l ->after_save();
 239   Register Itos_l1 = Otos_l1->after_save();
 240   Register Itos_l2 = Otos_l2->after_save();
 241   switch (type) {
 242     case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
 243     case T_CHAR   : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i);   break; // cannot use and3, 0xFFFF too big as immediate value!
 244     case T_BYTE   : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i);   break;
 245     case T_SHORT  : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i);   break;
 246     case T_LONG   :
 247 #ifndef _LP64
 248                     __ mov(O1, Itos_l2);  // move other half of long
 249 #endif              // ifdef or no ifdef, fall through to the T_INT case
 250     case T_INT    : __ mov(O0, Itos_i);                         break;
 251     case T_VOID   : /* nothing to do */                         break;
 252     case T_FLOAT  : assert(F0 == Ftos_f, "fix this code" );     break;
 253     case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" );     break;
 254     case T_OBJECT :
 255       __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
 256       __ verify_oop(Itos_i);
 257       break;
 258     default       : ShouldNotReachHere();
 259   }
 260   __ ret();                           // return from interpreter activation
 261   __ delayed()->restore(I5_savedSP, G0, SP);  // remove interpreter frame
 262   NOT_PRODUCT(__ emit_long(0);)       // marker for disassembly
 263   return entry;
 264 }
 265 
 266 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
 267   address entry = __ pc();
 268   __ push(state);
 269   __ call_VM(noreg, runtime_entry);
 270   __ dispatch_via(vtos, Interpreter::normal_table(vtos));
 271   return entry;
 272 }
 273 
 274 
 275 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
 276   address entry = __ pc();
 277   __ dispatch_next(state);
 278   return entry;
 279 }
 280 
 281 //
 282 // Helpers for commoning out cases in the various type of method entries.
 283 //
 284 
 285 // increment invocation count & check for overflow
 286 //
 287 // Note: checking for negative value instead of overflow
 288 //       so we have a 'sticky' overflow test
 289 //
 290 // Lmethod: method
 291 // ??: invocation counter
 292 //
 293 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
 294   // Note: In tiered we increment either counters in methodOop or in MDO depending if we're profiling or not.
 295   if (TieredCompilation) {
 296     const int increment = InvocationCounter::count_increment;
 297     const int mask = ((1 << Tier0InvokeNotifyFreqLog) - 1) << InvocationCounter::count_shift;
 298     Label no_mdo, done;
 299     if (ProfileInterpreter) {
 300       // If no method data exists, go to profile_continue.
 301       __ ld_ptr(Lmethod, methodOopDesc::method_data_offset(), G4_scratch);
 302       __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
 303       // Increment counter
 304       Address mdo_invocation_counter(G4_scratch,
 305                                      in_bytes(methodDataOopDesc::invocation_counter_offset()) +
 306                                      in_bytes(InvocationCounter::counter_offset()));
 307       __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
 308                                  G3_scratch, Lscratch,
 309                                  Assembler::zero, overflow);
 310       __ ba_short(done);
 311     }
 312 
 313     // Increment counter in methodOop
 314     __ bind(no_mdo);
 315     Address invocation_counter(Lmethod,
 316                                in_bytes(methodOopDesc::invocation_counter_offset()) +
 317                                in_bytes(InvocationCounter::counter_offset()));
 318     __ increment_mask_and_jump(invocation_counter, increment, mask,
 319                                G3_scratch, Lscratch,
 320                                Assembler::zero, overflow);
 321     __ bind(done);
 322   } else {
 323     // Update standard invocation counters
 324     __ increment_invocation_counter(O0, G3_scratch);
 325     if (ProfileInterpreter) {  // %%% Merge this into methodDataOop
 326       Address interpreter_invocation_counter(Lmethod,in_bytes(methodOopDesc::interpreter_invocation_counter_offset()));
 327       __ ld(interpreter_invocation_counter, G3_scratch);
 328       __ inc(G3_scratch);
 329       __ st(G3_scratch, interpreter_invocation_counter);
 330     }
 331 
 332     if (ProfileInterpreter && profile_method != NULL) {
 333       // Test to see if we should create a method data oop
 334       AddressLiteral profile_limit((address)&InvocationCounter::InterpreterProfileLimit);
 335       __ load_contents(profile_limit, G3_scratch);
 336       __ cmp_and_br_short(O0, G3_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
 337 
 338       // if no method data exists, go to profile_method
 339       __ test_method_data_pointer(*profile_method);
 340     }
 341 
 342     AddressLiteral invocation_limit((address)&InvocationCounter::InterpreterInvocationLimit);
 343     __ load_contents(invocation_limit, G3_scratch);
 344     __ cmp(O0, G3_scratch);
 345     __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
 346     __ delayed()->nop();
 347   }
 348 
 349 }
 350 
 351 // Allocate monitor and lock method (asm interpreter)
 352 // ebx - methodOop
 353 //
 354 void InterpreterGenerator::lock_method(void) {
 355   __ ld(Lmethod, in_bytes(methodOopDesc::access_flags_offset()), O0);  // Load access flags.
 356 
 357 #ifdef ASSERT
 358  { Label ok;
 359    __ btst(JVM_ACC_SYNCHRONIZED, O0);
 360    __ br( Assembler::notZero, false, Assembler::pt, ok);
 361    __ delayed()->nop();
 362    __ stop("method doesn't need synchronization");
 363    __ bind(ok);
 364   }
 365 #endif // ASSERT
 366 
 367   // get synchronization object to O0
 368   { Label done;
 369     const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
 370     __ btst(JVM_ACC_STATIC, O0);
 371     __ br( Assembler::zero, true, Assembler::pt, done);
 372     __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
 373 
 374     __ ld_ptr( Lmethod, in_bytes(methodOopDesc::constants_offset()), O0);
 375     __ ld_ptr( O0, constantPoolOopDesc::pool_holder_offset_in_bytes(), O0);
 376 
 377     // lock the mirror, not the klassOop
 378     __ ld_ptr( O0, mirror_offset, O0);
 379 
 380 #ifdef ASSERT
 381     __ tst(O0);
 382     __ breakpoint_trap(Assembler::zero);
 383 #endif // ASSERT
 384 
 385     __ bind(done);
 386   }
 387 
 388   __ add_monitor_to_stack(true, noreg, noreg);  // allocate monitor elem
 389   __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes());   // store object
 390   // __ untested("lock_object from method entry");
 391   __ lock_object(Lmonitors, O0);
 392 }
 393 
 394 
 395 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
 396                                                          Register Rscratch,
 397                                                          Register Rscratch2) {
 398   const int page_size = os::vm_page_size();
 399   Address saved_exception_pc(G2_thread, JavaThread::saved_exception_pc_offset());
 400   Label after_frame_check;
 401 
 402   assert_different_registers(Rframe_size, Rscratch, Rscratch2);
 403 
 404   __ set(page_size, Rscratch);
 405   __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
 406 
 407   // get the stack base, and in debug, verify it is non-zero
 408   __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch );
 409 #ifdef ASSERT
 410   Label base_not_zero;
 411   __ br_notnull_short(Rscratch, Assembler::pn, base_not_zero);
 412   __ stop("stack base is zero in generate_stack_overflow_check");
 413   __ bind(base_not_zero);
 414 #endif
 415 
 416   // get the stack size, and in debug, verify it is non-zero
 417   assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" );
 418   __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 );
 419 #ifdef ASSERT
 420   Label size_not_zero;
 421   __ br_notnull_short(Rscratch2, Assembler::pn, size_not_zero);
 422   __ stop("stack size is zero in generate_stack_overflow_check");
 423   __ bind(size_not_zero);
 424 #endif
 425 
 426   // compute the beginning of the protected zone minus the requested frame size
 427   __ sub( Rscratch, Rscratch2,   Rscratch );
 428   __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 );
 429   __ add( Rscratch, Rscratch2,   Rscratch );
 430 
 431   // Add in the size of the frame (which is the same as subtracting it from the
 432   // SP, which would take another register
 433   __ add( Rscratch, Rframe_size, Rscratch );
 434 
 435   // the frame is greater than one page in size, so check against
 436   // the bottom of the stack
 437   __ cmp_and_brx_short(SP, Rscratch, Assembler::greater, Assembler::pt, after_frame_check);
 438 
 439   // Save the return address as the exception pc
 440   __ st_ptr(O7, saved_exception_pc);
 441 
 442   // the stack will overflow, throw an exception
 443   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
 444 
 445   // if you get to here, then there is enough stack space
 446   __ bind( after_frame_check );
 447 }
 448 
 449 
 450 //
 451 // Generate a fixed interpreter frame. This is identical setup for interpreted
 452 // methods and for native methods hence the shared code.
 453 
 454 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
 455   //
 456   //
 457   // The entry code sets up a new interpreter frame in 4 steps:
 458   //
 459   // 1) Increase caller's SP by for the extra local space needed:
 460   //    (check for overflow)
 461   //    Efficient implementation of xload/xstore bytecodes requires
 462   //    that arguments and non-argument locals are in a contigously
 463   //    addressable memory block => non-argument locals must be
 464   //    allocated in the caller's frame.
 465   //
 466   // 2) Create a new stack frame and register window:
 467   //    The new stack frame must provide space for the standard
 468   //    register save area, the maximum java expression stack size,
 469   //    the monitor slots (0 slots initially), and some frame local
 470   //    scratch locations.
 471   //
 472   // 3) The following interpreter activation registers must be setup:
 473   //    Lesp       : expression stack pointer
 474   //    Lbcp       : bytecode pointer
 475   //    Lmethod    : method
 476   //    Llocals    : locals pointer
 477   //    Lmonitors  : monitor pointer
 478   //    LcpoolCache: constant pool cache
 479   //
 480   // 4) Initialize the non-argument locals if necessary:
 481   //    Non-argument locals may need to be initialized to NULL
 482   //    for GC to work. If the oop-map information is accurate
 483   //    (in the absence of the JSR problem), no initialization
 484   //    is necessary.
 485   //
 486   // (gri - 2/25/2000)
 487 
 488 
 489   const Address size_of_parameters(G5_method, methodOopDesc::size_of_parameters_offset());
 490   const Address size_of_locals    (G5_method, methodOopDesc::size_of_locals_offset());
 491   const Address max_stack         (G5_method, methodOopDesc::max_stack_offset());
 492   int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
 493 
 494   const int extra_space =
 495     rounded_vm_local_words +                   // frame local scratch space
 496     //6815692//methodOopDesc::extra_stack_words() +       // extra push slots for MH adapters
 497     frame::memory_parameter_word_sp_offset +   // register save area
 498     (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
 499 
 500   const Register Glocals_size = G3;
 501   const Register Otmp1 = O3;
 502   const Register Otmp2 = O4;
 503   // Lscratch can't be used as a temporary because the call_stub uses
 504   // it to assert that the stack frame was setup correctly.
 505 
 506   __ lduh( size_of_parameters, Glocals_size);
 507 
 508   // Gargs points to first local + BytesPerWord
 509   // Set the saved SP after the register window save
 510   //
 511   assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
 512   __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
 513   __ add(Gargs, Otmp1, Gargs);
 514 
 515   if (native_call) {
 516     __ calc_mem_param_words( Glocals_size, Gframe_size );
 517     __ add( Gframe_size,  extra_space, Gframe_size);
 518     __ round_to( Gframe_size, WordsPerLong );
 519     __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
 520   } else {
 521 
 522     //
 523     // Compute number of locals in method apart from incoming parameters
 524     //
 525     __ lduh( size_of_locals, Otmp1 );
 526     __ sub( Otmp1, Glocals_size, Glocals_size );
 527     __ round_to( Glocals_size, WordsPerLong );
 528     __ sll( Glocals_size, Interpreter::logStackElementSize, Glocals_size );
 529 
 530     // see if the frame is greater than one page in size. If so,
 531     // then we need to verify there is enough stack space remaining
 532     // Frame_size = (max_stack + extra_space) * BytesPerWord;
 533     __ lduh( max_stack, Gframe_size );
 534     __ add( Gframe_size, extra_space, Gframe_size );
 535     __ round_to( Gframe_size, WordsPerLong );
 536     __ sll( Gframe_size, Interpreter::logStackElementSize, Gframe_size);
 537 
 538     // Add in java locals size for stack overflow check only
 539     __ add( Gframe_size, Glocals_size, Gframe_size );
 540 
 541     const Register Otmp2 = O4;
 542     assert_different_registers(Otmp1, Otmp2, O5_savedSP);
 543     generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2);
 544 
 545     __ sub( Gframe_size, Glocals_size, Gframe_size);
 546 
 547     //
 548     // bump SP to accomodate the extra locals
 549     //
 550     __ sub( SP, Glocals_size, SP );
 551   }
 552 
 553   //
 554   // now set up a stack frame with the size computed above
 555   //
 556   __ neg( Gframe_size );
 557   __ save( SP, Gframe_size, SP );
 558 
 559   //
 560   // now set up all the local cache registers
 561   //
 562   // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
 563   // that all present references to Lbyte_code initialize the register
 564   // immediately before use
 565   if (native_call) {
 566     __ mov(G0, Lbcp);
 567   } else {
 568     __ ld_ptr(G5_method, methodOopDesc::const_offset(), Lbcp);
 569     __ add(Lbcp, in_bytes(constMethodOopDesc::codes_offset()), Lbcp);
 570   }
 571   __ mov( G5_method, Lmethod);                 // set Lmethod
 572   __ get_constant_pool_cache( LcpoolCache );   // set LcpoolCache
 573   __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
 574 #ifdef _LP64
 575   __ add( Lmonitors, STACK_BIAS, Lmonitors );   // Account for 64 bit stack bias
 576 #endif
 577   __ sub(Lmonitors, BytesPerWord, Lesp);       // set Lesp
 578 
 579   // setup interpreter activation registers
 580   __ sub(Gargs, BytesPerWord, Llocals);        // set Llocals
 581 
 582   if (ProfileInterpreter) {
 583 #ifdef FAST_DISPATCH
 584     // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
 585     // they both use I2.
 586     assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
 587 #endif // FAST_DISPATCH
 588     __ set_method_data_pointer();
 589   }
 590 
 591 }
 592 
 593 // Empty method, generate a very fast return.
 594 
 595 address InterpreterGenerator::generate_empty_entry(void) {
 596 
 597   // A method that does nother but return...
 598 
 599   address entry = __ pc();
 600   Label slow_path;
 601 
 602   __ verify_oop(G5_method);
 603 
 604   // do nothing for empty methods (do not even increment invocation counter)
 605   if ( UseFastEmptyMethods) {
 606     // If we need a safepoint check, generate full interpreter entry.
 607     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
 608     __ set(sync_state, G3_scratch);
 609     __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path);
 610 
 611     // Code: _return
 612     __ retl();
 613     __ delayed()->mov(O5_savedSP, SP);
 614 
 615     __ bind(slow_path);
 616     (void) generate_normal_entry(false);
 617 
 618     return entry;
 619   }
 620   return NULL;
 621 }
 622 
 623 // Call an accessor method (assuming it is resolved, otherwise drop into
 624 // vanilla (slow path) entry
 625 
 626 // Generates code to elide accessor methods
 627 // Uses G3_scratch and G1_scratch as scratch
 628 address InterpreterGenerator::generate_accessor_entry(void) {
 629 
 630   // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
 631   // parameter size = 1
 632   // Note: We can only use this code if the getfield has been resolved
 633   //       and if we don't have a null-pointer exception => check for
 634   //       these conditions first and use slow path if necessary.
 635   address entry = __ pc();
 636   Label slow_path;
 637 
 638 
 639   // XXX: for compressed oops pointer loading and decoding doesn't fit in
 640   // delay slot and damages G1
 641   if ( UseFastAccessorMethods && !UseCompressedOops ) {
 642     // Check if we need to reach a safepoint and generate full interpreter
 643     // frame if so.
 644     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
 645     __ load_contents(sync_state, G3_scratch);
 646     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 647     __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path);
 648 
 649     // Check if local 0 != NULL
 650     __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
 651     // check if local 0 == NULL and go the slow path
 652     __ br_null_short(Otos_i, Assembler::pn, slow_path);
 653 
 654 
 655     // read first instruction word and extract bytecode @ 1 and index @ 2
 656     // get first 4 bytes of the bytecodes (big endian!)
 657     __ ld_ptr(G5_method, methodOopDesc::const_offset(), G1_scratch);
 658     __ ld(G1_scratch, constMethodOopDesc::codes_offset(), G1_scratch);
 659 
 660     // move index @ 2 far left then to the right most two bytes.
 661     __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
 662     __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
 663                       ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
 664 
 665     // get constant pool cache
 666     __ ld_ptr(G5_method, methodOopDesc::constants_offset(), G3_scratch);
 667     __ ld_ptr(G3_scratch, constantPoolOopDesc::cache_offset_in_bytes(), G3_scratch);
 668 
 669     // get specific constant pool cache entry
 670     __ add(G3_scratch, G1_scratch, G3_scratch);
 671 
 672     // Check the constant Pool cache entry to see if it has been resolved.
 673     // If not, need the slow path.
 674     ByteSize cp_base_offset = constantPoolCacheOopDesc::base_offset();
 675     __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::indices_offset(), G1_scratch);
 676     __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
 677     __ and3(G1_scratch, 0xFF, G1_scratch);
 678     __ cmp_and_br_short(G1_scratch, Bytecodes::_getfield, Assembler::notEqual, Assembler::pn, slow_path);
 679 
 680     // Get the type and return field offset from the constant pool cache
 681     __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::flags_offset(), G1_scratch);
 682     __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), G3_scratch);
 683 
 684     Label xreturn_path;
 685     // Need to differentiate between igetfield, agetfield, bgetfield etc.
 686     // because they are different sizes.
 687     // Get the type from the constant pool cache
 688     __ srl(G1_scratch, ConstantPoolCacheEntry::tosBits, G1_scratch);
 689     // Make sure we don't need to mask G1_scratch for tosBits after the above shift
 690     ConstantPoolCacheEntry::verify_tosBits();
 691     __ cmp(G1_scratch, atos );
 692     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 693     __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
 694     __ cmp(G1_scratch, itos);
 695     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 696     __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
 697     __ cmp(G1_scratch, stos);
 698     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 699     __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
 700     __ cmp(G1_scratch, ctos);
 701     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 702     __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
 703 #ifdef ASSERT
 704     __ cmp(G1_scratch, btos);
 705     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 706     __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
 707     __ should_not_reach_here();
 708 #endif
 709     __ ldsb(Otos_i, G3_scratch, Otos_i);
 710     __ bind(xreturn_path);
 711 
 712     // _ireturn/_areturn
 713     __ retl();                      // return from leaf routine
 714     __ delayed()->mov(O5_savedSP, SP);
 715 
 716     // Generate regular method entry
 717     __ bind(slow_path);
 718     (void) generate_normal_entry(false);
 719     return entry;
 720   }
 721   return NULL;
 722 }
 723 
 724 // Method entry for java.lang.ref.Reference.get.
 725 address InterpreterGenerator::generate_Reference_get_entry(void) {
 726 #ifndef SERIALGC
 727   // Code: _aload_0, _getfield, _areturn
 728   // parameter size = 1
 729   //
 730   // The code that gets generated by this routine is split into 2 parts:
 731   //    1. The "intrinsified" code for G1 (or any SATB based GC),
 732   //    2. The slow path - which is an expansion of the regular method entry.
 733   //
 734   // Notes:-
 735   // * In the G1 code we do not check whether we need to block for
 736   //   a safepoint. If G1 is enabled then we must execute the specialized
 737   //   code for Reference.get (except when the Reference object is null)
 738   //   so that we can log the value in the referent field with an SATB
 739   //   update buffer.
 740   //   If the code for the getfield template is modified so that the
 741   //   G1 pre-barrier code is executed when the current method is
 742   //   Reference.get() then going through the normal method entry
 743   //   will be fine.
 744   // * The G1 code can, however, check the receiver object (the instance
 745   //   of java.lang.Reference) and jump to the slow path if null. If the
 746   //   Reference object is null then we obviously cannot fetch the referent
 747   //   and so we don't need to call the G1 pre-barrier. Thus we can use the
 748   //   regular method entry code to generate the NPE.
 749   //
 750   // This code is based on generate_accessor_enty.
 751 
 752   address entry = __ pc();
 753 
 754   const int referent_offset = java_lang_ref_Reference::referent_offset;
 755   guarantee(referent_offset > 0, "referent offset not initialized");
 756 
 757   if (UseG1GC) {
 758      Label slow_path;
 759 
 760     // In the G1 code we don't check if we need to reach a safepoint. We
 761     // continue and the thread will safepoint at the next bytecode dispatch.
 762 
 763     // Check if local 0 != NULL
 764     // If the receiver is null then it is OK to jump to the slow path.
 765     __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
 766     // check if local 0 == NULL and go the slow path
 767     __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
 768 
 769 
 770     // Load the value of the referent field.
 771     if (Assembler::is_simm13(referent_offset)) {
 772       __ load_heap_oop(Otos_i, referent_offset, Otos_i);
 773     } else {
 774       __ set(referent_offset, G3_scratch);
 775       __ load_heap_oop(Otos_i, G3_scratch, Otos_i);
 776     }
 777 
 778     // Generate the G1 pre-barrier code to log the value of
 779     // the referent field in an SATB buffer. Note with
 780     // these parameters the pre-barrier does not generate
 781     // the load of the previous value
 782 
 783     __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */,
 784                             Otos_i /* pre_val */,
 785                             G3_scratch /* tmp */,
 786                             true /* preserve_o_regs */);
 787 
 788     // _areturn
 789     __ retl();                      // return from leaf routine
 790     __ delayed()->mov(O5_savedSP, SP);
 791 
 792     // Generate regular method entry
 793     __ bind(slow_path);
 794     (void) generate_normal_entry(false);
 795     return entry;
 796   }
 797 #endif // SERIALGC
 798 
 799   // If G1 is not enabled then attempt to go through the accessor entry point
 800   // Reference.get is an accessor
 801   return generate_accessor_entry();
 802 }
 803 
 804 //
 805 // Interpreter stub for calling a native method. (asm interpreter)
 806 // This sets up a somewhat different looking stack for calling the native method
 807 // than the typical interpreter frame setup.
 808 //
 809 
 810 address InterpreterGenerator::generate_native_entry(bool synchronized) {
 811   address entry = __ pc();
 812 
 813   // the following temporary registers are used during frame creation
 814   const Register Gtmp1 = G3_scratch ;
 815   const Register Gtmp2 = G1_scratch;
 816   bool inc_counter  = UseCompiler || CountCompiledCalls;
 817 
 818   // make sure registers are different!
 819   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
 820 
 821   const Address Laccess_flags(Lmethod, methodOopDesc::access_flags_offset());
 822 
 823   __ verify_oop(G5_method);
 824 
 825   const Register Glocals_size = G3;
 826   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
 827 
 828   // make sure method is native & not abstract
 829   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
 830 #ifdef ASSERT
 831   __ ld(G5_method, methodOopDesc::access_flags_offset(), Gtmp1);
 832   {
 833     Label L;
 834     __ btst(JVM_ACC_NATIVE, Gtmp1);
 835     __ br(Assembler::notZero, false, Assembler::pt, L);
 836     __ delayed()->nop();
 837     __ stop("tried to execute non-native method as native");
 838     __ bind(L);
 839   }
 840   { Label L;
 841     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
 842     __ br(Assembler::zero, false, Assembler::pt, L);
 843     __ delayed()->nop();
 844     __ stop("tried to execute abstract method as non-abstract");
 845     __ bind(L);
 846   }
 847 #endif // ASSERT
 848 
 849  // generate the code to allocate the interpreter stack frame
 850   generate_fixed_frame(true);
 851 
 852   //
 853   // No locals to initialize for native method
 854   //
 855 
 856   // this slot will be set later, we initialize it to null here just in
 857   // case we get a GC before the actual value is stored later
 858   __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
 859 
 860   const Address do_not_unlock_if_synchronized(G2_thread,
 861     JavaThread::do_not_unlock_if_synchronized_offset());
 862   // Since at this point in the method invocation the exception handler
 863   // would try to exit the monitor of synchronized methods which hasn't
 864   // been entered yet, we set the thread local variable
 865   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
 866   // runtime, exception handling i.e. unlock_if_synchronized_method will
 867   // check this thread local flag.
 868   // This flag has two effects, one is to force an unwind in the topmost
 869   // interpreter frame and not perform an unlock while doing so.
 870 
 871   __ movbool(true, G3_scratch);
 872   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
 873 
 874   // increment invocation counter and check for overflow
 875   //
 876   // Note: checking for negative value instead of overflow
 877   //       so we have a 'sticky' overflow test (may be of
 878   //       importance as soon as we have true MT/MP)
 879   Label invocation_counter_overflow;
 880   Label Lcontinue;
 881   if (inc_counter) {
 882     generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
 883 
 884   }
 885   __ bind(Lcontinue);
 886 
 887   bang_stack_shadow_pages(true);
 888 
 889   // reset the _do_not_unlock_if_synchronized flag
 890   __ stbool(G0, do_not_unlock_if_synchronized);
 891 
 892   // check for synchronized methods
 893   // Must happen AFTER invocation_counter check and stack overflow check,
 894   // so method is not locked if overflows.
 895 
 896   if (synchronized) {
 897     lock_method();
 898   } else {
 899 #ifdef ASSERT
 900     { Label ok;
 901       __ ld(Laccess_flags, O0);
 902       __ btst(JVM_ACC_SYNCHRONIZED, O0);
 903       __ br( Assembler::zero, false, Assembler::pt, ok);
 904       __ delayed()->nop();
 905       __ stop("method needs synchronization");
 906       __ bind(ok);
 907     }
 908 #endif // ASSERT
 909   }
 910 
 911 
 912   // start execution
 913   __ verify_thread();
 914 
 915   // JVMTI support
 916   __ notify_method_entry();
 917 
 918   // native call
 919 
 920   // (note that O0 is never an oop--at most it is a handle)
 921   // It is important not to smash any handles created by this call,
 922   // until any oop handle in O0 is dereferenced.
 923 
 924   // (note that the space for outgoing params is preallocated)
 925 
 926   // get signature handler
 927   { Label L;
 928     Address signature_handler(Lmethod, methodOopDesc::signature_handler_offset());
 929     __ ld_ptr(signature_handler, G3_scratch);
 930     __ br_notnull_short(G3_scratch, Assembler::pt, L);
 931     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
 932     __ ld_ptr(signature_handler, G3_scratch);
 933     __ bind(L);
 934   }
 935 
 936   // Push a new frame so that the args will really be stored in
 937   // Copy a few locals across so the new frame has the variables
 938   // we need but these values will be dead at the jni call and
 939   // therefore not gc volatile like the values in the current
 940   // frame (Lmethod in particular)
 941 
 942   // Flush the method pointer to the register save area
 943   __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
 944   __ mov(Llocals, O1);
 945 
 946   // calculate where the mirror handle body is allocated in the interpreter frame:
 947   __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
 948 
 949   // Calculate current frame size
 950   __ sub(SP, FP, O3);         // Calculate negative of current frame size
 951   __ save(SP, O3, SP);        // Allocate an identical sized frame
 952 
 953   // Note I7 has leftover trash. Slow signature handler will fill it in
 954   // should we get there. Normal jni call will set reasonable last_Java_pc
 955   // below (and fix I7 so the stack trace doesn't have a meaningless frame
 956   // in it).
 957 
 958   // Load interpreter frame's Lmethod into same register here
 959 
 960   __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
 961 
 962   __ mov(I1, Llocals);
 963   __ mov(I2, Lscratch2);     // save the address of the mirror
 964 
 965 
 966   // ONLY Lmethod and Llocals are valid here!
 967 
 968   // call signature handler, It will move the arg properly since Llocals in current frame
 969   // matches that in outer frame
 970 
 971   __ callr(G3_scratch, 0);
 972   __ delayed()->nop();
 973 
 974   // Result handler is in Lscratch
 975 
 976   // Reload interpreter frame's Lmethod since slow signature handler may block
 977   __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
 978 
 979   { Label not_static;
 980 
 981     __ ld(Laccess_flags, O0);
 982     __ btst(JVM_ACC_STATIC, O0);
 983     __ br( Assembler::zero, false, Assembler::pt, not_static);
 984     // get native function entry point(O0 is a good temp until the very end)
 985     __ delayed()->ld_ptr(Lmethod, in_bytes(methodOopDesc::native_function_offset()), O0);
 986     // for static methods insert the mirror argument
 987     const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
 988 
 989     __ ld_ptr(Lmethod, methodOopDesc:: constants_offset(), O1);
 990     __ ld_ptr(O1, constantPoolOopDesc::pool_holder_offset_in_bytes(), O1);
 991     __ ld_ptr(O1, mirror_offset, O1);
 992 #ifdef ASSERT
 993     if (!PrintSignatureHandlers)  // do not dirty the output with this
 994     { Label L;
 995       __ br_notnull_short(O1, Assembler::pt, L);
 996       __ stop("mirror is missing");
 997       __ bind(L);
 998     }
 999 #endif // ASSERT
1000     __ st_ptr(O1, Lscratch2, 0);
1001     __ mov(Lscratch2, O1);
1002     __ bind(not_static);
1003   }
1004 
1005   // At this point, arguments have been copied off of stack into
1006   // their JNI positions, which are O1..O5 and SP[68..].
1007   // Oops are boxed in-place on the stack, with handles copied to arguments.
1008   // The result handler is in Lscratch.  O0 will shortly hold the JNIEnv*.
1009 
1010 #ifdef ASSERT
1011   { Label L;
1012     __ br_notnull_short(O0, Assembler::pt, L);
1013     __ stop("native entry point is missing");
1014     __ bind(L);
1015   }
1016 #endif // ASSERT
1017 
1018   //
1019   // setup the frame anchor
1020   //
1021   // The scavenge function only needs to know that the PC of this frame is
1022   // in the interpreter method entry code, it doesn't need to know the exact
1023   // PC and hence we can use O7 which points to the return address from the
1024   // previous call in the code stream (signature handler function)
1025   //
1026   // The other trick is we set last_Java_sp to FP instead of the usual SP because
1027   // we have pushed the extra frame in order to protect the volatile register(s)
1028   // in that frame when we return from the jni call
1029   //
1030 
1031   __ set_last_Java_frame(FP, O7);
1032   __ mov(O7, I7);  // make dummy interpreter frame look like one above,
1033                    // not meaningless information that'll confuse me.
1034 
1035   // flush the windows now. We don't care about the current (protection) frame
1036   // only the outer frames
1037 
1038   __ flush_windows();
1039 
1040   // mark windows as flushed
1041   Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1042   __ set(JavaFrameAnchor::flushed, G3_scratch);
1043   __ st(G3_scratch, flags);
1044 
1045   // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1046 
1047   Address thread_state(G2_thread, JavaThread::thread_state_offset());
1048 #ifdef ASSERT
1049   { Label L;
1050     __ ld(thread_state, G3_scratch);
1051     __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1052     __ stop("Wrong thread state in native stub");
1053     __ bind(L);
1054   }
1055 #endif // ASSERT
1056   __ set(_thread_in_native, G3_scratch);
1057   __ st(G3_scratch, thread_state);
1058 
1059   // Call the jni method, using the delay slot to set the JNIEnv* argument.
1060   __ save_thread(L7_thread_cache); // save Gthread
1061   __ callr(O0, 0);
1062   __ delayed()->
1063      add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1064 
1065   // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1066 
1067   __ restore_thread(L7_thread_cache); // restore G2_thread
1068   __ reinit_heapbase();
1069 
1070   // must we block?
1071 
1072   // Block, if necessary, before resuming in _thread_in_Java state.
1073   // In order for GC to work, don't clear the last_Java_sp until after blocking.
1074   { Label no_block;
1075     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1076 
1077     // Switch thread to "native transition" state before reading the synchronization state.
1078     // This additional state is necessary because reading and testing the synchronization
1079     // state is not atomic w.r.t. GC, as this scenario demonstrates:
1080     //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1081     //     VM thread changes sync state to synchronizing and suspends threads for GC.
1082     //     Thread A is resumed to finish this native method, but doesn't block here since it
1083     //     didn't see any synchronization is progress, and escapes.
1084     __ set(_thread_in_native_trans, G3_scratch);
1085     __ st(G3_scratch, thread_state);
1086     if(os::is_MP()) {
1087       if (UseMembar) {
1088         // Force this write out before the read below
1089         __ membar(Assembler::StoreLoad);
1090       } else {
1091         // Write serialization page so VM thread can do a pseudo remote membar.
1092         // We use the current thread pointer to calculate a thread specific
1093         // offset to write to within the page. This minimizes bus traffic
1094         // due to cache line collision.
1095         __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1096       }
1097     }
1098     __ load_contents(sync_state, G3_scratch);
1099     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1100 
1101     Label L;
1102     __ br(Assembler::notEqual, false, Assembler::pn, L);
1103     __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1104     __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1105     __ bind(L);
1106 
1107     // Block.  Save any potential method result value before the operation and
1108     // use a leaf call to leave the last_Java_frame setup undisturbed.
1109     save_native_result();
1110     __ call_VM_leaf(L7_thread_cache,
1111                     CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1112                     G2_thread);
1113 
1114     // Restore any method result value
1115     restore_native_result();
1116     __ bind(no_block);
1117   }
1118 
1119   // Clear the frame anchor now
1120 
1121   __ reset_last_Java_frame();
1122 
1123   // Move the result handler address
1124   __ mov(Lscratch, G3_scratch);
1125   // return possible result to the outer frame
1126 #ifndef __LP64
1127   __ mov(O0, I0);
1128   __ restore(O1, G0, O1);
1129 #else
1130   __ restore(O0, G0, O0);
1131 #endif /* __LP64 */
1132 
1133   // Move result handler to expected register
1134   __ mov(G3_scratch, Lscratch);
1135 
1136   // Back in normal (native) interpreter frame. State is thread_in_native_trans
1137   // switch to thread_in_Java.
1138 
1139   __ set(_thread_in_Java, G3_scratch);
1140   __ st(G3_scratch, thread_state);
1141 
1142   // reset handle block
1143   __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1144   __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1145 
1146   // If we have an oop result store it where it will be safe for any further gc
1147   // until we return now that we've released the handle it might be protected by
1148 
1149   {
1150     Label no_oop, store_result;
1151 
1152     __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1153     __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1154     __ addcc(G0, O0, O0);
1155     __ brx(Assembler::notZero, true, Assembler::pt, store_result);     // if result is not NULL:
1156     __ delayed()->ld_ptr(O0, 0, O0);                                   // unbox it
1157     __ mov(G0, O0);
1158 
1159     __ bind(store_result);
1160     // Store it where gc will look for it and result handler expects it.
1161     __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1162 
1163     __ bind(no_oop);
1164 
1165   }
1166 
1167 
1168   // handle exceptions (exception handling will handle unlocking!)
1169   { Label L;
1170     Address exception_addr(G2_thread, Thread::pending_exception_offset());
1171     __ ld_ptr(exception_addr, Gtemp);
1172     __ br_null_short(Gtemp, Assembler::pt, L);
1173     // Note: This could be handled more efficiently since we know that the native
1174     //       method doesn't have an exception handler. We could directly return
1175     //       to the exception handler for the caller.
1176     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1177     __ should_not_reach_here();
1178     __ bind(L);
1179   }
1180 
1181   // JVMTI support (preserves thread register)
1182   __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1183 
1184   if (synchronized) {
1185     // save and restore any potential method result value around the unlocking operation
1186     save_native_result();
1187 
1188     __ add( __ top_most_monitor(), O1);
1189     __ unlock_object(O1);
1190 
1191     restore_native_result();
1192   }
1193 
1194 #if defined(COMPILER2) && !defined(_LP64)
1195 
1196   // C2 expects long results in G1 we can't tell if we're returning to interpreted
1197   // or compiled so just be safe.
1198 
1199   __ sllx(O0, 32, G1);          // Shift bits into high G1
1200   __ srl (O1, 0, O1);           // Zero extend O1
1201   __ or3 (O1, G1, G1);          // OR 64 bits into G1
1202 
1203 #endif /* COMPILER2 && !_LP64 */
1204 
1205   // dispose of return address and remove activation
1206 #ifdef ASSERT
1207   {
1208     Label ok;
1209     __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1210     __ stop("bad I5_savedSP value");
1211     __ should_not_reach_here();
1212     __ bind(ok);
1213   }
1214 #endif
1215   if (TraceJumps) {
1216     // Move target to register that is recordable
1217     __ mov(Lscratch, G3_scratch);
1218     __ JMP(G3_scratch, 0);
1219   } else {
1220     __ jmp(Lscratch, 0);
1221   }
1222   __ delayed()->nop();
1223 
1224 
1225   if (inc_counter) {
1226     // handle invocation counter overflow
1227     __ bind(invocation_counter_overflow);
1228     generate_counter_overflow(Lcontinue);
1229   }
1230 
1231 
1232 
1233   return entry;
1234 }
1235 
1236 
1237 // Generic method entry to (asm) interpreter
1238 //------------------------------------------------------------------------------------------------------------------------
1239 //
1240 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1241   address entry = __ pc();
1242 
1243   bool inc_counter  = UseCompiler || CountCompiledCalls;
1244 
1245   // the following temporary registers are used during frame creation
1246   const Register Gtmp1 = G3_scratch ;
1247   const Register Gtmp2 = G1_scratch;
1248 
1249   // make sure registers are different!
1250   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1251 
1252   const Address size_of_parameters(G5_method, methodOopDesc::size_of_parameters_offset());
1253   const Address size_of_locals    (G5_method, methodOopDesc::size_of_locals_offset());
1254   // Seems like G5_method is live at the point this is used. So we could make this look consistent
1255   // and use in the asserts.
1256   const Address access_flags      (Lmethod,   methodOopDesc::access_flags_offset());
1257 
1258   __ verify_oop(G5_method);
1259 
1260   const Register Glocals_size = G3;
1261   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1262 
1263   // make sure method is not native & not abstract
1264   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1265 #ifdef ASSERT
1266   __ ld(G5_method, methodOopDesc::access_flags_offset(), Gtmp1);
1267   {
1268     Label L;
1269     __ btst(JVM_ACC_NATIVE, Gtmp1);
1270     __ br(Assembler::zero, false, Assembler::pt, L);
1271     __ delayed()->nop();
1272     __ stop("tried to execute native method as non-native");
1273     __ bind(L);
1274   }
1275   { Label L;
1276     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1277     __ br(Assembler::zero, false, Assembler::pt, L);
1278     __ delayed()->nop();
1279     __ stop("tried to execute abstract method as non-abstract");
1280     __ bind(L);
1281   }
1282 #endif // ASSERT
1283 
1284   // generate the code to allocate the interpreter stack frame
1285 
1286   generate_fixed_frame(false);
1287 
1288 #ifdef FAST_DISPATCH
1289   __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1290                                           // set bytecode dispatch table base
1291 #endif
1292 
1293   //
1294   // Code to initialize the extra (i.e. non-parm) locals
1295   //
1296   Register init_value = noreg;    // will be G0 if we must clear locals
1297   // The way the code was setup before zerolocals was always true for vanilla java entries.
1298   // It could only be false for the specialized entries like accessor or empty which have
1299   // no extra locals so the testing was a waste of time and the extra locals were always
1300   // initialized. We removed this extra complication to already over complicated code.
1301 
1302   init_value = G0;
1303   Label clear_loop;
1304 
1305   // NOTE: If you change the frame layout, this code will need to
1306   // be updated!
1307   __ lduh( size_of_locals, O2 );
1308   __ lduh( size_of_parameters, O1 );
1309   __ sll( O2, Interpreter::logStackElementSize, O2);
1310   __ sll( O1, Interpreter::logStackElementSize, O1 );
1311   __ sub( Llocals, O2, O2 );
1312   __ sub( Llocals, O1, O1 );
1313 
1314   __ bind( clear_loop );
1315   __ inc( O2, wordSize );
1316 
1317   __ cmp( O2, O1 );
1318   __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1319   __ delayed()->st_ptr( init_value, O2, 0 );
1320 
1321   const Address do_not_unlock_if_synchronized(G2_thread,
1322     JavaThread::do_not_unlock_if_synchronized_offset());
1323   // Since at this point in the method invocation the exception handler
1324   // would try to exit the monitor of synchronized methods which hasn't
1325   // been entered yet, we set the thread local variable
1326   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1327   // runtime, exception handling i.e. unlock_if_synchronized_method will
1328   // check this thread local flag.
1329   __ movbool(true, G3_scratch);
1330   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1331 
1332   // increment invocation counter and check for overflow
1333   //
1334   // Note: checking for negative value instead of overflow
1335   //       so we have a 'sticky' overflow test (may be of
1336   //       importance as soon as we have true MT/MP)
1337   Label invocation_counter_overflow;
1338   Label profile_method;
1339   Label profile_method_continue;
1340   Label Lcontinue;
1341   if (inc_counter) {
1342     generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1343     if (ProfileInterpreter) {
1344       __ bind(profile_method_continue);
1345     }
1346   }
1347   __ bind(Lcontinue);
1348 
1349   bang_stack_shadow_pages(false);
1350 
1351   // reset the _do_not_unlock_if_synchronized flag
1352   __ stbool(G0, do_not_unlock_if_synchronized);
1353 
1354   // check for synchronized methods
1355   // Must happen AFTER invocation_counter check and stack overflow check,
1356   // so method is not locked if overflows.
1357 
1358   if (synchronized) {
1359     lock_method();
1360   } else {
1361 #ifdef ASSERT
1362     { Label ok;
1363       __ ld(access_flags, O0);
1364       __ btst(JVM_ACC_SYNCHRONIZED, O0);
1365       __ br( Assembler::zero, false, Assembler::pt, ok);
1366       __ delayed()->nop();
1367       __ stop("method needs synchronization");
1368       __ bind(ok);
1369     }
1370 #endif // ASSERT
1371   }
1372 
1373   // start execution
1374 
1375   __ verify_thread();
1376 
1377   // jvmti support
1378   __ notify_method_entry();
1379 
1380   // start executing instructions
1381   __ dispatch_next(vtos);
1382 
1383 
1384   if (inc_counter) {
1385     if (ProfileInterpreter) {
1386       // We have decided to profile this method in the interpreter
1387       __ bind(profile_method);
1388 
1389       __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1390       __ set_method_data_pointer_for_bcp();
1391       __ ba_short(profile_method_continue);
1392     }
1393 
1394     // handle invocation counter overflow
1395     __ bind(invocation_counter_overflow);
1396     generate_counter_overflow(Lcontinue);
1397   }
1398 
1399 
1400   return entry;
1401 }
1402 
1403 
1404 //----------------------------------------------------------------------------------------------------
1405 // Entry points & stack frame layout
1406 //
1407 // Here we generate the various kind of entries into the interpreter.
1408 // The two main entry type are generic bytecode methods and native call method.
1409 // These both come in synchronized and non-synchronized versions but the
1410 // frame layout they create is very similar. The other method entry
1411 // types are really just special purpose entries that are really entry
1412 // and interpretation all in one. These are for trivial methods like
1413 // accessor, empty, or special math methods.
1414 //
1415 // When control flow reaches any of the entry types for the interpreter
1416 // the following holds ->
1417 //
1418 // C2 Calling Conventions:
1419 //
1420 // The entry code below assumes that the following registers are set
1421 // when coming in:
1422 //    G5_method: holds the methodOop of the method to call
1423 //    Lesp:    points to the TOS of the callers expression stack
1424 //             after having pushed all the parameters
1425 //
1426 // The entry code does the following to setup an interpreter frame
1427 //   pop parameters from the callers stack by adjusting Lesp
1428 //   set O0 to Lesp
1429 //   compute X = (max_locals - num_parameters)
1430 //   bump SP up by X to accomadate the extra locals
1431 //   compute X = max_expression_stack
1432 //               + vm_local_words
1433 //               + 16 words of register save area
1434 //   save frame doing a save sp, -X, sp growing towards lower addresses
1435 //   set Lbcp, Lmethod, LcpoolCache
1436 //   set Llocals to i0
1437 //   set Lmonitors to FP - rounded_vm_local_words
1438 //   set Lesp to Lmonitors - 4
1439 //
1440 //  The frame has now been setup to do the rest of the entry code
1441 
1442 // Try this optimization:  Most method entries could live in a
1443 // "one size fits all" stack frame without all the dynamic size
1444 // calculations.  It might be profitable to do all this calculation
1445 // statically and approximately for "small enough" methods.
1446 
1447 //-----------------------------------------------------------------------------------------------
1448 
1449 // C1 Calling conventions
1450 //
1451 // Upon method entry, the following registers are setup:
1452 //
1453 // g2 G2_thread: current thread
1454 // g5 G5_method: method to activate
1455 // g4 Gargs  : pointer to last argument
1456 //
1457 //
1458 // Stack:
1459 //
1460 // +---------------+ <--- sp
1461 // |               |
1462 // : reg save area :
1463 // |               |
1464 // +---------------+ <--- sp + 0x40
1465 // |               |
1466 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
1467 // |               |
1468 // +---------------+ <--- sp + 0x5c
1469 // |               |
1470 // :     free      :
1471 // |               |
1472 // +---------------+ <--- Gargs
1473 // |               |
1474 // :   arguments   :
1475 // |               |
1476 // +---------------+
1477 // |               |
1478 //
1479 //
1480 //
1481 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
1482 //
1483 // +---------------+ <--- sp
1484 // |               |
1485 // : reg save area :
1486 // |               |
1487 // +---------------+ <--- sp + 0x40
1488 // |               |
1489 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
1490 // |               |
1491 // +---------------+ <--- sp + 0x5c
1492 // |               |
1493 // :               :
1494 // |               | <--- Lesp
1495 // +---------------+ <--- Lmonitors (fp - 0x18)
1496 // |   VM locals   |
1497 // +---------------+ <--- fp
1498 // |               |
1499 // : reg save area :
1500 // |               |
1501 // +---------------+ <--- fp + 0x40
1502 // |               |
1503 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
1504 // |               |
1505 // +---------------+ <--- fp + 0x5c
1506 // |               |
1507 // :     free      :
1508 // |               |
1509 // +---------------+
1510 // |               |
1511 // : nonarg locals :
1512 // |               |
1513 // +---------------+
1514 // |               |
1515 // :   arguments   :
1516 // |               | <--- Llocals
1517 // +---------------+ <--- Gargs
1518 // |               |
1519 
1520 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
1521 
1522   // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
1523   // expression stack, the callee will have callee_extra_locals (so we can account for
1524   // frame extension) and monitor_size for monitors. Basically we need to calculate
1525   // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
1526   //
1527   //
1528   // The big complicating thing here is that we must ensure that the stack stays properly
1529   // aligned. This would be even uglier if monitor size wasn't modulo what the stack
1530   // needs to be aligned for). We are given that the sp (fp) is already aligned by
1531   // the caller so we must ensure that it is properly aligned for our callee.
1532   //
1533   const int rounded_vm_local_words =
1534        round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1535   // callee_locals and max_stack are counts, not the size in frame.
1536   const int locals_size =
1537        round_to(callee_extra_locals * Interpreter::stackElementWords, WordsPerLong);
1538   const int max_stack_words = max_stack * Interpreter::stackElementWords;
1539   return (round_to((max_stack_words
1540                    //6815692//+ methodOopDesc::extra_stack_words()
1541                    + rounded_vm_local_words
1542                    + frame::memory_parameter_word_sp_offset), WordsPerLong)
1543                    // already rounded
1544                    + locals_size + monitor_size);
1545 }
1546 
1547 // How much stack a method top interpreter activation needs in words.
1548 int AbstractInterpreter::size_top_interpreter_activation(methodOop method) {
1549 
1550   // See call_stub code
1551   int call_stub_size  = round_to(7 + frame::memory_parameter_word_sp_offset,
1552                                  WordsPerLong);    // 7 + register save area
1553 
1554   // Save space for one monitor to get into the interpreted method in case
1555   // the method is synchronized
1556   int monitor_size    = method->is_synchronized() ?
1557                                 1*frame::interpreter_frame_monitor_size() : 0;
1558   return size_activation_helper(method->max_locals(), method->max_stack(),
1559                                  monitor_size) + call_stub_size;
1560 }
1561 
1562 int AbstractInterpreter::layout_activation(methodOop method,
1563                                            int tempcount,
1564                                            int popframe_extra_args,
1565                                            int moncount,
1566                                            int caller_actual_parameters,
1567                                            int callee_param_count,
1568                                            int callee_local_count,
1569                                            frame* caller,
1570                                            frame* interpreter_frame,
1571                                            bool is_top_frame) {
1572   // Note: This calculation must exactly parallel the frame setup
1573   // in InterpreterGenerator::generate_fixed_frame.
1574   // If f!=NULL, set up the following variables:
1575   //   - Lmethod
1576   //   - Llocals
1577   //   - Lmonitors (to the indicated number of monitors)
1578   //   - Lesp (to the indicated number of temps)
1579   // The frame f (if not NULL) on entry is a description of the caller of the frame
1580   // we are about to layout. We are guaranteed that we will be able to fill in a
1581   // new interpreter frame as its callee (i.e. the stack space is allocated and
1582   // the amount was determined by an earlier call to this method with f == NULL).
1583   // On return f (if not NULL) while describe the interpreter frame we just layed out.
1584 
1585   int monitor_size           = moncount * frame::interpreter_frame_monitor_size();
1586   int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1587 
1588   assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1589   //
1590   // Note: if you look closely this appears to be doing something much different
1591   // than generate_fixed_frame. What is happening is this. On sparc we have to do
1592   // this dance with interpreter_sp_adjustment because the window save area would
1593   // appear just below the bottom (tos) of the caller's java expression stack. Because
1594   // the interpreter want to have the locals completely contiguous generate_fixed_frame
1595   // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size).
1596   // Now in generate_fixed_frame the extension of the caller's sp happens in the callee.
1597   // In this code the opposite occurs the caller adjusts it's own stack base on the callee.
1598   // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest)
1599   // because the oldest frame would have adjust its callers frame and yet that frame
1600   // already exists and isn't part of this array of frames we are unpacking. So at first
1601   // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper()
1602   // will after it calculates all of the frame's on_stack_size()'s will then figure out the
1603   // amount to adjust the caller of the initial (oldest) frame and the calculation will all
1604   // add up. It does seem like it simpler to account for the adjustment here (and remove the
1605   // callee... parameters here). However this would mean that this routine would have to take
1606   // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment)
1607   // and run the calling loop in the reverse order. This would also would appear to mean making
1608   // this code aware of what the interactions are when that initial caller fram was an osr or
1609   // other adapter frame. deoptimization is complicated enough and  hard enough to debug that
1610   // there is no sense in messing working code.
1611   //
1612 
1613   int rounded_cls = round_to((callee_local_count - callee_param_count), WordsPerLong);
1614   assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align");
1615 
1616   int raw_frame_size = size_activation_helper(rounded_cls, method->max_stack(),
1617                                               monitor_size);
1618 
1619   if (interpreter_frame != NULL) {
1620     // The skeleton frame must already look like an interpreter frame
1621     // even if not fully filled out.
1622     assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame");
1623 
1624     intptr_t* fp = interpreter_frame->fp();
1625 
1626     JavaThread* thread = JavaThread::current();
1627     RegisterMap map(thread, false);
1628     // More verification that skeleton frame is properly walkable
1629     assert(fp == caller->sp(), "fp must match");
1630 
1631     intptr_t* montop     = fp - rounded_vm_local_words;
1632 
1633     // preallocate monitors (cf. __ add_monitor_to_stack)
1634     intptr_t* monitors = montop - monitor_size;
1635 
1636     // preallocate stack space
1637     intptr_t*  esp = monitors - 1 -
1638                      (tempcount * Interpreter::stackElementWords) -
1639                      popframe_extra_args;
1640 
1641     int local_words = method->max_locals() * Interpreter::stackElementWords;
1642     NEEDS_CLEANUP;
1643     intptr_t* locals;
1644     if (caller->is_interpreted_frame()) {
1645       // Can force the locals area to end up properly overlapping the top of the expression stack.
1646       intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1;
1647       // Note that this computation means we replace size_of_parameters() values from the caller
1648       // interpreter frame's expression stack with our argument locals
1649       int parm_words  = caller_actual_parameters * Interpreter::stackElementWords;
1650       locals = Lesp_ptr + parm_words;
1651       int delta = local_words - parm_words;
1652       int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0;
1653       *interpreter_frame->register_addr(I5_savedSP)    = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS;
1654     } else {
1655       assert(caller->is_compiled_frame() || caller->is_entry_frame() || caller->is_ricochet_frame(), "only possible cases");
1656       // Don't have Lesp available; lay out locals block in the caller
1657       // adjacent to the register window save area.
1658       //
1659       // Compiled frames do not allocate a varargs area which is why this if
1660       // statement is needed.
1661       //
1662       if (caller->is_compiled_frame()) {
1663         locals = fp + frame::register_save_words + local_words - 1;
1664       } else {
1665         locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
1666       }
1667       if (!caller->is_entry_frame()) {
1668         // Caller wants his own SP back
1669         int caller_frame_size = caller->cb()->frame_size();
1670         *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS;
1671       }
1672     }
1673     if (TraceDeoptimization) {
1674       if (caller->is_entry_frame()) {
1675         // make sure I5_savedSP and the entry frames notion of saved SP
1676         // agree.  This assertion duplicate a check in entry frame code
1677         // but catches the failure earlier.
1678         assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP),
1679                "would change callers SP");
1680       }
1681       if (caller->is_entry_frame()) {
1682         tty->print("entry ");
1683       }
1684       if (caller->is_compiled_frame()) {
1685         tty->print("compiled ");
1686         if (caller->is_deoptimized_frame()) {
1687           tty->print("(deopt) ");
1688         }
1689       }
1690       if (caller->is_interpreted_frame()) {
1691         tty->print("interpreted ");
1692       }
1693       tty->print_cr("caller fp=0x%x sp=0x%x", caller->fp(), caller->sp());
1694       tty->print_cr("save area = 0x%x, 0x%x", caller->sp(), caller->sp() + 16);
1695       tty->print_cr("save area = 0x%x, 0x%x", caller->fp(), caller->fp() + 16);
1696       tty->print_cr("interpreter fp=0x%x sp=0x%x", interpreter_frame->fp(), interpreter_frame->sp());
1697       tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->sp(), interpreter_frame->sp() + 16);
1698       tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->fp(), interpreter_frame->fp() + 16);
1699       tty->print_cr("Llocals = 0x%x", locals);
1700       tty->print_cr("Lesp = 0x%x", esp);
1701       tty->print_cr("Lmonitors = 0x%x", monitors);
1702     }
1703 
1704     if (method->max_locals() > 0) {
1705       assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area");
1706       assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area");
1707       assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area");
1708       assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area");
1709     }
1710 #ifdef _LP64
1711     assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd");
1712 #endif
1713 
1714     *interpreter_frame->register_addr(Lmethod)     = (intptr_t) method;
1715     *interpreter_frame->register_addr(Llocals)     = (intptr_t) locals;
1716     *interpreter_frame->register_addr(Lmonitors)   = (intptr_t) monitors;
1717     *interpreter_frame->register_addr(Lesp)        = (intptr_t) esp;
1718     // Llast_SP will be same as SP as there is no adapter space
1719     *interpreter_frame->register_addr(Llast_SP)    = (intptr_t) interpreter_frame->sp() - STACK_BIAS;
1720     *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache();
1721 #ifdef FAST_DISPATCH
1722     *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table();
1723 #endif
1724 
1725 
1726 #ifdef ASSERT
1727     BasicObjectLock* mp = (BasicObjectLock*)monitors;
1728 
1729     assert(interpreter_frame->interpreter_frame_method() == method, "method matches");
1730     assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize)), "locals match");
1731     assert(interpreter_frame->interpreter_frame_monitor_end()   == mp, "monitor_end matches");
1732     assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches");
1733     assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches");
1734 
1735     // check bounds
1736     intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1);
1737     intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words;
1738     assert(lo < monitors && montop <= hi, "monitors in bounds");
1739     assert(lo <= esp && esp < monitors, "esp in bounds");
1740 #endif // ASSERT
1741   }
1742 
1743   return raw_frame_size;
1744 }
1745 
1746 //----------------------------------------------------------------------------------------------------
1747 // Exceptions
1748 void TemplateInterpreterGenerator::generate_throw_exception() {
1749 
1750   // Entry point in previous activation (i.e., if the caller was interpreted)
1751   Interpreter::_rethrow_exception_entry = __ pc();
1752   // O0: exception
1753 
1754   // entry point for exceptions thrown within interpreter code
1755   Interpreter::_throw_exception_entry = __ pc();
1756   __ verify_thread();
1757   // expression stack is undefined here
1758   // O0: exception, i.e. Oexception
1759   // Lbcp: exception bcx
1760   __ verify_oop(Oexception);
1761 
1762 
1763   // expression stack must be empty before entering the VM in case of an exception
1764   __ empty_expression_stack();
1765   // find exception handler address and preserve exception oop
1766   // call C routine to find handler and jump to it
1767   __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1768   __ push_ptr(O1); // push exception for exception handler bytecodes
1769 
1770   __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1771   __ delayed()->nop();
1772 
1773 
1774   // if the exception is not handled in the current frame
1775   // the frame is removed and the exception is rethrown
1776   // (i.e. exception continuation is _rethrow_exception)
1777   //
1778   // Note: At this point the bci is still the bxi for the instruction which caused
1779   //       the exception and the expression stack is empty. Thus, for any VM calls
1780   //       at this point, GC will find a legal oop map (with empty expression stack).
1781 
1782   // in current activation
1783   // tos: exception
1784   // Lbcp: exception bcp
1785 
1786   //
1787   // JVMTI PopFrame support
1788   //
1789 
1790   Interpreter::_remove_activation_preserving_args_entry = __ pc();
1791   Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1792   // Set the popframe_processing bit in popframe_condition indicating that we are
1793   // currently handling popframe, so that call_VMs that may happen later do not trigger new
1794   // popframe handling cycles.
1795 
1796   __ ld(popframe_condition_addr, G3_scratch);
1797   __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1798   __ stw(G3_scratch, popframe_condition_addr);
1799 
1800   // Empty the expression stack, as in normal exception handling
1801   __ empty_expression_stack();
1802   __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1803 
1804   {
1805     // Check to see whether we are returning to a deoptimized frame.
1806     // (The PopFrame call ensures that the caller of the popped frame is
1807     // either interpreted or compiled and deoptimizes it if compiled.)
1808     // In this case, we can't call dispatch_next() after the frame is
1809     // popped, but instead must save the incoming arguments and restore
1810     // them after deoptimization has occurred.
1811     //
1812     // Note that we don't compare the return PC against the
1813     // deoptimization blob's unpack entry because of the presence of
1814     // adapter frames in C2.
1815     Label caller_not_deoptimized;
1816     __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1817     __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1818 
1819     const Register Gtmp1 = G3_scratch;
1820     const Register Gtmp2 = G1_scratch;
1821 
1822     // Compute size of arguments for saving when returning to deoptimized caller
1823     __ lduh(Lmethod, in_bytes(methodOopDesc::size_of_parameters_offset()), Gtmp1);
1824     __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1825     __ sub(Llocals, Gtmp1, Gtmp2);
1826     __ add(Gtmp2, wordSize, Gtmp2);
1827     // Save these arguments
1828     __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1829     // Inform deoptimization that it is responsible for restoring these arguments
1830     __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1831     Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1832     __ st(Gtmp1, popframe_condition_addr);
1833 
1834     // Return from the current method
1835     // The caller's SP was adjusted upon method entry to accomodate
1836     // the callee's non-argument locals. Undo that adjustment.
1837     __ ret();
1838     __ delayed()->restore(I5_savedSP, G0, SP);
1839 
1840     __ bind(caller_not_deoptimized);
1841   }
1842 
1843   // Clear the popframe condition flag
1844   __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1845 
1846   // Get out of the current method (how this is done depends on the particular compiler calling
1847   // convention that the interpreter currently follows)
1848   // The caller's SP was adjusted upon method entry to accomodate
1849   // the callee's non-argument locals. Undo that adjustment.
1850   __ restore(I5_savedSP, G0, SP);
1851   // The method data pointer was incremented already during
1852   // call profiling. We have to restore the mdp for the current bcp.
1853   if (ProfileInterpreter) {
1854     __ set_method_data_pointer_for_bcp();
1855   }
1856   // Resume bytecode interpretation at the current bcp
1857   __ dispatch_next(vtos);
1858   // end of JVMTI PopFrame support
1859 
1860   Interpreter::_remove_activation_entry = __ pc();
1861 
1862   // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1863   __ pop_ptr(Oexception);                                  // get exception
1864 
1865   // Intel has the following comment:
1866   //// remove the activation (without doing throws on illegalMonitorExceptions)
1867   // They remove the activation without checking for bad monitor state.
1868   // %%% We should make sure this is the right semantics before implementing.
1869 
1870   // %%% changed set_vm_result_2 to set_vm_result and get_vm_result_2 to get_vm_result. Is there a bug here?
1871   __ set_vm_result(Oexception);
1872   __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1873 
1874   __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1875 
1876   __ get_vm_result(Oexception);
1877   __ verify_oop(Oexception);
1878 
1879     const int return_reg_adjustment = frame::pc_return_offset;
1880   Address issuing_pc_addr(I7, return_reg_adjustment);
1881 
1882   // We are done with this activation frame; find out where to go next.
1883   // The continuation point will be an exception handler, which expects
1884   // the following registers set up:
1885   //
1886   // Oexception: exception
1887   // Oissuing_pc: the local call that threw exception
1888   // Other On: garbage
1889   // In/Ln:  the contents of the caller's register window
1890   //
1891   // We do the required restore at the last possible moment, because we
1892   // need to preserve some state across a runtime call.
1893   // (Remember that the caller activation is unknown--it might not be
1894   // interpreted, so things like Lscratch are useless in the caller.)
1895 
1896   // Although the Intel version uses call_C, we can use the more
1897   // compact call_VM.  (The only real difference on SPARC is a
1898   // harmlessly ignored [re]set_last_Java_frame, compared with
1899   // the Intel code which lacks this.)
1900   __ mov(Oexception,      Oexception ->after_save());  // get exception in I0 so it will be on O0 after restore
1901   __ add(issuing_pc_addr, Oissuing_pc->after_save());  // likewise set I1 to a value local to the caller
1902   __ super_call_VM_leaf(L7_thread_cache,
1903                         CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1904                         G2_thread, Oissuing_pc->after_save());
1905 
1906   // The caller's SP was adjusted upon method entry to accomodate
1907   // the callee's non-argument locals. Undo that adjustment.
1908   __ JMP(O0, 0);                         // return exception handler in caller
1909   __ delayed()->restore(I5_savedSP, G0, SP);
1910 
1911   // (same old exception object is already in Oexception; see above)
1912   // Note that an "issuing PC" is actually the next PC after the call
1913 }
1914 
1915 
1916 //
1917 // JVMTI ForceEarlyReturn support
1918 //
1919 
1920 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1921   address entry = __ pc();
1922 
1923   __ empty_expression_stack();
1924   __ load_earlyret_value(state);
1925 
1926   __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1927   Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1928 
1929   // Clear the earlyret state
1930   __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1931 
1932   __ remove_activation(state,
1933                        /* throw_monitor_exception */ false,
1934                        /* install_monitor_exception */ false);
1935 
1936   // The caller's SP was adjusted upon method entry to accomodate
1937   // the callee's non-argument locals. Undo that adjustment.
1938   __ ret();                             // return to caller
1939   __ delayed()->restore(I5_savedSP, G0, SP);
1940 
1941   return entry;
1942 } // end of JVMTI ForceEarlyReturn support
1943 
1944 
1945 //------------------------------------------------------------------------------------------------------------------------
1946 // Helper for vtos entry point generation
1947 
1948 void TemplateInterpreterGenerator::set_vtos_entry_points(Template* t, address& bep, address& cep, address& sep, address& aep, address& iep, address& lep, address& fep, address& dep, address& vep) {
1949   assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1950   Label L;
1951   aep = __ pc(); __ push_ptr(); __ ba_short(L);
1952   fep = __ pc(); __ push_f();   __ ba_short(L);
1953   dep = __ pc(); __ push_d();   __ ba_short(L);
1954   lep = __ pc(); __ push_l();   __ ba_short(L);
1955   iep = __ pc(); __ push_i();
1956   bep = cep = sep = iep;                        // there aren't any
1957   vep = __ pc(); __ bind(L);                    // fall through
1958   generate_and_dispatch(t);
1959 }
1960 
1961 // --------------------------------------------------------------------------------
1962 
1963 
1964 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
1965  : TemplateInterpreterGenerator(code) {
1966    generate_all(); // down here so it can be "virtual"
1967 }
1968 
1969 // --------------------------------------------------------------------------------
1970 
1971 // Non-product code
1972 #ifndef PRODUCT
1973 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
1974   address entry = __ pc();
1975 
1976   __ push(state);
1977   __ mov(O7, Lscratch); // protect return address within interpreter
1978 
1979   // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
1980   __ mov( Otos_l2, G3_scratch );
1981   __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
1982   __ mov(Lscratch, O7); // restore return address
1983   __ pop(state);
1984   __ retl();
1985   __ delayed()->nop();
1986 
1987   return entry;
1988 }
1989 
1990 
1991 // helpers for generate_and_dispatch
1992 
1993 void TemplateInterpreterGenerator::count_bytecode() {
1994   __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
1995 }
1996 
1997 
1998 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
1999   __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
2000 }
2001 
2002 
2003 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2004   AddressLiteral index   (&BytecodePairHistogram::_index);
2005   AddressLiteral counters((address) &BytecodePairHistogram::_counters);
2006 
2007   // get index, shift out old bytecode, bring in new bytecode, and store it
2008   // _index = (_index >> log2_number_of_codes) |
2009   //          (bytecode << log2_number_of_codes);
2010 
2011   __ load_contents(index, G4_scratch);
2012   __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
2013   __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes,  G3_scratch );
2014   __ or3( G3_scratch,  G4_scratch, G4_scratch );
2015   __ store_contents(G4_scratch, index, G3_scratch);
2016 
2017   // bump bucket contents
2018   // _counters[_index] ++;
2019 
2020   __ set(counters, G3_scratch);                       // loads into G3_scratch
2021   __ sll( G4_scratch, LogBytesPerWord, G4_scratch );  // Index is word address
2022   __ add (G3_scratch, G4_scratch, G3_scratch);        // Add in index
2023   __ ld (G3_scratch, 0, G4_scratch);
2024   __ inc (G4_scratch);
2025   __ st (G4_scratch, 0, G3_scratch);
2026 }
2027 
2028 
2029 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2030   // Call a little run-time stub to avoid blow-up for each bytecode.
2031   // The run-time runtime saves the right registers, depending on
2032   // the tosca in-state for the given template.
2033   address entry = Interpreter::trace_code(t->tos_in());
2034   guarantee(entry != NULL, "entry must have been generated");
2035   __ call(entry, relocInfo::none);
2036   __ delayed()->nop();
2037 }
2038 
2039 
2040 void TemplateInterpreterGenerator::stop_interpreter_at() {
2041   AddressLiteral counter(&BytecodeCounter::_counter_value);
2042   __ load_contents(counter, G3_scratch);
2043   AddressLiteral stop_at(&StopInterpreterAt);
2044   __ load_ptr_contents(stop_at, G4_scratch);
2045   __ cmp(G3_scratch, G4_scratch);
2046   __ breakpoint_trap(Assembler::equal);
2047 }
2048 #endif // not PRODUCT
2049 #endif // !CC_INTERP