1 /* 2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved. 3 * Copyright 2012, 2013 SAP AG. All rights reserved. 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 5 * 6 * This code is free software; you can redistribute it and/or modify it 7 * under the terms of the GNU General Public License version 2 only, as 8 * published by the Free Software Foundation. 9 * 10 * This code is distributed in the hope that it will be useful, but WITHOUT 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 13 * version 2 for more details (a copy is included in the LICENSE file that 14 * accompanied this code). 15 * 16 * You should have received a copy of the GNU General Public License version 17 * 2 along with this work; if not, write to the Free Software Foundation, 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 19 * 20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 21 * or visit www.oracle.com if you need additional information or have any 22 * questions. 23 * 24 */ 25 26 #include "precompiled.hpp" 27 #include "asm/assembler.hpp" 28 #include "asm/macroAssembler.inline.hpp" 29 #include "interpreter/bytecodeHistogram.hpp" 30 #include "interpreter/cppInterpreter.hpp" 31 #include "interpreter/interpreter.hpp" 32 #include "interpreter/interpreterGenerator.hpp" 33 #include "interpreter/interpreterRuntime.hpp" 34 #include "oops/arrayOop.hpp" 35 #include "oops/methodData.hpp" 36 #include "oops/method.hpp" 37 #include "oops/oop.inline.hpp" 38 #include "prims/jvmtiExport.hpp" 39 #include "prims/jvmtiThreadState.hpp" 40 #include "runtime/arguments.hpp" 41 #include "runtime/deoptimization.hpp" 42 #include "runtime/frame.inline.hpp" 43 #include "runtime/interfaceSupport.hpp" 44 #include "runtime/sharedRuntime.hpp" 45 #include "runtime/stubRoutines.hpp" 46 #include "runtime/synchronizer.hpp" 47 #include "runtime/timer.hpp" 48 #include "runtime/vframeArray.hpp" 49 #include "utilities/debug.hpp" 50 #ifdef SHARK 51 #include "shark/shark_globals.hpp" 52 #endif 53 54 #ifdef CC_INTERP 55 56 #define __ _masm-> 57 58 // Contains is used for identifying interpreter frames during a stack-walk. 59 // A frame with a PC in InterpretMethod must be identified as a normal C frame. 60 bool CppInterpreter::contains(address pc) { 61 return _code->contains(pc); 62 } 63 64 #ifdef PRODUCT 65 #define BLOCK_COMMENT(str) // nothing 66 #else 67 #define BLOCK_COMMENT(str) __ block_comment(str) 68 #endif 69 70 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":") 71 72 static address interpreter_frame_manager = NULL; 73 static address frame_manager_specialized_return = NULL; 74 static address native_entry = NULL; 75 76 static address interpreter_return_address = NULL; 77 78 static address unctrap_frame_manager_entry = NULL; 79 80 static address deopt_frame_manager_return_atos = NULL; 81 static address deopt_frame_manager_return_btos = NULL; 82 static address deopt_frame_manager_return_itos = NULL; 83 static address deopt_frame_manager_return_ltos = NULL; 84 static address deopt_frame_manager_return_ftos = NULL; 85 static address deopt_frame_manager_return_dtos = NULL; 86 static address deopt_frame_manager_return_vtos = NULL; 87 88 // A result handler converts/unboxes a native call result into 89 // a java interpreter/compiler result. The current frame is an 90 // interpreter frame. 91 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) { 92 return AbstractInterpreterGenerator::generate_result_handler_for(type); 93 } 94 95 // tosca based result to c++ interpreter stack based result. 96 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) { 97 // 98 // A result is in the native abi result register from a native 99 // method call. We need to return this result to the interpreter by 100 // pushing the result on the interpreter's stack. 101 // 102 // Registers alive: 103 // R3_ARG1(R3_RET)/F1_ARG1(F1_RET) - result to move 104 // R4_ARG2 - address of tos 105 // LR 106 // 107 // Registers updated: 108 // R3_RET(R3_ARG1) - address of new tos (== R17_tos for T_VOID) 109 // 110 111 int number_of_used_slots = 1; 112 113 const Register tos = R4_ARG2; 114 Label done; 115 Label is_false; 116 117 address entry = __ pc(); 118 119 switch (type) { 120 case T_BOOLEAN: 121 __ cmpwi(CCR0, R3_RET, 0); 122 __ beq(CCR0, is_false); 123 __ li(R3_RET, 1); 124 __ stw(R3_RET, 0, tos); 125 __ b(done); 126 __ bind(is_false); 127 __ li(R3_RET, 0); 128 __ stw(R3_RET, 0, tos); 129 break; 130 case T_BYTE: 131 case T_CHAR: 132 case T_SHORT: 133 case T_INT: 134 __ stw(R3_RET, 0, tos); 135 break; 136 case T_LONG: 137 number_of_used_slots = 2; 138 // mark unused slot for debugging 139 // long goes to topmost slot 140 __ std(R3_RET, -BytesPerWord, tos); 141 __ li(R3_RET, 0); 142 __ std(R3_RET, 0, tos); 143 break; 144 case T_OBJECT: 145 __ verify_oop(R3_RET); 146 __ std(R3_RET, 0, tos); 147 break; 148 case T_FLOAT: 149 __ stfs(F1_RET, 0, tos); 150 break; 151 case T_DOUBLE: 152 number_of_used_slots = 2; 153 // mark unused slot for debugging 154 __ li(R3_RET, 0); 155 __ std(R3_RET, 0, tos); 156 // double goes to topmost slot 157 __ stfd(F1_RET, -BytesPerWord, tos); 158 break; 159 case T_VOID: 160 number_of_used_slots = 0; 161 break; 162 default: 163 ShouldNotReachHere(); 164 } 165 166 __ BIND(done); 167 168 // new expression stack top 169 __ addi(R3_RET, tos, -BytesPerWord * number_of_used_slots); 170 171 __ blr(); 172 173 return entry; 174 } 175 176 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) { 177 // 178 // Copy the result from the callee's stack to the caller's stack, 179 // caller and callee both being interpreted. 180 // 181 // Registers alive 182 // R3_ARG1 - address of callee's tos + BytesPerWord 183 // R4_ARG2 - address of caller's tos [i.e. free location] 184 // LR 185 // 186 // stack grows upwards, memory grows downwards. 187 // 188 // [ free ] <-- callee's tos 189 // [ optional result ] <-- R3_ARG1 190 // [ optional dummy ] 191 // ... 192 // [ free ] <-- caller's tos, R4_ARG2 193 // ... 194 // Registers updated 195 // R3_RET(R3_ARG1) - address of caller's new tos 196 // 197 // stack grows upwards, memory grows downwards. 198 // 199 // [ free ] <-- current tos, R3_RET 200 // [ optional result ] 201 // [ optional dummy ] 202 // ... 203 // 204 205 const Register from = R3_ARG1; 206 const Register ret = R3_ARG1; 207 const Register tos = R4_ARG2; 208 const Register tmp1 = R21_tmp1; 209 const Register tmp2 = R22_tmp2; 210 211 address entry = __ pc(); 212 213 switch (type) { 214 case T_BOOLEAN: 215 case T_BYTE: 216 case T_CHAR: 217 case T_SHORT: 218 case T_INT: 219 case T_FLOAT: 220 __ lwz(tmp1, 0, from); 221 __ stw(tmp1, 0, tos); 222 // New expression stack top. 223 __ addi(ret, tos, - BytesPerWord); 224 break; 225 case T_LONG: 226 case T_DOUBLE: 227 // Move both entries for debug purposes even though only one is live. 228 __ ld(tmp1, BytesPerWord, from); 229 __ ld(tmp2, 0, from); 230 __ std(tmp1, 0, tos); 231 __ std(tmp2, -BytesPerWord, tos); 232 // New expression stack top. 233 __ addi(ret, tos, - 2 * BytesPerWord); // two slots 234 break; 235 case T_OBJECT: 236 __ ld(tmp1, 0, from); 237 __ verify_oop(tmp1); 238 __ std(tmp1, 0, tos); 239 // New expression stack top. 240 __ addi(ret, tos, - BytesPerWord); 241 break; 242 case T_VOID: 243 // New expression stack top. 244 __ mr(ret, tos); 245 break; 246 default: 247 ShouldNotReachHere(); 248 } 249 250 __ blr(); 251 252 return entry; 253 } 254 255 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) { 256 // 257 // Load a result from the callee's stack into the caller's expecting 258 // return register, callee being interpreted, caller being call stub 259 // or jit code. 260 // 261 // Registers alive 262 // R3_ARG1 - callee expression tos + BytesPerWord 263 // LR 264 // 265 // stack grows upwards, memory grows downwards. 266 // 267 // [ free ] <-- callee's tos 268 // [ optional result ] <-- R3_ARG1 269 // [ optional dummy ] 270 // ... 271 // 272 // Registers updated 273 // R3_RET(R3_ARG1)/F1_RET - result 274 // 275 276 const Register from = R3_ARG1; 277 const Register ret = R3_ARG1; 278 const FloatRegister fret = F1_ARG1; 279 280 address entry = __ pc(); 281 282 // Implemented uniformly for both kinds of endianness. The interpreter 283 // implements boolean, byte, char, and short as jint (4 bytes). 284 switch (type) { 285 case T_BOOLEAN: 286 case T_CHAR: 287 // zero extension 288 __ lwz(ret, 0, from); 289 break; 290 case T_BYTE: 291 case T_SHORT: 292 case T_INT: 293 // sign extension 294 __ lwa(ret, 0, from); 295 break; 296 case T_LONG: 297 __ ld(ret, 0, from); 298 break; 299 case T_OBJECT: 300 __ ld(ret, 0, from); 301 __ verify_oop(ret); 302 break; 303 case T_FLOAT: 304 __ lfs(fret, 0, from); 305 break; 306 case T_DOUBLE: 307 __ lfd(fret, 0, from); 308 break; 309 case T_VOID: 310 break; 311 default: 312 ShouldNotReachHere(); 313 } 314 315 __ blr(); 316 317 return entry; 318 } 319 320 address CppInterpreter::return_entry(TosState state, int length, Bytecodes::Code code) { 321 assert(interpreter_return_address != NULL, "Not initialized"); 322 return interpreter_return_address; 323 } 324 325 address CppInterpreter::deopt_entry(TosState state, int length) { 326 address ret = NULL; 327 if (length != 0) { 328 switch (state) { 329 case atos: ret = deopt_frame_manager_return_atos; break; 330 case btos: ret = deopt_frame_manager_return_itos; break; 331 case ctos: 332 case stos: 333 case itos: ret = deopt_frame_manager_return_itos; break; 334 case ltos: ret = deopt_frame_manager_return_ltos; break; 335 case ftos: ret = deopt_frame_manager_return_ftos; break; 336 case dtos: ret = deopt_frame_manager_return_dtos; break; 337 case vtos: ret = deopt_frame_manager_return_vtos; break; 338 default: ShouldNotReachHere(); 339 } 340 } else { 341 ret = unctrap_frame_manager_entry; // re-execute the bytecode (e.g. uncommon trap, popframe) 342 } 343 assert(ret != NULL, "Not initialized"); 344 return ret; 345 } 346 347 // 348 // Helpers for commoning out cases in the various type of method entries. 349 // 350 351 // 352 // Registers alive 353 // R16_thread - JavaThread* 354 // R1_SP - old stack pointer 355 // R19_method - callee's Method 356 // R17_tos - address of caller's tos (prepushed) 357 // R15_prev_state - address of caller's BytecodeInterpreter or 0 358 // return_pc in R21_tmp15 (only when called within generate_native_entry) 359 // 360 // Registers updated 361 // R14_state - address of callee's interpreter state 362 // R1_SP - new stack pointer 363 // CCR4_is_synced - current method is synchronized 364 // 365 void CppInterpreterGenerator::generate_compute_interpreter_state(Label& stack_overflow_return) { 366 // 367 // Stack layout at this point: 368 // 369 // F1 [TOP_IJAVA_FRAME_ABI] <-- R1_SP 370 // alignment (optional) 371 // [F1's outgoing Java arguments] <-- R17_tos 372 // ... 373 // F2 [PARENT_IJAVA_FRAME_ABI] 374 // ... 375 376 //============================================================================= 377 // Allocate space for locals other than the parameters, the 378 // interpreter state, monitors, and the expression stack. 379 380 const Register local_count = R21_tmp1; 381 const Register parameter_count = R22_tmp2; 382 const Register max_stack = R23_tmp3; 383 // Must not be overwritten within this method! 384 // const Register return_pc = R29_tmp9; 385 386 const ConditionRegister is_synced = CCR4_is_synced; 387 const ConditionRegister is_native = CCR6; 388 const ConditionRegister is_static = CCR7; 389 390 assert(is_synced != is_native, "condition code registers must be distinct"); 391 assert(is_synced != is_static, "condition code registers must be distinct"); 392 assert(is_native != is_static, "condition code registers must be distinct"); 393 394 { 395 396 // Local registers 397 const Register top_frame_size = R24_tmp4; 398 const Register access_flags = R25_tmp5; 399 const Register state_offset = R26_tmp6; 400 Register mem_stack_limit = R27_tmp7; 401 const Register page_size = R28_tmp8; 402 403 BLOCK_COMMENT("compute_interpreter_state {"); 404 405 // access_flags = method->access_flags(); 406 // TODO: PPC port: assert(4 == methodOopDesc::sz_access_flags(), "unexpected field size"); 407 __ lwa(access_flags, method_(access_flags)); 408 409 // parameter_count = method->constMethod->size_of_parameters(); 410 // TODO: PPC port: assert(2 == ConstMethod::sz_size_of_parameters(), "unexpected field size"); 411 __ ld(max_stack, in_bytes(Method::const_offset()), R19_method); // Max_stack holds constMethod for a while. 412 __ lhz(parameter_count, in_bytes(ConstMethod::size_of_parameters_offset()), max_stack); 413 414 // local_count = method->constMethod()->max_locals(); 415 // TODO: PPC port: assert(2 == ConstMethod::sz_max_locals(), "unexpected field size"); 416 __ lhz(local_count, in_bytes(ConstMethod::size_of_locals_offset()), max_stack); 417 418 // max_stack = method->constMethod()->max_stack(); 419 // TODO: PPC port: assert(2 == ConstMethod::sz_max_stack(), "unexpected field size"); 420 __ lhz(max_stack, in_bytes(ConstMethod::max_stack_offset()), max_stack); 421 422 if (EnableInvokeDynamic) { 423 // Take into account 'extra_stack_entries' needed by method handles (see method.hpp). 424 __ addi(max_stack, max_stack, Method::extra_stack_entries()); 425 } 426 427 // mem_stack_limit = thread->stack_limit(); 428 __ ld(mem_stack_limit, thread_(stack_overflow_limit)); 429 430 // Point locals at the first argument. Method's locals are the 431 // parameters on top of caller's expression stack. 432 433 // tos points past last Java argument 434 __ sldi(R18_locals, parameter_count, Interpreter::logStackElementSize); 435 __ add(R18_locals, R17_tos, R18_locals); 436 437 // R18_locals - i*BytesPerWord points to i-th Java local (i starts at 0) 438 439 // Set is_native, is_synced, is_static - will be used later. 440 __ testbitdi(is_native, R0, access_flags, JVM_ACC_NATIVE_BIT); 441 __ testbitdi(is_synced, R0, access_flags, JVM_ACC_SYNCHRONIZED_BIT); 442 assert(is_synced->is_nonvolatile(), "is_synced must be non-volatile"); 443 __ testbitdi(is_static, R0, access_flags, JVM_ACC_STATIC_BIT); 444 445 // PARENT_IJAVA_FRAME_ABI 446 // 447 // frame_size = 448 // round_to((local_count - parameter_count)*BytesPerWord + 449 // 2*BytesPerWord + 450 // alignment + 451 // frame::interpreter_frame_cinterpreterstate_size_in_bytes() 452 // sizeof(PARENT_IJAVA_FRAME_ABI) 453 // method->is_synchronized() ? sizeof(BasicObjectLock) : 0 + 454 // max_stack*BytesPerWord, 455 // 16) 456 // 457 // Note that this calculation is exactly mirrored by 458 // AbstractInterpreter::layout_activation_impl() [ and 459 // AbstractInterpreter::size_activation() ]. Which is used by 460 // deoptimization so that it can allocate the proper sized 461 // frame. This only happens for interpreted frames so the extra 462 // notes below about max_stack below are not important. The other 463 // thing to note is that for interpreter frames other than the 464 // current activation the size of the stack is the size of the live 465 // portion of the stack at the particular bcp and NOT the maximum 466 // stack that the method might use. 467 // 468 // If we're calling a native method, we replace max_stack (which is 469 // zero) with space for the worst-case signature handler varargs 470 // vector, which is: 471 // 472 // max_stack = max(Argument::n_register_parameters, parameter_count+2); 473 // 474 // We add two slots to the parameter_count, one for the jni 475 // environment and one for a possible native mirror. We allocate 476 // space for at least the number of ABI registers, even though 477 // InterpreterRuntime::slow_signature_handler won't write more than 478 // parameter_count+2 words when it creates the varargs vector at the 479 // top of the stack. The generated slow signature handler will just 480 // load trash into registers beyond the necessary number. We're 481 // still going to cut the stack back by the ABI register parameter 482 // count so as to get SP+16 pointing at the ABI outgoing parameter 483 // area, so we need to allocate at least that much even though we're 484 // going to throw it away. 485 // 486 487 // Adjust max_stack for native methods: 488 Label skip_native_calculate_max_stack; 489 __ bfalse(is_native, skip_native_calculate_max_stack); 490 // if (is_native) { 491 // max_stack = max(Argument::n_register_parameters, parameter_count+2); 492 __ addi(max_stack, parameter_count, 2*Interpreter::stackElementWords); 493 __ cmpwi(CCR0, max_stack, Argument::n_register_parameters); 494 __ bge(CCR0, skip_native_calculate_max_stack); 495 __ li(max_stack, Argument::n_register_parameters); 496 // } 497 __ bind(skip_native_calculate_max_stack); 498 // max_stack is now in bytes 499 __ slwi(max_stack, max_stack, Interpreter::logStackElementSize); 500 501 // Calculate number of non-parameter locals (in slots): 502 Label not_java; 503 __ btrue(is_native, not_java); 504 // if (!is_native) { 505 // local_count = non-parameter local count 506 __ sub(local_count, local_count, parameter_count); 507 // } else { 508 // // nothing to do: method->max_locals() == 0 for native methods 509 // } 510 __ bind(not_java); 511 512 513 // Calculate top_frame_size and parent_frame_resize. 514 { 515 const Register parent_frame_resize = R12_scratch2; 516 517 BLOCK_COMMENT("Compute top_frame_size."); 518 // top_frame_size = TOP_IJAVA_FRAME_ABI 519 // + size of interpreter state 520 __ li(top_frame_size, frame::top_ijava_frame_abi_size 521 + frame::interpreter_frame_cinterpreterstate_size_in_bytes()); 522 // + max_stack 523 __ add(top_frame_size, top_frame_size, max_stack); 524 // + stack slots for a BasicObjectLock for synchronized methods 525 { 526 Label not_synced; 527 __ bfalse(is_synced, not_synced); 528 __ addi(top_frame_size, top_frame_size, frame::interpreter_frame_monitor_size_in_bytes()); 529 __ bind(not_synced); 530 } 531 // align 532 __ round_to(top_frame_size, frame::alignment_in_bytes); 533 534 535 BLOCK_COMMENT("Compute parent_frame_resize."); 536 // parent_frame_resize = R1_SP - R17_tos 537 __ sub(parent_frame_resize, R1_SP, R17_tos); 538 //__ li(parent_frame_resize, 0); 539 // + PARENT_IJAVA_FRAME_ABI 540 // + extra two slots for the no-parameter/no-locals 541 // method result 542 __ addi(parent_frame_resize, parent_frame_resize, 543 frame::parent_ijava_frame_abi_size 544 + 2*Interpreter::stackElementSize); 545 // + (locals_count - params_count) 546 __ sldi(R0, local_count, Interpreter::logStackElementSize); 547 __ add(parent_frame_resize, parent_frame_resize, R0); 548 // align 549 __ round_to(parent_frame_resize, frame::alignment_in_bytes); 550 551 // 552 // Stack layout at this point: 553 // 554 // The new frame F0 hasn't yet been pushed, F1 is still the top frame. 555 // 556 // F0 [TOP_IJAVA_FRAME_ABI] 557 // alignment (optional) 558 // [F0's full operand stack] 559 // [F0's monitors] (optional) 560 // [F0's BytecodeInterpreter object] 561 // F1 [PARENT_IJAVA_FRAME_ABI] 562 // alignment (optional) 563 // [F0's Java result] 564 // [F0's non-arg Java locals] 565 // [F1's outgoing Java arguments] <-- R17_tos 566 // ... 567 // F2 [PARENT_IJAVA_FRAME_ABI] 568 // ... 569 570 571 // Calculate new R14_state 572 // and 573 // test that the new memory stack pointer is above the limit, 574 // throw a StackOverflowError otherwise. 575 __ sub(R11_scratch1/*F1's SP*/, R1_SP, parent_frame_resize); 576 __ addi(R14_state, R11_scratch1/*F1's SP*/, 577 -frame::interpreter_frame_cinterpreterstate_size_in_bytes()); 578 __ sub(R11_scratch1/*F0's SP*/, 579 R11_scratch1/*F1's SP*/, top_frame_size); 580 581 BLOCK_COMMENT("Test for stack overflow:"); 582 __ cmpld(CCR0/*is_stack_overflow*/, R11_scratch1, mem_stack_limit); 583 __ blt(CCR0/*is_stack_overflow*/, stack_overflow_return); 584 585 586 //============================================================================= 587 // Frame_size doesn't overflow the stack. Allocate new frame and 588 // initialize interpreter state. 589 590 // Register state 591 // 592 // R15 - local_count 593 // R16 - parameter_count 594 // R17 - max_stack 595 // 596 // R18 - frame_size 597 // R19 - access_flags 598 // CCR4_is_synced - is_synced 599 // 600 // GR_Lstate - pointer to the uninitialized new BytecodeInterpreter. 601 602 // _last_Java_pc just needs to be close enough that we can identify 603 // the frame as an interpreted frame. It does not need to be the 604 // exact return address from either calling 605 // BytecodeInterpreter::InterpretMethod or the call to a jni native method. 606 // So we can initialize it here with a value of a bundle in this 607 // code fragment. We only do this initialization for java frames 608 // where InterpretMethod needs a a way to get a good pc value to 609 // store in the thread state. For interpreter frames used to call 610 // jni native code we just zero the value in the state and move an 611 // ip as needed in the native entry code. 612 // 613 // const Register last_Java_pc_addr = GR24_SCRATCH; // QQQ 27 614 // const Register last_Java_pc = GR26_SCRATCH; 615 616 // Must reference stack before setting new SP since Windows 617 // will not be able to deliver the exception on a bad SP. 618 // Windows also insists that we bang each page one at a time in order 619 // for the OS to map in the reserved pages. If we bang only 620 // the final page, Windows stops delivering exceptions to our 621 // VectoredExceptionHandler and terminates our program. 622 // Linux only requires a single bang but it's rare to have 623 // to bang more than 1 page so the code is enabled for both OS's. 624 625 // BANG THE STACK 626 // 627 // Nothing to do for PPC, because updating the SP will automatically 628 // bang the page. 629 630 // Up to here we have calculated the delta for the new C-frame and 631 // checked for a stack-overflow. Now we can savely update SP and 632 // resize the C-frame. 633 634 // R14_state has already been calculated. 635 __ push_interpreter_frame(top_frame_size, parent_frame_resize, 636 R25_tmp5, R26_tmp6, R27_tmp7, R28_tmp8); 637 638 } 639 640 // 641 // Stack layout at this point: 642 // 643 // F0 has been been pushed! 644 // 645 // F0 [TOP_IJAVA_FRAME_ABI] <-- R1_SP 646 // alignment (optional) (now it's here, if required) 647 // [F0's full operand stack] 648 // [F0's monitors] (optional) 649 // [F0's BytecodeInterpreter object] 650 // F1 [PARENT_IJAVA_FRAME_ABI] 651 // alignment (optional) (now it's here, if required) 652 // [F0's Java result] 653 // [F0's non-arg Java locals] 654 // [F1's outgoing Java arguments] 655 // ... 656 // F2 [PARENT_IJAVA_FRAME_ABI] 657 // ... 658 // 659 // R14_state points to F0's BytecodeInterpreter object. 660 // 661 662 } 663 664 //============================================================================= 665 // new BytecodeInterpreter-object is save, let's initialize it: 666 BLOCK_COMMENT("New BytecodeInterpreter-object is save."); 667 668 { 669 // Locals 670 const Register bytecode_addr = R24_tmp4; 671 const Register constants = R25_tmp5; 672 const Register tos = R26_tmp6; 673 const Register stack_base = R27_tmp7; 674 const Register local_addr = R28_tmp8; 675 { 676 Label L; 677 __ btrue(is_native, L); 678 // if (!is_native) { 679 // bytecode_addr = constMethod->codes(); 680 __ ld(bytecode_addr, method_(const)); 681 __ addi(bytecode_addr, bytecode_addr, in_bytes(ConstMethod::codes_offset())); 682 // } 683 __ bind(L); 684 } 685 686 __ ld(constants, in_bytes(Method::const_offset()), R19_method); 687 __ ld(constants, in_bytes(ConstMethod::constants_offset()), constants); 688 689 // state->_prev_link = prev_state; 690 __ std(R15_prev_state, state_(_prev_link)); 691 692 // For assertions only. 693 // TODO: not needed anyway because it coincides with `_monitor_base'. remove! 694 // state->_self_link = state; 695 DEBUG_ONLY(__ std(R14_state, state_(_self_link));) 696 697 // state->_thread = thread; 698 __ std(R16_thread, state_(_thread)); 699 700 // state->_method = method; 701 __ std(R19_method, state_(_method)); 702 703 // state->_locals = locals; 704 __ std(R18_locals, state_(_locals)); 705 706 // state->_oop_temp = NULL; 707 __ li(R0, 0); 708 __ std(R0, state_(_oop_temp)); 709 710 // state->_last_Java_fp = *R1_SP // Use *R1_SP as fp 711 __ ld(R0, _abi(callers_sp), R1_SP); 712 __ std(R0, state_(_last_Java_fp)); 713 714 BLOCK_COMMENT("load Stack base:"); 715 { 716 // Stack_base. 717 // if (!method->synchronized()) { 718 // stack_base = state; 719 // } else { 720 // stack_base = (uintptr_t)state - sizeof(BasicObjectLock); 721 // } 722 Label L; 723 __ mr(stack_base, R14_state); 724 __ bfalse(is_synced, L); 725 __ addi(stack_base, stack_base, -frame::interpreter_frame_monitor_size_in_bytes()); 726 __ bind(L); 727 } 728 729 // state->_mdx = NULL; 730 __ li(R0, 0); 731 __ std(R0, state_(_mdx)); 732 733 { 734 // if (method->is_native()) state->_bcp = NULL; 735 // else state->_bcp = bytecode_addr; 736 Label label1, label2; 737 __ bfalse(is_native, label1); 738 __ std(R0, state_(_bcp)); 739 __ b(label2); 740 __ bind(label1); 741 __ std(bytecode_addr, state_(_bcp)); 742 __ bind(label2); 743 } 744 745 746 // state->_result._to_call._callee = NULL; 747 __ std(R0, state_(_result._to_call._callee)); 748 749 // state->_monitor_base = state; 750 __ std(R14_state, state_(_monitor_base)); 751 752 // state->_msg = BytecodeInterpreter::method_entry; 753 __ li(R0, BytecodeInterpreter::method_entry); 754 __ stw(R0, state_(_msg)); 755 756 // state->_last_Java_sp = R1_SP; 757 __ std(R1_SP, state_(_last_Java_sp)); 758 759 // state->_stack_base = stack_base; 760 __ std(stack_base, state_(_stack_base)); 761 762 // tos = stack_base - 1 slot (prepushed); 763 // state->_stack.Tos(tos); 764 __ addi(tos, stack_base, - Interpreter::stackElementSize); 765 __ std(tos, state_(_stack)); 766 767 768 { 769 BLOCK_COMMENT("get last_Java_pc:"); 770 // if (!is_native) state->_last_Java_pc = <some_ip_in_this_code_buffer>; 771 // else state->_last_Java_pc = NULL; (just for neatness) 772 Label label1, label2; 773 __ btrue(is_native, label1); 774 __ get_PC_trash_LR(R0); 775 __ std(R0, state_(_last_Java_pc)); 776 __ b(label2); 777 __ bind(label1); 778 __ li(R0, 0); 779 __ std(R0, state_(_last_Java_pc)); 780 __ bind(label2); 781 } 782 783 784 // stack_limit = tos - max_stack; 785 __ sub(R0, tos, max_stack); 786 // state->_stack_limit = stack_limit; 787 __ std(R0, state_(_stack_limit)); 788 789 790 // cache = method->constants()->cache(); 791 __ ld(R0, ConstantPool::cache_offset_in_bytes(), constants); 792 // state->_constants = method->constants()->cache(); 793 __ std(R0, state_(_constants)); 794 795 796 797 //============================================================================= 798 // synchronized method, allocate and initialize method object lock. 799 // if (!method->is_synchronized()) goto fill_locals_with_0x0s; 800 Label fill_locals_with_0x0s; 801 __ bfalse(is_synced, fill_locals_with_0x0s); 802 803 // pool_holder = method->constants()->pool_holder(); 804 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 805 { 806 Label label1, label2; 807 // lockee = NULL; for java methods, correct value will be inserted in BytecodeInterpretMethod.hpp 808 __ li(R0,0); 809 __ bfalse(is_native, label2); 810 811 __ bfalse(is_static, label1); 812 // if (method->is_static()) lockee = 813 // pool_holder->klass_part()->java_mirror(); 814 __ ld(R11_scratch1/*pool_holder*/, ConstantPool::pool_holder_offset_in_bytes(), constants); 815 __ ld(R0/*lockee*/, mirror_offset, R11_scratch1/*pool_holder*/); 816 __ b(label2); 817 818 __ bind(label1); 819 // else lockee = *(oop*)locals; 820 __ ld(R0/*lockee*/, 0, R18_locals); 821 __ bind(label2); 822 823 // monitor->set_obj(lockee); 824 __ std(R0/*lockee*/, BasicObjectLock::obj_offset_in_bytes(), stack_base); 825 } 826 827 // See if we need to zero the locals 828 __ BIND(fill_locals_with_0x0s); 829 830 831 //============================================================================= 832 // fill locals with 0x0s 833 Label locals_zeroed; 834 __ btrue(is_native, locals_zeroed); 835 836 if (true /* zerolocals */ || ClearInterpreterLocals) { 837 // local_count is already num_locals_slots - num_param_slots 838 __ sldi(R0, parameter_count, Interpreter::logStackElementSize); 839 __ sub(local_addr, R18_locals, R0); 840 __ cmpdi(CCR0, local_count, 0); 841 __ ble(CCR0, locals_zeroed); 842 843 __ mtctr(local_count); 844 //__ ld_const_addr(R0, (address) 0xcafe0000babe); 845 __ li(R0, 0); 846 847 Label zero_slot; 848 __ bind(zero_slot); 849 850 // first local is at local_addr 851 __ std(R0, 0, local_addr); 852 __ addi(local_addr, local_addr, -BytesPerWord); 853 __ bdnz(zero_slot); 854 } 855 856 __ BIND(locals_zeroed); 857 858 } 859 BLOCK_COMMENT("} compute_interpreter_state"); 860 } 861 862 // Generate code to initiate compilation on invocation counter overflow. 863 void CppInterpreterGenerator::generate_counter_overflow(Label& continue_entry) { 864 // Registers alive 865 // R14_state 866 // R16_thread 867 // 868 // Registers updated 869 // R14_state 870 // R3_ARG1 (=R3_RET) 871 // R4_ARG2 872 873 // After entering the vm we remove the activation and retry the 874 // entry point in case the compilation is complete. 875 876 // InterpreterRuntime::frequency_counter_overflow takes one argument 877 // that indicates if the counter overflow occurs at a backwards 878 // branch (NULL bcp). We pass zero. The call returns the address 879 // of the verified entry point for the method or NULL if the 880 // compilation did not complete (either went background or bailed 881 // out). 882 __ li(R4_ARG2, 0); 883 884 // Pass false to call_VM so it doesn't check for pending exceptions, 885 // since at this point in the method invocation the exception 886 // handler would try to exit the monitor of synchronized methods 887 // which haven't been entered yet. 888 // 889 // Returns verified_entry_point or NULL, we don't care which. 890 // 891 // Do not use the variant `frequency_counter_overflow' that returns 892 // a structure, because this will change the argument list by a 893 // hidden parameter (gcc 4.1). 894 895 __ call_VM(noreg, 896 CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), 897 R4_ARG2, 898 false); 899 // Returns verified_entry_point or NULL, we don't care which as we ignore it 900 // and run interpreted. 901 902 // Reload method, it may have moved. 903 __ ld(R19_method, state_(_method)); 904 905 // We jump now to the label "continue_after_compile". 906 __ b(continue_entry); 907 } 908 909 // Increment invocation count and check for overflow. 910 // 911 // R19_method must contain Method* of method to profile. 912 void CppInterpreterGenerator::generate_counter_incr(Label& overflow) { 913 Label done; 914 const Register Rcounters = R12_scratch2; 915 const Register iv_be_count = R11_scratch1; 916 const Register invocation_limit = R12_scratch2; 917 const Register invocation_limit_addr = invocation_limit; 918 919 // Load and ev. allocate MethodCounters object. 920 __ get_method_counters(R19_method, Rcounters, done); 921 922 // Update standard invocation counters. 923 __ increment_invocation_counter(Rcounters, iv_be_count, R0); 924 925 // Compare against limit. 926 BLOCK_COMMENT("Compare counter against limit:"); 927 assert(4 == sizeof(InvocationCounter::InterpreterInvocationLimit), 928 "must be 4 bytes"); 929 __ load_const(invocation_limit_addr, (address)&InvocationCounter::InterpreterInvocationLimit); 930 __ lwa(invocation_limit, 0, invocation_limit_addr); 931 __ cmpw(CCR0, iv_be_count, invocation_limit); 932 __ bge(CCR0, overflow); 933 __ bind(done); 934 } 935 936 // 937 // Call a JNI method. 938 // 939 // Interpreter stub for calling a native method. (C++ interpreter) 940 // This sets up a somewhat different looking stack for calling the native method 941 // than the typical interpreter frame setup. 942 // 943 address CppInterpreterGenerator::generate_native_entry(void) { 944 if (native_entry != NULL) return native_entry; 945 address entry = __ pc(); 946 947 // Read 948 // R16_thread 949 // R15_prev_state - address of caller's BytecodeInterpreter, if this snippet 950 // gets called by the frame manager. 951 // R19_method - callee's Method 952 // R17_tos - address of caller's tos 953 // R1_SP - caller's stack pointer 954 // R21_sender_SP - initial caller sp 955 // 956 // Update 957 // R14_state - address of caller's BytecodeInterpreter 958 // R3_RET - integer result, if any. 959 // F1_RET - float result, if any. 960 // 961 // 962 // Stack layout at this point: 963 // 964 // 0 [TOP_IJAVA_FRAME_ABI] <-- R1_SP 965 // alignment (optional) 966 // [outgoing Java arguments] <-- R17_tos 967 // ... 968 // PARENT [PARENT_IJAVA_FRAME_ABI] 969 // ... 970 // 971 972 const bool inc_counter = UseCompiler || CountCompiledCalls; 973 974 const Register signature_handler_fd = R21_tmp1; 975 const Register pending_exception = R22_tmp2; 976 const Register result_handler_addr = R23_tmp3; 977 const Register native_method_fd = R24_tmp4; 978 const Register access_flags = R25_tmp5; 979 const Register active_handles = R26_tmp6; 980 const Register sync_state = R27_tmp7; 981 const Register sync_state_addr = sync_state; // Address is dead after use. 982 const Register suspend_flags = R24_tmp4; 983 984 const Register return_pc = R28_tmp8; // Register will be locked for some time. 985 986 const ConditionRegister is_synced = CCR4_is_synced; // Live-on-exit from compute_interpreter_state. 987 988 989 // R1_SP still points to caller's SP at this point. 990 991 // Save initial_caller_sp to caller's abi. The caller frame must be 992 // resized before returning to get rid of the c2i arguments (if 993 // any). 994 // Override the saved SP with the senderSP so we can pop c2i 995 // arguments (if any) off when we return 996 __ std(R21_sender_SP, _top_ijava_frame_abi(initial_caller_sp), R1_SP); 997 998 // Save LR to caller's frame. We don't use _abi(lr) here, because it is not safe. 999 __ mflr(return_pc); 1000 __ std(return_pc, _top_ijava_frame_abi(frame_manager_lr), R1_SP); 1001 1002 assert(return_pc->is_nonvolatile(), "return_pc must be a non-volatile register"); 1003 1004 __ verify_method_ptr(R19_method); 1005 1006 //============================================================================= 1007 1008 // If this snippet gets called by the frame manager (at label 1009 // `call_special'), then R15_prev_state is valid. If this snippet 1010 // is not called by the frame manager, but e.g. by the call stub or 1011 // by compiled code, then R15_prev_state is invalid. 1012 { 1013 // Set R15_prev_state to 0 if we don't return to the frame 1014 // manager; we will return to the call_stub or to compiled code 1015 // instead. If R15_prev_state is 0 there will be only one 1016 // interpreter frame (we will set this up later) in this C frame! 1017 // So we must take care about retrieving prev_state_(_prev_link) 1018 // and restoring R1_SP when popping that interpreter. 1019 Label prev_state_is_valid; 1020 1021 __ load_const(R11_scratch1/*frame_manager_returnpc_addr*/, (address)&frame_manager_specialized_return); 1022 __ ld(R12_scratch2/*frame_manager_returnpc*/, 0, R11_scratch1/*frame_manager_returnpc_addr*/); 1023 __ cmpd(CCR0, return_pc, R12_scratch2/*frame_manager_returnpc*/); 1024 __ beq(CCR0, prev_state_is_valid); 1025 1026 __ li(R15_prev_state, 0); 1027 1028 __ BIND(prev_state_is_valid); 1029 } 1030 1031 //============================================================================= 1032 // Allocate new frame and initialize interpreter state. 1033 1034 Label exception_return; 1035 Label exception_return_sync_check; 1036 Label stack_overflow_return; 1037 1038 // Generate new interpreter state and jump to stack_overflow_return in case of 1039 // a stack overflow. 1040 generate_compute_interpreter_state(stack_overflow_return); 1041 1042 //============================================================================= 1043 // Increment invocation counter. On overflow, entry to JNI method 1044 // will be compiled. 1045 Label invocation_counter_overflow; 1046 if (inc_counter) { 1047 generate_counter_incr(invocation_counter_overflow); 1048 } 1049 1050 Label continue_after_compile; 1051 __ BIND(continue_after_compile); 1052 1053 // access_flags = method->access_flags(); 1054 // Load access flags. 1055 assert(access_flags->is_nonvolatile(), 1056 "access_flags must be in a non-volatile register"); 1057 // Type check. 1058 // TODO: PPC port: assert(4 == methodOopDesc::sz_access_flags(), "unexpected field size"); 1059 __ lwz(access_flags, method_(access_flags)); 1060 1061 // We don't want to reload R19_method and access_flags after calls 1062 // to some helper functions. 1063 assert(R19_method->is_nonvolatile(), "R19_method must be a non-volatile register"); 1064 1065 // Check for synchronized methods. Must happen AFTER invocation counter 1066 // check, so method is not locked if counter overflows. 1067 1068 { 1069 Label method_is_not_synced; 1070 // Is_synced is still alive. 1071 assert(is_synced->is_nonvolatile(), "is_synced must be non-volatile"); 1072 __ bfalse(is_synced, method_is_not_synced); 1073 1074 lock_method(); 1075 // Reload method, it may have moved. 1076 __ ld(R19_method, state_(_method)); 1077 1078 __ BIND(method_is_not_synced); 1079 } 1080 1081 // jvmti/jvmpi support 1082 __ notify_method_entry(); 1083 1084 // Reload method, it may have moved. 1085 __ ld(R19_method, state_(_method)); 1086 1087 //============================================================================= 1088 // Get and call the signature handler 1089 1090 __ ld(signature_handler_fd, method_(signature_handler)); 1091 Label call_signature_handler; 1092 1093 __ cmpdi(CCR0, signature_handler_fd, 0); 1094 __ bne(CCR0, call_signature_handler); 1095 1096 // Method has never been called. Either generate a specialized 1097 // handler or point to the slow one. 1098 // 1099 // Pass parameter 'false' to avoid exception check in call_VM. 1100 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), R19_method, false); 1101 1102 // Check for an exception while looking up the target method. If we 1103 // incurred one, bail. 1104 __ ld(pending_exception, thread_(pending_exception)); 1105 __ cmpdi(CCR0, pending_exception, 0); 1106 __ bne(CCR0, exception_return_sync_check); // has pending exception 1107 1108 // reload method 1109 __ ld(R19_method, state_(_method)); 1110 1111 // Reload signature handler, it may have been created/assigned in the meanwhile 1112 __ ld(signature_handler_fd, method_(signature_handler)); 1113 1114 __ BIND(call_signature_handler); 1115 1116 // Before we call the signature handler we push a new frame to 1117 // protect the interpreter frame volatile registers when we return 1118 // from jni but before we can get back to Java. 1119 1120 // First set the frame anchor while the SP/FP registers are 1121 // convenient and the slow signature handler can use this same frame 1122 // anchor. 1123 1124 // We have a TOP_IJAVA_FRAME here, which belongs to us. 1125 __ set_top_ijava_frame_at_SP_as_last_Java_frame(R1_SP, R12_scratch2/*tmp*/); 1126 1127 // Now the interpreter frame (and its call chain) have been 1128 // invalidated and flushed. We are now protected against eager 1129 // being enabled in native code. Even if it goes eager the 1130 // registers will be reloaded as clean and we will invalidate after 1131 // the call so no spurious flush should be possible. 1132 1133 // Call signature handler and pass locals address. 1134 // 1135 // Our signature handlers copy required arguments to the C stack 1136 // (outgoing C args), R3_ARG1 to R10_ARG8, and F1_ARG1 to 1137 // F13_ARG13. 1138 __ mr(R3_ARG1, R18_locals); 1139 __ ld(signature_handler_fd, 0, signature_handler_fd); 1140 __ call_stub(signature_handler_fd); 1141 // reload method 1142 __ ld(R19_method, state_(_method)); 1143 1144 // Remove the register parameter varargs slots we allocated in 1145 // compute_interpreter_state. SP+16 ends up pointing to the ABI 1146 // outgoing argument area. 1147 // 1148 // Not needed on PPC64. 1149 //__ add(SP, SP, Argument::n_register_parameters*BytesPerWord); 1150 1151 assert(result_handler_addr->is_nonvolatile(), "result_handler_addr must be in a non-volatile register"); 1152 // Save across call to native method. 1153 __ mr(result_handler_addr, R3_RET); 1154 1155 // Set up fixed parameters and call the native method. 1156 // If the method is static, get mirror into R4_ARG2. 1157 1158 { 1159 Label method_is_not_static; 1160 // access_flags is non-volatile and still, no need to restore it 1161 1162 // restore access flags 1163 __ testbitdi(CCR0, R0, access_flags, JVM_ACC_STATIC_BIT); 1164 __ bfalse(CCR0, method_is_not_static); 1165 1166 // constants = method->constants(); 1167 __ ld(R11_scratch1, in_bytes(Method::const_offset()), R19_method); 1168 __ ld(R11_scratch1/*constants*/, in_bytes(ConstMethod::constants_offset()), R11_scratch1); 1169 // pool_holder = method->constants()->pool_holder(); 1170 __ ld(R11_scratch1/*pool_holder*/, ConstantPool::pool_holder_offset_in_bytes(), 1171 R11_scratch1/*constants*/); 1172 1173 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 1174 1175 // mirror = pool_holder->klass_part()->java_mirror(); 1176 __ ld(R0/*mirror*/, mirror_offset, R11_scratch1/*pool_holder*/); 1177 // state->_native_mirror = mirror; 1178 __ std(R0/*mirror*/, state_(_oop_temp)); 1179 // R4_ARG2 = &state->_oop_temp; 1180 __ addir(R4_ARG2, state_(_oop_temp)); 1181 1182 __ BIND(method_is_not_static); 1183 } 1184 1185 // At this point, arguments have been copied off the stack into 1186 // their JNI positions. Oops are boxed in-place on the stack, with 1187 // handles copied to arguments. The result handler address is in a 1188 // register. 1189 1190 // pass JNIEnv address as first parameter 1191 __ addir(R3_ARG1, thread_(jni_environment)); 1192 1193 // Load the native_method entry before we change the thread state. 1194 __ ld(native_method_fd, method_(native_function)); 1195 1196 //============================================================================= 1197 // Transition from _thread_in_Java to _thread_in_native. As soon as 1198 // we make this change the safepoint code needs to be certain that 1199 // the last Java frame we established is good. The pc in that frame 1200 // just needs to be near here not an actual return address. 1201 1202 // We use release_store_fence to update values like the thread state, where 1203 // we don't want the current thread to continue until all our prior memory 1204 // accesses (including the new thread state) are visible to other threads. 1205 __ li(R0, _thread_in_native); 1206 __ release(); 1207 1208 // TODO: PPC port: assert(4 == JavaThread::sz_thread_state(), "unexpected field size"); 1209 __ stw(R0, thread_(thread_state)); 1210 1211 if (UseMembar) { 1212 __ fence(); 1213 } 1214 1215 //============================================================================= 1216 // Call the native method. Argument registers must not have been 1217 // overwritten since "__ call_stub(signature_handler);" (except for 1218 // ARG1 and ARG2 for static methods) 1219 __ call_c(native_method_fd); 1220 1221 __ std(R3_RET, state_(_native_lresult)); 1222 __ stfd(F1_RET, state_(_native_fresult)); 1223 1224 // The frame_manager_lr field, which we use for setting the last 1225 // java frame, gets overwritten by the signature handler. Restore 1226 // it now. 1227 __ get_PC_trash_LR(R11_scratch1); 1228 __ std(R11_scratch1, _top_ijava_frame_abi(frame_manager_lr), R1_SP); 1229 1230 // Because of GC R19_method may no longer be valid. 1231 1232 // Block, if necessary, before resuming in _thread_in_Java state. 1233 // In order for GC to work, don't clear the last_Java_sp until after 1234 // blocking. 1235 1236 1237 1238 //============================================================================= 1239 // Switch thread to "native transition" state before reading the 1240 // synchronization state. This additional state is necessary 1241 // because reading and testing the synchronization state is not 1242 // atomic w.r.t. GC, as this scenario demonstrates: Java thread A, 1243 // in _thread_in_native state, loads _not_synchronized and is 1244 // preempted. VM thread changes sync state to synchronizing and 1245 // suspends threads for GC. Thread A is resumed to finish this 1246 // native method, but doesn't block here since it didn't see any 1247 // synchronization in progress, and escapes. 1248 1249 // We use release_store_fence to update values like the thread state, where 1250 // we don't want the current thread to continue until all our prior memory 1251 // accesses (including the new thread state) are visible to other threads. 1252 __ li(R0/*thread_state*/, _thread_in_native_trans); 1253 __ release(); 1254 __ stw(R0/*thread_state*/, thread_(thread_state)); 1255 if (UseMembar) { 1256 __ fence(); 1257 } 1258 // Write serialization page so that the VM thread can do a pseudo remote 1259 // membar. We use the current thread pointer to calculate a thread 1260 // specific offset to write to within the page. This minimizes bus 1261 // traffic due to cache line collision. 1262 else { 1263 __ serialize_memory(R16_thread, R11_scratch1, R12_scratch2); 1264 } 1265 1266 // Now before we return to java we must look for a current safepoint 1267 // (a new safepoint can not start since we entered native_trans). 1268 // We must check here because a current safepoint could be modifying 1269 // the callers registers right this moment. 1270 1271 // Acquire isn't strictly necessary here because of the fence, but 1272 // sync_state is declared to be volatile, so we do it anyway. 1273 __ load_const(sync_state_addr, SafepointSynchronize::address_of_state()); 1274 1275 // TODO: PPC port: assert(4 == SafepointSynchronize::sz_state(), "unexpected field size"); 1276 __ lwz(sync_state, 0, sync_state_addr); 1277 1278 // TODO: PPC port: assert(4 == Thread::sz_suspend_flags(), "unexpected field size"); 1279 __ lwz(suspend_flags, thread_(suspend_flags)); 1280 1281 __ acquire(); 1282 1283 Label sync_check_done; 1284 Label do_safepoint; 1285 // No synchronization in progress nor yet synchronized 1286 __ cmpwi(CCR0, sync_state, SafepointSynchronize::_not_synchronized); 1287 // not suspended 1288 __ cmpwi(CCR1, suspend_flags, 0); 1289 1290 __ bne(CCR0, do_safepoint); 1291 __ beq(CCR1, sync_check_done); 1292 __ bind(do_safepoint); 1293 // Block. We do the call directly and leave the current 1294 // last_Java_frame setup undisturbed. We must save any possible 1295 // native result acrosss the call. No oop is present 1296 1297 __ mr(R3_ARG1, R16_thread); 1298 __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, JavaThread::check_special_condition_for_native_trans), 1299 relocInfo::none); 1300 __ bind(sync_check_done); 1301 1302 //============================================================================= 1303 // <<<<<< Back in Interpreter Frame >>>>> 1304 1305 // We are in thread_in_native_trans here and back in the normal 1306 // interpreter frame. We don't have to do anything special about 1307 // safepoints and we can switch to Java mode anytime we are ready. 1308 1309 // Note: frame::interpreter_frame_result has a dependency on how the 1310 // method result is saved across the call to post_method_exit. For 1311 // native methods it assumes that the non-FPU/non-void result is 1312 // saved in _native_lresult and a FPU result in _native_fresult. If 1313 // this changes then the interpreter_frame_result implementation 1314 // will need to be updated too. 1315 1316 // On PPC64, we have stored the result directly after the native call. 1317 1318 //============================================================================= 1319 // back in Java 1320 1321 // We use release_store_fence to update values like the thread state, where 1322 // we don't want the current thread to continue until all our prior memory 1323 // accesses (including the new thread state) are visible to other threads. 1324 __ li(R0/*thread_state*/, _thread_in_Java); 1325 __ release(); 1326 __ stw(R0/*thread_state*/, thread_(thread_state)); 1327 if (UseMembar) { 1328 __ fence(); 1329 } 1330 1331 __ reset_last_Java_frame(); 1332 1333 // Reload GR27_method, call killed it. We can't look at 1334 // state->_method until we're back in java state because in java 1335 // state gc can't happen until we get to a safepoint. 1336 // 1337 // We've set thread_state to _thread_in_Java already, so restoring 1338 // R19_method from R14_state works; R19_method is invalid, because 1339 // GC may have happened. 1340 __ ld(R19_method, state_(_method)); // reload method, may have moved 1341 1342 // jvmdi/jvmpi support. Whether we've got an exception pending or 1343 // not, and whether unlocking throws an exception or not, we notify 1344 // on native method exit. If we do have an exception, we'll end up 1345 // in the caller's context to handle it, so if we don't do the 1346 // notify here, we'll drop it on the floor. 1347 1348 __ notify_method_exit(true/*native method*/, 1349 ilgl /*illegal state (not used for native methods)*/); 1350 1351 1352 1353 //============================================================================= 1354 // Handle exceptions 1355 1356 // See if we must unlock. 1357 // 1358 { 1359 Label method_is_not_synced; 1360 // is_synced is still alive 1361 assert(is_synced->is_nonvolatile(), "is_synced must be non-volatile"); 1362 __ bfalse(is_synced, method_is_not_synced); 1363 1364 unlock_method(); 1365 1366 __ bind(method_is_not_synced); 1367 } 1368 1369 // Reset active handles after returning from native. 1370 // thread->active_handles()->clear(); 1371 __ ld(active_handles, thread_(active_handles)); 1372 // JNIHandleBlock::_top is an int. 1373 // TODO: PPC port: assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size"); 1374 __ li(R0, 0); 1375 __ stw(R0, JNIHandleBlock::top_offset_in_bytes(), active_handles); 1376 1377 Label no_pending_exception_from_native_method; 1378 __ ld(R0/*pending_exception*/, thread_(pending_exception)); 1379 __ cmpdi(CCR0, R0/*pending_exception*/, 0); 1380 __ beq(CCR0, no_pending_exception_from_native_method); 1381 1382 1383 //----------------------------------------------------------------------------- 1384 // An exception is pending. We call into the runtime only if the 1385 // caller was not interpreted. If it was interpreted the 1386 // interpreter will do the correct thing. If it isn't interpreted 1387 // (call stub/compiled code) we will change our return and continue. 1388 __ BIND(exception_return); 1389 1390 Label return_to_initial_caller_with_pending_exception; 1391 __ cmpdi(CCR0, R15_prev_state, 0); 1392 __ beq(CCR0, return_to_initial_caller_with_pending_exception); 1393 1394 // We are returning to an interpreter activation, just pop the state, 1395 // pop our frame, leave the exception pending, and return. 1396 __ pop_interpreter_state(/*prev_state_may_be_0=*/false); 1397 __ pop_interpreter_frame(R11_scratch1, R12_scratch2, R21_tmp1 /* set to return pc */, R22_tmp2); 1398 __ mtlr(R21_tmp1); 1399 __ blr(); 1400 1401 __ BIND(exception_return_sync_check); 1402 1403 assert(is_synced->is_nonvolatile(), "is_synced must be non-volatile"); 1404 __ bfalse(is_synced, exception_return); 1405 unlock_method(); 1406 __ b(exception_return); 1407 1408 1409 __ BIND(return_to_initial_caller_with_pending_exception); 1410 // We are returning to a c2i-adapter / call-stub, get the address of the 1411 // exception handler, pop the frame and return to the handler. 1412 1413 // First, pop to caller's frame. 1414 __ pop_interpreter_frame(R11_scratch1, R12_scratch2, R21_tmp1 /* set to return pc */, R22_tmp2); 1415 1416 __ push_frame_abi112(0, R11_scratch1); 1417 // Get the address of the exception handler. 1418 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), 1419 R16_thread, 1420 R21_tmp1 /* return pc */); 1421 __ pop_frame(); 1422 1423 // Load the PC of the the exception handler into LR. 1424 __ mtlr(R3_RET); 1425 1426 // Load exception into R3_ARG1 and clear pending exception in thread. 1427 __ ld(R3_ARG1/*exception*/, thread_(pending_exception)); 1428 __ li(R4_ARG2, 0); 1429 __ std(R4_ARG2, thread_(pending_exception)); 1430 1431 // Load the original return pc into R4_ARG2. 1432 __ mr(R4_ARG2/*issuing_pc*/, R21_tmp1); 1433 1434 // Resize frame to get rid of a potential extension. 1435 __ resize_frame_to_initial_caller(R11_scratch1, R12_scratch2); 1436 1437 // Return to exception handler. 1438 __ blr(); 1439 1440 1441 //----------------------------------------------------------------------------- 1442 // No exception pending. 1443 __ BIND(no_pending_exception_from_native_method); 1444 1445 // Move native method result back into proper registers and return. 1446 // Invoke result handler (may unbox/promote). 1447 __ ld(R3_RET, state_(_native_lresult)); 1448 __ lfd(F1_RET, state_(_native_fresult)); 1449 __ call_stub(result_handler_addr); 1450 1451 // We have created a new BytecodeInterpreter object, now we must destroy it. 1452 // 1453 // Restore previous R14_state and caller's SP. R15_prev_state may 1454 // be 0 here, because our caller may be the call_stub or compiled 1455 // code. 1456 __ pop_interpreter_state(/*prev_state_may_be_0=*/true); 1457 __ pop_interpreter_frame(R11_scratch1, R12_scratch2, R21_tmp1 /* set to return pc */, R22_tmp2); 1458 // Resize frame to get rid of a potential extension. 1459 __ resize_frame_to_initial_caller(R11_scratch1, R12_scratch2); 1460 1461 // Must use the return pc which was loaded from the caller's frame 1462 // as the VM uses return-pc-patching for deoptimization. 1463 __ mtlr(R21_tmp1); 1464 __ blr(); 1465 1466 1467 1468 //============================================================================= 1469 // We encountered an exception while computing the interpreter 1470 // state, so R14_state isn't valid. Act as if we just returned from 1471 // the callee method with a pending exception. 1472 __ BIND(stack_overflow_return); 1473 1474 // 1475 // Register state: 1476 // R14_state invalid; trashed by compute_interpreter_state 1477 // R15_prev_state valid, but may be 0 1478 // 1479 // R1_SP valid, points to caller's SP; wasn't yet updated by 1480 // compute_interpreter_state 1481 // 1482 1483 // Create exception oop and make it pending. 1484 1485 // Throw the exception via RuntimeStub "throw_StackOverflowError_entry". 1486 // 1487 // Previously, we called C-Code directly. As a consequence, a 1488 // possible GC tried to process the argument oops of the top frame 1489 // (see RegisterMap::clear, which sets the corresponding flag to 1490 // true). This lead to crashes because: 1491 // 1. The top register map did not contain locations for the argument registers 1492 // 2. The arguments are dead anyway, could be already overwritten in the worst case 1493 // Solution: Call via special runtime stub that pushes it's own 1494 // frame. This runtime stub has the flag "CodeBlob::caller_must_gc_arguments()" 1495 // set to "false", what prevents the dead arguments getting GC'd. 1496 // 1497 // 2 cases exist: 1498 // 1. We were called by the c2i adapter / call stub 1499 // 2. We were called by the frame manager 1500 // 1501 // Both cases are handled by this code: 1502 // 1. - initial_caller_sp was saved in both cases on entry, so it's safe to load it back even if it was not changed. 1503 // - control flow will be: 1504 // throw_stackoverflow_stub->VM->throw_stackoverflow_stub->forward_excep->excp_blob of caller method 1505 // 2. - control flow will be: 1506 // throw_stackoverflow_stub->VM->throw_stackoverflow_stub->forward_excep->rethrow_excp_entry of frame manager->resume_method 1507 // Since we restored the caller SP above, the rethrow_excp_entry can restore the original interpreter state 1508 // registers using the stack and resume the calling method with a pending excp. 1509 1510 // Pop any c2i extension from the stack, restore LR just to be sure 1511 __ ld(R0, _top_ijava_frame_abi(frame_manager_lr), R1_SP); 1512 __ mtlr(R0); 1513 // Resize frame to get rid of a potential extension. 1514 __ resize_frame_to_initial_caller(R11_scratch1, R12_scratch2); 1515 1516 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "generated in wrong order"); 1517 // Load target address of the runtime stub. 1518 __ load_const(R12_scratch2, (StubRoutines::throw_StackOverflowError_entry())); 1519 __ mtctr(R12_scratch2); 1520 __ bctr(); 1521 1522 1523 //============================================================================= 1524 // Counter overflow. 1525 1526 if (inc_counter) { 1527 // Handle invocation counter overflow 1528 __ bind(invocation_counter_overflow); 1529 1530 generate_counter_overflow(continue_after_compile); 1531 } 1532 1533 native_entry = entry; 1534 return entry; 1535 } 1536 1537 bool AbstractInterpreter::can_be_compiled(methodHandle m) { 1538 // No special entry points that preclude compilation. 1539 return true; 1540 } 1541 1542 // Unlock the current method. 1543 // 1544 void CppInterpreterGenerator::unlock_method(void) { 1545 // Find preallocated monitor and unlock method. Method monitor is 1546 // the first one. 1547 1548 // Registers alive 1549 // R14_state 1550 // 1551 // Registers updated 1552 // volatiles 1553 // 1554 const Register monitor = R4_ARG2; 1555 1556 // Pass address of initial monitor we allocated. 1557 // 1558 // First monitor. 1559 __ addi(monitor, R14_state, -frame::interpreter_frame_monitor_size_in_bytes()); 1560 1561 // Unlock method 1562 __ unlock_object(monitor); 1563 } 1564 1565 // Lock the current method. 1566 // 1567 void CppInterpreterGenerator::lock_method(void) { 1568 // Find preallocated monitor and lock method. Method monitor is the 1569 // first one. 1570 1571 // 1572 // Registers alive 1573 // R14_state 1574 // 1575 // Registers updated 1576 // volatiles 1577 // 1578 1579 const Register monitor = R4_ARG2; 1580 const Register object = R5_ARG3; 1581 1582 // Pass address of initial monitor we allocated. 1583 __ addi(monitor, R14_state, -frame::interpreter_frame_monitor_size_in_bytes()); 1584 1585 // Pass object address. 1586 __ ld(object, BasicObjectLock::obj_offset_in_bytes(), monitor); 1587 1588 // Lock method. 1589 __ lock_object(monitor, object); 1590 } 1591 1592 // Generate code for handling resuming a deopted method. 1593 void CppInterpreterGenerator::generate_deopt_handling(Register result_index) { 1594 1595 //============================================================================= 1596 // Returning from a compiled method into a deopted method. The 1597 // bytecode at the bcp has completed. The result of the bytecode is 1598 // in the native abi (the tosca for the template based 1599 // interpreter). Any stack space that was used by the bytecode that 1600 // has completed has been removed (e.g. parameters for an invoke) so 1601 // all that we have to do is place any pending result on the 1602 // expression stack and resume execution on the next bytecode. 1603 1604 Label return_from_deopt_common; 1605 1606 // R3_RET and F1_RET are live here! Load the array index of the 1607 // required result stub address and continue at return_from_deopt_common. 1608 1609 // Deopt needs to jump to here to enter the interpreter (return a result). 1610 deopt_frame_manager_return_atos = __ pc(); 1611 __ li(result_index, AbstractInterpreter::BasicType_as_index(T_OBJECT)); 1612 __ b(return_from_deopt_common); 1613 1614 deopt_frame_manager_return_btos = __ pc(); 1615 __ li(result_index, AbstractInterpreter::BasicType_as_index(T_BOOLEAN)); 1616 __ b(return_from_deopt_common); 1617 1618 deopt_frame_manager_return_itos = __ pc(); 1619 __ li(result_index, AbstractInterpreter::BasicType_as_index(T_INT)); 1620 __ b(return_from_deopt_common); 1621 1622 deopt_frame_manager_return_ltos = __ pc(); 1623 __ li(result_index, AbstractInterpreter::BasicType_as_index(T_LONG)); 1624 __ b(return_from_deopt_common); 1625 1626 deopt_frame_manager_return_ftos = __ pc(); 1627 __ li(result_index, AbstractInterpreter::BasicType_as_index(T_FLOAT)); 1628 __ b(return_from_deopt_common); 1629 1630 deopt_frame_manager_return_dtos = __ pc(); 1631 __ li(result_index, AbstractInterpreter::BasicType_as_index(T_DOUBLE)); 1632 __ b(return_from_deopt_common); 1633 1634 deopt_frame_manager_return_vtos = __ pc(); 1635 __ li(result_index, AbstractInterpreter::BasicType_as_index(T_VOID)); 1636 // Last one, fall-through to return_from_deopt_common. 1637 1638 // Deopt return common. An index is present that lets us move any 1639 // possible result being return to the interpreter's stack. 1640 // 1641 __ BIND(return_from_deopt_common); 1642 1643 } 1644 1645 // Generate the code to handle a more_monitors message from the c++ interpreter. 1646 void CppInterpreterGenerator::generate_more_monitors() { 1647 1648 // 1649 // Registers alive 1650 // R16_thread - JavaThread* 1651 // R15_prev_state - previous BytecodeInterpreter or 0 1652 // R14_state - BytecodeInterpreter* address of receiver's interpreter state 1653 // R1_SP - old stack pointer 1654 // 1655 // Registers updated 1656 // R1_SP - new stack pointer 1657 // 1658 1659 // Very-local scratch registers. 1660 const Register old_tos = R21_tmp1; 1661 const Register new_tos = R22_tmp2; 1662 const Register stack_base = R23_tmp3; 1663 const Register stack_limit = R24_tmp4; 1664 const Register slot = R25_tmp5; 1665 const Register n_slots = R25_tmp5; 1666 1667 // Interpreter state fields. 1668 const Register msg = R24_tmp4; 1669 1670 // Load up relevant interpreter state. 1671 1672 __ ld(stack_base, state_(_stack_base)); // Old stack_base 1673 __ ld(old_tos, state_(_stack)); // Old tos 1674 __ ld(stack_limit, state_(_stack_limit)); // Old stack_limit 1675 1676 // extracted monitor_size 1677 int monitor_size = frame::interpreter_frame_monitor_size_in_bytes(); 1678 assert(Assembler::is_aligned((unsigned int)monitor_size, 1679 (unsigned int)frame::alignment_in_bytes), 1680 "size of a monitor must respect alignment of SP"); 1681 1682 // Save and restore top LR 1683 __ ld(R12_scratch2, _top_ijava_frame_abi(frame_manager_lr), R1_SP); 1684 __ resize_frame(-monitor_size, R11_scratch1);// Allocate space for new monitor 1685 __ std(R12_scratch2, _top_ijava_frame_abi(frame_manager_lr), R1_SP); 1686 // Initial_caller_sp is used as unextended_sp for non initial callers. 1687 __ std(R1_SP, _top_ijava_frame_abi(initial_caller_sp), R1_SP); 1688 __ addi(stack_base, stack_base, -monitor_size); // New stack_base 1689 __ addi(new_tos, old_tos, -monitor_size); // New tos 1690 __ addi(stack_limit, stack_limit, -monitor_size); // New stack_limit 1691 1692 __ std(R1_SP, state_(_last_Java_sp)); // Update frame_bottom 1693 1694 __ std(stack_base, state_(_stack_base)); // Update stack_base 1695 __ std(new_tos, state_(_stack)); // Update tos 1696 __ std(stack_limit, state_(_stack_limit)); // Update stack_limit 1697 1698 __ li(msg, BytecodeInterpreter::got_monitors); // Tell interpreter we allocated the lock 1699 __ stw(msg, state_(_msg)); 1700 1701 // Shuffle expression stack down. Recall that stack_base points 1702 // just above the new expression stack bottom. Old_tos and new_tos 1703 // are used to scan thru the old and new expression stacks. 1704 1705 Label copy_slot, copy_slot_finished; 1706 __ sub(n_slots, stack_base, new_tos); 1707 __ srdi_(n_slots, n_slots, LogBytesPerWord); // compute number of slots to copy 1708 assert(LogBytesPerWord == 3, "conflicts assembler instructions"); 1709 __ beq(CCR0, copy_slot_finished); // nothing to copy 1710 1711 __ mtctr(n_slots); 1712 1713 // loop 1714 __ bind(copy_slot); 1715 __ ldu(slot, BytesPerWord, old_tos); // slot = *++old_tos; 1716 __ stdu(slot, BytesPerWord, new_tos); // *++new_tos = slot; 1717 __ bdnz(copy_slot); 1718 1719 __ bind(copy_slot_finished); 1720 1721 // Restart interpreter 1722 __ li(R0, 0); 1723 __ std(R0, BasicObjectLock::obj_offset_in_bytes(), stack_base); // Mark lock as unused 1724 } 1725 1726 address CppInterpreterGenerator::generate_normal_entry(void) { 1727 if (interpreter_frame_manager != NULL) return interpreter_frame_manager; 1728 1729 address entry = __ pc(); 1730 1731 address return_from_native_pc = (address) NULL; 1732 1733 // Initial entry to frame manager (from call_stub or c2i_adapter) 1734 1735 // 1736 // Registers alive 1737 // R16_thread - JavaThread* 1738 // R19_method - callee's Method (method to be invoked) 1739 // R17_tos - address of sender tos (prepushed) 1740 // R1_SP - SP prepared by call stub such that caller's outgoing args are near top 1741 // LR - return address to caller (call_stub or c2i_adapter) 1742 // R21_sender_SP - initial caller sp 1743 // 1744 // Registers updated 1745 // R15_prev_state - 0 1746 // 1747 // Stack layout at this point: 1748 // 1749 // 0 [TOP_IJAVA_FRAME_ABI] <-- R1_SP 1750 // alignment (optional) 1751 // [outgoing Java arguments] <-- R17_tos 1752 // ... 1753 // PARENT [PARENT_IJAVA_FRAME_ABI] 1754 // ... 1755 // 1756 1757 // Save initial_caller_sp to caller's abi. 1758 // The caller frame must be resized before returning to get rid of 1759 // the c2i part on top of the calling compiled frame (if any). 1760 // R21_tmp1 must match sender_sp in gen_c2i_adapter. 1761 // Now override the saved SP with the senderSP so we can pop c2i 1762 // arguments (if any) off when we return. 1763 __ std(R21_sender_SP, _top_ijava_frame_abi(initial_caller_sp), R1_SP); 1764 1765 // Save LR to caller's frame. We don't use _abi(lr) here, 1766 // because it is not safe. 1767 __ mflr(R0); 1768 __ std(R0, _top_ijava_frame_abi(frame_manager_lr), R1_SP); 1769 1770 // If we come here, it is the first invocation of the frame manager. 1771 // So there is no previous interpreter state. 1772 __ li(R15_prev_state, 0); 1773 1774 1775 // Fall through to where "recursive" invocations go. 1776 1777 //============================================================================= 1778 // Dispatch an instance of the interpreter. Recursive activations 1779 // come here. 1780 1781 Label re_dispatch; 1782 __ BIND(re_dispatch); 1783 1784 // 1785 // Registers alive 1786 // R16_thread - JavaThread* 1787 // R19_method - callee's Method 1788 // R17_tos - address of caller's tos (prepushed) 1789 // R15_prev_state - address of caller's BytecodeInterpreter or 0 1790 // R1_SP - caller's SP trimmed such that caller's outgoing args are near top. 1791 // 1792 // Stack layout at this point: 1793 // 1794 // 0 [TOP_IJAVA_FRAME_ABI] 1795 // alignment (optional) 1796 // [outgoing Java arguments] 1797 // ... 1798 // PARENT [PARENT_IJAVA_FRAME_ABI] 1799 // ... 1800 1801 // fall through to interpreted execution 1802 1803 //============================================================================= 1804 // Allocate a new Java frame and initialize the new interpreter state. 1805 1806 Label stack_overflow_return; 1807 1808 // Create a suitable new Java frame plus a new BytecodeInterpreter instance 1809 // in the current (frame manager's) C frame. 1810 generate_compute_interpreter_state(stack_overflow_return); 1811 1812 // fall through 1813 1814 //============================================================================= 1815 // Interpreter dispatch. 1816 1817 Label call_interpreter; 1818 __ BIND(call_interpreter); 1819 1820 // 1821 // Registers alive 1822 // R16_thread - JavaThread* 1823 // R15_prev_state - previous BytecodeInterpreter or 0 1824 // R14_state - address of receiver's BytecodeInterpreter 1825 // R1_SP - receiver's stack pointer 1826 // 1827 1828 // Thread fields. 1829 const Register pending_exception = R21_tmp1; 1830 1831 // Interpreter state fields. 1832 const Register msg = R24_tmp4; 1833 1834 // MethodOop fields. 1835 const Register parameter_count = R25_tmp5; 1836 const Register result_index = R26_tmp6; 1837 1838 const Register dummy = R28_tmp8; 1839 1840 // Address of various interpreter stubs. 1841 // R29_tmp9 is reserved. 1842 const Register stub_addr = R27_tmp7; 1843 1844 // Uncommon trap needs to jump to here to enter the interpreter 1845 // (re-execute current bytecode). 1846 unctrap_frame_manager_entry = __ pc(); 1847 1848 // If we are profiling, store our fp (BSP) in the thread so we can 1849 // find it during a tick. 1850 if (Arguments::has_profile()) { 1851 // On PPC64 we store the pointer to the current BytecodeInterpreter, 1852 // instead of the bsp of ia64. This should suffice to be able to 1853 // find all interesting information. 1854 __ std(R14_state, thread_(last_interpreter_fp)); 1855 } 1856 1857 // R16_thread, R14_state and R15_prev_state are nonvolatile 1858 // registers. There is no need to save these. If we needed to save 1859 // some state in the current Java frame, this could be a place to do 1860 // so. 1861 1862 // Call Java bytecode dispatcher passing "BytecodeInterpreter* istate". 1863 __ call_VM_leaf(CAST_FROM_FN_PTR(address, 1864 JvmtiExport::can_post_interpreter_events() 1865 ? BytecodeInterpreter::runWithChecks 1866 : BytecodeInterpreter::run), 1867 R14_state); 1868 1869 interpreter_return_address = __ last_calls_return_pc(); 1870 1871 // R16_thread, R14_state and R15_prev_state have their values preserved. 1872 1873 // If we are profiling, clear the fp in the thread to tell 1874 // the profiler that we are no longer in the interpreter. 1875 if (Arguments::has_profile()) { 1876 __ li(R11_scratch1, 0); 1877 __ std(R11_scratch1, thread_(last_interpreter_fp)); 1878 } 1879 1880 // Load message from bytecode dispatcher. 1881 // TODO: PPC port: guarantee(4 == BytecodeInterpreter::sz_msg(), "unexpected field size"); 1882 __ lwz(msg, state_(_msg)); 1883 1884 1885 Label more_monitors; 1886 Label return_from_native; 1887 Label return_from_native_common; 1888 Label return_from_native_no_exception; 1889 Label return_from_interpreted_method; 1890 Label return_from_recursive_activation; 1891 Label unwind_recursive_activation; 1892 Label resume_interpreter; 1893 Label return_to_initial_caller; 1894 Label unwind_initial_activation; 1895 Label unwind_initial_activation_pending_exception; 1896 Label call_method; 1897 Label call_special; 1898 Label retry_method; 1899 Label retry_method_osr; 1900 Label popping_frame; 1901 Label throwing_exception; 1902 1903 // Branch according to the received message 1904 1905 __ cmpwi(CCR1, msg, BytecodeInterpreter::call_method); 1906 __ cmpwi(CCR2, msg, BytecodeInterpreter::return_from_method); 1907 1908 __ beq(CCR1, call_method); 1909 __ beq(CCR2, return_from_interpreted_method); 1910 1911 __ cmpwi(CCR3, msg, BytecodeInterpreter::more_monitors); 1912 __ cmpwi(CCR4, msg, BytecodeInterpreter::throwing_exception); 1913 1914 __ beq(CCR3, more_monitors); 1915 __ beq(CCR4, throwing_exception); 1916 1917 __ cmpwi(CCR5, msg, BytecodeInterpreter::popping_frame); 1918 __ cmpwi(CCR6, msg, BytecodeInterpreter::do_osr); 1919 1920 __ beq(CCR5, popping_frame); 1921 __ beq(CCR6, retry_method_osr); 1922 1923 __ stop("bad message from interpreter"); 1924 1925 1926 //============================================================================= 1927 // Add a monitor just below the existing one(s). State->_stack_base 1928 // points to the lowest existing one, so we insert the new one just 1929 // below it and shuffle the expression stack down. Ref. the above 1930 // stack layout picture, we must update _stack_base, _stack, _stack_limit 1931 // and _last_Java_sp in the interpreter state. 1932 1933 __ BIND(more_monitors); 1934 1935 generate_more_monitors(); 1936 __ b(call_interpreter); 1937 1938 generate_deopt_handling(result_index); 1939 1940 // Restoring the R14_state is already done by the deopt_blob. 1941 1942 // Current tos includes no parameter slots. 1943 __ ld(R17_tos, state_(_stack)); 1944 __ li(msg, BytecodeInterpreter::deopt_resume); 1945 __ b(return_from_native_common); 1946 1947 // We are sent here when we are unwinding from a native method or 1948 // adapter with an exception pending. We need to notify the interpreter 1949 // that there is an exception to process. 1950 // We arrive here also if the frame manager called an (interpreted) target 1951 // which returns with a StackOverflow exception. 1952 // The control flow is in this case is: 1953 // frame_manager->throw_excp_stub->forward_excp->rethrow_excp_entry 1954 1955 AbstractInterpreter::_rethrow_exception_entry = __ pc(); 1956 1957 // Restore R14_state. 1958 __ ld(R14_state, 0, R1_SP); 1959 __ addi(R14_state, R14_state, 1960 -frame::interpreter_frame_cinterpreterstate_size_in_bytes()); 1961 1962 // Store exception oop into thread object. 1963 __ std(R3_RET, thread_(pending_exception)); 1964 __ li(msg, BytecodeInterpreter::method_resume /*rethrow_exception*/); 1965 // 1966 // NOTE: the interpreter frame as setup be deopt does NOT include 1967 // any parameter slots (good thing since we have no callee here 1968 // and couldn't remove them) so we don't have to do any calculations 1969 // here to figure it out. 1970 // 1971 __ ld(R17_tos, state_(_stack)); 1972 __ b(return_from_native_common); 1973 1974 1975 //============================================================================= 1976 // Returning from a native method. Result is in the native abi 1977 // location so we must move it to the java expression stack. 1978 1979 __ BIND(return_from_native); 1980 guarantee(return_from_native_pc == (address) NULL, "precondition"); 1981 return_from_native_pc = __ pc(); 1982 1983 // Restore R14_state. 1984 __ ld(R14_state, 0, R1_SP); 1985 __ addi(R14_state, R14_state, -frame::interpreter_frame_cinterpreterstate_size_in_bytes()); 1986 1987 // 1988 // Registers alive 1989 // R16_thread 1990 // R14_state - address of caller's BytecodeInterpreter. 1991 // R3_RET - integer result, if any. 1992 // F1_RET - float result, if any. 1993 // 1994 // Registers updated 1995 // R19_method - callee's Method 1996 // R17_tos - caller's tos, with outgoing args popped 1997 // result_index - index of result handler. 1998 // msg - message for resuming interpreter. 1999 // 2000 2001 // Very-local scratch registers. 2002 2003 const ConditionRegister have_pending_exception = CCR0; 2004 2005 // Load callee Method, gc may have moved it. 2006 __ ld(R19_method, state_(_result._to_call._callee)); 2007 2008 // Load address of caller's tos. includes parameter slots. 2009 __ ld(R17_tos, state_(_stack)); 2010 2011 // Pop callee's parameters. 2012 2013 __ ld(parameter_count, in_bytes(Method::const_offset()), R19_method); 2014 __ lhz(parameter_count, in_bytes(ConstMethod::size_of_parameters_offset()), parameter_count); 2015 __ sldi(parameter_count, parameter_count, Interpreter::logStackElementSize); 2016 __ add(R17_tos, R17_tos, parameter_count); 2017 2018 // Result stub address array index 2019 // TODO: PPC port: assert(4 == methodOopDesc::sz_result_index(), "unexpected field size"); 2020 __ lwa(result_index, method_(result_index)); 2021 2022 __ li(msg, BytecodeInterpreter::method_resume); 2023 2024 // 2025 // Registers alive 2026 // R16_thread 2027 // R14_state - address of caller's BytecodeInterpreter. 2028 // R17_tos - address of caller's tos with outgoing args already popped 2029 // R3_RET - integer return value, if any. 2030 // F1_RET - float return value, if any. 2031 // result_index - index of result handler. 2032 // msg - message for resuming interpreter. 2033 // 2034 // Registers updated 2035 // R3_RET - new address of caller's tos, including result, if any 2036 // 2037 2038 __ BIND(return_from_native_common); 2039 2040 // Check for pending exception 2041 __ ld(pending_exception, thread_(pending_exception)); 2042 __ cmpdi(CCR0, pending_exception, 0); 2043 __ beq(CCR0, return_from_native_no_exception); 2044 2045 // If there's a pending exception, we really have no result, so 2046 // R3_RET is dead. Resume_interpreter assumes the new tos is in 2047 // R3_RET. 2048 __ mr(R3_RET, R17_tos); 2049 // `resume_interpreter' expects R15_prev_state to be alive. 2050 __ ld(R15_prev_state, state_(_prev_link)); 2051 __ b(resume_interpreter); 2052 2053 __ BIND(return_from_native_no_exception); 2054 2055 // No pending exception, copy method result from native ABI register 2056 // to tos. 2057 2058 // Address of stub descriptor address array. 2059 __ load_const(stub_addr, CppInterpreter::tosca_result_to_stack()); 2060 2061 // Pass address of tos to stub. 2062 __ mr(R4_ARG2, R17_tos); 2063 2064 // Address of stub descriptor address. 2065 __ sldi(result_index, result_index, LogBytesPerWord); 2066 __ add(stub_addr, stub_addr, result_index); 2067 2068 // Stub descriptor address. 2069 __ ld(stub_addr, 0, stub_addr); 2070 2071 // TODO: don't do this via a call, do it in place! 2072 // 2073 // call stub via descriptor 2074 // in R3_ARG1/F1_ARG1: result value (R3_RET or F1_RET) 2075 __ call_stub(stub_addr); 2076 2077 // new tos = result of call in R3_RET 2078 2079 // `resume_interpreter' expects R15_prev_state to be alive. 2080 __ ld(R15_prev_state, state_(_prev_link)); 2081 __ b(resume_interpreter); 2082 2083 //============================================================================= 2084 // We encountered an exception while computing the interpreter 2085 // state, so R14_state isn't valid. Act as if we just returned from 2086 // the callee method with a pending exception. 2087 __ BIND(stack_overflow_return); 2088 2089 // 2090 // Registers alive 2091 // R16_thread - JavaThread* 2092 // R1_SP - old stack pointer 2093 // R19_method - callee's Method 2094 // R17_tos - address of caller's tos (prepushed) 2095 // R15_prev_state - address of caller's BytecodeInterpreter or 0 2096 // R18_locals - address of callee's locals array 2097 // 2098 // Registers updated 2099 // R3_RET - address of resuming tos, if recursive unwind 2100 2101 Label Lskip_unextend_SP; 2102 2103 { 2104 const ConditionRegister is_initial_call = CCR0; 2105 const Register tos_save = R21_tmp1; 2106 const Register tmp = R22_tmp2; 2107 2108 assert(tos_save->is_nonvolatile(), "need a nonvolatile"); 2109 2110 // Is the exception thrown in the initial Java frame of this frame 2111 // manager frame? 2112 __ cmpdi(is_initial_call, R15_prev_state, 0); 2113 __ bne(is_initial_call, Lskip_unextend_SP); 2114 2115 // Pop any c2i extension from the stack. This is necessary in the 2116 // non-recursive case (that is we were called by the c2i adapter, 2117 // meaning we have to prev state). In this case we entered the frame 2118 // manager through a special entry which pushes the orignal 2119 // unextended SP to the stack. Here we load it back. 2120 __ ld(R0, _top_ijava_frame_abi(frame_manager_lr), R1_SP); 2121 __ mtlr(R0); 2122 // Resize frame to get rid of a potential extension. 2123 __ resize_frame_to_initial_caller(R11_scratch1, R12_scratch2); 2124 2125 // Fall through 2126 2127 __ bind(Lskip_unextend_SP); 2128 2129 // Throw the exception via RuntimeStub "throw_StackOverflowError_entry". 2130 // 2131 // Previously, we called C-Code directly. As a consequence, a 2132 // possible GC tried to process the argument oops of the top frame 2133 // (see RegisterMap::clear, which sets the corresponding flag to 2134 // true). This lead to crashes because: 2135 // 1. The top register map did not contain locations for the argument registers 2136 // 2. The arguments are dead anyway, could be already overwritten in the worst case 2137 // Solution: Call via special runtime stub that pushes it's own frame. This runtime stub has the flag 2138 // "CodeBlob::caller_must_gc_arguments()" set to "false", what prevents the dead arguments getting GC'd. 2139 // 2140 // 2 cases exist: 2141 // 1. We were called by the c2i adapter / call stub 2142 // 2. We were called by the frame manager 2143 // 2144 // Both cases are handled by this code: 2145 // 1. - initial_caller_sp was saved on stack => Load it back and we're ok 2146 // - control flow will be: 2147 // throw_stackoverflow_stub->VM->throw_stackoverflow_stub->forward_excep->excp_blob of calling method 2148 // 2. - control flow will be: 2149 // throw_stackoverflow_stub->VM->throw_stackoverflow_stub->forward_excep-> 2150 // ->rethrow_excp_entry of frame manager->resume_method 2151 // Since we restored the caller SP above, the rethrow_excp_entry can restore the original interpreter state 2152 // registers using the stack and resume the calling method with a pending excp. 2153 2154 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "generated in wrong order"); 2155 __ load_const(R3_ARG1, (StubRoutines::throw_StackOverflowError_entry())); 2156 __ mtctr(R3_ARG1); 2157 __ bctr(); 2158 } 2159 //============================================================================= 2160 // We have popped a frame from an interpreted call. We are assured 2161 // of returning to an interpreted call by the popframe abi. We have 2162 // no return value all we have to do is pop the current frame and 2163 // then make sure that the top of stack (of the caller) gets set to 2164 // where it was when we entered the callee (i.e. the args are still 2165 // in place). Or we are returning to the interpreter. In the first 2166 // case we must extract result (if any) from the java expression 2167 // stack and store it in the location the native abi would expect 2168 // for a call returning this type. In the second case we must simply 2169 // do a stack to stack move as we unwind. 2170 2171 __ BIND(popping_frame); 2172 2173 // Registers alive 2174 // R14_state 2175 // R15_prev_state 2176 // R17_tos 2177 // 2178 // Registers updated 2179 // R19_method 2180 // R3_RET 2181 // msg 2182 { 2183 Label L; 2184 2185 // Reload callee method, gc may have moved it. 2186 __ ld(R19_method, state_(_method)); 2187 2188 // We may be returning to a deoptimized frame in which case the 2189 // usual assumption of a recursive return is not true. 2190 2191 // not equal = is recursive call 2192 __ cmpdi(CCR0, R15_prev_state, 0); 2193 2194 __ bne(CCR0, L); 2195 2196 // Pop_frame capability. 2197 // The pop_frame api says that the underlying frame is a Java frame, in this case 2198 // (prev_state==null) it must be a compiled frame: 2199 // 2200 // Stack at this point: I, C2I + C, ... 2201 // 2202 // The outgoing arguments of the call have just been copied (popframe_preserve_args). 2203 // By the pop_frame api, we must end up in an interpreted frame. So the compiled frame 2204 // will be deoptimized. Deoptimization will restore the outgoing arguments from 2205 // popframe_preserve_args, adjust the tos such that it includes the popframe_preserve_args, 2206 // and adjust the bci such that the call will be executed again. 2207 // We have no results, just pop the interpreter frame, resize the compiled frame to get rid 2208 // of the c2i extension and return to the deopt_handler. 2209 __ b(unwind_initial_activation); 2210 2211 // is recursive call 2212 __ bind(L); 2213 2214 // Resume_interpreter expects the original tos in R3_RET. 2215 __ ld(R3_RET, prev_state_(_stack)); 2216 2217 // We're done. 2218 __ li(msg, BytecodeInterpreter::popping_frame); 2219 2220 __ b(unwind_recursive_activation); 2221 } 2222 2223 2224 //============================================================================= 2225 2226 // We have finished an interpreted call. We are either returning to 2227 // native (call_stub/c2) or we are returning to the interpreter. 2228 // When returning to native, we must extract the result (if any) 2229 // from the java expression stack and store it in the location the 2230 // native abi expects. When returning to the interpreter we must 2231 // simply do a stack to stack move as we unwind. 2232 2233 __ BIND(return_from_interpreted_method); 2234 2235 // 2236 // Registers alive 2237 // R16_thread - JavaThread* 2238 // R15_prev_state - address of caller's BytecodeInterpreter or 0 2239 // R14_state - address of callee's interpreter state 2240 // R1_SP - callee's stack pointer 2241 // 2242 // Registers updated 2243 // R19_method - callee's method 2244 // R3_RET - address of result (new caller's tos), 2245 // 2246 // if returning to interpreted 2247 // msg - message for interpreter, 2248 // if returning to interpreted 2249 // 2250 2251 // Check if this is the initial invocation of the frame manager. 2252 // If so, R15_prev_state will be null. 2253 __ cmpdi(CCR0, R15_prev_state, 0); 2254 2255 // Reload callee method, gc may have moved it. 2256 __ ld(R19_method, state_(_method)); 2257 2258 // Load the method's result type. 2259 __ lwz(result_index, method_(result_index)); 2260 2261 // Go to return_to_initial_caller if R15_prev_state is null. 2262 __ beq(CCR0, return_to_initial_caller); 2263 2264 // Copy callee's result to caller's expression stack via inline stack-to-stack 2265 // converters. 2266 { 2267 Register new_tos = R3_RET; 2268 Register from_temp = R4_ARG2; 2269 Register from = R5_ARG3; 2270 Register tos = R6_ARG4; 2271 Register tmp1 = R7_ARG5; 2272 Register tmp2 = R8_ARG6; 2273 2274 ConditionRegister result_type_is_void = CCR1; 2275 ConditionRegister result_type_is_long = CCR2; 2276 ConditionRegister result_type_is_double = CCR3; 2277 2278 Label stack_to_stack_void; 2279 Label stack_to_stack_double_slot; // T_LONG, T_DOUBLE 2280 Label stack_to_stack_single_slot; // T_BOOLEAN, T_BYTE, T_CHAR, T_SHORT, T_INT, T_FLOAT, T_OBJECT 2281 Label stack_to_stack_done; 2282 2283 // Pass callee's address of tos + BytesPerWord 2284 __ ld(from_temp, state_(_stack)); 2285 2286 // result type: void 2287 __ cmpwi(result_type_is_void, result_index, AbstractInterpreter::BasicType_as_index(T_VOID)); 2288 2289 // Pass caller's tos == callee's locals address 2290 __ ld(tos, state_(_locals)); 2291 2292 // result type: long 2293 __ cmpwi(result_type_is_long, result_index, AbstractInterpreter::BasicType_as_index(T_LONG)); 2294 2295 __ addi(from, from_temp, Interpreter::stackElementSize); 2296 2297 // !! don't branch above this line !! 2298 2299 // handle void 2300 __ beq(result_type_is_void, stack_to_stack_void); 2301 2302 // result type: double 2303 __ cmpwi(result_type_is_double, result_index, AbstractInterpreter::BasicType_as_index(T_DOUBLE)); 2304 2305 // handle long or double 2306 __ beq(result_type_is_long, stack_to_stack_double_slot); 2307 __ beq(result_type_is_double, stack_to_stack_double_slot); 2308 2309 // fall through to single slot types (incl. object) 2310 2311 { 2312 __ BIND(stack_to_stack_single_slot); 2313 // T_BOOLEAN, T_BYTE, T_CHAR, T_SHORT, T_INT, T_FLOAT, T_OBJECT 2314 2315 __ ld(tmp1, 0, from); 2316 __ std(tmp1, 0, tos); 2317 // New expression stack top 2318 __ addi(new_tos, tos, - BytesPerWord); 2319 2320 __ b(stack_to_stack_done); 2321 } 2322 2323 { 2324 __ BIND(stack_to_stack_double_slot); 2325 // T_LONG, T_DOUBLE 2326 2327 // Move both entries for debug purposes even though only one is live 2328 __ ld(tmp1, BytesPerWord, from); 2329 __ ld(tmp2, 0, from); 2330 __ std(tmp1, 0, tos); 2331 __ std(tmp2, -BytesPerWord, tos); 2332 2333 // new expression stack top 2334 __ addi(new_tos, tos, - 2 * BytesPerWord); // two slots 2335 __ b(stack_to_stack_done); 2336 } 2337 2338 { 2339 __ BIND(stack_to_stack_void); 2340 // T_VOID 2341 2342 // new expression stack top 2343 __ mr(new_tos, tos); 2344 // fall through to stack_to_stack_done 2345 } 2346 2347 __ BIND(stack_to_stack_done); 2348 } 2349 2350 // new tos = R3_RET 2351 2352 // Get the message for the interpreter 2353 __ li(msg, BytecodeInterpreter::method_resume); 2354 2355 // And fall thru 2356 2357 2358 //============================================================================= 2359 // Restore caller's interpreter state and pass pointer to caller's 2360 // new tos to caller. 2361 2362 __ BIND(unwind_recursive_activation); 2363 2364 // 2365 // Registers alive 2366 // R15_prev_state - address of caller's BytecodeInterpreter 2367 // R3_RET - address of caller's tos 2368 // msg - message for caller's BytecodeInterpreter 2369 // R1_SP - callee's stack pointer 2370 // 2371 // Registers updated 2372 // R14_state - address of caller's BytecodeInterpreter 2373 // R15_prev_state - address of its parent or 0 2374 // 2375 2376 // Pop callee's interpreter and set R14_state to caller's interpreter. 2377 __ pop_interpreter_state(/*prev_state_may_be_0=*/false); 2378 2379 // And fall thru 2380 2381 2382 //============================================================================= 2383 // Resume the (calling) interpreter after a call. 2384 2385 __ BIND(resume_interpreter); 2386 2387 // 2388 // Registers alive 2389 // R14_state - address of resuming BytecodeInterpreter 2390 // R15_prev_state - address of its parent or 0 2391 // R3_RET - address of resuming tos 2392 // msg - message for resuming interpreter 2393 // R1_SP - callee's stack pointer 2394 // 2395 // Registers updated 2396 // R1_SP - caller's stack pointer 2397 // 2398 2399 // Restore C stack pointer of caller (resuming interpreter), 2400 // R14_state already points to the resuming BytecodeInterpreter. 2401 __ pop_interpreter_frame_to_state(R14_state, R21_tmp1, R11_scratch1, R12_scratch2); 2402 2403 // Store new address of tos (holding return value) in interpreter state. 2404 __ std(R3_RET, state_(_stack)); 2405 2406 // Store message for interpreter. 2407 __ stw(msg, state_(_msg)); 2408 2409 __ b(call_interpreter); 2410 2411 //============================================================================= 2412 // Interpreter returning to native code (call_stub/c1/c2) from 2413 // initial activation. Convert stack result and unwind activation. 2414 2415 __ BIND(return_to_initial_caller); 2416 2417 // 2418 // Registers alive 2419 // R19_method - callee's Method 2420 // R14_state - address of callee's interpreter state 2421 // R16_thread - JavaThread 2422 // R1_SP - callee's stack pointer 2423 // 2424 // Registers updated 2425 // R3_RET/F1_RET - result in expected output register 2426 // 2427 2428 // If we have an exception pending we have no result and we 2429 // must figure out where to really return to. 2430 // 2431 __ ld(pending_exception, thread_(pending_exception)); 2432 __ cmpdi(CCR0, pending_exception, 0); 2433 __ bne(CCR0, unwind_initial_activation_pending_exception); 2434 2435 __ lwa(result_index, method_(result_index)); 2436 2437 // Address of stub descriptor address array. 2438 __ load_const(stub_addr, CppInterpreter::stack_result_to_native()); 2439 2440 // Pass address of callee's tos + BytesPerWord. 2441 // Will then point directly to result. 2442 __ ld(R3_ARG1, state_(_stack)); 2443 __ addi(R3_ARG1, R3_ARG1, Interpreter::stackElementSize); 2444 2445 // Address of stub descriptor address 2446 __ sldi(result_index, result_index, LogBytesPerWord); 2447 __ add(stub_addr, stub_addr, result_index); 2448 2449 // Stub descriptor address 2450 __ ld(stub_addr, 0, stub_addr); 2451 2452 // TODO: don't do this via a call, do it in place! 2453 // 2454 // call stub via descriptor 2455 __ call_stub(stub_addr); 2456 2457 __ BIND(unwind_initial_activation); 2458 2459 // Unwind from initial activation. No exception is pending. 2460 2461 // 2462 // Stack layout at this point: 2463 // 2464 // 0 [TOP_IJAVA_FRAME_ABI] <-- R1_SP 2465 // ... 2466 // CALLER [PARENT_IJAVA_FRAME_ABI] 2467 // ... 2468 // CALLER [unextended ABI] 2469 // ... 2470 // 2471 // The CALLER frame has a C2I adapter or is an entry-frame. 2472 // 2473 2474 // An interpreter frame exists, we may pop the TOP_IJAVA_FRAME and 2475 // turn the caller's PARENT_IJAVA_FRAME back into a TOP_IJAVA_FRAME. 2476 // But, we simply restore the return pc from the caller's frame and 2477 // use the caller's initial_caller_sp as the new SP which pops the 2478 // interpreter frame and "resizes" the caller's frame to its "unextended" 2479 // size. 2480 2481 // get rid of top frame 2482 __ pop_frame(); 2483 2484 // Load return PC from parent frame. 2485 __ ld(R21_tmp1, _parent_ijava_frame_abi(lr), R1_SP); 2486 2487 // Resize frame to get rid of a potential extension. 2488 __ resize_frame_to_initial_caller(R11_scratch1, R12_scratch2); 2489 2490 // update LR 2491 __ mtlr(R21_tmp1); 2492 2493 // return 2494 __ blr(); 2495 2496 //============================================================================= 2497 // Unwind from initial activation. An exception is pending 2498 2499 __ BIND(unwind_initial_activation_pending_exception); 2500 2501 // 2502 // Stack layout at this point: 2503 // 2504 // 0 [TOP_IJAVA_FRAME_ABI] <-- R1_SP 2505 // ... 2506 // CALLER [PARENT_IJAVA_FRAME_ABI] 2507 // ... 2508 // CALLER [unextended ABI] 2509 // ... 2510 // 2511 // The CALLER frame has a C2I adapter or is an entry-frame. 2512 // 2513 2514 // An interpreter frame exists, we may pop the TOP_IJAVA_FRAME and 2515 // turn the caller's PARENT_IJAVA_FRAME back into a TOP_IJAVA_FRAME. 2516 // But, we just pop the current TOP_IJAVA_FRAME and fall through 2517 2518 __ pop_frame(); 2519 __ ld(R3_ARG1, _top_ijava_frame_abi(lr), R1_SP); 2520 2521 // 2522 // Stack layout at this point: 2523 // 2524 // CALLER [PARENT_IJAVA_FRAME_ABI] <-- R1_SP 2525 // ... 2526 // CALLER [unextended ABI] 2527 // ... 2528 // 2529 // The CALLER frame has a C2I adapter or is an entry-frame. 2530 // 2531 // Registers alive 2532 // R16_thread 2533 // R3_ARG1 - return address to caller 2534 // 2535 // Registers updated 2536 // R3_ARG1 - address of pending exception 2537 // R4_ARG2 - issuing pc = return address to caller 2538 // LR - address of exception handler stub 2539 // 2540 2541 // Resize frame to get rid of a potential extension. 2542 __ resize_frame_to_initial_caller(R11_scratch1, R12_scratch2); 2543 2544 __ mr(R14, R3_ARG1); // R14 := ARG1 2545 __ mr(R4_ARG2, R3_ARG1); // ARG2 := ARG1 2546 2547 // Find the address of the "catch_exception" stub. 2548 __ push_frame_abi112(0, R11_scratch1); 2549 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), 2550 R16_thread, 2551 R4_ARG2); 2552 __ pop_frame(); 2553 2554 // Load continuation address into LR. 2555 __ mtlr(R3_RET); 2556 2557 // Load address of pending exception and clear it in thread object. 2558 __ ld(R3_ARG1/*R3_RET*/, thread_(pending_exception)); 2559 __ li(R4_ARG2, 0); 2560 __ std(R4_ARG2, thread_(pending_exception)); 2561 2562 // re-load issuing pc 2563 __ mr(R4_ARG2, R14); 2564 2565 // Branch to found exception handler. 2566 __ blr(); 2567 2568 //============================================================================= 2569 // Call a new method. Compute new args and trim the expression stack 2570 // to only what we are currently using and then recurse. 2571 2572 __ BIND(call_method); 2573 2574 // 2575 // Registers alive 2576 // R16_thread 2577 // R14_state - address of caller's BytecodeInterpreter 2578 // R1_SP - caller's stack pointer 2579 // 2580 // Registers updated 2581 // R15_prev_state - address of caller's BytecodeInterpreter 2582 // R17_tos - address of caller's tos 2583 // R19_method - callee's Method 2584 // R1_SP - trimmed back 2585 // 2586 2587 // Very-local scratch registers. 2588 2589 const Register offset = R21_tmp1; 2590 const Register tmp = R22_tmp2; 2591 const Register self_entry = R23_tmp3; 2592 const Register stub_entry = R24_tmp4; 2593 2594 const ConditionRegister cr = CCR0; 2595 2596 // Load the address of the frame manager. 2597 __ load_const(self_entry, &interpreter_frame_manager); 2598 __ ld(self_entry, 0, self_entry); 2599 2600 // Load BytecodeInterpreter._result._to_call._callee (callee's Method). 2601 __ ld(R19_method, state_(_result._to_call._callee)); 2602 // Load BytecodeInterpreter._stack (outgoing tos). 2603 __ ld(R17_tos, state_(_stack)); 2604 2605 // Save address of caller's BytecodeInterpreter. 2606 __ mr(R15_prev_state, R14_state); 2607 2608 // Load the callee's entry point. 2609 // Load BytecodeInterpreter._result._to_call._callee_entry_point. 2610 __ ld(stub_entry, state_(_result._to_call._callee_entry_point)); 2611 2612 // Check whether stub_entry is equal to self_entry. 2613 __ cmpd(cr, self_entry, stub_entry); 2614 // if (self_entry == stub_entry) 2615 // do a re-dispatch 2616 __ beq(cr, re_dispatch); 2617 // else 2618 // call the specialized entry (adapter for jni or compiled code) 2619 __ BIND(call_special); 2620 2621 // 2622 // Call the entry generated by `InterpreterGenerator::generate_native_entry'. 2623 // 2624 // Registers alive 2625 // R16_thread 2626 // R15_prev_state - address of caller's BytecodeInterpreter 2627 // R19_method - callee's Method 2628 // R17_tos - address of caller's tos 2629 // R1_SP - caller's stack pointer 2630 // 2631 2632 // Mark return from specialized entry for generate_native_entry. 2633 guarantee(return_from_native_pc != (address) NULL, "precondition"); 2634 frame_manager_specialized_return = return_from_native_pc; 2635 2636 // Set sender_SP in case we call interpreter native wrapper which 2637 // will expect it. Compiled code should not care. 2638 __ mr(R21_sender_SP, R1_SP); 2639 2640 // Do a tail call here, and let the link register point to 2641 // frame_manager_specialized_return which is return_from_native_pc. 2642 __ load_const(tmp, frame_manager_specialized_return); 2643 __ call_stub_and_return_to(stub_entry, tmp /* return_pc=tmp */); 2644 2645 2646 //============================================================================= 2647 // 2648 // InterpretMethod triggered OSR compilation of some Java method M 2649 // and now asks to run the compiled code. We call this code the 2650 // `callee'. 2651 // 2652 // This is our current idea on how OSR should look like on PPC64: 2653 // 2654 // While interpreting a Java method M the stack is: 2655 // 2656 // (InterpretMethod (M), IJAVA_FRAME (M), ANY_FRAME, ...). 2657 // 2658 // After having OSR compiled M, `InterpretMethod' returns to the 2659 // frame manager, sending the message `retry_method_osr'. The stack 2660 // is: 2661 // 2662 // (IJAVA_FRAME (M), ANY_FRAME, ...). 2663 // 2664 // The compiler will have generated an `nmethod' suitable for 2665 // continuing execution of M at the bytecode index at which OSR took 2666 // place. So now the frame manager calls the OSR entry. The OSR 2667 // entry sets up a JIT_FRAME for M and continues execution of M with 2668 // initial state determined by the IJAVA_FRAME. 2669 // 2670 // (JIT_FRAME (M), IJAVA_FRAME (M), ANY_FRAME, ...). 2671 // 2672 2673 __ BIND(retry_method_osr); 2674 { 2675 // 2676 // Registers alive 2677 // R16_thread 2678 // R15_prev_state - address of caller's BytecodeInterpreter 2679 // R14_state - address of callee's BytecodeInterpreter 2680 // R1_SP - callee's SP before call to InterpretMethod 2681 // 2682 // Registers updated 2683 // R17 - pointer to callee's locals array 2684 // (declared via `interpreter_arg_ptr_reg' in the AD file) 2685 // R19_method - callee's Method 2686 // R1_SP - callee's SP (will become SP of OSR adapter frame) 2687 // 2688 2689 // Provide a debugger breakpoint in the frame manager if breakpoints 2690 // in osr'd methods are requested. 2691 #ifdef COMPILER2 2692 NOT_PRODUCT( if (OptoBreakpointOSR) { __ illtrap(); } ) 2693 #endif 2694 2695 // Load callee's pointer to locals array from callee's state. 2696 // __ ld(R17, state_(_locals)); 2697 2698 // Load osr entry. 2699 __ ld(R12_scratch2, state_(_result._osr._osr_entry)); 2700 2701 // Load address of temporary osr buffer to arg1. 2702 __ ld(R3_ARG1, state_(_result._osr._osr_buf)); 2703 __ mtctr(R12_scratch2); 2704 2705 // Load method oop, gc may move it during execution of osr'd method. 2706 __ ld(R22_tmp2, state_(_method)); 2707 // Load message 'call_method'. 2708 __ li(R23_tmp3, BytecodeInterpreter::call_method); 2709 2710 { 2711 // Pop the IJAVA frame of the method which we are going to call osr'd. 2712 Label no_state, skip_no_state; 2713 __ pop_interpreter_state(/*prev_state_may_be_0=*/true); 2714 __ cmpdi(CCR0, R14_state,0); 2715 __ beq(CCR0, no_state); 2716 // return to interpreter 2717 __ pop_interpreter_frame_to_state(R14_state, R11_scratch1, R12_scratch2, R21_tmp1); 2718 2719 // Init _result._to_call._callee and tell gc that it contains a valid oop 2720 // by setting _msg to 'call_method'. 2721 __ std(R22_tmp2, state_(_result._to_call._callee)); 2722 // TODO: PPC port: assert(4 == BytecodeInterpreter::sz_msg(), "unexpected field size"); 2723 __ stw(R23_tmp3, state_(_msg)); 2724 2725 __ load_const(R21_tmp1, frame_manager_specialized_return); 2726 __ b(skip_no_state); 2727 __ bind(no_state); 2728 2729 // Return to initial caller. 2730 2731 // Get rid of top frame. 2732 __ pop_frame(); 2733 2734 // Load return PC from parent frame. 2735 __ ld(R21_tmp1, _parent_ijava_frame_abi(lr), R1_SP); 2736 2737 // Resize frame to get rid of a potential extension. 2738 __ resize_frame_to_initial_caller(R11_scratch1, R12_scratch2); 2739 2740 __ bind(skip_no_state); 2741 2742 // Update LR with return pc. 2743 __ mtlr(R21_tmp1); 2744 } 2745 // Jump to the osr entry point. 2746 __ bctr(); 2747 2748 } 2749 2750 //============================================================================= 2751 // Interpreted method "returned" with an exception, pass it on. 2752 // Pass no result, unwind activation and continue/return to 2753 // interpreter/call_stub/c2. 2754 2755 __ BIND(throwing_exception); 2756 2757 // Check if this is the initial invocation of the frame manager. If 2758 // so, previous interpreter state in R15_prev_state will be null. 2759 2760 // New tos of caller is callee's first parameter address, that is 2761 // callee's incoming arguments are popped. 2762 __ ld(R3_RET, state_(_locals)); 2763 2764 // Check whether this is an initial call. 2765 __ cmpdi(CCR0, R15_prev_state, 0); 2766 // Yes, called from the call stub or from generated code via a c2i frame. 2767 __ beq(CCR0, unwind_initial_activation_pending_exception); 2768 2769 // Send resume message, interpreter will see the exception first. 2770 2771 __ li(msg, BytecodeInterpreter::method_resume); 2772 __ b(unwind_recursive_activation); 2773 2774 2775 //============================================================================= 2776 // Push the last instruction out to the code buffer. 2777 2778 { 2779 __ unimplemented("end of InterpreterGenerator::generate_normal_entry", 128); 2780 } 2781 2782 interpreter_frame_manager = entry; 2783 return interpreter_frame_manager; 2784 } 2785 2786 // Generate code for various sorts of method entries 2787 // 2788 address AbstractInterpreterGenerator::generate_method_entry(AbstractInterpreter::MethodKind kind) { 2789 address entry_point = NULL; 2790 2791 switch (kind) { 2792 case Interpreter::zerolocals : break; 2793 case Interpreter::zerolocals_synchronized : break; 2794 case Interpreter::native : // Fall thru 2795 case Interpreter::native_synchronized : entry_point = ((CppInterpreterGenerator*)this)->generate_native_entry(); break; 2796 case Interpreter::empty : break; 2797 case Interpreter::accessor : entry_point = ((InterpreterGenerator*)this)->generate_accessor_entry(); break; 2798 case Interpreter::abstract : entry_point = ((InterpreterGenerator*)this)->generate_abstract_entry(); break; 2799 // These are special interpreter intrinsics which we don't support so far. 2800 case Interpreter::java_lang_math_sin : break; 2801 case Interpreter::java_lang_math_cos : break; 2802 case Interpreter::java_lang_math_tan : break; 2803 case Interpreter::java_lang_math_abs : break; 2804 case Interpreter::java_lang_math_log : break; 2805 case Interpreter::java_lang_math_log10 : break; 2806 case Interpreter::java_lang_math_sqrt : break; 2807 case Interpreter::java_lang_math_pow : break; 2808 case Interpreter::java_lang_math_exp : break; 2809 case Interpreter::java_lang_ref_reference_get: entry_point = ((InterpreterGenerator*)this)->generate_Reference_get_entry(); break; 2810 default : ShouldNotReachHere(); break; 2811 } 2812 2813 if (entry_point) { 2814 return entry_point; 2815 } 2816 return ((InterpreterGenerator*)this)->generate_normal_entry(); 2817 } 2818 2819 InterpreterGenerator::InterpreterGenerator(StubQueue* code) 2820 : CppInterpreterGenerator(code) { 2821 generate_all(); // down here so it can be "virtual" 2822 } 2823 2824 // How much stack a topmost interpreter method activation needs in words. 2825 int AbstractInterpreter::size_top_interpreter_activation(Method* method) { 2826 // Computation is in bytes not words to match layout_activation_impl 2827 // below, but the return is in words. 2828 2829 // 2830 // 0 [TOP_IJAVA_FRAME_ABI] \ 2831 // alignment (optional) \ | 2832 // [operand stack / Java parameters] > stack | | 2833 // [monitors] (optional) > monitors | | 2834 // [PARENT_IJAVA_FRAME_ABI] \ | | 2835 // [BytecodeInterpreter object] > interpreter \ | | | 2836 // alignment (optional) | round | parent | round | top 2837 // [Java result] (2 slots) > result | | | | 2838 // [Java non-arg locals] \ locals | | | | 2839 // [arg locals] / / / / / 2840 // 2841 2842 int locals = method->max_locals() * BytesPerWord; 2843 int interpreter = frame::interpreter_frame_cinterpreterstate_size_in_bytes(); 2844 int result = 2 * BytesPerWord; 2845 2846 int parent = round_to(interpreter + result + locals, 16) + frame::parent_ijava_frame_abi_size; 2847 2848 int stack = method->max_stack() * BytesPerWord; 2849 int monitors = method->is_synchronized() ? frame::interpreter_frame_monitor_size_in_bytes() : 0; 2850 int top = round_to(parent + monitors + stack, 16) + frame::top_ijava_frame_abi_size; 2851 2852 return (top / BytesPerWord); 2853 } 2854 2855 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill, 2856 frame* caller, 2857 frame* current, 2858 Method* method, 2859 intptr_t* locals, 2860 intptr_t* stack, 2861 intptr_t* stack_base, 2862 intptr_t* monitor_base, 2863 intptr_t* frame_sp, 2864 bool is_top_frame) { 2865 // What about any vtable? 2866 // 2867 to_fill->_thread = JavaThread::current(); 2868 // This gets filled in later but make it something recognizable for now. 2869 to_fill->_bcp = method->code_base(); 2870 to_fill->_locals = locals; 2871 to_fill->_constants = method->constants()->cache(); 2872 to_fill->_method = method; 2873 to_fill->_mdx = NULL; 2874 to_fill->_stack = stack; 2875 2876 if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution()) { 2877 to_fill->_msg = deopt_resume2; 2878 } else { 2879 to_fill->_msg = method_resume; 2880 } 2881 to_fill->_result._to_call._bcp_advance = 0; 2882 to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone 2883 to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone 2884 to_fill->_prev_link = NULL; 2885 2886 if (caller->is_interpreted_frame()) { 2887 interpreterState prev = caller->get_interpreterState(); 2888 2889 // Support MH calls. Make sure the interpreter will return the right address: 2890 // 1. Caller did ordinary interpreted->compiled call call: Set a prev_state 2891 // which makes the CPP interpreter return to frame manager "return_from_interpreted_method" 2892 // entry after finishing execution. 2893 // 2. Caller did a MH call: If the caller has a MethodHandleInvoke in it's 2894 // state (invariant: must be the caller of the bottom vframe) we used the 2895 // "call_special" entry to do the call, meaning the arguments have not been 2896 // popped from the stack. Therefore, don't enter a prev state in this case 2897 // in order to return to "return_from_native" frame manager entry which takes 2898 // care of popping arguments. Also, don't overwrite the MH.invoke Method in 2899 // the prev_state in order to be able to figure out the number of arguments to 2900 // pop. 2901 // The parameter method can represent MethodHandle.invokeExact(...). 2902 // The MethodHandleCompiler generates these synthetic Methods, 2903 // including bytecodes, if an invokedynamic call gets inlined. In 2904 // this case we want to return like from any other interpreted 2905 // Java call, so we set _prev_link. 2906 to_fill->_prev_link = prev; 2907 2908 if (*prev->_bcp == Bytecodes::_invokeinterface || *prev->_bcp == Bytecodes::_invokedynamic) { 2909 prev->_result._to_call._bcp_advance = 5; 2910 } else { 2911 prev->_result._to_call._bcp_advance = 3; 2912 } 2913 } 2914 to_fill->_oop_temp = NULL; 2915 to_fill->_stack_base = stack_base; 2916 // Need +1 here because stack_base points to the word just above the 2917 // first expr stack entry and stack_limit is supposed to point to 2918 // the word just below the last expr stack entry. See 2919 // generate_compute_interpreter_state. 2920 to_fill->_stack_limit = stack_base - (method->max_stack() + 1); 2921 to_fill->_monitor_base = (BasicObjectLock*) monitor_base; 2922 2923 to_fill->_frame_bottom = frame_sp; 2924 2925 // PPC64 specific 2926 to_fill->_last_Java_pc = NULL; 2927 to_fill->_last_Java_fp = NULL; 2928 to_fill->_last_Java_sp = frame_sp; 2929 #ifdef ASSERT 2930 to_fill->_self_link = to_fill; 2931 to_fill->_native_fresult = 123456.789; 2932 to_fill->_native_lresult = CONST64(0xdeafcafedeadc0de); 2933 #endif 2934 } 2935 2936 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, 2937 address last_Java_pc, 2938 intptr_t* last_Java_fp) { 2939 istate->_last_Java_pc = last_Java_pc; 2940 istate->_last_Java_fp = last_Java_fp; 2941 } 2942 2943 int AbstractInterpreter::layout_activation(Method* method, 2944 int temps, // Number of slots on java expression stack in use. 2945 int popframe_args, 2946 int monitors, // Number of active monitors. 2947 int caller_actual_parameters, 2948 int callee_params,// Number of slots for callee parameters. 2949 int callee_locals,// Number of slots for locals. 2950 frame* caller, 2951 frame* interpreter_frame, 2952 bool is_top_frame, 2953 bool is_bottom_frame) { 2954 2955 // NOTE this code must exactly mimic what 2956 // InterpreterGenerator::generate_compute_interpreter_state() does 2957 // as far as allocating an interpreter frame. However there is an 2958 // exception. With the C++ based interpreter only the top most frame 2959 // has a full sized expression stack. The 16 byte slop factor is 2960 // both the abi scratch area and a place to hold a result from a 2961 // callee on its way to the callers stack. 2962 2963 int monitor_size = frame::interpreter_frame_monitor_size_in_bytes() * monitors; 2964 int frame_size; 2965 int top_frame_size = round_to(frame::interpreter_frame_cinterpreterstate_size_in_bytes() 2966 + monitor_size 2967 + (method->max_stack() *Interpreter::stackElementWords * BytesPerWord) 2968 + 2*BytesPerWord, 2969 frame::alignment_in_bytes) 2970 + frame::top_ijava_frame_abi_size; 2971 if (is_top_frame) { 2972 frame_size = top_frame_size; 2973 } else { 2974 frame_size = round_to(frame::interpreter_frame_cinterpreterstate_size_in_bytes() 2975 + monitor_size 2976 + ((temps - callee_params + callee_locals) * 2977 Interpreter::stackElementWords * BytesPerWord) 2978 + 2*BytesPerWord, 2979 frame::alignment_in_bytes) 2980 + frame::parent_ijava_frame_abi_size; 2981 assert(popframe_args==0, "non-zero for top_frame only"); 2982 } 2983 2984 // If we actually have a frame to layout we must now fill in all the pieces. 2985 if (interpreter_frame != NULL) { 2986 2987 intptr_t sp = (intptr_t)interpreter_frame->sp(); 2988 intptr_t fp = *(intptr_t *)sp; 2989 assert(fp == (intptr_t)caller->sp(), "fp must match"); 2990 interpreterState cur_state = 2991 (interpreterState)(fp - frame::interpreter_frame_cinterpreterstate_size_in_bytes()); 2992 2993 // Now fill in the interpreterState object. 2994 2995 intptr_t* locals; 2996 if (caller->is_interpreted_frame()) { 2997 // Locals must agree with the caller because it will be used to set the 2998 // caller's tos when we return. 2999 interpreterState prev = caller->get_interpreterState(); 3000 // Calculate start of "locals" for MH calls. For MH calls, the 3001 // current method() (= MH target) and prev->callee() (= 3002 // MH.invoke*()) are different and especially have different 3003 // signatures. To pop the argumentsof the caller, we must use 3004 // the prev->callee()->size_of_arguments() because that's what 3005 // the caller actually pushed. Currently, for synthetic MH 3006 // calls (deoptimized from inlined MH calls), detected by 3007 // is_method_handle_invoke(), we use the callee's arguments 3008 // because here, the caller's and callee's signature match. 3009 if (true /*!caller->is_at_mh_callsite()*/) { 3010 locals = prev->stack() + method->size_of_parameters(); 3011 } else { 3012 // Normal MH call. 3013 locals = prev->stack() + prev->callee()->size_of_parameters(); 3014 } 3015 } else { 3016 bool is_deopted; 3017 locals = (intptr_t*) (fp + ((method->max_locals() - 1) * BytesPerWord) + 3018 frame::parent_ijava_frame_abi_size); 3019 } 3020 3021 intptr_t* monitor_base = (intptr_t*) cur_state; 3022 intptr_t* stack_base = (intptr_t*) ((intptr_t) monitor_base - monitor_size); 3023 3024 // Provide pop_frame capability on PPC64, add popframe_args. 3025 // +1 because stack is always prepushed. 3026 intptr_t* stack = (intptr_t*) ((intptr_t) stack_base - (temps + popframe_args + 1) * BytesPerWord); 3027 3028 BytecodeInterpreter::layout_interpreterState(cur_state, 3029 caller, 3030 interpreter_frame, 3031 method, 3032 locals, 3033 stack, 3034 stack_base, 3035 monitor_base, 3036 (intptr_t*)(((intptr_t)fp)-top_frame_size), 3037 is_top_frame); 3038 3039 BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, 3040 interpreter_frame->fp()); 3041 } 3042 return frame_size/BytesPerWord; 3043 } 3044 3045 #endif // CC_INTERP