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