1 /*
   2  * Copyright (c) 1997, 2016, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/macroAssembler.hpp"
  27 #include "interpreter/bytecodeHistogram.hpp"
  28 #include "interpreter/interpreter.hpp"
  29 #include "interpreter/interpreterGenerator.hpp"
  30 #include "interpreter/interpreterRuntime.hpp"
  31 #include "interpreter/templateTable.hpp"
  32 #include "oops/arrayOop.hpp"
  33 #include "oops/methodData.hpp"
  34 #include "oops/method.hpp"
  35 #include "oops/oop.inline.hpp"
  36 #include "prims/jvmtiExport.hpp"
  37 #include "prims/jvmtiThreadState.hpp"
  38 #include "runtime/arguments.hpp"
  39 #include "runtime/deoptimization.hpp"
  40 #include "runtime/frame.inline.hpp"
  41 #include "runtime/sharedRuntime.hpp"
  42 #include "runtime/stubRoutines.hpp"
  43 #include "runtime/synchronizer.hpp"
  44 #include "runtime/timer.hpp"
  45 #include "runtime/vframeArray.hpp"
  46 #include "utilities/debug.hpp"
  47 #include "utilities/macros.hpp"
  48 
  49 #ifndef CC_INTERP
  50 #ifndef FAST_DISPATCH
  51 #define FAST_DISPATCH 1
  52 #endif
  53 #undef FAST_DISPATCH
  54 
  55 
  56 // Generation of Interpreter
  57 //
  58 // The InterpreterGenerator generates the interpreter into Interpreter::_code.
  59 
  60 
  61 #define __ _masm->
  62 
  63 
  64 //----------------------------------------------------------------------------------------------------
  65 
  66 
  67 void InterpreterGenerator::save_native_result(void) {
  68   // result potentially in O0/O1: save it across calls
  69   const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
  70 
  71   // result potentially in F0/F1: save it across calls
  72   const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
  73 
  74   // save and restore any potential method result value around the unlocking operation
  75   __ stf(FloatRegisterImpl::D, F0, d_tmp);
  76 #ifdef _LP64
  77   __ stx(O0, l_tmp);
  78 #else
  79   __ std(O0, l_tmp);
  80 #endif
  81 }
  82 
  83 void InterpreterGenerator::restore_native_result(void) {
  84   const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
  85   const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
  86 
  87   // Restore any method result value
  88   __ ldf(FloatRegisterImpl::D, d_tmp, F0);
  89 #ifdef _LP64
  90   __ ldx(l_tmp, O0);
  91 #else
  92   __ ldd(l_tmp, O0);
  93 #endif
  94 }
  95 
  96 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
  97   assert(!pass_oop || message == NULL, "either oop or message but not both");
  98   address entry = __ pc();
  99   // expression stack must be empty before entering the VM if an exception happened
 100   __ empty_expression_stack();
 101   // load exception object
 102   __ set((intptr_t)name, G3_scratch);
 103   if (pass_oop) {
 104     __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
 105   } else {
 106     __ set((intptr_t)message, G4_scratch);
 107     __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
 108   }
 109   // throw exception
 110   assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
 111   AddressLiteral thrower(Interpreter::throw_exception_entry());
 112   __ jump_to(thrower, G3_scratch);
 113   __ delayed()->nop();
 114   return entry;
 115 }
 116 
 117 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
 118   address entry = __ pc();
 119   // expression stack must be empty before entering the VM if an exception
 120   // happened
 121   __ empty_expression_stack();
 122   // load exception object
 123   __ call_VM(Oexception,
 124              CAST_FROM_FN_PTR(address,
 125                               InterpreterRuntime::throw_ClassCastException),
 126              Otos_i);
 127   __ should_not_reach_here();
 128   return entry;
 129 }
 130 
 131 
 132 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
 133   address entry = __ pc();
 134   // expression stack must be empty before entering the VM if an exception happened
 135   __ empty_expression_stack();
 136   // convention: expect aberrant index in register G3_scratch, then shuffle the
 137   // index to G4_scratch for the VM call
 138   __ mov(G3_scratch, G4_scratch);
 139   __ set((intptr_t)name, G3_scratch);
 140   __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
 141   __ should_not_reach_here();
 142   return entry;
 143 }
 144 
 145 
 146 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
 147   address entry = __ pc();
 148   // expression stack must be empty before entering the VM if an exception happened
 149   __ empty_expression_stack();
 150   __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
 151   __ should_not_reach_here();
 152   return entry;
 153 }
 154 
 155 
 156 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
 157   address entry = __ pc();
 158 
 159   if (state == atos) {
 160     __ profile_return_type(O0, G3_scratch, G1_scratch);
 161   }
 162 
 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 (state == ltos) {
 174     __ srl (G1,  0, O1);
 175     __ srlx(G1, 32, O0);
 176   }
 177 #endif // !_LP64 && COMPILER2
 178 
 179   // The callee returns with the stack possibly adjusted by adapter transition
 180   // We remove that possible adjustment here.
 181   // All interpreter local registers are untouched. Any result is passed back
 182   // in the O0/O1 or float registers. Before continuing, the arguments must be
 183   // popped from the java expression stack; i.e., Lesp must be adjusted.
 184 
 185   __ mov(Llast_SP, SP);   // Remove any adapter added stack space.
 186 
 187   const Register cache = G3_scratch;
 188   const Register index  = G1_scratch;
 189   __ get_cache_and_index_at_bcp(cache, index, 1, index_size);
 190 
 191   const Register flags = cache;
 192   __ ld_ptr(cache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset(), flags);
 193   const Register parameter_size = flags;
 194   __ and3(flags, ConstantPoolCacheEntry::parameter_size_mask, parameter_size);  // argument size in words
 195   __ sll(parameter_size, Interpreter::logStackElementSize, parameter_size);     // each argument size in bytes
 196   __ add(Lesp, parameter_size, Lesp);                                           // pop arguments
 197   __ dispatch_next(state, step);
 198 
 199   return entry;
 200 }
 201 
 202 
 203 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
 204   address entry = __ pc();
 205   __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
 206   { Label L;
 207     Address exception_addr(G2_thread, Thread::pending_exception_offset());
 208     __ ld_ptr(exception_addr, Gtemp);  // Load pending exception.
 209     __ br_null_short(Gtemp, Assembler::pt, L);
 210     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
 211     __ should_not_reach_here();
 212     __ bind(L);
 213   }
 214   __ dispatch_next(state, step);
 215   return entry;
 216 }
 217 
 218 // A result handler converts/unboxes a native call result into
 219 // a java interpreter/compiler result. The current frame is an
 220 // interpreter frame. The activation frame unwind code must be
 221 // consistent with that of TemplateTable::_return(...). In the
 222 // case of native methods, the caller's SP was not modified.
 223 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
 224   address entry = __ pc();
 225   Register Itos_i  = Otos_i ->after_save();
 226   Register Itos_l  = Otos_l ->after_save();
 227   Register Itos_l1 = Otos_l1->after_save();
 228   Register Itos_l2 = Otos_l2->after_save();
 229   switch (type) {
 230     case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
 231     case T_CHAR   : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i);   break; // cannot use and3, 0xFFFF too big as immediate value!
 232     case T_BYTE   : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i);   break;
 233     case T_SHORT  : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i);   break;
 234     case T_LONG   :
 235 #ifndef _LP64
 236                     __ mov(O1, Itos_l2);  // move other half of long
 237 #endif              // ifdef or no ifdef, fall through to the T_INT case
 238     case T_INT    : __ mov(O0, Itos_i);                         break;
 239     case T_VOID   : /* nothing to do */                         break;
 240     case T_FLOAT  : assert(F0 == Ftos_f, "fix this code" );     break;
 241     case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" );     break;
 242     case T_OBJECT :
 243       __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
 244       __ verify_oop(Itos_i);
 245       break;
 246     default       : ShouldNotReachHere();
 247   }
 248   __ ret();                           // return from interpreter activation
 249   __ delayed()->restore(I5_savedSP, G0, SP);  // remove interpreter frame
 250   NOT_PRODUCT(__ emit_int32(0);)       // marker for disassembly
 251   return entry;
 252 }
 253 
 254 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
 255   address entry = __ pc();
 256   __ push(state);
 257   __ call_VM(noreg, runtime_entry);
 258   __ dispatch_via(vtos, Interpreter::normal_table(vtos));
 259   return entry;
 260 }
 261 
 262 
 263 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
 264   address entry = __ pc();
 265   __ dispatch_next(state);
 266   return entry;
 267 }
 268 
 269 //
 270 // Helpers for commoning out cases in the various type of method entries.
 271 //
 272 
 273 // increment invocation count & check for overflow
 274 //
 275 // Note: checking for negative value instead of overflow
 276 //       so we have a 'sticky' overflow test
 277 //
 278 // Lmethod: method
 279 // ??: invocation counter
 280 //
 281 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
 282   // Note: In tiered we increment either counters in MethodCounters* or in
 283   // MDO depending if we're profiling or not.
 284   const Register Rcounters = G3_scratch;
 285   Label done;
 286 
 287   if (TieredCompilation) {
 288     const int increment = InvocationCounter::count_increment;
 289     const int mask = ((1 << Tier0InvokeNotifyFreqLog) - 1) << InvocationCounter::count_shift;
 290     Label no_mdo;
 291     if (ProfileInterpreter) {
 292       // If no method data exists, go to profile_continue.
 293       __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
 294       __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
 295       // Increment counter
 296       Address mdo_invocation_counter(G4_scratch,
 297                                      in_bytes(MethodData::invocation_counter_offset()) +
 298                                      in_bytes(InvocationCounter::counter_offset()));
 299       __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
 300                                  G3_scratch, Lscratch,
 301                                  Assembler::zero, overflow);
 302       __ ba_short(done);
 303     }
 304 
 305     // Increment counter in MethodCounters*
 306     __ bind(no_mdo);
 307     Address invocation_counter(Rcounters,
 308             in_bytes(MethodCounters::invocation_counter_offset()) +
 309             in_bytes(InvocationCounter::counter_offset()));
 310     __ get_method_counters(Lmethod, Rcounters, done);
 311     __ increment_mask_and_jump(invocation_counter, increment, mask,
 312                                G4_scratch, Lscratch,
 313                                Assembler::zero, overflow);
 314     __ bind(done);
 315   } else {
 316     // Update standard invocation counters
 317     __ get_method_counters(Lmethod, Rcounters, done);
 318     __ increment_invocation_counter(Rcounters, O0, G4_scratch);
 319     if (ProfileInterpreter) {
 320       Address interpreter_invocation_counter(Rcounters,
 321             in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
 322       __ ld(interpreter_invocation_counter, G4_scratch);
 323       __ inc(G4_scratch);
 324       __ st(G4_scratch, interpreter_invocation_counter);
 325     }
 326 
 327     if (ProfileInterpreter && profile_method != NULL) {
 328       // Test to see if we should create a method data oop
 329       AddressLiteral profile_limit((address)&InvocationCounter::InterpreterProfileLimit);
 330       __ load_contents(profile_limit, G3_scratch);
 331       __ cmp_and_br_short(O0, G3_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
 332 
 333       // if no method data exists, go to profile_method
 334       __ test_method_data_pointer(*profile_method);
 335     }
 336 
 337     AddressLiteral invocation_limit((address)&InvocationCounter::InterpreterInvocationLimit);
 338     __ load_contents(invocation_limit, G3_scratch);
 339     __ cmp(O0, G3_scratch);
 340     __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
 341     __ delayed()->nop();
 342     __ bind(done);
 343   }
 344 
 345 }
 346 
 347 // Allocate monitor and lock method (asm interpreter)
 348 // ebx - Method*
 349 //
 350 void InterpreterGenerator::lock_method(void) {
 351   __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0);  // Load access flags.
 352 
 353 #ifdef ASSERT
 354  { Label ok;
 355    __ btst(JVM_ACC_SYNCHRONIZED, O0);
 356    __ br( Assembler::notZero, false, Assembler::pt, ok);
 357    __ delayed()->nop();
 358    __ stop("method doesn't need synchronization");
 359    __ bind(ok);
 360   }
 361 #endif // ASSERT
 362 
 363   // get synchronization object to O0
 364   { Label done;
 365     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
 366     __ btst(JVM_ACC_STATIC, O0);
 367     __ br( Assembler::zero, true, Assembler::pt, done);
 368     __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
 369 
 370     __ ld_ptr( Lmethod, in_bytes(Method::const_offset()), O0);
 371     __ ld_ptr( O0, in_bytes(ConstMethod::constants_offset()), O0);
 372     __ ld_ptr( O0, ConstantPool::pool_holder_offset_in_bytes(), O0);
 373 
 374     // lock the mirror, not the Klass*
 375     __ ld_ptr( O0, mirror_offset, O0);
 376 
 377 #ifdef ASSERT
 378     __ tst(O0);
 379     __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
 380 #endif // ASSERT
 381 
 382     __ bind(done);
 383   }
 384 
 385   __ add_monitor_to_stack(true, noreg, noreg);  // allocate monitor elem
 386   __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes());   // store object
 387   // __ untested("lock_object from method entry");
 388   __ lock_object(Lmonitors, O0);
 389 }
 390 
 391 
 392 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
 393                                                          Register Rscratch,
 394                                                          Register Rscratch2) {
 395   const int page_size = os::vm_page_size();
 396   Label after_frame_check;
 397 
 398   assert_different_registers(Rframe_size, Rscratch, Rscratch2);
 399 
 400   __ set(page_size, Rscratch);
 401   __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
 402 
 403   // get the stack base, and in debug, verify it is non-zero
 404   __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch );
 405 #ifdef ASSERT
 406   Label base_not_zero;
 407   __ br_notnull_short(Rscratch, Assembler::pn, base_not_zero);
 408   __ stop("stack base is zero in generate_stack_overflow_check");
 409   __ bind(base_not_zero);
 410 #endif
 411 
 412   // get the stack size, and in debug, verify it is non-zero
 413   assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" );
 414   __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 );
 415 #ifdef ASSERT
 416   Label size_not_zero;
 417   __ br_notnull_short(Rscratch2, Assembler::pn, size_not_zero);
 418   __ stop("stack size is zero in generate_stack_overflow_check");
 419   __ bind(size_not_zero);
 420 #endif
 421 
 422   // compute the beginning of the protected zone minus the requested frame size
 423   __ sub( Rscratch, Rscratch2,   Rscratch );
 424   __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 );
 425   __ add( Rscratch, Rscratch2,   Rscratch );
 426 
 427   // Add in the size of the frame (which is the same as subtracting it from the
 428   // SP, which would take another register
 429   __ add( Rscratch, Rframe_size, Rscratch );
 430 
 431   // the frame is greater than one page in size, so check against
 432   // the bottom of the stack
 433   __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check);
 434 
 435   // the stack will overflow, throw an exception
 436 
 437   // Note that SP is restored to sender's sp (in the delay slot). This
 438   // is necessary if the sender's frame is an extended compiled frame
 439   // (see gen_c2i_adapter()) and safer anyway in case of JSR292
 440   // adaptations.
 441 
 442   // Note also that the restored frame is not necessarily interpreted.
 443   // Use the shared runtime version of the StackOverflowError.
 444   assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
 445   AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
 446   __ jump_to(stub, Rscratch);
 447   __ delayed()->mov(O5_savedSP, SP);
 448 
 449   // if you get to here, then there is enough stack space
 450   __ bind( after_frame_check );
 451 }
 452 
 453 
 454 //
 455 // Generate a fixed interpreter frame. This is identical setup for interpreted
 456 // methods and for native methods hence the shared code.
 457 
 458 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
 459   //
 460   //
 461   // The entry code sets up a new interpreter frame in 4 steps:
 462   //
 463   // 1) Increase caller's SP by for the extra local space needed:
 464   //    (check for overflow)
 465   //    Efficient implementation of xload/xstore bytecodes requires
 466   //    that arguments and non-argument locals are in a contigously
 467   //    addressable memory block => non-argument locals must be
 468   //    allocated in the caller's frame.
 469   //
 470   // 2) Create a new stack frame and register window:
 471   //    The new stack frame must provide space for the standard
 472   //    register save area, the maximum java expression stack size,
 473   //    the monitor slots (0 slots initially), and some frame local
 474   //    scratch locations.
 475   //
 476   // 3) The following interpreter activation registers must be setup:
 477   //    Lesp       : expression stack pointer
 478   //    Lbcp       : bytecode pointer
 479   //    Lmethod    : method
 480   //    Llocals    : locals pointer
 481   //    Lmonitors  : monitor pointer
 482   //    LcpoolCache: constant pool cache
 483   //
 484   // 4) Initialize the non-argument locals if necessary:
 485   //    Non-argument locals may need to be initialized to NULL
 486   //    for GC to work. If the oop-map information is accurate
 487   //    (in the absence of the JSR problem), no initialization
 488   //    is necessary.
 489   //
 490   // (gri - 2/25/2000)
 491 
 492 
 493   int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
 494 
 495   const int extra_space =
 496     rounded_vm_local_words +                   // frame local scratch space
 497     Method::extra_stack_entries() +            // extra stack for jsr 292
 498     frame::memory_parameter_word_sp_offset +   // register save area
 499     (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
 500 
 501   const Register Glocals_size = G3;
 502   const Register RconstMethod = Glocals_size;
 503   const Register Otmp1 = O3;
 504   const Register Otmp2 = O4;
 505   // Lscratch can't be used as a temporary because the call_stub uses
 506   // it to assert that the stack frame was setup correctly.
 507   const Address constMethod       (G5_method, Method::const_offset());
 508   const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
 509 
 510   __ ld_ptr( constMethod, RconstMethod );
 511   __ lduh( size_of_parameters, Glocals_size);
 512 
 513   // Gargs points to first local + BytesPerWord
 514   // Set the saved SP after the register window save
 515   //
 516   assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
 517   __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
 518   __ add(Gargs, Otmp1, Gargs);
 519 
 520   if (native_call) {
 521     __ calc_mem_param_words( Glocals_size, Gframe_size );
 522     __ add( Gframe_size,  extra_space, Gframe_size);
 523     __ round_to( Gframe_size, WordsPerLong );
 524     __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
 525   } else {
 526 
 527     //
 528     // Compute number of locals in method apart from incoming parameters
 529     //
 530     const Address size_of_locals    (Otmp1, ConstMethod::size_of_locals_offset());
 531     __ ld_ptr( constMethod, Otmp1 );
 532     __ lduh( size_of_locals, Otmp1 );
 533     __ sub( Otmp1, Glocals_size, Glocals_size );
 534     __ round_to( Glocals_size, WordsPerLong );
 535     __ sll( Glocals_size, Interpreter::logStackElementSize, Glocals_size );
 536 
 537     // see if the frame is greater than one page in size. If so,
 538     // then we need to verify there is enough stack space remaining
 539     // Frame_size = (max_stack + extra_space) * BytesPerWord;
 540     __ ld_ptr( constMethod, Gframe_size );
 541     __ lduh( Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size );
 542     __ add( Gframe_size, extra_space, Gframe_size );
 543     __ round_to( Gframe_size, WordsPerLong );
 544     __ sll( Gframe_size, Interpreter::logStackElementSize, Gframe_size);
 545 
 546     // Add in java locals size for stack overflow check only
 547     __ add( Gframe_size, Glocals_size, Gframe_size );
 548 
 549     const Register Otmp2 = O4;
 550     assert_different_registers(Otmp1, Otmp2, O5_savedSP);
 551     generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2);
 552 
 553     __ sub( Gframe_size, Glocals_size, Gframe_size);
 554 
 555     //
 556     // bump SP to accomodate the extra locals
 557     //
 558     __ sub( SP, Glocals_size, SP );
 559   }
 560 
 561   //
 562   // now set up a stack frame with the size computed above
 563   //
 564   __ neg( Gframe_size );
 565   __ save( SP, Gframe_size, SP );
 566 
 567   //
 568   // now set up all the local cache registers
 569   //
 570   // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
 571   // that all present references to Lbyte_code initialize the register
 572   // immediately before use
 573   if (native_call) {
 574     __ mov(G0, Lbcp);
 575   } else {
 576     __ ld_ptr(G5_method, Method::const_offset(), Lbcp);
 577     __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp);
 578   }
 579   __ mov( G5_method, Lmethod);                 // set Lmethod
 580   __ get_constant_pool_cache( LcpoolCache );   // set LcpoolCache
 581   __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
 582 #ifdef _LP64
 583   __ add( Lmonitors, STACK_BIAS, Lmonitors );   // Account for 64 bit stack bias
 584 #endif
 585   __ sub(Lmonitors, BytesPerWord, Lesp);       // set Lesp
 586 
 587   // setup interpreter activation registers
 588   __ sub(Gargs, BytesPerWord, Llocals);        // set Llocals
 589 
 590   if (ProfileInterpreter) {
 591 #ifdef FAST_DISPATCH
 592     // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
 593     // they both use I2.
 594     assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
 595 #endif // FAST_DISPATCH
 596     __ set_method_data_pointer();
 597   }
 598 
 599 }
 600 
 601 // Empty method, generate a very fast return.
 602 
 603 address InterpreterGenerator::generate_empty_entry(void) {
 604 
 605   // A method that does nother but return...
 606 
 607   address entry = __ pc();
 608   Label slow_path;
 609 
 610   // do nothing for empty methods (do not even increment invocation counter)
 611   if ( UseFastEmptyMethods) {
 612     // If we need a safepoint check, generate full interpreter entry.
 613     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
 614     __ set(sync_state, G3_scratch);
 615     __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path);
 616 
 617     // Code: _return
 618     __ retl();
 619     __ delayed()->mov(O5_savedSP, SP);
 620 
 621     __ bind(slow_path);
 622     (void) generate_normal_entry(false);
 623 
 624     return entry;
 625   }
 626   return NULL;
 627 }
 628 
 629 // Call an accessor method (assuming it is resolved, otherwise drop into
 630 // vanilla (slow path) entry
 631 
 632 // Generates code to elide accessor methods
 633 // Uses G3_scratch and G1_scratch as scratch
 634 address InterpreterGenerator::generate_accessor_entry(void) {
 635 
 636   // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
 637   // parameter size = 1
 638   // Note: We can only use this code if the getfield has been resolved
 639   //       and if we don't have a null-pointer exception => check for
 640   //       these conditions first and use slow path if necessary.
 641   address entry = __ pc();
 642   Label slow_path;
 643 
 644 
 645   // XXX: for compressed oops pointer loading and decoding doesn't fit in
 646   // delay slot and damages G1
 647   if ( UseFastAccessorMethods && !UseCompressedOops ) {
 648     // Check if we need to reach a safepoint and generate full interpreter
 649     // frame if so.
 650     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
 651     __ load_contents(sync_state, G3_scratch);
 652     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
 653     __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path);
 654 
 655     // Check if local 0 != NULL
 656     __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
 657     // check if local 0 == NULL and go the slow path
 658     __ br_null_short(Otos_i, Assembler::pn, slow_path);
 659 
 660 
 661     // read first instruction word and extract bytecode @ 1 and index @ 2
 662     // get first 4 bytes of the bytecodes (big endian!)
 663     __ ld_ptr(G5_method, Method::const_offset(), G1_scratch);
 664     __ ld(G1_scratch, ConstMethod::codes_offset(), G1_scratch);
 665 
 666     // move index @ 2 far left then to the right most two bytes.
 667     __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
 668     __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
 669                       ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
 670 
 671     // get constant pool cache
 672     __ ld_ptr(G5_method, Method::const_offset(), G3_scratch);
 673     __ ld_ptr(G3_scratch, ConstMethod::constants_offset(), G3_scratch);
 674     __ ld_ptr(G3_scratch, ConstantPool::cache_offset_in_bytes(), G3_scratch);
 675 
 676     // get specific constant pool cache entry
 677     __ add(G3_scratch, G1_scratch, G3_scratch);
 678 
 679     // Check the constant Pool cache entry to see if it has been resolved.
 680     // If not, need the slow path.
 681     ByteSize cp_base_offset = ConstantPoolCache::base_offset();
 682     __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::indices_offset(), G1_scratch);
 683     __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
 684     __ and3(G1_scratch, 0xFF, G1_scratch);
 685     __ cmp_and_br_short(G1_scratch, Bytecodes::_getfield, Assembler::notEqual, Assembler::pn, slow_path);
 686 
 687     // Get the type and return field offset from the constant pool cache
 688     __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::flags_offset(), G1_scratch);
 689     __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), G3_scratch);
 690 
 691     Label xreturn_path;
 692     // Need to differentiate between igetfield, agetfield, bgetfield etc.
 693     // because they are different sizes.
 694     // Get the type from the constant pool cache
 695     __ srl(G1_scratch, ConstantPoolCacheEntry::tos_state_shift, G1_scratch);
 696     // Make sure we don't need to mask G1_scratch after the above shift
 697     ConstantPoolCacheEntry::verify_tos_state_shift();
 698     __ cmp(G1_scratch, atos );
 699     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 700     __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
 701     __ cmp(G1_scratch, itos);
 702     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 703     __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
 704     __ cmp(G1_scratch, stos);
 705     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 706     __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
 707     __ cmp(G1_scratch, ctos);
 708     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 709     __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
 710 #ifdef ASSERT
 711     __ cmp(G1_scratch, btos);
 712     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 713     __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
 714     __ cmp(G1_scratch, ztos);
 715     __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
 716     __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
 717     __ should_not_reach_here();
 718 #endif
 719     __ ldsb(Otos_i, G3_scratch, Otos_i);
 720     __ bind(xreturn_path);
 721 
 722     // _ireturn/_areturn
 723     __ retl();                      // return from leaf routine
 724     __ delayed()->mov(O5_savedSP, SP);
 725 
 726     // Generate regular method entry
 727     __ bind(slow_path);
 728     (void) generate_normal_entry(false);
 729     return entry;
 730   }
 731   return NULL;
 732 }
 733 
 734 // Method entry for java.lang.ref.Reference.get.
 735 address InterpreterGenerator::generate_Reference_get_entry(void) {
 736 #if INCLUDE_ALL_GCS
 737   // Code: _aload_0, _getfield, _areturn
 738   // parameter size = 1
 739   //
 740   // The code that gets generated by this routine is split into 2 parts:
 741   //    1. The "intrinsified" code for G1 (or any SATB based GC),
 742   //    2. The slow path - which is an expansion of the regular method entry.
 743   //
 744   // Notes:-
 745   // * In the G1 code we do not check whether we need to block for
 746   //   a safepoint. If G1 is enabled then we must execute the specialized
 747   //   code for Reference.get (except when the Reference object is null)
 748   //   so that we can log the value in the referent field with an SATB
 749   //   update buffer.
 750   //   If the code for the getfield template is modified so that the
 751   //   G1 pre-barrier code is executed when the current method is
 752   //   Reference.get() then going through the normal method entry
 753   //   will be fine.
 754   // * The G1 code can, however, check the receiver object (the instance
 755   //   of java.lang.Reference) and jump to the slow path if null. If the
 756   //   Reference object is null then we obviously cannot fetch the referent
 757   //   and so we don't need to call the G1 pre-barrier. Thus we can use the
 758   //   regular method entry code to generate the NPE.
 759   //
 760   // This code is based on generate_accessor_enty.
 761 
 762   address entry = __ pc();
 763 
 764   const int referent_offset = java_lang_ref_Reference::referent_offset;
 765   guarantee(referent_offset > 0, "referent offset not initialized");
 766 
 767   if (UseG1GC) {
 768      Label slow_path;
 769 
 770     // In the G1 code we don't check if we need to reach a safepoint. We
 771     // continue and the thread will safepoint at the next bytecode dispatch.
 772 
 773     // Check if local 0 != NULL
 774     // If the receiver is null then it is OK to jump to the slow path.
 775     __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
 776     // check if local 0 == NULL and go the slow path
 777     __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
 778 
 779 
 780     // Load the value of the referent field.
 781     if (Assembler::is_simm13(referent_offset)) {
 782       __ load_heap_oop(Otos_i, referent_offset, Otos_i);
 783     } else {
 784       __ set(referent_offset, G3_scratch);
 785       __ load_heap_oop(Otos_i, G3_scratch, Otos_i);
 786     }
 787 
 788     // Generate the G1 pre-barrier code to log the value of
 789     // the referent field in an SATB buffer. Note with
 790     // these parameters the pre-barrier does not generate
 791     // the load of the previous value
 792 
 793     __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */,
 794                             Otos_i /* pre_val */,
 795                             G3_scratch /* tmp */,
 796                             true /* preserve_o_regs */);
 797 
 798     // _areturn
 799     __ retl();                      // return from leaf routine
 800     __ delayed()->mov(O5_savedSP, SP);
 801 
 802     // Generate regular method entry
 803     __ bind(slow_path);
 804     (void) generate_normal_entry(false);
 805     return entry;
 806   }
 807 #endif // INCLUDE_ALL_GCS
 808 
 809   // If G1 is not enabled then attempt to go through the accessor entry point
 810   // Reference.get is an accessor
 811   return generate_accessor_entry();
 812 }
 813 
 814 //
 815 // Interpreter stub for calling a native method. (asm interpreter)
 816 // This sets up a somewhat different looking stack for calling the native method
 817 // than the typical interpreter frame setup.
 818 //
 819 
 820 address InterpreterGenerator::generate_native_entry(bool synchronized) {
 821   address entry = __ pc();
 822 
 823   // the following temporary registers are used during frame creation
 824   const Register Gtmp1 = G3_scratch ;
 825   const Register Gtmp2 = G1_scratch;
 826   bool inc_counter  = UseCompiler || CountCompiledCalls;
 827 
 828   // make sure registers are different!
 829   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
 830 
 831   const Address Laccess_flags(Lmethod, Method::access_flags_offset());
 832 
 833   const Register Glocals_size = G3;
 834   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
 835 
 836   // make sure method is native & not abstract
 837   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
 838 #ifdef ASSERT
 839   __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
 840   {
 841     Label L;
 842     __ btst(JVM_ACC_NATIVE, Gtmp1);
 843     __ br(Assembler::notZero, false, Assembler::pt, L);
 844     __ delayed()->nop();
 845     __ stop("tried to execute non-native method as native");
 846     __ bind(L);
 847   }
 848   { Label L;
 849     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
 850     __ br(Assembler::zero, false, Assembler::pt, L);
 851     __ delayed()->nop();
 852     __ stop("tried to execute abstract method as non-abstract");
 853     __ bind(L);
 854   }
 855 #endif // ASSERT
 856 
 857  // generate the code to allocate the interpreter stack frame
 858   generate_fixed_frame(true);
 859 
 860   //
 861   // No locals to initialize for native method
 862   //
 863 
 864   // this slot will be set later, we initialize it to null here just in
 865   // case we get a GC before the actual value is stored later
 866   __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
 867 
 868   const Address do_not_unlock_if_synchronized(G2_thread,
 869     JavaThread::do_not_unlock_if_synchronized_offset());
 870   // Since at this point in the method invocation the exception handler
 871   // would try to exit the monitor of synchronized methods which hasn't
 872   // been entered yet, we set the thread local variable
 873   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
 874   // runtime, exception handling i.e. unlock_if_synchronized_method will
 875   // check this thread local flag.
 876   // This flag has two effects, one is to force an unwind in the topmost
 877   // interpreter frame and not perform an unlock while doing so.
 878 
 879   __ movbool(true, G3_scratch);
 880   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
 881 
 882   // increment invocation counter and check for overflow
 883   //
 884   // Note: checking for negative value instead of overflow
 885   //       so we have a 'sticky' overflow test (may be of
 886   //       importance as soon as we have true MT/MP)
 887   Label invocation_counter_overflow;
 888   Label Lcontinue;
 889   if (inc_counter) {
 890     generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
 891 
 892   }
 893   __ bind(Lcontinue);
 894 
 895   bang_stack_shadow_pages(true);
 896 
 897   // reset the _do_not_unlock_if_synchronized flag
 898   __ stbool(G0, do_not_unlock_if_synchronized);
 899 
 900   // check for synchronized methods
 901   // Must happen AFTER invocation_counter check and stack overflow check,
 902   // so method is not locked if overflows.
 903 
 904   if (synchronized) {
 905     lock_method();
 906   } else {
 907 #ifdef ASSERT
 908     { Label ok;
 909       __ ld(Laccess_flags, O0);
 910       __ btst(JVM_ACC_SYNCHRONIZED, O0);
 911       __ br( Assembler::zero, false, Assembler::pt, ok);
 912       __ delayed()->nop();
 913       __ stop("method needs synchronization");
 914       __ bind(ok);
 915     }
 916 #endif // ASSERT
 917   }
 918 
 919 
 920   // start execution
 921   __ verify_thread();
 922 
 923   // JVMTI support
 924   __ notify_method_entry();
 925 
 926   // native call
 927 
 928   // (note that O0 is never an oop--at most it is a handle)
 929   // It is important not to smash any handles created by this call,
 930   // until any oop handle in O0 is dereferenced.
 931 
 932   // (note that the space for outgoing params is preallocated)
 933 
 934   // get signature handler
 935   { Label L;
 936     Address signature_handler(Lmethod, Method::signature_handler_offset());
 937     __ ld_ptr(signature_handler, G3_scratch);
 938     __ br_notnull_short(G3_scratch, Assembler::pt, L);
 939     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
 940     __ ld_ptr(signature_handler, G3_scratch);
 941     __ bind(L);
 942   }
 943 
 944   // Push a new frame so that the args will really be stored in
 945   // Copy a few locals across so the new frame has the variables
 946   // we need but these values will be dead at the jni call and
 947   // therefore not gc volatile like the values in the current
 948   // frame (Lmethod in particular)
 949 
 950   // Flush the method pointer to the register save area
 951   __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
 952   __ mov(Llocals, O1);
 953 
 954   // calculate where the mirror handle body is allocated in the interpreter frame:
 955   __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
 956 
 957   // Calculate current frame size
 958   __ sub(SP, FP, O3);         // Calculate negative of current frame size
 959   __ save(SP, O3, SP);        // Allocate an identical sized frame
 960 
 961   // Note I7 has leftover trash. Slow signature handler will fill it in
 962   // should we get there. Normal jni call will set reasonable last_Java_pc
 963   // below (and fix I7 so the stack trace doesn't have a meaningless frame
 964   // in it).
 965 
 966   // Load interpreter frame's Lmethod into same register here
 967 
 968   __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
 969 
 970   __ mov(I1, Llocals);
 971   __ mov(I2, Lscratch2);     // save the address of the mirror
 972 
 973 
 974   // ONLY Lmethod and Llocals are valid here!
 975 
 976   // call signature handler, It will move the arg properly since Llocals in current frame
 977   // matches that in outer frame
 978 
 979   __ callr(G3_scratch, 0);
 980   __ delayed()->nop();
 981 
 982   // Result handler is in Lscratch
 983 
 984   // Reload interpreter frame's Lmethod since slow signature handler may block
 985   __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
 986 
 987   { Label not_static;
 988 
 989     __ ld(Laccess_flags, O0);
 990     __ btst(JVM_ACC_STATIC, O0);
 991     __ br( Assembler::zero, false, Assembler::pt, not_static);
 992     // get native function entry point(O0 is a good temp until the very end)
 993     __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0);
 994     // for static methods insert the mirror argument
 995     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
 996 
 997     __ ld_ptr(Lmethod, Method:: const_offset(), O1);
 998     __ ld_ptr(O1, ConstMethod::constants_offset(), O1);
 999     __ ld_ptr(O1, ConstantPool::pool_holder_offset_in_bytes(), O1);
1000     __ ld_ptr(O1, mirror_offset, O1);
1001 #ifdef ASSERT
1002     if (!PrintSignatureHandlers)  // do not dirty the output with this
1003     { Label L;
1004       __ br_notnull_short(O1, Assembler::pt, L);
1005       __ stop("mirror is missing");
1006       __ bind(L);
1007     }
1008 #endif // ASSERT
1009     __ st_ptr(O1, Lscratch2, 0);
1010     __ mov(Lscratch2, O1);
1011     __ bind(not_static);
1012   }
1013 
1014   // At this point, arguments have been copied off of stack into
1015   // their JNI positions, which are O1..O5 and SP[68..].
1016   // Oops are boxed in-place on the stack, with handles copied to arguments.
1017   // The result handler is in Lscratch.  O0 will shortly hold the JNIEnv*.
1018 
1019 #ifdef ASSERT
1020   { Label L;
1021     __ br_notnull_short(O0, Assembler::pt, L);
1022     __ stop("native entry point is missing");
1023     __ bind(L);
1024   }
1025 #endif // ASSERT
1026 
1027   //
1028   // setup the frame anchor
1029   //
1030   // The scavenge function only needs to know that the PC of this frame is
1031   // in the interpreter method entry code, it doesn't need to know the exact
1032   // PC and hence we can use O7 which points to the return address from the
1033   // previous call in the code stream (signature handler function)
1034   //
1035   // The other trick is we set last_Java_sp to FP instead of the usual SP because
1036   // we have pushed the extra frame in order to protect the volatile register(s)
1037   // in that frame when we return from the jni call
1038   //
1039 
1040   __ set_last_Java_frame(FP, O7);
1041   __ mov(O7, I7);  // make dummy interpreter frame look like one above,
1042                    // not meaningless information that'll confuse me.
1043 
1044   // flush the windows now. We don't care about the current (protection) frame
1045   // only the outer frames
1046 
1047   __ flushw();
1048 
1049   // mark windows as flushed
1050   Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1051   __ set(JavaFrameAnchor::flushed, G3_scratch);
1052   __ st(G3_scratch, flags);
1053 
1054   // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1055 
1056   Address thread_state(G2_thread, JavaThread::thread_state_offset());
1057 #ifdef ASSERT
1058   { Label L;
1059     __ ld(thread_state, G3_scratch);
1060     __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1061     __ stop("Wrong thread state in native stub");
1062     __ bind(L);
1063   }
1064 #endif // ASSERT
1065   __ set(_thread_in_native, G3_scratch);
1066   __ st(G3_scratch, thread_state);
1067 
1068   // Call the jni method, using the delay slot to set the JNIEnv* argument.
1069   __ save_thread(L7_thread_cache); // save Gthread
1070   __ callr(O0, 0);
1071   __ delayed()->
1072      add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1073 
1074   // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1075 
1076   __ restore_thread(L7_thread_cache); // restore G2_thread
1077   __ reinit_heapbase();
1078 
1079   // must we block?
1080 
1081   // Block, if necessary, before resuming in _thread_in_Java state.
1082   // In order for GC to work, don't clear the last_Java_sp until after blocking.
1083   { Label no_block;
1084     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1085 
1086     // Switch thread to "native transition" state before reading the synchronization state.
1087     // This additional state is necessary because reading and testing the synchronization
1088     // state is not atomic w.r.t. GC, as this scenario demonstrates:
1089     //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1090     //     VM thread changes sync state to synchronizing and suspends threads for GC.
1091     //     Thread A is resumed to finish this native method, but doesn't block here since it
1092     //     didn't see any synchronization is progress, and escapes.
1093     __ set(_thread_in_native_trans, G3_scratch);
1094     __ st(G3_scratch, thread_state);
1095     if(os::is_MP()) {
1096       if (UseMembar) {
1097         // Force this write out before the read below
1098         __ membar(Assembler::StoreLoad);
1099       } else {
1100         // Write serialization page so VM thread can do a pseudo remote membar.
1101         // We use the current thread pointer to calculate a thread specific
1102         // offset to write to within the page. This minimizes bus traffic
1103         // due to cache line collision.
1104         __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1105       }
1106     }
1107     __ load_contents(sync_state, G3_scratch);
1108     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1109 
1110     Label L;
1111     __ br(Assembler::notEqual, false, Assembler::pn, L);
1112     __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1113     __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1114     __ bind(L);
1115 
1116     // Block.  Save any potential method result value before the operation and
1117     // use a leaf call to leave the last_Java_frame setup undisturbed.
1118     save_native_result();
1119     __ call_VM_leaf(L7_thread_cache,
1120                     CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1121                     G2_thread);
1122 
1123     // Restore any method result value
1124     restore_native_result();
1125     __ bind(no_block);
1126   }
1127 
1128   // Clear the frame anchor now
1129 
1130   __ reset_last_Java_frame();
1131 
1132   // Move the result handler address
1133   __ mov(Lscratch, G3_scratch);
1134   // return possible result to the outer frame
1135 #ifndef __LP64
1136   __ mov(O0, I0);
1137   __ restore(O1, G0, O1);
1138 #else
1139   __ restore(O0, G0, O0);
1140 #endif /* __LP64 */
1141 
1142   // Move result handler to expected register
1143   __ mov(G3_scratch, Lscratch);
1144 
1145   // Back in normal (native) interpreter frame. State is thread_in_native_trans
1146   // switch to thread_in_Java.
1147 
1148   __ set(_thread_in_Java, G3_scratch);
1149   __ st(G3_scratch, thread_state);
1150 
1151   // reset handle block
1152   __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1153   __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1154 
1155   // If we have an oop result store it where it will be safe for any further gc
1156   // until we return now that we've released the handle it might be protected by
1157 
1158   {
1159     Label no_oop, store_result;
1160 
1161     __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1162     __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1163     __ addcc(G0, O0, O0);
1164     __ brx(Assembler::notZero, true, Assembler::pt, store_result);     // if result is not NULL:
1165     __ delayed()->ld_ptr(O0, 0, O0);                                   // unbox it
1166     __ mov(G0, O0);
1167 
1168     __ bind(store_result);
1169     // Store it where gc will look for it and result handler expects it.
1170     __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1171 
1172     __ bind(no_oop);
1173 
1174   }
1175 
1176 
1177   // handle exceptions (exception handling will handle unlocking!)
1178   { Label L;
1179     Address exception_addr(G2_thread, Thread::pending_exception_offset());
1180     __ ld_ptr(exception_addr, Gtemp);
1181     __ br_null_short(Gtemp, Assembler::pt, L);
1182     // Note: This could be handled more efficiently since we know that the native
1183     //       method doesn't have an exception handler. We could directly return
1184     //       to the exception handler for the caller.
1185     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1186     __ should_not_reach_here();
1187     __ bind(L);
1188   }
1189 
1190   // JVMTI support (preserves thread register)
1191   __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1192 
1193   if (synchronized) {
1194     // save and restore any potential method result value around the unlocking operation
1195     save_native_result();
1196 
1197     __ add( __ top_most_monitor(), O1);
1198     __ unlock_object(O1);
1199 
1200     restore_native_result();
1201   }
1202 
1203 #if defined(COMPILER2) && !defined(_LP64)
1204 
1205   // C2 expects long results in G1 we can't tell if we're returning to interpreted
1206   // or compiled so just be safe.
1207 
1208   __ sllx(O0, 32, G1);          // Shift bits into high G1
1209   __ srl (O1, 0, O1);           // Zero extend O1
1210   __ or3 (O1, G1, G1);          // OR 64 bits into G1
1211 
1212 #endif /* COMPILER2 && !_LP64 */
1213 
1214   // dispose of return address and remove activation
1215 #ifdef ASSERT
1216   {
1217     Label ok;
1218     __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1219     __ stop("bad I5_savedSP value");
1220     __ should_not_reach_here();
1221     __ bind(ok);
1222   }
1223 #endif
1224   if (TraceJumps) {
1225     // Move target to register that is recordable
1226     __ mov(Lscratch, G3_scratch);
1227     __ JMP(G3_scratch, 0);
1228   } else {
1229     __ jmp(Lscratch, 0);
1230   }
1231   __ delayed()->nop();
1232 
1233 
1234   if (inc_counter) {
1235     // handle invocation counter overflow
1236     __ bind(invocation_counter_overflow);
1237     generate_counter_overflow(Lcontinue);
1238   }
1239 
1240 
1241 
1242   return entry;
1243 }
1244 
1245 
1246 // Generic method entry to (asm) interpreter
1247 //------------------------------------------------------------------------------------------------------------------------
1248 //
1249 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1250   address entry = __ pc();
1251 
1252   bool inc_counter  = UseCompiler || CountCompiledCalls;
1253 
1254   // the following temporary registers are used during frame creation
1255   const Register Gtmp1 = G3_scratch ;
1256   const Register Gtmp2 = G1_scratch;
1257 
1258   // make sure registers are different!
1259   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1260 
1261   const Address constMethod       (G5_method, Method::const_offset());
1262   // Seems like G5_method is live at the point this is used. So we could make this look consistent
1263   // and use in the asserts.
1264   const Address access_flags      (Lmethod,   Method::access_flags_offset());
1265 
1266   const Register Glocals_size = G3;
1267   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1268 
1269   // make sure method is not native & not abstract
1270   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1271 #ifdef ASSERT
1272   __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1273   {
1274     Label L;
1275     __ btst(JVM_ACC_NATIVE, Gtmp1);
1276     __ br(Assembler::zero, false, Assembler::pt, L);
1277     __ delayed()->nop();
1278     __ stop("tried to execute native method as non-native");
1279     __ bind(L);
1280   }
1281   { Label L;
1282     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1283     __ br(Assembler::zero, false, Assembler::pt, L);
1284     __ delayed()->nop();
1285     __ stop("tried to execute abstract method as non-abstract");
1286     __ bind(L);
1287   }
1288 #endif // ASSERT
1289 
1290   // generate the code to allocate the interpreter stack frame
1291 
1292   generate_fixed_frame(false);
1293 
1294 #ifdef FAST_DISPATCH
1295   __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1296                                           // set bytecode dispatch table base
1297 #endif
1298 
1299   //
1300   // Code to initialize the extra (i.e. non-parm) locals
1301   //
1302   Register init_value = noreg;    // will be G0 if we must clear locals
1303   // The way the code was setup before zerolocals was always true for vanilla java entries.
1304   // It could only be false for the specialized entries like accessor or empty which have
1305   // no extra locals so the testing was a waste of time and the extra locals were always
1306   // initialized. We removed this extra complication to already over complicated code.
1307 
1308   init_value = G0;
1309   Label clear_loop;
1310 
1311   const Register RconstMethod = O1;
1312   const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1313   const Address size_of_locals    (RconstMethod, ConstMethod::size_of_locals_offset());
1314 
1315   // NOTE: If you change the frame layout, this code will need to
1316   // be updated!
1317   __ ld_ptr( constMethod, RconstMethod );
1318   __ lduh( size_of_locals, O2 );
1319   __ lduh( size_of_parameters, O1 );
1320   __ sll( O2, Interpreter::logStackElementSize, O2);
1321   __ sll( O1, Interpreter::logStackElementSize, O1 );
1322   __ sub( Llocals, O2, O2 );
1323   __ sub( Llocals, O1, O1 );
1324 
1325   __ bind( clear_loop );
1326   __ inc( O2, wordSize );
1327 
1328   __ cmp( O2, O1 );
1329   __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1330   __ delayed()->st_ptr( init_value, O2, 0 );
1331 
1332   const Address do_not_unlock_if_synchronized(G2_thread,
1333     JavaThread::do_not_unlock_if_synchronized_offset());
1334   // Since at this point in the method invocation the exception handler
1335   // would try to exit the monitor of synchronized methods which hasn't
1336   // been entered yet, we set the thread local variable
1337   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1338   // runtime, exception handling i.e. unlock_if_synchronized_method will
1339   // check this thread local flag.
1340   __ movbool(true, G3_scratch);
1341   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1342 
1343   __ profile_parameters_type(G1_scratch, G3_scratch, G4_scratch, Lscratch);
1344   // increment invocation counter and check for overflow
1345   //
1346   // Note: checking for negative value instead of overflow
1347   //       so we have a 'sticky' overflow test (may be of
1348   //       importance as soon as we have true MT/MP)
1349   Label invocation_counter_overflow;
1350   Label profile_method;
1351   Label profile_method_continue;
1352   Label Lcontinue;
1353   if (inc_counter) {
1354     generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1355     if (ProfileInterpreter) {
1356       __ bind(profile_method_continue);
1357     }
1358   }
1359   __ bind(Lcontinue);
1360 
1361   bang_stack_shadow_pages(false);
1362 
1363   // reset the _do_not_unlock_if_synchronized flag
1364   __ stbool(G0, do_not_unlock_if_synchronized);
1365 
1366   // check for synchronized methods
1367   // Must happen AFTER invocation_counter check and stack overflow check,
1368   // so method is not locked if overflows.
1369 
1370   if (synchronized) {
1371     lock_method();
1372   } else {
1373 #ifdef ASSERT
1374     { Label ok;
1375       __ ld(access_flags, O0);
1376       __ btst(JVM_ACC_SYNCHRONIZED, O0);
1377       __ br( Assembler::zero, false, Assembler::pt, ok);
1378       __ delayed()->nop();
1379       __ stop("method needs synchronization");
1380       __ bind(ok);
1381     }
1382 #endif // ASSERT
1383   }
1384 
1385   // start execution
1386 
1387   __ verify_thread();
1388 
1389   // jvmti support
1390   __ notify_method_entry();
1391 
1392   // start executing instructions
1393   __ dispatch_next(vtos);
1394 
1395 
1396   if (inc_counter) {
1397     if (ProfileInterpreter) {
1398       // We have decided to profile this method in the interpreter
1399       __ bind(profile_method);
1400 
1401       __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1402       __ set_method_data_pointer_for_bcp();
1403       __ ba_short(profile_method_continue);
1404     }
1405 
1406     // handle invocation counter overflow
1407     __ bind(invocation_counter_overflow);
1408     generate_counter_overflow(Lcontinue);
1409   }
1410 
1411 
1412   return entry;
1413 }
1414 
1415 
1416 //----------------------------------------------------------------------------------------------------
1417 // Entry points & stack frame layout
1418 //
1419 // Here we generate the various kind of entries into the interpreter.
1420 // The two main entry type are generic bytecode methods and native call method.
1421 // These both come in synchronized and non-synchronized versions but the
1422 // frame layout they create is very similar. The other method entry
1423 // types are really just special purpose entries that are really entry
1424 // and interpretation all in one. These are for trivial methods like
1425 // accessor, empty, or special math methods.
1426 //
1427 // When control flow reaches any of the entry types for the interpreter
1428 // the following holds ->
1429 //
1430 // C2 Calling Conventions:
1431 //
1432 // The entry code below assumes that the following registers are set
1433 // when coming in:
1434 //    G5_method: holds the Method* of the method to call
1435 //    Lesp:    points to the TOS of the callers expression stack
1436 //             after having pushed all the parameters
1437 //
1438 // The entry code does the following to setup an interpreter frame
1439 //   pop parameters from the callers stack by adjusting Lesp
1440 //   set O0 to Lesp
1441 //   compute X = (max_locals - num_parameters)
1442 //   bump SP up by X to accomadate the extra locals
1443 //   compute X = max_expression_stack
1444 //               + vm_local_words
1445 //               + 16 words of register save area
1446 //   save frame doing a save sp, -X, sp growing towards lower addresses
1447 //   set Lbcp, Lmethod, LcpoolCache
1448 //   set Llocals to i0
1449 //   set Lmonitors to FP - rounded_vm_local_words
1450 //   set Lesp to Lmonitors - 4
1451 //
1452 //  The frame has now been setup to do the rest of the entry code
1453 
1454 // Try this optimization:  Most method entries could live in a
1455 // "one size fits all" stack frame without all the dynamic size
1456 // calculations.  It might be profitable to do all this calculation
1457 // statically and approximately for "small enough" methods.
1458 
1459 //-----------------------------------------------------------------------------------------------
1460 
1461 // C1 Calling conventions
1462 //
1463 // Upon method entry, the following registers are setup:
1464 //
1465 // g2 G2_thread: current thread
1466 // g5 G5_method: method to activate
1467 // g4 Gargs  : pointer to last argument
1468 //
1469 //
1470 // Stack:
1471 //
1472 // +---------------+ <--- sp
1473 // |               |
1474 // : reg save area :
1475 // |               |
1476 // +---------------+ <--- sp + 0x40
1477 // |               |
1478 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
1479 // |               |
1480 // +---------------+ <--- sp + 0x5c
1481 // |               |
1482 // :     free      :
1483 // |               |
1484 // +---------------+ <--- Gargs
1485 // |               |
1486 // :   arguments   :
1487 // |               |
1488 // +---------------+
1489 // |               |
1490 //
1491 //
1492 //
1493 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
1494 //
1495 // +---------------+ <--- sp
1496 // |               |
1497 // : reg save area :
1498 // |               |
1499 // +---------------+ <--- sp + 0x40
1500 // |               |
1501 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
1502 // |               |
1503 // +---------------+ <--- sp + 0x5c
1504 // |               |
1505 // :               :
1506 // |               | <--- Lesp
1507 // +---------------+ <--- Lmonitors (fp - 0x18)
1508 // |   VM locals   |
1509 // +---------------+ <--- fp
1510 // |               |
1511 // : reg save area :
1512 // |               |
1513 // +---------------+ <--- fp + 0x40
1514 // |               |
1515 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
1516 // |               |
1517 // +---------------+ <--- fp + 0x5c
1518 // |               |
1519 // :     free      :
1520 // |               |
1521 // +---------------+
1522 // |               |
1523 // : nonarg locals :
1524 // |               |
1525 // +---------------+
1526 // |               |
1527 // :   arguments   :
1528 // |               | <--- Llocals
1529 // +---------------+ <--- Gargs
1530 // |               |
1531 
1532 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
1533 
1534   // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
1535   // expression stack, the callee will have callee_extra_locals (so we can account for
1536   // frame extension) and monitor_size for monitors. Basically we need to calculate
1537   // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
1538   //
1539   //
1540   // The big complicating thing here is that we must ensure that the stack stays properly
1541   // aligned. This would be even uglier if monitor size wasn't modulo what the stack
1542   // needs to be aligned for). We are given that the sp (fp) is already aligned by
1543   // the caller so we must ensure that it is properly aligned for our callee.
1544   //
1545   const int rounded_vm_local_words =
1546        round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1547   // callee_locals and max_stack are counts, not the size in frame.
1548   const int locals_size =
1549        round_to(callee_extra_locals * Interpreter::stackElementWords, WordsPerLong);
1550   const int max_stack_words = max_stack * Interpreter::stackElementWords;
1551   return (round_to((max_stack_words
1552                    + rounded_vm_local_words
1553                    + frame::memory_parameter_word_sp_offset), WordsPerLong)
1554                    // already rounded
1555                    + locals_size + monitor_size);
1556 }
1557 
1558 // How much stack a method top interpreter activation needs in words.
1559 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
1560 
1561   // See call_stub code
1562   int call_stub_size  = round_to(7 + frame::memory_parameter_word_sp_offset,
1563                                  WordsPerLong);    // 7 + register save area
1564 
1565   // Save space for one monitor to get into the interpreted method in case
1566   // the method is synchronized
1567   int monitor_size    = method->is_synchronized() ?
1568                                 1*frame::interpreter_frame_monitor_size() : 0;
1569   return size_activation_helper(method->max_locals(), method->max_stack(),
1570                                 monitor_size) + call_stub_size;
1571 }
1572 
1573 int AbstractInterpreter::size_activation(int max_stack,
1574                                          int temps,
1575                                          int extra_args,
1576                                          int monitors,
1577                                          int callee_params,
1578                                          int callee_locals,
1579                                          bool is_top_frame) {
1580   // Note: This calculation must exactly parallel the frame setup
1581   // in InterpreterGenerator::generate_fixed_frame.
1582 
1583   int monitor_size           = monitors * frame::interpreter_frame_monitor_size();
1584 
1585   assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1586 
1587   //
1588   // Note: if you look closely this appears to be doing something much different
1589   // than generate_fixed_frame. What is happening is this. On sparc we have to do
1590   // this dance with interpreter_sp_adjustment because the window save area would
1591   // appear just below the bottom (tos) of the caller's java expression stack. Because
1592   // the interpreter want to have the locals completely contiguous generate_fixed_frame
1593   // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size).
1594   // Now in generate_fixed_frame the extension of the caller's sp happens in the callee.
1595   // In this code the opposite occurs the caller adjusts it's own stack base on the callee.
1596   // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest)
1597   // because the oldest frame would have adjust its callers frame and yet that frame
1598   // already exists and isn't part of this array of frames we are unpacking. So at first
1599   // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper()
1600   // will after it calculates all of the frame's on_stack_size()'s will then figure out the
1601   // amount to adjust the caller of the initial (oldest) frame and the calculation will all
1602   // add up. It does seem like it simpler to account for the adjustment here (and remove the
1603   // callee... parameters here). However this would mean that this routine would have to take
1604   // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment)
1605   // and run the calling loop in the reverse order. This would also would appear to mean making
1606   // this code aware of what the interactions are when that initial caller fram was an osr or
1607   // other adapter frame. deoptimization is complicated enough and  hard enough to debug that
1608   // there is no sense in messing working code.
1609   //
1610 
1611   int rounded_cls = round_to((callee_locals - callee_params), WordsPerLong);
1612   assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align");
1613 
1614   int raw_frame_size = size_activation_helper(rounded_cls, max_stack, monitor_size);
1615 
1616   return raw_frame_size;
1617 }
1618 
1619 void AbstractInterpreter::layout_activation(Method* method,
1620                                             int tempcount,
1621                                             int popframe_extra_args,
1622                                             int moncount,
1623                                             int caller_actual_parameters,
1624                                             int callee_param_count,
1625                                             int callee_local_count,
1626                                             frame* caller,
1627                                             frame* interpreter_frame,
1628                                             bool is_top_frame,
1629                                             bool is_bottom_frame) {
1630   // Set up the following variables:
1631   //   - Lmethod
1632   //   - Llocals
1633   //   - Lmonitors (to the indicated number of monitors)
1634   //   - Lesp (to the indicated number of temps)
1635   // The frame caller on entry is a description of the caller of the
1636   // frame we are about to layout. We are guaranteed that we will be
1637   // able to fill in a new interpreter frame as its callee (i.e. the
1638   // stack space is allocated and the amount was determined by an
1639   // earlier call to the size_activation() method).  On return caller
1640   // while describe the interpreter frame we just layed out.
1641 
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   int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1647   int monitor_size           = moncount * frame::interpreter_frame_monitor_size();
1648   assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1649 
1650   intptr_t* fp = interpreter_frame->fp();
1651 
1652   JavaThread* thread = JavaThread::current();
1653   RegisterMap map(thread, false);
1654   // More verification that skeleton frame is properly walkable
1655   assert(fp == caller->sp(), "fp must match");
1656 
1657   intptr_t* montop     = fp - rounded_vm_local_words;
1658 
1659   // preallocate monitors (cf. __ add_monitor_to_stack)
1660   intptr_t* monitors = montop - monitor_size;
1661 
1662   // preallocate stack space
1663   intptr_t*  esp = monitors - 1 -
1664     (tempcount * Interpreter::stackElementWords) -
1665     popframe_extra_args;
1666 
1667   int local_words = method->max_locals() * Interpreter::stackElementWords;
1668   NEEDS_CLEANUP;
1669   intptr_t* locals;
1670   if (caller->is_interpreted_frame()) {
1671     // Can force the locals area to end up properly overlapping the top of the expression stack.
1672     intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1;
1673     // Note that this computation means we replace size_of_parameters() values from the caller
1674     // interpreter frame's expression stack with our argument locals
1675     int parm_words  = caller_actual_parameters * Interpreter::stackElementWords;
1676     locals = Lesp_ptr + parm_words;
1677     int delta = local_words - parm_words;
1678     int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0;
1679     *interpreter_frame->register_addr(I5_savedSP)    = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS;
1680     if (!is_bottom_frame) {
1681       // Llast_SP is set below for the current frame to SP (with the
1682       // extra space for the callee's locals). Here we adjust
1683       // Llast_SP for the caller's frame, removing the extra space
1684       // for the current method's locals.
1685       *caller->register_addr(Llast_SP) = *interpreter_frame->register_addr(I5_savedSP);
1686     } else {
1687       assert(*caller->register_addr(Llast_SP) >= *interpreter_frame->register_addr(I5_savedSP), "strange Llast_SP");
1688     }
1689   } else {
1690     assert(caller->is_compiled_frame() || caller->is_entry_frame(), "only possible cases");
1691     // Don't have Lesp available; lay out locals block in the caller
1692     // adjacent to the register window save area.
1693     //
1694     // Compiled frames do not allocate a varargs area which is why this if
1695     // statement is needed.
1696     //
1697     if (caller->is_compiled_frame()) {
1698       locals = fp + frame::register_save_words + local_words - 1;
1699     } else {
1700       locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
1701     }
1702     if (!caller->is_entry_frame()) {
1703       // Caller wants his own SP back
1704       int caller_frame_size = caller->cb()->frame_size();
1705       *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS;
1706     }
1707   }
1708   if (TraceDeoptimization) {
1709     if (caller->is_entry_frame()) {
1710       // make sure I5_savedSP and the entry frames notion of saved SP
1711       // agree.  This assertion duplicate a check in entry frame code
1712       // but catches the failure earlier.
1713       assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP),
1714              "would change callers SP");
1715     }
1716     if (caller->is_entry_frame()) {
1717       tty->print("entry ");
1718     }
1719     if (caller->is_compiled_frame()) {
1720       tty->print("compiled ");
1721       if (caller->is_deoptimized_frame()) {
1722         tty->print("(deopt) ");
1723       }
1724     }
1725     if (caller->is_interpreted_frame()) {
1726       tty->print("interpreted ");
1727     }
1728     tty->print_cr("caller fp=0x%x sp=0x%x", caller->fp(), caller->sp());
1729     tty->print_cr("save area = 0x%x, 0x%x", caller->sp(), caller->sp() + 16);
1730     tty->print_cr("save area = 0x%x, 0x%x", caller->fp(), caller->fp() + 16);
1731     tty->print_cr("interpreter fp=0x%x sp=0x%x", interpreter_frame->fp(), interpreter_frame->sp());
1732     tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->sp(), interpreter_frame->sp() + 16);
1733     tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->fp(), interpreter_frame->fp() + 16);
1734     tty->print_cr("Llocals = 0x%x", locals);
1735     tty->print_cr("Lesp = 0x%x", esp);
1736     tty->print_cr("Lmonitors = 0x%x", monitors);
1737   }
1738 
1739   if (method->max_locals() > 0) {
1740     assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area");
1741     assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area");
1742     assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area");
1743     assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area");
1744   }
1745 #ifdef _LP64
1746   assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd");
1747 #endif
1748 
1749   *interpreter_frame->register_addr(Lmethod)     = (intptr_t) method;
1750   *interpreter_frame->register_addr(Llocals)     = (intptr_t) locals;
1751   *interpreter_frame->register_addr(Lmonitors)   = (intptr_t) monitors;
1752   *interpreter_frame->register_addr(Lesp)        = (intptr_t) esp;
1753   // Llast_SP will be same as SP as there is no adapter space
1754   *interpreter_frame->register_addr(Llast_SP)    = (intptr_t) interpreter_frame->sp() - STACK_BIAS;
1755   *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache();
1756 #ifdef FAST_DISPATCH
1757   *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table();
1758 #endif
1759 
1760 
1761 #ifdef ASSERT
1762   BasicObjectLock* mp = (BasicObjectLock*)monitors;
1763 
1764   assert(interpreter_frame->interpreter_frame_method() == method, "method matches");
1765   assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize)), "locals match");
1766   assert(interpreter_frame->interpreter_frame_monitor_end()   == mp, "monitor_end matches");
1767   assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches");
1768   assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches");
1769 
1770   // check bounds
1771   intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1);
1772   intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words;
1773   assert(lo < monitors && montop <= hi, "monitors in bounds");
1774   assert(lo <= esp && esp < monitors, "esp in bounds");
1775 #endif // ASSERT
1776 }
1777 
1778 //----------------------------------------------------------------------------------------------------
1779 // Exceptions
1780 void TemplateInterpreterGenerator::generate_throw_exception() {
1781 
1782   // Entry point in previous activation (i.e., if the caller was interpreted)
1783   Interpreter::_rethrow_exception_entry = __ pc();
1784   // O0: exception
1785 
1786   // entry point for exceptions thrown within interpreter code
1787   Interpreter::_throw_exception_entry = __ pc();
1788   __ verify_thread();
1789   // expression stack is undefined here
1790   // O0: exception, i.e. Oexception
1791   // Lbcp: exception bcx
1792   __ verify_oop(Oexception);
1793 
1794 
1795   // expression stack must be empty before entering the VM in case of an exception
1796   __ empty_expression_stack();
1797   // find exception handler address and preserve exception oop
1798   // call C routine to find handler and jump to it
1799   __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1800   __ push_ptr(O1); // push exception for exception handler bytecodes
1801 
1802   __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1803   __ delayed()->nop();
1804 
1805 
1806   // if the exception is not handled in the current frame
1807   // the frame is removed and the exception is rethrown
1808   // (i.e. exception continuation is _rethrow_exception)
1809   //
1810   // Note: At this point the bci is still the bxi for the instruction which caused
1811   //       the exception and the expression stack is empty. Thus, for any VM calls
1812   //       at this point, GC will find a legal oop map (with empty expression stack).
1813 
1814   // in current activation
1815   // tos: exception
1816   // Lbcp: exception bcp
1817 
1818   //
1819   // JVMTI PopFrame support
1820   //
1821 
1822   Interpreter::_remove_activation_preserving_args_entry = __ pc();
1823   Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1824   // Set the popframe_processing bit in popframe_condition indicating that we are
1825   // currently handling popframe, so that call_VMs that may happen later do not trigger new
1826   // popframe handling cycles.
1827 
1828   __ ld(popframe_condition_addr, G3_scratch);
1829   __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1830   __ stw(G3_scratch, popframe_condition_addr);
1831 
1832   // Empty the expression stack, as in normal exception handling
1833   __ empty_expression_stack();
1834   __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1835 
1836   {
1837     // Check to see whether we are returning to a deoptimized frame.
1838     // (The PopFrame call ensures that the caller of the popped frame is
1839     // either interpreted or compiled and deoptimizes it if compiled.)
1840     // In this case, we can't call dispatch_next() after the frame is
1841     // popped, but instead must save the incoming arguments and restore
1842     // them after deoptimization has occurred.
1843     //
1844     // Note that we don't compare the return PC against the
1845     // deoptimization blob's unpack entry because of the presence of
1846     // adapter frames in C2.
1847     Label caller_not_deoptimized;
1848     __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1849     __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1850 
1851     const Register Gtmp1 = G3_scratch;
1852     const Register Gtmp2 = G1_scratch;
1853     const Register RconstMethod = Gtmp1;
1854     const Address constMethod(Lmethod, Method::const_offset());
1855     const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1856 
1857     // Compute size of arguments for saving when returning to deoptimized caller
1858     __ ld_ptr(constMethod, RconstMethod);
1859     __ lduh(size_of_parameters, Gtmp1);
1860     __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1861     __ sub(Llocals, Gtmp1, Gtmp2);
1862     __ add(Gtmp2, wordSize, Gtmp2);
1863     // Save these arguments
1864     __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1865     // Inform deoptimization that it is responsible for restoring these arguments
1866     __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1867     Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1868     __ st(Gtmp1, popframe_condition_addr);
1869 
1870     // Return from the current method
1871     // The caller's SP was adjusted upon method entry to accomodate
1872     // the callee's non-argument locals. Undo that adjustment.
1873     __ ret();
1874     __ delayed()->restore(I5_savedSP, G0, SP);
1875 
1876     __ bind(caller_not_deoptimized);
1877   }
1878 
1879   // Clear the popframe condition flag
1880   __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1881 
1882   // Get out of the current method (how this is done depends on the particular compiler calling
1883   // convention that the interpreter currently follows)
1884   // The caller's SP was adjusted upon method entry to accomodate
1885   // the callee's non-argument locals. Undo that adjustment.
1886   __ restore(I5_savedSP, G0, SP);
1887   // The method data pointer was incremented already during
1888   // call profiling. We have to restore the mdp for the current bcp.
1889   if (ProfileInterpreter) {
1890     __ set_method_data_pointer_for_bcp();
1891   }
1892 
1893 #if INCLUDE_JVMTI
1894   if (EnableInvokeDynamic) {
1895     Label L_done;
1896 
1897     __ ldub(Address(Lbcp, 0), G1_scratch);  // Load current bytecode
1898     __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done);
1899 
1900     // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
1901     // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
1902 
1903     __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp);
1904 
1905     __ br_null(G1_scratch, false, Assembler::pn, L_done);
1906     __ delayed()->nop();
1907 
1908     __ st_ptr(G1_scratch, Lesp, wordSize);
1909     __ bind(L_done);
1910   }
1911 #endif // INCLUDE_JVMTI
1912 
1913   // Resume bytecode interpretation at the current bcp
1914   __ dispatch_next(vtos);
1915   // end of JVMTI PopFrame support
1916 
1917   Interpreter::_remove_activation_entry = __ pc();
1918 
1919   // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1920   __ pop_ptr(Oexception);                                  // get exception
1921 
1922   // Intel has the following comment:
1923   //// remove the activation (without doing throws on illegalMonitorExceptions)
1924   // They remove the activation without checking for bad monitor state.
1925   // %%% We should make sure this is the right semantics before implementing.
1926 
1927   __ set_vm_result(Oexception);
1928   __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1929 
1930   __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1931 
1932   __ get_vm_result(Oexception);
1933   __ verify_oop(Oexception);
1934 
1935     const int return_reg_adjustment = frame::pc_return_offset;
1936   Address issuing_pc_addr(I7, return_reg_adjustment);
1937 
1938   // We are done with this activation frame; find out where to go next.
1939   // The continuation point will be an exception handler, which expects
1940   // the following registers set up:
1941   //
1942   // Oexception: exception
1943   // Oissuing_pc: the local call that threw exception
1944   // Other On: garbage
1945   // In/Ln:  the contents of the caller's register window
1946   //
1947   // We do the required restore at the last possible moment, because we
1948   // need to preserve some state across a runtime call.
1949   // (Remember that the caller activation is unknown--it might not be
1950   // interpreted, so things like Lscratch are useless in the caller.)
1951 
1952   // Although the Intel version uses call_C, we can use the more
1953   // compact call_VM.  (The only real difference on SPARC is a
1954   // harmlessly ignored [re]set_last_Java_frame, compared with
1955   // the Intel code which lacks this.)
1956   __ mov(Oexception,      Oexception ->after_save());  // get exception in I0 so it will be on O0 after restore
1957   __ add(issuing_pc_addr, Oissuing_pc->after_save());  // likewise set I1 to a value local to the caller
1958   __ super_call_VM_leaf(L7_thread_cache,
1959                         CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1960                         G2_thread, Oissuing_pc->after_save());
1961 
1962   // The caller's SP was adjusted upon method entry to accomodate
1963   // the callee's non-argument locals. Undo that adjustment.
1964   __ JMP(O0, 0);                         // return exception handler in caller
1965   __ delayed()->restore(I5_savedSP, G0, SP);
1966 
1967   // (same old exception object is already in Oexception; see above)
1968   // Note that an "issuing PC" is actually the next PC after the call
1969 }
1970 
1971 
1972 //
1973 // JVMTI ForceEarlyReturn support
1974 //
1975 
1976 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1977   address entry = __ pc();
1978 
1979   __ empty_expression_stack();
1980   __ load_earlyret_value(state);
1981 
1982   __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1983   Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1984 
1985   // Clear the earlyret state
1986   __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1987 
1988   __ remove_activation(state,
1989                        /* throw_monitor_exception */ false,
1990                        /* install_monitor_exception */ false);
1991 
1992   // The caller's SP was adjusted upon method entry to accomodate
1993   // the callee's non-argument locals. Undo that adjustment.
1994   __ ret();                             // return to caller
1995   __ delayed()->restore(I5_savedSP, G0, SP);
1996 
1997   return entry;
1998 } // end of JVMTI ForceEarlyReturn support
1999 
2000 
2001 //------------------------------------------------------------------------------------------------------------------------
2002 // Helper for vtos entry point generation
2003 
2004 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) {
2005   assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
2006   Label L;
2007   aep = __ pc(); __ push_ptr(); __ ba_short(L);
2008   fep = __ pc(); __ push_f();   __ ba_short(L);
2009   dep = __ pc(); __ push_d();   __ ba_short(L);
2010   lep = __ pc(); __ push_l();   __ ba_short(L);
2011   iep = __ pc(); __ push_i();
2012   bep = cep = sep = iep;                        // there aren't any
2013   vep = __ pc(); __ bind(L);                    // fall through
2014   generate_and_dispatch(t);
2015 }
2016 
2017 // --------------------------------------------------------------------------------
2018 
2019 
2020 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2021  : TemplateInterpreterGenerator(code) {
2022    generate_all(); // down here so it can be "virtual"
2023 }
2024 
2025 // --------------------------------------------------------------------------------
2026 
2027 // Non-product code
2028 #ifndef PRODUCT
2029 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
2030   address entry = __ pc();
2031 
2032   __ push(state);
2033   __ mov(O7, Lscratch); // protect return address within interpreter
2034 
2035   // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
2036   __ mov( Otos_l2, G3_scratch );
2037   __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
2038   __ mov(Lscratch, O7); // restore return address
2039   __ pop(state);
2040   __ retl();
2041   __ delayed()->nop();
2042 
2043   return entry;
2044 }
2045 
2046 
2047 // helpers for generate_and_dispatch
2048 
2049 void TemplateInterpreterGenerator::count_bytecode() {
2050   __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
2051 }
2052 
2053 
2054 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
2055   __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
2056 }
2057 
2058 
2059 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2060   AddressLiteral index   (&BytecodePairHistogram::_index);
2061   AddressLiteral counters((address) &BytecodePairHistogram::_counters);
2062 
2063   // get index, shift out old bytecode, bring in new bytecode, and store it
2064   // _index = (_index >> log2_number_of_codes) |
2065   //          (bytecode << log2_number_of_codes);
2066 
2067   __ load_contents(index, G4_scratch);
2068   __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
2069   __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes,  G3_scratch );
2070   __ or3( G3_scratch,  G4_scratch, G4_scratch );
2071   __ store_contents(G4_scratch, index, G3_scratch);
2072 
2073   // bump bucket contents
2074   // _counters[_index] ++;
2075 
2076   __ set(counters, G3_scratch);                       // loads into G3_scratch
2077   __ sll( G4_scratch, LogBytesPerWord, G4_scratch );  // Index is word address
2078   __ add (G3_scratch, G4_scratch, G3_scratch);        // Add in index
2079   __ ld (G3_scratch, 0, G4_scratch);
2080   __ inc (G4_scratch);
2081   __ st (G4_scratch, 0, G3_scratch);
2082 }
2083 
2084 
2085 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2086   // Call a little run-time stub to avoid blow-up for each bytecode.
2087   // The run-time runtime saves the right registers, depending on
2088   // the tosca in-state for the given template.
2089   address entry = Interpreter::trace_code(t->tos_in());
2090   guarantee(entry != NULL, "entry must have been generated");
2091   __ call(entry, relocInfo::none);
2092   __ delayed()->nop();
2093 }
2094 
2095 
2096 void TemplateInterpreterGenerator::stop_interpreter_at() {
2097   AddressLiteral counter(&BytecodeCounter::_counter_value);
2098   __ load_contents(counter, G3_scratch);
2099   AddressLiteral stop_at(&StopInterpreterAt);
2100   __ load_ptr_contents(stop_at, G4_scratch);
2101   __ cmp(G3_scratch, G4_scratch);
2102   __ breakpoint_trap(Assembler::equal, Assembler::icc);
2103 }
2104 #endif // not PRODUCT
2105 #endif // !CC_INTERP