1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "interpreter/bytecodeHistogram.hpp" 28 #include "interpreter/interpreter.hpp" 29 #include "interpreter/interpreterRuntime.hpp" 30 #include "interpreter/interp_masm.hpp" 31 #include "interpreter/templateInterpreterGenerator.hpp" 32 #include "interpreter/templateTable.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/sharedRuntime.hpp" 43 #include "runtime/stubRoutines.hpp" 44 #include "runtime/synchronizer.hpp" 45 #include "runtime/timer.hpp" 46 #include "runtime/vframeArray.hpp" 47 #include "utilities/debug.hpp" 48 #include "utilities/macros.hpp" 49 50 #ifndef FAST_DISPATCH 51 #define FAST_DISPATCH 1 52 #endif 53 #undef FAST_DISPATCH 54 55 // Size of interpreter code. Increase if too small. Interpreter will 56 // fail with a guarantee ("not enough space for interpreter generation"); 57 // if too small. 58 // Run with +PrintInterpreter to get the VM to print out the size. 59 // Max size with JVMTI 60 #ifdef _LP64 61 // The sethi() instruction generates lots more instructions when shell 62 // stack limit is unlimited, so that's why this is much bigger. 63 int TemplateInterpreter::InterpreterCodeSize = 260 * K; 64 #else 65 int TemplateInterpreter::InterpreterCodeSize = 230 * K; 66 #endif 67 68 // Generation of Interpreter 69 // 70 // The TemplateInterpreterGenerator generates the interpreter into Interpreter::_code. 71 72 73 #define __ _masm-> 74 75 76 //---------------------------------------------------------------------------------------------------- 77 78 #ifndef _LP64 79 address TemplateInterpreterGenerator::generate_slow_signature_handler() { 80 address entry = __ pc(); 81 Argument argv(0, true); 82 83 // We are in the jni transition frame. Save the last_java_frame corresponding to the 84 // outer interpreter frame 85 // 86 __ set_last_Java_frame(FP, noreg); 87 // make sure the interpreter frame we've pushed has a valid return pc 88 __ mov(O7, I7); 89 __ mov(Lmethod, G3_scratch); 90 __ mov(Llocals, G4_scratch); 91 __ save_frame(0); 92 __ mov(G2_thread, L7_thread_cache); 93 __ add(argv.address_in_frame(), O3); 94 __ mov(G2_thread, O0); 95 __ mov(G3_scratch, O1); 96 __ call(CAST_FROM_FN_PTR(address, InterpreterRuntime::slow_signature_handler), relocInfo::runtime_call_type); 97 __ delayed()->mov(G4_scratch, O2); 98 __ mov(L7_thread_cache, G2_thread); 99 __ reset_last_Java_frame(); 100 101 // load the register arguments (the C code packed them as varargs) 102 for (Argument ldarg = argv.successor(); ldarg.is_register(); ldarg = ldarg.successor()) { 103 __ ld_ptr(ldarg.address_in_frame(), ldarg.as_register()); 104 } 105 __ ret(); 106 __ delayed()-> 107 restore(O0, 0, Lscratch); // caller's Lscratch gets the result handler 108 return entry; 109 } 110 111 112 #else 113 // LP64 passes floating point arguments in F1, F3, F5, etc. instead of 114 // O0, O1, O2 etc.. 115 // Doubles are passed in D0, D2, D4 116 // We store the signature of the first 16 arguments in the first argument 117 // slot because it will be overwritten prior to calling the native 118 // function, with the pointer to the JNIEnv. 119 // If LP64 there can be up to 16 floating point arguments in registers 120 // or 6 integer registers. 121 address TemplateInterpreterGenerator::generate_slow_signature_handler() { 122 123 enum { 124 non_float = 0, 125 float_sig = 1, 126 double_sig = 2, 127 sig_mask = 3 128 }; 129 130 address entry = __ pc(); 131 Argument argv(0, true); 132 133 // We are in the jni transition frame. Save the last_java_frame corresponding to the 134 // outer interpreter frame 135 // 136 __ set_last_Java_frame(FP, noreg); 137 // make sure the interpreter frame we've pushed has a valid return pc 138 __ mov(O7, I7); 139 __ mov(Lmethod, G3_scratch); 140 __ mov(Llocals, G4_scratch); 141 __ save_frame(0); 142 __ mov(G2_thread, L7_thread_cache); 143 __ add(argv.address_in_frame(), O3); 144 __ mov(G2_thread, O0); 145 __ mov(G3_scratch, O1); 146 __ call(CAST_FROM_FN_PTR(address, InterpreterRuntime::slow_signature_handler), relocInfo::runtime_call_type); 147 __ delayed()->mov(G4_scratch, O2); 148 __ mov(L7_thread_cache, G2_thread); 149 __ reset_last_Java_frame(); 150 151 152 // load the register arguments (the C code packed them as varargs) 153 Address Sig = argv.address_in_frame(); // Argument 0 holds the signature 154 __ ld_ptr( Sig, G3_scratch ); // Get register argument signature word into G3_scratch 155 __ mov( G3_scratch, G4_scratch); 156 __ srl( G4_scratch, 2, G4_scratch); // Skip Arg 0 157 Label done; 158 for (Argument ldarg = argv.successor(); ldarg.is_float_register(); ldarg = ldarg.successor()) { 159 Label NonFloatArg; 160 Label LoadFloatArg; 161 Label LoadDoubleArg; 162 Label NextArg; 163 Address a = ldarg.address_in_frame(); 164 __ andcc(G4_scratch, sig_mask, G3_scratch); 165 __ br(Assembler::zero, false, Assembler::pt, NonFloatArg); 166 __ delayed()->nop(); 167 168 __ cmp(G3_scratch, float_sig ); 169 __ br(Assembler::equal, false, Assembler::pt, LoadFloatArg); 170 __ delayed()->nop(); 171 172 __ cmp(G3_scratch, double_sig ); 173 __ br(Assembler::equal, false, Assembler::pt, LoadDoubleArg); 174 __ delayed()->nop(); 175 176 __ bind(NonFloatArg); 177 // There are only 6 integer register arguments! 178 if ( ldarg.is_register() ) 179 __ ld_ptr(ldarg.address_in_frame(), ldarg.as_register()); 180 else { 181 // Optimization, see if there are any more args and get out prior to checking 182 // all 16 float registers. My guess is that this is rare. 183 // If is_register is false, then we are done the first six integer args. 184 __ br_null_short(G4_scratch, Assembler::pt, done); 185 } 186 __ ba(NextArg); 187 __ delayed()->srl( G4_scratch, 2, G4_scratch ); 188 189 __ bind(LoadFloatArg); 190 __ ldf( FloatRegisterImpl::S, a, ldarg.as_float_register(), 4); 191 __ ba(NextArg); 192 __ delayed()->srl( G4_scratch, 2, G4_scratch ); 193 194 __ bind(LoadDoubleArg); 195 __ ldf( FloatRegisterImpl::D, a, ldarg.as_double_register() ); 196 __ ba(NextArg); 197 __ delayed()->srl( G4_scratch, 2, G4_scratch ); 198 199 __ bind(NextArg); 200 201 } 202 203 __ bind(done); 204 __ ret(); 205 __ delayed()-> 206 restore(O0, 0, Lscratch); // caller's Lscratch gets the result handler 207 return entry; 208 } 209 #endif 210 211 void TemplateInterpreterGenerator::generate_counter_overflow(Label& Lcontinue) { 212 213 // Generate code to initiate compilation on the counter overflow. 214 215 // InterpreterRuntime::frequency_counter_overflow takes two arguments, 216 // the first indicates if the counter overflow occurs at a backwards branch (NULL bcp) 217 // and the second is only used when the first is true. We pass zero for both. 218 // The call returns the address of the verified entry point for the method or NULL 219 // if the compilation did not complete (either went background or bailed out). 220 __ set((int)false, O2); 221 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), O2, O2, true); 222 // returns verified_entry_point or NULL 223 // we ignore it in any case 224 __ ba_short(Lcontinue); 225 226 } 227 228 229 // End of helpers 230 231 // Various method entries 232 233 // Abstract method entry 234 // Attempt to execute abstract method. Throw exception 235 // 236 address TemplateInterpreterGenerator::generate_abstract_entry(void) { 237 address entry = __ pc(); 238 // abstract method entry 239 // throw exception 240 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodError)); 241 // the call_VM checks for exception, so we should never return here. 242 __ should_not_reach_here(); 243 return entry; 244 245 } 246 247 void TemplateInterpreterGenerator::save_native_result(void) { 248 // result potentially in O0/O1: save it across calls 249 const Address& l_tmp = InterpreterMacroAssembler::l_tmp; 250 251 // result potentially in F0/F1: save it across calls 252 const Address& d_tmp = InterpreterMacroAssembler::d_tmp; 253 254 // save and restore any potential method result value around the unlocking operation 255 __ stf(FloatRegisterImpl::D, F0, d_tmp); 256 #ifdef _LP64 257 __ stx(O0, l_tmp); 258 #else 259 __ std(O0, l_tmp); 260 #endif 261 } 262 263 void TemplateInterpreterGenerator::restore_native_result(void) { 264 const Address& l_tmp = InterpreterMacroAssembler::l_tmp; 265 const Address& d_tmp = InterpreterMacroAssembler::d_tmp; 266 267 // Restore any method result value 268 __ ldf(FloatRegisterImpl::D, d_tmp, F0); 269 #ifdef _LP64 270 __ ldx(l_tmp, O0); 271 #else 272 __ ldd(l_tmp, O0); 273 #endif 274 } 275 276 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) { 277 assert(!pass_oop || message == NULL, "either oop or message but not both"); 278 address entry = __ pc(); 279 // expression stack must be empty before entering the VM if an exception happened 280 __ empty_expression_stack(); 281 // load exception object 282 __ set((intptr_t)name, G3_scratch); 283 if (pass_oop) { 284 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i); 285 } else { 286 __ set((intptr_t)message, G4_scratch); 287 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch); 288 } 289 // throw exception 290 assert(Interpreter::throw_exception_entry() != NULL, "generate it first"); 291 AddressLiteral thrower(Interpreter::throw_exception_entry()); 292 __ jump_to(thrower, G3_scratch); 293 __ delayed()->nop(); 294 return entry; 295 } 296 297 address TemplateInterpreterGenerator::generate_ClassCastException_handler() { 298 address entry = __ pc(); 299 // expression stack must be empty before entering the VM if an exception 300 // happened 301 __ empty_expression_stack(); 302 // load exception object 303 __ call_VM(Oexception, 304 CAST_FROM_FN_PTR(address, 305 InterpreterRuntime::throw_ClassCastException), 306 Otos_i); 307 __ should_not_reach_here(); 308 return entry; 309 } 310 311 312 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) { 313 address entry = __ pc(); 314 // expression stack must be empty before entering the VM if an exception happened 315 __ empty_expression_stack(); 316 // convention: expect aberrant index in register G3_scratch, then shuffle the 317 // index to G4_scratch for the VM call 318 __ mov(G3_scratch, G4_scratch); 319 __ set((intptr_t)name, G3_scratch); 320 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch); 321 __ should_not_reach_here(); 322 return entry; 323 } 324 325 326 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() { 327 address entry = __ pc(); 328 // expression stack must be empty before entering the VM if an exception happened 329 __ empty_expression_stack(); 330 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError)); 331 __ should_not_reach_here(); 332 return entry; 333 } 334 335 336 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) { 337 address entry = __ pc(); 338 339 if (state == atos) { 340 __ profile_return_type(O0, G3_scratch, G1_scratch); 341 } 342 343 #if !defined(_LP64) && defined(COMPILER2) 344 // All return values are where we want them, except for Longs. C2 returns 345 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1. 346 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit 347 // build even if we are returning from interpreted we just do a little 348 // stupid shuffing. 349 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to 350 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node 351 // first which would move g1 -> O0/O1 and destroy the exception we were throwing. 352 353 if (state == ltos) { 354 __ srl (G1, 0, O1); 355 __ srlx(G1, 32, O0); 356 } 357 #endif // !_LP64 && COMPILER2 358 359 // The callee returns with the stack possibly adjusted by adapter transition 360 // We remove that possible adjustment here. 361 // All interpreter local registers are untouched. Any result is passed back 362 // in the O0/O1 or float registers. Before continuing, the arguments must be 363 // popped from the java expression stack; i.e., Lesp must be adjusted. 364 365 __ mov(Llast_SP, SP); // Remove any adapter added stack space. 366 367 const Register cache = G3_scratch; 368 const Register index = G1_scratch; 369 __ get_cache_and_index_at_bcp(cache, index, 1, index_size); 370 371 const Register flags = cache; 372 __ ld_ptr(cache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset(), flags); 373 const Register parameter_size = flags; 374 __ and3(flags, ConstantPoolCacheEntry::parameter_size_mask, parameter_size); // argument size in words 375 __ sll(parameter_size, Interpreter::logStackElementSize, parameter_size); // each argument size in bytes 376 __ add(Lesp, parameter_size, Lesp); // pop arguments 377 __ dispatch_next(state, step); 378 379 return entry; 380 } 381 382 383 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) { 384 address entry = __ pc(); 385 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache 386 #if INCLUDE_JVMCI 387 // Check if we need to take lock at entry of synchronized method. 388 if (UseJVMCICompiler) { 389 Label L; 390 Address pending_monitor_enter_addr(G2_thread, JavaThread::pending_monitorenter_offset()); 391 __ ldbool(pending_monitor_enter_addr, Gtemp); // Load if pending monitor enter 392 __ cmp_and_br_short(Gtemp, G0, Assembler::equal, Assembler::pn, L); 393 // Clear flag. 394 __ stbool(G0, pending_monitor_enter_addr); 395 // Take lock. 396 lock_method(); 397 __ bind(L); 398 } 399 #endif 400 { Label L; 401 Address exception_addr(G2_thread, Thread::pending_exception_offset()); 402 __ ld_ptr(exception_addr, Gtemp); // Load pending exception. 403 __ br_null_short(Gtemp, Assembler::pt, L); 404 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception)); 405 __ should_not_reach_here(); 406 __ bind(L); 407 } 408 __ dispatch_next(state, step); 409 return entry; 410 } 411 412 // A result handler converts/unboxes a native call result into 413 // a java interpreter/compiler result. The current frame is an 414 // interpreter frame. The activation frame unwind code must be 415 // consistent with that of TemplateTable::_return(...). In the 416 // case of native methods, the caller's SP was not modified. 417 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) { 418 address entry = __ pc(); 419 Register Itos_i = Otos_i ->after_save(); 420 Register Itos_l = Otos_l ->after_save(); 421 Register Itos_l1 = Otos_l1->after_save(); 422 Register Itos_l2 = Otos_l2->after_save(); 423 switch (type) { 424 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false 425 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value! 426 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break; 427 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break; 428 case T_LONG : 429 #ifndef _LP64 430 __ mov(O1, Itos_l2); // move other half of long 431 #endif // ifdef or no ifdef, fall through to the T_INT case 432 case T_INT : __ mov(O0, Itos_i); break; 433 case T_VOID : /* nothing to do */ break; 434 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break; 435 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break; 436 case T_OBJECT : 437 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i); 438 __ verify_oop(Itos_i); 439 break; 440 default : ShouldNotReachHere(); 441 } 442 __ ret(); // return from interpreter activation 443 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame 444 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly 445 return entry; 446 } 447 448 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) { 449 address entry = __ pc(); 450 __ push(state); 451 __ call_VM(noreg, runtime_entry); 452 __ dispatch_via(vtos, Interpreter::normal_table(vtos)); 453 return entry; 454 } 455 456 457 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) { 458 address entry = __ pc(); 459 __ dispatch_next(state); 460 return entry; 461 } 462 463 // 464 // Helpers for commoning out cases in the various type of method entries. 465 // 466 467 // increment invocation count & check for overflow 468 // 469 // Note: checking for negative value instead of overflow 470 // so we have a 'sticky' overflow test 471 // 472 // Lmethod: method 473 // ??: invocation counter 474 // 475 void TemplateInterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) { 476 // Note: In tiered we increment either counters in MethodCounters* or in 477 // MDO depending if we're profiling or not. 478 const Register G3_method_counters = G3_scratch; 479 Label done; 480 481 if (TieredCompilation) { 482 const int increment = InvocationCounter::count_increment; 483 Label no_mdo; 484 if (ProfileInterpreter) { 485 // If no method data exists, go to profile_continue. 486 __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch); 487 __ br_null_short(G4_scratch, Assembler::pn, no_mdo); 488 // Increment counter 489 Address mdo_invocation_counter(G4_scratch, 490 in_bytes(MethodData::invocation_counter_offset()) + 491 in_bytes(InvocationCounter::counter_offset())); 492 Address mask(G4_scratch, in_bytes(MethodData::invoke_mask_offset())); 493 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask, 494 G3_scratch, Lscratch, 495 Assembler::zero, overflow); 496 __ ba_short(done); 497 } 498 499 // Increment counter in MethodCounters* 500 __ bind(no_mdo); 501 Address invocation_counter(G3_method_counters, 502 in_bytes(MethodCounters::invocation_counter_offset()) + 503 in_bytes(InvocationCounter::counter_offset())); 504 __ get_method_counters(Lmethod, G3_method_counters, done); 505 Address mask(G3_method_counters, in_bytes(MethodCounters::invoke_mask_offset())); 506 __ increment_mask_and_jump(invocation_counter, increment, mask, 507 G4_scratch, Lscratch, 508 Assembler::zero, overflow); 509 __ bind(done); 510 } else { // not TieredCompilation 511 // Update standard invocation counters 512 __ get_method_counters(Lmethod, G3_method_counters, done); 513 __ increment_invocation_counter(G3_method_counters, O0, G4_scratch); 514 if (ProfileInterpreter) { 515 Address interpreter_invocation_counter(G3_method_counters, 516 in_bytes(MethodCounters::interpreter_invocation_counter_offset())); 517 __ ld(interpreter_invocation_counter, G4_scratch); 518 __ inc(G4_scratch); 519 __ st(G4_scratch, interpreter_invocation_counter); 520 } 521 522 if (ProfileInterpreter && profile_method != NULL) { 523 // Test to see if we should create a method data oop 524 Address profile_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_profile_limit_offset())); 525 __ ld(profile_limit, G1_scratch); 526 __ cmp_and_br_short(O0, G1_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue); 527 528 // if no method data exists, go to profile_method 529 __ test_method_data_pointer(*profile_method); 530 } 531 532 Address invocation_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_invocation_limit_offset())); 533 __ ld(invocation_limit, G3_scratch); 534 __ cmp(O0, G3_scratch); 535 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance 536 __ delayed()->nop(); 537 __ bind(done); 538 } 539 540 } 541 542 // Allocate monitor and lock method (asm interpreter) 543 // ebx - Method* 544 // 545 void TemplateInterpreterGenerator::lock_method() { 546 __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0); // Load access flags. 547 548 #ifdef ASSERT 549 { Label ok; 550 __ btst(JVM_ACC_SYNCHRONIZED, O0); 551 __ br( Assembler::notZero, false, Assembler::pt, ok); 552 __ delayed()->nop(); 553 __ stop("method doesn't need synchronization"); 554 __ bind(ok); 555 } 556 #endif // ASSERT 557 558 // get synchronization object to O0 559 { Label done; 560 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 561 __ btst(JVM_ACC_STATIC, O0); 562 __ br( Assembler::zero, true, Assembler::pt, done); 563 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case 564 565 __ ld_ptr( Lmethod, in_bytes(Method::const_offset()), O0); 566 __ ld_ptr( O0, in_bytes(ConstMethod::constants_offset()), O0); 567 __ ld_ptr( O0, ConstantPool::pool_holder_offset_in_bytes(), O0); 568 569 // lock the mirror, not the Klass* 570 __ ld_ptr( O0, mirror_offset, O0); 571 572 #ifdef ASSERT 573 __ tst(O0); 574 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc); 575 #endif // ASSERT 576 577 __ bind(done); 578 } 579 580 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem 581 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object 582 // __ untested("lock_object from method entry"); 583 __ lock_object(Lmonitors, O0); 584 } 585 586 587 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size, 588 Register Rscratch, 589 Register Rscratch2) { 590 const int page_size = os::vm_page_size(); 591 Label after_frame_check; 592 593 assert_different_registers(Rframe_size, Rscratch, Rscratch2); 594 595 __ set(page_size, Rscratch); 596 __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check); 597 598 // get the stack base, and in debug, verify it is non-zero 599 __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch ); 600 #ifdef ASSERT 601 Label base_not_zero; 602 __ br_notnull_short(Rscratch, Assembler::pn, base_not_zero); 603 __ stop("stack base is zero in generate_stack_overflow_check"); 604 __ bind(base_not_zero); 605 #endif 606 607 // get the stack size, and in debug, verify it is non-zero 608 assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" ); 609 __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 ); 610 #ifdef ASSERT 611 Label size_not_zero; 612 __ br_notnull_short(Rscratch2, Assembler::pn, size_not_zero); 613 __ stop("stack size is zero in generate_stack_overflow_check"); 614 __ bind(size_not_zero); 615 #endif 616 617 // compute the beginning of the protected zone minus the requested frame size 618 __ sub( Rscratch, Rscratch2, Rscratch ); 619 __ set( JavaThread::stack_red_zone_size() + JavaThread::stack_yellow_zone_size(), Rscratch2 ); 620 __ add( Rscratch, Rscratch2, Rscratch ); 621 622 // Add in the size of the frame (which is the same as subtracting it from the 623 // SP, which would take another register 624 __ add( Rscratch, Rframe_size, Rscratch ); 625 626 // the frame is greater than one page in size, so check against 627 // the bottom of the stack 628 __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check); 629 630 // the stack will overflow, throw an exception 631 632 // Note that SP is restored to sender's sp (in the delay slot). This 633 // is necessary if the sender's frame is an extended compiled frame 634 // (see gen_c2i_adapter()) and safer anyway in case of JSR292 635 // adaptations. 636 637 // Note also that the restored frame is not necessarily interpreted. 638 // Use the shared runtime version of the StackOverflowError. 639 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated"); 640 AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry()); 641 __ jump_to(stub, Rscratch); 642 __ delayed()->mov(O5_savedSP, SP); 643 644 // if you get to here, then there is enough stack space 645 __ bind( after_frame_check ); 646 } 647 648 649 // 650 // Generate a fixed interpreter frame. This is identical setup for interpreted 651 // methods and for native methods hence the shared code. 652 653 654 //---------------------------------------------------------------------------------------------------- 655 // Stack frame layout 656 // 657 // When control flow reaches any of the entry types for the interpreter 658 // the following holds -> 659 // 660 // C2 Calling Conventions: 661 // 662 // The entry code below assumes that the following registers are set 663 // when coming in: 664 // G5_method: holds the Method* of the method to call 665 // Lesp: points to the TOS of the callers expression stack 666 // after having pushed all the parameters 667 // 668 // The entry code does the following to setup an interpreter frame 669 // pop parameters from the callers stack by adjusting Lesp 670 // set O0 to Lesp 671 // compute X = (max_locals - num_parameters) 672 // bump SP up by X to accomadate the extra locals 673 // compute X = max_expression_stack 674 // + vm_local_words 675 // + 16 words of register save area 676 // save frame doing a save sp, -X, sp growing towards lower addresses 677 // set Lbcp, Lmethod, LcpoolCache 678 // set Llocals to i0 679 // set Lmonitors to FP - rounded_vm_local_words 680 // set Lesp to Lmonitors - 4 681 // 682 // The frame has now been setup to do the rest of the entry code 683 684 // Try this optimization: Most method entries could live in a 685 // "one size fits all" stack frame without all the dynamic size 686 // calculations. It might be profitable to do all this calculation 687 // statically and approximately for "small enough" methods. 688 689 //----------------------------------------------------------------------------------------------- 690 691 // C1 Calling conventions 692 // 693 // Upon method entry, the following registers are setup: 694 // 695 // g2 G2_thread: current thread 696 // g5 G5_method: method to activate 697 // g4 Gargs : pointer to last argument 698 // 699 // 700 // Stack: 701 // 702 // +---------------+ <--- sp 703 // | | 704 // : reg save area : 705 // | | 706 // +---------------+ <--- sp + 0x40 707 // | | 708 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later) 709 // | | 710 // +---------------+ <--- sp + 0x5c 711 // | | 712 // : free : 713 // | | 714 // +---------------+ <--- Gargs 715 // | | 716 // : arguments : 717 // | | 718 // +---------------+ 719 // | | 720 // 721 // 722 // 723 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like: 724 // 725 // +---------------+ <--- sp 726 // | | 727 // : reg save area : 728 // | | 729 // +---------------+ <--- sp + 0x40 730 // | | 731 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later) 732 // | | 733 // +---------------+ <--- sp + 0x5c 734 // | | 735 // : : 736 // | | <--- Lesp 737 // +---------------+ <--- Lmonitors (fp - 0x18) 738 // | VM locals | 739 // +---------------+ <--- fp 740 // | | 741 // : reg save area : 742 // | | 743 // +---------------+ <--- fp + 0x40 744 // | | 745 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later) 746 // | | 747 // +---------------+ <--- fp + 0x5c 748 // | | 749 // : free : 750 // | | 751 // +---------------+ 752 // | | 753 // : nonarg locals : 754 // | | 755 // +---------------+ 756 // | | 757 // : arguments : 758 // | | <--- Llocals 759 // +---------------+ <--- Gargs 760 // | | 761 762 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) { 763 // 764 // 765 // The entry code sets up a new interpreter frame in 4 steps: 766 // 767 // 1) Increase caller's SP by for the extra local space needed: 768 // (check for overflow) 769 // Efficient implementation of xload/xstore bytecodes requires 770 // that arguments and non-argument locals are in a contigously 771 // addressable memory block => non-argument locals must be 772 // allocated in the caller's frame. 773 // 774 // 2) Create a new stack frame and register window: 775 // The new stack frame must provide space for the standard 776 // register save area, the maximum java expression stack size, 777 // the monitor slots (0 slots initially), and some frame local 778 // scratch locations. 779 // 780 // 3) The following interpreter activation registers must be setup: 781 // Lesp : expression stack pointer 782 // Lbcp : bytecode pointer 783 // Lmethod : method 784 // Llocals : locals pointer 785 // Lmonitors : monitor pointer 786 // LcpoolCache: constant pool cache 787 // 788 // 4) Initialize the non-argument locals if necessary: 789 // Non-argument locals may need to be initialized to NULL 790 // for GC to work. If the oop-map information is accurate 791 // (in the absence of the JSR problem), no initialization 792 // is necessary. 793 // 794 // (gri - 2/25/2000) 795 796 797 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong ); 798 799 const int extra_space = 800 rounded_vm_local_words + // frame local scratch space 801 Method::extra_stack_entries() + // extra stack for jsr 292 802 frame::memory_parameter_word_sp_offset + // register save area 803 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0); 804 805 const Register Glocals_size = G3; 806 const Register RconstMethod = Glocals_size; 807 const Register Otmp1 = O3; 808 const Register Otmp2 = O4; 809 // Lscratch can't be used as a temporary because the call_stub uses 810 // it to assert that the stack frame was setup correctly. 811 const Address constMethod (G5_method, Method::const_offset()); 812 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset()); 813 814 __ ld_ptr( constMethod, RconstMethod ); 815 __ lduh( size_of_parameters, Glocals_size); 816 817 // Gargs points to first local + BytesPerWord 818 // Set the saved SP after the register window save 819 // 820 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP); 821 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1); 822 __ add(Gargs, Otmp1, Gargs); 823 824 if (native_call) { 825 __ calc_mem_param_words( Glocals_size, Gframe_size ); 826 __ add( Gframe_size, extra_space, Gframe_size); 827 __ round_to( Gframe_size, WordsPerLong ); 828 __ sll( Gframe_size, LogBytesPerWord, Gframe_size ); 829 } else { 830 831 // 832 // Compute number of locals in method apart from incoming parameters 833 // 834 const Address size_of_locals (Otmp1, ConstMethod::size_of_locals_offset()); 835 __ ld_ptr( constMethod, Otmp1 ); 836 __ lduh( size_of_locals, Otmp1 ); 837 __ sub( Otmp1, Glocals_size, Glocals_size ); 838 __ round_to( Glocals_size, WordsPerLong ); 839 __ sll( Glocals_size, Interpreter::logStackElementSize, Glocals_size ); 840 841 // see if the frame is greater than one page in size. If so, 842 // then we need to verify there is enough stack space remaining 843 // Frame_size = (max_stack + extra_space) * BytesPerWord; 844 __ ld_ptr( constMethod, Gframe_size ); 845 __ lduh( Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size ); 846 __ add( Gframe_size, extra_space, Gframe_size ); 847 __ round_to( Gframe_size, WordsPerLong ); 848 __ sll( Gframe_size, Interpreter::logStackElementSize, Gframe_size); 849 850 // Add in java locals size for stack overflow check only 851 __ add( Gframe_size, Glocals_size, Gframe_size ); 852 853 const Register Otmp2 = O4; 854 assert_different_registers(Otmp1, Otmp2, O5_savedSP); 855 generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2); 856 857 __ sub( Gframe_size, Glocals_size, Gframe_size); 858 859 // 860 // bump SP to accomodate the extra locals 861 // 862 __ sub( SP, Glocals_size, SP ); 863 } 864 865 // 866 // now set up a stack frame with the size computed above 867 // 868 __ neg( Gframe_size ); 869 __ save( SP, Gframe_size, SP ); 870 871 // 872 // now set up all the local cache registers 873 // 874 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note 875 // that all present references to Lbyte_code initialize the register 876 // immediately before use 877 if (native_call) { 878 __ mov(G0, Lbcp); 879 } else { 880 __ ld_ptr(G5_method, Method::const_offset(), Lbcp); 881 __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp); 882 } 883 __ mov( G5_method, Lmethod); // set Lmethod 884 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache 885 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors 886 #ifdef _LP64 887 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias 888 #endif 889 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp 890 891 // setup interpreter activation registers 892 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals 893 894 if (ProfileInterpreter) { 895 #ifdef FAST_DISPATCH 896 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since 897 // they both use I2. 898 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive"); 899 #endif // FAST_DISPATCH 900 __ set_method_data_pointer(); 901 } 902 903 } 904 905 // Method entry for java.lang.ref.Reference.get. 906 address TemplateInterpreterGenerator::generate_Reference_get_entry(void) { 907 #if INCLUDE_ALL_GCS 908 // Code: _aload_0, _getfield, _areturn 909 // parameter size = 1 910 // 911 // The code that gets generated by this routine is split into 2 parts: 912 // 1. The "intrinsified" code for G1 (or any SATB based GC), 913 // 2. The slow path - which is an expansion of the regular method entry. 914 // 915 // Notes:- 916 // * In the G1 code we do not check whether we need to block for 917 // a safepoint. If G1 is enabled then we must execute the specialized 918 // code for Reference.get (except when the Reference object is null) 919 // so that we can log the value in the referent field with an SATB 920 // update buffer. 921 // If the code for the getfield template is modified so that the 922 // G1 pre-barrier code is executed when the current method is 923 // Reference.get() then going through the normal method entry 924 // will be fine. 925 // * The G1 code can, however, check the receiver object (the instance 926 // of java.lang.Reference) and jump to the slow path if null. If the 927 // Reference object is null then we obviously cannot fetch the referent 928 // and so we don't need to call the G1 pre-barrier. Thus we can use the 929 // regular method entry code to generate the NPE. 930 // 931 // This code is based on generate_accessor_enty. 932 933 address entry = __ pc(); 934 935 const int referent_offset = java_lang_ref_Reference::referent_offset; 936 guarantee(referent_offset > 0, "referent offset not initialized"); 937 938 if (UseG1GC) { 939 Label slow_path; 940 941 // In the G1 code we don't check if we need to reach a safepoint. We 942 // continue and the thread will safepoint at the next bytecode dispatch. 943 944 // Check if local 0 != NULL 945 // If the receiver is null then it is OK to jump to the slow path. 946 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0 947 // check if local 0 == NULL and go the slow path 948 __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path); 949 950 951 // Load the value of the referent field. 952 if (Assembler::is_simm13(referent_offset)) { 953 __ load_heap_oop(Otos_i, referent_offset, Otos_i); 954 } else { 955 __ set(referent_offset, G3_scratch); 956 __ load_heap_oop(Otos_i, G3_scratch, Otos_i); 957 } 958 959 // Generate the G1 pre-barrier code to log the value of 960 // the referent field in an SATB buffer. Note with 961 // these parameters the pre-barrier does not generate 962 // the load of the previous value 963 964 __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */, 965 Otos_i /* pre_val */, 966 G3_scratch /* tmp */, 967 true /* preserve_o_regs */); 968 969 // _areturn 970 __ retl(); // return from leaf routine 971 __ delayed()->mov(O5_savedSP, SP); 972 973 // Generate regular method entry 974 __ bind(slow_path); 975 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals)); 976 return entry; 977 } 978 #endif // INCLUDE_ALL_GCS 979 980 // If G1 is not enabled then attempt to go through the accessor entry point 981 // Reference.get is an accessor 982 return NULL; 983 } 984 985 /** 986 * Method entry for static native methods: 987 * int java.util.zip.CRC32.update(int crc, int b) 988 */ 989 address TemplateInterpreterGenerator::generate_CRC32_update_entry() { 990 991 if (UseCRC32Intrinsics) { 992 address entry = __ pc(); 993 994 Label L_slow_path; 995 // If we need a safepoint check, generate full interpreter entry. 996 ExternalAddress state(SafepointSynchronize::address_of_state()); 997 __ set(ExternalAddress(SafepointSynchronize::address_of_state()), O2); 998 __ set(SafepointSynchronize::_not_synchronized, O3); 999 __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pt, L_slow_path); 1000 1001 // Load parameters 1002 const Register crc = O0; // initial crc 1003 const Register val = O1; // byte to update with 1004 const Register table = O2; // address of 256-entry lookup table 1005 1006 __ ldub(Gargs, 3, val); 1007 __ lduw(Gargs, 8, crc); 1008 1009 __ set(ExternalAddress(StubRoutines::crc_table_addr()), table); 1010 1011 __ not1(crc); // ~crc 1012 __ clruwu(crc); 1013 __ update_byte_crc32(crc, val, table); 1014 __ not1(crc); // ~crc 1015 1016 // result in O0 1017 __ retl(); 1018 __ delayed()->nop(); 1019 1020 // generate a vanilla native entry as the slow path 1021 __ bind(L_slow_path); 1022 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native)); 1023 return entry; 1024 } 1025 return NULL; 1026 } 1027 1028 /** 1029 * Method entry for static native methods: 1030 * int java.util.zip.CRC32.updateBytes(int crc, byte[] b, int off, int len) 1031 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 1032 */ 1033 address TemplateInterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractInterpreter::MethodKind kind) { 1034 1035 if (UseCRC32Intrinsics) { 1036 address entry = __ pc(); 1037 1038 Label L_slow_path; 1039 // If we need a safepoint check, generate full interpreter entry. 1040 ExternalAddress state(SafepointSynchronize::address_of_state()); 1041 __ set(ExternalAddress(SafepointSynchronize::address_of_state()), O2); 1042 __ set(SafepointSynchronize::_not_synchronized, O3); 1043 __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pt, L_slow_path); 1044 1045 // Load parameters from the stack 1046 const Register crc = O0; // initial crc 1047 const Register buf = O1; // source java byte array address 1048 const Register len = O2; // len 1049 const Register offset = O3; // offset 1050 1051 // Arguments are reversed on java expression stack 1052 // Calculate address of start element 1053 if (kind == Interpreter::java_util_zip_CRC32_updateByteBuffer) { 1054 __ lduw(Gargs, 0, len); 1055 __ lduw(Gargs, 8, offset); 1056 __ ldx( Gargs, 16, buf); 1057 __ lduw(Gargs, 32, crc); 1058 __ add(buf, offset, buf); 1059 } else { 1060 __ lduw(Gargs, 0, len); 1061 __ lduw(Gargs, 8, offset); 1062 __ ldx( Gargs, 16, buf); 1063 __ lduw(Gargs, 24, crc); 1064 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size 1065 __ add(buf ,offset, buf); 1066 } 1067 1068 // Call the crc32 kernel 1069 __ MacroAssembler::save_thread(L7_thread_cache); 1070 __ kernel_crc32(crc, buf, len, O3); 1071 __ MacroAssembler::restore_thread(L7_thread_cache); 1072 1073 // result in O0 1074 __ retl(); 1075 __ delayed()->nop(); 1076 1077 // generate a vanilla native entry as the slow path 1078 __ bind(L_slow_path); 1079 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native)); 1080 return entry; 1081 } 1082 return NULL; 1083 } 1084 1085 // Not supported 1086 address TemplateInterpreterGenerator::generate_CRC32C_updateBytes_entry(AbstractInterpreter::MethodKind kind) { 1087 return NULL; 1088 } 1089 1090 // Not supported 1091 address TemplateInterpreterGenerator::generate_math_entry(AbstractInterpreter::MethodKind kind) { 1092 return NULL; 1093 } 1094 // 1095 // Interpreter stub for calling a native method. (asm interpreter) 1096 // This sets up a somewhat different looking stack for calling the native method 1097 // than the typical interpreter frame setup. 1098 // 1099 1100 address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) { 1101 address entry = __ pc(); 1102 1103 // the following temporary registers are used during frame creation 1104 const Register Gtmp1 = G3_scratch ; 1105 const Register Gtmp2 = G1_scratch; 1106 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods; 1107 1108 // make sure registers are different! 1109 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2); 1110 1111 const Address Laccess_flags(Lmethod, Method::access_flags_offset()); 1112 1113 const Register Glocals_size = G3; 1114 assert_different_registers(Glocals_size, G4_scratch, Gframe_size); 1115 1116 // make sure method is native & not abstract 1117 // rethink these assertions - they can be simplified and shared (gri 2/25/2000) 1118 #ifdef ASSERT 1119 __ ld(G5_method, Method::access_flags_offset(), Gtmp1); 1120 { 1121 Label L; 1122 __ btst(JVM_ACC_NATIVE, Gtmp1); 1123 __ br(Assembler::notZero, false, Assembler::pt, L); 1124 __ delayed()->nop(); 1125 __ stop("tried to execute non-native method as native"); 1126 __ bind(L); 1127 } 1128 { Label L; 1129 __ btst(JVM_ACC_ABSTRACT, Gtmp1); 1130 __ br(Assembler::zero, false, Assembler::pt, L); 1131 __ delayed()->nop(); 1132 __ stop("tried to execute abstract method as non-abstract"); 1133 __ bind(L); 1134 } 1135 #endif // ASSERT 1136 1137 // generate the code to allocate the interpreter stack frame 1138 generate_fixed_frame(true); 1139 1140 // 1141 // No locals to initialize for native method 1142 // 1143 1144 // this slot will be set later, we initialize it to null here just in 1145 // case we get a GC before the actual value is stored later 1146 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS); 1147 1148 const Address do_not_unlock_if_synchronized(G2_thread, 1149 JavaThread::do_not_unlock_if_synchronized_offset()); 1150 // Since at this point in the method invocation the exception handler 1151 // would try to exit the monitor of synchronized methods which hasn't 1152 // been entered yet, we set the thread local variable 1153 // _do_not_unlock_if_synchronized to true. If any exception was thrown by 1154 // runtime, exception handling i.e. unlock_if_synchronized_method will 1155 // check this thread local flag. 1156 // This flag has two effects, one is to force an unwind in the topmost 1157 // interpreter frame and not perform an unlock while doing so. 1158 1159 __ movbool(true, G3_scratch); 1160 __ stbool(G3_scratch, do_not_unlock_if_synchronized); 1161 1162 // increment invocation counter and check for overflow 1163 // 1164 // Note: checking for negative value instead of overflow 1165 // so we have a 'sticky' overflow test (may be of 1166 // importance as soon as we have true MT/MP) 1167 Label invocation_counter_overflow; 1168 Label Lcontinue; 1169 if (inc_counter) { 1170 generate_counter_incr(&invocation_counter_overflow, NULL, NULL); 1171 1172 } 1173 __ bind(Lcontinue); 1174 1175 bang_stack_shadow_pages(true); 1176 1177 // reset the _do_not_unlock_if_synchronized flag 1178 __ stbool(G0, do_not_unlock_if_synchronized); 1179 1180 // check for synchronized methods 1181 // Must happen AFTER invocation_counter check and stack overflow check, 1182 // so method is not locked if overflows. 1183 1184 if (synchronized) { 1185 lock_method(); 1186 } else { 1187 #ifdef ASSERT 1188 { Label ok; 1189 __ ld(Laccess_flags, O0); 1190 __ btst(JVM_ACC_SYNCHRONIZED, O0); 1191 __ br( Assembler::zero, false, Assembler::pt, ok); 1192 __ delayed()->nop(); 1193 __ stop("method needs synchronization"); 1194 __ bind(ok); 1195 } 1196 #endif // ASSERT 1197 } 1198 1199 1200 // start execution 1201 __ verify_thread(); 1202 1203 // JVMTI support 1204 __ notify_method_entry(); 1205 1206 // native call 1207 1208 // (note that O0 is never an oop--at most it is a handle) 1209 // It is important not to smash any handles created by this call, 1210 // until any oop handle in O0 is dereferenced. 1211 1212 // (note that the space for outgoing params is preallocated) 1213 1214 // get signature handler 1215 { Label L; 1216 Address signature_handler(Lmethod, Method::signature_handler_offset()); 1217 __ ld_ptr(signature_handler, G3_scratch); 1218 __ br_notnull_short(G3_scratch, Assembler::pt, L); 1219 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod); 1220 __ ld_ptr(signature_handler, G3_scratch); 1221 __ bind(L); 1222 } 1223 1224 // Push a new frame so that the args will really be stored in 1225 // Copy a few locals across so the new frame has the variables 1226 // we need but these values will be dead at the jni call and 1227 // therefore not gc volatile like the values in the current 1228 // frame (Lmethod in particular) 1229 1230 // Flush the method pointer to the register save area 1231 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS); 1232 __ mov(Llocals, O1); 1233 1234 // calculate where the mirror handle body is allocated in the interpreter frame: 1235 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2); 1236 1237 // Calculate current frame size 1238 __ sub(SP, FP, O3); // Calculate negative of current frame size 1239 __ save(SP, O3, SP); // Allocate an identical sized frame 1240 1241 // Note I7 has leftover trash. Slow signature handler will fill it in 1242 // should we get there. Normal jni call will set reasonable last_Java_pc 1243 // below (and fix I7 so the stack trace doesn't have a meaningless frame 1244 // in it). 1245 1246 // Load interpreter frame's Lmethod into same register here 1247 1248 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod); 1249 1250 __ mov(I1, Llocals); 1251 __ mov(I2, Lscratch2); // save the address of the mirror 1252 1253 1254 // ONLY Lmethod and Llocals are valid here! 1255 1256 // call signature handler, It will move the arg properly since Llocals in current frame 1257 // matches that in outer frame 1258 1259 __ callr(G3_scratch, 0); 1260 __ delayed()->nop(); 1261 1262 // Result handler is in Lscratch 1263 1264 // Reload interpreter frame's Lmethod since slow signature handler may block 1265 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod); 1266 1267 { Label not_static; 1268 1269 __ ld(Laccess_flags, O0); 1270 __ btst(JVM_ACC_STATIC, O0); 1271 __ br( Assembler::zero, false, Assembler::pt, not_static); 1272 // get native function entry point(O0 is a good temp until the very end) 1273 __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0); 1274 // for static methods insert the mirror argument 1275 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 1276 1277 __ ld_ptr(Lmethod, Method:: const_offset(), O1); 1278 __ ld_ptr(O1, ConstMethod::constants_offset(), O1); 1279 __ ld_ptr(O1, ConstantPool::pool_holder_offset_in_bytes(), O1); 1280 __ ld_ptr(O1, mirror_offset, O1); 1281 #ifdef ASSERT 1282 if (!PrintSignatureHandlers) // do not dirty the output with this 1283 { Label L; 1284 __ br_notnull_short(O1, Assembler::pt, L); 1285 __ stop("mirror is missing"); 1286 __ bind(L); 1287 } 1288 #endif // ASSERT 1289 __ st_ptr(O1, Lscratch2, 0); 1290 __ mov(Lscratch2, O1); 1291 __ bind(not_static); 1292 } 1293 1294 // At this point, arguments have been copied off of stack into 1295 // their JNI positions, which are O1..O5 and SP[68..]. 1296 // Oops are boxed in-place on the stack, with handles copied to arguments. 1297 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*. 1298 1299 #ifdef ASSERT 1300 { Label L; 1301 __ br_notnull_short(O0, Assembler::pt, L); 1302 __ stop("native entry point is missing"); 1303 __ bind(L); 1304 } 1305 #endif // ASSERT 1306 1307 // 1308 // setup the frame anchor 1309 // 1310 // The scavenge function only needs to know that the PC of this frame is 1311 // in the interpreter method entry code, it doesn't need to know the exact 1312 // PC and hence we can use O7 which points to the return address from the 1313 // previous call in the code stream (signature handler function) 1314 // 1315 // The other trick is we set last_Java_sp to FP instead of the usual SP because 1316 // we have pushed the extra frame in order to protect the volatile register(s) 1317 // in that frame when we return from the jni call 1318 // 1319 1320 __ set_last_Java_frame(FP, O7); 1321 __ mov(O7, I7); // make dummy interpreter frame look like one above, 1322 // not meaningless information that'll confuse me. 1323 1324 // flush the windows now. We don't care about the current (protection) frame 1325 // only the outer frames 1326 1327 __ flushw(); 1328 1329 // mark windows as flushed 1330 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset()); 1331 __ set(JavaFrameAnchor::flushed, G3_scratch); 1332 __ st(G3_scratch, flags); 1333 1334 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready. 1335 1336 Address thread_state(G2_thread, JavaThread::thread_state_offset()); 1337 #ifdef ASSERT 1338 { Label L; 1339 __ ld(thread_state, G3_scratch); 1340 __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L); 1341 __ stop("Wrong thread state in native stub"); 1342 __ bind(L); 1343 } 1344 #endif // ASSERT 1345 __ set(_thread_in_native, G3_scratch); 1346 __ st(G3_scratch, thread_state); 1347 1348 // Call the jni method, using the delay slot to set the JNIEnv* argument. 1349 __ save_thread(L7_thread_cache); // save Gthread 1350 __ callr(O0, 0); 1351 __ delayed()-> 1352 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0); 1353 1354 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD 1355 1356 __ restore_thread(L7_thread_cache); // restore G2_thread 1357 __ reinit_heapbase(); 1358 1359 // must we block? 1360 1361 // Block, if necessary, before resuming in _thread_in_Java state. 1362 // In order for GC to work, don't clear the last_Java_sp until after blocking. 1363 { Label no_block; 1364 AddressLiteral sync_state(SafepointSynchronize::address_of_state()); 1365 1366 // Switch thread to "native transition" state before reading the synchronization state. 1367 // This additional state is necessary because reading and testing the synchronization 1368 // state is not atomic w.r.t. GC, as this scenario demonstrates: 1369 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 1370 // VM thread changes sync state to synchronizing and suspends threads for GC. 1371 // Thread A is resumed to finish this native method, but doesn't block here since it 1372 // didn't see any synchronization is progress, and escapes. 1373 __ set(_thread_in_native_trans, G3_scratch); 1374 __ st(G3_scratch, thread_state); 1375 if(os::is_MP()) { 1376 if (UseMembar) { 1377 // Force this write out before the read below 1378 __ membar(Assembler::StoreLoad); 1379 } else { 1380 // Write serialization page so VM thread can do a pseudo remote membar. 1381 // We use the current thread pointer to calculate a thread specific 1382 // offset to write to within the page. This minimizes bus traffic 1383 // due to cache line collision. 1384 __ serialize_memory(G2_thread, G1_scratch, G3_scratch); 1385 } 1386 } 1387 __ load_contents(sync_state, G3_scratch); 1388 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized); 1389 1390 Label L; 1391 __ br(Assembler::notEqual, false, Assembler::pn, L); 1392 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch); 1393 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block); 1394 __ bind(L); 1395 1396 // Block. Save any potential method result value before the operation and 1397 // use a leaf call to leave the last_Java_frame setup undisturbed. 1398 save_native_result(); 1399 __ call_VM_leaf(L7_thread_cache, 1400 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans), 1401 G2_thread); 1402 1403 // Restore any method result value 1404 restore_native_result(); 1405 __ bind(no_block); 1406 } 1407 1408 // Clear the frame anchor now 1409 1410 __ reset_last_Java_frame(); 1411 1412 // Move the result handler address 1413 __ mov(Lscratch, G3_scratch); 1414 // return possible result to the outer frame 1415 #ifndef __LP64 1416 __ mov(O0, I0); 1417 __ restore(O1, G0, O1); 1418 #else 1419 __ restore(O0, G0, O0); 1420 #endif /* __LP64 */ 1421 1422 // Move result handler to expected register 1423 __ mov(G3_scratch, Lscratch); 1424 1425 // Back in normal (native) interpreter frame. State is thread_in_native_trans 1426 // switch to thread_in_Java. 1427 1428 __ set(_thread_in_Java, G3_scratch); 1429 __ st(G3_scratch, thread_state); 1430 1431 // reset handle block 1432 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch); 1433 __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes()); 1434 1435 // If we have an oop result store it where it will be safe for any further gc 1436 // until we return now that we've released the handle it might be protected by 1437 1438 { 1439 Label no_oop, store_result; 1440 1441 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch); 1442 __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop); 1443 __ addcc(G0, O0, O0); 1444 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL: 1445 __ delayed()->ld_ptr(O0, 0, O0); // unbox it 1446 __ mov(G0, O0); 1447 1448 __ bind(store_result); 1449 // Store it where gc will look for it and result handler expects it. 1450 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS); 1451 1452 __ bind(no_oop); 1453 1454 } 1455 1456 1457 // handle exceptions (exception handling will handle unlocking!) 1458 { Label L; 1459 Address exception_addr(G2_thread, Thread::pending_exception_offset()); 1460 __ ld_ptr(exception_addr, Gtemp); 1461 __ br_null_short(Gtemp, Assembler::pt, L); 1462 // Note: This could be handled more efficiently since we know that the native 1463 // method doesn't have an exception handler. We could directly return 1464 // to the exception handler for the caller. 1465 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception)); 1466 __ should_not_reach_here(); 1467 __ bind(L); 1468 } 1469 1470 // JVMTI support (preserves thread register) 1471 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI); 1472 1473 if (synchronized) { 1474 // save and restore any potential method result value around the unlocking operation 1475 save_native_result(); 1476 1477 __ add( __ top_most_monitor(), O1); 1478 __ unlock_object(O1); 1479 1480 restore_native_result(); 1481 } 1482 1483 #if defined(COMPILER2) && !defined(_LP64) 1484 1485 // C2 expects long results in G1 we can't tell if we're returning to interpreted 1486 // or compiled so just be safe. 1487 1488 __ sllx(O0, 32, G1); // Shift bits into high G1 1489 __ srl (O1, 0, O1); // Zero extend O1 1490 __ or3 (O1, G1, G1); // OR 64 bits into G1 1491 1492 #endif /* COMPILER2 && !_LP64 */ 1493 1494 // dispose of return address and remove activation 1495 #ifdef ASSERT 1496 { 1497 Label ok; 1498 __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok); 1499 __ stop("bad I5_savedSP value"); 1500 __ should_not_reach_here(); 1501 __ bind(ok); 1502 } 1503 #endif 1504 if (TraceJumps) { 1505 // Move target to register that is recordable 1506 __ mov(Lscratch, G3_scratch); 1507 __ JMP(G3_scratch, 0); 1508 } else { 1509 __ jmp(Lscratch, 0); 1510 } 1511 __ delayed()->nop(); 1512 1513 1514 if (inc_counter) { 1515 // handle invocation counter overflow 1516 __ bind(invocation_counter_overflow); 1517 generate_counter_overflow(Lcontinue); 1518 } 1519 1520 1521 1522 return entry; 1523 } 1524 1525 1526 // Generic method entry to (asm) interpreter 1527 address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) { 1528 address entry = __ pc(); 1529 1530 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods; 1531 1532 // the following temporary registers are used during frame creation 1533 const Register Gtmp1 = G3_scratch ; 1534 const Register Gtmp2 = G1_scratch; 1535 1536 // make sure registers are different! 1537 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2); 1538 1539 const Address constMethod (G5_method, Method::const_offset()); 1540 // Seems like G5_method is live at the point this is used. So we could make this look consistent 1541 // and use in the asserts. 1542 const Address access_flags (Lmethod, Method::access_flags_offset()); 1543 1544 const Register Glocals_size = G3; 1545 assert_different_registers(Glocals_size, G4_scratch, Gframe_size); 1546 1547 // make sure method is not native & not abstract 1548 // rethink these assertions - they can be simplified and shared (gri 2/25/2000) 1549 #ifdef ASSERT 1550 __ ld(G5_method, Method::access_flags_offset(), Gtmp1); 1551 { 1552 Label L; 1553 __ btst(JVM_ACC_NATIVE, Gtmp1); 1554 __ br(Assembler::zero, false, Assembler::pt, L); 1555 __ delayed()->nop(); 1556 __ stop("tried to execute native method as non-native"); 1557 __ bind(L); 1558 } 1559 { Label L; 1560 __ btst(JVM_ACC_ABSTRACT, Gtmp1); 1561 __ br(Assembler::zero, false, Assembler::pt, L); 1562 __ delayed()->nop(); 1563 __ stop("tried to execute abstract method as non-abstract"); 1564 __ bind(L); 1565 } 1566 #endif // ASSERT 1567 1568 // generate the code to allocate the interpreter stack frame 1569 1570 generate_fixed_frame(false); 1571 1572 #ifdef FAST_DISPATCH 1573 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables); 1574 // set bytecode dispatch table base 1575 #endif 1576 1577 // 1578 // Code to initialize the extra (i.e. non-parm) locals 1579 // 1580 Register init_value = noreg; // will be G0 if we must clear locals 1581 // The way the code was setup before zerolocals was always true for vanilla java entries. 1582 // It could only be false for the specialized entries like accessor or empty which have 1583 // no extra locals so the testing was a waste of time and the extra locals were always 1584 // initialized. We removed this extra complication to already over complicated code. 1585 1586 init_value = G0; 1587 Label clear_loop; 1588 1589 const Register RconstMethod = O1; 1590 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset()); 1591 const Address size_of_locals (RconstMethod, ConstMethod::size_of_locals_offset()); 1592 1593 // NOTE: If you change the frame layout, this code will need to 1594 // be updated! 1595 __ ld_ptr( constMethod, RconstMethod ); 1596 __ lduh( size_of_locals, O2 ); 1597 __ lduh( size_of_parameters, O1 ); 1598 __ sll( O2, Interpreter::logStackElementSize, O2); 1599 __ sll( O1, Interpreter::logStackElementSize, O1 ); 1600 __ sub( Llocals, O2, O2 ); 1601 __ sub( Llocals, O1, O1 ); 1602 1603 __ bind( clear_loop ); 1604 __ inc( O2, wordSize ); 1605 1606 __ cmp( O2, O1 ); 1607 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop ); 1608 __ delayed()->st_ptr( init_value, O2, 0 ); 1609 1610 const Address do_not_unlock_if_synchronized(G2_thread, 1611 JavaThread::do_not_unlock_if_synchronized_offset()); 1612 // Since at this point in the method invocation the exception handler 1613 // would try to exit the monitor of synchronized methods which hasn't 1614 // been entered yet, we set the thread local variable 1615 // _do_not_unlock_if_synchronized to true. If any exception was thrown by 1616 // runtime, exception handling i.e. unlock_if_synchronized_method will 1617 // check this thread local flag. 1618 __ movbool(true, G3_scratch); 1619 __ stbool(G3_scratch, do_not_unlock_if_synchronized); 1620 1621 __ profile_parameters_type(G1_scratch, G3_scratch, G4_scratch, Lscratch); 1622 // increment invocation counter and check for overflow 1623 // 1624 // Note: checking for negative value instead of overflow 1625 // so we have a 'sticky' overflow test (may be of 1626 // importance as soon as we have true MT/MP) 1627 Label invocation_counter_overflow; 1628 Label profile_method; 1629 Label profile_method_continue; 1630 Label Lcontinue; 1631 if (inc_counter) { 1632 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue); 1633 if (ProfileInterpreter) { 1634 __ bind(profile_method_continue); 1635 } 1636 } 1637 __ bind(Lcontinue); 1638 1639 bang_stack_shadow_pages(false); 1640 1641 // reset the _do_not_unlock_if_synchronized flag 1642 __ stbool(G0, do_not_unlock_if_synchronized); 1643 1644 // check for synchronized methods 1645 // Must happen AFTER invocation_counter check and stack overflow check, 1646 // so method is not locked if overflows. 1647 1648 if (synchronized) { 1649 lock_method(); 1650 } else { 1651 #ifdef ASSERT 1652 { Label ok; 1653 __ ld(access_flags, O0); 1654 __ btst(JVM_ACC_SYNCHRONIZED, O0); 1655 __ br( Assembler::zero, false, Assembler::pt, ok); 1656 __ delayed()->nop(); 1657 __ stop("method needs synchronization"); 1658 __ bind(ok); 1659 } 1660 #endif // ASSERT 1661 } 1662 1663 // start execution 1664 1665 __ verify_thread(); 1666 1667 // jvmti support 1668 __ notify_method_entry(); 1669 1670 // start executing instructions 1671 __ dispatch_next(vtos); 1672 1673 1674 if (inc_counter) { 1675 if (ProfileInterpreter) { 1676 // We have decided to profile this method in the interpreter 1677 __ bind(profile_method); 1678 1679 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method)); 1680 __ set_method_data_pointer_for_bcp(); 1681 __ ba_short(profile_method_continue); 1682 } 1683 1684 // handle invocation counter overflow 1685 __ bind(invocation_counter_overflow); 1686 generate_counter_overflow(Lcontinue); 1687 } 1688 1689 1690 return entry; 1691 } 1692 1693 //---------------------------------------------------------------------------------------------------- 1694 // Exceptions 1695 void TemplateInterpreterGenerator::generate_throw_exception() { 1696 1697 // Entry point in previous activation (i.e., if the caller was interpreted) 1698 Interpreter::_rethrow_exception_entry = __ pc(); 1699 // O0: exception 1700 1701 // entry point for exceptions thrown within interpreter code 1702 Interpreter::_throw_exception_entry = __ pc(); 1703 __ verify_thread(); 1704 // expression stack is undefined here 1705 // O0: exception, i.e. Oexception 1706 // Lbcp: exception bcp 1707 __ verify_oop(Oexception); 1708 1709 1710 // expression stack must be empty before entering the VM in case of an exception 1711 __ empty_expression_stack(); 1712 // find exception handler address and preserve exception oop 1713 // call C routine to find handler and jump to it 1714 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception); 1715 __ push_ptr(O1); // push exception for exception handler bytecodes 1716 1717 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!) 1718 __ delayed()->nop(); 1719 1720 1721 // if the exception is not handled in the current frame 1722 // the frame is removed and the exception is rethrown 1723 // (i.e. exception continuation is _rethrow_exception) 1724 // 1725 // Note: At this point the bci is still the bxi for the instruction which caused 1726 // the exception and the expression stack is empty. Thus, for any VM calls 1727 // at this point, GC will find a legal oop map (with empty expression stack). 1728 1729 // in current activation 1730 // tos: exception 1731 // Lbcp: exception bcp 1732 1733 // 1734 // JVMTI PopFrame support 1735 // 1736 1737 Interpreter::_remove_activation_preserving_args_entry = __ pc(); 1738 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset()); 1739 // Set the popframe_processing bit in popframe_condition indicating that we are 1740 // currently handling popframe, so that call_VMs that may happen later do not trigger new 1741 // popframe handling cycles. 1742 1743 __ ld(popframe_condition_addr, G3_scratch); 1744 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch); 1745 __ stw(G3_scratch, popframe_condition_addr); 1746 1747 // Empty the expression stack, as in normal exception handling 1748 __ empty_expression_stack(); 1749 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false); 1750 1751 { 1752 // Check to see whether we are returning to a deoptimized frame. 1753 // (The PopFrame call ensures that the caller of the popped frame is 1754 // either interpreted or compiled and deoptimizes it if compiled.) 1755 // In this case, we can't call dispatch_next() after the frame is 1756 // popped, but instead must save the incoming arguments and restore 1757 // them after deoptimization has occurred. 1758 // 1759 // Note that we don't compare the return PC against the 1760 // deoptimization blob's unpack entry because of the presence of 1761 // adapter frames in C2. 1762 Label caller_not_deoptimized; 1763 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7); 1764 __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized); 1765 1766 const Register Gtmp1 = G3_scratch; 1767 const Register Gtmp2 = G1_scratch; 1768 const Register RconstMethod = Gtmp1; 1769 const Address constMethod(Lmethod, Method::const_offset()); 1770 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset()); 1771 1772 // Compute size of arguments for saving when returning to deoptimized caller 1773 __ ld_ptr(constMethod, RconstMethod); 1774 __ lduh(size_of_parameters, Gtmp1); 1775 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1); 1776 __ sub(Llocals, Gtmp1, Gtmp2); 1777 __ add(Gtmp2, wordSize, Gtmp2); 1778 // Save these arguments 1779 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2); 1780 // Inform deoptimization that it is responsible for restoring these arguments 1781 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1); 1782 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset()); 1783 __ st(Gtmp1, popframe_condition_addr); 1784 1785 // Return from the current method 1786 // The caller's SP was adjusted upon method entry to accomodate 1787 // the callee's non-argument locals. Undo that adjustment. 1788 __ ret(); 1789 __ delayed()->restore(I5_savedSP, G0, SP); 1790 1791 __ bind(caller_not_deoptimized); 1792 } 1793 1794 // Clear the popframe condition flag 1795 __ stw(G0 /* popframe_inactive */, popframe_condition_addr); 1796 1797 // Get out of the current method (how this is done depends on the particular compiler calling 1798 // convention that the interpreter currently follows) 1799 // The caller's SP was adjusted upon method entry to accomodate 1800 // the callee's non-argument locals. Undo that adjustment. 1801 __ restore(I5_savedSP, G0, SP); 1802 // The method data pointer was incremented already during 1803 // call profiling. We have to restore the mdp for the current bcp. 1804 if (ProfileInterpreter) { 1805 __ set_method_data_pointer_for_bcp(); 1806 } 1807 1808 #if INCLUDE_JVMTI 1809 { 1810 Label L_done; 1811 1812 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode 1813 __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done); 1814 1815 // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call. 1816 // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL. 1817 1818 __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp); 1819 1820 __ br_null(G1_scratch, false, Assembler::pn, L_done); 1821 __ delayed()->nop(); 1822 1823 __ st_ptr(G1_scratch, Lesp, wordSize); 1824 __ bind(L_done); 1825 } 1826 #endif // INCLUDE_JVMTI 1827 1828 // Resume bytecode interpretation at the current bcp 1829 __ dispatch_next(vtos); 1830 // end of JVMTI PopFrame support 1831 1832 Interpreter::_remove_activation_entry = __ pc(); 1833 1834 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here) 1835 __ pop_ptr(Oexception); // get exception 1836 1837 // Intel has the following comment: 1838 //// remove the activation (without doing throws on illegalMonitorExceptions) 1839 // They remove the activation without checking for bad monitor state. 1840 // %%% We should make sure this is the right semantics before implementing. 1841 1842 __ set_vm_result(Oexception); 1843 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false); 1844 1845 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI); 1846 1847 __ get_vm_result(Oexception); 1848 __ verify_oop(Oexception); 1849 1850 const int return_reg_adjustment = frame::pc_return_offset; 1851 Address issuing_pc_addr(I7, return_reg_adjustment); 1852 1853 // We are done with this activation frame; find out where to go next. 1854 // The continuation point will be an exception handler, which expects 1855 // the following registers set up: 1856 // 1857 // Oexception: exception 1858 // Oissuing_pc: the local call that threw exception 1859 // Other On: garbage 1860 // In/Ln: the contents of the caller's register window 1861 // 1862 // We do the required restore at the last possible moment, because we 1863 // need to preserve some state across a runtime call. 1864 // (Remember that the caller activation is unknown--it might not be 1865 // interpreted, so things like Lscratch are useless in the caller.) 1866 1867 // Although the Intel version uses call_C, we can use the more 1868 // compact call_VM. (The only real difference on SPARC is a 1869 // harmlessly ignored [re]set_last_Java_frame, compared with 1870 // the Intel code which lacks this.) 1871 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore 1872 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller 1873 __ super_call_VM_leaf(L7_thread_cache, 1874 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), 1875 G2_thread, Oissuing_pc->after_save()); 1876 1877 // The caller's SP was adjusted upon method entry to accomodate 1878 // the callee's non-argument locals. Undo that adjustment. 1879 __ JMP(O0, 0); // return exception handler in caller 1880 __ delayed()->restore(I5_savedSP, G0, SP); 1881 1882 // (same old exception object is already in Oexception; see above) 1883 // Note that an "issuing PC" is actually the next PC after the call 1884 } 1885 1886 1887 // 1888 // JVMTI ForceEarlyReturn support 1889 // 1890 1891 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) { 1892 address entry = __ pc(); 1893 1894 __ empty_expression_stack(); 1895 __ load_earlyret_value(state); 1896 1897 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch); 1898 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset()); 1899 1900 // Clear the earlyret state 1901 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr); 1902 1903 __ remove_activation(state, 1904 /* throw_monitor_exception */ false, 1905 /* install_monitor_exception */ false); 1906 1907 // The caller's SP was adjusted upon method entry to accomodate 1908 // the callee's non-argument locals. Undo that adjustment. 1909 __ ret(); // return to caller 1910 __ delayed()->restore(I5_savedSP, G0, SP); 1911 1912 return entry; 1913 } // end of JVMTI ForceEarlyReturn support 1914 1915 1916 //------------------------------------------------------------------------------------------------------------------------ 1917 // Helper for vtos entry point generation 1918 1919 void TemplateInterpreterGenerator::set_vtos_entry_points(Template* t, address& bep, address& cep, address& sep, address& aep, address& iep, address& lep, address& fep, address& dep, address& vep) { 1920 assert(t->is_valid() && t->tos_in() == vtos, "illegal template"); 1921 Label L; 1922 aep = __ pc(); __ push_ptr(); __ ba_short(L); 1923 fep = __ pc(); __ push_f(); __ ba_short(L); 1924 dep = __ pc(); __ push_d(); __ ba_short(L); 1925 lep = __ pc(); __ push_l(); __ ba_short(L); 1926 iep = __ pc(); __ push_i(); 1927 bep = cep = sep = iep; // there aren't any 1928 vep = __ pc(); __ bind(L); // fall through 1929 generate_and_dispatch(t); 1930 } 1931 1932 // -------------------------------------------------------------------------------- 1933 1934 // Non-product code 1935 #ifndef PRODUCT 1936 address TemplateInterpreterGenerator::generate_trace_code(TosState state) { 1937 address entry = __ pc(); 1938 1939 __ push(state); 1940 __ mov(O7, Lscratch); // protect return address within interpreter 1941 1942 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer 1943 __ mov( Otos_l2, G3_scratch ); 1944 __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch); 1945 __ mov(Lscratch, O7); // restore return address 1946 __ pop(state); 1947 __ retl(); 1948 __ delayed()->nop(); 1949 1950 return entry; 1951 } 1952 1953 1954 // helpers for generate_and_dispatch 1955 1956 void TemplateInterpreterGenerator::count_bytecode() { 1957 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch); 1958 } 1959 1960 1961 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) { 1962 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch); 1963 } 1964 1965 1966 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) { 1967 AddressLiteral index (&BytecodePairHistogram::_index); 1968 AddressLiteral counters((address) &BytecodePairHistogram::_counters); 1969 1970 // get index, shift out old bytecode, bring in new bytecode, and store it 1971 // _index = (_index >> log2_number_of_codes) | 1972 // (bytecode << log2_number_of_codes); 1973 1974 __ load_contents(index, G4_scratch); 1975 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch ); 1976 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch ); 1977 __ or3( G3_scratch, G4_scratch, G4_scratch ); 1978 __ store_contents(G4_scratch, index, G3_scratch); 1979 1980 // bump bucket contents 1981 // _counters[_index] ++; 1982 1983 __ set(counters, G3_scratch); // loads into G3_scratch 1984 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address 1985 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index 1986 __ ld (G3_scratch, 0, G4_scratch); 1987 __ inc (G4_scratch); 1988 __ st (G4_scratch, 0, G3_scratch); 1989 } 1990 1991 1992 void TemplateInterpreterGenerator::trace_bytecode(Template* t) { 1993 // Call a little run-time stub to avoid blow-up for each bytecode. 1994 // The run-time runtime saves the right registers, depending on 1995 // the tosca in-state for the given template. 1996 address entry = Interpreter::trace_code(t->tos_in()); 1997 guarantee(entry != NULL, "entry must have been generated"); 1998 __ call(entry, relocInfo::none); 1999 __ delayed()->nop(); 2000 } 2001 2002 2003 void TemplateInterpreterGenerator::stop_interpreter_at() { 2004 AddressLiteral counter(&BytecodeCounter::_counter_value); 2005 __ load_contents(counter, G3_scratch); 2006 AddressLiteral stop_at(&StopInterpreterAt); 2007 __ load_ptr_contents(stop_at, G4_scratch); 2008 __ cmp(G3_scratch, G4_scratch); 2009 __ breakpoint_trap(Assembler::equal, Assembler::icc); 2010 } 2011 #endif // not PRODUCT