1 /* 2 * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "interpreter/bytecodeHistogram.hpp" 28 #include "interpreter/interpreter.hpp" 29 #include "interpreter/interpreterGenerator.hpp" 30 #include "interpreter/interpreterRuntime.hpp" 31 #include "interpreter/templateTable.hpp" 32 #include "oops/arrayOop.hpp" 33 #include "oops/methodData.hpp" 34 #include "oops/method.hpp" 35 #include "oops/oop.inline.hpp" 36 #include "prims/jvmtiExport.hpp" 37 #include "prims/jvmtiThreadState.hpp" 38 #include "runtime/arguments.hpp" 39 #include "runtime/deoptimization.hpp" 40 #include "runtime/frame.inline.hpp" 41 #include "runtime/sharedRuntime.hpp" 42 #include "runtime/stubRoutines.hpp" 43 #include "runtime/synchronizer.hpp" 44 #include "runtime/timer.hpp" 45 #include "runtime/vframeArray.hpp" 46 #include "utilities/debug.hpp" 47 48 #ifndef CC_INTERP 49 #ifndef FAST_DISPATCH 50 #define FAST_DISPATCH 1 51 #endif 52 #undef FAST_DISPATCH 53 54 55 // Generation of Interpreter 56 // 57 // The InterpreterGenerator generates the interpreter into Interpreter::_code. 58 59 60 #define __ _masm-> 61 62 63 //---------------------------------------------------------------------------------------------------- 64 65 66 void InterpreterGenerator::save_native_result(void) { 67 // result potentially in O0/O1: save it across calls 68 const Address& l_tmp = InterpreterMacroAssembler::l_tmp; 69 70 // result potentially in F0/F1: save it across calls 71 const Address& d_tmp = InterpreterMacroAssembler::d_tmp; 72 73 // save and restore any potential method result value around the unlocking operation 74 __ stf(FloatRegisterImpl::D, F0, d_tmp); 75 #ifdef _LP64 76 __ stx(O0, l_tmp); 77 #else 78 __ std(O0, l_tmp); 79 #endif 80 } 81 82 void InterpreterGenerator::restore_native_result(void) { 83 const Address& l_tmp = InterpreterMacroAssembler::l_tmp; 84 const Address& d_tmp = InterpreterMacroAssembler::d_tmp; 85 86 // Restore any method result value 87 __ ldf(FloatRegisterImpl::D, d_tmp, F0); 88 #ifdef _LP64 89 __ ldx(l_tmp, O0); 90 #else 91 __ ldd(l_tmp, O0); 92 #endif 93 } 94 95 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) { 96 assert(!pass_oop || message == NULL, "either oop or message but not both"); 97 address entry = __ pc(); 98 // expression stack must be empty before entering the VM if an exception happened 99 __ empty_expression_stack(); 100 // load exception object 101 __ set((intptr_t)name, G3_scratch); 102 if (pass_oop) { 103 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i); 104 } else { 105 __ set((intptr_t)message, G4_scratch); 106 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch); 107 } 108 // throw exception 109 assert(Interpreter::throw_exception_entry() != NULL, "generate it first"); 110 AddressLiteral thrower(Interpreter::throw_exception_entry()); 111 __ jump_to(thrower, G3_scratch); 112 __ delayed()->nop(); 113 return entry; 114 } 115 116 address TemplateInterpreterGenerator::generate_ClassCastException_handler() { 117 address entry = __ pc(); 118 // expression stack must be empty before entering the VM if an exception 119 // happened 120 __ empty_expression_stack(); 121 // load exception object 122 __ call_VM(Oexception, 123 CAST_FROM_FN_PTR(address, 124 InterpreterRuntime::throw_ClassCastException), 125 Otos_i); 126 __ should_not_reach_here(); 127 return entry; 128 } 129 130 131 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) { 132 address entry = __ pc(); 133 // expression stack must be empty before entering the VM if an exception happened 134 __ empty_expression_stack(); 135 // convention: expect aberrant index in register G3_scratch, then shuffle the 136 // index to G4_scratch for the VM call 137 __ mov(G3_scratch, G4_scratch); 138 __ set((intptr_t)name, G3_scratch); 139 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch); 140 __ should_not_reach_here(); 141 return entry; 142 } 143 144 145 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() { 146 address entry = __ pc(); 147 // expression stack must be empty before entering the VM if an exception happened 148 __ empty_expression_stack(); 149 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError)); 150 __ should_not_reach_here(); 151 return entry; 152 } 153 154 155 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step) { 156 TosState incoming_state = state; 157 158 Label cont; 159 address compiled_entry = __ pc(); 160 161 address entry = __ pc(); 162 #if !defined(_LP64) && defined(COMPILER2) 163 // All return values are where we want them, except for Longs. C2 returns 164 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1. 165 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit 166 // build even if we are returning from interpreted we just do a little 167 // stupid shuffing. 168 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to 169 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node 170 // first which would move g1 -> O0/O1 and destroy the exception we were throwing. 171 172 if (incoming_state == ltos) { 173 __ srl (G1, 0, O1); 174 __ srlx(G1, 32, O0); 175 } 176 #endif // !_LP64 && COMPILER2 177 178 __ bind(cont); 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 Label L_got_cache, L_giant_index; 189 const Register cache = G3_scratch; 190 const Register size = G1_scratch; 191 if (EnableInvokeDynamic) { 192 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode. 193 __ cmp_and_br_short(G1_scratch, Bytecodes::_invokedynamic, Assembler::equal, Assembler::pn, L_giant_index); 194 } 195 __ get_cache_and_index_at_bcp(cache, G1_scratch, 1); 196 __ bind(L_got_cache); 197 __ ld_ptr(cache, ConstantPoolCache::base_offset() + 198 ConstantPoolCacheEntry::flags_offset(), size); 199 __ and3(size, 0xFF, size); // argument size in words 200 __ sll(size, Interpreter::logStackElementSize, size); // each argument size in bytes 201 __ add(Lesp, size, Lesp); // pop arguments 202 __ dispatch_next(state, step); 203 204 // out of the main line of code... 205 if (EnableInvokeDynamic) { 206 __ bind(L_giant_index); 207 __ get_cache_and_index_at_bcp(cache, G1_scratch, 1, sizeof(u4)); 208 __ ba_short(L_got_cache); 209 } 210 211 return entry; 212 } 213 214 215 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) { 216 address entry = __ pc(); 217 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache 218 { Label L; 219 Address exception_addr(G2_thread, Thread::pending_exception_offset()); 220 __ ld_ptr(exception_addr, Gtemp); // Load pending exception. 221 __ br_null_short(Gtemp, Assembler::pt, L); 222 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception)); 223 __ should_not_reach_here(); 224 __ bind(L); 225 } 226 __ dispatch_next(state, step); 227 return entry; 228 } 229 230 // A result handler converts/unboxes a native call result into 231 // a java interpreter/compiler result. The current frame is an 232 // interpreter frame. The activation frame unwind code must be 233 // consistent with that of TemplateTable::_return(...). In the 234 // case of native methods, the caller's SP was not modified. 235 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) { 236 address entry = __ pc(); 237 Register Itos_i = Otos_i ->after_save(); 238 Register Itos_l = Otos_l ->after_save(); 239 Register Itos_l1 = Otos_l1->after_save(); 240 Register Itos_l2 = Otos_l2->after_save(); 241 switch (type) { 242 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false 243 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value! 244 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break; 245 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break; 246 case T_LONG : 247 #ifndef _LP64 248 __ mov(O1, Itos_l2); // move other half of long 249 #endif // ifdef or no ifdef, fall through to the T_INT case 250 case T_INT : __ mov(O0, Itos_i); break; 251 case T_VOID : /* nothing to do */ break; 252 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break; 253 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break; 254 case T_OBJECT : 255 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i); 256 __ verify_oop(Itos_i); 257 break; 258 default : ShouldNotReachHere(); 259 } 260 __ ret(); // return from interpreter activation 261 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame 262 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly 263 return entry; 264 } 265 266 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) { 267 address entry = __ pc(); 268 __ push(state); 269 __ call_VM(noreg, runtime_entry); 270 __ dispatch_via(vtos, Interpreter::normal_table(vtos)); 271 return entry; 272 } 273 274 275 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) { 276 address entry = __ pc(); 277 __ dispatch_next(state); 278 return entry; 279 } 280 281 // 282 // Helpers for commoning out cases in the various type of method entries. 283 // 284 285 // increment invocation count & check for overflow 286 // 287 // Note: checking for negative value instead of overflow 288 // so we have a 'sticky' overflow test 289 // 290 // Lmethod: method 291 // ??: invocation counter 292 // 293 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) { 294 // Note: In tiered we increment either counters in Method* or in MDO depending if we're profiling or not. 295 if (TieredCompilation) { 296 const int increment = InvocationCounter::count_increment; 297 const int mask = ((1 << Tier0InvokeNotifyFreqLog) - 1) << InvocationCounter::count_shift; 298 Label no_mdo, done; 299 if (ProfileInterpreter) { 300 // If no method data exists, go to profile_continue. 301 __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch); 302 __ br_null_short(G4_scratch, Assembler::pn, no_mdo); 303 // Increment counter 304 Address mdo_invocation_counter(G4_scratch, 305 in_bytes(MethodData::invocation_counter_offset()) + 306 in_bytes(InvocationCounter::counter_offset())); 307 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask, 308 G3_scratch, Lscratch, 309 Assembler::zero, overflow); 310 __ ba_short(done); 311 } 312 313 // Increment counter in Method* 314 __ bind(no_mdo); 315 Address invocation_counter(Lmethod, 316 in_bytes(Method::invocation_counter_offset()) + 317 in_bytes(InvocationCounter::counter_offset())); 318 __ increment_mask_and_jump(invocation_counter, increment, mask, 319 G3_scratch, Lscratch, 320 Assembler::zero, overflow); 321 __ bind(done); 322 } else { 323 // Update standard invocation counters 324 __ increment_invocation_counter(O0, G3_scratch); 325 if (ProfileInterpreter) { // %%% Merge this into MethodData* 326 Address interpreter_invocation_counter(Lmethod,in_bytes(Method::interpreter_invocation_counter_offset())); 327 __ ld(interpreter_invocation_counter, G3_scratch); 328 __ inc(G3_scratch); 329 __ st(G3_scratch, interpreter_invocation_counter); 330 } 331 332 if (ProfileInterpreter && profile_method != NULL) { 333 // Test to see if we should create a method data oop 334 AddressLiteral profile_limit((address)&InvocationCounter::InterpreterProfileLimit); 335 __ load_contents(profile_limit, G3_scratch); 336 __ cmp_and_br_short(O0, G3_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue); 337 338 // if no method data exists, go to profile_method 339 __ test_method_data_pointer(*profile_method); 340 } 341 342 AddressLiteral invocation_limit((address)&InvocationCounter::InterpreterInvocationLimit); 343 __ load_contents(invocation_limit, G3_scratch); 344 __ cmp(O0, G3_scratch); 345 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance 346 __ delayed()->nop(); 347 } 348 349 } 350 351 // Allocate monitor and lock method (asm interpreter) 352 // ebx - Method* 353 // 354 void InterpreterGenerator::lock_method(void) { 355 __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0); // Load access flags. 356 357 #ifdef ASSERT 358 { Label ok; 359 __ btst(JVM_ACC_SYNCHRONIZED, O0); 360 __ br( Assembler::notZero, false, Assembler::pt, ok); 361 __ delayed()->nop(); 362 __ stop("method doesn't need synchronization"); 363 __ bind(ok); 364 } 365 #endif // ASSERT 366 367 // get synchronization object to O0 368 { Label done; 369 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 370 __ btst(JVM_ACC_STATIC, O0); 371 __ br( Assembler::zero, true, Assembler::pt, done); 372 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case 373 374 __ ld_ptr( Lmethod, in_bytes(Method::const_offset()), O0); 375 __ ld_ptr( O0, in_bytes(ConstMethod::constants_offset()), O0); 376 __ ld_ptr( O0, ConstantPool::pool_holder_offset_in_bytes(), O0); 377 378 // lock the mirror, not the Klass* 379 __ ld_ptr( O0, mirror_offset, O0); 380 381 #ifdef ASSERT 382 __ tst(O0); 383 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc); 384 #endif // ASSERT 385 386 __ bind(done); 387 } 388 389 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem 390 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object 391 // __ untested("lock_object from method entry"); 392 __ lock_object(Lmonitors, O0); 393 } 394 395 396 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size, 397 Register Rscratch, 398 Register Rscratch2) { 399 const int page_size = os::vm_page_size(); 400 Label after_frame_check; 401 402 assert_different_registers(Rframe_size, Rscratch, Rscratch2); 403 404 __ set(page_size, Rscratch); 405 __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check); 406 407 // get the stack base, and in debug, verify it is non-zero 408 __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch ); 409 #ifdef ASSERT 410 Label base_not_zero; 411 __ br_notnull_short(Rscratch, Assembler::pn, base_not_zero); 412 __ stop("stack base is zero in generate_stack_overflow_check"); 413 __ bind(base_not_zero); 414 #endif 415 416 // get the stack size, and in debug, verify it is non-zero 417 assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" ); 418 __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 ); 419 #ifdef ASSERT 420 Label size_not_zero; 421 __ br_notnull_short(Rscratch2, Assembler::pn, size_not_zero); 422 __ stop("stack size is zero in generate_stack_overflow_check"); 423 __ bind(size_not_zero); 424 #endif 425 426 // compute the beginning of the protected zone minus the requested frame size 427 __ sub( Rscratch, Rscratch2, Rscratch ); 428 __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 ); 429 __ add( Rscratch, Rscratch2, Rscratch ); 430 431 // Add in the size of the frame (which is the same as subtracting it from the 432 // SP, which would take another register 433 __ add( Rscratch, Rframe_size, Rscratch ); 434 435 // the frame is greater than one page in size, so check against 436 // the bottom of the stack 437 __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check); 438 439 // the stack will overflow, throw an exception 440 441 // Note that SP is restored to sender's sp (in the delay slot). This 442 // is necessary if the sender's frame is an extended compiled frame 443 // (see gen_c2i_adapter()) and safer anyway in case of JSR292 444 // adaptations. 445 446 // Note also that the restored frame is not necessarily interpreted. 447 // Use the shared runtime version of the StackOverflowError. 448 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated"); 449 AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry()); 450 __ jump_to(stub, Rscratch); 451 __ delayed()->mov(O5_savedSP, SP); 452 453 // if you get to here, then there is enough stack space 454 __ bind( after_frame_check ); 455 } 456 457 458 // 459 // Generate a fixed interpreter frame. This is identical setup for interpreted 460 // methods and for native methods hence the shared code. 461 462 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) { 463 // 464 // 465 // The entry code sets up a new interpreter frame in 4 steps: 466 // 467 // 1) Increase caller's SP by for the extra local space needed: 468 // (check for overflow) 469 // Efficient implementation of xload/xstore bytecodes requires 470 // that arguments and non-argument locals are in a contigously 471 // addressable memory block => non-argument locals must be 472 // allocated in the caller's frame. 473 // 474 // 2) Create a new stack frame and register window: 475 // The new stack frame must provide space for the standard 476 // register save area, the maximum java expression stack size, 477 // the monitor slots (0 slots initially), and some frame local 478 // scratch locations. 479 // 480 // 3) The following interpreter activation registers must be setup: 481 // Lesp : expression stack pointer 482 // Lbcp : bytecode pointer 483 // Lmethod : method 484 // Llocals : locals pointer 485 // Lmonitors : monitor pointer 486 // LcpoolCache: constant pool cache 487 // 488 // 4) Initialize the non-argument locals if necessary: 489 // Non-argument locals may need to be initialized to NULL 490 // for GC to work. If the oop-map information is accurate 491 // (in the absence of the JSR problem), no initialization 492 // is necessary. 493 // 494 // (gri - 2/25/2000) 495 496 497 const Address size_of_parameters(G5_method, Method::size_of_parameters_offset()); 498 const Address size_of_locals (G5_method, Method::size_of_locals_offset()); 499 const Address constMethod (G5_method, Method::const_offset()); 500 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong ); 501 502 const int extra_space = 503 rounded_vm_local_words + // frame local scratch space 504 //6815692//Method::extra_stack_words() + // extra push slots for MH adapters 505 frame::memory_parameter_word_sp_offset + // register save area 506 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0); 507 508 const Register Glocals_size = G3; 509 const Register Otmp1 = O3; 510 const Register Otmp2 = O4; 511 // Lscratch can't be used as a temporary because the call_stub uses 512 // it to assert that the stack frame was setup correctly. 513 514 __ lduh( size_of_parameters, Glocals_size); 515 516 // Gargs points to first local + BytesPerWord 517 // Set the saved SP after the register window save 518 // 519 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP); 520 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1); 521 __ add(Gargs, Otmp1, Gargs); 522 523 if (native_call) { 524 __ calc_mem_param_words( Glocals_size, Gframe_size ); 525 __ add( Gframe_size, extra_space, Gframe_size); 526 __ round_to( Gframe_size, WordsPerLong ); 527 __ sll( Gframe_size, LogBytesPerWord, Gframe_size ); 528 } else { 529 530 // 531 // Compute number of locals in method apart from incoming parameters 532 // 533 __ lduh( size_of_locals, Otmp1 ); 534 __ sub( Otmp1, Glocals_size, Glocals_size ); 535 __ round_to( Glocals_size, WordsPerLong ); 536 __ sll( Glocals_size, Interpreter::logStackElementSize, Glocals_size ); 537 538 // see if the frame is greater than one page in size. If so, 539 // then we need to verify there is enough stack space remaining 540 // Frame_size = (max_stack + extra_space) * BytesPerWord; 541 __ ld_ptr( constMethod, Gframe_size ); 542 __ lduh( Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size ); 543 __ add( Gframe_size, extra_space, Gframe_size ); 544 __ round_to( Gframe_size, WordsPerLong ); 545 __ sll( Gframe_size, Interpreter::logStackElementSize, Gframe_size); 546 547 // Add in java locals size for stack overflow check only 548 __ add( Gframe_size, Glocals_size, Gframe_size ); 549 550 const Register Otmp2 = O4; 551 assert_different_registers(Otmp1, Otmp2, O5_savedSP); 552 generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2); 553 554 __ sub( Gframe_size, Glocals_size, Gframe_size); 555 556 // 557 // bump SP to accomodate the extra locals 558 // 559 __ sub( SP, Glocals_size, SP ); 560 } 561 562 // 563 // now set up a stack frame with the size computed above 564 // 565 __ neg( Gframe_size ); 566 __ save( SP, Gframe_size, SP ); 567 568 // 569 // now set up all the local cache registers 570 // 571 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note 572 // that all present references to Lbyte_code initialize the register 573 // immediately before use 574 if (native_call) { 575 __ mov(G0, Lbcp); 576 } else { 577 __ ld_ptr(G5_method, Method::const_offset(), Lbcp); 578 __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp); 579 } 580 __ mov( G5_method, Lmethod); // set Lmethod 581 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache 582 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors 583 #ifdef _LP64 584 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias 585 #endif 586 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp 587 588 // setup interpreter activation registers 589 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals 590 591 if (ProfileInterpreter) { 592 #ifdef FAST_DISPATCH 593 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since 594 // they both use I2. 595 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive"); 596 #endif // FAST_DISPATCH 597 __ set_method_data_pointer(); 598 } 599 600 } 601 602 // Empty method, generate a very fast return. 603 604 address InterpreterGenerator::generate_empty_entry(void) { 605 606 // A method that does nother but return... 607 608 address entry = __ pc(); 609 Label slow_path; 610 611 // do nothing for empty methods (do not even increment invocation counter) 612 if ( UseFastEmptyMethods) { 613 // If we need a safepoint check, generate full interpreter entry. 614 AddressLiteral sync_state(SafepointSynchronize::address_of_state()); 615 __ set(sync_state, G3_scratch); 616 __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path); 617 618 // Code: _return 619 __ retl(); 620 __ delayed()->mov(O5_savedSP, SP); 621 622 __ bind(slow_path); 623 (void) generate_normal_entry(false); 624 625 return entry; 626 } 627 return NULL; 628 } 629 630 // Call an accessor method (assuming it is resolved, otherwise drop into 631 // vanilla (slow path) entry 632 633 // Generates code to elide accessor methods 634 // Uses G3_scratch and G1_scratch as scratch 635 address InterpreterGenerator::generate_accessor_entry(void) { 636 637 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof; 638 // parameter size = 1 639 // Note: We can only use this code if the getfield has been resolved 640 // and if we don't have a null-pointer exception => check for 641 // these conditions first and use slow path if necessary. 642 address entry = __ pc(); 643 Label slow_path; 644 645 646 // XXX: for compressed oops pointer loading and decoding doesn't fit in 647 // delay slot and damages G1 648 if ( UseFastAccessorMethods && !UseCompressedOops ) { 649 // Check if we need to reach a safepoint and generate full interpreter 650 // frame if so. 651 AddressLiteral sync_state(SafepointSynchronize::address_of_state()); 652 __ load_contents(sync_state, G3_scratch); 653 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized); 654 __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path); 655 656 // Check if local 0 != NULL 657 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0 658 // check if local 0 == NULL and go the slow path 659 __ br_null_short(Otos_i, Assembler::pn, slow_path); 660 661 662 // read first instruction word and extract bytecode @ 1 and index @ 2 663 // get first 4 bytes of the bytecodes (big endian!) 664 __ ld_ptr(G5_method, Method::const_offset(), G1_scratch); 665 __ ld(G1_scratch, ConstMethod::codes_offset(), G1_scratch); 666 667 // move index @ 2 far left then to the right most two bytes. 668 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch); 669 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words( 670 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch); 671 672 // get constant pool cache 673 __ ld_ptr(G5_method, Method::const_offset(), G3_scratch); 674 __ ld_ptr(G3_scratch, ConstMethod::constants_offset(), G3_scratch); 675 __ ld_ptr(G3_scratch, ConstantPool::cache_offset_in_bytes(), G3_scratch); 676 677 // get specific constant pool cache entry 678 __ add(G3_scratch, G1_scratch, G3_scratch); 679 680 // Check the constant Pool cache entry to see if it has been resolved. 681 // If not, need the slow path. 682 ByteSize cp_base_offset = ConstantPoolCache::base_offset(); 683 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::indices_offset(), G1_scratch); 684 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch); 685 __ and3(G1_scratch, 0xFF, G1_scratch); 686 __ cmp_and_br_short(G1_scratch, Bytecodes::_getfield, Assembler::notEqual, Assembler::pn, slow_path); 687 688 // Get the type and return field offset from the constant pool cache 689 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::flags_offset(), G1_scratch); 690 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), G3_scratch); 691 692 Label xreturn_path; 693 // Need to differentiate between igetfield, agetfield, bgetfield etc. 694 // because they are different sizes. 695 // Get the type from the constant pool cache 696 __ srl(G1_scratch, ConstantPoolCacheEntry::tos_state_shift, G1_scratch); 697 // Make sure we don't need to mask G1_scratch after the above shift 698 ConstantPoolCacheEntry::verify_tos_state_shift(); 699 __ cmp(G1_scratch, atos ); 700 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 701 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i); 702 __ cmp(G1_scratch, itos); 703 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 704 __ delayed()->ld(Otos_i, G3_scratch, Otos_i); 705 __ cmp(G1_scratch, stos); 706 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 707 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i); 708 __ cmp(G1_scratch, ctos); 709 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 710 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i); 711 #ifdef ASSERT 712 __ cmp(G1_scratch, btos); 713 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 714 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i); 715 __ should_not_reach_here(); 716 #endif 717 __ ldsb(Otos_i, G3_scratch, Otos_i); 718 __ bind(xreturn_path); 719 720 // _ireturn/_areturn 721 __ retl(); // return from leaf routine 722 __ delayed()->mov(O5_savedSP, SP); 723 724 // Generate regular method entry 725 __ bind(slow_path); 726 (void) generate_normal_entry(false); 727 return entry; 728 } 729 return NULL; 730 } 731 732 // Method entry for java.lang.ref.Reference.get. 733 address InterpreterGenerator::generate_Reference_get_entry(void) { 734 #ifndef SERIALGC 735 // Code: _aload_0, _getfield, _areturn 736 // parameter size = 1 737 // 738 // The code that gets generated by this routine is split into 2 parts: 739 // 1. The "intrinsified" code for G1 (or any SATB based GC), 740 // 2. The slow path - which is an expansion of the regular method entry. 741 // 742 // Notes:- 743 // * In the G1 code we do not check whether we need to block for 744 // a safepoint. If G1 is enabled then we must execute the specialized 745 // code for Reference.get (except when the Reference object is null) 746 // so that we can log the value in the referent field with an SATB 747 // update buffer. 748 // If the code for the getfield template is modified so that the 749 // G1 pre-barrier code is executed when the current method is 750 // Reference.get() then going through the normal method entry 751 // will be fine. 752 // * The G1 code can, however, check the receiver object (the instance 753 // of java.lang.Reference) and jump to the slow path if null. If the 754 // Reference object is null then we obviously cannot fetch the referent 755 // and so we don't need to call the G1 pre-barrier. Thus we can use the 756 // regular method entry code to generate the NPE. 757 // 758 // This code is based on generate_accessor_enty. 759 760 address entry = __ pc(); 761 762 const int referent_offset = java_lang_ref_Reference::referent_offset; 763 guarantee(referent_offset > 0, "referent offset not initialized"); 764 765 if (UseG1GC) { 766 Label slow_path; 767 768 // In the G1 code we don't check if we need to reach a safepoint. We 769 // continue and the thread will safepoint at the next bytecode dispatch. 770 771 // Check if local 0 != NULL 772 // If the receiver is null then it is OK to jump to the slow path. 773 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0 774 // check if local 0 == NULL and go the slow path 775 __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path); 776 777 778 // Load the value of the referent field. 779 if (Assembler::is_simm13(referent_offset)) { 780 __ load_heap_oop(Otos_i, referent_offset, Otos_i); 781 } else { 782 __ set(referent_offset, G3_scratch); 783 __ load_heap_oop(Otos_i, G3_scratch, Otos_i); 784 } 785 786 // Generate the G1 pre-barrier code to log the value of 787 // the referent field in an SATB buffer. Note with 788 // these parameters the pre-barrier does not generate 789 // the load of the previous value 790 791 __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */, 792 Otos_i /* pre_val */, 793 G3_scratch /* tmp */, 794 true /* preserve_o_regs */); 795 796 // _areturn 797 __ retl(); // return from leaf routine 798 __ delayed()->mov(O5_savedSP, SP); 799 800 // Generate regular method entry 801 __ bind(slow_path); 802 (void) generate_normal_entry(false); 803 return entry; 804 } 805 #endif // SERIALGC 806 807 // If G1 is not enabled then attempt to go through the accessor entry point 808 // Reference.get is an accessor 809 return generate_accessor_entry(); 810 } 811 812 // 813 // Interpreter stub for calling a native method. (asm interpreter) 814 // This sets up a somewhat different looking stack for calling the native method 815 // than the typical interpreter frame setup. 816 // 817 818 address InterpreterGenerator::generate_native_entry(bool synchronized) { 819 address entry = __ pc(); 820 821 // the following temporary registers are used during frame creation 822 const Register Gtmp1 = G3_scratch ; 823 const Register Gtmp2 = G1_scratch; 824 bool inc_counter = UseCompiler || CountCompiledCalls; 825 826 // make sure registers are different! 827 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2); 828 829 const Address Laccess_flags(Lmethod, Method::access_flags_offset()); 830 831 const Register Glocals_size = G3; 832 assert_different_registers(Glocals_size, G4_scratch, Gframe_size); 833 834 // make sure method is native & not abstract 835 // rethink these assertions - they can be simplified and shared (gri 2/25/2000) 836 #ifdef ASSERT 837 __ ld(G5_method, Method::access_flags_offset(), Gtmp1); 838 { 839 Label L; 840 __ btst(JVM_ACC_NATIVE, Gtmp1); 841 __ br(Assembler::notZero, false, Assembler::pt, L); 842 __ delayed()->nop(); 843 __ stop("tried to execute non-native method as native"); 844 __ bind(L); 845 } 846 { Label L; 847 __ btst(JVM_ACC_ABSTRACT, Gtmp1); 848 __ br(Assembler::zero, false, Assembler::pt, L); 849 __ delayed()->nop(); 850 __ stop("tried to execute abstract method as non-abstract"); 851 __ bind(L); 852 } 853 #endif // ASSERT 854 855 // generate the code to allocate the interpreter stack frame 856 generate_fixed_frame(true); 857 858 // 859 // No locals to initialize for native method 860 // 861 862 // this slot will be set later, we initialize it to null here just in 863 // case we get a GC before the actual value is stored later 864 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS); 865 866 const Address do_not_unlock_if_synchronized(G2_thread, 867 JavaThread::do_not_unlock_if_synchronized_offset()); 868 // Since at this point in the method invocation the exception handler 869 // would try to exit the monitor of synchronized methods which hasn't 870 // been entered yet, we set the thread local variable 871 // _do_not_unlock_if_synchronized to true. If any exception was thrown by 872 // runtime, exception handling i.e. unlock_if_synchronized_method will 873 // check this thread local flag. 874 // This flag has two effects, one is to force an unwind in the topmost 875 // interpreter frame and not perform an unlock while doing so. 876 877 __ movbool(true, G3_scratch); 878 __ stbool(G3_scratch, do_not_unlock_if_synchronized); 879 880 // increment invocation counter and check for overflow 881 // 882 // Note: checking for negative value instead of overflow 883 // so we have a 'sticky' overflow test (may be of 884 // importance as soon as we have true MT/MP) 885 Label invocation_counter_overflow; 886 Label Lcontinue; 887 if (inc_counter) { 888 generate_counter_incr(&invocation_counter_overflow, NULL, NULL); 889 890 } 891 __ bind(Lcontinue); 892 893 bang_stack_shadow_pages(true); 894 895 // reset the _do_not_unlock_if_synchronized flag 896 __ stbool(G0, do_not_unlock_if_synchronized); 897 898 // check for synchronized methods 899 // Must happen AFTER invocation_counter check and stack overflow check, 900 // so method is not locked if overflows. 901 902 if (synchronized) { 903 lock_method(); 904 } else { 905 #ifdef ASSERT 906 { Label ok; 907 __ ld(Laccess_flags, O0); 908 __ btst(JVM_ACC_SYNCHRONIZED, O0); 909 __ br( Assembler::zero, false, Assembler::pt, ok); 910 __ delayed()->nop(); 911 __ stop("method needs synchronization"); 912 __ bind(ok); 913 } 914 #endif // ASSERT 915 } 916 917 918 // start execution 919 __ verify_thread(); 920 921 // JVMTI support 922 __ notify_method_entry(); 923 924 // native call 925 926 // (note that O0 is never an oop--at most it is a handle) 927 // It is important not to smash any handles created by this call, 928 // until any oop handle in O0 is dereferenced. 929 930 // (note that the space for outgoing params is preallocated) 931 932 // get signature handler 933 { Label L; 934 Address signature_handler(Lmethod, Method::signature_handler_offset()); 935 __ ld_ptr(signature_handler, G3_scratch); 936 __ br_notnull_short(G3_scratch, Assembler::pt, L); 937 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod); 938 __ ld_ptr(signature_handler, G3_scratch); 939 __ bind(L); 940 } 941 942 // Push a new frame so that the args will really be stored in 943 // Copy a few locals across so the new frame has the variables 944 // we need but these values will be dead at the jni call and 945 // therefore not gc volatile like the values in the current 946 // frame (Lmethod in particular) 947 948 // Flush the method pointer to the register save area 949 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS); 950 __ mov(Llocals, O1); 951 952 // calculate where the mirror handle body is allocated in the interpreter frame: 953 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2); 954 955 // Calculate current frame size 956 __ sub(SP, FP, O3); // Calculate negative of current frame size 957 __ save(SP, O3, SP); // Allocate an identical sized frame 958 959 // Note I7 has leftover trash. Slow signature handler will fill it in 960 // should we get there. Normal jni call will set reasonable last_Java_pc 961 // below (and fix I7 so the stack trace doesn't have a meaningless frame 962 // in it). 963 964 // Load interpreter frame's Lmethod into same register here 965 966 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod); 967 968 __ mov(I1, Llocals); 969 __ mov(I2, Lscratch2); // save the address of the mirror 970 971 972 // ONLY Lmethod and Llocals are valid here! 973 974 // call signature handler, It will move the arg properly since Llocals in current frame 975 // matches that in outer frame 976 977 __ callr(G3_scratch, 0); 978 __ delayed()->nop(); 979 980 // Result handler is in Lscratch 981 982 // Reload interpreter frame's Lmethod since slow signature handler may block 983 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod); 984 985 { Label not_static; 986 987 __ ld(Laccess_flags, O0); 988 __ btst(JVM_ACC_STATIC, O0); 989 __ br( Assembler::zero, false, Assembler::pt, not_static); 990 // get native function entry point(O0 is a good temp until the very end) 991 __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0); 992 // for static methods insert the mirror argument 993 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 994 995 __ ld_ptr(Lmethod, Method:: const_offset(), O1); 996 __ ld_ptr(O1, ConstMethod::constants_offset(), O1); 997 __ ld_ptr(O1, ConstantPool::pool_holder_offset_in_bytes(), O1); 998 __ ld_ptr(O1, mirror_offset, O1); 999 #ifdef ASSERT 1000 if (!PrintSignatureHandlers) // do not dirty the output with this 1001 { Label L; 1002 __ br_notnull_short(O1, Assembler::pt, L); 1003 __ stop("mirror is missing"); 1004 __ bind(L); 1005 } 1006 #endif // ASSERT 1007 __ st_ptr(O1, Lscratch2, 0); 1008 __ mov(Lscratch2, O1); 1009 __ bind(not_static); 1010 } 1011 1012 // At this point, arguments have been copied off of stack into 1013 // their JNI positions, which are O1..O5 and SP[68..]. 1014 // Oops are boxed in-place on the stack, with handles copied to arguments. 1015 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*. 1016 1017 #ifdef ASSERT 1018 { Label L; 1019 __ br_notnull_short(O0, Assembler::pt, L); 1020 __ stop("native entry point is missing"); 1021 __ bind(L); 1022 } 1023 #endif // ASSERT 1024 1025 // 1026 // setup the frame anchor 1027 // 1028 // The scavenge function only needs to know that the PC of this frame is 1029 // in the interpreter method entry code, it doesn't need to know the exact 1030 // PC and hence we can use O7 which points to the return address from the 1031 // previous call in the code stream (signature handler function) 1032 // 1033 // The other trick is we set last_Java_sp to FP instead of the usual SP because 1034 // we have pushed the extra frame in order to protect the volatile register(s) 1035 // in that frame when we return from the jni call 1036 // 1037 1038 __ set_last_Java_frame(FP, O7); 1039 __ mov(O7, I7); // make dummy interpreter frame look like one above, 1040 // not meaningless information that'll confuse me. 1041 1042 // flush the windows now. We don't care about the current (protection) frame 1043 // only the outer frames 1044 1045 __ flush_windows(); 1046 1047 // mark windows as flushed 1048 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset()); 1049 __ set(JavaFrameAnchor::flushed, G3_scratch); 1050 __ st(G3_scratch, flags); 1051 1052 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready. 1053 1054 Address thread_state(G2_thread, JavaThread::thread_state_offset()); 1055 #ifdef ASSERT 1056 { Label L; 1057 __ ld(thread_state, G3_scratch); 1058 __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L); 1059 __ stop("Wrong thread state in native stub"); 1060 __ bind(L); 1061 } 1062 #endif // ASSERT 1063 __ set(_thread_in_native, G3_scratch); 1064 __ st(G3_scratch, thread_state); 1065 1066 // Call the jni method, using the delay slot to set the JNIEnv* argument. 1067 __ save_thread(L7_thread_cache); // save Gthread 1068 __ callr(O0, 0); 1069 __ delayed()-> 1070 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0); 1071 1072 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD 1073 1074 __ restore_thread(L7_thread_cache); // restore G2_thread 1075 __ reinit_heapbase(); 1076 1077 // must we block? 1078 1079 // Block, if necessary, before resuming in _thread_in_Java state. 1080 // In order for GC to work, don't clear the last_Java_sp until after blocking. 1081 { Label no_block; 1082 AddressLiteral sync_state(SafepointSynchronize::address_of_state()); 1083 1084 // Switch thread to "native transition" state before reading the synchronization state. 1085 // This additional state is necessary because reading and testing the synchronization 1086 // state is not atomic w.r.t. GC, as this scenario demonstrates: 1087 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 1088 // VM thread changes sync state to synchronizing and suspends threads for GC. 1089 // Thread A is resumed to finish this native method, but doesn't block here since it 1090 // didn't see any synchronization is progress, and escapes. 1091 __ set(_thread_in_native_trans, G3_scratch); 1092 __ st(G3_scratch, thread_state); 1093 if(os::is_MP()) { 1094 if (UseMembar) { 1095 // Force this write out before the read below 1096 __ membar(Assembler::StoreLoad); 1097 } else { 1098 // Write serialization page so VM thread can do a pseudo remote membar. 1099 // We use the current thread pointer to calculate a thread specific 1100 // offset to write to within the page. This minimizes bus traffic 1101 // due to cache line collision. 1102 __ serialize_memory(G2_thread, G1_scratch, G3_scratch); 1103 } 1104 } 1105 __ load_contents(sync_state, G3_scratch); 1106 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized); 1107 1108 Label L; 1109 __ br(Assembler::notEqual, false, Assembler::pn, L); 1110 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch); 1111 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block); 1112 __ bind(L); 1113 1114 // Block. Save any potential method result value before the operation and 1115 // use a leaf call to leave the last_Java_frame setup undisturbed. 1116 save_native_result(); 1117 __ call_VM_leaf(L7_thread_cache, 1118 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans), 1119 G2_thread); 1120 1121 // Restore any method result value 1122 restore_native_result(); 1123 __ bind(no_block); 1124 } 1125 1126 // Clear the frame anchor now 1127 1128 __ reset_last_Java_frame(); 1129 1130 // Move the result handler address 1131 __ mov(Lscratch, G3_scratch); 1132 // return possible result to the outer frame 1133 #ifndef __LP64 1134 __ mov(O0, I0); 1135 __ restore(O1, G0, O1); 1136 #else 1137 __ restore(O0, G0, O0); 1138 #endif /* __LP64 */ 1139 1140 // Move result handler to expected register 1141 __ mov(G3_scratch, Lscratch); 1142 1143 // Back in normal (native) interpreter frame. State is thread_in_native_trans 1144 // switch to thread_in_Java. 1145 1146 __ set(_thread_in_Java, G3_scratch); 1147 __ st(G3_scratch, thread_state); 1148 1149 // reset handle block 1150 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch); 1151 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes()); 1152 1153 // If we have an oop result store it where it will be safe for any further gc 1154 // until we return now that we've released the handle it might be protected by 1155 1156 { 1157 Label no_oop, store_result; 1158 1159 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch); 1160 __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop); 1161 __ addcc(G0, O0, O0); 1162 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL: 1163 __ delayed()->ld_ptr(O0, 0, O0); // unbox it 1164 __ mov(G0, O0); 1165 1166 __ bind(store_result); 1167 // Store it where gc will look for it and result handler expects it. 1168 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS); 1169 1170 __ bind(no_oop); 1171 1172 } 1173 1174 1175 // handle exceptions (exception handling will handle unlocking!) 1176 { Label L; 1177 Address exception_addr(G2_thread, Thread::pending_exception_offset()); 1178 __ ld_ptr(exception_addr, Gtemp); 1179 __ br_null_short(Gtemp, Assembler::pt, L); 1180 // Note: This could be handled more efficiently since we know that the native 1181 // method doesn't have an exception handler. We could directly return 1182 // to the exception handler for the caller. 1183 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception)); 1184 __ should_not_reach_here(); 1185 __ bind(L); 1186 } 1187 1188 // JVMTI support (preserves thread register) 1189 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI); 1190 1191 if (synchronized) { 1192 // save and restore any potential method result value around the unlocking operation 1193 save_native_result(); 1194 1195 __ add( __ top_most_monitor(), O1); 1196 __ unlock_object(O1); 1197 1198 restore_native_result(); 1199 } 1200 1201 #if defined(COMPILER2) && !defined(_LP64) 1202 1203 // C2 expects long results in G1 we can't tell if we're returning to interpreted 1204 // or compiled so just be safe. 1205 1206 __ sllx(O0, 32, G1); // Shift bits into high G1 1207 __ srl (O1, 0, O1); // Zero extend O1 1208 __ or3 (O1, G1, G1); // OR 64 bits into G1 1209 1210 #endif /* COMPILER2 && !_LP64 */ 1211 1212 // dispose of return address and remove activation 1213 #ifdef ASSERT 1214 { 1215 Label ok; 1216 __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok); 1217 __ stop("bad I5_savedSP value"); 1218 __ should_not_reach_here(); 1219 __ bind(ok); 1220 } 1221 #endif 1222 if (TraceJumps) { 1223 // Move target to register that is recordable 1224 __ mov(Lscratch, G3_scratch); 1225 __ JMP(G3_scratch, 0); 1226 } else { 1227 __ jmp(Lscratch, 0); 1228 } 1229 __ delayed()->nop(); 1230 1231 1232 if (inc_counter) { 1233 // handle invocation counter overflow 1234 __ bind(invocation_counter_overflow); 1235 generate_counter_overflow(Lcontinue); 1236 } 1237 1238 1239 1240 return entry; 1241 } 1242 1243 1244 // Generic method entry to (asm) interpreter 1245 //------------------------------------------------------------------------------------------------------------------------ 1246 // 1247 address InterpreterGenerator::generate_normal_entry(bool synchronized) { 1248 address entry = __ pc(); 1249 1250 bool inc_counter = UseCompiler || CountCompiledCalls; 1251 1252 // the following temporary registers are used during frame creation 1253 const Register Gtmp1 = G3_scratch ; 1254 const Register Gtmp2 = G1_scratch; 1255 1256 // make sure registers are different! 1257 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2); 1258 1259 const Address size_of_parameters(G5_method, Method::size_of_parameters_offset()); 1260 const Address size_of_locals (G5_method, Method::size_of_locals_offset()); 1261 // Seems like G5_method is live at the point this is used. So we could make this look consistent 1262 // and use in the asserts. 1263 const Address access_flags (Lmethod, Method::access_flags_offset()); 1264 1265 const Register Glocals_size = G3; 1266 assert_different_registers(Glocals_size, G4_scratch, Gframe_size); 1267 1268 // make sure method is not native & not abstract 1269 // rethink these assertions - they can be simplified and shared (gri 2/25/2000) 1270 #ifdef ASSERT 1271 __ ld(G5_method, Method::access_flags_offset(), Gtmp1); 1272 { 1273 Label L; 1274 __ btst(JVM_ACC_NATIVE, Gtmp1); 1275 __ br(Assembler::zero, false, Assembler::pt, L); 1276 __ delayed()->nop(); 1277 __ stop("tried to execute native method as non-native"); 1278 __ bind(L); 1279 } 1280 { Label L; 1281 __ btst(JVM_ACC_ABSTRACT, Gtmp1); 1282 __ br(Assembler::zero, false, Assembler::pt, L); 1283 __ delayed()->nop(); 1284 __ stop("tried to execute abstract method as non-abstract"); 1285 __ bind(L); 1286 } 1287 #endif // ASSERT 1288 1289 // generate the code to allocate the interpreter stack frame 1290 1291 generate_fixed_frame(false); 1292 1293 #ifdef FAST_DISPATCH 1294 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables); 1295 // set bytecode dispatch table base 1296 #endif 1297 1298 // 1299 // Code to initialize the extra (i.e. non-parm) locals 1300 // 1301 Register init_value = noreg; // will be G0 if we must clear locals 1302 // The way the code was setup before zerolocals was always true for vanilla java entries. 1303 // It could only be false for the specialized entries like accessor or empty which have 1304 // no extra locals so the testing was a waste of time and the extra locals were always 1305 // initialized. We removed this extra complication to already over complicated code. 1306 1307 init_value = G0; 1308 Label clear_loop; 1309 1310 // NOTE: If you change the frame layout, this code will need to 1311 // be updated! 1312 __ lduh( size_of_locals, O2 ); 1313 __ lduh( size_of_parameters, O1 ); 1314 __ sll( O2, Interpreter::logStackElementSize, O2); 1315 __ sll( O1, Interpreter::logStackElementSize, O1 ); 1316 __ sub( Llocals, O2, O2 ); 1317 __ sub( Llocals, O1, O1 ); 1318 1319 __ bind( clear_loop ); 1320 __ inc( O2, wordSize ); 1321 1322 __ cmp( O2, O1 ); 1323 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop ); 1324 __ delayed()->st_ptr( init_value, O2, 0 ); 1325 1326 const Address do_not_unlock_if_synchronized(G2_thread, 1327 JavaThread::do_not_unlock_if_synchronized_offset()); 1328 // Since at this point in the method invocation the exception handler 1329 // would try to exit the monitor of synchronized methods which hasn't 1330 // been entered yet, we set the thread local variable 1331 // _do_not_unlock_if_synchronized to true. If any exception was thrown by 1332 // runtime, exception handling i.e. unlock_if_synchronized_method will 1333 // check this thread local flag. 1334 __ movbool(true, G3_scratch); 1335 __ stbool(G3_scratch, do_not_unlock_if_synchronized); 1336 1337 // increment invocation counter and check for overflow 1338 // 1339 // Note: checking for negative value instead of overflow 1340 // so we have a 'sticky' overflow test (may be of 1341 // importance as soon as we have true MT/MP) 1342 Label invocation_counter_overflow; 1343 Label profile_method; 1344 Label profile_method_continue; 1345 Label Lcontinue; 1346 if (inc_counter) { 1347 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue); 1348 if (ProfileInterpreter) { 1349 __ bind(profile_method_continue); 1350 } 1351 } 1352 __ bind(Lcontinue); 1353 1354 bang_stack_shadow_pages(false); 1355 1356 // reset the _do_not_unlock_if_synchronized flag 1357 __ stbool(G0, do_not_unlock_if_synchronized); 1358 1359 // check for synchronized methods 1360 // Must happen AFTER invocation_counter check and stack overflow check, 1361 // so method is not locked if overflows. 1362 1363 if (synchronized) { 1364 lock_method(); 1365 } else { 1366 #ifdef ASSERT 1367 { Label ok; 1368 __ ld(access_flags, O0); 1369 __ btst(JVM_ACC_SYNCHRONIZED, O0); 1370 __ br( Assembler::zero, false, Assembler::pt, ok); 1371 __ delayed()->nop(); 1372 __ stop("method needs synchronization"); 1373 __ bind(ok); 1374 } 1375 #endif // ASSERT 1376 } 1377 1378 // start execution 1379 1380 __ verify_thread(); 1381 1382 // jvmti support 1383 __ notify_method_entry(); 1384 1385 // start executing instructions 1386 __ dispatch_next(vtos); 1387 1388 1389 if (inc_counter) { 1390 if (ProfileInterpreter) { 1391 // We have decided to profile this method in the interpreter 1392 __ bind(profile_method); 1393 1394 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method)); 1395 __ set_method_data_pointer_for_bcp(); 1396 __ ba_short(profile_method_continue); 1397 } 1398 1399 // handle invocation counter overflow 1400 __ bind(invocation_counter_overflow); 1401 generate_counter_overflow(Lcontinue); 1402 } 1403 1404 1405 return entry; 1406 } 1407 1408 1409 //---------------------------------------------------------------------------------------------------- 1410 // Entry points & stack frame layout 1411 // 1412 // Here we generate the various kind of entries into the interpreter. 1413 // The two main entry type are generic bytecode methods and native call method. 1414 // These both come in synchronized and non-synchronized versions but the 1415 // frame layout they create is very similar. The other method entry 1416 // types are really just special purpose entries that are really entry 1417 // and interpretation all in one. These are for trivial methods like 1418 // accessor, empty, or special math methods. 1419 // 1420 // When control flow reaches any of the entry types for the interpreter 1421 // the following holds -> 1422 // 1423 // C2 Calling Conventions: 1424 // 1425 // The entry code below assumes that the following registers are set 1426 // when coming in: 1427 // G5_method: holds the Method* of the method to call 1428 // Lesp: points to the TOS of the callers expression stack 1429 // after having pushed all the parameters 1430 // 1431 // The entry code does the following to setup an interpreter frame 1432 // pop parameters from the callers stack by adjusting Lesp 1433 // set O0 to Lesp 1434 // compute X = (max_locals - num_parameters) 1435 // bump SP up by X to accomadate the extra locals 1436 // compute X = max_expression_stack 1437 // + vm_local_words 1438 // + 16 words of register save area 1439 // save frame doing a save sp, -X, sp growing towards lower addresses 1440 // set Lbcp, Lmethod, LcpoolCache 1441 // set Llocals to i0 1442 // set Lmonitors to FP - rounded_vm_local_words 1443 // set Lesp to Lmonitors - 4 1444 // 1445 // The frame has now been setup to do the rest of the entry code 1446 1447 // Try this optimization: Most method entries could live in a 1448 // "one size fits all" stack frame without all the dynamic size 1449 // calculations. It might be profitable to do all this calculation 1450 // statically and approximately for "small enough" methods. 1451 1452 //----------------------------------------------------------------------------------------------- 1453 1454 // C1 Calling conventions 1455 // 1456 // Upon method entry, the following registers are setup: 1457 // 1458 // g2 G2_thread: current thread 1459 // g5 G5_method: method to activate 1460 // g4 Gargs : pointer to last argument 1461 // 1462 // 1463 // Stack: 1464 // 1465 // +---------------+ <--- sp 1466 // | | 1467 // : reg save area : 1468 // | | 1469 // +---------------+ <--- sp + 0x40 1470 // | | 1471 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later) 1472 // | | 1473 // +---------------+ <--- sp + 0x5c 1474 // | | 1475 // : free : 1476 // | | 1477 // +---------------+ <--- Gargs 1478 // | | 1479 // : arguments : 1480 // | | 1481 // +---------------+ 1482 // | | 1483 // 1484 // 1485 // 1486 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like: 1487 // 1488 // +---------------+ <--- sp 1489 // | | 1490 // : reg save area : 1491 // | | 1492 // +---------------+ <--- sp + 0x40 1493 // | | 1494 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later) 1495 // | | 1496 // +---------------+ <--- sp + 0x5c 1497 // | | 1498 // : : 1499 // | | <--- Lesp 1500 // +---------------+ <--- Lmonitors (fp - 0x18) 1501 // | VM locals | 1502 // +---------------+ <--- fp 1503 // | | 1504 // : reg save area : 1505 // | | 1506 // +---------------+ <--- fp + 0x40 1507 // | | 1508 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later) 1509 // | | 1510 // +---------------+ <--- fp + 0x5c 1511 // | | 1512 // : free : 1513 // | | 1514 // +---------------+ 1515 // | | 1516 // : nonarg locals : 1517 // | | 1518 // +---------------+ 1519 // | | 1520 // : arguments : 1521 // | | <--- Llocals 1522 // +---------------+ <--- Gargs 1523 // | | 1524 1525 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) { 1526 1527 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated 1528 // expression stack, the callee will have callee_extra_locals (so we can account for 1529 // frame extension) and monitor_size for monitors. Basically we need to calculate 1530 // this exactly like generate_fixed_frame/generate_compute_interpreter_state. 1531 // 1532 // 1533 // The big complicating thing here is that we must ensure that the stack stays properly 1534 // aligned. This would be even uglier if monitor size wasn't modulo what the stack 1535 // needs to be aligned for). We are given that the sp (fp) is already aligned by 1536 // the caller so we must ensure that it is properly aligned for our callee. 1537 // 1538 const int rounded_vm_local_words = 1539 round_to(frame::interpreter_frame_vm_local_words,WordsPerLong); 1540 // callee_locals and max_stack are counts, not the size in frame. 1541 const int locals_size = 1542 round_to(callee_extra_locals * Interpreter::stackElementWords, WordsPerLong); 1543 const int max_stack_words = max_stack * Interpreter::stackElementWords; 1544 return (round_to((max_stack_words 1545 //6815692//+ Method::extra_stack_words() 1546 + rounded_vm_local_words 1547 + frame::memory_parameter_word_sp_offset), WordsPerLong) 1548 // already rounded 1549 + locals_size + monitor_size); 1550 } 1551 1552 // How much stack a method top interpreter activation needs in words. 1553 int AbstractInterpreter::size_top_interpreter_activation(Method* method) { 1554 1555 // See call_stub code 1556 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset, 1557 WordsPerLong); // 7 + register save area 1558 1559 // Save space for one monitor to get into the interpreted method in case 1560 // the method is synchronized 1561 int monitor_size = method->is_synchronized() ? 1562 1*frame::interpreter_frame_monitor_size() : 0; 1563 return size_activation_helper(method->max_locals(), method->max_stack(), 1564 monitor_size) + call_stub_size; 1565 } 1566 1567 int AbstractInterpreter::layout_activation(Method* method, 1568 int tempcount, 1569 int popframe_extra_args, 1570 int moncount, 1571 int caller_actual_parameters, 1572 int callee_param_count, 1573 int callee_local_count, 1574 frame* caller, 1575 frame* interpreter_frame, 1576 bool is_top_frame) { 1577 // Note: This calculation must exactly parallel the frame setup 1578 // in InterpreterGenerator::generate_fixed_frame. 1579 // If f!=NULL, set up the following variables: 1580 // - Lmethod 1581 // - Llocals 1582 // - Lmonitors (to the indicated number of monitors) 1583 // - Lesp (to the indicated number of temps) 1584 // The frame f (if not NULL) on entry is a description of the caller of the frame 1585 // we are about to layout. We are guaranteed that we will be able to fill in a 1586 // new interpreter frame as its callee (i.e. the stack space is allocated and 1587 // the amount was determined by an earlier call to this method with f == NULL). 1588 // On return f (if not NULL) while describe the interpreter frame we just layed out. 1589 1590 int monitor_size = moncount * frame::interpreter_frame_monitor_size(); 1591 int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong); 1592 1593 assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align"); 1594 // 1595 // Note: if you look closely this appears to be doing something much different 1596 // than generate_fixed_frame. What is happening is this. On sparc we have to do 1597 // this dance with interpreter_sp_adjustment because the window save area would 1598 // appear just below the bottom (tos) of the caller's java expression stack. Because 1599 // the interpreter want to have the locals completely contiguous generate_fixed_frame 1600 // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size). 1601 // Now in generate_fixed_frame the extension of the caller's sp happens in the callee. 1602 // In this code the opposite occurs the caller adjusts it's own stack base on the callee. 1603 // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest) 1604 // because the oldest frame would have adjust its callers frame and yet that frame 1605 // already exists and isn't part of this array of frames we are unpacking. So at first 1606 // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper() 1607 // will after it calculates all of the frame's on_stack_size()'s will then figure out the 1608 // amount to adjust the caller of the initial (oldest) frame and the calculation will all 1609 // add up. It does seem like it simpler to account for the adjustment here (and remove the 1610 // callee... parameters here). However this would mean that this routine would have to take 1611 // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment) 1612 // and run the calling loop in the reverse order. This would also would appear to mean making 1613 // this code aware of what the interactions are when that initial caller fram was an osr or 1614 // other adapter frame. deoptimization is complicated enough and hard enough to debug that 1615 // there is no sense in messing working code. 1616 // 1617 1618 int rounded_cls = round_to((callee_local_count - callee_param_count), WordsPerLong); 1619 assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align"); 1620 1621 int raw_frame_size = size_activation_helper(rounded_cls, method->max_stack(), 1622 monitor_size); 1623 1624 if (interpreter_frame != NULL) { 1625 // The skeleton frame must already look like an interpreter frame 1626 // even if not fully filled out. 1627 assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame"); 1628 1629 intptr_t* fp = interpreter_frame->fp(); 1630 1631 JavaThread* thread = JavaThread::current(); 1632 RegisterMap map(thread, false); 1633 // More verification that skeleton frame is properly walkable 1634 assert(fp == caller->sp(), "fp must match"); 1635 1636 intptr_t* montop = fp - rounded_vm_local_words; 1637 1638 // preallocate monitors (cf. __ add_monitor_to_stack) 1639 intptr_t* monitors = montop - monitor_size; 1640 1641 // preallocate stack space 1642 intptr_t* esp = monitors - 1 - 1643 (tempcount * Interpreter::stackElementWords) - 1644 popframe_extra_args; 1645 1646 int local_words = method->max_locals() * Interpreter::stackElementWords; 1647 NEEDS_CLEANUP; 1648 intptr_t* locals; 1649 if (caller->is_interpreted_frame()) { 1650 // Can force the locals area to end up properly overlapping the top of the expression stack. 1651 intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1; 1652 // Note that this computation means we replace size_of_parameters() values from the caller 1653 // interpreter frame's expression stack with our argument locals 1654 int parm_words = caller_actual_parameters * Interpreter::stackElementWords; 1655 locals = Lesp_ptr + parm_words; 1656 int delta = local_words - parm_words; 1657 int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0; 1658 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS; 1659 } else { 1660 assert(caller->is_compiled_frame() || caller->is_entry_frame(), "only possible cases"); 1661 // Don't have Lesp available; lay out locals block in the caller 1662 // adjacent to the register window save area. 1663 // 1664 // Compiled frames do not allocate a varargs area which is why this if 1665 // statement is needed. 1666 // 1667 if (caller->is_compiled_frame()) { 1668 locals = fp + frame::register_save_words + local_words - 1; 1669 } else { 1670 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1; 1671 } 1672 if (!caller->is_entry_frame()) { 1673 // Caller wants his own SP back 1674 int caller_frame_size = caller->cb()->frame_size(); 1675 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS; 1676 } 1677 } 1678 if (TraceDeoptimization) { 1679 if (caller->is_entry_frame()) { 1680 // make sure I5_savedSP and the entry frames notion of saved SP 1681 // agree. This assertion duplicate a check in entry frame code 1682 // but catches the failure earlier. 1683 assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP), 1684 "would change callers SP"); 1685 } 1686 if (caller->is_entry_frame()) { 1687 tty->print("entry "); 1688 } 1689 if (caller->is_compiled_frame()) { 1690 tty->print("compiled "); 1691 if (caller->is_deoptimized_frame()) { 1692 tty->print("(deopt) "); 1693 } 1694 } 1695 if (caller->is_interpreted_frame()) { 1696 tty->print("interpreted "); 1697 } 1698 tty->print_cr("caller fp=0x%x sp=0x%x", caller->fp(), caller->sp()); 1699 tty->print_cr("save area = 0x%x, 0x%x", caller->sp(), caller->sp() + 16); 1700 tty->print_cr("save area = 0x%x, 0x%x", caller->fp(), caller->fp() + 16); 1701 tty->print_cr("interpreter fp=0x%x sp=0x%x", interpreter_frame->fp(), interpreter_frame->sp()); 1702 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->sp(), interpreter_frame->sp() + 16); 1703 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->fp(), interpreter_frame->fp() + 16); 1704 tty->print_cr("Llocals = 0x%x", locals); 1705 tty->print_cr("Lesp = 0x%x", esp); 1706 tty->print_cr("Lmonitors = 0x%x", monitors); 1707 } 1708 1709 if (method->max_locals() > 0) { 1710 assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area"); 1711 assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area"); 1712 assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area"); 1713 assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area"); 1714 } 1715 #ifdef _LP64 1716 assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd"); 1717 #endif 1718 1719 *interpreter_frame->register_addr(Lmethod) = (intptr_t) method; 1720 *interpreter_frame->register_addr(Llocals) = (intptr_t) locals; 1721 *interpreter_frame->register_addr(Lmonitors) = (intptr_t) monitors; 1722 *interpreter_frame->register_addr(Lesp) = (intptr_t) esp; 1723 // Llast_SP will be same as SP as there is no adapter space 1724 *interpreter_frame->register_addr(Llast_SP) = (intptr_t) interpreter_frame->sp() - STACK_BIAS; 1725 *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache(); 1726 #ifdef FAST_DISPATCH 1727 *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table(); 1728 #endif 1729 1730 1731 #ifdef ASSERT 1732 BasicObjectLock* mp = (BasicObjectLock*)monitors; 1733 1734 assert(interpreter_frame->interpreter_frame_method() == method, "method matches"); 1735 assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize)), "locals match"); 1736 assert(interpreter_frame->interpreter_frame_monitor_end() == mp, "monitor_end matches"); 1737 assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches"); 1738 assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches"); 1739 1740 // check bounds 1741 intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1); 1742 intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words; 1743 assert(lo < monitors && montop <= hi, "monitors in bounds"); 1744 assert(lo <= esp && esp < monitors, "esp in bounds"); 1745 #endif // ASSERT 1746 } 1747 1748 return raw_frame_size; 1749 } 1750 1751 //---------------------------------------------------------------------------------------------------- 1752 // Exceptions 1753 void TemplateInterpreterGenerator::generate_throw_exception() { 1754 1755 // Entry point in previous activation (i.e., if the caller was interpreted) 1756 Interpreter::_rethrow_exception_entry = __ pc(); 1757 // O0: exception 1758 1759 // entry point for exceptions thrown within interpreter code 1760 Interpreter::_throw_exception_entry = __ pc(); 1761 __ verify_thread(); 1762 // expression stack is undefined here 1763 // O0: exception, i.e. Oexception 1764 // Lbcp: exception bcx 1765 __ verify_oop(Oexception); 1766 1767 1768 // expression stack must be empty before entering the VM in case of an exception 1769 __ empty_expression_stack(); 1770 // find exception handler address and preserve exception oop 1771 // call C routine to find handler and jump to it 1772 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception); 1773 __ push_ptr(O1); // push exception for exception handler bytecodes 1774 1775 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!) 1776 __ delayed()->nop(); 1777 1778 1779 // if the exception is not handled in the current frame 1780 // the frame is removed and the exception is rethrown 1781 // (i.e. exception continuation is _rethrow_exception) 1782 // 1783 // Note: At this point the bci is still the bxi for the instruction which caused 1784 // the exception and the expression stack is empty. Thus, for any VM calls 1785 // at this point, GC will find a legal oop map (with empty expression stack). 1786 1787 // in current activation 1788 // tos: exception 1789 // Lbcp: exception bcp 1790 1791 // 1792 // JVMTI PopFrame support 1793 // 1794 1795 Interpreter::_remove_activation_preserving_args_entry = __ pc(); 1796 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset()); 1797 // Set the popframe_processing bit in popframe_condition indicating that we are 1798 // currently handling popframe, so that call_VMs that may happen later do not trigger new 1799 // popframe handling cycles. 1800 1801 __ ld(popframe_condition_addr, G3_scratch); 1802 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch); 1803 __ stw(G3_scratch, popframe_condition_addr); 1804 1805 // Empty the expression stack, as in normal exception handling 1806 __ empty_expression_stack(); 1807 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false); 1808 1809 { 1810 // Check to see whether we are returning to a deoptimized frame. 1811 // (The PopFrame call ensures that the caller of the popped frame is 1812 // either interpreted or compiled and deoptimizes it if compiled.) 1813 // In this case, we can't call dispatch_next() after the frame is 1814 // popped, but instead must save the incoming arguments and restore 1815 // them after deoptimization has occurred. 1816 // 1817 // Note that we don't compare the return PC against the 1818 // deoptimization blob's unpack entry because of the presence of 1819 // adapter frames in C2. 1820 Label caller_not_deoptimized; 1821 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7); 1822 __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized); 1823 1824 const Register Gtmp1 = G3_scratch; 1825 const Register Gtmp2 = G1_scratch; 1826 1827 // Compute size of arguments for saving when returning to deoptimized caller 1828 __ lduh(Lmethod, in_bytes(Method::size_of_parameters_offset()), Gtmp1); 1829 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1); 1830 __ sub(Llocals, Gtmp1, Gtmp2); 1831 __ add(Gtmp2, wordSize, Gtmp2); 1832 // Save these arguments 1833 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2); 1834 // Inform deoptimization that it is responsible for restoring these arguments 1835 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1); 1836 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset()); 1837 __ st(Gtmp1, popframe_condition_addr); 1838 1839 // Return from the current method 1840 // The caller's SP was adjusted upon method entry to accomodate 1841 // the callee's non-argument locals. Undo that adjustment. 1842 __ ret(); 1843 __ delayed()->restore(I5_savedSP, G0, SP); 1844 1845 __ bind(caller_not_deoptimized); 1846 } 1847 1848 // Clear the popframe condition flag 1849 __ stw(G0 /* popframe_inactive */, popframe_condition_addr); 1850 1851 // Get out of the current method (how this is done depends on the particular compiler calling 1852 // convention that the interpreter currently follows) 1853 // The caller's SP was adjusted upon method entry to accomodate 1854 // the callee's non-argument locals. Undo that adjustment. 1855 __ restore(I5_savedSP, G0, SP); 1856 // The method data pointer was incremented already during 1857 // call profiling. We have to restore the mdp for the current bcp. 1858 if (ProfileInterpreter) { 1859 __ set_method_data_pointer_for_bcp(); 1860 } 1861 // Resume bytecode interpretation at the current bcp 1862 __ dispatch_next(vtos); 1863 // end of JVMTI PopFrame support 1864 1865 Interpreter::_remove_activation_entry = __ pc(); 1866 1867 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here) 1868 __ pop_ptr(Oexception); // get exception 1869 1870 // Intel has the following comment: 1871 //// remove the activation (without doing throws on illegalMonitorExceptions) 1872 // They remove the activation without checking for bad monitor state. 1873 // %%% We should make sure this is the right semantics before implementing. 1874 1875 __ set_vm_result(Oexception); 1876 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false); 1877 1878 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI); 1879 1880 __ get_vm_result(Oexception); 1881 __ verify_oop(Oexception); 1882 1883 const int return_reg_adjustment = frame::pc_return_offset; 1884 Address issuing_pc_addr(I7, return_reg_adjustment); 1885 1886 // We are done with this activation frame; find out where to go next. 1887 // The continuation point will be an exception handler, which expects 1888 // the following registers set up: 1889 // 1890 // Oexception: exception 1891 // Oissuing_pc: the local call that threw exception 1892 // Other On: garbage 1893 // In/Ln: the contents of the caller's register window 1894 // 1895 // We do the required restore at the last possible moment, because we 1896 // need to preserve some state across a runtime call. 1897 // (Remember that the caller activation is unknown--it might not be 1898 // interpreted, so things like Lscratch are useless in the caller.) 1899 1900 // Although the Intel version uses call_C, we can use the more 1901 // compact call_VM. (The only real difference on SPARC is a 1902 // harmlessly ignored [re]set_last_Java_frame, compared with 1903 // the Intel code which lacks this.) 1904 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore 1905 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller 1906 __ super_call_VM_leaf(L7_thread_cache, 1907 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), 1908 G2_thread, Oissuing_pc->after_save()); 1909 1910 // The caller's SP was adjusted upon method entry to accomodate 1911 // the callee's non-argument locals. Undo that adjustment. 1912 __ JMP(O0, 0); // return exception handler in caller 1913 __ delayed()->restore(I5_savedSP, G0, SP); 1914 1915 // (same old exception object is already in Oexception; see above) 1916 // Note that an "issuing PC" is actually the next PC after the call 1917 } 1918 1919 1920 // 1921 // JVMTI ForceEarlyReturn support 1922 // 1923 1924 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) { 1925 address entry = __ pc(); 1926 1927 __ empty_expression_stack(); 1928 __ load_earlyret_value(state); 1929 1930 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch); 1931 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset()); 1932 1933 // Clear the earlyret state 1934 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr); 1935 1936 __ remove_activation(state, 1937 /* throw_monitor_exception */ false, 1938 /* install_monitor_exception */ false); 1939 1940 // The caller's SP was adjusted upon method entry to accomodate 1941 // the callee's non-argument locals. Undo that adjustment. 1942 __ ret(); // return to caller 1943 __ delayed()->restore(I5_savedSP, G0, SP); 1944 1945 return entry; 1946 } // end of JVMTI ForceEarlyReturn support 1947 1948 1949 //------------------------------------------------------------------------------------------------------------------------ 1950 // Helper for vtos entry point generation 1951 1952 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) { 1953 assert(t->is_valid() && t->tos_in() == vtos, "illegal template"); 1954 Label L; 1955 aep = __ pc(); __ push_ptr(); __ ba_short(L); 1956 fep = __ pc(); __ push_f(); __ ba_short(L); 1957 dep = __ pc(); __ push_d(); __ ba_short(L); 1958 lep = __ pc(); __ push_l(); __ ba_short(L); 1959 iep = __ pc(); __ push_i(); 1960 bep = cep = sep = iep; // there aren't any 1961 vep = __ pc(); __ bind(L); // fall through 1962 generate_and_dispatch(t); 1963 } 1964 1965 // -------------------------------------------------------------------------------- 1966 1967 1968 InterpreterGenerator::InterpreterGenerator(StubQueue* code) 1969 : TemplateInterpreterGenerator(code) { 1970 generate_all(); // down here so it can be "virtual" 1971 } 1972 1973 // -------------------------------------------------------------------------------- 1974 1975 // Non-product code 1976 #ifndef PRODUCT 1977 address TemplateInterpreterGenerator::generate_trace_code(TosState state) { 1978 address entry = __ pc(); 1979 1980 __ push(state); 1981 __ mov(O7, Lscratch); // protect return address within interpreter 1982 1983 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer 1984 __ mov( Otos_l2, G3_scratch ); 1985 __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch); 1986 __ mov(Lscratch, O7); // restore return address 1987 __ pop(state); 1988 __ retl(); 1989 __ delayed()->nop(); 1990 1991 return entry; 1992 } 1993 1994 1995 // helpers for generate_and_dispatch 1996 1997 void TemplateInterpreterGenerator::count_bytecode() { 1998 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch); 1999 } 2000 2001 2002 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) { 2003 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch); 2004 } 2005 2006 2007 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) { 2008 AddressLiteral index (&BytecodePairHistogram::_index); 2009 AddressLiteral counters((address) &BytecodePairHistogram::_counters); 2010 2011 // get index, shift out old bytecode, bring in new bytecode, and store it 2012 // _index = (_index >> log2_number_of_codes) | 2013 // (bytecode << log2_number_of_codes); 2014 2015 __ load_contents(index, G4_scratch); 2016 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch ); 2017 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch ); 2018 __ or3( G3_scratch, G4_scratch, G4_scratch ); 2019 __ store_contents(G4_scratch, index, G3_scratch); 2020 2021 // bump bucket contents 2022 // _counters[_index] ++; 2023 2024 __ set(counters, G3_scratch); // loads into G3_scratch 2025 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address 2026 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index 2027 __ ld (G3_scratch, 0, G4_scratch); 2028 __ inc (G4_scratch); 2029 __ st (G4_scratch, 0, G3_scratch); 2030 } 2031 2032 2033 void TemplateInterpreterGenerator::trace_bytecode(Template* t) { 2034 // Call a little run-time stub to avoid blow-up for each bytecode. 2035 // The run-time runtime saves the right registers, depending on 2036 // the tosca in-state for the given template. 2037 address entry = Interpreter::trace_code(t->tos_in()); 2038 guarantee(entry != NULL, "entry must have been generated"); 2039 __ call(entry, relocInfo::none); 2040 __ delayed()->nop(); 2041 } 2042 2043 2044 void TemplateInterpreterGenerator::stop_interpreter_at() { 2045 AddressLiteral counter(&BytecodeCounter::_counter_value); 2046 __ load_contents(counter, G3_scratch); 2047 AddressLiteral stop_at(&StopInterpreterAt); 2048 __ load_ptr_contents(stop_at, G4_scratch); 2049 __ cmp(G3_scratch, G4_scratch); 2050 __ breakpoint_trap(Assembler::equal, Assembler::icc); 2051 } 2052 #endif // not PRODUCT 2053 #endif // !CC_INTERP