1 /*
   2  * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/macroAssembler.hpp"
  27 #include "interpreter/bytecodeHistogram.hpp"
  28 #include "interpreter/interpreter.hpp"
  29 #include "interpreter/interpreterGenerator.hpp"
  30 #include "interpreter/interpreterRuntime.hpp"
  31 #include "interpreter/interp_masm.hpp"
  32 #include "interpreter/templateTable.hpp"
  33 #include "oops/arrayOop.hpp"
  34 #include "oops/methodData.hpp"
  35 #include "oops/method.hpp"
  36 #include "oops/oop.inline.hpp"
  37 #include "prims/jvmtiExport.hpp"
  38 #include "prims/jvmtiThreadState.hpp"
  39 #include "runtime/arguments.hpp"
  40 #include "runtime/deoptimization.hpp"
  41 #include "runtime/frame.inline.hpp"
  42 #include "runtime/sharedRuntime.hpp"
  43 #include "runtime/stubRoutines.hpp"
  44 #include "runtime/synchronizer.hpp"
  45 #include "runtime/timer.hpp"
  46 #include "runtime/vframeArray.hpp"
  47 #include "utilities/debug.hpp"
  48 #include "utilities/macros.hpp"
  49 
  50 #ifndef CC_INTERP
  51 #ifndef FAST_DISPATCH
  52 #define FAST_DISPATCH 1
  53 #endif
  54 #undef FAST_DISPATCH
  55 
  56 
  57 // Generation of Interpreter
  58 //
  59 // The InterpreterGenerator generates the interpreter into Interpreter::_code.
  60 
  61 
  62 #define __ _masm->
  63 
  64 
  65 //----------------------------------------------------------------------------------------------------
  66 
  67 
  68 void InterpreterGenerator::save_native_result(void) {
  69   // result potentially in O0/O1: save it across calls
  70   const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
  71 
  72   // result potentially in F0/F1: save it across calls
  73   const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
  74 
  75   // save and restore any potential method result value around the unlocking operation
  76   __ stf(FloatRegisterImpl::D, F0, d_tmp);
  77 #ifdef _LP64
  78   __ stx(O0, l_tmp);
  79 #else
  80   __ std(O0, l_tmp);
  81 #endif
  82 }
  83 
  84 void InterpreterGenerator::restore_native_result(void) {
  85   const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
  86   const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
  87 
  88   // Restore any method result value
  89   __ ldf(FloatRegisterImpl::D, d_tmp, F0);
  90 #ifdef _LP64
  91   __ ldx(l_tmp, O0);
  92 #else
  93   __ ldd(l_tmp, O0);
  94 #endif
  95 }
  96 
  97 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
  98   assert(!pass_oop || message == NULL, "either oop or message but not both");
  99   address entry = __ pc();
 100   // expression stack must be empty before entering the VM if an exception happened
 101   __ empty_expression_stack();
 102   // load exception object
 103   __ set((intptr_t)name, G3_scratch);
 104   if (pass_oop) {
 105     __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
 106   } else {
 107     __ set((intptr_t)message, G4_scratch);
 108     __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
 109   }
 110   // throw exception
 111   assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
 112   AddressLiteral thrower(Interpreter::throw_exception_entry());
 113   __ jump_to(thrower, G3_scratch);
 114   __ delayed()->nop();
 115   return entry;
 116 }
 117 
 118 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
 119   address entry = __ pc();
 120   // expression stack must be empty before entering the VM if an exception
 121   // happened
 122   __ empty_expression_stack();
 123   // load exception object
 124   __ call_VM(Oexception,
 125              CAST_FROM_FN_PTR(address,
 126                               InterpreterRuntime::throw_ClassCastException),
 127              Otos_i);
 128   __ should_not_reach_here();
 129   return entry;
 130 }
 131 
 132 
 133 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
 134   address entry = __ pc();
 135   // expression stack must be empty before entering the VM if an exception happened
 136   __ empty_expression_stack();
 137   // convention: expect aberrant index in register G3_scratch, then shuffle the
 138   // index to G4_scratch for the VM call
 139   __ mov(G3_scratch, G4_scratch);
 140   __ set((intptr_t)name, G3_scratch);
 141   __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
 142   __ should_not_reach_here();
 143   return entry;
 144 }
 145 
 146 
 147 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
 148   address entry = __ pc();
 149   // expression stack must be empty before entering the VM if an exception happened
 150   __ empty_expression_stack();
 151   __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
 152   __ should_not_reach_here();
 153   return entry;
 154 }
 155 
 156 
 157 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
 158   address entry = __ pc();
 159 
 160   if (state == atos) {
 161     __ profile_return_type(O0, G3_scratch, G1_scratch);
 162   }
 163 
 164 #if !defined(_LP64) && defined(COMPILER2)
 165   // All return values are where we want them, except for Longs.  C2 returns
 166   // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
 167   // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
 168   // build even if we are returning from interpreted we just do a little
 169   // stupid shuffing.
 170   // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
 171   // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
 172   // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
 173 
 174   if (state == ltos) {
 175     __ srl (G1,  0, O1);
 176     __ srlx(G1, 32, O0);
 177   }
 178 #endif // !_LP64 && COMPILER2
 179 
 180   // The callee returns with the stack possibly adjusted by adapter transition
 181   // We remove that possible adjustment here.
 182   // All interpreter local registers are untouched. Any result is passed back
 183   // in the O0/O1 or float registers. Before continuing, the arguments must be
 184   // popped from the java expression stack; i.e., Lesp must be adjusted.
 185 
 186   __ mov(Llast_SP, SP);   // Remove any adapter added stack space.
 187 
 188   const Register cache = G3_scratch;
 189   const Register index  = G1_scratch;
 190   __ get_cache_and_index_at_bcp(cache, index, 1, index_size);
 191 
 192   const Register flags = cache;
 193   __ ld_ptr(cache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset(), flags);
 194   const Register parameter_size = flags;
 195   __ and3(flags, ConstantPoolCacheEntry::parameter_size_mask, parameter_size);  // argument size in words
 196   __ sll(parameter_size, Interpreter::logStackElementSize, parameter_size);     // each argument size in bytes
 197   __ add(Lesp, parameter_size, Lesp);                                           // pop arguments
 198   __ dispatch_next(state, step);
 199 
 200   return entry;
 201 }
 202 
 203 
 204 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
 205   address entry = __ pc();
 206   __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
 207   { Label L;
 208     Address exception_addr(G2_thread, Thread::pending_exception_offset());
 209     __ ld_ptr(exception_addr, Gtemp);  // Load pending exception.
 210     __ br_null_short(Gtemp, Assembler::pt, L);
 211     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
 212     __ should_not_reach_here();
 213     __ bind(L);
 214   }
 215   __ dispatch_next(state, step);
 216   return entry;
 217 }
 218 
 219 // A result handler converts/unboxes a native call result into
 220 // a java interpreter/compiler result. The current frame is an
 221 // interpreter frame. The activation frame unwind code must be
 222 // consistent with that of TemplateTable::_return(...). In the
 223 // case of native methods, the caller's SP was not modified.
 224 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
 225   address entry = __ pc();
 226   Register Itos_i  = Otos_i ->after_save();
 227   Register Itos_l  = Otos_l ->after_save();
 228   Register Itos_l1 = Otos_l1->after_save();
 229   Register Itos_l2 = Otos_l2->after_save();
 230   switch (type) {
 231     case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
 232     case T_CHAR   : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i);   break; // cannot use and3, 0xFFFF too big as immediate value!
 233     case T_BYTE   : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i);   break;
 234     case T_SHORT  : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i);   break;
 235     case T_LONG   :
 236 #ifndef _LP64
 237                     __ mov(O1, Itos_l2);  // move other half of long
 238 #endif              // ifdef or no ifdef, fall through to the T_INT case
 239     case T_INT    : __ mov(O0, Itos_i);                         break;
 240     case T_VOID   : /* nothing to do */                         break;
 241     case T_FLOAT  : assert(F0 == Ftos_f, "fix this code" );     break;
 242     case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" );     break;
 243     case T_OBJECT :
 244       __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
 245       __ verify_oop(Itos_i);
 246       break;
 247     default       : ShouldNotReachHere();
 248   }
 249   __ ret();                           // return from interpreter activation
 250   __ delayed()->restore(I5_savedSP, G0, SP);  // remove interpreter frame
 251   NOT_PRODUCT(__ emit_int32(0);)       // marker for disassembly
 252   return entry;
 253 }
 254 
 255 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
 256   address entry = __ pc();
 257   __ push(state);
 258   __ call_VM(noreg, runtime_entry);
 259   __ dispatch_via(vtos, Interpreter::normal_table(vtos));
 260   return entry;
 261 }
 262 
 263 
 264 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
 265   address entry = __ pc();
 266   __ dispatch_next(state);
 267   return entry;
 268 }
 269 
 270 //
 271 // Helpers for commoning out cases in the various type of method entries.
 272 //
 273 
 274 // increment invocation count & check for overflow
 275 //
 276 // Note: checking for negative value instead of overflow
 277 //       so we have a 'sticky' overflow test
 278 //
 279 // Lmethod: method
 280 // ??: invocation counter
 281 //
 282 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
 283   // Note: In tiered we increment either counters in MethodCounters* or in
 284   // MDO depending if we're profiling or not.
 285   const Register G3_method_counters = G3_scratch;
 286   Label done;
 287 
 288   if (TieredCompilation) {
 289     const int increment = InvocationCounter::count_increment;
 290     Label no_mdo;
 291     if (ProfileInterpreter) {
 292       // If no method data exists, go to profile_continue.
 293       __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
 294       __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
 295       // Increment counter
 296       Address mdo_invocation_counter(G4_scratch,
 297                                      in_bytes(MethodData::invocation_counter_offset()) +
 298                                      in_bytes(InvocationCounter::counter_offset()));
 299       Address mask(G4_scratch, in_bytes(MethodData::invoke_mask_offset()));
 300       __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
 301                                  G3_scratch, Lscratch,
 302                                  Assembler::zero, overflow);
 303       __ ba_short(done);
 304     }
 305 
 306     // Increment counter in MethodCounters*
 307     __ bind(no_mdo);
 308     Address invocation_counter(G3_method_counters,
 309             in_bytes(MethodCounters::invocation_counter_offset()) +
 310             in_bytes(InvocationCounter::counter_offset()));
 311     __ get_method_counters(Lmethod, G3_method_counters, done);
 312     Address mask(G3_method_counters, in_bytes(MethodCounters::invoke_mask_offset()));
 313     __ increment_mask_and_jump(invocation_counter, increment, mask,
 314                                G4_scratch, Lscratch,
 315                                Assembler::zero, overflow);
 316     __ bind(done);
 317   } else { // not TieredCompilation
 318     // Update standard invocation counters
 319     __ get_method_counters(Lmethod, G3_method_counters, done);
 320     __ increment_invocation_counter(G3_method_counters, O0, G4_scratch);
 321     if (ProfileInterpreter) {
 322       Address interpreter_invocation_counter(G3_method_counters,
 323             in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
 324       __ ld(interpreter_invocation_counter, G4_scratch);
 325       __ inc(G4_scratch);
 326       __ st(G4_scratch, interpreter_invocation_counter);
 327     }
 328 
 329     if (ProfileInterpreter && profile_method != NULL) {
 330       // Test to see if we should create a method data oop
 331       Address profile_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_profile_limit_offset()));
 332       __ ld(profile_limit, G1_scratch);
 333       __ cmp_and_br_short(O0, G1_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
 334 
 335       // if no method data exists, go to profile_method
 336       __ test_method_data_pointer(*profile_method);
 337     }
 338 
 339     Address invocation_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_invocation_limit_offset()));
 340     __ ld(invocation_limit, G3_scratch);
 341     __ cmp(O0, G3_scratch);
 342     __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
 343     __ delayed()->nop();
 344     __ bind(done);
 345   }
 346 
 347 }
 348 
 349 // Allocate monitor and lock method (asm interpreter)
 350 // ebx - Method*
 351 //
 352 void InterpreterGenerator::lock_method(void) {
 353   __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0);  // Load access flags.
 354 
 355 #ifdef ASSERT
 356  { Label ok;
 357    __ btst(JVM_ACC_SYNCHRONIZED, O0);
 358    __ br( Assembler::notZero, false, Assembler::pt, ok);
 359    __ delayed()->nop();
 360    __ stop("method doesn't need synchronization");
 361    __ bind(ok);
 362   }
 363 #endif // ASSERT
 364 
 365   // get synchronization object to O0
 366   { Label done;
 367     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
 368     __ btst(JVM_ACC_STATIC, O0);
 369     __ br( Assembler::zero, true, Assembler::pt, done);
 370     __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
 371 
 372     __ ld_ptr( Lmethod, in_bytes(Method::const_offset()), O0);
 373     __ ld_ptr( O0, in_bytes(ConstMethod::constants_offset()), O0);
 374     __ ld_ptr( O0, ConstantPool::pool_holder_offset_in_bytes(), O0);
 375 
 376     // lock the mirror, not the Klass*
 377     __ ld_ptr( O0, mirror_offset, O0);
 378 
 379 #ifdef ASSERT
 380     __ tst(O0);
 381     __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
 382 #endif // ASSERT
 383 
 384     __ bind(done);
 385   }
 386 
 387   __ add_monitor_to_stack(true, noreg, noreg);  // allocate monitor elem
 388   __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes());   // store object
 389   // __ untested("lock_object from method entry");
 390   __ lock_object(Lmonitors, O0);
 391 }
 392 
 393 
 394 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
 395                                                          Register Rscratch,
 396                                                          Register Rscratch2) {
 397   const int page_size = os::vm_page_size();
 398   Label after_frame_check;
 399 
 400   assert_different_registers(Rframe_size, Rscratch, Rscratch2);
 401 
 402   __ set(page_size, Rscratch);
 403   __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
 404 
 405   // get the stack base, and in debug, verify it is non-zero
 406   __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch );
 407 #ifdef ASSERT
 408   Label base_not_zero;
 409   __ br_notnull_short(Rscratch, Assembler::pn, base_not_zero);
 410   __ stop("stack base is zero in generate_stack_overflow_check");
 411   __ bind(base_not_zero);
 412 #endif
 413 
 414   // get the stack size, and in debug, verify it is non-zero
 415   assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" );
 416   __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 );
 417 #ifdef ASSERT
 418   Label size_not_zero;
 419   __ br_notnull_short(Rscratch2, Assembler::pn, size_not_zero);
 420   __ stop("stack size is zero in generate_stack_overflow_check");
 421   __ bind(size_not_zero);
 422 #endif
 423 
 424   // compute the beginning of the protected zone minus the requested frame size
 425   __ sub( Rscratch, Rscratch2,   Rscratch );
 426   __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 );
 427   __ add( Rscratch, Rscratch2,   Rscratch );
 428 
 429   // Add in the size of the frame (which is the same as subtracting it from the
 430   // SP, which would take another register
 431   __ add( Rscratch, Rframe_size, Rscratch );
 432 
 433   // the frame is greater than one page in size, so check against
 434   // the bottom of the stack
 435   __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check);
 436 
 437   // the stack will overflow, throw an exception
 438 
 439   // Note that SP is restored to sender's sp (in the delay slot). This
 440   // is necessary if the sender's frame is an extended compiled frame
 441   // (see gen_c2i_adapter()) and safer anyway in case of JSR292
 442   // adaptations.
 443 
 444   // Note also that the restored frame is not necessarily interpreted.
 445   // Use the shared runtime version of the StackOverflowError.
 446   assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
 447   AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
 448   __ jump_to(stub, Rscratch);
 449   __ delayed()->mov(O5_savedSP, SP);
 450 
 451   // if you get to here, then there is enough stack space
 452   __ bind( after_frame_check );
 453 }
 454 
 455 
 456 //
 457 // Generate a fixed interpreter frame. This is identical setup for interpreted
 458 // methods and for native methods hence the shared code.
 459 
 460 
 461 //----------------------------------------------------------------------------------------------------
 462 // Stack frame layout
 463 //
 464 // When control flow reaches any of the entry types for the interpreter
 465 // the following holds ->
 466 //
 467 // C2 Calling Conventions:
 468 //
 469 // The entry code below assumes that the following registers are set
 470 // when coming in:
 471 //    G5_method: holds the Method* of the method to call
 472 //    Lesp:    points to the TOS of the callers expression stack
 473 //             after having pushed all the parameters
 474 //
 475 // The entry code does the following to setup an interpreter frame
 476 //   pop parameters from the callers stack by adjusting Lesp
 477 //   set O0 to Lesp
 478 //   compute X = (max_locals - num_parameters)
 479 //   bump SP up by X to accomadate the extra locals
 480 //   compute X = max_expression_stack
 481 //               + vm_local_words
 482 //               + 16 words of register save area
 483 //   save frame doing a save sp, -X, sp growing towards lower addresses
 484 //   set Lbcp, Lmethod, LcpoolCache
 485 //   set Llocals to i0
 486 //   set Lmonitors to FP - rounded_vm_local_words
 487 //   set Lesp to Lmonitors - 4
 488 //
 489 //  The frame has now been setup to do the rest of the entry code
 490 
 491 // Try this optimization:  Most method entries could live in a
 492 // "one size fits all" stack frame without all the dynamic size
 493 // calculations.  It might be profitable to do all this calculation
 494 // statically and approximately for "small enough" methods.
 495 
 496 //-----------------------------------------------------------------------------------------------
 497 
 498 // C1 Calling conventions
 499 //
 500 // Upon method entry, the following registers are setup:
 501 //
 502 // g2 G2_thread: current thread
 503 // g5 G5_method: method to activate
 504 // g4 Gargs  : pointer to last argument
 505 //
 506 //
 507 // Stack:
 508 //
 509 // +---------------+ <--- sp
 510 // |               |
 511 // : reg save area :
 512 // |               |
 513 // +---------------+ <--- sp + 0x40
 514 // |               |
 515 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
 516 // |               |
 517 // +---------------+ <--- sp + 0x5c
 518 // |               |
 519 // :     free      :
 520 // |               |
 521 // +---------------+ <--- Gargs
 522 // |               |
 523 // :   arguments   :
 524 // |               |
 525 // +---------------+
 526 // |               |
 527 //
 528 //
 529 //
 530 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
 531 //
 532 // +---------------+ <--- sp
 533 // |               |
 534 // : reg save area :
 535 // |               |
 536 // +---------------+ <--- sp + 0x40
 537 // |               |
 538 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
 539 // |               |
 540 // +---------------+ <--- sp + 0x5c
 541 // |               |
 542 // :               :
 543 // |               | <--- Lesp
 544 // +---------------+ <--- Lmonitors (fp - 0x18)
 545 // |   VM locals   |
 546 // +---------------+ <--- fp
 547 // |               |
 548 // : reg save area :
 549 // |               |
 550 // +---------------+ <--- fp + 0x40
 551 // |               |
 552 // : extra 7 slots :      note: these slots are not really needed for the interpreter (fix later)
 553 // |               |
 554 // +---------------+ <--- fp + 0x5c
 555 // |               |
 556 // :     free      :
 557 // |               |
 558 // +---------------+
 559 // |               |
 560 // : nonarg locals :
 561 // |               |
 562 // +---------------+
 563 // |               |
 564 // :   arguments   :
 565 // |               | <--- Llocals
 566 // +---------------+ <--- Gargs
 567 // |               |
 568 
 569 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
 570   //
 571   //
 572   // The entry code sets up a new interpreter frame in 4 steps:
 573   //
 574   // 1) Increase caller's SP by for the extra local space needed:
 575   //    (check for overflow)
 576   //    Efficient implementation of xload/xstore bytecodes requires
 577   //    that arguments and non-argument locals are in a contigously
 578   //    addressable memory block => non-argument locals must be
 579   //    allocated in the caller's frame.
 580   //
 581   // 2) Create a new stack frame and register window:
 582   //    The new stack frame must provide space for the standard
 583   //    register save area, the maximum java expression stack size,
 584   //    the monitor slots (0 slots initially), and some frame local
 585   //    scratch locations.
 586   //
 587   // 3) The following interpreter activation registers must be setup:
 588   //    Lesp       : expression stack pointer
 589   //    Lbcp       : bytecode pointer
 590   //    Lmethod    : method
 591   //    Llocals    : locals pointer
 592   //    Lmonitors  : monitor pointer
 593   //    LcpoolCache: constant pool cache
 594   //
 595   // 4) Initialize the non-argument locals if necessary:
 596   //    Non-argument locals may need to be initialized to NULL
 597   //    for GC to work. If the oop-map information is accurate
 598   //    (in the absence of the JSR problem), no initialization
 599   //    is necessary.
 600   //
 601   // (gri - 2/25/2000)
 602 
 603 
 604   int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
 605 
 606   const int extra_space =
 607     rounded_vm_local_words +                   // frame local scratch space
 608     Method::extra_stack_entries() +            // extra stack for jsr 292
 609     frame::memory_parameter_word_sp_offset +   // register save area
 610     (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
 611 
 612   const Register Glocals_size = G3;
 613   const Register RconstMethod = Glocals_size;
 614   const Register Otmp1 = O3;
 615   const Register Otmp2 = O4;
 616   // Lscratch can't be used as a temporary because the call_stub uses
 617   // it to assert that the stack frame was setup correctly.
 618   const Address constMethod       (G5_method, Method::const_offset());
 619   const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
 620 
 621   __ ld_ptr( constMethod, RconstMethod );
 622   __ lduh( size_of_parameters, Glocals_size);
 623 
 624   // Gargs points to first local + BytesPerWord
 625   // Set the saved SP after the register window save
 626   //
 627   assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
 628   __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
 629   __ add(Gargs, Otmp1, Gargs);
 630 
 631   if (native_call) {
 632     __ calc_mem_param_words( Glocals_size, Gframe_size );
 633     __ add( Gframe_size,  extra_space, Gframe_size);
 634     __ round_to( Gframe_size, WordsPerLong );
 635     __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
 636   } else {
 637 
 638     //
 639     // Compute number of locals in method apart from incoming parameters
 640     //
 641     const Address size_of_locals    (Otmp1, ConstMethod::size_of_locals_offset());
 642     __ ld_ptr( constMethod, Otmp1 );
 643     __ lduh( size_of_locals, Otmp1 );
 644     __ sub( Otmp1, Glocals_size, Glocals_size );
 645     __ round_to( Glocals_size, WordsPerLong );
 646     __ sll( Glocals_size, Interpreter::logStackElementSize, Glocals_size );
 647 
 648     // see if the frame is greater than one page in size. If so,
 649     // then we need to verify there is enough stack space remaining
 650     // Frame_size = (max_stack + extra_space) * BytesPerWord;
 651     __ ld_ptr( constMethod, Gframe_size );
 652     __ lduh( Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size );
 653     __ add( Gframe_size, extra_space, Gframe_size );
 654     __ round_to( Gframe_size, WordsPerLong );
 655     __ sll( Gframe_size, Interpreter::logStackElementSize, Gframe_size);
 656 
 657     // Add in java locals size for stack overflow check only
 658     __ add( Gframe_size, Glocals_size, Gframe_size );
 659 
 660     const Register Otmp2 = O4;
 661     assert_different_registers(Otmp1, Otmp2, O5_savedSP);
 662     generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2);
 663 
 664     __ sub( Gframe_size, Glocals_size, Gframe_size);
 665 
 666     //
 667     // bump SP to accomodate the extra locals
 668     //
 669     __ sub( SP, Glocals_size, SP );
 670   }
 671 
 672   //
 673   // now set up a stack frame with the size computed above
 674   //
 675   __ neg( Gframe_size );
 676   __ save( SP, Gframe_size, SP );
 677 
 678   //
 679   // now set up all the local cache registers
 680   //
 681   // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
 682   // that all present references to Lbyte_code initialize the register
 683   // immediately before use
 684   if (native_call) {
 685     __ mov(G0, Lbcp);
 686   } else {
 687     __ ld_ptr(G5_method, Method::const_offset(), Lbcp);
 688     __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp);
 689   }
 690   __ mov( G5_method, Lmethod);                 // set Lmethod
 691   __ get_constant_pool_cache( LcpoolCache );   // set LcpoolCache
 692   __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
 693 #ifdef _LP64
 694   __ add( Lmonitors, STACK_BIAS, Lmonitors );   // Account for 64 bit stack bias
 695 #endif
 696   __ sub(Lmonitors, BytesPerWord, Lesp);       // set Lesp
 697 
 698   // setup interpreter activation registers
 699   __ sub(Gargs, BytesPerWord, Llocals);        // set Llocals
 700 
 701   if (ProfileInterpreter) {
 702 #ifdef FAST_DISPATCH
 703     // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
 704     // they both use I2.
 705     assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
 706 #endif // FAST_DISPATCH
 707     __ set_method_data_pointer();
 708   }
 709 
 710 }
 711 
 712 // Method entry for java.lang.ref.Reference.get.
 713 address InterpreterGenerator::generate_Reference_get_entry(void) {
 714 #if INCLUDE_ALL_GCS
 715   // Code: _aload_0, _getfield, _areturn
 716   // parameter size = 1
 717   //
 718   // The code that gets generated by this routine is split into 2 parts:
 719   //    1. The "intrinsified" code for G1 (or any SATB based GC),
 720   //    2. The slow path - which is an expansion of the regular method entry.
 721   //
 722   // Notes:-
 723   // * In the G1 code we do not check whether we need to block for
 724   //   a safepoint. If G1 is enabled then we must execute the specialized
 725   //   code for Reference.get (except when the Reference object is null)
 726   //   so that we can log the value in the referent field with an SATB
 727   //   update buffer.
 728   //   If the code for the getfield template is modified so that the
 729   //   G1 pre-barrier code is executed when the current method is
 730   //   Reference.get() then going through the normal method entry
 731   //   will be fine.
 732   // * The G1 code can, however, check the receiver object (the instance
 733   //   of java.lang.Reference) and jump to the slow path if null. If the
 734   //   Reference object is null then we obviously cannot fetch the referent
 735   //   and so we don't need to call the G1 pre-barrier. Thus we can use the
 736   //   regular method entry code to generate the NPE.
 737   //
 738   // This code is based on generate_accessor_enty.
 739 
 740   address entry = __ pc();
 741 
 742   const int referent_offset = java_lang_ref_Reference::referent_offset;
 743   guarantee(referent_offset > 0, "referent offset not initialized");
 744 
 745   if (UseG1GC) {
 746      Label slow_path;
 747 
 748     // In the G1 code we don't check if we need to reach a safepoint. We
 749     // continue and the thread will safepoint at the next bytecode dispatch.
 750 
 751     // Check if local 0 != NULL
 752     // If the receiver is null then it is OK to jump to the slow path.
 753     __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
 754     // check if local 0 == NULL and go the slow path
 755     __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
 756 
 757 
 758     // Load the value of the referent field.
 759     if (Assembler::is_simm13(referent_offset)) {
 760       __ load_heap_oop(Otos_i, referent_offset, Otos_i);
 761     } else {
 762       __ set(referent_offset, G3_scratch);
 763       __ load_heap_oop(Otos_i, G3_scratch, Otos_i);
 764     }
 765 
 766     // Generate the G1 pre-barrier code to log the value of
 767     // the referent field in an SATB buffer. Note with
 768     // these parameters the pre-barrier does not generate
 769     // the load of the previous value
 770 
 771     __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */,
 772                             Otos_i /* pre_val */,
 773                             G3_scratch /* tmp */,
 774                             true /* preserve_o_regs */);
 775 
 776     // _areturn
 777     __ retl();                      // return from leaf routine
 778     __ delayed()->mov(O5_savedSP, SP);
 779 
 780     // Generate regular method entry
 781     __ bind(slow_path);
 782     __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals));
 783     return entry;
 784   }
 785 #endif // INCLUDE_ALL_GCS
 786 
 787   // If G1 is not enabled then attempt to go through the accessor entry point
 788   // Reference.get is an accessor
 789   return NULL;
 790 }
 791 
 792 //
 793 // Interpreter stub for calling a native method. (asm interpreter)
 794 // This sets up a somewhat different looking stack for calling the native method
 795 // than the typical interpreter frame setup.
 796 //
 797 
 798 address InterpreterGenerator::generate_native_entry(bool synchronized) {
 799   address entry = __ pc();
 800 
 801   // the following temporary registers are used during frame creation
 802   const Register Gtmp1 = G3_scratch ;
 803   const Register Gtmp2 = G1_scratch;
 804   bool inc_counter  = UseCompiler || CountCompiledCalls || LogTouchedMethods;
 805 
 806   // make sure registers are different!
 807   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
 808 
 809   const Address Laccess_flags(Lmethod, Method::access_flags_offset());
 810 
 811   const Register Glocals_size = G3;
 812   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
 813 
 814   // make sure method is native & not abstract
 815   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
 816 #ifdef ASSERT
 817   __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
 818   {
 819     Label L;
 820     __ btst(JVM_ACC_NATIVE, Gtmp1);
 821     __ br(Assembler::notZero, false, Assembler::pt, L);
 822     __ delayed()->nop();
 823     __ stop("tried to execute non-native method as native");
 824     __ bind(L);
 825   }
 826   { Label L;
 827     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
 828     __ br(Assembler::zero, false, Assembler::pt, L);
 829     __ delayed()->nop();
 830     __ stop("tried to execute abstract method as non-abstract");
 831     __ bind(L);
 832   }
 833 #endif // ASSERT
 834 
 835  // generate the code to allocate the interpreter stack frame
 836   generate_fixed_frame(true);
 837 
 838   //
 839   // No locals to initialize for native method
 840   //
 841 
 842   // this slot will be set later, we initialize it to null here just in
 843   // case we get a GC before the actual value is stored later
 844   __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
 845 
 846   const Address do_not_unlock_if_synchronized(G2_thread,
 847     JavaThread::do_not_unlock_if_synchronized_offset());
 848   // Since at this point in the method invocation the exception handler
 849   // would try to exit the monitor of synchronized methods which hasn't
 850   // been entered yet, we set the thread local variable
 851   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
 852   // runtime, exception handling i.e. unlock_if_synchronized_method will
 853   // check this thread local flag.
 854   // This flag has two effects, one is to force an unwind in the topmost
 855   // interpreter frame and not perform an unlock while doing so.
 856 
 857   __ movbool(true, G3_scratch);
 858   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
 859 
 860   // increment invocation counter and check for overflow
 861   //
 862   // Note: checking for negative value instead of overflow
 863   //       so we have a 'sticky' overflow test (may be of
 864   //       importance as soon as we have true MT/MP)
 865   Label invocation_counter_overflow;
 866   Label Lcontinue;
 867   if (inc_counter) {
 868     generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
 869 
 870   }
 871   __ bind(Lcontinue);
 872 
 873   bang_stack_shadow_pages(true);
 874 
 875   // reset the _do_not_unlock_if_synchronized flag
 876   __ stbool(G0, do_not_unlock_if_synchronized);
 877 
 878   // check for synchronized methods
 879   // Must happen AFTER invocation_counter check and stack overflow check,
 880   // so method is not locked if overflows.
 881 
 882   if (synchronized) {
 883     lock_method();
 884   } else {
 885 #ifdef ASSERT
 886     { Label ok;
 887       __ ld(Laccess_flags, O0);
 888       __ btst(JVM_ACC_SYNCHRONIZED, O0);
 889       __ br( Assembler::zero, false, Assembler::pt, ok);
 890       __ delayed()->nop();
 891       __ stop("method needs synchronization");
 892       __ bind(ok);
 893     }
 894 #endif // ASSERT
 895   }
 896 
 897 
 898   // start execution
 899   __ verify_thread();
 900 
 901   // JVMTI support
 902   __ notify_method_entry();
 903 
 904   // native call
 905 
 906   // (note that O0 is never an oop--at most it is a handle)
 907   // It is important not to smash any handles created by this call,
 908   // until any oop handle in O0 is dereferenced.
 909 
 910   // (note that the space for outgoing params is preallocated)
 911 
 912   // get signature handler
 913   { Label L;
 914     Address signature_handler(Lmethod, Method::signature_handler_offset());
 915     __ ld_ptr(signature_handler, G3_scratch);
 916     __ br_notnull_short(G3_scratch, Assembler::pt, L);
 917     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
 918     __ ld_ptr(signature_handler, G3_scratch);
 919     __ bind(L);
 920   }
 921 
 922   // Push a new frame so that the args will really be stored in
 923   // Copy a few locals across so the new frame has the variables
 924   // we need but these values will be dead at the jni call and
 925   // therefore not gc volatile like the values in the current
 926   // frame (Lmethod in particular)
 927 
 928   // Flush the method pointer to the register save area
 929   __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
 930   __ mov(Llocals, O1);
 931 
 932   // calculate where the mirror handle body is allocated in the interpreter frame:
 933   __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
 934 
 935   // Calculate current frame size
 936   __ sub(SP, FP, O3);         // Calculate negative of current frame size
 937   __ save(SP, O3, SP);        // Allocate an identical sized frame
 938 
 939   // Note I7 has leftover trash. Slow signature handler will fill it in
 940   // should we get there. Normal jni call will set reasonable last_Java_pc
 941   // below (and fix I7 so the stack trace doesn't have a meaningless frame
 942   // in it).
 943 
 944   // Load interpreter frame's Lmethod into same register here
 945 
 946   __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
 947 
 948   __ mov(I1, Llocals);
 949   __ mov(I2, Lscratch2);     // save the address of the mirror
 950 
 951 
 952   // ONLY Lmethod and Llocals are valid here!
 953 
 954   // call signature handler, It will move the arg properly since Llocals in current frame
 955   // matches that in outer frame
 956 
 957   __ callr(G3_scratch, 0);
 958   __ delayed()->nop();
 959 
 960   // Result handler is in Lscratch
 961 
 962   // Reload interpreter frame's Lmethod since slow signature handler may block
 963   __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
 964 
 965   { Label not_static;
 966 
 967     __ ld(Laccess_flags, O0);
 968     __ btst(JVM_ACC_STATIC, O0);
 969     __ br( Assembler::zero, false, Assembler::pt, not_static);
 970     // get native function entry point(O0 is a good temp until the very end)
 971     __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0);
 972     // for static methods insert the mirror argument
 973     const int mirror_offset = in_bytes(Klass::java_mirror_offset());
 974 
 975     __ ld_ptr(Lmethod, Method:: const_offset(), O1);
 976     __ ld_ptr(O1, ConstMethod::constants_offset(), O1);
 977     __ ld_ptr(O1, ConstantPool::pool_holder_offset_in_bytes(), O1);
 978     __ ld_ptr(O1, mirror_offset, O1);
 979 #ifdef ASSERT
 980     if (!PrintSignatureHandlers)  // do not dirty the output with this
 981     { Label L;
 982       __ br_notnull_short(O1, Assembler::pt, L);
 983       __ stop("mirror is missing");
 984       __ bind(L);
 985     }
 986 #endif // ASSERT
 987     __ st_ptr(O1, Lscratch2, 0);
 988     __ mov(Lscratch2, O1);
 989     __ bind(not_static);
 990   }
 991 
 992   // At this point, arguments have been copied off of stack into
 993   // their JNI positions, which are O1..O5 and SP[68..].
 994   // Oops are boxed in-place on the stack, with handles copied to arguments.
 995   // The result handler is in Lscratch.  O0 will shortly hold the JNIEnv*.
 996 
 997 #ifdef ASSERT
 998   { Label L;
 999     __ br_notnull_short(O0, Assembler::pt, L);
1000     __ stop("native entry point is missing");
1001     __ bind(L);
1002   }
1003 #endif // ASSERT
1004 
1005   //
1006   // setup the frame anchor
1007   //
1008   // The scavenge function only needs to know that the PC of this frame is
1009   // in the interpreter method entry code, it doesn't need to know the exact
1010   // PC and hence we can use O7 which points to the return address from the
1011   // previous call in the code stream (signature handler function)
1012   //
1013   // The other trick is we set last_Java_sp to FP instead of the usual SP because
1014   // we have pushed the extra frame in order to protect the volatile register(s)
1015   // in that frame when we return from the jni call
1016   //
1017 
1018   __ set_last_Java_frame(FP, O7);
1019   __ mov(O7, I7);  // make dummy interpreter frame look like one above,
1020                    // not meaningless information that'll confuse me.
1021 
1022   // flush the windows now. We don't care about the current (protection) frame
1023   // only the outer frames
1024 
1025   __ flushw();
1026 
1027   // mark windows as flushed
1028   Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1029   __ set(JavaFrameAnchor::flushed, G3_scratch);
1030   __ st(G3_scratch, flags);
1031 
1032   // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1033 
1034   Address thread_state(G2_thread, JavaThread::thread_state_offset());
1035 #ifdef ASSERT
1036   { Label L;
1037     __ ld(thread_state, G3_scratch);
1038     __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1039     __ stop("Wrong thread state in native stub");
1040     __ bind(L);
1041   }
1042 #endif // ASSERT
1043   __ set(_thread_in_native, G3_scratch);
1044   __ st(G3_scratch, thread_state);
1045 
1046   // Call the jni method, using the delay slot to set the JNIEnv* argument.
1047   __ save_thread(L7_thread_cache); // save Gthread
1048   __ callr(O0, 0);
1049   __ delayed()->
1050      add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1051 
1052   // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1053 
1054   __ restore_thread(L7_thread_cache); // restore G2_thread
1055   __ reinit_heapbase();
1056 
1057   // must we block?
1058 
1059   // Block, if necessary, before resuming in _thread_in_Java state.
1060   // In order for GC to work, don't clear the last_Java_sp until after blocking.
1061   { Label no_block;
1062     AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1063 
1064     // Switch thread to "native transition" state before reading the synchronization state.
1065     // This additional state is necessary because reading and testing the synchronization
1066     // state is not atomic w.r.t. GC, as this scenario demonstrates:
1067     //     Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1068     //     VM thread changes sync state to synchronizing and suspends threads for GC.
1069     //     Thread A is resumed to finish this native method, but doesn't block here since it
1070     //     didn't see any synchronization is progress, and escapes.
1071     __ set(_thread_in_native_trans, G3_scratch);
1072     __ st(G3_scratch, thread_state);
1073     if(os::is_MP()) {
1074       if (UseMembar) {
1075         // Force this write out before the read below
1076         __ membar(Assembler::StoreLoad);
1077       } else {
1078         // Write serialization page so VM thread can do a pseudo remote membar.
1079         // We use the current thread pointer to calculate a thread specific
1080         // offset to write to within the page. This minimizes bus traffic
1081         // due to cache line collision.
1082         __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1083       }
1084     }
1085     __ load_contents(sync_state, G3_scratch);
1086     __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1087 
1088     Label L;
1089     __ br(Assembler::notEqual, false, Assembler::pn, L);
1090     __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1091     __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1092     __ bind(L);
1093 
1094     // Block.  Save any potential method result value before the operation and
1095     // use a leaf call to leave the last_Java_frame setup undisturbed.
1096     save_native_result();
1097     __ call_VM_leaf(L7_thread_cache,
1098                     CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1099                     G2_thread);
1100 
1101     // Restore any method result value
1102     restore_native_result();
1103     __ bind(no_block);
1104   }
1105 
1106   // Clear the frame anchor now
1107 
1108   __ reset_last_Java_frame();
1109 
1110   // Move the result handler address
1111   __ mov(Lscratch, G3_scratch);
1112   // return possible result to the outer frame
1113 #ifndef __LP64
1114   __ mov(O0, I0);
1115   __ restore(O1, G0, O1);
1116 #else
1117   __ restore(O0, G0, O0);
1118 #endif /* __LP64 */
1119 
1120   // Move result handler to expected register
1121   __ mov(G3_scratch, Lscratch);
1122 
1123   // Back in normal (native) interpreter frame. State is thread_in_native_trans
1124   // switch to thread_in_Java.
1125 
1126   __ set(_thread_in_Java, G3_scratch);
1127   __ st(G3_scratch, thread_state);
1128 
1129   // reset handle block
1130   __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1131   __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1132 
1133   // If we have an oop result store it where it will be safe for any further gc
1134   // until we return now that we've released the handle it might be protected by
1135 
1136   {
1137     Label no_oop, store_result;
1138 
1139     __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1140     __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1141     __ addcc(G0, O0, O0);
1142     __ brx(Assembler::notZero, true, Assembler::pt, store_result);     // if result is not NULL:
1143     __ delayed()->ld_ptr(O0, 0, O0);                                   // unbox it
1144     __ mov(G0, O0);
1145 
1146     __ bind(store_result);
1147     // Store it where gc will look for it and result handler expects it.
1148     __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1149 
1150     __ bind(no_oop);
1151 
1152   }
1153 
1154 
1155   // handle exceptions (exception handling will handle unlocking!)
1156   { Label L;
1157     Address exception_addr(G2_thread, Thread::pending_exception_offset());
1158     __ ld_ptr(exception_addr, Gtemp);
1159     __ br_null_short(Gtemp, Assembler::pt, L);
1160     // Note: This could be handled more efficiently since we know that the native
1161     //       method doesn't have an exception handler. We could directly return
1162     //       to the exception handler for the caller.
1163     __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1164     __ should_not_reach_here();
1165     __ bind(L);
1166   }
1167 
1168   // JVMTI support (preserves thread register)
1169   __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1170 
1171   if (synchronized) {
1172     // save and restore any potential method result value around the unlocking operation
1173     save_native_result();
1174 
1175     __ add( __ top_most_monitor(), O1);
1176     __ unlock_object(O1);
1177 
1178     restore_native_result();
1179   }
1180 
1181 #if defined(COMPILER2) && !defined(_LP64)
1182 
1183   // C2 expects long results in G1 we can't tell if we're returning to interpreted
1184   // or compiled so just be safe.
1185 
1186   __ sllx(O0, 32, G1);          // Shift bits into high G1
1187   __ srl (O1, 0, O1);           // Zero extend O1
1188   __ or3 (O1, G1, G1);          // OR 64 bits into G1
1189 
1190 #endif /* COMPILER2 && !_LP64 */
1191 
1192   // dispose of return address and remove activation
1193 #ifdef ASSERT
1194   {
1195     Label ok;
1196     __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1197     __ stop("bad I5_savedSP value");
1198     __ should_not_reach_here();
1199     __ bind(ok);
1200   }
1201 #endif
1202   if (TraceJumps) {
1203     // Move target to register that is recordable
1204     __ mov(Lscratch, G3_scratch);
1205     __ JMP(G3_scratch, 0);
1206   } else {
1207     __ jmp(Lscratch, 0);
1208   }
1209   __ delayed()->nop();
1210 
1211 
1212   if (inc_counter) {
1213     // handle invocation counter overflow
1214     __ bind(invocation_counter_overflow);
1215     generate_counter_overflow(Lcontinue);
1216   }
1217 
1218 
1219 
1220   return entry;
1221 }
1222 
1223 
1224 // Generic method entry to (asm) interpreter
1225 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1226   address entry = __ pc();
1227 
1228   bool inc_counter  = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1229 
1230   // the following temporary registers are used during frame creation
1231   const Register Gtmp1 = G3_scratch ;
1232   const Register Gtmp2 = G1_scratch;
1233 
1234   // make sure registers are different!
1235   assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1236 
1237   const Address constMethod       (G5_method, Method::const_offset());
1238   // Seems like G5_method is live at the point this is used. So we could make this look consistent
1239   // and use in the asserts.
1240   const Address access_flags      (Lmethod,   Method::access_flags_offset());
1241 
1242   const Register Glocals_size = G3;
1243   assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1244 
1245   // make sure method is not native & not abstract
1246   // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1247 #ifdef ASSERT
1248   __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1249   {
1250     Label L;
1251     __ btst(JVM_ACC_NATIVE, Gtmp1);
1252     __ br(Assembler::zero, false, Assembler::pt, L);
1253     __ delayed()->nop();
1254     __ stop("tried to execute native method as non-native");
1255     __ bind(L);
1256   }
1257   { Label L;
1258     __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1259     __ br(Assembler::zero, false, Assembler::pt, L);
1260     __ delayed()->nop();
1261     __ stop("tried to execute abstract method as non-abstract");
1262     __ bind(L);
1263   }
1264 #endif // ASSERT
1265 
1266   // generate the code to allocate the interpreter stack frame
1267 
1268   generate_fixed_frame(false);
1269 
1270 #ifdef FAST_DISPATCH
1271   __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1272                                           // set bytecode dispatch table base
1273 #endif
1274 
1275   //
1276   // Code to initialize the extra (i.e. non-parm) locals
1277   //
1278   Register init_value = noreg;    // will be G0 if we must clear locals
1279   // The way the code was setup before zerolocals was always true for vanilla java entries.
1280   // It could only be false for the specialized entries like accessor or empty which have
1281   // no extra locals so the testing was a waste of time and the extra locals were always
1282   // initialized. We removed this extra complication to already over complicated code.
1283 
1284   init_value = G0;
1285   Label clear_loop;
1286 
1287   const Register RconstMethod = O1;
1288   const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1289   const Address size_of_locals    (RconstMethod, ConstMethod::size_of_locals_offset());
1290 
1291   // NOTE: If you change the frame layout, this code will need to
1292   // be updated!
1293   __ ld_ptr( constMethod, RconstMethod );
1294   __ lduh( size_of_locals, O2 );
1295   __ lduh( size_of_parameters, O1 );
1296   __ sll( O2, Interpreter::logStackElementSize, O2);
1297   __ sll( O1, Interpreter::logStackElementSize, O1 );
1298   __ sub( Llocals, O2, O2 );
1299   __ sub( Llocals, O1, O1 );
1300 
1301   __ bind( clear_loop );
1302   __ inc( O2, wordSize );
1303 
1304   __ cmp( O2, O1 );
1305   __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1306   __ delayed()->st_ptr( init_value, O2, 0 );
1307 
1308   const Address do_not_unlock_if_synchronized(G2_thread,
1309     JavaThread::do_not_unlock_if_synchronized_offset());
1310   // Since at this point in the method invocation the exception handler
1311   // would try to exit the monitor of synchronized methods which hasn't
1312   // been entered yet, we set the thread local variable
1313   // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1314   // runtime, exception handling i.e. unlock_if_synchronized_method will
1315   // check this thread local flag.
1316   __ movbool(true, G3_scratch);
1317   __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1318 
1319   __ profile_parameters_type(G1_scratch, G3_scratch, G4_scratch, Lscratch);
1320   // increment invocation counter and check for overflow
1321   //
1322   // Note: checking for negative value instead of overflow
1323   //       so we have a 'sticky' overflow test (may be of
1324   //       importance as soon as we have true MT/MP)
1325   Label invocation_counter_overflow;
1326   Label profile_method;
1327   Label profile_method_continue;
1328   Label Lcontinue;
1329   if (inc_counter) {
1330     generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1331     if (ProfileInterpreter) {
1332       __ bind(profile_method_continue);
1333     }
1334   }
1335   __ bind(Lcontinue);
1336 
1337   bang_stack_shadow_pages(false);
1338 
1339   // reset the _do_not_unlock_if_synchronized flag
1340   __ stbool(G0, do_not_unlock_if_synchronized);
1341 
1342   // check for synchronized methods
1343   // Must happen AFTER invocation_counter check and stack overflow check,
1344   // so method is not locked if overflows.
1345 
1346   if (synchronized) {
1347     lock_method();
1348   } else {
1349 #ifdef ASSERT
1350     { Label ok;
1351       __ ld(access_flags, O0);
1352       __ btst(JVM_ACC_SYNCHRONIZED, O0);
1353       __ br( Assembler::zero, false, Assembler::pt, ok);
1354       __ delayed()->nop();
1355       __ stop("method needs synchronization");
1356       __ bind(ok);
1357     }
1358 #endif // ASSERT
1359   }
1360 
1361   // start execution
1362 
1363   __ verify_thread();
1364 
1365   // jvmti support
1366   __ notify_method_entry();
1367 
1368   // start executing instructions
1369   __ dispatch_next(vtos);
1370 
1371 
1372   if (inc_counter) {
1373     if (ProfileInterpreter) {
1374       // We have decided to profile this method in the interpreter
1375       __ bind(profile_method);
1376 
1377       __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1378       __ set_method_data_pointer_for_bcp();
1379       __ ba_short(profile_method_continue);
1380     }
1381 
1382     // handle invocation counter overflow
1383     __ bind(invocation_counter_overflow);
1384     generate_counter_overflow(Lcontinue);
1385   }
1386 
1387 
1388   return entry;
1389 }
1390 
1391 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
1392 
1393   // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
1394   // expression stack, the callee will have callee_extra_locals (so we can account for
1395   // frame extension) and monitor_size for monitors. Basically we need to calculate
1396   // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
1397   //
1398   //
1399   // The big complicating thing here is that we must ensure that the stack stays properly
1400   // aligned. This would be even uglier if monitor size wasn't modulo what the stack
1401   // needs to be aligned for). We are given that the sp (fp) is already aligned by
1402   // the caller so we must ensure that it is properly aligned for our callee.
1403   //
1404   const int rounded_vm_local_words =
1405        round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1406   // callee_locals and max_stack are counts, not the size in frame.
1407   const int locals_size =
1408        round_to(callee_extra_locals * Interpreter::stackElementWords, WordsPerLong);
1409   const int max_stack_words = max_stack * Interpreter::stackElementWords;
1410   return (round_to((max_stack_words
1411                    + rounded_vm_local_words
1412                    + frame::memory_parameter_word_sp_offset), WordsPerLong)
1413                    // already rounded
1414                    + locals_size + monitor_size);
1415 }
1416 
1417 // How much stack a method top interpreter activation needs in words.
1418 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
1419 
1420   // See call_stub code
1421   int call_stub_size  = round_to(7 + frame::memory_parameter_word_sp_offset,
1422                                  WordsPerLong);    // 7 + register save area
1423 
1424   // Save space for one monitor to get into the interpreted method in case
1425   // the method is synchronized
1426   int monitor_size    = method->is_synchronized() ?
1427                                 1*frame::interpreter_frame_monitor_size() : 0;
1428   return size_activation_helper(method->max_locals(), method->max_stack(),
1429                                 monitor_size) + call_stub_size;
1430 }
1431 
1432 int AbstractInterpreter::size_activation(int max_stack,
1433                                          int temps,
1434                                          int extra_args,
1435                                          int monitors,
1436                                          int callee_params,
1437                                          int callee_locals,
1438                                          bool is_top_frame) {
1439   // Note: This calculation must exactly parallel the frame setup
1440   // in InterpreterGenerator::generate_fixed_frame.
1441 
1442   int monitor_size           = monitors * frame::interpreter_frame_monitor_size();
1443 
1444   assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1445 
1446   //
1447   // Note: if you look closely this appears to be doing something much different
1448   // than generate_fixed_frame. What is happening is this. On sparc we have to do
1449   // this dance with interpreter_sp_adjustment because the window save area would
1450   // appear just below the bottom (tos) of the caller's java expression stack. Because
1451   // the interpreter want to have the locals completely contiguous generate_fixed_frame
1452   // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size).
1453   // Now in generate_fixed_frame the extension of the caller's sp happens in the callee.
1454   // In this code the opposite occurs the caller adjusts it's own stack base on the callee.
1455   // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest)
1456   // because the oldest frame would have adjust its callers frame and yet that frame
1457   // already exists and isn't part of this array of frames we are unpacking. So at first
1458   // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper()
1459   // will after it calculates all of the frame's on_stack_size()'s will then figure out the
1460   // amount to adjust the caller of the initial (oldest) frame and the calculation will all
1461   // add up. It does seem like it simpler to account for the adjustment here (and remove the
1462   // callee... parameters here). However this would mean that this routine would have to take
1463   // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment)
1464   // and run the calling loop in the reverse order. This would also would appear to mean making
1465   // this code aware of what the interactions are when that initial caller fram was an osr or
1466   // other adapter frame. deoptimization is complicated enough and  hard enough to debug that
1467   // there is no sense in messing working code.
1468   //
1469 
1470   int rounded_cls = round_to((callee_locals - callee_params), WordsPerLong);
1471   assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align");
1472 
1473   int raw_frame_size = size_activation_helper(rounded_cls, max_stack, monitor_size);
1474 
1475   return raw_frame_size;
1476 }
1477 
1478 void AbstractInterpreter::layout_activation(Method* method,
1479                                             int tempcount,
1480                                             int popframe_extra_args,
1481                                             int moncount,
1482                                             int caller_actual_parameters,
1483                                             int callee_param_count,
1484                                             int callee_local_count,
1485                                             frame* caller,
1486                                             frame* interpreter_frame,
1487                                             bool is_top_frame,
1488                                             bool is_bottom_frame) {
1489   // Set up the following variables:
1490   //   - Lmethod
1491   //   - Llocals
1492   //   - Lmonitors (to the indicated number of monitors)
1493   //   - Lesp (to the indicated number of temps)
1494   // The frame caller on entry is a description of the caller of the
1495   // frame we are about to layout. We are guaranteed that we will be
1496   // able to fill in a new interpreter frame as its callee (i.e. the
1497   // stack space is allocated and the amount was determined by an
1498   // earlier call to the size_activation() method).  On return caller
1499   // while describe the interpreter frame we just layed out.
1500 
1501   // The skeleton frame must already look like an interpreter frame
1502   // even if not fully filled out.
1503   assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame");
1504 
1505   int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1506   int monitor_size           = moncount * frame::interpreter_frame_monitor_size();
1507   assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1508 
1509   intptr_t* fp = interpreter_frame->fp();
1510 
1511   JavaThread* thread = JavaThread::current();
1512   RegisterMap map(thread, false);
1513   // More verification that skeleton frame is properly walkable
1514   assert(fp == caller->sp(), "fp must match");
1515 
1516   intptr_t* montop     = fp - rounded_vm_local_words;
1517 
1518   // preallocate monitors (cf. __ add_monitor_to_stack)
1519   intptr_t* monitors = montop - monitor_size;
1520 
1521   // preallocate stack space
1522   intptr_t*  esp = monitors - 1 -
1523     (tempcount * Interpreter::stackElementWords) -
1524     popframe_extra_args;
1525 
1526   int local_words = method->max_locals() * Interpreter::stackElementWords;
1527   NEEDS_CLEANUP;
1528   intptr_t* locals;
1529   if (caller->is_interpreted_frame()) {
1530     // Can force the locals area to end up properly overlapping the top of the expression stack.
1531     intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1;
1532     // Note that this computation means we replace size_of_parameters() values from the caller
1533     // interpreter frame's expression stack with our argument locals
1534     int parm_words  = caller_actual_parameters * Interpreter::stackElementWords;
1535     locals = Lesp_ptr + parm_words;
1536     int delta = local_words - parm_words;
1537     int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0;
1538     *interpreter_frame->register_addr(I5_savedSP)    = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS;
1539     if (!is_bottom_frame) {
1540       // Llast_SP is set below for the current frame to SP (with the
1541       // extra space for the callee's locals). Here we adjust
1542       // Llast_SP for the caller's frame, removing the extra space
1543       // for the current method's locals.
1544       *caller->register_addr(Llast_SP) = *interpreter_frame->register_addr(I5_savedSP);
1545     } else {
1546       assert(*caller->register_addr(Llast_SP) >= *interpreter_frame->register_addr(I5_savedSP), "strange Llast_SP");
1547     }
1548   } else {
1549     assert(caller->is_compiled_frame() || caller->is_entry_frame(), "only possible cases");
1550     // Don't have Lesp available; lay out locals block in the caller
1551     // adjacent to the register window save area.
1552     //
1553     // Compiled frames do not allocate a varargs area which is why this if
1554     // statement is needed.
1555     //
1556     if (caller->is_compiled_frame()) {
1557       locals = fp + frame::register_save_words + local_words - 1;
1558     } else {
1559       locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
1560     }
1561     if (!caller->is_entry_frame()) {
1562       // Caller wants his own SP back
1563       int caller_frame_size = caller->cb()->frame_size();
1564       *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS;
1565     }
1566   }
1567   if (TraceDeoptimization) {
1568     if (caller->is_entry_frame()) {
1569       // make sure I5_savedSP and the entry frames notion of saved SP
1570       // agree.  This assertion duplicate a check in entry frame code
1571       // but catches the failure earlier.
1572       assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP),
1573              "would change callers SP");
1574     }
1575     if (caller->is_entry_frame()) {
1576       tty->print("entry ");
1577     }
1578     if (caller->is_compiled_frame()) {
1579       tty->print("compiled ");
1580       if (caller->is_deoptimized_frame()) {
1581         tty->print("(deopt) ");
1582       }
1583     }
1584     if (caller->is_interpreted_frame()) {
1585       tty->print("interpreted ");
1586     }
1587     tty->print_cr("caller fp=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(caller->fp()), p2i(caller->sp()));
1588     tty->print_cr("save area = " INTPTR_FORMAT ", " INTPTR_FORMAT, p2i(caller->sp()), p2i(caller->sp() + 16));
1589     tty->print_cr("save area = " INTPTR_FORMAT ", " INTPTR_FORMAT, p2i(caller->fp()), p2i(caller->fp() + 16));
1590     tty->print_cr("interpreter fp=" INTPTR_FORMAT ", " INTPTR_FORMAT, p2i(interpreter_frame->fp()), p2i(interpreter_frame->sp()));
1591     tty->print_cr("save area = " INTPTR_FORMAT ", " INTPTR_FORMAT, p2i(interpreter_frame->sp()), p2i(interpreter_frame->sp() + 16));
1592     tty->print_cr("save area = " INTPTR_FORMAT ", " INTPTR_FORMAT, p2i(interpreter_frame->fp()), p2i(interpreter_frame->fp() + 16));
1593     tty->print_cr("Llocals = " INTPTR_FORMAT, p2i(locals));
1594     tty->print_cr("Lesp = " INTPTR_FORMAT, p2i(esp));
1595     tty->print_cr("Lmonitors = " INTPTR_FORMAT, p2i(monitors));
1596   }
1597 
1598   if (method->max_locals() > 0) {
1599     assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area");
1600     assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area");
1601     assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area");
1602     assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area");
1603   }
1604 #ifdef _LP64
1605   assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd");
1606 #endif
1607 
1608   *interpreter_frame->register_addr(Lmethod)     = (intptr_t) method;
1609   *interpreter_frame->register_addr(Llocals)     = (intptr_t) locals;
1610   *interpreter_frame->register_addr(Lmonitors)   = (intptr_t) monitors;
1611   *interpreter_frame->register_addr(Lesp)        = (intptr_t) esp;
1612   // Llast_SP will be same as SP as there is no adapter space
1613   *interpreter_frame->register_addr(Llast_SP)    = (intptr_t) interpreter_frame->sp() - STACK_BIAS;
1614   *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache();
1615 #ifdef FAST_DISPATCH
1616   *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table();
1617 #endif
1618 
1619 
1620 #ifdef ASSERT
1621   BasicObjectLock* mp = (BasicObjectLock*)monitors;
1622 
1623   assert(interpreter_frame->interpreter_frame_method() == method, "method matches");
1624   assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize)), "locals match");
1625   assert(interpreter_frame->interpreter_frame_monitor_end()   == mp, "monitor_end matches");
1626   assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches");
1627   assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches");
1628 
1629   // check bounds
1630   intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1);
1631   intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words;
1632   assert(lo < monitors && montop <= hi, "monitors in bounds");
1633   assert(lo <= esp && esp < monitors, "esp in bounds");
1634 #endif // ASSERT
1635 }
1636 
1637 //----------------------------------------------------------------------------------------------------
1638 // Exceptions
1639 void TemplateInterpreterGenerator::generate_throw_exception() {
1640 
1641   // Entry point in previous activation (i.e., if the caller was interpreted)
1642   Interpreter::_rethrow_exception_entry = __ pc();
1643   // O0: exception
1644 
1645   // entry point for exceptions thrown within interpreter code
1646   Interpreter::_throw_exception_entry = __ pc();
1647   __ verify_thread();
1648   // expression stack is undefined here
1649   // O0: exception, i.e. Oexception
1650   // Lbcp: exception bcp
1651   __ verify_oop(Oexception);
1652 
1653 
1654   // expression stack must be empty before entering the VM in case of an exception
1655   __ empty_expression_stack();
1656   // find exception handler address and preserve exception oop
1657   // call C routine to find handler and jump to it
1658   __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1659   __ push_ptr(O1); // push exception for exception handler bytecodes
1660 
1661   __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1662   __ delayed()->nop();
1663 
1664 
1665   // if the exception is not handled in the current frame
1666   // the frame is removed and the exception is rethrown
1667   // (i.e. exception continuation is _rethrow_exception)
1668   //
1669   // Note: At this point the bci is still the bxi for the instruction which caused
1670   //       the exception and the expression stack is empty. Thus, for any VM calls
1671   //       at this point, GC will find a legal oop map (with empty expression stack).
1672 
1673   // in current activation
1674   // tos: exception
1675   // Lbcp: exception bcp
1676 
1677   //
1678   // JVMTI PopFrame support
1679   //
1680 
1681   Interpreter::_remove_activation_preserving_args_entry = __ pc();
1682   Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1683   // Set the popframe_processing bit in popframe_condition indicating that we are
1684   // currently handling popframe, so that call_VMs that may happen later do not trigger new
1685   // popframe handling cycles.
1686 
1687   __ ld(popframe_condition_addr, G3_scratch);
1688   __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1689   __ stw(G3_scratch, popframe_condition_addr);
1690 
1691   // Empty the expression stack, as in normal exception handling
1692   __ empty_expression_stack();
1693   __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1694 
1695   {
1696     // Check to see whether we are returning to a deoptimized frame.
1697     // (The PopFrame call ensures that the caller of the popped frame is
1698     // either interpreted or compiled and deoptimizes it if compiled.)
1699     // In this case, we can't call dispatch_next() after the frame is
1700     // popped, but instead must save the incoming arguments and restore
1701     // them after deoptimization has occurred.
1702     //
1703     // Note that we don't compare the return PC against the
1704     // deoptimization blob's unpack entry because of the presence of
1705     // adapter frames in C2.
1706     Label caller_not_deoptimized;
1707     __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1708     __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1709 
1710     const Register Gtmp1 = G3_scratch;
1711     const Register Gtmp2 = G1_scratch;
1712     const Register RconstMethod = Gtmp1;
1713     const Address constMethod(Lmethod, Method::const_offset());
1714     const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1715 
1716     // Compute size of arguments for saving when returning to deoptimized caller
1717     __ ld_ptr(constMethod, RconstMethod);
1718     __ lduh(size_of_parameters, Gtmp1);
1719     __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1720     __ sub(Llocals, Gtmp1, Gtmp2);
1721     __ add(Gtmp2, wordSize, Gtmp2);
1722     // Save these arguments
1723     __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1724     // Inform deoptimization that it is responsible for restoring these arguments
1725     __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1726     Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1727     __ st(Gtmp1, popframe_condition_addr);
1728 
1729     // Return from the current method
1730     // The caller's SP was adjusted upon method entry to accomodate
1731     // the callee's non-argument locals. Undo that adjustment.
1732     __ ret();
1733     __ delayed()->restore(I5_savedSP, G0, SP);
1734 
1735     __ bind(caller_not_deoptimized);
1736   }
1737 
1738   // Clear the popframe condition flag
1739   __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1740 
1741   // Get out of the current method (how this is done depends on the particular compiler calling
1742   // convention that the interpreter currently follows)
1743   // The caller's SP was adjusted upon method entry to accomodate
1744   // the callee's non-argument locals. Undo that adjustment.
1745   __ restore(I5_savedSP, G0, SP);
1746   // The method data pointer was incremented already during
1747   // call profiling. We have to restore the mdp for the current bcp.
1748   if (ProfileInterpreter) {
1749     __ set_method_data_pointer_for_bcp();
1750   }
1751 
1752 #if INCLUDE_JVMTI
1753   {
1754     Label L_done;
1755 
1756     __ ldub(Address(Lbcp, 0), G1_scratch);  // Load current bytecode
1757     __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done);
1758 
1759     // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
1760     // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
1761 
1762     __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp);
1763 
1764     __ br_null(G1_scratch, false, Assembler::pn, L_done);
1765     __ delayed()->nop();
1766 
1767     __ st_ptr(G1_scratch, Lesp, wordSize);
1768     __ bind(L_done);
1769   }
1770 #endif // INCLUDE_JVMTI
1771 
1772   // Resume bytecode interpretation at the current bcp
1773   __ dispatch_next(vtos);
1774   // end of JVMTI PopFrame support
1775 
1776   Interpreter::_remove_activation_entry = __ pc();
1777 
1778   // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1779   __ pop_ptr(Oexception);                                  // get exception
1780 
1781   // Intel has the following comment:
1782   //// remove the activation (without doing throws on illegalMonitorExceptions)
1783   // They remove the activation without checking for bad monitor state.
1784   // %%% We should make sure this is the right semantics before implementing.
1785 
1786   __ set_vm_result(Oexception);
1787   __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1788 
1789   __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1790 
1791   __ get_vm_result(Oexception);
1792   __ verify_oop(Oexception);
1793 
1794     const int return_reg_adjustment = frame::pc_return_offset;
1795   Address issuing_pc_addr(I7, return_reg_adjustment);
1796 
1797   // We are done with this activation frame; find out where to go next.
1798   // The continuation point will be an exception handler, which expects
1799   // the following registers set up:
1800   //
1801   // Oexception: exception
1802   // Oissuing_pc: the local call that threw exception
1803   // Other On: garbage
1804   // In/Ln:  the contents of the caller's register window
1805   //
1806   // We do the required restore at the last possible moment, because we
1807   // need to preserve some state across a runtime call.
1808   // (Remember that the caller activation is unknown--it might not be
1809   // interpreted, so things like Lscratch are useless in the caller.)
1810 
1811   // Although the Intel version uses call_C, we can use the more
1812   // compact call_VM.  (The only real difference on SPARC is a
1813   // harmlessly ignored [re]set_last_Java_frame, compared with
1814   // the Intel code which lacks this.)
1815   __ mov(Oexception,      Oexception ->after_save());  // get exception in I0 so it will be on O0 after restore
1816   __ add(issuing_pc_addr, Oissuing_pc->after_save());  // likewise set I1 to a value local to the caller
1817   __ super_call_VM_leaf(L7_thread_cache,
1818                         CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1819                         G2_thread, Oissuing_pc->after_save());
1820 
1821   // The caller's SP was adjusted upon method entry to accomodate
1822   // the callee's non-argument locals. Undo that adjustment.
1823   __ JMP(O0, 0);                         // return exception handler in caller
1824   __ delayed()->restore(I5_savedSP, G0, SP);
1825 
1826   // (same old exception object is already in Oexception; see above)
1827   // Note that an "issuing PC" is actually the next PC after the call
1828 }
1829 
1830 
1831 //
1832 // JVMTI ForceEarlyReturn support
1833 //
1834 
1835 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1836   address entry = __ pc();
1837 
1838   __ empty_expression_stack();
1839   __ load_earlyret_value(state);
1840 
1841   __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1842   Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1843 
1844   // Clear the earlyret state
1845   __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1846 
1847   __ remove_activation(state,
1848                        /* throw_monitor_exception */ false,
1849                        /* install_monitor_exception */ false);
1850 
1851   // The caller's SP was adjusted upon method entry to accomodate
1852   // the callee's non-argument locals. Undo that adjustment.
1853   __ ret();                             // return to caller
1854   __ delayed()->restore(I5_savedSP, G0, SP);
1855 
1856   return entry;
1857 } // end of JVMTI ForceEarlyReturn support
1858 
1859 
1860 //------------------------------------------------------------------------------------------------------------------------
1861 // Helper for vtos entry point generation
1862 
1863 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) {
1864   assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1865   Label L;
1866   aep = __ pc(); __ push_ptr(); __ ba_short(L);
1867   fep = __ pc(); __ push_f();   __ ba_short(L);
1868   dep = __ pc(); __ push_d();   __ ba_short(L);
1869   lep = __ pc(); __ push_l();   __ ba_short(L);
1870   iep = __ pc(); __ push_i();
1871   bep = cep = sep = iep;                        // there aren't any
1872   vep = __ pc(); __ bind(L);                    // fall through
1873   generate_and_dispatch(t);
1874 }
1875 
1876 // --------------------------------------------------------------------------------
1877 
1878 
1879 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
1880  : TemplateInterpreterGenerator(code) {
1881    generate_all(); // down here so it can be "virtual"
1882 }
1883 
1884 // --------------------------------------------------------------------------------
1885 
1886 // Non-product code
1887 #ifndef PRODUCT
1888 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
1889   address entry = __ pc();
1890 
1891   __ push(state);
1892   __ mov(O7, Lscratch); // protect return address within interpreter
1893 
1894   // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
1895   __ mov( Otos_l2, G3_scratch );
1896   __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
1897   __ mov(Lscratch, O7); // restore return address
1898   __ pop(state);
1899   __ retl();
1900   __ delayed()->nop();
1901 
1902   return entry;
1903 }
1904 
1905 
1906 // helpers for generate_and_dispatch
1907 
1908 void TemplateInterpreterGenerator::count_bytecode() {
1909   __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
1910 }
1911 
1912 
1913 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
1914   __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
1915 }
1916 
1917 
1918 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
1919   AddressLiteral index   (&BytecodePairHistogram::_index);
1920   AddressLiteral counters((address) &BytecodePairHistogram::_counters);
1921 
1922   // get index, shift out old bytecode, bring in new bytecode, and store it
1923   // _index = (_index >> log2_number_of_codes) |
1924   //          (bytecode << log2_number_of_codes);
1925 
1926   __ load_contents(index, G4_scratch);
1927   __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
1928   __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes,  G3_scratch );
1929   __ or3( G3_scratch,  G4_scratch, G4_scratch );
1930   __ store_contents(G4_scratch, index, G3_scratch);
1931 
1932   // bump bucket contents
1933   // _counters[_index] ++;
1934 
1935   __ set(counters, G3_scratch);                       // loads into G3_scratch
1936   __ sll( G4_scratch, LogBytesPerWord, G4_scratch );  // Index is word address
1937   __ add (G3_scratch, G4_scratch, G3_scratch);        // Add in index
1938   __ ld (G3_scratch, 0, G4_scratch);
1939   __ inc (G4_scratch);
1940   __ st (G4_scratch, 0, G3_scratch);
1941 }
1942 
1943 
1944 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
1945   // Call a little run-time stub to avoid blow-up for each bytecode.
1946   // The run-time runtime saves the right registers, depending on
1947   // the tosca in-state for the given template.
1948   address entry = Interpreter::trace_code(t->tos_in());
1949   guarantee(entry != NULL, "entry must have been generated");
1950   __ call(entry, relocInfo::none);
1951   __ delayed()->nop();
1952 }
1953 
1954 
1955 void TemplateInterpreterGenerator::stop_interpreter_at() {
1956   AddressLiteral counter(&BytecodeCounter::_counter_value);
1957   __ load_contents(counter, G3_scratch);
1958   AddressLiteral stop_at(&StopInterpreterAt);
1959   __ load_ptr_contents(stop_at, G4_scratch);
1960   __ cmp(G3_scratch, G4_scratch);
1961   __ breakpoint_trap(Assembler::equal, Assembler::icc);
1962 }
1963 #endif // not PRODUCT
1964 #endif // !CC_INTERP