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