1 /* 2 * Copyright (c) 2003, 2015, 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 "logging/log.hpp" 31 #include "oops/arrayOop.hpp" 32 #include "oops/markOop.hpp" 33 #include "oops/methodData.hpp" 34 #include "oops/method.hpp" 35 #include "prims/jvmtiExport.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 void InterpreterMacroAssembler::narrow(Register result) { 44 45 // Get method->_constMethod->_result_type 46 ldr(rscratch1, Address(rfp, frame::interpreter_frame_method_offset * wordSize)); 47 ldr(rscratch1, Address(rscratch1, Method::const_offset())); 48 ldrb(rscratch1, Address(rscratch1, ConstMethod::result_type_offset())); 49 50 Label done, notBool, notByte, notChar; 51 52 // common case first 53 cmpw(rscratch1, T_INT); 54 br(Assembler::EQ, done); 55 56 // mask integer result to narrower return type. 57 cmpw(rscratch1, T_BOOLEAN); 58 br(Assembler::NE, notBool); 59 andw(result, result, 0x1); 60 b(done); 61 62 bind(notBool); 63 cmpw(rscratch1, T_BYTE); 64 br(Assembler::NE, notByte); 65 sbfx(result, result, 0, 8); 66 b(done); 67 68 bind(notByte); 69 cmpw(rscratch1, T_CHAR); 70 br(Assembler::NE, notChar); 71 ubfx(result, result, 0, 16); // truncate upper 16 bits 72 b(done); 73 74 bind(notChar); 75 sbfx(result, result, 0, 16); // sign-extend short 76 77 // Nothing to do for T_INT 78 bind(done); 79 } 80 81 void InterpreterMacroAssembler::jump_to_entry(address entry) { 82 assert(entry, "Entry must have been generated by now"); 83 b(entry); 84 } 85 86 void InterpreterMacroAssembler::check_and_handle_popframe(Register java_thread) { 87 if (JvmtiExport::can_pop_frame()) { 88 Label L; 89 // Initiate popframe handling only if it is not already being 90 // processed. If the flag has the popframe_processing bit set, it 91 // means that this code is called *during* popframe handling - we 92 // don't want to reenter. 93 // This method is only called just after the call into the vm in 94 // call_VM_base, so the arg registers are available. 95 ldrw(rscratch1, Address(rthread, JavaThread::popframe_condition_offset())); 96 tbz(rscratch1, exact_log2(JavaThread::popframe_pending_bit), L); 97 tbnz(rscratch1, exact_log2(JavaThread::popframe_processing_bit), L); 98 // Call Interpreter::remove_activation_preserving_args_entry() to get the 99 // address of the same-named entrypoint in the generated interpreter code. 100 call_VM_leaf(CAST_FROM_FN_PTR(address, Interpreter::remove_activation_preserving_args_entry)); 101 br(r0); 102 bind(L); 103 } 104 } 105 106 107 void InterpreterMacroAssembler::load_earlyret_value(TosState state) { 108 ldr(r2, Address(rthread, JavaThread::jvmti_thread_state_offset())); 109 const Address tos_addr(r2, JvmtiThreadState::earlyret_tos_offset()); 110 const Address oop_addr(r2, JvmtiThreadState::earlyret_oop_offset()); 111 const Address val_addr(r2, JvmtiThreadState::earlyret_value_offset()); 112 switch (state) { 113 case atos: ldr(r0, oop_addr); 114 str(zr, oop_addr); 115 verify_oop(r0, state); break; 116 case ltos: ldr(r0, val_addr); break; 117 case btos: // fall through 118 case ztos: // fall through 119 case ctos: // fall through 120 case stos: // fall through 121 case itos: ldrw(r0, val_addr); break; 122 case ftos: ldrs(v0, val_addr); break; 123 case dtos: ldrd(v0, val_addr); break; 124 case vtos: /* nothing to do */ break; 125 default : ShouldNotReachHere(); 126 } 127 // Clean up tos value in the thread object 128 movw(rscratch1, (int) ilgl); 129 strw(rscratch1, tos_addr); 130 strw(zr, val_addr); 131 } 132 133 134 void InterpreterMacroAssembler::check_and_handle_earlyret(Register java_thread) { 135 if (JvmtiExport::can_force_early_return()) { 136 Label L; 137 ldr(rscratch1, Address(rthread, JavaThread::jvmti_thread_state_offset())); 138 cbz(rscratch1, L); // if (thread->jvmti_thread_state() == NULL) exit; 139 140 // Initiate earlyret handling only if it is not already being processed. 141 // If the flag has the earlyret_processing bit set, it means that this code 142 // is called *during* earlyret handling - we don't want to reenter. 143 ldrw(rscratch1, Address(rscratch1, JvmtiThreadState::earlyret_state_offset())); 144 cmpw(rscratch1, JvmtiThreadState::earlyret_pending); 145 br(Assembler::NE, L); 146 147 // Call Interpreter::remove_activation_early_entry() to get the address of the 148 // same-named entrypoint in the generated interpreter code. 149 ldr(rscratch1, Address(rthread, JavaThread::jvmti_thread_state_offset())); 150 ldrw(rscratch1, Address(rscratch1, JvmtiThreadState::earlyret_tos_offset())); 151 call_VM_leaf(CAST_FROM_FN_PTR(address, Interpreter::remove_activation_early_entry), rscratch1); 152 br(r0); 153 bind(L); 154 } 155 } 156 157 void InterpreterMacroAssembler::get_unsigned_2_byte_index_at_bcp( 158 Register reg, 159 int bcp_offset) { 160 assert(bcp_offset >= 0, "bcp is still pointing to start of bytecode"); 161 ldrh(reg, Address(rbcp, bcp_offset)); 162 rev16(reg, reg); 163 } 164 165 void InterpreterMacroAssembler::get_dispatch() { 166 unsigned long offset; 167 adrp(rdispatch, ExternalAddress((address)Interpreter::dispatch_table()), offset); 168 lea(rdispatch, Address(rdispatch, offset)); 169 } 170 171 void InterpreterMacroAssembler::get_cache_index_at_bcp(Register index, 172 int bcp_offset, 173 size_t index_size) { 174 assert(bcp_offset > 0, "bcp is still pointing to start of bytecode"); 175 if (index_size == sizeof(u2)) { 176 load_unsigned_short(index, Address(rbcp, bcp_offset)); 177 } else if (index_size == sizeof(u4)) { 178 // assert(EnableInvokeDynamic, "giant index used only for JSR 292"); 179 ldrw(index, Address(rbcp, bcp_offset)); 180 // Check if the secondary index definition is still ~x, otherwise 181 // we have to change the following assembler code to calculate the 182 // plain index. 183 assert(ConstantPool::decode_invokedynamic_index(~123) == 123, "else change next line"); 184 eonw(index, index, zr); // convert to plain index 185 } else if (index_size == sizeof(u1)) { 186 load_unsigned_byte(index, Address(rbcp, bcp_offset)); 187 } else { 188 ShouldNotReachHere(); 189 } 190 } 191 192 // Return 193 // Rindex: index into constant pool 194 // Rcache: address of cache entry - ConstantPoolCache::base_offset() 195 // 196 // A caller must add ConstantPoolCache::base_offset() to Rcache to get 197 // the true address of the cache entry. 198 // 199 void InterpreterMacroAssembler::get_cache_and_index_at_bcp(Register cache, 200 Register index, 201 int bcp_offset, 202 size_t index_size) { 203 assert_different_registers(cache, index); 204 assert_different_registers(cache, rcpool); 205 get_cache_index_at_bcp(index, bcp_offset, index_size); 206 assert(sizeof(ConstantPoolCacheEntry) == 4 * wordSize, "adjust code below"); 207 // convert from field index to ConstantPoolCacheEntry 208 // aarch64 already has the cache in rcpool so there is no need to 209 // install it in cache. instead we pre-add the indexed offset to 210 // rcpool and return it in cache. All clients of this method need to 211 // be modified accordingly. 212 add(cache, rcpool, index, Assembler::LSL, 5); 213 } 214 215 216 void InterpreterMacroAssembler::get_cache_and_index_and_bytecode_at_bcp(Register cache, 217 Register index, 218 Register bytecode, 219 int byte_no, 220 int bcp_offset, 221 size_t index_size) { 222 get_cache_and_index_at_bcp(cache, index, bcp_offset, index_size); 223 // We use a 32-bit load here since the layout of 64-bit words on 224 // little-endian machines allow us that. 225 // n.b. unlike x86 cache already includes the index offset 226 lea(bytecode, Address(cache, 227 ConstantPoolCache::base_offset() 228 + ConstantPoolCacheEntry::indices_offset())); 229 ldarw(bytecode, bytecode); 230 const int shift_count = (1 + byte_no) * BitsPerByte; 231 ubfx(bytecode, bytecode, shift_count, BitsPerByte); 232 } 233 234 void InterpreterMacroAssembler::get_cache_entry_pointer_at_bcp(Register cache, 235 Register tmp, 236 int bcp_offset, 237 size_t index_size) { 238 assert(cache != tmp, "must use different register"); 239 get_cache_index_at_bcp(tmp, bcp_offset, index_size); 240 assert(sizeof(ConstantPoolCacheEntry) == 4 * wordSize, "adjust code below"); 241 // convert from field index to ConstantPoolCacheEntry index 242 // and from word offset to byte offset 243 assert(exact_log2(in_bytes(ConstantPoolCacheEntry::size_in_bytes())) == 2 + LogBytesPerWord, "else change next line"); 244 ldr(cache, Address(rfp, frame::interpreter_frame_cache_offset * wordSize)); 245 // skip past the header 246 add(cache, cache, in_bytes(ConstantPoolCache::base_offset())); 247 add(cache, cache, tmp, Assembler::LSL, 2 + LogBytesPerWord); // construct pointer to cache entry 248 } 249 250 void InterpreterMacroAssembler::get_method_counters(Register method, 251 Register mcs, Label& skip) { 252 Label has_counters; 253 ldr(mcs, Address(method, Method::method_counters_offset())); 254 cbnz(mcs, has_counters); 255 call_VM(noreg, CAST_FROM_FN_PTR(address, 256 InterpreterRuntime::build_method_counters), method); 257 ldr(mcs, Address(method, Method::method_counters_offset())); 258 cbz(mcs, skip); // No MethodCounters allocated, OutOfMemory 259 bind(has_counters); 260 } 261 262 // Load object from cpool->resolved_references(index) 263 void InterpreterMacroAssembler::load_resolved_reference_at_index( 264 Register result, Register index) { 265 assert_different_registers(result, index); 266 // convert from field index to resolved_references() index and from 267 // word index to byte offset. Since this is a java object, it can be compressed 268 Register tmp = index; // reuse 269 lslw(tmp, tmp, LogBytesPerHeapOop); 270 271 get_constant_pool(result); 272 // load pointer for resolved_references[] objArray 273 ldr(result, Address(result, ConstantPool::resolved_references_offset_in_bytes())); 274 // JNIHandles::resolve(obj); 275 ldr(result, Address(result, 0)); 276 // Add in the index 277 add(result, result, tmp); 278 load_heap_oop(result, Address(result, arrayOopDesc::base_offset_in_bytes(T_OBJECT))); 279 } 280 281 // Generate a subtype check: branch to ok_is_subtype if sub_klass is a 282 // subtype of super_klass. 283 // 284 // Args: 285 // r0: superklass 286 // Rsub_klass: subklass 287 // 288 // Kills: 289 // r2, r5 290 void InterpreterMacroAssembler::gen_subtype_check(Register Rsub_klass, 291 Label& ok_is_subtype) { 292 assert(Rsub_klass != r0, "r0 holds superklass"); 293 assert(Rsub_klass != r2, "r2 holds 2ndary super array length"); 294 assert(Rsub_klass != r5, "r5 holds 2ndary super array scan ptr"); 295 296 // Profile the not-null value's klass. 297 profile_typecheck(r2, Rsub_klass, r5); // blows r2, reloads r5 298 299 // Do the check. 300 check_klass_subtype(Rsub_klass, r0, r2, ok_is_subtype); // blows r2 301 302 // Profile the failure of the check. 303 profile_typecheck_failed(r2); // blows r2 304 } 305 306 // Java Expression Stack 307 308 void InterpreterMacroAssembler::pop_ptr(Register r) { 309 ldr(r, post(esp, wordSize)); 310 } 311 312 void InterpreterMacroAssembler::pop_i(Register r) { 313 ldrw(r, post(esp, wordSize)); 314 } 315 316 void InterpreterMacroAssembler::pop_l(Register r) { 317 ldr(r, post(esp, 2 * Interpreter::stackElementSize)); 318 } 319 320 void InterpreterMacroAssembler::push_ptr(Register r) { 321 str(r, pre(esp, -wordSize)); 322 } 323 324 void InterpreterMacroAssembler::push_i(Register r) { 325 str(r, pre(esp, -wordSize)); 326 } 327 328 void InterpreterMacroAssembler::push_l(Register r) { 329 str(zr, pre(esp, -wordSize)); 330 str(r, pre(esp, - wordSize)); 331 } 332 333 void InterpreterMacroAssembler::pop_f(FloatRegister r) { 334 ldrs(r, post(esp, wordSize)); 335 } 336 337 void InterpreterMacroAssembler::pop_d(FloatRegister r) { 338 ldrd(r, post(esp, 2 * Interpreter::stackElementSize)); 339 } 340 341 void InterpreterMacroAssembler::push_f(FloatRegister r) { 342 strs(r, pre(esp, -wordSize)); 343 } 344 345 void InterpreterMacroAssembler::push_d(FloatRegister r) { 346 strd(r, pre(esp, 2* -wordSize)); 347 } 348 349 void InterpreterMacroAssembler::pop(TosState state) { 350 switch (state) { 351 case atos: pop_ptr(); break; 352 case btos: 353 case ztos: 354 case ctos: 355 case stos: 356 case itos: pop_i(); break; 357 case ltos: pop_l(); break; 358 case ftos: pop_f(); break; 359 case dtos: pop_d(); break; 360 case vtos: /* nothing to do */ break; 361 default: ShouldNotReachHere(); 362 } 363 verify_oop(r0, state); 364 } 365 366 void InterpreterMacroAssembler::push(TosState state) { 367 verify_oop(r0, state); 368 switch (state) { 369 case atos: push_ptr(); break; 370 case btos: 371 case ztos: 372 case ctos: 373 case stos: 374 case itos: push_i(); break; 375 case ltos: push_l(); break; 376 case ftos: push_f(); break; 377 case dtos: push_d(); break; 378 case vtos: /* nothing to do */ break; 379 default : ShouldNotReachHere(); 380 } 381 } 382 383 // Helpers for swap and dup 384 void InterpreterMacroAssembler::load_ptr(int n, Register val) { 385 ldr(val, Address(esp, Interpreter::expr_offset_in_bytes(n))); 386 } 387 388 void InterpreterMacroAssembler::store_ptr(int n, Register val) { 389 str(val, Address(esp, Interpreter::expr_offset_in_bytes(n))); 390 } 391 392 393 void InterpreterMacroAssembler::prepare_to_jump_from_interpreted() { 394 // set sender sp 395 mov(r13, sp); 396 // record last_sp 397 str(esp, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 398 } 399 400 // Jump to from_interpreted entry of a call unless single stepping is possible 401 // in this thread in which case we must call the i2i entry 402 void InterpreterMacroAssembler::jump_from_interpreted(Register method, Register temp) { 403 prepare_to_jump_from_interpreted(); 404 405 if (JvmtiExport::can_post_interpreter_events()) { 406 Label run_compiled_code; 407 // JVMTI events, such as single-stepping, are implemented partly by avoiding running 408 // compiled code in threads for which the event is enabled. Check here for 409 // interp_only_mode if these events CAN be enabled. 410 ldrw(rscratch1, Address(rthread, JavaThread::interp_only_mode_offset())); 411 cbzw(rscratch1, run_compiled_code); 412 ldr(rscratch1, Address(method, Method::interpreter_entry_offset())); 413 br(rscratch1); 414 bind(run_compiled_code); 415 } 416 417 ldr(rscratch1, Address(method, Method::from_interpreted_offset())); 418 br(rscratch1); 419 } 420 421 // The following two routines provide a hook so that an implementation 422 // can schedule the dispatch in two parts. amd64 does not do this. 423 void InterpreterMacroAssembler::dispatch_prolog(TosState state, int step) { 424 } 425 426 void InterpreterMacroAssembler::dispatch_epilog(TosState state, int step) { 427 dispatch_next(state, step); 428 } 429 430 void InterpreterMacroAssembler::dispatch_base(TosState state, 431 address* table, 432 bool verifyoop) { 433 if (VerifyActivationFrameSize) { 434 Unimplemented(); 435 } 436 if (verifyoop) { 437 verify_oop(r0, state); 438 } 439 if (table == Interpreter::dispatch_table(state)) { 440 addw(rscratch2, rscratch1, Interpreter::distance_from_dispatch_table(state)); 441 ldr(rscratch2, Address(rdispatch, rscratch2, Address::uxtw(3))); 442 } else { 443 mov(rscratch2, (address)table); 444 ldr(rscratch2, Address(rscratch2, rscratch1, Address::uxtw(3))); 445 } 446 br(rscratch2); 447 } 448 449 void InterpreterMacroAssembler::dispatch_only(TosState state) { 450 dispatch_base(state, Interpreter::dispatch_table(state)); 451 } 452 453 void InterpreterMacroAssembler::dispatch_only_normal(TosState state) { 454 dispatch_base(state, Interpreter::normal_table(state)); 455 } 456 457 void InterpreterMacroAssembler::dispatch_only_noverify(TosState state) { 458 dispatch_base(state, Interpreter::normal_table(state), false); 459 } 460 461 462 void InterpreterMacroAssembler::dispatch_next(TosState state, int step) { 463 // load next bytecode 464 ldrb(rscratch1, Address(pre(rbcp, step))); 465 dispatch_base(state, Interpreter::dispatch_table(state)); 466 } 467 468 void InterpreterMacroAssembler::dispatch_via(TosState state, address* table) { 469 // load current bytecode 470 ldrb(rscratch1, Address(rbcp, 0)); 471 dispatch_base(state, table); 472 } 473 474 // remove activation 475 // 476 // Unlock the receiver if this is a synchronized method. 477 // Unlock any Java monitors from syncronized blocks. 478 // Remove the activation from the stack. 479 // 480 // If there are locked Java monitors 481 // If throw_monitor_exception 482 // throws IllegalMonitorStateException 483 // Else if install_monitor_exception 484 // installs IllegalMonitorStateException 485 // Else 486 // no error processing 487 void InterpreterMacroAssembler::remove_activation( 488 TosState state, 489 bool throw_monitor_exception, 490 bool install_monitor_exception, 491 bool notify_jvmdi) { 492 // Note: Registers r3 xmm0 may be in use for the 493 // result check if synchronized method 494 Label unlocked, unlock, no_unlock; 495 496 // get the value of _do_not_unlock_if_synchronized into r3 497 const Address do_not_unlock_if_synchronized(rthread, 498 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset())); 499 ldrb(r3, do_not_unlock_if_synchronized); 500 strb(zr, do_not_unlock_if_synchronized); // reset the flag 501 502 // get method access flags 503 ldr(r1, Address(rfp, frame::interpreter_frame_method_offset * wordSize)); 504 ldr(r2, Address(r1, Method::access_flags_offset())); 505 tbz(r2, exact_log2(JVM_ACC_SYNCHRONIZED), unlocked); 506 507 // Don't unlock anything if the _do_not_unlock_if_synchronized flag 508 // is set. 509 cbnz(r3, no_unlock); 510 511 // unlock monitor 512 push(state); // save result 513 514 // BasicObjectLock will be first in list, since this is a 515 // synchronized method. However, need to check that the object has 516 // not been unlocked by an explicit monitorexit bytecode. 517 const Address monitor(rfp, frame::interpreter_frame_initial_sp_offset * 518 wordSize - (int) sizeof(BasicObjectLock)); 519 // We use c_rarg1 so that if we go slow path it will be the correct 520 // register for unlock_object to pass to VM directly 521 lea(c_rarg1, monitor); // address of first monitor 522 523 ldr(r0, Address(c_rarg1, BasicObjectLock::obj_offset_in_bytes())); 524 cbnz(r0, unlock); 525 526 pop(state); 527 if (throw_monitor_exception) { 528 // Entry already unlocked, need to throw exception 529 call_VM(noreg, CAST_FROM_FN_PTR(address, 530 InterpreterRuntime::throw_illegal_monitor_state_exception)); 531 should_not_reach_here(); 532 } else { 533 // Monitor already unlocked during a stack unroll. If requested, 534 // install an illegal_monitor_state_exception. Continue with 535 // stack unrolling. 536 if (install_monitor_exception) { 537 call_VM(noreg, CAST_FROM_FN_PTR(address, 538 InterpreterRuntime::new_illegal_monitor_state_exception)); 539 } 540 b(unlocked); 541 } 542 543 bind(unlock); 544 unlock_object(c_rarg1); 545 pop(state); 546 547 // Check that for block-structured locking (i.e., that all locked 548 // objects has been unlocked) 549 bind(unlocked); 550 551 // r0: Might contain return value 552 553 // Check that all monitors are unlocked 554 { 555 Label loop, exception, entry, restart; 556 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize; 557 const Address monitor_block_top( 558 rfp, frame::interpreter_frame_monitor_block_top_offset * wordSize); 559 const Address monitor_block_bot( 560 rfp, frame::interpreter_frame_initial_sp_offset * wordSize); 561 562 bind(restart); 563 // We use c_rarg1 so that if we go slow path it will be the correct 564 // register for unlock_object to pass to VM directly 565 ldr(c_rarg1, monitor_block_top); // points to current entry, starting 566 // with top-most entry 567 lea(r19, monitor_block_bot); // points to word before bottom of 568 // monitor block 569 b(entry); 570 571 // Entry already locked, need to throw exception 572 bind(exception); 573 574 if (throw_monitor_exception) { 575 // Throw exception 576 MacroAssembler::call_VM(noreg, 577 CAST_FROM_FN_PTR(address, InterpreterRuntime:: 578 throw_illegal_monitor_state_exception)); 579 should_not_reach_here(); 580 } else { 581 // Stack unrolling. Unlock object and install illegal_monitor_exception. 582 // Unlock does not block, so don't have to worry about the frame. 583 // We don't have to preserve c_rarg1 since we are going to throw an exception. 584 585 push(state); 586 unlock_object(c_rarg1); 587 pop(state); 588 589 if (install_monitor_exception) { 590 call_VM(noreg, CAST_FROM_FN_PTR(address, 591 InterpreterRuntime:: 592 new_illegal_monitor_state_exception)); 593 } 594 595 b(restart); 596 } 597 598 bind(loop); 599 // check if current entry is used 600 ldr(rscratch1, Address(c_rarg1, BasicObjectLock::obj_offset_in_bytes())); 601 cbnz(rscratch1, exception); 602 603 add(c_rarg1, c_rarg1, entry_size); // otherwise advance to next entry 604 bind(entry); 605 cmp(c_rarg1, r19); // check if bottom reached 606 br(Assembler::NE, loop); // if not at bottom then check this entry 607 } 608 609 bind(no_unlock); 610 611 // jvmti support 612 if (notify_jvmdi) { 613 notify_method_exit(state, NotifyJVMTI); // preserve TOSCA 614 } else { 615 notify_method_exit(state, SkipNotifyJVMTI); // preserve TOSCA 616 } 617 618 // remove activation 619 // get sender esp 620 ldr(esp, 621 Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); 622 if (StackReservedPages > 0) { 623 // testing if reserved zone needs to be re-enabled 624 Label no_reserved_zone_enabling; 625 626 ldr(rscratch1, Address(rthread, JavaThread::reserved_stack_activation_offset())); 627 cmp(esp, rscratch1); 628 br(Assembler::LS, no_reserved_zone_enabling); 629 630 call_VM_leaf( 631 CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), rthread); 632 call_VM(noreg, CAST_FROM_FN_PTR(address, 633 InterpreterRuntime::throw_delayed_StackOverflowError)); 634 should_not_reach_here(); 635 636 bind(no_reserved_zone_enabling); 637 } 638 // remove frame anchor 639 leave(); 640 // If we're returning to interpreted code we will shortly be 641 // adjusting SP to allow some space for ESP. If we're returning to 642 // compiled code the saved sender SP was saved in sender_sp, so this 643 // restores it. 644 andr(sp, esp, -16); 645 } 646 647 // Lock object 648 // 649 // Args: 650 // c_rarg1: BasicObjectLock to be used for locking 651 // 652 // Kills: 653 // r0 654 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, .. (param regs) 655 // rscratch1, rscratch2 (scratch regs) 656 void InterpreterMacroAssembler::lock_object(Register lock_reg) 657 { 658 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be c_rarg1"); 659 if (UseHeavyMonitors) { 660 call_VM(noreg, 661 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter), 662 lock_reg); 663 } else { 664 Label done; 665 666 const Register swap_reg = r0; 667 const Register tmp = c_rarg2; 668 const Register obj_reg = c_rarg3; // Will contain the oop 669 670 const int obj_offset = BasicObjectLock::obj_offset_in_bytes(); 671 const int lock_offset = BasicObjectLock::lock_offset_in_bytes (); 672 const int mark_offset = lock_offset + 673 BasicLock::displaced_header_offset_in_bytes(); 674 675 Label slow_case; 676 677 // Load object pointer into obj_reg %c_rarg3 678 ldr(obj_reg, Address(lock_reg, obj_offset)); 679 680 if (UseBiasedLocking) { 681 biased_locking_enter(lock_reg, obj_reg, swap_reg, tmp, false, done, &slow_case); 682 } 683 684 // Load (object->mark() | 1) into swap_reg 685 ldr(rscratch1, Address(obj_reg, 0)); 686 orr(swap_reg, rscratch1, 1); 687 688 // Save (object->mark() | 1) into BasicLock's displaced header 689 str(swap_reg, Address(lock_reg, mark_offset)); 690 691 assert(lock_offset == 0, 692 "displached header must be first word in BasicObjectLock"); 693 694 Label fail; 695 if (PrintBiasedLockingStatistics) { 696 Label fast; 697 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, fast, &fail); 698 bind(fast); 699 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()), 700 rscratch2, rscratch1, tmp); 701 b(done); 702 bind(fail); 703 } else { 704 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL); 705 } 706 707 // Test if the oopMark is an obvious stack pointer, i.e., 708 // 1) (mark & 7) == 0, and 709 // 2) rsp <= mark < mark + os::pagesize() 710 // 711 // These 3 tests can be done by evaluating the following 712 // expression: ((mark - rsp) & (7 - os::vm_page_size())), 713 // assuming both stack pointer and pagesize have their 714 // least significant 3 bits clear. 715 // NOTE: the oopMark is in swap_reg %r0 as the result of cmpxchg 716 // NOTE2: aarch64 does not like to subtract sp from rn so take a 717 // copy 718 mov(rscratch1, sp); 719 sub(swap_reg, swap_reg, rscratch1); 720 ands(swap_reg, swap_reg, (unsigned long)(7 - os::vm_page_size())); 721 722 // Save the test result, for recursive case, the result is zero 723 str(swap_reg, Address(lock_reg, mark_offset)); 724 725 if (PrintBiasedLockingStatistics) { 726 br(Assembler::NE, slow_case); 727 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()), 728 rscratch2, rscratch1, tmp); 729 } 730 br(Assembler::EQ, done); 731 732 bind(slow_case); 733 734 // Call the runtime routine for slow case 735 call_VM(noreg, 736 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter), 737 lock_reg); 738 739 bind(done); 740 } 741 } 742 743 744 // Unlocks an object. Used in monitorexit bytecode and 745 // remove_activation. Throws an IllegalMonitorException if object is 746 // not locked by current thread. 747 // 748 // Args: 749 // c_rarg1: BasicObjectLock for lock 750 // 751 // Kills: 752 // r0 753 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, ... (param regs) 754 // rscratch1, rscratch2 (scratch regs) 755 void InterpreterMacroAssembler::unlock_object(Register lock_reg) 756 { 757 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be rarg1"); 758 759 if (UseHeavyMonitors) { 760 call_VM(noreg, 761 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit), 762 lock_reg); 763 } else { 764 Label done; 765 766 const Register swap_reg = r0; 767 const Register header_reg = c_rarg2; // Will contain the old oopMark 768 const Register obj_reg = c_rarg3; // Will contain the oop 769 770 save_bcp(); // Save in case of exception 771 772 // Convert from BasicObjectLock structure to object and BasicLock 773 // structure Store the BasicLock address into %r0 774 lea(swap_reg, Address(lock_reg, BasicObjectLock::lock_offset_in_bytes())); 775 776 // Load oop into obj_reg(%c_rarg3) 777 ldr(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); 778 779 // Free entry 780 str(zr, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); 781 782 if (UseBiasedLocking) { 783 biased_locking_exit(obj_reg, header_reg, done); 784 } 785 786 // Load the old header from BasicLock structure 787 ldr(header_reg, Address(swap_reg, 788 BasicLock::displaced_header_offset_in_bytes())); 789 790 // Test for recursion 791 cbz(header_reg, done); 792 793 // Atomic swap back the old header 794 cmpxchgptr(swap_reg, header_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL); 795 796 // Call the runtime routine for slow case. 797 str(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); // restore obj 798 call_VM(noreg, 799 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit), 800 lock_reg); 801 802 bind(done); 803 804 restore_bcp(); 805 } 806 } 807 808 void InterpreterMacroAssembler::test_method_data_pointer(Register mdp, 809 Label& zero_continue) { 810 assert(ProfileInterpreter, "must be profiling interpreter"); 811 ldr(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 812 cbz(mdp, zero_continue); 813 } 814 815 // Set the method data pointer for the current bcp. 816 void InterpreterMacroAssembler::set_method_data_pointer_for_bcp() { 817 assert(ProfileInterpreter, "must be profiling interpreter"); 818 Label set_mdp; 819 stp(r0, r1, Address(pre(sp, -2 * wordSize))); 820 821 // Test MDO to avoid the call if it is NULL. 822 ldr(r0, Address(rmethod, in_bytes(Method::method_data_offset()))); 823 cbz(r0, set_mdp); 824 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::bcp_to_di), rmethod, rbcp); 825 // r0: mdi 826 // mdo is guaranteed to be non-zero here, we checked for it before the call. 827 ldr(r1, Address(rmethod, in_bytes(Method::method_data_offset()))); 828 lea(r1, Address(r1, in_bytes(MethodData::data_offset()))); 829 add(r0, r1, r0); 830 str(r0, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 831 bind(set_mdp); 832 ldp(r0, r1, Address(post(sp, 2 * wordSize))); 833 } 834 835 void InterpreterMacroAssembler::verify_method_data_pointer() { 836 assert(ProfileInterpreter, "must be profiling interpreter"); 837 #ifdef ASSERT 838 Label verify_continue; 839 stp(r0, r1, Address(pre(sp, -2 * wordSize))); 840 stp(r2, r3, Address(pre(sp, -2 * wordSize))); 841 test_method_data_pointer(r3, verify_continue); // If mdp is zero, continue 842 get_method(r1); 843 844 // If the mdp is valid, it will point to a DataLayout header which is 845 // consistent with the bcp. The converse is highly probable also. 846 ldrsh(r2, Address(r3, in_bytes(DataLayout::bci_offset()))); 847 ldr(rscratch1, Address(r1, Method::const_offset())); 848 add(r2, r2, rscratch1, Assembler::LSL); 849 lea(r2, Address(r2, ConstMethod::codes_offset())); 850 cmp(r2, rbcp); 851 br(Assembler::EQ, verify_continue); 852 // r1: method 853 // rbcp: bcp // rbcp == 22 854 // r3: mdp 855 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::verify_mdp), 856 r1, rbcp, r3); 857 bind(verify_continue); 858 ldp(r2, r3, Address(post(sp, 2 * wordSize))); 859 ldp(r0, r1, Address(post(sp, 2 * wordSize))); 860 #endif // ASSERT 861 } 862 863 864 void InterpreterMacroAssembler::set_mdp_data_at(Register mdp_in, 865 int constant, 866 Register value) { 867 assert(ProfileInterpreter, "must be profiling interpreter"); 868 Address data(mdp_in, constant); 869 str(value, data); 870 } 871 872 873 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in, 874 int constant, 875 bool decrement) { 876 increment_mdp_data_at(mdp_in, noreg, constant, decrement); 877 } 878 879 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in, 880 Register reg, 881 int constant, 882 bool decrement) { 883 assert(ProfileInterpreter, "must be profiling interpreter"); 884 // %%% this does 64bit counters at best it is wasting space 885 // at worst it is a rare bug when counters overflow 886 887 assert_different_registers(rscratch2, rscratch1, mdp_in, reg); 888 889 Address addr1(mdp_in, constant); 890 Address addr2(rscratch2, reg, Address::lsl(0)); 891 Address &addr = addr1; 892 if (reg != noreg) { 893 lea(rscratch2, addr1); 894 addr = addr2; 895 } 896 897 if (decrement) { 898 // Decrement the register. Set condition codes. 899 // Intel does this 900 // addptr(data, (int32_t) -DataLayout::counter_increment); 901 // If the decrement causes the counter to overflow, stay negative 902 // Label L; 903 // jcc(Assembler::negative, L); 904 // addptr(data, (int32_t) DataLayout::counter_increment); 905 // so we do this 906 ldr(rscratch1, addr); 907 subs(rscratch1, rscratch1, (unsigned)DataLayout::counter_increment); 908 Label L; 909 br(Assembler::LO, L); // skip store if counter underflow 910 str(rscratch1, addr); 911 bind(L); 912 } else { 913 assert(DataLayout::counter_increment == 1, 914 "flow-free idiom only works with 1"); 915 // Intel does this 916 // Increment the register. Set carry flag. 917 // addptr(data, DataLayout::counter_increment); 918 // If the increment causes the counter to overflow, pull back by 1. 919 // sbbptr(data, (int32_t)0); 920 // so we do this 921 ldr(rscratch1, addr); 922 adds(rscratch1, rscratch1, DataLayout::counter_increment); 923 Label L; 924 br(Assembler::CS, L); // skip store if counter overflow 925 str(rscratch1, addr); 926 bind(L); 927 } 928 } 929 930 void InterpreterMacroAssembler::set_mdp_flag_at(Register mdp_in, 931 int flag_byte_constant) { 932 assert(ProfileInterpreter, "must be profiling interpreter"); 933 int header_offset = in_bytes(DataLayout::header_offset()); 934 int header_bits = DataLayout::flag_mask_to_header_mask(flag_byte_constant); 935 // Set the flag 936 ldr(rscratch1, Address(mdp_in, header_offset)); 937 orr(rscratch1, rscratch1, header_bits); 938 str(rscratch1, Address(mdp_in, header_offset)); 939 } 940 941 942 void InterpreterMacroAssembler::test_mdp_data_at(Register mdp_in, 943 int offset, 944 Register value, 945 Register test_value_out, 946 Label& not_equal_continue) { 947 assert(ProfileInterpreter, "must be profiling interpreter"); 948 if (test_value_out == noreg) { 949 ldr(rscratch1, Address(mdp_in, offset)); 950 cmp(value, rscratch1); 951 } else { 952 // Put the test value into a register, so caller can use it: 953 ldr(test_value_out, Address(mdp_in, offset)); 954 cmp(value, test_value_out); 955 } 956 br(Assembler::NE, not_equal_continue); 957 } 958 959 960 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in, 961 int offset_of_disp) { 962 assert(ProfileInterpreter, "must be profiling interpreter"); 963 ldr(rscratch1, Address(mdp_in, offset_of_disp)); 964 add(mdp_in, mdp_in, rscratch1, LSL); 965 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 966 } 967 968 969 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in, 970 Register reg, 971 int offset_of_disp) { 972 assert(ProfileInterpreter, "must be profiling interpreter"); 973 lea(rscratch1, Address(mdp_in, offset_of_disp)); 974 ldr(rscratch1, Address(rscratch1, reg, Address::lsl(0))); 975 add(mdp_in, mdp_in, rscratch1, LSL); 976 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 977 } 978 979 980 void InterpreterMacroAssembler::update_mdp_by_constant(Register mdp_in, 981 int constant) { 982 assert(ProfileInterpreter, "must be profiling interpreter"); 983 add(mdp_in, mdp_in, (unsigned)constant); 984 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 985 } 986 987 988 void InterpreterMacroAssembler::update_mdp_for_ret(Register return_bci) { 989 assert(ProfileInterpreter, "must be profiling interpreter"); 990 // save/restore across call_VM 991 stp(zr, return_bci, Address(pre(sp, -2 * wordSize))); 992 call_VM(noreg, 993 CAST_FROM_FN_PTR(address, InterpreterRuntime::update_mdp_for_ret), 994 return_bci); 995 ldp(zr, return_bci, Address(post(sp, 2 * wordSize))); 996 } 997 998 999 void InterpreterMacroAssembler::profile_taken_branch(Register mdp, 1000 Register bumped_count) { 1001 if (ProfileInterpreter) { 1002 Label profile_continue; 1003 1004 // If no method data exists, go to profile_continue. 1005 // Otherwise, assign to mdp 1006 test_method_data_pointer(mdp, profile_continue); 1007 1008 // We are taking a branch. Increment the taken count. 1009 // We inline increment_mdp_data_at to return bumped_count in a register 1010 //increment_mdp_data_at(mdp, in_bytes(JumpData::taken_offset())); 1011 Address data(mdp, in_bytes(JumpData::taken_offset())); 1012 ldr(bumped_count, data); 1013 assert(DataLayout::counter_increment == 1, 1014 "flow-free idiom only works with 1"); 1015 // Intel does this to catch overflow 1016 // addptr(bumped_count, DataLayout::counter_increment); 1017 // sbbptr(bumped_count, 0); 1018 // so we do this 1019 adds(bumped_count, bumped_count, DataLayout::counter_increment); 1020 Label L; 1021 br(Assembler::CS, L); // skip store if counter overflow 1022 str(bumped_count, data); 1023 bind(L); 1024 // The method data pointer needs to be updated to reflect the new target. 1025 update_mdp_by_offset(mdp, in_bytes(JumpData::displacement_offset())); 1026 bind(profile_continue); 1027 } 1028 } 1029 1030 1031 void InterpreterMacroAssembler::profile_not_taken_branch(Register mdp) { 1032 if (ProfileInterpreter) { 1033 Label profile_continue; 1034 1035 // If no method data exists, go to profile_continue. 1036 test_method_data_pointer(mdp, profile_continue); 1037 1038 // We are taking a branch. Increment the not taken count. 1039 increment_mdp_data_at(mdp, in_bytes(BranchData::not_taken_offset())); 1040 1041 // The method data pointer needs to be updated to correspond to 1042 // the next bytecode 1043 update_mdp_by_constant(mdp, in_bytes(BranchData::branch_data_size())); 1044 bind(profile_continue); 1045 } 1046 } 1047 1048 1049 void InterpreterMacroAssembler::profile_call(Register mdp) { 1050 if (ProfileInterpreter) { 1051 Label profile_continue; 1052 1053 // If no method data exists, go to profile_continue. 1054 test_method_data_pointer(mdp, profile_continue); 1055 1056 // We are making a call. Increment the count. 1057 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1058 1059 // The method data pointer needs to be updated to reflect the new target. 1060 update_mdp_by_constant(mdp, in_bytes(CounterData::counter_data_size())); 1061 bind(profile_continue); 1062 } 1063 } 1064 1065 void InterpreterMacroAssembler::profile_final_call(Register mdp) { 1066 if (ProfileInterpreter) { 1067 Label profile_continue; 1068 1069 // If no method data exists, go to profile_continue. 1070 test_method_data_pointer(mdp, profile_continue); 1071 1072 // We are making a call. Increment the count. 1073 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1074 1075 // The method data pointer needs to be updated to reflect the new target. 1076 update_mdp_by_constant(mdp, 1077 in_bytes(VirtualCallData:: 1078 virtual_call_data_size())); 1079 bind(profile_continue); 1080 } 1081 } 1082 1083 1084 void InterpreterMacroAssembler::profile_virtual_call(Register receiver, 1085 Register mdp, 1086 Register reg2, 1087 bool receiver_can_be_null) { 1088 if (ProfileInterpreter) { 1089 Label profile_continue; 1090 1091 // If no method data exists, go to profile_continue. 1092 test_method_data_pointer(mdp, profile_continue); 1093 1094 Label skip_receiver_profile; 1095 if (receiver_can_be_null) { 1096 Label not_null; 1097 // We are making a call. Increment the count for null receiver. 1098 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1099 b(skip_receiver_profile); 1100 bind(not_null); 1101 } 1102 1103 // Record the receiver type. 1104 record_klass_in_profile(receiver, mdp, reg2, true); 1105 bind(skip_receiver_profile); 1106 1107 // The method data pointer needs to be updated to reflect the new target. 1108 #if INCLUDE_JVMCI 1109 if (MethodProfileWidth == 0) { 1110 update_mdp_by_constant(mdp, in_bytes(VirtualCallData::virtual_call_data_size())); 1111 } 1112 #else // INCLUDE_JVMCI 1113 update_mdp_by_constant(mdp, 1114 in_bytes(VirtualCallData:: 1115 virtual_call_data_size())); 1116 #endif // INCLUDE_JVMCI 1117 bind(profile_continue); 1118 } 1119 } 1120 1121 #if INCLUDE_JVMCI 1122 void InterpreterMacroAssembler::profile_called_method(Register method, Register mdp, Register reg2) { 1123 assert_different_registers(method, mdp, reg2); 1124 if (ProfileInterpreter && MethodProfileWidth > 0) { 1125 Label profile_continue; 1126 1127 // If no method data exists, go to profile_continue. 1128 test_method_data_pointer(mdp, profile_continue); 1129 1130 Label done; 1131 record_item_in_profile_helper(method, mdp, reg2, 0, done, MethodProfileWidth, 1132 &VirtualCallData::method_offset, &VirtualCallData::method_count_offset, in_bytes(VirtualCallData::nonprofiled_receiver_count_offset())); 1133 bind(done); 1134 1135 update_mdp_by_constant(mdp, in_bytes(VirtualCallData::virtual_call_data_size())); 1136 bind(profile_continue); 1137 } 1138 } 1139 #endif // INCLUDE_JVMCI 1140 1141 // This routine creates a state machine for updating the multi-row 1142 // type profile at a virtual call site (or other type-sensitive bytecode). 1143 // The machine visits each row (of receiver/count) until the receiver type 1144 // is found, or until it runs out of rows. At the same time, it remembers 1145 // the location of the first empty row. (An empty row records null for its 1146 // receiver, and can be allocated for a newly-observed receiver type.) 1147 // Because there are two degrees of freedom in the state, a simple linear 1148 // search will not work; it must be a decision tree. Hence this helper 1149 // function is recursive, to generate the required tree structured code. 1150 // It's the interpreter, so we are trading off code space for speed. 1151 // See below for example code. 1152 void InterpreterMacroAssembler::record_klass_in_profile_helper( 1153 Register receiver, Register mdp, 1154 Register reg2, int start_row, 1155 Label& done, bool is_virtual_call) { 1156 if (TypeProfileWidth == 0) { 1157 if (is_virtual_call) { 1158 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1159 } 1160 #if INCLUDE_JVMCI 1161 else if (EnableJVMCI) { 1162 increment_mdp_data_at(mdp, in_bytes(ReceiverTypeData::nonprofiled_receiver_count_offset())); 1163 } 1164 #endif // INCLUDE_JVMCI 1165 } else { 1166 int non_profiled_offset = -1; 1167 if (is_virtual_call) { 1168 non_profiled_offset = in_bytes(CounterData::count_offset()); 1169 } 1170 #if INCLUDE_JVMCI 1171 else if (EnableJVMCI) { 1172 non_profiled_offset = in_bytes(ReceiverTypeData::nonprofiled_receiver_count_offset()); 1173 } 1174 #endif // INCLUDE_JVMCI 1175 1176 record_item_in_profile_helper(receiver, mdp, reg2, 0, done, TypeProfileWidth, 1177 &VirtualCallData::receiver_offset, &VirtualCallData::receiver_count_offset, non_profiled_offset); 1178 } 1179 } 1180 1181 void InterpreterMacroAssembler::record_item_in_profile_helper(Register item, Register mdp, 1182 Register reg2, int start_row, Label& done, int total_rows, 1183 OffsetFunction item_offset_fn, OffsetFunction item_count_offset_fn, 1184 int non_profiled_offset) { 1185 int last_row = total_rows - 1; 1186 assert(start_row <= last_row, "must be work left to do"); 1187 // Test this row for both the item and for null. 1188 // Take any of three different outcomes: 1189 // 1. found item => increment count and goto done 1190 // 2. found null => keep looking for case 1, maybe allocate this cell 1191 // 3. found something else => keep looking for cases 1 and 2 1192 // Case 3 is handled by a recursive call. 1193 for (int row = start_row; row <= last_row; row++) { 1194 Label next_test; 1195 bool test_for_null_also = (row == start_row); 1196 1197 // See if the item is item[n]. 1198 int item_offset = in_bytes(item_offset_fn(row)); 1199 test_mdp_data_at(mdp, item_offset, item, 1200 (test_for_null_also ? reg2 : noreg), 1201 next_test); 1202 // (Reg2 now contains the item from the CallData.) 1203 1204 // The item is item[n]. Increment count[n]. 1205 int count_offset = in_bytes(item_count_offset_fn(row)); 1206 increment_mdp_data_at(mdp, count_offset); 1207 b(done); 1208 bind(next_test); 1209 1210 if (test_for_null_also) { 1211 Label found_null; 1212 // Failed the equality check on item[n]... Test for null. 1213 if (start_row == last_row) { 1214 // The only thing left to do is handle the null case. 1215 if (non_profiled_offset >= 0) { 1216 cbz(reg2, found_null); 1217 // Item did not match any saved item and there is no empty row for it. 1218 // Increment total counter to indicate polymorphic case. 1219 increment_mdp_data_at(mdp, non_profiled_offset); 1220 b(done); 1221 bind(found_null); 1222 } else { 1223 cbnz(reg2, done); 1224 } 1225 break; 1226 } 1227 // Since null is rare, make it be the branch-taken case. 1228 cbz(reg2, found_null); 1229 1230 // Put all the "Case 3" tests here. 1231 record_item_in_profile_helper(item, mdp, reg2, start_row + 1, done, total_rows, 1232 item_offset_fn, item_count_offset_fn, non_profiled_offset); 1233 1234 // Found a null. Keep searching for a matching item, 1235 // but remember that this is an empty (unused) slot. 1236 bind(found_null); 1237 } 1238 } 1239 1240 // In the fall-through case, we found no matching item, but we 1241 // observed the item[start_row] is NULL. 1242 1243 // Fill in the item field and increment the count. 1244 int item_offset = in_bytes(item_offset_fn(start_row)); 1245 set_mdp_data_at(mdp, item_offset, item); 1246 int count_offset = in_bytes(item_count_offset_fn(start_row)); 1247 mov(reg2, DataLayout::counter_increment); 1248 set_mdp_data_at(mdp, count_offset, reg2); 1249 if (start_row > 0) { 1250 b(done); 1251 } 1252 } 1253 1254 // Example state machine code for three profile rows: 1255 // // main copy of decision tree, rooted at row[1] 1256 // if (row[0].rec == rec) { row[0].incr(); goto done; } 1257 // if (row[0].rec != NULL) { 1258 // // inner copy of decision tree, rooted at row[1] 1259 // if (row[1].rec == rec) { row[1].incr(); goto done; } 1260 // if (row[1].rec != NULL) { 1261 // // degenerate decision tree, rooted at row[2] 1262 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1263 // if (row[2].rec != NULL) { count.incr(); goto done; } // overflow 1264 // row[2].init(rec); goto done; 1265 // } else { 1266 // // remember row[1] is empty 1267 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1268 // row[1].init(rec); goto done; 1269 // } 1270 // } else { 1271 // // remember row[0] is empty 1272 // if (row[1].rec == rec) { row[1].incr(); goto done; } 1273 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1274 // row[0].init(rec); goto done; 1275 // } 1276 // done: 1277 1278 void InterpreterMacroAssembler::record_klass_in_profile(Register receiver, 1279 Register mdp, Register reg2, 1280 bool is_virtual_call) { 1281 assert(ProfileInterpreter, "must be profiling"); 1282 Label done; 1283 1284 record_klass_in_profile_helper(receiver, mdp, reg2, 0, done, is_virtual_call); 1285 1286 bind (done); 1287 } 1288 1289 void InterpreterMacroAssembler::profile_ret(Register return_bci, 1290 Register mdp) { 1291 if (ProfileInterpreter) { 1292 Label profile_continue; 1293 uint row; 1294 1295 // If no method data exists, go to profile_continue. 1296 test_method_data_pointer(mdp, profile_continue); 1297 1298 // Update the total ret count. 1299 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1300 1301 for (row = 0; row < RetData::row_limit(); row++) { 1302 Label next_test; 1303 1304 // See if return_bci is equal to bci[n]: 1305 test_mdp_data_at(mdp, 1306 in_bytes(RetData::bci_offset(row)), 1307 return_bci, noreg, 1308 next_test); 1309 1310 // return_bci is equal to bci[n]. Increment the count. 1311 increment_mdp_data_at(mdp, in_bytes(RetData::bci_count_offset(row))); 1312 1313 // The method data pointer needs to be updated to reflect the new target. 1314 update_mdp_by_offset(mdp, 1315 in_bytes(RetData::bci_displacement_offset(row))); 1316 b(profile_continue); 1317 bind(next_test); 1318 } 1319 1320 update_mdp_for_ret(return_bci); 1321 1322 bind(profile_continue); 1323 } 1324 } 1325 1326 void InterpreterMacroAssembler::profile_null_seen(Register mdp) { 1327 if (ProfileInterpreter) { 1328 Label profile_continue; 1329 1330 // If no method data exists, go to profile_continue. 1331 test_method_data_pointer(mdp, profile_continue); 1332 1333 set_mdp_flag_at(mdp, BitData::null_seen_byte_constant()); 1334 1335 // The method data pointer needs to be updated. 1336 int mdp_delta = in_bytes(BitData::bit_data_size()); 1337 if (TypeProfileCasts) { 1338 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size()); 1339 } 1340 update_mdp_by_constant(mdp, mdp_delta); 1341 1342 bind(profile_continue); 1343 } 1344 } 1345 1346 void InterpreterMacroAssembler::profile_typecheck_failed(Register mdp) { 1347 if (ProfileInterpreter && TypeProfileCasts) { 1348 Label profile_continue; 1349 1350 // If no method data exists, go to profile_continue. 1351 test_method_data_pointer(mdp, profile_continue); 1352 1353 int count_offset = in_bytes(CounterData::count_offset()); 1354 // Back up the address, since we have already bumped the mdp. 1355 count_offset -= in_bytes(VirtualCallData::virtual_call_data_size()); 1356 1357 // *Decrement* the counter. We expect to see zero or small negatives. 1358 increment_mdp_data_at(mdp, count_offset, true); 1359 1360 bind (profile_continue); 1361 } 1362 } 1363 1364 void InterpreterMacroAssembler::profile_typecheck(Register mdp, Register klass, Register reg2) { 1365 if (ProfileInterpreter) { 1366 Label profile_continue; 1367 1368 // If no method data exists, go to profile_continue. 1369 test_method_data_pointer(mdp, profile_continue); 1370 1371 // The method data pointer needs to be updated. 1372 int mdp_delta = in_bytes(BitData::bit_data_size()); 1373 if (TypeProfileCasts) { 1374 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size()); 1375 1376 // Record the object type. 1377 record_klass_in_profile(klass, mdp, reg2, false); 1378 } 1379 update_mdp_by_constant(mdp, mdp_delta); 1380 1381 bind(profile_continue); 1382 } 1383 } 1384 1385 void InterpreterMacroAssembler::profile_switch_default(Register mdp) { 1386 if (ProfileInterpreter) { 1387 Label profile_continue; 1388 1389 // If no method data exists, go to profile_continue. 1390 test_method_data_pointer(mdp, profile_continue); 1391 1392 // Update the default case count 1393 increment_mdp_data_at(mdp, 1394 in_bytes(MultiBranchData::default_count_offset())); 1395 1396 // The method data pointer needs to be updated. 1397 update_mdp_by_offset(mdp, 1398 in_bytes(MultiBranchData:: 1399 default_displacement_offset())); 1400 1401 bind(profile_continue); 1402 } 1403 } 1404 1405 void InterpreterMacroAssembler::profile_switch_case(Register index, 1406 Register mdp, 1407 Register reg2) { 1408 if (ProfileInterpreter) { 1409 Label profile_continue; 1410 1411 // If no method data exists, go to profile_continue. 1412 test_method_data_pointer(mdp, profile_continue); 1413 1414 // Build the base (index * per_case_size_in_bytes()) + 1415 // case_array_offset_in_bytes() 1416 movw(reg2, in_bytes(MultiBranchData::per_case_size())); 1417 movw(rscratch1, in_bytes(MultiBranchData::case_array_offset())); 1418 Assembler::maddw(index, index, reg2, rscratch1); 1419 1420 // Update the case count 1421 increment_mdp_data_at(mdp, 1422 index, 1423 in_bytes(MultiBranchData::relative_count_offset())); 1424 1425 // The method data pointer needs to be updated. 1426 update_mdp_by_offset(mdp, 1427 index, 1428 in_bytes(MultiBranchData:: 1429 relative_displacement_offset())); 1430 1431 bind(profile_continue); 1432 } 1433 } 1434 1435 void InterpreterMacroAssembler::verify_oop(Register reg, TosState state) { 1436 if (state == atos) { 1437 MacroAssembler::verify_oop(reg); 1438 } 1439 } 1440 1441 void InterpreterMacroAssembler::verify_FPU(int stack_depth, TosState state) { ; } 1442 1443 1444 void InterpreterMacroAssembler::notify_method_entry() { 1445 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to 1446 // track stack depth. If it is possible to enter interp_only_mode we add 1447 // the code to check if the event should be sent. 1448 if (JvmtiExport::can_post_interpreter_events()) { 1449 Label L; 1450 ldrw(r3, Address(rthread, JavaThread::interp_only_mode_offset())); 1451 cbzw(r3, L); 1452 call_VM(noreg, CAST_FROM_FN_PTR(address, 1453 InterpreterRuntime::post_method_entry)); 1454 bind(L); 1455 } 1456 1457 { 1458 SkipIfEqual skip(this, &DTraceMethodProbes, false); 1459 get_method(c_rarg1); 1460 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 1461 rthread, c_rarg1); 1462 } 1463 1464 // RedefineClasses() tracing support for obsolete method entry 1465 if (log_is_enabled(Trace, redefine, class, obsolete)) { 1466 get_method(c_rarg1); 1467 call_VM_leaf( 1468 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 1469 rthread, c_rarg1); 1470 } 1471 1472 } 1473 1474 1475 void InterpreterMacroAssembler::notify_method_exit( 1476 TosState state, NotifyMethodExitMode mode) { 1477 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to 1478 // track stack depth. If it is possible to enter interp_only_mode we add 1479 // the code to check if the event should be sent. 1480 if (mode == NotifyJVMTI && JvmtiExport::can_post_interpreter_events()) { 1481 Label L; 1482 // Note: frame::interpreter_frame_result has a dependency on how the 1483 // method result is saved across the call to post_method_exit. If this 1484 // is changed then the interpreter_frame_result implementation will 1485 // need to be updated too. 1486 1487 // template interpreter will leave the result on the top of the stack. 1488 push(state); 1489 ldrw(r3, Address(rthread, JavaThread::interp_only_mode_offset())); 1490 cbz(r3, L); 1491 call_VM(noreg, 1492 CAST_FROM_FN_PTR(address, InterpreterRuntime::post_method_exit)); 1493 bind(L); 1494 pop(state); 1495 } 1496 1497 { 1498 SkipIfEqual skip(this, &DTraceMethodProbes, false); 1499 push(state); 1500 get_method(c_rarg1); 1501 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 1502 rthread, c_rarg1); 1503 pop(state); 1504 } 1505 } 1506 1507 1508 // Jump if ((*counter_addr += increment) & mask) satisfies the condition. 1509 void InterpreterMacroAssembler::increment_mask_and_jump(Address counter_addr, 1510 int increment, Address mask, 1511 Register scratch, Register scratch2, 1512 bool preloaded, Condition cond, 1513 Label* where) { 1514 if (!preloaded) { 1515 ldrw(scratch, counter_addr); 1516 } 1517 add(scratch, scratch, increment); 1518 strw(scratch, counter_addr); 1519 ldrw(scratch2, mask); 1520 ands(scratch, scratch, scratch2); 1521 br(cond, *where); 1522 } 1523 1524 void InterpreterMacroAssembler::call_VM_leaf_base(address entry_point, 1525 int number_of_arguments) { 1526 // interpreter specific 1527 // 1528 // Note: No need to save/restore rbcp & rlocals pointer since these 1529 // are callee saved registers and no blocking/ GC can happen 1530 // in leaf calls. 1531 #ifdef ASSERT 1532 { 1533 Label L; 1534 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 1535 cbz(rscratch1, L); 1536 stop("InterpreterMacroAssembler::call_VM_leaf_base:" 1537 " last_sp != NULL"); 1538 bind(L); 1539 } 1540 #endif /* ASSERT */ 1541 // super call 1542 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments); 1543 } 1544 1545 void InterpreterMacroAssembler::call_VM_base(Register oop_result, 1546 Register java_thread, 1547 Register last_java_sp, 1548 address entry_point, 1549 int number_of_arguments, 1550 bool check_exceptions) { 1551 // interpreter specific 1552 // 1553 // Note: Could avoid restoring locals ptr (callee saved) - however doesn't 1554 // really make a difference for these runtime calls, since they are 1555 // slow anyway. Btw., bcp must be saved/restored since it may change 1556 // due to GC. 1557 // assert(java_thread == noreg , "not expecting a precomputed java thread"); 1558 save_bcp(); 1559 #ifdef ASSERT 1560 { 1561 Label L; 1562 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 1563 cbz(rscratch1, L); 1564 stop("InterpreterMacroAssembler::call_VM_leaf_base:" 1565 " last_sp != NULL"); 1566 bind(L); 1567 } 1568 #endif /* ASSERT */ 1569 // super call 1570 MacroAssembler::call_VM_base(oop_result, noreg, last_java_sp, 1571 entry_point, number_of_arguments, 1572 check_exceptions); 1573 // interpreter specific 1574 restore_bcp(); 1575 restore_locals(); 1576 } 1577 1578 void InterpreterMacroAssembler::profile_obj_type(Register obj, const Address& mdo_addr) { 1579 Label update, next, none; 1580 1581 verify_oop(obj); 1582 1583 cbnz(obj, update); 1584 orptr(mdo_addr, TypeEntries::null_seen); 1585 b(next); 1586 1587 bind(update); 1588 load_klass(obj, obj); 1589 1590 ldr(rscratch1, mdo_addr); 1591 eor(obj, obj, rscratch1); 1592 tst(obj, TypeEntries::type_klass_mask); 1593 br(Assembler::EQ, next); // klass seen before, nothing to 1594 // do. The unknown bit may have been 1595 // set already but no need to check. 1596 1597 tbnz(obj, exact_log2(TypeEntries::type_unknown), next); 1598 // already unknown. Nothing to do anymore. 1599 1600 ldr(rscratch1, mdo_addr); 1601 cbz(rscratch1, none); 1602 cmp(rscratch1, TypeEntries::null_seen); 1603 br(Assembler::EQ, none); 1604 // There is a chance that the checks above (re-reading profiling 1605 // data from memory) fail if another thread has just set the 1606 // profiling to this obj's klass 1607 ldr(rscratch1, mdo_addr); 1608 eor(obj, obj, rscratch1); 1609 tst(obj, TypeEntries::type_klass_mask); 1610 br(Assembler::EQ, next); 1611 1612 // different than before. Cannot keep accurate profile. 1613 orptr(mdo_addr, TypeEntries::type_unknown); 1614 b(next); 1615 1616 bind(none); 1617 // first time here. Set profile type. 1618 str(obj, mdo_addr); 1619 1620 bind(next); 1621 } 1622 1623 void InterpreterMacroAssembler::profile_arguments_type(Register mdp, Register callee, Register tmp, bool is_virtual) { 1624 if (!ProfileInterpreter) { 1625 return; 1626 } 1627 1628 if (MethodData::profile_arguments() || MethodData::profile_return()) { 1629 Label profile_continue; 1630 1631 test_method_data_pointer(mdp, profile_continue); 1632 1633 int off_to_start = is_virtual ? in_bytes(VirtualCallData::virtual_call_data_size()) : in_bytes(CounterData::counter_data_size()); 1634 1635 ldrb(rscratch1, Address(mdp, in_bytes(DataLayout::tag_offset()) - off_to_start)); 1636 cmp(rscratch1, is_virtual ? DataLayout::virtual_call_type_data_tag : DataLayout::call_type_data_tag); 1637 br(Assembler::NE, profile_continue); 1638 1639 if (MethodData::profile_arguments()) { 1640 Label done; 1641 int off_to_args = in_bytes(TypeEntriesAtCall::args_data_offset()); 1642 1643 for (int i = 0; i < TypeProfileArgsLimit; i++) { 1644 if (i > 0 || MethodData::profile_return()) { 1645 // If return value type is profiled we may have no argument to profile 1646 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset()))); 1647 sub(tmp, tmp, i*TypeStackSlotEntries::per_arg_count()); 1648 cmp(tmp, TypeStackSlotEntries::per_arg_count()); 1649 add(rscratch1, mdp, off_to_args); 1650 br(Assembler::LT, done); 1651 } 1652 ldr(tmp, Address(callee, Method::const_offset())); 1653 load_unsigned_short(tmp, Address(tmp, ConstMethod::size_of_parameters_offset())); 1654 // stack offset o (zero based) from the start of the argument 1655 // list, for n arguments translates into offset n - o - 1 from 1656 // the end of the argument list 1657 ldr(rscratch1, Address(mdp, in_bytes(TypeEntriesAtCall::stack_slot_offset(i)))); 1658 sub(tmp, tmp, rscratch1); 1659 sub(tmp, tmp, 1); 1660 Address arg_addr = argument_address(tmp); 1661 ldr(tmp, arg_addr); 1662 1663 Address mdo_arg_addr(mdp, in_bytes(TypeEntriesAtCall::argument_type_offset(i))); 1664 profile_obj_type(tmp, mdo_arg_addr); 1665 1666 int to_add = in_bytes(TypeStackSlotEntries::per_arg_size()); 1667 off_to_args += to_add; 1668 } 1669 1670 if (MethodData::profile_return()) { 1671 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset()))); 1672 sub(tmp, tmp, TypeProfileArgsLimit*TypeStackSlotEntries::per_arg_count()); 1673 } 1674 1675 add(rscratch1, mdp, off_to_args); 1676 bind(done); 1677 mov(mdp, rscratch1); 1678 1679 if (MethodData::profile_return()) { 1680 // We're right after the type profile for the last 1681 // argument. tmp is the number of cells left in the 1682 // CallTypeData/VirtualCallTypeData to reach its end. Non null 1683 // if there's a return to profile. 1684 assert(ReturnTypeEntry::static_cell_count() < TypeStackSlotEntries::per_arg_count(), "can't move past ret type"); 1685 add(mdp, mdp, tmp, LSL, exact_log2(DataLayout::cell_size)); 1686 } 1687 str(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 1688 } else { 1689 assert(MethodData::profile_return(), "either profile call args or call ret"); 1690 update_mdp_by_constant(mdp, in_bytes(TypeEntriesAtCall::return_only_size())); 1691 } 1692 1693 // mdp points right after the end of the 1694 // CallTypeData/VirtualCallTypeData, right after the cells for the 1695 // return value type if there's one 1696 1697 bind(profile_continue); 1698 } 1699 } 1700 1701 void InterpreterMacroAssembler::profile_return_type(Register mdp, Register ret, Register tmp) { 1702 assert_different_registers(mdp, ret, tmp, rbcp); 1703 if (ProfileInterpreter && MethodData::profile_return()) { 1704 Label profile_continue, done; 1705 1706 test_method_data_pointer(mdp, profile_continue); 1707 1708 if (MethodData::profile_return_jsr292_only()) { 1709 assert(Method::intrinsic_id_size_in_bytes() == 2, "assuming Method::_intrinsic_id is u2"); 1710 1711 // If we don't profile all invoke bytecodes we must make sure 1712 // it's a bytecode we indeed profile. We can't go back to the 1713 // begining of the ProfileData we intend to update to check its 1714 // type because we're right after it and we don't known its 1715 // length 1716 Label do_profile; 1717 ldrb(rscratch1, Address(rbcp, 0)); 1718 cmp(rscratch1, Bytecodes::_invokedynamic); 1719 br(Assembler::EQ, do_profile); 1720 cmp(rscratch1, Bytecodes::_invokehandle); 1721 br(Assembler::EQ, do_profile); 1722 get_method(tmp); 1723 ldrh(rscratch1, Address(tmp, Method::intrinsic_id_offset_in_bytes())); 1724 cmp(rscratch1, vmIntrinsics::_compiledLambdaForm); 1725 br(Assembler::NE, profile_continue); 1726 1727 bind(do_profile); 1728 } 1729 1730 Address mdo_ret_addr(mdp, -in_bytes(ReturnTypeEntry::size())); 1731 mov(tmp, ret); 1732 profile_obj_type(tmp, mdo_ret_addr); 1733 1734 bind(profile_continue); 1735 } 1736 } 1737 1738 void InterpreterMacroAssembler::profile_parameters_type(Register mdp, Register tmp1, Register tmp2) { 1739 if (ProfileInterpreter && MethodData::profile_parameters()) { 1740 Label profile_continue, done; 1741 1742 test_method_data_pointer(mdp, profile_continue); 1743 1744 // Load the offset of the area within the MDO used for 1745 // parameters. If it's negative we're not profiling any parameters 1746 ldr(tmp1, Address(mdp, in_bytes(MethodData::parameters_type_data_di_offset()) - in_bytes(MethodData::data_offset()))); 1747 cmp(tmp1, 0u); 1748 br(Assembler::LT, profile_continue); 1749 1750 // Compute a pointer to the area for parameters from the offset 1751 // and move the pointer to the slot for the last 1752 // parameters. Collect profiling from last parameter down. 1753 // mdo start + parameters offset + array length - 1 1754 add(mdp, mdp, tmp1); 1755 ldr(tmp1, Address(mdp, ArrayData::array_len_offset())); 1756 sub(tmp1, tmp1, TypeStackSlotEntries::per_arg_count()); 1757 1758 Label loop; 1759 bind(loop); 1760 1761 int off_base = in_bytes(ParametersTypeData::stack_slot_offset(0)); 1762 int type_base = in_bytes(ParametersTypeData::type_offset(0)); 1763 int per_arg_scale = exact_log2(DataLayout::cell_size); 1764 add(rscratch1, mdp, off_base); 1765 add(rscratch2, mdp, type_base); 1766 1767 Address arg_off(rscratch1, tmp1, Address::lsl(per_arg_scale)); 1768 Address arg_type(rscratch2, tmp1, Address::lsl(per_arg_scale)); 1769 1770 // load offset on the stack from the slot for this parameter 1771 ldr(tmp2, arg_off); 1772 neg(tmp2, tmp2); 1773 // read the parameter from the local area 1774 ldr(tmp2, Address(rlocals, tmp2, Address::lsl(Interpreter::logStackElementSize))); 1775 1776 // profile the parameter 1777 profile_obj_type(tmp2, arg_type); 1778 1779 // go to next parameter 1780 subs(tmp1, tmp1, TypeStackSlotEntries::per_arg_count()); 1781 br(Assembler::GE, loop); 1782 1783 bind(profile_continue); 1784 } 1785 }