1 /* 2 * Copyright (c) 2003, 2017, 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::cache_offset_in_bytes())); 274 ldr(result, Address(result, ConstantPoolCache::resolved_references_offset_in_bytes())); 275 // JNIHandles::resolve(obj); 276 ldr(result, Address(result, 0)); 277 // Add in the index 278 add(result, result, tmp); 279 load_heap_oop(result, Address(result, arrayOopDesc::base_offset_in_bytes(T_OBJECT))); 280 } 281 282 // Generate a subtype check: branch to ok_is_subtype if sub_klass is a 283 // subtype of super_klass. 284 // 285 // Args: 286 // r0: superklass 287 // Rsub_klass: subklass 288 // 289 // Kills: 290 // r2, r5 291 void InterpreterMacroAssembler::gen_subtype_check(Register Rsub_klass, 292 Label& ok_is_subtype) { 293 assert(Rsub_klass != r0, "r0 holds superklass"); 294 assert(Rsub_klass != r2, "r2 holds 2ndary super array length"); 295 assert(Rsub_klass != r5, "r5 holds 2ndary super array scan ptr"); 296 297 // Profile the not-null value's klass. 298 profile_typecheck(r2, Rsub_klass, r5); // blows r2, reloads r5 299 300 // Do the check. 301 check_klass_subtype(Rsub_klass, r0, r2, ok_is_subtype); // blows r2 302 303 // Profile the failure of the check. 304 profile_typecheck_failed(r2); // blows r2 305 } 306 307 // Java Expression Stack 308 309 void InterpreterMacroAssembler::pop_ptr(Register r) { 310 ldr(r, post(esp, wordSize)); 311 } 312 313 void InterpreterMacroAssembler::pop_i(Register r) { 314 ldrw(r, post(esp, wordSize)); 315 } 316 317 void InterpreterMacroAssembler::pop_l(Register r) { 318 ldr(r, post(esp, 2 * Interpreter::stackElementSize)); 319 } 320 321 void InterpreterMacroAssembler::push_ptr(Register r) { 322 str(r, pre(esp, -wordSize)); 323 } 324 325 void InterpreterMacroAssembler::push_i(Register r) { 326 str(r, pre(esp, -wordSize)); 327 } 328 329 void InterpreterMacroAssembler::push_l(Register r) { 330 str(zr, pre(esp, -wordSize)); 331 str(r, pre(esp, - wordSize)); 332 } 333 334 void InterpreterMacroAssembler::pop_f(FloatRegister r) { 335 ldrs(r, post(esp, wordSize)); 336 } 337 338 void InterpreterMacroAssembler::pop_d(FloatRegister r) { 339 ldrd(r, post(esp, 2 * Interpreter::stackElementSize)); 340 } 341 342 void InterpreterMacroAssembler::push_f(FloatRegister r) { 343 strs(r, pre(esp, -wordSize)); 344 } 345 346 void InterpreterMacroAssembler::push_d(FloatRegister r) { 347 strd(r, pre(esp, 2* -wordSize)); 348 } 349 350 void InterpreterMacroAssembler::pop(TosState state) { 351 switch (state) { 352 case atos: pop_ptr(); break; 353 case btos: 354 case ztos: 355 case ctos: 356 case stos: 357 case itos: pop_i(); break; 358 case ltos: pop_l(); break; 359 case ftos: pop_f(); break; 360 case dtos: pop_d(); break; 361 case vtos: /* nothing to do */ break; 362 default: ShouldNotReachHere(); 363 } 364 verify_oop(r0, state); 365 } 366 367 void InterpreterMacroAssembler::push(TosState state) { 368 verify_oop(r0, state); 369 switch (state) { 370 case atos: push_ptr(); break; 371 case btos: 372 case ztos: 373 case ctos: 374 case stos: 375 case itos: push_i(); break; 376 case ltos: push_l(); break; 377 case ftos: push_f(); break; 378 case dtos: push_d(); break; 379 case vtos: /* nothing to do */ break; 380 default : ShouldNotReachHere(); 381 } 382 } 383 384 // Helpers for swap and dup 385 void InterpreterMacroAssembler::load_ptr(int n, Register val) { 386 ldr(val, Address(esp, Interpreter::expr_offset_in_bytes(n))); 387 } 388 389 void InterpreterMacroAssembler::store_ptr(int n, Register val) { 390 str(val, Address(esp, Interpreter::expr_offset_in_bytes(n))); 391 } 392 393 394 void InterpreterMacroAssembler::prepare_to_jump_from_interpreted() { 395 // set sender sp 396 mov(r13, sp); 397 // record last_sp 398 str(esp, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 399 } 400 401 // Jump to from_interpreted entry of a call unless single stepping is possible 402 // in this thread in which case we must call the i2i entry 403 void InterpreterMacroAssembler::jump_from_interpreted(Register method, Register temp) { 404 prepare_to_jump_from_interpreted(); 405 406 if (JvmtiExport::can_post_interpreter_events()) { 407 Label run_compiled_code; 408 // JVMTI events, such as single-stepping, are implemented partly by avoiding running 409 // compiled code in threads for which the event is enabled. Check here for 410 // interp_only_mode if these events CAN be enabled. 411 ldrw(rscratch1, Address(rthread, JavaThread::interp_only_mode_offset())); 412 cbzw(rscratch1, run_compiled_code); 413 ldr(rscratch1, Address(method, Method::interpreter_entry_offset())); 414 br(rscratch1); 415 bind(run_compiled_code); 416 } 417 418 ldr(rscratch1, Address(method, Method::from_interpreted_offset())); 419 br(rscratch1); 420 } 421 422 // The following two routines provide a hook so that an implementation 423 // can schedule the dispatch in two parts. amd64 does not do this. 424 void InterpreterMacroAssembler::dispatch_prolog(TosState state, int step) { 425 } 426 427 void InterpreterMacroAssembler::dispatch_epilog(TosState state, int step) { 428 dispatch_next(state, step); 429 } 430 431 void InterpreterMacroAssembler::dispatch_base(TosState state, 432 address* table, 433 bool verifyoop) { 434 if (VerifyActivationFrameSize) { 435 Unimplemented(); 436 } 437 if (verifyoop) { 438 verify_oop(r0, state); 439 } 440 if (table == Interpreter::dispatch_table(state)) { 441 addw(rscratch2, rscratch1, Interpreter::distance_from_dispatch_table(state)); 442 ldr(rscratch2, Address(rdispatch, rscratch2, Address::uxtw(3))); 443 } else { 444 mov(rscratch2, (address)table); 445 ldr(rscratch2, Address(rscratch2, rscratch1, Address::uxtw(3))); 446 } 447 br(rscratch2); 448 } 449 450 void InterpreterMacroAssembler::dispatch_only(TosState state) { 451 dispatch_base(state, Interpreter::dispatch_table(state)); 452 } 453 454 void InterpreterMacroAssembler::dispatch_only_normal(TosState state) { 455 dispatch_base(state, Interpreter::normal_table(state)); 456 } 457 458 void InterpreterMacroAssembler::dispatch_only_noverify(TosState state) { 459 dispatch_base(state, Interpreter::normal_table(state), false); 460 } 461 462 463 void InterpreterMacroAssembler::dispatch_next(TosState state, int step) { 464 // load next bytecode 465 ldrb(rscratch1, Address(pre(rbcp, step))); 466 dispatch_base(state, Interpreter::dispatch_table(state)); 467 } 468 469 void InterpreterMacroAssembler::dispatch_via(TosState state, address* table) { 470 // load current bytecode 471 ldrb(rscratch1, Address(rbcp, 0)); 472 dispatch_base(state, table); 473 } 474 475 // remove activation 476 // 477 // Unlock the receiver if this is a synchronized method. 478 // Unlock any Java monitors from syncronized blocks. 479 // Remove the activation from the stack. 480 // 481 // If there are locked Java monitors 482 // If throw_monitor_exception 483 // throws IllegalMonitorStateException 484 // Else if install_monitor_exception 485 // installs IllegalMonitorStateException 486 // Else 487 // no error processing 488 void InterpreterMacroAssembler::remove_activation( 489 TosState state, 490 bool throw_monitor_exception, 491 bool install_monitor_exception, 492 bool notify_jvmdi) { 493 // Note: Registers r3 xmm0 may be in use for the 494 // result check if synchronized method 495 Label unlocked, unlock, no_unlock; 496 497 // get the value of _do_not_unlock_if_synchronized into r3 498 const Address do_not_unlock_if_synchronized(rthread, 499 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset())); 500 ldrb(r3, do_not_unlock_if_synchronized); 501 strb(zr, do_not_unlock_if_synchronized); // reset the flag 502 503 // get method access flags 504 ldr(r1, Address(rfp, frame::interpreter_frame_method_offset * wordSize)); 505 ldr(r2, Address(r1, Method::access_flags_offset())); 506 tbz(r2, exact_log2(JVM_ACC_SYNCHRONIZED), unlocked); 507 508 // Don't unlock anything if the _do_not_unlock_if_synchronized flag 509 // is set. 510 cbnz(r3, no_unlock); 511 512 // unlock monitor 513 push(state); // save result 514 515 // BasicObjectLock will be first in list, since this is a 516 // synchronized method. However, need to check that the object has 517 // not been unlocked by an explicit monitorexit bytecode. 518 const Address monitor(rfp, frame::interpreter_frame_initial_sp_offset * 519 wordSize - (int) sizeof(BasicObjectLock)); 520 // We use c_rarg1 so that if we go slow path it will be the correct 521 // register for unlock_object to pass to VM directly 522 lea(c_rarg1, monitor); // address of first monitor 523 524 ldr(r0, Address(c_rarg1, BasicObjectLock::obj_offset_in_bytes())); 525 cbnz(r0, unlock); 526 527 pop(state); 528 if (throw_monitor_exception) { 529 // Entry already unlocked, need to throw exception 530 call_VM(noreg, CAST_FROM_FN_PTR(address, 531 InterpreterRuntime::throw_illegal_monitor_state_exception)); 532 should_not_reach_here(); 533 } else { 534 // Monitor already unlocked during a stack unroll. If requested, 535 // install an illegal_monitor_state_exception. Continue with 536 // stack unrolling. 537 if (install_monitor_exception) { 538 call_VM(noreg, CAST_FROM_FN_PTR(address, 539 InterpreterRuntime::new_illegal_monitor_state_exception)); 540 } 541 b(unlocked); 542 } 543 544 bind(unlock); 545 unlock_object(c_rarg1); 546 pop(state); 547 548 // Check that for block-structured locking (i.e., that all locked 549 // objects has been unlocked) 550 bind(unlocked); 551 552 // r0: Might contain return value 553 554 // Check that all monitors are unlocked 555 { 556 Label loop, exception, entry, restart; 557 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize; 558 const Address monitor_block_top( 559 rfp, frame::interpreter_frame_monitor_block_top_offset * wordSize); 560 const Address monitor_block_bot( 561 rfp, frame::interpreter_frame_initial_sp_offset * wordSize); 562 563 bind(restart); 564 // We use c_rarg1 so that if we go slow path it will be the correct 565 // register for unlock_object to pass to VM directly 566 ldr(c_rarg1, monitor_block_top); // points to current entry, starting 567 // with top-most entry 568 lea(r19, monitor_block_bot); // points to word before bottom of 569 // monitor block 570 b(entry); 571 572 // Entry already locked, need to throw exception 573 bind(exception); 574 575 if (throw_monitor_exception) { 576 // Throw exception 577 MacroAssembler::call_VM(noreg, 578 CAST_FROM_FN_PTR(address, InterpreterRuntime:: 579 throw_illegal_monitor_state_exception)); 580 should_not_reach_here(); 581 } else { 582 // Stack unrolling. Unlock object and install illegal_monitor_exception. 583 // Unlock does not block, so don't have to worry about the frame. 584 // We don't have to preserve c_rarg1 since we are going to throw an exception. 585 586 push(state); 587 unlock_object(c_rarg1); 588 pop(state); 589 590 if (install_monitor_exception) { 591 call_VM(noreg, CAST_FROM_FN_PTR(address, 592 InterpreterRuntime:: 593 new_illegal_monitor_state_exception)); 594 } 595 596 b(restart); 597 } 598 599 bind(loop); 600 // check if current entry is used 601 ldr(rscratch1, Address(c_rarg1, BasicObjectLock::obj_offset_in_bytes())); 602 cbnz(rscratch1, exception); 603 604 add(c_rarg1, c_rarg1, entry_size); // otherwise advance to next entry 605 bind(entry); 606 cmp(c_rarg1, r19); // check if bottom reached 607 br(Assembler::NE, loop); // if not at bottom then check this entry 608 } 609 610 bind(no_unlock); 611 612 // jvmti support 613 if (notify_jvmdi) { 614 notify_method_exit(state, NotifyJVMTI); // preserve TOSCA 615 } else { 616 notify_method_exit(state, SkipNotifyJVMTI); // preserve TOSCA 617 } 618 619 // remove activation 620 // get sender esp 621 ldr(esp, 622 Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); 623 if (StackReservedPages > 0) { 624 // testing if reserved zone needs to be re-enabled 625 Label no_reserved_zone_enabling; 626 627 ldr(rscratch1, Address(rthread, JavaThread::reserved_stack_activation_offset())); 628 cmp(esp, rscratch1); 629 br(Assembler::LS, no_reserved_zone_enabling); 630 631 call_VM_leaf( 632 CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), rthread); 633 call_VM(noreg, CAST_FROM_FN_PTR(address, 634 InterpreterRuntime::throw_delayed_StackOverflowError)); 635 should_not_reach_here(); 636 637 bind(no_reserved_zone_enabling); 638 } 639 // remove frame anchor 640 leave(); 641 // If we're returning to interpreted code we will shortly be 642 // adjusting SP to allow some space for ESP. If we're returning to 643 // compiled code the saved sender SP was saved in sender_sp, so this 644 // restores it. 645 andr(sp, esp, -16); 646 } 647 648 // Lock object 649 // 650 // Args: 651 // c_rarg1: BasicObjectLock to be used for locking 652 // 653 // Kills: 654 // r0 655 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, .. (param regs) 656 // rscratch1, rscratch2 (scratch regs) 657 void InterpreterMacroAssembler::lock_object(Register lock_reg) 658 { 659 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be c_rarg1"); 660 if (UseHeavyMonitors) { 661 call_VM(noreg, 662 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter), 663 lock_reg); 664 } else { 665 Label done; 666 667 const Register swap_reg = r0; 668 const Register tmp = c_rarg2; 669 const Register obj_reg = c_rarg3; // Will contain the oop 670 671 const int obj_offset = BasicObjectLock::obj_offset_in_bytes(); 672 const int lock_offset = BasicObjectLock::lock_offset_in_bytes (); 673 const int mark_offset = lock_offset + 674 BasicLock::displaced_header_offset_in_bytes(); 675 676 Label slow_case; 677 678 // Load object pointer into obj_reg %c_rarg3 679 ldr(obj_reg, Address(lock_reg, obj_offset)); 680 681 if (UseBiasedLocking) { 682 biased_locking_enter(lock_reg, obj_reg, swap_reg, tmp, false, done, &slow_case); 683 } 684 685 // Load (object->mark() | 1) into swap_reg 686 ldr(rscratch1, Address(obj_reg, 0)); 687 orr(swap_reg, rscratch1, 1); 688 689 // Save (object->mark() | 1) into BasicLock's displaced header 690 str(swap_reg, Address(lock_reg, mark_offset)); 691 692 assert(lock_offset == 0, 693 "displached header must be first word in BasicObjectLock"); 694 695 Label fail; 696 if (PrintBiasedLockingStatistics) { 697 Label fast; 698 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, fast, &fail); 699 bind(fast); 700 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()), 701 rscratch2, rscratch1, tmp); 702 b(done); 703 bind(fail); 704 } else { 705 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL); 706 } 707 708 // Test if the oopMark is an obvious stack pointer, i.e., 709 // 1) (mark & 7) == 0, and 710 // 2) rsp <= mark < mark + os::pagesize() 711 // 712 // These 3 tests can be done by evaluating the following 713 // expression: ((mark - rsp) & (7 - os::vm_page_size())), 714 // assuming both stack pointer and pagesize have their 715 // least significant 3 bits clear. 716 // NOTE: the oopMark is in swap_reg %r0 as the result of cmpxchg 717 // NOTE2: aarch64 does not like to subtract sp from rn so take a 718 // copy 719 mov(rscratch1, sp); 720 sub(swap_reg, swap_reg, rscratch1); 721 ands(swap_reg, swap_reg, (unsigned long)(7 - os::vm_page_size())); 722 723 // Save the test result, for recursive case, the result is zero 724 str(swap_reg, Address(lock_reg, mark_offset)); 725 726 if (PrintBiasedLockingStatistics) { 727 br(Assembler::NE, slow_case); 728 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()), 729 rscratch2, rscratch1, tmp); 730 } 731 br(Assembler::EQ, done); 732 733 bind(slow_case); 734 735 // Call the runtime routine for slow case 736 call_VM(noreg, 737 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter), 738 lock_reg); 739 740 bind(done); 741 } 742 } 743 744 745 // Unlocks an object. Used in monitorexit bytecode and 746 // remove_activation. Throws an IllegalMonitorException if object is 747 // not locked by current thread. 748 // 749 // Args: 750 // c_rarg1: BasicObjectLock for lock 751 // 752 // Kills: 753 // r0 754 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, ... (param regs) 755 // rscratch1, rscratch2 (scratch regs) 756 void InterpreterMacroAssembler::unlock_object(Register lock_reg) 757 { 758 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be rarg1"); 759 760 if (UseHeavyMonitors) { 761 call_VM(noreg, 762 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit), 763 lock_reg); 764 } else { 765 Label done; 766 767 const Register swap_reg = r0; 768 const Register header_reg = c_rarg2; // Will contain the old oopMark 769 const Register obj_reg = c_rarg3; // Will contain the oop 770 771 save_bcp(); // Save in case of exception 772 773 // Convert from BasicObjectLock structure to object and BasicLock 774 // structure Store the BasicLock address into %r0 775 lea(swap_reg, Address(lock_reg, BasicObjectLock::lock_offset_in_bytes())); 776 777 // Load oop into obj_reg(%c_rarg3) 778 ldr(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); 779 780 // Free entry 781 str(zr, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); 782 783 if (UseBiasedLocking) { 784 biased_locking_exit(obj_reg, header_reg, done); 785 } 786 787 // Load the old header from BasicLock structure 788 ldr(header_reg, Address(swap_reg, 789 BasicLock::displaced_header_offset_in_bytes())); 790 791 // Test for recursion 792 cbz(header_reg, done); 793 794 // Atomic swap back the old header 795 cmpxchgptr(swap_reg, header_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL); 796 797 // Call the runtime routine for slow case. 798 str(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); // restore obj 799 call_VM(noreg, 800 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit), 801 lock_reg); 802 803 bind(done); 804 805 restore_bcp(); 806 } 807 } 808 809 void InterpreterMacroAssembler::test_method_data_pointer(Register mdp, 810 Label& zero_continue) { 811 assert(ProfileInterpreter, "must be profiling interpreter"); 812 ldr(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 813 cbz(mdp, zero_continue); 814 } 815 816 // Set the method data pointer for the current bcp. 817 void InterpreterMacroAssembler::set_method_data_pointer_for_bcp() { 818 assert(ProfileInterpreter, "must be profiling interpreter"); 819 Label set_mdp; 820 stp(r0, r1, Address(pre(sp, -2 * wordSize))); 821 822 // Test MDO to avoid the call if it is NULL. 823 ldr(r0, Address(rmethod, in_bytes(Method::method_data_offset()))); 824 cbz(r0, set_mdp); 825 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::bcp_to_di), rmethod, rbcp); 826 // r0: mdi 827 // mdo is guaranteed to be non-zero here, we checked for it before the call. 828 ldr(r1, Address(rmethod, in_bytes(Method::method_data_offset()))); 829 lea(r1, Address(r1, in_bytes(MethodData::data_offset()))); 830 add(r0, r1, r0); 831 str(r0, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 832 bind(set_mdp); 833 ldp(r0, r1, Address(post(sp, 2 * wordSize))); 834 } 835 836 void InterpreterMacroAssembler::verify_method_data_pointer() { 837 assert(ProfileInterpreter, "must be profiling interpreter"); 838 #ifdef ASSERT 839 Label verify_continue; 840 stp(r0, r1, Address(pre(sp, -2 * wordSize))); 841 stp(r2, r3, Address(pre(sp, -2 * wordSize))); 842 test_method_data_pointer(r3, verify_continue); // If mdp is zero, continue 843 get_method(r1); 844 845 // If the mdp is valid, it will point to a DataLayout header which is 846 // consistent with the bcp. The converse is highly probable also. 847 ldrsh(r2, Address(r3, in_bytes(DataLayout::bci_offset()))); 848 ldr(rscratch1, Address(r1, Method::const_offset())); 849 add(r2, r2, rscratch1, Assembler::LSL); 850 lea(r2, Address(r2, ConstMethod::codes_offset())); 851 cmp(r2, rbcp); 852 br(Assembler::EQ, verify_continue); 853 // r1: method 854 // rbcp: bcp // rbcp == 22 855 // r3: mdp 856 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::verify_mdp), 857 r1, rbcp, r3); 858 bind(verify_continue); 859 ldp(r2, r3, Address(post(sp, 2 * wordSize))); 860 ldp(r0, r1, Address(post(sp, 2 * wordSize))); 861 #endif // ASSERT 862 } 863 864 865 void InterpreterMacroAssembler::set_mdp_data_at(Register mdp_in, 866 int constant, 867 Register value) { 868 assert(ProfileInterpreter, "must be profiling interpreter"); 869 Address data(mdp_in, constant); 870 str(value, data); 871 } 872 873 874 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in, 875 int constant, 876 bool decrement) { 877 increment_mdp_data_at(mdp_in, noreg, constant, decrement); 878 } 879 880 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in, 881 Register reg, 882 int constant, 883 bool decrement) { 884 assert(ProfileInterpreter, "must be profiling interpreter"); 885 // %%% this does 64bit counters at best it is wasting space 886 // at worst it is a rare bug when counters overflow 887 888 assert_different_registers(rscratch2, rscratch1, mdp_in, reg); 889 890 Address addr1(mdp_in, constant); 891 Address addr2(rscratch2, reg, Address::lsl(0)); 892 Address &addr = addr1; 893 if (reg != noreg) { 894 lea(rscratch2, addr1); 895 addr = addr2; 896 } 897 898 if (decrement) { 899 // Decrement the register. Set condition codes. 900 // Intel does this 901 // addptr(data, (int32_t) -DataLayout::counter_increment); 902 // If the decrement causes the counter to overflow, stay negative 903 // Label L; 904 // jcc(Assembler::negative, L); 905 // addptr(data, (int32_t) DataLayout::counter_increment); 906 // so we do this 907 ldr(rscratch1, addr); 908 subs(rscratch1, rscratch1, (unsigned)DataLayout::counter_increment); 909 Label L; 910 br(Assembler::LO, L); // skip store if counter underflow 911 str(rscratch1, addr); 912 bind(L); 913 } else { 914 assert(DataLayout::counter_increment == 1, 915 "flow-free idiom only works with 1"); 916 // Intel does this 917 // Increment the register. Set carry flag. 918 // addptr(data, DataLayout::counter_increment); 919 // If the increment causes the counter to overflow, pull back by 1. 920 // sbbptr(data, (int32_t)0); 921 // so we do this 922 ldr(rscratch1, addr); 923 adds(rscratch1, rscratch1, DataLayout::counter_increment); 924 Label L; 925 br(Assembler::CS, L); // skip store if counter overflow 926 str(rscratch1, addr); 927 bind(L); 928 } 929 } 930 931 void InterpreterMacroAssembler::set_mdp_flag_at(Register mdp_in, 932 int flag_byte_constant) { 933 assert(ProfileInterpreter, "must be profiling interpreter"); 934 int header_offset = in_bytes(DataLayout::header_offset()); 935 int header_bits = DataLayout::flag_mask_to_header_mask(flag_byte_constant); 936 // Set the flag 937 ldr(rscratch1, Address(mdp_in, header_offset)); 938 orr(rscratch1, rscratch1, header_bits); 939 str(rscratch1, Address(mdp_in, header_offset)); 940 } 941 942 943 void InterpreterMacroAssembler::test_mdp_data_at(Register mdp_in, 944 int offset, 945 Register value, 946 Register test_value_out, 947 Label& not_equal_continue) { 948 assert(ProfileInterpreter, "must be profiling interpreter"); 949 if (test_value_out == noreg) { 950 ldr(rscratch1, Address(mdp_in, offset)); 951 cmp(value, rscratch1); 952 } else { 953 // Put the test value into a register, so caller can use it: 954 ldr(test_value_out, Address(mdp_in, offset)); 955 cmp(value, test_value_out); 956 } 957 br(Assembler::NE, not_equal_continue); 958 } 959 960 961 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in, 962 int offset_of_disp) { 963 assert(ProfileInterpreter, "must be profiling interpreter"); 964 ldr(rscratch1, Address(mdp_in, offset_of_disp)); 965 add(mdp_in, mdp_in, rscratch1, LSL); 966 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 967 } 968 969 970 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in, 971 Register reg, 972 int offset_of_disp) { 973 assert(ProfileInterpreter, "must be profiling interpreter"); 974 lea(rscratch1, Address(mdp_in, offset_of_disp)); 975 ldr(rscratch1, Address(rscratch1, reg, Address::lsl(0))); 976 add(mdp_in, mdp_in, rscratch1, LSL); 977 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 978 } 979 980 981 void InterpreterMacroAssembler::update_mdp_by_constant(Register mdp_in, 982 int constant) { 983 assert(ProfileInterpreter, "must be profiling interpreter"); 984 add(mdp_in, mdp_in, (unsigned)constant); 985 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 986 } 987 988 989 void InterpreterMacroAssembler::update_mdp_for_ret(Register return_bci) { 990 assert(ProfileInterpreter, "must be profiling interpreter"); 991 // save/restore across call_VM 992 stp(zr, return_bci, Address(pre(sp, -2 * wordSize))); 993 call_VM(noreg, 994 CAST_FROM_FN_PTR(address, InterpreterRuntime::update_mdp_for_ret), 995 return_bci); 996 ldp(zr, return_bci, Address(post(sp, 2 * wordSize))); 997 } 998 999 1000 void InterpreterMacroAssembler::profile_taken_branch(Register mdp, 1001 Register bumped_count) { 1002 if (ProfileInterpreter) { 1003 Label profile_continue; 1004 1005 // If no method data exists, go to profile_continue. 1006 // Otherwise, assign to mdp 1007 test_method_data_pointer(mdp, profile_continue); 1008 1009 // We are taking a branch. Increment the taken count. 1010 // We inline increment_mdp_data_at to return bumped_count in a register 1011 //increment_mdp_data_at(mdp, in_bytes(JumpData::taken_offset())); 1012 Address data(mdp, in_bytes(JumpData::taken_offset())); 1013 ldr(bumped_count, data); 1014 assert(DataLayout::counter_increment == 1, 1015 "flow-free idiom only works with 1"); 1016 // Intel does this to catch overflow 1017 // addptr(bumped_count, DataLayout::counter_increment); 1018 // sbbptr(bumped_count, 0); 1019 // so we do this 1020 adds(bumped_count, bumped_count, DataLayout::counter_increment); 1021 Label L; 1022 br(Assembler::CS, L); // skip store if counter overflow 1023 str(bumped_count, data); 1024 bind(L); 1025 // The method data pointer needs to be updated to reflect the new target. 1026 update_mdp_by_offset(mdp, in_bytes(JumpData::displacement_offset())); 1027 bind(profile_continue); 1028 } 1029 } 1030 1031 1032 void InterpreterMacroAssembler::profile_not_taken_branch(Register mdp) { 1033 if (ProfileInterpreter) { 1034 Label profile_continue; 1035 1036 // If no method data exists, go to profile_continue. 1037 test_method_data_pointer(mdp, profile_continue); 1038 1039 // We are taking a branch. Increment the not taken count. 1040 increment_mdp_data_at(mdp, in_bytes(BranchData::not_taken_offset())); 1041 1042 // The method data pointer needs to be updated to correspond to 1043 // the next bytecode 1044 update_mdp_by_constant(mdp, in_bytes(BranchData::branch_data_size())); 1045 bind(profile_continue); 1046 } 1047 } 1048 1049 1050 void InterpreterMacroAssembler::profile_call(Register mdp) { 1051 if (ProfileInterpreter) { 1052 Label profile_continue; 1053 1054 // If no method data exists, go to profile_continue. 1055 test_method_data_pointer(mdp, profile_continue); 1056 1057 // We are making a call. Increment the count. 1058 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1059 1060 // The method data pointer needs to be updated to reflect the new target. 1061 update_mdp_by_constant(mdp, in_bytes(CounterData::counter_data_size())); 1062 bind(profile_continue); 1063 } 1064 } 1065 1066 void InterpreterMacroAssembler::profile_final_call(Register mdp) { 1067 if (ProfileInterpreter) { 1068 Label profile_continue; 1069 1070 // If no method data exists, go to profile_continue. 1071 test_method_data_pointer(mdp, profile_continue); 1072 1073 // We are making a call. Increment the count. 1074 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1075 1076 // The method data pointer needs to be updated to reflect the new target. 1077 update_mdp_by_constant(mdp, 1078 in_bytes(VirtualCallData:: 1079 virtual_call_data_size())); 1080 bind(profile_continue); 1081 } 1082 } 1083 1084 1085 void InterpreterMacroAssembler::profile_virtual_call(Register receiver, 1086 Register mdp, 1087 Register reg2, 1088 bool receiver_can_be_null) { 1089 if (ProfileInterpreter) { 1090 Label profile_continue; 1091 1092 // If no method data exists, go to profile_continue. 1093 test_method_data_pointer(mdp, profile_continue); 1094 1095 Label skip_receiver_profile; 1096 if (receiver_can_be_null) { 1097 Label not_null; 1098 // We are making a call. Increment the count for null receiver. 1099 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1100 b(skip_receiver_profile); 1101 bind(not_null); 1102 } 1103 1104 // Record the receiver type. 1105 record_klass_in_profile(receiver, mdp, reg2, true); 1106 bind(skip_receiver_profile); 1107 1108 // The method data pointer needs to be updated to reflect the new target. 1109 #if INCLUDE_JVMCI 1110 if (MethodProfileWidth == 0) { 1111 update_mdp_by_constant(mdp, in_bytes(VirtualCallData::virtual_call_data_size())); 1112 } 1113 #else // INCLUDE_JVMCI 1114 update_mdp_by_constant(mdp, 1115 in_bytes(VirtualCallData:: 1116 virtual_call_data_size())); 1117 #endif // INCLUDE_JVMCI 1118 bind(profile_continue); 1119 } 1120 } 1121 1122 #if INCLUDE_JVMCI 1123 void InterpreterMacroAssembler::profile_called_method(Register method, Register mdp, Register reg2) { 1124 assert_different_registers(method, mdp, reg2); 1125 if (ProfileInterpreter && MethodProfileWidth > 0) { 1126 Label profile_continue; 1127 1128 // If no method data exists, go to profile_continue. 1129 test_method_data_pointer(mdp, profile_continue); 1130 1131 Label done; 1132 record_item_in_profile_helper(method, mdp, reg2, 0, done, MethodProfileWidth, 1133 &VirtualCallData::method_offset, &VirtualCallData::method_count_offset, in_bytes(VirtualCallData::nonprofiled_receiver_count_offset())); 1134 bind(done); 1135 1136 update_mdp_by_constant(mdp, in_bytes(VirtualCallData::virtual_call_data_size())); 1137 bind(profile_continue); 1138 } 1139 } 1140 #endif // INCLUDE_JVMCI 1141 1142 // This routine creates a state machine for updating the multi-row 1143 // type profile at a virtual call site (or other type-sensitive bytecode). 1144 // The machine visits each row (of receiver/count) until the receiver type 1145 // is found, or until it runs out of rows. At the same time, it remembers 1146 // the location of the first empty row. (An empty row records null for its 1147 // receiver, and can be allocated for a newly-observed receiver type.) 1148 // Because there are two degrees of freedom in the state, a simple linear 1149 // search will not work; it must be a decision tree. Hence this helper 1150 // function is recursive, to generate the required tree structured code. 1151 // It's the interpreter, so we are trading off code space for speed. 1152 // See below for example code. 1153 void InterpreterMacroAssembler::record_klass_in_profile_helper( 1154 Register receiver, Register mdp, 1155 Register reg2, int start_row, 1156 Label& done, bool is_virtual_call) { 1157 if (TypeProfileWidth == 0) { 1158 if (is_virtual_call) { 1159 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1160 } 1161 #if INCLUDE_JVMCI 1162 else if (EnableJVMCI) { 1163 increment_mdp_data_at(mdp, in_bytes(ReceiverTypeData::nonprofiled_receiver_count_offset())); 1164 } 1165 #endif // INCLUDE_JVMCI 1166 } else { 1167 int non_profiled_offset = -1; 1168 if (is_virtual_call) { 1169 non_profiled_offset = in_bytes(CounterData::count_offset()); 1170 } 1171 #if INCLUDE_JVMCI 1172 else if (EnableJVMCI) { 1173 non_profiled_offset = in_bytes(ReceiverTypeData::nonprofiled_receiver_count_offset()); 1174 } 1175 #endif // INCLUDE_JVMCI 1176 1177 record_item_in_profile_helper(receiver, mdp, reg2, 0, done, TypeProfileWidth, 1178 &VirtualCallData::receiver_offset, &VirtualCallData::receiver_count_offset, non_profiled_offset); 1179 } 1180 } 1181 1182 void InterpreterMacroAssembler::record_item_in_profile_helper(Register item, Register mdp, 1183 Register reg2, int start_row, Label& done, int total_rows, 1184 OffsetFunction item_offset_fn, OffsetFunction item_count_offset_fn, 1185 int non_profiled_offset) { 1186 int last_row = total_rows - 1; 1187 assert(start_row <= last_row, "must be work left to do"); 1188 // Test this row for both the item and for null. 1189 // Take any of three different outcomes: 1190 // 1. found item => increment count and goto done 1191 // 2. found null => keep looking for case 1, maybe allocate this cell 1192 // 3. found something else => keep looking for cases 1 and 2 1193 // Case 3 is handled by a recursive call. 1194 for (int row = start_row; row <= last_row; row++) { 1195 Label next_test; 1196 bool test_for_null_also = (row == start_row); 1197 1198 // See if the item is item[n]. 1199 int item_offset = in_bytes(item_offset_fn(row)); 1200 test_mdp_data_at(mdp, item_offset, item, 1201 (test_for_null_also ? reg2 : noreg), 1202 next_test); 1203 // (Reg2 now contains the item from the CallData.) 1204 1205 // The item is item[n]. Increment count[n]. 1206 int count_offset = in_bytes(item_count_offset_fn(row)); 1207 increment_mdp_data_at(mdp, count_offset); 1208 b(done); 1209 bind(next_test); 1210 1211 if (test_for_null_also) { 1212 Label found_null; 1213 // Failed the equality check on item[n]... Test for null. 1214 if (start_row == last_row) { 1215 // The only thing left to do is handle the null case. 1216 if (non_profiled_offset >= 0) { 1217 cbz(reg2, found_null); 1218 // Item did not match any saved item and there is no empty row for it. 1219 // Increment total counter to indicate polymorphic case. 1220 increment_mdp_data_at(mdp, non_profiled_offset); 1221 b(done); 1222 bind(found_null); 1223 } else { 1224 cbnz(reg2, done); 1225 } 1226 break; 1227 } 1228 // Since null is rare, make it be the branch-taken case. 1229 cbz(reg2, found_null); 1230 1231 // Put all the "Case 3" tests here. 1232 record_item_in_profile_helper(item, mdp, reg2, start_row + 1, done, total_rows, 1233 item_offset_fn, item_count_offset_fn, non_profiled_offset); 1234 1235 // Found a null. Keep searching for a matching item, 1236 // but remember that this is an empty (unused) slot. 1237 bind(found_null); 1238 } 1239 } 1240 1241 // In the fall-through case, we found no matching item, but we 1242 // observed the item[start_row] is NULL. 1243 1244 // Fill in the item field and increment the count. 1245 int item_offset = in_bytes(item_offset_fn(start_row)); 1246 set_mdp_data_at(mdp, item_offset, item); 1247 int count_offset = in_bytes(item_count_offset_fn(start_row)); 1248 mov(reg2, DataLayout::counter_increment); 1249 set_mdp_data_at(mdp, count_offset, reg2); 1250 if (start_row > 0) { 1251 b(done); 1252 } 1253 } 1254 1255 // Example state machine code for three profile rows: 1256 // // main copy of decision tree, rooted at row[1] 1257 // if (row[0].rec == rec) { row[0].incr(); goto done; } 1258 // if (row[0].rec != NULL) { 1259 // // inner copy of decision tree, rooted at row[1] 1260 // if (row[1].rec == rec) { row[1].incr(); goto done; } 1261 // if (row[1].rec != NULL) { 1262 // // degenerate decision tree, rooted at row[2] 1263 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1264 // if (row[2].rec != NULL) { count.incr(); goto done; } // overflow 1265 // row[2].init(rec); goto done; 1266 // } else { 1267 // // remember row[1] is empty 1268 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1269 // row[1].init(rec); goto done; 1270 // } 1271 // } else { 1272 // // remember row[0] is empty 1273 // if (row[1].rec == rec) { row[1].incr(); goto done; } 1274 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1275 // row[0].init(rec); goto done; 1276 // } 1277 // done: 1278 1279 void InterpreterMacroAssembler::record_klass_in_profile(Register receiver, 1280 Register mdp, Register reg2, 1281 bool is_virtual_call) { 1282 assert(ProfileInterpreter, "must be profiling"); 1283 Label done; 1284 1285 record_klass_in_profile_helper(receiver, mdp, reg2, 0, done, is_virtual_call); 1286 1287 bind (done); 1288 } 1289 1290 void InterpreterMacroAssembler::profile_ret(Register return_bci, 1291 Register mdp) { 1292 if (ProfileInterpreter) { 1293 Label profile_continue; 1294 uint row; 1295 1296 // If no method data exists, go to profile_continue. 1297 test_method_data_pointer(mdp, profile_continue); 1298 1299 // Update the total ret count. 1300 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1301 1302 for (row = 0; row < RetData::row_limit(); row++) { 1303 Label next_test; 1304 1305 // See if return_bci is equal to bci[n]: 1306 test_mdp_data_at(mdp, 1307 in_bytes(RetData::bci_offset(row)), 1308 return_bci, noreg, 1309 next_test); 1310 1311 // return_bci is equal to bci[n]. Increment the count. 1312 increment_mdp_data_at(mdp, in_bytes(RetData::bci_count_offset(row))); 1313 1314 // The method data pointer needs to be updated to reflect the new target. 1315 update_mdp_by_offset(mdp, 1316 in_bytes(RetData::bci_displacement_offset(row))); 1317 b(profile_continue); 1318 bind(next_test); 1319 } 1320 1321 update_mdp_for_ret(return_bci); 1322 1323 bind(profile_continue); 1324 } 1325 } 1326 1327 void InterpreterMacroAssembler::profile_null_seen(Register mdp) { 1328 if (ProfileInterpreter) { 1329 Label profile_continue; 1330 1331 // If no method data exists, go to profile_continue. 1332 test_method_data_pointer(mdp, profile_continue); 1333 1334 set_mdp_flag_at(mdp, BitData::null_seen_byte_constant()); 1335 1336 // The method data pointer needs to be updated. 1337 int mdp_delta = in_bytes(BitData::bit_data_size()); 1338 if (TypeProfileCasts) { 1339 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size()); 1340 } 1341 update_mdp_by_constant(mdp, mdp_delta); 1342 1343 bind(profile_continue); 1344 } 1345 } 1346 1347 void InterpreterMacroAssembler::profile_typecheck_failed(Register mdp) { 1348 if (ProfileInterpreter && TypeProfileCasts) { 1349 Label profile_continue; 1350 1351 // If no method data exists, go to profile_continue. 1352 test_method_data_pointer(mdp, profile_continue); 1353 1354 int count_offset = in_bytes(CounterData::count_offset()); 1355 // Back up the address, since we have already bumped the mdp. 1356 count_offset -= in_bytes(VirtualCallData::virtual_call_data_size()); 1357 1358 // *Decrement* the counter. We expect to see zero or small negatives. 1359 increment_mdp_data_at(mdp, count_offset, true); 1360 1361 bind (profile_continue); 1362 } 1363 } 1364 1365 void InterpreterMacroAssembler::profile_typecheck(Register mdp, Register klass, Register reg2) { 1366 if (ProfileInterpreter) { 1367 Label profile_continue; 1368 1369 // If no method data exists, go to profile_continue. 1370 test_method_data_pointer(mdp, profile_continue); 1371 1372 // The method data pointer needs to be updated. 1373 int mdp_delta = in_bytes(BitData::bit_data_size()); 1374 if (TypeProfileCasts) { 1375 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size()); 1376 1377 // Record the object type. 1378 record_klass_in_profile(klass, mdp, reg2, false); 1379 } 1380 update_mdp_by_constant(mdp, mdp_delta); 1381 1382 bind(profile_continue); 1383 } 1384 } 1385 1386 void InterpreterMacroAssembler::profile_switch_default(Register mdp) { 1387 if (ProfileInterpreter) { 1388 Label profile_continue; 1389 1390 // If no method data exists, go to profile_continue. 1391 test_method_data_pointer(mdp, profile_continue); 1392 1393 // Update the default case count 1394 increment_mdp_data_at(mdp, 1395 in_bytes(MultiBranchData::default_count_offset())); 1396 1397 // The method data pointer needs to be updated. 1398 update_mdp_by_offset(mdp, 1399 in_bytes(MultiBranchData:: 1400 default_displacement_offset())); 1401 1402 bind(profile_continue); 1403 } 1404 } 1405 1406 void InterpreterMacroAssembler::profile_switch_case(Register index, 1407 Register mdp, 1408 Register reg2) { 1409 if (ProfileInterpreter) { 1410 Label profile_continue; 1411 1412 // If no method data exists, go to profile_continue. 1413 test_method_data_pointer(mdp, profile_continue); 1414 1415 // Build the base (index * per_case_size_in_bytes()) + 1416 // case_array_offset_in_bytes() 1417 movw(reg2, in_bytes(MultiBranchData::per_case_size())); 1418 movw(rscratch1, in_bytes(MultiBranchData::case_array_offset())); 1419 Assembler::maddw(index, index, reg2, rscratch1); 1420 1421 // Update the case count 1422 increment_mdp_data_at(mdp, 1423 index, 1424 in_bytes(MultiBranchData::relative_count_offset())); 1425 1426 // The method data pointer needs to be updated. 1427 update_mdp_by_offset(mdp, 1428 index, 1429 in_bytes(MultiBranchData:: 1430 relative_displacement_offset())); 1431 1432 bind(profile_continue); 1433 } 1434 } 1435 1436 void InterpreterMacroAssembler::verify_oop(Register reg, TosState state) { 1437 if (state == atos) { 1438 MacroAssembler::verify_oop(reg); 1439 } 1440 } 1441 1442 void InterpreterMacroAssembler::verify_FPU(int stack_depth, TosState state) { ; } 1443 1444 1445 void InterpreterMacroAssembler::notify_method_entry() { 1446 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to 1447 // track stack depth. If it is possible to enter interp_only_mode we add 1448 // the code to check if the event should be sent. 1449 if (JvmtiExport::can_post_interpreter_events()) { 1450 Label L; 1451 ldrw(r3, Address(rthread, JavaThread::interp_only_mode_offset())); 1452 cbzw(r3, L); 1453 call_VM(noreg, CAST_FROM_FN_PTR(address, 1454 InterpreterRuntime::post_method_entry)); 1455 bind(L); 1456 } 1457 1458 { 1459 SkipIfEqual skip(this, &DTraceMethodProbes, false); 1460 get_method(c_rarg1); 1461 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 1462 rthread, c_rarg1); 1463 } 1464 1465 // RedefineClasses() tracing support for obsolete method entry 1466 if (log_is_enabled(Trace, redefine, class, obsolete)) { 1467 get_method(c_rarg1); 1468 call_VM_leaf( 1469 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 1470 rthread, c_rarg1); 1471 } 1472 1473 } 1474 1475 1476 void InterpreterMacroAssembler::notify_method_exit( 1477 TosState state, NotifyMethodExitMode mode) { 1478 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to 1479 // track stack depth. If it is possible to enter interp_only_mode we add 1480 // the code to check if the event should be sent. 1481 if (mode == NotifyJVMTI && JvmtiExport::can_post_interpreter_events()) { 1482 Label L; 1483 // Note: frame::interpreter_frame_result has a dependency on how the 1484 // method result is saved across the call to post_method_exit. If this 1485 // is changed then the interpreter_frame_result implementation will 1486 // need to be updated too. 1487 1488 // template interpreter will leave the result on the top of the stack. 1489 push(state); 1490 ldrw(r3, Address(rthread, JavaThread::interp_only_mode_offset())); 1491 cbz(r3, L); 1492 call_VM(noreg, 1493 CAST_FROM_FN_PTR(address, InterpreterRuntime::post_method_exit)); 1494 bind(L); 1495 pop(state); 1496 } 1497 1498 { 1499 SkipIfEqual skip(this, &DTraceMethodProbes, false); 1500 push(state); 1501 get_method(c_rarg1); 1502 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 1503 rthread, c_rarg1); 1504 pop(state); 1505 } 1506 } 1507 1508 1509 // Jump if ((*counter_addr += increment) & mask) satisfies the condition. 1510 void InterpreterMacroAssembler::increment_mask_and_jump(Address counter_addr, 1511 int increment, Address mask, 1512 Register scratch, Register scratch2, 1513 bool preloaded, Condition cond, 1514 Label* where) { 1515 if (!preloaded) { 1516 ldrw(scratch, counter_addr); 1517 } 1518 add(scratch, scratch, increment); 1519 strw(scratch, counter_addr); 1520 ldrw(scratch2, mask); 1521 ands(scratch, scratch, scratch2); 1522 br(cond, *where); 1523 } 1524 1525 void InterpreterMacroAssembler::call_VM_leaf_base(address entry_point, 1526 int number_of_arguments) { 1527 // interpreter specific 1528 // 1529 // Note: No need to save/restore rbcp & rlocals pointer since these 1530 // are callee saved registers and no blocking/ GC can happen 1531 // in leaf calls. 1532 #ifdef ASSERT 1533 { 1534 Label L; 1535 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 1536 cbz(rscratch1, L); 1537 stop("InterpreterMacroAssembler::call_VM_leaf_base:" 1538 " last_sp != NULL"); 1539 bind(L); 1540 } 1541 #endif /* ASSERT */ 1542 // super call 1543 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments); 1544 } 1545 1546 void InterpreterMacroAssembler::call_VM_base(Register oop_result, 1547 Register java_thread, 1548 Register last_java_sp, 1549 address entry_point, 1550 int number_of_arguments, 1551 bool check_exceptions) { 1552 // interpreter specific 1553 // 1554 // Note: Could avoid restoring locals ptr (callee saved) - however doesn't 1555 // really make a difference for these runtime calls, since they are 1556 // slow anyway. Btw., bcp must be saved/restored since it may change 1557 // due to GC. 1558 // assert(java_thread == noreg , "not expecting a precomputed java thread"); 1559 save_bcp(); 1560 #ifdef ASSERT 1561 { 1562 Label L; 1563 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 1564 cbz(rscratch1, L); 1565 stop("InterpreterMacroAssembler::call_VM_leaf_base:" 1566 " last_sp != NULL"); 1567 bind(L); 1568 } 1569 #endif /* ASSERT */ 1570 // super call 1571 MacroAssembler::call_VM_base(oop_result, noreg, last_java_sp, 1572 entry_point, number_of_arguments, 1573 check_exceptions); 1574 // interpreter specific 1575 restore_bcp(); 1576 restore_locals(); 1577 } 1578 1579 void InterpreterMacroAssembler::profile_obj_type(Register obj, const Address& mdo_addr) { 1580 Label update, next, none; 1581 1582 verify_oop(obj); 1583 1584 cbnz(obj, update); 1585 orptr(mdo_addr, TypeEntries::null_seen); 1586 b(next); 1587 1588 bind(update); 1589 load_klass(obj, obj); 1590 1591 ldr(rscratch1, mdo_addr); 1592 eor(obj, obj, rscratch1); 1593 tst(obj, TypeEntries::type_klass_mask); 1594 br(Assembler::EQ, next); // klass seen before, nothing to 1595 // do. The unknown bit may have been 1596 // set already but no need to check. 1597 1598 tbnz(obj, exact_log2(TypeEntries::type_unknown), next); 1599 // already unknown. Nothing to do anymore. 1600 1601 ldr(rscratch1, mdo_addr); 1602 cbz(rscratch1, none); 1603 cmp(rscratch1, TypeEntries::null_seen); 1604 br(Assembler::EQ, none); 1605 // There is a chance that the checks above (re-reading profiling 1606 // data from memory) fail if another thread has just set the 1607 // profiling to this obj's klass 1608 ldr(rscratch1, mdo_addr); 1609 eor(obj, obj, rscratch1); 1610 tst(obj, TypeEntries::type_klass_mask); 1611 br(Assembler::EQ, next); 1612 1613 // different than before. Cannot keep accurate profile. 1614 orptr(mdo_addr, TypeEntries::type_unknown); 1615 b(next); 1616 1617 bind(none); 1618 // first time here. Set profile type. 1619 str(obj, mdo_addr); 1620 1621 bind(next); 1622 } 1623 1624 void InterpreterMacroAssembler::profile_arguments_type(Register mdp, Register callee, Register tmp, bool is_virtual) { 1625 if (!ProfileInterpreter) { 1626 return; 1627 } 1628 1629 if (MethodData::profile_arguments() || MethodData::profile_return()) { 1630 Label profile_continue; 1631 1632 test_method_data_pointer(mdp, profile_continue); 1633 1634 int off_to_start = is_virtual ? in_bytes(VirtualCallData::virtual_call_data_size()) : in_bytes(CounterData::counter_data_size()); 1635 1636 ldrb(rscratch1, Address(mdp, in_bytes(DataLayout::tag_offset()) - off_to_start)); 1637 cmp(rscratch1, is_virtual ? DataLayout::virtual_call_type_data_tag : DataLayout::call_type_data_tag); 1638 br(Assembler::NE, profile_continue); 1639 1640 if (MethodData::profile_arguments()) { 1641 Label done; 1642 int off_to_args = in_bytes(TypeEntriesAtCall::args_data_offset()); 1643 1644 for (int i = 0; i < TypeProfileArgsLimit; i++) { 1645 if (i > 0 || MethodData::profile_return()) { 1646 // If return value type is profiled we may have no argument to profile 1647 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset()))); 1648 sub(tmp, tmp, i*TypeStackSlotEntries::per_arg_count()); 1649 cmp(tmp, TypeStackSlotEntries::per_arg_count()); 1650 add(rscratch1, mdp, off_to_args); 1651 br(Assembler::LT, done); 1652 } 1653 ldr(tmp, Address(callee, Method::const_offset())); 1654 load_unsigned_short(tmp, Address(tmp, ConstMethod::size_of_parameters_offset())); 1655 // stack offset o (zero based) from the start of the argument 1656 // list, for n arguments translates into offset n - o - 1 from 1657 // the end of the argument list 1658 ldr(rscratch1, Address(mdp, in_bytes(TypeEntriesAtCall::stack_slot_offset(i)))); 1659 sub(tmp, tmp, rscratch1); 1660 sub(tmp, tmp, 1); 1661 Address arg_addr = argument_address(tmp); 1662 ldr(tmp, arg_addr); 1663 1664 Address mdo_arg_addr(mdp, in_bytes(TypeEntriesAtCall::argument_type_offset(i))); 1665 profile_obj_type(tmp, mdo_arg_addr); 1666 1667 int to_add = in_bytes(TypeStackSlotEntries::per_arg_size()); 1668 off_to_args += to_add; 1669 } 1670 1671 if (MethodData::profile_return()) { 1672 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset()))); 1673 sub(tmp, tmp, TypeProfileArgsLimit*TypeStackSlotEntries::per_arg_count()); 1674 } 1675 1676 add(rscratch1, mdp, off_to_args); 1677 bind(done); 1678 mov(mdp, rscratch1); 1679 1680 if (MethodData::profile_return()) { 1681 // We're right after the type profile for the last 1682 // argument. tmp is the number of cells left in the 1683 // CallTypeData/VirtualCallTypeData to reach its end. Non null 1684 // if there's a return to profile. 1685 assert(ReturnTypeEntry::static_cell_count() < TypeStackSlotEntries::per_arg_count(), "can't move past ret type"); 1686 add(mdp, mdp, tmp, LSL, exact_log2(DataLayout::cell_size)); 1687 } 1688 str(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 1689 } else { 1690 assert(MethodData::profile_return(), "either profile call args or call ret"); 1691 update_mdp_by_constant(mdp, in_bytes(TypeEntriesAtCall::return_only_size())); 1692 } 1693 1694 // mdp points right after the end of the 1695 // CallTypeData/VirtualCallTypeData, right after the cells for the 1696 // return value type if there's one 1697 1698 bind(profile_continue); 1699 } 1700 } 1701 1702 void InterpreterMacroAssembler::profile_return_type(Register mdp, Register ret, Register tmp) { 1703 assert_different_registers(mdp, ret, tmp, rbcp); 1704 if (ProfileInterpreter && MethodData::profile_return()) { 1705 Label profile_continue, done; 1706 1707 test_method_data_pointer(mdp, profile_continue); 1708 1709 if (MethodData::profile_return_jsr292_only()) { 1710 assert(Method::intrinsic_id_size_in_bytes() == 2, "assuming Method::_intrinsic_id is u2"); 1711 1712 // If we don't profile all invoke bytecodes we must make sure 1713 // it's a bytecode we indeed profile. We can't go back to the 1714 // begining of the ProfileData we intend to update to check its 1715 // type because we're right after it and we don't known its 1716 // length 1717 Label do_profile; 1718 ldrb(rscratch1, Address(rbcp, 0)); 1719 cmp(rscratch1, Bytecodes::_invokedynamic); 1720 br(Assembler::EQ, do_profile); 1721 cmp(rscratch1, Bytecodes::_invokehandle); 1722 br(Assembler::EQ, do_profile); 1723 get_method(tmp); 1724 ldrh(rscratch1, Address(tmp, Method::intrinsic_id_offset_in_bytes())); 1725 cmp(rscratch1, vmIntrinsics::_compiledLambdaForm); 1726 br(Assembler::NE, profile_continue); 1727 1728 bind(do_profile); 1729 } 1730 1731 Address mdo_ret_addr(mdp, -in_bytes(ReturnTypeEntry::size())); 1732 mov(tmp, ret); 1733 profile_obj_type(tmp, mdo_ret_addr); 1734 1735 bind(profile_continue); 1736 } 1737 } 1738 1739 void InterpreterMacroAssembler::profile_parameters_type(Register mdp, Register tmp1, Register tmp2) { 1740 if (ProfileInterpreter && MethodData::profile_parameters()) { 1741 Label profile_continue, done; 1742 1743 test_method_data_pointer(mdp, profile_continue); 1744 1745 // Load the offset of the area within the MDO used for 1746 // parameters. If it's negative we're not profiling any parameters 1747 ldr(tmp1, Address(mdp, in_bytes(MethodData::parameters_type_data_di_offset()) - in_bytes(MethodData::data_offset()))); 1748 cmp(tmp1, 0u); 1749 br(Assembler::LT, profile_continue); 1750 1751 // Compute a pointer to the area for parameters from the offset 1752 // and move the pointer to the slot for the last 1753 // parameters. Collect profiling from last parameter down. 1754 // mdo start + parameters offset + array length - 1 1755 add(mdp, mdp, tmp1); 1756 ldr(tmp1, Address(mdp, ArrayData::array_len_offset())); 1757 sub(tmp1, tmp1, TypeStackSlotEntries::per_arg_count()); 1758 1759 Label loop; 1760 bind(loop); 1761 1762 int off_base = in_bytes(ParametersTypeData::stack_slot_offset(0)); 1763 int type_base = in_bytes(ParametersTypeData::type_offset(0)); 1764 int per_arg_scale = exact_log2(DataLayout::cell_size); 1765 add(rscratch1, mdp, off_base); 1766 add(rscratch2, mdp, type_base); 1767 1768 Address arg_off(rscratch1, tmp1, Address::lsl(per_arg_scale)); 1769 Address arg_type(rscratch2, tmp1, Address::lsl(per_arg_scale)); 1770 1771 // load offset on the stack from the slot for this parameter 1772 ldr(tmp2, arg_off); 1773 neg(tmp2, tmp2); 1774 // read the parameter from the local area 1775 ldr(tmp2, Address(rlocals, tmp2, Address::lsl(Interpreter::logStackElementSize))); 1776 1777 // profile the parameter 1778 profile_obj_type(tmp2, arg_type); 1779 1780 // go to next parameter 1781 subs(tmp1, tmp1, TypeStackSlotEntries::per_arg_count()); 1782 br(Assembler::GE, loop); 1783 1784 bind(profile_continue); 1785 } 1786 }