1 /* 2 * Copyright (c) 1997, 2017, 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/align.hpp" 48 #include "utilities/debug.hpp" 49 #include "utilities/macros.hpp" 50 51 // Size of interpreter code. Increase if too small. Interpreter will 52 // fail with a guarantee ("not enough space for interpreter generation"); 53 // if too small. 54 // Run with +PrintInterpreter to get the VM to print out the size. 55 // Max size with JVMTI 56 // The sethi() instruction generates lots more instructions when shell 57 // stack limit is unlimited, so that's why this is much bigger. 58 int TemplateInterpreter::InterpreterCodeSize = 260 * K; 59 60 // Generation of Interpreter 61 // 62 // The TemplateInterpreterGenerator generates the interpreter into Interpreter::_code. 63 64 65 #define __ _masm-> 66 67 68 //---------------------------------------------------------------------------------------------------- 69 70 // LP64 passes floating point arguments in F1, F3, F5, etc. instead of 71 // O0, O1, O2 etc.. 72 // Doubles are passed in D0, D2, D4 73 // We store the signature of the first 16 arguments in the first argument 74 // slot because it will be overwritten prior to calling the native 75 // function, with the pointer to the JNIEnv. 76 // If LP64 there can be up to 16 floating point arguments in registers 77 // or 6 integer registers. 78 address TemplateInterpreterGenerator::generate_slow_signature_handler() { 79 80 enum { 81 non_float = 0, 82 float_sig = 1, 83 double_sig = 2, 84 sig_mask = 3 85 }; 86 87 address entry = __ pc(); 88 Argument argv(0, true); 89 90 // We are in the jni transition frame. Save the last_java_frame corresponding to the 91 // outer interpreter frame 92 // 93 __ set_last_Java_frame(FP, noreg); 94 // make sure the interpreter frame we've pushed has a valid return pc 95 __ mov(O7, I7); 96 __ mov(Lmethod, G3_scratch); 97 __ mov(Llocals, G4_scratch); 98 __ save_frame(0); 99 __ mov(G2_thread, L7_thread_cache); 100 __ add(argv.address_in_frame(), O3); 101 __ mov(G2_thread, O0); 102 __ mov(G3_scratch, O1); 103 __ call(CAST_FROM_FN_PTR(address, InterpreterRuntime::slow_signature_handler), relocInfo::runtime_call_type); 104 __ delayed()->mov(G4_scratch, O2); 105 __ mov(L7_thread_cache, G2_thread); 106 __ reset_last_Java_frame(); 107 108 109 // load the register arguments (the C code packed them as varargs) 110 Address Sig = argv.address_in_frame(); // Argument 0 holds the signature 111 __ ld_ptr( Sig, G3_scratch ); // Get register argument signature word into G3_scratch 112 __ mov( G3_scratch, G4_scratch); 113 __ srl( G4_scratch, 2, G4_scratch); // Skip Arg 0 114 Label done; 115 for (Argument ldarg = argv.successor(); ldarg.is_float_register(); ldarg = ldarg.successor()) { 116 Label NonFloatArg; 117 Label LoadFloatArg; 118 Label LoadDoubleArg; 119 Label NextArg; 120 Address a = ldarg.address_in_frame(); 121 __ andcc(G4_scratch, sig_mask, G3_scratch); 122 __ br(Assembler::zero, false, Assembler::pt, NonFloatArg); 123 __ delayed()->nop(); 124 125 __ cmp(G3_scratch, float_sig ); 126 __ br(Assembler::equal, false, Assembler::pt, LoadFloatArg); 127 __ delayed()->nop(); 128 129 __ cmp(G3_scratch, double_sig ); 130 __ br(Assembler::equal, false, Assembler::pt, LoadDoubleArg); 131 __ delayed()->nop(); 132 133 __ bind(NonFloatArg); 134 // There are only 6 integer register arguments! 135 if ( ldarg.is_register() ) 136 __ ld_ptr(ldarg.address_in_frame(), ldarg.as_register()); 137 else { 138 // Optimization, see if there are any more args and get out prior to checking 139 // all 16 float registers. My guess is that this is rare. 140 // If is_register is false, then we are done the first six integer args. 141 __ br_null_short(G4_scratch, Assembler::pt, done); 142 } 143 __ ba(NextArg); 144 __ delayed()->srl( G4_scratch, 2, G4_scratch ); 145 146 __ bind(LoadFloatArg); 147 __ ldf( FloatRegisterImpl::S, a, ldarg.as_float_register(), 4); 148 __ ba(NextArg); 149 __ delayed()->srl( G4_scratch, 2, G4_scratch ); 150 151 __ bind(LoadDoubleArg); 152 __ ldf( FloatRegisterImpl::D, a, ldarg.as_double_register() ); 153 __ ba(NextArg); 154 __ delayed()->srl( G4_scratch, 2, G4_scratch ); 155 156 __ bind(NextArg); 157 } 158 159 __ bind(done); 160 __ ret(); 161 __ delayed()->restore(O0, 0, Lscratch); // caller's Lscratch gets the result handler 162 163 return entry; 164 } 165 166 void TemplateInterpreterGenerator::generate_counter_overflow(Label& Lcontinue) { 167 168 // Generate code to initiate compilation on the counter overflow. 169 170 // InterpreterRuntime::frequency_counter_overflow takes two arguments, 171 // the first indicates if the counter overflow occurs at a backwards branch (NULL bcp) 172 // and the second is only used when the first is true. We pass zero for both. 173 // The call returns the address of the verified entry point for the method or NULL 174 // if the compilation did not complete (either went background or bailed out). 175 __ set((int)false, O2); 176 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), O2, O2, true); 177 // returns verified_entry_point or NULL 178 // we ignore it in any case 179 __ ba_short(Lcontinue); 180 } 181 182 183 // End of helpers 184 185 // Various method entries 186 187 // Abstract method entry 188 // Attempt to execute abstract method. Throw exception 189 // 190 address TemplateInterpreterGenerator::generate_abstract_entry(void) { 191 address entry = __ pc(); 192 // abstract method entry 193 // throw exception 194 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodError)); 195 // the call_VM checks for exception, so we should never return here. 196 __ should_not_reach_here(); 197 return entry; 198 } 199 200 void TemplateInterpreterGenerator::save_native_result(void) { 201 // result potentially in O0/O1: save it across calls 202 const Address& l_tmp = InterpreterMacroAssembler::l_tmp; 203 204 // result potentially in F0/F1: save it across calls 205 const Address& d_tmp = InterpreterMacroAssembler::d_tmp; 206 207 // save and restore any potential method result value around the unlocking operation 208 __ stf(FloatRegisterImpl::D, F0, d_tmp); 209 __ stx(O0, l_tmp); 210 } 211 212 void TemplateInterpreterGenerator::restore_native_result(void) { 213 const Address& l_tmp = InterpreterMacroAssembler::l_tmp; 214 const Address& d_tmp = InterpreterMacroAssembler::d_tmp; 215 216 // Restore any method result value 217 __ ldf(FloatRegisterImpl::D, d_tmp, F0); 218 __ ldx(l_tmp, O0); 219 } 220 221 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) { 222 assert(!pass_oop || message == NULL, "either oop or message but not both"); 223 address entry = __ pc(); 224 // expression stack must be empty before entering the VM if an exception happened 225 __ empty_expression_stack(); 226 // load exception object 227 __ set((intptr_t)name, G3_scratch); 228 if (pass_oop) { 229 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i); 230 } else { 231 __ set((intptr_t)message, G4_scratch); 232 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch); 233 } 234 // throw exception 235 assert(Interpreter::throw_exception_entry() != NULL, "generate it first"); 236 AddressLiteral thrower(Interpreter::throw_exception_entry()); 237 __ jump_to(thrower, G3_scratch); 238 __ delayed()->nop(); 239 return entry; 240 } 241 242 address TemplateInterpreterGenerator::generate_ClassCastException_handler() { 243 address entry = __ pc(); 244 // expression stack must be empty before entering the VM if an exception 245 // happened 246 __ empty_expression_stack(); 247 // load exception object 248 __ call_VM(Oexception, 249 CAST_FROM_FN_PTR(address, 250 InterpreterRuntime::throw_ClassCastException), 251 Otos_i); 252 __ should_not_reach_here(); 253 return entry; 254 } 255 256 257 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) { 258 address entry = __ pc(); 259 // expression stack must be empty before entering the VM if an exception happened 260 __ empty_expression_stack(); 261 // convention: expect aberrant index in register G3_scratch, then shuffle the 262 // index to G4_scratch for the VM call 263 __ mov(G3_scratch, G4_scratch); 264 __ set((intptr_t)name, G3_scratch); 265 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch); 266 __ should_not_reach_here(); 267 return entry; 268 } 269 270 271 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() { 272 address entry = __ pc(); 273 // expression stack must be empty before entering the VM if an exception happened 274 __ empty_expression_stack(); 275 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError)); 276 __ should_not_reach_here(); 277 return entry; 278 } 279 280 281 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) { 282 address entry = __ pc(); 283 284 if (state == atos) { 285 __ profile_return_type(O0, G3_scratch, G1_scratch); 286 } 287 288 // The callee returns with the stack possibly adjusted by adapter transition 289 // We remove that possible adjustment here. 290 // All interpreter local registers are untouched. Any result is passed back 291 // in the O0/O1 or float registers. Before continuing, the arguments must be 292 // popped from the java expression stack; i.e., Lesp must be adjusted. 293 294 __ mov(Llast_SP, SP); // Remove any adapter added stack space. 295 296 const Register cache = G3_scratch; 297 const Register index = G1_scratch; 298 __ get_cache_and_index_at_bcp(cache, index, 1, index_size); 299 300 const Register flags = cache; 301 __ ld_ptr(cache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset(), flags); 302 const Register parameter_size = flags; 303 __ and3(flags, ConstantPoolCacheEntry::parameter_size_mask, parameter_size); // argument size in words 304 __ sll(parameter_size, Interpreter::logStackElementSize, parameter_size); // each argument size in bytes 305 __ add(Lesp, parameter_size, Lesp); // pop arguments 306 307 __ check_and_handle_popframe(Gtemp); 308 __ check_and_handle_earlyret(Gtemp); 309 310 __ dispatch_next(state, step); 311 312 return entry; 313 } 314 315 316 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) { 317 address entry = __ pc(); 318 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache 319 #if INCLUDE_JVMCI 320 // Check if we need to take lock at entry of synchronized method. This can 321 // only occur on method entry so emit it only for vtos with step 0. 322 if (UseJVMCICompiler && state == vtos && step == 0) { 323 Label L; 324 Address pending_monitor_enter_addr(G2_thread, JavaThread::pending_monitorenter_offset()); 325 __ ldbool(pending_monitor_enter_addr, Gtemp); // Load if pending monitor enter 326 __ cmp_and_br_short(Gtemp, G0, Assembler::equal, Assembler::pn, L); 327 // Clear flag. 328 __ stbool(G0, pending_monitor_enter_addr); 329 // Take lock. 330 lock_method(); 331 __ bind(L); 332 } else { 333 #ifdef ASSERT 334 if (UseJVMCICompiler) { 335 Label L; 336 Address pending_monitor_enter_addr(G2_thread, JavaThread::pending_monitorenter_offset()); 337 __ ldbool(pending_monitor_enter_addr, Gtemp); // Load if pending monitor enter 338 __ cmp_and_br_short(Gtemp, G0, Assembler::equal, Assembler::pn, L); 339 __ stop("unexpected pending monitor in deopt entry"); 340 __ bind(L); 341 } 342 #endif 343 } 344 #endif 345 { Label L; 346 Address exception_addr(G2_thread, Thread::pending_exception_offset()); 347 __ ld_ptr(exception_addr, Gtemp); // Load pending exception. 348 __ br_null_short(Gtemp, Assembler::pt, L); 349 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception)); 350 __ should_not_reach_here(); 351 __ bind(L); 352 } 353 __ dispatch_next(state, step); 354 return entry; 355 } 356 357 // A result handler converts/unboxes a native call result into 358 // a java interpreter/compiler result. The current frame is an 359 // interpreter frame. The activation frame unwind code must be 360 // consistent with that of TemplateTable::_return(...). In the 361 // case of native methods, the caller's SP was not modified. 362 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) { 363 address entry = __ pc(); 364 Register Itos_i = Otos_i ->after_save(); 365 Register Itos_l = Otos_l ->after_save(); 366 Register Itos_l1 = Otos_l1->after_save(); 367 Register Itos_l2 = Otos_l2->after_save(); 368 switch (type) { 369 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false 370 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value! 371 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break; 372 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break; 373 case T_LONG : 374 case T_INT : __ mov(O0, Itos_i); break; 375 case T_VOID : /* nothing to do */ break; 376 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break; 377 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break; 378 case T_OBJECT : 379 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i); 380 __ verify_oop(Itos_i); 381 break; 382 default : ShouldNotReachHere(); 383 } 384 __ ret(); // return from interpreter activation 385 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame 386 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly 387 return entry; 388 } 389 390 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) { 391 address entry = __ pc(); 392 __ push(state); 393 __ call_VM(noreg, runtime_entry); 394 __ dispatch_via(vtos, Interpreter::normal_table(vtos)); 395 return entry; 396 } 397 398 399 // 400 // Helpers for commoning out cases in the various type of method entries. 401 // 402 403 // increment invocation count & check for overflow 404 // 405 // Note: checking for negative value instead of overflow 406 // so we have a 'sticky' overflow test 407 // 408 // Lmethod: method 409 // ??: invocation counter 410 // 411 void TemplateInterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) { 412 // Note: In tiered we increment either counters in MethodCounters* or in 413 // MDO depending if we're profiling or not. 414 const Register G3_method_counters = G3_scratch; 415 Label done; 416 417 if (TieredCompilation) { 418 const int increment = InvocationCounter::count_increment; 419 Label no_mdo; 420 if (ProfileInterpreter) { 421 // If no method data exists, go to profile_continue. 422 __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch); 423 __ br_null_short(G4_scratch, Assembler::pn, no_mdo); 424 // Increment counter 425 Address mdo_invocation_counter(G4_scratch, 426 in_bytes(MethodData::invocation_counter_offset()) + 427 in_bytes(InvocationCounter::counter_offset())); 428 Address mask(G4_scratch, in_bytes(MethodData::invoke_mask_offset())); 429 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask, 430 G3_scratch, Lscratch, 431 Assembler::zero, overflow); 432 __ ba_short(done); 433 } 434 435 // Increment counter in MethodCounters* 436 __ bind(no_mdo); 437 Address invocation_counter(G3_method_counters, 438 in_bytes(MethodCounters::invocation_counter_offset()) + 439 in_bytes(InvocationCounter::counter_offset())); 440 __ get_method_counters(Lmethod, G3_method_counters, done); 441 Address mask(G3_method_counters, in_bytes(MethodCounters::invoke_mask_offset())); 442 __ increment_mask_and_jump(invocation_counter, increment, mask, 443 G4_scratch, Lscratch, 444 Assembler::zero, overflow); 445 __ bind(done); 446 } else { // not TieredCompilation 447 // Update standard invocation counters 448 __ get_method_counters(Lmethod, G3_method_counters, done); 449 __ increment_invocation_counter(G3_method_counters, O0, G4_scratch); 450 if (ProfileInterpreter) { 451 Address interpreter_invocation_counter(G3_method_counters, 452 in_bytes(MethodCounters::interpreter_invocation_counter_offset())); 453 __ ld(interpreter_invocation_counter, G4_scratch); 454 __ inc(G4_scratch); 455 __ st(G4_scratch, interpreter_invocation_counter); 456 } 457 458 if (ProfileInterpreter && profile_method != NULL) { 459 // Test to see if we should create a method data oop 460 Address profile_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_profile_limit_offset())); 461 __ ld(profile_limit, G1_scratch); 462 __ cmp_and_br_short(O0, G1_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue); 463 464 // if no method data exists, go to profile_method 465 __ test_method_data_pointer(*profile_method); 466 } 467 468 Address invocation_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_invocation_limit_offset())); 469 __ ld(invocation_limit, G3_scratch); 470 __ cmp(O0, G3_scratch); 471 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance 472 __ delayed()->nop(); 473 __ bind(done); 474 } 475 } 476 477 // Allocate monitor and lock method (asm interpreter) 478 // ebx - Method* 479 // 480 void TemplateInterpreterGenerator::lock_method() { 481 __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0); // Load access flags. 482 483 #ifdef ASSERT 484 { Label ok; 485 __ btst(JVM_ACC_SYNCHRONIZED, O0); 486 __ br( Assembler::notZero, false, Assembler::pt, ok); 487 __ delayed()->nop(); 488 __ stop("method doesn't need synchronization"); 489 __ bind(ok); 490 } 491 #endif // ASSERT 492 493 // get synchronization object to O0 494 { Label done; 495 __ btst(JVM_ACC_STATIC, O0); 496 __ br( Assembler::zero, true, Assembler::pt, done); 497 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case 498 499 // lock the mirror, not the Klass* 500 __ load_mirror(O0, Lmethod); 501 502 #ifdef ASSERT 503 __ tst(O0); 504 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc); 505 #endif // ASSERT 506 507 __ bind(done); 508 } 509 510 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem 511 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object 512 // __ untested("lock_object from method entry"); 513 __ lock_object(Lmonitors, O0); 514 } 515 516 // See if we've got enough room on the stack for locals plus overhead below 517 // JavaThread::stack_overflow_limit(). If not, throw a StackOverflowError 518 // without going through the signal handler, i.e., reserved and yellow zones 519 // will not be made usable. The shadow zone must suffice to handle the 520 // overflow. 521 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size, 522 Register Rscratch) { 523 const int page_size = os::vm_page_size(); 524 Label after_frame_check; 525 526 assert_different_registers(Rframe_size, Rscratch); 527 528 __ set(page_size, Rscratch); 529 __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check); 530 531 // Get the stack overflow limit, and in debug, verify it is non-zero. 532 __ ld_ptr(G2_thread, JavaThread::stack_overflow_limit_offset(), Rscratch); 533 #ifdef ASSERT 534 Label limit_ok; 535 __ br_notnull_short(Rscratch, Assembler::pn, limit_ok); 536 __ stop("stack overflow limit is zero in generate_stack_overflow_check"); 537 __ bind(limit_ok); 538 #endif 539 540 // Add in the size of the frame (which is the same as subtracting it from the 541 // SP, which would take another register. 542 __ add(Rscratch, Rframe_size, Rscratch); 543 544 // The frame is greater than one page in size, so check against 545 // the bottom of the stack. 546 __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check); 547 548 // The stack will overflow, throw an exception. 549 550 // Note that SP is restored to sender's sp (in the delay slot). This 551 // is necessary if the sender's frame is an extended compiled frame 552 // (see gen_c2i_adapter()) and safer anyway in case of JSR292 553 // adaptations. 554 555 // Note also that the restored frame is not necessarily interpreted. 556 // Use the shared runtime version of the StackOverflowError. 557 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated"); 558 AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry()); 559 __ jump_to(stub, Rscratch); 560 __ delayed()->mov(O5_savedSP, SP); 561 562 // If you get to here, then there is enough stack space. 563 __ bind(after_frame_check); 564 } 565 566 567 // 568 // Generate a fixed interpreter frame. This is identical setup for interpreted 569 // methods and for native methods hence the shared code. 570 571 572 //---------------------------------------------------------------------------------------------------- 573 // Stack frame layout 574 // 575 // When control flow reaches any of the entry types for the interpreter 576 // the following holds -> 577 // 578 // C2 Calling Conventions: 579 // 580 // The entry code below assumes that the following registers are set 581 // when coming in: 582 // G5_method: holds the Method* of the method to call 583 // Lesp: points to the TOS of the callers expression stack 584 // after having pushed all the parameters 585 // 586 // The entry code does the following to setup an interpreter frame 587 // pop parameters from the callers stack by adjusting Lesp 588 // set O0 to Lesp 589 // compute X = (max_locals - num_parameters) 590 // bump SP up by X to accommodate the extra locals 591 // compute X = max_expression_stack 592 // + vm_local_words 593 // + 16 words of register save area 594 // save frame doing a save sp, -X, sp growing towards lower addresses 595 // set Lbcp, Lmethod, LcpoolCache 596 // set Llocals to i0 597 // set Lmonitors to FP - rounded_vm_local_words 598 // set Lesp to Lmonitors - 4 599 // 600 // The frame has now been setup to do the rest of the entry code 601 602 // Try this optimization: Most method entries could live in a 603 // "one size fits all" stack frame without all the dynamic size 604 // calculations. It might be profitable to do all this calculation 605 // statically and approximately for "small enough" methods. 606 607 //----------------------------------------------------------------------------------------------- 608 609 // C1 Calling conventions 610 // 611 // Upon method entry, the following registers are setup: 612 // 613 // g2 G2_thread: current thread 614 // g5 G5_method: method to activate 615 // g4 Gargs : pointer to last argument 616 // 617 // 618 // Stack: 619 // 620 // +---------------+ <--- sp 621 // | | 622 // : reg save area : 623 // | | 624 // +---------------+ <--- sp + 0x40 625 // | | 626 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later) 627 // | | 628 // +---------------+ <--- sp + 0x5c 629 // | | 630 // : free : 631 // | | 632 // +---------------+ <--- Gargs 633 // | | 634 // : arguments : 635 // | | 636 // +---------------+ 637 // | | 638 // 639 // 640 // 641 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like: 642 // 643 // +---------------+ <--- sp 644 // | | 645 // : reg save area : 646 // | | 647 // +---------------+ <--- sp + 0x40 648 // | | 649 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later) 650 // | | 651 // +---------------+ <--- sp + 0x5c 652 // | | 653 // : : 654 // | | <--- Lesp 655 // +---------------+ <--- Lmonitors (fp - 0x18) 656 // | VM locals | 657 // +---------------+ <--- fp 658 // | | 659 // : reg save area : 660 // | | 661 // +---------------+ <--- fp + 0x40 662 // | | 663 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later) 664 // | | 665 // +---------------+ <--- fp + 0x5c 666 // | | 667 // : free : 668 // | | 669 // +---------------+ 670 // | | 671 // : nonarg locals : 672 // | | 673 // +---------------+ 674 // | | 675 // : arguments : 676 // | | <--- Llocals 677 // +---------------+ <--- Gargs 678 // | | 679 680 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) { 681 // 682 // 683 // The entry code sets up a new interpreter frame in 4 steps: 684 // 685 // 1) Increase caller's SP by for the extra local space needed: 686 // (check for overflow) 687 // Efficient implementation of xload/xstore bytecodes requires 688 // that arguments and non-argument locals are in a contiguously 689 // addressable memory block => non-argument locals must be 690 // allocated in the caller's frame. 691 // 692 // 2) Create a new stack frame and register window: 693 // The new stack frame must provide space for the standard 694 // register save area, the maximum java expression stack size, 695 // the monitor slots (0 slots initially), and some frame local 696 // scratch locations. 697 // 698 // 3) The following interpreter activation registers must be setup: 699 // Lesp : expression stack pointer 700 // Lbcp : bytecode pointer 701 // Lmethod : method 702 // Llocals : locals pointer 703 // Lmonitors : monitor pointer 704 // LcpoolCache: constant pool cache 705 // 706 // 4) Initialize the non-argument locals if necessary: 707 // Non-argument locals may need to be initialized to NULL 708 // for GC to work. If the oop-map information is accurate 709 // (in the absence of the JSR problem), no initialization 710 // is necessary. 711 // 712 // (gri - 2/25/2000) 713 714 715 int rounded_vm_local_words = align_up((int)frame::interpreter_frame_vm_local_words, WordsPerLong ); 716 717 const int extra_space = 718 rounded_vm_local_words + // frame local scratch space 719 Method::extra_stack_entries() + // extra stack for jsr 292 720 frame::memory_parameter_word_sp_offset + // register save area 721 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0); 722 723 const Register Glocals_size = G3; 724 const Register RconstMethod = Glocals_size; 725 const Register Otmp1 = O3; 726 const Register Otmp2 = O4; 727 // Lscratch can't be used as a temporary because the call_stub uses 728 // it to assert that the stack frame was setup correctly. 729 const Address constMethod (G5_method, Method::const_offset()); 730 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset()); 731 732 __ ld_ptr( constMethod, RconstMethod ); 733 __ lduh( size_of_parameters, Glocals_size); 734 735 // Gargs points to first local + BytesPerWord 736 // Set the saved SP after the register window save 737 // 738 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP); 739 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1); 740 __ add(Gargs, Otmp1, Gargs); 741 742 if (native_call) { 743 __ calc_mem_param_words( Glocals_size, Gframe_size ); 744 __ add( Gframe_size, extra_space, Gframe_size); 745 __ round_to( Gframe_size, WordsPerLong ); 746 __ sll( Gframe_size, LogBytesPerWord, Gframe_size ); 747 748 // Native calls don't need the stack size check since they have no 749 // expression stack and the arguments are already on the stack and 750 // we only add a handful of words to the stack. 751 } else { 752 753 // 754 // Compute number of locals in method apart from incoming parameters 755 // 756 const Address size_of_locals(Otmp1, ConstMethod::size_of_locals_offset()); 757 __ ld_ptr(constMethod, Otmp1); 758 __ lduh(size_of_locals, Otmp1); 759 __ sub(Otmp1, Glocals_size, Glocals_size); 760 __ round_to(Glocals_size, WordsPerLong); 761 __ sll(Glocals_size, Interpreter::logStackElementSize, Glocals_size); 762 763 // See if the frame is greater than one page in size. If so, 764 // then we need to verify there is enough stack space remaining. 765 // Frame_size = (max_stack + extra_space) * BytesPerWord; 766 __ ld_ptr(constMethod, Gframe_size); 767 __ lduh(Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size); 768 __ add(Gframe_size, extra_space, Gframe_size); 769 __ round_to(Gframe_size, WordsPerLong); 770 __ sll(Gframe_size, Interpreter::logStackElementSize, Gframe_size); 771 772 // Add in java locals size for stack overflow check only 773 __ add(Gframe_size, Glocals_size, Gframe_size); 774 775 const Register Otmp2 = O4; 776 assert_different_registers(Otmp1, Otmp2, O5_savedSP); 777 generate_stack_overflow_check(Gframe_size, Otmp1); 778 779 __ sub(Gframe_size, Glocals_size, Gframe_size); 780 781 // 782 // bump SP to accommodate the extra locals 783 // 784 __ sub(SP, Glocals_size, SP); 785 } 786 787 // 788 // now set up a stack frame with the size computed above 789 // 790 __ neg( Gframe_size ); 791 __ save( SP, Gframe_size, SP ); 792 793 // 794 // now set up all the local cache registers 795 // 796 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note 797 // that all present references to Lbyte_code initialize the register 798 // immediately before use 799 if (native_call) { 800 __ mov(G0, Lbcp); 801 } else { 802 __ ld_ptr(G5_method, Method::const_offset(), Lbcp); 803 __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp); 804 } 805 __ mov( G5_method, Lmethod); // set Lmethod 806 // Get mirror and store it in the frame as GC root for this Method* 807 Register mirror = LcpoolCache; 808 __ load_mirror(mirror, Lmethod); 809 __ st_ptr(mirror, FP, (frame::interpreter_frame_mirror_offset * wordSize) + STACK_BIAS); 810 __ get_constant_pool_cache(LcpoolCache); // set LcpoolCache 811 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors 812 __ add(Lmonitors, STACK_BIAS, Lmonitors); // Account for 64 bit stack bias 813 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp 814 815 // setup interpreter activation registers 816 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals 817 818 if (ProfileInterpreter) { 819 __ set_method_data_pointer(); 820 } 821 822 } 823 824 // Method entry for java.lang.ref.Reference.get. 825 address TemplateInterpreterGenerator::generate_Reference_get_entry(void) { 826 #if INCLUDE_ALL_GCS 827 // Code: _aload_0, _getfield, _areturn 828 // parameter size = 1 829 // 830 // The code that gets generated by this routine is split into 2 parts: 831 // 1. The "intrinsified" code for G1 (or any SATB based GC), 832 // 2. The slow path - which is an expansion of the regular method entry. 833 // 834 // Notes:- 835 // * In the G1 code we do not check whether we need to block for 836 // a safepoint. If G1 is enabled then we must execute the specialized 837 // code for Reference.get (except when the Reference object is null) 838 // so that we can log the value in the referent field with an SATB 839 // update buffer. 840 // If the code for the getfield template is modified so that the 841 // G1 pre-barrier code is executed when the current method is 842 // Reference.get() then going through the normal method entry 843 // will be fine. 844 // * The G1 code can, however, check the receiver object (the instance 845 // of java.lang.Reference) and jump to the slow path if null. If the 846 // Reference object is null then we obviously cannot fetch the referent 847 // and so we don't need to call the G1 pre-barrier. Thus we can use the 848 // regular method entry code to generate the NPE. 849 // 850 // This code is based on generate_accessor_enty. 851 852 address entry = __ pc(); 853 854 const int referent_offset = java_lang_ref_Reference::referent_offset; 855 guarantee(referent_offset > 0, "referent offset not initialized"); 856 857 if (UseG1GC) { 858 Label slow_path; 859 860 // In the G1 code we don't check if we need to reach a safepoint. We 861 // continue and the thread will safepoint at the next bytecode dispatch. 862 863 // Check if local 0 != NULL 864 // If the receiver is null then it is OK to jump to the slow path. 865 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0 866 // check if local 0 == NULL and go the slow path 867 __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path); 868 869 870 // Load the value of the referent field. 871 if (Assembler::is_simm13(referent_offset)) { 872 __ load_heap_oop(Otos_i, referent_offset, Otos_i); 873 } else { 874 __ set(referent_offset, G3_scratch); 875 __ load_heap_oop(Otos_i, G3_scratch, Otos_i); 876 } 877 878 // Generate the G1 pre-barrier code to log the value of 879 // the referent field in an SATB buffer. Note with 880 // these parameters the pre-barrier does not generate 881 // the load of the previous value 882 883 __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */, 884 Otos_i /* pre_val */, 885 G3_scratch /* tmp */, 886 true /* preserve_o_regs */); 887 888 // _areturn 889 __ retl(); // return from leaf routine 890 __ delayed()->mov(O5_savedSP, SP); 891 892 // Generate regular method entry 893 __ bind(slow_path); 894 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals)); 895 return entry; 896 } 897 #endif // INCLUDE_ALL_GCS 898 899 // If G1 is not enabled then attempt to go through the accessor entry point 900 // Reference.get is an accessor 901 return NULL; 902 } 903 904 /** 905 * Method entry for static native methods: 906 * int java.util.zip.CRC32.update(int crc, int b) 907 */ 908 address TemplateInterpreterGenerator::generate_CRC32_update_entry() { 909 910 if (UseCRC32Intrinsics) { 911 address entry = __ pc(); 912 913 Label L_slow_path; 914 // If we need a safepoint check, generate full interpreter entry. 915 ExternalAddress state(SafepointSynchronize::address_of_state()); 916 __ set(ExternalAddress(SafepointSynchronize::address_of_state()), O2); 917 __ set(SafepointSynchronize::_not_synchronized, O3); 918 __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pt, L_slow_path); 919 920 // Load parameters 921 const Register crc = O0; // initial crc 922 const Register val = O1; // byte to update with 923 const Register table = O2; // address of 256-entry lookup table 924 925 __ ldub(Gargs, 3, val); 926 __ lduw(Gargs, 8, crc); 927 928 __ set(ExternalAddress(StubRoutines::crc_table_addr()), table); 929 930 __ not1(crc); // ~crc 931 __ clruwu(crc); 932 __ update_byte_crc32(crc, val, table); 933 __ not1(crc); // ~crc 934 935 // result in O0 936 __ retl(); 937 __ delayed()->nop(); 938 939 // generate a vanilla native entry as the slow path 940 __ bind(L_slow_path); 941 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native)); 942 return entry; 943 } 944 return NULL; 945 } 946 947 /** 948 * Method entry for static native methods: 949 * int java.util.zip.CRC32.updateBytes(int crc, byte[] b, int off, int len) 950 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 951 */ 952 address TemplateInterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractInterpreter::MethodKind kind) { 953 954 if (UseCRC32Intrinsics) { 955 address entry = __ pc(); 956 957 Label L_slow_path; 958 // If we need a safepoint check, generate full interpreter entry. 959 ExternalAddress state(SafepointSynchronize::address_of_state()); 960 __ set(ExternalAddress(SafepointSynchronize::address_of_state()), O2); 961 __ set(SafepointSynchronize::_not_synchronized, O3); 962 __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pt, L_slow_path); 963 964 // Load parameters from the stack 965 const Register crc = O0; // initial crc 966 const Register buf = O1; // source java byte array address 967 const Register len = O2; // len 968 const Register offset = O3; // offset 969 970 // Arguments are reversed on java expression stack 971 // Calculate address of start element 972 if (kind == Interpreter::java_util_zip_CRC32_updateByteBuffer) { 973 __ lduw(Gargs, 0, len); 974 __ lduw(Gargs, 8, offset); 975 __ ldx( Gargs, 16, buf); 976 __ lduw(Gargs, 32, crc); 977 __ add(buf, offset, buf); 978 } else { 979 __ lduw(Gargs, 0, len); 980 __ lduw(Gargs, 8, offset); 981 __ ldx( Gargs, 16, buf); 982 __ lduw(Gargs, 24, crc); 983 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size 984 __ add(buf, offset, buf); 985 } 986 987 // Call the crc32 kernel 988 __ MacroAssembler::save_thread(L7_thread_cache); 989 __ kernel_crc32(crc, buf, len, O3); 990 __ MacroAssembler::restore_thread(L7_thread_cache); 991 992 // result in O0 993 __ retl(); 994 __ delayed()->nop(); 995 996 // generate a vanilla native entry as the slow path 997 __ bind(L_slow_path); 998 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native)); 999 return entry; 1000 } 1001 return NULL; 1002 } 1003 1004 /** 1005 * Method entry for intrinsic-candidate (non-native) methods: 1006 * int java.util.zip.CRC32C.updateBytes(int crc, byte[] b, int off, int end) 1007 * int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 1008 * Unlike CRC32, CRC32C does not have any methods marked as native 1009 * CRC32C also uses an "end" variable instead of the length variable CRC32 uses 1010 */ 1011 address TemplateInterpreterGenerator::generate_CRC32C_updateBytes_entry(AbstractInterpreter::MethodKind kind) { 1012 1013 if (UseCRC32CIntrinsics) { 1014 address entry = __ pc(); 1015 1016 // Load parameters from the stack 1017 const Register crc = O0; // initial crc 1018 const Register buf = O1; // source java byte array address 1019 const Register offset = O2; // offset 1020 const Register end = O3; // index of last element to process 1021 const Register len = O2; // len argument to the kernel 1022 const Register table = O3; // crc32c lookup table address 1023 1024 // Arguments are reversed on java expression stack 1025 // Calculate address of start element 1026 if (kind == Interpreter::java_util_zip_CRC32C_updateDirectByteBuffer) { 1027 __ lduw(Gargs, 0, end); 1028 __ lduw(Gargs, 8, offset); 1029 __ ldx( Gargs, 16, buf); 1030 __ lduw(Gargs, 32, crc); 1031 __ add(buf, offset, buf); 1032 __ sub(end, offset, len); 1033 } else { 1034 __ lduw(Gargs, 0, end); 1035 __ lduw(Gargs, 8, offset); 1036 __ ldx( Gargs, 16, buf); 1037 __ lduw(Gargs, 24, crc); 1038 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size 1039 __ add(buf, offset, buf); 1040 __ sub(end, offset, len); 1041 } 1042 1043 // Call the crc32c kernel 1044 __ MacroAssembler::save_thread(L7_thread_cache); 1045 __ kernel_crc32c(crc, buf, len, table); 1046 __ MacroAssembler::restore_thread(L7_thread_cache); 1047 1048 // result in O0 1049 __ retl(); 1050 __ delayed()->nop(); 1051 1052 return entry; 1053 } 1054 return NULL; 1055 } 1056 1057 /* Math routines only partially supported. 1058 * 1059 * Providing support for fma (float/double) only. 1060 */ 1061 address TemplateInterpreterGenerator::generate_math_entry(AbstractInterpreter::MethodKind kind) 1062 { 1063 if (!InlineIntrinsics) return NULL; // Generate a vanilla entry 1064 1065 address entry = __ pc(); 1066 1067 switch (kind) { 1068 case Interpreter::java_lang_math_fmaF: 1069 if (UseFMA) { 1070 // float .fma(float a, float b, float c) 1071 const FloatRegister ra = F1; 1072 const FloatRegister rb = F2; 1073 const FloatRegister rc = F3; 1074 const FloatRegister rd = F0; // Result. 1075 1076 __ ldf(FloatRegisterImpl::S, Gargs, 0, rc); 1077 __ ldf(FloatRegisterImpl::S, Gargs, 8, rb); 1078 __ ldf(FloatRegisterImpl::S, Gargs, 16, ra); 1079 1080 __ fmadd(FloatRegisterImpl::S, ra, rb, rc, rd); 1081 __ retl(); // Result in F0 (rd). 1082 __ delayed()->mov(O5_savedSP, SP); 1083 1084 return entry; 1085 } 1086 break; 1087 case Interpreter::java_lang_math_fmaD: 1088 if (UseFMA) { 1089 // double .fma(double a, double b, double c) 1090 const FloatRegister ra = F2; // D1 1091 const FloatRegister rb = F4; // D2 1092 const FloatRegister rc = F6; // D3 1093 const FloatRegister rd = F0; // D0 Result. 1094 1095 __ ldf(FloatRegisterImpl::D, Gargs, 0, rc); 1096 __ ldf(FloatRegisterImpl::D, Gargs, 16, rb); 1097 __ ldf(FloatRegisterImpl::D, Gargs, 32, ra); 1098 1099 __ fmadd(FloatRegisterImpl::D, ra, rb, rc, rd); 1100 __ retl(); // Result in D0 (rd). 1101 __ delayed()->mov(O5_savedSP, SP); 1102 1103 return entry; 1104 } 1105 break; 1106 default: 1107 break; 1108 } 1109 return NULL; 1110 } 1111 1112 // TODO: rather than touching all pages, check against stack_overflow_limit and bang yellow page to 1113 // generate exception 1114 void TemplateInterpreterGenerator::bang_stack_shadow_pages(bool native_call) { 1115 // Quick & dirty stack overflow checking: bang the stack & handle trap. 1116 // Note that we do the banging after the frame is setup, since the exception 1117 // handling code expects to find a valid interpreter frame on the stack. 1118 // Doing the banging earlier fails if the caller frame is not an interpreter 1119 // frame. 1120 // (Also, the exception throwing code expects to unlock any synchronized 1121 // method receiver, so do the banging after locking the receiver.) 1122 1123 // Bang each page in the shadow zone. We can't assume it's been done for 1124 // an interpreter frame with greater than a page of locals, so each page 1125 // needs to be checked. Only true for non-native. 1126 if (UseStackBanging) { 1127 const int page_size = os::vm_page_size(); 1128 const int n_shadow_pages = ((int)JavaThread::stack_shadow_zone_size()) / page_size; 1129 const int start_page = native_call ? n_shadow_pages : 1; 1130 for (int pages = start_page; pages <= n_shadow_pages; pages++) { 1131 __ bang_stack_with_offset(pages*page_size); 1132 } 1133 } 1134 } 1135 1136 // 1137 // Interpreter stub for calling a native method. (asm interpreter) 1138 // This sets up a somewhat different looking stack for calling the native method 1139 // than the typical interpreter frame setup. 1140 // 1141 1142 address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) { 1143 address entry = __ pc(); 1144 1145 // the following temporary registers are used during frame creation 1146 const Register Gtmp1 = G3_scratch ; 1147 const Register Gtmp2 = G1_scratch; 1148 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods; 1149 1150 // make sure registers are different! 1151 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2); 1152 1153 const Address Laccess_flags(Lmethod, Method::access_flags_offset()); 1154 1155 const Register Glocals_size = G3; 1156 assert_different_registers(Glocals_size, G4_scratch, Gframe_size); 1157 1158 // make sure method is native & not abstract 1159 // rethink these assertions - they can be simplified and shared (gri 2/25/2000) 1160 #ifdef ASSERT 1161 __ ld(G5_method, Method::access_flags_offset(), Gtmp1); 1162 { Label L; 1163 __ btst(JVM_ACC_NATIVE, Gtmp1); 1164 __ br(Assembler::notZero, false, Assembler::pt, L); 1165 __ delayed()->nop(); 1166 __ stop("tried to execute non-native method as native"); 1167 __ bind(L); 1168 } 1169 { Label L; 1170 __ btst(JVM_ACC_ABSTRACT, Gtmp1); 1171 __ br(Assembler::zero, false, Assembler::pt, L); 1172 __ delayed()->nop(); 1173 __ stop("tried to execute abstract method as non-abstract"); 1174 __ bind(L); 1175 } 1176 #endif // ASSERT 1177 1178 // generate the code to allocate the interpreter stack frame 1179 generate_fixed_frame(true); 1180 1181 // 1182 // No locals to initialize for native method 1183 // 1184 1185 // this slot will be set later, we initialize it to null here just in 1186 // case we get a GC before the actual value is stored later 1187 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS); 1188 1189 const Address do_not_unlock_if_synchronized(G2_thread, 1190 JavaThread::do_not_unlock_if_synchronized_offset()); 1191 // Since at this point in the method invocation the exception handler 1192 // would try to exit the monitor of synchronized methods which hasn't 1193 // been entered yet, we set the thread local variable 1194 // _do_not_unlock_if_synchronized to true. If any exception was thrown by 1195 // runtime, exception handling i.e. unlock_if_synchronized_method will 1196 // check this thread local flag. 1197 // This flag has two effects, one is to force an unwind in the topmost 1198 // interpreter frame and not perform an unlock while doing so. 1199 1200 __ movbool(true, G3_scratch); 1201 __ stbool(G3_scratch, do_not_unlock_if_synchronized); 1202 1203 // increment invocation counter and check for overflow 1204 // 1205 // Note: checking for negative value instead of overflow 1206 // so we have a 'sticky' overflow test (may be of 1207 // importance as soon as we have true MT/MP) 1208 Label invocation_counter_overflow; 1209 Label Lcontinue; 1210 if (inc_counter) { 1211 generate_counter_incr(&invocation_counter_overflow, NULL, NULL); 1212 1213 } 1214 __ bind(Lcontinue); 1215 1216 bang_stack_shadow_pages(true); 1217 1218 // reset the _do_not_unlock_if_synchronized flag 1219 __ stbool(G0, do_not_unlock_if_synchronized); 1220 1221 // check for synchronized methods 1222 // Must happen AFTER invocation_counter check and stack overflow check, 1223 // so method is not locked if overflows. 1224 1225 if (synchronized) { 1226 lock_method(); 1227 } else { 1228 #ifdef ASSERT 1229 { Label ok; 1230 __ ld(Laccess_flags, O0); 1231 __ btst(JVM_ACC_SYNCHRONIZED, O0); 1232 __ br( Assembler::zero, false, Assembler::pt, ok); 1233 __ delayed()->nop(); 1234 __ stop("method needs synchronization"); 1235 __ bind(ok); 1236 } 1237 #endif // ASSERT 1238 } 1239 1240 1241 // start execution 1242 __ verify_thread(); 1243 1244 // JVMTI support 1245 __ notify_method_entry(); 1246 1247 // native call 1248 1249 // (note that O0 is never an oop--at most it is a handle) 1250 // It is important not to smash any handles created by this call, 1251 // until any oop handle in O0 is dereferenced. 1252 1253 // (note that the space for outgoing params is preallocated) 1254 1255 // get signature handler 1256 { Label L; 1257 Address signature_handler(Lmethod, Method::signature_handler_offset()); 1258 __ ld_ptr(signature_handler, G3_scratch); 1259 __ br_notnull_short(G3_scratch, Assembler::pt, L); 1260 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod); 1261 __ ld_ptr(signature_handler, G3_scratch); 1262 __ bind(L); 1263 } 1264 1265 // Push a new frame so that the args will really be stored in 1266 // Copy a few locals across so the new frame has the variables 1267 // we need but these values will be dead at the jni call and 1268 // therefore not gc volatile like the values in the current 1269 // frame (Lmethod in particular) 1270 1271 // Flush the method pointer to the register save area 1272 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS); 1273 __ mov(Llocals, O1); 1274 1275 // calculate where the mirror handle body is allocated in the interpreter frame: 1276 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2); 1277 1278 // Calculate current frame size 1279 __ sub(SP, FP, O3); // Calculate negative of current frame size 1280 __ save(SP, O3, SP); // Allocate an identical sized frame 1281 1282 // Note I7 has leftover trash. Slow signature handler will fill it in 1283 // should we get there. Normal jni call will set reasonable last_Java_pc 1284 // below (and fix I7 so the stack trace doesn't have a meaningless frame 1285 // in it). 1286 1287 // Load interpreter frame's Lmethod into same register here 1288 1289 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod); 1290 1291 __ mov(I1, Llocals); 1292 __ mov(I2, Lscratch2); // save the address of the mirror 1293 1294 1295 // ONLY Lmethod and Llocals are valid here! 1296 1297 // call signature handler, It will move the arg properly since Llocals in current frame 1298 // matches that in outer frame 1299 1300 __ callr(G3_scratch, 0); 1301 __ delayed()->nop(); 1302 1303 // Result handler is in Lscratch 1304 1305 // Reload interpreter frame's Lmethod since slow signature handler may block 1306 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod); 1307 1308 { Label not_static; 1309 1310 __ ld(Laccess_flags, O0); 1311 __ btst(JVM_ACC_STATIC, O0); 1312 __ br( Assembler::zero, false, Assembler::pt, not_static); 1313 // get native function entry point(O0 is a good temp until the very end) 1314 __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0); 1315 // for static methods insert the mirror argument 1316 __ load_mirror(O1, Lmethod); 1317 #ifdef ASSERT 1318 if (!PrintSignatureHandlers) // do not dirty the output with this 1319 { Label L; 1320 __ br_notnull_short(O1, Assembler::pt, L); 1321 __ stop("mirror is missing"); 1322 __ bind(L); 1323 } 1324 #endif // ASSERT 1325 __ st_ptr(O1, Lscratch2, 0); 1326 __ mov(Lscratch2, O1); 1327 __ bind(not_static); 1328 } 1329 1330 // At this point, arguments have been copied off of stack into 1331 // their JNI positions, which are O1..O5 and SP[68..]. 1332 // Oops are boxed in-place on the stack, with handles copied to arguments. 1333 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*. 1334 1335 #ifdef ASSERT 1336 { Label L; 1337 __ br_notnull_short(O0, Assembler::pt, L); 1338 __ stop("native entry point is missing"); 1339 __ bind(L); 1340 } 1341 #endif // ASSERT 1342 1343 // 1344 // setup the frame anchor 1345 // 1346 // The scavenge function only needs to know that the PC of this frame is 1347 // in the interpreter method entry code, it doesn't need to know the exact 1348 // PC and hence we can use O7 which points to the return address from the 1349 // previous call in the code stream (signature handler function) 1350 // 1351 // The other trick is we set last_Java_sp to FP instead of the usual SP because 1352 // we have pushed the extra frame in order to protect the volatile register(s) 1353 // in that frame when we return from the jni call 1354 // 1355 1356 __ set_last_Java_frame(FP, O7); 1357 __ mov(O7, I7); // make dummy interpreter frame look like one above, 1358 // not meaningless information that'll confuse me. 1359 1360 // flush the windows now. We don't care about the current (protection) frame 1361 // only the outer frames 1362 1363 __ flushw(); 1364 1365 // mark windows as flushed 1366 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset()); 1367 __ set(JavaFrameAnchor::flushed, G3_scratch); 1368 __ st(G3_scratch, flags); 1369 1370 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready. 1371 1372 Address thread_state(G2_thread, JavaThread::thread_state_offset()); 1373 #ifdef ASSERT 1374 { Label L; 1375 __ ld(thread_state, G3_scratch); 1376 __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L); 1377 __ stop("Wrong thread state in native stub"); 1378 __ bind(L); 1379 } 1380 #endif // ASSERT 1381 __ set(_thread_in_native, G3_scratch); 1382 __ st(G3_scratch, thread_state); 1383 1384 // Call the jni method, using the delay slot to set the JNIEnv* argument. 1385 __ save_thread(L7_thread_cache); // save Gthread 1386 __ callr(O0, 0); 1387 __ delayed()-> 1388 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0); 1389 1390 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD 1391 1392 __ restore_thread(L7_thread_cache); // restore G2_thread 1393 __ reinit_heapbase(); 1394 1395 // must we block? 1396 1397 // Block, if necessary, before resuming in _thread_in_Java state. 1398 // In order for GC to work, don't clear the last_Java_sp until after blocking. 1399 { Label no_block; 1400 AddressLiteral sync_state(SafepointSynchronize::address_of_state()); 1401 1402 // Switch thread to "native transition" state before reading the synchronization state. 1403 // This additional state is necessary because reading and testing the synchronization 1404 // state is not atomic w.r.t. GC, as this scenario demonstrates: 1405 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 1406 // VM thread changes sync state to synchronizing and suspends threads for GC. 1407 // Thread A is resumed to finish this native method, but doesn't block here since it 1408 // didn't see any synchronization is progress, and escapes. 1409 __ set(_thread_in_native_trans, G3_scratch); 1410 __ st(G3_scratch, thread_state); 1411 if (os::is_MP()) { 1412 if (UseMembar) { 1413 // Force this write out before the read below 1414 __ membar(Assembler::StoreLoad); 1415 } else { 1416 // Write serialization page so VM thread can do a pseudo remote membar. 1417 // We use the current thread pointer to calculate a thread specific 1418 // offset to write to within the page. This minimizes bus traffic 1419 // due to cache line collision. 1420 __ serialize_memory(G2_thread, G1_scratch, G3_scratch); 1421 } 1422 } 1423 __ load_contents(sync_state, G3_scratch); 1424 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized); 1425 1426 Label L; 1427 __ br(Assembler::notEqual, false, Assembler::pn, L); 1428 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch); 1429 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block); 1430 __ bind(L); 1431 1432 // Block. Save any potential method result value before the operation and 1433 // use a leaf call to leave the last_Java_frame setup undisturbed. 1434 save_native_result(); 1435 __ call_VM_leaf(L7_thread_cache, 1436 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans), 1437 G2_thread); 1438 1439 // Restore any method result value 1440 restore_native_result(); 1441 __ bind(no_block); 1442 } 1443 1444 // Clear the frame anchor now 1445 1446 __ reset_last_Java_frame(); 1447 1448 // Move the result handler address 1449 __ mov(Lscratch, G3_scratch); 1450 // return possible result to the outer frame 1451 __ restore(O0, G0, O0); 1452 1453 // Move result handler to expected register 1454 __ mov(G3_scratch, Lscratch); 1455 1456 // Back in normal (native) interpreter frame. State is thread_in_native_trans 1457 // switch to thread_in_Java. 1458 1459 __ set(_thread_in_Java, G3_scratch); 1460 __ st(G3_scratch, thread_state); 1461 1462 if (CheckJNICalls) { 1463 // clear_pending_jni_exception_check 1464 __ st_ptr(G0, G2_thread, JavaThread::pending_jni_exception_check_fn_offset()); 1465 } 1466 1467 // reset handle block 1468 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch); 1469 __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes()); 1470 1471 // If we have an oop result store it where it will be safe for any further gc 1472 // until we return now that we've released the handle it might be protected by 1473 1474 { Label no_oop, store_result; 1475 1476 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch); 1477 __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop); 1478 // Unbox oop result, e.g. JNIHandles::resolve value in O0. 1479 __ br_null(O0, false, Assembler::pn, store_result); // Use NULL as-is. 1480 __ delayed()->andcc(O0, JNIHandles::weak_tag_mask, G0); // Test for jweak 1481 __ brx(Assembler::zero, true, Assembler::pt, store_result); 1482 __ delayed()->ld_ptr(O0, 0, O0); // Maybe resolve (untagged) jobject. 1483 // Resolve jweak. 1484 __ ld_ptr(O0, -JNIHandles::weak_tag_value, O0); 1485 #if INCLUDE_ALL_GCS 1486 if (UseG1GC) { 1487 __ g1_write_barrier_pre(noreg /* obj */, 1488 noreg /* index */, 1489 0 /* offset */, 1490 O0 /* pre_val */, 1491 G3_scratch /* tmp */, 1492 true /* preserve_o_regs */); 1493 } 1494 #endif // INCLUDE_ALL_GCS 1495 __ bind(store_result); 1496 // Store it where gc will look for it and result handler expects it. 1497 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS); 1498 1499 __ bind(no_oop); 1500 1501 } 1502 1503 1504 // handle exceptions (exception handling will handle unlocking!) 1505 { Label L; 1506 Address exception_addr(G2_thread, Thread::pending_exception_offset()); 1507 __ ld_ptr(exception_addr, Gtemp); 1508 __ br_null_short(Gtemp, Assembler::pt, L); 1509 // Note: This could be handled more efficiently since we know that the native 1510 // method doesn't have an exception handler. We could directly return 1511 // to the exception handler for the caller. 1512 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception)); 1513 __ should_not_reach_here(); 1514 __ bind(L); 1515 } 1516 1517 // JVMTI support (preserves thread register) 1518 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI); 1519 1520 if (synchronized) { 1521 // save and restore any potential method result value around the unlocking operation 1522 save_native_result(); 1523 1524 __ add( __ top_most_monitor(), O1); 1525 __ unlock_object(O1); 1526 1527 restore_native_result(); 1528 } 1529 1530 // dispose of return address and remove activation 1531 #ifdef ASSERT 1532 { Label ok; 1533 __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok); 1534 __ stop("bad I5_savedSP value"); 1535 __ should_not_reach_here(); 1536 __ bind(ok); 1537 } 1538 #endif 1539 __ jmp(Lscratch, 0); 1540 __ delayed()->nop(); 1541 1542 if (inc_counter) { 1543 // handle invocation counter overflow 1544 __ bind(invocation_counter_overflow); 1545 generate_counter_overflow(Lcontinue); 1546 } 1547 1548 return entry; 1549 } 1550 1551 1552 // Generic method entry to (asm) interpreter 1553 address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) { 1554 address entry = __ pc(); 1555 1556 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods; 1557 1558 // the following temporary registers are used during frame creation 1559 const Register Gtmp1 = G3_scratch ; 1560 const Register Gtmp2 = G1_scratch; 1561 1562 // make sure registers are different! 1563 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2); 1564 1565 const Address constMethod (G5_method, Method::const_offset()); 1566 // Seems like G5_method is live at the point this is used. So we could make this look consistent 1567 // and use in the asserts. 1568 const Address access_flags (Lmethod, Method::access_flags_offset()); 1569 1570 const Register Glocals_size = G3; 1571 assert_different_registers(Glocals_size, G4_scratch, Gframe_size); 1572 1573 // make sure method is not native & not abstract 1574 // rethink these assertions - they can be simplified and shared (gri 2/25/2000) 1575 #ifdef ASSERT 1576 __ ld(G5_method, Method::access_flags_offset(), Gtmp1); 1577 { Label L; 1578 __ btst(JVM_ACC_NATIVE, Gtmp1); 1579 __ br(Assembler::zero, false, Assembler::pt, L); 1580 __ delayed()->nop(); 1581 __ stop("tried to execute native method as non-native"); 1582 __ bind(L); 1583 } 1584 { Label L; 1585 __ btst(JVM_ACC_ABSTRACT, Gtmp1); 1586 __ br(Assembler::zero, false, Assembler::pt, L); 1587 __ delayed()->nop(); 1588 __ stop("tried to execute abstract method as non-abstract"); 1589 __ bind(L); 1590 } 1591 #endif // ASSERT 1592 1593 // generate the code to allocate the interpreter stack frame 1594 1595 generate_fixed_frame(false); 1596 1597 // 1598 // Code to initialize the extra (i.e. non-parm) locals 1599 // 1600 Register init_value = noreg; // will be G0 if we must clear locals 1601 // The way the code was setup before zerolocals was always true for vanilla java entries. 1602 // It could only be false for the specialized entries like accessor or empty which have 1603 // no extra locals so the testing was a waste of time and the extra locals were always 1604 // initialized. We removed this extra complication to already over complicated code. 1605 1606 init_value = G0; 1607 Label clear_loop; 1608 1609 const Register RconstMethod = O1; 1610 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset()); 1611 const Address size_of_locals (RconstMethod, ConstMethod::size_of_locals_offset()); 1612 1613 // NOTE: If you change the frame layout, this code will need to 1614 // be updated! 1615 __ ld_ptr( constMethod, RconstMethod ); 1616 __ lduh( size_of_locals, O2 ); 1617 __ lduh( size_of_parameters, O1 ); 1618 __ sll( O2, Interpreter::logStackElementSize, O2); 1619 __ sll( O1, Interpreter::logStackElementSize, O1 ); 1620 __ sub( Llocals, O2, O2 ); 1621 __ sub( Llocals, O1, O1 ); 1622 1623 __ bind( clear_loop ); 1624 __ inc( O2, wordSize ); 1625 1626 __ cmp( O2, O1 ); 1627 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop ); 1628 __ delayed()->st_ptr( init_value, O2, 0 ); 1629 1630 const Address do_not_unlock_if_synchronized(G2_thread, 1631 JavaThread::do_not_unlock_if_synchronized_offset()); 1632 // Since at this point in the method invocation the exception handler 1633 // would try to exit the monitor of synchronized methods which hasn't 1634 // been entered yet, we set the thread local variable 1635 // _do_not_unlock_if_synchronized to true. If any exception was thrown by 1636 // runtime, exception handling i.e. unlock_if_synchronized_method will 1637 // check this thread local flag. 1638 __ movbool(true, G3_scratch); 1639 __ stbool(G3_scratch, do_not_unlock_if_synchronized); 1640 1641 __ profile_parameters_type(G1_scratch, G3_scratch, G4_scratch, Lscratch); 1642 // increment invocation counter and check for overflow 1643 // 1644 // Note: checking for negative value instead of overflow 1645 // so we have a 'sticky' overflow test (may be of 1646 // importance as soon as we have true MT/MP) 1647 Label invocation_counter_overflow; 1648 Label profile_method; 1649 Label profile_method_continue; 1650 Label Lcontinue; 1651 if (inc_counter) { 1652 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue); 1653 if (ProfileInterpreter) { 1654 __ bind(profile_method_continue); 1655 } 1656 } 1657 __ bind(Lcontinue); 1658 1659 bang_stack_shadow_pages(false); 1660 1661 // reset the _do_not_unlock_if_synchronized flag 1662 __ stbool(G0, do_not_unlock_if_synchronized); 1663 1664 // check for synchronized methods 1665 // Must happen AFTER invocation_counter check and stack overflow check, 1666 // so method is not locked if overflows. 1667 1668 if (synchronized) { 1669 lock_method(); 1670 } else { 1671 #ifdef ASSERT 1672 { Label ok; 1673 __ ld(access_flags, O0); 1674 __ btst(JVM_ACC_SYNCHRONIZED, O0); 1675 __ br( Assembler::zero, false, Assembler::pt, ok); 1676 __ delayed()->nop(); 1677 __ stop("method needs synchronization"); 1678 __ bind(ok); 1679 } 1680 #endif // ASSERT 1681 } 1682 1683 // start execution 1684 1685 __ verify_thread(); 1686 1687 // jvmti support 1688 __ notify_method_entry(); 1689 1690 // start executing instructions 1691 __ dispatch_next(vtos); 1692 1693 1694 if (inc_counter) { 1695 if (ProfileInterpreter) { 1696 // We have decided to profile this method in the interpreter 1697 __ bind(profile_method); 1698 1699 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method)); 1700 __ set_method_data_pointer_for_bcp(); 1701 __ ba_short(profile_method_continue); 1702 } 1703 1704 // handle invocation counter overflow 1705 __ bind(invocation_counter_overflow); 1706 generate_counter_overflow(Lcontinue); 1707 } 1708 1709 return entry; 1710 } 1711 1712 //---------------------------------------------------------------------------------------------------- 1713 // Exceptions 1714 void TemplateInterpreterGenerator::generate_throw_exception() { 1715 1716 // Entry point in previous activation (i.e., if the caller was interpreted) 1717 Interpreter::_rethrow_exception_entry = __ pc(); 1718 // O0: exception 1719 1720 // entry point for exceptions thrown within interpreter code 1721 Interpreter::_throw_exception_entry = __ pc(); 1722 __ verify_thread(); 1723 // expression stack is undefined here 1724 // O0: exception, i.e. Oexception 1725 // Lbcp: exception bcp 1726 __ verify_oop(Oexception); 1727 1728 1729 // expression stack must be empty before entering the VM in case of an exception 1730 __ empty_expression_stack(); 1731 // find exception handler address and preserve exception oop 1732 // call C routine to find handler and jump to it 1733 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception); 1734 __ push_ptr(O1); // push exception for exception handler bytecodes 1735 1736 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!) 1737 __ delayed()->nop(); 1738 1739 1740 // if the exception is not handled in the current frame 1741 // the frame is removed and the exception is rethrown 1742 // (i.e. exception continuation is _rethrow_exception) 1743 // 1744 // Note: At this point the bci is still the bxi for the instruction which caused 1745 // the exception and the expression stack is empty. Thus, for any VM calls 1746 // at this point, GC will find a legal oop map (with empty expression stack). 1747 1748 // in current activation 1749 // tos: exception 1750 // Lbcp: exception bcp 1751 1752 // 1753 // JVMTI PopFrame support 1754 // 1755 1756 Interpreter::_remove_activation_preserving_args_entry = __ pc(); 1757 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset()); 1758 // Set the popframe_processing bit in popframe_condition indicating that we are 1759 // currently handling popframe, so that call_VMs that may happen later do not trigger new 1760 // popframe handling cycles. 1761 1762 __ ld(popframe_condition_addr, G3_scratch); 1763 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch); 1764 __ stw(G3_scratch, popframe_condition_addr); 1765 1766 // Empty the expression stack, as in normal exception handling 1767 __ empty_expression_stack(); 1768 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false); 1769 1770 { 1771 // Check to see whether we are returning to a deoptimized frame. 1772 // (The PopFrame call ensures that the caller of the popped frame is 1773 // either interpreted or compiled and deoptimizes it if compiled.) 1774 // In this case, we can't call dispatch_next() after the frame is 1775 // popped, but instead must save the incoming arguments and restore 1776 // them after deoptimization has occurred. 1777 // 1778 // Note that we don't compare the return PC against the 1779 // deoptimization blob's unpack entry because of the presence of 1780 // adapter frames in C2. 1781 Label caller_not_deoptimized; 1782 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7); 1783 __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized); 1784 1785 const Register Gtmp1 = G3_scratch; 1786 const Register Gtmp2 = G1_scratch; 1787 const Register RconstMethod = Gtmp1; 1788 const Address constMethod(Lmethod, Method::const_offset()); 1789 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset()); 1790 1791 // Compute size of arguments for saving when returning to deoptimized caller 1792 __ ld_ptr(constMethod, RconstMethod); 1793 __ lduh(size_of_parameters, Gtmp1); 1794 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1); 1795 __ sub(Llocals, Gtmp1, Gtmp2); 1796 __ add(Gtmp2, wordSize, Gtmp2); 1797 // Save these arguments 1798 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2); 1799 // Inform deoptimization that it is responsible for restoring these arguments 1800 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1); 1801 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset()); 1802 __ st(Gtmp1, popframe_condition_addr); 1803 1804 // Return from the current method 1805 // The caller's SP was adjusted upon method entry to accomodate 1806 // the callee's non-argument locals. Undo that adjustment. 1807 __ ret(); 1808 __ delayed()->restore(I5_savedSP, G0, SP); 1809 1810 __ bind(caller_not_deoptimized); 1811 } 1812 1813 // Clear the popframe condition flag 1814 __ stw(G0 /* popframe_inactive */, popframe_condition_addr); 1815 1816 // Get out of the current method (how this is done depends on the particular compiler calling 1817 // convention that the interpreter currently follows) 1818 // The caller's SP was adjusted upon method entry to accomodate 1819 // the callee's non-argument locals. Undo that adjustment. 1820 __ restore(I5_savedSP, G0, SP); 1821 // The method data pointer was incremented already during 1822 // call profiling. We have to restore the mdp for the current bcp. 1823 if (ProfileInterpreter) { 1824 __ set_method_data_pointer_for_bcp(); 1825 } 1826 1827 #if INCLUDE_JVMTI 1828 { Label L_done; 1829 1830 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode 1831 __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done); 1832 1833 // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call. 1834 // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL. 1835 1836 __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp); 1837 1838 __ br_null(G1_scratch, false, Assembler::pn, L_done); 1839 __ delayed()->nop(); 1840 1841 __ st_ptr(G1_scratch, Lesp, wordSize); 1842 __ bind(L_done); 1843 } 1844 #endif // INCLUDE_JVMTI 1845 1846 // Resume bytecode interpretation at the current bcp 1847 __ dispatch_next(vtos); 1848 // end of JVMTI PopFrame support 1849 1850 Interpreter::_remove_activation_entry = __ pc(); 1851 1852 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here) 1853 __ pop_ptr(Oexception); // get exception 1854 1855 // Intel has the following comment: 1856 //// remove the activation (without doing throws on illegalMonitorExceptions) 1857 // They remove the activation without checking for bad monitor state. 1858 // %%% We should make sure this is the right semantics before implementing. 1859 1860 __ set_vm_result(Oexception); 1861 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false); 1862 1863 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI); 1864 1865 __ get_vm_result(Oexception); 1866 __ verify_oop(Oexception); 1867 1868 const int return_reg_adjustment = frame::pc_return_offset; 1869 Address issuing_pc_addr(I7, return_reg_adjustment); 1870 1871 // We are done with this activation frame; find out where to go next. 1872 // The continuation point will be an exception handler, which expects 1873 // the following registers set up: 1874 // 1875 // Oexception: exception 1876 // Oissuing_pc: the local call that threw exception 1877 // Other On: garbage 1878 // In/Ln: the contents of the caller's register window 1879 // 1880 // We do the required restore at the last possible moment, because we 1881 // need to preserve some state across a runtime call. 1882 // (Remember that the caller activation is unknown--it might not be 1883 // interpreted, so things like Lscratch are useless in the caller.) 1884 1885 // Although the Intel version uses call_C, we can use the more 1886 // compact call_VM. (The only real difference on SPARC is a 1887 // harmlessly ignored [re]set_last_Java_frame, compared with 1888 // the Intel code which lacks this.) 1889 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore 1890 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller 1891 __ super_call_VM_leaf(L7_thread_cache, 1892 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), 1893 G2_thread, Oissuing_pc->after_save()); 1894 1895 // The caller's SP was adjusted upon method entry to accomodate 1896 // the callee's non-argument locals. Undo that adjustment. 1897 __ JMP(O0, 0); // return exception handler in caller 1898 __ delayed()->restore(I5_savedSP, G0, SP); 1899 1900 // (same old exception object is already in Oexception; see above) 1901 // Note that an "issuing PC" is actually the next PC after the call 1902 } 1903 1904 1905 // 1906 // JVMTI ForceEarlyReturn support 1907 // 1908 1909 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) { 1910 address entry = __ pc(); 1911 1912 __ empty_expression_stack(); 1913 __ load_earlyret_value(state); 1914 1915 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch); 1916 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset()); 1917 1918 // Clear the earlyret state 1919 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr); 1920 1921 __ remove_activation(state, 1922 /* throw_monitor_exception */ false, 1923 /* install_monitor_exception */ false); 1924 1925 // The caller's SP was adjusted upon method entry to accomodate 1926 // the callee's non-argument locals. Undo that adjustment. 1927 __ ret(); // return to caller 1928 __ delayed()->restore(I5_savedSP, G0, SP); 1929 1930 return entry; 1931 } // end of JVMTI ForceEarlyReturn support 1932 1933 1934 //------------------------------------------------------------------------------------------------------------------------ 1935 // Helper for vtos entry point generation 1936 1937 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) { 1938 assert(t->is_valid() && t->tos_in() == vtos, "illegal template"); 1939 Label L; 1940 aep = __ pc(); __ push_ptr(); __ ba_short(L); 1941 fep = __ pc(); __ push_f(); __ ba_short(L); 1942 dep = __ pc(); __ push_d(); __ ba_short(L); 1943 lep = __ pc(); __ push_l(); __ ba_short(L); 1944 iep = __ pc(); __ push_i(); 1945 bep = cep = sep = iep; // there aren't any 1946 vep = __ pc(); __ bind(L); // fall through 1947 generate_and_dispatch(t); 1948 } 1949 1950 // -------------------------------------------------------------------------------- 1951 1952 // Non-product code 1953 #ifndef PRODUCT 1954 address TemplateInterpreterGenerator::generate_trace_code(TosState state) { 1955 address entry = __ pc(); 1956 1957 __ push(state); 1958 __ mov(O7, Lscratch); // protect return address within interpreter 1959 1960 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer 1961 __ mov( Otos_l2, G3_scratch ); 1962 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::trace_bytecode), G0, Otos_l1, G3_scratch); 1963 __ mov(Lscratch, O7); // restore return address 1964 __ pop(state); 1965 __ retl(); 1966 __ delayed()->nop(); 1967 1968 return entry; 1969 } 1970 1971 1972 // helpers for generate_and_dispatch 1973 1974 void TemplateInterpreterGenerator::count_bytecode() { 1975 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch); 1976 } 1977 1978 1979 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) { 1980 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch); 1981 } 1982 1983 1984 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) { 1985 AddressLiteral index (&BytecodePairHistogram::_index); 1986 AddressLiteral counters((address) &BytecodePairHistogram::_counters); 1987 1988 // get index, shift out old bytecode, bring in new bytecode, and store it 1989 // _index = (_index >> log2_number_of_codes) | 1990 // (bytecode << log2_number_of_codes); 1991 1992 __ load_contents(index, G4_scratch); 1993 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch ); 1994 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch ); 1995 __ or3( G3_scratch, G4_scratch, G4_scratch ); 1996 __ store_contents(G4_scratch, index, G3_scratch); 1997 1998 // bump bucket contents 1999 // _counters[_index] ++; 2000 2001 __ set(counters, G3_scratch); // loads into G3_scratch 2002 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address 2003 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index 2004 __ ld (G3_scratch, 0, G4_scratch); 2005 __ inc (G4_scratch); 2006 __ st (G4_scratch, 0, G3_scratch); 2007 } 2008 2009 2010 void TemplateInterpreterGenerator::trace_bytecode(Template* t) { 2011 // Call a little run-time stub to avoid blow-up for each bytecode. 2012 // The run-time runtime saves the right registers, depending on 2013 // the tosca in-state for the given template. 2014 address entry = Interpreter::trace_code(t->tos_in()); 2015 guarantee(entry != NULL, "entry must have been generated"); 2016 __ call(entry, relocInfo::none); 2017 __ delayed()->nop(); 2018 } 2019 2020 2021 void TemplateInterpreterGenerator::stop_interpreter_at() { 2022 AddressLiteral counter(&BytecodeCounter::_counter_value); 2023 __ load_contents(counter, G3_scratch); 2024 AddressLiteral stop_at(&StopInterpreterAt); 2025 __ load_ptr_contents(stop_at, G4_scratch); 2026 __ cmp(G3_scratch, G4_scratch); 2027 __ breakpoint_trap(Assembler::equal, Assembler::icc); 2028 } 2029 #endif // not PRODUCT