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 // Implementation of InterpreterMacroAssembler 44 45 void InterpreterMacroAssembler::jump_to_entry(address entry) { 46 assert(entry, "Entry must have been generated by now"); 47 b(entry); 48 } 49 50 #ifndef CC_INTERP 51 52 void InterpreterMacroAssembler::check_and_handle_popframe(Register java_thread) { 53 if (JvmtiExport::can_pop_frame()) { 54 Label L; 55 // Initiate popframe handling only if it is not already being 56 // processed. If the flag has the popframe_processing bit set, it 57 // means that this code is called *during* popframe handling - we 58 // don't want to reenter. 59 // This method is only called just after the call into the vm in 60 // call_VM_base, so the arg registers are available. 61 ldrw(rscratch1, Address(rthread, JavaThread::popframe_condition_offset())); 62 tstw(rscratch1, JavaThread::popframe_pending_bit); 63 br(Assembler::EQ, L); 64 tstw(rscratch1, JavaThread::popframe_processing_bit); 65 br(Assembler::NE, L); 66 // Call Interpreter::remove_activation_preserving_args_entry() to get the 67 // address of the same-named entrypoint in the generated interpreter code. 68 call_VM_leaf(CAST_FROM_FN_PTR(address, Interpreter::remove_activation_preserving_args_entry)); 69 br(r0); 70 bind(L); 71 } 72 } 73 74 75 void InterpreterMacroAssembler::load_earlyret_value(TosState state) { 76 ldr(r2, Address(rthread, JavaThread::jvmti_thread_state_offset())); 77 const Address tos_addr(r2, JvmtiThreadState::earlyret_tos_offset()); 78 const Address oop_addr(r2, JvmtiThreadState::earlyret_oop_offset()); 79 const Address val_addr(r2, JvmtiThreadState::earlyret_value_offset()); 80 switch (state) { 81 case atos: ldr(r0, oop_addr); 82 str(zr, oop_addr); 83 verify_oop(r0, state); break; 84 case ltos: ldr(r0, val_addr); break; 85 case btos: // fall through 86 case ctos: // fall through 87 case stos: // fall through 88 case itos: ldrw(r0, val_addr); break; 89 case ftos: ldrs(v0, val_addr); break; 90 case dtos: ldrd(v0, val_addr); break; 91 case vtos: /* nothing to do */ break; 92 default : ShouldNotReachHere(); 93 } 94 // Clean up tos value in the thread object 95 movw(rscratch1, (int) ilgl); 96 strw(rscratch1, tos_addr); 97 strw(zr, val_addr); 98 } 99 100 101 void InterpreterMacroAssembler::check_and_handle_earlyret(Register java_thread) { 102 if (JvmtiExport::can_force_early_return()) { 103 Label L; 104 ldr(rscratch1, Address(rthread, JavaThread::jvmti_thread_state_offset())); 105 cbz(rscratch1, L); // if (thread->jvmti_thread_state() == NULL) exit; 106 107 // Initiate earlyret handling only if it is not already being processed. 108 // If the flag has the earlyret_processing bit set, it means that this code 109 // is called *during* earlyret handling - we don't want to reenter. 110 ldrw(rscratch1, Address(rscratch1, JvmtiThreadState::earlyret_state_offset())); 111 cmpw(rscratch1, JvmtiThreadState::earlyret_pending); 112 br(Assembler::NE, L); 113 114 // Call Interpreter::remove_activation_early_entry() to get the address of the 115 // same-named entrypoint in the generated interpreter code. 116 ldr(rscratch1, Address(rthread, JavaThread::jvmti_thread_state_offset())); 117 ldrw(rscratch1, Address(rscratch1, JvmtiThreadState::earlyret_tos_offset())); 118 call_VM_leaf(CAST_FROM_FN_PTR(address, Interpreter::remove_activation_early_entry), rscratch1); 119 br(r0); 120 bind(L); 121 } 122 } 123 124 void InterpreterMacroAssembler::get_unsigned_2_byte_index_at_bcp( 125 Register reg, 126 int bcp_offset) { 127 assert(bcp_offset >= 0, "bcp is still pointing to start of bytecode"); 128 ldrh(reg, Address(rbcp, bcp_offset)); 129 rev16(reg, reg); 130 } 131 132 void InterpreterMacroAssembler::get_dispatch() { 133 unsigned long offset; 134 adrp(rdispatch, ExternalAddress((address)Interpreter::dispatch_table()), offset); 135 lea(rdispatch, Address(rdispatch, offset)); 136 } 137 138 void InterpreterMacroAssembler::get_cache_index_at_bcp(Register index, 139 int bcp_offset, 140 size_t index_size) { 141 assert(bcp_offset > 0, "bcp is still pointing to start of bytecode"); 142 if (index_size == sizeof(u2)) { 143 load_unsigned_short(index, Address(rbcp, bcp_offset)); 144 } else if (index_size == sizeof(u4)) { 145 // assert(EnableInvokeDynamic, "giant index used only for JSR 292"); 146 ldrw(index, Address(rbcp, bcp_offset)); 147 // Check if the secondary index definition is still ~x, otherwise 148 // we have to change the following assembler code to calculate the 149 // plain index. 150 assert(ConstantPool::decode_invokedynamic_index(~123) == 123, "else change next line"); 151 eonw(index, index, zr); // convert to plain index 152 } else if (index_size == sizeof(u1)) { 153 load_unsigned_byte(index, Address(rbcp, bcp_offset)); 154 } else { 155 ShouldNotReachHere(); 156 } 157 } 158 159 // Return 160 // Rindex: index into constant pool 161 // Rcache: address of cache entry - ConstantPoolCache::base_offset() 162 // 163 // A caller must add ConstantPoolCache::base_offset() to Rcache to get 164 // the true address of the cache entry. 165 // 166 void InterpreterMacroAssembler::get_cache_and_index_at_bcp(Register cache, 167 Register index, 168 int bcp_offset, 169 size_t index_size) { 170 assert_different_registers(cache, index); 171 assert_different_registers(cache, rcpool); 172 get_cache_index_at_bcp(index, bcp_offset, index_size); 173 assert(sizeof(ConstantPoolCacheEntry) == 4 * wordSize, "adjust code below"); 174 // convert from field index to ConstantPoolCacheEntry 175 // aarch64 already has the cache in rcpool so there is no need to 176 // install it in cache. instead we pre-add the indexed offset to 177 // rcpool and return it in cache. All clients of this method need to 178 // be modified accordingly. 179 add(cache, rcpool, index, Assembler::LSL, 5); 180 } 181 182 183 void InterpreterMacroAssembler::get_cache_and_index_and_bytecode_at_bcp(Register cache, 184 Register index, 185 Register bytecode, 186 int byte_no, 187 int bcp_offset, 188 size_t index_size) { 189 get_cache_and_index_at_bcp(cache, index, bcp_offset, index_size); 190 // We use a 32-bit load here since the layout of 64-bit words on 191 // little-endian machines allow us that. 192 // n.b. unlike x86 cache already includes the index offset 193 lea(bytecode, Address(cache, 194 ConstantPoolCache::base_offset() 195 + ConstantPoolCacheEntry::indices_offset())); 196 ldarw(bytecode, bytecode); 197 const int shift_count = (1 + byte_no) * BitsPerByte; 198 ubfx(bytecode, bytecode, shift_count, BitsPerByte); 199 } 200 201 void InterpreterMacroAssembler::get_cache_entry_pointer_at_bcp(Register cache, 202 Register tmp, 203 int bcp_offset, 204 size_t index_size) { 205 assert(cache != tmp, "must use different register"); 206 get_cache_index_at_bcp(tmp, bcp_offset, index_size); 207 assert(sizeof(ConstantPoolCacheEntry) == 4 * wordSize, "adjust code below"); 208 // convert from field index to ConstantPoolCacheEntry index 209 // and from word offset to byte offset 210 assert(exact_log2(in_bytes(ConstantPoolCacheEntry::size_in_bytes())) == 2 + LogBytesPerWord, "else change next line"); 211 ldr(cache, Address(rfp, frame::interpreter_frame_cache_offset * wordSize)); 212 // skip past the header 213 add(cache, cache, in_bytes(ConstantPoolCache::base_offset())); 214 add(cache, cache, tmp, Assembler::LSL, 2 + LogBytesPerWord); // construct pointer to cache entry 215 } 216 217 void InterpreterMacroAssembler::get_method_counters(Register method, 218 Register mcs, Label& skip) { 219 Label has_counters; 220 ldr(mcs, Address(method, Method::method_counters_offset())); 221 cbnz(mcs, has_counters); 222 call_VM(noreg, CAST_FROM_FN_PTR(address, 223 InterpreterRuntime::build_method_counters), method); 224 ldr(mcs, Address(method, Method::method_counters_offset())); 225 cbz(mcs, skip); // No MethodCounters allocated, OutOfMemory 226 bind(has_counters); 227 } 228 229 // Load object from cpool->resolved_references(index) 230 void InterpreterMacroAssembler::load_resolved_reference_at_index( 231 Register result, Register index) { 232 assert_different_registers(result, index); 233 // convert from field index to resolved_references() index and from 234 // word index to byte offset. Since this is a java object, it can be compressed 235 Register tmp = index; // reuse 236 lslw(tmp, tmp, LogBytesPerHeapOop); 237 238 get_constant_pool(result); 239 // load pointer for resolved_references[] objArray 240 ldr(result, Address(result, ConstantPool::resolved_references_offset_in_bytes())); 241 // JNIHandles::resolve(obj); 242 ldr(result, Address(result, 0)); 243 // Add in the index 244 add(result, result, tmp); 245 load_heap_oop(result, Address(result, arrayOopDesc::base_offset_in_bytes(T_OBJECT))); 246 } 247 248 // Generate a subtype check: branch to ok_is_subtype if sub_klass is a 249 // subtype of super_klass. 250 // 251 // Args: 252 // r0: superklass 253 // Rsub_klass: subklass 254 // 255 // Kills: 256 // r2, r5 257 void InterpreterMacroAssembler::gen_subtype_check(Register Rsub_klass, 258 Label& ok_is_subtype) { 259 assert(Rsub_klass != r0, "r0 holds superklass"); 260 assert(Rsub_klass != r2, "r2 holds 2ndary super array length"); 261 assert(Rsub_klass != r5, "r5 holds 2ndary super array scan ptr"); 262 263 // Profile the not-null value's klass. 264 profile_typecheck(r2, Rsub_klass, r5); // blows r2, reloads r5 265 266 // Do the check. 267 check_klass_subtype(Rsub_klass, r0, r2, ok_is_subtype); // blows r2 268 269 // Profile the failure of the check. 270 profile_typecheck_failed(r2); // blows r2 271 } 272 273 // Java Expression Stack 274 275 void InterpreterMacroAssembler::pop_ptr(Register r) { 276 ldr(r, post(esp, wordSize)); 277 } 278 279 void InterpreterMacroAssembler::pop_i(Register r) { 280 ldrw(r, post(esp, wordSize)); 281 } 282 283 void InterpreterMacroAssembler::pop_l(Register r) { 284 ldr(r, post(esp, 2 * Interpreter::stackElementSize)); 285 } 286 287 void InterpreterMacroAssembler::push_ptr(Register r) { 288 str(r, pre(esp, -wordSize)); 289 } 290 291 void InterpreterMacroAssembler::push_i(Register r) { 292 str(r, pre(esp, -wordSize)); 293 } 294 295 void InterpreterMacroAssembler::push_l(Register r) { 296 str(r, pre(esp, 2 * -wordSize)); 297 } 298 299 void InterpreterMacroAssembler::pop_f(FloatRegister r) { 300 ldrs(r, post(esp, wordSize)); 301 } 302 303 void InterpreterMacroAssembler::pop_d(FloatRegister r) { 304 ldrd(r, post(esp, 2 * Interpreter::stackElementSize)); 305 } 306 307 void InterpreterMacroAssembler::push_f(FloatRegister r) { 308 strs(r, pre(esp, -wordSize)); 309 } 310 311 void InterpreterMacroAssembler::push_d(FloatRegister r) { 312 strd(r, pre(esp, 2* -wordSize)); 313 } 314 315 void InterpreterMacroAssembler::pop(TosState state) { 316 switch (state) { 317 case atos: pop_ptr(); break; 318 case btos: 319 case ctos: 320 case stos: 321 case itos: pop_i(); break; 322 case ltos: pop_l(); break; 323 case ftos: pop_f(); break; 324 case dtos: pop_d(); break; 325 case vtos: /* nothing to do */ break; 326 default: ShouldNotReachHere(); 327 } 328 verify_oop(r0, state); 329 } 330 331 void InterpreterMacroAssembler::push(TosState state) { 332 verify_oop(r0, state); 333 switch (state) { 334 case atos: push_ptr(); break; 335 case btos: 336 case ctos: 337 case stos: 338 case itos: push_i(); break; 339 case ltos: push_l(); break; 340 case ftos: push_f(); break; 341 case dtos: push_d(); break; 342 case vtos: /* nothing to do */ break; 343 default : ShouldNotReachHere(); 344 } 345 } 346 347 // Helpers for swap and dup 348 void InterpreterMacroAssembler::load_ptr(int n, Register val) { 349 ldr(val, Address(esp, Interpreter::expr_offset_in_bytes(n))); 350 } 351 352 void InterpreterMacroAssembler::store_ptr(int n, Register val) { 353 str(val, Address(esp, Interpreter::expr_offset_in_bytes(n))); 354 } 355 356 357 void InterpreterMacroAssembler::prepare_to_jump_from_interpreted() { 358 // set sender sp 359 mov(r13, sp); 360 // record last_sp 361 str(esp, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 362 } 363 364 // Jump to from_interpreted entry of a call unless single stepping is possible 365 // in this thread in which case we must call the i2i entry 366 void InterpreterMacroAssembler::jump_from_interpreted(Register method, Register temp) { 367 prepare_to_jump_from_interpreted(); 368 369 if (JvmtiExport::can_post_interpreter_events()) { 370 Label run_compiled_code; 371 // JVMTI events, such as single-stepping, are implemented partly by avoiding running 372 // compiled code in threads for which the event is enabled. Check here for 373 // interp_only_mode if these events CAN be enabled. 374 // interp_only is an int, on little endian it is sufficient to test the byte only 375 // Is a cmpl faster? 376 ldr(rscratch1, Address(rthread, JavaThread::interp_only_mode_offset())); 377 cbz(rscratch1, run_compiled_code); 378 ldr(rscratch1, Address(method, Method::interpreter_entry_offset())); 379 br(rscratch1); 380 bind(run_compiled_code); 381 } 382 383 ldr(rscratch1, Address(method, Method::from_interpreted_offset())); 384 br(rscratch1); 385 } 386 387 // The following two routines provide a hook so that an implementation 388 // can schedule the dispatch in two parts. amd64 does not do this. 389 void InterpreterMacroAssembler::dispatch_prolog(TosState state, int step) { 390 } 391 392 void InterpreterMacroAssembler::dispatch_epilog(TosState state, int step) { 393 dispatch_next(state, step); 394 } 395 396 void InterpreterMacroAssembler::dispatch_base(TosState state, 397 address* table, 398 bool verifyoop) { 399 if (VerifyActivationFrameSize) { 400 Unimplemented(); 401 } 402 if (verifyoop) { 403 verify_oop(r0, state); 404 } 405 if (table == Interpreter::dispatch_table(state)) { 406 addw(rscratch2, rscratch1, Interpreter::distance_from_dispatch_table(state)); 407 ldr(rscratch2, Address(rdispatch, rscratch2, Address::uxtw(3))); 408 } else { 409 mov(rscratch2, (address)table); 410 ldr(rscratch2, Address(rscratch2, rscratch1, Address::uxtw(3))); 411 } 412 br(rscratch2); 413 } 414 415 void InterpreterMacroAssembler::dispatch_only(TosState state) { 416 dispatch_base(state, Interpreter::dispatch_table(state)); 417 } 418 419 void InterpreterMacroAssembler::dispatch_only_normal(TosState state) { 420 dispatch_base(state, Interpreter::normal_table(state)); 421 } 422 423 void InterpreterMacroAssembler::dispatch_only_noverify(TosState state) { 424 dispatch_base(state, Interpreter::normal_table(state), false); 425 } 426 427 428 void InterpreterMacroAssembler::dispatch_next(TosState state, int step) { 429 // load next bytecode 430 ldrb(rscratch1, Address(pre(rbcp, step))); 431 dispatch_base(state, Interpreter::dispatch_table(state)); 432 } 433 434 void InterpreterMacroAssembler::dispatch_via(TosState state, address* table) { 435 // load current bytecode 436 ldrb(rscratch1, Address(rbcp, 0)); 437 dispatch_base(state, table); 438 } 439 440 // remove activation 441 // 442 // Unlock the receiver if this is a synchronized method. 443 // Unlock any Java monitors from syncronized blocks. 444 // Remove the activation from the stack. 445 // 446 // If there are locked Java monitors 447 // If throw_monitor_exception 448 // throws IllegalMonitorStateException 449 // Else if install_monitor_exception 450 // installs IllegalMonitorStateException 451 // Else 452 // no error processing 453 void InterpreterMacroAssembler::remove_activation( 454 TosState state, 455 bool throw_monitor_exception, 456 bool install_monitor_exception, 457 bool notify_jvmdi) { 458 // Note: Registers r3 xmm0 may be in use for the 459 // result check if synchronized method 460 Label unlocked, unlock, no_unlock; 461 462 // get the value of _do_not_unlock_if_synchronized into r3 463 const Address do_not_unlock_if_synchronized(rthread, 464 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset())); 465 ldrb(r3, do_not_unlock_if_synchronized); 466 strb(zr, do_not_unlock_if_synchronized); // reset the flag 467 468 // get method access flags 469 ldr(r1, Address(rfp, frame::interpreter_frame_method_offset * wordSize)); 470 ldr(r2, Address(r1, Method::access_flags_offset())); 471 tst(r2, JVM_ACC_SYNCHRONIZED); 472 br(Assembler::EQ, unlocked); 473 474 // Don't unlock anything if the _do_not_unlock_if_synchronized flag 475 // is set. 476 cbnz(r3, no_unlock); 477 478 // unlock monitor 479 push(state); // save result 480 481 // BasicObjectLock will be first in list, since this is a 482 // synchronized method. However, need to check that the object has 483 // not been unlocked by an explicit monitorexit bytecode. 484 const Address monitor(rfp, frame::interpreter_frame_initial_sp_offset * 485 wordSize - (int) sizeof(BasicObjectLock)); 486 // We use c_rarg1 so that if we go slow path it will be the correct 487 // register for unlock_object to pass to VM directly 488 lea(c_rarg1, monitor); // address of first monitor 489 490 ldr(r0, Address(c_rarg1, BasicObjectLock::obj_offset_in_bytes())); 491 cbnz(r0, unlock); 492 493 pop(state); 494 if (throw_monitor_exception) { 495 // Entry already unlocked, need to throw exception 496 call_VM(noreg, CAST_FROM_FN_PTR(address, 497 InterpreterRuntime::throw_illegal_monitor_state_exception)); 498 should_not_reach_here(); 499 } else { 500 // Monitor already unlocked during a stack unroll. If requested, 501 // install an illegal_monitor_state_exception. Continue with 502 // stack unrolling. 503 if (install_monitor_exception) { 504 call_VM(noreg, CAST_FROM_FN_PTR(address, 505 InterpreterRuntime::new_illegal_monitor_state_exception)); 506 } 507 b(unlocked); 508 } 509 510 bind(unlock); 511 unlock_object(c_rarg1); 512 pop(state); 513 514 // Check that for block-structured locking (i.e., that all locked 515 // objects has been unlocked) 516 bind(unlocked); 517 518 // r0: Might contain return value 519 520 // Check that all monitors are unlocked 521 { 522 Label loop, exception, entry, restart; 523 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize; 524 const Address monitor_block_top( 525 rfp, frame::interpreter_frame_monitor_block_top_offset * wordSize); 526 const Address monitor_block_bot( 527 rfp, frame::interpreter_frame_initial_sp_offset * wordSize); 528 529 bind(restart); 530 // We use c_rarg1 so that if we go slow path it will be the correct 531 // register for unlock_object to pass to VM directly 532 ldr(c_rarg1, monitor_block_top); // points to current entry, starting 533 // with top-most entry 534 lea(r19, monitor_block_bot); // points to word before bottom of 535 // monitor block 536 b(entry); 537 538 // Entry already locked, need to throw exception 539 bind(exception); 540 541 if (throw_monitor_exception) { 542 // Throw exception 543 MacroAssembler::call_VM(noreg, 544 CAST_FROM_FN_PTR(address, InterpreterRuntime:: 545 throw_illegal_monitor_state_exception)); 546 should_not_reach_here(); 547 } else { 548 // Stack unrolling. Unlock object and install illegal_monitor_exception. 549 // Unlock does not block, so don't have to worry about the frame. 550 // We don't have to preserve c_rarg1 since we are going to throw an exception. 551 552 push(state); 553 unlock_object(c_rarg1); 554 pop(state); 555 556 if (install_monitor_exception) { 557 call_VM(noreg, CAST_FROM_FN_PTR(address, 558 InterpreterRuntime:: 559 new_illegal_monitor_state_exception)); 560 } 561 562 b(restart); 563 } 564 565 bind(loop); 566 // check if current entry is used 567 ldr(rscratch1, Address(c_rarg1, BasicObjectLock::obj_offset_in_bytes())); 568 cbnz(rscratch1, exception); 569 570 add(c_rarg1, c_rarg1, entry_size); // otherwise advance to next entry 571 bind(entry); 572 cmp(c_rarg1, r19); // check if bottom reached 573 br(Assembler::NE, loop); // if not at bottom then check this entry 574 } 575 576 bind(no_unlock); 577 578 // jvmti support 579 if (notify_jvmdi) { 580 notify_method_exit(state, NotifyJVMTI); // preserve TOSCA 581 } else { 582 notify_method_exit(state, SkipNotifyJVMTI); // preserve TOSCA 583 } 584 585 // remove activation 586 // get sender esp 587 ldr(esp, 588 Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); 589 // remove frame anchor 590 leave(); 591 // If we're returning to interpreted code we will shortly be 592 // adjusting SP to allow some space for ESP. If we're returning to 593 // compiled code the saved sender SP was saved in sender_sp, so this 594 // restores it. 595 andr(sp, esp, -16); 596 } 597 598 #endif // C_INTERP 599 600 // Lock object 601 // 602 // Args: 603 // c_rarg1: BasicObjectLock to be used for locking 604 // 605 // Kills: 606 // r0 607 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, .. (param regs) 608 // rscratch1, rscratch2 (scratch regs) 609 void InterpreterMacroAssembler::lock_object(Register lock_reg) 610 { 611 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be c_rarg1"); 612 if (UseHeavyMonitors) { 613 call_VM(noreg, 614 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter), 615 lock_reg); 616 } else { 617 Label done; 618 619 const Register swap_reg = r0; 620 const Register tmp = c_rarg2; 621 const Register obj_reg = c_rarg3; // Will contain the oop 622 623 const int obj_offset = BasicObjectLock::obj_offset_in_bytes(); 624 const int lock_offset = BasicObjectLock::lock_offset_in_bytes (); 625 const int mark_offset = lock_offset + 626 BasicLock::displaced_header_offset_in_bytes(); 627 628 Label slow_case; 629 630 // Load object pointer into obj_reg %c_rarg3 631 ldr(obj_reg, Address(lock_reg, obj_offset)); 632 633 if (UseBiasedLocking) { 634 biased_locking_enter(lock_reg, obj_reg, swap_reg, tmp, false, done, &slow_case); 635 } 636 637 // Load (object->mark() | 1) into swap_reg 638 ldr(rscratch1, Address(obj_reg, 0)); 639 orr(swap_reg, rscratch1, 1); 640 641 // Save (object->mark() | 1) into BasicLock's displaced header 642 str(swap_reg, Address(lock_reg, mark_offset)); 643 644 assert(lock_offset == 0, 645 "displached header must be first word in BasicObjectLock"); 646 647 Label fail; 648 if (PrintBiasedLockingStatistics) { 649 Label fast; 650 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, fast, &fail); 651 bind(fast); 652 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()), 653 rscratch2, rscratch1, tmp); 654 b(done); 655 bind(fail); 656 } else { 657 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL); 658 } 659 660 // Test if the oopMark is an obvious stack pointer, i.e., 661 // 1) (mark & 7) == 0, and 662 // 2) rsp <= mark < mark + os::pagesize() 663 // 664 // These 3 tests can be done by evaluating the following 665 // expression: ((mark - rsp) & (7 - os::vm_page_size())), 666 // assuming both stack pointer and pagesize have their 667 // least significant 3 bits clear. 668 // NOTE: the oopMark is in swap_reg %r0 as the result of cmpxchg 669 // NOTE2: aarch64 does not like to subtract sp from rn so take a 670 // copy 671 mov(rscratch1, sp); 672 sub(swap_reg, swap_reg, rscratch1); 673 ands(swap_reg, swap_reg, (unsigned long)(7 - os::vm_page_size())); 674 675 // Save the test result, for recursive case, the result is zero 676 str(swap_reg, Address(lock_reg, mark_offset)); 677 678 if (PrintBiasedLockingStatistics) { 679 br(Assembler::NE, slow_case); 680 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()), 681 rscratch2, rscratch1, tmp); 682 } 683 br(Assembler::EQ, done); 684 685 bind(slow_case); 686 687 // Call the runtime routine for slow case 688 call_VM(noreg, 689 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter), 690 lock_reg); 691 692 bind(done); 693 } 694 } 695 696 697 // Unlocks an object. Used in monitorexit bytecode and 698 // remove_activation. Throws an IllegalMonitorException if object is 699 // not locked by current thread. 700 // 701 // Args: 702 // c_rarg1: BasicObjectLock for lock 703 // 704 // Kills: 705 // r0 706 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, ... (param regs) 707 // rscratch1, rscratch2 (scratch regs) 708 void InterpreterMacroAssembler::unlock_object(Register lock_reg) 709 { 710 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be rarg1"); 711 712 if (UseHeavyMonitors) { 713 call_VM(noreg, 714 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit), 715 lock_reg); 716 } else { 717 Label done; 718 719 const Register swap_reg = r0; 720 const Register header_reg = c_rarg2; // Will contain the old oopMark 721 const Register obj_reg = c_rarg3; // Will contain the oop 722 723 save_bcp(); // Save in case of exception 724 725 // Convert from BasicObjectLock structure to object and BasicLock 726 // structure Store the BasicLock address into %r0 727 lea(swap_reg, Address(lock_reg, BasicObjectLock::lock_offset_in_bytes())); 728 729 // Load oop into obj_reg(%c_rarg3) 730 ldr(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); 731 732 // Free entry 733 str(zr, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); 734 735 if (UseBiasedLocking) { 736 biased_locking_exit(obj_reg, header_reg, done); 737 } 738 739 // Load the old header from BasicLock structure 740 ldr(header_reg, Address(swap_reg, 741 BasicLock::displaced_header_offset_in_bytes())); 742 743 // Test for recursion 744 cbz(header_reg, done); 745 746 // Atomic swap back the old header 747 cmpxchgptr(swap_reg, header_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL); 748 749 // Call the runtime routine for slow case. 750 str(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); // restore obj 751 call_VM(noreg, 752 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit), 753 lock_reg); 754 755 bind(done); 756 757 restore_bcp(); 758 } 759 } 760 761 #ifndef CC_INTERP 762 763 void InterpreterMacroAssembler::test_method_data_pointer(Register mdp, 764 Label& zero_continue) { 765 assert(ProfileInterpreter, "must be profiling interpreter"); 766 ldr(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 767 cbz(mdp, zero_continue); 768 } 769 770 // Set the method data pointer for the current bcp. 771 void InterpreterMacroAssembler::set_method_data_pointer_for_bcp() { 772 assert(ProfileInterpreter, "must be profiling interpreter"); 773 Label set_mdp; 774 stp(r0, r1, Address(pre(sp, -2 * wordSize))); 775 776 // Test MDO to avoid the call if it is NULL. 777 ldr(r0, Address(rmethod, in_bytes(Method::method_data_offset()))); 778 cbz(r0, set_mdp); 779 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::bcp_to_di), rmethod, rbcp); 780 // r0: mdi 781 // mdo is guaranteed to be non-zero here, we checked for it before the call. 782 ldr(r1, Address(rmethod, in_bytes(Method::method_data_offset()))); 783 lea(r1, Address(r1, in_bytes(MethodData::data_offset()))); 784 add(r0, r1, r0); 785 str(r0, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 786 bind(set_mdp); 787 ldp(r0, r1, Address(post(sp, 2 * wordSize))); 788 } 789 790 void InterpreterMacroAssembler::verify_method_data_pointer() { 791 assert(ProfileInterpreter, "must be profiling interpreter"); 792 #ifdef ASSERT 793 Label verify_continue; 794 stp(r0, r1, Address(pre(sp, -2 * wordSize))); 795 stp(r2, r3, Address(pre(sp, -2 * wordSize))); 796 test_method_data_pointer(r3, verify_continue); // If mdp is zero, continue 797 get_method(r1); 798 799 // If the mdp is valid, it will point to a DataLayout header which is 800 // consistent with the bcp. The converse is highly probable also. 801 ldrsh(r2, Address(r3, in_bytes(DataLayout::bci_offset()))); 802 ldr(rscratch1, Address(r1, Method::const_offset())); 803 add(r2, r2, rscratch1, Assembler::LSL); 804 lea(r2, Address(r2, ConstMethod::codes_offset())); 805 cmp(r2, rbcp); 806 br(Assembler::EQ, verify_continue); 807 // r1: method 808 // rbcp: bcp // rbcp == 22 809 // r3: mdp 810 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::verify_mdp), 811 r1, rbcp, r3); 812 bind(verify_continue); 813 ldp(r2, r3, Address(post(sp, 2 * wordSize))); 814 ldp(r0, r1, Address(post(sp, 2 * wordSize))); 815 #endif // ASSERT 816 } 817 818 819 void InterpreterMacroAssembler::set_mdp_data_at(Register mdp_in, 820 int constant, 821 Register value) { 822 assert(ProfileInterpreter, "must be profiling interpreter"); 823 Address data(mdp_in, constant); 824 str(value, data); 825 } 826 827 828 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in, 829 int constant, 830 bool decrement) { 831 increment_mdp_data_at(mdp_in, noreg, constant, decrement); 832 } 833 834 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in, 835 Register reg, 836 int constant, 837 bool decrement) { 838 assert(ProfileInterpreter, "must be profiling interpreter"); 839 // %%% this does 64bit counters at best it is wasting space 840 // at worst it is a rare bug when counters overflow 841 842 assert_different_registers(rscratch2, rscratch1, mdp_in, reg); 843 844 Address addr1(mdp_in, constant); 845 Address addr2(rscratch2, reg, Address::lsl(0)); 846 Address &addr = addr1; 847 if (reg != noreg) { 848 lea(rscratch2, addr1); 849 addr = addr2; 850 } 851 852 if (decrement) { 853 // Decrement the register. Set condition codes. 854 // Intel does this 855 // addptr(data, (int32_t) -DataLayout::counter_increment); 856 // If the decrement causes the counter to overflow, stay negative 857 // Label L; 858 // jcc(Assembler::negative, L); 859 // addptr(data, (int32_t) DataLayout::counter_increment); 860 // so we do this 861 ldr(rscratch1, addr); 862 subs(rscratch1, rscratch1, (unsigned)DataLayout::counter_increment); 863 Label L; 864 br(Assembler::LO, L); // skip store if counter underflow 865 str(rscratch1, addr); 866 bind(L); 867 } else { 868 assert(DataLayout::counter_increment == 1, 869 "flow-free idiom only works with 1"); 870 // Intel does this 871 // Increment the register. Set carry flag. 872 // addptr(data, DataLayout::counter_increment); 873 // If the increment causes the counter to overflow, pull back by 1. 874 // sbbptr(data, (int32_t)0); 875 // so we do this 876 ldr(rscratch1, addr); 877 adds(rscratch1, rscratch1, DataLayout::counter_increment); 878 Label L; 879 br(Assembler::CS, L); // skip store if counter overflow 880 str(rscratch1, addr); 881 bind(L); 882 } 883 } 884 885 void InterpreterMacroAssembler::set_mdp_flag_at(Register mdp_in, 886 int flag_byte_constant) { 887 assert(ProfileInterpreter, "must be profiling interpreter"); 888 int header_offset = in_bytes(DataLayout::header_offset()); 889 int header_bits = DataLayout::flag_mask_to_header_mask(flag_byte_constant); 890 // Set the flag 891 ldr(rscratch1, Address(mdp_in, header_offset)); 892 orr(rscratch1, rscratch1, header_bits); 893 str(rscratch1, Address(mdp_in, header_offset)); 894 } 895 896 897 void InterpreterMacroAssembler::test_mdp_data_at(Register mdp_in, 898 int offset, 899 Register value, 900 Register test_value_out, 901 Label& not_equal_continue) { 902 assert(ProfileInterpreter, "must be profiling interpreter"); 903 if (test_value_out == noreg) { 904 ldr(rscratch1, Address(mdp_in, offset)); 905 cmp(value, rscratch1); 906 } else { 907 // Put the test value into a register, so caller can use it: 908 ldr(test_value_out, Address(mdp_in, offset)); 909 cmp(value, test_value_out); 910 } 911 br(Assembler::NE, not_equal_continue); 912 } 913 914 915 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in, 916 int offset_of_disp) { 917 assert(ProfileInterpreter, "must be profiling interpreter"); 918 ldr(rscratch1, Address(mdp_in, offset_of_disp)); 919 add(mdp_in, mdp_in, rscratch1, LSL); 920 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 921 } 922 923 924 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in, 925 Register reg, 926 int offset_of_disp) { 927 assert(ProfileInterpreter, "must be profiling interpreter"); 928 lea(rscratch1, Address(mdp_in, offset_of_disp)); 929 ldr(rscratch1, Address(rscratch1, reg, Address::lsl(0))); 930 add(mdp_in, mdp_in, rscratch1, LSL); 931 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 932 } 933 934 935 void InterpreterMacroAssembler::update_mdp_by_constant(Register mdp_in, 936 int constant) { 937 assert(ProfileInterpreter, "must be profiling interpreter"); 938 add(mdp_in, mdp_in, (unsigned)constant); 939 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 940 } 941 942 943 void InterpreterMacroAssembler::update_mdp_for_ret(Register return_bci) { 944 assert(ProfileInterpreter, "must be profiling interpreter"); 945 // save/restore across call_VM 946 stp(zr, return_bci, Address(pre(sp, -2 * wordSize))); 947 call_VM(noreg, 948 CAST_FROM_FN_PTR(address, InterpreterRuntime::update_mdp_for_ret), 949 return_bci); 950 ldp(zr, return_bci, Address(post(sp, 2 * wordSize))); 951 } 952 953 954 void InterpreterMacroAssembler::profile_taken_branch(Register mdp, 955 Register bumped_count) { 956 if (ProfileInterpreter) { 957 Label profile_continue; 958 959 // If no method data exists, go to profile_continue. 960 // Otherwise, assign to mdp 961 test_method_data_pointer(mdp, profile_continue); 962 963 // We are taking a branch. Increment the taken count. 964 // We inline increment_mdp_data_at to return bumped_count in a register 965 //increment_mdp_data_at(mdp, in_bytes(JumpData::taken_offset())); 966 Address data(mdp, in_bytes(JumpData::taken_offset())); 967 ldr(bumped_count, data); 968 assert(DataLayout::counter_increment == 1, 969 "flow-free idiom only works with 1"); 970 // Intel does this to catch overflow 971 // addptr(bumped_count, DataLayout::counter_increment); 972 // sbbptr(bumped_count, 0); 973 // so we do this 974 adds(bumped_count, bumped_count, DataLayout::counter_increment); 975 Label L; 976 br(Assembler::CS, L); // skip store if counter overflow 977 str(bumped_count, data); 978 bind(L); 979 // The method data pointer needs to be updated to reflect the new target. 980 update_mdp_by_offset(mdp, in_bytes(JumpData::displacement_offset())); 981 bind(profile_continue); 982 } 983 } 984 985 986 void InterpreterMacroAssembler::profile_not_taken_branch(Register mdp) { 987 if (ProfileInterpreter) { 988 Label profile_continue; 989 990 // If no method data exists, go to profile_continue. 991 test_method_data_pointer(mdp, profile_continue); 992 993 // We are taking a branch. Increment the not taken count. 994 increment_mdp_data_at(mdp, in_bytes(BranchData::not_taken_offset())); 995 996 // The method data pointer needs to be updated to correspond to 997 // the next bytecode 998 update_mdp_by_constant(mdp, in_bytes(BranchData::branch_data_size())); 999 bind(profile_continue); 1000 } 1001 } 1002 1003 1004 void InterpreterMacroAssembler::profile_call(Register mdp) { 1005 if (ProfileInterpreter) { 1006 Label profile_continue; 1007 1008 // If no method data exists, go to profile_continue. 1009 test_method_data_pointer(mdp, profile_continue); 1010 1011 // We are making a call. Increment the count. 1012 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1013 1014 // The method data pointer needs to be updated to reflect the new target. 1015 update_mdp_by_constant(mdp, in_bytes(CounterData::counter_data_size())); 1016 bind(profile_continue); 1017 } 1018 } 1019 1020 void InterpreterMacroAssembler::profile_final_call(Register mdp) { 1021 if (ProfileInterpreter) { 1022 Label profile_continue; 1023 1024 // If no method data exists, go to profile_continue. 1025 test_method_data_pointer(mdp, profile_continue); 1026 1027 // We are making a call. Increment the count. 1028 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1029 1030 // The method data pointer needs to be updated to reflect the new target. 1031 update_mdp_by_constant(mdp, 1032 in_bytes(VirtualCallData:: 1033 virtual_call_data_size())); 1034 bind(profile_continue); 1035 } 1036 } 1037 1038 1039 void InterpreterMacroAssembler::profile_virtual_call(Register receiver, 1040 Register mdp, 1041 Register reg2, 1042 bool receiver_can_be_null) { 1043 if (ProfileInterpreter) { 1044 Label profile_continue; 1045 1046 // If no method data exists, go to profile_continue. 1047 test_method_data_pointer(mdp, profile_continue); 1048 1049 Label skip_receiver_profile; 1050 if (receiver_can_be_null) { 1051 Label not_null; 1052 // We are making a call. Increment the count for null receiver. 1053 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1054 b(skip_receiver_profile); 1055 bind(not_null); 1056 } 1057 1058 // Record the receiver type. 1059 record_klass_in_profile(receiver, mdp, reg2, true); 1060 bind(skip_receiver_profile); 1061 1062 // The method data pointer needs to be updated to reflect the new target. 1063 update_mdp_by_constant(mdp, 1064 in_bytes(VirtualCallData:: 1065 virtual_call_data_size())); 1066 bind(profile_continue); 1067 } 1068 } 1069 1070 // This routine creates a state machine for updating the multi-row 1071 // type profile at a virtual call site (or other type-sensitive bytecode). 1072 // The machine visits each row (of receiver/count) until the receiver type 1073 // is found, or until it runs out of rows. At the same time, it remembers 1074 // the location of the first empty row. (An empty row records null for its 1075 // receiver, and can be allocated for a newly-observed receiver type.) 1076 // Because there are two degrees of freedom in the state, a simple linear 1077 // search will not work; it must be a decision tree. Hence this helper 1078 // function is recursive, to generate the required tree structured code. 1079 // It's the interpreter, so we are trading off code space for speed. 1080 // See below for example code. 1081 void InterpreterMacroAssembler::record_klass_in_profile_helper( 1082 Register receiver, Register mdp, 1083 Register reg2, int start_row, 1084 Label& done, bool is_virtual_call) { 1085 if (TypeProfileWidth == 0) { 1086 if (is_virtual_call) { 1087 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1088 } 1089 return; 1090 } 1091 1092 int last_row = VirtualCallData::row_limit() - 1; 1093 assert(start_row <= last_row, "must be work left to do"); 1094 // Test this row for both the receiver and for null. 1095 // Take any of three different outcomes: 1096 // 1. found receiver => increment count and goto done 1097 // 2. found null => keep looking for case 1, maybe allocate this cell 1098 // 3. found something else => keep looking for cases 1 and 2 1099 // Case 3 is handled by a recursive call. 1100 for (int row = start_row; row <= last_row; row++) { 1101 Label next_test; 1102 bool test_for_null_also = (row == start_row); 1103 1104 // See if the receiver is receiver[n]. 1105 int recvr_offset = in_bytes(VirtualCallData::receiver_offset(row)); 1106 test_mdp_data_at(mdp, recvr_offset, receiver, 1107 (test_for_null_also ? reg2 : noreg), 1108 next_test); 1109 // (Reg2 now contains the receiver from the CallData.) 1110 1111 // The receiver is receiver[n]. Increment count[n]. 1112 int count_offset = in_bytes(VirtualCallData::receiver_count_offset(row)); 1113 increment_mdp_data_at(mdp, count_offset); 1114 b(done); 1115 bind(next_test); 1116 1117 if (test_for_null_also) { 1118 Label found_null; 1119 // Failed the equality check on receiver[n]... Test for null. 1120 if (start_row == last_row) { 1121 // The only thing left to do is handle the null case. 1122 if (is_virtual_call) { 1123 cbz(reg2, found_null); 1124 // Receiver did not match any saved receiver and there is no empty row for it. 1125 // Increment total counter to indicate polymorphic case. 1126 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1127 b(done); 1128 bind(found_null); 1129 } else { 1130 cbz(reg2, done); 1131 } 1132 break; 1133 } 1134 // Since null is rare, make it be the branch-taken case. 1135 cbz(reg2,found_null); 1136 1137 // Put all the "Case 3" tests here. 1138 record_klass_in_profile_helper(receiver, mdp, reg2, start_row + 1, done, is_virtual_call); 1139 1140 // Found a null. Keep searching for a matching receiver, 1141 // but remember that this is an empty (unused) slot. 1142 bind(found_null); 1143 } 1144 } 1145 1146 // In the fall-through case, we found no matching receiver, but we 1147 // observed the receiver[start_row] is NULL. 1148 1149 // Fill in the receiver field and increment the count. 1150 int recvr_offset = in_bytes(VirtualCallData::receiver_offset(start_row)); 1151 set_mdp_data_at(mdp, recvr_offset, receiver); 1152 int count_offset = in_bytes(VirtualCallData::receiver_count_offset(start_row)); 1153 mov(reg2, DataLayout::counter_increment); 1154 set_mdp_data_at(mdp, count_offset, reg2); 1155 if (start_row > 0) { 1156 b(done); 1157 } 1158 } 1159 1160 // Example state machine code for three profile rows: 1161 // // main copy of decision tree, rooted at row[1] 1162 // if (row[0].rec == rec) { row[0].incr(); goto done; } 1163 // if (row[0].rec != NULL) { 1164 // // inner copy of decision tree, rooted at row[1] 1165 // if (row[1].rec == rec) { row[1].incr(); goto done; } 1166 // if (row[1].rec != NULL) { 1167 // // degenerate decision tree, rooted at row[2] 1168 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1169 // if (row[2].rec != NULL) { count.incr(); goto done; } // overflow 1170 // row[2].init(rec); goto done; 1171 // } else { 1172 // // remember row[1] is empty 1173 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1174 // row[1].init(rec); goto done; 1175 // } 1176 // } else { 1177 // // remember row[0] is empty 1178 // if (row[1].rec == rec) { row[1].incr(); goto done; } 1179 // if (row[2].rec == rec) { row[2].incr(); goto done; } 1180 // row[0].init(rec); goto done; 1181 // } 1182 // done: 1183 1184 void InterpreterMacroAssembler::record_klass_in_profile(Register receiver, 1185 Register mdp, Register reg2, 1186 bool is_virtual_call) { 1187 assert(ProfileInterpreter, "must be profiling"); 1188 Label done; 1189 1190 record_klass_in_profile_helper(receiver, mdp, reg2, 0, done, is_virtual_call); 1191 1192 bind (done); 1193 } 1194 1195 void InterpreterMacroAssembler::profile_ret(Register return_bci, 1196 Register mdp) { 1197 if (ProfileInterpreter) { 1198 Label profile_continue; 1199 uint row; 1200 1201 // If no method data exists, go to profile_continue. 1202 test_method_data_pointer(mdp, profile_continue); 1203 1204 // Update the total ret count. 1205 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset())); 1206 1207 for (row = 0; row < RetData::row_limit(); row++) { 1208 Label next_test; 1209 1210 // See if return_bci is equal to bci[n]: 1211 test_mdp_data_at(mdp, 1212 in_bytes(RetData::bci_offset(row)), 1213 return_bci, noreg, 1214 next_test); 1215 1216 // return_bci is equal to bci[n]. Increment the count. 1217 increment_mdp_data_at(mdp, in_bytes(RetData::bci_count_offset(row))); 1218 1219 // The method data pointer needs to be updated to reflect the new target. 1220 update_mdp_by_offset(mdp, 1221 in_bytes(RetData::bci_displacement_offset(row))); 1222 b(profile_continue); 1223 bind(next_test); 1224 } 1225 1226 update_mdp_for_ret(return_bci); 1227 1228 bind(profile_continue); 1229 } 1230 } 1231 1232 void InterpreterMacroAssembler::profile_null_seen(Register mdp) { 1233 if (ProfileInterpreter) { 1234 Label profile_continue; 1235 1236 // If no method data exists, go to profile_continue. 1237 test_method_data_pointer(mdp, profile_continue); 1238 1239 set_mdp_flag_at(mdp, BitData::null_seen_byte_constant()); 1240 1241 // The method data pointer needs to be updated. 1242 int mdp_delta = in_bytes(BitData::bit_data_size()); 1243 if (TypeProfileCasts) { 1244 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size()); 1245 } 1246 update_mdp_by_constant(mdp, mdp_delta); 1247 1248 bind(profile_continue); 1249 } 1250 } 1251 1252 void InterpreterMacroAssembler::profile_typecheck_failed(Register mdp) { 1253 if (ProfileInterpreter && TypeProfileCasts) { 1254 Label profile_continue; 1255 1256 // If no method data exists, go to profile_continue. 1257 test_method_data_pointer(mdp, profile_continue); 1258 1259 int count_offset = in_bytes(CounterData::count_offset()); 1260 // Back up the address, since we have already bumped the mdp. 1261 count_offset -= in_bytes(VirtualCallData::virtual_call_data_size()); 1262 1263 // *Decrement* the counter. We expect to see zero or small negatives. 1264 increment_mdp_data_at(mdp, count_offset, true); 1265 1266 bind (profile_continue); 1267 } 1268 } 1269 1270 void InterpreterMacroAssembler::profile_typecheck(Register mdp, Register klass, Register reg2) { 1271 if (ProfileInterpreter) { 1272 Label profile_continue; 1273 1274 // If no method data exists, go to profile_continue. 1275 test_method_data_pointer(mdp, profile_continue); 1276 1277 // The method data pointer needs to be updated. 1278 int mdp_delta = in_bytes(BitData::bit_data_size()); 1279 if (TypeProfileCasts) { 1280 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size()); 1281 1282 // Record the object type. 1283 record_klass_in_profile(klass, mdp, reg2, false); 1284 } 1285 update_mdp_by_constant(mdp, mdp_delta); 1286 1287 bind(profile_continue); 1288 } 1289 } 1290 1291 void InterpreterMacroAssembler::profile_switch_default(Register mdp) { 1292 if (ProfileInterpreter) { 1293 Label profile_continue; 1294 1295 // If no method data exists, go to profile_continue. 1296 test_method_data_pointer(mdp, profile_continue); 1297 1298 // Update the default case count 1299 increment_mdp_data_at(mdp, 1300 in_bytes(MultiBranchData::default_count_offset())); 1301 1302 // The method data pointer needs to be updated. 1303 update_mdp_by_offset(mdp, 1304 in_bytes(MultiBranchData:: 1305 default_displacement_offset())); 1306 1307 bind(profile_continue); 1308 } 1309 } 1310 1311 void InterpreterMacroAssembler::profile_switch_case(Register index, 1312 Register mdp, 1313 Register reg2) { 1314 if (ProfileInterpreter) { 1315 Label profile_continue; 1316 1317 // If no method data exists, go to profile_continue. 1318 test_method_data_pointer(mdp, profile_continue); 1319 1320 // Build the base (index * per_case_size_in_bytes()) + 1321 // case_array_offset_in_bytes() 1322 movw(reg2, in_bytes(MultiBranchData::per_case_size())); 1323 movw(rscratch1, in_bytes(MultiBranchData::case_array_offset())); 1324 Assembler::maddw(index, index, reg2, rscratch1); 1325 1326 // Update the case count 1327 increment_mdp_data_at(mdp, 1328 index, 1329 in_bytes(MultiBranchData::relative_count_offset())); 1330 1331 // The method data pointer needs to be updated. 1332 update_mdp_by_offset(mdp, 1333 index, 1334 in_bytes(MultiBranchData:: 1335 relative_displacement_offset())); 1336 1337 bind(profile_continue); 1338 } 1339 } 1340 1341 void InterpreterMacroAssembler::verify_oop(Register reg, TosState state) { 1342 if (state == atos) { 1343 MacroAssembler::verify_oop(reg); 1344 } 1345 } 1346 1347 void InterpreterMacroAssembler::verify_FPU(int stack_depth, TosState state) { ; } 1348 #endif // !CC_INTERP 1349 1350 1351 void InterpreterMacroAssembler::notify_method_entry() { 1352 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to 1353 // track stack depth. If it is possible to enter interp_only_mode we add 1354 // the code to check if the event should be sent. 1355 if (JvmtiExport::can_post_interpreter_events()) { 1356 Label L; 1357 ldr(r3, Address(rthread, JavaThread::interp_only_mode_offset())); 1358 tst(r3, ~0); 1359 br(Assembler::EQ, L); 1360 call_VM(noreg, CAST_FROM_FN_PTR(address, 1361 InterpreterRuntime::post_method_entry)); 1362 bind(L); 1363 } 1364 1365 { 1366 SkipIfEqual skip(this, &DTraceMethodProbes, false); 1367 get_method(c_rarg1); 1368 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 1369 rthread, c_rarg1); 1370 } 1371 1372 // RedefineClasses() tracing support for obsolete method entry 1373 if (RC_TRACE_IN_RANGE(0x00001000, 0x00002000)) { 1374 get_method(c_rarg1); 1375 call_VM_leaf( 1376 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 1377 rthread, c_rarg1); 1378 } 1379 1380 } 1381 1382 1383 void InterpreterMacroAssembler::notify_method_exit( 1384 TosState state, NotifyMethodExitMode mode) { 1385 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to 1386 // track stack depth. If it is possible to enter interp_only_mode we add 1387 // the code to check if the event should be sent. 1388 if (mode == NotifyJVMTI && JvmtiExport::can_post_interpreter_events()) { 1389 Label L; 1390 // Note: frame::interpreter_frame_result has a dependency on how the 1391 // method result is saved across the call to post_method_exit. If this 1392 // is changed then the interpreter_frame_result implementation will 1393 // need to be updated too. 1394 1395 // For c++ interpreter the result is always stored at a known location in the frame 1396 // template interpreter will leave it on the top of the stack. 1397 NOT_CC_INTERP(push(state);) 1398 ldrw(r3, Address(rthread, JavaThread::interp_only_mode_offset())); 1399 cbz(r3, L); 1400 call_VM(noreg, 1401 CAST_FROM_FN_PTR(address, InterpreterRuntime::post_method_exit)); 1402 bind(L); 1403 NOT_CC_INTERP(pop(state)); 1404 } 1405 1406 { 1407 SkipIfEqual skip(this, &DTraceMethodProbes, false); 1408 NOT_CC_INTERP(push(state)); 1409 get_method(c_rarg1); 1410 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 1411 rthread, c_rarg1); 1412 NOT_CC_INTERP(pop(state)); 1413 } 1414 } 1415 1416 1417 // Jump if ((*counter_addr += increment) & mask) satisfies the condition. 1418 void InterpreterMacroAssembler::increment_mask_and_jump(Address counter_addr, 1419 int increment, Address mask, 1420 Register scratch, Register scratch2, 1421 bool preloaded, Condition cond, 1422 Label* where) { 1423 if (!preloaded) { 1424 ldrw(scratch, counter_addr); 1425 } 1426 add(scratch, scratch, increment); 1427 strw(scratch, counter_addr); 1428 ldrw(scratch2, mask); 1429 ands(scratch, scratch, scratch2); 1430 br(cond, *where); 1431 } 1432 1433 void InterpreterMacroAssembler::call_VM_leaf_base(address entry_point, 1434 int number_of_arguments) { 1435 // interpreter specific 1436 // 1437 // Note: No need to save/restore rbcp & rlocals pointer since these 1438 // are callee saved registers and no blocking/ GC can happen 1439 // in leaf calls. 1440 #ifdef ASSERT 1441 { 1442 Label L; 1443 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 1444 cbz(rscratch1, L); 1445 stop("InterpreterMacroAssembler::call_VM_leaf_base:" 1446 " last_sp != NULL"); 1447 bind(L); 1448 } 1449 #endif /* ASSERT */ 1450 // super call 1451 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments); 1452 } 1453 1454 void InterpreterMacroAssembler::call_VM_base(Register oop_result, 1455 Register java_thread, 1456 Register last_java_sp, 1457 address entry_point, 1458 int number_of_arguments, 1459 bool check_exceptions) { 1460 // interpreter specific 1461 // 1462 // Note: Could avoid restoring locals ptr (callee saved) - however doesn't 1463 // really make a difference for these runtime calls, since they are 1464 // slow anyway. Btw., bcp must be saved/restored since it may change 1465 // due to GC. 1466 // assert(java_thread == noreg , "not expecting a precomputed java thread"); 1467 save_bcp(); 1468 #ifdef ASSERT 1469 { 1470 Label L; 1471 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 1472 cbz(rscratch1, L); 1473 stop("InterpreterMacroAssembler::call_VM_leaf_base:" 1474 " last_sp != NULL"); 1475 bind(L); 1476 } 1477 #endif /* ASSERT */ 1478 // super call 1479 MacroAssembler::call_VM_base(oop_result, noreg, last_java_sp, 1480 entry_point, number_of_arguments, 1481 check_exceptions); 1482 // interpreter specific 1483 restore_bcp(); 1484 restore_locals(); 1485 } 1486 1487 void InterpreterMacroAssembler::profile_obj_type(Register obj, const Address& mdo_addr) { 1488 Label update, next, none; 1489 1490 verify_oop(obj); 1491 1492 cbnz(obj, update); 1493 orptr(mdo_addr, TypeEntries::null_seen); 1494 b(next); 1495 1496 bind(update); 1497 load_klass(obj, obj); 1498 1499 ldr(rscratch1, mdo_addr); 1500 eor(obj, obj, rscratch1); 1501 tst(obj, TypeEntries::type_klass_mask); 1502 br(Assembler::EQ, next); // klass seen before, nothing to 1503 // do. The unknown bit may have been 1504 // set already but no need to check. 1505 1506 tst(obj, TypeEntries::type_unknown); 1507 br(Assembler::NE, next); // already unknown. Nothing to do anymore. 1508 1509 ldr(rscratch1, mdo_addr); 1510 cbz(rscratch1, none); 1511 cmp(rscratch1, TypeEntries::null_seen); 1512 br(Assembler::EQ, none); 1513 // There is a chance that the checks above (re-reading profiling 1514 // data from memory) fail if another thread has just set the 1515 // profiling to this obj's klass 1516 ldr(rscratch1, mdo_addr); 1517 eor(obj, obj, rscratch1); 1518 tst(obj, TypeEntries::type_klass_mask); 1519 br(Assembler::EQ, next); 1520 1521 // different than before. Cannot keep accurate profile. 1522 orptr(mdo_addr, TypeEntries::type_unknown); 1523 b(next); 1524 1525 bind(none); 1526 // first time here. Set profile type. 1527 str(obj, mdo_addr); 1528 1529 bind(next); 1530 } 1531 1532 void InterpreterMacroAssembler::profile_arguments_type(Register mdp, Register callee, Register tmp, bool is_virtual) { 1533 if (!ProfileInterpreter) { 1534 return; 1535 } 1536 1537 if (MethodData::profile_arguments() || MethodData::profile_return()) { 1538 Label profile_continue; 1539 1540 test_method_data_pointer(mdp, profile_continue); 1541 1542 int off_to_start = is_virtual ? in_bytes(VirtualCallData::virtual_call_data_size()) : in_bytes(CounterData::counter_data_size()); 1543 1544 ldrb(rscratch1, Address(mdp, in_bytes(DataLayout::tag_offset()) - off_to_start)); 1545 cmp(rscratch1, is_virtual ? DataLayout::virtual_call_type_data_tag : DataLayout::call_type_data_tag); 1546 br(Assembler::NE, profile_continue); 1547 1548 if (MethodData::profile_arguments()) { 1549 Label done; 1550 int off_to_args = in_bytes(TypeEntriesAtCall::args_data_offset()); 1551 1552 for (int i = 0; i < TypeProfileArgsLimit; i++) { 1553 if (i > 0 || MethodData::profile_return()) { 1554 // If return value type is profiled we may have no argument to profile 1555 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset()))); 1556 sub(tmp, tmp, i*TypeStackSlotEntries::per_arg_count()); 1557 cmp(tmp, TypeStackSlotEntries::per_arg_count()); 1558 add(rscratch1, mdp, off_to_args); 1559 br(Assembler::LT, done); 1560 } 1561 ldr(tmp, Address(callee, Method::const_offset())); 1562 load_unsigned_short(tmp, Address(tmp, ConstMethod::size_of_parameters_offset())); 1563 // stack offset o (zero based) from the start of the argument 1564 // list, for n arguments translates into offset n - o - 1 from 1565 // the end of the argument list 1566 ldr(rscratch1, Address(mdp, in_bytes(TypeEntriesAtCall::stack_slot_offset(i)))); 1567 sub(tmp, tmp, rscratch1); 1568 sub(tmp, tmp, 1); 1569 Address arg_addr = argument_address(tmp); 1570 ldr(tmp, arg_addr); 1571 1572 Address mdo_arg_addr(mdp, in_bytes(TypeEntriesAtCall::argument_type_offset(i))); 1573 profile_obj_type(tmp, mdo_arg_addr); 1574 1575 int to_add = in_bytes(TypeStackSlotEntries::per_arg_size()); 1576 off_to_args += to_add; 1577 } 1578 1579 if (MethodData::profile_return()) { 1580 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset()))); 1581 sub(tmp, tmp, TypeProfileArgsLimit*TypeStackSlotEntries::per_arg_count()); 1582 } 1583 1584 add(rscratch1, mdp, off_to_args); 1585 bind(done); 1586 mov(mdp, rscratch1); 1587 1588 if (MethodData::profile_return()) { 1589 // We're right after the type profile for the last 1590 // argument. tmp is the number of cells left in the 1591 // CallTypeData/VirtualCallTypeData to reach its end. Non null 1592 // if there's a return to profile. 1593 assert(ReturnTypeEntry::static_cell_count() < TypeStackSlotEntries::per_arg_count(), "can't move past ret type"); 1594 add(mdp, mdp, tmp, LSL, exact_log2(DataLayout::cell_size)); 1595 } 1596 str(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize)); 1597 } else { 1598 assert(MethodData::profile_return(), "either profile call args or call ret"); 1599 update_mdp_by_constant(mdp, in_bytes(TypeEntriesAtCall::return_only_size())); 1600 } 1601 1602 // mdp points right after the end of the 1603 // CallTypeData/VirtualCallTypeData, right after the cells for the 1604 // return value type if there's one 1605 1606 bind(profile_continue); 1607 } 1608 } 1609 1610 void InterpreterMacroAssembler::profile_return_type(Register mdp, Register ret, Register tmp) { 1611 assert_different_registers(mdp, ret, tmp, rbcp); 1612 if (ProfileInterpreter && MethodData::profile_return()) { 1613 Label profile_continue, done; 1614 1615 test_method_data_pointer(mdp, profile_continue); 1616 1617 if (MethodData::profile_return_jsr292_only()) { 1618 assert(Method::intrinsic_id_size_in_bytes() == 2, "assuming Method::_intrinsic_id is u2"); 1619 1620 // If we don't profile all invoke bytecodes we must make sure 1621 // it's a bytecode we indeed profile. We can't go back to the 1622 // begining of the ProfileData we intend to update to check its 1623 // type because we're right after it and we don't known its 1624 // length 1625 Label do_profile; 1626 ldrb(rscratch1, Address(rbcp, 0)); 1627 cmp(rscratch1, Bytecodes::_invokedynamic); 1628 br(Assembler::EQ, do_profile); 1629 cmp(rscratch1, Bytecodes::_invokehandle); 1630 br(Assembler::EQ, do_profile); 1631 get_method(tmp); 1632 ldrh(rscratch1, Address(tmp, Method::intrinsic_id_offset_in_bytes())); 1633 cmp(rscratch1, vmIntrinsics::_compiledLambdaForm); 1634 br(Assembler::NE, profile_continue); 1635 1636 bind(do_profile); 1637 } 1638 1639 Address mdo_ret_addr(mdp, -in_bytes(ReturnTypeEntry::size())); 1640 mov(tmp, ret); 1641 profile_obj_type(tmp, mdo_ret_addr); 1642 1643 bind(profile_continue); 1644 } 1645 } 1646 1647 void InterpreterMacroAssembler::profile_parameters_type(Register mdp, Register tmp1, Register tmp2) { 1648 if (ProfileInterpreter && MethodData::profile_parameters()) { 1649 Label profile_continue, done; 1650 1651 test_method_data_pointer(mdp, profile_continue); 1652 1653 // Load the offset of the area within the MDO used for 1654 // parameters. If it's negative we're not profiling any parameters 1655 ldr(tmp1, Address(mdp, in_bytes(MethodData::parameters_type_data_di_offset()) - in_bytes(MethodData::data_offset()))); 1656 cmp(tmp1, 0u); 1657 br(Assembler::LT, profile_continue); 1658 1659 // Compute a pointer to the area for parameters from the offset 1660 // and move the pointer to the slot for the last 1661 // parameters. Collect profiling from last parameter down. 1662 // mdo start + parameters offset + array length - 1 1663 add(mdp, mdp, tmp1); 1664 ldr(tmp1, Address(mdp, ArrayData::array_len_offset())); 1665 sub(tmp1, tmp1, TypeStackSlotEntries::per_arg_count()); 1666 1667 Label loop; 1668 bind(loop); 1669 1670 int off_base = in_bytes(ParametersTypeData::stack_slot_offset(0)); 1671 int type_base = in_bytes(ParametersTypeData::type_offset(0)); 1672 int per_arg_scale = exact_log2(DataLayout::cell_size); 1673 add(rscratch1, mdp, off_base); 1674 add(rscratch2, mdp, type_base); 1675 1676 Address arg_off(rscratch1, tmp1, Address::lsl(per_arg_scale)); 1677 Address arg_type(rscratch2, tmp1, Address::lsl(per_arg_scale)); 1678 1679 // load offset on the stack from the slot for this parameter 1680 ldr(tmp2, arg_off); 1681 neg(tmp2, tmp2); 1682 // read the parameter from the local area 1683 ldr(tmp2, Address(rlocals, tmp2, Address::lsl(Interpreter::logStackElementSize))); 1684 1685 // profile the parameter 1686 profile_obj_type(tmp2, arg_type); 1687 1688 // go to next parameter 1689 subs(tmp1, tmp1, TypeStackSlotEntries::per_arg_count()); 1690 br(Assembler::GE, loop); 1691 1692 bind(profile_continue); 1693 } 1694 }