1 /*
   2  * Copyright (c) 1997, 2018, 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 "gc/shared/barrierSetAssembler.hpp"
  28 #include "interpreter/bytecodeHistogram.hpp"
  29 #include "interpreter/interpreter.hpp"
  30 #include "interpreter/interpreterRuntime.hpp"
  31 #include "interpreter/interp_masm.hpp"
  32 #include "interpreter/templateInterpreterGenerator.hpp"
  33 #include "interpreter/templateTable.hpp"
  34 #include "oops/arrayOop.hpp"
  35 #include "oops/methodData.hpp"
  36 #include "oops/method.hpp"
  37 #include "oops/oop.inline.hpp"
  38 #include "prims/jvmtiExport.hpp"
  39 #include "prims/jvmtiThreadState.hpp"
  40 #include "runtime/arguments.hpp"
  41 #include "runtime/deoptimization.hpp"
  42 #include "runtime/frame.inline.hpp"
  43 #include "runtime/sharedRuntime.hpp"
  44 #include "runtime/stubRoutines.hpp"
  45 #include "runtime/synchronizer.hpp"
  46 #include "runtime/timer.hpp"
  47 #include "runtime/vframeArray.hpp"
  48 #include "utilities/align.hpp"
  49 #include "utilities/debug.hpp"
  50 #include "utilities/macros.hpp"
  51 
  52 // Size of interpreter code.  Increase if too small.  Interpreter will
  53 // fail with a guarantee ("not enough space for interpreter generation");
  54 // if too small.
  55 // Run with +PrintInterpreter to get the VM to print out the size.
  56 // Max size with JVMTI
  57 // The sethi() instruction generates lots more instructions when shell
  58 // stack limit is unlimited, so that's why this is much bigger.
  59 int TemplateInterpreter::InterpreterCodeSize = 260 * K;
  60 
  61 // Generation of Interpreter
  62 //
  63 // The TemplateInterpreterGenerator generates the interpreter into Interpreter::_code.
  64 
  65 
  66 #define __ _masm->
  67 
  68 
  69 //----------------------------------------------------------------------------------------------------
  70 
  71 // LP64 passes floating point arguments in F1, F3, F5, etc. instead of
  72 // O0, O1, O2 etc..
  73 // Doubles are passed in D0, D2, D4
  74 // We store the signature of the first 16 arguments in the first argument
  75 // slot because it will be overwritten prior to calling the native
  76 // function, with the pointer to the JNIEnv.
  77 // If LP64 there can be up to 16 floating point arguments in registers
  78 // or 6 integer registers.
  79 address TemplateInterpreterGenerator::generate_slow_signature_handler() {
  80 
  81   enum {
  82     non_float  = 0,
  83     float_sig  = 1,
  84     double_sig = 2,
  85     sig_mask   = 3
  86   };
  87 
  88   address entry = __ pc();
  89   Argument argv(0, true);
  90 
  91   // We are in the jni transition frame. Save the last_java_frame corresponding to the
  92   // outer interpreter frame
  93   //
  94   __ set_last_Java_frame(FP, noreg);
  95   // make sure the interpreter frame we've pushed has a valid return pc
  96   __ mov(O7, I7);
  97   __ mov(Lmethod, G3_scratch);
  98   __ mov(Llocals, G4_scratch);
  99   __ save_frame(0);
 100   __ mov(G2_thread, L7_thread_cache);
 101   __ add(argv.address_in_frame(), O3);
 102   __ mov(G2_thread, O0);
 103   __ mov(G3_scratch, O1);
 104   __ call(CAST_FROM_FN_PTR(address, InterpreterRuntime::slow_signature_handler), relocInfo::runtime_call_type);
 105   __ delayed()->mov(G4_scratch, O2);
 106   __ mov(L7_thread_cache, G2_thread);
 107   __ reset_last_Java_frame();
 108 
 109 
 110   // load the register arguments (the C code packed them as varargs)
 111   Address Sig = argv.address_in_frame();        // Argument 0 holds the signature
 112   __ ld_ptr( Sig, G3_scratch );                   // Get register argument signature word into G3_scratch
 113   __ mov( G3_scratch, G4_scratch);
 114   __ srl( G4_scratch, 2, G4_scratch);             // Skip Arg 0
 115   Label done;
 116   for (Argument ldarg = argv.successor(); ldarg.is_float_register(); ldarg = ldarg.successor()) {
 117     Label NonFloatArg;
 118     Label LoadFloatArg;
 119     Label LoadDoubleArg;
 120     Label NextArg;
 121     Address a = ldarg.address_in_frame();
 122     __ andcc(G4_scratch, sig_mask, G3_scratch);
 123     __ br(Assembler::zero, false, Assembler::pt, NonFloatArg);
 124     __ delayed()->nop();
 125 
 126     __ cmp(G3_scratch, float_sig );
 127     __ br(Assembler::equal, false, Assembler::pt, LoadFloatArg);
 128     __ delayed()->nop();
 129 
 130     __ cmp(G3_scratch, double_sig );
 131     __ br(Assembler::equal, false, Assembler::pt, LoadDoubleArg);
 132     __ delayed()->nop();
 133 
 134     __ bind(NonFloatArg);
 135     // There are only 6 integer register arguments!
 136     if ( ldarg.is_register() )
 137       __ ld_ptr(ldarg.address_in_frame(), ldarg.as_register());
 138     else {
 139     // Optimization, see if there are any more args and get out prior to checking
 140     // all 16 float registers.  My guess is that this is rare.
 141     // If is_register is false, then we are done the first six integer args.
 142       __ br_null_short(G4_scratch, Assembler::pt, done);
 143     }
 144     __ ba(NextArg);
 145     __ delayed()->srl( G4_scratch, 2, G4_scratch );
 146 
 147     __ bind(LoadFloatArg);
 148     __ ldf( FloatRegisterImpl::S, a, ldarg.as_float_register(), 4);
 149     __ ba(NextArg);
 150     __ delayed()->srl( G4_scratch, 2, G4_scratch );
 151 
 152     __ bind(LoadDoubleArg);
 153     __ ldf( FloatRegisterImpl::D, a, ldarg.as_double_register() );
 154     __ ba(NextArg);
 155     __ delayed()->srl( G4_scratch, 2, G4_scratch );
 156 
 157     __ bind(NextArg);
 158   }
 159 
 160   __ bind(done);
 161   __ ret();
 162   __ delayed()->restore(O0, 0, Lscratch);  // caller's Lscratch gets the result handler
 163 
 164   return entry;
 165 }
 166 
 167 void TemplateInterpreterGenerator::generate_counter_overflow(Label& Lcontinue) {
 168 
 169   // Generate code to initiate compilation on the counter overflow.
 170 
 171   // InterpreterRuntime::frequency_counter_overflow takes two arguments,
 172   // the first indicates if the counter overflow occurs at a backwards branch (NULL bcp)
 173   // and the second is only used when the first is true.  We pass zero for both.
 174   // The call returns the address of the verified entry point for the method or NULL
 175   // if the compilation did not complete (either went background or bailed out).
 176   __ set((int)false, O2);
 177   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), O2, O2, true);
 178   // returns verified_entry_point or NULL
 179   // we ignore it in any case
 180   __ ba_short(Lcontinue);
 181 }
 182 
 183 
 184 // End of helpers
 185 
 186 // Various method entries
 187 
 188 // Abstract method entry
 189 // Attempt to execute abstract method. Throw exception
 190 //
 191 address TemplateInterpreterGenerator::generate_abstract_entry(void) {
 192   address entry = __ pc();
 193   // abstract method entry
 194   // throw exception
 195   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodErrorWithMethod), G5_method);
 196   // the call_VM checks for exception, so we should never return here.
 197   __ should_not_reach_here();
 198   return entry;
 199 }
 200 
 201 void TemplateInterpreterGenerator::save_native_result(void) {
 202   // result potentially in O0/O1: save it across calls
 203   const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
 204 
 205   // result potentially in F0/F1: save it across calls
 206   const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
 207 
 208   // save and restore any potential method result value around the unlocking operation
 209   __ stf(FloatRegisterImpl::D, F0, d_tmp);
 210   __ stx(O0, l_tmp);
 211 }
 212 
 213 void TemplateInterpreterGenerator::restore_native_result(void) {
 214   const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
 215   const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
 216 
 217   // Restore any method result value
 218   __ ldf(FloatRegisterImpl::D, d_tmp, F0);
 219   __ ldx(l_tmp, O0);
 220 }
 221 
 222 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
 223   assert(!pass_oop || message == NULL, "either oop or message but not both");
 224   address entry = __ pc();
 225   // expression stack must be empty before entering the VM if an exception happened
 226   __ empty_expression_stack();
 227   // load exception object
 228   __ set((intptr_t)name, G3_scratch);
 229   if (pass_oop) {
 230     __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
 231   } else {
 232     __ set((intptr_t)message, G4_scratch);
 233     __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
 234   }
 235   // throw exception
 236   assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
 237   AddressLiteral thrower(Interpreter::throw_exception_entry());
 238   __ jump_to(thrower, G3_scratch);
 239   __ delayed()->nop();
 240   return entry;
 241 }
 242 
 243 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
 244   address entry = __ pc();
 245   // expression stack must be empty before entering the VM if an exception
 246   // happened
 247   __ empty_expression_stack();
 248   // load exception object
 249   __ call_VM(Oexception,
 250              CAST_FROM_FN_PTR(address,
 251                               InterpreterRuntime::throw_ClassCastException),
 252              Otos_i);
 253   __ should_not_reach_here();
 254   return entry;
 255 }
 256 
 257 
 258 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
 259   address entry = __ pc();
 260   // expression stack must be empty before entering the VM if an exception happened
 261   __ empty_expression_stack();
 262   // convention: expect aberrant index in register G3_scratch, then shuffle the
 263   // index to G4_scratch for the VM call
 264   __ mov(G3_scratch, G4_scratch);
 265   __ set((intptr_t)name, G3_scratch);
 266   __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
 267   __ should_not_reach_here();
 268   return entry;
 269 }
 270 
 271 
 272 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
 273   address entry = __ pc();
 274   // expression stack must be empty before entering the VM if an exception happened
 275   __ empty_expression_stack();
 276   __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
 277   __ should_not_reach_here();
 278   return entry;
 279 }
 280 
 281 
 282 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
 283   address entry = __ pc();
 284 
 285   if (state == atos) {
 286     __ profile_return_type(O0, G3_scratch, G1_scratch);
 287   }
 288 
 289   // The callee returns with the stack possibly adjusted by adapter transition
 290   // We remove that possible adjustment here.
 291   // All interpreter local registers are untouched. Any result is passed back
 292   // in the O0/O1 or float registers. Before continuing, the arguments must be
 293   // popped from the java expression stack; i.e., Lesp must be adjusted.
 294 
 295   __ mov(Llast_SP, SP);   // Remove any adapter added stack space.
 296 
 297   const Register cache = G3_scratch;
 298   const Register index  = G1_scratch;
 299   __ get_cache_and_index_at_bcp(cache, index, 1, index_size);
 300 
 301   const Register flags = cache;
 302   __ ld_ptr(cache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset(), flags);
 303   const Register parameter_size = flags;
 304   __ and3(flags, ConstantPoolCacheEntry::parameter_size_mask, parameter_size);  // argument size in words
 305   __ sll(parameter_size, Interpreter::logStackElementSize, parameter_size);     // each argument size in bytes
 306   __ add(Lesp, parameter_size, Lesp);                                           // pop arguments
 307 
 308   __ check_and_handle_popframe(Gtemp);
 309   __ check_and_handle_earlyret(Gtemp);
 310 
 311   __ dispatch_next(state, step);
 312 
 313   return entry;
 314 }
 315 
 316 
 317 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step, address continuation) {
 318   address entry = __ pc();
 319   __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
 320 #if INCLUDE_JVMCI
 321   // Check if we need to take lock at entry of synchronized method.  This can
 322   // only occur on method entry so emit it only for vtos with step 0.
 323   if (EnableJVMCI && state == vtos && step == 0) {
 324     Label L;
 325     Address pending_monitor_enter_addr(G2_thread, JavaThread::pending_monitorenter_offset());
 326     __ ldbool(pending_monitor_enter_addr, Gtemp);  // Load if pending monitor enter
 327     __ cmp_and_br_short(Gtemp, G0, Assembler::equal, Assembler::pn, L);
 328     // Clear flag.
 329     __ stbool(G0, pending_monitor_enter_addr);
 330     // Take lock.
 331     lock_method();
 332     __ bind(L);
 333   } else {
 334 #ifdef ASSERT
 335     if (EnableJVMCI) {
 336       Label L;
 337       Address pending_monitor_enter_addr(G2_thread, JavaThread::pending_monitorenter_offset());
 338       __ ldbool(pending_monitor_enter_addr, Gtemp);  // Load if pending monitor enter
 339       __ cmp_and_br_short(Gtemp, G0, Assembler::equal, Assembler::pn, L);
 340       __ stop("unexpected pending monitor in deopt entry");
 341       __ bind(L);
 342     }
 343 #endif
 344   }
 345 #endif
 346   { Label L;
 347     Address exception_addr(G2_thread, Thread::pending_exception_offset());
 348     __ ld_ptr(exception_addr, Gtemp);  // Load pending exception.
 349     __ br_null_short(Gtemp, Assembler::pt, L);
 350     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
 351     __ should_not_reach_here();
 352     __ bind(L);
 353   }
 354   if (continuation == NULL) {
 355     __ dispatch_next(state, step);
 356   } else {
 357     __ jump_to_entry(continuation);
 358   }
 359   return entry;
 360 }
 361 
 362 // A result handler converts/unboxes a native call result into
 363 // a java interpreter/compiler result. The current frame is an
 364 // interpreter frame. The activation frame unwind code must be
 365 // consistent with that of TemplateTable::_return(...). In the
 366 // case of native methods, the caller's SP was not modified.
 367 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
 368   address entry = __ pc();
 369   Register Itos_i  = Otos_i ->after_save();
 370   Register Itos_l  = Otos_l ->after_save();
 371   Register Itos_l1 = Otos_l1->after_save();
 372   Register Itos_l2 = Otos_l2->after_save();
 373   switch (type) {
 374     case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
 375     case T_CHAR   : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i);   break; // cannot use and3, 0xFFFF too big as immediate value!
 376     case T_BYTE   : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i);   break;
 377     case T_SHORT  : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i);   break;
 378     case T_LONG   :
 379     case T_INT    : __ mov(O0, Itos_i);                         break;
 380     case T_VOID   : /* nothing to do */                         break;
 381     case T_FLOAT  : assert(F0 == Ftos_f, "fix this code" );     break;
 382     case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" );     break;
 383     case T_OBJECT :
 384       __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
 385       __ verify_oop(Itos_i);
 386       break;
 387     default       : ShouldNotReachHere();
 388   }
 389   __ ret();                           // return from interpreter activation
 390   __ delayed()->restore(I5_savedSP, G0, SP);  // remove interpreter frame
 391   NOT_PRODUCT(__ emit_int32(0);)       // marker for disassembly
 392   return entry;
 393 }
 394 
 395 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
 396   address entry = __ pc();
 397   __ push(state);
 398   __ call_VM(noreg, runtime_entry);
 399   __ dispatch_via(vtos, Interpreter::normal_table(vtos));
 400   return entry;
 401 }
 402 
 403 
 404 //
 405 // Helpers for commoning out cases in the various type of method entries.
 406 //
 407 
 408 // increment invocation count & check for overflow
 409 //
 410 // Note: checking for negative value instead of overflow
 411 //       so we have a 'sticky' overflow test
 412 //
 413 // Lmethod: method
 414 // ??: invocation counter
 415 //
 416 void TemplateInterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
 417   // Note: In tiered we increment either counters in MethodCounters* or in
 418   // MDO depending if we're profiling or not.
 419   const Register G3_method_counters = G3_scratch;
 420   Label done;
 421 
 422   if (TieredCompilation) {
 423     const int increment = InvocationCounter::count_increment;
 424     Label no_mdo;
 425     if (ProfileInterpreter) {
 426       // If no method data exists, go to profile_continue.
 427       __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
 428       __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
 429       // Increment counter
 430       Address mdo_invocation_counter(G4_scratch,
 431                                      in_bytes(MethodData::invocation_counter_offset()) +
 432                                      in_bytes(InvocationCounter::counter_offset()));
 433       Address mask(G4_scratch, in_bytes(MethodData::invoke_mask_offset()));
 434       __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
 435                                  G3_scratch, Lscratch,
 436                                  Assembler::zero, overflow);
 437       __ ba_short(done);
 438     }
 439 
 440     // Increment counter in MethodCounters*
 441     __ bind(no_mdo);
 442     Address invocation_counter(G3_method_counters,
 443             in_bytes(MethodCounters::invocation_counter_offset()) +
 444             in_bytes(InvocationCounter::counter_offset()));
 445     __ get_method_counters(Lmethod, G3_method_counters, done);
 446     Address mask(G3_method_counters, in_bytes(MethodCounters::invoke_mask_offset()));
 447     __ increment_mask_and_jump(invocation_counter, increment, mask,
 448                                G4_scratch, Lscratch,
 449                                Assembler::zero, overflow);
 450     __ bind(done);
 451   } else { // not TieredCompilation
 452     // Update standard invocation counters
 453     __ get_method_counters(Lmethod, G3_method_counters, done);
 454     __ increment_invocation_counter(G3_method_counters, O0, G4_scratch);
 455     if (ProfileInterpreter) {
 456       Address interpreter_invocation_counter(G3_method_counters,
 457             in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
 458       __ ld(interpreter_invocation_counter, G4_scratch);
 459       __ inc(G4_scratch);
 460       __ st(G4_scratch, interpreter_invocation_counter);
 461     }
 462 
 463     if (ProfileInterpreter && profile_method != NULL) {
 464       // Test to see if we should create a method data oop
 465       Address profile_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_profile_limit_offset()));
 466       __ ld(profile_limit, G1_scratch);
 467       __ cmp_and_br_short(O0, G1_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
 468 
 469       // if no method data exists, go to profile_method
 470       __ test_method_data_pointer(*profile_method);
 471     }
 472 
 473     Address invocation_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_invocation_limit_offset()));
 474     __ ld(invocation_limit, G3_scratch);
 475     __ cmp(O0, G3_scratch);
 476     __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
 477     __ delayed()->nop();
 478     __ bind(done);
 479   }
 480 }
 481 
 482 // Allocate monitor and lock method (asm interpreter)
 483 // ebx - Method*
 484 //
 485 void TemplateInterpreterGenerator::lock_method() {
 486   __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0);  // Load access flags.
 487 
 488 #ifdef ASSERT
 489  { Label ok;
 490    __ btst(JVM_ACC_SYNCHRONIZED, O0);
 491    __ br( Assembler::notZero, false, Assembler::pt, ok);
 492    __ delayed()->nop();
 493    __ stop("method doesn't need synchronization");
 494    __ bind(ok);
 495   }
 496 #endif // ASSERT
 497 
 498   // get synchronization object to O0
 499   { Label done;
 500     __ btst(JVM_ACC_STATIC, O0);
 501     __ br( Assembler::zero, true, Assembler::pt, done);
 502     __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
 503 
 504     // lock the mirror, not the Klass*
 505     __ load_mirror(O0, Lmethod, Lscratch);
 506 
 507 #ifdef ASSERT
 508     __ tst(O0);
 509     __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
 510 #endif // ASSERT
 511 
 512     __ bind(done);
 513   }
 514 
 515   __ add_monitor_to_stack(true, noreg, noreg);  // allocate monitor elem
 516   __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes());   // store object
 517   // __ untested("lock_object from method entry");
 518   __ lock_object(Lmonitors, O0);
 519 }
 520 
 521 // See if we've got enough room on the stack for locals plus overhead below
 522 // JavaThread::stack_overflow_limit(). If not, throw a StackOverflowError
 523 // without going through the signal handler, i.e., reserved and yellow zones
 524 // will not be made usable. The shadow zone must suffice to handle the
 525 // overflow.
 526 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
 527                                                                  Register Rscratch) {
 528   const int page_size = os::vm_page_size();
 529   Label after_frame_check;
 530 
 531   assert_different_registers(Rframe_size, Rscratch);
 532 
 533   __ set(page_size, Rscratch);
 534   __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
 535 
 536   // Get the stack overflow limit, and in debug, verify it is non-zero.
 537   __ ld_ptr(G2_thread, JavaThread::stack_overflow_limit_offset(), Rscratch);
 538 #ifdef ASSERT
 539   Label limit_ok;
 540   __ br_notnull_short(Rscratch, Assembler::pn, limit_ok);
 541   __ stop("stack overflow limit is zero in generate_stack_overflow_check");
 542   __ bind(limit_ok);
 543 #endif
 544 
 545   // Add in the size of the frame (which is the same as subtracting it from the
 546   // SP, which would take another register.
 547   __ add(Rscratch, Rframe_size, Rscratch);
 548 
 549   // The frame is greater than one page in size, so check against
 550   // the bottom of the stack.
 551   __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check);
 552 
 553   // The stack will overflow, throw an exception.
 554 
 555   // Note that SP is restored to sender's sp (in the delay slot). This
 556   // is necessary if the sender's frame is an extended compiled frame
 557   // (see gen_c2i_adapter()) and safer anyway in case of JSR292
 558   // adaptations.
 559 
 560   // Note also that the restored frame is not necessarily interpreted.
 561   // Use the shared runtime version of the StackOverflowError.
 562   assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
 563   AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
 564   __ jump_to(stub, Rscratch);
 565   __ delayed()->mov(O5_savedSP, SP);
 566 
 567   // If you get to here, then there is enough stack space.
 568   __ bind(after_frame_check);
 569 }
 570 
 571 
 572 //
 573 // Generate a fixed interpreter frame. This is identical setup for interpreted
 574 // methods and for native methods hence the shared code.
 575 
 576 
 577 //----------------------------------------------------------------------------------------------------
 578 // Stack frame layout
 579 //
 580 // When control flow reaches any of the entry types for the interpreter
 581 // the following holds ->
 582 //
 583 // C2 Calling Conventions:
 584 //
 585 // The entry code below assumes that the following registers are set
 586 // when coming in:
 587 //    G5_method: holds the Method* of the method to call
 588 //    Lesp:    points to the TOS of the callers expression stack
 589 //             after having pushed all the parameters
 590 //
 591 // The entry code does the following to setup an interpreter frame
 592 //   pop parameters from the callers stack by adjusting Lesp
 593 //   set O0 to Lesp
 594 //   compute X = (max_locals - num_parameters)
 595 //   bump SP up by X to accommodate the extra locals
 596 //   compute X = max_expression_stack
 597 //               + vm_local_words
 598 //               + 16 words of register save area
 599 //   save frame doing a save sp, -X, sp growing towards lower addresses
 600 //   set Lbcp, Lmethod, LcpoolCache
 601 //   set Llocals to i0
 602 //   set Lmonitors to FP - rounded_vm_local_words
 603 //   set Lesp to Lmonitors - 4
 604 //
 605 //  The frame has now been setup to do the rest of the entry code
 606 
 607 // Try this optimization:  Most method entries could live in a
 608 // "one size fits all" stack frame without all the dynamic size
 609 // calculations.  It might be profitable to do all this calculation
 610 // statically and approximately for "small enough" methods.
 611 
 612 //-----------------------------------------------------------------------------------------------
 613 
 614 // C1 Calling conventions
 615 //
 616 // Upon method entry, the following registers are setup:
 617 //
 618 // g2 G2_thread: current thread
 619 // g5 G5_method: method to activate
 620 // g4 Gargs  : pointer to last argument
 621 //
 622 //
 623 // Stack:
 624 //
 625 // +---------------+ <--- sp
 626 // |               |
 627 // : reg save area :
 628 // |               |
 629 // +---------------+ <--- sp + 0x40
 630 // |               |
 631 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
 632 // |               |
 633 // +---------------+ <--- sp + 0x5c
 634 // |               |
 635 // :     free      :
 636 // |               |
 637 // +---------------+ <--- Gargs
 638 // |               |
 639 // :   arguments   :
 640 // |               |
 641 // +---------------+
 642 // |               |
 643 //
 644 //
 645 //
 646 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
 647 //
 648 // +---------------+ <--- sp
 649 // |               |
 650 // : reg save area :
 651 // |               |
 652 // +---------------+ <--- sp + 0x40
 653 // |               |
 654 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
 655 // |               |
 656 // +---------------+ <--- sp + 0x5c
 657 // |               |
 658 // :               :
 659 // |               | <--- Lesp
 660 // +---------------+ <--- Lmonitors (fp - 0x18)
 661 // |   VM locals   |
 662 // +---------------+ <--- fp
 663 // |               |
 664 // : reg save area :
 665 // |               |
 666 // +---------------+ <--- fp + 0x40
 667 // |               |
 668 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
 669 // |               |
 670 // +---------------+ <--- fp + 0x5c
 671 // |               |
 672 // :     free      :
 673 // |               |
 674 // +---------------+
 675 // |               |
 676 // : nonarg locals :
 677 // |               |
 678 // +---------------+
 679 // |               |
 680 // :   arguments   :
 681 // |               | <--- Llocals
 682 // +---------------+ <--- Gargs
 683 // |               |
 684 
 685 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
 686   //
 687   //
 688   // The entry code sets up a new interpreter frame in 4 steps:
 689   //
 690   // 1) Increase caller's SP by for the extra local space needed:
 691   //    (check for overflow)
 692   //    Efficient implementation of xload/xstore bytecodes requires
 693   //    that arguments and non-argument locals are in a contiguously
 694   //    addressable memory block => non-argument locals must be
 695   //    allocated in the caller's frame.
 696   //
 697   // 2) Create a new stack frame and register window:
 698   //    The new stack frame must provide space for the standard
 699   //    register save area, the maximum java expression stack size,
 700   //    the monitor slots (0 slots initially), and some frame local
 701   //    scratch locations.
 702   //
 703   // 3) The following interpreter activation registers must be setup:
 704   //    Lesp       : expression stack pointer
 705   //    Lbcp       : bytecode pointer
 706   //    Lmethod    : method
 707   //    Llocals    : locals pointer
 708   //    Lmonitors  : monitor pointer
 709   //    LcpoolCache: constant pool cache
 710   //
 711   // 4) Initialize the non-argument locals if necessary:
 712   //    Non-argument locals may need to be initialized to NULL
 713   //    for GC to work. If the oop-map information is accurate
 714   //    (in the absence of the JSR problem), no initialization
 715   //    is necessary.
 716   //
 717   // (gri - 2/25/2000)
 718 
 719 
 720   int rounded_vm_local_words = align_up((int)frame::interpreter_frame_vm_local_words, WordsPerLong );
 721 
 722   const int extra_space =
 723     rounded_vm_local_words +                   // frame local scratch space
 724     Method::extra_stack_entries() +            // extra stack for jsr 292
 725     frame::memory_parameter_word_sp_offset +   // register save area
 726     (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
 727 
 728   const Register Glocals_size = G3;
 729   const Register RconstMethod = Glocals_size;
 730   const Register Otmp1 = O3;
 731   const Register Otmp2 = O4;
 732   // Lscratch can't be used as a temporary because the call_stub uses
 733   // it to assert that the stack frame was setup correctly.
 734   const Address constMethod       (G5_method, Method::const_offset());
 735   const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
 736 
 737   __ ld_ptr( constMethod, RconstMethod );
 738   __ lduh( size_of_parameters, Glocals_size);
 739 
 740   // Gargs points to first local + BytesPerWord
 741   // Set the saved SP after the register window save
 742   //
 743   assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
 744   __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
 745   __ add(Gargs, Otmp1, Gargs);
 746 
 747   if (native_call) {
 748     __ calc_mem_param_words( Glocals_size, Gframe_size );
 749     __ add( Gframe_size,  extra_space, Gframe_size);
 750     __ round_to( Gframe_size, WordsPerLong );
 751     __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
 752 
 753     // Native calls don't need the stack size check since they have no
 754     // expression stack and the arguments are already on the stack and
 755     // we only add a handful of words to the stack.
 756   } else {
 757 
 758     //
 759     // Compute number of locals in method apart from incoming parameters
 760     //
 761     const Address size_of_locals(Otmp1, ConstMethod::size_of_locals_offset());
 762     __ ld_ptr(constMethod, Otmp1);
 763     __ lduh(size_of_locals, Otmp1);
 764     __ sub(Otmp1, Glocals_size, Glocals_size);
 765     __ round_to(Glocals_size, WordsPerLong);
 766     __ sll(Glocals_size, Interpreter::logStackElementSize, Glocals_size);
 767 
 768     // See if the frame is greater than one page in size. If so,
 769     // then we need to verify there is enough stack space remaining.
 770     // Frame_size = (max_stack + extra_space) * BytesPerWord;
 771     __ ld_ptr(constMethod, Gframe_size);
 772     __ lduh(Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size);
 773     __ add(Gframe_size, extra_space, Gframe_size);
 774     __ round_to(Gframe_size, WordsPerLong);
 775     __ sll(Gframe_size, Interpreter::logStackElementSize, Gframe_size);
 776 
 777     // Add in java locals size for stack overflow check only
 778     __ add(Gframe_size, Glocals_size, Gframe_size);
 779 
 780     const Register Otmp2 = O4;
 781     assert_different_registers(Otmp1, Otmp2, O5_savedSP);
 782     generate_stack_overflow_check(Gframe_size, Otmp1);
 783 
 784     __ sub(Gframe_size, Glocals_size, Gframe_size);
 785 
 786     //
 787     // bump SP to accommodate the extra locals
 788     //
 789     __ sub(SP, Glocals_size, SP);
 790   }
 791 
 792   //
 793   // now set up a stack frame with the size computed above
 794   //
 795   __ neg( Gframe_size );
 796   __ save( SP, Gframe_size, SP );
 797 
 798   //
 799   // now set up all the local cache registers
 800   //
 801   // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
 802   // that all present references to Lbyte_code initialize the register
 803   // immediately before use
 804   if (native_call) {
 805     __ mov(G0, Lbcp);
 806   } else {
 807     __ ld_ptr(G5_method, Method::const_offset(), Lbcp);
 808     __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp);
 809   }
 810   __ mov( G5_method, Lmethod);                 // set Lmethod
 811   // Get mirror and store it in the frame as GC root for this Method*
 812   Register mirror = LcpoolCache;
 813   __ load_mirror(mirror, Lmethod, Lscratch);
 814   __ st_ptr(mirror, FP, (frame::interpreter_frame_mirror_offset * wordSize) + STACK_BIAS);
 815   __ get_constant_pool_cache(LcpoolCache);     // set LcpoolCache
 816   __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
 817   __ add(Lmonitors, STACK_BIAS, Lmonitors);    // Account for 64 bit stack bias
 818   __ sub(Lmonitors, BytesPerWord, Lesp);       // set Lesp
 819 
 820   // setup interpreter activation registers
 821   __ sub(Gargs, BytesPerWord, Llocals);        // set Llocals
 822 
 823   if (ProfileInterpreter) {
 824     __ set_method_data_pointer();
 825   }
 826 
 827 }
 828 
 829 // Method entry for java.lang.ref.Reference.get.
 830 address TemplateInterpreterGenerator::generate_Reference_get_entry(void) {
 831   // Code: _aload_0, _getfield, _areturn
 832   // parameter size = 1
 833   //
 834   // The code that gets generated by this routine is split into 2 parts:
 835   //    1. The "intrinsified" code performing an ON_WEAK_OOP_REF load,
 836   //    2. The slow path - which is an expansion of the regular method entry.
 837   //
 838   // Notes:-
 839   // * An intrinsic is always executed, where an ON_WEAK_OOP_REF load is performed.
 840   // * We may jump to the slow path iff the receiver is null. If the
 841   //   Reference object is null then we no longer perform an ON_WEAK_OOP_REF load
 842   //   Thus we can use the regular method entry code to generate the NPE.
 843   //
 844   // This code is based on generate_accessor_enty.
 845 
 846   address entry = __ pc();
 847 
 848   const int referent_offset = java_lang_ref_Reference::referent_offset;
 849   guarantee(referent_offset > 0, "referent offset not initialized");
 850 
 851   Label slow_path;
 852 
 853   // In the G1 code we don't check if we need to reach a safepoint. We
 854   // continue and the thread will safepoint at the next bytecode dispatch.
 855 
 856   // Check if local 0 != NULL
 857   // If the receiver is null then it is OK to jump to the slow path.
 858   __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
 859   // check if local 0 == NULL and go the slow path
 860   __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
 861 
 862   __ load_heap_oop(Otos_i, referent_offset, Otos_i, G3_scratch, ON_WEAK_OOP_REF);
 863 
 864   // _areturn
 865   __ retl();                      // return from leaf routine
 866   __ delayed()->mov(O5_savedSP, SP);
 867 
 868   // Generate regular method entry
 869   __ bind(slow_path);
 870   __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals));
 871   return entry;
 872 }
 873 
 874 /**
 875  * Method entry for static native methods:
 876  *   int java.util.zip.CRC32.update(int crc, int b)
 877  */
 878 address TemplateInterpreterGenerator::generate_CRC32_update_entry() {
 879 
 880   if (UseCRC32Intrinsics) {
 881     address entry = __ pc();
 882 
 883     Label L_slow_path;
 884     // If we need a safepoint check, generate full interpreter entry.
 885     __ safepoint_poll(L_slow_path, false, G2_thread, O2);
 886     __ delayed()->nop();
 887 
 888     // Load parameters
 889     const Register crc   = O0; // initial crc
 890     const Register val   = O1; // byte to update with
 891     const Register table = O2; // address of 256-entry lookup table
 892 
 893     __ ldub(Gargs, 3, val);
 894     __ lduw(Gargs, 8, crc);
 895 
 896     __ set(ExternalAddress(StubRoutines::crc_table_addr()), table);
 897 
 898     __ not1(crc); // ~crc
 899     __ clruwu(crc);
 900     __ update_byte_crc32(crc, val, table);
 901     __ not1(crc); // ~crc
 902 
 903     // result in O0
 904     __ retl();
 905     __ delayed()->nop();
 906 
 907     // generate a vanilla native entry as the slow path
 908     __ bind(L_slow_path);
 909     __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native));
 910     return entry;
 911   }
 912   return NULL;
 913 }
 914 
 915 /**
 916  * Method entry for static native methods:
 917  *   int java.util.zip.CRC32.updateBytes(int crc, byte[] b, int off, int len)
 918  *   int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
 919  */
 920 address TemplateInterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
 921 
 922   if (UseCRC32Intrinsics) {
 923     address entry = __ pc();
 924 
 925     Label L_slow_path;
 926     // If we need a safepoint check, generate full interpreter entry.
 927 
 928     __ safepoint_poll(L_slow_path, false, G2_thread, O2);
 929     __ delayed()->nop();
 930 
 931     // Load parameters from the stack
 932     const Register crc    = O0; // initial crc
 933     const Register buf    = O1; // source java byte array address
 934     const Register len    = O2; // len
 935     const Register offset = O3; // offset
 936 
 937     // Arguments are reversed on java expression stack
 938     // Calculate address of start element
 939     if (kind == Interpreter::java_util_zip_CRC32_updateByteBuffer) {
 940       __ lduw(Gargs, 0,  len);
 941       __ lduw(Gargs, 8,  offset);
 942       __ ldx( Gargs, 16, buf);
 943       __ lduw(Gargs, 32, crc);
 944       __ add(buf, offset, buf);
 945     } else {
 946       __ lduw(Gargs, 0,  len);
 947       __ lduw(Gargs, 8,  offset);
 948       __ ldx( Gargs, 16, buf);
 949       __ lduw(Gargs, 24, crc);
 950       __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size
 951       __ add(buf, offset, buf);
 952     }
 953 
 954     // Call the crc32 kernel
 955     __ MacroAssembler::save_thread(L7_thread_cache);
 956     __ kernel_crc32(crc, buf, len, O3);
 957     __ MacroAssembler::restore_thread(L7_thread_cache);
 958 
 959     // result in O0
 960     __ retl();
 961     __ delayed()->nop();
 962 
 963     // generate a vanilla native entry as the slow path
 964     __ bind(L_slow_path);
 965     __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native));
 966     return entry;
 967   }
 968   return NULL;
 969 }
 970 
 971 /**
 972  * Method entry for intrinsic-candidate (non-native) methods:
 973  *   int java.util.zip.CRC32C.updateBytes(int crc, byte[] b, int off, int end)
 974  *   int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
 975  * Unlike CRC32, CRC32C does not have any methods marked as native
 976  * CRC32C also uses an "end" variable instead of the length variable CRC32 uses
 977  */
 978 address TemplateInterpreterGenerator::generate_CRC32C_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
 979 
 980   if (UseCRC32CIntrinsics) {
 981     address entry = __ pc();
 982 
 983     // Load parameters from the stack
 984     const Register crc    = O0; // initial crc
 985     const Register buf    = O1; // source java byte array address
 986     const Register offset = O2; // offset
 987     const Register end    = O3; // index of last element to process
 988     const Register len    = O2; // len argument to the kernel
 989     const Register table  = O3; // crc32c lookup table address
 990 
 991     // Arguments are reversed on java expression stack
 992     // Calculate address of start element
 993     if (kind == Interpreter::java_util_zip_CRC32C_updateDirectByteBuffer) {
 994       __ lduw(Gargs, 0,  end);
 995       __ lduw(Gargs, 8,  offset);
 996       __ ldx( Gargs, 16, buf);
 997       __ lduw(Gargs, 32, crc);
 998       __ add(buf, offset, buf);
 999       __ sub(end, offset, len);
1000     } else {
1001       __ lduw(Gargs, 0,  end);
1002       __ lduw(Gargs, 8,  offset);
1003       __ ldx( Gargs, 16, buf);
1004       __ lduw(Gargs, 24, crc);
1005       __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size
1006       __ add(buf, offset, buf);
1007       __ sub(end, offset, len);
1008     }
1009 
1010     // Call the crc32c kernel
1011     __ MacroAssembler::save_thread(L7_thread_cache);
1012     __ kernel_crc32c(crc, buf, len, table);
1013     __ MacroAssembler::restore_thread(L7_thread_cache);
1014 
1015     // result in O0
1016     __ retl();
1017     __ delayed()->nop();
1018 
1019     return entry;
1020   }
1021   return NULL;
1022 }
1023 
1024 /* Math routines only partially supported.
1025  *
1026  *   Providing support for fma (float/double) only.
1027  */
1028 address TemplateInterpreterGenerator::generate_math_entry(AbstractInterpreter::MethodKind kind)
1029 {
1030   if (!InlineIntrinsics) return NULL; // Generate a vanilla entry
1031 
1032   address entry = __ pc();
1033 
1034   switch (kind) {
1035     case Interpreter::java_lang_math_fmaF:
1036       if (UseFMA) {
1037         // float .fma(float a, float b, float c)
1038         const FloatRegister ra = F1;
1039         const FloatRegister rb = F2;
1040         const FloatRegister rc = F3;
1041         const FloatRegister rd = F0; // Result.
1042 
1043         __ ldf(FloatRegisterImpl::S, Gargs,  0, rc);
1044         __ ldf(FloatRegisterImpl::S, Gargs,  8, rb);
1045         __ ldf(FloatRegisterImpl::S, Gargs, 16, ra);
1046 
1047         __ fmadd(FloatRegisterImpl::S, ra, rb, rc, rd);
1048         __ retl();  // Result in F0 (rd).
1049         __ delayed()->mov(O5_savedSP, SP);
1050 
1051         return entry;
1052       }
1053       break;
1054     case Interpreter::java_lang_math_fmaD:
1055       if (UseFMA) {
1056         // double .fma(double a, double b, double c)
1057         const FloatRegister ra = F2; // D1
1058         const FloatRegister rb = F4; // D2
1059         const FloatRegister rc = F6; // D3
1060         const FloatRegister rd = F0; // D0 Result.
1061 
1062         __ ldf(FloatRegisterImpl::D, Gargs,  0, rc);
1063         __ ldf(FloatRegisterImpl::D, Gargs, 16, rb);
1064         __ ldf(FloatRegisterImpl::D, Gargs, 32, ra);
1065 
1066         __ fmadd(FloatRegisterImpl::D, ra, rb, rc, rd);
1067         __ retl();  // Result in D0 (rd).
1068         __ delayed()->mov(O5_savedSP, SP);
1069 
1070         return entry;
1071       }
1072       break;
1073     default:
1074       break;
1075   }
1076   return NULL;
1077 }
1078 
1079 // TODO: rather than touching all pages, check against stack_overflow_limit and bang yellow page to
1080 // generate exception
1081 void TemplateInterpreterGenerator::bang_stack_shadow_pages(bool native_call) {
1082   // Quick & dirty stack overflow checking: bang the stack & handle trap.
1083   // Note that we do the banging after the frame is setup, since the exception
1084   // handling code expects to find a valid interpreter frame on the stack.
1085   // Doing the banging earlier fails if the caller frame is not an interpreter
1086   // frame.
1087   // (Also, the exception throwing code expects to unlock any synchronized
1088   // method receiver, so do the banging after locking the receiver.)
1089 
1090   // Bang each page in the shadow zone. We can't assume it's been done for
1091   // an interpreter frame with greater than a page of locals, so each page
1092   // needs to be checked.  Only true for non-native.
1093   if (UseStackBanging) {
1094     const int page_size = os::vm_page_size();
1095     const int n_shadow_pages = ((int)JavaThread::stack_shadow_zone_size()) / page_size;
1096     const int start_page = native_call ? n_shadow_pages : 1;
1097     for (int pages = start_page; pages <= n_shadow_pages; pages++) {
1098       __ bang_stack_with_offset(pages*page_size);
1099     }
1100   }
1101 }
1102 
1103 //
1104 // Interpreter stub for calling a native method. (asm interpreter)
1105 // This sets up a somewhat different looking stack for calling the native method
1106 // than the typical interpreter frame setup.
1107 //
1108 
1109 address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
1110   address entry = __ pc();
1111 
1112   // the following temporary registers are used during frame creation
1113   const Register Gtmp1 = G3_scratch ;
1114   const Register Gtmp2 = G1_scratch;
1115   bool inc_counter  = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1116 
1117   // make sure registers are different!
1118   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1119 
1120   const Address Laccess_flags(Lmethod, Method::access_flags_offset());
1121 
1122   const Register Glocals_size = G3;
1123   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1124 
1125   // make sure method is native & not abstract
1126   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1127 #ifdef ASSERT
1128   __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1129   { Label L;
1130     __ btst(JVM_ACC_NATIVE, Gtmp1);
1131     __ br(Assembler::notZero, false, Assembler::pt, L);
1132     __ delayed()->nop();
1133     __ stop("tried to execute non-native method as native");
1134     __ bind(L);
1135   }
1136   { Label L;
1137     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1138     __ br(Assembler::zero, false, Assembler::pt, L);
1139     __ delayed()->nop();
1140     __ stop("tried to execute abstract method as non-abstract");
1141     __ bind(L);
1142   }
1143 #endif // ASSERT
1144 
1145  // generate the code to allocate the interpreter stack frame
1146   generate_fixed_frame(true);
1147 
1148   //
1149   // No locals to initialize for native method
1150   //
1151 
1152   // this slot will be set later, we initialize it to null here just in
1153   // case we get a GC before the actual value is stored later
1154   __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
1155 
1156   const Address do_not_unlock_if_synchronized(G2_thread,
1157     JavaThread::do_not_unlock_if_synchronized_offset());
1158   // Since at this point in the method invocation the exception handler
1159   // would try to exit the monitor of synchronized methods which hasn't
1160   // been entered yet, we set the thread local variable
1161   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1162   // runtime, exception handling i.e. unlock_if_synchronized_method will
1163   // check this thread local flag.
1164   // This flag has two effects, one is to force an unwind in the topmost
1165   // interpreter frame and not perform an unlock while doing so.
1166 
1167   __ movbool(true, G3_scratch);
1168   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1169 
1170   // increment invocation counter and check for overflow
1171   //
1172   // Note: checking for negative value instead of overflow
1173   //       so we have a 'sticky' overflow test (may be of
1174   //       importance as soon as we have true MT/MP)
1175   Label invocation_counter_overflow;
1176   Label Lcontinue;
1177   if (inc_counter) {
1178     generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
1179 
1180   }
1181   __ bind(Lcontinue);
1182 
1183   bang_stack_shadow_pages(true);
1184 
1185   // reset the _do_not_unlock_if_synchronized flag
1186   __ stbool(G0, do_not_unlock_if_synchronized);
1187 
1188   // check for synchronized methods
1189   // Must happen AFTER invocation_counter check and stack overflow check,
1190   // so method is not locked if overflows.
1191 
1192   if (synchronized) {
1193     lock_method();
1194   } else {
1195 #ifdef ASSERT
1196     { Label ok;
1197       __ ld(Laccess_flags, O0);
1198       __ btst(JVM_ACC_SYNCHRONIZED, O0);
1199       __ br( Assembler::zero, false, Assembler::pt, ok);
1200       __ delayed()->nop();
1201       __ stop("method needs synchronization");
1202       __ bind(ok);
1203     }
1204 #endif // ASSERT
1205   }
1206 
1207 
1208   // start execution
1209   __ verify_thread();
1210 
1211   // JVMTI support
1212   __ notify_method_entry();
1213 
1214   // native call
1215 
1216   // (note that O0 is never an oop--at most it is a handle)
1217   // It is important not to smash any handles created by this call,
1218   // until any oop handle in O0 is dereferenced.
1219 
1220   // (note that the space for outgoing params is preallocated)
1221 
1222   // get signature handler
1223   { Label L;
1224     Address signature_handler(Lmethod, Method::signature_handler_offset());
1225     __ ld_ptr(signature_handler, G3_scratch);
1226     __ br_notnull_short(G3_scratch, Assembler::pt, L);
1227     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
1228     __ ld_ptr(signature_handler, G3_scratch);
1229     __ bind(L);
1230   }
1231 
1232   // Push a new frame so that the args will really be stored in
1233   // Copy a few locals across so the new frame has the variables
1234   // we need but these values will be dead at the jni call and
1235   // therefore not gc volatile like the values in the current
1236   // frame (Lmethod in particular)
1237 
1238   // Flush the method pointer to the register save area
1239   __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
1240   __ mov(Llocals, O1);
1241 
1242   // calculate where the mirror handle body is allocated in the interpreter frame:
1243   __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
1244 
1245   // Calculate current frame size
1246   __ sub(SP, FP, O3);         // Calculate negative of current frame size
1247   __ save(SP, O3, SP);        // Allocate an identical sized frame
1248 
1249   // Note I7 has leftover trash. Slow signature handler will fill it in
1250   // should we get there. Normal jni call will set reasonable last_Java_pc
1251   // below (and fix I7 so the stack trace doesn't have a meaningless frame
1252   // in it).
1253 
1254   // Load interpreter frame's Lmethod into same register here
1255 
1256   __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1257 
1258   __ mov(I1, Llocals);
1259   __ mov(I2, Lscratch2);     // save the address of the mirror
1260 
1261 
1262   // ONLY Lmethod and Llocals are valid here!
1263 
1264   // call signature handler, It will move the arg properly since Llocals in current frame
1265   // matches that in outer frame
1266 
1267   __ callr(G3_scratch, 0);
1268   __ delayed()->nop();
1269 
1270   // Result handler is in Lscratch
1271 
1272   // Reload interpreter frame's Lmethod since slow signature handler may block
1273   __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1274 
1275   { Label not_static;
1276 
1277     __ ld(Laccess_flags, O0);
1278     __ btst(JVM_ACC_STATIC, O0);
1279     __ br( Assembler::zero, false, Assembler::pt, not_static);
1280     // get native function entry point(O0 is a good temp until the very end)
1281     __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0);
1282     // for static methods insert the mirror argument
1283     __ load_mirror(O1, Lmethod, G3_scratch);
1284 #ifdef ASSERT
1285     if (!PrintSignatureHandlers)  // do not dirty the output with this
1286     { Label L;
1287       __ br_notnull_short(O1, Assembler::pt, L);
1288       __ stop("mirror is missing");
1289       __ bind(L);
1290     }
1291 #endif // ASSERT
1292     __ st_ptr(O1, Lscratch2, 0);
1293     __ mov(Lscratch2, O1);
1294     __ bind(not_static);
1295   }
1296 
1297   // At this point, arguments have been copied off of stack into
1298   // their JNI positions, which are O1..O5 and SP[68..].
1299   // Oops are boxed in-place on the stack, with handles copied to arguments.
1300   // The result handler is in Lscratch.  O0 will shortly hold the JNIEnv*.
1301 
1302 #ifdef ASSERT
1303   { Label L;
1304     __ br_notnull_short(O0, Assembler::pt, L);
1305     __ stop("native entry point is missing");
1306     __ bind(L);
1307   }
1308 #endif // ASSERT
1309 
1310   //
1311   // setup the frame anchor
1312   //
1313   // The scavenge function only needs to know that the PC of this frame is
1314   // in the interpreter method entry code, it doesn't need to know the exact
1315   // PC and hence we can use O7 which points to the return address from the
1316   // previous call in the code stream (signature handler function)
1317   //
1318   // The other trick is we set last_Java_sp to FP instead of the usual SP because
1319   // we have pushed the extra frame in order to protect the volatile register(s)
1320   // in that frame when we return from the jni call
1321   //
1322 
1323   __ set_last_Java_frame(FP, O7);
1324   __ mov(O7, I7);  // make dummy interpreter frame look like one above,
1325                    // not meaningless information that'll confuse me.
1326 
1327   // flush the windows now. We don't care about the current (protection) frame
1328   // only the outer frames
1329 
1330   __ flushw();
1331 
1332   // mark windows as flushed
1333   Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1334   __ set(JavaFrameAnchor::flushed, G3_scratch);
1335   __ st(G3_scratch, flags);
1336 
1337   // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1338 
1339   Address thread_state(G2_thread, JavaThread::thread_state_offset());
1340 #ifdef ASSERT
1341   { Label L;
1342     __ ld(thread_state, G3_scratch);
1343     __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1344     __ stop("Wrong thread state in native stub");
1345     __ bind(L);
1346   }
1347 #endif // ASSERT
1348   __ set(_thread_in_native, G3_scratch);
1349   __ st(G3_scratch, thread_state);
1350 
1351   // Call the jni method, using the delay slot to set the JNIEnv* argument.
1352   __ save_thread(L7_thread_cache); // save Gthread
1353   __ callr(O0, 0);
1354   __ delayed()->
1355      add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1356 
1357   // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1358 
1359   __ restore_thread(L7_thread_cache); // restore G2_thread
1360   __ reinit_heapbase();
1361 
1362   // must we block?
1363 
1364   // Block, if necessary, before resuming in _thread_in_Java state.
1365   // In order for GC to work, don't clear the last_Java_sp until after blocking.
1366   { Label no_block;
1367 
1368     // Switch thread to "native transition" state before reading the synchronization state.
1369     // This additional state is necessary because reading and testing the synchronization
1370     // state is not atomic w.r.t. GC, as this scenario demonstrates:
1371     //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1372     //     VM thread changes sync state to synchronizing and suspends threads for GC.
1373     //     Thread A is resumed to finish this native method, but doesn't block here since it
1374     //     didn't see any synchronization is progress, and escapes.
1375     __ set(_thread_in_native_trans, G3_scratch);
1376     __ st(G3_scratch, thread_state);
1377     if (os::is_MP()) {
1378       if (UseMembar) {
1379         // Force this write out before the read below
1380         __ membar(Assembler::StoreLoad);
1381       } else {
1382         // Write serialization page so VM thread can do a pseudo remote membar.
1383         // We use the current thread pointer to calculate a thread specific
1384         // offset to write to within the page. This minimizes bus traffic
1385         // due to cache line collision.
1386         __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1387       }
1388     }
1389 
1390     Label L;
1391     __ safepoint_poll(L, false, G2_thread, G3_scratch);
1392     __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1393     __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1394     __ bind(L);
1395 
1396     // Block.  Save any potential method result value before the operation and
1397     // use a leaf call to leave the last_Java_frame setup undisturbed.
1398     save_native_result();
1399     __ call_VM_leaf(L7_thread_cache,
1400                     CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1401                     G2_thread);
1402 
1403     // Restore any method result value
1404     restore_native_result();
1405     __ bind(no_block);
1406   }
1407 
1408   // Clear the frame anchor now
1409 
1410   __ reset_last_Java_frame();
1411 
1412   // Move the result handler address
1413   __ mov(Lscratch, G3_scratch);
1414   // return possible result to the outer frame
1415   __ restore(O0, G0, O0);
1416 
1417   // Move result handler to expected register
1418   __ mov(G3_scratch, Lscratch);
1419 
1420   // Back in normal (native) interpreter frame. State is thread_in_native_trans
1421   // switch to thread_in_Java.
1422 
1423   __ set(_thread_in_Java, G3_scratch);
1424   __ st(G3_scratch, thread_state);
1425 
1426   if (CheckJNICalls) {
1427     // clear_pending_jni_exception_check
1428     __ st_ptr(G0, G2_thread, JavaThread::pending_jni_exception_check_fn_offset());
1429   }
1430 
1431   // reset handle block
1432   __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1433   __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1434 
1435   // If we have an oop result store it where it will be safe for any further gc
1436   // until we return now that we've released the handle it might be protected by
1437 
1438   { Label no_oop;
1439 
1440     __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1441     __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1442     __ resolve_jobject(O0, G3_scratch);
1443     // Store it where gc will look for it and result handler expects it.
1444     __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1445 
1446     __ bind(no_oop);
1447   }
1448 
1449 
1450   // handle exceptions (exception handling will handle unlocking!)
1451   { Label L;
1452     Address exception_addr(G2_thread, Thread::pending_exception_offset());
1453     __ ld_ptr(exception_addr, Gtemp);
1454     __ br_null_short(Gtemp, Assembler::pt, L);
1455     // Note: This could be handled more efficiently since we know that the native
1456     //       method doesn't have an exception handler. We could directly return
1457     //       to the exception handler for the caller.
1458     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1459     __ should_not_reach_here();
1460     __ bind(L);
1461   }
1462 
1463   // JVMTI support (preserves thread register)
1464   __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1465 
1466   if (synchronized) {
1467     // save and restore any potential method result value around the unlocking operation
1468     save_native_result();
1469 
1470     __ add( __ top_most_monitor(), O1);
1471     __ unlock_object(O1);
1472 
1473     restore_native_result();
1474   }
1475 
1476   // dispose of return address and remove activation
1477 #ifdef ASSERT
1478   { Label ok;
1479     __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1480     __ stop("bad I5_savedSP value");
1481     __ should_not_reach_here();
1482     __ bind(ok);
1483   }
1484 #endif
1485   __ jmp(Lscratch, 0);
1486   __ delayed()->nop();
1487 
1488   if (inc_counter) {
1489     // handle invocation counter overflow
1490     __ bind(invocation_counter_overflow);
1491     generate_counter_overflow(Lcontinue);
1492   }
1493 
1494   return entry;
1495 }
1496 
1497 
1498 // Generic method entry to (asm) interpreter
1499 address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) {
1500   address entry = __ pc();
1501 
1502   bool inc_counter  = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1503 
1504   // the following temporary registers are used during frame creation
1505   const Register Gtmp1 = G3_scratch ;
1506   const Register Gtmp2 = G1_scratch;
1507 
1508   // make sure registers are different!
1509   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1510 
1511   const Address constMethod       (G5_method, Method::const_offset());
1512   // Seems like G5_method is live at the point this is used. So we could make this look consistent
1513   // and use in the asserts.
1514   const Address access_flags      (Lmethod,   Method::access_flags_offset());
1515 
1516   const Register Glocals_size = G3;
1517   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1518 
1519   // make sure method is not native & not abstract
1520   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1521 #ifdef ASSERT
1522   __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1523   { Label L;
1524     __ btst(JVM_ACC_NATIVE, Gtmp1);
1525     __ br(Assembler::zero, false, Assembler::pt, L);
1526     __ delayed()->nop();
1527     __ stop("tried to execute native method as non-native");
1528     __ bind(L);
1529   }
1530   { Label L;
1531     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1532     __ br(Assembler::zero, false, Assembler::pt, L);
1533     __ delayed()->nop();
1534     __ stop("tried to execute abstract method as non-abstract");
1535     __ bind(L);
1536   }
1537 #endif // ASSERT
1538 
1539   // generate the code to allocate the interpreter stack frame
1540 
1541   generate_fixed_frame(false);
1542 
1543   //
1544   // Code to initialize the extra (i.e. non-parm) locals
1545   //
1546   Register init_value = noreg;    // will be G0 if we must clear locals
1547   // The way the code was setup before zerolocals was always true for vanilla java entries.
1548   // It could only be false for the specialized entries like accessor or empty which have
1549   // no extra locals so the testing was a waste of time and the extra locals were always
1550   // initialized. We removed this extra complication to already over complicated code.
1551 
1552   init_value = G0;
1553   Label clear_loop;
1554 
1555   const Register RconstMethod = O1;
1556   const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1557   const Address size_of_locals    (RconstMethod, ConstMethod::size_of_locals_offset());
1558 
1559   // NOTE: If you change the frame layout, this code will need to
1560   // be updated!
1561   __ ld_ptr( constMethod, RconstMethod );
1562   __ lduh( size_of_locals, O2 );
1563   __ lduh( size_of_parameters, O1 );
1564   __ sll( O2, Interpreter::logStackElementSize, O2);
1565   __ sll( O1, Interpreter::logStackElementSize, O1 );
1566   __ sub( Llocals, O2, O2 );
1567   __ sub( Llocals, O1, O1 );
1568 
1569   __ bind( clear_loop );
1570   __ inc( O2, wordSize );
1571 
1572   __ cmp( O2, O1 );
1573   __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1574   __ delayed()->st_ptr( init_value, O2, 0 );
1575 
1576   const Address do_not_unlock_if_synchronized(G2_thread,
1577     JavaThread::do_not_unlock_if_synchronized_offset());
1578   // Since at this point in the method invocation the exception handler
1579   // would try to exit the monitor of synchronized methods which hasn't
1580   // been entered yet, we set the thread local variable
1581   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1582   // runtime, exception handling i.e. unlock_if_synchronized_method will
1583   // check this thread local flag.
1584   __ movbool(true, G3_scratch);
1585   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1586 
1587   __ profile_parameters_type(G1_scratch, G3_scratch, G4_scratch, Lscratch);
1588   // increment invocation counter and check for overflow
1589   //
1590   // Note: checking for negative value instead of overflow
1591   //       so we have a 'sticky' overflow test (may be of
1592   //       importance as soon as we have true MT/MP)
1593   Label invocation_counter_overflow;
1594   Label profile_method;
1595   Label profile_method_continue;
1596   Label Lcontinue;
1597   if (inc_counter) {
1598     generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1599     if (ProfileInterpreter) {
1600       __ bind(profile_method_continue);
1601     }
1602   }
1603   __ bind(Lcontinue);
1604 
1605   bang_stack_shadow_pages(false);
1606 
1607   // reset the _do_not_unlock_if_synchronized flag
1608   __ stbool(G0, do_not_unlock_if_synchronized);
1609 
1610   // check for synchronized methods
1611   // Must happen AFTER invocation_counter check and stack overflow check,
1612   // so method is not locked if overflows.
1613 
1614   if (synchronized) {
1615     lock_method();
1616   } else {
1617 #ifdef ASSERT
1618     { Label ok;
1619       __ ld(access_flags, O0);
1620       __ btst(JVM_ACC_SYNCHRONIZED, O0);
1621       __ br( Assembler::zero, false, Assembler::pt, ok);
1622       __ delayed()->nop();
1623       __ stop("method needs synchronization");
1624       __ bind(ok);
1625     }
1626 #endif // ASSERT
1627   }
1628 
1629   // start execution
1630 
1631   __ verify_thread();
1632 
1633   // jvmti support
1634   __ notify_method_entry();
1635 
1636   // start executing instructions
1637   __ dispatch_next(vtos);
1638 
1639 
1640   if (inc_counter) {
1641     if (ProfileInterpreter) {
1642       // We have decided to profile this method in the interpreter
1643       __ bind(profile_method);
1644 
1645       __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1646       __ set_method_data_pointer_for_bcp();
1647       __ ba_short(profile_method_continue);
1648     }
1649 
1650     // handle invocation counter overflow
1651     __ bind(invocation_counter_overflow);
1652     generate_counter_overflow(Lcontinue);
1653   }
1654 
1655   return entry;
1656 }
1657 
1658 //----------------------------------------------------------------------------------------------------
1659 // Exceptions
1660 void TemplateInterpreterGenerator::generate_throw_exception() {
1661 
1662   // Entry point in previous activation (i.e., if the caller was interpreted)
1663   Interpreter::_rethrow_exception_entry = __ pc();
1664   // O0: exception
1665 
1666   // entry point for exceptions thrown within interpreter code
1667   Interpreter::_throw_exception_entry = __ pc();
1668   __ verify_thread();
1669   // expression stack is undefined here
1670   // O0: exception, i.e. Oexception
1671   // Lbcp: exception bcp
1672   __ verify_oop(Oexception);
1673 
1674 
1675   // expression stack must be empty before entering the VM in case of an exception
1676   __ empty_expression_stack();
1677   // find exception handler address and preserve exception oop
1678   // call C routine to find handler and jump to it
1679   __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1680   __ push_ptr(O1); // push exception for exception handler bytecodes
1681 
1682   __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1683   __ delayed()->nop();
1684 
1685 
1686   // if the exception is not handled in the current frame
1687   // the frame is removed and the exception is rethrown
1688   // (i.e. exception continuation is _rethrow_exception)
1689   //
1690   // Note: At this point the bci is still the bxi for the instruction which caused
1691   //       the exception and the expression stack is empty. Thus, for any VM calls
1692   //       at this point, GC will find a legal oop map (with empty expression stack).
1693 
1694   // in current activation
1695   // tos: exception
1696   // Lbcp: exception bcp
1697 
1698   //
1699   // JVMTI PopFrame support
1700   //
1701 
1702   Interpreter::_remove_activation_preserving_args_entry = __ pc();
1703   Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1704   // Set the popframe_processing bit in popframe_condition indicating that we are
1705   // currently handling popframe, so that call_VMs that may happen later do not trigger new
1706   // popframe handling cycles.
1707 
1708   __ ld(popframe_condition_addr, G3_scratch);
1709   __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1710   __ stw(G3_scratch, popframe_condition_addr);
1711 
1712   // Empty the expression stack, as in normal exception handling
1713   __ empty_expression_stack();
1714   __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1715 
1716   {
1717     // Check to see whether we are returning to a deoptimized frame.
1718     // (The PopFrame call ensures that the caller of the popped frame is
1719     // either interpreted or compiled and deoptimizes it if compiled.)
1720     // In this case, we can't call dispatch_next() after the frame is
1721     // popped, but instead must save the incoming arguments and restore
1722     // them after deoptimization has occurred.
1723     //
1724     // Note that we don't compare the return PC against the
1725     // deoptimization blob's unpack entry because of the presence of
1726     // adapter frames in C2.
1727     Label caller_not_deoptimized;
1728     __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1729     __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1730 
1731     const Register Gtmp1 = G3_scratch;
1732     const Register Gtmp2 = G1_scratch;
1733     const Register RconstMethod = Gtmp1;
1734     const Address constMethod(Lmethod, Method::const_offset());
1735     const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1736 
1737     // Compute size of arguments for saving when returning to deoptimized caller
1738     __ ld_ptr(constMethod, RconstMethod);
1739     __ lduh(size_of_parameters, Gtmp1);
1740     __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1741     __ sub(Llocals, Gtmp1, Gtmp2);
1742     __ add(Gtmp2, wordSize, Gtmp2);
1743     // Save these arguments
1744     __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1745     // Inform deoptimization that it is responsible for restoring these arguments
1746     __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1747     Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1748     __ st(Gtmp1, popframe_condition_addr);
1749 
1750     // Return from the current method
1751     // The caller's SP was adjusted upon method entry to accomodate
1752     // the callee's non-argument locals. Undo that adjustment.
1753     __ ret();
1754     __ delayed()->restore(I5_savedSP, G0, SP);
1755 
1756     __ bind(caller_not_deoptimized);
1757   }
1758 
1759   // Clear the popframe condition flag
1760   __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1761 
1762   // Get out of the current method (how this is done depends on the particular compiler calling
1763   // convention that the interpreter currently follows)
1764   // The caller's SP was adjusted upon method entry to accomodate
1765   // the callee's non-argument locals. Undo that adjustment.
1766   __ restore(I5_savedSP, G0, SP);
1767   // The method data pointer was incremented already during
1768   // call profiling. We have to restore the mdp for the current bcp.
1769   if (ProfileInterpreter) {
1770     __ set_method_data_pointer_for_bcp();
1771   }
1772 
1773 #if INCLUDE_JVMTI
1774   { Label L_done;
1775 
1776     __ ldub(Address(Lbcp, 0), G1_scratch);  // Load current bytecode
1777     __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done);
1778 
1779     // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
1780     // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
1781 
1782     __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp);
1783 
1784     __ br_null(G1_scratch, false, Assembler::pn, L_done);
1785     __ delayed()->nop();
1786 
1787     __ st_ptr(G1_scratch, Lesp, wordSize);
1788     __ bind(L_done);
1789   }
1790 #endif // INCLUDE_JVMTI
1791 
1792   // Resume bytecode interpretation at the current bcp
1793   __ dispatch_next(vtos);
1794   // end of JVMTI PopFrame support
1795 
1796   Interpreter::_remove_activation_entry = __ pc();
1797 
1798   // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1799   __ pop_ptr(Oexception);                                  // get exception
1800 
1801   // Intel has the following comment:
1802   //// remove the activation (without doing throws on illegalMonitorExceptions)
1803   // They remove the activation without checking for bad monitor state.
1804   // %%% We should make sure this is the right semantics before implementing.
1805 
1806   __ set_vm_result(Oexception);
1807   __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1808 
1809   __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1810 
1811   __ get_vm_result(Oexception);
1812   __ verify_oop(Oexception);
1813 
1814   const int return_reg_adjustment = frame::pc_return_offset;
1815   Address issuing_pc_addr(I7, return_reg_adjustment);
1816 
1817   // We are done with this activation frame; find out where to go next.
1818   // The continuation point will be an exception handler, which expects
1819   // the following registers set up:
1820   //
1821   // Oexception: exception
1822   // Oissuing_pc: the local call that threw exception
1823   // Other On: garbage
1824   // In/Ln:  the contents of the caller's register window
1825   //
1826   // We do the required restore at the last possible moment, because we
1827   // need to preserve some state across a runtime call.
1828   // (Remember that the caller activation is unknown--it might not be
1829   // interpreted, so things like Lscratch are useless in the caller.)
1830 
1831   // Although the Intel version uses call_C, we can use the more
1832   // compact call_VM.  (The only real difference on SPARC is a
1833   // harmlessly ignored [re]set_last_Java_frame, compared with
1834   // the Intel code which lacks this.)
1835   __ mov(Oexception,      Oexception ->after_save());  // get exception in I0 so it will be on O0 after restore
1836   __ add(issuing_pc_addr, Oissuing_pc->after_save());  // likewise set I1 to a value local to the caller
1837   __ super_call_VM_leaf(L7_thread_cache,
1838                         CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1839                         G2_thread, Oissuing_pc->after_save());
1840 
1841   // The caller's SP was adjusted upon method entry to accomodate
1842   // the callee's non-argument locals. Undo that adjustment.
1843   __ JMP(O0, 0);                         // return exception handler in caller
1844   __ delayed()->restore(I5_savedSP, G0, SP);
1845 
1846   // (same old exception object is already in Oexception; see above)
1847   // Note that an "issuing PC" is actually the next PC after the call
1848 }
1849 
1850 
1851 //
1852 // JVMTI ForceEarlyReturn support
1853 //
1854 
1855 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1856   address entry = __ pc();
1857 
1858   __ empty_expression_stack();
1859   __ load_earlyret_value(state);
1860 
1861   __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1862   Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1863 
1864   // Clear the earlyret state
1865   __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1866 
1867   __ remove_activation(state,
1868                        /* throw_monitor_exception */ false,
1869                        /* install_monitor_exception */ false);
1870 
1871   // The caller's SP was adjusted upon method entry to accomodate
1872   // the callee's non-argument locals. Undo that adjustment.
1873   __ ret();                             // return to caller
1874   __ delayed()->restore(I5_savedSP, G0, SP);
1875 
1876   return entry;
1877 } // end of JVMTI ForceEarlyReturn support
1878 
1879 
1880 //------------------------------------------------------------------------------------------------------------------------
1881 // Helper for vtos entry point generation
1882 
1883 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) {
1884   assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1885   Label L;
1886   aep = __ pc(); __ push_ptr(); __ ba_short(L);
1887   fep = __ pc(); __ push_f();   __ ba_short(L);
1888   dep = __ pc(); __ push_d();   __ ba_short(L);
1889   lep = __ pc(); __ push_l();   __ ba_short(L);
1890   iep = __ pc(); __ push_i();
1891   bep = cep = sep = iep;                        // there aren't any
1892   vep = __ pc(); __ bind(L);                    // fall through
1893   generate_and_dispatch(t);
1894 }
1895 
1896 // --------------------------------------------------------------------------------
1897 
1898 // Non-product code
1899 #ifndef PRODUCT
1900 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
1901   address entry = __ pc();
1902 
1903   __ push(state);
1904   __ mov(O7, Lscratch); // protect return address within interpreter
1905 
1906   // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
1907   __ mov( Otos_l2, G3_scratch );
1908   __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
1909   __ mov(Lscratch, O7); // restore return address
1910   __ pop(state);
1911   __ retl();
1912   __ delayed()->nop();
1913 
1914   return entry;
1915 }
1916 
1917 
1918 // helpers for generate_and_dispatch
1919 
1920 void TemplateInterpreterGenerator::count_bytecode() {
1921   __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
1922 }
1923 
1924 
1925 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
1926   __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
1927 }
1928 
1929 
1930 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
1931   AddressLiteral index   (&BytecodePairHistogram::_index);
1932   AddressLiteral counters((address) &BytecodePairHistogram::_counters);
1933 
1934   // get index, shift out old bytecode, bring in new bytecode, and store it
1935   // _index = (_index >> log2_number_of_codes) |
1936   //          (bytecode << log2_number_of_codes);
1937 
1938   __ load_contents(index, G4_scratch);
1939   __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
1940   __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes,  G3_scratch );
1941   __ or3( G3_scratch,  G4_scratch, G4_scratch );
1942   __ store_contents(G4_scratch, index, G3_scratch);
1943 
1944   // bump bucket contents
1945   // _counters[_index] ++;
1946 
1947   __ set(counters, G3_scratch);                       // loads into G3_scratch
1948   __ sll( G4_scratch, LogBytesPerWord, G4_scratch );  // Index is word address
1949   __ add (G3_scratch, G4_scratch, G3_scratch);        // Add in index
1950   __ ld (G3_scratch, 0, G4_scratch);
1951   __ inc (G4_scratch);
1952   __ st (G4_scratch, 0, G3_scratch);
1953 }
1954 
1955 
1956 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
1957   // Call a little run-time stub to avoid blow-up for each bytecode.
1958   // The run-time runtime saves the right registers, depending on
1959   // the tosca in-state for the given template.
1960   address entry = Interpreter::trace_code(t->tos_in());
1961   guarantee(entry != NULL, "entry must have been generated");
1962   __ call(entry, relocInfo::none);
1963   __ delayed()->nop();
1964 }
1965 
1966 
1967 void TemplateInterpreterGenerator::stop_interpreter_at() {
1968   AddressLiteral counter(&BytecodeCounter::_counter_value);
1969   __ load_contents(counter, G3_scratch);
1970   AddressLiteral stop_at(&StopInterpreterAt);
1971   __ load_ptr_contents(stop_at, G4_scratch);
1972   __ cmp(G3_scratch, G4_scratch);
1973   __ breakpoint_trap(Assembler::equal, Assembler::icc);
1974 }
1975 #endif // not PRODUCT