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 #include "utilities/macros.hpp" 48 49 #ifndef CC_INTERP 50 #ifndef FAST_DISPATCH 51 #define FAST_DISPATCH 1 52 #endif 53 #undef FAST_DISPATCH 54 55 56 // Generation of Interpreter 57 // 58 // The InterpreterGenerator generates the interpreter into Interpreter::_code. 59 60 61 #define __ _masm-> 62 63 64 //---------------------------------------------------------------------------------------------------- 65 66 67 void InterpreterGenerator::save_native_result(void) { 68 // result potentially in O0/O1: save it across calls 69 const Address& l_tmp = InterpreterMacroAssembler::l_tmp; 70 71 // result potentially in F0/F1: save it across calls 72 const Address& d_tmp = InterpreterMacroAssembler::d_tmp; 73 74 // save and restore any potential method result value around the unlocking operation 75 __ stf(FloatRegisterImpl::D, F0, d_tmp); 76 #ifdef _LP64 77 __ stx(O0, l_tmp); 78 #else 79 __ std(O0, l_tmp); 80 #endif 81 } 82 83 void InterpreterGenerator::restore_native_result(void) { 84 const Address& l_tmp = InterpreterMacroAssembler::l_tmp; 85 const Address& d_tmp = InterpreterMacroAssembler::d_tmp; 86 87 // Restore any method result value 88 __ ldf(FloatRegisterImpl::D, d_tmp, F0); 89 #ifdef _LP64 90 __ ldx(l_tmp, O0); 91 #else 92 __ ldd(l_tmp, O0); 93 #endif 94 } 95 96 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) { 97 assert(!pass_oop || message == NULL, "either oop or message but not both"); 98 address entry = __ pc(); 99 // expression stack must be empty before entering the VM if an exception happened 100 __ empty_expression_stack(); 101 // load exception object 102 __ set((intptr_t)name, G3_scratch); 103 if (pass_oop) { 104 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i); 105 } else { 106 __ set((intptr_t)message, G4_scratch); 107 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch); 108 } 109 // throw exception 110 assert(Interpreter::throw_exception_entry() != NULL, "generate it first"); 111 AddressLiteral thrower(Interpreter::throw_exception_entry()); 112 __ jump_to(thrower, G3_scratch); 113 __ delayed()->nop(); 114 return entry; 115 } 116 117 address TemplateInterpreterGenerator::generate_ClassCastException_handler() { 118 address entry = __ pc(); 119 // expression stack must be empty before entering the VM if an exception 120 // happened 121 __ empty_expression_stack(); 122 // load exception object 123 __ call_VM(Oexception, 124 CAST_FROM_FN_PTR(address, 125 InterpreterRuntime::throw_ClassCastException), 126 Otos_i); 127 __ should_not_reach_here(); 128 return entry; 129 } 130 131 132 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) { 133 address entry = __ pc(); 134 // expression stack must be empty before entering the VM if an exception happened 135 __ empty_expression_stack(); 136 // convention: expect aberrant index in register G3_scratch, then shuffle the 137 // index to G4_scratch for the VM call 138 __ mov(G3_scratch, G4_scratch); 139 __ set((intptr_t)name, G3_scratch); 140 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch); 141 __ should_not_reach_here(); 142 return entry; 143 } 144 145 146 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() { 147 address entry = __ pc(); 148 // expression stack must be empty before entering the VM if an exception happened 149 __ empty_expression_stack(); 150 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError)); 151 __ should_not_reach_here(); 152 return entry; 153 } 154 155 156 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step) { 157 TosState incoming_state = state; 158 159 Label cont; 160 address compiled_entry = __ pc(); 161 162 address entry = __ pc(); 163 #if !defined(_LP64) && defined(COMPILER2) 164 // All return values are where we want them, except for Longs. C2 returns 165 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1. 166 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit 167 // build even if we are returning from interpreted we just do a little 168 // stupid shuffing. 169 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to 170 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node 171 // first which would move g1 -> O0/O1 and destroy the exception we were throwing. 172 173 if (incoming_state == ltos) { 174 __ srl (G1, 0, O1); 175 __ srlx(G1, 32, O0); 176 } 177 #endif // !_LP64 && COMPILER2 178 179 __ bind(cont); 180 181 // The callee returns with the stack possibly adjusted by adapter transition 182 // We remove that possible adjustment here. 183 // All interpreter local registers are untouched. Any result is passed back 184 // in the O0/O1 or float registers. Before continuing, the arguments must be 185 // popped from the java expression stack; i.e., Lesp must be adjusted. 186 187 __ mov(Llast_SP, SP); // Remove any adapter added stack space. 188 189 Label L_got_cache, L_giant_index; 190 const Register cache = G3_scratch; 191 const Register size = G1_scratch; 192 if (EnableInvokeDynamic) { 193 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode. 194 __ cmp_and_br_short(G1_scratch, Bytecodes::_invokedynamic, Assembler::equal, Assembler::pn, L_giant_index); 195 } 196 __ get_cache_and_index_at_bcp(cache, G1_scratch, 1); 197 __ bind(L_got_cache); 198 __ ld_ptr(cache, ConstantPoolCache::base_offset() + 199 ConstantPoolCacheEntry::flags_offset(), size); 200 __ and3(size, 0xFF, size); // argument size in words 201 __ sll(size, Interpreter::logStackElementSize, size); // each argument size in bytes 202 __ add(Lesp, size, Lesp); // pop arguments 203 __ dispatch_next(state, step); 204 205 // out of the main line of code... 206 if (EnableInvokeDynamic) { 207 __ bind(L_giant_index); 208 __ get_cache_and_index_at_bcp(cache, G1_scratch, 1, sizeof(u4)); 209 __ ba_short(L_got_cache); 210 } 211 212 return entry; 213 } 214 215 216 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) { 217 address entry = __ pc(); 218 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache 219 { Label L; 220 Address exception_addr(G2_thread, Thread::pending_exception_offset()); 221 __ ld_ptr(exception_addr, Gtemp); // Load pending exception. 222 __ br_null_short(Gtemp, Assembler::pt, L); 223 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception)); 224 __ should_not_reach_here(); 225 __ bind(L); 226 } 227 __ dispatch_next(state, step); 228 return entry; 229 } 230 231 // A result handler converts/unboxes a native call result into 232 // a java interpreter/compiler result. The current frame is an 233 // interpreter frame. The activation frame unwind code must be 234 // consistent with that of TemplateTable::_return(...). In the 235 // case of native methods, the caller's SP was not modified. 236 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) { 237 address entry = __ pc(); 238 Register Itos_i = Otos_i ->after_save(); 239 Register Itos_l = Otos_l ->after_save(); 240 Register Itos_l1 = Otos_l1->after_save(); 241 Register Itos_l2 = Otos_l2->after_save(); 242 switch (type) { 243 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false 244 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value! 245 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break; 246 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break; 247 case T_LONG : 248 #ifndef _LP64 249 __ mov(O1, Itos_l2); // move other half of long 250 #endif // ifdef or no ifdef, fall through to the T_INT case 251 case T_INT : __ mov(O0, Itos_i); break; 252 case T_VOID : /* nothing to do */ break; 253 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break; 254 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break; 255 case T_OBJECT : 256 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i); 257 __ verify_oop(Itos_i); 258 break; 259 default : ShouldNotReachHere(); 260 } 261 __ ret(); // return from interpreter activation 262 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame 263 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly 264 return entry; 265 } 266 267 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) { 268 address entry = __ pc(); 269 __ push(state); 270 __ call_VM(noreg, runtime_entry); 271 __ dispatch_via(vtos, Interpreter::normal_table(vtos)); 272 return entry; 273 } 274 275 276 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) { 277 address entry = __ pc(); 278 __ dispatch_next(state); 279 return entry; 280 } 281 282 // 283 // Helpers for commoning out cases in the various type of method entries. 284 // 285 286 // increment invocation count & check for overflow 287 // 288 // Note: checking for negative value instead of overflow 289 // so we have a 'sticky' overflow test 290 // 291 // Lmethod: method 292 // ??: invocation counter 293 // 294 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) { 295 // Note: In tiered we increment either counters in Method* or in MDO depending if we're profiling or not. 296 if (TieredCompilation) { 297 const int increment = InvocationCounter::count_increment; 298 const int mask = ((1 << Tier0InvokeNotifyFreqLog) - 1) << InvocationCounter::count_shift; 299 Label no_mdo, done; 300 if (ProfileInterpreter) { 301 // If no method data exists, go to profile_continue. 302 __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch); 303 __ br_null_short(G4_scratch, Assembler::pn, no_mdo); 304 // Increment counter 305 Address mdo_invocation_counter(G4_scratch, 306 in_bytes(MethodData::invocation_counter_offset()) + 307 in_bytes(InvocationCounter::counter_offset())); 308 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask, 309 G3_scratch, Lscratch, 310 Assembler::zero, overflow); 311 __ ba_short(done); 312 } 313 314 // Increment counter in Method* 315 __ bind(no_mdo); 316 Address invocation_counter(Lmethod, 317 in_bytes(Method::invocation_counter_offset()) + 318 in_bytes(InvocationCounter::counter_offset())); 319 __ increment_mask_and_jump(invocation_counter, increment, mask, 320 G3_scratch, Lscratch, 321 Assembler::zero, overflow); 322 __ bind(done); 323 } else { 324 // Update standard invocation counters 325 __ increment_invocation_counter(O0, G3_scratch); 326 if (ProfileInterpreter) { // %%% Merge this into MethodData* 327 Address interpreter_invocation_counter(Lmethod,in_bytes(Method::interpreter_invocation_counter_offset())); 328 __ ld(interpreter_invocation_counter, G3_scratch); 329 __ inc(G3_scratch); 330 __ st(G3_scratch, interpreter_invocation_counter); 331 } 332 333 if (ProfileInterpreter && profile_method != NULL) { 334 // Test to see if we should create a method data oop 335 AddressLiteral profile_limit((address)&InvocationCounter::InterpreterProfileLimit); 336 __ load_contents(profile_limit, G3_scratch); 337 __ cmp_and_br_short(O0, G3_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue); 338 339 // if no method data exists, go to profile_method 340 __ test_method_data_pointer(*profile_method); 341 } 342 343 AddressLiteral invocation_limit((address)&InvocationCounter::InterpreterInvocationLimit); 344 __ load_contents(invocation_limit, G3_scratch); 345 __ cmp(O0, G3_scratch); 346 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance 347 __ delayed()->nop(); 348 } 349 350 } 351 352 // Allocate monitor and lock method (asm interpreter) 353 // ebx - Method* 354 // 355 void InterpreterGenerator::lock_method(void) { 356 __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0); // Load access flags. 357 358 #ifdef ASSERT 359 { Label ok; 360 __ btst(JVM_ACC_SYNCHRONIZED, O0); 361 __ br( Assembler::notZero, false, Assembler::pt, ok); 362 __ delayed()->nop(); 363 __ stop("method doesn't need synchronization"); 364 __ bind(ok); 365 } 366 #endif // ASSERT 367 368 // get synchronization object to O0 369 { Label done; 370 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 371 __ btst(JVM_ACC_STATIC, O0); 372 __ br( Assembler::zero, true, Assembler::pt, done); 373 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case 374 375 __ ld_ptr( Lmethod, in_bytes(Method::const_offset()), O0); 376 __ ld_ptr( O0, in_bytes(ConstMethod::constants_offset()), O0); 377 __ ld_ptr( O0, ConstantPool::pool_holder_offset_in_bytes(), O0); 378 379 // lock the mirror, not the Klass* 380 __ ld_ptr( O0, mirror_offset, O0); 381 382 #ifdef ASSERT 383 __ tst(O0); 384 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc); 385 #endif // ASSERT 386 387 __ bind(done); 388 } 389 390 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem 391 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object 392 // __ untested("lock_object from method entry"); 393 __ lock_object(Lmonitors, O0); 394 } 395 396 397 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size, 398 Register Rscratch, 399 Register Rscratch2) { 400 const int page_size = os::vm_page_size(); 401 Label after_frame_check; 402 403 assert_different_registers(Rframe_size, Rscratch, Rscratch2); 404 405 __ set(page_size, Rscratch); 406 __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check); 407 408 // get the stack base, and in debug, verify it is non-zero 409 __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch ); 410 #ifdef ASSERT 411 Label base_not_zero; 412 __ br_notnull_short(Rscratch, Assembler::pn, base_not_zero); 413 __ stop("stack base is zero in generate_stack_overflow_check"); 414 __ bind(base_not_zero); 415 #endif 416 417 // get the stack size, and in debug, verify it is non-zero 418 assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" ); 419 __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 ); 420 #ifdef ASSERT 421 Label size_not_zero; 422 __ br_notnull_short(Rscratch2, Assembler::pn, size_not_zero); 423 __ stop("stack size is zero in generate_stack_overflow_check"); 424 __ bind(size_not_zero); 425 #endif 426 427 // compute the beginning of the protected zone minus the requested frame size 428 __ sub( Rscratch, Rscratch2, Rscratch ); 429 __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 ); 430 __ add( Rscratch, Rscratch2, Rscratch ); 431 432 // Add in the size of the frame (which is the same as subtracting it from the 433 // SP, which would take another register 434 __ add( Rscratch, Rframe_size, Rscratch ); 435 436 // the frame is greater than one page in size, so check against 437 // the bottom of the stack 438 __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check); 439 440 // the stack will overflow, throw an exception 441 442 // Note that SP is restored to sender's sp (in the delay slot). This 443 // is necessary if the sender's frame is an extended compiled frame 444 // (see gen_c2i_adapter()) and safer anyway in case of JSR292 445 // adaptations. 446 447 // Note also that the restored frame is not necessarily interpreted. 448 // Use the shared runtime version of the StackOverflowError. 449 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated"); 450 AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry()); 451 __ jump_to(stub, Rscratch); 452 __ delayed()->mov(O5_savedSP, SP); 453 454 // if you get to here, then there is enough stack space 455 __ bind( after_frame_check ); 456 } 457 458 459 // 460 // Generate a fixed interpreter frame. This is identical setup for interpreted 461 // methods and for native methods hence the shared code. 462 463 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) { 464 // 465 // 466 // The entry code sets up a new interpreter frame in 4 steps: 467 // 468 // 1) Increase caller's SP by for the extra local space needed: 469 // (check for overflow) 470 // Efficient implementation of xload/xstore bytecodes requires 471 // that arguments and non-argument locals are in a contigously 472 // addressable memory block => non-argument locals must be 473 // allocated in the caller's frame. 474 // 475 // 2) Create a new stack frame and register window: 476 // The new stack frame must provide space for the standard 477 // register save area, the maximum java expression stack size, 478 // the monitor slots (0 slots initially), and some frame local 479 // scratch locations. 480 // 481 // 3) The following interpreter activation registers must be setup: 482 // Lesp : expression stack pointer 483 // Lbcp : bytecode pointer 484 // Lmethod : method 485 // Llocals : locals pointer 486 // Lmonitors : monitor pointer 487 // LcpoolCache: constant pool cache 488 // 489 // 4) Initialize the non-argument locals if necessary: 490 // Non-argument locals may need to be initialized to NULL 491 // for GC to work. If the oop-map information is accurate 492 // (in the absence of the JSR problem), no initialization 493 // is necessary. 494 // 495 // (gri - 2/25/2000) 496 497 498 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong ); 499 500 const int extra_space = 501 rounded_vm_local_words + // frame local scratch space 502 frame::memory_parameter_word_sp_offset + // register save area 503 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0); 504 505 const Register Glocals_size = G3; 506 const Register RconstMethod = Glocals_size; 507 const Register Otmp1 = O3; 508 const Register Otmp2 = O4; 509 // Lscratch can't be used as a temporary because the call_stub uses 510 // it to assert that the stack frame was setup correctly. 511 const Address constMethod (G5_method, Method::const_offset()); 512 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset()); 513 514 __ ld_ptr( constMethod, RconstMethod ); 515 __ lduh( size_of_parameters, Glocals_size); 516 517 // Gargs points to first local + BytesPerWord 518 // Set the saved SP after the register window save 519 // 520 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP); 521 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1); 522 __ add(Gargs, Otmp1, Gargs); 523 524 if (native_call) { 525 __ calc_mem_param_words( Glocals_size, Gframe_size ); 526 __ add( Gframe_size, extra_space, Gframe_size); 527 __ round_to( Gframe_size, WordsPerLong ); 528 __ sll( Gframe_size, LogBytesPerWord, Gframe_size ); 529 } else { 530 531 // 532 // Compute number of locals in method apart from incoming parameters 533 // 534 const Address size_of_locals (Otmp1, ConstMethod::size_of_locals_offset()); 535 __ ld_ptr( constMethod, Otmp1 ); 536 __ lduh( size_of_locals, Otmp1 ); 537 __ sub( Otmp1, Glocals_size, Glocals_size ); 538 __ round_to( Glocals_size, WordsPerLong ); 539 __ sll( Glocals_size, Interpreter::logStackElementSize, Glocals_size ); 540 541 // see if the frame is greater than one page in size. If so, 542 // then we need to verify there is enough stack space remaining 543 // Frame_size = (max_stack + extra_space) * BytesPerWord; 544 __ ld_ptr( constMethod, Gframe_size ); 545 __ lduh( Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size ); 546 __ add( Gframe_size, extra_space, Gframe_size ); 547 __ round_to( Gframe_size, WordsPerLong ); 548 __ sll( Gframe_size, Interpreter::logStackElementSize, Gframe_size); 549 550 // Add in java locals size for stack overflow check only 551 __ add( Gframe_size, Glocals_size, Gframe_size ); 552 553 const Register Otmp2 = O4; 554 assert_different_registers(Otmp1, Otmp2, O5_savedSP); 555 generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2); 556 557 __ sub( Gframe_size, Glocals_size, Gframe_size); 558 559 // 560 // bump SP to accomodate the extra locals 561 // 562 __ sub( SP, Glocals_size, SP ); 563 } 564 565 // 566 // now set up a stack frame with the size computed above 567 // 568 __ neg( Gframe_size ); 569 __ save( SP, Gframe_size, SP ); 570 571 // 572 // now set up all the local cache registers 573 // 574 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note 575 // that all present references to Lbyte_code initialize the register 576 // immediately before use 577 if (native_call) { 578 __ mov(G0, Lbcp); 579 } else { 580 __ ld_ptr(G5_method, Method::const_offset(), Lbcp); 581 __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp); 582 } 583 __ mov( G5_method, Lmethod); // set Lmethod 584 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache 585 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors 586 #ifdef _LP64 587 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias 588 #endif 589 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp 590 591 // setup interpreter activation registers 592 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals 593 594 if (ProfileInterpreter) { 595 #ifdef FAST_DISPATCH 596 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since 597 // they both use I2. 598 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive"); 599 #endif // FAST_DISPATCH 600 __ set_method_data_pointer(); 601 } 602 603 } 604 605 // Empty method, generate a very fast return. 606 607 address InterpreterGenerator::generate_empty_entry(void) { 608 609 // A method that does nother but return... 610 611 address entry = __ pc(); 612 Label slow_path; 613 614 // do nothing for empty methods (do not even increment invocation counter) 615 if ( UseFastEmptyMethods) { 616 // If we need a safepoint check, generate full interpreter entry. 617 AddressLiteral sync_state(SafepointSynchronize::address_of_state()); 618 __ set(sync_state, G3_scratch); 619 __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path); 620 621 // Code: _return 622 __ retl(); 623 __ delayed()->mov(O5_savedSP, SP); 624 625 __ bind(slow_path); 626 (void) generate_normal_entry(false); 627 628 return entry; 629 } 630 return NULL; 631 } 632 633 // Call an accessor method (assuming it is resolved, otherwise drop into 634 // vanilla (slow path) entry 635 636 // Generates code to elide accessor methods 637 // Uses G3_scratch and G1_scratch as scratch 638 address InterpreterGenerator::generate_accessor_entry(void) { 639 640 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof; 641 // parameter size = 1 642 // Note: We can only use this code if the getfield has been resolved 643 // and if we don't have a null-pointer exception => check for 644 // these conditions first and use slow path if necessary. 645 address entry = __ pc(); 646 Label slow_path; 647 648 649 // XXX: for compressed oops pointer loading and decoding doesn't fit in 650 // delay slot and damages G1 651 if ( UseFastAccessorMethods && !UseCompressedOops ) { 652 // Check if we need to reach a safepoint and generate full interpreter 653 // frame if so. 654 AddressLiteral sync_state(SafepointSynchronize::address_of_state()); 655 __ load_contents(sync_state, G3_scratch); 656 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized); 657 __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path); 658 659 // Check if local 0 != NULL 660 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0 661 // check if local 0 == NULL and go the slow path 662 __ br_null_short(Otos_i, Assembler::pn, slow_path); 663 664 665 // read first instruction word and extract bytecode @ 1 and index @ 2 666 // get first 4 bytes of the bytecodes (big endian!) 667 __ ld_ptr(G5_method, Method::const_offset(), G1_scratch); 668 __ ld(G1_scratch, ConstMethod::codes_offset(), G1_scratch); 669 670 // move index @ 2 far left then to the right most two bytes. 671 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch); 672 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words( 673 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch); 674 675 // get constant pool cache 676 __ ld_ptr(G5_method, Method::const_offset(), G3_scratch); 677 __ ld_ptr(G3_scratch, ConstMethod::constants_offset(), G3_scratch); 678 __ ld_ptr(G3_scratch, ConstantPool::cache_offset_in_bytes(), G3_scratch); 679 680 // get specific constant pool cache entry 681 __ add(G3_scratch, G1_scratch, G3_scratch); 682 683 // Check the constant Pool cache entry to see if it has been resolved. 684 // If not, need the slow path. 685 ByteSize cp_base_offset = ConstantPoolCache::base_offset(); 686 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::indices_offset(), G1_scratch); 687 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch); 688 __ and3(G1_scratch, 0xFF, G1_scratch); 689 __ cmp_and_br_short(G1_scratch, Bytecodes::_getfield, Assembler::notEqual, Assembler::pn, slow_path); 690 691 // Get the type and return field offset from the constant pool cache 692 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::flags_offset(), G1_scratch); 693 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), G3_scratch); 694 695 Label xreturn_path; 696 // Need to differentiate between igetfield, agetfield, bgetfield etc. 697 // because they are different sizes. 698 // Get the type from the constant pool cache 699 __ srl(G1_scratch, ConstantPoolCacheEntry::tos_state_shift, G1_scratch); 700 // Make sure we don't need to mask G1_scratch after the above shift 701 ConstantPoolCacheEntry::verify_tos_state_shift(); 702 __ cmp(G1_scratch, atos ); 703 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 704 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i); 705 __ cmp(G1_scratch, itos); 706 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 707 __ delayed()->ld(Otos_i, G3_scratch, Otos_i); 708 __ cmp(G1_scratch, stos); 709 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 710 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i); 711 __ cmp(G1_scratch, ctos); 712 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 713 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i); 714 #ifdef ASSERT 715 __ cmp(G1_scratch, btos); 716 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 717 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i); 718 __ should_not_reach_here(); 719 #endif 720 __ ldsb(Otos_i, G3_scratch, Otos_i); 721 __ bind(xreturn_path); 722 723 // _ireturn/_areturn 724 __ retl(); // return from leaf routine 725 __ delayed()->mov(O5_savedSP, SP); 726 727 // Generate regular method entry 728 __ bind(slow_path); 729 (void) generate_normal_entry(false); 730 return entry; 731 } 732 return NULL; 733 } 734 735 // Method entry for java.lang.ref.Reference.get. 736 address InterpreterGenerator::generate_Reference_get_entry(void) { 737 #if INCLUDE_ALL_GCS 738 // Code: _aload_0, _getfield, _areturn 739 // parameter size = 1 740 // 741 // The code that gets generated by this routine is split into 2 parts: 742 // 1. The "intrinsified" code for G1 (or any SATB based GC), 743 // 2. The slow path - which is an expansion of the regular method entry. 744 // 745 // Notes:- 746 // * In the G1 code we do not check whether we need to block for 747 // a safepoint. If G1 is enabled then we must execute the specialized 748 // code for Reference.get (except when the Reference object is null) 749 // so that we can log the value in the referent field with an SATB 750 // update buffer. 751 // If the code for the getfield template is modified so that the 752 // G1 pre-barrier code is executed when the current method is 753 // Reference.get() then going through the normal method entry 754 // will be fine. 755 // * The G1 code can, however, check the receiver object (the instance 756 // of java.lang.Reference) and jump to the slow path if null. If the 757 // Reference object is null then we obviously cannot fetch the referent 758 // and so we don't need to call the G1 pre-barrier. Thus we can use the 759 // regular method entry code to generate the NPE. 760 // 761 // This code is based on generate_accessor_enty. 762 763 address entry = __ pc(); 764 765 const int referent_offset = java_lang_ref_Reference::referent_offset; 766 guarantee(referent_offset > 0, "referent offset not initialized"); 767 768 if (UseG1GC) { 769 Label slow_path; 770 771 // In the G1 code we don't check if we need to reach a safepoint. We 772 // continue and the thread will safepoint at the next bytecode dispatch. 773 774 // Check if local 0 != NULL 775 // If the receiver is null then it is OK to jump to the slow path. 776 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0 777 // check if local 0 == NULL and go the slow path 778 __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path); 779 780 781 // Load the value of the referent field. 782 if (Assembler::is_simm13(referent_offset)) { 783 __ load_heap_oop(Otos_i, referent_offset, Otos_i); 784 } else { 785 __ set(referent_offset, G3_scratch); 786 __ load_heap_oop(Otos_i, G3_scratch, Otos_i); 787 } 788 789 // Generate the G1 pre-barrier code to log the value of 790 // the referent field in an SATB buffer. Note with 791 // these parameters the pre-barrier does not generate 792 // the load of the previous value 793 794 __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */, 795 Otos_i /* pre_val */, 796 G3_scratch /* tmp */, 797 true /* preserve_o_regs */); 798 799 // _areturn 800 __ retl(); // return from leaf routine 801 __ delayed()->mov(O5_savedSP, SP); 802 803 // Generate regular method entry 804 __ bind(slow_path); 805 (void) generate_normal_entry(false); 806 return entry; 807 } 808 #endif // INCLUDE_ALL_GCS 809 810 // If G1 is not enabled then attempt to go through the accessor entry point 811 // Reference.get is an accessor 812 return generate_accessor_entry(); 813 } 814 815 // 816 // Interpreter stub for calling a native method. (asm interpreter) 817 // This sets up a somewhat different looking stack for calling the native method 818 // than the typical interpreter frame setup. 819 // 820 821 address InterpreterGenerator::generate_native_entry(bool synchronized) { 822 address entry = __ pc(); 823 824 // the following temporary registers are used during frame creation 825 const Register Gtmp1 = G3_scratch ; 826 const Register Gtmp2 = G1_scratch; 827 bool inc_counter = UseCompiler || CountCompiledCalls; 828 829 // make sure registers are different! 830 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2); 831 832 const Address Laccess_flags(Lmethod, Method::access_flags_offset()); 833 834 const Register Glocals_size = G3; 835 assert_different_registers(Glocals_size, G4_scratch, Gframe_size); 836 837 // make sure method is native & not abstract 838 // rethink these assertions - they can be simplified and shared (gri 2/25/2000) 839 #ifdef ASSERT 840 __ ld(G5_method, Method::access_flags_offset(), Gtmp1); 841 { 842 Label L; 843 __ btst(JVM_ACC_NATIVE, Gtmp1); 844 __ br(Assembler::notZero, false, Assembler::pt, L); 845 __ delayed()->nop(); 846 __ stop("tried to execute non-native method as native"); 847 __ bind(L); 848 } 849 { Label L; 850 __ btst(JVM_ACC_ABSTRACT, Gtmp1); 851 __ br(Assembler::zero, false, Assembler::pt, L); 852 __ delayed()->nop(); 853 __ stop("tried to execute abstract method as non-abstract"); 854 __ bind(L); 855 } 856 #endif // ASSERT 857 858 // generate the code to allocate the interpreter stack frame 859 generate_fixed_frame(true); 860 861 // 862 // No locals to initialize for native method 863 // 864 865 // this slot will be set later, we initialize it to null here just in 866 // case we get a GC before the actual value is stored later 867 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS); 868 869 const Address do_not_unlock_if_synchronized(G2_thread, 870 JavaThread::do_not_unlock_if_synchronized_offset()); 871 // Since at this point in the method invocation the exception handler 872 // would try to exit the monitor of synchronized methods which hasn't 873 // been entered yet, we set the thread local variable 874 // _do_not_unlock_if_synchronized to true. If any exception was thrown by 875 // runtime, exception handling i.e. unlock_if_synchronized_method will 876 // check this thread local flag. 877 // This flag has two effects, one is to force an unwind in the topmost 878 // interpreter frame and not perform an unlock while doing so. 879 880 __ movbool(true, G3_scratch); 881 __ stbool(G3_scratch, do_not_unlock_if_synchronized); 882 883 // increment invocation counter and check for overflow 884 // 885 // Note: checking for negative value instead of overflow 886 // so we have a 'sticky' overflow test (may be of 887 // importance as soon as we have true MT/MP) 888 Label invocation_counter_overflow; 889 Label Lcontinue; 890 if (inc_counter) { 891 generate_counter_incr(&invocation_counter_overflow, NULL, NULL); 892 893 } 894 __ bind(Lcontinue); 895 896 bang_stack_shadow_pages(true); 897 898 // reset the _do_not_unlock_if_synchronized flag 899 __ stbool(G0, do_not_unlock_if_synchronized); 900 901 // check for synchronized methods 902 // Must happen AFTER invocation_counter check and stack overflow check, 903 // so method is not locked if overflows. 904 905 if (synchronized) { 906 lock_method(); 907 } else { 908 #ifdef ASSERT 909 { Label ok; 910 __ ld(Laccess_flags, O0); 911 __ btst(JVM_ACC_SYNCHRONIZED, O0); 912 __ br( Assembler::zero, false, Assembler::pt, ok); 913 __ delayed()->nop(); 914 __ stop("method needs synchronization"); 915 __ bind(ok); 916 } 917 #endif // ASSERT 918 } 919 920 921 // start execution 922 __ verify_thread(); 923 924 // JVMTI support 925 __ notify_method_entry(); 926 927 // native call 928 929 // (note that O0 is never an oop--at most it is a handle) 930 // It is important not to smash any handles created by this call, 931 // until any oop handle in O0 is dereferenced. 932 933 // (note that the space for outgoing params is preallocated) 934 935 // get signature handler 936 { Label L; 937 Address signature_handler(Lmethod, Method::signature_handler_offset()); 938 __ ld_ptr(signature_handler, G3_scratch); 939 __ br_notnull_short(G3_scratch, Assembler::pt, L); 940 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod); 941 __ ld_ptr(signature_handler, G3_scratch); 942 __ bind(L); 943 } 944 945 // Push a new frame so that the args will really be stored in 946 // Copy a few locals across so the new frame has the variables 947 // we need but these values will be dead at the jni call and 948 // therefore not gc volatile like the values in the current 949 // frame (Lmethod in particular) 950 951 // Flush the method pointer to the register save area 952 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS); 953 __ mov(Llocals, O1); 954 955 // calculate where the mirror handle body is allocated in the interpreter frame: 956 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2); 957 958 // Calculate current frame size 959 __ sub(SP, FP, O3); // Calculate negative of current frame size 960 __ save(SP, O3, SP); // Allocate an identical sized frame 961 962 // Note I7 has leftover trash. Slow signature handler will fill it in 963 // should we get there. Normal jni call will set reasonable last_Java_pc 964 // below (and fix I7 so the stack trace doesn't have a meaningless frame 965 // in it). 966 967 // Load interpreter frame's Lmethod into same register here 968 969 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod); 970 971 __ mov(I1, Llocals); 972 __ mov(I2, Lscratch2); // save the address of the mirror 973 974 975 // ONLY Lmethod and Llocals are valid here! 976 977 // call signature handler, It will move the arg properly since Llocals in current frame 978 // matches that in outer frame 979 980 __ callr(G3_scratch, 0); 981 __ delayed()->nop(); 982 983 // Result handler is in Lscratch 984 985 // Reload interpreter frame's Lmethod since slow signature handler may block 986 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod); 987 988 { Label not_static; 989 990 __ ld(Laccess_flags, O0); 991 __ btst(JVM_ACC_STATIC, O0); 992 __ br( Assembler::zero, false, Assembler::pt, not_static); 993 // get native function entry point(O0 is a good temp until the very end) 994 __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0); 995 // for static methods insert the mirror argument 996 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 997 998 __ ld_ptr(Lmethod, Method:: const_offset(), O1); 999 __ ld_ptr(O1, ConstMethod::constants_offset(), O1); 1000 __ ld_ptr(O1, ConstantPool::pool_holder_offset_in_bytes(), O1); 1001 __ ld_ptr(O1, mirror_offset, O1); 1002 #ifdef ASSERT 1003 if (!PrintSignatureHandlers) // do not dirty the output with this 1004 { Label L; 1005 __ br_notnull_short(O1, Assembler::pt, L); 1006 __ stop("mirror is missing"); 1007 __ bind(L); 1008 } 1009 #endif // ASSERT 1010 __ st_ptr(O1, Lscratch2, 0); 1011 __ mov(Lscratch2, O1); 1012 __ bind(not_static); 1013 } 1014 1015 // At this point, arguments have been copied off of stack into 1016 // their JNI positions, which are O1..O5 and SP[68..]. 1017 // Oops are boxed in-place on the stack, with handles copied to arguments. 1018 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*. 1019 1020 #ifdef ASSERT 1021 { Label L; 1022 __ br_notnull_short(O0, Assembler::pt, L); 1023 __ stop("native entry point is missing"); 1024 __ bind(L); 1025 } 1026 #endif // ASSERT 1027 1028 // 1029 // setup the frame anchor 1030 // 1031 // The scavenge function only needs to know that the PC of this frame is 1032 // in the interpreter method entry code, it doesn't need to know the exact 1033 // PC and hence we can use O7 which points to the return address from the 1034 // previous call in the code stream (signature handler function) 1035 // 1036 // The other trick is we set last_Java_sp to FP instead of the usual SP because 1037 // we have pushed the extra frame in order to protect the volatile register(s) 1038 // in that frame when we return from the jni call 1039 // 1040 1041 __ set_last_Java_frame(FP, O7); 1042 __ mov(O7, I7); // make dummy interpreter frame look like one above, 1043 // not meaningless information that'll confuse me. 1044 1045 // flush the windows now. We don't care about the current (protection) frame 1046 // only the outer frames 1047 1048 __ flush_windows(); 1049 1050 // mark windows as flushed 1051 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset()); 1052 __ set(JavaFrameAnchor::flushed, G3_scratch); 1053 __ st(G3_scratch, flags); 1054 1055 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready. 1056 1057 Address thread_state(G2_thread, JavaThread::thread_state_offset()); 1058 #ifdef ASSERT 1059 { Label L; 1060 __ ld(thread_state, G3_scratch); 1061 __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L); 1062 __ stop("Wrong thread state in native stub"); 1063 __ bind(L); 1064 } 1065 #endif // ASSERT 1066 __ set(_thread_in_native, G3_scratch); 1067 __ st(G3_scratch, thread_state); 1068 1069 // Call the jni method, using the delay slot to set the JNIEnv* argument. 1070 __ save_thread(L7_thread_cache); // save Gthread 1071 __ callr(O0, 0); 1072 __ delayed()-> 1073 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0); 1074 1075 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD 1076 1077 __ restore_thread(L7_thread_cache); // restore G2_thread 1078 __ reinit_heapbase(); 1079 1080 // must we block? 1081 1082 // Block, if necessary, before resuming in _thread_in_Java state. 1083 // In order for GC to work, don't clear the last_Java_sp until after blocking. 1084 { Label no_block; 1085 AddressLiteral sync_state(SafepointSynchronize::address_of_state()); 1086 1087 // Switch thread to "native transition" state before reading the synchronization state. 1088 // This additional state is necessary because reading and testing the synchronization 1089 // state is not atomic w.r.t. GC, as this scenario demonstrates: 1090 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 1091 // VM thread changes sync state to synchronizing and suspends threads for GC. 1092 // Thread A is resumed to finish this native method, but doesn't block here since it 1093 // didn't see any synchronization is progress, and escapes. 1094 __ set(_thread_in_native_trans, G3_scratch); 1095 __ st(G3_scratch, thread_state); 1096 if(os::is_MP()) { 1097 if (UseMembar) { 1098 // Force this write out before the read below 1099 __ membar(Assembler::StoreLoad); 1100 } else { 1101 // Write serialization page so VM thread can do a pseudo remote membar. 1102 // We use the current thread pointer to calculate a thread specific 1103 // offset to write to within the page. This minimizes bus traffic 1104 // due to cache line collision. 1105 __ serialize_memory(G2_thread, G1_scratch, G3_scratch); 1106 } 1107 } 1108 __ load_contents(sync_state, G3_scratch); 1109 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized); 1110 1111 Label L; 1112 __ br(Assembler::notEqual, false, Assembler::pn, L); 1113 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch); 1114 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block); 1115 __ bind(L); 1116 1117 // Block. Save any potential method result value before the operation and 1118 // use a leaf call to leave the last_Java_frame setup undisturbed. 1119 save_native_result(); 1120 __ call_VM_leaf(L7_thread_cache, 1121 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans), 1122 G2_thread); 1123 1124 // Restore any method result value 1125 restore_native_result(); 1126 __ bind(no_block); 1127 } 1128 1129 // Clear the frame anchor now 1130 1131 __ reset_last_Java_frame(); 1132 1133 // Move the result handler address 1134 __ mov(Lscratch, G3_scratch); 1135 // return possible result to the outer frame 1136 #ifndef __LP64 1137 __ mov(O0, I0); 1138 __ restore(O1, G0, O1); 1139 #else 1140 __ restore(O0, G0, O0); 1141 #endif /* __LP64 */ 1142 1143 // Move result handler to expected register 1144 __ mov(G3_scratch, Lscratch); 1145 1146 // Back in normal (native) interpreter frame. State is thread_in_native_trans 1147 // switch to thread_in_Java. 1148 1149 __ set(_thread_in_Java, G3_scratch); 1150 __ st(G3_scratch, thread_state); 1151 1152 // reset handle block 1153 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch); 1154 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes()); 1155 1156 // If we have an oop result store it where it will be safe for any further gc 1157 // until we return now that we've released the handle it might be protected by 1158 1159 { 1160 Label no_oop, store_result; 1161 1162 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch); 1163 __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop); 1164 __ addcc(G0, O0, O0); 1165 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL: 1166 __ delayed()->ld_ptr(O0, 0, O0); // unbox it 1167 __ mov(G0, O0); 1168 1169 __ bind(store_result); 1170 // Store it where gc will look for it and result handler expects it. 1171 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS); 1172 1173 __ bind(no_oop); 1174 1175 } 1176 1177 1178 // handle exceptions (exception handling will handle unlocking!) 1179 { Label L; 1180 Address exception_addr(G2_thread, Thread::pending_exception_offset()); 1181 __ ld_ptr(exception_addr, Gtemp); 1182 __ br_null_short(Gtemp, Assembler::pt, L); 1183 // Note: This could be handled more efficiently since we know that the native 1184 // method doesn't have an exception handler. We could directly return 1185 // to the exception handler for the caller. 1186 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception)); 1187 __ should_not_reach_here(); 1188 __ bind(L); 1189 } 1190 1191 // JVMTI support (preserves thread register) 1192 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI); 1193 1194 if (synchronized) { 1195 // save and restore any potential method result value around the unlocking operation 1196 save_native_result(); 1197 1198 __ add( __ top_most_monitor(), O1); 1199 __ unlock_object(O1); 1200 1201 restore_native_result(); 1202 } 1203 1204 #if defined(COMPILER2) && !defined(_LP64) 1205 1206 // C2 expects long results in G1 we can't tell if we're returning to interpreted 1207 // or compiled so just be safe. 1208 1209 __ sllx(O0, 32, G1); // Shift bits into high G1 1210 __ srl (O1, 0, O1); // Zero extend O1 1211 __ or3 (O1, G1, G1); // OR 64 bits into G1 1212 1213 #endif /* COMPILER2 && !_LP64 */ 1214 1215 // dispose of return address and remove activation 1216 #ifdef ASSERT 1217 { 1218 Label ok; 1219 __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok); 1220 __ stop("bad I5_savedSP value"); 1221 __ should_not_reach_here(); 1222 __ bind(ok); 1223 } 1224 #endif 1225 if (TraceJumps) { 1226 // Move target to register that is recordable 1227 __ mov(Lscratch, G3_scratch); 1228 __ JMP(G3_scratch, 0); 1229 } else { 1230 __ jmp(Lscratch, 0); 1231 } 1232 __ delayed()->nop(); 1233 1234 1235 if (inc_counter) { 1236 // handle invocation counter overflow 1237 __ bind(invocation_counter_overflow); 1238 generate_counter_overflow(Lcontinue); 1239 } 1240 1241 1242 1243 return entry; 1244 } 1245 1246 1247 // Generic method entry to (asm) interpreter 1248 //------------------------------------------------------------------------------------------------------------------------ 1249 // 1250 address InterpreterGenerator::generate_normal_entry(bool synchronized) { 1251 address entry = __ pc(); 1252 1253 bool inc_counter = UseCompiler || CountCompiledCalls; 1254 1255 // the following temporary registers are used during frame creation 1256 const Register Gtmp1 = G3_scratch ; 1257 const Register Gtmp2 = G1_scratch; 1258 1259 // make sure registers are different! 1260 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2); 1261 1262 const Address constMethod (G5_method, Method::const_offset()); 1263 // Seems like G5_method is live at the point this is used. So we could make this look consistent 1264 // and use in the asserts. 1265 const Address access_flags (Lmethod, Method::access_flags_offset()); 1266 1267 const Register Glocals_size = G3; 1268 assert_different_registers(Glocals_size, G4_scratch, Gframe_size); 1269 1270 // make sure method is not native & not abstract 1271 // rethink these assertions - they can be simplified and shared (gri 2/25/2000) 1272 #ifdef ASSERT 1273 __ ld(G5_method, Method::access_flags_offset(), Gtmp1); 1274 { 1275 Label L; 1276 __ btst(JVM_ACC_NATIVE, Gtmp1); 1277 __ br(Assembler::zero, false, Assembler::pt, L); 1278 __ delayed()->nop(); 1279 __ stop("tried to execute native method as non-native"); 1280 __ bind(L); 1281 } 1282 { Label L; 1283 __ btst(JVM_ACC_ABSTRACT, Gtmp1); 1284 __ br(Assembler::zero, false, Assembler::pt, L); 1285 __ delayed()->nop(); 1286 __ stop("tried to execute abstract method as non-abstract"); 1287 __ bind(L); 1288 } 1289 #endif // ASSERT 1290 1291 // generate the code to allocate the interpreter stack frame 1292 1293 generate_fixed_frame(false); 1294 1295 #ifdef FAST_DISPATCH 1296 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables); 1297 // set bytecode dispatch table base 1298 #endif 1299 1300 // 1301 // Code to initialize the extra (i.e. non-parm) locals 1302 // 1303 Register init_value = noreg; // will be G0 if we must clear locals 1304 // The way the code was setup before zerolocals was always true for vanilla java entries. 1305 // It could only be false for the specialized entries like accessor or empty which have 1306 // no extra locals so the testing was a waste of time and the extra locals were always 1307 // initialized. We removed this extra complication to already over complicated code. 1308 1309 init_value = G0; 1310 Label clear_loop; 1311 1312 const Register RconstMethod = O1; 1313 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset()); 1314 const Address size_of_locals (RconstMethod, ConstMethod::size_of_locals_offset()); 1315 1316 // NOTE: If you change the frame layout, this code will need to 1317 // be updated! 1318 __ ld_ptr( constMethod, RconstMethod ); 1319 __ lduh( size_of_locals, O2 ); 1320 __ lduh( size_of_parameters, O1 ); 1321 __ sll( O2, Interpreter::logStackElementSize, O2); 1322 __ sll( O1, Interpreter::logStackElementSize, O1 ); 1323 __ sub( Llocals, O2, O2 ); 1324 __ sub( Llocals, O1, O1 ); 1325 1326 __ bind( clear_loop ); 1327 __ inc( O2, wordSize ); 1328 1329 __ cmp( O2, O1 ); 1330 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop ); 1331 __ delayed()->st_ptr( init_value, O2, 0 ); 1332 1333 const Address do_not_unlock_if_synchronized(G2_thread, 1334 JavaThread::do_not_unlock_if_synchronized_offset()); 1335 // Since at this point in the method invocation the exception handler 1336 // would try to exit the monitor of synchronized methods which hasn't 1337 // been entered yet, we set the thread local variable 1338 // _do_not_unlock_if_synchronized to true. If any exception was thrown by 1339 // runtime, exception handling i.e. unlock_if_synchronized_method will 1340 // check this thread local flag. 1341 __ movbool(true, G3_scratch); 1342 __ stbool(G3_scratch, do_not_unlock_if_synchronized); 1343 1344 // increment invocation counter and check for overflow 1345 // 1346 // Note: checking for negative value instead of overflow 1347 // so we have a 'sticky' overflow test (may be of 1348 // importance as soon as we have true MT/MP) 1349 Label invocation_counter_overflow; 1350 Label profile_method; 1351 Label profile_method_continue; 1352 Label Lcontinue; 1353 if (inc_counter) { 1354 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue); 1355 if (ProfileInterpreter) { 1356 __ bind(profile_method_continue); 1357 } 1358 } 1359 __ bind(Lcontinue); 1360 1361 bang_stack_shadow_pages(false); 1362 1363 // reset the _do_not_unlock_if_synchronized flag 1364 __ stbool(G0, do_not_unlock_if_synchronized); 1365 1366 // check for synchronized methods 1367 // Must happen AFTER invocation_counter check and stack overflow check, 1368 // so method is not locked if overflows. 1369 1370 if (synchronized) { 1371 lock_method(); 1372 } else { 1373 #ifdef ASSERT 1374 { Label ok; 1375 __ ld(access_flags, O0); 1376 __ btst(JVM_ACC_SYNCHRONIZED, O0); 1377 __ br( Assembler::zero, false, Assembler::pt, ok); 1378 __ delayed()->nop(); 1379 __ stop("method needs synchronization"); 1380 __ bind(ok); 1381 } 1382 #endif // ASSERT 1383 } 1384 1385 // start execution 1386 1387 __ verify_thread(); 1388 1389 // jvmti support 1390 __ notify_method_entry(); 1391 1392 // start executing instructions 1393 __ dispatch_next(vtos); 1394 1395 1396 if (inc_counter) { 1397 if (ProfileInterpreter) { 1398 // We have decided to profile this method in the interpreter 1399 __ bind(profile_method); 1400 1401 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method)); 1402 __ set_method_data_pointer_for_bcp(); 1403 __ ba_short(profile_method_continue); 1404 } 1405 1406 // handle invocation counter overflow 1407 __ bind(invocation_counter_overflow); 1408 generate_counter_overflow(Lcontinue); 1409 } 1410 1411 1412 return entry; 1413 } 1414 1415 1416 //---------------------------------------------------------------------------------------------------- 1417 // Entry points & stack frame layout 1418 // 1419 // Here we generate the various kind of entries into the interpreter. 1420 // The two main entry type are generic bytecode methods and native call method. 1421 // These both come in synchronized and non-synchronized versions but the 1422 // frame layout they create is very similar. The other method entry 1423 // types are really just special purpose entries that are really entry 1424 // and interpretation all in one. These are for trivial methods like 1425 // accessor, empty, or special math methods. 1426 // 1427 // When control flow reaches any of the entry types for the interpreter 1428 // the following holds -> 1429 // 1430 // C2 Calling Conventions: 1431 // 1432 // The entry code below assumes that the following registers are set 1433 // when coming in: 1434 // G5_method: holds the Method* of the method to call 1435 // Lesp: points to the TOS of the callers expression stack 1436 // after having pushed all the parameters 1437 // 1438 // The entry code does the following to setup an interpreter frame 1439 // pop parameters from the callers stack by adjusting Lesp 1440 // set O0 to Lesp 1441 // compute X = (max_locals - num_parameters) 1442 // bump SP up by X to accomadate the extra locals 1443 // compute X = max_expression_stack 1444 // + vm_local_words 1445 // + 16 words of register save area 1446 // save frame doing a save sp, -X, sp growing towards lower addresses 1447 // set Lbcp, Lmethod, LcpoolCache 1448 // set Llocals to i0 1449 // set Lmonitors to FP - rounded_vm_local_words 1450 // set Lesp to Lmonitors - 4 1451 // 1452 // The frame has now been setup to do the rest of the entry code 1453 1454 // Try this optimization: Most method entries could live in a 1455 // "one size fits all" stack frame without all the dynamic size 1456 // calculations. It might be profitable to do all this calculation 1457 // statically and approximately for "small enough" methods. 1458 1459 //----------------------------------------------------------------------------------------------- 1460 1461 // C1 Calling conventions 1462 // 1463 // Upon method entry, the following registers are setup: 1464 // 1465 // g2 G2_thread: current thread 1466 // g5 G5_method: method to activate 1467 // g4 Gargs : pointer to last argument 1468 // 1469 // 1470 // Stack: 1471 // 1472 // +---------------+ <--- sp 1473 // | | 1474 // : reg save area : 1475 // | | 1476 // +---------------+ <--- sp + 0x40 1477 // | | 1478 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later) 1479 // | | 1480 // +---------------+ <--- sp + 0x5c 1481 // | | 1482 // : free : 1483 // | | 1484 // +---------------+ <--- Gargs 1485 // | | 1486 // : arguments : 1487 // | | 1488 // +---------------+ 1489 // | | 1490 // 1491 // 1492 // 1493 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like: 1494 // 1495 // +---------------+ <--- sp 1496 // | | 1497 // : reg save area : 1498 // | | 1499 // +---------------+ <--- sp + 0x40 1500 // | | 1501 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later) 1502 // | | 1503 // +---------------+ <--- sp + 0x5c 1504 // | | 1505 // : : 1506 // | | <--- Lesp 1507 // +---------------+ <--- Lmonitors (fp - 0x18) 1508 // | VM locals | 1509 // +---------------+ <--- fp 1510 // | | 1511 // : reg save area : 1512 // | | 1513 // +---------------+ <--- fp + 0x40 1514 // | | 1515 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later) 1516 // | | 1517 // +---------------+ <--- fp + 0x5c 1518 // | | 1519 // : free : 1520 // | | 1521 // +---------------+ 1522 // | | 1523 // : nonarg locals : 1524 // | | 1525 // +---------------+ 1526 // | | 1527 // : arguments : 1528 // | | <--- Llocals 1529 // +---------------+ <--- Gargs 1530 // | | 1531 1532 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) { 1533 1534 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated 1535 // expression stack, the callee will have callee_extra_locals (so we can account for 1536 // frame extension) and monitor_size for monitors. Basically we need to calculate 1537 // this exactly like generate_fixed_frame/generate_compute_interpreter_state. 1538 // 1539 // 1540 // The big complicating thing here is that we must ensure that the stack stays properly 1541 // aligned. This would be even uglier if monitor size wasn't modulo what the stack 1542 // needs to be aligned for). We are given that the sp (fp) is already aligned by 1543 // the caller so we must ensure that it is properly aligned for our callee. 1544 // 1545 const int rounded_vm_local_words = 1546 round_to(frame::interpreter_frame_vm_local_words,WordsPerLong); 1547 // callee_locals and max_stack are counts, not the size in frame. 1548 const int locals_size = 1549 round_to(callee_extra_locals * Interpreter::stackElementWords, WordsPerLong); 1550 const int max_stack_words = max_stack * Interpreter::stackElementWords; 1551 return (round_to((max_stack_words 1552 + rounded_vm_local_words 1553 + frame::memory_parameter_word_sp_offset), WordsPerLong) 1554 // already rounded 1555 + locals_size + monitor_size); 1556 } 1557 1558 // How much stack a method top interpreter activation needs in words. 1559 int AbstractInterpreter::size_top_interpreter_activation(Method* method) { 1560 1561 // See call_stub code 1562 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset, 1563 WordsPerLong); // 7 + register save area 1564 1565 // Save space for one monitor to get into the interpreted method in case 1566 // the method is synchronized 1567 int monitor_size = method->is_synchronized() ? 1568 1*frame::interpreter_frame_monitor_size() : 0; 1569 return size_activation_helper(method->max_locals(), method->max_stack(), 1570 monitor_size) + call_stub_size; 1571 } 1572 1573 int AbstractInterpreter::layout_activation(Method* method, 1574 int tempcount, 1575 int popframe_extra_args, 1576 int moncount, 1577 int caller_actual_parameters, 1578 int callee_param_count, 1579 int callee_local_count, 1580 frame* caller, 1581 frame* interpreter_frame, 1582 bool is_top_frame, 1583 bool is_bottom_frame) { 1584 // Note: This calculation must exactly parallel the frame setup 1585 // in InterpreterGenerator::generate_fixed_frame. 1586 // If f!=NULL, set up the following variables: 1587 // - Lmethod 1588 // - Llocals 1589 // - Lmonitors (to the indicated number of monitors) 1590 // - Lesp (to the indicated number of temps) 1591 // The frame f (if not NULL) on entry is a description of the caller of the frame 1592 // we are about to layout. We are guaranteed that we will be able to fill in a 1593 // new interpreter frame as its callee (i.e. the stack space is allocated and 1594 // the amount was determined by an earlier call to this method with f == NULL). 1595 // On return f (if not NULL) while describe the interpreter frame we just layed out. 1596 1597 int monitor_size = moncount * frame::interpreter_frame_monitor_size(); 1598 int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong); 1599 1600 assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align"); 1601 // 1602 // Note: if you look closely this appears to be doing something much different 1603 // than generate_fixed_frame. What is happening is this. On sparc we have to do 1604 // this dance with interpreter_sp_adjustment because the window save area would 1605 // appear just below the bottom (tos) of the caller's java expression stack. Because 1606 // the interpreter want to have the locals completely contiguous generate_fixed_frame 1607 // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size). 1608 // Now in generate_fixed_frame the extension of the caller's sp happens in the callee. 1609 // In this code the opposite occurs the caller adjusts it's own stack base on the callee. 1610 // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest) 1611 // because the oldest frame would have adjust its callers frame and yet that frame 1612 // already exists and isn't part of this array of frames we are unpacking. So at first 1613 // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper() 1614 // will after it calculates all of the frame's on_stack_size()'s will then figure out the 1615 // amount to adjust the caller of the initial (oldest) frame and the calculation will all 1616 // add up. It does seem like it simpler to account for the adjustment here (and remove the 1617 // callee... parameters here). However this would mean that this routine would have to take 1618 // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment) 1619 // and run the calling loop in the reverse order. This would also would appear to mean making 1620 // this code aware of what the interactions are when that initial caller fram was an osr or 1621 // other adapter frame. deoptimization is complicated enough and hard enough to debug that 1622 // there is no sense in messing working code. 1623 // 1624 1625 int rounded_cls = round_to((callee_local_count - callee_param_count), WordsPerLong); 1626 assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align"); 1627 1628 int raw_frame_size = size_activation_helper(rounded_cls, method->max_stack(), 1629 monitor_size); 1630 1631 if (interpreter_frame != NULL) { 1632 // The skeleton frame must already look like an interpreter frame 1633 // even if not fully filled out. 1634 assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame"); 1635 1636 intptr_t* fp = interpreter_frame->fp(); 1637 1638 JavaThread* thread = JavaThread::current(); 1639 RegisterMap map(thread, false); 1640 // More verification that skeleton frame is properly walkable 1641 assert(fp == caller->sp(), "fp must match"); 1642 1643 intptr_t* montop = fp - rounded_vm_local_words; 1644 1645 // preallocate monitors (cf. __ add_monitor_to_stack) 1646 intptr_t* monitors = montop - monitor_size; 1647 1648 // preallocate stack space 1649 intptr_t* esp = monitors - 1 - 1650 (tempcount * Interpreter::stackElementWords) - 1651 popframe_extra_args; 1652 1653 int local_words = method->max_locals() * Interpreter::stackElementWords; 1654 NEEDS_CLEANUP; 1655 intptr_t* locals; 1656 if (caller->is_interpreted_frame()) { 1657 // Can force the locals area to end up properly overlapping the top of the expression stack. 1658 intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1; 1659 // Note that this computation means we replace size_of_parameters() values from the caller 1660 // interpreter frame's expression stack with our argument locals 1661 int parm_words = caller_actual_parameters * Interpreter::stackElementWords; 1662 locals = Lesp_ptr + parm_words; 1663 int delta = local_words - parm_words; 1664 int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0; 1665 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS; 1666 if (!is_bottom_frame) { 1667 // Llast_SP is set below for the current frame to SP (with the 1668 // extra space for the callee's locals). Here we adjust 1669 // Llast_SP for the caller's frame, removing the extra space 1670 // for the current method's locals. 1671 *caller->register_addr(Llast_SP) = *interpreter_frame->register_addr(I5_savedSP); 1672 } else { 1673 assert(*caller->register_addr(Llast_SP) >= *interpreter_frame->register_addr(I5_savedSP), "strange Llast_SP"); 1674 } 1675 } else { 1676 assert(caller->is_compiled_frame() || caller->is_entry_frame(), "only possible cases"); 1677 // Don't have Lesp available; lay out locals block in the caller 1678 // adjacent to the register window save area. 1679 // 1680 // Compiled frames do not allocate a varargs area which is why this if 1681 // statement is needed. 1682 // 1683 if (caller->is_compiled_frame()) { 1684 locals = fp + frame::register_save_words + local_words - 1; 1685 } else { 1686 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1; 1687 } 1688 if (!caller->is_entry_frame()) { 1689 // Caller wants his own SP back 1690 int caller_frame_size = caller->cb()->frame_size(); 1691 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS; 1692 } 1693 } 1694 if (TraceDeoptimization) { 1695 if (caller->is_entry_frame()) { 1696 // make sure I5_savedSP and the entry frames notion of saved SP 1697 // agree. This assertion duplicate a check in entry frame code 1698 // but catches the failure earlier. 1699 assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP), 1700 "would change callers SP"); 1701 } 1702 if (caller->is_entry_frame()) { 1703 tty->print("entry "); 1704 } 1705 if (caller->is_compiled_frame()) { 1706 tty->print("compiled "); 1707 if (caller->is_deoptimized_frame()) { 1708 tty->print("(deopt) "); 1709 } 1710 } 1711 if (caller->is_interpreted_frame()) { 1712 tty->print("interpreted "); 1713 } 1714 tty->print_cr("caller fp=0x%x sp=0x%x", caller->fp(), caller->sp()); 1715 tty->print_cr("save area = 0x%x, 0x%x", caller->sp(), caller->sp() + 16); 1716 tty->print_cr("save area = 0x%x, 0x%x", caller->fp(), caller->fp() + 16); 1717 tty->print_cr("interpreter fp=0x%x sp=0x%x", interpreter_frame->fp(), interpreter_frame->sp()); 1718 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->sp(), interpreter_frame->sp() + 16); 1719 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->fp(), interpreter_frame->fp() + 16); 1720 tty->print_cr("Llocals = 0x%x", locals); 1721 tty->print_cr("Lesp = 0x%x", esp); 1722 tty->print_cr("Lmonitors = 0x%x", monitors); 1723 } 1724 1725 if (method->max_locals() > 0) { 1726 assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area"); 1727 assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area"); 1728 assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area"); 1729 assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area"); 1730 } 1731 #ifdef _LP64 1732 assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd"); 1733 #endif 1734 1735 *interpreter_frame->register_addr(Lmethod) = (intptr_t) method; 1736 *interpreter_frame->register_addr(Llocals) = (intptr_t) locals; 1737 *interpreter_frame->register_addr(Lmonitors) = (intptr_t) monitors; 1738 *interpreter_frame->register_addr(Lesp) = (intptr_t) esp; 1739 // Llast_SP will be same as SP as there is no adapter space 1740 *interpreter_frame->register_addr(Llast_SP) = (intptr_t) interpreter_frame->sp() - STACK_BIAS; 1741 *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache(); 1742 #ifdef FAST_DISPATCH 1743 *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table(); 1744 #endif 1745 1746 1747 #ifdef ASSERT 1748 BasicObjectLock* mp = (BasicObjectLock*)monitors; 1749 1750 assert(interpreter_frame->interpreter_frame_method() == method, "method matches"); 1751 assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize)), "locals match"); 1752 assert(interpreter_frame->interpreter_frame_monitor_end() == mp, "monitor_end matches"); 1753 assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches"); 1754 assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches"); 1755 1756 // check bounds 1757 intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1); 1758 intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words; 1759 assert(lo < monitors && montop <= hi, "monitors in bounds"); 1760 assert(lo <= esp && esp < monitors, "esp in bounds"); 1761 #endif // ASSERT 1762 } 1763 1764 return raw_frame_size; 1765 } 1766 1767 //---------------------------------------------------------------------------------------------------- 1768 // Exceptions 1769 void TemplateInterpreterGenerator::generate_throw_exception() { 1770 1771 // Entry point in previous activation (i.e., if the caller was interpreted) 1772 Interpreter::_rethrow_exception_entry = __ pc(); 1773 // O0: exception 1774 1775 // entry point for exceptions thrown within interpreter code 1776 Interpreter::_throw_exception_entry = __ pc(); 1777 __ verify_thread(); 1778 // expression stack is undefined here 1779 // O0: exception, i.e. Oexception 1780 // Lbcp: exception bcx 1781 __ verify_oop(Oexception); 1782 1783 1784 // expression stack must be empty before entering the VM in case of an exception 1785 __ empty_expression_stack(); 1786 // find exception handler address and preserve exception oop 1787 // call C routine to find handler and jump to it 1788 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception); 1789 __ push_ptr(O1); // push exception for exception handler bytecodes 1790 1791 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!) 1792 __ delayed()->nop(); 1793 1794 1795 // if the exception is not handled in the current frame 1796 // the frame is removed and the exception is rethrown 1797 // (i.e. exception continuation is _rethrow_exception) 1798 // 1799 // Note: At this point the bci is still the bxi for the instruction which caused 1800 // the exception and the expression stack is empty. Thus, for any VM calls 1801 // at this point, GC will find a legal oop map (with empty expression stack). 1802 1803 // in current activation 1804 // tos: exception 1805 // Lbcp: exception bcp 1806 1807 // 1808 // JVMTI PopFrame support 1809 // 1810 1811 Interpreter::_remove_activation_preserving_args_entry = __ pc(); 1812 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset()); 1813 // Set the popframe_processing bit in popframe_condition indicating that we are 1814 // currently handling popframe, so that call_VMs that may happen later do not trigger new 1815 // popframe handling cycles. 1816 1817 __ ld(popframe_condition_addr, G3_scratch); 1818 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch); 1819 __ stw(G3_scratch, popframe_condition_addr); 1820 1821 // Empty the expression stack, as in normal exception handling 1822 __ empty_expression_stack(); 1823 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false); 1824 1825 { 1826 // Check to see whether we are returning to a deoptimized frame. 1827 // (The PopFrame call ensures that the caller of the popped frame is 1828 // either interpreted or compiled and deoptimizes it if compiled.) 1829 // In this case, we can't call dispatch_next() after the frame is 1830 // popped, but instead must save the incoming arguments and restore 1831 // them after deoptimization has occurred. 1832 // 1833 // Note that we don't compare the return PC against the 1834 // deoptimization blob's unpack entry because of the presence of 1835 // adapter frames in C2. 1836 Label caller_not_deoptimized; 1837 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7); 1838 __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized); 1839 1840 const Register Gtmp1 = G3_scratch; 1841 const Register Gtmp2 = G1_scratch; 1842 const Register RconstMethod = Gtmp1; 1843 const Address constMethod(Lmethod, Method::const_offset()); 1844 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset()); 1845 1846 // Compute size of arguments for saving when returning to deoptimized caller 1847 __ ld_ptr(constMethod, RconstMethod); 1848 __ lduh(size_of_parameters, Gtmp1); 1849 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1); 1850 __ sub(Llocals, Gtmp1, Gtmp2); 1851 __ add(Gtmp2, wordSize, Gtmp2); 1852 // Save these arguments 1853 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2); 1854 // Inform deoptimization that it is responsible for restoring these arguments 1855 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1); 1856 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset()); 1857 __ st(Gtmp1, popframe_condition_addr); 1858 1859 // Return from the current method 1860 // The caller's SP was adjusted upon method entry to accomodate 1861 // the callee's non-argument locals. Undo that adjustment. 1862 __ ret(); 1863 __ delayed()->restore(I5_savedSP, G0, SP); 1864 1865 __ bind(caller_not_deoptimized); 1866 } 1867 1868 // Clear the popframe condition flag 1869 __ stw(G0 /* popframe_inactive */, popframe_condition_addr); 1870 1871 // Get out of the current method (how this is done depends on the particular compiler calling 1872 // convention that the interpreter currently follows) 1873 // The caller's SP was adjusted upon method entry to accomodate 1874 // the callee's non-argument locals. Undo that adjustment. 1875 __ restore(I5_savedSP, G0, SP); 1876 // The method data pointer was incremented already during 1877 // call profiling. We have to restore the mdp for the current bcp. 1878 if (ProfileInterpreter) { 1879 __ set_method_data_pointer_for_bcp(); 1880 } 1881 // Resume bytecode interpretation at the current bcp 1882 __ dispatch_next(vtos); 1883 // end of JVMTI PopFrame support 1884 1885 Interpreter::_remove_activation_entry = __ pc(); 1886 1887 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here) 1888 __ pop_ptr(Oexception); // get exception 1889 1890 // Intel has the following comment: 1891 //// remove the activation (without doing throws on illegalMonitorExceptions) 1892 // They remove the activation without checking for bad monitor state. 1893 // %%% We should make sure this is the right semantics before implementing. 1894 1895 __ set_vm_result(Oexception); 1896 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false); 1897 1898 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI); 1899 1900 __ get_vm_result(Oexception); 1901 __ verify_oop(Oexception); 1902 1903 const int return_reg_adjustment = frame::pc_return_offset; 1904 Address issuing_pc_addr(I7, return_reg_adjustment); 1905 1906 // We are done with this activation frame; find out where to go next. 1907 // The continuation point will be an exception handler, which expects 1908 // the following registers set up: 1909 // 1910 // Oexception: exception 1911 // Oissuing_pc: the local call that threw exception 1912 // Other On: garbage 1913 // In/Ln: the contents of the caller's register window 1914 // 1915 // We do the required restore at the last possible moment, because we 1916 // need to preserve some state across a runtime call. 1917 // (Remember that the caller activation is unknown--it might not be 1918 // interpreted, so things like Lscratch are useless in the caller.) 1919 1920 // Although the Intel version uses call_C, we can use the more 1921 // compact call_VM. (The only real difference on SPARC is a 1922 // harmlessly ignored [re]set_last_Java_frame, compared with 1923 // the Intel code which lacks this.) 1924 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore 1925 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller 1926 __ super_call_VM_leaf(L7_thread_cache, 1927 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), 1928 G2_thread, Oissuing_pc->after_save()); 1929 1930 // The caller's SP was adjusted upon method entry to accomodate 1931 // the callee's non-argument locals. Undo that adjustment. 1932 __ JMP(O0, 0); // return exception handler in caller 1933 __ delayed()->restore(I5_savedSP, G0, SP); 1934 1935 // (same old exception object is already in Oexception; see above) 1936 // Note that an "issuing PC" is actually the next PC after the call 1937 } 1938 1939 1940 // 1941 // JVMTI ForceEarlyReturn support 1942 // 1943 1944 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) { 1945 address entry = __ pc(); 1946 1947 __ empty_expression_stack(); 1948 __ load_earlyret_value(state); 1949 1950 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch); 1951 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset()); 1952 1953 // Clear the earlyret state 1954 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr); 1955 1956 __ remove_activation(state, 1957 /* throw_monitor_exception */ false, 1958 /* install_monitor_exception */ false); 1959 1960 // The caller's SP was adjusted upon method entry to accomodate 1961 // the callee's non-argument locals. Undo that adjustment. 1962 __ ret(); // return to caller 1963 __ delayed()->restore(I5_savedSP, G0, SP); 1964 1965 return entry; 1966 } // end of JVMTI ForceEarlyReturn support 1967 1968 1969 //------------------------------------------------------------------------------------------------------------------------ 1970 // Helper for vtos entry point generation 1971 1972 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) { 1973 assert(t->is_valid() && t->tos_in() == vtos, "illegal template"); 1974 Label L; 1975 aep = __ pc(); __ push_ptr(); __ ba_short(L); 1976 fep = __ pc(); __ push_f(); __ ba_short(L); 1977 dep = __ pc(); __ push_d(); __ ba_short(L); 1978 lep = __ pc(); __ push_l(); __ ba_short(L); 1979 iep = __ pc(); __ push_i(); 1980 bep = cep = sep = iep; // there aren't any 1981 vep = __ pc(); __ bind(L); // fall through 1982 generate_and_dispatch(t); 1983 } 1984 1985 // -------------------------------------------------------------------------------- 1986 1987 1988 InterpreterGenerator::InterpreterGenerator(StubQueue* code) 1989 : TemplateInterpreterGenerator(code) { 1990 generate_all(); // down here so it can be "virtual" 1991 } 1992 1993 // -------------------------------------------------------------------------------- 1994 1995 // Non-product code 1996 #ifndef PRODUCT 1997 address TemplateInterpreterGenerator::generate_trace_code(TosState state) { 1998 address entry = __ pc(); 1999 2000 __ push(state); 2001 __ mov(O7, Lscratch); // protect return address within interpreter 2002 2003 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer 2004 __ mov( Otos_l2, G3_scratch ); 2005 __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch); 2006 __ mov(Lscratch, O7); // restore return address 2007 __ pop(state); 2008 __ retl(); 2009 __ delayed()->nop(); 2010 2011 return entry; 2012 } 2013 2014 2015 // helpers for generate_and_dispatch 2016 2017 void TemplateInterpreterGenerator::count_bytecode() { 2018 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch); 2019 } 2020 2021 2022 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) { 2023 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch); 2024 } 2025 2026 2027 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) { 2028 AddressLiteral index (&BytecodePairHistogram::_index); 2029 AddressLiteral counters((address) &BytecodePairHistogram::_counters); 2030 2031 // get index, shift out old bytecode, bring in new bytecode, and store it 2032 // _index = (_index >> log2_number_of_codes) | 2033 // (bytecode << log2_number_of_codes); 2034 2035 __ load_contents(index, G4_scratch); 2036 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch ); 2037 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch ); 2038 __ or3( G3_scratch, G4_scratch, G4_scratch ); 2039 __ store_contents(G4_scratch, index, G3_scratch); 2040 2041 // bump bucket contents 2042 // _counters[_index] ++; 2043 2044 __ set(counters, G3_scratch); // loads into G3_scratch 2045 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address 2046 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index 2047 __ ld (G3_scratch, 0, G4_scratch); 2048 __ inc (G4_scratch); 2049 __ st (G4_scratch, 0, G3_scratch); 2050 } 2051 2052 2053 void TemplateInterpreterGenerator::trace_bytecode(Template* t) { 2054 // Call a little run-time stub to avoid blow-up for each bytecode. 2055 // The run-time runtime saves the right registers, depending on 2056 // the tosca in-state for the given template. 2057 address entry = Interpreter::trace_code(t->tos_in()); 2058 guarantee(entry != NULL, "entry must have been generated"); 2059 __ call(entry, relocInfo::none); 2060 __ delayed()->nop(); 2061 } 2062 2063 2064 void TemplateInterpreterGenerator::stop_interpreter_at() { 2065 AddressLiteral counter(&BytecodeCounter::_counter_value); 2066 __ load_contents(counter, G3_scratch); 2067 AddressLiteral stop_at(&StopInterpreterAt); 2068 __ load_ptr_contents(stop_at, G4_scratch); 2069 __ cmp(G3_scratch, G4_scratch); 2070 __ breakpoint_trap(Assembler::equal, Assembler::icc); 2071 } 2072 #endif // not PRODUCT 2073 #endif // !CC_INTERP