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) {
 157   TosState incoming_state = state;
 158 
 159   Label cont;
 160   address compiled_entry = __ pc();
 161 
 162   address entry = __ pc();
 163 #if !defined(_LP64) && defined(COMPILER2)
 164   // All return values are where we want them, except for Longs.  C2 returns
 165   // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
 166   // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
 167   // build even if we are returning from interpreted we just do a little
 168   // stupid shuffing.
 169   // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
 170   // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
 171   // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
 172 
 173   if (incoming_state == ltos) {
 174     __ srl (G1,  0, O1);
 175     __ srlx(G1, 32, O0);
 176   }
 177 #endif // !_LP64 && COMPILER2
 178 
 179   __ bind(cont);
 180 
 181   // The callee returns with the stack possibly adjusted by adapter transition
 182   // We remove that possible adjustment here.
 183   // All interpreter local registers are untouched. Any result is passed back
 184   // in the O0/O1 or float registers. Before continuing, the arguments must be
 185   // popped from the java expression stack; i.e., Lesp must be adjusted.
 186 
 187   __ mov(Llast_SP, SP);   // Remove any adapter added stack space.
 188 
 189   Label L_got_cache, L_giant_index;
 190   const Register cache = G3_scratch;
 191   const Register size  = G1_scratch;
 192   if (EnableInvokeDynamic) {
 193     __ ldub(Address(Lbcp, 0), G1_scratch);  // Load current bytecode.
 194     __ cmp_and_br_short(G1_scratch, Bytecodes::_invokedynamic, Assembler::equal, Assembler::pn, L_giant_index);
 195   }
 196   __ get_cache_and_index_at_bcp(cache, G1_scratch, 1);
 197   __ bind(L_got_cache);
 198   __ ld_ptr(cache, ConstantPoolCache::base_offset() +
 199                    ConstantPoolCacheEntry::flags_offset(), size);
 200   __ and3(size, 0xFF, size);                   // argument size in words
 201   __ sll(size, Interpreter::logStackElementSize, size); // each argument size in bytes
 202   __ add(Lesp, size, Lesp);                    // pop arguments
 203   __ dispatch_next(state, step);
 204 
 205   // out of the main line of code...
 206   if (EnableInvokeDynamic) {
 207     __ bind(L_giant_index);
 208     __ get_cache_and_index_at_bcp(cache, G1_scratch, 1, sizeof(u4));
 209     __ ba_short(L_got_cache);
 210   }
 211 
 212   return entry;
 213 }
 214 
 215 
 216 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
 217   address entry = __ pc();
 218   __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
 219   { Label L;
 220     Address exception_addr(G2_thread, Thread::pending_exception_offset());
 221     __ ld_ptr(exception_addr, Gtemp);  // Load pending exception.
 222     __ br_null_short(Gtemp, Assembler::pt, L);
 223     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
 224     __ should_not_reach_here();
 225     __ bind(L);
 226   }
 227   __ dispatch_next(state, step);
 228   return entry;
 229 }
 230 
 231 // A result handler converts/unboxes a native call result into
 232 // a java interpreter/compiler result. The current frame is an
 233 // interpreter frame. The activation frame unwind code must be
 234 // consistent with that of TemplateTable::_return(...). In the
 235 // case of native methods, the caller's SP was not modified.
 236 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
 237   address entry = __ pc();
 238   Register Itos_i  = Otos_i ->after_save();
 239   Register Itos_l  = Otos_l ->after_save();
 240   Register Itos_l1 = Otos_l1->after_save();
 241   Register Itos_l2 = Otos_l2->after_save();
 242   switch (type) {
 243     case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
 244     case T_CHAR   : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i);   break; // cannot use and3, 0xFFFF too big as immediate value!
 245     case T_BYTE   : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i);   break;
 246     case T_SHORT  : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i);   break;
 247     case T_LONG   :
 248 #ifndef _LP64
 249                     __ mov(O1, Itos_l2);  // move other half of long
 250 #endif              // ifdef or no ifdef, fall through to the T_INT case
 251     case T_INT    : __ mov(O0, Itos_i);                         break;
 252     case T_VOID   : /* nothing to do */                         break;
 253     case T_FLOAT  : assert(F0 == Ftos_f, "fix this code" );     break;
 254     case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" );     break;
 255     case T_OBJECT :
 256       __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
 257       __ verify_oop(Itos_i);
 258       break;
 259     default       : ShouldNotReachHere();
 260   }
 261   __ ret();                           // return from interpreter activation
 262   __ delayed()->restore(I5_savedSP, G0, SP);  // remove interpreter frame
 263   NOT_PRODUCT(__ emit_int32(0);)       // marker for disassembly
 264   return entry;
 265 }
 266 
 267 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
 268   address entry = __ pc();
 269   __ push(state);
 270   __ call_VM(noreg, runtime_entry);
 271   __ dispatch_via(vtos, Interpreter::normal_table(vtos));
 272   return entry;
 273 }
 274 
 275 
 276 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
 277   address entry = __ pc();
 278   __ dispatch_next(state);
 279   return entry;
 280 }
 281 
 282 //
 283 // Helpers for commoning out cases in the various type of method entries.
 284 //
 285 
 286 // increment invocation count & check for overflow
 287 //
 288 // Note: checking for negative value instead of overflow
 289 //       so we have a 'sticky' overflow test
 290 //
 291 // Lmethod: method
 292 // ??: invocation counter
 293 //
 294 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
 295   // Note: In tiered we increment either counters in MethodCounters* or in
 296   // MDO depending if we're profiling or not.
 297   const Register Rcounters = G3_scratch;
 298   Label done;
 299 
 300   if (TieredCompilation) {
 301     const int increment = InvocationCounter::count_increment;
 302     const int mask = ((1 << Tier0InvokeNotifyFreqLog) - 1) << InvocationCounter::count_shift;
 303     Label no_mdo;
 304     if (ProfileInterpreter) {
 305       // If no method data exists, go to profile_continue.
 306       __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
 307       __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
 308       // Increment counter
 309       Address mdo_invocation_counter(G4_scratch,
 310                                      in_bytes(MethodData::invocation_counter_offset()) +
 311                                      in_bytes(InvocationCounter::counter_offset()));
 312       __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
 313                                  G3_scratch, Lscratch,
 314                                  Assembler::zero, overflow);
 315       __ ba_short(done);
 316     }
 317 
 318     // Increment counter in MethodCounters*
 319     __ bind(no_mdo);
 320     Address invocation_counter(Rcounters,
 321             in_bytes(MethodCounters::invocation_counter_offset()) +
 322             in_bytes(InvocationCounter::counter_offset()));
 323     __ get_method_counters(Lmethod, Rcounters, done);
 324     __ increment_mask_and_jump(invocation_counter, increment, mask,
 325                                G4_scratch, Lscratch,
 326                                Assembler::zero, overflow);
 327     __ bind(done);
 328   } else {
 329     // Update standard invocation counters
 330     __ get_method_counters(Lmethod, Rcounters, done);
 331     __ increment_invocation_counter(Rcounters, O0, G4_scratch);
 332     if (ProfileInterpreter) {
 333       Address interpreter_invocation_counter(Rcounters,
 334             in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
 335       __ ld(interpreter_invocation_counter, G4_scratch);
 336       __ inc(G4_scratch);
 337       __ st(G4_scratch, interpreter_invocation_counter);
 338     }
 339 
 340     if (ProfileInterpreter && profile_method != NULL) {
 341       // Test to see if we should create a method data oop
 342       AddressLiteral profile_limit((address)&InvocationCounter::InterpreterProfileLimit);
 343       __ load_contents(profile_limit, G3_scratch);
 344       __ cmp_and_br_short(O0, G3_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
 345 
 346       // if no method data exists, go to profile_method
 347       __ test_method_data_pointer(*profile_method);
 348     }
 349 
 350     AddressLiteral invocation_limit((address)&InvocationCounter::InterpreterInvocationLimit);
 351     __ load_contents(invocation_limit, G3_scratch);
 352     __ cmp(O0, G3_scratch);
 353     __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
 354     __ delayed()->nop();
 355     __ bind(done);
 356   }
 357 
 358 }
 359 
 360 // Allocate monitor and lock method (asm interpreter)
 361 // ebx - Method*
 362 //
 363 void InterpreterGenerator::lock_method(void) {
 364   __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0);  // Load access flags.
 365 
 366 #ifdef ASSERT
 367  { Label ok;
 368    __ btst(JVM_ACC_SYNCHRONIZED, O0);
 369    __ br( Assembler::notZero, false, Assembler::pt, ok);
 370    __ delayed()->nop();
 371    __ stop("method doesn't need synchronization");
 372    __ bind(ok);
 373   }
 374 #endif // ASSERT
 375 
 376   // get synchronization object to O0
 377   { Label done;
 378     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
 379     __ btst(JVM_ACC_STATIC, O0);
 380     __ br( Assembler::zero, true, Assembler::pt, done);
 381     __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
 382 
 383     __ ld_ptr( Lmethod, in_bytes(Method::const_offset()), O0);
 384     __ ld_ptr( O0, in_bytes(ConstMethod::constants_offset()), O0);
 385     __ ld_ptr( O0, ConstantPool::pool_holder_offset_in_bytes(), O0);
 386 
 387     // lock the mirror, not the Klass*
 388     __ ld_ptr( O0, mirror_offset, O0);
 389 
 390 #ifdef ASSERT
 391     __ tst(O0);
 392     __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
 393 #endif // ASSERT
 394 
 395     __ bind(done);
 396   }
 397 
 398   __ add_monitor_to_stack(true, noreg, noreg);  // allocate monitor elem
 399   __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes());   // store object
 400   // __ untested("lock_object from method entry");
 401   __ lock_object(Lmonitors, O0);
 402 }
 403 
 404 
 405 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
 406                                                          Register Rscratch,
 407                                                          Register Rscratch2) {
 408   const int page_size = os::vm_page_size();
 409   Label after_frame_check;
 410 
 411   assert_different_registers(Rframe_size, Rscratch, Rscratch2);
 412 
 413   __ set(page_size, Rscratch);
 414   __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
 415 
 416   // get the stack base, and in debug, verify it is non-zero
 417   __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch );
 418 #ifdef ASSERT
 419   Label base_not_zero;
 420   __ br_notnull_short(Rscratch, Assembler::pn, base_not_zero);
 421   __ stop("stack base is zero in generate_stack_overflow_check");
 422   __ bind(base_not_zero);
 423 #endif
 424 
 425   // get the stack size, and in debug, verify it is non-zero
 426   assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" );
 427   __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 );
 428 #ifdef ASSERT
 429   Label size_not_zero;
 430   __ br_notnull_short(Rscratch2, Assembler::pn, size_not_zero);
 431   __ stop("stack size is zero in generate_stack_overflow_check");
 432   __ bind(size_not_zero);
 433 #endif
 434 
 435   // compute the beginning of the protected zone minus the requested frame size
 436   __ sub( Rscratch, Rscratch2,   Rscratch );
 437   __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 );
 438   __ add( Rscratch, Rscratch2,   Rscratch );
 439 
 440   // Add in the size of the frame (which is the same as subtracting it from the
 441   // SP, which would take another register
 442   __ add( Rscratch, Rframe_size, Rscratch );
 443 
 444   // the frame is greater than one page in size, so check against
 445   // the bottom of the stack
 446   __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check);
 447 
 448   // the stack will overflow, throw an exception
 449 
 450   // Note that SP is restored to sender's sp (in the delay slot). This
 451   // is necessary if the sender's frame is an extended compiled frame
 452   // (see gen_c2i_adapter()) and safer anyway in case of JSR292
 453   // adaptations.
 454 
 455   // Note also that the restored frame is not necessarily interpreted.
 456   // Use the shared runtime version of the StackOverflowError.
 457   assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
 458   AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
 459   __ jump_to(stub, Rscratch);
 460   __ delayed()->mov(O5_savedSP, SP);
 461 
 462   // if you get to here, then there is enough stack space
 463   __ bind( after_frame_check );
 464 }
 465 
 466 
 467 //
 468 // Generate a fixed interpreter frame. This is identical setup for interpreted
 469 // methods and for native methods hence the shared code.
 470 
 471 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
 472   //
 473   //
 474   // The entry code sets up a new interpreter frame in 4 steps:
 475   //
 476   // 1) Increase caller's SP by for the extra local space needed:
 477   //    (check for overflow)
 478   //    Efficient implementation of xload/xstore bytecodes requires
 479   //    that arguments and non-argument locals are in a contigously
 480   //    addressable memory block => non-argument locals must be
 481   //    allocated in the caller's frame.
 482   //
 483   // 2) Create a new stack frame and register window:
 484   //    The new stack frame must provide space for the standard
 485   //    register save area, the maximum java expression stack size,
 486   //    the monitor slots (0 slots initially), and some frame local
 487   //    scratch locations.
 488   //
 489   // 3) The following interpreter activation registers must be setup:
 490   //    Lesp       : expression stack pointer
 491   //    Lbcp       : bytecode pointer
 492   //    Lmethod    : method
 493   //    Llocals    : locals pointer
 494   //    Lmonitors  : monitor pointer
 495   //    LcpoolCache: constant pool cache
 496   //
 497   // 4) Initialize the non-argument locals if necessary:
 498   //    Non-argument locals may need to be initialized to NULL
 499   //    for GC to work. If the oop-map information is accurate
 500   //    (in the absence of the JSR problem), no initialization
 501   //    is necessary.
 502   //
 503   // (gri - 2/25/2000)
 504 
 505 
 506   int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
 507 
 508   const int extra_space =
 509     rounded_vm_local_words +                   // frame local scratch space
 510     //6815692//Method::extra_stack_words() +       // extra push slots for MH adapters
 511     frame::memory_parameter_word_sp_offset +   // register save area
 512     (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
 513 
 514   const Register Glocals_size = G3;
 515   const Register RconstMethod = Glocals_size;
 516   const Register Otmp1 = O3;
 517   const Register Otmp2 = O4;
 518   // Lscratch can't be used as a temporary because the call_stub uses
 519   // it to assert that the stack frame was setup correctly.
 520   const Address constMethod       (G5_method, Method::const_offset());
 521   const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
 522 
 523   __ ld_ptr( constMethod, RconstMethod );
 524   __ lduh( size_of_parameters, Glocals_size);
 525 
 526   // Gargs points to first local + BytesPerWord
 527   // Set the saved SP after the register window save
 528   //
 529   assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
 530   __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
 531   __ add(Gargs, Otmp1, Gargs);
 532 
 533   if (native_call) {
 534     __ calc_mem_param_words( Glocals_size, Gframe_size );
 535     __ add( Gframe_size,  extra_space, Gframe_size);
 536     __ round_to( Gframe_size, WordsPerLong );
 537     __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
 538   } else {
 539 
 540     //
 541     // Compute number of locals in method apart from incoming parameters
 542     //
 543     const Address size_of_locals    (Otmp1, ConstMethod::size_of_locals_offset());
 544     __ ld_ptr( constMethod, Otmp1 );
 545     __ lduh( size_of_locals, Otmp1 );
 546     __ sub( Otmp1, Glocals_size, Glocals_size );
 547     __ round_to( Glocals_size, WordsPerLong );
 548     __ sll( Glocals_size, Interpreter::logStackElementSize, Glocals_size );
 549 
 550     // see if the frame is greater than one page in size. If so,
 551     // then we need to verify there is enough stack space remaining
 552     // Frame_size = (max_stack + extra_space) * BytesPerWord;
 553     __ ld_ptr( constMethod, Gframe_size );
 554     __ lduh( Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size );
 555     __ add( Gframe_size, extra_space, Gframe_size );
 556     __ round_to( Gframe_size, WordsPerLong );
 557     __ sll( Gframe_size, Interpreter::logStackElementSize, Gframe_size);
 558 
 559     // Add in java locals size for stack overflow check only
 560     __ add( Gframe_size, Glocals_size, Gframe_size );
 561 
 562     const Register Otmp2 = O4;
 563     assert_different_registers(Otmp1, Otmp2, O5_savedSP);
 564     generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2);
 565 
 566     __ sub( Gframe_size, Glocals_size, Gframe_size);
 567 
 568     //
 569     // bump SP to accomodate the extra locals
 570     //
 571     __ sub( SP, Glocals_size, SP );
 572   }
 573 
 574   //
 575   // now set up a stack frame with the size computed above
 576   //
 577   __ neg( Gframe_size );
 578   __ save( SP, Gframe_size, SP );
 579 
 580   //
 581   // now set up all the local cache registers
 582   //
 583   // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
 584   // that all present references to Lbyte_code initialize the register
 585   // immediately before use
 586   if (native_call) {
 587     __ mov(G0, Lbcp);
 588   } else {
 589     __ ld_ptr(G5_method, Method::const_offset(), Lbcp);
 590     __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp);
 591   }
 592   __ mov( G5_method, Lmethod);                 // set Lmethod
 593   __ get_constant_pool_cache( LcpoolCache );   // set LcpoolCache
 594   __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
 595 #ifdef _LP64
 596   __ add( Lmonitors, STACK_BIAS, Lmonitors );   // Account for 64 bit stack bias
 597 #endif
 598   __ sub(Lmonitors, BytesPerWord, Lesp);       // set Lesp
 599 
 600   // setup interpreter activation registers
 601   __ sub(Gargs, BytesPerWord, Llocals);        // set Llocals
 602 
 603   if (ProfileInterpreter) {
 604 #ifdef FAST_DISPATCH
 605     // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
 606     // they both use I2.
 607     assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
 608 #endif // FAST_DISPATCH
 609     __ set_method_data_pointer();
 610   }
 611 
 612 }
 613 
 614 // Empty method, generate a very fast return.
 615 
 616 address InterpreterGenerator::generate_empty_entry(void) {
 617 
 618   // A method that does nother but return...
 619 
 620   address entry = __ pc();
 621   Label slow_path;
 622 
 623   // do nothing for empty methods (do not even increment invocation counter)
 624   if ( UseFastEmptyMethods) {
 625     // If we need a safepoint check, generate full interpreter entry.
 626     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
 627     __ set(sync_state, G3_scratch);
 628     __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path);
 629 
 630     // Code: _return
 631     __ retl();
 632     __ delayed()->mov(O5_savedSP, SP);
 633 
 634     __ bind(slow_path);
 635     (void) generate_normal_entry(false);
 636 
 637     return entry;
 638   }
 639   return NULL;
 640 }
 641 
 642 // Call an accessor method (assuming it is resolved, otherwise drop into
 643 // vanilla (slow path) entry
 644 
 645 // Generates code to elide accessor methods
 646 // Uses G3_scratch and G1_scratch as scratch
 647 address InterpreterGenerator::generate_accessor_entry(void) {
 648 
 649   // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
 650   // parameter size = 1
 651   // Note: We can only use this code if the getfield has been resolved
 652   //       and if we don't have a null-pointer exception => check for
 653   //       these conditions first and use slow path if necessary.
 654   address entry = __ pc();
 655   Label slow_path;
 656 
 657 
 658   // XXX: for compressed oops pointer loading and decoding doesn't fit in
 659   // delay slot and damages G1
 660   if ( UseFastAccessorMethods && !UseCompressedOops ) {
 661     // Check if we need to reach a safepoint and generate full interpreter
 662     // frame if so.
 663     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
 664     __ load_contents(sync_state, G3_scratch);
 665     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 666     __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path);
 667 
 668     // Check if local 0 != NULL
 669     __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
 670     // check if local 0 == NULL and go the slow path
 671     __ br_null_short(Otos_i, Assembler::pn, slow_path);
 672 
 673 
 674     // read first instruction word and extract bytecode @ 1 and index @ 2
 675     // get first 4 bytes of the bytecodes (big endian!)
 676     __ ld_ptr(G5_method, Method::const_offset(), G1_scratch);
 677     __ ld(G1_scratch, ConstMethod::codes_offset(), G1_scratch);
 678 
 679     // move index @ 2 far left then to the right most two bytes.
 680     __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
 681     __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
 682                       ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
 683 
 684     // get constant pool cache
 685     __ ld_ptr(G5_method, Method::const_offset(), G3_scratch);
 686     __ ld_ptr(G3_scratch, ConstMethod::constants_offset(), G3_scratch);
 687     __ ld_ptr(G3_scratch, ConstantPool::cache_offset_in_bytes(), G3_scratch);
 688 
 689     // get specific constant pool cache entry
 690     __ add(G3_scratch, G1_scratch, G3_scratch);
 691 
 692     // Check the constant Pool cache entry to see if it has been resolved.
 693     // If not, need the slow path.
 694     ByteSize cp_base_offset = ConstantPoolCache::base_offset();
 695     __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::indices_offset(), G1_scratch);
 696     __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
 697     __ and3(G1_scratch, 0xFF, G1_scratch);
 698     __ cmp_and_br_short(G1_scratch, Bytecodes::_getfield, Assembler::notEqual, Assembler::pn, slow_path);
 699 
 700     // Get the type and return field offset from the constant pool cache
 701     __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::flags_offset(), G1_scratch);
 702     __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), G3_scratch);
 703 
 704     Label xreturn_path;
 705     // Need to differentiate between igetfield, agetfield, bgetfield etc.
 706     // because they are different sizes.
 707     // Get the type from the constant pool cache
 708     __ srl(G1_scratch, ConstantPoolCacheEntry::tos_state_shift, G1_scratch);
 709     // Make sure we don't need to mask G1_scratch after the above shift
 710     ConstantPoolCacheEntry::verify_tos_state_shift();
 711     __ cmp(G1_scratch, atos );
 712     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 713     __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
 714     __ cmp(G1_scratch, itos);
 715     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 716     __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
 717     __ cmp(G1_scratch, stos);
 718     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 719     __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
 720     __ cmp(G1_scratch, ctos);
 721     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 722     __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
 723 #ifdef ASSERT
 724     __ cmp(G1_scratch, btos);
 725     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 726     __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
 727     __ should_not_reach_here();
 728 #endif
 729     __ ldsb(Otos_i, G3_scratch, Otos_i);
 730     __ bind(xreturn_path);
 731 
 732     // _ireturn/_areturn
 733     __ retl();                      // return from leaf routine
 734     __ delayed()->mov(O5_savedSP, SP);
 735 
 736     // Generate regular method entry
 737     __ bind(slow_path);
 738     (void) generate_normal_entry(false);
 739     return entry;
 740   }
 741   return NULL;
 742 }
 743 
 744 // Method entry for java.lang.ref.Reference.get.
 745 address InterpreterGenerator::generate_Reference_get_entry(void) {
 746 #if INCLUDE_ALL_GCS
 747   // Code: _aload_0, _getfield, _areturn
 748   // parameter size = 1
 749   //
 750   // The code that gets generated by this routine is split into 2 parts:
 751   //    1. The "intrinsified" code for G1 (or any SATB based GC),
 752   //    2. The slow path - which is an expansion of the regular method entry.
 753   //
 754   // Notes:-
 755   // * In the G1 code we do not check whether we need to block for
 756   //   a safepoint. If G1 is enabled then we must execute the specialized
 757   //   code for Reference.get (except when the Reference object is null)
 758   //   so that we can log the value in the referent field with an SATB
 759   //   update buffer.
 760   //   If the code for the getfield template is modified so that the
 761   //   G1 pre-barrier code is executed when the current method is
 762   //   Reference.get() then going through the normal method entry
 763   //   will be fine.
 764   // * The G1 code can, however, check the receiver object (the instance
 765   //   of java.lang.Reference) and jump to the slow path if null. If the
 766   //   Reference object is null then we obviously cannot fetch the referent
 767   //   and so we don't need to call the G1 pre-barrier. Thus we can use the
 768   //   regular method entry code to generate the NPE.
 769   //
 770   // This code is based on generate_accessor_enty.
 771 
 772   address entry = __ pc();
 773 
 774   const int referent_offset = java_lang_ref_Reference::referent_offset;
 775   guarantee(referent_offset > 0, "referent offset not initialized");
 776 
 777   if (UseG1GC) {
 778      Label slow_path;
 779 
 780     // In the G1 code we don't check if we need to reach a safepoint. We
 781     // continue and the thread will safepoint at the next bytecode dispatch.
 782 
 783     // Check if local 0 != NULL
 784     // If the receiver is null then it is OK to jump to the slow path.
 785     __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
 786     // check if local 0 == NULL and go the slow path
 787     __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
 788 
 789 
 790     // Load the value of the referent field.
 791     if (Assembler::is_simm13(referent_offset)) {
 792       __ load_heap_oop(Otos_i, referent_offset, Otos_i);
 793     } else {
 794       __ set(referent_offset, G3_scratch);
 795       __ load_heap_oop(Otos_i, G3_scratch, Otos_i);
 796     }
 797 
 798     // Generate the G1 pre-barrier code to log the value of
 799     // the referent field in an SATB buffer. Note with
 800     // these parameters the pre-barrier does not generate
 801     // the load of the previous value
 802 
 803     __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */,
 804                             Otos_i /* pre_val */,
 805                             G3_scratch /* tmp */,
 806                             true /* preserve_o_regs */);
 807 
 808     // _areturn
 809     __ retl();                      // return from leaf routine
 810     __ delayed()->mov(O5_savedSP, SP);
 811 
 812     // Generate regular method entry
 813     __ bind(slow_path);
 814     (void) generate_normal_entry(false);
 815     return entry;
 816   }
 817 #endif // INCLUDE_ALL_GCS
 818 
 819   // If G1 is not enabled then attempt to go through the accessor entry point
 820   // Reference.get is an accessor
 821   return generate_accessor_entry();
 822 }
 823 
 824 //
 825 // Interpreter stub for calling a native method. (asm interpreter)
 826 // This sets up a somewhat different looking stack for calling the native method
 827 // than the typical interpreter frame setup.
 828 //
 829 
 830 address InterpreterGenerator::generate_native_entry(bool synchronized) {
 831   address entry = __ pc();
 832 
 833   // the following temporary registers are used during frame creation
 834   const Register Gtmp1 = G3_scratch ;
 835   const Register Gtmp2 = G1_scratch;
 836   bool inc_counter  = UseCompiler || CountCompiledCalls;
 837 
 838   // make sure registers are different!
 839   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
 840 
 841   const Address Laccess_flags(Lmethod, Method::access_flags_offset());
 842 
 843   const Register Glocals_size = G3;
 844   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
 845 
 846   // make sure method is native & not abstract
 847   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
 848 #ifdef ASSERT
 849   __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
 850   {
 851     Label L;
 852     __ btst(JVM_ACC_NATIVE, Gtmp1);
 853     __ br(Assembler::notZero, false, Assembler::pt, L);
 854     __ delayed()->nop();
 855     __ stop("tried to execute non-native method as native");
 856     __ bind(L);
 857   }
 858   { Label L;
 859     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
 860     __ br(Assembler::zero, false, Assembler::pt, L);
 861     __ delayed()->nop();
 862     __ stop("tried to execute abstract method as non-abstract");
 863     __ bind(L);
 864   }
 865 #endif // ASSERT
 866 
 867  // generate the code to allocate the interpreter stack frame
 868   generate_fixed_frame(true);
 869 
 870   //
 871   // No locals to initialize for native method
 872   //
 873 
 874   // this slot will be set later, we initialize it to null here just in
 875   // case we get a GC before the actual value is stored later
 876   __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
 877 
 878   const Address do_not_unlock_if_synchronized(G2_thread,
 879     JavaThread::do_not_unlock_if_synchronized_offset());
 880   // Since at this point in the method invocation the exception handler
 881   // would try to exit the monitor of synchronized methods which hasn't
 882   // been entered yet, we set the thread local variable
 883   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
 884   // runtime, exception handling i.e. unlock_if_synchronized_method will
 885   // check this thread local flag.
 886   // This flag has two effects, one is to force an unwind in the topmost
 887   // interpreter frame and not perform an unlock while doing so.
 888 
 889   __ movbool(true, G3_scratch);
 890   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
 891 
 892   // increment invocation counter and check for overflow
 893   //
 894   // Note: checking for negative value instead of overflow
 895   //       so we have a 'sticky' overflow test (may be of
 896   //       importance as soon as we have true MT/MP)
 897   Label invocation_counter_overflow;
 898   Label Lcontinue;
 899   if (inc_counter) {
 900     generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
 901 
 902   }
 903   __ bind(Lcontinue);
 904 
 905   bang_stack_shadow_pages(true);
 906 
 907   // reset the _do_not_unlock_if_synchronized flag
 908   __ stbool(G0, do_not_unlock_if_synchronized);
 909 
 910   // check for synchronized methods
 911   // Must happen AFTER invocation_counter check and stack overflow check,
 912   // so method is not locked if overflows.
 913 
 914   if (synchronized) {
 915     lock_method();
 916   } else {
 917 #ifdef ASSERT
 918     { Label ok;
 919       __ ld(Laccess_flags, O0);
 920       __ btst(JVM_ACC_SYNCHRONIZED, O0);
 921       __ br( Assembler::zero, false, Assembler::pt, ok);
 922       __ delayed()->nop();
 923       __ stop("method needs synchronization");
 924       __ bind(ok);
 925     }
 926 #endif // ASSERT
 927   }
 928 
 929 
 930   // start execution
 931   __ verify_thread();
 932 
 933   // JVMTI support
 934   __ notify_method_entry();
 935 
 936   // native call
 937 
 938   // (note that O0 is never an oop--at most it is a handle)
 939   // It is important not to smash any handles created by this call,
 940   // until any oop handle in O0 is dereferenced.
 941 
 942   // (note that the space for outgoing params is preallocated)
 943 
 944   // get signature handler
 945   { Label L;
 946     Address signature_handler(Lmethod, Method::signature_handler_offset());
 947     __ ld_ptr(signature_handler, G3_scratch);
 948     __ br_notnull_short(G3_scratch, Assembler::pt, L);
 949     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
 950     __ ld_ptr(signature_handler, G3_scratch);
 951     __ bind(L);
 952   }
 953 
 954   // Push a new frame so that the args will really be stored in
 955   // Copy a few locals across so the new frame has the variables
 956   // we need but these values will be dead at the jni call and
 957   // therefore not gc volatile like the values in the current
 958   // frame (Lmethod in particular)
 959 
 960   // Flush the method pointer to the register save area
 961   __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
 962   __ mov(Llocals, O1);
 963 
 964   // calculate where the mirror handle body is allocated in the interpreter frame:
 965   __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
 966 
 967   // Calculate current frame size
 968   __ sub(SP, FP, O3);         // Calculate negative of current frame size
 969   __ save(SP, O3, SP);        // Allocate an identical sized frame
 970 
 971   // Note I7 has leftover trash. Slow signature handler will fill it in
 972   // should we get there. Normal jni call will set reasonable last_Java_pc
 973   // below (and fix I7 so the stack trace doesn't have a meaningless frame
 974   // in it).
 975 
 976   // Load interpreter frame's Lmethod into same register here
 977 
 978   __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
 979 
 980   __ mov(I1, Llocals);
 981   __ mov(I2, Lscratch2);     // save the address of the mirror
 982 
 983 
 984   // ONLY Lmethod and Llocals are valid here!
 985 
 986   // call signature handler, It will move the arg properly since Llocals in current frame
 987   // matches that in outer frame
 988 
 989   __ callr(G3_scratch, 0);
 990   __ delayed()->nop();
 991 
 992   // Result handler is in Lscratch
 993 
 994   // Reload interpreter frame's Lmethod since slow signature handler may block
 995   __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
 996 
 997   { Label not_static;
 998 
 999     __ ld(Laccess_flags, O0);
1000     __ btst(JVM_ACC_STATIC, O0);
1001     __ br( Assembler::zero, false, Assembler::pt, not_static);
1002     // get native function entry point(O0 is a good temp until the very end)
1003     __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0);
1004     // for static methods insert the mirror argument
1005     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
1006 
1007     __ ld_ptr(Lmethod, Method:: const_offset(), O1);
1008     __ ld_ptr(O1, ConstMethod::constants_offset(), O1);
1009     __ ld_ptr(O1, ConstantPool::pool_holder_offset_in_bytes(), O1);
1010     __ ld_ptr(O1, mirror_offset, O1);
1011 #ifdef ASSERT
1012     if (!PrintSignatureHandlers)  // do not dirty the output with this
1013     { Label L;
1014       __ br_notnull_short(O1, Assembler::pt, L);
1015       __ stop("mirror is missing");
1016       __ bind(L);
1017     }
1018 #endif // ASSERT
1019     __ st_ptr(O1, Lscratch2, 0);
1020     __ mov(Lscratch2, O1);
1021     __ bind(not_static);
1022   }
1023 
1024   // At this point, arguments have been copied off of stack into
1025   // their JNI positions, which are O1..O5 and SP[68..].
1026   // Oops are boxed in-place on the stack, with handles copied to arguments.
1027   // The result handler is in Lscratch.  O0 will shortly hold the JNIEnv*.
1028 
1029 #ifdef ASSERT
1030   { Label L;
1031     __ br_notnull_short(O0, Assembler::pt, L);
1032     __ stop("native entry point is missing");
1033     __ bind(L);
1034   }
1035 #endif // ASSERT
1036 
1037   //
1038   // setup the frame anchor
1039   //
1040   // The scavenge function only needs to know that the PC of this frame is
1041   // in the interpreter method entry code, it doesn't need to know the exact
1042   // PC and hence we can use O7 which points to the return address from the
1043   // previous call in the code stream (signature handler function)
1044   //
1045   // The other trick is we set last_Java_sp to FP instead of the usual SP because
1046   // we have pushed the extra frame in order to protect the volatile register(s)
1047   // in that frame when we return from the jni call
1048   //
1049 
1050   __ set_last_Java_frame(FP, O7);
1051   __ mov(O7, I7);  // make dummy interpreter frame look like one above,
1052                    // not meaningless information that'll confuse me.
1053 
1054   // flush the windows now. We don't care about the current (protection) frame
1055   // only the outer frames
1056 
1057   __ flushw();
1058 
1059   // mark windows as flushed
1060   Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1061   __ set(JavaFrameAnchor::flushed, G3_scratch);
1062   __ st(G3_scratch, flags);
1063 
1064   // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1065 
1066   Address thread_state(G2_thread, JavaThread::thread_state_offset());
1067 #ifdef ASSERT
1068   { Label L;
1069     __ ld(thread_state, G3_scratch);
1070     __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1071     __ stop("Wrong thread state in native stub");
1072     __ bind(L);
1073   }
1074 #endif // ASSERT
1075   __ set(_thread_in_native, G3_scratch);
1076   __ st(G3_scratch, thread_state);
1077 
1078   // Call the jni method, using the delay slot to set the JNIEnv* argument.
1079   __ save_thread(L7_thread_cache); // save Gthread
1080   __ callr(O0, 0);
1081   __ delayed()->
1082      add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1083 
1084   // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1085 
1086   __ restore_thread(L7_thread_cache); // restore G2_thread
1087   __ reinit_heapbase();
1088 
1089   // must we block?
1090 
1091   // Block, if necessary, before resuming in _thread_in_Java state.
1092   // In order for GC to work, don't clear the last_Java_sp until after blocking.
1093   { Label no_block;
1094     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1095 
1096     // Switch thread to "native transition" state before reading the synchronization state.
1097     // This additional state is necessary because reading and testing the synchronization
1098     // state is not atomic w.r.t. GC, as this scenario demonstrates:
1099     //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1100     //     VM thread changes sync state to synchronizing and suspends threads for GC.
1101     //     Thread A is resumed to finish this native method, but doesn't block here since it
1102     //     didn't see any synchronization is progress, and escapes.
1103     __ set(_thread_in_native_trans, G3_scratch);
1104     __ st(G3_scratch, thread_state);
1105     if(os::is_MP()) {
1106       if (UseMembar) {
1107         // Force this write out before the read below
1108         __ membar(Assembler::StoreLoad);
1109       } else {
1110         // Write serialization page so VM thread can do a pseudo remote membar.
1111         // We use the current thread pointer to calculate a thread specific
1112         // offset to write to within the page. This minimizes bus traffic
1113         // due to cache line collision.
1114         __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1115       }
1116     }
1117     __ load_contents(sync_state, G3_scratch);
1118     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1119 
1120     Label L;
1121     __ br(Assembler::notEqual, false, Assembler::pn, L);
1122     __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1123     __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1124     __ bind(L);
1125 
1126     // Block.  Save any potential method result value before the operation and
1127     // use a leaf call to leave the last_Java_frame setup undisturbed.
1128     save_native_result();
1129     __ call_VM_leaf(L7_thread_cache,
1130                     CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1131                     G2_thread);
1132 
1133     // Restore any method result value
1134     restore_native_result();
1135     __ bind(no_block);
1136   }
1137 
1138   // Clear the frame anchor now
1139 
1140   __ reset_last_Java_frame();
1141 
1142   // Move the result handler address
1143   __ mov(Lscratch, G3_scratch);
1144   // return possible result to the outer frame
1145 #ifndef __LP64
1146   __ mov(O0, I0);
1147   __ restore(O1, G0, O1);
1148 #else
1149   __ restore(O0, G0, O0);
1150 #endif /* __LP64 */
1151 
1152   // Move result handler to expected register
1153   __ mov(G3_scratch, Lscratch);
1154 
1155   // Back in normal (native) interpreter frame. State is thread_in_native_trans
1156   // switch to thread_in_Java.
1157 
1158   __ set(_thread_in_Java, G3_scratch);
1159   __ st(G3_scratch, thread_state);
1160 
1161   // reset handle block
1162   __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1163   __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1164 
1165   // If we have an oop result store it where it will be safe for any further gc
1166   // until we return now that we've released the handle it might be protected by
1167 
1168   {
1169     Label no_oop, store_result;
1170 
1171     __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1172     __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1173     __ addcc(G0, O0, O0);
1174     __ brx(Assembler::notZero, true, Assembler::pt, store_result);     // if result is not NULL:
1175     __ delayed()->ld_ptr(O0, 0, O0);                                   // unbox it
1176     __ mov(G0, O0);
1177 
1178     __ bind(store_result);
1179     // Store it where gc will look for it and result handler expects it.
1180     __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1181 
1182     __ bind(no_oop);
1183 
1184   }
1185 
1186 
1187   // handle exceptions (exception handling will handle unlocking!)
1188   { Label L;
1189     Address exception_addr(G2_thread, Thread::pending_exception_offset());
1190     __ ld_ptr(exception_addr, Gtemp);
1191     __ br_null_short(Gtemp, Assembler::pt, L);
1192     // Note: This could be handled more efficiently since we know that the native
1193     //       method doesn't have an exception handler. We could directly return
1194     //       to the exception handler for the caller.
1195     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1196     __ should_not_reach_here();
1197     __ bind(L);
1198   }
1199 
1200   // JVMTI support (preserves thread register)
1201   __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1202 
1203   if (synchronized) {
1204     // save and restore any potential method result value around the unlocking operation
1205     save_native_result();
1206 
1207     __ add( __ top_most_monitor(), O1);
1208     __ unlock_object(O1);
1209 
1210     restore_native_result();
1211   }
1212 
1213 #if defined(COMPILER2) && !defined(_LP64)
1214 
1215   // C2 expects long results in G1 we can't tell if we're returning to interpreted
1216   // or compiled so just be safe.
1217 
1218   __ sllx(O0, 32, G1);          // Shift bits into high G1
1219   __ srl (O1, 0, O1);           // Zero extend O1
1220   __ or3 (O1, G1, G1);          // OR 64 bits into G1
1221 
1222 #endif /* COMPILER2 && !_LP64 */
1223 
1224   // dispose of return address and remove activation
1225 #ifdef ASSERT
1226   {
1227     Label ok;
1228     __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1229     __ stop("bad I5_savedSP value");
1230     __ should_not_reach_here();
1231     __ bind(ok);
1232   }
1233 #endif
1234   if (TraceJumps) {
1235     // Move target to register that is recordable
1236     __ mov(Lscratch, G3_scratch);
1237     __ JMP(G3_scratch, 0);
1238   } else {
1239     __ jmp(Lscratch, 0);
1240   }
1241   __ delayed()->nop();
1242 
1243 
1244   if (inc_counter) {
1245     // handle invocation counter overflow
1246     __ bind(invocation_counter_overflow);
1247     generate_counter_overflow(Lcontinue);
1248   }
1249 
1250 
1251 
1252   return entry;
1253 }
1254 
1255 
1256 // Generic method entry to (asm) interpreter
1257 //------------------------------------------------------------------------------------------------------------------------
1258 //
1259 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1260   address entry = __ pc();
1261 
1262   bool inc_counter  = UseCompiler || CountCompiledCalls;
1263 
1264   // the following temporary registers are used during frame creation
1265   const Register Gtmp1 = G3_scratch ;
1266   const Register Gtmp2 = G1_scratch;
1267 
1268   // make sure registers are different!
1269   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1270 
1271   const Address constMethod       (G5_method, Method::const_offset());
1272   // Seems like G5_method is live at the point this is used. So we could make this look consistent
1273   // and use in the asserts.
1274   const Address access_flags      (Lmethod,   Method::access_flags_offset());
1275 
1276   const Register Glocals_size = G3;
1277   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1278 
1279   // make sure method is not native & not abstract
1280   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1281 #ifdef ASSERT
1282   __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1283   {
1284     Label L;
1285     __ btst(JVM_ACC_NATIVE, Gtmp1);
1286     __ br(Assembler::zero, false, Assembler::pt, L);
1287     __ delayed()->nop();
1288     __ stop("tried to execute native method as non-native");
1289     __ bind(L);
1290   }
1291   { Label L;
1292     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1293     __ br(Assembler::zero, false, Assembler::pt, L);
1294     __ delayed()->nop();
1295     __ stop("tried to execute abstract method as non-abstract");
1296     __ bind(L);
1297   }
1298 #endif // ASSERT
1299 
1300   // generate the code to allocate the interpreter stack frame
1301 
1302   generate_fixed_frame(false);
1303 
1304 #ifdef FAST_DISPATCH
1305   __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1306                                           // set bytecode dispatch table base
1307 #endif
1308 
1309   //
1310   // Code to initialize the extra (i.e. non-parm) locals
1311   //
1312   Register init_value = noreg;    // will be G0 if we must clear locals
1313   // The way the code was setup before zerolocals was always true for vanilla java entries.
1314   // It could only be false for the specialized entries like accessor or empty which have
1315   // no extra locals so the testing was a waste of time and the extra locals were always
1316   // initialized. We removed this extra complication to already over complicated code.
1317 
1318   init_value = G0;
1319   Label clear_loop;
1320 
1321   const Register RconstMethod = O1;
1322   const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1323   const Address size_of_locals    (RconstMethod, ConstMethod::size_of_locals_offset());
1324 
1325   // NOTE: If you change the frame layout, this code will need to
1326   // be updated!
1327   __ ld_ptr( constMethod, RconstMethod );
1328   __ lduh( size_of_locals, O2 );
1329   __ lduh( size_of_parameters, O1 );
1330   __ sll( O2, Interpreter::logStackElementSize, O2);
1331   __ sll( O1, Interpreter::logStackElementSize, O1 );
1332   __ sub( Llocals, O2, O2 );
1333   __ sub( Llocals, O1, O1 );
1334 
1335   __ bind( clear_loop );
1336   __ inc( O2, wordSize );
1337 
1338   __ cmp( O2, O1 );
1339   __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1340   __ delayed()->st_ptr( init_value, O2, 0 );
1341 
1342   const Address do_not_unlock_if_synchronized(G2_thread,
1343     JavaThread::do_not_unlock_if_synchronized_offset());
1344   // Since at this point in the method invocation the exception handler
1345   // would try to exit the monitor of synchronized methods which hasn't
1346   // been entered yet, we set the thread local variable
1347   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1348   // runtime, exception handling i.e. unlock_if_synchronized_method will
1349   // check this thread local flag.
1350   __ movbool(true, G3_scratch);
1351   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1352 
1353   // increment invocation counter and check for overflow
1354   //
1355   // Note: checking for negative value instead of overflow
1356   //       so we have a 'sticky' overflow test (may be of
1357   //       importance as soon as we have true MT/MP)
1358   Label invocation_counter_overflow;
1359   Label profile_method;
1360   Label profile_method_continue;
1361   Label Lcontinue;
1362   if (inc_counter) {
1363     generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1364     if (ProfileInterpreter) {
1365       __ bind(profile_method_continue);
1366     }
1367   }
1368   __ bind(Lcontinue);
1369 
1370   bang_stack_shadow_pages(false);
1371 
1372   // reset the _do_not_unlock_if_synchronized flag
1373   __ stbool(G0, do_not_unlock_if_synchronized);
1374 
1375   // check for synchronized methods
1376   // Must happen AFTER invocation_counter check and stack overflow check,
1377   // so method is not locked if overflows.
1378 
1379   if (synchronized) {
1380     lock_method();
1381   } else {
1382 #ifdef ASSERT
1383     { Label ok;
1384       __ ld(access_flags, O0);
1385       __ btst(JVM_ACC_SYNCHRONIZED, O0);
1386       __ br( Assembler::zero, false, Assembler::pt, ok);
1387       __ delayed()->nop();
1388       __ stop("method needs synchronization");
1389       __ bind(ok);
1390     }
1391 #endif // ASSERT
1392   }
1393 
1394   // start execution
1395 
1396   __ verify_thread();
1397 
1398   // jvmti support
1399   __ notify_method_entry();
1400 
1401   // start executing instructions
1402   __ dispatch_next(vtos);
1403 
1404 
1405   if (inc_counter) {
1406     if (ProfileInterpreter) {
1407       // We have decided to profile this method in the interpreter
1408       __ bind(profile_method);
1409 
1410       __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1411       __ set_method_data_pointer_for_bcp();
1412       __ ba_short(profile_method_continue);
1413     }
1414 
1415     // handle invocation counter overflow
1416     __ bind(invocation_counter_overflow);
1417     generate_counter_overflow(Lcontinue);
1418   }
1419 
1420 
1421   return entry;
1422 }
1423 
1424 
1425 //----------------------------------------------------------------------------------------------------
1426 // Entry points & stack frame layout
1427 //
1428 // Here we generate the various kind of entries into the interpreter.
1429 // The two main entry type are generic bytecode methods and native call method.
1430 // These both come in synchronized and non-synchronized versions but the
1431 // frame layout they create is very similar. The other method entry
1432 // types are really just special purpose entries that are really entry
1433 // and interpretation all in one. These are for trivial methods like
1434 // accessor, empty, or special math methods.
1435 //
1436 // When control flow reaches any of the entry types for the interpreter
1437 // the following holds ->
1438 //
1439 // C2 Calling Conventions:
1440 //
1441 // The entry code below assumes that the following registers are set
1442 // when coming in:
1443 //    G5_method: holds the Method* of the method to call
1444 //    Lesp:    points to the TOS of the callers expression stack
1445 //             after having pushed all the parameters
1446 //
1447 // The entry code does the following to setup an interpreter frame
1448 //   pop parameters from the callers stack by adjusting Lesp
1449 //   set O0 to Lesp
1450 //   compute X = (max_locals - num_parameters)
1451 //   bump SP up by X to accomadate the extra locals
1452 //   compute X = max_expression_stack
1453 //               + vm_local_words
1454 //               + 16 words of register save area
1455 //   save frame doing a save sp, -X, sp growing towards lower addresses
1456 //   set Lbcp, Lmethod, LcpoolCache
1457 //   set Llocals to i0
1458 //   set Lmonitors to FP - rounded_vm_local_words
1459 //   set Lesp to Lmonitors - 4
1460 //
1461 //  The frame has now been setup to do the rest of the entry code
1462 
1463 // Try this optimization:  Most method entries could live in a
1464 // "one size fits all" stack frame without all the dynamic size
1465 // calculations.  It might be profitable to do all this calculation
1466 // statically and approximately for "small enough" methods.
1467 
1468 //-----------------------------------------------------------------------------------------------
1469 
1470 // C1 Calling conventions
1471 //
1472 // Upon method entry, the following registers are setup:
1473 //
1474 // g2 G2_thread: current thread
1475 // g5 G5_method: method to activate
1476 // g4 Gargs  : pointer to last argument
1477 //
1478 //
1479 // Stack:
1480 //
1481 // +---------------+ <--- sp
1482 // |               |
1483 // : reg save area :
1484 // |               |
1485 // +---------------+ <--- sp + 0x40
1486 // |               |
1487 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
1488 // |               |
1489 // +---------------+ <--- sp + 0x5c
1490 // |               |
1491 // :     free      :
1492 // |               |
1493 // +---------------+ <--- Gargs
1494 // |               |
1495 // :   arguments   :
1496 // |               |
1497 // +---------------+
1498 // |               |
1499 //
1500 //
1501 //
1502 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
1503 //
1504 // +---------------+ <--- sp
1505 // |               |
1506 // : reg save area :
1507 // |               |
1508 // +---------------+ <--- sp + 0x40
1509 // |               |
1510 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
1511 // |               |
1512 // +---------------+ <--- sp + 0x5c
1513 // |               |
1514 // :               :
1515 // |               | <--- Lesp
1516 // +---------------+ <--- Lmonitors (fp - 0x18)
1517 // |   VM locals   |
1518 // +---------------+ <--- fp
1519 // |               |
1520 // : reg save area :
1521 // |               |
1522 // +---------------+ <--- fp + 0x40
1523 // |               |
1524 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
1525 // |               |
1526 // +---------------+ <--- fp + 0x5c
1527 // |               |
1528 // :     free      :
1529 // |               |
1530 // +---------------+
1531 // |               |
1532 // : nonarg locals :
1533 // |               |
1534 // +---------------+
1535 // |               |
1536 // :   arguments   :
1537 // |               | <--- Llocals
1538 // +---------------+ <--- Gargs
1539 // |               |
1540 
1541 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
1542 
1543   // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
1544   // expression stack, the callee will have callee_extra_locals (so we can account for
1545   // frame extension) and monitor_size for monitors. Basically we need to calculate
1546   // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
1547   //
1548   //
1549   // The big complicating thing here is that we must ensure that the stack stays properly
1550   // aligned. This would be even uglier if monitor size wasn't modulo what the stack
1551   // needs to be aligned for). We are given that the sp (fp) is already aligned by
1552   // the caller so we must ensure that it is properly aligned for our callee.
1553   //
1554   const int rounded_vm_local_words =
1555        round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1556   // callee_locals and max_stack are counts, not the size in frame.
1557   const int locals_size =
1558        round_to(callee_extra_locals * Interpreter::stackElementWords, WordsPerLong);
1559   const int max_stack_words = max_stack * Interpreter::stackElementWords;
1560   return (round_to((max_stack_words
1561                    //6815692//+ Method::extra_stack_words()
1562                    + rounded_vm_local_words
1563                    + frame::memory_parameter_word_sp_offset), WordsPerLong)
1564                    // already rounded
1565                    + locals_size + monitor_size);
1566 }
1567 
1568 // How much stack a method top interpreter activation needs in words.
1569 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
1570 
1571   // See call_stub code
1572   int call_stub_size  = round_to(7 + frame::memory_parameter_word_sp_offset,
1573                                  WordsPerLong);    // 7 + register save area
1574 
1575   // Save space for one monitor to get into the interpreted method in case
1576   // the method is synchronized
1577   int monitor_size    = method->is_synchronized() ?
1578                                 1*frame::interpreter_frame_monitor_size() : 0;
1579   return size_activation_helper(method->max_locals(), method->max_stack(),
1580                                  monitor_size) + call_stub_size;
1581 }
1582 
1583 int AbstractInterpreter::layout_activation(Method* method,
1584                                            int tempcount,
1585                                            int popframe_extra_args,
1586                                            int moncount,
1587                                            int caller_actual_parameters,
1588                                            int callee_param_count,
1589                                            int callee_local_count,
1590                                            frame* caller,
1591                                            frame* interpreter_frame,
1592                                            bool is_top_frame,
1593                                            bool is_bottom_frame) {
1594   // Note: This calculation must exactly parallel the frame setup
1595   // in InterpreterGenerator::generate_fixed_frame.
1596   // If f!=NULL, set up the following variables:
1597   //   - Lmethod
1598   //   - Llocals
1599   //   - Lmonitors (to the indicated number of monitors)
1600   //   - Lesp (to the indicated number of temps)
1601   // The frame f (if not NULL) on entry is a description of the caller of the frame
1602   // we are about to layout. We are guaranteed that we will be able to fill in a
1603   // new interpreter frame as its callee (i.e. the stack space is allocated and
1604   // the amount was determined by an earlier call to this method with f == NULL).
1605   // On return f (if not NULL) while describe the interpreter frame we just layed out.
1606 
1607   int monitor_size           = moncount * frame::interpreter_frame_monitor_size();
1608   int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1609 
1610   assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1611   //
1612   // Note: if you look closely this appears to be doing something much different
1613   // than generate_fixed_frame. What is happening is this. On sparc we have to do
1614   // this dance with interpreter_sp_adjustment because the window save area would
1615   // appear just below the bottom (tos) of the caller's java expression stack. Because
1616   // the interpreter want to have the locals completely contiguous generate_fixed_frame
1617   // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size).
1618   // Now in generate_fixed_frame the extension of the caller's sp happens in the callee.
1619   // In this code the opposite occurs the caller adjusts it's own stack base on the callee.
1620   // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest)
1621   // because the oldest frame would have adjust its callers frame and yet that frame
1622   // already exists and isn't part of this array of frames we are unpacking. So at first
1623   // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper()
1624   // will after it calculates all of the frame's on_stack_size()'s will then figure out the
1625   // amount to adjust the caller of the initial (oldest) frame and the calculation will all
1626   // add up. It does seem like it simpler to account for the adjustment here (and remove the
1627   // callee... parameters here). However this would mean that this routine would have to take
1628   // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment)
1629   // and run the calling loop in the reverse order. This would also would appear to mean making
1630   // this code aware of what the interactions are when that initial caller fram was an osr or
1631   // other adapter frame. deoptimization is complicated enough and  hard enough to debug that
1632   // there is no sense in messing working code.
1633   //
1634 
1635   int rounded_cls = round_to((callee_local_count - callee_param_count), WordsPerLong);
1636   assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align");
1637 
1638   int raw_frame_size = size_activation_helper(rounded_cls, method->max_stack(),
1639                                               monitor_size);
1640 
1641   if (interpreter_frame != NULL) {
1642     // The skeleton frame must already look like an interpreter frame
1643     // even if not fully filled out.
1644     assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame");
1645 
1646     intptr_t* fp = interpreter_frame->fp();
1647 
1648     JavaThread* thread = JavaThread::current();
1649     RegisterMap map(thread, false);
1650     // More verification that skeleton frame is properly walkable
1651     assert(fp == caller->sp(), "fp must match");
1652 
1653     intptr_t* montop     = fp - rounded_vm_local_words;
1654 
1655     // preallocate monitors (cf. __ add_monitor_to_stack)
1656     intptr_t* monitors = montop - monitor_size;
1657 
1658     // preallocate stack space
1659     intptr_t*  esp = monitors - 1 -
1660                      (tempcount * Interpreter::stackElementWords) -
1661                      popframe_extra_args;
1662 
1663     int local_words = method->max_locals() * Interpreter::stackElementWords;
1664     NEEDS_CLEANUP;
1665     intptr_t* locals;
1666     if (caller->is_interpreted_frame()) {
1667       // Can force the locals area to end up properly overlapping the top of the expression stack.
1668       intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1;
1669       // Note that this computation means we replace size_of_parameters() values from the caller
1670       // interpreter frame's expression stack with our argument locals
1671       int parm_words  = caller_actual_parameters * Interpreter::stackElementWords;
1672       locals = Lesp_ptr + parm_words;
1673       int delta = local_words - parm_words;
1674       int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0;
1675       *interpreter_frame->register_addr(I5_savedSP)    = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS;
1676       if (!is_bottom_frame) {
1677         // Llast_SP is set below for the current frame to SP (with the
1678         // extra space for the callee's locals). Here we adjust
1679         // Llast_SP for the caller's frame, removing the extra space
1680         // for the current method's locals.
1681         *caller->register_addr(Llast_SP) = *interpreter_frame->register_addr(I5_savedSP);
1682       } else {
1683         assert(*caller->register_addr(Llast_SP) >= *interpreter_frame->register_addr(I5_savedSP), "strange Llast_SP");
1684       }
1685     } else {
1686       assert(caller->is_compiled_frame() || caller->is_entry_frame(), "only possible cases");
1687       // Don't have Lesp available; lay out locals block in the caller
1688       // adjacent to the register window save area.
1689       //
1690       // Compiled frames do not allocate a varargs area which is why this if
1691       // statement is needed.
1692       //
1693       if (caller->is_compiled_frame()) {
1694         locals = fp + frame::register_save_words + local_words - 1;
1695       } else {
1696         locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
1697       }
1698       if (!caller->is_entry_frame()) {
1699         // Caller wants his own SP back
1700         int caller_frame_size = caller->cb()->frame_size();
1701         *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS;
1702       }
1703     }
1704     if (TraceDeoptimization) {
1705       if (caller->is_entry_frame()) {
1706         // make sure I5_savedSP and the entry frames notion of saved SP
1707         // agree.  This assertion duplicate a check in entry frame code
1708         // but catches the failure earlier.
1709         assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP),
1710                "would change callers SP");
1711       }
1712       if (caller->is_entry_frame()) {
1713         tty->print("entry ");
1714       }
1715       if (caller->is_compiled_frame()) {
1716         tty->print("compiled ");
1717         if (caller->is_deoptimized_frame()) {
1718           tty->print("(deopt) ");
1719         }
1720       }
1721       if (caller->is_interpreted_frame()) {
1722         tty->print("interpreted ");
1723       }
1724       tty->print_cr("caller fp=0x%x sp=0x%x", caller->fp(), caller->sp());
1725       tty->print_cr("save area = 0x%x, 0x%x", caller->sp(), caller->sp() + 16);
1726       tty->print_cr("save area = 0x%x, 0x%x", caller->fp(), caller->fp() + 16);
1727       tty->print_cr("interpreter fp=0x%x sp=0x%x", interpreter_frame->fp(), interpreter_frame->sp());
1728       tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->sp(), interpreter_frame->sp() + 16);
1729       tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->fp(), interpreter_frame->fp() + 16);
1730       tty->print_cr("Llocals = 0x%x", locals);
1731       tty->print_cr("Lesp = 0x%x", esp);
1732       tty->print_cr("Lmonitors = 0x%x", monitors);
1733     }
1734 
1735     if (method->max_locals() > 0) {
1736       assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area");
1737       assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area");
1738       assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area");
1739       assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area");
1740     }
1741 #ifdef _LP64
1742     assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd");
1743 #endif
1744 
1745     *interpreter_frame->register_addr(Lmethod)     = (intptr_t) method;
1746     *interpreter_frame->register_addr(Llocals)     = (intptr_t) locals;
1747     *interpreter_frame->register_addr(Lmonitors)   = (intptr_t) monitors;
1748     *interpreter_frame->register_addr(Lesp)        = (intptr_t) esp;
1749     // Llast_SP will be same as SP as there is no adapter space
1750     *interpreter_frame->register_addr(Llast_SP)    = (intptr_t) interpreter_frame->sp() - STACK_BIAS;
1751     *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache();
1752 #ifdef FAST_DISPATCH
1753     *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table();
1754 #endif
1755 
1756 
1757 #ifdef ASSERT
1758     BasicObjectLock* mp = (BasicObjectLock*)monitors;
1759 
1760     assert(interpreter_frame->interpreter_frame_method() == method, "method matches");
1761     assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize)), "locals match");
1762     assert(interpreter_frame->interpreter_frame_monitor_end()   == mp, "monitor_end matches");
1763     assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches");
1764     assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches");
1765 
1766     // check bounds
1767     intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1);
1768     intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words;
1769     assert(lo < monitors && montop <= hi, "monitors in bounds");
1770     assert(lo <= esp && esp < monitors, "esp in bounds");
1771 #endif // ASSERT
1772   }
1773 
1774   return raw_frame_size;
1775 }
1776 
1777 //----------------------------------------------------------------------------------------------------
1778 // Exceptions
1779 void TemplateInterpreterGenerator::generate_throw_exception() {
1780 
1781   // Entry point in previous activation (i.e., if the caller was interpreted)
1782   Interpreter::_rethrow_exception_entry = __ pc();
1783   // O0: exception
1784 
1785   // entry point for exceptions thrown within interpreter code
1786   Interpreter::_throw_exception_entry = __ pc();
1787   __ verify_thread();
1788   // expression stack is undefined here
1789   // O0: exception, i.e. Oexception
1790   // Lbcp: exception bcx
1791   __ verify_oop(Oexception);
1792 
1793 
1794   // expression stack must be empty before entering the VM in case of an exception
1795   __ empty_expression_stack();
1796   // find exception handler address and preserve exception oop
1797   // call C routine to find handler and jump to it
1798   __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1799   __ push_ptr(O1); // push exception for exception handler bytecodes
1800 
1801   __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1802   __ delayed()->nop();
1803 
1804 
1805   // if the exception is not handled in the current frame
1806   // the frame is removed and the exception is rethrown
1807   // (i.e. exception continuation is _rethrow_exception)
1808   //
1809   // Note: At this point the bci is still the bxi for the instruction which caused
1810   //       the exception and the expression stack is empty. Thus, for any VM calls
1811   //       at this point, GC will find a legal oop map (with empty expression stack).
1812 
1813   // in current activation
1814   // tos: exception
1815   // Lbcp: exception bcp
1816 
1817   //
1818   // JVMTI PopFrame support
1819   //
1820 
1821   Interpreter::_remove_activation_preserving_args_entry = __ pc();
1822   Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1823   // Set the popframe_processing bit in popframe_condition indicating that we are
1824   // currently handling popframe, so that call_VMs that may happen later do not trigger new
1825   // popframe handling cycles.
1826 
1827   __ ld(popframe_condition_addr, G3_scratch);
1828   __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1829   __ stw(G3_scratch, popframe_condition_addr);
1830 
1831   // Empty the expression stack, as in normal exception handling
1832   __ empty_expression_stack();
1833   __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1834 
1835   {
1836     // Check to see whether we are returning to a deoptimized frame.
1837     // (The PopFrame call ensures that the caller of the popped frame is
1838     // either interpreted or compiled and deoptimizes it if compiled.)
1839     // In this case, we can't call dispatch_next() after the frame is
1840     // popped, but instead must save the incoming arguments and restore
1841     // them after deoptimization has occurred.
1842     //
1843     // Note that we don't compare the return PC against the
1844     // deoptimization blob's unpack entry because of the presence of
1845     // adapter frames in C2.
1846     Label caller_not_deoptimized;
1847     __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1848     __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1849 
1850     const Register Gtmp1 = G3_scratch;
1851     const Register Gtmp2 = G1_scratch;
1852     const Register RconstMethod = Gtmp1;
1853     const Address constMethod(Lmethod, Method::const_offset());
1854     const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1855 
1856     // Compute size of arguments for saving when returning to deoptimized caller
1857     __ ld_ptr(constMethod, RconstMethod);
1858     __ lduh(size_of_parameters, Gtmp1);
1859     __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1860     __ sub(Llocals, Gtmp1, Gtmp2);
1861     __ add(Gtmp2, wordSize, Gtmp2);
1862     // Save these arguments
1863     __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1864     // Inform deoptimization that it is responsible for restoring these arguments
1865     __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1866     Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1867     __ st(Gtmp1, popframe_condition_addr);
1868 
1869     // Return from the current method
1870     // The caller's SP was adjusted upon method entry to accomodate
1871     // the callee's non-argument locals. Undo that adjustment.
1872     __ ret();
1873     __ delayed()->restore(I5_savedSP, G0, SP);
1874 
1875     __ bind(caller_not_deoptimized);
1876   }
1877 
1878   // Clear the popframe condition flag
1879   __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1880 
1881   // Get out of the current method (how this is done depends on the particular compiler calling
1882   // convention that the interpreter currently follows)
1883   // The caller's SP was adjusted upon method entry to accomodate
1884   // the callee's non-argument locals. Undo that adjustment.
1885   __ restore(I5_savedSP, G0, SP);
1886   // The method data pointer was incremented already during
1887   // call profiling. We have to restore the mdp for the current bcp.
1888   if (ProfileInterpreter) {
1889     __ set_method_data_pointer_for_bcp();
1890   }
1891   // Resume bytecode interpretation at the current bcp
1892   __ dispatch_next(vtos);
1893   // end of JVMTI PopFrame support
1894 
1895   Interpreter::_remove_activation_entry = __ pc();
1896 
1897   // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1898   __ pop_ptr(Oexception);                                  // get exception
1899 
1900   // Intel has the following comment:
1901   //// remove the activation (without doing throws on illegalMonitorExceptions)
1902   // They remove the activation without checking for bad monitor state.
1903   // %%% We should make sure this is the right semantics before implementing.
1904 
1905   __ set_vm_result(Oexception);
1906   __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1907 
1908   __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1909 
1910   __ get_vm_result(Oexception);
1911   __ verify_oop(Oexception);
1912 
1913     const int return_reg_adjustment = frame::pc_return_offset;
1914   Address issuing_pc_addr(I7, return_reg_adjustment);
1915 
1916   // We are done with this activation frame; find out where to go next.
1917   // The continuation point will be an exception handler, which expects
1918   // the following registers set up:
1919   //
1920   // Oexception: exception
1921   // Oissuing_pc: the local call that threw exception
1922   // Other On: garbage
1923   // In/Ln:  the contents of the caller's register window
1924   //
1925   // We do the required restore at the last possible moment, because we
1926   // need to preserve some state across a runtime call.
1927   // (Remember that the caller activation is unknown--it might not be
1928   // interpreted, so things like Lscratch are useless in the caller.)
1929 
1930   // Although the Intel version uses call_C, we can use the more
1931   // compact call_VM.  (The only real difference on SPARC is a
1932   // harmlessly ignored [re]set_last_Java_frame, compared with
1933   // the Intel code which lacks this.)
1934   __ mov(Oexception,      Oexception ->after_save());  // get exception in I0 so it will be on O0 after restore
1935   __ add(issuing_pc_addr, Oissuing_pc->after_save());  // likewise set I1 to a value local to the caller
1936   __ super_call_VM_leaf(L7_thread_cache,
1937                         CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1938                         G2_thread, Oissuing_pc->after_save());
1939 
1940   // The caller's SP was adjusted upon method entry to accomodate
1941   // the callee's non-argument locals. Undo that adjustment.
1942   __ JMP(O0, 0);                         // return exception handler in caller
1943   __ delayed()->restore(I5_savedSP, G0, SP);
1944 
1945   // (same old exception object is already in Oexception; see above)
1946   // Note that an "issuing PC" is actually the next PC after the call
1947 }
1948 
1949 
1950 //
1951 // JVMTI ForceEarlyReturn support
1952 //
1953 
1954 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1955   address entry = __ pc();
1956 
1957   __ empty_expression_stack();
1958   __ load_earlyret_value(state);
1959 
1960   __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1961   Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1962 
1963   // Clear the earlyret state
1964   __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1965 
1966   __ remove_activation(state,
1967                        /* throw_monitor_exception */ false,
1968                        /* install_monitor_exception */ false);
1969 
1970   // The caller's SP was adjusted upon method entry to accomodate
1971   // the callee's non-argument locals. Undo that adjustment.
1972   __ ret();                             // return to caller
1973   __ delayed()->restore(I5_savedSP, G0, SP);
1974 
1975   return entry;
1976 } // end of JVMTI ForceEarlyReturn support
1977 
1978 
1979 //------------------------------------------------------------------------------------------------------------------------
1980 // Helper for vtos entry point generation
1981 
1982 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) {
1983   assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1984   Label L;
1985   aep = __ pc(); __ push_ptr(); __ ba_short(L);
1986   fep = __ pc(); __ push_f();   __ ba_short(L);
1987   dep = __ pc(); __ push_d();   __ ba_short(L);
1988   lep = __ pc(); __ push_l();   __ ba_short(L);
1989   iep = __ pc(); __ push_i();
1990   bep = cep = sep = iep;                        // there aren't any
1991   vep = __ pc(); __ bind(L);                    // fall through
1992   generate_and_dispatch(t);
1993 }
1994 
1995 // --------------------------------------------------------------------------------
1996 
1997 
1998 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
1999  : TemplateInterpreterGenerator(code) {
2000    generate_all(); // down here so it can be "virtual"
2001 }
2002 
2003 // --------------------------------------------------------------------------------
2004 
2005 // Non-product code
2006 #ifndef PRODUCT
2007 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
2008   address entry = __ pc();
2009 
2010   __ push(state);
2011   __ mov(O7, Lscratch); // protect return address within interpreter
2012 
2013   // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
2014   __ mov( Otos_l2, G3_scratch );
2015   __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
2016   __ mov(Lscratch, O7); // restore return address
2017   __ pop(state);
2018   __ retl();
2019   __ delayed()->nop();
2020 
2021   return entry;
2022 }
2023 
2024 
2025 // helpers for generate_and_dispatch
2026 
2027 void TemplateInterpreterGenerator::count_bytecode() {
2028   __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
2029 }
2030 
2031 
2032 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
2033   __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
2034 }
2035 
2036 
2037 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2038   AddressLiteral index   (&BytecodePairHistogram::_index);
2039   AddressLiteral counters((address) &BytecodePairHistogram::_counters);
2040 
2041   // get index, shift out old bytecode, bring in new bytecode, and store it
2042   // _index = (_index >> log2_number_of_codes) |
2043   //          (bytecode << log2_number_of_codes);
2044 
2045   __ load_contents(index, G4_scratch);
2046   __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
2047   __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes,  G3_scratch );
2048   __ or3( G3_scratch,  G4_scratch, G4_scratch );
2049   __ store_contents(G4_scratch, index, G3_scratch);
2050 
2051   // bump bucket contents
2052   // _counters[_index] ++;
2053 
2054   __ set(counters, G3_scratch);                       // loads into G3_scratch
2055   __ sll( G4_scratch, LogBytesPerWord, G4_scratch );  // Index is word address
2056   __ add (G3_scratch, G4_scratch, G3_scratch);        // Add in index
2057   __ ld (G3_scratch, 0, G4_scratch);
2058   __ inc (G4_scratch);
2059   __ st (G4_scratch, 0, G3_scratch);
2060 }
2061 
2062 
2063 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2064   // Call a little run-time stub to avoid blow-up for each bytecode.
2065   // The run-time runtime saves the right registers, depending on
2066   // the tosca in-state for the given template.
2067   address entry = Interpreter::trace_code(t->tos_in());
2068   guarantee(entry != NULL, "entry must have been generated");
2069   __ call(entry, relocInfo::none);
2070   __ delayed()->nop();
2071 }
2072 
2073 
2074 void TemplateInterpreterGenerator::stop_interpreter_at() {
2075   AddressLiteral counter(&BytecodeCounter::_counter_value);
2076   __ load_contents(counter, G3_scratch);
2077   AddressLiteral stop_at(&StopInterpreterAt);
2078   __ load_ptr_contents(stop_at, G4_scratch);
2079   __ cmp(G3_scratch, G4_scratch);
2080   __ breakpoint_trap(Assembler::equal, Assembler::icc);
2081 }
2082 #endif // not PRODUCT
2083 #endif // !CC_INTERP