1 /* 2 * Copyright (c) 2007, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/assembler.hpp" 27 #include "ci/ciMethod.hpp" 28 #include "interpreter/bytecodeHistogram.hpp" 29 #include "interpreter/cppInterpreter.hpp" 30 #include "interpreter/interpreter.hpp" 31 #include "interpreter/interpreterGenerator.hpp" 32 #include "interpreter/interpreterRuntime.hpp" 33 #include "oops/arrayOop.hpp" 34 #include "oops/methodData.hpp" 35 #include "oops/method.hpp" 36 #include "oops/oop.inline.hpp" 37 #include "prims/jvmtiExport.hpp" 38 #include "prims/jvmtiThreadState.hpp" 39 #include "runtime/arguments.hpp" 40 #include "runtime/deoptimization.hpp" 41 #include "runtime/frame.inline.hpp" 42 #include "runtime/interfaceSupport.hpp" 43 #include "runtime/sharedRuntime.hpp" 44 #include "runtime/stubRoutines.hpp" 45 #include "runtime/synchronizer.hpp" 46 #include "runtime/timer.hpp" 47 #include "runtime/vframeArray.hpp" 48 #include "utilities/debug.hpp" 49 #include "utilities/macros.hpp" 50 #ifdef SHARK 51 #include "shark/shark_globals.hpp" 52 #endif 53 54 #ifdef CC_INTERP 55 56 // Routine exists to make tracebacks look decent in debugger 57 // while "shadow" interpreter frames are on stack. It is also 58 // used to distinguish interpreter frames. 59 60 extern "C" void RecursiveInterpreterActivation(interpreterState istate) { 61 ShouldNotReachHere(); 62 } 63 64 bool CppInterpreter::contains(address pc) { 65 return ( _code->contains(pc) || 66 ( pc == (CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset))); 67 } 68 69 #define STATE(field_name) Lstate, in_bytes(byte_offset_of(BytecodeInterpreter, field_name)) 70 #define __ _masm-> 71 72 Label frame_manager_entry; 73 Label fast_accessor_slow_entry_path; // fast accessor methods need to be able to jmp to unsynchronized 74 // c++ interpreter entry point this holds that entry point label. 75 76 static address unctrap_frame_manager_entry = NULL; 77 78 static address interpreter_return_address = NULL; 79 static address deopt_frame_manager_return_atos = NULL; 80 static address deopt_frame_manager_return_btos = NULL; 81 static address deopt_frame_manager_return_itos = NULL; 82 static address deopt_frame_manager_return_ltos = NULL; 83 static address deopt_frame_manager_return_ftos = NULL; 84 static address deopt_frame_manager_return_dtos = NULL; 85 static address deopt_frame_manager_return_vtos = NULL; 86 87 const Register prevState = G1_scratch; 88 89 void InterpreterGenerator::save_native_result(void) { 90 // result potentially in O0/O1: save it across calls 91 __ stf(FloatRegisterImpl::D, F0, STATE(_native_fresult)); 92 #ifdef _LP64 93 __ stx(O0, STATE(_native_lresult)); 94 #else 95 __ std(O0, STATE(_native_lresult)); 96 #endif 97 } 98 99 void InterpreterGenerator::restore_native_result(void) { 100 101 // Restore any method result value 102 __ ldf(FloatRegisterImpl::D, STATE(_native_fresult), F0); 103 #ifdef _LP64 104 __ ldx(STATE(_native_lresult), O0); 105 #else 106 __ ldd(STATE(_native_lresult), O0); 107 #endif 108 } 109 110 // A result handler converts/unboxes a native call result into 111 // a java interpreter/compiler result. The current frame is an 112 // interpreter frame. The activation frame unwind code must be 113 // consistent with that of TemplateTable::_return(...). In the 114 // case of native methods, the caller's SP was not modified. 115 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) { 116 address entry = __ pc(); 117 Register Itos_i = Otos_i ->after_save(); 118 Register Itos_l = Otos_l ->after_save(); 119 Register Itos_l1 = Otos_l1->after_save(); 120 Register Itos_l2 = Otos_l2->after_save(); 121 switch (type) { 122 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false 123 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value! 124 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break; 125 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break; 126 case T_LONG : 127 #ifndef _LP64 128 __ mov(O1, Itos_l2); // move other half of long 129 #endif // ifdef or no ifdef, fall through to the T_INT case 130 case T_INT : __ mov(O0, Itos_i); break; 131 case T_VOID : /* nothing to do */ break; 132 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break; 133 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break; 134 case T_OBJECT : 135 __ ld_ptr(STATE(_oop_temp), Itos_i); 136 __ verify_oop(Itos_i); 137 break; 138 default : ShouldNotReachHere(); 139 } 140 __ ret(); // return from interpreter activation 141 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame 142 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly 143 return entry; 144 } 145 146 // tosca based result to c++ interpreter stack based result. 147 // Result goes to address in L1_scratch 148 149 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) { 150 // A result is in the native abi result register from a native method call. 151 // We need to return this result to the interpreter by pushing the result on the interpreter's 152 // stack. This is relatively simple the destination is in L1_scratch 153 // i.e. L1_scratch is the first free element on the stack. If we "push" a return value we must 154 // adjust L1_scratch 155 address entry = __ pc(); 156 switch (type) { 157 case T_BOOLEAN: 158 // !0 => true; 0 => false 159 __ subcc(G0, O0, G0); 160 __ addc(G0, 0, O0); 161 __ st(O0, L1_scratch, 0); 162 __ sub(L1_scratch, wordSize, L1_scratch); 163 break; 164 165 // cannot use and3, 0xFFFF too big as immediate value! 166 case T_CHAR : 167 __ sll(O0, 16, O0); 168 __ srl(O0, 16, O0); 169 __ st(O0, L1_scratch, 0); 170 __ sub(L1_scratch, wordSize, L1_scratch); 171 break; 172 173 case T_BYTE : 174 __ sll(O0, 24, O0); 175 __ sra(O0, 24, O0); 176 __ st(O0, L1_scratch, 0); 177 __ sub(L1_scratch, wordSize, L1_scratch); 178 break; 179 180 case T_SHORT : 181 __ sll(O0, 16, O0); 182 __ sra(O0, 16, O0); 183 __ st(O0, L1_scratch, 0); 184 __ sub(L1_scratch, wordSize, L1_scratch); 185 break; 186 case T_LONG : 187 #ifndef _LP64 188 #if defined(COMPILER2) 189 // All return values are where we want them, except for Longs. C2 returns 190 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1. 191 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit 192 // build even if we are returning from interpreted we just do a little 193 // stupid shuffing. 194 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to 195 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node 196 // first which would move g1 -> O0/O1 and destroy the exception we were throwing. 197 __ stx(G1, L1_scratch, -wordSize); 198 #else 199 // native result is in O0, O1 200 __ st(O1, L1_scratch, 0); // Low order 201 __ st(O0, L1_scratch, -wordSize); // High order 202 #endif /* COMPILER2 */ 203 #else 204 __ stx(O0, L1_scratch, -wordSize); 205 #endif 206 __ sub(L1_scratch, 2*wordSize, L1_scratch); 207 break; 208 209 case T_INT : 210 __ st(O0, L1_scratch, 0); 211 __ sub(L1_scratch, wordSize, L1_scratch); 212 break; 213 214 case T_VOID : /* nothing to do */ 215 break; 216 217 case T_FLOAT : 218 __ stf(FloatRegisterImpl::S, F0, L1_scratch, 0); 219 __ sub(L1_scratch, wordSize, L1_scratch); 220 break; 221 222 case T_DOUBLE : 223 // Every stack slot is aligned on 64 bit, However is this 224 // the correct stack slot on 64bit?? QQQ 225 __ stf(FloatRegisterImpl::D, F0, L1_scratch, -wordSize); 226 __ sub(L1_scratch, 2*wordSize, L1_scratch); 227 break; 228 case T_OBJECT : 229 __ verify_oop(O0); 230 __ st_ptr(O0, L1_scratch, 0); 231 __ sub(L1_scratch, wordSize, L1_scratch); 232 break; 233 default : ShouldNotReachHere(); 234 } 235 __ retl(); // return from interpreter activation 236 __ delayed()->nop(); // schedule this better 237 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly 238 return entry; 239 } 240 241 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) { 242 // A result is in the java expression stack of the interpreted method that has just 243 // returned. Place this result on the java expression stack of the caller. 244 // 245 // The current interpreter activation in Lstate is for the method just returning its 246 // result. So we know that the result of this method is on the top of the current 247 // execution stack (which is pre-pushed) and will be return to the top of the caller 248 // stack. The top of the callers stack is the bottom of the locals of the current 249 // activation. 250 // Because of the way activation are managed by the frame manager the value of esp is 251 // below both the stack top of the current activation and naturally the stack top 252 // of the calling activation. This enable this routine to leave the return address 253 // to the frame manager on the stack and do a vanilla return. 254 // 255 // On entry: O0 - points to source (callee stack top) 256 // O1 - points to destination (caller stack top [i.e. free location]) 257 // destroys O2, O3 258 // 259 260 address entry = __ pc(); 261 switch (type) { 262 case T_VOID: break; 263 break; 264 case T_FLOAT : 265 case T_BOOLEAN: 266 case T_CHAR : 267 case T_BYTE : 268 case T_SHORT : 269 case T_INT : 270 // 1 word result 271 __ ld(O0, 0, O2); 272 __ st(O2, O1, 0); 273 __ sub(O1, wordSize, O1); 274 break; 275 case T_DOUBLE : 276 case T_LONG : 277 // return top two words on current expression stack to caller's expression stack 278 // The caller's expression stack is adjacent to the current frame manager's intepretState 279 // except we allocated one extra word for this intepretState so we won't overwrite it 280 // when we return a two word result. 281 #ifdef _LP64 282 __ ld_ptr(O0, 0, O2); 283 __ st_ptr(O2, O1, -wordSize); 284 #else 285 __ ld(O0, 0, O2); 286 __ ld(O0, wordSize, O3); 287 __ st(O3, O1, 0); 288 __ st(O2, O1, -wordSize); 289 #endif 290 __ sub(O1, 2*wordSize, O1); 291 break; 292 case T_OBJECT : 293 __ ld_ptr(O0, 0, O2); 294 __ verify_oop(O2); // verify it 295 __ st_ptr(O2, O1, 0); 296 __ sub(O1, wordSize, O1); 297 break; 298 default : ShouldNotReachHere(); 299 } 300 __ retl(); 301 __ delayed()->nop(); // QQ schedule this better 302 return entry; 303 } 304 305 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) { 306 // A result is in the java expression stack of the interpreted method that has just 307 // returned. Place this result in the native abi that the caller expects. 308 // We are in a new frame registers we set must be in caller (i.e. callstub) frame. 309 // 310 // Similar to generate_stack_to_stack_converter above. Called at a similar time from the 311 // frame manager execept in this situation the caller is native code (c1/c2/call_stub) 312 // and so rather than return result onto caller's java expression stack we return the 313 // result in the expected location based on the native abi. 314 // On entry: O0 - source (stack top) 315 // On exit result in expected output register 316 // QQQ schedule this better 317 318 address entry = __ pc(); 319 switch (type) { 320 case T_VOID: break; 321 break; 322 case T_FLOAT : 323 __ ldf(FloatRegisterImpl::S, O0, 0, F0); 324 break; 325 case T_BOOLEAN: 326 case T_CHAR : 327 case T_BYTE : 328 case T_SHORT : 329 case T_INT : 330 // 1 word result 331 __ ld(O0, 0, O0->after_save()); 332 break; 333 case T_DOUBLE : 334 __ ldf(FloatRegisterImpl::D, O0, 0, F0); 335 break; 336 case T_LONG : 337 // return top two words on current expression stack to caller's expression stack 338 // The caller's expression stack is adjacent to the current frame manager's interpretState 339 // except we allocated one extra word for this intepretState so we won't overwrite it 340 // when we return a two word result. 341 #ifdef _LP64 342 __ ld_ptr(O0, 0, O0->after_save()); 343 #else 344 __ ld(O0, wordSize, O1->after_save()); 345 __ ld(O0, 0, O0->after_save()); 346 #endif 347 #if defined(COMPILER2) && !defined(_LP64) 348 // C2 expects long results in G1 we can't tell if we're returning to interpreted 349 // or compiled so just be safe use G1 and O0/O1 350 351 // Shift bits into high (msb) of G1 352 __ sllx(Otos_l1->after_save(), 32, G1); 353 // Zero extend low bits 354 __ srl (Otos_l2->after_save(), 0, Otos_l2->after_save()); 355 __ or3 (Otos_l2->after_save(), G1, G1); 356 #endif /* COMPILER2 */ 357 break; 358 case T_OBJECT : 359 __ ld_ptr(O0, 0, O0->after_save()); 360 __ verify_oop(O0->after_save()); // verify it 361 break; 362 default : ShouldNotReachHere(); 363 } 364 __ retl(); 365 __ delayed()->nop(); 366 return entry; 367 } 368 369 address CppInterpreter::return_entry(TosState state, int length, Bytecodes::Code code) { 370 // make it look good in the debugger 371 return CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset; 372 } 373 374 address CppInterpreter::deopt_entry(TosState state, int length) { 375 address ret = NULL; 376 if (length != 0) { 377 switch (state) { 378 case atos: ret = deopt_frame_manager_return_atos; break; 379 case btos: ret = deopt_frame_manager_return_btos; break; 380 case ctos: 381 case stos: 382 case itos: ret = deopt_frame_manager_return_itos; break; 383 case ltos: ret = deopt_frame_manager_return_ltos; break; 384 case ftos: ret = deopt_frame_manager_return_ftos; break; 385 case dtos: ret = deopt_frame_manager_return_dtos; break; 386 case vtos: ret = deopt_frame_manager_return_vtos; break; 387 } 388 } else { 389 ret = unctrap_frame_manager_entry; // re-execute the bytecode ( e.g. uncommon trap) 390 } 391 assert(ret != NULL, "Not initialized"); 392 return ret; 393 } 394 395 // 396 // Helpers for commoning out cases in the various type of method entries. 397 // 398 399 // increment invocation count & check for overflow 400 // 401 // Note: checking for negative value instead of overflow 402 // so we have a 'sticky' overflow test 403 // 404 // Lmethod: method 405 // ??: invocation counter 406 // 407 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) { 408 Label done; 409 const Register Rcounters = G3_scratch; 410 411 __ ld_ptr(STATE(_method), G5_method); 412 __ get_method_counters(G5_method, Rcounters, done); 413 414 // Update standard invocation counters 415 __ increment_invocation_counter(Rcounters, O0, G4_scratch); 416 if (ProfileInterpreter) { 417 Address interpreter_invocation_counter(Rcounters, 418 in_bytes(MethodCounters::interpreter_invocation_counter_offset())); 419 __ ld(interpreter_invocation_counter, G4_scratch); 420 __ inc(G4_scratch); 421 __ st(G4_scratch, interpreter_invocation_counter); 422 } 423 424 AddressLiteral invocation_limit((address)&InvocationCounter::InterpreterInvocationLimit); 425 __ load_contents(invocation_limit, G3_scratch); 426 __ cmp(O0, G3_scratch); 427 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); 428 __ delayed()->nop(); 429 __ bind(done); 430 } 431 432 address InterpreterGenerator::generate_empty_entry(void) { 433 434 // A method that does nothing but return... 435 436 address entry = __ pc(); 437 Label slow_path; 438 439 // do nothing for empty methods (do not even increment invocation counter) 440 if ( UseFastEmptyMethods) { 441 // If we need a safepoint check, generate full interpreter entry. 442 AddressLiteral sync_state(SafepointSynchronize::address_of_state()); 443 __ load_contents(sync_state, G3_scratch); 444 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized); 445 __ br(Assembler::notEqual, false, Assembler::pn, frame_manager_entry); 446 __ delayed()->nop(); 447 448 // Code: _return 449 __ retl(); 450 __ delayed()->mov(O5_savedSP, SP); 451 return entry; 452 } 453 return NULL; 454 } 455 456 // Call an accessor method (assuming it is resolved, otherwise drop into 457 // vanilla (slow path) entry 458 459 // Generates code to elide accessor methods 460 // Uses G3_scratch and G1_scratch as scratch 461 address InterpreterGenerator::generate_accessor_entry(void) { 462 463 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof; 464 // parameter size = 1 465 // Note: We can only use this code if the getfield has been resolved 466 // and if we don't have a null-pointer exception => check for 467 // these conditions first and use slow path if necessary. 468 address entry = __ pc(); 469 Label slow_path; 470 471 if ( UseFastAccessorMethods) { 472 // Check if we need to reach a safepoint and generate full interpreter 473 // frame if so. 474 AddressLiteral sync_state(SafepointSynchronize::address_of_state()); 475 __ load_contents(sync_state, G3_scratch); 476 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized); 477 __ br(Assembler::notEqual, false, Assembler::pn, slow_path); 478 __ delayed()->nop(); 479 480 // Check if local 0 != NULL 481 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0 482 __ tst(Otos_i); // check if local 0 == NULL and go the slow path 483 __ brx(Assembler::zero, false, Assembler::pn, slow_path); 484 __ delayed()->nop(); 485 486 487 // read first instruction word and extract bytecode @ 1 and index @ 2 488 // get first 4 bytes of the bytecodes (big endian!) 489 __ ld_ptr(Address(G5_method, in_bytes(Method::const_offset())), G1_scratch); 490 __ ld(Address(G1_scratch, in_bytes(ConstMethod::codes_offset())), G1_scratch); 491 492 // move index @ 2 far left then to the right most two bytes. 493 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch); 494 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words( 495 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch); 496 497 // get constant pool cache 498 __ ld_ptr(G5_method, in_bytes(Method::const_offset()), G3_scratch); 499 __ ld_ptr(G3_scratch, in_bytes(ConstMethod::constants_offset()), G3_scratch); 500 __ ld_ptr(G3_scratch, ConstantPool::cache_offset_in_bytes(), G3_scratch); 501 502 // get specific constant pool cache entry 503 __ add(G3_scratch, G1_scratch, G3_scratch); 504 505 // Check the constant Pool cache entry to see if it has been resolved. 506 // If not, need the slow path. 507 ByteSize cp_base_offset = ConstantPoolCache::base_offset(); 508 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::indices_offset()), G1_scratch); 509 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch); 510 __ and3(G1_scratch, 0xFF, G1_scratch); 511 __ cmp(G1_scratch, Bytecodes::_getfield); 512 __ br(Assembler::notEqual, false, Assembler::pn, slow_path); 513 __ delayed()->nop(); 514 515 // Get the type and return field offset from the constant pool cache 516 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset()), G1_scratch); 517 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset()), G3_scratch); 518 519 Label xreturn_path; 520 // Need to differentiate between igetfield, agetfield, bgetfield etc. 521 // because they are different sizes. 522 // Get the type from the constant pool cache 523 __ srl(G1_scratch, ConstantPoolCacheEntry::tos_state_shift, G1_scratch); 524 // Make sure we don't need to mask G1_scratch after the above shift 525 ConstantPoolCacheEntry::verify_tos_state_shift(); 526 __ cmp(G1_scratch, atos ); 527 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 528 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i); 529 __ cmp(G1_scratch, itos); 530 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 531 __ delayed()->ld(Otos_i, G3_scratch, Otos_i); 532 __ cmp(G1_scratch, stos); 533 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 534 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i); 535 __ cmp(G1_scratch, ctos); 536 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 537 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i); 538 #ifdef ASSERT 539 __ cmp(G1_scratch, btos); 540 __ br(Assembler::equal, true, Assembler::pt, xreturn_path); 541 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i); 542 __ should_not_reach_here(); 543 #endif 544 __ ldsb(Otos_i, G3_scratch, Otos_i); 545 __ bind(xreturn_path); 546 547 // _ireturn/_areturn 548 __ retl(); // return from leaf routine 549 __ delayed()->mov(O5_savedSP, SP); 550 551 // Generate regular method entry 552 __ bind(slow_path); 553 __ ba(fast_accessor_slow_entry_path); 554 __ delayed()->nop(); 555 return entry; 556 } 557 return NULL; 558 } 559 560 address InterpreterGenerator::generate_Reference_get_entry(void) { 561 #if INCLUDE_ALL_GCS 562 if (UseG1GC) { 563 // We need to generate have a routine that generates code to: 564 // * load the value in the referent field 565 // * passes that value to the pre-barrier. 566 // 567 // In the case of G1 this will record the value of the 568 // referent in an SATB buffer if marking is active. 569 // This will cause concurrent marking to mark the referent 570 // field as live. 571 Unimplemented(); 572 } 573 #endif // INCLUDE_ALL_GCS 574 575 // If G1 is not enabled then attempt to go through the accessor entry point 576 // Reference.get is an accessor 577 return generate_accessor_entry(); 578 } 579 580 // 581 // Interpreter stub for calling a native method. (C++ interpreter) 582 // This sets up a somewhat different looking stack for calling the native method 583 // than the typical interpreter frame setup. 584 // 585 586 address InterpreterGenerator::generate_native_entry(bool synchronized) { 587 address entry = __ pc(); 588 589 // the following temporary registers are used during frame creation 590 const Register Gtmp1 = G3_scratch ; 591 const Register Gtmp2 = G1_scratch; 592 const Register RconstMethod = Gtmp1; 593 const Address constMethod(G5_method, in_bytes(Method::const_offset())); 594 const Address size_of_parameters(RconstMethod, in_bytes(ConstMethod::size_of_parameters_offset())); 595 596 bool inc_counter = UseCompiler || CountCompiledCalls; 597 598 // make sure registers are different! 599 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2); 600 601 const Address access_flags (G5_method, in_bytes(Method::access_flags_offset())); 602 603 Label Lentry; 604 __ bind(Lentry); 605 606 const Register Glocals_size = G3; 607 assert_different_registers(Glocals_size, G4_scratch, Gframe_size); 608 609 // make sure method is native & not abstract 610 // rethink these assertions - they can be simplified and shared (gri 2/25/2000) 611 #ifdef ASSERT 612 __ ld(access_flags, Gtmp1); 613 { 614 Label L; 615 __ btst(JVM_ACC_NATIVE, Gtmp1); 616 __ br(Assembler::notZero, false, Assembler::pt, L); 617 __ delayed()->nop(); 618 __ stop("tried to execute non-native method as native"); 619 __ bind(L); 620 } 621 { Label L; 622 __ btst(JVM_ACC_ABSTRACT, Gtmp1); 623 __ br(Assembler::zero, false, Assembler::pt, L); 624 __ delayed()->nop(); 625 __ stop("tried to execute abstract method as non-abstract"); 626 __ bind(L); 627 } 628 #endif // ASSERT 629 630 __ ld_ptr(constMethod, RconstMethod); 631 __ lduh(size_of_parameters, Gtmp1); 632 __ sll(Gtmp1, LogBytesPerWord, Gtmp2); // parameter size in bytes 633 __ add(Gargs, Gtmp2, Gargs); // points to first local + BytesPerWord 634 // NEW 635 __ add(Gargs, -wordSize, Gargs); // points to first local[0] 636 // generate the code to allocate the interpreter stack frame 637 // NEW FRAME ALLOCATED HERE 638 // save callers original sp 639 // __ mov(SP, I5_savedSP->after_restore()); 640 641 generate_compute_interpreter_state(Lstate, G0, true); 642 643 // At this point Lstate points to new interpreter state 644 // 645 646 const Address do_not_unlock_if_synchronized(G2_thread, 647 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset())); 648 // Since at this point in the method invocation the exception handler 649 // would try to exit the monitor of synchronized methods which hasn't 650 // been entered yet, we set the thread local variable 651 // _do_not_unlock_if_synchronized to true. If any exception was thrown by 652 // runtime, exception handling i.e. unlock_if_synchronized_method will 653 // check this thread local flag. 654 // This flag has two effects, one is to force an unwind in the topmost 655 // interpreter frame and not perform an unlock while doing so. 656 657 __ movbool(true, G3_scratch); 658 __ stbool(G3_scratch, do_not_unlock_if_synchronized); 659 660 661 // increment invocation counter and check for overflow 662 // 663 // Note: checking for negative value instead of overflow 664 // so we have a 'sticky' overflow test (may be of 665 // importance as soon as we have true MT/MP) 666 Label invocation_counter_overflow; 667 if (inc_counter) { 668 generate_counter_incr(&invocation_counter_overflow, NULL, NULL); 669 } 670 Label Lcontinue; 671 __ bind(Lcontinue); 672 673 bang_stack_shadow_pages(true); 674 // reset the _do_not_unlock_if_synchronized flag 675 __ stbool(G0, do_not_unlock_if_synchronized); 676 677 // check for synchronized methods 678 // Must happen AFTER invocation_counter check, so method is not locked 679 // if counter overflows. 680 681 if (synchronized) { 682 lock_method(); 683 // Don't see how G2_thread is preserved here... 684 // __ verify_thread(); QQQ destroys L0,L1 can't use 685 } else { 686 #ifdef ASSERT 687 { Label ok; 688 __ ld_ptr(STATE(_method), G5_method); 689 __ ld(access_flags, O0); 690 __ btst(JVM_ACC_SYNCHRONIZED, O0); 691 __ br( Assembler::zero, false, Assembler::pt, ok); 692 __ delayed()->nop(); 693 __ stop("method needs synchronization"); 694 __ bind(ok); 695 } 696 #endif // ASSERT 697 } 698 699 // start execution 700 701 // __ verify_thread(); kills L1,L2 can't use at the moment 702 703 // jvmti/jvmpi support 704 __ notify_method_entry(); 705 706 // native call 707 708 // (note that O0 is never an oop--at most it is a handle) 709 // It is important not to smash any handles created by this call, 710 // until any oop handle in O0 is dereferenced. 711 712 // (note that the space for outgoing params is preallocated) 713 714 // get signature handler 715 716 Label pending_exception_present; 717 718 { Label L; 719 __ ld_ptr(STATE(_method), G5_method); 720 __ ld_ptr(Address(G5_method, in_bytes(Method::signature_handler_offset())), G3_scratch); 721 __ tst(G3_scratch); 722 __ brx(Assembler::notZero, false, Assembler::pt, L); 723 __ delayed()->nop(); 724 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), G5_method, false); 725 __ ld_ptr(STATE(_method), G5_method); 726 727 Address exception_addr(G2_thread, in_bytes(Thread::pending_exception_offset())); 728 __ ld_ptr(exception_addr, G3_scratch); 729 __ br_notnull_short(G3_scratch, Assembler::pn, pending_exception_present); 730 __ ld_ptr(Address(G5_method, in_bytes(Method::signature_handler_offset())), G3_scratch); 731 __ bind(L); 732 } 733 734 // Push a new frame so that the args will really be stored in 735 // Copy a few locals across so the new frame has the variables 736 // we need but these values will be dead at the jni call and 737 // therefore not gc volatile like the values in the current 738 // frame (Lstate in particular) 739 740 // Flush the state pointer to the register save area 741 // Which is the only register we need for a stack walk. 742 __ st_ptr(Lstate, SP, (Lstate->sp_offset_in_saved_window() * wordSize) + STACK_BIAS); 743 744 __ mov(Lstate, O1); // Need to pass the state pointer across the frame 745 746 // Calculate current frame size 747 __ sub(SP, FP, O3); // Calculate negative of current frame size 748 __ save(SP, O3, SP); // Allocate an identical sized frame 749 750 __ mov(I1, Lstate); // In the "natural" register. 751 752 // Note I7 has leftover trash. Slow signature handler will fill it in 753 // should we get there. Normal jni call will set reasonable last_Java_pc 754 // below (and fix I7 so the stack trace doesn't have a meaningless frame 755 // in it). 756 757 758 // call signature handler 759 __ ld_ptr(STATE(_method), Lmethod); 760 __ ld_ptr(STATE(_locals), Llocals); 761 762 __ callr(G3_scratch, 0); 763 __ delayed()->nop(); 764 __ ld_ptr(STATE(_thread), G2_thread); // restore thread (shouldn't be needed) 765 766 { Label not_static; 767 768 __ ld_ptr(STATE(_method), G5_method); 769 __ ld(access_flags, O0); 770 __ btst(JVM_ACC_STATIC, O0); 771 __ br( Assembler::zero, false, Assembler::pt, not_static); 772 __ delayed()-> 773 // get native function entry point(O0 is a good temp until the very end) 774 ld_ptr(Address(G5_method, in_bytes(Method::native_function_offset())), O0); 775 // for static methods insert the mirror argument 776 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 777 778 __ ld_ptr(Address(G5_method, in_bytes(Method:: const_offset())), O1); 779 __ ld_ptr(Address(O1, in_bytes(ConstMethod::constants_offset())), O1); 780 __ ld_ptr(Address(O1, ConstantPool::pool_holder_offset_in_bytes()), O1); 781 __ ld_ptr(O1, mirror_offset, O1); 782 // where the mirror handle body is allocated: 783 #ifdef ASSERT 784 if (!PrintSignatureHandlers) // do not dirty the output with this 785 { Label L; 786 __ tst(O1); 787 __ brx(Assembler::notZero, false, Assembler::pt, L); 788 __ delayed()->nop(); 789 __ stop("mirror is missing"); 790 __ bind(L); 791 } 792 #endif // ASSERT 793 __ st_ptr(O1, STATE(_oop_temp)); 794 __ add(STATE(_oop_temp), O1); // this is really an LEA not an add 795 __ bind(not_static); 796 } 797 798 // At this point, arguments have been copied off of stack into 799 // their JNI positions, which are O1..O5 and SP[68..]. 800 // Oops are boxed in-place on the stack, with handles copied to arguments. 801 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*. 802 803 #ifdef ASSERT 804 { Label L; 805 __ tst(O0); 806 __ brx(Assembler::notZero, false, Assembler::pt, L); 807 __ delayed()->nop(); 808 __ stop("native entry point is missing"); 809 __ bind(L); 810 } 811 #endif // ASSERT 812 813 // 814 // setup the java frame anchor 815 // 816 // The scavenge function only needs to know that the PC of this frame is 817 // in the interpreter method entry code, it doesn't need to know the exact 818 // PC and hence we can use O7 which points to the return address from the 819 // previous call in the code stream (signature handler function) 820 // 821 // The other trick is we set last_Java_sp to FP instead of the usual SP because 822 // we have pushed the extra frame in order to protect the volatile register(s) 823 // in that frame when we return from the jni call 824 // 825 826 827 __ set_last_Java_frame(FP, O7); 828 __ mov(O7, I7); // make dummy interpreter frame look like one above, 829 // not meaningless information that'll confuse me. 830 831 // flush the windows now. We don't care about the current (protection) frame 832 // only the outer frames 833 834 __ flushw(); 835 836 // mark windows as flushed 837 Address flags(G2_thread, 838 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset())); 839 __ set(JavaFrameAnchor::flushed, G3_scratch); 840 __ st(G3_scratch, flags); 841 842 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready. 843 844 Address thread_state(G2_thread, in_bytes(JavaThread::thread_state_offset())); 845 #ifdef ASSERT 846 { Label L; 847 __ ld(thread_state, G3_scratch); 848 __ cmp(G3_scratch, _thread_in_Java); 849 __ br(Assembler::equal, false, Assembler::pt, L); 850 __ delayed()->nop(); 851 __ stop("Wrong thread state in native stub"); 852 __ bind(L); 853 } 854 #endif // ASSERT 855 __ set(_thread_in_native, G3_scratch); 856 __ st(G3_scratch, thread_state); 857 858 // Call the jni method, using the delay slot to set the JNIEnv* argument. 859 __ callr(O0, 0); 860 __ delayed()-> 861 add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0); 862 __ ld_ptr(STATE(_thread), G2_thread); // restore thread 863 864 // must we block? 865 866 // Block, if necessary, before resuming in _thread_in_Java state. 867 // In order for GC to work, don't clear the last_Java_sp until after blocking. 868 { Label no_block; 869 AddressLiteral sync_state(SafepointSynchronize::address_of_state()); 870 871 // Switch thread to "native transition" state before reading the synchronization state. 872 // This additional state is necessary because reading and testing the synchronization 873 // state is not atomic w.r.t. GC, as this scenario demonstrates: 874 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 875 // VM thread changes sync state to synchronizing and suspends threads for GC. 876 // Thread A is resumed to finish this native method, but doesn't block here since it 877 // didn't see any synchronization is progress, and escapes. 878 __ set(_thread_in_native_trans, G3_scratch); 879 __ st(G3_scratch, thread_state); 880 if(os::is_MP()) { 881 // Write serialization page so VM thread can do a pseudo remote membar. 882 // We use the current thread pointer to calculate a thread specific 883 // offset to write to within the page. This minimizes bus traffic 884 // due to cache line collision. 885 __ serialize_memory(G2_thread, G1_scratch, G3_scratch); 886 } 887 __ load_contents(sync_state, G3_scratch); 888 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized); 889 890 891 Label L; 892 Address suspend_state(G2_thread, in_bytes(JavaThread::suspend_flags_offset())); 893 __ br(Assembler::notEqual, false, Assembler::pn, L); 894 __ delayed()-> 895 ld(suspend_state, G3_scratch); 896 __ cmp(G3_scratch, 0); 897 __ br(Assembler::equal, false, Assembler::pt, no_block); 898 __ delayed()->nop(); 899 __ bind(L); 900 901 // Block. Save any potential method result value before the operation and 902 // use a leaf call to leave the last_Java_frame setup undisturbed. 903 save_native_result(); 904 __ call_VM_leaf(noreg, 905 CAST_FROM_FN_PTR(address, JavaThread::check_safepoint_and_suspend_for_native_trans), 906 G2_thread); 907 __ ld_ptr(STATE(_thread), G2_thread); // restore thread 908 // Restore any method result value 909 restore_native_result(); 910 __ bind(no_block); 911 } 912 913 // Clear the frame anchor now 914 915 __ reset_last_Java_frame(); 916 917 // Move the result handler address 918 __ mov(Lscratch, G3_scratch); 919 // return possible result to the outer frame 920 #ifndef __LP64 921 __ mov(O0, I0); 922 __ restore(O1, G0, O1); 923 #else 924 __ restore(O0, G0, O0); 925 #endif /* __LP64 */ 926 927 // Move result handler to expected register 928 __ mov(G3_scratch, Lscratch); 929 930 931 // thread state is thread_in_native_trans. Any safepoint blocking has 932 // happened in the trampoline we are ready to switch to thread_in_Java. 933 934 __ set(_thread_in_Java, G3_scratch); 935 __ st(G3_scratch, thread_state); 936 937 // If we have an oop result store it where it will be safe for any further gc 938 // until we return now that we've released the handle it might be protected by 939 940 { 941 Label no_oop, store_result; 942 943 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch); 944 __ cmp(G3_scratch, Lscratch); 945 __ brx(Assembler::notEqual, false, Assembler::pt, no_oop); 946 __ delayed()->nop(); 947 __ addcc(G0, O0, O0); 948 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL: 949 __ delayed()->ld_ptr(O0, 0, O0); // unbox it 950 __ mov(G0, O0); 951 952 __ bind(store_result); 953 // Store it where gc will look for it and result handler expects it. 954 __ st_ptr(O0, STATE(_oop_temp)); 955 956 __ bind(no_oop); 957 958 } 959 960 // reset handle block 961 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch); 962 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes()); 963 964 965 // handle exceptions (exception handling will handle unlocking!) 966 { Label L; 967 Address exception_addr (G2_thread, in_bytes(Thread::pending_exception_offset())); 968 969 __ ld_ptr(exception_addr, Gtemp); 970 __ tst(Gtemp); 971 __ brx(Assembler::equal, false, Assembler::pt, L); 972 __ delayed()->nop(); 973 __ bind(pending_exception_present); 974 // With c++ interpreter we just leave it pending caller will do the correct thing. However... 975 // Like x86 we ignore the result of the native call and leave the method locked. This 976 // seems wrong to leave things locked. 977 978 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type); 979 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame 980 981 __ bind(L); 982 } 983 984 // jvmdi/jvmpi support (preserves thread register) 985 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI); 986 987 if (synchronized) { 988 // save and restore any potential method result value around the unlocking operation 989 save_native_result(); 990 991 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize; 992 // Get the initial monitor we allocated 993 __ sub(Lstate, entry_size, O1); // initial monitor 994 __ unlock_object(O1); 995 restore_native_result(); 996 } 997 998 #if defined(COMPILER2) && !defined(_LP64) 999 1000 // C2 expects long results in G1 we can't tell if we're returning to interpreted 1001 // or compiled so just be safe. 1002 1003 __ sllx(O0, 32, G1); // Shift bits into high G1 1004 __ srl (O1, 0, O1); // Zero extend O1 1005 __ or3 (O1, G1, G1); // OR 64 bits into G1 1006 1007 #endif /* COMPILER2 && !_LP64 */ 1008 1009 #ifdef ASSERT 1010 { 1011 Label ok; 1012 __ cmp(I5_savedSP, FP); 1013 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok); 1014 __ delayed()->nop(); 1015 __ stop("bad I5_savedSP value"); 1016 __ should_not_reach_here(); 1017 __ bind(ok); 1018 } 1019 #endif 1020 // Calls result handler which POPS FRAME 1021 if (TraceJumps) { 1022 // Move target to register that is recordable 1023 __ mov(Lscratch, G3_scratch); 1024 __ JMP(G3_scratch, 0); 1025 } else { 1026 __ jmp(Lscratch, 0); 1027 } 1028 __ delayed()->nop(); 1029 1030 if (inc_counter) { 1031 // handle invocation counter overflow 1032 __ bind(invocation_counter_overflow); 1033 generate_counter_overflow(Lcontinue); 1034 } 1035 1036 1037 return entry; 1038 } 1039 1040 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state, 1041 const Register prev_state, 1042 bool native) { 1043 1044 // On entry 1045 // G5_method - caller's method 1046 // Gargs - points to initial parameters (i.e. locals[0]) 1047 // G2_thread - valid? (C1 only??) 1048 // "prev_state" - contains any previous frame manager state which we must save a link 1049 // 1050 // On return 1051 // "state" is a pointer to the newly allocated state object. We must allocate and initialize 1052 // a new interpretState object and the method expression stack. 1053 1054 assert_different_registers(state, prev_state); 1055 assert_different_registers(prev_state, G3_scratch); 1056 const Register Gtmp = G3_scratch; 1057 const Address constMethod (G5_method, in_bytes(Method::const_offset())); 1058 const Address access_flags (G5_method, in_bytes(Method::access_flags_offset())); 1059 1060 // slop factor is two extra slots on the expression stack so that 1061 // we always have room to store a result when returning from a call without parameters 1062 // that returns a result. 1063 1064 const int slop_factor = 2*wordSize; 1065 1066 const int fixed_size = ((sizeof(BytecodeInterpreter) + slop_factor) >> LogBytesPerWord) + // what is the slop factor? 1067 Method::extra_stack_entries() + // extra stack for jsr 292 1068 frame::memory_parameter_word_sp_offset + // register save area + param window 1069 (native ? frame::interpreter_frame_extra_outgoing_argument_words : 0); // JNI, class 1070 1071 // XXX G5_method valid 1072 1073 // Now compute new frame size 1074 1075 if (native) { 1076 const Register RconstMethod = Gtmp; 1077 const Address size_of_parameters(RconstMethod, in_bytes(ConstMethod::size_of_parameters_offset())); 1078 __ ld_ptr(constMethod, RconstMethod); 1079 __ lduh( size_of_parameters, Gtmp ); 1080 __ calc_mem_param_words(Gtmp, Gtmp); // space for native call parameters passed on the stack in words 1081 } else { 1082 // Full size expression stack 1083 __ ld_ptr(constMethod, Gtmp); 1084 __ lduh(Gtmp, in_bytes(ConstMethod::max_stack_offset()), Gtmp); 1085 } 1086 __ add(Gtmp, fixed_size, Gtmp); // plus the fixed portion 1087 1088 __ neg(Gtmp); // negative space for stack/parameters in words 1089 __ and3(Gtmp, -WordsPerLong, Gtmp); // make multiple of 2 (SP must be 2-word aligned) 1090 __ sll(Gtmp, LogBytesPerWord, Gtmp); // negative space for frame in bytes 1091 1092 // Need to do stack size check here before we fault on large frames 1093 1094 Label stack_ok; 1095 1096 const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages : 1097 (StackRedPages+StackYellowPages); 1098 1099 1100 __ ld_ptr(G2_thread, in_bytes(Thread::stack_base_offset()), O0); 1101 __ ld_ptr(G2_thread, in_bytes(Thread::stack_size_offset()), O1); 1102 // compute stack bottom 1103 __ sub(O0, O1, O0); 1104 1105 // Avoid touching the guard pages 1106 // Also a fudge for frame size of BytecodeInterpreter::run 1107 // It varies from 1k->4k depending on build type 1108 const int fudge = 6 * K; 1109 1110 __ set(fudge + (max_pages * os::vm_page_size()), O1); 1111 1112 __ add(O0, O1, O0); 1113 __ sub(O0, Gtmp, O0); 1114 __ cmp(SP, O0); 1115 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, stack_ok); 1116 __ delayed()->nop(); 1117 1118 // throw exception return address becomes throwing pc 1119 1120 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError)); 1121 __ stop("never reached"); 1122 1123 __ bind(stack_ok); 1124 1125 __ save(SP, Gtmp, SP); // setup new frame and register window 1126 1127 // New window I7 call_stub or previous activation 1128 // O6 - register save area, BytecodeInterpreter just below it, args/locals just above that 1129 // 1130 __ sub(FP, sizeof(BytecodeInterpreter), state); // Point to new Interpreter state 1131 __ add(state, STACK_BIAS, state ); // Account for 64bit bias 1132 1133 #define XXX_STATE(field_name) state, in_bytes(byte_offset_of(BytecodeInterpreter, field_name)) 1134 1135 // Initialize a new Interpreter state 1136 // orig_sp - caller's original sp 1137 // G2_thread - thread 1138 // Gargs - &locals[0] (unbiased?) 1139 // G5_method - method 1140 // SP (biased) - accounts for full size java stack, BytecodeInterpreter object, register save area, and register parameter save window 1141 1142 1143 __ set(0xdead0004, O1); 1144 1145 1146 __ st_ptr(Gargs, XXX_STATE(_locals)); 1147 __ st_ptr(G0, XXX_STATE(_oop_temp)); 1148 1149 __ st_ptr(state, XXX_STATE(_self_link)); // point to self 1150 __ st_ptr(prev_state->after_save(), XXX_STATE(_prev_link)); // Chain interpreter states 1151 __ st_ptr(G2_thread, XXX_STATE(_thread)); // Store javathread 1152 1153 if (native) { 1154 __ st_ptr(G0, XXX_STATE(_bcp)); 1155 } else { 1156 __ ld_ptr(G5_method, in_bytes(Method::const_offset()), O2); // get ConstMethod* 1157 __ add(O2, in_bytes(ConstMethod::codes_offset()), O2); // get bcp 1158 __ st_ptr(O2, XXX_STATE(_bcp)); 1159 } 1160 1161 __ st_ptr(G0, XXX_STATE(_mdx)); 1162 __ st_ptr(G5_method, XXX_STATE(_method)); 1163 1164 __ set((int) BytecodeInterpreter::method_entry, O1); 1165 __ st(O1, XXX_STATE(_msg)); 1166 1167 __ ld_ptr(constMethod, O3); 1168 __ ld_ptr(O3, in_bytes(ConstMethod::constants_offset()), O3); 1169 __ ld_ptr(O3, ConstantPool::cache_offset_in_bytes(), O2); 1170 __ st_ptr(O2, XXX_STATE(_constants)); 1171 1172 __ st_ptr(G0, XXX_STATE(_result._to_call._callee)); 1173 1174 // Monitor base is just start of BytecodeInterpreter object; 1175 __ mov(state, O2); 1176 __ st_ptr(O2, XXX_STATE(_monitor_base)); 1177 1178 // Do we need a monitor for synchonized method? 1179 { 1180 __ ld(access_flags, O1); 1181 Label done; 1182 Label got_obj; 1183 __ btst(JVM_ACC_SYNCHRONIZED, O1); 1184 __ br( Assembler::zero, false, Assembler::pt, done); 1185 1186 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 1187 __ delayed()->btst(JVM_ACC_STATIC, O1); 1188 __ ld_ptr(XXX_STATE(_locals), O1); 1189 __ br( Assembler::zero, true, Assembler::pt, got_obj); 1190 __ delayed()->ld_ptr(O1, 0, O1); // get receiver for not-static case 1191 __ ld_ptr(constMethod, O1); 1192 __ ld_ptr( O1, in_bytes(ConstMethod::constants_offset()), O1); 1193 __ ld_ptr( O1, ConstantPool::pool_holder_offset_in_bytes(), O1); 1194 // lock the mirror, not the Klass* 1195 __ ld_ptr( O1, mirror_offset, O1); 1196 1197 __ bind(got_obj); 1198 1199 #ifdef ASSERT 1200 __ tst(O1); 1201 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc); 1202 #endif // ASSERT 1203 1204 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize; 1205 __ sub(SP, entry_size, SP); // account for initial monitor 1206 __ sub(O2, entry_size, O2); // initial monitor 1207 __ st_ptr(O1, O2, BasicObjectLock::obj_offset_in_bytes()); // and allocate it for interpreter use 1208 __ bind(done); 1209 } 1210 1211 // Remember initial frame bottom 1212 1213 __ st_ptr(SP, XXX_STATE(_frame_bottom)); 1214 1215 __ st_ptr(O2, XXX_STATE(_stack_base)); 1216 1217 __ sub(O2, wordSize, O2); // prepush 1218 __ st_ptr(O2, XXX_STATE(_stack)); // PREPUSH 1219 1220 // Full size expression stack 1221 __ ld_ptr(constMethod, O3); 1222 __ lduh(O3, in_bytes(ConstMethod::max_stack_offset()), O3); 1223 __ inc(O3, Method::extra_stack_entries()); 1224 __ sll(O3, LogBytesPerWord, O3); 1225 __ sub(O2, O3, O3); 1226 // __ sub(O3, wordSize, O3); // so prepush doesn't look out of bounds 1227 __ st_ptr(O3, XXX_STATE(_stack_limit)); 1228 1229 if (!native) { 1230 // 1231 // Code to initialize locals 1232 // 1233 Register init_value = noreg; // will be G0 if we must clear locals 1234 // Now zero locals 1235 if (true /* zerolocals */ || ClearInterpreterLocals) { 1236 // explicitly initialize locals 1237 init_value = G0; 1238 } else { 1239 #ifdef ASSERT 1240 // initialize locals to a garbage pattern for better debugging 1241 init_value = O3; 1242 __ set( 0x0F0F0F0F, init_value ); 1243 #endif // ASSERT 1244 } 1245 if (init_value != noreg) { 1246 Label clear_loop; 1247 const Register RconstMethod = O1; 1248 const Address size_of_parameters(RconstMethod, in_bytes(ConstMethod::size_of_parameters_offset())); 1249 const Address size_of_locals (RconstMethod, in_bytes(ConstMethod::size_of_locals_offset())); 1250 1251 // NOTE: If you change the frame layout, this code will need to 1252 // be updated! 1253 __ ld_ptr( constMethod, RconstMethod ); 1254 __ lduh( size_of_locals, O2 ); 1255 __ lduh( size_of_parameters, O1 ); 1256 __ sll( O2, LogBytesPerWord, O2); 1257 __ sll( O1, LogBytesPerWord, O1 ); 1258 __ ld_ptr(XXX_STATE(_locals), L2_scratch); 1259 __ sub( L2_scratch, O2, O2 ); 1260 __ sub( L2_scratch, O1, O1 ); 1261 1262 __ bind( clear_loop ); 1263 __ inc( O2, wordSize ); 1264 1265 __ cmp( O2, O1 ); 1266 __ br( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop ); 1267 __ delayed()->st_ptr( init_value, O2, 0 ); 1268 } 1269 } 1270 } 1271 // Find preallocated monitor and lock method (C++ interpreter) 1272 // 1273 void InterpreterGenerator::lock_method(void) { 1274 // Lock the current method. 1275 // Destroys registers L2_scratch, L3_scratch, O0 1276 // 1277 // Find everything relative to Lstate 1278 1279 #ifdef ASSERT 1280 __ ld_ptr(STATE(_method), L2_scratch); 1281 __ ld(L2_scratch, in_bytes(Method::access_flags_offset()), O0); 1282 1283 { Label ok; 1284 __ btst(JVM_ACC_SYNCHRONIZED, O0); 1285 __ br( Assembler::notZero, false, Assembler::pt, ok); 1286 __ delayed()->nop(); 1287 __ stop("method doesn't need synchronization"); 1288 __ bind(ok); 1289 } 1290 #endif // ASSERT 1291 1292 // monitor is already allocated at stack base 1293 // and the lockee is already present 1294 __ ld_ptr(STATE(_stack_base), L2_scratch); 1295 __ ld_ptr(L2_scratch, BasicObjectLock::obj_offset_in_bytes(), O0); // get object 1296 __ lock_object(L2_scratch, O0); 1297 1298 } 1299 1300 // Generate code for handling resuming a deopted method 1301 void CppInterpreterGenerator::generate_deopt_handling() { 1302 1303 Label return_from_deopt_common; 1304 1305 // deopt needs to jump to here to enter the interpreter (return a result) 1306 deopt_frame_manager_return_atos = __ pc(); 1307 1308 // O0/O1 live 1309 __ ba(return_from_deopt_common); 1310 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_OBJECT), L3_scratch); // Result stub address array index 1311 1312 1313 // deopt needs to jump to here to enter the interpreter (return a result) 1314 deopt_frame_manager_return_btos = __ pc(); 1315 1316 // O0/O1 live 1317 __ ba(return_from_deopt_common); 1318 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_BOOLEAN), L3_scratch); // Result stub address array index 1319 1320 // deopt needs to jump to here to enter the interpreter (return a result) 1321 deopt_frame_manager_return_itos = __ pc(); 1322 1323 // O0/O1 live 1324 __ ba(return_from_deopt_common); 1325 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_INT), L3_scratch); // Result stub address array index 1326 1327 // deopt needs to jump to here to enter the interpreter (return a result) 1328 1329 deopt_frame_manager_return_ltos = __ pc(); 1330 #if !defined(_LP64) && defined(COMPILER2) 1331 // All return values are where we want them, except for Longs. C2 returns 1332 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1. 1333 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit 1334 // build even if we are returning from interpreted we just do a little 1335 // stupid shuffing. 1336 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to 1337 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node 1338 // first which would move g1 -> O0/O1 and destroy the exception we were throwing. 1339 1340 __ srl (G1, 0,O1); 1341 __ srlx(G1,32,O0); 1342 #endif /* !_LP64 && COMPILER2 */ 1343 // O0/O1 live 1344 __ ba(return_from_deopt_common); 1345 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_LONG), L3_scratch); // Result stub address array index 1346 1347 // deopt needs to jump to here to enter the interpreter (return a result) 1348 1349 deopt_frame_manager_return_ftos = __ pc(); 1350 // O0/O1 live 1351 __ ba(return_from_deopt_common); 1352 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_FLOAT), L3_scratch); // Result stub address array index 1353 1354 // deopt needs to jump to here to enter the interpreter (return a result) 1355 deopt_frame_manager_return_dtos = __ pc(); 1356 1357 // O0/O1 live 1358 __ ba(return_from_deopt_common); 1359 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_DOUBLE), L3_scratch); // Result stub address array index 1360 1361 // deopt needs to jump to here to enter the interpreter (return a result) 1362 deopt_frame_manager_return_vtos = __ pc(); 1363 1364 // O0/O1 live 1365 __ set(AbstractInterpreter::BasicType_as_index(T_VOID), L3_scratch); 1366 1367 // Deopt return common 1368 // an index is present that lets us move any possible result being 1369 // return to the interpreter's stack 1370 // 1371 __ bind(return_from_deopt_common); 1372 1373 // Result if any is in native abi result (O0..O1/F0..F1). The java expression 1374 // stack is in the state that the calling convention left it. 1375 // Copy the result from native abi result and place it on java expression stack. 1376 1377 // Current interpreter state is present in Lstate 1378 1379 // Get current pre-pushed top of interpreter stack 1380 // Any result (if any) is in native abi 1381 // result type index is in L3_scratch 1382 1383 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack 1384 1385 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch); 1386 __ sll(L3_scratch, LogBytesPerWord, L3_scratch); 1387 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address 1388 __ jmpl(Lscratch, G0, O7); // and convert it 1389 __ delayed()->nop(); 1390 1391 // L1_scratch points to top of stack (prepushed) 1392 __ st_ptr(L1_scratch, STATE(_stack)); 1393 } 1394 1395 // Generate the code to handle a more_monitors message from the c++ interpreter 1396 void CppInterpreterGenerator::generate_more_monitors() { 1397 1398 Label entry, loop; 1399 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize; 1400 // 1. compute new pointers // esp: old expression stack top 1401 __ delayed()->ld_ptr(STATE(_stack_base), L4_scratch); // current expression stack bottom 1402 __ sub(L4_scratch, entry_size, L4_scratch); 1403 __ st_ptr(L4_scratch, STATE(_stack_base)); 1404 1405 __ sub(SP, entry_size, SP); // Grow stack 1406 __ st_ptr(SP, STATE(_frame_bottom)); 1407 1408 __ ld_ptr(STATE(_stack_limit), L2_scratch); 1409 __ sub(L2_scratch, entry_size, L2_scratch); 1410 __ st_ptr(L2_scratch, STATE(_stack_limit)); 1411 1412 __ ld_ptr(STATE(_stack), L1_scratch); // Get current stack top 1413 __ sub(L1_scratch, entry_size, L1_scratch); 1414 __ st_ptr(L1_scratch, STATE(_stack)); 1415 __ ba(entry); 1416 __ delayed()->add(L1_scratch, wordSize, L1_scratch); // first real entry (undo prepush) 1417 1418 // 2. move expression stack 1419 1420 __ bind(loop); 1421 __ st_ptr(L3_scratch, Address(L1_scratch, 0)); 1422 __ add(L1_scratch, wordSize, L1_scratch); 1423 __ bind(entry); 1424 __ cmp(L1_scratch, L4_scratch); 1425 __ br(Assembler::notEqual, false, Assembler::pt, loop); 1426 __ delayed()->ld_ptr(L1_scratch, entry_size, L3_scratch); 1427 1428 // now zero the slot so we can find it. 1429 __ st_ptr(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes()); 1430 1431 } 1432 1433 // Initial entry to C++ interpreter from the call_stub. 1434 // This entry point is called the frame manager since it handles the generation 1435 // of interpreter activation frames via requests directly from the vm (via call_stub) 1436 // and via requests from the interpreter. The requests from the call_stub happen 1437 // directly thru the entry point. Requests from the interpreter happen via returning 1438 // from the interpreter and examining the message the interpreter has returned to 1439 // the frame manager. The frame manager can take the following requests: 1440 1441 // NO_REQUEST - error, should never happen. 1442 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and 1443 // allocate a new monitor. 1444 // CALL_METHOD - setup a new activation to call a new method. Very similar to what 1445 // happens during entry during the entry via the call stub. 1446 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub. 1447 // 1448 // Arguments: 1449 // 1450 // ebx: Method* 1451 // ecx: receiver - unused (retrieved from stack as needed) 1452 // esi: previous frame manager state (NULL from the call_stub/c1/c2) 1453 // 1454 // 1455 // Stack layout at entry 1456 // 1457 // [ return address ] <--- esp 1458 // [ parameter n ] 1459 // ... 1460 // [ parameter 1 ] 1461 // [ expression stack ] 1462 // 1463 // 1464 // We are free to blow any registers we like because the call_stub which brought us here 1465 // initially has preserved the callee save registers already. 1466 // 1467 // 1468 1469 static address interpreter_frame_manager = NULL; 1470 1471 #ifdef ASSERT 1472 #define VALIDATE_STATE(scratch, marker) \ 1473 { \ 1474 Label skip; \ 1475 __ ld_ptr(STATE(_self_link), scratch); \ 1476 __ cmp(Lstate, scratch); \ 1477 __ brx(Assembler::equal, false, Assembler::pt, skip); \ 1478 __ delayed()->nop(); \ 1479 __ breakpoint_trap(); \ 1480 __ emit_int32(marker); \ 1481 __ bind(skip); \ 1482 } 1483 #else 1484 #define VALIDATE_STATE(scratch, marker) 1485 #endif /* ASSERT */ 1486 1487 void CppInterpreterGenerator::adjust_callers_stack(Register args) { 1488 // 1489 // Adjust caller's stack so that all the locals can be contiguous with 1490 // the parameters. 1491 // Worries about stack overflow make this a pain. 1492 // 1493 // Destroys args, G3_scratch, G3_scratch 1494 // In/Out O5_savedSP (sender's original SP) 1495 // 1496 // assert_different_registers(state, prev_state); 1497 const Register Gtmp = G3_scratch; 1498 const Register RconstMethod = G3_scratch; 1499 const Register tmp = O2; 1500 const Address constMethod(G5_method, in_bytes(Method::const_offset())); 1501 const Address size_of_parameters(RconstMethod, in_bytes(ConstMethod::size_of_parameters_offset())); 1502 const Address size_of_locals (RconstMethod, in_bytes(ConstMethod::size_of_locals_offset())); 1503 1504 __ ld_ptr(constMethod, RconstMethod); 1505 __ lduh(size_of_parameters, tmp); 1506 __ sll(tmp, LogBytesPerWord, Gargs); // parameter size in bytes 1507 __ add(args, Gargs, Gargs); // points to first local + BytesPerWord 1508 // NEW 1509 __ add(Gargs, -wordSize, Gargs); // points to first local[0] 1510 // determine extra space for non-argument locals & adjust caller's SP 1511 // Gtmp1: parameter size in words 1512 __ lduh(size_of_locals, Gtmp); 1513 __ compute_extra_locals_size_in_bytes(tmp, Gtmp, Gtmp); 1514 1515 #if 1 1516 // c2i adapters place the final interpreter argument in the register save area for O0/I0 1517 // the call_stub will place the final interpreter argument at 1518 // frame::memory_parameter_word_sp_offset. This is mostly not noticable for either asm 1519 // or c++ interpreter. However with the c++ interpreter when we do a recursive call 1520 // and try to make it look good in the debugger we will store the argument to 1521 // RecursiveInterpreterActivation in the register argument save area. Without allocating 1522 // extra space for the compiler this will overwrite locals in the local array of the 1523 // interpreter. 1524 // QQQ still needed with frameless adapters??? 1525 1526 const int c2i_adjust_words = frame::memory_parameter_word_sp_offset - frame::callee_register_argument_save_area_sp_offset; 1527 1528 __ add(Gtmp, c2i_adjust_words*wordSize, Gtmp); 1529 #endif // 1 1530 1531 1532 __ sub(SP, Gtmp, SP); // just caller's frame for the additional space we need. 1533 } 1534 1535 address InterpreterGenerator::generate_normal_entry(bool synchronized) { 1536 1537 // G5_method: Method* 1538 // G2_thread: thread (unused) 1539 // Gargs: bottom of args (sender_sp) 1540 // O5: sender's sp 1541 1542 // A single frame manager is plenty as we don't specialize for synchronized. We could and 1543 // the code is pretty much ready. Would need to change the test below and for good measure 1544 // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized 1545 // routines. Not clear this is worth it yet. 1546 1547 if (interpreter_frame_manager) { 1548 return interpreter_frame_manager; 1549 } 1550 1551 __ bind(frame_manager_entry); 1552 1553 // the following temporary registers are used during frame creation 1554 const Register Gtmp1 = G3_scratch; 1555 // const Register Lmirror = L1; // native mirror (native calls only) 1556 1557 const Address constMethod (G5_method, in_bytes(Method::const_offset())); 1558 const Address access_flags (G5_method, in_bytes(Method::access_flags_offset())); 1559 1560 address entry_point = __ pc(); 1561 __ mov(G0, prevState); // no current activation 1562 1563 1564 Label re_dispatch; 1565 1566 __ bind(re_dispatch); 1567 1568 // Interpreter needs to have locals completely contiguous. In order to do that 1569 // We must adjust the caller's stack pointer for any locals beyond just the 1570 // parameters 1571 adjust_callers_stack(Gargs); 1572 1573 // O5_savedSP still contains sender's sp 1574 1575 // NEW FRAME 1576 1577 generate_compute_interpreter_state(Lstate, prevState, false); 1578 1579 // At this point a new interpreter frame and state object are created and initialized 1580 // Lstate has the pointer to the new activation 1581 // Any stack banging or limit check should already be done. 1582 1583 Label call_interpreter; 1584 1585 __ bind(call_interpreter); 1586 1587 1588 #if 1 1589 __ set(0xdead002, Lmirror); 1590 __ set(0xdead002, L2_scratch); 1591 __ set(0xdead003, L3_scratch); 1592 __ set(0xdead004, L4_scratch); 1593 __ set(0xdead005, Lscratch); 1594 __ set(0xdead006, Lscratch2); 1595 __ set(0xdead007, L7_scratch); 1596 1597 __ set(0xdeaf002, O2); 1598 __ set(0xdeaf003, O3); 1599 __ set(0xdeaf004, O4); 1600 __ set(0xdeaf005, O5); 1601 #endif 1602 1603 // Call interpreter (stack bang complete) enter here if message is 1604 // set and we know stack size is valid 1605 1606 Label call_interpreter_2; 1607 1608 __ bind(call_interpreter_2); 1609 1610 #ifdef ASSERT 1611 { 1612 Label skip; 1613 __ ld_ptr(STATE(_frame_bottom), G3_scratch); 1614 __ cmp(G3_scratch, SP); 1615 __ brx(Assembler::equal, false, Assembler::pt, skip); 1616 __ delayed()->nop(); 1617 __ stop("SP not restored to frame bottom"); 1618 __ bind(skip); 1619 } 1620 #endif 1621 1622 VALIDATE_STATE(G3_scratch, 4); 1623 __ set_last_Java_frame(SP, noreg); 1624 __ mov(Lstate, O0); // (arg) pointer to current state 1625 1626 __ call(CAST_FROM_FN_PTR(address, 1627 JvmtiExport::can_post_interpreter_events() ? 1628 BytecodeInterpreter::runWithChecks 1629 : BytecodeInterpreter::run), 1630 relocInfo::runtime_call_type); 1631 1632 __ delayed()->nop(); 1633 1634 __ ld_ptr(STATE(_thread), G2_thread); 1635 __ reset_last_Java_frame(); 1636 1637 // examine msg from interpreter to determine next action 1638 __ ld_ptr(STATE(_thread), G2_thread); // restore G2_thread 1639 1640 __ ld(STATE(_msg), L1_scratch); // Get new message 1641 1642 Label call_method; 1643 Label return_from_interpreted_method; 1644 Label throw_exception; 1645 Label do_OSR; 1646 Label bad_msg; 1647 Label resume_interpreter; 1648 1649 __ cmp(L1_scratch, (int)BytecodeInterpreter::call_method); 1650 __ br(Assembler::equal, false, Assembler::pt, call_method); 1651 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::return_from_method); 1652 __ br(Assembler::equal, false, Assembler::pt, return_from_interpreted_method); 1653 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::throwing_exception); 1654 __ br(Assembler::equal, false, Assembler::pt, throw_exception); 1655 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::do_osr); 1656 __ br(Assembler::equal, false, Assembler::pt, do_OSR); 1657 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::more_monitors); 1658 __ br(Assembler::notEqual, false, Assembler::pt, bad_msg); 1659 1660 // Allocate more monitor space, shuffle expression stack.... 1661 1662 generate_more_monitors(); 1663 1664 // new monitor slot allocated, resume the interpreter. 1665 1666 __ set((int)BytecodeInterpreter::got_monitors, L1_scratch); 1667 VALIDATE_STATE(G3_scratch, 5); 1668 __ ba(call_interpreter); 1669 __ delayed()->st(L1_scratch, STATE(_msg)); 1670 1671 // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode) 1672 unctrap_frame_manager_entry = __ pc(); 1673 1674 // QQQ what message do we send 1675 1676 __ ba(call_interpreter); 1677 __ delayed()->ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame 1678 1679 //============================================================================= 1680 // Returning from a compiled method into a deopted method. The bytecode at the 1681 // bcp has completed. The result of the bytecode is in the native abi (the tosca 1682 // for the template based interpreter). Any stack space that was used by the 1683 // bytecode that has completed has been removed (e.g. parameters for an invoke) 1684 // so all that we have to do is place any pending result on the expression stack 1685 // and resume execution on the next bytecode. 1686 1687 generate_deopt_handling(); 1688 1689 // ready to resume the interpreter 1690 1691 __ set((int)BytecodeInterpreter::deopt_resume, L1_scratch); 1692 __ ba(call_interpreter); 1693 __ delayed()->st(L1_scratch, STATE(_msg)); 1694 1695 // Current frame has caught an exception we need to dispatch to the 1696 // handler. We can get here because a native interpreter frame caught 1697 // an exception in which case there is no handler and we must rethrow 1698 // If it is a vanilla interpreted frame the we simply drop into the 1699 // interpreter and let it do the lookup. 1700 1701 Interpreter::_rethrow_exception_entry = __ pc(); 1702 1703 Label return_with_exception; 1704 Label unwind_and_forward; 1705 1706 // O0: exception 1707 // O7: throwing pc 1708 1709 // We want exception in the thread no matter what we ultimately decide about frame type. 1710 1711 Address exception_addr (G2_thread, in_bytes(Thread::pending_exception_offset())); 1712 __ verify_thread(); 1713 __ st_ptr(O0, exception_addr); 1714 1715 // get the Method* 1716 __ ld_ptr(STATE(_method), G5_method); 1717 1718 // if this current frame vanilla or native? 1719 1720 __ ld(access_flags, Gtmp1); 1721 __ btst(JVM_ACC_NATIVE, Gtmp1); 1722 __ br(Assembler::zero, false, Assembler::pt, return_with_exception); // vanilla interpreted frame handle directly 1723 __ delayed()->nop(); 1724 1725 // We drop thru to unwind a native interpreted frame with a pending exception 1726 // We jump here for the initial interpreter frame with exception pending 1727 // We unwind the current acivation and forward it to our caller. 1728 1729 __ bind(unwind_and_forward); 1730 1731 // Unwind frame and jump to forward exception. unwinding will place throwing pc in O7 1732 // as expected by forward_exception. 1733 1734 __ restore(FP, G0, SP); // unwind interpreter state frame 1735 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type); 1736 __ delayed()->mov(I5_savedSP->after_restore(), SP); 1737 1738 // Return point from a call which returns a result in the native abi 1739 // (c1/c2/jni-native). This result must be processed onto the java 1740 // expression stack. 1741 // 1742 // A pending exception may be present in which case there is no result present 1743 1744 address return_from_native_method = __ pc(); 1745 1746 VALIDATE_STATE(G3_scratch, 6); 1747 1748 // Result if any is in native abi result (O0..O1/F0..F1). The java expression 1749 // stack is in the state that the calling convention left it. 1750 // Copy the result from native abi result and place it on java expression stack. 1751 1752 // Current interpreter state is present in Lstate 1753 1754 // Exception pending? 1755 1756 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame 1757 __ ld_ptr(exception_addr, Lscratch); // get any pending exception 1758 __ tst(Lscratch); // exception pending? 1759 __ brx(Assembler::notZero, false, Assembler::pt, return_with_exception); 1760 __ delayed()->nop(); 1761 1762 // Process the native abi result to java expression stack 1763 1764 __ ld_ptr(STATE(_result._to_call._callee), L4_scratch); // called method 1765 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack 1766 // get parameter size 1767 __ ld_ptr(L4_scratch, in_bytes(Method::const_offset()), L2_scratch); 1768 __ lduh(L2_scratch, in_bytes(ConstMethod::size_of_parameters_offset()), L2_scratch); 1769 __ sll(L2_scratch, LogBytesPerWord, L2_scratch ); // parameter size in bytes 1770 __ add(L1_scratch, L2_scratch, L1_scratch); // stack destination for result 1771 __ ld(L4_scratch, in_bytes(Method::result_index_offset()), L3_scratch); // called method result type index 1772 1773 // tosca is really just native abi 1774 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch); 1775 __ sll(L3_scratch, LogBytesPerWord, L3_scratch); 1776 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address 1777 __ jmpl(Lscratch, G0, O7); // and convert it 1778 __ delayed()->nop(); 1779 1780 // L1_scratch points to top of stack (prepushed) 1781 1782 __ ba(resume_interpreter); 1783 __ delayed()->mov(L1_scratch, O1); 1784 1785 // An exception is being caught on return to a vanilla interpreter frame. 1786 // Empty the stack and resume interpreter 1787 1788 __ bind(return_with_exception); 1789 1790 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame 1791 __ ld_ptr(STATE(_stack_base), O1); // empty java expression stack 1792 __ ba(resume_interpreter); 1793 __ delayed()->sub(O1, wordSize, O1); // account for prepush 1794 1795 // Return from interpreted method we return result appropriate to the caller (i.e. "recursive" 1796 // interpreter call, or native) and unwind this interpreter activation. 1797 // All monitors should be unlocked. 1798 1799 __ bind(return_from_interpreted_method); 1800 1801 VALIDATE_STATE(G3_scratch, 7); 1802 1803 Label return_to_initial_caller; 1804 1805 // Interpreted result is on the top of the completed activation expression stack. 1806 // We must return it to the top of the callers stack if caller was interpreted 1807 // otherwise we convert to native abi result and return to call_stub/c1/c2 1808 // The caller's expression stack was truncated by the call however the current activation 1809 // has enough stuff on the stack that we have usable space there no matter what. The 1810 // other thing that makes it easy is that the top of the caller's stack is stored in STATE(_locals) 1811 // for the current activation 1812 1813 __ ld_ptr(STATE(_prev_link), L1_scratch); 1814 __ ld_ptr(STATE(_method), L2_scratch); // get method just executed 1815 __ ld(L2_scratch, in_bytes(Method::result_index_offset()), L2_scratch); 1816 __ tst(L1_scratch); 1817 __ brx(Assembler::zero, false, Assembler::pt, return_to_initial_caller); 1818 __ delayed()->sll(L2_scratch, LogBytesPerWord, L2_scratch); 1819 1820 // Copy result to callers java stack 1821 1822 __ set((intptr_t)CppInterpreter::_stack_to_stack, L4_scratch); 1823 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address 1824 __ ld_ptr(STATE(_stack), O0); // current top (prepushed) 1825 __ ld_ptr(STATE(_locals), O1); // stack destination 1826 1827 // O0 - will be source, O1 - will be destination (preserved) 1828 __ jmpl(Lscratch, G0, O7); // and convert it 1829 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack) 1830 1831 // O1 == &locals[0] 1832 1833 // Result is now on caller's stack. Just unwind current activation and resume 1834 1835 Label unwind_recursive_activation; 1836 1837 1838 __ bind(unwind_recursive_activation); 1839 1840 // O1 == &locals[0] (really callers stacktop) for activation now returning 1841 // returning to interpreter method from "recursive" interpreter call 1842 // result converter left O1 pointing to top of the( prepushed) java stack for method we are returning 1843 // to. Now all we must do is unwind the state from the completed call 1844 1845 // Must restore stack 1846 VALIDATE_STATE(G3_scratch, 8); 1847 1848 // Return to interpreter method after a method call (interpreted/native/c1/c2) has completed. 1849 // Result if any is already on the caller's stack. All we must do now is remove the now dead 1850 // frame and tell interpreter to resume. 1851 1852 1853 __ mov(O1, I1); // pass back new stack top across activation 1854 // POP FRAME HERE ================================== 1855 __ restore(FP, G0, SP); // unwind interpreter state frame 1856 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame 1857 1858 1859 // Resume the interpreter. The current frame contains the current interpreter 1860 // state object. 1861 // 1862 // O1 == new java stack pointer 1863 1864 __ bind(resume_interpreter); 1865 VALIDATE_STATE(G3_scratch, 10); 1866 1867 // A frame we have already used before so no need to bang stack so use call_interpreter_2 entry 1868 1869 __ set((int)BytecodeInterpreter::method_resume, L1_scratch); 1870 __ st(L1_scratch, STATE(_msg)); 1871 __ ba(call_interpreter_2); 1872 __ delayed()->st_ptr(O1, STATE(_stack)); 1873 1874 1875 // Fast accessor methods share this entry point. 1876 // This works because frame manager is in the same codelet 1877 // This can either be an entry via call_stub/c1/c2 or a recursive interpreter call 1878 // we need to do a little register fixup here once we distinguish the two of them 1879 if (UseFastAccessorMethods && !synchronized) { 1880 // Call stub_return address still in O7 1881 __ bind(fast_accessor_slow_entry_path); 1882 __ set((intptr_t)return_from_native_method - 8, Gtmp1); 1883 __ cmp(Gtmp1, O7); // returning to interpreter? 1884 __ brx(Assembler::equal, true, Assembler::pt, re_dispatch); // yep 1885 __ delayed()->nop(); 1886 __ ba(re_dispatch); 1887 __ delayed()->mov(G0, prevState); // initial entry 1888 1889 } 1890 1891 // interpreter returning to native code (call_stub/c1/c2) 1892 // convert result and unwind initial activation 1893 // L2_scratch - scaled result type index 1894 1895 __ bind(return_to_initial_caller); 1896 1897 __ set((intptr_t)CppInterpreter::_stack_to_native_abi, L4_scratch); 1898 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address 1899 __ ld_ptr(STATE(_stack), O0); // current top (prepushed) 1900 __ jmpl(Lscratch, G0, O7); // and convert it 1901 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack) 1902 1903 Label unwind_initial_activation; 1904 __ bind(unwind_initial_activation); 1905 1906 // RETURN TO CALL_STUB/C1/C2 code (result if any in I0..I1/(F0/..F1) 1907 // we can return here with an exception that wasn't handled by interpreted code 1908 // how does c1/c2 see it on return? 1909 1910 // compute resulting sp before/after args popped depending upon calling convention 1911 // __ ld_ptr(STATE(_saved_sp), Gtmp1); 1912 // 1913 // POP FRAME HERE ================================== 1914 __ restore(FP, G0, SP); 1915 __ retl(); 1916 __ delayed()->mov(I5_savedSP->after_restore(), SP); 1917 1918 // OSR request, unwind the current frame and transfer to the OSR entry 1919 // and enter OSR nmethod 1920 1921 __ bind(do_OSR); 1922 Label remove_initial_frame; 1923 __ ld_ptr(STATE(_prev_link), L1_scratch); 1924 __ ld_ptr(STATE(_result._osr._osr_buf), G1_scratch); 1925 1926 // We are going to pop this frame. Is there another interpreter frame underneath 1927 // it or is it callstub/compiled? 1928 1929 __ tst(L1_scratch); 1930 __ brx(Assembler::zero, false, Assembler::pt, remove_initial_frame); 1931 __ delayed()->ld_ptr(STATE(_result._osr._osr_entry), G3_scratch); 1932 1933 // Frame underneath is an interpreter frame simply unwind 1934 // POP FRAME HERE ================================== 1935 __ restore(FP, G0, SP); // unwind interpreter state frame 1936 __ mov(I5_savedSP->after_restore(), SP); 1937 1938 // Since we are now calling native need to change our "return address" from the 1939 // dummy RecursiveInterpreterActivation to a return from native 1940 1941 __ set((intptr_t)return_from_native_method - 8, O7); 1942 1943 __ jmpl(G3_scratch, G0, G0); 1944 __ delayed()->mov(G1_scratch, O0); 1945 1946 __ bind(remove_initial_frame); 1947 1948 // POP FRAME HERE ================================== 1949 __ restore(FP, G0, SP); 1950 __ mov(I5_savedSP->after_restore(), SP); 1951 __ jmpl(G3_scratch, G0, G0); 1952 __ delayed()->mov(G1_scratch, O0); 1953 1954 // Call a new method. All we do is (temporarily) trim the expression stack 1955 // push a return address to bring us back to here and leap to the new entry. 1956 // At this point we have a topmost frame that was allocated by the frame manager 1957 // which contains the current method interpreted state. We trim this frame 1958 // of excess java expression stack entries and then recurse. 1959 1960 __ bind(call_method); 1961 1962 // stack points to next free location and not top element on expression stack 1963 // method expects sp to be pointing to topmost element 1964 1965 __ ld_ptr(STATE(_thread), G2_thread); 1966 __ ld_ptr(STATE(_result._to_call._callee), G5_method); 1967 1968 1969 // SP already takes in to account the 2 extra words we use for slop 1970 // when we call a "static long no_params()" method. So if 1971 // we trim back sp by the amount of unused java expression stack 1972 // there will be automagically the 2 extra words we need. 1973 // We also have to worry about keeping SP aligned. 1974 1975 __ ld_ptr(STATE(_stack), Gargs); 1976 __ ld_ptr(STATE(_stack_limit), L1_scratch); 1977 1978 // compute the unused java stack size 1979 __ sub(Gargs, L1_scratch, L2_scratch); // compute unused space 1980 1981 // Round down the unused space to that stack is always 16-byte aligned 1982 // by making the unused space a multiple of the size of two longs. 1983 1984 __ and3(L2_scratch, -2*BytesPerLong, L2_scratch); 1985 1986 // Now trim the stack 1987 __ add(SP, L2_scratch, SP); 1988 1989 1990 // Now point to the final argument (account for prepush) 1991 __ add(Gargs, wordSize, Gargs); 1992 #ifdef ASSERT 1993 // Make sure we have space for the window 1994 __ sub(Gargs, SP, L1_scratch); 1995 __ cmp(L1_scratch, 16*wordSize); 1996 { 1997 Label skip; 1998 __ brx(Assembler::greaterEqual, false, Assembler::pt, skip); 1999 __ delayed()->nop(); 2000 __ stop("killed stack"); 2001 __ bind(skip); 2002 } 2003 #endif // ASSERT 2004 2005 // Create a new frame where we can store values that make it look like the interpreter 2006 // really recursed. 2007 2008 // prepare to recurse or call specialized entry 2009 2010 // First link the registers we need 2011 2012 // make the pc look good in debugger 2013 __ set(CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation), O7); 2014 // argument too 2015 __ mov(Lstate, I0); 2016 2017 // Record our sending SP 2018 __ mov(SP, O5_savedSP); 2019 2020 __ ld_ptr(STATE(_result._to_call._callee_entry_point), L2_scratch); 2021 __ set((intptr_t) entry_point, L1_scratch); 2022 __ cmp(L1_scratch, L2_scratch); 2023 __ brx(Assembler::equal, false, Assembler::pt, re_dispatch); 2024 __ delayed()->mov(Lstate, prevState); // link activations 2025 2026 // method uses specialized entry, push a return so we look like call stub setup 2027 // this path will handle fact that result is returned in registers and not 2028 // on the java stack. 2029 2030 __ set((intptr_t)return_from_native_method - 8, O7); 2031 __ jmpl(L2_scratch, G0, G0); // Do specialized entry 2032 __ delayed()->nop(); 2033 2034 // 2035 // Bad Message from interpreter 2036 // 2037 __ bind(bad_msg); 2038 __ stop("Bad message from interpreter"); 2039 2040 // Interpreted method "returned" with an exception pass it on... 2041 // Pass result, unwind activation and continue/return to interpreter/call_stub 2042 // We handle result (if any) differently based on return to interpreter or call_stub 2043 2044 __ bind(throw_exception); 2045 __ ld_ptr(STATE(_prev_link), L1_scratch); 2046 __ tst(L1_scratch); 2047 __ brx(Assembler::zero, false, Assembler::pt, unwind_and_forward); 2048 __ delayed()->nop(); 2049 2050 __ ld_ptr(STATE(_locals), O1); // get result of popping callee's args 2051 __ ba(unwind_recursive_activation); 2052 __ delayed()->nop(); 2053 2054 interpreter_frame_manager = entry_point; 2055 return entry_point; 2056 } 2057 2058 InterpreterGenerator::InterpreterGenerator(StubQueue* code) 2059 : CppInterpreterGenerator(code) { 2060 generate_all(); // down here so it can be "virtual" 2061 } 2062 2063 2064 template<class M> static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) { 2065 2066 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated 2067 // expression stack, the callee will have callee_extra_locals (so we can account for 2068 // frame extension) and monitor_size for monitors. Basically we need to calculate 2069 // this exactly like generate_fixed_frame/generate_compute_interpreter_state. 2070 // 2071 // 2072 // The big complicating thing here is that we must ensure that the stack stays properly 2073 // aligned. This would be even uglier if monitor size wasn't modulo what the stack 2074 // needs to be aligned for). We are given that the sp (fp) is already aligned by 2075 // the caller so we must ensure that it is properly aligned for our callee. 2076 // 2077 // Ths c++ interpreter always makes sure that we have a enough extra space on the 2078 // stack at all times to deal with the "stack long no_params()" method issue. This 2079 // is "slop_factor" here. 2080 const int slop_factor = 2; 2081 2082 const int fixed_size = sizeof(BytecodeInterpreter)/wordSize + // interpreter state object 2083 frame::memory_parameter_word_sp_offset; // register save area + param window 2084 return (round_to(max_stack + 2085 slop_factor + 2086 fixed_size + 2087 monitor_size + 2088 (callee_extra_locals * Interpreter::stackElementWords), WordsPerLong)); 2089 2090 } 2091 2092 int AbstractInterpreter::size_top_interpreter_activation(Method* method) { 2093 2094 // See call_stub code 2095 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset, 2096 WordsPerLong); // 7 + register save area 2097 2098 // Save space for one monitor to get into the interpreted method in case 2099 // the method is synchronized 2100 int monitor_size = method->is_synchronized() ? 2101 1*frame::interpreter_frame_monitor_size() : 0; 2102 return size_activation_helper<Method>(method->max_locals(), method->max_stack(), 2103 monitor_size) + call_stub_size; 2104 } 2105 2106 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill, 2107 frame* caller, 2108 frame* current, 2109 Method* method, 2110 intptr_t* locals, 2111 intptr_t* stack, 2112 intptr_t* stack_base, 2113 intptr_t* monitor_base, 2114 intptr_t* frame_bottom, 2115 bool is_top_frame 2116 ) 2117 { 2118 // What about any vtable? 2119 // 2120 to_fill->_thread = JavaThread::current(); 2121 // This gets filled in later but make it something recognizable for now 2122 to_fill->_bcp = method->code_base(); 2123 to_fill->_locals = locals; 2124 to_fill->_constants = method->constants()->cache(); 2125 to_fill->_method = method; 2126 to_fill->_mdx = NULL; 2127 to_fill->_stack = stack; 2128 if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) { 2129 to_fill->_msg = deopt_resume2; 2130 } else { 2131 to_fill->_msg = method_resume; 2132 } 2133 to_fill->_result._to_call._bcp_advance = 0; 2134 to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone 2135 to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone 2136 to_fill->_prev_link = NULL; 2137 2138 // Fill in the registers for the frame 2139 2140 // Need to install _sender_sp. Actually not too hard in C++! 2141 // When the skeletal frames are layed out we fill in a value 2142 // for _sender_sp. That value is only correct for the oldest 2143 // skeletal frame constructed (because there is only a single 2144 // entry for "caller_adjustment". While the skeletal frames 2145 // exist that is good enough. We correct that calculation 2146 // here and get all the frames correct. 2147 2148 // to_fill->_sender_sp = locals - (method->size_of_parameters() - 1); 2149 2150 *current->register_addr(Lstate) = (intptr_t) to_fill; 2151 // skeletal already places a useful value here and this doesn't account 2152 // for alignment so don't bother. 2153 // *current->register_addr(I5_savedSP) = (intptr_t) locals - (method->size_of_parameters() - 1); 2154 2155 if (caller->is_interpreted_frame()) { 2156 interpreterState prev = caller->get_interpreterState(); 2157 to_fill->_prev_link = prev; 2158 // Make the prev callee look proper 2159 prev->_result._to_call._callee = method; 2160 if (*prev->_bcp == Bytecodes::_invokeinterface) { 2161 prev->_result._to_call._bcp_advance = 5; 2162 } else { 2163 prev->_result._to_call._bcp_advance = 3; 2164 } 2165 } 2166 to_fill->_oop_temp = NULL; 2167 to_fill->_stack_base = stack_base; 2168 // Need +1 here because stack_base points to the word just above the first expr stack entry 2169 // and stack_limit is supposed to point to the word just below the last expr stack entry. 2170 // See generate_compute_interpreter_state. 2171 to_fill->_stack_limit = stack_base - (method->max_stack() + 1); 2172 to_fill->_monitor_base = (BasicObjectLock*) monitor_base; 2173 2174 // sparc specific 2175 to_fill->_frame_bottom = frame_bottom; 2176 to_fill->_self_link = to_fill; 2177 #ifdef ASSERT 2178 to_fill->_native_fresult = 123456.789; 2179 to_fill->_native_lresult = CONST64(0xdeadcafedeafcafe); 2180 #endif 2181 } 2182 2183 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, address last_Java_pc, intptr_t* last_Java_fp) { 2184 istate->_last_Java_pc = (intptr_t*) last_Java_pc; 2185 } 2186 2187 template<class M> static int frame_size_helper(M* method, 2188 int moncount, 2189 int callee_param_size, 2190 int callee_locals_size, 2191 bool is_top_frame, 2192 int& monitor_size, 2193 int& full_frame_words) { 2194 int extra_locals_size = callee_locals_size - callee_param_size; 2195 monitor_size = (sizeof(BasicObjectLock) * moncount) / wordSize; 2196 full_frame_words = size_activation_helper<M>(extra_locals_size, method->max_stack(), monitor_size); 2197 int short_frame_words = size_activation_helper<M>(extra_locals_size, method->max_stack(), monitor_size); 2198 int frame_words = is_top_frame ? full_frame_words : short_frame_words; 2199 2200 return frame_words; 2201 } 2202 2203 template<class M> int AbstractInterpreter::size_activation(M* method, 2204 int tempcount, 2205 int popframe_extra_args, 2206 int moncount, 2207 int callee_param_size, 2208 int callee_locals_size, 2209 bool is_top_frame) { 2210 assert(popframe_extra_args == 0, "NEED TO FIX"); 2211 // NOTE: return size is in words not bytes 2212 // Calculate the amount our frame will be adjust by the callee. For top frame 2213 // this is zero. 2214 2215 // NOTE: ia64 seems to do this wrong (or at least backwards) in that it 2216 // calculates the extra locals based on itself. Not what the callee does 2217 // to it. So it ignores last_frame_adjust value. Seems suspicious as far 2218 // as getting sender_sp correct. 2219 2220 int unused_monitor_size = 0; 2221 int unused_full_frame_words = 0; 2222 return frame_size_helper(method, moncount, callee_param_size, callee_locals_size, is_top_frame, 2223 unused_monitor_size, unused_full_frame_words); 2224 } 2225 2226 template int AbstractInterpreter::size_activation<Method>(Method* method, 2227 int temps, 2228 int popframe_args, 2229 int monitors, 2230 int callee_params, 2231 int callee_locals, 2232 bool is_top_frame); 2233 2234 template int AbstractInterpreter::size_activation<ciMethod>(ciMethod* method, 2235 int temps, 2236 int popframe_args, 2237 int monitors, 2238 int callee_params, 2239 int callee_locals, 2240 bool is_top_frame); 2241 2242 void AbstractInterpreter::layout_activation(Method* method, 2243 int tempcount, // Number of slots on java expression stack in use 2244 int popframe_extra_args, 2245 int moncount, // Number of active monitors 2246 int caller_actual_parameters, 2247 int callee_param_size, 2248 int callee_locals_size, 2249 frame* caller, 2250 frame* interpreter_frame, 2251 bool is_top_frame, 2252 bool is_bottom_frame) { 2253 assert(popframe_extra_args == 0, "NEED TO FIX"); 2254 // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state() 2255 // does as far as allocating an interpreter frame. 2256 // Set up the method, locals, and monitors. 2257 // The frame interpreter_frame is guaranteed to be the right size, 2258 // as determined by a previous call to the size_activation() method. 2259 // It is also guaranteed to be walkable even though it is in a skeletal state 2260 // NOTE: tempcount is the current size of the java expression stack. For top most 2261 // frames we will allocate a full sized expression stack and not the curback 2262 // version that non-top frames have. 2263 2264 int monitor_size = 0; 2265 int full_frame_words = 0; 2266 int frame_words = frame_size_helper<Method>(method, moncount, callee_param_size, callee_locals_size, 2267 is_top_frame, monitor_size, full_frame_words); 2268 2269 /* 2270 We must now fill in all the pieces of the frame. This means both 2271 the interpreterState and the registers. 2272 */ 2273 2274 // MUCHO HACK 2275 2276 intptr_t* frame_bottom = interpreter_frame->sp() - (full_frame_words - frame_words); 2277 // 'interpreter_frame->sp()' is unbiased while 'frame_bottom' must be a biased value in 64bit mode. 2278 assert(((intptr_t)frame_bottom & 0xf) == 0, "SP biased in layout_activation"); 2279 frame_bottom = (intptr_t*)((intptr_t)frame_bottom - STACK_BIAS); 2280 2281 /* Now fillin the interpreterState object */ 2282 2283 interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() - sizeof(BytecodeInterpreter)); 2284 2285 2286 intptr_t* locals; 2287 2288 // Calculate the postion of locals[0]. This is painful because of 2289 // stack alignment (same as ia64). The problem is that we can 2290 // not compute the location of locals from fp(). fp() will account 2291 // for the extra locals but it also accounts for aligning the stack 2292 // and we can't determine if the locals[0] was misaligned but max_locals 2293 // was enough to have the 2294 // calculate postion of locals. fp already accounts for extra locals. 2295 // +2 for the static long no_params() issue. 2296 2297 if (caller->is_interpreted_frame()) { 2298 // locals must agree with the caller because it will be used to set the 2299 // caller's tos when we return. 2300 interpreterState prev = caller->get_interpreterState(); 2301 // stack() is prepushed. 2302 locals = prev->stack() + method->size_of_parameters(); 2303 } else { 2304 // Lay out locals block in the caller adjacent to the register window save area. 2305 // 2306 // Compiled frames do not allocate a varargs area which is why this if 2307 // statement is needed. 2308 // 2309 intptr_t* fp = interpreter_frame->fp(); 2310 int local_words = method->max_locals() * Interpreter::stackElementWords; 2311 2312 if (caller->is_compiled_frame()) { 2313 locals = fp + frame::register_save_words + local_words - 1; 2314 } else { 2315 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1; 2316 } 2317 2318 } 2319 // END MUCHO HACK 2320 2321 intptr_t* monitor_base = (intptr_t*) cur_state; 2322 intptr_t* stack_base = monitor_base - monitor_size; 2323 /* +1 because stack is always prepushed */ 2324 intptr_t* stack = stack_base - (tempcount + 1); 2325 2326 2327 BytecodeInterpreter::layout_interpreterState(cur_state, 2328 caller, 2329 interpreter_frame, 2330 method, 2331 locals, 2332 stack, 2333 stack_base, 2334 monitor_base, 2335 frame_bottom, 2336 is_top_frame); 2337 2338 BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp()); 2339 } 2340 2341 #endif // CC_INTERP