1 /*
   2  * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/macroAssembler.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/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/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 #include "utilities/macros.hpp"
  48 
  49 #ifndef CC_INTERP
  50 #ifndef FAST_DISPATCH
  51 #define FAST_DISPATCH 1
  52 #endif
  53 #undef FAST_DISPATCH
  54 
  55 
  56 // Generation of Interpreter
  57 //
  58 // The InterpreterGenerator generates the interpreter into Interpreter::_code.
  59 
  60 
  61 #define __ _masm->
  62 
  63 
  64 //----------------------------------------------------------------------------------------------------
  65 
  66 
  67 void InterpreterGenerator::save_native_result(void) {
  68   // result potentially in O0/O1: save it across calls
  69   const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
  70 
  71   // result potentially in F0/F1: save it across calls
  72   const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
  73 
  74   // save and restore any potential method result value around the unlocking operation
  75   __ stf(FloatRegisterImpl::D, F0, d_tmp);
  76 #ifdef _LP64
  77   __ stx(O0, l_tmp);
  78 #else
  79   __ std(O0, l_tmp);
  80 #endif
  81 }
  82 
  83 void InterpreterGenerator::restore_native_result(void) {
  84   const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
  85   const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
  86 
  87   // Restore any method result value
  88   __ ldf(FloatRegisterImpl::D, d_tmp, F0);
  89 #ifdef _LP64
  90   __ ldx(l_tmp, O0);
  91 #else
  92   __ ldd(l_tmp, O0);
  93 #endif
  94 }
  95 
  96 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
  97   assert(!pass_oop || message == NULL, "either oop or message but not both");
  98   address entry = __ pc();
  99   // expression stack must be empty before entering the VM if an exception happened
 100   __ empty_expression_stack();
 101   // load exception object
 102   __ set((intptr_t)name, G3_scratch);
 103   if (pass_oop) {
 104     __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
 105   } else {
 106     __ set((intptr_t)message, G4_scratch);
 107     __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
 108   }
 109   // throw exception
 110   assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
 111   AddressLiteral thrower(Interpreter::throw_exception_entry());
 112   __ jump_to(thrower, G3_scratch);
 113   __ delayed()->nop();
 114   return entry;
 115 }
 116 
 117 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
 118   address entry = __ pc();
 119   // expression stack must be empty before entering the VM if an exception
 120   // happened
 121   __ empty_expression_stack();
 122   // load exception object
 123   __ call_VM(Oexception,
 124              CAST_FROM_FN_PTR(address,
 125                               InterpreterRuntime::throw_ClassCastException),
 126              Otos_i);
 127   __ should_not_reach_here();
 128   return entry;
 129 }
 130 
 131 
 132 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
 133   address entry = __ pc();
 134   // expression stack must be empty before entering the VM if an exception happened
 135   __ empty_expression_stack();
 136   // convention: expect aberrant index in register G3_scratch, then shuffle the
 137   // index to G4_scratch for the VM call
 138   __ mov(G3_scratch, G4_scratch);
 139   __ set((intptr_t)name, G3_scratch);
 140   __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
 141   __ should_not_reach_here();
 142   return entry;
 143 }
 144 
 145 
 146 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
 147   address entry = __ pc();
 148   // expression stack must be empty before entering the VM if an exception happened
 149   __ empty_expression_stack();
 150   __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
 151   __ should_not_reach_here();
 152   return entry;
 153 }
 154 
 155 
 156 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
 157   address entry = __ pc();
 158 
 159 #if !defined(_LP64) && defined(COMPILER2)
 160   // All return values are where we want them, except for Longs.  C2 returns
 161   // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
 162   // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
 163   // build even if we are returning from interpreted we just do a little
 164   // stupid shuffing.
 165   // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
 166   // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
 167   // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
 168 
 169   if (state == ltos) {
 170     __ srl (G1,  0, O1);
 171     __ srlx(G1, 32, O0);
 172   }
 173 #endif // !_LP64 && COMPILER2
 174 
 175   // The callee returns with the stack possibly adjusted by adapter transition
 176   // We remove that possible adjustment here.
 177   // All interpreter local registers are untouched. Any result is passed back
 178   // in the O0/O1 or float registers. Before continuing, the arguments must be
 179   // popped from the java expression stack; i.e., Lesp must be adjusted.
 180 
 181   __ mov(Llast_SP, SP);   // Remove any adapter added stack space.
 182 
 183   const Register cache = G3_scratch;
 184   const Register index  = G1_scratch;
 185   __ get_cache_and_index_at_bcp(cache, index, 1, index_size);
 186 
 187   const Register flags = cache;
 188   __ ld_ptr(cache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset(), flags);
 189   const Register parameter_size = flags;
 190   __ and3(flags, ConstantPoolCacheEntry::parameter_size_mask, parameter_size);  // argument size in words
 191   __ sll(parameter_size, Interpreter::logStackElementSize, parameter_size);     // each argument size in bytes
 192   __ add(Lesp, parameter_size, Lesp);                                           // pop arguments
 193   __ dispatch_next(state, step);
 194 
 195   return entry;
 196 }
 197 
 198 
 199 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
 200   address entry = __ pc();
 201   __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
 202   { Label L;
 203     Address exception_addr(G2_thread, Thread::pending_exception_offset());
 204     __ ld_ptr(exception_addr, Gtemp);  // Load pending exception.
 205     __ br_null_short(Gtemp, Assembler::pt, L);
 206     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
 207     __ should_not_reach_here();
 208     __ bind(L);
 209   }
 210   __ dispatch_next(state, step);
 211   return entry;
 212 }
 213 
 214 // A result handler converts/unboxes a native call result into
 215 // a java interpreter/compiler result. The current frame is an
 216 // interpreter frame. The activation frame unwind code must be
 217 // consistent with that of TemplateTable::_return(...). In the
 218 // case of native methods, the caller's SP was not modified.
 219 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
 220   address entry = __ pc();
 221   Register Itos_i  = Otos_i ->after_save();
 222   Register Itos_l  = Otos_l ->after_save();
 223   Register Itos_l1 = Otos_l1->after_save();
 224   Register Itos_l2 = Otos_l2->after_save();
 225   switch (type) {
 226     case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
 227     case T_CHAR   : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i);   break; // cannot use and3, 0xFFFF too big as immediate value!
 228     case T_BYTE   : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i);   break;
 229     case T_SHORT  : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i);   break;
 230     case T_LONG   :
 231 #ifndef _LP64
 232                     __ mov(O1, Itos_l2);  // move other half of long
 233 #endif              // ifdef or no ifdef, fall through to the T_INT case
 234     case T_INT    : __ mov(O0, Itos_i);                         break;
 235     case T_VOID   : /* nothing to do */                         break;
 236     case T_FLOAT  : assert(F0 == Ftos_f, "fix this code" );     break;
 237     case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" );     break;
 238     case T_OBJECT :
 239       __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
 240       __ verify_oop(Itos_i);
 241       break;
 242     default       : ShouldNotReachHere();
 243   }
 244   __ ret();                           // return from interpreter activation
 245   __ delayed()->restore(I5_savedSP, G0, SP);  // remove interpreter frame
 246   NOT_PRODUCT(__ emit_int32(0);)       // marker for disassembly
 247   return entry;
 248 }
 249 
 250 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
 251   address entry = __ pc();
 252   __ push(state);
 253   __ call_VM(noreg, runtime_entry);
 254   __ dispatch_via(vtos, Interpreter::normal_table(vtos));
 255   return entry;
 256 }
 257 
 258 
 259 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
 260   address entry = __ pc();
 261   __ dispatch_next(state);
 262   return entry;
 263 }
 264 
 265 //
 266 // Helpers for commoning out cases in the various type of method entries.
 267 //
 268 
 269 // increment invocation count & check for overflow
 270 //
 271 // Note: checking for negative value instead of overflow
 272 //       so we have a 'sticky' overflow test
 273 //
 274 // Lmethod: method
 275 // ??: invocation counter
 276 //
 277 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
 278   // Note: In tiered we increment either counters in MethodCounters* or in
 279   // MDO depending if we're profiling or not.
 280   const Register Rcounters = G3_scratch;
 281   Label done;
 282 
 283   if (TieredCompilation) {
 284     const int increment = InvocationCounter::count_increment;
 285     const int mask = ((1 << Tier0InvokeNotifyFreqLog) - 1) << InvocationCounter::count_shift;
 286     Label no_mdo;
 287     if (ProfileInterpreter) {
 288       // If no method data exists, go to profile_continue.
 289       __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
 290       __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
 291       // Increment counter
 292       Address mdo_invocation_counter(G4_scratch,
 293                                      in_bytes(MethodData::invocation_counter_offset()) +
 294                                      in_bytes(InvocationCounter::counter_offset()));
 295       __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
 296                                  G3_scratch, Lscratch,
 297                                  Assembler::zero, overflow);
 298       __ ba_short(done);
 299     }
 300 
 301     // Increment counter in MethodCounters*
 302     __ bind(no_mdo);
 303     Address invocation_counter(Rcounters,
 304             in_bytes(MethodCounters::invocation_counter_offset()) +
 305             in_bytes(InvocationCounter::counter_offset()));
 306     __ get_method_counters(Lmethod, Rcounters, done);
 307     __ increment_mask_and_jump(invocation_counter, increment, mask,
 308                                G4_scratch, Lscratch,
 309                                Assembler::zero, overflow);
 310     __ bind(done);
 311   } else {
 312     // Update standard invocation counters
 313     __ get_method_counters(Lmethod, Rcounters, done);
 314     __ increment_invocation_counter(Rcounters, O0, G4_scratch);
 315     if (ProfileInterpreter) {
 316       Address interpreter_invocation_counter(Rcounters,
 317             in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
 318       __ ld(interpreter_invocation_counter, G4_scratch);
 319       __ inc(G4_scratch);
 320       __ st(G4_scratch, interpreter_invocation_counter);
 321     }
 322 
 323     if (ProfileInterpreter && profile_method != NULL) {
 324       // Test to see if we should create a method data oop
 325       AddressLiteral profile_limit((address)&InvocationCounter::InterpreterProfileLimit);
 326       __ load_contents(profile_limit, G3_scratch);
 327       __ cmp_and_br_short(O0, G3_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
 328 
 329       // if no method data exists, go to profile_method
 330       __ test_method_data_pointer(*profile_method);
 331     }
 332 
 333     AddressLiteral invocation_limit((address)&InvocationCounter::InterpreterInvocationLimit);
 334     __ load_contents(invocation_limit, G3_scratch);
 335     __ cmp(O0, G3_scratch);
 336     __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
 337     __ delayed()->nop();
 338     __ bind(done);
 339   }
 340 
 341 }
 342 
 343 // Allocate monitor and lock method (asm interpreter)
 344 // ebx - Method*
 345 //
 346 void InterpreterGenerator::lock_method(void) {
 347   __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0);  // Load access flags.
 348 
 349 #ifdef ASSERT
 350  { Label ok;
 351    __ btst(JVM_ACC_SYNCHRONIZED, O0);
 352    __ br( Assembler::notZero, false, Assembler::pt, ok);
 353    __ delayed()->nop();
 354    __ stop("method doesn't need synchronization");
 355    __ bind(ok);
 356   }
 357 #endif // ASSERT
 358 
 359   // get synchronization object to O0
 360   { Label done;
 361     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
 362     __ btst(JVM_ACC_STATIC, O0);
 363     __ br( Assembler::zero, true, Assembler::pt, done);
 364     __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
 365 
 366     __ ld_ptr( Lmethod, in_bytes(Method::const_offset()), O0);
 367     __ ld_ptr( O0, in_bytes(ConstMethod::constants_offset()), O0);
 368     __ ld_ptr( O0, ConstantPool::pool_holder_offset_in_bytes(), O0);
 369 
 370     // lock the mirror, not the Klass*
 371     __ ld_ptr( O0, mirror_offset, O0);
 372 
 373 #ifdef ASSERT
 374     __ tst(O0);
 375     __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
 376 #endif // ASSERT
 377 
 378     __ bind(done);
 379   }
 380 
 381   __ add_monitor_to_stack(true, noreg, noreg);  // allocate monitor elem
 382   __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes());   // store object
 383   // __ untested("lock_object from method entry");
 384   __ lock_object(Lmonitors, O0);
 385 }
 386 
 387 
 388 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
 389                                                          Register Rscratch,
 390                                                          Register Rscratch2) {
 391   const int page_size = os::vm_page_size();
 392   Label after_frame_check;
 393 
 394   assert_different_registers(Rframe_size, Rscratch, Rscratch2);
 395 
 396   __ set(page_size, Rscratch);
 397   __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
 398 
 399   // get the stack base, and in debug, verify it is non-zero
 400   __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch );
 401 #ifdef ASSERT
 402   Label base_not_zero;
 403   __ br_notnull_short(Rscratch, Assembler::pn, base_not_zero);
 404   __ stop("stack base is zero in generate_stack_overflow_check");
 405   __ bind(base_not_zero);
 406 #endif
 407 
 408   // get the stack size, and in debug, verify it is non-zero
 409   assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" );
 410   __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 );
 411 #ifdef ASSERT
 412   Label size_not_zero;
 413   __ br_notnull_short(Rscratch2, Assembler::pn, size_not_zero);
 414   __ stop("stack size is zero in generate_stack_overflow_check");
 415   __ bind(size_not_zero);
 416 #endif
 417 
 418   // compute the beginning of the protected zone minus the requested frame size
 419   __ sub( Rscratch, Rscratch2,   Rscratch );
 420   __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 );
 421   __ add( Rscratch, Rscratch2,   Rscratch );
 422 
 423   // Add in the size of the frame (which is the same as subtracting it from the
 424   // SP, which would take another register
 425   __ add( Rscratch, Rframe_size, Rscratch );
 426 
 427   // the frame is greater than one page in size, so check against
 428   // the bottom of the stack
 429   __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check);
 430 
 431   // the stack will overflow, throw an exception
 432 
 433   // Note that SP is restored to sender's sp (in the delay slot). This
 434   // is necessary if the sender's frame is an extended compiled frame
 435   // (see gen_c2i_adapter()) and safer anyway in case of JSR292
 436   // adaptations.
 437 
 438   // Note also that the restored frame is not necessarily interpreted.
 439   // Use the shared runtime version of the StackOverflowError.
 440   assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
 441   AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
 442   __ jump_to(stub, Rscratch);
 443   __ delayed()->mov(O5_savedSP, SP);
 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   int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
 490 
 491   const int extra_space =
 492     rounded_vm_local_words +                   // frame local scratch space
 493     Method::extra_stack_entries() +            // extra stack for jsr 292
 494     frame::memory_parameter_word_sp_offset +   // register save area
 495     (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
 496 
 497   const Register Glocals_size = G3;
 498   const Register RconstMethod = Glocals_size;
 499   const Register Otmp1 = O3;
 500   const Register Otmp2 = O4;
 501   // Lscratch can't be used as a temporary because the call_stub uses
 502   // it to assert that the stack frame was setup correctly.
 503   const Address constMethod       (G5_method, Method::const_offset());
 504   const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
 505 
 506   __ ld_ptr( constMethod, RconstMethod );
 507   __ lduh( size_of_parameters, Glocals_size);
 508 
 509   // Gargs points to first local + BytesPerWord
 510   // Set the saved SP after the register window save
 511   //
 512   assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
 513   __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
 514   __ add(Gargs, Otmp1, Gargs);
 515 
 516   if (native_call) {
 517     __ calc_mem_param_words( Glocals_size, Gframe_size );
 518     __ add( Gframe_size,  extra_space, Gframe_size);
 519     __ round_to( Gframe_size, WordsPerLong );
 520     __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
 521   } else {
 522 
 523     //
 524     // Compute number of locals in method apart from incoming parameters
 525     //
 526     const Address size_of_locals    (Otmp1, ConstMethod::size_of_locals_offset());
 527     __ ld_ptr( constMethod, Otmp1 );
 528     __ lduh( size_of_locals, Otmp1 );
 529     __ sub( Otmp1, Glocals_size, Glocals_size );
 530     __ round_to( Glocals_size, WordsPerLong );
 531     __ sll( Glocals_size, Interpreter::logStackElementSize, Glocals_size );
 532 
 533     // see if the frame is greater than one page in size. If so,
 534     // then we need to verify there is enough stack space remaining
 535     // Frame_size = (max_stack + extra_space) * BytesPerWord;
 536     __ ld_ptr( constMethod, Gframe_size );
 537     __ lduh( Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size );
 538     __ add( Gframe_size, extra_space, Gframe_size );
 539     __ round_to( Gframe_size, WordsPerLong );
 540     __ sll( Gframe_size, Interpreter::logStackElementSize, Gframe_size);
 541 
 542     // Add in java locals size for stack overflow check only
 543     __ add( Gframe_size, Glocals_size, Gframe_size );
 544 
 545     const Register Otmp2 = O4;
 546     assert_different_registers(Otmp1, Otmp2, O5_savedSP);
 547     generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2);
 548 
 549     __ sub( Gframe_size, Glocals_size, Gframe_size);
 550 
 551     //
 552     // bump SP to accomodate the extra locals
 553     //
 554     __ sub( SP, Glocals_size, SP );
 555   }
 556 
 557   //
 558   // now set up a stack frame with the size computed above
 559   //
 560   __ neg( Gframe_size );
 561   __ save( SP, Gframe_size, SP );
 562 
 563   //
 564   // now set up all the local cache registers
 565   //
 566   // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
 567   // that all present references to Lbyte_code initialize the register
 568   // immediately before use
 569   if (native_call) {
 570     __ mov(G0, Lbcp);
 571   } else {
 572     __ ld_ptr(G5_method, Method::const_offset(), Lbcp);
 573     __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp);
 574   }
 575   __ mov( G5_method, Lmethod);                 // set Lmethod
 576   __ get_constant_pool_cache( LcpoolCache );   // set LcpoolCache
 577   __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
 578 #ifdef _LP64
 579   __ add( Lmonitors, STACK_BIAS, Lmonitors );   // Account for 64 bit stack bias
 580 #endif
 581   __ sub(Lmonitors, BytesPerWord, Lesp);       // set Lesp
 582 
 583   // setup interpreter activation registers
 584   __ sub(Gargs, BytesPerWord, Llocals);        // set Llocals
 585 
 586   if (ProfileInterpreter) {
 587 #ifdef FAST_DISPATCH
 588     // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
 589     // they both use I2.
 590     assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
 591 #endif // FAST_DISPATCH
 592     __ set_method_data_pointer();
 593   }
 594 
 595 }
 596 
 597 // Empty method, generate a very fast return.
 598 
 599 address InterpreterGenerator::generate_empty_entry(void) {
 600 
 601   // A method that does nother but return...
 602 
 603   address entry = __ pc();
 604   Label slow_path;
 605 
 606   // do nothing for empty methods (do not even increment invocation counter)
 607   if ( UseFastEmptyMethods) {
 608     // If we need a safepoint check, generate full interpreter entry.
 609     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
 610     __ set(sync_state, G3_scratch);
 611     __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path);
 612 
 613     // Code: _return
 614     __ retl();
 615     __ delayed()->mov(O5_savedSP, SP);
 616 
 617     __ bind(slow_path);
 618     (void) generate_normal_entry(false);
 619 
 620     return entry;
 621   }
 622   return NULL;
 623 }
 624 
 625 // Call an accessor method (assuming it is resolved, otherwise drop into
 626 // vanilla (slow path) entry
 627 
 628 // Generates code to elide accessor methods
 629 // Uses G3_scratch and G1_scratch as scratch
 630 address InterpreterGenerator::generate_accessor_entry(void) {
 631 
 632   // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
 633   // parameter size = 1
 634   // Note: We can only use this code if the getfield has been resolved
 635   //       and if we don't have a null-pointer exception => check for
 636   //       these conditions first and use slow path if necessary.
 637   address entry = __ pc();
 638   Label slow_path;
 639 
 640 
 641   // XXX: for compressed oops pointer loading and decoding doesn't fit in
 642   // delay slot and damages G1
 643   if ( UseFastAccessorMethods && !UseCompressedOops ) {
 644     // Check if we need to reach a safepoint and generate full interpreter
 645     // frame if so.
 646     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
 647     __ load_contents(sync_state, G3_scratch);
 648     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 649     __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path);
 650 
 651     // Check if local 0 != NULL
 652     __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
 653     // check if local 0 == NULL and go the slow path
 654     __ br_null_short(Otos_i, Assembler::pn, slow_path);
 655 
 656 
 657     // read first instruction word and extract bytecode @ 1 and index @ 2
 658     // get first 4 bytes of the bytecodes (big endian!)
 659     __ ld_ptr(G5_method, Method::const_offset(), G1_scratch);
 660     __ ld(G1_scratch, ConstMethod::codes_offset(), G1_scratch);
 661 
 662     // move index @ 2 far left then to the right most two bytes.
 663     __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
 664     __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
 665                       ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
 666 
 667     // get constant pool cache
 668     __ ld_ptr(G5_method, Method::const_offset(), G3_scratch);
 669     __ ld_ptr(G3_scratch, ConstMethod::constants_offset(), G3_scratch);
 670     __ ld_ptr(G3_scratch, ConstantPool::cache_offset_in_bytes(), G3_scratch);
 671 
 672     // get specific constant pool cache entry
 673     __ add(G3_scratch, G1_scratch, G3_scratch);
 674 
 675     // Check the constant Pool cache entry to see if it has been resolved.
 676     // If not, need the slow path.
 677     ByteSize cp_base_offset = ConstantPoolCache::base_offset();
 678     __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::indices_offset(), G1_scratch);
 679     __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
 680     __ and3(G1_scratch, 0xFF, G1_scratch);
 681     __ cmp_and_br_short(G1_scratch, Bytecodes::_getfield, Assembler::notEqual, Assembler::pn, slow_path);
 682 
 683     // Get the type and return field offset from the constant pool cache
 684     __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::flags_offset(), G1_scratch);
 685     __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), G3_scratch);
 686 
 687     Label xreturn_path;
 688     // Need to differentiate between igetfield, agetfield, bgetfield etc.
 689     // because they are different sizes.
 690     // Get the type from the constant pool cache
 691     __ srl(G1_scratch, ConstantPoolCacheEntry::tos_state_shift, G1_scratch);
 692     // Make sure we don't need to mask G1_scratch after the above shift
 693     ConstantPoolCacheEntry::verify_tos_state_shift();
 694     __ cmp(G1_scratch, atos );
 695     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 696     __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
 697     __ cmp(G1_scratch, itos);
 698     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 699     __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
 700     __ cmp(G1_scratch, stos);
 701     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 702     __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
 703     __ cmp(G1_scratch, ctos);
 704     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 705     __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
 706 #ifdef ASSERT
 707     __ cmp(G1_scratch, btos);
 708     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 709     __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
 710     __ should_not_reach_here();
 711 #endif
 712     __ ldsb(Otos_i, G3_scratch, Otos_i);
 713     __ bind(xreturn_path);
 714 
 715     // _ireturn/_areturn
 716     __ retl();                      // return from leaf routine
 717     __ delayed()->mov(O5_savedSP, SP);
 718 
 719     // Generate regular method entry
 720     __ bind(slow_path);
 721     (void) generate_normal_entry(false);
 722     return entry;
 723   }
 724   return NULL;
 725 }
 726 
 727 // Method entry for java.lang.ref.Reference.get.
 728 address InterpreterGenerator::generate_Reference_get_entry(void) {
 729 #if INCLUDE_ALL_GCS
 730   // Code: _aload_0, _getfield, _areturn
 731   // parameter size = 1
 732   //
 733   // The code that gets generated by this routine is split into 2 parts:
 734   //    1. The "intrinsified" code for G1 (or any SATB based GC),
 735   //    2. The slow path - which is an expansion of the regular method entry.
 736   //
 737   // Notes:-
 738   // * In the G1 code we do not check whether we need to block for
 739   //   a safepoint. If G1 is enabled then we must execute the specialized
 740   //   code for Reference.get (except when the Reference object is null)
 741   //   so that we can log the value in the referent field with an SATB
 742   //   update buffer.
 743   //   If the code for the getfield template is modified so that the
 744   //   G1 pre-barrier code is executed when the current method is
 745   //   Reference.get() then going through the normal method entry
 746   //   will be fine.
 747   // * The G1 code can, however, check the receiver object (the instance
 748   //   of java.lang.Reference) and jump to the slow path if null. If the
 749   //   Reference object is null then we obviously cannot fetch the referent
 750   //   and so we don't need to call the G1 pre-barrier. Thus we can use the
 751   //   regular method entry code to generate the NPE.
 752   //
 753   // This code is based on generate_accessor_enty.
 754 
 755   address entry = __ pc();
 756 
 757   const int referent_offset = java_lang_ref_Reference::referent_offset;
 758   guarantee(referent_offset > 0, "referent offset not initialized");
 759 
 760   if (UseG1GC) {
 761      Label slow_path;
 762 
 763     // In the G1 code we don't check if we need to reach a safepoint. We
 764     // continue and the thread will safepoint at the next bytecode dispatch.
 765 
 766     // Check if local 0 != NULL
 767     // If the receiver is null then it is OK to jump to the slow path.
 768     __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
 769     // check if local 0 == NULL and go the slow path
 770     __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
 771 
 772 
 773     // Load the value of the referent field.
 774     if (Assembler::is_simm13(referent_offset)) {
 775       __ load_heap_oop(Otos_i, referent_offset, Otos_i);
 776     } else {
 777       __ set(referent_offset, G3_scratch);
 778       __ load_heap_oop(Otos_i, G3_scratch, Otos_i);
 779     }
 780 
 781     // Generate the G1 pre-barrier code to log the value of
 782     // the referent field in an SATB buffer. Note with
 783     // these parameters the pre-barrier does not generate
 784     // the load of the previous value
 785 
 786     __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */,
 787                             Otos_i /* pre_val */,
 788                             G3_scratch /* tmp */,
 789                             true /* preserve_o_regs */);
 790 
 791     // _areturn
 792     __ retl();                      // return from leaf routine
 793     __ delayed()->mov(O5_savedSP, SP);
 794 
 795     // Generate regular method entry
 796     __ bind(slow_path);
 797     (void) generate_normal_entry(false);
 798     return entry;
 799   }
 800 #endif // INCLUDE_ALL_GCS
 801 
 802   // If G1 is not enabled then attempt to go through the accessor entry point
 803   // Reference.get is an accessor
 804   return generate_accessor_entry();
 805 }
 806 
 807 //
 808 // Interpreter stub for calling a native method. (asm interpreter)
 809 // This sets up a somewhat different looking stack for calling the native method
 810 // than the typical interpreter frame setup.
 811 //
 812 
 813 address InterpreterGenerator::generate_native_entry(bool synchronized) {
 814   address entry = __ pc();
 815 
 816   // the following temporary registers are used during frame creation
 817   const Register Gtmp1 = G3_scratch ;
 818   const Register Gtmp2 = G1_scratch;
 819   bool inc_counter  = UseCompiler || CountCompiledCalls;
 820 
 821   // make sure registers are different!
 822   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
 823 
 824   const Address Laccess_flags(Lmethod, Method::access_flags_offset());
 825 
 826   const Register Glocals_size = G3;
 827   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
 828 
 829   // make sure method is native & not abstract
 830   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
 831 #ifdef ASSERT
 832   __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
 833   {
 834     Label L;
 835     __ btst(JVM_ACC_NATIVE, Gtmp1);
 836     __ br(Assembler::notZero, false, Assembler::pt, L);
 837     __ delayed()->nop();
 838     __ stop("tried to execute non-native method as native");
 839     __ bind(L);
 840   }
 841   { Label L;
 842     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
 843     __ br(Assembler::zero, false, Assembler::pt, L);
 844     __ delayed()->nop();
 845     __ stop("tried to execute abstract method as non-abstract");
 846     __ bind(L);
 847   }
 848 #endif // ASSERT
 849 
 850  // generate the code to allocate the interpreter stack frame
 851   generate_fixed_frame(true);
 852 
 853   //
 854   // No locals to initialize for native method
 855   //
 856 
 857   // this slot will be set later, we initialize it to null here just in
 858   // case we get a GC before the actual value is stored later
 859   __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
 860 
 861   const Address do_not_unlock_if_synchronized(G2_thread,
 862     JavaThread::do_not_unlock_if_synchronized_offset());
 863   // Since at this point in the method invocation the exception handler
 864   // would try to exit the monitor of synchronized methods which hasn't
 865   // been entered yet, we set the thread local variable
 866   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
 867   // runtime, exception handling i.e. unlock_if_synchronized_method will
 868   // check this thread local flag.
 869   // This flag has two effects, one is to force an unwind in the topmost
 870   // interpreter frame and not perform an unlock while doing so.
 871 
 872   __ movbool(true, G3_scratch);
 873   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
 874 
 875   // increment invocation counter and check for overflow
 876   //
 877   // Note: checking for negative value instead of overflow
 878   //       so we have a 'sticky' overflow test (may be of
 879   //       importance as soon as we have true MT/MP)
 880   Label invocation_counter_overflow;
 881   Label Lcontinue;
 882   if (inc_counter) {
 883     generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
 884 
 885   }
 886   __ bind(Lcontinue);
 887 
 888   bang_stack_shadow_pages(true);
 889 
 890   // reset the _do_not_unlock_if_synchronized flag
 891   __ stbool(G0, do_not_unlock_if_synchronized);
 892 
 893   // check for synchronized methods
 894   // Must happen AFTER invocation_counter check and stack overflow check,
 895   // so method is not locked if overflows.
 896 
 897   if (synchronized) {
 898     lock_method();
 899   } else {
 900 #ifdef ASSERT
 901     { Label ok;
 902       __ ld(Laccess_flags, O0);
 903       __ btst(JVM_ACC_SYNCHRONIZED, O0);
 904       __ br( Assembler::zero, false, Assembler::pt, ok);
 905       __ delayed()->nop();
 906       __ stop("method needs synchronization");
 907       __ bind(ok);
 908     }
 909 #endif // ASSERT
 910   }
 911 
 912 
 913   // start execution
 914   __ verify_thread();
 915 
 916   // JVMTI support
 917   __ notify_method_entry();
 918 
 919   // native call
 920 
 921   // (note that O0 is never an oop--at most it is a handle)
 922   // It is important not to smash any handles created by this call,
 923   // until any oop handle in O0 is dereferenced.
 924 
 925   // (note that the space for outgoing params is preallocated)
 926 
 927   // get signature handler
 928   { Label L;
 929     Address signature_handler(Lmethod, Method::signature_handler_offset());
 930     __ ld_ptr(signature_handler, G3_scratch);
 931     __ br_notnull_short(G3_scratch, Assembler::pt, L);
 932     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
 933     __ ld_ptr(signature_handler, G3_scratch);
 934     __ bind(L);
 935   }
 936 
 937   // Push a new frame so that the args will really be stored in
 938   // Copy a few locals across so the new frame has the variables
 939   // we need but these values will be dead at the jni call and
 940   // therefore not gc volatile like the values in the current
 941   // frame (Lmethod in particular)
 942 
 943   // Flush the method pointer to the register save area
 944   __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
 945   __ mov(Llocals, O1);
 946 
 947   // calculate where the mirror handle body is allocated in the interpreter frame:
 948   __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
 949 
 950   // Calculate current frame size
 951   __ sub(SP, FP, O3);         // Calculate negative of current frame size
 952   __ save(SP, O3, SP);        // Allocate an identical sized frame
 953 
 954   // Note I7 has leftover trash. Slow signature handler will fill it in
 955   // should we get there. Normal jni call will set reasonable last_Java_pc
 956   // below (and fix I7 so the stack trace doesn't have a meaningless frame
 957   // in it).
 958 
 959   // Load interpreter frame's Lmethod into same register here
 960 
 961   __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
 962 
 963   __ mov(I1, Llocals);
 964   __ mov(I2, Lscratch2);     // save the address of the mirror
 965 
 966 
 967   // ONLY Lmethod and Llocals are valid here!
 968 
 969   // call signature handler, It will move the arg properly since Llocals in current frame
 970   // matches that in outer frame
 971 
 972   __ callr(G3_scratch, 0);
 973   __ delayed()->nop();
 974 
 975   // Result handler is in Lscratch
 976 
 977   // Reload interpreter frame's Lmethod since slow signature handler may block
 978   __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
 979 
 980   { Label not_static;
 981 
 982     __ ld(Laccess_flags, O0);
 983     __ btst(JVM_ACC_STATIC, O0);
 984     __ br( Assembler::zero, false, Assembler::pt, not_static);
 985     // get native function entry point(O0 is a good temp until the very end)
 986     __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0);
 987     // for static methods insert the mirror argument
 988     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
 989 
 990     __ ld_ptr(Lmethod, Method:: const_offset(), O1);
 991     __ ld_ptr(O1, ConstMethod::constants_offset(), O1);
 992     __ ld_ptr(O1, ConstantPool::pool_holder_offset_in_bytes(), O1);
 993     __ ld_ptr(O1, mirror_offset, O1);
 994 #ifdef ASSERT
 995     if (!PrintSignatureHandlers)  // do not dirty the output with this
 996     { Label L;
 997       __ br_notnull_short(O1, Assembler::pt, L);
 998       __ stop("mirror is missing");
 999       __ bind(L);
1000     }
1001 #endif // ASSERT
1002     __ st_ptr(O1, Lscratch2, 0);
1003     __ mov(Lscratch2, O1);
1004     __ bind(not_static);
1005   }
1006 
1007   // At this point, arguments have been copied off of stack into
1008   // their JNI positions, which are O1..O5 and SP[68..].
1009   // Oops are boxed in-place on the stack, with handles copied to arguments.
1010   // The result handler is in Lscratch.  O0 will shortly hold the JNIEnv*.
1011 
1012 #ifdef ASSERT
1013   { Label L;
1014     __ br_notnull_short(O0, Assembler::pt, L);
1015     __ stop("native entry point is missing");
1016     __ bind(L);
1017   }
1018 #endif // ASSERT
1019 
1020   //
1021   // setup the frame anchor
1022   //
1023   // The scavenge function only needs to know that the PC of this frame is
1024   // in the interpreter method entry code, it doesn't need to know the exact
1025   // PC and hence we can use O7 which points to the return address from the
1026   // previous call in the code stream (signature handler function)
1027   //
1028   // The other trick is we set last_Java_sp to FP instead of the usual SP because
1029   // we have pushed the extra frame in order to protect the volatile register(s)
1030   // in that frame when we return from the jni call
1031   //
1032 
1033   __ set_last_Java_frame(FP, O7);
1034   __ mov(O7, I7);  // make dummy interpreter frame look like one above,
1035                    // not meaningless information that'll confuse me.
1036 
1037   // flush the windows now. We don't care about the current (protection) frame
1038   // only the outer frames
1039 
1040   __ flushw();
1041 
1042   // mark windows as flushed
1043   Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1044   __ set(JavaFrameAnchor::flushed, G3_scratch);
1045   __ st(G3_scratch, flags);
1046 
1047   // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1048 
1049   Address thread_state(G2_thread, JavaThread::thread_state_offset());
1050 #ifdef ASSERT
1051   { Label L;
1052     __ ld(thread_state, G3_scratch);
1053     __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1054     __ stop("Wrong thread state in native stub");
1055     __ bind(L);
1056   }
1057 #endif // ASSERT
1058   __ set(_thread_in_native, G3_scratch);
1059   __ st(G3_scratch, thread_state);
1060 
1061   // Call the jni method, using the delay slot to set the JNIEnv* argument.
1062   __ save_thread(L7_thread_cache); // save Gthread
1063   __ callr(O0, 0);
1064   __ delayed()->
1065      add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1066 
1067   // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1068 
1069   __ restore_thread(L7_thread_cache); // restore G2_thread
1070   __ reinit_heapbase();
1071 
1072   // must we block?
1073 
1074   // Block, if necessary, before resuming in _thread_in_Java state.
1075   // In order for GC to work, don't clear the last_Java_sp until after blocking.
1076   { Label no_block;
1077     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1078 
1079     // Switch thread to "native transition" state before reading the synchronization state.
1080     // This additional state is necessary because reading and testing the synchronization
1081     // state is not atomic w.r.t. GC, as this scenario demonstrates:
1082     //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1083     //     VM thread changes sync state to synchronizing and suspends threads for GC.
1084     //     Thread A is resumed to finish this native method, but doesn't block here since it
1085     //     didn't see any synchronization is progress, and escapes.
1086     __ set(_thread_in_native_trans, G3_scratch);
1087     __ st(G3_scratch, thread_state);
1088     if(os::is_MP()) {
1089       if (UseMembar) {
1090         // Force this write out before the read below
1091         __ membar(Assembler::StoreLoad);
1092       } else {
1093         // Write serialization page so VM thread can do a pseudo remote membar.
1094         // We use the current thread pointer to calculate a thread specific
1095         // offset to write to within the page. This minimizes bus traffic
1096         // due to cache line collision.
1097         __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1098       }
1099     }
1100     __ load_contents(sync_state, G3_scratch);
1101     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1102 
1103     Label L;
1104     __ br(Assembler::notEqual, false, Assembler::pn, L);
1105     __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1106     __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1107     __ bind(L);
1108 
1109     // Block.  Save any potential method result value before the operation and
1110     // use a leaf call to leave the last_Java_frame setup undisturbed.
1111     save_native_result();
1112     __ call_VM_leaf(L7_thread_cache,
1113                     CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1114                     G2_thread);
1115 
1116     // Restore any method result value
1117     restore_native_result();
1118     __ bind(no_block);
1119   }
1120 
1121   // Clear the frame anchor now
1122 
1123   __ reset_last_Java_frame();
1124 
1125   // Move the result handler address
1126   __ mov(Lscratch, G3_scratch);
1127   // return possible result to the outer frame
1128 #ifndef __LP64
1129   __ mov(O0, I0);
1130   __ restore(O1, G0, O1);
1131 #else
1132   __ restore(O0, G0, O0);
1133 #endif /* __LP64 */
1134 
1135   // Move result handler to expected register
1136   __ mov(G3_scratch, Lscratch);
1137 
1138   // Back in normal (native) interpreter frame. State is thread_in_native_trans
1139   // switch to thread_in_Java.
1140 
1141   __ set(_thread_in_Java, G3_scratch);
1142   __ st(G3_scratch, thread_state);
1143 
1144   // reset handle block
1145   __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1146   __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1147 
1148   // If we have an oop result store it where it will be safe for any further gc
1149   // until we return now that we've released the handle it might be protected by
1150 
1151   {
1152     Label no_oop, store_result;
1153 
1154     __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1155     __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1156     __ addcc(G0, O0, O0);
1157     __ brx(Assembler::notZero, true, Assembler::pt, store_result);     // if result is not NULL:
1158     __ delayed()->ld_ptr(O0, 0, O0);                                   // unbox it
1159     __ mov(G0, O0);
1160 
1161     __ bind(store_result);
1162     // Store it where gc will look for it and result handler expects it.
1163     __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1164 
1165     __ bind(no_oop);
1166 
1167   }
1168 
1169 
1170   // handle exceptions (exception handling will handle unlocking!)
1171   { Label L;
1172     Address exception_addr(G2_thread, Thread::pending_exception_offset());
1173     __ ld_ptr(exception_addr, Gtemp);
1174     __ br_null_short(Gtemp, Assembler::pt, L);
1175     // Note: This could be handled more efficiently since we know that the native
1176     //       method doesn't have an exception handler. We could directly return
1177     //       to the exception handler for the caller.
1178     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1179     __ should_not_reach_here();
1180     __ bind(L);
1181   }
1182 
1183   // JVMTI support (preserves thread register)
1184   __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1185 
1186   if (synchronized) {
1187     // save and restore any potential method result value around the unlocking operation
1188     save_native_result();
1189 
1190     __ add( __ top_most_monitor(), O1);
1191     __ unlock_object(O1);
1192 
1193     restore_native_result();
1194   }
1195 
1196 #if defined(COMPILER2) && !defined(_LP64)
1197 
1198   // C2 expects long results in G1 we can't tell if we're returning to interpreted
1199   // or compiled so just be safe.
1200 
1201   __ sllx(O0, 32, G1);          // Shift bits into high G1
1202   __ srl (O1, 0, O1);           // Zero extend O1
1203   __ or3 (O1, G1, G1);          // OR 64 bits into G1
1204 
1205 #endif /* COMPILER2 && !_LP64 */
1206 
1207   // dispose of return address and remove activation
1208 #ifdef ASSERT
1209   {
1210     Label ok;
1211     __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1212     __ stop("bad I5_savedSP value");
1213     __ should_not_reach_here();
1214     __ bind(ok);
1215   }
1216 #endif
1217   if (TraceJumps) {
1218     // Move target to register that is recordable
1219     __ mov(Lscratch, G3_scratch);
1220     __ JMP(G3_scratch, 0);
1221   } else {
1222     __ jmp(Lscratch, 0);
1223   }
1224   __ delayed()->nop();
1225 
1226 
1227   if (inc_counter) {
1228     // handle invocation counter overflow
1229     __ bind(invocation_counter_overflow);
1230     generate_counter_overflow(Lcontinue);
1231   }
1232 
1233 
1234 
1235   return entry;
1236 }
1237 
1238 
1239 // Generic method entry to (asm) interpreter
1240 //------------------------------------------------------------------------------------------------------------------------
1241 //
1242 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1243   address entry = __ pc();
1244 
1245   bool inc_counter  = UseCompiler || CountCompiledCalls;
1246 
1247   // the following temporary registers are used during frame creation
1248   const Register Gtmp1 = G3_scratch ;
1249   const Register Gtmp2 = G1_scratch;
1250 
1251   // make sure registers are different!
1252   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1253 
1254   const Address constMethod       (G5_method, Method::const_offset());
1255   // Seems like G5_method is live at the point this is used. So we could make this look consistent
1256   // and use in the asserts.
1257   const Address access_flags      (Lmethod,   Method::access_flags_offset());
1258 
1259   const Register Glocals_size = G3;
1260   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1261 
1262   // make sure method is not native & not abstract
1263   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1264 #ifdef ASSERT
1265   __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1266   {
1267     Label L;
1268     __ btst(JVM_ACC_NATIVE, Gtmp1);
1269     __ br(Assembler::zero, false, Assembler::pt, L);
1270     __ delayed()->nop();
1271     __ stop("tried to execute native method as non-native");
1272     __ bind(L);
1273   }
1274   { Label L;
1275     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1276     __ br(Assembler::zero, false, Assembler::pt, L);
1277     __ delayed()->nop();
1278     __ stop("tried to execute abstract method as non-abstract");
1279     __ bind(L);
1280   }
1281 #endif // ASSERT
1282 
1283   // generate the code to allocate the interpreter stack frame
1284 
1285   generate_fixed_frame(false);
1286 
1287 #ifdef FAST_DISPATCH
1288   __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1289                                           // set bytecode dispatch table base
1290 #endif
1291 
1292   //
1293   // Code to initialize the extra (i.e. non-parm) locals
1294   //
1295   Register init_value = noreg;    // will be G0 if we must clear locals
1296   // The way the code was setup before zerolocals was always true for vanilla java entries.
1297   // It could only be false for the specialized entries like accessor or empty which have
1298   // no extra locals so the testing was a waste of time and the extra locals were always
1299   // initialized. We removed this extra complication to already over complicated code.
1300 
1301   init_value = G0;
1302   Label clear_loop;
1303 
1304   const Register RconstMethod = O1;
1305   const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1306   const Address size_of_locals    (RconstMethod, ConstMethod::size_of_locals_offset());
1307 
1308   // NOTE: If you change the frame layout, this code will need to
1309   // be updated!
1310   __ ld_ptr( constMethod, RconstMethod );
1311   __ lduh( size_of_locals, O2 );
1312   __ lduh( size_of_parameters, O1 );
1313   __ sll( O2, Interpreter::logStackElementSize, O2);
1314   __ sll( O1, Interpreter::logStackElementSize, O1 );
1315   __ sub( Llocals, O2, O2 );
1316   __ sub( Llocals, O1, O1 );
1317 
1318   __ bind( clear_loop );
1319   __ inc( O2, wordSize );
1320 
1321   __ cmp( O2, O1 );
1322   __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1323   __ delayed()->st_ptr( init_value, O2, 0 );
1324 
1325   const Address do_not_unlock_if_synchronized(G2_thread,
1326     JavaThread::do_not_unlock_if_synchronized_offset());
1327   // Since at this point in the method invocation the exception handler
1328   // would try to exit the monitor of synchronized methods which hasn't
1329   // been entered yet, we set the thread local variable
1330   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1331   // runtime, exception handling i.e. unlock_if_synchronized_method will
1332   // check this thread local flag.
1333   __ movbool(true, G3_scratch);
1334   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1335 
1336   // increment invocation counter and check for overflow
1337   //
1338   // Note: checking for negative value instead of overflow
1339   //       so we have a 'sticky' overflow test (may be of
1340   //       importance as soon as we have true MT/MP)
1341   Label invocation_counter_overflow;
1342   Label profile_method;
1343   Label profile_method_continue;
1344   Label Lcontinue;
1345   if (inc_counter) {
1346     generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1347     if (ProfileInterpreter) {
1348       __ bind(profile_method_continue);
1349     }
1350   }
1351   __ bind(Lcontinue);
1352 
1353   bang_stack_shadow_pages(false);
1354 
1355   // reset the _do_not_unlock_if_synchronized flag
1356   __ stbool(G0, do_not_unlock_if_synchronized);
1357 
1358   // check for synchronized methods
1359   // Must happen AFTER invocation_counter check and stack overflow check,
1360   // so method is not locked if overflows.
1361 
1362   if (synchronized) {
1363     lock_method();
1364   } else {
1365 #ifdef ASSERT
1366     { Label ok;
1367       __ ld(access_flags, O0);
1368       __ btst(JVM_ACC_SYNCHRONIZED, O0);
1369       __ br( Assembler::zero, false, Assembler::pt, ok);
1370       __ delayed()->nop();
1371       __ stop("method needs synchronization");
1372       __ bind(ok);
1373     }
1374 #endif // ASSERT
1375   }
1376 
1377   // start execution
1378 
1379   __ verify_thread();
1380 
1381   // jvmti support
1382   __ notify_method_entry();
1383 
1384   // start executing instructions
1385   __ dispatch_next(vtos);
1386 
1387 
1388   if (inc_counter) {
1389     if (ProfileInterpreter) {
1390       // We have decided to profile this method in the interpreter
1391       __ bind(profile_method);
1392 
1393       __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1394       __ set_method_data_pointer_for_bcp();
1395       __ ba_short(profile_method_continue);
1396     }
1397 
1398     // handle invocation counter overflow
1399     __ bind(invocation_counter_overflow);
1400     generate_counter_overflow(Lcontinue);
1401   }
1402 
1403 
1404   return entry;
1405 }
1406 
1407 
1408 //----------------------------------------------------------------------------------------------------
1409 // Entry points & stack frame layout
1410 //
1411 // Here we generate the various kind of entries into the interpreter.
1412 // The two main entry type are generic bytecode methods and native call method.
1413 // These both come in synchronized and non-synchronized versions but the
1414 // frame layout they create is very similar. The other method entry
1415 // types are really just special purpose entries that are really entry
1416 // and interpretation all in one. These are for trivial methods like
1417 // accessor, empty, or special math methods.
1418 //
1419 // When control flow reaches any of the entry types for the interpreter
1420 // the following holds ->
1421 //
1422 // C2 Calling Conventions:
1423 //
1424 // The entry code below assumes that the following registers are set
1425 // when coming in:
1426 //    G5_method: holds the Method* of the method to call
1427 //    Lesp:    points to the TOS of the callers expression stack
1428 //             after having pushed all the parameters
1429 //
1430 // The entry code does the following to setup an interpreter frame
1431 //   pop parameters from the callers stack by adjusting Lesp
1432 //   set O0 to Lesp
1433 //   compute X = (max_locals - num_parameters)
1434 //   bump SP up by X to accomadate the extra locals
1435 //   compute X = max_expression_stack
1436 //               + vm_local_words
1437 //               + 16 words of register save area
1438 //   save frame doing a save sp, -X, sp growing towards lower addresses
1439 //   set Lbcp, Lmethod, LcpoolCache
1440 //   set Llocals to i0
1441 //   set Lmonitors to FP - rounded_vm_local_words
1442 //   set Lesp to Lmonitors - 4
1443 //
1444 //  The frame has now been setup to do the rest of the entry code
1445 
1446 // Try this optimization:  Most method entries could live in a
1447 // "one size fits all" stack frame without all the dynamic size
1448 // calculations.  It might be profitable to do all this calculation
1449 // statically and approximately for "small enough" methods.
1450 
1451 //-----------------------------------------------------------------------------------------------
1452 
1453 // C1 Calling conventions
1454 //
1455 // Upon method entry, the following registers are setup:
1456 //
1457 // g2 G2_thread: current thread
1458 // g5 G5_method: method to activate
1459 // g4 Gargs  : pointer to last argument
1460 //
1461 //
1462 // Stack:
1463 //
1464 // +---------------+ <--- sp
1465 // |               |
1466 // : reg save area :
1467 // |               |
1468 // +---------------+ <--- sp + 0x40
1469 // |               |
1470 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
1471 // |               |
1472 // +---------------+ <--- sp + 0x5c
1473 // |               |
1474 // :     free      :
1475 // |               |
1476 // +---------------+ <--- Gargs
1477 // |               |
1478 // :   arguments   :
1479 // |               |
1480 // +---------------+
1481 // |               |
1482 //
1483 //
1484 //
1485 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
1486 //
1487 // +---------------+ <--- sp
1488 // |               |
1489 // : reg save area :
1490 // |               |
1491 // +---------------+ <--- sp + 0x40
1492 // |               |
1493 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
1494 // |               |
1495 // +---------------+ <--- sp + 0x5c
1496 // |               |
1497 // :               :
1498 // |               | <--- Lesp
1499 // +---------------+ <--- Lmonitors (fp - 0x18)
1500 // |   VM locals   |
1501 // +---------------+ <--- fp
1502 // |               |
1503 // : reg save area :
1504 // |               |
1505 // +---------------+ <--- fp + 0x40
1506 // |               |
1507 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
1508 // |               |
1509 // +---------------+ <--- fp + 0x5c
1510 // |               |
1511 // :     free      :
1512 // |               |
1513 // +---------------+
1514 // |               |
1515 // : nonarg locals :
1516 // |               |
1517 // +---------------+
1518 // |               |
1519 // :   arguments   :
1520 // |               | <--- Llocals
1521 // +---------------+ <--- Gargs
1522 // |               |
1523 
1524 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
1525 
1526   // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
1527   // expression stack, the callee will have callee_extra_locals (so we can account for
1528   // frame extension) and monitor_size for monitors. Basically we need to calculate
1529   // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
1530   //
1531   //
1532   // The big complicating thing here is that we must ensure that the stack stays properly
1533   // aligned. This would be even uglier if monitor size wasn't modulo what the stack
1534   // needs to be aligned for). We are given that the sp (fp) is already aligned by
1535   // the caller so we must ensure that it is properly aligned for our callee.
1536   //
1537   const int rounded_vm_local_words =
1538        round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1539   // callee_locals and max_stack are counts, not the size in frame.
1540   const int locals_size =
1541        round_to(callee_extra_locals * Interpreter::stackElementWords, WordsPerLong);
1542   const int max_stack_words = max_stack * Interpreter::stackElementWords;
1543   return (round_to((max_stack_words
1544                    + rounded_vm_local_words
1545                    + frame::memory_parameter_word_sp_offset), WordsPerLong)
1546                    // already rounded
1547                    + locals_size + monitor_size);
1548 }
1549 
1550 // How much stack a method top interpreter activation needs in words.
1551 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
1552 
1553   // See call_stub code
1554   int call_stub_size  = round_to(7 + frame::memory_parameter_word_sp_offset,
1555                                  WordsPerLong);    // 7 + register save area
1556 
1557   // Save space for one monitor to get into the interpreted method in case
1558   // the method is synchronized
1559   int monitor_size    = method->is_synchronized() ?
1560                                 1*frame::interpreter_frame_monitor_size() : 0;
1561   return size_activation_helper(method->max_locals(), method->max_stack(),
1562                                  monitor_size) + call_stub_size;
1563 }
1564 
1565 int AbstractInterpreter::layout_activation(Method* method,
1566                                            int tempcount,
1567                                            int popframe_extra_args,
1568                                            int moncount,
1569                                            int caller_actual_parameters,
1570                                            int callee_param_count,
1571                                            int callee_local_count,
1572                                            frame* caller,
1573                                            frame* interpreter_frame,
1574                                            bool is_top_frame,
1575                                            bool is_bottom_frame) {
1576   // Note: This calculation must exactly parallel the frame setup
1577   // in InterpreterGenerator::generate_fixed_frame.
1578   // If f!=NULL, set up the following variables:
1579   //   - Lmethod
1580   //   - Llocals
1581   //   - Lmonitors (to the indicated number of monitors)
1582   //   - Lesp (to the indicated number of temps)
1583   // The frame f (if not NULL) on entry is a description of the caller of the frame
1584   // we are about to layout. We are guaranteed that we will be able to fill in a
1585   // new interpreter frame as its callee (i.e. the stack space is allocated and
1586   // the amount was determined by an earlier call to this method with f == NULL).
1587   // On return f (if not NULL) while describe the interpreter frame we just layed out.
1588 
1589   int monitor_size           = moncount * frame::interpreter_frame_monitor_size();
1590   int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1591 
1592   assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1593   //
1594   // Note: if you look closely this appears to be doing something much different
1595   // than generate_fixed_frame. What is happening is this. On sparc we have to do
1596   // this dance with interpreter_sp_adjustment because the window save area would
1597   // appear just below the bottom (tos) of the caller's java expression stack. Because
1598   // the interpreter want to have the locals completely contiguous generate_fixed_frame
1599   // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size).
1600   // Now in generate_fixed_frame the extension of the caller's sp happens in the callee.
1601   // In this code the opposite occurs the caller adjusts it's own stack base on the callee.
1602   // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest)
1603   // because the oldest frame would have adjust its callers frame and yet that frame
1604   // already exists and isn't part of this array of frames we are unpacking. So at first
1605   // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper()
1606   // will after it calculates all of the frame's on_stack_size()'s will then figure out the
1607   // amount to adjust the caller of the initial (oldest) frame and the calculation will all
1608   // add up. It does seem like it simpler to account for the adjustment here (and remove the
1609   // callee... parameters here). However this would mean that this routine would have to take
1610   // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment)
1611   // and run the calling loop in the reverse order. This would also would appear to mean making
1612   // this code aware of what the interactions are when that initial caller fram was an osr or
1613   // other adapter frame. deoptimization is complicated enough and  hard enough to debug that
1614   // there is no sense in messing working code.
1615   //
1616 
1617   int rounded_cls = round_to((callee_local_count - callee_param_count), WordsPerLong);
1618   assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align");
1619 
1620   int raw_frame_size = size_activation_helper(rounded_cls, method->max_stack(),
1621                                               monitor_size);
1622 
1623   if (interpreter_frame != NULL) {
1624     // The skeleton frame must already look like an interpreter frame
1625     // even if not fully filled out.
1626     assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame");
1627 
1628     intptr_t* fp = interpreter_frame->fp();
1629 
1630     JavaThread* thread = JavaThread::current();
1631     RegisterMap map(thread, false);
1632     // More verification that skeleton frame is properly walkable
1633     assert(fp == caller->sp(), "fp must match");
1634 
1635     intptr_t* montop     = fp - rounded_vm_local_words;
1636 
1637     // preallocate monitors (cf. __ add_monitor_to_stack)
1638     intptr_t* monitors = montop - monitor_size;
1639 
1640     // preallocate stack space
1641     intptr_t*  esp = monitors - 1 -
1642                      (tempcount * Interpreter::stackElementWords) -
1643                      popframe_extra_args;
1644 
1645     int local_words = method->max_locals() * Interpreter::stackElementWords;
1646     NEEDS_CLEANUP;
1647     intptr_t* locals;
1648     if (caller->is_interpreted_frame()) {
1649       // Can force the locals area to end up properly overlapping the top of the expression stack.
1650       intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1;
1651       // Note that this computation means we replace size_of_parameters() values from the caller
1652       // interpreter frame's expression stack with our argument locals
1653       int parm_words  = caller_actual_parameters * Interpreter::stackElementWords;
1654       locals = Lesp_ptr + parm_words;
1655       int delta = local_words - parm_words;
1656       int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0;
1657       *interpreter_frame->register_addr(I5_savedSP)    = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS;
1658       if (!is_bottom_frame) {
1659         // Llast_SP is set below for the current frame to SP (with the
1660         // extra space for the callee's locals). Here we adjust
1661         // Llast_SP for the caller's frame, removing the extra space
1662         // for the current method's locals.
1663         *caller->register_addr(Llast_SP) = *interpreter_frame->register_addr(I5_savedSP);
1664       } else {
1665         assert(*caller->register_addr(Llast_SP) >= *interpreter_frame->register_addr(I5_savedSP), "strange Llast_SP");
1666       }
1667     } else {
1668       assert(caller->is_compiled_frame() || caller->is_entry_frame(), "only possible cases");
1669       // Don't have Lesp available; lay out locals block in the caller
1670       // adjacent to the register window save area.
1671       //
1672       // Compiled frames do not allocate a varargs area which is why this if
1673       // statement is needed.
1674       //
1675       if (caller->is_compiled_frame()) {
1676         locals = fp + frame::register_save_words + local_words - 1;
1677       } else {
1678         locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
1679       }
1680       if (!caller->is_entry_frame()) {
1681         // Caller wants his own SP back
1682         int caller_frame_size = caller->cb()->frame_size();
1683         *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS;
1684       }
1685     }
1686     if (TraceDeoptimization) {
1687       if (caller->is_entry_frame()) {
1688         // make sure I5_savedSP and the entry frames notion of saved SP
1689         // agree.  This assertion duplicate a check in entry frame code
1690         // but catches the failure earlier.
1691         assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP),
1692                "would change callers SP");
1693       }
1694       if (caller->is_entry_frame()) {
1695         tty->print("entry ");
1696       }
1697       if (caller->is_compiled_frame()) {
1698         tty->print("compiled ");
1699         if (caller->is_deoptimized_frame()) {
1700           tty->print("(deopt) ");
1701         }
1702       }
1703       if (caller->is_interpreted_frame()) {
1704         tty->print("interpreted ");
1705       }
1706       tty->print_cr("caller fp=0x%x sp=0x%x", caller->fp(), caller->sp());
1707       tty->print_cr("save area = 0x%x, 0x%x", caller->sp(), caller->sp() + 16);
1708       tty->print_cr("save area = 0x%x, 0x%x", caller->fp(), caller->fp() + 16);
1709       tty->print_cr("interpreter fp=0x%x sp=0x%x", interpreter_frame->fp(), interpreter_frame->sp());
1710       tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->sp(), interpreter_frame->sp() + 16);
1711       tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->fp(), interpreter_frame->fp() + 16);
1712       tty->print_cr("Llocals = 0x%x", locals);
1713       tty->print_cr("Lesp = 0x%x", esp);
1714       tty->print_cr("Lmonitors = 0x%x", monitors);
1715     }
1716 
1717     if (method->max_locals() > 0) {
1718       assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area");
1719       assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area");
1720       assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area");
1721       assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area");
1722     }
1723 #ifdef _LP64
1724     assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd");
1725 #endif
1726 
1727     *interpreter_frame->register_addr(Lmethod)     = (intptr_t) method;
1728     *interpreter_frame->register_addr(Llocals)     = (intptr_t) locals;
1729     *interpreter_frame->register_addr(Lmonitors)   = (intptr_t) monitors;
1730     *interpreter_frame->register_addr(Lesp)        = (intptr_t) esp;
1731     // Llast_SP will be same as SP as there is no adapter space
1732     *interpreter_frame->register_addr(Llast_SP)    = (intptr_t) interpreter_frame->sp() - STACK_BIAS;
1733     *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache();
1734 #ifdef FAST_DISPATCH
1735     *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table();
1736 #endif
1737 
1738 
1739 #ifdef ASSERT
1740     BasicObjectLock* mp = (BasicObjectLock*)monitors;
1741 
1742     assert(interpreter_frame->interpreter_frame_method() == method, "method matches");
1743     assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize)), "locals match");
1744     assert(interpreter_frame->interpreter_frame_monitor_end()   == mp, "monitor_end matches");
1745     assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches");
1746     assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches");
1747 
1748     // check bounds
1749     intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1);
1750     intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words;
1751     assert(lo < monitors && montop <= hi, "monitors in bounds");
1752     assert(lo <= esp && esp < monitors, "esp in bounds");
1753 #endif // ASSERT
1754   }
1755 
1756   return raw_frame_size;
1757 }
1758 
1759 //----------------------------------------------------------------------------------------------------
1760 // Exceptions
1761 void TemplateInterpreterGenerator::generate_throw_exception() {
1762 
1763   // Entry point in previous activation (i.e., if the caller was interpreted)
1764   Interpreter::_rethrow_exception_entry = __ pc();
1765   // O0: exception
1766 
1767   // entry point for exceptions thrown within interpreter code
1768   Interpreter::_throw_exception_entry = __ pc();
1769   __ verify_thread();
1770   // expression stack is undefined here
1771   // O0: exception, i.e. Oexception
1772   // Lbcp: exception bcx
1773   __ verify_oop(Oexception);
1774 
1775 
1776   // expression stack must be empty before entering the VM in case of an exception
1777   __ empty_expression_stack();
1778   // find exception handler address and preserve exception oop
1779   // call C routine to find handler and jump to it
1780   __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1781   __ push_ptr(O1); // push exception for exception handler bytecodes
1782 
1783   __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1784   __ delayed()->nop();
1785 
1786 
1787   // if the exception is not handled in the current frame
1788   // the frame is removed and the exception is rethrown
1789   // (i.e. exception continuation is _rethrow_exception)
1790   //
1791   // Note: At this point the bci is still the bxi for the instruction which caused
1792   //       the exception and the expression stack is empty. Thus, for any VM calls
1793   //       at this point, GC will find a legal oop map (with empty expression stack).
1794 
1795   // in current activation
1796   // tos: exception
1797   // Lbcp: exception bcp
1798 
1799   //
1800   // JVMTI PopFrame support
1801   //
1802 
1803   Interpreter::_remove_activation_preserving_args_entry = __ pc();
1804   Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1805   // Set the popframe_processing bit in popframe_condition indicating that we are
1806   // currently handling popframe, so that call_VMs that may happen later do not trigger new
1807   // popframe handling cycles.
1808 
1809   __ ld(popframe_condition_addr, G3_scratch);
1810   __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1811   __ stw(G3_scratch, popframe_condition_addr);
1812 
1813   // Empty the expression stack, as in normal exception handling
1814   __ empty_expression_stack();
1815   __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1816 
1817   {
1818     // Check to see whether we are returning to a deoptimized frame.
1819     // (The PopFrame call ensures that the caller of the popped frame is
1820     // either interpreted or compiled and deoptimizes it if compiled.)
1821     // In this case, we can't call dispatch_next() after the frame is
1822     // popped, but instead must save the incoming arguments and restore
1823     // them after deoptimization has occurred.
1824     //
1825     // Note that we don't compare the return PC against the
1826     // deoptimization blob's unpack entry because of the presence of
1827     // adapter frames in C2.
1828     Label caller_not_deoptimized;
1829     __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1830     __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1831 
1832     const Register Gtmp1 = G3_scratch;
1833     const Register Gtmp2 = G1_scratch;
1834     const Register RconstMethod = Gtmp1;
1835     const Address constMethod(Lmethod, Method::const_offset());
1836     const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1837 
1838     // Compute size of arguments for saving when returning to deoptimized caller
1839     __ ld_ptr(constMethod, RconstMethod);
1840     __ lduh(size_of_parameters, Gtmp1);
1841     __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1842     __ sub(Llocals, Gtmp1, Gtmp2);
1843     __ add(Gtmp2, wordSize, Gtmp2);
1844     // Save these arguments
1845     __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1846     // Inform deoptimization that it is responsible for restoring these arguments
1847     __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1848     Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1849     __ st(Gtmp1, popframe_condition_addr);
1850 
1851     // Return from the current method
1852     // The caller's SP was adjusted upon method entry to accomodate
1853     // the callee's non-argument locals. Undo that adjustment.
1854     __ ret();
1855     __ delayed()->restore(I5_savedSP, G0, SP);
1856 
1857     __ bind(caller_not_deoptimized);
1858   }
1859 
1860   // Clear the popframe condition flag
1861   __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1862 
1863   // Get out of the current method (how this is done depends on the particular compiler calling
1864   // convention that the interpreter currently follows)
1865   // The caller's SP was adjusted upon method entry to accomodate
1866   // the callee's non-argument locals. Undo that adjustment.
1867   __ restore(I5_savedSP, G0, SP);
1868   // The method data pointer was incremented already during
1869   // call profiling. We have to restore the mdp for the current bcp.
1870   if (ProfileInterpreter) {
1871     __ set_method_data_pointer_for_bcp();
1872   }
1873 
1874 #if INCLUDE_JVMTI
1875   if (EnableInvokeDynamic) {
1876     Label L_done;
1877 
1878     __ ldub(Address(Lbcp, 0), G1_scratch);  // Load current bytecode
1879     __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done);
1880 
1881     // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
1882     // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
1883 
1884     __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp);
1885 
1886     __ br_null(G1_scratch, false, Assembler::pn, L_done);
1887     __ delayed()->nop();
1888 
1889     __ st_ptr(G1_scratch, Lesp, wordSize);
1890     __ bind(L_done);
1891   }
1892 #endif // INCLUDE_JVMTI
1893 
1894   // Resume bytecode interpretation at the current bcp
1895   __ dispatch_next(vtos);
1896   // end of JVMTI PopFrame support
1897 
1898   Interpreter::_remove_activation_entry = __ pc();
1899 
1900   // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1901   __ pop_ptr(Oexception);                                  // get exception
1902 
1903   // Intel has the following comment:
1904   //// remove the activation (without doing throws on illegalMonitorExceptions)
1905   // They remove the activation without checking for bad monitor state.
1906   // %%% We should make sure this is the right semantics before implementing.
1907 
1908   __ set_vm_result(Oexception);
1909   __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1910 
1911   __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1912 
1913   __ get_vm_result(Oexception);
1914   __ verify_oop(Oexception);
1915 
1916     const int return_reg_adjustment = frame::pc_return_offset;
1917   Address issuing_pc_addr(I7, return_reg_adjustment);
1918 
1919   // We are done with this activation frame; find out where to go next.
1920   // The continuation point will be an exception handler, which expects
1921   // the following registers set up:
1922   //
1923   // Oexception: exception
1924   // Oissuing_pc: the local call that threw exception
1925   // Other On: garbage
1926   // In/Ln:  the contents of the caller's register window
1927   //
1928   // We do the required restore at the last possible moment, because we
1929   // need to preserve some state across a runtime call.
1930   // (Remember that the caller activation is unknown--it might not be
1931   // interpreted, so things like Lscratch are useless in the caller.)
1932 
1933   // Although the Intel version uses call_C, we can use the more
1934   // compact call_VM.  (The only real difference on SPARC is a
1935   // harmlessly ignored [re]set_last_Java_frame, compared with
1936   // the Intel code which lacks this.)
1937   __ mov(Oexception,      Oexception ->after_save());  // get exception in I0 so it will be on O0 after restore
1938   __ add(issuing_pc_addr, Oissuing_pc->after_save());  // likewise set I1 to a value local to the caller
1939   __ super_call_VM_leaf(L7_thread_cache,
1940                         CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1941                         G2_thread, Oissuing_pc->after_save());
1942 
1943   // The caller's SP was adjusted upon method entry to accomodate
1944   // the callee's non-argument locals. Undo that adjustment.
1945   __ JMP(O0, 0);                         // return exception handler in caller
1946   __ delayed()->restore(I5_savedSP, G0, SP);
1947 
1948   // (same old exception object is already in Oexception; see above)
1949   // Note that an "issuing PC" is actually the next PC after the call
1950 }
1951 
1952 
1953 //
1954 // JVMTI ForceEarlyReturn support
1955 //
1956 
1957 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1958   address entry = __ pc();
1959 
1960   __ empty_expression_stack();
1961   __ load_earlyret_value(state);
1962 
1963   __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1964   Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1965 
1966   // Clear the earlyret state
1967   __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1968 
1969   __ remove_activation(state,
1970                        /* throw_monitor_exception */ false,
1971                        /* install_monitor_exception */ false);
1972 
1973   // The caller's SP was adjusted upon method entry to accomodate
1974   // the callee's non-argument locals. Undo that adjustment.
1975   __ ret();                             // return to caller
1976   __ delayed()->restore(I5_savedSP, G0, SP);
1977 
1978   return entry;
1979 } // end of JVMTI ForceEarlyReturn support
1980 
1981 
1982 //------------------------------------------------------------------------------------------------------------------------
1983 // Helper for vtos entry point generation
1984 
1985 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) {
1986   assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1987   Label L;
1988   aep = __ pc(); __ push_ptr(); __ ba_short(L);
1989   fep = __ pc(); __ push_f();   __ ba_short(L);
1990   dep = __ pc(); __ push_d();   __ ba_short(L);
1991   lep = __ pc(); __ push_l();   __ ba_short(L);
1992   iep = __ pc(); __ push_i();
1993   bep = cep = sep = iep;                        // there aren't any
1994   vep = __ pc(); __ bind(L);                    // fall through
1995   generate_and_dispatch(t);
1996 }
1997 
1998 // --------------------------------------------------------------------------------
1999 
2000 
2001 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2002  : TemplateInterpreterGenerator(code) {
2003    generate_all(); // down here so it can be "virtual"
2004 }
2005 
2006 // --------------------------------------------------------------------------------
2007 
2008 // Non-product code
2009 #ifndef PRODUCT
2010 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
2011   address entry = __ pc();
2012 
2013   __ push(state);
2014   __ mov(O7, Lscratch); // protect return address within interpreter
2015 
2016   // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
2017   __ mov( Otos_l2, G3_scratch );
2018   __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
2019   __ mov(Lscratch, O7); // restore return address
2020   __ pop(state);
2021   __ retl();
2022   __ delayed()->nop();
2023 
2024   return entry;
2025 }
2026 
2027 
2028 // helpers for generate_and_dispatch
2029 
2030 void TemplateInterpreterGenerator::count_bytecode() {
2031   __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
2032 }
2033 
2034 
2035 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
2036   __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
2037 }
2038 
2039 
2040 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2041   AddressLiteral index   (&BytecodePairHistogram::_index);
2042   AddressLiteral counters((address) &BytecodePairHistogram::_counters);
2043 
2044   // get index, shift out old bytecode, bring in new bytecode, and store it
2045   // _index = (_index >> log2_number_of_codes) |
2046   //          (bytecode << log2_number_of_codes);
2047 
2048   __ load_contents(index, G4_scratch);
2049   __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
2050   __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes,  G3_scratch );
2051   __ or3( G3_scratch,  G4_scratch, G4_scratch );
2052   __ store_contents(G4_scratch, index, G3_scratch);
2053 
2054   // bump bucket contents
2055   // _counters[_index] ++;
2056 
2057   __ set(counters, G3_scratch);                       // loads into G3_scratch
2058   __ sll( G4_scratch, LogBytesPerWord, G4_scratch );  // Index is word address
2059   __ add (G3_scratch, G4_scratch, G3_scratch);        // Add in index
2060   __ ld (G3_scratch, 0, G4_scratch);
2061   __ inc (G4_scratch);
2062   __ st (G4_scratch, 0, G3_scratch);
2063 }
2064 
2065 
2066 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2067   // Call a little run-time stub to avoid blow-up for each bytecode.
2068   // The run-time runtime saves the right registers, depending on
2069   // the tosca in-state for the given template.
2070   address entry = Interpreter::trace_code(t->tos_in());
2071   guarantee(entry != NULL, "entry must have been generated");
2072   __ call(entry, relocInfo::none);
2073   __ delayed()->nop();
2074 }
2075 
2076 
2077 void TemplateInterpreterGenerator::stop_interpreter_at() {
2078   AddressLiteral counter(&BytecodeCounter::_counter_value);
2079   __ load_contents(counter, G3_scratch);
2080   AddressLiteral stop_at(&StopInterpreterAt);
2081   __ load_ptr_contents(stop_at, G4_scratch);
2082   __ cmp(G3_scratch, G4_scratch);
2083   __ breakpoint_trap(Assembler::equal, Assembler::icc);
2084 }
2085 #endif // not PRODUCT
2086 #endif // !CC_INTERP