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 "oops/arrayOop.hpp" 31 #include "oops/markOop.hpp" 32 #include "oops/methodData.hpp" 33 #include "oops/method.hpp" 34 #include "prims/jvmtiExport.hpp" 35 #include "prims/jvmtiRedefineClassesTrace.hpp" 36 #include "prims/jvmtiThreadState.hpp" 37 #include "runtime/basicLock.hpp" 38 #include "runtime/biasedLocking.hpp" 39 #include "runtime/sharedRuntime.hpp" 40 #include "runtime/thread.inline.hpp" 41 42 43 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 tstw(rscratch1, JavaThread::popframe_pending_bit); 97 br(Assembler::EQ, L); 98 tstw(rscratch1, JavaThread::popframe_processing_bit); 99 br(Assembler::NE, L); 100 // Call Interpreter::remove_activation_preserving_args_entry() to get the 101 // address of the same-named entrypoint in the generated interpreter code. 102 call_VM_leaf(CAST_FROM_FN_PTR(address, Interpreter::remove_activation_preserving_args_entry)); 103 br(r0); 104 bind(L); 105 } 106 } 107 108 109 void InterpreterMacroAssembler::load_earlyret_value(TosState state) { 110 ldr(r2, Address(rthread, JavaThread::jvmti_thread_state_offset())); 111 const Address tos_addr(r2, JvmtiThreadState::earlyret_tos_offset()); 112 const Address oop_addr(r2, JvmtiThreadState::earlyret_oop_offset()); 113 const Address val_addr(r2, JvmtiThreadState::earlyret_value_offset()); 114 switch (state) { 115 case atos: ldr(r0, oop_addr); 116 str(zr, oop_addr); 117 verify_oop(r0, state); break; 118 case ltos: ldr(r0, val_addr); break; 119 case btos: // fall through 120 case ztos: // fall through 121 case ctos: // fall through 122 case stos: // fall through 123 case itos: ldrw(r0, val_addr); break; 124 case ftos: ldrs(v0, val_addr); break; 125 case dtos: ldrd(v0, val_addr); break; 126 case vtos: /* nothing to do */ break; 127 default : ShouldNotReachHere(); 128 } 129 // Clean up tos value in the thread object 130 movw(rscratch1, (int) ilgl); 131 strw(rscratch1, tos_addr); 132 strw(zr, val_addr); 133 } 134 135 136 void InterpreterMacroAssembler::check_and_handle_earlyret(Register java_thread) { 137 if (JvmtiExport::can_force_early_return()) { 138 Label L; 139 ldr(rscratch1, Address(rthread, JavaThread::jvmti_thread_state_offset())); 140 cbz(rscratch1, L); // if (thread->jvmti_thread_state() == NULL) exit; 141 142 // Initiate earlyret handling only if it is not already being processed. 143 // If the flag has the earlyret_processing bit set, it means that this code 144 // is called *during* earlyret handling - we don't want to reenter. 145 ldrw(rscratch1, Address(rscratch1, JvmtiThreadState::earlyret_state_offset())); 146 cmpw(rscratch1, JvmtiThreadState::earlyret_pending); 147 br(Assembler::NE, L); 148 149 // Call Interpreter::remove_activation_early_entry() to get the address of the 150 // same-named entrypoint in the generated interpreter code. 151 ldr(rscratch1, Address(rthread, JavaThread::jvmti_thread_state_offset())); 152 ldrw(rscratch1, Address(rscratch1, JvmtiThreadState::earlyret_tos_offset())); 153 call_VM_leaf(CAST_FROM_FN_PTR(address, Interpreter::remove_activation_early_entry), rscratch1); 154 br(r0); 155 bind(L); 156 } 157 } 158 159 void InterpreterMacroAssembler::get_unsigned_2_byte_index_at_bcp( 160 Register reg, 161 int bcp_offset) { 162 assert(bcp_offset >= 0, "bcp is still pointing to start of bytecode"); 163 ldrh(reg, Address(rbcp, bcp_offset)); 164 rev16(reg, reg); 165 } 166 167 void InterpreterMacroAssembler::get_dispatch() { 168 unsigned long offset; 169 adrp(rdispatch, ExternalAddress((address)Interpreter::dispatch_table()), offset); 170 lea(rdispatch, Address(rdispatch, offset)); 171 } 172 173 void InterpreterMacroAssembler::get_cache_index_at_bcp(Register index, 174 int bcp_offset, 175 size_t index_size) { 176 assert(bcp_offset > 0, "bcp is still pointing to start of bytecode"); 177 if (index_size == sizeof(u2)) { 178 load_unsigned_short(index, Address(rbcp, bcp_offset)); 179 } else if (index_size == sizeof(u4)) { 180 // assert(EnableInvokeDynamic, "giant index used only for JSR 292"); 181 ldrw(index, Address(rbcp, bcp_offset)); 182 // Check if the secondary index definition is still ~x, otherwise 183 // we have to change the following assembler code to calculate the 184 // plain index. 185 assert(ConstantPool::decode_invokedynamic_index(~123) == 123, "else change next line"); 186 eonw(index, index, zr); // convert to plain index 187 } else if (index_size == sizeof(u1)) { 188 load_unsigned_byte(index, Address(rbcp, bcp_offset)); 189 } else { 190 ShouldNotReachHere(); 191 } 192 } 193 194 // Return 195 // Rindex: index into constant pool 196 // Rcache: address of cache entry - ConstantPoolCache::base_offset() 197 // 198 // A caller must add ConstantPoolCache::base_offset() to Rcache to get 199 // the true address of the cache entry. 200 // 201 void InterpreterMacroAssembler::get_cache_and_index_at_bcp(Register cache, 202 Register index, 203 int bcp_offset, 204 size_t index_size) { 205 assert_different_registers(cache, index); 206 assert_different_registers(cache, rcpool); 207 get_cache_index_at_bcp(index, bcp_offset, index_size); 208 assert(sizeof(ConstantPoolCacheEntry) == 4 * wordSize, "adjust code below"); 209 // convert from field index to ConstantPoolCacheEntry 210 // aarch64 already has the cache in rcpool so there is no need to 211 // install it in cache. instead we pre-add the indexed offset to 212 // rcpool and return it in cache. All clients of this method need to 213 // be modified accordingly. 214 add(cache, rcpool, index, Assembler::LSL, 5); 215 } 216 217 218 void InterpreterMacroAssembler::get_cache_and_index_and_bytecode_at_bcp(Register cache, 219 Register index, 220 Register bytecode, 221 int byte_no, 222 int bcp_offset, 223 size_t index_size) { 224 get_cache_and_index_at_bcp(cache, index, bcp_offset, index_size); 225 // We use a 32-bit load here since the layout of 64-bit words on 226 // little-endian machines allow us that. 227 // n.b. unlike x86 cache already includes the index offset 228 lea(bytecode, Address(cache, 229 ConstantPoolCache::base_offset() 230 + ConstantPoolCacheEntry::indices_offset())); 231 ldarw(bytecode, bytecode); 232 const int shift_count = (1 + byte_no) * BitsPerByte; 233 ubfx(bytecode, bytecode, shift_count, BitsPerByte); 234 } 235 236 void InterpreterMacroAssembler::get_cache_entry_pointer_at_bcp(Register cache, 237 Register tmp, 238 int bcp_offset, 239 size_t index_size) { 240 assert(cache != tmp, "must use different register"); 241 get_cache_index_at_bcp(tmp, bcp_offset, index_size); 242 assert(sizeof(ConstantPoolCacheEntry) == 4 * wordSize, "adjust code below"); 243 // convert from field index to ConstantPoolCacheEntry index 244 // and from word offset to byte offset 245 assert(exact_log2(in_bytes(ConstantPoolCacheEntry::size_in_bytes())) == 2 + LogBytesPerWord, "else change next line"); 246 ldr(cache, Address(rfp, frame::interpreter_frame_cache_offset * wordSize)); 247 // skip past the header 248 add(cache, cache, in_bytes(ConstantPoolCache::base_offset())); 249 add(cache, cache, tmp, Assembler::LSL, 2 + LogBytesPerWord); // construct pointer to cache entry 250 } 251 252 void InterpreterMacroAssembler::get_method_counters(Register method, 253 Register mcs, Label& skip) { 254 Label has_counters; 255 ldr(mcs, Address(method, Method::method_counters_offset())); 256 cbnz(mcs, has_counters); 257 call_VM(noreg, CAST_FROM_FN_PTR(address, 258 InterpreterRuntime::build_method_counters), method); 259 ldr(mcs, Address(method, Method::method_counters_offset())); 260 cbz(mcs, skip); // No MethodCounters allocated, OutOfMemory 261 bind(has_counters); 262 } 263 264 // Load object from cpool->resolved_references(index) 265 void InterpreterMacroAssembler::load_resolved_reference_at_index( 266 Register result, Register index) { 267 assert_different_registers(result, index); 268 // convert from field index to resolved_references() index and from 269 // word index to byte offset. Since this is a java object, it can be compressed 270 Register tmp = index; // reuse 271 lslw(tmp, tmp, LogBytesPerHeapOop); 272 273 get_constant_pool(result); 274 // load pointer for resolved_references[] objArray 275 ldr(result, Address(result, ConstantPool::resolved_references_offset_in_bytes())); 276 // JNIHandles::resolve(obj); 277 ldr(result, Address(result, 0)); 278 // Add in the index 279 add(result, result, tmp); 280 load_heap_oop(result, Address(result, arrayOopDesc::base_offset_in_bytes(T_OBJECT))); 281 } 282 283 // Generate a subtype check: branch to ok_is_subtype if sub_klass is a 284 // subtype of super_klass. 285 // 286 // Args: 287 // r0: superklass 288 // Rsub_klass: subklass 289 // 290 // Kills: 291 // r2, r5 292 void InterpreterMacroAssembler::gen_subtype_check(Register Rsub_klass, 293 Label& ok_is_subtype) { 294 assert(Rsub_klass != r0, "r0 holds superklass"); 295 assert(Rsub_klass != r2, "r2 holds 2ndary super array length"); 296 assert(Rsub_klass != r5, "r5 holds 2ndary super array scan ptr"); 297 298 // Profile the not-null value's klass. 299 profile_typecheck(r2, Rsub_klass, r5); // blows r2, reloads r5 300 301 // Do the check. 302 check_klass_subtype(Rsub_klass, r0, r2, ok_is_subtype); // blows r2 303 304 // Profile the failure of the check. 305 profile_typecheck_failed(r2); // blows r2 306 } 307 308 // Java Expression Stack 309 310 void InterpreterMacroAssembler::pop_ptr(Register r) { 311 ldr(r, post(esp, wordSize)); 312 } 313 314 void InterpreterMacroAssembler::pop_i(Register r) { 315 ldrw(r, post(esp, wordSize)); 316 } 317 318 void InterpreterMacroAssembler::pop_l(Register r) { 319 ldr(r, post(esp, 2 * Interpreter::stackElementSize)); 320 } 321 322 void InterpreterMacroAssembler::push_ptr(Register r) { 323 str(r, pre(esp, -wordSize)); 324 } 325 326 void InterpreterMacroAssembler::push_i(Register r) { 327 str(r, pre(esp, -wordSize)); 328 } 329 330 void InterpreterMacroAssembler::push_l(Register r) { 331 str(r, pre(esp, 2 * -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 // interp_only is an int, on little endian it is sufficient to test the byte only 412 // Is a cmpl faster? 413 ldr(rscratch1, Address(rthread, JavaThread::interp_only_mode_offset())); 414 cbz(rscratch1, run_compiled_code); 415 ldr(rscratch1, Address(method, Method::interpreter_entry_offset())); 416 br(rscratch1); 417 bind(run_compiled_code); 418 } 419 420 ldr(rscratch1, Address(method, Method::from_interpreted_offset())); 421 br(rscratch1); 422 } 423 424 // The following two routines provide a hook so that an implementation 425 // can schedule the dispatch in two parts. amd64 does not do this. 426 void InterpreterMacroAssembler::dispatch_prolog(TosState state, int step) { 427 } 428 429 void InterpreterMacroAssembler::dispatch_epilog(TosState state, int step) { 430 dispatch_next(state, step); 431 } 432 433 void InterpreterMacroAssembler::dispatch_base(TosState state, 434 address* table, 435 bool verifyoop) { 436 if (VerifyActivationFrameSize) { 437 Unimplemented(); 438 } 439 if (verifyoop) { 440 verify_oop(r0, state); 441 } 442 if (table == Interpreter::dispatch_table(state)) { 443 addw(rscratch2, rscratch1, Interpreter::distance_from_dispatch_table(state)); 444 ldr(rscratch2, Address(rdispatch, rscratch2, Address::uxtw(3))); 445 } else { 446 mov(rscratch2, (address)table); 447 ldr(rscratch2, Address(rscratch2, rscratch1, Address::uxtw(3))); 448 } 449 br(rscratch2); 450 } 451 452 void InterpreterMacroAssembler::dispatch_only(TosState state) { 453 dispatch_base(state, Interpreter::dispatch_table(state)); 454 } 455 456 void InterpreterMacroAssembler::dispatch_only_normal(TosState state) { 457 dispatch_base(state, Interpreter::normal_table(state)); 458 } 459 460 void InterpreterMacroAssembler::dispatch_only_noverify(TosState state) { 461 dispatch_base(state, Interpreter::normal_table(state), false); 462 } 463 464 465 void InterpreterMacroAssembler::dispatch_next(TosState state, int step) { 466 // load next bytecode 467 ldrb(rscratch1, Address(pre(rbcp, step))); 468 dispatch_base(state, Interpreter::dispatch_table(state)); 469 } 470 471 void InterpreterMacroAssembler::dispatch_via(TosState state, address* table) { 472 // load current bytecode 473 ldrb(rscratch1, Address(rbcp, 0)); 474 dispatch_base(state, table); 475 } 476 477 // remove activation 478 // 479 // Unlock the receiver if this is a synchronized method. 480 // Unlock any Java monitors from syncronized blocks. 481 // Remove the activation from the stack. 482 // 483 // If there are locked Java monitors 484 // If throw_monitor_exception 485 // throws IllegalMonitorStateException 486 // Else if install_monitor_exception 487 // installs IllegalMonitorStateException 488 // Else 489 // no error processing 490 void InterpreterMacroAssembler::remove_activation( 491 TosState state, 492 bool throw_monitor_exception, 493 bool install_monitor_exception, 494 bool notify_jvmdi) { 495 // Note: Registers r3 xmm0 may be in use for the 496 // result check if synchronized method 497 Label unlocked, unlock, no_unlock; 498 499 // get the value of _do_not_unlock_if_synchronized into r3 500 const Address do_not_unlock_if_synchronized(rthread, 501 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset())); 502 ldrb(r3, do_not_unlock_if_synchronized); 503 strb(zr, do_not_unlock_if_synchronized); // reset the flag 504 505 // get method access flags 506 ldr(r1, Address(rfp, frame::interpreter_frame_method_offset * wordSize)); 507 ldr(r2, Address(r1, Method::access_flags_offset())); 508 tst(r2, JVM_ACC_SYNCHRONIZED); 509 br(Assembler::EQ, unlocked); 510 511 // Don't unlock anything if the _do_not_unlock_if_synchronized flag 512 // is set. 513 cbnz(r3, no_unlock); 514 515 // unlock monitor 516 push(state); // save result 517 518 // BasicObjectLock will be first in list, since this is a 519 // synchronized method. However, need to check that the object has 520 // not been unlocked by an explicit monitorexit bytecode. 521 const Address monitor(rfp, frame::interpreter_frame_initial_sp_offset * 522 wordSize - (int) sizeof(BasicObjectLock)); 523 // We use c_rarg1 so that if we go slow path it will be the correct 524 // register for unlock_object to pass to VM directly 525 lea(c_rarg1, monitor); // address of first monitor 526 527 ldr(r0, Address(c_rarg1, BasicObjectLock::obj_offset_in_bytes())); 528 cbnz(r0, unlock); 529 530 pop(state); 531 if (throw_monitor_exception) { 532 // Entry already unlocked, need to throw exception 533 call_VM(noreg, CAST_FROM_FN_PTR(address, 534 InterpreterRuntime::throw_illegal_monitor_state_exception)); 535 should_not_reach_here(); 536 } else { 537 // Monitor already unlocked during a stack unroll. If requested, 538 // install an illegal_monitor_state_exception. Continue with 539 // stack unrolling. 540 if (install_monitor_exception) { 541 call_VM(noreg, CAST_FROM_FN_PTR(address, 542 InterpreterRuntime::new_illegal_monitor_state_exception)); 543 } 544 b(unlocked); 545 } 546 547 bind(unlock); 548 unlock_object(c_rarg1); 549 pop(state); 550 551 // Check that for block-structured locking (i.e., that all locked 552 // objects has been unlocked) 553 bind(unlocked); 554 555 // r0: Might contain return value 556 557 // Check that all monitors are unlocked 558 { 559 Label loop, exception, entry, restart; 560 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize; 561 const Address monitor_block_top( 562 rfp, frame::interpreter_frame_monitor_block_top_offset * wordSize); 563 const Address monitor_block_bot( 564 rfp, frame::interpreter_frame_initial_sp_offset * wordSize); 565 566 bind(restart); 567 // We use c_rarg1 so that if we go slow path it will be the correct 568 // register for unlock_object to pass to VM directly 569 ldr(c_rarg1, monitor_block_top); // points to current entry, starting 570 // with top-most entry 571 lea(r19, monitor_block_bot); // points to word before bottom of 572 // monitor block 573 b(entry); 574 575 // Entry already locked, need to throw exception 576 bind(exception); 577 578 if (throw_monitor_exception) { 579 // Throw exception 580 MacroAssembler::call_VM(noreg, 581 CAST_FROM_FN_PTR(address, InterpreterRuntime:: 582 throw_illegal_monitor_state_exception)); 583 should_not_reach_here(); 584 } else { 585 // Stack unrolling. Unlock object and install illegal_monitor_exception. 586 // Unlock does not block, so don't have to worry about the frame. 587 // We don't have to preserve c_rarg1 since we are going to throw an exception. 588 589 push(state); 590 unlock_object(c_rarg1); 591 pop(state); 592 593 if (install_monitor_exception) { 594 call_VM(noreg, CAST_FROM_FN_PTR(address, 595 InterpreterRuntime:: 596 new_illegal_monitor_state_exception)); 597 } 598 599 b(restart); 600 } 601 602 bind(loop); 603 // check if current entry is used 604 ldr(rscratch1, Address(c_rarg1, BasicObjectLock::obj_offset_in_bytes())); 605 cbnz(rscratch1, exception); 606 607 add(c_rarg1, c_rarg1, entry_size); // otherwise advance to next entry 608 bind(entry); 609 cmp(c_rarg1, r19); // check if bottom reached 610 br(Assembler::NE, loop); // if not at bottom then check this entry 611 } 612 613 bind(no_unlock); 614 615 // jvmti support 616 if (notify_jvmdi) { 617 notify_method_exit(state, NotifyJVMTI); // preserve TOSCA 618 } else { 619 notify_method_exit(state, SkipNotifyJVMTI); // preserve TOSCA 620 } 621 622 // remove activation 623 // get sender esp 624 ldr(esp, 625 Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); 626 // remove frame anchor 627 leave(); 628 // If we're returning to interpreted code we will shortly be 629 // adjusting SP to allow some space for ESP. If we're returning to 630 // compiled code the saved sender SP was saved in sender_sp, so this 631 // restores it. 632 andr(sp, esp, -16); 633 } 634 635 // Lock object 636 // 637 // Args: 638 // c_rarg1: BasicObjectLock to be used for locking 639 // 640 // Kills: 641 // r0 642 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, .. (param regs) 643 // rscratch1, rscratch2 (scratch regs) 644 void InterpreterMacroAssembler::lock_object(Register lock_reg) 645 { 646 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be c_rarg1"); 647 if (UseHeavyMonitors) { 648 call_VM(noreg, 649 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter), 650 lock_reg); 651 } else { 652 Label done; 653 654 const Register swap_reg = r0; 655 const Register tmp = c_rarg2; 656 const Register obj_reg = c_rarg3; // Will contain the oop 657 658 const int obj_offset = BasicObjectLock::obj_offset_in_bytes(); 659 const int lock_offset = BasicObjectLock::lock_offset_in_bytes (); 660 const int mark_offset = lock_offset + 661 BasicLock::displaced_header_offset_in_bytes(); 662 663 Label slow_case; 664 665 // Load object pointer into obj_reg %c_rarg3 666 ldr(obj_reg, Address(lock_reg, obj_offset)); 667 668 if (UseBiasedLocking) { 669 biased_locking_enter(lock_reg, obj_reg, swap_reg, tmp, false, done, &slow_case); 670 } 671 672 // Load (object->mark() | 1) into swap_reg 673 ldr(rscratch1, Address(obj_reg, 0)); 674 orr(swap_reg, rscratch1, 1); 675 676 // Save (object->mark() | 1) into BasicLock's displaced header 677 str(swap_reg, Address(lock_reg, mark_offset)); 678 679 assert(lock_offset == 0, 680 "displached header must be first word in BasicObjectLock"); 681 682 Label fail; 683 if (PrintBiasedLockingStatistics) { 684 Label fast; 685 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, fast, &fail); 686 bind(fast); 687 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()), 688 rscratch2, rscratch1, tmp); 689 b(done); 690 bind(fail); 691 } else { 692 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL); 693 } 694 695 // Test if the oopMark is an obvious stack pointer, i.e., 696 // 1) (mark & 7) == 0, and 697 // 2) rsp <= mark < mark + os::pagesize() 698 // 699 // These 3 tests can be done by evaluating the following 700 // expression: ((mark - rsp) & (7 - os::vm_page_size())), 701 // assuming both stack pointer and pagesize have their 702 // least significant 3 bits clear. 703 // NOTE: the oopMark is in swap_reg %r0 as the result of cmpxchg 704 // NOTE2: aarch64 does not like to subtract sp from rn so take a 705 // copy 706 mov(rscratch1, sp); 707 sub(swap_reg, swap_reg, rscratch1); 708 ands(swap_reg, swap_reg, (unsigned long)(7 - os::vm_page_size())); 709 710 // Save the test result, for recursive case, the result is zero 711 str(swap_reg, Address(lock_reg, mark_offset)); 712 713 if (PrintBiasedLockingStatistics) { 714 br(Assembler::NE, slow_case); 715 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()), 716 rscratch2, rscratch1, tmp); 717 } 718 br(Assembler::EQ, done); 719 720 bind(slow_case); 721 722 // Call the runtime routine for slow case 723 call_VM(noreg, 724 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter), 725 lock_reg); 726 727 bind(done); 728 } 729 } 730 731 732 // Unlocks an object. Used in monitorexit bytecode and 733 // remove_activation. Throws an IllegalMonitorException if object is 734 // not locked by current thread. 735 // 736 // Args: 737 // c_rarg1: BasicObjectLock for lock 738 // 739 // Kills: 740 // r0 741 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, ... (param regs) 742 // rscratch1, rscratch2 (scratch regs) 743 void InterpreterMacroAssembler::unlock_object(Register lock_reg) 744 { 745 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be rarg1"); 746 747 if (UseHeavyMonitors) { 748 call_VM(noreg, 749 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit), 750 lock_reg); 751 } else { 752 Label done; 753 754 const Register swap_reg = r0; 755 const Register header_reg = c_rarg2; // Will contain the old oopMark 756 const Register obj_reg = c_rarg3; // Will contain the oop 757 758 save_bcp(); // Save in case of exception 759 760 // Convert from BasicObjectLock structure to object and BasicLock 761 // structure Store the BasicLock address into %r0 762 lea(swap_reg, Address(lock_reg, BasicObjectLock::lock_offset_in_bytes())); 763 764 // Load oop into obj_reg(%c_rarg3) 765 ldr(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); 766 767 // Free entry 768 str(zr, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); 769 770 if (UseBiasedLocking) { 771 biased_locking_exit(obj_reg, header_reg, done); 772 } 773 774 // Load the old header from BasicLock structure 775 ldr(header_reg, Address(swap_reg, 776 BasicLock::displaced_header_offset_in_bytes())); 777 778 // Test for recursion 779 cbz(header_reg, done); 780 781 // Atomic swap back the old header 782 cmpxchgptr(swap_reg, header_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL); 783 784 // Call the runtime routine for slow case. 785 str(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); // restore obj 786 call_VM(noreg, 787 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit), 788 lock_reg); 789 790 bind(done); 791 792 restore_bcp(); 793 } 794 } 795 796 void InterpreterMacroAssembler::test_method_data_pointer(Register mdp, 797 Label& zero_continue) { 798 assert(ProfileInterpreter, "must be profiling interpreter"); 799 ldr(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 800 cbz(mdp, zero_continue); 801 } 802 803 // Set the method data pointer for the current bcp. 804 void InterpreterMacroAssembler::set_method_data_pointer_for_bcp() { 805 assert(ProfileInterpreter, "must be profiling interpreter"); 806 Label set_mdp; 807 stp(r0, r1, Address(pre(sp, -2 * wordSize))); 808 809 // Test MDO to avoid the call if it is NULL. 810 ldr(r0, Address(rmethod, in_bytes(Method::method_data_offset()))); 811 cbz(r0, set_mdp); 812 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::bcp_to_di), rmethod, rbcp); 813 // r0: mdi 814 // mdo is guaranteed to be non-zero here, we checked for it before the call. 815 ldr(r1, Address(rmethod, in_bytes(Method::method_data_offset()))); 816 lea(r1, Address(r1, in_bytes(MethodData::data_offset()))); 817 add(r0, r1, r0); 818 str(r0, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 819 bind(set_mdp); 820 ldp(r0, r1, Address(post(sp, 2 * wordSize))); 821 } 822 823 void InterpreterMacroAssembler::verify_method_data_pointer() { 824 assert(ProfileInterpreter, "must be profiling interpreter"); 825 #ifdef ASSERT 826 Label verify_continue; 827 stp(r0, r1, Address(pre(sp, -2 * wordSize))); 828 stp(r2, r3, Address(pre(sp, -2 * wordSize))); 829 test_method_data_pointer(r3, verify_continue); // If mdp is zero, continue 830 get_method(r1); 831 832 // If the mdp is valid, it will point to a DataLayout header which is 833 // consistent with the bcp. The converse is highly probable also. 834 ldrsh(r2, Address(r3, in_bytes(DataLayout::bci_offset()))); 835 ldr(rscratch1, Address(r1, Method::const_offset())); 836 add(r2, r2, rscratch1, Assembler::LSL); 837 lea(r2, Address(r2, ConstMethod::codes_offset())); 838 cmp(r2, rbcp); 839 br(Assembler::EQ, verify_continue); 840 // r1: method 841 // rbcp: bcp // rbcp == 22 842 // r3: mdp 843 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::verify_mdp), 844 r1, rbcp, r3); 845 bind(verify_continue); 846 ldp(r2, r3, Address(post(sp, 2 * wordSize))); 847 ldp(r0, r1, Address(post(sp, 2 * wordSize))); 848 #endif // ASSERT 849 } 850 851 852 void InterpreterMacroAssembler::set_mdp_data_at(Register mdp_in, 853 int constant, 854 Register value) { 855 assert(ProfileInterpreter, "must be profiling interpreter"); 856 Address data(mdp_in, constant); 857 str(value, data); 858 } 859 860 861 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in, 862 int constant, 863 bool decrement) { 864 increment_mdp_data_at(mdp_in, noreg, constant, decrement); 865 } 866 867 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in, 868 Register reg, 869 int constant, 870 bool decrement) { 871 assert(ProfileInterpreter, "must be profiling interpreter"); 872 // %%% this does 64bit counters at best it is wasting space 873 // at worst it is a rare bug when counters overflow 874 875 assert_different_registers(rscratch2, rscratch1, mdp_in, reg); 876 877 Address addr1(mdp_in, constant); 878 Address addr2(rscratch2, reg, Address::lsl(0)); 879 Address &addr = addr1; 880 if (reg != noreg) { 881 lea(rscratch2, addr1); 882 addr = addr2; 883 } 884 885 if (decrement) { 886 // Decrement the register. Set condition codes. 887 // Intel does this 888 // addptr(data, (int32_t) -DataLayout::counter_increment); 889 // If the decrement causes the counter to overflow, stay negative 890 // Label L; 891 // jcc(Assembler::negative, L); 892 // addptr(data, (int32_t) DataLayout::counter_increment); 893 // so we do this 894 ldr(rscratch1, addr); 895 subs(rscratch1, rscratch1, (unsigned)DataLayout::counter_increment); 896 Label L; 897 br(Assembler::LO, L); // skip store if counter underflow 898 str(rscratch1, addr); 899 bind(L); 900 } else { 901 assert(DataLayout::counter_increment == 1, 902 "flow-free idiom only works with 1"); 903 // Intel does this 904 // Increment the register. Set carry flag. 905 // addptr(data, DataLayout::counter_increment); 906 // If the increment causes the counter to overflow, pull back by 1. 907 // sbbptr(data, (int32_t)0); 908 // so we do this 909 ldr(rscratch1, addr); 910 adds(rscratch1, rscratch1, DataLayout::counter_increment); 911 Label L; 912 br(Assembler::CS, L); // skip store if counter overflow 913 str(rscratch1, addr); 914 bind(L); 915 } 916 } 917 918 void InterpreterMacroAssembler::set_mdp_flag_at(Register mdp_in, 919 int flag_byte_constant) { 920 assert(ProfileInterpreter, "must be profiling interpreter"); 921 int header_offset = in_bytes(DataLayout::header_offset()); 922 int header_bits = DataLayout::flag_mask_to_header_mask(flag_byte_constant); 923 // Set the flag 924 ldr(rscratch1, Address(mdp_in, header_offset)); 925 orr(rscratch1, rscratch1, header_bits); 926 str(rscratch1, Address(mdp_in, header_offset)); 927 } 928 929 930 void InterpreterMacroAssembler::test_mdp_data_at(Register mdp_in, 931 int offset, 932 Register value, 933 Register test_value_out, 934 Label& not_equal_continue) { 935 assert(ProfileInterpreter, "must be profiling interpreter"); 936 if (test_value_out == noreg) { 937 ldr(rscratch1, Address(mdp_in, offset)); 938 cmp(value, rscratch1); 939 } else { 940 // Put the test value into a register, so caller can use it: 941 ldr(test_value_out, Address(mdp_in, offset)); 942 cmp(value, test_value_out); 943 } 944 br(Assembler::NE, not_equal_continue); 945 } 946 947 948 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in, 949 int offset_of_disp) { 950 assert(ProfileInterpreter, "must be profiling interpreter"); 951 ldr(rscratch1, Address(mdp_in, offset_of_disp)); 952 add(mdp_in, mdp_in, rscratch1, LSL); 953 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 954 } 955 956 957 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in, 958 Register reg, 959 int offset_of_disp) { 960 assert(ProfileInterpreter, "must be profiling interpreter"); 961 lea(rscratch1, Address(mdp_in, offset_of_disp)); 962 ldr(rscratch1, Address(rscratch1, reg, Address::lsl(0))); 963 add(mdp_in, mdp_in, rscratch1, LSL); 964 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 965 } 966 967 968 void InterpreterMacroAssembler::update_mdp_by_constant(Register mdp_in, 969 int constant) { 970 assert(ProfileInterpreter, "must be profiling interpreter"); 971 add(mdp_in, mdp_in, (unsigned)constant); 972 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 973 } 974 975 976 void InterpreterMacroAssembler::update_mdp_for_ret(Register return_bci) { 977 assert(ProfileInterpreter, "must be profiling interpreter"); 978 // save/restore across call_VM 979 stp(zr, return_bci, Address(pre(sp, -2 * wordSize))); 980 call_VM(noreg, 981 CAST_FROM_FN_PTR(address, InterpreterRuntime::update_mdp_for_ret), 982 return_bci); 983 ldp(zr, return_bci, Address(post(sp, 2 * wordSize))); 984 } 985 986 987 void InterpreterMacroAssembler::profile_taken_branch(Register mdp, 988 Register bumped_count) { 989 if (ProfileInterpreter) { 990 Label profile_continue; 991 992 // If no method data exists, go to profile_continue. 993 // Otherwise, assign to mdp 994 test_method_data_pointer(mdp, profile_continue); 995 996 // We are taking a branch. Increment the taken count. 997 // We inline increment_mdp_data_at to return bumped_count in a register 998 //increment_mdp_data_at(mdp, in_bytes(JumpData::taken_offset())); 999 Address data(mdp, in_bytes(JumpData::taken_offset())); 1000 ldr(bumped_count, data); 1001 assert(DataLayout::counter_increment == 1, 1002 "flow-free idiom only works with 1"); 1003 // Intel does this to catch overflow 1004 // addptr(bumped_count, DataLayout::counter_increment); 1005 // sbbptr(bumped_count, 0); 1006 // so we do this 1007 adds(bumped_count, bumped_count, DataLayout::counter_increment); 1008 Label L; 1009 br(Assembler::CS, L); // skip store if counter overflow 1010 str(bumped_count, data); 1011 bind(L); 1012 // The method data pointer needs to be updated to reflect the new target. 1013 update_mdp_by_offset(mdp, in_bytes(JumpData::displacement_offset())); 1014 bind(profile_continue); 1015 } 1016 } 1017 1018 1019 void InterpreterMacroAssembler::profile_not_taken_branch(Register mdp) { 1020 if (ProfileInterpreter) { 1021 Label profile_continue; 1022 1023 // If no method data exists, go to profile_continue. 1024 test_method_data_pointer(mdp, profile_continue); 1025 1026 // We are taking a branch. Increment the not taken count. 1027 increment_mdp_data_at(mdp, in_bytes(BranchData::not_taken_offset())); 1028 1029 // The method data pointer needs to be updated to correspond to 1030 // the next bytecode 1031 update_mdp_by_constant(mdp, in_bytes(BranchData::branch_data_size())); 1032 bind(profile_continue); 1033 } 1034 } 1035 1036 1037 void InterpreterMacroAssembler::profile_call(Register mdp) { 1038 if (ProfileInterpreter) { 1039 Label profile_continue; 1040 1041 // If no method data exists, go to profile_continue. 1042 test_method_data_pointer(mdp, profile_continue); 1043 1044 // We are making a call. Increment the count. 1045 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1046 1047 // The method data pointer needs to be updated to reflect the new target. 1048 update_mdp_by_constant(mdp, in_bytes(CounterData::counter_data_size())); 1049 bind(profile_continue); 1050 } 1051 } 1052 1053 void InterpreterMacroAssembler::profile_final_call(Register mdp) { 1054 if (ProfileInterpreter) { 1055 Label profile_continue; 1056 1057 // If no method data exists, go to profile_continue. 1058 test_method_data_pointer(mdp, profile_continue); 1059 1060 // We are making a call. Increment the count. 1061 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1062 1063 // The method data pointer needs to be updated to reflect the new target. 1064 update_mdp_by_constant(mdp, 1065 in_bytes(VirtualCallData:: 1066 virtual_call_data_size())); 1067 bind(profile_continue); 1068 } 1069 } 1070 1071 1072 void InterpreterMacroAssembler::profile_virtual_call(Register receiver, 1073 Register mdp, 1074 Register reg2, 1075 bool receiver_can_be_null) { 1076 if (ProfileInterpreter) { 1077 Label profile_continue; 1078 1079 // If no method data exists, go to profile_continue. 1080 test_method_data_pointer(mdp, profile_continue); 1081 1082 Label skip_receiver_profile; 1083 if (receiver_can_be_null) { 1084 Label not_null; 1085 // We are making a call. Increment the count for null receiver. 1086 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1087 b(skip_receiver_profile); 1088 bind(not_null); 1089 } 1090 1091 // Record the receiver type. 1092 record_klass_in_profile(receiver, mdp, reg2, true); 1093 bind(skip_receiver_profile); 1094 1095 // The method data pointer needs to be updated to reflect the new target. 1096 #if INCLUDE_JVMCI 1097 if (MethodProfileWidth == 0) { 1098 update_mdp_by_constant(mdp, in_bytes(VirtualCallData::virtual_call_data_size())); 1099 } 1100 #else // INCLUDE_JVMCI 1101 update_mdp_by_constant(mdp, 1102 in_bytes(VirtualCallData:: 1103 virtual_call_data_size())); 1104 #endif // INCLUDE_JVMCI 1105 bind(profile_continue); 1106 } 1107 } 1108 1109 #if INCLUDE_JVMCI 1110 void InterpreterMacroAssembler::profile_called_method(Register method, Register mdp, Register reg2) { 1111 assert_different_registers(method, mdp, reg2); 1112 if (ProfileInterpreter && MethodProfileWidth > 0) { 1113 Label profile_continue; 1114 1115 // If no method data exists, go to profile_continue. 1116 test_method_data_pointer(mdp, profile_continue); 1117 1118 Label done; 1119 record_item_in_profile_helper(method, mdp, reg2, 0, done, MethodProfileWidth, 1120 &VirtualCallData::method_offset, &VirtualCallData::method_count_offset, in_bytes(VirtualCallData::nonprofiled_receiver_count_offset())); 1121 bind(done); 1122 1123 update_mdp_by_constant(mdp, in_bytes(VirtualCallData::virtual_call_data_size())); 1124 bind(profile_continue); 1125 } 1126 } 1127 #endif // INCLUDE_JVMCI 1128 1129 // This routine creates a state machine for updating the multi-row 1130 // type profile at a virtual call site (or other type-sensitive bytecode). 1131 // The machine visits each row (of receiver/count) until the receiver type 1132 // is found, or until it runs out of rows. At the same time, it remembers 1133 // the location of the first empty row. (An empty row records null for its 1134 // receiver, and can be allocated for a newly-observed receiver type.) 1135 // Because there are two degrees of freedom in the state, a simple linear 1136 // search will not work; it must be a decision tree. Hence this helper 1137 // function is recursive, to generate the required tree structured code. 1138 // It's the interpreter, so we are trading off code space for speed. 1139 // See below for example code. 1140 void InterpreterMacroAssembler::record_klass_in_profile_helper( 1141 Register receiver, Register mdp, 1142 Register reg2, int start_row, 1143 Label& done, bool is_virtual_call) { 1144 if (TypeProfileWidth == 0) { 1145 if (is_virtual_call) { 1146 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1147 } 1148 #if INCLUDE_JVMCI 1149 else if (EnableJVMCI) { 1150 increment_mdp_data_at(mdp, in_bytes(ReceiverTypeData::nonprofiled_receiver_count_offset())); 1151 } 1152 #endif // INCLUDE_JVMCI 1153 } else { 1154 int non_profiled_offset = -1; 1155 if (is_virtual_call) { 1156 non_profiled_offset = in_bytes(CounterData::count_offset()); 1157 } 1158 #if INCLUDE_JVMCI 1159 else if (EnableJVMCI) { 1160 non_profiled_offset = in_bytes(ReceiverTypeData::nonprofiled_receiver_count_offset()); 1161 } 1162 #endif // INCLUDE_JVMCI 1163 1164 record_item_in_profile_helper(receiver, mdp, reg2, 0, done, TypeProfileWidth, 1165 &VirtualCallData::receiver_offset, &VirtualCallData::receiver_count_offset, non_profiled_offset); 1166 } 1167 } 1168 1169 void InterpreterMacroAssembler::record_item_in_profile_helper(Register item, Register mdp, 1170 Register reg2, int start_row, Label& done, int total_rows, 1171 OffsetFunction item_offset_fn, OffsetFunction item_count_offset_fn, 1172 int non_profiled_offset) { 1173 int last_row = total_rows - 1; 1174 assert(start_row <= last_row, "must be work left to do"); 1175 // Test this row for both the item and for null. 1176 // Take any of three different outcomes: 1177 // 1. found item => increment count and goto done 1178 // 2. found null => keep looking for case 1, maybe allocate this cell 1179 // 3. found something else => keep looking for cases 1 and 2 1180 // Case 3 is handled by a recursive call. 1181 for (int row = start_row; row <= last_row; row++) { 1182 Label next_test; 1183 bool test_for_null_also = (row == start_row); 1184 1185 // See if the item is item[n]. 1186 int item_offset = in_bytes(item_offset_fn(row)); 1187 test_mdp_data_at(mdp, item_offset, item, 1188 (test_for_null_also ? reg2 : noreg), 1189 next_test); 1190 // (Reg2 now contains the item from the CallData.) 1191 1192 // The item is item[n]. Increment count[n]. 1193 int count_offset = in_bytes(item_count_offset_fn(row)); 1194 increment_mdp_data_at(mdp, count_offset); 1195 b(done); 1196 bind(next_test); 1197 1198 if (test_for_null_also) { 1199 Label found_null; 1200 // Failed the equality check on item[n]... Test for null. 1201 if (start_row == last_row) { 1202 // The only thing left to do is handle the null case. 1203 if (non_profiled_offset >= 0) { 1204 cbz(reg2, found_null); 1205 // Item did not match any saved item and there is no empty row for it. 1206 // Increment total counter to indicate polymorphic case. 1207 increment_mdp_data_at(mdp, non_profiled_offset); 1208 b(done); 1209 bind(found_null); 1210 } else { 1211 cbnz(reg2, done); 1212 } 1213 break; 1214 } 1215 // Since null is rare, make it be the branch-taken case. 1216 cbz(reg2, found_null); 1217 1218 // Put all the "Case 3" tests here. 1219 record_item_in_profile_helper(item, mdp, reg2, start_row + 1, done, total_rows, 1220 item_offset_fn, item_count_offset_fn, non_profiled_offset); 1221 1222 // Found a null. Keep searching for a matching item, 1223 // but remember that this is an empty (unused) slot. 1224 bind(found_null); 1225 } 1226 } 1227 1228 // In the fall-through case, we found no matching item, but we 1229 // observed the item[start_row] is NULL. 1230 1231 // Fill in the item field and increment the count. 1232 int item_offset = in_bytes(item_offset_fn(start_row)); 1233 set_mdp_data_at(mdp, item_offset, item); 1234 int count_offset = in_bytes(item_count_offset_fn(start_row)); 1235 mov(reg2, DataLayout::counter_increment); 1236 set_mdp_data_at(mdp, count_offset, reg2); 1237 if (start_row > 0) { 1238 b(done); 1239 } 1240 } 1241 1242 // Example state machine code for three profile rows: 1243 // // main copy of decision tree, rooted at row[1] 1244 // if (row[0].rec == rec) { row[0].incr(); goto done; } 1245 // if (row[0].rec != NULL) { 1246 // // inner copy of decision tree, rooted at row[1] 1247 // if (row[1].rec == rec) { row[1].incr(); goto done; } 1248 // if (row[1].rec != NULL) { 1249 // // degenerate decision tree, rooted at row[2] 1250 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1251 // if (row[2].rec != NULL) { count.incr(); goto done; } // overflow 1252 // row[2].init(rec); goto done; 1253 // } else { 1254 // // remember row[1] is empty 1255 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1256 // row[1].init(rec); goto done; 1257 // } 1258 // } else { 1259 // // remember row[0] is empty 1260 // if (row[1].rec == rec) { row[1].incr(); goto done; } 1261 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1262 // row[0].init(rec); goto done; 1263 // } 1264 // done: 1265 1266 void InterpreterMacroAssembler::record_klass_in_profile(Register receiver, 1267 Register mdp, Register reg2, 1268 bool is_virtual_call) { 1269 assert(ProfileInterpreter, "must be profiling"); 1270 Label done; 1271 1272 record_klass_in_profile_helper(receiver, mdp, reg2, 0, done, is_virtual_call); 1273 1274 bind (done); 1275 } 1276 1277 void InterpreterMacroAssembler::profile_ret(Register return_bci, 1278 Register mdp) { 1279 if (ProfileInterpreter) { 1280 Label profile_continue; 1281 uint row; 1282 1283 // If no method data exists, go to profile_continue. 1284 test_method_data_pointer(mdp, profile_continue); 1285 1286 // Update the total ret count. 1287 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1288 1289 for (row = 0; row < RetData::row_limit(); row++) { 1290 Label next_test; 1291 1292 // See if return_bci is equal to bci[n]: 1293 test_mdp_data_at(mdp, 1294 in_bytes(RetData::bci_offset(row)), 1295 return_bci, noreg, 1296 next_test); 1297 1298 // return_bci is equal to bci[n]. Increment the count. 1299 increment_mdp_data_at(mdp, in_bytes(RetData::bci_count_offset(row))); 1300 1301 // The method data pointer needs to be updated to reflect the new target. 1302 update_mdp_by_offset(mdp, 1303 in_bytes(RetData::bci_displacement_offset(row))); 1304 b(profile_continue); 1305 bind(next_test); 1306 } 1307 1308 update_mdp_for_ret(return_bci); 1309 1310 bind(profile_continue); 1311 } 1312 } 1313 1314 void InterpreterMacroAssembler::profile_null_seen(Register mdp) { 1315 if (ProfileInterpreter) { 1316 Label profile_continue; 1317 1318 // If no method data exists, go to profile_continue. 1319 test_method_data_pointer(mdp, profile_continue); 1320 1321 set_mdp_flag_at(mdp, BitData::null_seen_byte_constant()); 1322 1323 // The method data pointer needs to be updated. 1324 int mdp_delta = in_bytes(BitData::bit_data_size()); 1325 if (TypeProfileCasts) { 1326 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size()); 1327 } 1328 update_mdp_by_constant(mdp, mdp_delta); 1329 1330 bind(profile_continue); 1331 } 1332 } 1333 1334 void InterpreterMacroAssembler::profile_typecheck_failed(Register mdp) { 1335 if (ProfileInterpreter && TypeProfileCasts) { 1336 Label profile_continue; 1337 1338 // If no method data exists, go to profile_continue. 1339 test_method_data_pointer(mdp, profile_continue); 1340 1341 int count_offset = in_bytes(CounterData::count_offset()); 1342 // Back up the address, since we have already bumped the mdp. 1343 count_offset -= in_bytes(VirtualCallData::virtual_call_data_size()); 1344 1345 // *Decrement* the counter. We expect to see zero or small negatives. 1346 increment_mdp_data_at(mdp, count_offset, true); 1347 1348 bind (profile_continue); 1349 } 1350 } 1351 1352 void InterpreterMacroAssembler::profile_typecheck(Register mdp, Register klass, Register reg2) { 1353 if (ProfileInterpreter) { 1354 Label profile_continue; 1355 1356 // If no method data exists, go to profile_continue. 1357 test_method_data_pointer(mdp, profile_continue); 1358 1359 // The method data pointer needs to be updated. 1360 int mdp_delta = in_bytes(BitData::bit_data_size()); 1361 if (TypeProfileCasts) { 1362 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size()); 1363 1364 // Record the object type. 1365 record_klass_in_profile(klass, mdp, reg2, false); 1366 } 1367 update_mdp_by_constant(mdp, mdp_delta); 1368 1369 bind(profile_continue); 1370 } 1371 } 1372 1373 void InterpreterMacroAssembler::profile_switch_default(Register mdp) { 1374 if (ProfileInterpreter) { 1375 Label profile_continue; 1376 1377 // If no method data exists, go to profile_continue. 1378 test_method_data_pointer(mdp, profile_continue); 1379 1380 // Update the default case count 1381 increment_mdp_data_at(mdp, 1382 in_bytes(MultiBranchData::default_count_offset())); 1383 1384 // The method data pointer needs to be updated. 1385 update_mdp_by_offset(mdp, 1386 in_bytes(MultiBranchData:: 1387 default_displacement_offset())); 1388 1389 bind(profile_continue); 1390 } 1391 } 1392 1393 void InterpreterMacroAssembler::profile_switch_case(Register index, 1394 Register mdp, 1395 Register reg2) { 1396 if (ProfileInterpreter) { 1397 Label profile_continue; 1398 1399 // If no method data exists, go to profile_continue. 1400 test_method_data_pointer(mdp, profile_continue); 1401 1402 // Build the base (index * per_case_size_in_bytes()) + 1403 // case_array_offset_in_bytes() 1404 movw(reg2, in_bytes(MultiBranchData::per_case_size())); 1405 movw(rscratch1, in_bytes(MultiBranchData::case_array_offset())); 1406 Assembler::maddw(index, index, reg2, rscratch1); 1407 1408 // Update the case count 1409 increment_mdp_data_at(mdp, 1410 index, 1411 in_bytes(MultiBranchData::relative_count_offset())); 1412 1413 // The method data pointer needs to be updated. 1414 update_mdp_by_offset(mdp, 1415 index, 1416 in_bytes(MultiBranchData:: 1417 relative_displacement_offset())); 1418 1419 bind(profile_continue); 1420 } 1421 } 1422 1423 void InterpreterMacroAssembler::verify_oop(Register reg, TosState state) { 1424 if (state == atos) { 1425 MacroAssembler::verify_oop(reg); 1426 } 1427 } 1428 1429 void InterpreterMacroAssembler::verify_FPU(int stack_depth, TosState state) { ; } 1430 1431 1432 void InterpreterMacroAssembler::notify_method_entry() { 1433 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to 1434 // track stack depth. If it is possible to enter interp_only_mode we add 1435 // the code to check if the event should be sent. 1436 if (JvmtiExport::can_post_interpreter_events()) { 1437 Label L; 1438 ldrw(r3, Address(rthread, JavaThread::interp_only_mode_offset())); 1439 cbzw(r3, L); 1440 call_VM(noreg, CAST_FROM_FN_PTR(address, 1441 InterpreterRuntime::post_method_entry)); 1442 bind(L); 1443 } 1444 1445 { 1446 SkipIfEqual skip(this, &DTraceMethodProbes, false); 1447 get_method(c_rarg1); 1448 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 1449 rthread, c_rarg1); 1450 } 1451 1452 // RedefineClasses() tracing support for obsolete method entry 1453 if (RC_TRACE_IN_RANGE(0x00001000, 0x00002000)) { 1454 get_method(c_rarg1); 1455 call_VM_leaf( 1456 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 1457 rthread, c_rarg1); 1458 } 1459 1460 } 1461 1462 1463 void InterpreterMacroAssembler::notify_method_exit( 1464 TosState state, NotifyMethodExitMode mode) { 1465 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to 1466 // track stack depth. If it is possible to enter interp_only_mode we add 1467 // the code to check if the event should be sent. 1468 if (mode == NotifyJVMTI && JvmtiExport::can_post_interpreter_events()) { 1469 Label L; 1470 // Note: frame::interpreter_frame_result has a dependency on how the 1471 // method result is saved across the call to post_method_exit. If this 1472 // is changed then the interpreter_frame_result implementation will 1473 // need to be updated too. 1474 1475 // template interpreter will leave the result on the top of the stack. 1476 push(state); 1477 ldrw(r3, Address(rthread, JavaThread::interp_only_mode_offset())); 1478 cbz(r3, L); 1479 call_VM(noreg, 1480 CAST_FROM_FN_PTR(address, InterpreterRuntime::post_method_exit)); 1481 bind(L); 1482 pop(state); 1483 } 1484 1485 { 1486 SkipIfEqual skip(this, &DTraceMethodProbes, false); 1487 push(state); 1488 get_method(c_rarg1); 1489 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 1490 rthread, c_rarg1); 1491 pop(state); 1492 } 1493 } 1494 1495 1496 // Jump if ((*counter_addr += increment) & mask) satisfies the condition. 1497 void InterpreterMacroAssembler::increment_mask_and_jump(Address counter_addr, 1498 int increment, Address mask, 1499 Register scratch, Register scratch2, 1500 bool preloaded, Condition cond, 1501 Label* where) { 1502 if (!preloaded) { 1503 ldrw(scratch, counter_addr); 1504 } 1505 add(scratch, scratch, increment); 1506 strw(scratch, counter_addr); 1507 ldrw(scratch2, mask); 1508 ands(scratch, scratch, scratch2); 1509 br(cond, *where); 1510 } 1511 1512 void InterpreterMacroAssembler::call_VM_leaf_base(address entry_point, 1513 int number_of_arguments) { 1514 // interpreter specific 1515 // 1516 // Note: No need to save/restore rbcp & rlocals pointer since these 1517 // are callee saved registers and no blocking/ GC can happen 1518 // in leaf calls. 1519 #ifdef ASSERT 1520 { 1521 Label L; 1522 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 1523 cbz(rscratch1, L); 1524 stop("InterpreterMacroAssembler::call_VM_leaf_base:" 1525 " last_sp != NULL"); 1526 bind(L); 1527 } 1528 #endif /* ASSERT */ 1529 // super call 1530 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments); 1531 } 1532 1533 void InterpreterMacroAssembler::call_VM_base(Register oop_result, 1534 Register java_thread, 1535 Register last_java_sp, 1536 address entry_point, 1537 int number_of_arguments, 1538 bool check_exceptions) { 1539 // interpreter specific 1540 // 1541 // Note: Could avoid restoring locals ptr (callee saved) - however doesn't 1542 // really make a difference for these runtime calls, since they are 1543 // slow anyway. Btw., bcp must be saved/restored since it may change 1544 // due to GC. 1545 // assert(java_thread == noreg , "not expecting a precomputed java thread"); 1546 save_bcp(); 1547 #ifdef ASSERT 1548 { 1549 Label L; 1550 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 1551 cbz(rscratch1, L); 1552 stop("InterpreterMacroAssembler::call_VM_leaf_base:" 1553 " last_sp != NULL"); 1554 bind(L); 1555 } 1556 #endif /* ASSERT */ 1557 // super call 1558 MacroAssembler::call_VM_base(oop_result, noreg, last_java_sp, 1559 entry_point, number_of_arguments, 1560 check_exceptions); 1561 // interpreter specific 1562 restore_bcp(); 1563 restore_locals(); 1564 } 1565 1566 void InterpreterMacroAssembler::profile_obj_type(Register obj, const Address& mdo_addr) { 1567 Label update, next, none; 1568 1569 verify_oop(obj); 1570 1571 cbnz(obj, update); 1572 orptr(mdo_addr, TypeEntries::null_seen); 1573 b(next); 1574 1575 bind(update); 1576 load_klass(obj, obj); 1577 1578 ldr(rscratch1, mdo_addr); 1579 eor(obj, obj, rscratch1); 1580 tst(obj, TypeEntries::type_klass_mask); 1581 br(Assembler::EQ, next); // klass seen before, nothing to 1582 // do. The unknown bit may have been 1583 // set already but no need to check. 1584 1585 tst(obj, TypeEntries::type_unknown); 1586 br(Assembler::NE, next); // already unknown. Nothing to do anymore. 1587 1588 ldr(rscratch1, mdo_addr); 1589 cbz(rscratch1, none); 1590 cmp(rscratch1, TypeEntries::null_seen); 1591 br(Assembler::EQ, none); 1592 // There is a chance that the checks above (re-reading profiling 1593 // data from memory) fail if another thread has just set the 1594 // profiling to this obj's klass 1595 ldr(rscratch1, mdo_addr); 1596 eor(obj, obj, rscratch1); 1597 tst(obj, TypeEntries::type_klass_mask); 1598 br(Assembler::EQ, next); 1599 1600 // different than before. Cannot keep accurate profile. 1601 orptr(mdo_addr, TypeEntries::type_unknown); 1602 b(next); 1603 1604 bind(none); 1605 // first time here. Set profile type. 1606 str(obj, mdo_addr); 1607 1608 bind(next); 1609 } 1610 1611 void InterpreterMacroAssembler::profile_arguments_type(Register mdp, Register callee, Register tmp, bool is_virtual) { 1612 if (!ProfileInterpreter) { 1613 return; 1614 } 1615 1616 if (MethodData::profile_arguments() || MethodData::profile_return()) { 1617 Label profile_continue; 1618 1619 test_method_data_pointer(mdp, profile_continue); 1620 1621 int off_to_start = is_virtual ? in_bytes(VirtualCallData::virtual_call_data_size()) : in_bytes(CounterData::counter_data_size()); 1622 1623 ldrb(rscratch1, Address(mdp, in_bytes(DataLayout::tag_offset()) - off_to_start)); 1624 cmp(rscratch1, is_virtual ? DataLayout::virtual_call_type_data_tag : DataLayout::call_type_data_tag); 1625 br(Assembler::NE, profile_continue); 1626 1627 if (MethodData::profile_arguments()) { 1628 Label done; 1629 int off_to_args = in_bytes(TypeEntriesAtCall::args_data_offset()); 1630 1631 for (int i = 0; i < TypeProfileArgsLimit; i++) { 1632 if (i > 0 || MethodData::profile_return()) { 1633 // If return value type is profiled we may have no argument to profile 1634 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset()))); 1635 sub(tmp, tmp, i*TypeStackSlotEntries::per_arg_count()); 1636 cmp(tmp, TypeStackSlotEntries::per_arg_count()); 1637 add(rscratch1, mdp, off_to_args); 1638 br(Assembler::LT, done); 1639 } 1640 ldr(tmp, Address(callee, Method::const_offset())); 1641 load_unsigned_short(tmp, Address(tmp, ConstMethod::size_of_parameters_offset())); 1642 // stack offset o (zero based) from the start of the argument 1643 // list, for n arguments translates into offset n - o - 1 from 1644 // the end of the argument list 1645 ldr(rscratch1, Address(mdp, in_bytes(TypeEntriesAtCall::stack_slot_offset(i)))); 1646 sub(tmp, tmp, rscratch1); 1647 sub(tmp, tmp, 1); 1648 Address arg_addr = argument_address(tmp); 1649 ldr(tmp, arg_addr); 1650 1651 Address mdo_arg_addr(mdp, in_bytes(TypeEntriesAtCall::argument_type_offset(i))); 1652 profile_obj_type(tmp, mdo_arg_addr); 1653 1654 int to_add = in_bytes(TypeStackSlotEntries::per_arg_size()); 1655 off_to_args += to_add; 1656 } 1657 1658 if (MethodData::profile_return()) { 1659 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset()))); 1660 sub(tmp, tmp, TypeProfileArgsLimit*TypeStackSlotEntries::per_arg_count()); 1661 } 1662 1663 add(rscratch1, mdp, off_to_args); 1664 bind(done); 1665 mov(mdp, rscratch1); 1666 1667 if (MethodData::profile_return()) { 1668 // We're right after the type profile for the last 1669 // argument. tmp is the number of cells left in the 1670 // CallTypeData/VirtualCallTypeData to reach its end. Non null 1671 // if there's a return to profile. 1672 assert(ReturnTypeEntry::static_cell_count() < TypeStackSlotEntries::per_arg_count(), "can't move past ret type"); 1673 add(mdp, mdp, tmp, LSL, exact_log2(DataLayout::cell_size)); 1674 } 1675 str(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 1676 } else { 1677 assert(MethodData::profile_return(), "either profile call args or call ret"); 1678 update_mdp_by_constant(mdp, in_bytes(TypeEntriesAtCall::return_only_size())); 1679 } 1680 1681 // mdp points right after the end of the 1682 // CallTypeData/VirtualCallTypeData, right after the cells for the 1683 // return value type if there's one 1684 1685 bind(profile_continue); 1686 } 1687 } 1688 1689 void InterpreterMacroAssembler::profile_return_type(Register mdp, Register ret, Register tmp) { 1690 assert_different_registers(mdp, ret, tmp, rbcp); 1691 if (ProfileInterpreter && MethodData::profile_return()) { 1692 Label profile_continue, done; 1693 1694 test_method_data_pointer(mdp, profile_continue); 1695 1696 if (MethodData::profile_return_jsr292_only()) { 1697 assert(Method::intrinsic_id_size_in_bytes() == 2, "assuming Method::_intrinsic_id is u2"); 1698 1699 // If we don't profile all invoke bytecodes we must make sure 1700 // it's a bytecode we indeed profile. We can't go back to the 1701 // begining of the ProfileData we intend to update to check its 1702 // type because we're right after it and we don't known its 1703 // length 1704 Label do_profile; 1705 ldrb(rscratch1, Address(rbcp, 0)); 1706 cmp(rscratch1, Bytecodes::_invokedynamic); 1707 br(Assembler::EQ, do_profile); 1708 cmp(rscratch1, Bytecodes::_invokehandle); 1709 br(Assembler::EQ, do_profile); 1710 get_method(tmp); 1711 ldrh(rscratch1, Address(tmp, Method::intrinsic_id_offset_in_bytes())); 1712 cmp(rscratch1, vmIntrinsics::_compiledLambdaForm); 1713 br(Assembler::NE, profile_continue); 1714 1715 bind(do_profile); 1716 } 1717 1718 Address mdo_ret_addr(mdp, -in_bytes(ReturnTypeEntry::size())); 1719 mov(tmp, ret); 1720 profile_obj_type(tmp, mdo_ret_addr); 1721 1722 bind(profile_continue); 1723 } 1724 } 1725 1726 void InterpreterMacroAssembler::profile_parameters_type(Register mdp, Register tmp1, Register tmp2) { 1727 if (ProfileInterpreter && MethodData::profile_parameters()) { 1728 Label profile_continue, done; 1729 1730 test_method_data_pointer(mdp, profile_continue); 1731 1732 // Load the offset of the area within the MDO used for 1733 // parameters. If it's negative we're not profiling any parameters 1734 ldr(tmp1, Address(mdp, in_bytes(MethodData::parameters_type_data_di_offset()) - in_bytes(MethodData::data_offset()))); 1735 cmp(tmp1, 0u); 1736 br(Assembler::LT, profile_continue); 1737 1738 // Compute a pointer to the area for parameters from the offset 1739 // and move the pointer to the slot for the last 1740 // parameters. Collect profiling from last parameter down. 1741 // mdo start + parameters offset + array length - 1 1742 add(mdp, mdp, tmp1); 1743 ldr(tmp1, Address(mdp, ArrayData::array_len_offset())); 1744 sub(tmp1, tmp1, TypeStackSlotEntries::per_arg_count()); 1745 1746 Label loop; 1747 bind(loop); 1748 1749 int off_base = in_bytes(ParametersTypeData::stack_slot_offset(0)); 1750 int type_base = in_bytes(ParametersTypeData::type_offset(0)); 1751 int per_arg_scale = exact_log2(DataLayout::cell_size); 1752 add(rscratch1, mdp, off_base); 1753 add(rscratch2, mdp, type_base); 1754 1755 Address arg_off(rscratch1, tmp1, Address::lsl(per_arg_scale)); 1756 Address arg_type(rscratch2, tmp1, Address::lsl(per_arg_scale)); 1757 1758 // load offset on the stack from the slot for this parameter 1759 ldr(tmp2, arg_off); 1760 neg(tmp2, tmp2); 1761 // read the parameter from the local area 1762 ldr(tmp2, Address(rlocals, tmp2, Address::lsl(Interpreter::logStackElementSize))); 1763 1764 // profile the parameter 1765 profile_obj_type(tmp2, arg_type); 1766 1767 // go to next parameter 1768 subs(tmp1, tmp1, TypeStackSlotEntries::per_arg_count()); 1769 br(Assembler::GE, loop); 1770 1771 bind(profile_continue); 1772 } 1773 }