1 /* 2 * Copyright (c) 2007, 2012, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "interpreter/bytecodeHistogram.hpp" 28 #include "interpreter/cppInterpreter.hpp" 29 #include "interpreter/interpreter.hpp" 30 #include "interpreter/interpreterGenerator.hpp" 31 #include "interpreter/interpreterRuntime.hpp" 32 #include "oops/arrayOop.hpp" 33 #include "oops/methodData.hpp" 34 #include "oops/method.hpp" 35 #include "oops/oop.inline.hpp" 36 #include "prims/jvmtiExport.hpp" 37 #include "prims/jvmtiThreadState.hpp" 38 #include "runtime/arguments.hpp" 39 #include "runtime/deoptimization.hpp" 40 #include "runtime/frame.inline.hpp" 41 #include "runtime/interfaceSupport.hpp" 42 #include "runtime/sharedRuntime.hpp" 43 #include "runtime/stubRoutines.hpp" 44 #include "runtime/synchronizer.hpp" 45 #include "runtime/timer.hpp" 46 #include "runtime/vframeArray.hpp" 47 #include "utilities/debug.hpp" 48 #include "utilities/macros.hpp" 49 #ifdef SHARK 50 #include "shark/shark_globals.hpp" 51 #endif 52 53 #ifdef CC_INTERP 54 55 // Routine exists to make tracebacks look decent in debugger 56 // while we are recursed in the frame manager/c++ interpreter. 57 // We could use an address in the frame manager but having 58 // frames look natural in the debugger is a plus. 59 extern "C" void RecursiveInterpreterActivation(interpreterState istate ) 60 { 61 // 62 ShouldNotReachHere(); 63 } 64 65 66 #define __ _masm-> 67 #define STATE(field_name) (Address(state, byte_offset_of(BytecodeInterpreter, field_name))) 68 69 Label fast_accessor_slow_entry_path; // fast accessor methods need to be able to jmp to unsynchronized 70 // c++ interpreter entry point this holds that entry point label. 71 72 // default registers for state and sender_sp 73 // state and sender_sp are the same on 32bit because we have no choice. 74 // state could be rsi on 64bit but it is an arg reg and not callee save 75 // so r13 is better choice. 76 77 const Register state = NOT_LP64(rsi) LP64_ONLY(r13); 78 const Register sender_sp_on_entry = NOT_LP64(rsi) LP64_ONLY(r13); 79 80 // NEEDED for JVMTI? 81 // address AbstractInterpreter::_remove_activation_preserving_args_entry; 82 83 static address unctrap_frame_manager_entry = NULL; 84 85 static address deopt_frame_manager_return_atos = NULL; 86 static address deopt_frame_manager_return_btos = NULL; 87 static address deopt_frame_manager_return_itos = NULL; 88 static address deopt_frame_manager_return_ltos = NULL; 89 static address deopt_frame_manager_return_ftos = NULL; 90 static address deopt_frame_manager_return_dtos = NULL; 91 static address deopt_frame_manager_return_vtos = NULL; 92 93 int AbstractInterpreter::BasicType_as_index(BasicType type) { 94 int i = 0; 95 switch (type) { 96 case T_BOOLEAN: i = 0; break; 97 case T_CHAR : i = 1; break; 98 case T_BYTE : i = 2; break; 99 case T_SHORT : i = 3; break; 100 case T_INT : i = 4; break; 101 case T_VOID : i = 5; break; 102 case T_FLOAT : i = 8; break; 103 case T_LONG : i = 9; break; 104 case T_DOUBLE : i = 6; break; 105 case T_OBJECT : // fall through 106 case T_ARRAY : i = 7; break; 107 default : ShouldNotReachHere(); 108 } 109 assert(0 <= i && i < AbstractInterpreter::number_of_result_handlers, "index out of bounds"); 110 return i; 111 } 112 113 // Is this pc anywhere within code owned by the interpreter? 114 // This only works for pc that might possibly be exposed to frame 115 // walkers. It clearly misses all of the actual c++ interpreter 116 // implementation 117 bool CppInterpreter::contains(address pc) { 118 return (_code->contains(pc) || 119 pc == CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation)); 120 } 121 122 123 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) { 124 address entry = __ pc(); 125 switch (type) { 126 case T_BOOLEAN: __ c2bool(rax); break; 127 case T_CHAR : __ andl(rax, 0xFFFF); break; 128 case T_BYTE : __ sign_extend_byte (rax); break; 129 case T_SHORT : __ sign_extend_short(rax); break; 130 case T_VOID : // fall thru 131 case T_LONG : // fall thru 132 case T_INT : /* nothing to do */ break; 133 134 case T_DOUBLE : 135 case T_FLOAT : 136 { 137 const Register t = InterpreterRuntime::SignatureHandlerGenerator::temp(); 138 __ pop(t); // remove return address first 139 // Must return a result for interpreter or compiler. In SSE 140 // mode, results are returned in xmm0 and the FPU stack must 141 // be empty. 142 if (type == T_FLOAT && UseSSE >= 1) { 143 #ifndef _LP64 144 // Load ST0 145 __ fld_d(Address(rsp, 0)); 146 // Store as float and empty fpu stack 147 __ fstp_s(Address(rsp, 0)); 148 #endif // !_LP64 149 // and reload 150 __ movflt(xmm0, Address(rsp, 0)); 151 } else if (type == T_DOUBLE && UseSSE >= 2 ) { 152 __ movdbl(xmm0, Address(rsp, 0)); 153 } else { 154 // restore ST0 155 __ fld_d(Address(rsp, 0)); 156 } 157 // and pop the temp 158 __ addptr(rsp, 2 * wordSize); 159 __ push(t); // restore return address 160 } 161 break; 162 case T_OBJECT : 163 // retrieve result from frame 164 __ movptr(rax, STATE(_oop_temp)); 165 // and verify it 166 __ verify_oop(rax); 167 break; 168 default : ShouldNotReachHere(); 169 } 170 __ ret(0); // return from result handler 171 return entry; 172 } 173 174 // tosca based result to c++ interpreter stack based result. 175 // Result goes to top of native stack. 176 177 #undef EXTEND // SHOULD NOT BE NEEDED 178 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) { 179 // A result is in the tosca (abi result) from either a native method call or compiled 180 // code. Place this result on the java expression stack so C++ interpreter can use it. 181 address entry = __ pc(); 182 183 const Register t = InterpreterRuntime::SignatureHandlerGenerator::temp(); 184 __ pop(t); // remove return address first 185 switch (type) { 186 case T_VOID: 187 break; 188 case T_BOOLEAN: 189 #ifdef EXTEND 190 __ c2bool(rax); 191 #endif 192 __ push(rax); 193 break; 194 case T_CHAR : 195 #ifdef EXTEND 196 __ andl(rax, 0xFFFF); 197 #endif 198 __ push(rax); 199 break; 200 case T_BYTE : 201 #ifdef EXTEND 202 __ sign_extend_byte (rax); 203 #endif 204 __ push(rax); 205 break; 206 case T_SHORT : 207 #ifdef EXTEND 208 __ sign_extend_short(rax); 209 #endif 210 __ push(rax); 211 break; 212 case T_LONG : 213 __ push(rdx); // pushes useless junk on 64bit 214 __ push(rax); 215 break; 216 case T_INT : 217 __ push(rax); 218 break; 219 case T_FLOAT : 220 // Result is in ST(0)/xmm0 221 __ subptr(rsp, wordSize); 222 if ( UseSSE < 1) { 223 __ fstp_s(Address(rsp, 0)); 224 } else { 225 __ movflt(Address(rsp, 0), xmm0); 226 } 227 break; 228 case T_DOUBLE : 229 __ subptr(rsp, 2*wordSize); 230 if ( UseSSE < 2 ) { 231 __ fstp_d(Address(rsp, 0)); 232 } else { 233 __ movdbl(Address(rsp, 0), xmm0); 234 } 235 break; 236 case T_OBJECT : 237 __ verify_oop(rax); // verify it 238 __ push(rax); 239 break; 240 default : ShouldNotReachHere(); 241 } 242 __ jmp(t); // return from result handler 243 return entry; 244 } 245 246 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) { 247 // A result is in the java expression stack of the interpreted method that has just 248 // returned. Place this result on the java expression stack of the caller. 249 // 250 // The current interpreter activation in rsi/r13 is for the method just returning its 251 // result. So we know that the result of this method is on the top of the current 252 // execution stack (which is pre-pushed) and will be return to the top of the caller 253 // stack. The top of the callers stack is the bottom of the locals of the current 254 // activation. 255 // Because of the way activation are managed by the frame manager the value of rsp is 256 // below both the stack top of the current activation and naturally the stack top 257 // of the calling activation. This enable this routine to leave the return address 258 // to the frame manager on the stack and do a vanilla return. 259 // 260 // On entry: rsi/r13 - interpreter state of activation returning a (potential) result 261 // On Return: rsi/r13 - unchanged 262 // rax - new stack top for caller activation (i.e. activation in _prev_link) 263 // 264 // Can destroy rdx, rcx. 265 // 266 267 address entry = __ pc(); 268 const Register t = InterpreterRuntime::SignatureHandlerGenerator::temp(); 269 switch (type) { 270 case T_VOID: 271 __ movptr(rax, STATE(_locals)); // pop parameters get new stack value 272 __ addptr(rax, wordSize); // account for prepush before we return 273 break; 274 case T_FLOAT : 275 case T_BOOLEAN: 276 case T_CHAR : 277 case T_BYTE : 278 case T_SHORT : 279 case T_INT : 280 // 1 word result 281 __ movptr(rdx, STATE(_stack)); 282 __ movptr(rax, STATE(_locals)); // address for result 283 __ movl(rdx, Address(rdx, wordSize)); // get result 284 __ movptr(Address(rax, 0), rdx); // and store it 285 break; 286 case T_LONG : 287 case T_DOUBLE : 288 // return top two words on current expression stack to caller's expression stack 289 // The caller's expression stack is adjacent to the current frame manager's intepretState 290 // except we allocated one extra word for this intepretState so we won't overwrite it 291 // when we return a two word result. 292 293 __ movptr(rax, STATE(_locals)); // address for result 294 __ movptr(rcx, STATE(_stack)); 295 __ subptr(rax, wordSize); // need addition word besides locals[0] 296 __ movptr(rdx, Address(rcx, 2*wordSize)); // get result word (junk in 64bit) 297 __ movptr(Address(rax, wordSize), rdx); // and store it 298 __ movptr(rdx, Address(rcx, wordSize)); // get result word 299 __ movptr(Address(rax, 0), rdx); // and store it 300 break; 301 case T_OBJECT : 302 __ movptr(rdx, STATE(_stack)); 303 __ movptr(rax, STATE(_locals)); // address for result 304 __ movptr(rdx, Address(rdx, wordSize)); // get result 305 __ verify_oop(rdx); // verify it 306 __ movptr(Address(rax, 0), rdx); // and store it 307 break; 308 default : ShouldNotReachHere(); 309 } 310 __ ret(0); 311 return entry; 312 } 313 314 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) { 315 // A result is in the java expression stack of the interpreted method that has just 316 // returned. Place this result in the native abi that the caller expects. 317 // 318 // Similar to generate_stack_to_stack_converter above. Called at a similar time from the 319 // frame manager execept in this situation the caller is native code (c1/c2/call_stub) 320 // and so rather than return result onto caller's java expression stack we return the 321 // result in the expected location based on the native abi. 322 // On entry: rsi/r13 - interpreter state of activation returning a (potential) result 323 // On Return: rsi/r13 - unchanged 324 // Other registers changed [rax/rdx/ST(0) as needed for the result returned] 325 326 address entry = __ pc(); 327 switch (type) { 328 case T_VOID: 329 break; 330 case T_BOOLEAN: 331 case T_CHAR : 332 case T_BYTE : 333 case T_SHORT : 334 case T_INT : 335 __ movptr(rdx, STATE(_stack)); // get top of stack 336 __ movl(rax, Address(rdx, wordSize)); // get result word 1 337 break; 338 case T_LONG : 339 __ movptr(rdx, STATE(_stack)); // get top of stack 340 __ movptr(rax, Address(rdx, wordSize)); // get result low word 341 NOT_LP64(__ movl(rdx, Address(rdx, 2*wordSize));) // get result high word 342 break; 343 case T_FLOAT : 344 __ movptr(rdx, STATE(_stack)); // get top of stack 345 if ( UseSSE >= 1) { 346 __ movflt(xmm0, Address(rdx, wordSize)); 347 } else { 348 __ fld_s(Address(rdx, wordSize)); // pushd float result 349 } 350 break; 351 case T_DOUBLE : 352 __ movptr(rdx, STATE(_stack)); // get top of stack 353 if ( UseSSE > 1) { 354 __ movdbl(xmm0, Address(rdx, wordSize)); 355 } else { 356 __ fld_d(Address(rdx, wordSize)); // push double result 357 } 358 break; 359 case T_OBJECT : 360 __ movptr(rdx, STATE(_stack)); // get top of stack 361 __ movptr(rax, Address(rdx, wordSize)); // get result word 1 362 __ verify_oop(rax); // verify it 363 break; 364 default : ShouldNotReachHere(); 365 } 366 __ ret(0); 367 return entry; 368 } 369 370 address CppInterpreter::return_entry(TosState state, int length) { 371 // make it look good in the debugger 372 return CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation); 373 } 374 375 address CppInterpreter::deopt_entry(TosState state, int length) { 376 address ret = NULL; 377 if (length != 0) { 378 switch (state) { 379 case atos: ret = deopt_frame_manager_return_atos; break; 380 case btos: ret = deopt_frame_manager_return_btos; break; 381 case ctos: 382 case stos: 383 case itos: ret = deopt_frame_manager_return_itos; break; 384 case ltos: ret = deopt_frame_manager_return_ltos; break; 385 case ftos: ret = deopt_frame_manager_return_ftos; break; 386 case dtos: ret = deopt_frame_manager_return_dtos; break; 387 case vtos: ret = deopt_frame_manager_return_vtos; break; 388 } 389 } else { 390 ret = unctrap_frame_manager_entry; // re-execute the bytecode ( e.g. uncommon trap) 391 } 392 assert(ret != NULL, "Not initialized"); 393 return ret; 394 } 395 396 // C++ Interpreter 397 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state, 398 const Register locals, 399 const Register sender_sp, 400 bool native) { 401 402 // On entry the "locals" argument points to locals[0] (or where it would be in case no locals in 403 // a static method). "state" contains any previous frame manager state which we must save a link 404 // to in the newly generated state object. On return "state" is a pointer to the newly allocated 405 // state object. We must allocate and initialize a new interpretState object and the method 406 // expression stack. Because the returned result (if any) of the method will be placed on the caller's 407 // expression stack and this will overlap with locals[0] (and locals[1] if double/long) we must 408 // be sure to leave space on the caller's stack so that this result will not overwrite values when 409 // locals[0] and locals[1] do not exist (and in fact are return address and saved rbp). So when 410 // we are non-native we in essence ensure that locals[0-1] exist. We play an extra trick in 411 // non-product builds and initialize this last local with the previous interpreterState as 412 // this makes things look real nice in the debugger. 413 414 // State on entry 415 // Assumes locals == &locals[0] 416 // Assumes state == any previous frame manager state (assuming call path from c++ interpreter) 417 // Assumes rax = return address 418 // rcx == senders_sp 419 // rbx == method 420 // Modifies rcx, rdx, rax 421 // Returns: 422 // state == address of new interpreterState 423 // rsp == bottom of method's expression stack. 424 425 const Address const_offset (rbx, Method::const_offset()); 426 427 428 // On entry sp is the sender's sp. This includes the space for the arguments 429 // that the sender pushed. If the sender pushed no args (a static) and the 430 // caller returns a long then we need two words on the sender's stack which 431 // are not present (although when we return a restore full size stack the 432 // space will be present). If we didn't allocate two words here then when 433 // we "push" the result of the caller's stack we would overwrite the return 434 // address and the saved rbp. Not good. So simply allocate 2 words now 435 // just to be safe. This is the "static long no_params() method" issue. 436 // See Lo.java for a testcase. 437 // We don't need this for native calls because they return result in 438 // register and the stack is expanded in the caller before we store 439 // the results on the stack. 440 441 if (!native) { 442 #ifdef PRODUCT 443 __ subptr(rsp, 2*wordSize); 444 #else /* PRODUCT */ 445 __ push((int32_t)NULL_WORD); 446 __ push(state); // make it look like a real argument 447 #endif /* PRODUCT */ 448 } 449 450 // Now that we are assure of space for stack result, setup typical linkage 451 452 __ push(rax); 453 __ enter(); 454 455 __ mov(rax, state); // save current state 456 457 __ lea(rsp, Address(rsp, -(int)sizeof(BytecodeInterpreter))); 458 __ mov(state, rsp); 459 460 // rsi/r13 == state/locals rax == prevstate 461 462 // initialize the "shadow" frame so that use since C++ interpreter not directly 463 // recursive. Simpler to recurse but we can't trim expression stack as we call 464 // new methods. 465 __ movptr(STATE(_locals), locals); // state->_locals = locals() 466 __ movptr(STATE(_self_link), state); // point to self 467 __ movptr(STATE(_prev_link), rax); // state->_link = state on entry (NULL or previous state) 468 __ movptr(STATE(_sender_sp), sender_sp); // state->_sender_sp = sender_sp 469 #ifdef _LP64 470 __ movptr(STATE(_thread), r15_thread); // state->_bcp = codes() 471 #else 472 __ get_thread(rax); // get vm's javathread* 473 __ movptr(STATE(_thread), rax); // state->_bcp = codes() 474 #endif // _LP64 475 __ movptr(rdx, Address(rbx, Method::const_offset())); // get constantMethodOop 476 __ lea(rdx, Address(rdx, ConstMethod::codes_offset())); // get code base 477 if (native) { 478 __ movptr(STATE(_bcp), (int32_t)NULL_WORD); // state->_bcp = NULL 479 } else { 480 __ movptr(STATE(_bcp), rdx); // state->_bcp = codes() 481 } 482 __ xorptr(rdx, rdx); 483 __ movptr(STATE(_oop_temp), rdx); // state->_oop_temp = NULL (only really needed for native) 484 __ movptr(STATE(_mdx), rdx); // state->_mdx = NULL 485 __ movptr(rdx, Address(rbx, Method::const_offset())); 486 __ movptr(rdx, Address(rdx, ConstMethod::constants_offset())); 487 __ movptr(rdx, Address(rdx, ConstantPool::cache_offset_in_bytes())); 488 __ movptr(STATE(_constants), rdx); // state->_constants = constants() 489 490 __ movptr(STATE(_method), rbx); // state->_method = method() 491 __ movl(STATE(_msg), (int32_t) BytecodeInterpreter::method_entry); // state->_msg = initial method entry 492 __ movptr(STATE(_result._to_call._callee), (int32_t) NULL_WORD); // state->_result._to_call._callee_callee = NULL 493 494 495 __ movptr(STATE(_monitor_base), rsp); // set monitor block bottom (grows down) this would point to entry [0] 496 // entries run from -1..x where &monitor[x] == 497 498 { 499 // Must not attempt to lock method until we enter interpreter as gc won't be able to find the 500 // initial frame. However we allocate a free monitor so we don't have to shuffle the expression stack 501 // immediately. 502 503 // synchronize method 504 const Address access_flags (rbx, Method::access_flags_offset()); 505 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize; 506 Label not_synced; 507 508 __ movl(rax, access_flags); 509 __ testl(rax, JVM_ACC_SYNCHRONIZED); 510 __ jcc(Assembler::zero, not_synced); 511 512 // Allocate initial monitor and pre initialize it 513 // get synchronization object 514 515 Label done; 516 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 517 __ movl(rax, access_flags); 518 __ testl(rax, JVM_ACC_STATIC); 519 __ movptr(rax, Address(locals, 0)); // get receiver (assume this is frequent case) 520 __ jcc(Assembler::zero, done); 521 __ movptr(rax, Address(rbx, Method::const_offset())); 522 __ movptr(rax, Address(rax, ConstMethod::constants_offset())); 523 __ movptr(rax, Address(rax, ConstantPool::pool_holder_offset_in_bytes())); 524 __ movptr(rax, Address(rax, mirror_offset)); 525 __ bind(done); 526 // add space for monitor & lock 527 __ subptr(rsp, entry_size); // add space for a monitor entry 528 __ movptr(Address(rsp, BasicObjectLock::obj_offset_in_bytes()), rax); // store object 529 __ bind(not_synced); 530 } 531 532 __ movptr(STATE(_stack_base), rsp); // set expression stack base ( == &monitors[-count]) 533 if (native) { 534 __ movptr(STATE(_stack), rsp); // set current expression stack tos 535 __ movptr(STATE(_stack_limit), rsp); 536 } else { 537 __ subptr(rsp, wordSize); // pre-push stack 538 __ movptr(STATE(_stack), rsp); // set current expression stack tos 539 540 // compute full expression stack limit 541 542 __ movptr(rdx, Address(rbx, Method::const_offset())); 543 __ load_unsigned_short(rdx, Address(rdx, ConstMethod::max_stack_offset())); // get size of expression stack in words 544 __ negptr(rdx); // so we can subtract in next step 545 // Allocate expression stack 546 __ lea(rsp, Address(rsp, rdx, Address::times_ptr)); 547 __ movptr(STATE(_stack_limit), rsp); 548 } 549 550 #ifdef _LP64 551 // Make sure stack is properly aligned and sized for the abi 552 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows 553 __ andptr(rsp, -16); // must be 16 byte boundary (see amd64 ABI) 554 #endif // _LP64 555 556 557 558 } 559 560 // Helpers for commoning out cases in the various type of method entries. 561 // 562 563 // increment invocation count & check for overflow 564 // 565 // Note: checking for negative value instead of overflow 566 // so we have a 'sticky' overflow test 567 // 568 // rbx,: method 569 // rcx: invocation counter 570 // 571 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) { 572 573 const Address invocation_counter(rbx, Method::invocation_counter_offset() + InvocationCounter::counter_offset()); 574 const Address backedge_counter (rbx, Method::backedge_counter_offset() + InvocationCounter::counter_offset()); 575 576 if (ProfileInterpreter) { // %%% Merge this into MethodData* 577 __ incrementl(Address(rbx,Method::interpreter_invocation_counter_offset())); 578 } 579 // Update standard invocation counters 580 __ movl(rax, backedge_counter); // load backedge counter 581 582 __ increment(rcx, InvocationCounter::count_increment); 583 __ andl(rax, InvocationCounter::count_mask_value); // mask out the status bits 584 585 __ movl(invocation_counter, rcx); // save invocation count 586 __ addl(rcx, rax); // add both counters 587 588 // profile_method is non-null only for interpreted method so 589 // profile_method != NULL == !native_call 590 // BytecodeInterpreter only calls for native so code is elided. 591 592 __ cmp32(rcx, 593 ExternalAddress((address)&InvocationCounter::InterpreterInvocationLimit)); 594 __ jcc(Assembler::aboveEqual, *overflow); 595 596 } 597 598 void InterpreterGenerator::generate_counter_overflow(Label* do_continue) { 599 600 // C++ interpreter on entry 601 // rsi/r13 - new interpreter state pointer 602 // rbp - interpreter frame pointer 603 // rbx - method 604 605 // On return (i.e. jump to entry_point) [ back to invocation of interpreter ] 606 // rbx, - method 607 // rcx - rcvr (assuming there is one) 608 // top of stack return address of interpreter caller 609 // rsp - sender_sp 610 611 // C++ interpreter only 612 // rsi/r13 - previous interpreter state pointer 613 614 // InterpreterRuntime::frequency_counter_overflow takes one argument 615 // indicating if the counter overflow occurs at a backwards branch (non-NULL bcp). 616 // The call returns the address of the verified entry point for the method or NULL 617 // if the compilation did not complete (either went background or bailed out). 618 __ movptr(rax, (int32_t)false); 619 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), rax); 620 621 // for c++ interpreter can rsi really be munged? 622 __ lea(state, Address(rbp, -(int)sizeof(BytecodeInterpreter))); // restore state 623 __ movptr(rbx, Address(state, byte_offset_of(BytecodeInterpreter, _method))); // restore method 624 __ movptr(rdi, Address(state, byte_offset_of(BytecodeInterpreter, _locals))); // get locals pointer 625 626 __ jmp(*do_continue, relocInfo::none); 627 628 } 629 630 void InterpreterGenerator::generate_stack_overflow_check(void) { 631 // see if we've got enough room on the stack for locals plus overhead. 632 // the expression stack grows down incrementally, so the normal guard 633 // page mechanism will work for that. 634 // 635 // Registers live on entry: 636 // 637 // Asm interpreter 638 // rdx: number of additional locals this frame needs (what we must check) 639 // rbx,: Method* 640 641 // C++ Interpreter 642 // rsi/r13: previous interpreter frame state object 643 // rdi: &locals[0] 644 // rcx: # of locals 645 // rdx: number of additional locals this frame needs (what we must check) 646 // rbx: Method* 647 648 // destroyed on exit 649 // rax, 650 651 // NOTE: since the additional locals are also always pushed (wasn't obvious in 652 // generate_method_entry) so the guard should work for them too. 653 // 654 655 // monitor entry size: see picture of stack set (generate_method_entry) and frame_i486.hpp 656 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize; 657 658 // total overhead size: entry_size + (saved rbp, thru expr stack bottom). 659 // be sure to change this if you add/subtract anything to/from the overhead area 660 const int overhead_size = (int)sizeof(BytecodeInterpreter); 661 662 const int page_size = os::vm_page_size(); 663 664 Label after_frame_check; 665 666 // compute rsp as if this were going to be the last frame on 667 // the stack before the red zone 668 669 Label after_frame_check_pop; 670 671 // save rsi == caller's bytecode ptr (c++ previous interp. state) 672 // QQQ problem here?? rsi overload???? 673 __ push(state); 674 675 const Register thread = LP64_ONLY(r15_thread) NOT_LP64(rsi); 676 677 NOT_LP64(__ get_thread(thread)); 678 679 const Address stack_base(thread, Thread::stack_base_offset()); 680 const Address stack_size(thread, Thread::stack_size_offset()); 681 682 // locals + overhead, in bytes 683 // Always give one monitor to allow us to start interp if sync method. 684 // Any additional monitors need a check when moving the expression stack 685 const int one_monitor = frame::interpreter_frame_monitor_size() * wordSize; 686 __ movptr(rax, Address(rbx, Method::const_offset())); 687 __ load_unsigned_short(rax, Address(rax, ConstMethod::max_stack_offset())); // get size of expression stack in words 688 __ lea(rax, Address(noreg, rax, Interpreter::stackElementScale(), one_monitor)); 689 __ lea(rax, Address(rax, rdx, Interpreter::stackElementScale(), overhead_size)); 690 691 #ifdef ASSERT 692 Label stack_base_okay, stack_size_okay; 693 // verify that thread stack base is non-zero 694 __ cmpptr(stack_base, (int32_t)0); 695 __ jcc(Assembler::notEqual, stack_base_okay); 696 __ stop("stack base is zero"); 697 __ bind(stack_base_okay); 698 // verify that thread stack size is non-zero 699 __ cmpptr(stack_size, (int32_t)0); 700 __ jcc(Assembler::notEqual, stack_size_okay); 701 __ stop("stack size is zero"); 702 __ bind(stack_size_okay); 703 #endif 704 705 // Add stack base to locals and subtract stack size 706 __ addptr(rax, stack_base); 707 __ subptr(rax, stack_size); 708 709 // We should have a magic number here for the size of the c++ interpreter frame. 710 // We can't actually tell this ahead of time. The debug version size is around 3k 711 // product is 1k and fastdebug is 4k 712 const int slop = 6 * K; 713 714 // Use the maximum number of pages we might bang. 715 const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages : 716 (StackRedPages+StackYellowPages); 717 // Only need this if we are stack banging which is temporary while 718 // we're debugging. 719 __ addptr(rax, slop + 2*max_pages * page_size); 720 721 // check against the current stack bottom 722 __ cmpptr(rsp, rax); 723 __ jcc(Assembler::above, after_frame_check_pop); 724 725 __ pop(state); // get c++ prev state. 726 727 // throw exception return address becomes throwing pc 728 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError)); 729 730 // all done with frame size check 731 __ bind(after_frame_check_pop); 732 __ pop(state); 733 734 __ bind(after_frame_check); 735 } 736 737 // Find preallocated monitor and lock method (C++ interpreter) 738 // rbx - Method* 739 // 740 void InterpreterGenerator::lock_method(void) { 741 // assumes state == rsi/r13 == pointer to current interpreterState 742 // minimally destroys rax, rdx|c_rarg1, rdi 743 // 744 // synchronize method 745 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize; 746 const Address access_flags (rbx, Method::access_flags_offset()); 747 748 const Register monitor = NOT_LP64(rdx) LP64_ONLY(c_rarg1); 749 750 // find initial monitor i.e. monitors[-1] 751 __ movptr(monitor, STATE(_monitor_base)); // get monitor bottom limit 752 __ subptr(monitor, entry_size); // point to initial monitor 753 754 #ifdef ASSERT 755 { Label L; 756 __ movl(rax, access_flags); 757 __ testl(rax, JVM_ACC_SYNCHRONIZED); 758 __ jcc(Assembler::notZero, L); 759 __ stop("method doesn't need synchronization"); 760 __ bind(L); 761 } 762 #endif // ASSERT 763 // get synchronization object 764 { Label done; 765 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 766 __ movl(rax, access_flags); 767 __ movptr(rdi, STATE(_locals)); // prepare to get receiver (assume common case) 768 __ testl(rax, JVM_ACC_STATIC); 769 __ movptr(rax, Address(rdi, 0)); // get receiver (assume this is frequent case) 770 __ jcc(Assembler::zero, done); 771 __ movptr(rax, Address(rbx, Method::const_offset())); 772 __ movptr(rax, Address(rax, ConstMethod::constants_offset())); 773 __ movptr(rax, Address(rax, ConstantPool::pool_holder_offset_in_bytes())); 774 __ movptr(rax, Address(rax, mirror_offset)); 775 __ bind(done); 776 } 777 #ifdef ASSERT 778 { Label L; 779 __ cmpptr(rax, Address(monitor, BasicObjectLock::obj_offset_in_bytes())); // correct object? 780 __ jcc(Assembler::equal, L); 781 __ stop("wrong synchronization lobject"); 782 __ bind(L); 783 } 784 #endif // ASSERT 785 // can destroy rax, rdx|c_rarg1, rcx, and (via call_VM) rdi! 786 __ lock_object(monitor); 787 } 788 789 // Call an accessor method (assuming it is resolved, otherwise drop into vanilla (slow path) entry 790 791 address InterpreterGenerator::generate_accessor_entry(void) { 792 793 // rbx: Method* 794 795 // rsi/r13: senderSP must preserved for slow path, set SP to it on fast path 796 797 Label xreturn_path; 798 799 // do fastpath for resolved accessor methods 800 if (UseFastAccessorMethods) { 801 802 address entry_point = __ pc(); 803 804 Label slow_path; 805 // If we need a safepoint check, generate full interpreter entry. 806 ExternalAddress state(SafepointSynchronize::address_of_state()); 807 __ cmp32(ExternalAddress(SafepointSynchronize::address_of_state()), 808 SafepointSynchronize::_not_synchronized); 809 810 __ jcc(Assembler::notEqual, slow_path); 811 // ASM/C++ Interpreter 812 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof; parameter size = 1 813 // Note: We can only use this code if the getfield has been resolved 814 // and if we don't have a null-pointer exception => check for 815 // these conditions first and use slow path if necessary. 816 // rbx,: method 817 // rcx: receiver 818 __ movptr(rax, Address(rsp, wordSize)); 819 820 // check if local 0 != NULL and read field 821 __ testptr(rax, rax); 822 __ jcc(Assembler::zero, slow_path); 823 824 // read first instruction word and extract bytecode @ 1 and index @ 2 825 __ movptr(rdx, Address(rbx, Method::const_offset())); 826 __ movptr(rdi, Address(rdx, ConstMethod::constants_offset())); 827 __ movl(rdx, Address(rdx, ConstMethod::codes_offset())); 828 // Shift codes right to get the index on the right. 829 // The bytecode fetched looks like <index><0xb4><0x2a> 830 __ shrl(rdx, 2*BitsPerByte); 831 __ shll(rdx, exact_log2(in_words(ConstantPoolCacheEntry::size()))); 832 __ movptr(rdi, Address(rdi, ConstantPool::cache_offset_in_bytes())); 833 834 // rax,: local 0 835 // rbx,: method 836 // rcx: receiver - do not destroy since it is needed for slow path! 837 // rcx: scratch 838 // rdx: constant pool cache index 839 // rdi: constant pool cache 840 // rsi/r13: sender sp 841 842 // check if getfield has been resolved and read constant pool cache entry 843 // check the validity of the cache entry by testing whether _indices field 844 // contains Bytecode::_getfield in b1 byte. 845 assert(in_words(ConstantPoolCacheEntry::size()) == 4, "adjust shift below"); 846 __ movl(rcx, 847 Address(rdi, 848 rdx, 849 Address::times_ptr, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::indices_offset())); 850 __ shrl(rcx, 2*BitsPerByte); 851 __ andl(rcx, 0xFF); 852 __ cmpl(rcx, Bytecodes::_getfield); 853 __ jcc(Assembler::notEqual, slow_path); 854 855 // Note: constant pool entry is not valid before bytecode is resolved 856 __ movptr(rcx, 857 Address(rdi, 858 rdx, 859 Address::times_ptr, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::f2_offset())); 860 __ movl(rdx, 861 Address(rdi, 862 rdx, 863 Address::times_ptr, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset())); 864 865 Label notByte, notShort, notChar; 866 const Address field_address (rax, rcx, Address::times_1); 867 868 // Need to differentiate between igetfield, agetfield, bgetfield etc. 869 // because they are different sizes. 870 // Use the type from the constant pool cache 871 __ shrl(rdx, ConstantPoolCacheEntry::tos_state_shift); 872 // Make sure we don't need to mask rdx after the above shift 873 ConstantPoolCacheEntry::verify_tos_state_shift(); 874 #ifdef _LP64 875 Label notObj; 876 __ cmpl(rdx, atos); 877 __ jcc(Assembler::notEqual, notObj); 878 // atos 879 __ movptr(rax, field_address); 880 __ jmp(xreturn_path); 881 882 __ bind(notObj); 883 #endif // _LP64 884 __ cmpl(rdx, btos); 885 __ jcc(Assembler::notEqual, notByte); 886 __ load_signed_byte(rax, field_address); 887 __ jmp(xreturn_path); 888 889 __ bind(notByte); 890 __ cmpl(rdx, stos); 891 __ jcc(Assembler::notEqual, notShort); 892 __ load_signed_short(rax, field_address); 893 __ jmp(xreturn_path); 894 895 __ bind(notShort); 896 __ cmpl(rdx, ctos); 897 __ jcc(Assembler::notEqual, notChar); 898 __ load_unsigned_short(rax, field_address); 899 __ jmp(xreturn_path); 900 901 __ bind(notChar); 902 #ifdef ASSERT 903 Label okay; 904 #ifndef _LP64 905 __ cmpl(rdx, atos); 906 __ jcc(Assembler::equal, okay); 907 #endif // _LP64 908 __ cmpl(rdx, itos); 909 __ jcc(Assembler::equal, okay); 910 __ stop("what type is this?"); 911 __ bind(okay); 912 #endif // ASSERT 913 // All the rest are a 32 bit wordsize 914 __ movl(rax, field_address); 915 916 __ bind(xreturn_path); 917 918 // _ireturn/_areturn 919 __ pop(rdi); // get return address 920 __ mov(rsp, sender_sp_on_entry); // set sp to sender sp 921 __ jmp(rdi); 922 923 // generate a vanilla interpreter entry as the slow path 924 __ bind(slow_path); 925 // We will enter c++ interpreter looking like it was 926 // called by the call_stub this will cause it to return 927 // a tosca result to the invoker which might have been 928 // the c++ interpreter itself. 929 930 __ jmp(fast_accessor_slow_entry_path); 931 return entry_point; 932 933 } else { 934 return NULL; 935 } 936 937 } 938 939 address InterpreterGenerator::generate_Reference_get_entry(void) { 940 #if INCLUDE_ALL_GCS 941 if (UseG1GC) { 942 // We need to generate have a routine that generates code to: 943 // * load the value in the referent field 944 // * passes that value to the pre-barrier. 945 // 946 // In the case of G1 this will record the value of the 947 // referent in an SATB buffer if marking is active. 948 // This will cause concurrent marking to mark the referent 949 // field as live. 950 Unimplemented(); 951 } 952 #endif // INCLUDE_ALL_GCS 953 954 // If G1 is not enabled then attempt to go through the accessor entry point 955 // Reference.get is an accessor 956 return generate_accessor_entry(); 957 } 958 959 // 960 // C++ Interpreter stub for calling a native method. 961 // This sets up a somewhat different looking stack for calling the native method 962 // than the typical interpreter frame setup but still has the pointer to 963 // an interpreter state. 964 // 965 966 address InterpreterGenerator::generate_native_entry(bool synchronized) { 967 // determine code generation flags 968 bool inc_counter = UseCompiler || CountCompiledCalls; 969 970 // rbx: Method* 971 // rcx: receiver (unused) 972 // rsi/r13: previous interpreter state (if called from C++ interpreter) must preserve 973 // in any case. If called via c1/c2/call_stub rsi/r13 is junk (to use) but harmless 974 // to save/restore. 975 address entry_point = __ pc(); 976 977 const Address constMethod (rbx, Method::const_offset()); 978 const Address invocation_counter(rbx, Method::invocation_counter_offset() + InvocationCounter::counter_offset()); 979 const Address access_flags (rbx, Method::access_flags_offset()); 980 const Address size_of_parameters(rcx, ConstMethod::size_of_parameters_offset()); 981 982 // rsi/r13 == state/locals rdi == prevstate 983 const Register locals = rdi; 984 985 // get parameter size (always needed) 986 __ movptr(rcx, constMethod); 987 __ load_unsigned_short(rcx, size_of_parameters); 988 989 // rbx: Method* 990 // rcx: size of parameters 991 __ pop(rax); // get return address 992 // for natives the size of locals is zero 993 994 // compute beginning of parameters /locals 995 996 __ lea(locals, Address(rsp, rcx, Address::times_ptr, -wordSize)); 997 998 // initialize fixed part of activation frame 999 1000 // Assumes rax = return address 1001 1002 // allocate and initialize new interpreterState and method expression stack 1003 // IN(locals) -> locals 1004 // IN(state) -> previous frame manager state (NULL from stub/c1/c2) 1005 // destroys rax, rcx, rdx 1006 // OUT (state) -> new interpreterState 1007 // OUT(rsp) -> bottom of methods expression stack 1008 1009 // save sender_sp 1010 __ mov(rcx, sender_sp_on_entry); 1011 // start with NULL previous state 1012 __ movptr(state, (int32_t)NULL_WORD); 1013 generate_compute_interpreter_state(state, locals, rcx, true); 1014 1015 #ifdef ASSERT 1016 { Label L; 1017 __ movptr(rax, STATE(_stack_base)); 1018 #ifdef _LP64 1019 // duplicate the alignment rsp got after setting stack_base 1020 __ subptr(rax, frame::arg_reg_save_area_bytes); // windows 1021 __ andptr(rax, -16); // must be 16 byte boundary (see amd64 ABI) 1022 #endif // _LP64 1023 __ cmpptr(rax, rsp); 1024 __ jcc(Assembler::equal, L); 1025 __ stop("broken stack frame setup in interpreter"); 1026 __ bind(L); 1027 } 1028 #endif 1029 1030 if (inc_counter) __ movl(rcx, invocation_counter); // (pre-)fetch invocation count 1031 1032 const Register unlock_thread = LP64_ONLY(r15_thread) NOT_LP64(rax); 1033 NOT_LP64(__ movptr(unlock_thread, STATE(_thread));) // get thread 1034 // Since at this point in the method invocation the exception handler 1035 // would try to exit the monitor of synchronized methods which hasn't 1036 // been entered yet, we set the thread local variable 1037 // _do_not_unlock_if_synchronized to true. The remove_activation will 1038 // check this flag. 1039 1040 const Address do_not_unlock_if_synchronized(unlock_thread, 1041 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset())); 1042 __ movbool(do_not_unlock_if_synchronized, true); 1043 1044 // make sure method is native & not abstract 1045 #ifdef ASSERT 1046 __ movl(rax, access_flags); 1047 { 1048 Label L; 1049 __ testl(rax, JVM_ACC_NATIVE); 1050 __ jcc(Assembler::notZero, L); 1051 __ stop("tried to execute non-native method as native"); 1052 __ bind(L); 1053 } 1054 { Label L; 1055 __ testl(rax, JVM_ACC_ABSTRACT); 1056 __ jcc(Assembler::zero, L); 1057 __ stop("tried to execute abstract method in interpreter"); 1058 __ bind(L); 1059 } 1060 #endif 1061 1062 1063 // increment invocation count & check for overflow 1064 Label invocation_counter_overflow; 1065 if (inc_counter) { 1066 generate_counter_incr(&invocation_counter_overflow, NULL, NULL); 1067 } 1068 1069 Label continue_after_compile; 1070 1071 __ bind(continue_after_compile); 1072 1073 bang_stack_shadow_pages(true); 1074 1075 // reset the _do_not_unlock_if_synchronized flag 1076 NOT_LP64(__ movl(rax, STATE(_thread));) // get thread 1077 __ movbool(do_not_unlock_if_synchronized, false); 1078 1079 1080 // check for synchronized native methods 1081 // 1082 // Note: This must happen *after* invocation counter check, since 1083 // when overflow happens, the method should not be locked. 1084 if (synchronized) { 1085 // potentially kills rax, rcx, rdx, rdi 1086 lock_method(); 1087 } else { 1088 // no synchronization necessary 1089 #ifdef ASSERT 1090 { Label L; 1091 __ movl(rax, access_flags); 1092 __ testl(rax, JVM_ACC_SYNCHRONIZED); 1093 __ jcc(Assembler::zero, L); 1094 __ stop("method needs synchronization"); 1095 __ bind(L); 1096 } 1097 #endif 1098 } 1099 1100 // start execution 1101 1102 // jvmti support 1103 __ notify_method_entry(); 1104 1105 // work registers 1106 const Register method = rbx; 1107 const Register thread = LP64_ONLY(r15_thread) NOT_LP64(rdi); 1108 const Register t = InterpreterRuntime::SignatureHandlerGenerator::temp(); // rcx|rscratch1 1109 const Address constMethod (method, Method::const_offset()); 1110 const Address size_of_parameters(t, ConstMethod::size_of_parameters_offset()); 1111 1112 // allocate space for parameters 1113 __ movptr(method, STATE(_method)); 1114 __ verify_method_ptr(method); 1115 __ movptr(t, constMethod); 1116 __ load_unsigned_short(t, size_of_parameters); 1117 __ shll(t, 2); 1118 #ifdef _LP64 1119 __ subptr(rsp, t); 1120 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows 1121 __ andptr(rsp, -16); // must be 16 byte boundary (see amd64 ABI) 1122 #else 1123 __ addptr(t, 2*wordSize); // allocate two more slots for JNIEnv and possible mirror 1124 __ subptr(rsp, t); 1125 __ andptr(rsp, -(StackAlignmentInBytes)); // gcc needs 16 byte aligned stacks to do XMM intrinsics 1126 #endif // _LP64 1127 1128 // get signature handler 1129 Label pending_exception_present; 1130 1131 { Label L; 1132 __ movptr(t, Address(method, Method::signature_handler_offset())); 1133 __ testptr(t, t); 1134 __ jcc(Assembler::notZero, L); 1135 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), method, false); 1136 __ movptr(method, STATE(_method)); 1137 __ cmpptr(Address(thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD); 1138 __ jcc(Assembler::notEqual, pending_exception_present); 1139 __ verify_method_ptr(method); 1140 __ movptr(t, Address(method, Method::signature_handler_offset())); 1141 __ bind(L); 1142 } 1143 #ifdef ASSERT 1144 { 1145 Label L; 1146 __ push(t); 1147 __ get_thread(t); // get vm's javathread* 1148 __ cmpptr(t, STATE(_thread)); 1149 __ jcc(Assembler::equal, L); 1150 __ int3(); 1151 __ bind(L); 1152 __ pop(t); 1153 } 1154 #endif // 1155 1156 const Register from_ptr = InterpreterRuntime::SignatureHandlerGenerator::from(); 1157 // call signature handler 1158 assert(InterpreterRuntime::SignatureHandlerGenerator::to () == rsp, "adjust this code"); 1159 1160 // The generated handlers do not touch RBX (the method oop). 1161 // However, large signatures cannot be cached and are generated 1162 // each time here. The slow-path generator will blow RBX 1163 // sometime, so we must reload it after the call. 1164 __ movptr(from_ptr, STATE(_locals)); // get the from pointer 1165 __ call(t); 1166 __ movptr(method, STATE(_method)); 1167 __ verify_method_ptr(method); 1168 1169 // result handler is in rax 1170 // set result handler 1171 __ movptr(STATE(_result_handler), rax); 1172 1173 1174 // get native function entry point 1175 { Label L; 1176 __ movptr(rax, Address(method, Method::native_function_offset())); 1177 __ testptr(rax, rax); 1178 __ jcc(Assembler::notZero, L); 1179 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), method); 1180 __ movptr(method, STATE(_method)); 1181 __ verify_method_ptr(method); 1182 __ movptr(rax, Address(method, Method::native_function_offset())); 1183 __ bind(L); 1184 } 1185 1186 // pass mirror handle if static call 1187 { Label L; 1188 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 1189 __ movl(t, Address(method, Method::access_flags_offset())); 1190 __ testl(t, JVM_ACC_STATIC); 1191 __ jcc(Assembler::zero, L); 1192 // get mirror 1193 __ movptr(t, Address(method, Method:: const_offset())); 1194 __ movptr(t, Address(t, ConstMethod::constants_offset())); 1195 __ movptr(t, Address(t, ConstantPool::pool_holder_offset_in_bytes())); 1196 __ movptr(t, Address(t, mirror_offset)); 1197 // copy mirror into activation object 1198 __ movptr(STATE(_oop_temp), t); 1199 // pass handle to mirror 1200 #ifdef _LP64 1201 __ lea(c_rarg1, STATE(_oop_temp)); 1202 #else 1203 __ lea(t, STATE(_oop_temp)); 1204 __ movptr(Address(rsp, wordSize), t); 1205 #endif // _LP64 1206 __ bind(L); 1207 } 1208 #ifdef ASSERT 1209 { 1210 Label L; 1211 __ push(t); 1212 __ get_thread(t); // get vm's javathread* 1213 __ cmpptr(t, STATE(_thread)); 1214 __ jcc(Assembler::equal, L); 1215 __ int3(); 1216 __ bind(L); 1217 __ pop(t); 1218 } 1219 #endif // 1220 1221 // pass JNIEnv 1222 #ifdef _LP64 1223 __ lea(c_rarg0, Address(thread, JavaThread::jni_environment_offset())); 1224 #else 1225 __ movptr(thread, STATE(_thread)); // get thread 1226 __ lea(t, Address(thread, JavaThread::jni_environment_offset())); 1227 1228 __ movptr(Address(rsp, 0), t); 1229 #endif // _LP64 1230 1231 #ifdef ASSERT 1232 { 1233 Label L; 1234 __ push(t); 1235 __ get_thread(t); // get vm's javathread* 1236 __ cmpptr(t, STATE(_thread)); 1237 __ jcc(Assembler::equal, L); 1238 __ int3(); 1239 __ bind(L); 1240 __ pop(t); 1241 } 1242 #endif // 1243 1244 #ifdef ASSERT 1245 { Label L; 1246 __ movl(t, Address(thread, JavaThread::thread_state_offset())); 1247 __ cmpl(t, _thread_in_Java); 1248 __ jcc(Assembler::equal, L); 1249 __ stop("Wrong thread state in native stub"); 1250 __ bind(L); 1251 } 1252 #endif 1253 1254 // Change state to native (we save the return address in the thread, since it might not 1255 // be pushed on the stack when we do a a stack traversal). It is enough that the pc() 1256 // points into the right code segment. It does not have to be the correct return pc. 1257 1258 __ set_last_Java_frame(thread, noreg, rbp, __ pc()); 1259 1260 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_native); 1261 1262 __ call(rax); 1263 1264 // result potentially in rdx:rax or ST0 1265 __ movptr(method, STATE(_method)); 1266 NOT_LP64(__ movptr(thread, STATE(_thread));) // get thread 1267 1268 // The potential result is in ST(0) & rdx:rax 1269 // With C++ interpreter we leave any possible result in ST(0) until we are in result handler and then 1270 // we do the appropriate stuff for returning the result. rdx:rax must always be saved because just about 1271 // anything we do here will destroy it, st(0) is only saved if we re-enter the vm where it would 1272 // be destroyed. 1273 // It is safe to do these pushes because state is _thread_in_native and return address will be found 1274 // via _last_native_pc and not via _last_jave_sp 1275 1276 // Must save the value of ST(0)/xmm0 since it could be destroyed before we get to result handler 1277 { Label Lpush, Lskip; 1278 ExternalAddress float_handler(AbstractInterpreter::result_handler(T_FLOAT)); 1279 ExternalAddress double_handler(AbstractInterpreter::result_handler(T_DOUBLE)); 1280 __ cmpptr(STATE(_result_handler), float_handler.addr()); 1281 __ jcc(Assembler::equal, Lpush); 1282 __ cmpptr(STATE(_result_handler), double_handler.addr()); 1283 __ jcc(Assembler::notEqual, Lskip); 1284 __ bind(Lpush); 1285 __ subptr(rsp, 2*wordSize); 1286 if ( UseSSE < 2 ) { 1287 __ fstp_d(Address(rsp, 0)); 1288 } else { 1289 __ movdbl(Address(rsp, 0), xmm0); 1290 } 1291 __ bind(Lskip); 1292 } 1293 1294 // save rax:rdx for potential use by result handler. 1295 __ push(rax); 1296 #ifndef _LP64 1297 __ push(rdx); 1298 #endif // _LP64 1299 1300 // Either restore the MXCSR register after returning from the JNI Call 1301 // or verify that it wasn't changed. 1302 if (VM_Version::supports_sse()) { 1303 if (RestoreMXCSROnJNICalls) { 1304 __ ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std())); 1305 } 1306 else if (CheckJNICalls ) { 1307 __ call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry())); 1308 } 1309 } 1310 1311 #ifndef _LP64 1312 // Either restore the x87 floating pointer control word after returning 1313 // from the JNI call or verify that it wasn't changed. 1314 if (CheckJNICalls) { 1315 __ call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry())); 1316 } 1317 #endif // _LP64 1318 1319 1320 // change thread state 1321 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_native_trans); 1322 if(os::is_MP()) { 1323 // Write serialization page so VM thread can do a pseudo remote membar. 1324 // We use the current thread pointer to calculate a thread specific 1325 // offset to write to within the page. This minimizes bus traffic 1326 // due to cache line collision. 1327 __ serialize_memory(thread, rcx); 1328 } 1329 1330 // check for safepoint operation in progress and/or pending suspend requests 1331 { Label Continue; 1332 1333 __ cmp32(ExternalAddress(SafepointSynchronize::address_of_state()), 1334 SafepointSynchronize::_not_synchronized); 1335 1336 // threads running native code and they are expected to self-suspend 1337 // when leaving the _thread_in_native state. We need to check for 1338 // pending suspend requests here. 1339 Label L; 1340 __ jcc(Assembler::notEqual, L); 1341 __ cmpl(Address(thread, JavaThread::suspend_flags_offset()), 0); 1342 __ jcc(Assembler::equal, Continue); 1343 __ bind(L); 1344 1345 // Don't use call_VM as it will see a possible pending exception and forward it 1346 // and never return here preventing us from clearing _last_native_pc down below. 1347 // Also can't use call_VM_leaf either as it will check to see if rsi & rdi are 1348 // preserved and correspond to the bcp/locals pointers. 1349 // 1350 1351 ((MacroAssembler*)_masm)->call_VM_leaf(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans), 1352 thread); 1353 __ increment(rsp, wordSize); 1354 1355 __ movptr(method, STATE(_method)); 1356 __ verify_method_ptr(method); 1357 __ movptr(thread, STATE(_thread)); // get thread 1358 1359 __ bind(Continue); 1360 } 1361 1362 // change thread state 1363 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_Java); 1364 1365 __ reset_last_Java_frame(thread, true, true); 1366 1367 // reset handle block 1368 __ movptr(t, Address(thread, JavaThread::active_handles_offset())); 1369 __ movptr(Address(t, JNIHandleBlock::top_offset_in_bytes()), (int32_t)NULL_WORD); 1370 1371 // If result was an oop then unbox and save it in the frame 1372 { Label L; 1373 Label no_oop, store_result; 1374 ExternalAddress oop_handler(AbstractInterpreter::result_handler(T_OBJECT)); 1375 __ cmpptr(STATE(_result_handler), oop_handler.addr()); 1376 __ jcc(Assembler::notEqual, no_oop); 1377 #ifndef _LP64 1378 __ pop(rdx); 1379 #endif // _LP64 1380 __ pop(rax); 1381 __ testptr(rax, rax); 1382 __ jcc(Assembler::zero, store_result); 1383 // unbox 1384 __ movptr(rax, Address(rax, 0)); 1385 __ bind(store_result); 1386 __ movptr(STATE(_oop_temp), rax); 1387 // keep stack depth as expected by pushing oop which will eventually be discarded 1388 __ push(rax); 1389 #ifndef _LP64 1390 __ push(rdx); 1391 #endif // _LP64 1392 __ bind(no_oop); 1393 } 1394 1395 { 1396 Label no_reguard; 1397 __ cmpl(Address(thread, JavaThread::stack_guard_state_offset()), JavaThread::stack_guard_yellow_disabled); 1398 __ jcc(Assembler::notEqual, no_reguard); 1399 1400 __ pusha(); 1401 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages))); 1402 __ popa(); 1403 1404 __ bind(no_reguard); 1405 } 1406 1407 1408 // QQQ Seems like for native methods we simply return and the caller will see the pending 1409 // exception and do the right thing. Certainly the interpreter will, don't know about 1410 // compiled methods. 1411 // Seems that the answer to above is no this is wrong. The old code would see the exception 1412 // and forward it before doing the unlocking and notifying jvmdi that method has exited. 1413 // This seems wrong need to investigate the spec. 1414 1415 // handle exceptions (exception handling will handle unlocking!) 1416 { Label L; 1417 __ cmpptr(Address(thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD); 1418 __ jcc(Assembler::zero, L); 1419 __ bind(pending_exception_present); 1420 1421 // There are potential results on the stack (rax/rdx, ST(0)) we ignore these and simply 1422 // return and let caller deal with exception. This skips the unlocking here which 1423 // seems wrong but seems to be what asm interpreter did. Can't find this in the spec. 1424 // Note: must preverve method in rbx 1425 // 1426 1427 // remove activation 1428 1429 __ movptr(t, STATE(_sender_sp)); 1430 __ leave(); // remove frame anchor 1431 __ pop(rdi); // get return address 1432 __ movptr(state, STATE(_prev_link)); // get previous state for return 1433 __ mov(rsp, t); // set sp to sender sp 1434 __ push(rdi); // push throwing pc 1435 // The skips unlocking!! This seems to be what asm interpreter does but seems 1436 // very wrong. Not clear if this violates the spec. 1437 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 1438 __ bind(L); 1439 } 1440 1441 // do unlocking if necessary 1442 { Label L; 1443 __ movl(t, Address(method, Method::access_flags_offset())); 1444 __ testl(t, JVM_ACC_SYNCHRONIZED); 1445 __ jcc(Assembler::zero, L); 1446 // the code below should be shared with interpreter macro assembler implementation 1447 { Label unlock; 1448 const Register monitor = NOT_LP64(rdx) LP64_ONLY(c_rarg1); 1449 // BasicObjectLock will be first in list, since this is a synchronized method. However, need 1450 // to check that the object has not been unlocked by an explicit monitorexit bytecode. 1451 __ movptr(monitor, STATE(_monitor_base)); 1452 __ subptr(monitor, frame::interpreter_frame_monitor_size() * wordSize); // address of initial monitor 1453 1454 __ movptr(t, Address(monitor, BasicObjectLock::obj_offset_in_bytes())); 1455 __ testptr(t, t); 1456 __ jcc(Assembler::notZero, unlock); 1457 1458 // Entry already unlocked, need to throw exception 1459 __ MacroAssembler::call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_illegal_monitor_state_exception)); 1460 __ should_not_reach_here(); 1461 1462 __ bind(unlock); 1463 __ unlock_object(monitor); 1464 // unlock can blow rbx so restore it for path that needs it below 1465 __ movptr(method, STATE(_method)); 1466 } 1467 __ bind(L); 1468 } 1469 1470 // jvmti support 1471 // Note: This must happen _after_ handling/throwing any exceptions since 1472 // the exception handler code notifies the runtime of method exits 1473 // too. If this happens before, method entry/exit notifications are 1474 // not properly paired (was bug - gri 11/22/99). 1475 __ notify_method_exit(vtos, InterpreterMacroAssembler::NotifyJVMTI); 1476 1477 // restore potential result in rdx:rax, call result handler to restore potential result in ST0 & handle result 1478 #ifndef _LP64 1479 __ pop(rdx); 1480 #endif // _LP64 1481 __ pop(rax); 1482 __ movptr(t, STATE(_result_handler)); // get result handler 1483 __ call(t); // call result handler to convert to tosca form 1484 1485 // remove activation 1486 1487 __ movptr(t, STATE(_sender_sp)); 1488 1489 __ leave(); // remove frame anchor 1490 __ pop(rdi); // get return address 1491 __ movptr(state, STATE(_prev_link)); // get previous state for return (if c++ interpreter was caller) 1492 __ mov(rsp, t); // set sp to sender sp 1493 __ jmp(rdi); 1494 1495 // invocation counter overflow 1496 if (inc_counter) { 1497 // Handle overflow of counter and compile method 1498 __ bind(invocation_counter_overflow); 1499 generate_counter_overflow(&continue_after_compile); 1500 } 1501 1502 return entry_point; 1503 } 1504 1505 // Generate entries that will put a result type index into rcx 1506 void CppInterpreterGenerator::generate_deopt_handling() { 1507 1508 Label return_from_deopt_common; 1509 1510 // Generate entries that will put a result type index into rcx 1511 // deopt needs to jump to here to enter the interpreter (return a result) 1512 deopt_frame_manager_return_atos = __ pc(); 1513 1514 // rax is live here 1515 __ movl(rcx, AbstractInterpreter::BasicType_as_index(T_OBJECT)); // Result stub address array index 1516 __ jmp(return_from_deopt_common); 1517 1518 1519 // deopt needs to jump to here to enter the interpreter (return a result) 1520 deopt_frame_manager_return_btos = __ pc(); 1521 1522 // rax is live here 1523 __ movl(rcx, AbstractInterpreter::BasicType_as_index(T_BOOLEAN)); // Result stub address array index 1524 __ jmp(return_from_deopt_common); 1525 1526 // deopt needs to jump to here to enter the interpreter (return a result) 1527 deopt_frame_manager_return_itos = __ pc(); 1528 1529 // rax is live here 1530 __ movl(rcx, AbstractInterpreter::BasicType_as_index(T_INT)); // Result stub address array index 1531 __ jmp(return_from_deopt_common); 1532 1533 // deopt needs to jump to here to enter the interpreter (return a result) 1534 1535 deopt_frame_manager_return_ltos = __ pc(); 1536 // rax,rdx are live here 1537 __ movl(rcx, AbstractInterpreter::BasicType_as_index(T_LONG)); // Result stub address array index 1538 __ jmp(return_from_deopt_common); 1539 1540 // deopt needs to jump to here to enter the interpreter (return a result) 1541 1542 deopt_frame_manager_return_ftos = __ pc(); 1543 // st(0) is live here 1544 __ movl(rcx, AbstractInterpreter::BasicType_as_index(T_FLOAT)); // Result stub address array index 1545 __ jmp(return_from_deopt_common); 1546 1547 // deopt needs to jump to here to enter the interpreter (return a result) 1548 deopt_frame_manager_return_dtos = __ pc(); 1549 1550 // st(0) is live here 1551 __ movl(rcx, AbstractInterpreter::BasicType_as_index(T_DOUBLE)); // Result stub address array index 1552 __ jmp(return_from_deopt_common); 1553 1554 // deopt needs to jump to here to enter the interpreter (return a result) 1555 deopt_frame_manager_return_vtos = __ pc(); 1556 1557 __ movl(rcx, AbstractInterpreter::BasicType_as_index(T_VOID)); 1558 1559 // Deopt return common 1560 // an index is present in rcx that lets us move any possible result being 1561 // return to the interpreter's stack 1562 // 1563 // Because we have a full sized interpreter frame on the youngest 1564 // activation the stack is pushed too deep to share the tosca to 1565 // stack converters directly. We shrink the stack to the desired 1566 // amount and then push result and then re-extend the stack. 1567 // We could have the code in size_activation layout a short 1568 // frame for the top activation but that would look different 1569 // than say sparc (which needs a full size activation because 1570 // the windows are in the way. Really it could be short? QQQ 1571 // 1572 __ bind(return_from_deopt_common); 1573 1574 __ lea(state, Address(rbp, -(int)sizeof(BytecodeInterpreter))); 1575 1576 // setup rsp so we can push the "result" as needed. 1577 __ movptr(rsp, STATE(_stack)); // trim stack (is prepushed) 1578 __ addptr(rsp, wordSize); // undo prepush 1579 1580 ExternalAddress tosca_to_stack((address)CppInterpreter::_tosca_to_stack); 1581 // Address index(noreg, rcx, Address::times_ptr); 1582 __ movptr(rcx, ArrayAddress(tosca_to_stack, Address(noreg, rcx, Address::times_ptr))); 1583 // __ movl(rcx, Address(noreg, rcx, Address::times_ptr, int(AbstractInterpreter::_tosca_to_stack))); 1584 __ call(rcx); // call result converter 1585 1586 __ movl(STATE(_msg), (int)BytecodeInterpreter::deopt_resume); 1587 __ lea(rsp, Address(rsp, -wordSize)); // prepush stack (result if any already present) 1588 __ movptr(STATE(_stack), rsp); // inform interpreter of new stack depth (parameters removed, 1589 // result if any on stack already ) 1590 __ movptr(rsp, STATE(_stack_limit)); // restore expression stack to full depth 1591 } 1592 1593 // Generate the code to handle a more_monitors message from the c++ interpreter 1594 void CppInterpreterGenerator::generate_more_monitors() { 1595 1596 1597 Label entry, loop; 1598 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize; 1599 // 1. compute new pointers // rsp: old expression stack top 1600 __ movptr(rdx, STATE(_stack_base)); // rdx: old expression stack bottom 1601 __ subptr(rsp, entry_size); // move expression stack top limit 1602 __ subptr(STATE(_stack), entry_size); // update interpreter stack top 1603 __ subptr(STATE(_stack_limit), entry_size); // inform interpreter 1604 __ subptr(rdx, entry_size); // move expression stack bottom 1605 __ movptr(STATE(_stack_base), rdx); // inform interpreter 1606 __ movptr(rcx, STATE(_stack)); // set start value for copy loop 1607 __ jmp(entry); 1608 // 2. move expression stack contents 1609 __ bind(loop); 1610 __ movptr(rbx, Address(rcx, entry_size)); // load expression stack word from old location 1611 __ movptr(Address(rcx, 0), rbx); // and store it at new location 1612 __ addptr(rcx, wordSize); // advance to next word 1613 __ bind(entry); 1614 __ cmpptr(rcx, rdx); // check if bottom reached 1615 __ jcc(Assembler::notEqual, loop); // if not at bottom then copy next word 1616 // now zero the slot so we can find it. 1617 __ movptr(Address(rdx, BasicObjectLock::obj_offset_in_bytes()), (int32_t) NULL_WORD); 1618 __ movl(STATE(_msg), (int)BytecodeInterpreter::got_monitors); 1619 } 1620 1621 1622 // Initial entry to C++ interpreter from the call_stub. 1623 // This entry point is called the frame manager since it handles the generation 1624 // of interpreter activation frames via requests directly from the vm (via call_stub) 1625 // and via requests from the interpreter. The requests from the call_stub happen 1626 // directly thru the entry point. Requests from the interpreter happen via returning 1627 // from the interpreter and examining the message the interpreter has returned to 1628 // the frame manager. The frame manager can take the following requests: 1629 1630 // NO_REQUEST - error, should never happen. 1631 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and 1632 // allocate a new monitor. 1633 // CALL_METHOD - setup a new activation to call a new method. Very similar to what 1634 // happens during entry during the entry via the call stub. 1635 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub. 1636 // 1637 // Arguments: 1638 // 1639 // rbx: Method* 1640 // rcx: receiver - unused (retrieved from stack as needed) 1641 // rsi/r13: previous frame manager state (NULL from the call_stub/c1/c2) 1642 // 1643 // 1644 // Stack layout at entry 1645 // 1646 // [ return address ] <--- rsp 1647 // [ parameter n ] 1648 // ... 1649 // [ parameter 1 ] 1650 // [ expression stack ] 1651 // 1652 // 1653 // We are free to blow any registers we like because the call_stub which brought us here 1654 // initially has preserved the callee save registers already. 1655 // 1656 // 1657 1658 static address interpreter_frame_manager = NULL; 1659 1660 address InterpreterGenerator::generate_normal_entry(bool synchronized) { 1661 1662 // rbx: Method* 1663 // rsi/r13: sender sp 1664 1665 // Because we redispatch "recursive" interpreter entries thru this same entry point 1666 // the "input" register usage is a little strange and not what you expect coming 1667 // from the call_stub. From the call stub rsi/rdi (current/previous) interpreter 1668 // state are NULL but on "recursive" dispatches they are what you'd expect. 1669 // rsi: current interpreter state (C++ interpreter) must preserve (null from call_stub/c1/c2) 1670 1671 1672 // A single frame manager is plenty as we don't specialize for synchronized. We could and 1673 // the code is pretty much ready. Would need to change the test below and for good measure 1674 // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized 1675 // routines. Not clear this is worth it yet. 1676 1677 if (interpreter_frame_manager) return interpreter_frame_manager; 1678 1679 address entry_point = __ pc(); 1680 1681 // Fast accessor methods share this entry point. 1682 // This works because frame manager is in the same codelet 1683 if (UseFastAccessorMethods && !synchronized) __ bind(fast_accessor_slow_entry_path); 1684 1685 Label dispatch_entry_2; 1686 __ movptr(rcx, sender_sp_on_entry); 1687 __ movptr(state, (int32_t)NULL_WORD); // no current activation 1688 1689 __ jmp(dispatch_entry_2); 1690 1691 const Register locals = rdi; 1692 1693 Label re_dispatch; 1694 1695 __ bind(re_dispatch); 1696 1697 // save sender sp (doesn't include return address 1698 __ lea(rcx, Address(rsp, wordSize)); 1699 1700 __ bind(dispatch_entry_2); 1701 1702 // save sender sp 1703 __ push(rcx); 1704 1705 const Address constMethod (rbx, Method::const_offset()); 1706 const Address access_flags (rbx, Method::access_flags_offset()); 1707 const Address size_of_parameters(rdx, ConstMethod::size_of_parameters_offset()); 1708 const Address size_of_locals (rdx, ConstMethod::size_of_locals_offset()); 1709 1710 // const Address monitor_block_top (rbp, frame::interpreter_frame_monitor_block_top_offset * wordSize); 1711 // const Address monitor_block_bot (rbp, frame::interpreter_frame_initial_sp_offset * wordSize); 1712 // const Address monitor(rbp, frame::interpreter_frame_initial_sp_offset * wordSize - (int)sizeof(BasicObjectLock)); 1713 1714 // get parameter size (always needed) 1715 __ movptr(rdx, constMethod); 1716 __ load_unsigned_short(rcx, size_of_parameters); 1717 1718 // rbx: Method* 1719 // rcx: size of parameters 1720 __ load_unsigned_short(rdx, size_of_locals); // get size of locals in words 1721 1722 __ subptr(rdx, rcx); // rdx = no. of additional locals 1723 1724 // see if we've got enough room on the stack for locals plus overhead. 1725 generate_stack_overflow_check(); // C++ 1726 1727 // c++ interpreter does not use stack banging or any implicit exceptions 1728 // leave for now to verify that check is proper. 1729 bang_stack_shadow_pages(false); 1730 1731 1732 1733 // compute beginning of parameters (rdi) 1734 __ lea(locals, Address(rsp, rcx, Address::times_ptr, wordSize)); 1735 1736 // save sender's sp 1737 // __ movl(rcx, rsp); 1738 1739 // get sender's sp 1740 __ pop(rcx); 1741 1742 // get return address 1743 __ pop(rax); 1744 1745 // rdx - # of additional locals 1746 // allocate space for locals 1747 // explicitly initialize locals 1748 { 1749 Label exit, loop; 1750 __ testl(rdx, rdx); // (32bit ok) 1751 __ jcc(Assembler::lessEqual, exit); // do nothing if rdx <= 0 1752 __ bind(loop); 1753 __ push((int32_t)NULL_WORD); // initialize local variables 1754 __ decrement(rdx); // until everything initialized 1755 __ jcc(Assembler::greater, loop); 1756 __ bind(exit); 1757 } 1758 1759 1760 // Assumes rax = return address 1761 1762 // allocate and initialize new interpreterState and method expression stack 1763 // IN(locals) -> locals 1764 // IN(state) -> any current interpreter activation 1765 // destroys rax, rcx, rdx, rdi 1766 // OUT (state) -> new interpreterState 1767 // OUT(rsp) -> bottom of methods expression stack 1768 1769 generate_compute_interpreter_state(state, locals, rcx, false); 1770 1771 // Call interpreter 1772 1773 Label call_interpreter; 1774 __ bind(call_interpreter); 1775 1776 // c++ interpreter does not use stack banging or any implicit exceptions 1777 // leave for now to verify that check is proper. 1778 bang_stack_shadow_pages(false); 1779 1780 1781 // Call interpreter enter here if message is 1782 // set and we know stack size is valid 1783 1784 Label call_interpreter_2; 1785 1786 __ bind(call_interpreter_2); 1787 1788 { 1789 const Register thread = NOT_LP64(rcx) LP64_ONLY(r15_thread); 1790 1791 #ifdef _LP64 1792 __ mov(c_rarg0, state); 1793 #else 1794 __ push(state); // push arg to interpreter 1795 __ movptr(thread, STATE(_thread)); 1796 #endif // _LP64 1797 1798 // We can setup the frame anchor with everything we want at this point 1799 // as we are thread_in_Java and no safepoints can occur until we go to 1800 // vm mode. We do have to clear flags on return from vm but that is it 1801 // 1802 __ movptr(Address(thread, JavaThread::last_Java_fp_offset()), rbp); 1803 __ movptr(Address(thread, JavaThread::last_Java_sp_offset()), rsp); 1804 1805 // Call the interpreter 1806 1807 RuntimeAddress normal(CAST_FROM_FN_PTR(address, BytecodeInterpreter::run)); 1808 RuntimeAddress checking(CAST_FROM_FN_PTR(address, BytecodeInterpreter::runWithChecks)); 1809 1810 __ call(JvmtiExport::can_post_interpreter_events() ? checking : normal); 1811 NOT_LP64(__ pop(rax);) // discard parameter to run 1812 // 1813 // state is preserved since it is callee saved 1814 // 1815 1816 // reset_last_Java_frame 1817 1818 NOT_LP64(__ movl(thread, STATE(_thread));) 1819 __ reset_last_Java_frame(thread, true, true); 1820 } 1821 1822 // examine msg from interpreter to determine next action 1823 1824 __ movl(rdx, STATE(_msg)); // Get new message 1825 1826 Label call_method; 1827 Label return_from_interpreted_method; 1828 Label throw_exception; 1829 Label bad_msg; 1830 Label do_OSR; 1831 1832 __ cmpl(rdx, (int32_t)BytecodeInterpreter::call_method); 1833 __ jcc(Assembler::equal, call_method); 1834 __ cmpl(rdx, (int32_t)BytecodeInterpreter::return_from_method); 1835 __ jcc(Assembler::equal, return_from_interpreted_method); 1836 __ cmpl(rdx, (int32_t)BytecodeInterpreter::do_osr); 1837 __ jcc(Assembler::equal, do_OSR); 1838 __ cmpl(rdx, (int32_t)BytecodeInterpreter::throwing_exception); 1839 __ jcc(Assembler::equal, throw_exception); 1840 __ cmpl(rdx, (int32_t)BytecodeInterpreter::more_monitors); 1841 __ jcc(Assembler::notEqual, bad_msg); 1842 1843 // Allocate more monitor space, shuffle expression stack.... 1844 1845 generate_more_monitors(); 1846 1847 __ jmp(call_interpreter); 1848 1849 // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode) 1850 unctrap_frame_manager_entry = __ pc(); 1851 // 1852 // Load the registers we need. 1853 __ lea(state, Address(rbp, -(int)sizeof(BytecodeInterpreter))); 1854 __ movptr(rsp, STATE(_stack_limit)); // restore expression stack to full depth 1855 __ jmp(call_interpreter_2); 1856 1857 1858 1859 //============================================================================= 1860 // Returning from a compiled method into a deopted method. The bytecode at the 1861 // bcp has completed. The result of the bytecode is in the native abi (the tosca 1862 // for the template based interpreter). Any stack space that was used by the 1863 // bytecode that has completed has been removed (e.g. parameters for an invoke) 1864 // so all that we have to do is place any pending result on the expression stack 1865 // and resume execution on the next bytecode. 1866 1867 1868 generate_deopt_handling(); 1869 __ jmp(call_interpreter); 1870 1871 1872 // Current frame has caught an exception we need to dispatch to the 1873 // handler. We can get here because a native interpreter frame caught 1874 // an exception in which case there is no handler and we must rethrow 1875 // If it is a vanilla interpreted frame the we simply drop into the 1876 // interpreter and let it do the lookup. 1877 1878 Interpreter::_rethrow_exception_entry = __ pc(); 1879 // rax: exception 1880 // rdx: return address/pc that threw exception 1881 1882 Label return_with_exception; 1883 Label unwind_and_forward; 1884 1885 // restore state pointer. 1886 __ lea(state, Address(rbp, -(int)sizeof(BytecodeInterpreter))); 1887 1888 __ movptr(rbx, STATE(_method)); // get method 1889 #ifdef _LP64 1890 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), rax); 1891 #else 1892 __ movl(rcx, STATE(_thread)); // get thread 1893 1894 // Store exception with interpreter will expect it 1895 __ movptr(Address(rcx, Thread::pending_exception_offset()), rax); 1896 #endif // _LP64 1897 1898 // is current frame vanilla or native? 1899 1900 __ movl(rdx, access_flags); 1901 __ testl(rdx, JVM_ACC_NATIVE); 1902 __ jcc(Assembler::zero, return_with_exception); // vanilla interpreted frame, handle directly 1903 1904 // We drop thru to unwind a native interpreted frame with a pending exception 1905 // We jump here for the initial interpreter frame with exception pending 1906 // We unwind the current acivation and forward it to our caller. 1907 1908 __ bind(unwind_and_forward); 1909 1910 // unwind rbp, return stack to unextended value and re-push return address 1911 1912 __ movptr(rcx, STATE(_sender_sp)); 1913 __ leave(); 1914 __ pop(rdx); 1915 __ mov(rsp, rcx); 1916 __ push(rdx); 1917 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 1918 1919 // Return point from a call which returns a result in the native abi 1920 // (c1/c2/jni-native). This result must be processed onto the java 1921 // expression stack. 1922 // 1923 // A pending exception may be present in which case there is no result present 1924 1925 Label resume_interpreter; 1926 Label do_float; 1927 Label do_double; 1928 Label done_conv; 1929 1930 // The FPU stack is clean if UseSSE >= 2 but must be cleaned in other cases 1931 if (UseSSE < 2) { 1932 __ lea(state, Address(rbp, -(int)sizeof(BytecodeInterpreter))); 1933 __ movptr(rbx, STATE(_result._to_call._callee)); // get method just executed 1934 __ movl(rcx, Address(rbx, Method::result_index_offset())); 1935 __ cmpl(rcx, AbstractInterpreter::BasicType_as_index(T_FLOAT)); // Result stub address array index 1936 __ jcc(Assembler::equal, do_float); 1937 __ cmpl(rcx, AbstractInterpreter::BasicType_as_index(T_DOUBLE)); // Result stub address array index 1938 __ jcc(Assembler::equal, do_double); 1939 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2) 1940 __ empty_FPU_stack(); 1941 #endif // COMPILER2 1942 __ jmp(done_conv); 1943 1944 __ bind(do_float); 1945 #ifdef COMPILER2 1946 for (int i = 1; i < 8; i++) { 1947 __ ffree(i); 1948 } 1949 #endif // COMPILER2 1950 __ jmp(done_conv); 1951 __ bind(do_double); 1952 #ifdef COMPILER2 1953 for (int i = 1; i < 8; i++) { 1954 __ ffree(i); 1955 } 1956 #endif // COMPILER2 1957 __ jmp(done_conv); 1958 } else { 1959 __ MacroAssembler::verify_FPU(0, "generate_return_entry_for compiled"); 1960 __ jmp(done_conv); 1961 } 1962 1963 // Return point to interpreter from compiled/native method 1964 InternalAddress return_from_native_method(__ pc()); 1965 1966 __ bind(done_conv); 1967 1968 1969 // Result if any is in tosca. The java expression stack is in the state that the 1970 // calling convention left it (i.e. params may or may not be present) 1971 // Copy the result from tosca and place it on java expression stack. 1972 1973 // Restore rsi/r13 as compiled code may not preserve it 1974 1975 __ lea(state, Address(rbp, -(int)sizeof(BytecodeInterpreter))); 1976 1977 // restore stack to what we had when we left (in case i2c extended it) 1978 1979 __ movptr(rsp, STATE(_stack)); 1980 __ lea(rsp, Address(rsp, wordSize)); 1981 1982 // If there is a pending exception then we don't really have a result to process 1983 1984 #ifdef _LP64 1985 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD); 1986 #else 1987 __ movptr(rcx, STATE(_thread)); // get thread 1988 __ cmpptr(Address(rcx, Thread::pending_exception_offset()), (int32_t)NULL_WORD); 1989 #endif // _LP64 1990 __ jcc(Assembler::notZero, return_with_exception); 1991 1992 // get method just executed 1993 __ movptr(rbx, STATE(_result._to_call._callee)); 1994 1995 // callee left args on top of expression stack, remove them 1996 __ movptr(rcx, constMethod); 1997 __ load_unsigned_short(rcx, Address(rcx, ConstMethod::size_of_parameters_offset())); 1998 1999 __ lea(rsp, Address(rsp, rcx, Address::times_ptr)); 2000 2001 __ movl(rcx, Address(rbx, Method::result_index_offset())); 2002 ExternalAddress tosca_to_stack((address)CppInterpreter::_tosca_to_stack); 2003 // Address index(noreg, rax, Address::times_ptr); 2004 __ movptr(rcx, ArrayAddress(tosca_to_stack, Address(noreg, rcx, Address::times_ptr))); 2005 // __ movl(rcx, Address(noreg, rcx, Address::times_ptr, int(AbstractInterpreter::_tosca_to_stack))); 2006 __ call(rcx); // call result converter 2007 __ jmp(resume_interpreter); 2008 2009 // An exception is being caught on return to a vanilla interpreter frame. 2010 // Empty the stack and resume interpreter 2011 2012 __ bind(return_with_exception); 2013 2014 // Exception present, empty stack 2015 __ movptr(rsp, STATE(_stack_base)); 2016 __ jmp(resume_interpreter); 2017 2018 // Return from interpreted method we return result appropriate to the caller (i.e. "recursive" 2019 // interpreter call, or native) and unwind this interpreter activation. 2020 // All monitors should be unlocked. 2021 2022 __ bind(return_from_interpreted_method); 2023 2024 Label return_to_initial_caller; 2025 2026 __ movptr(rbx, STATE(_method)); // get method just executed 2027 __ cmpptr(STATE(_prev_link), (int32_t)NULL_WORD); // returning from "recursive" interpreter call? 2028 __ movl(rax, Address(rbx, Method::result_index_offset())); // get result type index 2029 __ jcc(Assembler::equal, return_to_initial_caller); // back to native code (call_stub/c1/c2) 2030 2031 // Copy result to callers java stack 2032 ExternalAddress stack_to_stack((address)CppInterpreter::_stack_to_stack); 2033 // Address index(noreg, rax, Address::times_ptr); 2034 2035 __ movptr(rax, ArrayAddress(stack_to_stack, Address(noreg, rax, Address::times_ptr))); 2036 // __ movl(rax, Address(noreg, rax, Address::times_ptr, int(AbstractInterpreter::_stack_to_stack))); 2037 __ call(rax); // call result converter 2038 2039 Label unwind_recursive_activation; 2040 __ bind(unwind_recursive_activation); 2041 2042 // returning to interpreter method from "recursive" interpreter call 2043 // result converter left rax pointing to top of the java stack for method we are returning 2044 // to. Now all we must do is unwind the state from the completed call 2045 2046 __ movptr(state, STATE(_prev_link)); // unwind state 2047 __ leave(); // pop the frame 2048 __ mov(rsp, rax); // unwind stack to remove args 2049 2050 // Resume the interpreter. The current frame contains the current interpreter 2051 // state object. 2052 // 2053 2054 __ bind(resume_interpreter); 2055 2056 // state == interpreterState object for method we are resuming 2057 2058 __ movl(STATE(_msg), (int)BytecodeInterpreter::method_resume); 2059 __ lea(rsp, Address(rsp, -wordSize)); // prepush stack (result if any already present) 2060 __ movptr(STATE(_stack), rsp); // inform interpreter of new stack depth (parameters removed, 2061 // result if any on stack already ) 2062 __ movptr(rsp, STATE(_stack_limit)); // restore expression stack to full depth 2063 __ jmp(call_interpreter_2); // No need to bang 2064 2065 // interpreter returning to native code (call_stub/c1/c2) 2066 // convert result and unwind initial activation 2067 // rax - result index 2068 2069 __ bind(return_to_initial_caller); 2070 ExternalAddress stack_to_native((address)CppInterpreter::_stack_to_native_abi); 2071 // Address index(noreg, rax, Address::times_ptr); 2072 2073 __ movptr(rax, ArrayAddress(stack_to_native, Address(noreg, rax, Address::times_ptr))); 2074 __ call(rax); // call result converter 2075 2076 Label unwind_initial_activation; 2077 __ bind(unwind_initial_activation); 2078 2079 // RETURN TO CALL_STUB/C1/C2 code (result if any in rax/rdx ST(0)) 2080 2081 /* Current stack picture 2082 2083 [ incoming parameters ] 2084 [ extra locals ] 2085 [ return address to CALL_STUB/C1/C2] 2086 fp -> [ CALL_STUB/C1/C2 fp ] 2087 BytecodeInterpreter object 2088 expression stack 2089 sp -> 2090 2091 */ 2092 2093 // return restoring the stack to the original sender_sp value 2094 2095 __ movptr(rcx, STATE(_sender_sp)); 2096 __ leave(); 2097 __ pop(rdi); // get return address 2098 // set stack to sender's sp 2099 __ mov(rsp, rcx); 2100 __ jmp(rdi); // return to call_stub 2101 2102 // OSR request, adjust return address to make current frame into adapter frame 2103 // and enter OSR nmethod 2104 2105 __ bind(do_OSR); 2106 2107 Label remove_initial_frame; 2108 2109 // We are going to pop this frame. Is there another interpreter frame underneath 2110 // it or is it callstub/compiled? 2111 2112 // Move buffer to the expected parameter location 2113 __ movptr(rcx, STATE(_result._osr._osr_buf)); 2114 2115 __ movptr(rax, STATE(_result._osr._osr_entry)); 2116 2117 __ cmpptr(STATE(_prev_link), (int32_t)NULL_WORD); // returning from "recursive" interpreter call? 2118 __ jcc(Assembler::equal, remove_initial_frame); // back to native code (call_stub/c1/c2) 2119 2120 __ movptr(sender_sp_on_entry, STATE(_sender_sp)); // get sender's sp in expected register 2121 __ leave(); // pop the frame 2122 __ mov(rsp, sender_sp_on_entry); // trim any stack expansion 2123 2124 2125 // We know we are calling compiled so push specialized return 2126 // method uses specialized entry, push a return so we look like call stub setup 2127 // this path will handle fact that result is returned in registers and not 2128 // on the java stack. 2129 2130 __ pushptr(return_from_native_method.addr()); 2131 2132 __ jmp(rax); 2133 2134 __ bind(remove_initial_frame); 2135 2136 __ movptr(rdx, STATE(_sender_sp)); 2137 __ leave(); 2138 // get real return 2139 __ pop(rsi); 2140 // set stack to sender's sp 2141 __ mov(rsp, rdx); 2142 // repush real return 2143 __ push(rsi); 2144 // Enter OSR nmethod 2145 __ jmp(rax); 2146 2147 2148 2149 2150 // Call a new method. All we do is (temporarily) trim the expression stack 2151 // push a return address to bring us back to here and leap to the new entry. 2152 2153 __ bind(call_method); 2154 2155 // stack points to next free location and not top element on expression stack 2156 // method expects sp to be pointing to topmost element 2157 2158 __ movptr(rsp, STATE(_stack)); // pop args to c++ interpreter, set sp to java stack top 2159 __ lea(rsp, Address(rsp, wordSize)); 2160 2161 __ movptr(rbx, STATE(_result._to_call._callee)); // get method to execute 2162 2163 // don't need a return address if reinvoking interpreter 2164 2165 // Make it look like call_stub calling conventions 2166 2167 // Get (potential) receiver 2168 // get size of parameters in words 2169 __ movptr(rcx, constMethod); 2170 __ load_unsigned_short(rcx, Address(rcx, ConstMethod::size_of_parameters_offset())); 2171 2172 ExternalAddress recursive(CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation)); 2173 __ pushptr(recursive.addr()); // make it look good in the debugger 2174 2175 InternalAddress entry(entry_point); 2176 __ cmpptr(STATE(_result._to_call._callee_entry_point), entry.addr()); // returning to interpreter? 2177 __ jcc(Assembler::equal, re_dispatch); // yes 2178 2179 __ pop(rax); // pop dummy address 2180 2181 2182 // get specialized entry 2183 __ movptr(rax, STATE(_result._to_call._callee_entry_point)); 2184 // set sender SP 2185 __ mov(sender_sp_on_entry, rsp); 2186 2187 // method uses specialized entry, push a return so we look like call stub setup 2188 // this path will handle fact that result is returned in registers and not 2189 // on the java stack. 2190 2191 __ pushptr(return_from_native_method.addr()); 2192 2193 __ jmp(rax); 2194 2195 __ bind(bad_msg); 2196 __ stop("Bad message from interpreter"); 2197 2198 // Interpreted method "returned" with an exception pass it on... 2199 // Pass result, unwind activation and continue/return to interpreter/call_stub 2200 // We handle result (if any) differently based on return to interpreter or call_stub 2201 2202 Label unwind_initial_with_pending_exception; 2203 2204 __ bind(throw_exception); 2205 __ cmpptr(STATE(_prev_link), (int32_t)NULL_WORD); // returning from recursive interpreter call? 2206 __ jcc(Assembler::equal, unwind_initial_with_pending_exception); // no, back to native code (call_stub/c1/c2) 2207 __ movptr(rax, STATE(_locals)); // pop parameters get new stack value 2208 __ addptr(rax, wordSize); // account for prepush before we return 2209 __ jmp(unwind_recursive_activation); 2210 2211 __ bind(unwind_initial_with_pending_exception); 2212 2213 // We will unwind the current (initial) interpreter frame and forward 2214 // the exception to the caller. We must put the exception in the 2215 // expected register and clear pending exception and then forward. 2216 2217 __ jmp(unwind_and_forward); 2218 2219 interpreter_frame_manager = entry_point; 2220 return entry_point; 2221 } 2222 2223 address AbstractInterpreterGenerator::generate_method_entry(AbstractInterpreter::MethodKind kind) { 2224 // determine code generation flags 2225 bool synchronized = false; 2226 address entry_point = NULL; 2227 2228 switch (kind) { 2229 case Interpreter::zerolocals : break; 2230 case Interpreter::zerolocals_synchronized: synchronized = true; break; 2231 case Interpreter::native : entry_point = ((InterpreterGenerator*)this)->generate_native_entry(false); break; 2232 case Interpreter::native_synchronized : entry_point = ((InterpreterGenerator*)this)->generate_native_entry(true); break; 2233 case Interpreter::empty : entry_point = ((InterpreterGenerator*)this)->generate_empty_entry(); break; 2234 case Interpreter::accessor : entry_point = ((InterpreterGenerator*)this)->generate_accessor_entry(); break; 2235 case Interpreter::abstract : entry_point = ((InterpreterGenerator*)this)->generate_abstract_entry(); break; 2236 case Interpreter::method_handle : entry_point = ((InterpreterGenerator*)this)->generate_method_handle_entry(); break; 2237 2238 case Interpreter::java_lang_math_sin : // fall thru 2239 case Interpreter::java_lang_math_cos : // fall thru 2240 case Interpreter::java_lang_math_tan : // fall thru 2241 case Interpreter::java_lang_math_abs : // fall thru 2242 case Interpreter::java_lang_math_log : // fall thru 2243 case Interpreter::java_lang_math_log10 : // fall thru 2244 case Interpreter::java_lang_math_sqrt : entry_point = ((InterpreterGenerator*)this)->generate_math_entry(kind); break; 2245 case Interpreter::java_lang_ref_reference_get 2246 : entry_point = ((InterpreterGenerator*)this)->generate_Reference_get_entry(); break; 2247 default : ShouldNotReachHere(); break; 2248 } 2249 2250 if (entry_point) return entry_point; 2251 2252 return ((InterpreterGenerator*)this)->generate_normal_entry(synchronized); 2253 2254 } 2255 2256 InterpreterGenerator::InterpreterGenerator(StubQueue* code) 2257 : CppInterpreterGenerator(code) { 2258 generate_all(); // down here so it can be "virtual" 2259 } 2260 2261 // Deoptimization helpers for C++ interpreter 2262 2263 // How much stack a method activation needs in words. 2264 int AbstractInterpreter::size_top_interpreter_activation(Method* method) { 2265 2266 const int stub_code = 4; // see generate_call_stub 2267 // Save space for one monitor to get into the interpreted method in case 2268 // the method is synchronized 2269 int monitor_size = method->is_synchronized() ? 2270 1*frame::interpreter_frame_monitor_size() : 0; 2271 2272 // total static overhead size. Account for interpreter state object, return 2273 // address, saved rbp and 2 words for a "static long no_params() method" issue. 2274 2275 const int overhead_size = sizeof(BytecodeInterpreter)/wordSize + 2276 ( frame::sender_sp_offset - frame::link_offset) + 2; 2277 2278 const int method_stack = (method->max_locals() + method->max_stack()) * 2279 Interpreter::stackElementWords; 2280 return overhead_size + method_stack + stub_code; 2281 } 2282 2283 // returns the activation size. 2284 static int size_activation_helper(int extra_locals_size, int monitor_size) { 2285 return (extra_locals_size + // the addition space for locals 2286 2*BytesPerWord + // return address and saved rbp 2287 2*BytesPerWord + // "static long no_params() method" issue 2288 sizeof(BytecodeInterpreter) + // interpreterState 2289 monitor_size); // monitors 2290 } 2291 2292 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill, 2293 frame* caller, 2294 frame* current, 2295 Method* method, 2296 intptr_t* locals, 2297 intptr_t* stack, 2298 intptr_t* stack_base, 2299 intptr_t* monitor_base, 2300 intptr_t* frame_bottom, 2301 bool is_top_frame 2302 ) 2303 { 2304 // What about any vtable? 2305 // 2306 to_fill->_thread = JavaThread::current(); 2307 // This gets filled in later but make it something recognizable for now 2308 to_fill->_bcp = method->code_base(); 2309 to_fill->_locals = locals; 2310 to_fill->_constants = method->constants()->cache(); 2311 to_fill->_method = method; 2312 to_fill->_mdx = NULL; 2313 to_fill->_stack = stack; 2314 if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) { 2315 to_fill->_msg = deopt_resume2; 2316 } else { 2317 to_fill->_msg = method_resume; 2318 } 2319 to_fill->_result._to_call._bcp_advance = 0; 2320 to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone 2321 to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone 2322 to_fill->_prev_link = NULL; 2323 2324 to_fill->_sender_sp = caller->unextended_sp(); 2325 2326 if (caller->is_interpreted_frame()) { 2327 interpreterState prev = caller->get_interpreterState(); 2328 to_fill->_prev_link = prev; 2329 // *current->register_addr(GR_Iprev_state) = (intptr_t) prev; 2330 // Make the prev callee look proper 2331 prev->_result._to_call._callee = method; 2332 if (*prev->_bcp == Bytecodes::_invokeinterface) { 2333 prev->_result._to_call._bcp_advance = 5; 2334 } else { 2335 prev->_result._to_call._bcp_advance = 3; 2336 } 2337 } 2338 to_fill->_oop_temp = NULL; 2339 to_fill->_stack_base = stack_base; 2340 // Need +1 here because stack_base points to the word just above the first expr stack entry 2341 // and stack_limit is supposed to point to the word just below the last expr stack entry. 2342 // See generate_compute_interpreter_state. 2343 to_fill->_stack_limit = stack_base - (method->max_stack() + 1); 2344 to_fill->_monitor_base = (BasicObjectLock*) monitor_base; 2345 2346 to_fill->_self_link = to_fill; 2347 assert(stack >= to_fill->_stack_limit && stack < to_fill->_stack_base, 2348 "Stack top out of range"); 2349 } 2350 2351 int AbstractInterpreter::layout_activation(Method* method, 2352 int tempcount, // 2353 int popframe_extra_args, 2354 int moncount, 2355 int caller_actual_parameters, 2356 int callee_param_count, 2357 int callee_locals, 2358 frame* caller, 2359 frame* interpreter_frame, 2360 bool is_top_frame, 2361 bool is_bottom_frame) { 2362 2363 assert(popframe_extra_args == 0, "FIX ME"); 2364 // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state() 2365 // does as far as allocating an interpreter frame. 2366 // If interpreter_frame!=NULL, set up the method, locals, and monitors. 2367 // The frame interpreter_frame, if not NULL, is guaranteed to be the right size, 2368 // as determined by a previous call to this method. 2369 // It is also guaranteed to be walkable even though it is in a skeletal state 2370 // NOTE: return size is in words not bytes 2371 // NOTE: tempcount is the current size of the java expression stack. For top most 2372 // frames we will allocate a full sized expression stack and not the curback 2373 // version that non-top frames have. 2374 2375 // Calculate the amount our frame will be adjust by the callee. For top frame 2376 // this is zero. 2377 2378 // NOTE: ia64 seems to do this wrong (or at least backwards) in that it 2379 // calculates the extra locals based on itself. Not what the callee does 2380 // to it. So it ignores last_frame_adjust value. Seems suspicious as far 2381 // as getting sender_sp correct. 2382 2383 int extra_locals_size = (callee_locals - callee_param_count) * BytesPerWord; 2384 int monitor_size = sizeof(BasicObjectLock) * moncount; 2385 2386 // First calculate the frame size without any java expression stack 2387 int short_frame_size = size_activation_helper(extra_locals_size, 2388 monitor_size); 2389 2390 // Now with full size expression stack 2391 int full_frame_size = short_frame_size + method->max_stack() * BytesPerWord; 2392 2393 // and now with only live portion of the expression stack 2394 short_frame_size = short_frame_size + tempcount * BytesPerWord; 2395 2396 // the size the activation is right now. Only top frame is full size 2397 int frame_size = (is_top_frame ? full_frame_size : short_frame_size); 2398 2399 if (interpreter_frame != NULL) { 2400 #ifdef ASSERT 2401 assert(caller->unextended_sp() == interpreter_frame->interpreter_frame_sender_sp(), "Frame not properly walkable"); 2402 #endif 2403 2404 // MUCHO HACK 2405 2406 intptr_t* frame_bottom = (intptr_t*) ((intptr_t)interpreter_frame->sp() - (full_frame_size - frame_size)); 2407 2408 /* Now fillin the interpreterState object */ 2409 2410 // The state object is the first thing on the frame and easily located 2411 2412 interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() - sizeof(BytecodeInterpreter)); 2413 2414 2415 // Find the locals pointer. This is rather simple on x86 because there is no 2416 // confusing rounding at the callee to account for. We can trivially locate 2417 // our locals based on the current fp(). 2418 // Note: the + 2 is for handling the "static long no_params() method" issue. 2419 // (too bad I don't really remember that issue well...) 2420 2421 intptr_t* locals; 2422 // If the caller is interpreted we need to make sure that locals points to the first 2423 // argument that the caller passed and not in an area where the stack might have been extended. 2424 // because the stack to stack to converter needs a proper locals value in order to remove the 2425 // arguments from the caller and place the result in the proper location. Hmm maybe it'd be 2426 // simpler if we simply stored the result in the BytecodeInterpreter object and let the c++ code 2427 // adjust the stack?? HMMM QQQ 2428 // 2429 if (caller->is_interpreted_frame()) { 2430 // locals must agree with the caller because it will be used to set the 2431 // caller's tos when we return. 2432 interpreterState prev = caller->get_interpreterState(); 2433 // stack() is prepushed. 2434 locals = prev->stack() + method->size_of_parameters(); 2435 // locals = caller->unextended_sp() + (method->size_of_parameters() - 1); 2436 if (locals != interpreter_frame->fp() + frame::sender_sp_offset + (method->max_locals() - 1) + 2) { 2437 // os::breakpoint(); 2438 } 2439 } else { 2440 // this is where a c2i would have placed locals (except for the +2) 2441 locals = interpreter_frame->fp() + frame::sender_sp_offset + (method->max_locals() - 1) + 2; 2442 } 2443 2444 intptr_t* monitor_base = (intptr_t*) cur_state; 2445 intptr_t* stack_base = (intptr_t*) ((intptr_t) monitor_base - monitor_size); 2446 /* +1 because stack is always prepushed */ 2447 intptr_t* stack = (intptr_t*) ((intptr_t) stack_base - (tempcount + 1) * BytesPerWord); 2448 2449 2450 BytecodeInterpreter::layout_interpreterState(cur_state, 2451 caller, 2452 interpreter_frame, 2453 method, 2454 locals, 2455 stack, 2456 stack_base, 2457 monitor_base, 2458 frame_bottom, 2459 is_top_frame); 2460 2461 // BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp()); 2462 } 2463 return frame_size/BytesPerWord; 2464 } 2465 2466 #endif // CC_INTERP (all)