1 /* 2 * Copyright (c) 2003, 2011, Oracle and/or its affiliates. All rights reserved. 3 * Copyright (c) 2014, Red Hat Inc. All rights reserved. 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 5 * 6 * This code is free software; you can redistribute it and/or modify it 7 * under the terms of the GNU General Public License version 2 only, as 8 * published by the Free Software Foundation. 9 * 10 * This code is distributed in the hope that it will be useful, but WITHOUT 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 13 * version 2 for more details (a copy is included in the LICENSE file that 14 * accompanied this code). 15 * 16 * You should have received a copy of the GNU General Public License version 17 * 2 along with this work; if not, write to the Free Software Foundation, 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 19 * 20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 21 * or visit www.oracle.com if you need additional information or have any 22 * questions. 23 * 24 */ 25 26 #include "precompiled.hpp" 27 #include "interp_masm_aarch64.hpp" 28 #include "interpreter/interpreter.hpp" 29 #include "interpreter/interpreterRuntime.hpp" 30 #include "oops/arrayOop.hpp" 31 #include "oops/markOop.hpp" 32 #include "oops/methodData.hpp" 33 #include "oops/method.hpp" 34 #include "prims/jvmtiExport.hpp" 35 #include "prims/jvmtiRedefineClassesTrace.hpp" 36 #include "prims/jvmtiThreadState.hpp" 37 #include "runtime/basicLock.hpp" 38 #include "runtime/biasedLocking.hpp" 39 #include "runtime/sharedRuntime.hpp" 40 #include "runtime/thread.inline.hpp" 41 42 43 // Implementation of InterpreterMacroAssembler 44 45 #ifndef CC_INTERP 46 47 void InterpreterMacroAssembler::check_and_handle_popframe(Register java_thread) { 48 if (JvmtiExport::can_pop_frame()) { 49 Label L; 50 // Initiate popframe handling only if it is not already being 51 // processed. If the flag has the popframe_processing bit set, it 52 // means that this code is called *during* popframe handling - we 53 // don't want to reenter. 54 // This method is only called just after the call into the vm in 55 // call_VM_base, so the arg registers are available. 56 ldrw(rscratch1, Address(rthread, JavaThread::popframe_condition_offset())); 57 tstw(rscratch1, JavaThread::popframe_pending_bit); 58 br(Assembler::EQ, L); 59 tstw(rscratch1, JavaThread::popframe_processing_bit); 60 br(Assembler::NE, L); 61 // Call Interpreter::remove_activation_preserving_args_entry() to get the 62 // address of the same-named entrypoint in the generated interpreter code. 63 call_VM_leaf(CAST_FROM_FN_PTR(address, Interpreter::remove_activation_preserving_args_entry)); 64 br(r0); 65 bind(L); 66 } 67 } 68 69 70 void InterpreterMacroAssembler::load_earlyret_value(TosState state) { 71 ldr(r2, Address(rthread, JavaThread::jvmti_thread_state_offset())); 72 const Address tos_addr(r2, JvmtiThreadState::earlyret_tos_offset()); 73 const Address oop_addr(r2, JvmtiThreadState::earlyret_oop_offset()); 74 const Address val_addr(r2, JvmtiThreadState::earlyret_value_offset()); 75 switch (state) { 76 case atos: ldr(r0, oop_addr); 77 str(zr, oop_addr); 78 verify_oop(r0, state); break; 79 case ltos: ldr(r0, val_addr); break; 80 case btos: // fall through 81 case ctos: // fall through 82 case stos: // fall through 83 case itos: ldrw(r0, val_addr); break; 84 case ftos: ldrs(v0, val_addr); break; 85 case dtos: ldrd(v0, val_addr); break; 86 case vtos: /* nothing to do */ break; 87 default : ShouldNotReachHere(); 88 } 89 // Clean up tos value in the thread object 90 movw(rscratch1, (int) ilgl); 91 strw(rscratch1, tos_addr); 92 strw(zr, val_addr); 93 } 94 95 96 void InterpreterMacroAssembler::check_and_handle_earlyret(Register java_thread) { 97 if (JvmtiExport::can_force_early_return()) { 98 Label L; 99 ldr(rscratch1, Address(rthread, JavaThread::jvmti_thread_state_offset())); 100 cbz(rscratch1, L); // if (thread->jvmti_thread_state() == NULL) exit; 101 102 // Initiate earlyret handling only if it is not already being processed. 103 // If the flag has the earlyret_processing bit set, it means that this code 104 // is called *during* earlyret handling - we don't want to reenter. 105 ldrw(rscratch1, Address(rscratch1, JvmtiThreadState::earlyret_state_offset())); 106 cmpw(rscratch1, JvmtiThreadState::earlyret_pending); 107 br(Assembler::NE, L); 108 109 // Call Interpreter::remove_activation_early_entry() to get the address of the 110 // same-named entrypoint in the generated interpreter code. 111 ldr(rscratch1, Address(rthread, JavaThread::jvmti_thread_state_offset())); 112 ldrw(rscratch1, Address(rscratch1, JvmtiThreadState::earlyret_tos_offset())); 113 call_VM_leaf(CAST_FROM_FN_PTR(address, Interpreter::remove_activation_early_entry), rscratch1); 114 br(r0); 115 bind(L); 116 } 117 } 118 119 void InterpreterMacroAssembler::get_unsigned_2_byte_index_at_bcp( 120 Register reg, 121 int bcp_offset) { 122 assert(bcp_offset >= 0, "bcp is still pointing to start of bytecode"); 123 ldrh(reg, Address(rbcp, bcp_offset)); 124 rev16(reg, reg); 125 } 126 127 void InterpreterMacroAssembler::get_dispatch() { 128 unsigned long offset; 129 adrp(rdispatch, ExternalAddress((address)Interpreter::dispatch_table()), offset); 130 lea(rdispatch, Address(rdispatch, offset)); 131 } 132 133 void InterpreterMacroAssembler::get_cache_index_at_bcp(Register index, 134 int bcp_offset, 135 size_t index_size) { 136 assert(bcp_offset > 0, "bcp is still pointing to start of bytecode"); 137 if (index_size == sizeof(u2)) { 138 load_unsigned_short(index, Address(rbcp, bcp_offset)); 139 } else if (index_size == sizeof(u4)) { 140 // assert(EnableInvokeDynamic, "giant index used only for JSR 292"); 141 ldrw(index, Address(rbcp, bcp_offset)); 142 // Check if the secondary index definition is still ~x, otherwise 143 // we have to change the following assembler code to calculate the 144 // plain index. 145 assert(ConstantPool::decode_invokedynamic_index(~123) == 123, "else change next line"); 146 eonw(index, index, zr); // convert to plain index 147 } else if (index_size == sizeof(u1)) { 148 load_unsigned_byte(index, Address(rbcp, bcp_offset)); 149 } else { 150 ShouldNotReachHere(); 151 } 152 } 153 154 // Return 155 // Rindex: index into constant pool 156 // Rcache: address of cache entry - ConstantPoolCache::base_offset() 157 // 158 // A caller must add ConstantPoolCache::base_offset() to Rcache to get 159 // the true address of the cache entry. 160 // 161 void InterpreterMacroAssembler::get_cache_and_index_at_bcp(Register cache, 162 Register index, 163 int bcp_offset, 164 size_t index_size) { 165 assert_different_registers(cache, index); 166 assert_different_registers(cache, rcpool); 167 get_cache_index_at_bcp(index, bcp_offset, index_size); 168 assert(sizeof(ConstantPoolCacheEntry) == 4 * wordSize, "adjust code below"); 169 // convert from field index to ConstantPoolCacheEntry 170 // aarch64 already has the cache in rcpool so there is no need to 171 // install it in cache. instead we pre-add the indexed offset to 172 // rcpool and return it in cache. All clients of this method need to 173 // be modified accordingly. 174 add(cache, rcpool, index, Assembler::LSL, 5); 175 } 176 177 178 void InterpreterMacroAssembler::get_cache_and_index_and_bytecode_at_bcp(Register cache, 179 Register index, 180 Register bytecode, 181 int byte_no, 182 int bcp_offset, 183 size_t index_size) { 184 get_cache_and_index_at_bcp(cache, index, bcp_offset, index_size); 185 // We use a 32-bit load here since the layout of 64-bit words on 186 // little-endian machines allow us that. 187 // n.b. unlike x86 cache alreeady includes the index offset 188 ldrw(bytecode, Address(cache, 189 ConstantPoolCache::base_offset() 190 + ConstantPoolCacheEntry::indices_offset())); 191 const int shift_count = (1 + byte_no) * BitsPerByte; 192 ubfx(bytecode, bytecode, shift_count, BitsPerByte); 193 } 194 195 void InterpreterMacroAssembler::get_cache_entry_pointer_at_bcp(Register cache, 196 Register tmp, 197 int bcp_offset, 198 size_t index_size) { 199 assert(cache != tmp, "must use different register"); 200 get_cache_index_at_bcp(tmp, bcp_offset, index_size); 201 assert(sizeof(ConstantPoolCacheEntry) == 4 * wordSize, "adjust code below"); 202 // convert from field index to ConstantPoolCacheEntry index 203 // and from word offset to byte offset 204 assert(exact_log2(in_bytes(ConstantPoolCacheEntry::size_in_bytes())) == 2 + LogBytesPerWord, "else change next line"); 205 ldr(cache, Address(rfp, frame::interpreter_frame_cache_offset * wordSize)); 206 // skip past the header 207 add(cache, cache, in_bytes(ConstantPoolCache::base_offset())); 208 add(cache, cache, tmp, Assembler::LSL, 2 + LogBytesPerWord); // construct pointer to cache entry 209 } 210 211 void InterpreterMacroAssembler::get_method_counters(Register method, 212 Register mcs, Label& skip) { 213 Label has_counters; 214 ldr(mcs, Address(method, Method::method_counters_offset())); 215 cbnz(mcs, has_counters); 216 call_VM(noreg, CAST_FROM_FN_PTR(address, 217 InterpreterRuntime::build_method_counters), method); 218 ldr(mcs, Address(method, Method::method_counters_offset())); 219 cbz(mcs, skip); // No MethodCounters allocated, OutOfMemory 220 bind(has_counters); 221 } 222 223 // Load object from cpool->resolved_references(index) 224 void InterpreterMacroAssembler::load_resolved_reference_at_index( 225 Register result, Register index) { 226 assert_different_registers(result, index); 227 // convert from field index to resolved_references() index and from 228 // word index to byte offset. Since this is a java object, it can be compressed 229 Register tmp = index; // reuse 230 lslw(tmp, tmp, LogBytesPerHeapOop); 231 232 get_constant_pool(result); 233 // load pointer for resolved_references[] objArray 234 ldr(result, Address(result, ConstantPool::resolved_references_offset_in_bytes())); 235 // JNIHandles::resolve(obj); 236 ldr(result, Address(result, 0)); 237 // Add in the index 238 add(result, result, tmp); 239 load_heap_oop(result, Address(result, arrayOopDesc::base_offset_in_bytes(T_OBJECT))); 240 } 241 242 // Generate a subtype check: branch to ok_is_subtype if sub_klass is a 243 // subtype of super_klass. 244 // 245 // Args: 246 // r0: superklass 247 // Rsub_klass: subklass 248 // 249 // Kills: 250 // r2, r5 251 void InterpreterMacroAssembler::gen_subtype_check(Register Rsub_klass, 252 Label& ok_is_subtype) { 253 assert(Rsub_klass != r0, "r0 holds superklass"); 254 assert(Rsub_klass != r2, "r2 holds 2ndary super array length"); 255 assert(Rsub_klass != r5, "r5 holds 2ndary super array scan ptr"); 256 257 // Profile the not-null value's klass. 258 profile_typecheck(r2, Rsub_klass, r5); // blows r2, reloads r5 259 260 // Do the check. 261 check_klass_subtype(Rsub_klass, r0, r2, ok_is_subtype); // blows r2 262 263 // Profile the failure of the check. 264 profile_typecheck_failed(r2); // blows r2 265 } 266 267 // Java Expression Stack 268 269 void InterpreterMacroAssembler::pop_ptr(Register r) { 270 ldr(r, post(esp, wordSize)); 271 } 272 273 void InterpreterMacroAssembler::pop_i(Register r) { 274 ldrw(r, post(esp, wordSize)); 275 } 276 277 void InterpreterMacroAssembler::pop_l(Register r) { 278 ldr(r, post(esp, 2 * Interpreter::stackElementSize)); 279 } 280 281 void InterpreterMacroAssembler::push_ptr(Register r) { 282 str(r, pre(esp, -wordSize)); 283 } 284 285 void InterpreterMacroAssembler::push_i(Register r) { 286 str(r, pre(esp, -wordSize)); 287 } 288 289 void InterpreterMacroAssembler::push_l(Register r) { 290 str(r, pre(esp, 2 * -wordSize)); 291 } 292 293 void InterpreterMacroAssembler::pop_f(FloatRegister r) { 294 ldrs(r, post(esp, wordSize)); 295 } 296 297 void InterpreterMacroAssembler::pop_d(FloatRegister r) { 298 ldrd(r, post(esp, 2 * Interpreter::stackElementSize)); 299 } 300 301 void InterpreterMacroAssembler::push_f(FloatRegister r) { 302 strs(r, pre(esp, -wordSize)); 303 } 304 305 void InterpreterMacroAssembler::push_d(FloatRegister r) { 306 strd(r, pre(esp, 2* -wordSize)); 307 } 308 309 void InterpreterMacroAssembler::pop(TosState state) { 310 switch (state) { 311 case atos: pop_ptr(); break; 312 case btos: 313 case ctos: 314 case stos: 315 case itos: pop_i(); break; 316 case ltos: pop_l(); break; 317 case ftos: pop_f(); break; 318 case dtos: pop_d(); break; 319 case vtos: /* nothing to do */ break; 320 default: ShouldNotReachHere(); 321 } 322 verify_oop(r0, state); 323 } 324 325 void InterpreterMacroAssembler::push(TosState state) { 326 verify_oop(r0, state); 327 switch (state) { 328 case atos: push_ptr(); break; 329 case btos: 330 case ctos: 331 case stos: 332 case itos: push_i(); break; 333 case ltos: push_l(); break; 334 case ftos: push_f(); break; 335 case dtos: push_d(); break; 336 case vtos: /* nothing to do */ break; 337 default : ShouldNotReachHere(); 338 } 339 } 340 341 // Helpers for swap and dup 342 void InterpreterMacroAssembler::load_ptr(int n, Register val) { 343 ldr(val, Address(esp, Interpreter::expr_offset_in_bytes(n))); 344 } 345 346 void InterpreterMacroAssembler::store_ptr(int n, Register val) { 347 str(val, Address(esp, Interpreter::expr_offset_in_bytes(n))); 348 } 349 350 351 void InterpreterMacroAssembler::prepare_to_jump_from_interpreted() { 352 // set sender sp 353 mov(r13, sp); 354 // record last_sp 355 str(esp, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 356 } 357 358 // Jump to from_interpreted entry of a call unless single stepping is possible 359 // in this thread in which case we must call the i2i entry 360 void InterpreterMacroAssembler::jump_from_interpreted(Register method, Register temp) { 361 prepare_to_jump_from_interpreted(); 362 363 if (JvmtiExport::can_post_interpreter_events()) { 364 Label run_compiled_code; 365 // JVMTI events, such as single-stepping, are implemented partly by avoiding running 366 // compiled code in threads for which the event is enabled. Check here for 367 // interp_only_mode if these events CAN be enabled. 368 // interp_only is an int, on little endian it is sufficient to test the byte only 369 // Is a cmpl faster? 370 ldr(rscratch1, Address(rthread, JavaThread::interp_only_mode_offset())); 371 cbz(rscratch1, run_compiled_code); 372 ldr(rscratch1, Address(method, Method::interpreter_entry_offset())); 373 br(rscratch1); 374 bind(run_compiled_code); 375 } 376 377 ldr(rscratch1, Address(method, Method::from_interpreted_offset())); 378 br(rscratch1); 379 } 380 381 // The following two routines provide a hook so that an implementation 382 // can schedule the dispatch in two parts. amd64 does not do this. 383 void InterpreterMacroAssembler::dispatch_prolog(TosState state, int step) { 384 } 385 386 void InterpreterMacroAssembler::dispatch_epilog(TosState state, int step) { 387 dispatch_next(state, step); 388 } 389 390 void InterpreterMacroAssembler::dispatch_base(TosState state, 391 address* table, 392 bool verifyoop) { 393 if (VerifyActivationFrameSize) { 394 Unimplemented(); 395 } 396 if (verifyoop) { 397 verify_oop(r0, state); 398 } 399 if (table == Interpreter::dispatch_table(state)) { 400 addw(rscratch2, rscratch1, Interpreter::distance_from_dispatch_table(state)); 401 ldr(rscratch2, Address(rdispatch, rscratch2, Address::uxtw(3))); 402 } else { 403 mov(rscratch2, (address)table); 404 ldr(rscratch2, Address(rscratch2, rscratch1, Address::uxtw(3))); 405 } 406 br(rscratch2); 407 } 408 409 void InterpreterMacroAssembler::dispatch_only(TosState state) { 410 dispatch_base(state, Interpreter::dispatch_table(state)); 411 } 412 413 void InterpreterMacroAssembler::dispatch_only_normal(TosState state) { 414 dispatch_base(state, Interpreter::normal_table(state)); 415 } 416 417 void InterpreterMacroAssembler::dispatch_only_noverify(TosState state) { 418 dispatch_base(state, Interpreter::normal_table(state), false); 419 } 420 421 422 void InterpreterMacroAssembler::dispatch_next(TosState state, int step) { 423 // load next bytecode 424 ldrb(rscratch1, Address(pre(rbcp, step))); 425 dispatch_base(state, Interpreter::dispatch_table(state)); 426 } 427 428 void InterpreterMacroAssembler::dispatch_via(TosState state, address* table) { 429 // load current bytecode 430 ldrb(rscratch1, Address(rbcp, 0)); 431 dispatch_base(state, table); 432 } 433 434 // remove activation 435 // 436 // Unlock the receiver if this is a synchronized method. 437 // Unlock any Java monitors from syncronized blocks. 438 // Remove the activation from the stack. 439 // 440 // If there are locked Java monitors 441 // If throw_monitor_exception 442 // throws IllegalMonitorStateException 443 // Else if install_monitor_exception 444 // installs IllegalMonitorStateException 445 // Else 446 // no error processing 447 void InterpreterMacroAssembler::remove_activation( 448 TosState state, 449 bool throw_monitor_exception, 450 bool install_monitor_exception, 451 bool notify_jvmdi) { 452 // Note: Registers r3 xmm0 may be in use for the 453 // result check if synchronized method 454 Label unlocked, unlock, no_unlock; 455 456 // get the value of _do_not_unlock_if_synchronized into r3 457 const Address do_not_unlock_if_synchronized(rthread, 458 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset())); 459 ldrb(r3, do_not_unlock_if_synchronized); 460 strb(zr, do_not_unlock_if_synchronized); // reset the flag 461 462 // get method access flags 463 ldr(r1, Address(rfp, frame::interpreter_frame_method_offset * wordSize)); 464 ldr(r2, Address(r1, Method::access_flags_offset())); 465 tst(r2, JVM_ACC_SYNCHRONIZED); 466 br(Assembler::EQ, unlocked); 467 468 // Don't unlock anything if the _do_not_unlock_if_synchronized flag 469 // is set. 470 cbnz(r3, no_unlock); 471 472 // unlock monitor 473 push(state); // save result 474 475 // BasicObjectLock will be first in list, since this is a 476 // synchronized method. However, need to check that the object has 477 // not been unlocked by an explicit monitorexit bytecode. 478 const Address monitor(rfp, frame::interpreter_frame_initial_sp_offset * 479 wordSize - (int) sizeof(BasicObjectLock)); 480 // We use c_rarg1 so that if we go slow path it will be the correct 481 // register for unlock_object to pass to VM directly 482 lea(c_rarg1, monitor); // address of first monitor 483 484 ldr(r0, Address(c_rarg1, BasicObjectLock::obj_offset_in_bytes())); 485 cbnz(r0, unlock); 486 487 pop(state); 488 if (throw_monitor_exception) { 489 // Entry already unlocked, need to throw exception 490 call_VM(noreg, CAST_FROM_FN_PTR(address, 491 InterpreterRuntime::throw_illegal_monitor_state_exception)); 492 should_not_reach_here(); 493 } else { 494 // Monitor already unlocked during a stack unroll. If requested, 495 // install an illegal_monitor_state_exception. Continue with 496 // stack unrolling. 497 if (install_monitor_exception) { 498 call_VM(noreg, CAST_FROM_FN_PTR(address, 499 InterpreterRuntime::new_illegal_monitor_state_exception)); 500 } 501 b(unlocked); 502 } 503 504 bind(unlock); 505 unlock_object(c_rarg1); 506 pop(state); 507 508 // Check that for block-structured locking (i.e., that all locked 509 // objects has been unlocked) 510 bind(unlocked); 511 512 // r0: Might contain return value 513 514 // Check that all monitors are unlocked 515 { 516 Label loop, exception, entry, restart; 517 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize; 518 const Address monitor_block_top( 519 rfp, frame::interpreter_frame_monitor_block_top_offset * wordSize); 520 const Address monitor_block_bot( 521 rfp, frame::interpreter_frame_initial_sp_offset * wordSize); 522 523 bind(restart); 524 // We use c_rarg1 so that if we go slow path it will be the correct 525 // register for unlock_object to pass to VM directly 526 ldr(c_rarg1, monitor_block_top); // points to current entry, starting 527 // with top-most entry 528 lea(r19, monitor_block_bot); // points to word before bottom of 529 // monitor block 530 b(entry); 531 532 // Entry already locked, need to throw exception 533 bind(exception); 534 535 if (throw_monitor_exception) { 536 // Throw exception 537 MacroAssembler::call_VM(noreg, 538 CAST_FROM_FN_PTR(address, InterpreterRuntime:: 539 throw_illegal_monitor_state_exception)); 540 should_not_reach_here(); 541 } else { 542 // Stack unrolling. Unlock object and install illegal_monitor_exception. 543 // Unlock does not block, so don't have to worry about the frame. 544 // We don't have to preserve c_rarg1 since we are going to throw an exception. 545 546 push(state); 547 unlock_object(c_rarg1); 548 pop(state); 549 550 if (install_monitor_exception) { 551 call_VM(noreg, CAST_FROM_FN_PTR(address, 552 InterpreterRuntime:: 553 new_illegal_monitor_state_exception)); 554 } 555 556 b(restart); 557 } 558 559 bind(loop); 560 // check if current entry is used 561 ldr(rscratch1, Address(c_rarg1, BasicObjectLock::obj_offset_in_bytes())); 562 cbnz(rscratch1, exception); 563 564 add(c_rarg1, c_rarg1, entry_size); // otherwise advance to next entry 565 bind(entry); 566 cmp(c_rarg1, r19); // check if bottom reached 567 br(Assembler::NE, loop); // if not at bottom then check this entry 568 } 569 570 bind(no_unlock); 571 572 // jvmti support 573 if (notify_jvmdi) { 574 notify_method_exit(state, NotifyJVMTI); // preserve TOSCA 575 } else { 576 notify_method_exit(state, SkipNotifyJVMTI); // preserve TOSCA 577 } 578 579 // remove activation 580 // get sender esp 581 ldr(esp, 582 Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); 583 // remove frame anchor 584 leave(); 585 // If we're returning to interpreted code we will shortly be 586 // adjusting SP to allow some space for ESP. If we're returning to 587 // compiled code the saved sender SP was saved in sender_sp, so this 588 // restores it. 589 andr(sp, esp, -16); 590 } 591 592 #endif // C_INTERP 593 594 // Lock object 595 // 596 // Args: 597 // c_rarg1: BasicObjectLock to be used for locking 598 // 599 // Kills: 600 // r0 601 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, .. (param regs) 602 // rscratch1, rscratch2 (scratch regs) 603 void InterpreterMacroAssembler::lock_object(Register lock_reg) 604 { 605 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be c_rarg1"); 606 if (UseHeavyMonitors) { 607 call_VM(noreg, 608 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter), 609 lock_reg); 610 } else { 611 Label done; 612 613 const Register swap_reg = r0; 614 const Register obj_reg = c_rarg3; // Will contain the oop 615 616 const int obj_offset = BasicObjectLock::obj_offset_in_bytes(); 617 const int lock_offset = BasicObjectLock::lock_offset_in_bytes (); 618 const int mark_offset = lock_offset + 619 BasicLock::displaced_header_offset_in_bytes(); 620 621 Label slow_case; 622 623 // Load object pointer into obj_reg %c_rarg3 624 ldr(obj_reg, Address(lock_reg, obj_offset)); 625 626 if (UseBiasedLocking) { 627 biased_locking_enter(lock_reg, obj_reg, swap_reg, rscratch2, false, done, &slow_case); 628 } 629 630 // Load (object->mark() | 1) into swap_reg 631 ldr(rscratch1, Address(obj_reg, 0)); 632 orr(swap_reg, rscratch1, 1); 633 634 // Save (object->mark() | 1) into BasicLock's displaced header 635 str(swap_reg, Address(lock_reg, mark_offset)); 636 637 assert(lock_offset == 0, 638 "displached header must be first word in BasicObjectLock"); 639 640 Label fail; 641 if (PrintBiasedLockingStatistics) { 642 Label fast; 643 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, fast, &fail); 644 bind(fast); 645 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()), 646 rscratch2, rscratch1); 647 b(done); 648 bind(fail); 649 } else { 650 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL); 651 } 652 653 // Test if the oopMark is an obvious stack pointer, i.e., 654 // 1) (mark & 7) == 0, and 655 // 2) rsp <= mark < mark + os::pagesize() 656 // 657 // These 3 tests can be done by evaluating the following 658 // expression: ((mark - rsp) & (7 - os::vm_page_size())), 659 // assuming both stack pointer and pagesize have their 660 // least significant 3 bits clear. 661 // NOTE: the oopMark is in swap_reg %r0 as the result of cmpxchg 662 // NOTE2: aarch64 does not like to subtract sp from rn so take a 663 // copy 664 mov(rscratch1, sp); 665 sub(swap_reg, swap_reg, rscratch1); 666 ands(swap_reg, swap_reg, (unsigned long)(7 - os::vm_page_size())); 667 668 // Save the test result, for recursive case, the result is zero 669 str(swap_reg, Address(lock_reg, mark_offset)); 670 671 if (PrintBiasedLockingStatistics) { 672 br(Assembler::NE, slow_case); 673 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()), 674 rscratch2, rscratch1); 675 } 676 br(Assembler::EQ, done); 677 678 bind(slow_case); 679 680 // Call the runtime routine for slow case 681 call_VM(noreg, 682 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter), 683 lock_reg); 684 685 bind(done); 686 } 687 } 688 689 690 // Unlocks an object. Used in monitorexit bytecode and 691 // remove_activation. Throws an IllegalMonitorException if object is 692 // not locked by current thread. 693 // 694 // Args: 695 // c_rarg1: BasicObjectLock for lock 696 // 697 // Kills: 698 // r0 699 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, ... (param regs) 700 // rscratch1, rscratch2 (scratch regs) 701 void InterpreterMacroAssembler::unlock_object(Register lock_reg) 702 { 703 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be rarg1"); 704 705 if (UseHeavyMonitors) { 706 call_VM(noreg, 707 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit), 708 lock_reg); 709 } else { 710 Label done; 711 712 const Register swap_reg = r0; 713 const Register header_reg = c_rarg2; // Will contain the old oopMark 714 const Register obj_reg = c_rarg3; // Will contain the oop 715 716 save_bcp(); // Save in case of exception 717 718 // Convert from BasicObjectLock structure to object and BasicLock 719 // structure Store the BasicLock address into %r0 720 lea(swap_reg, Address(lock_reg, BasicObjectLock::lock_offset_in_bytes())); 721 722 // Load oop into obj_reg(%c_rarg3) 723 ldr(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); 724 725 // Free entry 726 str(zr, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); 727 728 if (UseBiasedLocking) { 729 biased_locking_exit(obj_reg, header_reg, done); 730 } 731 732 // Load the old header from BasicLock structure 733 ldr(header_reg, Address(swap_reg, 734 BasicLock::displaced_header_offset_in_bytes())); 735 736 // Test for recursion 737 cbz(header_reg, done); 738 739 // Atomic swap back the old header 740 cmpxchgptr(swap_reg, header_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL); 741 742 // Call the runtime routine for slow case. 743 str(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); // restore obj 744 call_VM(noreg, 745 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit), 746 lock_reg); 747 748 bind(done); 749 750 restore_bcp(); 751 } 752 } 753 754 #ifndef CC_INTERP 755 756 void InterpreterMacroAssembler::test_method_data_pointer(Register mdp, 757 Label& zero_continue) { 758 assert(ProfileInterpreter, "must be profiling interpreter"); 759 ldr(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 760 cbz(mdp, zero_continue); 761 } 762 763 // Set the method data pointer for the current bcp. 764 void InterpreterMacroAssembler::set_method_data_pointer_for_bcp() { 765 assert(ProfileInterpreter, "must be profiling interpreter"); 766 Label set_mdp; 767 stp(r0, r1, Address(pre(sp, -2 * wordSize))); 768 769 // Test MDO to avoid the call if it is NULL. 770 ldr(r0, Address(rmethod, in_bytes(Method::method_data_offset()))); 771 cbz(r0, set_mdp); 772 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::bcp_to_di), rmethod, rbcp); 773 // r0: mdi 774 // mdo is guaranteed to be non-zero here, we checked for it before the call. 775 ldr(r1, Address(rmethod, in_bytes(Method::method_data_offset()))); 776 lea(r1, Address(r1, in_bytes(MethodData::data_offset()))); 777 add(r0, r1, r0); 778 str(r0, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 779 bind(set_mdp); 780 ldp(r0, r1, Address(post(sp, 2 * wordSize))); 781 } 782 783 void InterpreterMacroAssembler::verify_method_data_pointer() { 784 assert(ProfileInterpreter, "must be profiling interpreter"); 785 #ifdef ASSERT 786 Label verify_continue; 787 stp(r0, r1, Address(pre(sp, -2 * wordSize))); 788 stp(r2, r3, Address(pre(sp, -2 * wordSize))); 789 test_method_data_pointer(r3, verify_continue); // If mdp is zero, continue 790 get_method(r1); 791 792 // If the mdp is valid, it will point to a DataLayout header which is 793 // consistent with the bcp. The converse is highly probable also. 794 ldrsh(r2, Address(r3, in_bytes(DataLayout::bci_offset()))); 795 ldr(rscratch1, Address(r1, Method::const_offset())); 796 add(r2, r2, rscratch1, Assembler::LSL); 797 lea(r2, Address(r2, ConstMethod::codes_offset())); 798 cmp(r2, rbcp); 799 br(Assembler::EQ, verify_continue); 800 // r1: method 801 // rbcp: bcp // rbcp == 22 802 // r3: mdp 803 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::verify_mdp), 804 r1, rbcp, r3); 805 bind(verify_continue); 806 ldp(r2, r3, Address(post(sp, 2 * wordSize))); 807 ldp(r0, r1, Address(post(sp, 2 * wordSize))); 808 #endif // ASSERT 809 } 810 811 812 void InterpreterMacroAssembler::set_mdp_data_at(Register mdp_in, 813 int constant, 814 Register value) { 815 assert(ProfileInterpreter, "must be profiling interpreter"); 816 Address data(mdp_in, constant); 817 str(value, data); 818 } 819 820 821 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in, 822 int constant, 823 bool decrement) { 824 increment_mdp_data_at(mdp_in, noreg, constant, decrement); 825 } 826 827 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in, 828 Register reg, 829 int constant, 830 bool decrement) { 831 assert(ProfileInterpreter, "must be profiling interpreter"); 832 // %%% this does 64bit counters at best it is wasting space 833 // at worst it is a rare bug when counters overflow 834 835 assert_different_registers(rscratch2, rscratch1, mdp_in, reg); 836 837 Address addr1(mdp_in, constant); 838 Address addr2(rscratch2, reg, Address::lsl(0)); 839 Address &addr = addr1; 840 if (reg != noreg) { 841 lea(rscratch2, addr1); 842 addr = addr2; 843 } 844 845 if (decrement) { 846 // Decrement the register. Set condition codes. 847 // Intel does this 848 // addptr(data, (int32_t) -DataLayout::counter_increment); 849 // If the decrement causes the counter to overflow, stay negative 850 // Label L; 851 // jcc(Assembler::negative, L); 852 // addptr(data, (int32_t) DataLayout::counter_increment); 853 // so we do this 854 ldr(rscratch1, addr); 855 subs(rscratch1, rscratch1, (unsigned)DataLayout::counter_increment); 856 Label L; 857 br(Assembler::LO, L); // skip store if counter underflow 858 str(rscratch1, addr); 859 bind(L); 860 } else { 861 assert(DataLayout::counter_increment == 1, 862 "flow-free idiom only works with 1"); 863 // Intel does this 864 // Increment the register. Set carry flag. 865 // addptr(data, DataLayout::counter_increment); 866 // If the increment causes the counter to overflow, pull back by 1. 867 // sbbptr(data, (int32_t)0); 868 // so we do this 869 ldr(rscratch1, addr); 870 adds(rscratch1, rscratch1, DataLayout::counter_increment); 871 Label L; 872 br(Assembler::CS, L); // skip store if counter overflow 873 str(rscratch1, addr); 874 bind(L); 875 } 876 } 877 878 void InterpreterMacroAssembler::set_mdp_flag_at(Register mdp_in, 879 int flag_byte_constant) { 880 assert(ProfileInterpreter, "must be profiling interpreter"); 881 int header_offset = in_bytes(DataLayout::header_offset()); 882 int header_bits = DataLayout::flag_mask_to_header_mask(flag_byte_constant); 883 // Set the flag 884 ldr(rscratch1, Address(mdp_in, header_offset)); 885 orr(rscratch1, rscratch1, header_bits); 886 str(rscratch1, Address(mdp_in, header_offset)); 887 } 888 889 890 void InterpreterMacroAssembler::test_mdp_data_at(Register mdp_in, 891 int offset, 892 Register value, 893 Register test_value_out, 894 Label& not_equal_continue) { 895 assert(ProfileInterpreter, "must be profiling interpreter"); 896 if (test_value_out == noreg) { 897 ldr(rscratch1, Address(mdp_in, offset)); 898 cmp(value, rscratch1); 899 } else { 900 // Put the test value into a register, so caller can use it: 901 ldr(test_value_out, Address(mdp_in, offset)); 902 cmp(value, test_value_out); 903 } 904 br(Assembler::NE, not_equal_continue); 905 } 906 907 908 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in, 909 int offset_of_disp) { 910 assert(ProfileInterpreter, "must be profiling interpreter"); 911 ldr(rscratch1, Address(mdp_in, offset_of_disp)); 912 add(mdp_in, mdp_in, rscratch1, LSL); 913 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 914 } 915 916 917 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in, 918 Register reg, 919 int offset_of_disp) { 920 assert(ProfileInterpreter, "must be profiling interpreter"); 921 lea(rscratch1, Address(mdp_in, offset_of_disp)); 922 ldr(rscratch1, Address(rscratch1, reg, Address::lsl(0))); 923 add(mdp_in, mdp_in, rscratch1, LSL); 924 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 925 } 926 927 928 void InterpreterMacroAssembler::update_mdp_by_constant(Register mdp_in, 929 int constant) { 930 assert(ProfileInterpreter, "must be profiling interpreter"); 931 add(mdp_in, mdp_in, (unsigned)constant); 932 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 933 } 934 935 936 void InterpreterMacroAssembler::update_mdp_for_ret(Register return_bci) { 937 assert(ProfileInterpreter, "must be profiling interpreter"); 938 // save/restore across call_VM 939 stp(zr, return_bci, Address(pre(sp, -2 * wordSize))); 940 call_VM(noreg, 941 CAST_FROM_FN_PTR(address, InterpreterRuntime::update_mdp_for_ret), 942 return_bci); 943 ldp(zr, return_bci, Address(post(sp, 2 * wordSize))); 944 } 945 946 947 void InterpreterMacroAssembler::profile_taken_branch(Register mdp, 948 Register bumped_count) { 949 if (ProfileInterpreter) { 950 Label profile_continue; 951 952 // If no method data exists, go to profile_continue. 953 // Otherwise, assign to mdp 954 test_method_data_pointer(mdp, profile_continue); 955 956 // We are taking a branch. Increment the taken count. 957 // We inline increment_mdp_data_at to return bumped_count in a register 958 //increment_mdp_data_at(mdp, in_bytes(JumpData::taken_offset())); 959 Address data(mdp, in_bytes(JumpData::taken_offset())); 960 ldr(bumped_count, data); 961 assert(DataLayout::counter_increment == 1, 962 "flow-free idiom only works with 1"); 963 // Intel does this to catch overflow 964 // addptr(bumped_count, DataLayout::counter_increment); 965 // sbbptr(bumped_count, 0); 966 // so we do this 967 adds(bumped_count, bumped_count, DataLayout::counter_increment); 968 Label L; 969 br(Assembler::CS, L); // skip store if counter overflow 970 str(bumped_count, data); 971 bind(L); 972 // The method data pointer needs to be updated to reflect the new target. 973 update_mdp_by_offset(mdp, in_bytes(JumpData::displacement_offset())); 974 bind(profile_continue); 975 } 976 } 977 978 979 void InterpreterMacroAssembler::profile_not_taken_branch(Register mdp) { 980 if (ProfileInterpreter) { 981 Label profile_continue; 982 983 // If no method data exists, go to profile_continue. 984 test_method_data_pointer(mdp, profile_continue); 985 986 // We are taking a branch. Increment the not taken count. 987 increment_mdp_data_at(mdp, in_bytes(BranchData::not_taken_offset())); 988 989 // The method data pointer needs to be updated to correspond to 990 // the next bytecode 991 update_mdp_by_constant(mdp, in_bytes(BranchData::branch_data_size())); 992 bind(profile_continue); 993 } 994 } 995 996 997 void InterpreterMacroAssembler::profile_call(Register mdp) { 998 if (ProfileInterpreter) { 999 Label profile_continue; 1000 1001 // If no method data exists, go to profile_continue. 1002 test_method_data_pointer(mdp, profile_continue); 1003 1004 // We are making a call. Increment the count. 1005 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1006 1007 // The method data pointer needs to be updated to reflect the new target. 1008 update_mdp_by_constant(mdp, in_bytes(CounterData::counter_data_size())); 1009 bind(profile_continue); 1010 } 1011 } 1012 1013 void InterpreterMacroAssembler::profile_final_call(Register mdp) { 1014 if (ProfileInterpreter) { 1015 Label profile_continue; 1016 1017 // If no method data exists, go to profile_continue. 1018 test_method_data_pointer(mdp, profile_continue); 1019 1020 // We are making a call. Increment the count. 1021 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1022 1023 // The method data pointer needs to be updated to reflect the new target. 1024 update_mdp_by_constant(mdp, 1025 in_bytes(VirtualCallData:: 1026 virtual_call_data_size())); 1027 bind(profile_continue); 1028 } 1029 } 1030 1031 1032 void InterpreterMacroAssembler::profile_virtual_call(Register receiver, 1033 Register mdp, 1034 Register reg2, 1035 bool receiver_can_be_null) { 1036 if (ProfileInterpreter) { 1037 Label profile_continue; 1038 1039 // If no method data exists, go to profile_continue. 1040 test_method_data_pointer(mdp, profile_continue); 1041 1042 Label skip_receiver_profile; 1043 if (receiver_can_be_null) { 1044 Label not_null; 1045 // We are making a call. Increment the count for null receiver. 1046 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1047 b(skip_receiver_profile); 1048 bind(not_null); 1049 } 1050 1051 // Record the receiver type. 1052 record_klass_in_profile(receiver, mdp, reg2, true); 1053 bind(skip_receiver_profile); 1054 1055 // The method data pointer needs to be updated to reflect the new target. 1056 update_mdp_by_constant(mdp, 1057 in_bytes(VirtualCallData:: 1058 virtual_call_data_size())); 1059 bind(profile_continue); 1060 } 1061 } 1062 1063 // This routine creates a state machine for updating the multi-row 1064 // type profile at a virtual call site (or other type-sensitive bytecode). 1065 // The machine visits each row (of receiver/count) until the receiver type 1066 // is found, or until it runs out of rows. At the same time, it remembers 1067 // the location of the first empty row. (An empty row records null for its 1068 // receiver, and can be allocated for a newly-observed receiver type.) 1069 // Because there are two degrees of freedom in the state, a simple linear 1070 // search will not work; it must be a decision tree. Hence this helper 1071 // function is recursive, to generate the required tree structured code. 1072 // It's the interpreter, so we are trading off code space for speed. 1073 // See below for example code. 1074 void InterpreterMacroAssembler::record_klass_in_profile_helper( 1075 Register receiver, Register mdp, 1076 Register reg2, int start_row, 1077 Label& done, bool is_virtual_call) { 1078 if (TypeProfileWidth == 0) { 1079 if (is_virtual_call) { 1080 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1081 } 1082 return; 1083 } 1084 1085 int last_row = VirtualCallData::row_limit() - 1; 1086 assert(start_row <= last_row, "must be work left to do"); 1087 // Test this row for both the receiver and for null. 1088 // Take any of three different outcomes: 1089 // 1. found receiver => increment count and goto done 1090 // 2. found null => keep looking for case 1, maybe allocate this cell 1091 // 3. found something else => keep looking for cases 1 and 2 1092 // Case 3 is handled by a recursive call. 1093 for (int row = start_row; row <= last_row; row++) { 1094 Label next_test; 1095 bool test_for_null_also = (row == start_row); 1096 1097 // See if the receiver is receiver[n]. 1098 int recvr_offset = in_bytes(VirtualCallData::receiver_offset(row)); 1099 test_mdp_data_at(mdp, recvr_offset, receiver, 1100 (test_for_null_also ? reg2 : noreg), 1101 next_test); 1102 // (Reg2 now contains the receiver from the CallData.) 1103 1104 // The receiver is receiver[n]. Increment count[n]. 1105 int count_offset = in_bytes(VirtualCallData::receiver_count_offset(row)); 1106 increment_mdp_data_at(mdp, count_offset); 1107 b(done); 1108 bind(next_test); 1109 1110 if (test_for_null_also) { 1111 Label found_null; 1112 // Failed the equality check on receiver[n]... Test for null. 1113 if (start_row == last_row) { 1114 // The only thing left to do is handle the null case. 1115 if (is_virtual_call) { 1116 cbz(reg2, found_null); 1117 // Receiver did not match any saved receiver and there is no empty row for it. 1118 // Increment total counter to indicate polymorphic case. 1119 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1120 b(done); 1121 bind(found_null); 1122 } else { 1123 cbz(reg2, done); 1124 } 1125 break; 1126 } 1127 // Since null is rare, make it be the branch-taken case. 1128 cbz(reg2,found_null); 1129 1130 // Put all the "Case 3" tests here. 1131 record_klass_in_profile_helper(receiver, mdp, reg2, start_row + 1, done, is_virtual_call); 1132 1133 // Found a null. Keep searching for a matching receiver, 1134 // but remember that this is an empty (unused) slot. 1135 bind(found_null); 1136 } 1137 } 1138 1139 // In the fall-through case, we found no matching receiver, but we 1140 // observed the receiver[start_row] is NULL. 1141 1142 // Fill in the receiver field and increment the count. 1143 int recvr_offset = in_bytes(VirtualCallData::receiver_offset(start_row)); 1144 set_mdp_data_at(mdp, recvr_offset, receiver); 1145 int count_offset = in_bytes(VirtualCallData::receiver_count_offset(start_row)); 1146 mov(reg2, DataLayout::counter_increment); 1147 set_mdp_data_at(mdp, count_offset, reg2); 1148 if (start_row > 0) { 1149 b(done); 1150 } 1151 } 1152 1153 // Example state machine code for three profile rows: 1154 // // main copy of decision tree, rooted at row[1] 1155 // if (row[0].rec == rec) { row[0].incr(); goto done; } 1156 // if (row[0].rec != NULL) { 1157 // // inner copy of decision tree, rooted at row[1] 1158 // if (row[1].rec == rec) { row[1].incr(); goto done; } 1159 // if (row[1].rec != NULL) { 1160 // // degenerate decision tree, rooted at row[2] 1161 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1162 // if (row[2].rec != NULL) { count.incr(); goto done; } // overflow 1163 // row[2].init(rec); goto done; 1164 // } else { 1165 // // remember row[1] is empty 1166 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1167 // row[1].init(rec); goto done; 1168 // } 1169 // } else { 1170 // // remember row[0] is empty 1171 // if (row[1].rec == rec) { row[1].incr(); goto done; } 1172 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1173 // row[0].init(rec); goto done; 1174 // } 1175 // done: 1176 1177 void InterpreterMacroAssembler::record_klass_in_profile(Register receiver, 1178 Register mdp, Register reg2, 1179 bool is_virtual_call) { 1180 assert(ProfileInterpreter, "must be profiling"); 1181 Label done; 1182 1183 record_klass_in_profile_helper(receiver, mdp, reg2, 0, done, is_virtual_call); 1184 1185 bind (done); 1186 } 1187 1188 void InterpreterMacroAssembler::profile_ret(Register return_bci, 1189 Register mdp) { 1190 if (ProfileInterpreter) { 1191 Label profile_continue; 1192 uint row; 1193 1194 // If no method data exists, go to profile_continue. 1195 test_method_data_pointer(mdp, profile_continue); 1196 1197 // Update the total ret count. 1198 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1199 1200 for (row = 0; row < RetData::row_limit(); row++) { 1201 Label next_test; 1202 1203 // See if return_bci is equal to bci[n]: 1204 test_mdp_data_at(mdp, 1205 in_bytes(RetData::bci_offset(row)), 1206 return_bci, noreg, 1207 next_test); 1208 1209 // return_bci is equal to bci[n]. Increment the count. 1210 increment_mdp_data_at(mdp, in_bytes(RetData::bci_count_offset(row))); 1211 1212 // The method data pointer needs to be updated to reflect the new target. 1213 update_mdp_by_offset(mdp, 1214 in_bytes(RetData::bci_displacement_offset(row))); 1215 b(profile_continue); 1216 bind(next_test); 1217 } 1218 1219 update_mdp_for_ret(return_bci); 1220 1221 bind(profile_continue); 1222 } 1223 } 1224 1225 void InterpreterMacroAssembler::profile_null_seen(Register mdp) { 1226 if (ProfileInterpreter) { 1227 Label profile_continue; 1228 1229 // If no method data exists, go to profile_continue. 1230 test_method_data_pointer(mdp, profile_continue); 1231 1232 set_mdp_flag_at(mdp, BitData::null_seen_byte_constant()); 1233 1234 // The method data pointer needs to be updated. 1235 int mdp_delta = in_bytes(BitData::bit_data_size()); 1236 if (TypeProfileCasts) { 1237 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size()); 1238 } 1239 update_mdp_by_constant(mdp, mdp_delta); 1240 1241 bind(profile_continue); 1242 } 1243 } 1244 1245 void InterpreterMacroAssembler::profile_typecheck_failed(Register mdp) { 1246 if (ProfileInterpreter && TypeProfileCasts) { 1247 Label profile_continue; 1248 1249 // If no method data exists, go to profile_continue. 1250 test_method_data_pointer(mdp, profile_continue); 1251 1252 int count_offset = in_bytes(CounterData::count_offset()); 1253 // Back up the address, since we have already bumped the mdp. 1254 count_offset -= in_bytes(VirtualCallData::virtual_call_data_size()); 1255 1256 // *Decrement* the counter. We expect to see zero or small negatives. 1257 increment_mdp_data_at(mdp, count_offset, true); 1258 1259 bind (profile_continue); 1260 } 1261 } 1262 1263 void InterpreterMacroAssembler::profile_typecheck(Register mdp, Register klass, Register reg2) { 1264 if (ProfileInterpreter) { 1265 Label profile_continue; 1266 1267 // If no method data exists, go to profile_continue. 1268 test_method_data_pointer(mdp, profile_continue); 1269 1270 // The method data pointer needs to be updated. 1271 int mdp_delta = in_bytes(BitData::bit_data_size()); 1272 if (TypeProfileCasts) { 1273 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size()); 1274 1275 // Record the object type. 1276 record_klass_in_profile(klass, mdp, reg2, false); 1277 } 1278 update_mdp_by_constant(mdp, mdp_delta); 1279 1280 bind(profile_continue); 1281 } 1282 } 1283 1284 void InterpreterMacroAssembler::profile_switch_default(Register mdp) { 1285 if (ProfileInterpreter) { 1286 Label profile_continue; 1287 1288 // If no method data exists, go to profile_continue. 1289 test_method_data_pointer(mdp, profile_continue); 1290 1291 // Update the default case count 1292 increment_mdp_data_at(mdp, 1293 in_bytes(MultiBranchData::default_count_offset())); 1294 1295 // The method data pointer needs to be updated. 1296 update_mdp_by_offset(mdp, 1297 in_bytes(MultiBranchData:: 1298 default_displacement_offset())); 1299 1300 bind(profile_continue); 1301 } 1302 } 1303 1304 void InterpreterMacroAssembler::profile_switch_case(Register index, 1305 Register mdp, 1306 Register reg2) { 1307 if (ProfileInterpreter) { 1308 Label profile_continue; 1309 1310 // If no method data exists, go to profile_continue. 1311 test_method_data_pointer(mdp, profile_continue); 1312 1313 // Build the base (index * per_case_size_in_bytes()) + 1314 // case_array_offset_in_bytes() 1315 movw(reg2, in_bytes(MultiBranchData::per_case_size())); 1316 movw(rscratch1, in_bytes(MultiBranchData::case_array_offset())); 1317 Assembler::maddw(index, index, reg2, rscratch1); 1318 1319 // Update the case count 1320 increment_mdp_data_at(mdp, 1321 index, 1322 in_bytes(MultiBranchData::relative_count_offset())); 1323 1324 // The method data pointer needs to be updated. 1325 update_mdp_by_offset(mdp, 1326 index, 1327 in_bytes(MultiBranchData:: 1328 relative_displacement_offset())); 1329 1330 bind(profile_continue); 1331 } 1332 } 1333 1334 void InterpreterMacroAssembler::verify_oop(Register reg, TosState state) { 1335 if (state == atos) { 1336 MacroAssembler::verify_oop(reg); 1337 } 1338 } 1339 1340 void InterpreterMacroAssembler::verify_FPU(int stack_depth, TosState state) { ; } 1341 #endif // !CC_INTERP 1342 1343 1344 void InterpreterMacroAssembler::notify_method_entry() { 1345 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to 1346 // track stack depth. If it is possible to enter interp_only_mode we add 1347 // the code to check if the event should be sent. 1348 if (JvmtiExport::can_post_interpreter_events()) { 1349 Label L; 1350 ldr(r3, Address(rthread, JavaThread::interp_only_mode_offset())); 1351 tst(r3, ~0); 1352 br(Assembler::EQ, L); 1353 call_VM(noreg, CAST_FROM_FN_PTR(address, 1354 InterpreterRuntime::post_method_entry)); 1355 bind(L); 1356 } 1357 1358 { 1359 SkipIfEqual skip(this, &DTraceMethodProbes, false); 1360 get_method(c_rarg1); 1361 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 1362 rthread, c_rarg1); 1363 } 1364 1365 // RedefineClasses() tracing support for obsolete method entry 1366 if (RC_TRACE_IN_RANGE(0x00001000, 0x00002000)) { 1367 get_method(c_rarg1); 1368 call_VM_leaf( 1369 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 1370 rthread, c_rarg1); 1371 } 1372 1373 } 1374 1375 1376 void InterpreterMacroAssembler::notify_method_exit( 1377 TosState state, NotifyMethodExitMode mode) { 1378 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to 1379 // track stack depth. If it is possible to enter interp_only_mode we add 1380 // the code to check if the event should be sent. 1381 if (mode == NotifyJVMTI && JvmtiExport::can_post_interpreter_events()) { 1382 Label L; 1383 // Note: frame::interpreter_frame_result has a dependency on how the 1384 // method result is saved across the call to post_method_exit. If this 1385 // is changed then the interpreter_frame_result implementation will 1386 // need to be updated too. 1387 1388 // For c++ interpreter the result is always stored at a known location in the frame 1389 // template interpreter will leave it on the top of the stack. 1390 NOT_CC_INTERP(push(state);) 1391 ldrw(r3, Address(rthread, JavaThread::interp_only_mode_offset())); 1392 cbz(r3, L); 1393 call_VM(noreg, 1394 CAST_FROM_FN_PTR(address, InterpreterRuntime::post_method_exit)); 1395 bind(L); 1396 NOT_CC_INTERP(pop(state)); 1397 } 1398 1399 { 1400 SkipIfEqual skip(this, &DTraceMethodProbes, false); 1401 NOT_CC_INTERP(push(state)); 1402 get_method(c_rarg1); 1403 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 1404 rthread, c_rarg1); 1405 NOT_CC_INTERP(pop(state)); 1406 } 1407 } 1408 1409 1410 // Jump if ((*counter_addr += increment) & mask) satisfies the condition. 1411 void InterpreterMacroAssembler::increment_mask_and_jump(Address counter_addr, 1412 int increment, Address mask, 1413 Register scratch, Register scratch2, 1414 bool preloaded, Condition cond, 1415 Label* where) { 1416 if (!preloaded) { 1417 ldrw(scratch, counter_addr); 1418 } 1419 add(scratch, scratch, increment); 1420 strw(scratch, counter_addr); 1421 ldrw(scratch2, mask); 1422 ands(scratch, scratch, scratch2); 1423 br(cond, *where); 1424 } 1425 1426 void InterpreterMacroAssembler::call_VM_leaf_base(address entry_point, 1427 int number_of_arguments) { 1428 // interpreter specific 1429 // 1430 // Note: No need to save/restore rbcp & rlocals pointer since these 1431 // are callee saved registers and no blocking/ GC can happen 1432 // in leaf calls. 1433 #ifdef ASSERT 1434 { 1435 Label L; 1436 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 1437 cbz(rscratch1, L); 1438 stop("InterpreterMacroAssembler::call_VM_leaf_base:" 1439 " last_sp != NULL"); 1440 bind(L); 1441 } 1442 #endif /* ASSERT */ 1443 // super call 1444 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments); 1445 } 1446 1447 void InterpreterMacroAssembler::call_VM_base(Register oop_result, 1448 Register java_thread, 1449 Register last_java_sp, 1450 address entry_point, 1451 int number_of_arguments, 1452 bool check_exceptions) { 1453 // interpreter specific 1454 // 1455 // Note: Could avoid restoring locals ptr (callee saved) - however doesn't 1456 // really make a difference for these runtime calls, since they are 1457 // slow anyway. Btw., bcp must be saved/restored since it may change 1458 // due to GC. 1459 // assert(java_thread == noreg , "not expecting a precomputed java thread"); 1460 save_bcp(); 1461 #ifdef ASSERT 1462 { 1463 Label L; 1464 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 1465 cbz(rscratch1, L); 1466 stop("InterpreterMacroAssembler::call_VM_leaf_base:" 1467 " last_sp != NULL"); 1468 bind(L); 1469 } 1470 #endif /* ASSERT */ 1471 // super call 1472 MacroAssembler::call_VM_base(oop_result, noreg, last_java_sp, 1473 entry_point, number_of_arguments, 1474 check_exceptions); 1475 // interpreter specific 1476 restore_bcp(); 1477 restore_locals(); 1478 } 1479 1480 void InterpreterMacroAssembler::profile_obj_type(Register obj, const Address& mdo_addr) { 1481 Label update, next, none; 1482 1483 verify_oop(obj); 1484 1485 cbnz(obj, update); 1486 orptr(mdo_addr, TypeEntries::null_seen); 1487 b(next); 1488 1489 bind(update); 1490 load_klass(obj, obj); 1491 1492 ldr(rscratch1, mdo_addr); 1493 eor(obj, obj, rscratch1); 1494 tst(obj, TypeEntries::type_klass_mask); 1495 br(Assembler::EQ, next); // klass seen before, nothing to 1496 // do. The unknown bit may have been 1497 // set already but no need to check. 1498 1499 tst(obj, TypeEntries::type_unknown); 1500 br(Assembler::NE, next); // already unknown. Nothing to do anymore. 1501 1502 ldr(rscratch1, mdo_addr); 1503 cbz(rscratch1, none); 1504 cmp(rscratch1, TypeEntries::null_seen); 1505 br(Assembler::EQ, none); 1506 // There is a chance that the checks above (re-reading profiling 1507 // data from memory) fail if another thread has just set the 1508 // profiling to this obj's klass 1509 ldr(rscratch1, mdo_addr); 1510 eor(obj, obj, rscratch1); 1511 tst(obj, TypeEntries::type_klass_mask); 1512 br(Assembler::EQ, next); 1513 1514 // different than before. Cannot keep accurate profile. 1515 orptr(mdo_addr, TypeEntries::type_unknown); 1516 b(next); 1517 1518 bind(none); 1519 // first time here. Set profile type. 1520 str(obj, mdo_addr); 1521 1522 bind(next); 1523 } 1524 1525 void InterpreterMacroAssembler::profile_arguments_type(Register mdp, Register callee, Register tmp, bool is_virtual) { 1526 if (!ProfileInterpreter) { 1527 return; 1528 } 1529 1530 if (MethodData::profile_arguments() || MethodData::profile_return()) { 1531 Label profile_continue; 1532 1533 test_method_data_pointer(mdp, profile_continue); 1534 1535 int off_to_start = is_virtual ? in_bytes(VirtualCallData::virtual_call_data_size()) : in_bytes(CounterData::counter_data_size()); 1536 1537 ldrb(rscratch1, Address(mdp, in_bytes(DataLayout::tag_offset()) - off_to_start)); 1538 cmp(rscratch1, is_virtual ? DataLayout::virtual_call_type_data_tag : DataLayout::call_type_data_tag); 1539 br(Assembler::NE, profile_continue); 1540 1541 if (MethodData::profile_arguments()) { 1542 Label done; 1543 int off_to_args = in_bytes(TypeEntriesAtCall::args_data_offset()); 1544 add(mdp, mdp, off_to_args); 1545 1546 for (int i = 0; i < TypeProfileArgsLimit; i++) { 1547 if (i > 0 || MethodData::profile_return()) { 1548 // If return value type is profiled we may have no argument to profile 1549 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset())-off_to_args)); 1550 sub(tmp, tmp, i*TypeStackSlotEntries::per_arg_count()); 1551 cmp(tmp, TypeStackSlotEntries::per_arg_count()); 1552 br(Assembler::LT, done); 1553 } 1554 ldr(tmp, Address(callee, Method::const_offset())); 1555 load_unsigned_short(tmp, Address(tmp, ConstMethod::size_of_parameters_offset())); 1556 // stack offset o (zero based) from the start of the argument 1557 // list, for n arguments translates into offset n - o - 1 from 1558 // the end of the argument list 1559 ldr(rscratch1, Address(mdp, in_bytes(TypeEntriesAtCall::stack_slot_offset(i))-off_to_args)); 1560 sub(tmp, tmp, rscratch1); 1561 sub(tmp, tmp, 1); 1562 Address arg_addr = argument_address(tmp); 1563 ldr(tmp, arg_addr); 1564 1565 Address mdo_arg_addr(mdp, in_bytes(TypeEntriesAtCall::argument_type_offset(i))-off_to_args); 1566 profile_obj_type(tmp, mdo_arg_addr); 1567 1568 int to_add = in_bytes(TypeStackSlotEntries::per_arg_size()); 1569 add(mdp, mdp, to_add); 1570 off_to_args += to_add; 1571 } 1572 1573 if (MethodData::profile_return()) { 1574 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset())-off_to_args)); 1575 sub(tmp, tmp, TypeProfileArgsLimit*TypeStackSlotEntries::per_arg_count()); 1576 } 1577 1578 bind(done); 1579 1580 if (MethodData::profile_return()) { 1581 // We're right after the type profile for the last 1582 // argument. tmp is the number of cells left in the 1583 // CallTypeData/VirtualCallTypeData to reach its end. Non null 1584 // if there's a return to profile. 1585 assert(ReturnTypeEntry::static_cell_count() < TypeStackSlotEntries::per_arg_count(), "can't move past ret type"); 1586 add(mdp, mdp, tmp, LSL, exact_log2(DataLayout::cell_size)); 1587 } 1588 str(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 1589 } else { 1590 assert(MethodData::profile_return(), "either profile call args or call ret"); 1591 update_mdp_by_constant(mdp, in_bytes(TypeEntriesAtCall::return_only_size())); 1592 } 1593 1594 // mdp points right after the end of the 1595 // CallTypeData/VirtualCallTypeData, right after the cells for the 1596 // return value type if there's one 1597 1598 bind(profile_continue); 1599 } 1600 } 1601 1602 void InterpreterMacroAssembler::profile_return_type(Register mdp, Register ret, Register tmp) { 1603 assert_different_registers(mdp, ret, tmp, rbcp); 1604 if (ProfileInterpreter && MethodData::profile_return()) { 1605 Label profile_continue, done; 1606 1607 test_method_data_pointer(mdp, profile_continue); 1608 1609 if (MethodData::profile_return_jsr292_only()) { 1610 // If we don't profile all invoke bytecodes we must make sure 1611 // it's a bytecode we indeed profile. We can't go back to the 1612 // begining of the ProfileData we intend to update to check its 1613 // type because we're right after it and we don't known its 1614 // length 1615 Label do_profile; 1616 ldrb(rscratch1, Address(rbcp, 0)); 1617 cmp(rscratch1, Bytecodes::_invokedynamic); 1618 br(Assembler::EQ, do_profile); 1619 cmp(rscratch1, Bytecodes::_invokehandle); 1620 br(Assembler::EQ, do_profile); 1621 get_method(tmp); 1622 ldrb(rscratch1, Address(tmp, Method::intrinsic_id_offset_in_bytes())); 1623 cmp(rscratch1, vmIntrinsics::_compiledLambdaForm); 1624 br(Assembler::NE, profile_continue); 1625 1626 bind(do_profile); 1627 } 1628 1629 Address mdo_ret_addr(mdp, -in_bytes(ReturnTypeEntry::size())); 1630 mov(tmp, ret); 1631 profile_obj_type(tmp, mdo_ret_addr); 1632 1633 bind(profile_continue); 1634 } 1635 } 1636 1637 void InterpreterMacroAssembler::profile_parameters_type(Register mdp, Register tmp1, Register tmp2) { 1638 if (ProfileInterpreter && MethodData::profile_parameters()) { 1639 Label profile_continue, done; 1640 1641 test_method_data_pointer(mdp, profile_continue); 1642 1643 // Load the offset of the area within the MDO used for 1644 // parameters. If it's negative we're not profiling any parameters 1645 ldr(tmp1, Address(mdp, in_bytes(MethodData::parameters_type_data_di_offset()) - in_bytes(MethodData::data_offset()))); 1646 cmp(tmp1, 0u); 1647 br(Assembler::LT, profile_continue); 1648 1649 // Compute a pointer to the area for parameters from the offset 1650 // and move the pointer to the slot for the last 1651 // parameters. Collect profiling from last parameter down. 1652 // mdo start + parameters offset + array length - 1 1653 add(mdp, mdp, tmp1); 1654 ldr(tmp1, Address(mdp, ArrayData::array_len_offset())); 1655 sub(tmp1, tmp1, TypeStackSlotEntries::per_arg_count()); 1656 1657 Label loop; 1658 bind(loop); 1659 1660 int off_base = in_bytes(ParametersTypeData::stack_slot_offset(0)); 1661 int type_base = in_bytes(ParametersTypeData::type_offset(0)); 1662 int per_arg_scale = exact_log2(DataLayout::cell_size); 1663 add(rscratch1, mdp, off_base); 1664 add(rscratch2, mdp, type_base); 1665 1666 Address arg_off(rscratch1, tmp1, Address::lsl(per_arg_scale)); 1667 Address arg_type(rscratch2, tmp1, Address::lsl(per_arg_scale)); 1668 1669 // load offset on the stack from the slot for this parameter 1670 ldr(tmp2, arg_off); 1671 neg(tmp2, tmp2); 1672 // read the parameter from the local area 1673 ldr(tmp2, Address(rlocals, tmp2, Address::lsl(Interpreter::logStackElementSize))); 1674 1675 // profile the parameter 1676 profile_obj_type(tmp2, arg_type); 1677 1678 // go to next parameter 1679 subs(tmp1, tmp1, TypeStackSlotEntries::per_arg_count()); 1680 br(Assembler::GE, loop); 1681 1682 bind(profile_continue); 1683 } 1684 }