1 /* 2 * Copyright (c) 2014, Oracle and/or its affiliates. All rights reserved. 3 * Copyright 2013, 2014 SAP AG. All rights reserved. 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 5 * 6 * This code is free software; you can redistribute it and/or modify it 7 * under the terms of the GNU General Public License version 2 only, as 8 * published by the Free Software Foundation. 9 * 10 * This code is distributed in the hope that it will be useful, but WITHOUT 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 13 * version 2 for more details (a copy is included in the LICENSE file that 14 * accompanied this code). 15 * 16 * You should have received a copy of the GNU General Public License version 17 * 2 along with this work; if not, write to the Free Software Foundation, 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 19 * 20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 21 * or visit www.oracle.com if you need additional information or have any 22 * questions. 23 * 24 */ 25 26 #include "precompiled.hpp" 27 #ifndef CC_INTERP 28 #include "asm/macroAssembler.inline.hpp" 29 #include "interpreter/bytecodeHistogram.hpp" 30 #include "interpreter/interpreter.hpp" 31 #include "interpreter/interpreterGenerator.hpp" 32 #include "interpreter/interpreterRuntime.hpp" 33 #include "interpreter/templateTable.hpp" 34 #include "oops/arrayOop.hpp" 35 #include "oops/methodData.hpp" 36 #include "oops/method.hpp" 37 #include "oops/oop.inline.hpp" 38 #include "prims/jvmtiExport.hpp" 39 #include "prims/jvmtiThreadState.hpp" 40 #include "runtime/arguments.hpp" 41 #include "runtime/deoptimization.hpp" 42 #include "runtime/frame.inline.hpp" 43 #include "runtime/sharedRuntime.hpp" 44 #include "runtime/stubRoutines.hpp" 45 #include "runtime/synchronizer.hpp" 46 #include "runtime/timer.hpp" 47 #include "runtime/vframeArray.hpp" 48 #include "utilities/debug.hpp" 49 #include "utilities/macros.hpp" 50 51 #undef __ 52 #define __ _masm-> 53 54 #ifdef PRODUCT 55 #define BLOCK_COMMENT(str) /* nothing */ 56 #else 57 #define BLOCK_COMMENT(str) __ block_comment(str) 58 #endif 59 60 #define BIND(label) __ bind(label); BLOCK_COMMENT(#label ":") 61 62 //----------------------------------------------------------------------------- 63 64 // Actually we should never reach here since we do stack overflow checks before pushing any frame. 65 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() { 66 address entry = __ pc(); 67 __ unimplemented("generate_StackOverflowError_handler"); 68 return entry; 69 } 70 71 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) { 72 address entry = __ pc(); 73 __ empty_expression_stack(); 74 __ load_const_optimized(R4_ARG2, (address) name); 75 // Index is in R17_tos. 76 __ mr(R5_ARG3, R17_tos); 77 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException)); 78 return entry; 79 } 80 81 #if 0 82 // Call special ClassCastException constructor taking object to cast 83 // and target class as arguments. 84 address TemplateInterpreterGenerator::generate_ClassCastException_verbose_handler() { 85 address entry = __ pc(); 86 87 // Expression stack must be empty before entering the VM if an 88 // exception happened. 89 __ empty_expression_stack(); 90 91 // Thread will be loaded to R3_ARG1. 92 // Target class oop is in register R5_ARG3 by convention! 93 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ClassCastException_verbose), R17_tos, R5_ARG3); 94 // Above call must not return here since exception pending. 95 DEBUG_ONLY(__ should_not_reach_here();) 96 return entry; 97 } 98 #endif 99 100 address TemplateInterpreterGenerator::generate_ClassCastException_handler() { 101 address entry = __ pc(); 102 // Expression stack must be empty before entering the VM if an 103 // exception happened. 104 __ empty_expression_stack(); 105 106 // Load exception object. 107 // Thread will be loaded to R3_ARG1. 108 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ClassCastException), R17_tos); 109 #ifdef ASSERT 110 // Above call must not return here since exception pending. 111 __ should_not_reach_here(); 112 #endif 113 return entry; 114 } 115 116 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) { 117 address entry = __ pc(); 118 //__ untested("generate_exception_handler_common"); 119 Register Rexception = R17_tos; 120 121 // Expression stack must be empty before entering the VM if an exception happened. 122 __ empty_expression_stack(); 123 124 __ load_const_optimized(R4_ARG2, (address) name, R11_scratch1); 125 if (pass_oop) { 126 __ mr(R5_ARG3, Rexception); 127 __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), false); 128 } else { 129 __ load_const_optimized(R5_ARG3, (address) message, R11_scratch1); 130 __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), false); 131 } 132 133 // Throw exception. 134 __ mr(R3_ARG1, Rexception); 135 __ load_const_optimized(R11_scratch1, Interpreter::throw_exception_entry(), R12_scratch2); 136 __ mtctr(R11_scratch1); 137 __ bctr(); 138 139 return entry; 140 } 141 142 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) { 143 address entry = __ pc(); 144 __ unimplemented("generate_continuation_for"); 145 return entry; 146 } 147 148 // This entry is returned to when a call returns to the interpreter. 149 // When we arrive here, we expect that the callee stack frame is already popped. 150 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) { 151 address entry = __ pc(); 152 153 // Move the value out of the return register back to the TOS cache of current frame. 154 switch (state) { 155 case ltos: 156 case btos: 157 case ztos: 158 case ctos: 159 case stos: 160 case atos: 161 case itos: __ mr(R17_tos, R3_RET); break; // RET -> TOS cache 162 case ftos: 163 case dtos: __ fmr(F15_ftos, F1_RET); break; // TOS cache -> GR_FRET 164 case vtos: break; // Nothing to do, this was a void return. 165 default : ShouldNotReachHere(); 166 } 167 168 __ restore_interpreter_state(R11_scratch1); // Sets R11_scratch1 = fp. 169 __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1); 170 __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0); 171 172 // Compiled code destroys templateTableBase, reload. 173 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R12_scratch2); 174 175 if (state == atos) { 176 __ profile_return_type(R3_RET, R11_scratch1, R12_scratch2); 177 } 178 179 const Register cache = R11_scratch1; 180 const Register size = R12_scratch2; 181 __ get_cache_and_index_at_bcp(cache, 1, index_size); 182 183 // Get least significant byte of 64 bit value: 184 #if defined(VM_LITTLE_ENDIAN) 185 __ lbz(size, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset()), cache); 186 #else 187 __ lbz(size, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset()) + 7, cache); 188 #endif 189 __ sldi(size, size, Interpreter::logStackElementSize); 190 __ add(R15_esp, R15_esp, size); 191 __ dispatch_next(state, step); 192 return entry; 193 } 194 195 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) { 196 address entry = __ pc(); 197 // If state != vtos, we're returning from a native method, which put it's result 198 // into the result register. So move the value out of the return register back 199 // to the TOS cache of current frame. 200 201 switch (state) { 202 case ltos: 203 case btos: 204 case ztos: 205 case ctos: 206 case stos: 207 case atos: 208 case itos: __ mr(R17_tos, R3_RET); break; // GR_RET -> TOS cache 209 case ftos: 210 case dtos: __ fmr(F15_ftos, F1_RET); break; // TOS cache -> GR_FRET 211 case vtos: break; // Nothing to do, this was a void return. 212 default : ShouldNotReachHere(); 213 } 214 215 // Load LcpoolCache @@@ should be already set! 216 __ get_constant_pool_cache(R27_constPoolCache); 217 218 // Handle a pending exception, fall through if none. 219 __ check_and_forward_exception(R11_scratch1, R12_scratch2); 220 221 // Start executing bytecodes. 222 __ dispatch_next(state, step); 223 224 return entry; 225 } 226 227 // A result handler converts the native result into java format. 228 // Use the shared code between c++ and template interpreter. 229 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) { 230 return AbstractInterpreterGenerator::generate_result_handler_for(type); 231 } 232 233 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) { 234 address entry = __ pc(); 235 236 __ push(state); 237 __ call_VM(noreg, runtime_entry); 238 __ dispatch_via(vtos, Interpreter::_normal_table.table_for(vtos)); 239 240 return entry; 241 } 242 243 // Helpers for commoning out cases in the various type of method entries. 244 245 // Increment invocation count & check for overflow. 246 // 247 // Note: checking for negative value instead of overflow 248 // so we have a 'sticky' overflow test. 249 // 250 void TemplateInterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) { 251 // Note: In tiered we increment either counters in method or in MDO depending if we're profiling or not. 252 Register Rscratch1 = R11_scratch1; 253 Register Rscratch2 = R12_scratch2; 254 Register R3_counters = R3_ARG1; 255 Label done; 256 257 if (TieredCompilation) { 258 const int increment = InvocationCounter::count_increment; 259 const int mask = ((1 << Tier0InvokeNotifyFreqLog) - 1) << InvocationCounter::count_shift; 260 Label no_mdo; 261 if (ProfileInterpreter) { 262 const Register Rmdo = Rscratch1; 263 // If no method data exists, go to profile_continue. 264 __ ld(Rmdo, in_bytes(Method::method_data_offset()), R19_method); 265 __ cmpdi(CCR0, Rmdo, 0); 266 __ beq(CCR0, no_mdo); 267 268 // Increment invocation counter in the MDO. 269 const int mdo_bc_offs = in_bytes(MethodData::backedge_counter_offset()) + in_bytes(InvocationCounter::counter_offset()); 270 __ lwz(Rscratch2, mdo_bc_offs, Rmdo); 271 __ addi(Rscratch2, Rscratch2, increment); 272 __ stw(Rscratch2, mdo_bc_offs, Rmdo); 273 __ load_const_optimized(Rscratch1, mask, R0); 274 __ and_(Rscratch1, Rscratch2, Rscratch1); 275 __ bne(CCR0, done); 276 __ b(*overflow); 277 } 278 279 // Increment counter in MethodCounters*. 280 const int mo_ic_offs = in_bytes(MethodCounters::invocation_counter_offset()) + in_bytes(InvocationCounter::counter_offset()); 281 __ bind(no_mdo); 282 __ get_method_counters(R19_method, R3_counters, done); 283 __ lwz(Rscratch2, mo_ic_offs, R3_counters); 284 __ addi(Rscratch2, Rscratch2, increment); 285 __ stw(Rscratch2, mo_ic_offs, R3_counters); 286 __ load_const_optimized(Rscratch1, mask, R0); 287 __ and_(Rscratch1, Rscratch2, Rscratch1); 288 __ beq(CCR0, *overflow); 289 290 __ bind(done); 291 292 } else { 293 294 // Update standard invocation counters. 295 Register Rsum_ivc_bec = R4_ARG2; 296 __ get_method_counters(R19_method, R3_counters, done); 297 __ increment_invocation_counter(R3_counters, Rsum_ivc_bec, R12_scratch2); 298 // Increment interpreter invocation counter. 299 if (ProfileInterpreter) { // %%% Merge this into methodDataOop. 300 __ lwz(R12_scratch2, in_bytes(MethodCounters::interpreter_invocation_counter_offset()), R3_counters); 301 __ addi(R12_scratch2, R12_scratch2, 1); 302 __ stw(R12_scratch2, in_bytes(MethodCounters::interpreter_invocation_counter_offset()), R3_counters); 303 } 304 // Check if we must create a method data obj. 305 if (ProfileInterpreter && profile_method != NULL) { 306 const Register profile_limit = Rscratch1; 307 int pl_offs = __ load_const_optimized(profile_limit, &InvocationCounter::InterpreterProfileLimit, R0, true); 308 __ lwz(profile_limit, pl_offs, profile_limit); 309 // Test to see if we should create a method data oop. 310 __ cmpw(CCR0, Rsum_ivc_bec, profile_limit); 311 __ blt(CCR0, *profile_method_continue); 312 // If no method data exists, go to profile_method. 313 __ test_method_data_pointer(*profile_method); 314 } 315 // Finally check for counter overflow. 316 if (overflow) { 317 const Register invocation_limit = Rscratch1; 318 int il_offs = __ load_const_optimized(invocation_limit, &InvocationCounter::InterpreterInvocationLimit, R0, true); 319 __ lwz(invocation_limit, il_offs, invocation_limit); 320 assert(4 == sizeof(InvocationCounter::InterpreterInvocationLimit), "unexpected field size"); 321 __ cmpw(CCR0, Rsum_ivc_bec, invocation_limit); 322 __ bge(CCR0, *overflow); 323 } 324 325 __ bind(done); 326 } 327 } 328 329 // Generate code to initiate compilation on invocation counter overflow. 330 void TemplateInterpreterGenerator::generate_counter_overflow(Label& continue_entry) { 331 // Generate code to initiate compilation on the counter overflow. 332 333 // InterpreterRuntime::frequency_counter_overflow takes one arguments, 334 // which indicates if the counter overflow occurs at a backwards branch (NULL bcp) 335 // We pass zero in. 336 // The call returns the address of the verified entry point for the method or NULL 337 // if the compilation did not complete (either went background or bailed out). 338 // 339 // Unlike the C++ interpreter above: Check exceptions! 340 // Assumption: Caller must set the flag "do_not_unlock_if_sychronized" if the monitor of a sync'ed 341 // method has not yet been created. Thus, no unlocking of a non-existing monitor can occur. 342 343 __ li(R4_ARG2, 0); 344 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), R4_ARG2, true); 345 346 // Returns verified_entry_point or NULL. 347 // We ignore it in any case. 348 __ b(continue_entry); 349 } 350 351 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rmem_frame_size, Register Rscratch1) { 352 assert_different_registers(Rmem_frame_size, Rscratch1); 353 __ generate_stack_overflow_check_with_compare_and_throw(Rmem_frame_size, Rscratch1); 354 } 355 356 void TemplateInterpreterGenerator::unlock_method(bool check_exceptions) { 357 __ unlock_object(R26_monitor, check_exceptions); 358 } 359 360 // Lock the current method, interpreter register window must be set up! 361 void TemplateInterpreterGenerator::lock_method(Register Rflags, Register Rscratch1, Register Rscratch2, bool flags_preloaded) { 362 const Register Robj_to_lock = Rscratch2; 363 364 { 365 if (!flags_preloaded) { 366 __ lwz(Rflags, method_(access_flags)); 367 } 368 369 #ifdef ASSERT 370 // Check if methods needs synchronization. 371 { 372 Label Lok; 373 __ testbitdi(CCR0, R0, Rflags, JVM_ACC_SYNCHRONIZED_BIT); 374 __ btrue(CCR0,Lok); 375 __ stop("method doesn't need synchronization"); 376 __ bind(Lok); 377 } 378 #endif // ASSERT 379 } 380 381 // Get synchronization object to Rscratch2. 382 { 383 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 384 Label Lstatic; 385 Label Ldone; 386 387 __ testbitdi(CCR0, R0, Rflags, JVM_ACC_STATIC_BIT); 388 __ btrue(CCR0, Lstatic); 389 390 // Non-static case: load receiver obj from stack and we're done. 391 __ ld(Robj_to_lock, R18_locals); 392 __ b(Ldone); 393 394 __ bind(Lstatic); // Static case: Lock the java mirror 395 __ ld(Robj_to_lock, in_bytes(Method::const_offset()), R19_method); 396 __ ld(Robj_to_lock, in_bytes(ConstMethod::constants_offset()), Robj_to_lock); 397 __ ld(Robj_to_lock, ConstantPool::pool_holder_offset_in_bytes(), Robj_to_lock); 398 __ ld(Robj_to_lock, mirror_offset, Robj_to_lock); 399 400 __ bind(Ldone); 401 __ verify_oop(Robj_to_lock); 402 } 403 404 // Got the oop to lock => execute! 405 __ add_monitor_to_stack(true, Rscratch1, R0); 406 407 __ std(Robj_to_lock, BasicObjectLock::obj_offset_in_bytes(), R26_monitor); 408 __ lock_object(R26_monitor, Robj_to_lock); 409 } 410 411 // Generate a fixed interpreter frame for pure interpreter 412 // and I2N native transition frames. 413 // 414 // Before (stack grows downwards): 415 // 416 // | ... | 417 // |------------- | 418 // | java arg0 | 419 // | ... | 420 // | java argn | 421 // | | <- R15_esp 422 // | | 423 // |--------------| 424 // | abi_112 | 425 // | | <- R1_SP 426 // |==============| 427 // 428 // 429 // After: 430 // 431 // | ... | 432 // | java arg0 |<- R18_locals 433 // | ... | 434 // | java argn | 435 // |--------------| 436 // | | 437 // | java locals | 438 // | | 439 // |--------------| 440 // | abi_48 | 441 // |==============| 442 // | | 443 // | istate | 444 // | | 445 // |--------------| 446 // | monitor |<- R26_monitor 447 // |--------------| 448 // | |<- R15_esp 449 // | expression | 450 // | stack | 451 // | | 452 // |--------------| 453 // | | 454 // | abi_112 |<- R1_SP 455 // |==============| 456 // 457 // The top most frame needs an abi space of 112 bytes. This space is needed, 458 // since we call to c. The c function may spill their arguments to the caller 459 // frame. When we call to java, we don't need these spill slots. In order to save 460 // space on the stack, we resize the caller. However, java local reside in 461 // the caller frame and the frame has to be increased. The frame_size for the 462 // current frame was calculated based on max_stack as size for the expression 463 // stack. At the call, just a part of the expression stack might be used. 464 // We don't want to waste this space and cut the frame back accordingly. 465 // The resulting amount for resizing is calculated as follows: 466 // resize = (number_of_locals - number_of_arguments) * slot_size 467 // + (R1_SP - R15_esp) + 48 468 // 469 // The size for the callee frame is calculated: 470 // framesize = 112 + max_stack + monitor + state_size 471 // 472 // maxstack: Max number of slots on the expression stack, loaded from the method. 473 // monitor: We statically reserve room for one monitor object. 474 // state_size: We save the current state of the interpreter to this area. 475 // 476 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call, Register Rsize_of_parameters, Register Rsize_of_locals) { 477 Register parent_frame_resize = R6_ARG4, // Frame will grow by this number of bytes. 478 top_frame_size = R7_ARG5, 479 Rconst_method = R8_ARG6; 480 481 assert_different_registers(Rsize_of_parameters, Rsize_of_locals, parent_frame_resize, top_frame_size); 482 483 __ ld(Rconst_method, method_(const)); 484 __ lhz(Rsize_of_parameters /* number of params */, 485 in_bytes(ConstMethod::size_of_parameters_offset()), Rconst_method); 486 if (native_call) { 487 // If we're calling a native method, we reserve space for the worst-case signature 488 // handler varargs vector, which is max(Argument::n_register_parameters, parameter_count+2). 489 // We add two slots to the parameter_count, one for the jni 490 // environment and one for a possible native mirror. 491 Label skip_native_calculate_max_stack; 492 __ addi(top_frame_size, Rsize_of_parameters, 2); 493 __ cmpwi(CCR0, top_frame_size, Argument::n_register_parameters); 494 __ bge(CCR0, skip_native_calculate_max_stack); 495 __ li(top_frame_size, Argument::n_register_parameters); 496 __ bind(skip_native_calculate_max_stack); 497 __ sldi(Rsize_of_parameters, Rsize_of_parameters, Interpreter::logStackElementSize); 498 __ sldi(top_frame_size, top_frame_size, Interpreter::logStackElementSize); 499 __ sub(parent_frame_resize, R1_SP, R15_esp); // <0, off by Interpreter::stackElementSize! 500 assert(Rsize_of_locals == noreg, "Rsize_of_locals not initialized"); // Only relevant value is Rsize_of_parameters. 501 } else { 502 __ lhz(Rsize_of_locals /* number of params */, in_bytes(ConstMethod::size_of_locals_offset()), Rconst_method); 503 __ sldi(Rsize_of_parameters, Rsize_of_parameters, Interpreter::logStackElementSize); 504 __ sldi(Rsize_of_locals, Rsize_of_locals, Interpreter::logStackElementSize); 505 __ lhz(top_frame_size, in_bytes(ConstMethod::max_stack_offset()), Rconst_method); 506 __ sub(R11_scratch1, Rsize_of_locals, Rsize_of_parameters); // >=0 507 __ sub(parent_frame_resize, R1_SP, R15_esp); // <0, off by Interpreter::stackElementSize! 508 __ sldi(top_frame_size, top_frame_size, Interpreter::logStackElementSize); 509 __ add(parent_frame_resize, parent_frame_resize, R11_scratch1); 510 } 511 512 // Compute top frame size. 513 __ addi(top_frame_size, top_frame_size, frame::abi_reg_args_size + frame::ijava_state_size); 514 515 // Cut back area between esp and max_stack. 516 __ addi(parent_frame_resize, parent_frame_resize, frame::abi_minframe_size - Interpreter::stackElementSize); 517 518 __ round_to(top_frame_size, frame::alignment_in_bytes); 519 __ round_to(parent_frame_resize, frame::alignment_in_bytes); 520 // parent_frame_resize = (locals-parameters) - (ESP-SP-ABI48) Rounded to frame alignment size. 521 // Enlarge by locals-parameters (not in case of native_call), shrink by ESP-SP-ABI48. 522 523 { 524 // -------------------------------------------------------------------------- 525 // Stack overflow check 526 527 Label cont; 528 __ add(R11_scratch1, parent_frame_resize, top_frame_size); 529 generate_stack_overflow_check(R11_scratch1, R12_scratch2); 530 } 531 532 // Set up interpreter state registers. 533 534 __ add(R18_locals, R15_esp, Rsize_of_parameters); 535 __ ld(R27_constPoolCache, in_bytes(ConstMethod::constants_offset()), Rconst_method); 536 __ ld(R27_constPoolCache, ConstantPool::cache_offset_in_bytes(), R27_constPoolCache); 537 538 // Set method data pointer. 539 if (ProfileInterpreter) { 540 Label zero_continue; 541 __ ld(R28_mdx, method_(method_data)); 542 __ cmpdi(CCR0, R28_mdx, 0); 543 __ beq(CCR0, zero_continue); 544 __ addi(R28_mdx, R28_mdx, in_bytes(MethodData::data_offset())); 545 __ bind(zero_continue); 546 } 547 548 if (native_call) { 549 __ li(R14_bcp, 0); // Must initialize. 550 } else { 551 __ add(R14_bcp, in_bytes(ConstMethod::codes_offset()), Rconst_method); 552 } 553 554 // Resize parent frame. 555 __ mflr(R12_scratch2); 556 __ neg(parent_frame_resize, parent_frame_resize); 557 __ resize_frame(parent_frame_resize, R11_scratch1); 558 __ std(R12_scratch2, _abi(lr), R1_SP); 559 560 __ addi(R26_monitor, R1_SP, - frame::ijava_state_size); 561 __ addi(R15_esp, R26_monitor, - Interpreter::stackElementSize); 562 563 // Store values. 564 // R15_esp, R14_bcp, R26_monitor, R28_mdx are saved at java calls 565 // in InterpreterMacroAssembler::call_from_interpreter. 566 __ std(R19_method, _ijava_state_neg(method), R1_SP); 567 __ std(R21_sender_SP, _ijava_state_neg(sender_sp), R1_SP); 568 __ std(R27_constPoolCache, _ijava_state_neg(cpoolCache), R1_SP); 569 __ std(R18_locals, _ijava_state_neg(locals), R1_SP); 570 571 // Note: esp, bcp, monitor, mdx live in registers. Hence, the correct version can only 572 // be found in the frame after save_interpreter_state is done. This is always true 573 // for non-top frames. But when a signal occurs, dumping the top frame can go wrong, 574 // because e.g. frame::interpreter_frame_bcp() will not access the correct value 575 // (Enhanced Stack Trace). 576 // The signal handler does not save the interpreter state into the frame. 577 __ li(R0, 0); 578 #ifdef ASSERT 579 // Fill remaining slots with constants. 580 __ load_const_optimized(R11_scratch1, 0x5afe); 581 __ load_const_optimized(R12_scratch2, 0xdead); 582 #endif 583 // We have to initialize some frame slots for native calls (accessed by GC). 584 if (native_call) { 585 __ std(R26_monitor, _ijava_state_neg(monitors), R1_SP); 586 __ std(R14_bcp, _ijava_state_neg(bcp), R1_SP); 587 if (ProfileInterpreter) { __ std(R28_mdx, _ijava_state_neg(mdx), R1_SP); } 588 } 589 #ifdef ASSERT 590 else { 591 __ std(R12_scratch2, _ijava_state_neg(monitors), R1_SP); 592 __ std(R12_scratch2, _ijava_state_neg(bcp), R1_SP); 593 __ std(R12_scratch2, _ijava_state_neg(mdx), R1_SP); 594 } 595 __ std(R11_scratch1, _ijava_state_neg(ijava_reserved), R1_SP); 596 __ std(R12_scratch2, _ijava_state_neg(esp), R1_SP); 597 __ std(R12_scratch2, _ijava_state_neg(lresult), R1_SP); 598 __ std(R12_scratch2, _ijava_state_neg(fresult), R1_SP); 599 #endif 600 __ subf(R12_scratch2, top_frame_size, R1_SP); 601 __ std(R0, _ijava_state_neg(oop_tmp), R1_SP); 602 __ std(R12_scratch2, _ijava_state_neg(top_frame_sp), R1_SP); 603 604 // Push top frame. 605 __ push_frame(top_frame_size, R11_scratch1); 606 } 607 608 // End of helpers 609 610 // ============================================================================ 611 // Various method entries 612 // 613 614 // Empty method, generate a very fast return. We must skip this entry if 615 // someone's debugging, indicated by the flag 616 // "interp_mode" in the Thread obj. 617 // Note: empty methods are generated mostly methods that do assertions, which are 618 // disabled in the "java opt build". 619 address TemplateInterpreterGenerator::generate_empty_entry(void) { 620 if (!UseFastEmptyMethods) { 621 NOT_PRODUCT(__ should_not_reach_here();) 622 return Interpreter::entry_for_kind(Interpreter::zerolocals); 623 } 624 625 Label Lslow_path; 626 const Register Rjvmti_mode = R11_scratch1; 627 address entry = __ pc(); 628 629 __ lwz(Rjvmti_mode, thread_(interp_only_mode)); 630 __ cmpwi(CCR0, Rjvmti_mode, 0); 631 __ bne(CCR0, Lslow_path); // jvmti_mode!=0 632 633 // Noone's debuggin: Simply return. 634 // Pop c2i arguments (if any) off when we return. 635 #ifdef ASSERT 636 __ ld(R9_ARG7, 0, R1_SP); 637 __ ld(R10_ARG8, 0, R21_sender_SP); 638 __ cmpd(CCR0, R9_ARG7, R10_ARG8); 639 __ asm_assert_eq("backlink", 0x545); 640 #endif // ASSERT 641 __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started. 642 643 // And we're done. 644 __ blr(); 645 646 __ bind(Lslow_path); 647 __ branch_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals), R11_scratch1); 648 __ flush(); 649 650 return entry; 651 } 652 653 // Support abs and sqrt like in compiler. 654 // For others we can use a normal (native) entry. 655 656 inline bool math_entry_available(AbstractInterpreter::MethodKind kind) { 657 // Provide math entry with debugging on demand. 658 // Note: Debugging changes which code will get executed: 659 // Debugging or disabled InlineIntrinsics: java method will get interpreted and performs a native call. 660 // Not debugging and enabled InlineIntrinics: processor instruction will get used. 661 // Result might differ slightly due to rounding etc. 662 if (!InlineIntrinsics && (!FLAG_IS_ERGO(InlineIntrinsics))) return false; // Generate a vanilla entry. 663 664 return ((kind==Interpreter::java_lang_math_sqrt && VM_Version::has_fsqrt()) || 665 (kind==Interpreter::java_lang_math_abs)); 666 } 667 668 address TemplateInterpreterGenerator::generate_math_entry(AbstractInterpreter::MethodKind kind) { 669 if (!math_entry_available(kind)) { 670 NOT_PRODUCT(__ should_not_reach_here();) 671 return Interpreter::entry_for_kind(Interpreter::zerolocals); 672 } 673 674 Label Lslow_path; 675 const Register Rjvmti_mode = R11_scratch1; 676 address entry = __ pc(); 677 678 // Provide math entry with debugging on demand. 679 __ lwz(Rjvmti_mode, thread_(interp_only_mode)); 680 __ cmpwi(CCR0, Rjvmti_mode, 0); 681 __ bne(CCR0, Lslow_path); // jvmti_mode!=0 682 683 __ lfd(F1_RET, Interpreter::stackElementSize, R15_esp); 684 685 // Pop c2i arguments (if any) off when we return. 686 #ifdef ASSERT 687 __ ld(R9_ARG7, 0, R1_SP); 688 __ ld(R10_ARG8, 0, R21_sender_SP); 689 __ cmpd(CCR0, R9_ARG7, R10_ARG8); 690 __ asm_assert_eq("backlink", 0x545); 691 #endif // ASSERT 692 __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started. 693 694 if (kind == Interpreter::java_lang_math_sqrt) { 695 __ fsqrt(F1_RET, F1_RET); 696 } else if (kind == Interpreter::java_lang_math_abs) { 697 __ fabs(F1_RET, F1_RET); 698 } else { 699 ShouldNotReachHere(); 700 } 701 702 // And we're done. 703 __ blr(); 704 705 // Provide slow path for JVMTI case. 706 __ bind(Lslow_path); 707 __ branch_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals), R12_scratch2); 708 __ flush(); 709 710 return entry; 711 } 712 713 // Interpreter stub for calling a native method. (asm interpreter) 714 // This sets up a somewhat different looking stack for calling the 715 // native method than the typical interpreter frame setup. 716 // 717 // On entry: 718 // R19_method - method 719 // R16_thread - JavaThread* 720 // R15_esp - intptr_t* sender tos 721 // 722 // abstract stack (grows up) 723 // [ IJava (caller of JNI callee) ] <-- ASP 724 // ... 725 address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) { 726 727 address entry = __ pc(); 728 729 const bool inc_counter = UseCompiler || CountCompiledCalls; 730 731 // ----------------------------------------------------------------------------- 732 // Allocate a new frame that represents the native callee (i2n frame). 733 // This is not a full-blown interpreter frame, but in particular, the 734 // following registers are valid after this: 735 // - R19_method 736 // - R18_local (points to start of argumuments to native function) 737 // 738 // abstract stack (grows up) 739 // [ IJava (caller of JNI callee) ] <-- ASP 740 // ... 741 742 const Register signature_handler_fd = R11_scratch1; 743 const Register pending_exception = R0; 744 const Register result_handler_addr = R31; 745 const Register native_method_fd = R11_scratch1; 746 const Register access_flags = R22_tmp2; 747 const Register active_handles = R11_scratch1; // R26_monitor saved to state. 748 const Register sync_state = R12_scratch2; 749 const Register sync_state_addr = sync_state; // Address is dead after use. 750 const Register suspend_flags = R11_scratch1; 751 752 //============================================================================= 753 // Allocate new frame and initialize interpreter state. 754 755 Label exception_return; 756 Label exception_return_sync_check; 757 Label stack_overflow_return; 758 759 // Generate new interpreter state and jump to stack_overflow_return in case of 760 // a stack overflow. 761 //generate_compute_interpreter_state(stack_overflow_return); 762 763 Register size_of_parameters = R22_tmp2; 764 765 generate_fixed_frame(true, size_of_parameters, noreg /* unused */); 766 767 //============================================================================= 768 // Increment invocation counter. On overflow, entry to JNI method 769 // will be compiled. 770 Label invocation_counter_overflow, continue_after_compile; 771 if (inc_counter) { 772 if (synchronized) { 773 // Since at this point in the method invocation the exception handler 774 // would try to exit the monitor of synchronized methods which hasn't 775 // been entered yet, we set the thread local variable 776 // _do_not_unlock_if_synchronized to true. If any exception was thrown by 777 // runtime, exception handling i.e. unlock_if_synchronized_method will 778 // check this thread local flag. 779 // This flag has two effects, one is to force an unwind in the topmost 780 // interpreter frame and not perform an unlock while doing so. 781 __ li(R0, 1); 782 __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread); 783 } 784 generate_counter_incr(&invocation_counter_overflow, NULL, NULL); 785 786 BIND(continue_after_compile); 787 // Reset the _do_not_unlock_if_synchronized flag. 788 if (synchronized) { 789 __ li(R0, 0); 790 __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread); 791 } 792 } 793 794 // access_flags = method->access_flags(); 795 // Load access flags. 796 assert(access_flags->is_nonvolatile(), 797 "access_flags must be in a non-volatile register"); 798 // Type check. 799 assert(4 == sizeof(AccessFlags), "unexpected field size"); 800 __ lwz(access_flags, method_(access_flags)); 801 802 // We don't want to reload R19_method and access_flags after calls 803 // to some helper functions. 804 assert(R19_method->is_nonvolatile(), 805 "R19_method must be a non-volatile register"); 806 807 // Check for synchronized methods. Must happen AFTER invocation counter 808 // check, so method is not locked if counter overflows. 809 810 if (synchronized) { 811 lock_method(access_flags, R11_scratch1, R12_scratch2, true); 812 813 // Update monitor in state. 814 __ ld(R11_scratch1, 0, R1_SP); 815 __ std(R26_monitor, _ijava_state_neg(monitors), R11_scratch1); 816 } 817 818 // jvmti/jvmpi support 819 __ notify_method_entry(); 820 821 //============================================================================= 822 // Get and call the signature handler. 823 824 __ ld(signature_handler_fd, method_(signature_handler)); 825 Label call_signature_handler; 826 827 __ cmpdi(CCR0, signature_handler_fd, 0); 828 __ bne(CCR0, call_signature_handler); 829 830 // Method has never been called. Either generate a specialized 831 // handler or point to the slow one. 832 // 833 // Pass parameter 'false' to avoid exception check in call_VM. 834 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), R19_method, false); 835 836 // Check for an exception while looking up the target method. If we 837 // incurred one, bail. 838 __ ld(pending_exception, thread_(pending_exception)); 839 __ cmpdi(CCR0, pending_exception, 0); 840 __ bne(CCR0, exception_return_sync_check); // Has pending exception. 841 842 // Reload signature handler, it may have been created/assigned in the meanwhile. 843 __ ld(signature_handler_fd, method_(signature_handler)); 844 __ twi_0(signature_handler_fd); // Order wrt. load of klass mirror and entry point (isync is below). 845 846 BIND(call_signature_handler); 847 848 // Before we call the signature handler we push a new frame to 849 // protect the interpreter frame volatile registers when we return 850 // from jni but before we can get back to Java. 851 852 // First set the frame anchor while the SP/FP registers are 853 // convenient and the slow signature handler can use this same frame 854 // anchor. 855 856 // We have a TOP_IJAVA_FRAME here, which belongs to us. 857 __ set_top_ijava_frame_at_SP_as_last_Java_frame(R1_SP, R12_scratch2/*tmp*/); 858 859 // Now the interpreter frame (and its call chain) have been 860 // invalidated and flushed. We are now protected against eager 861 // being enabled in native code. Even if it goes eager the 862 // registers will be reloaded as clean and we will invalidate after 863 // the call so no spurious flush should be possible. 864 865 // Call signature handler and pass locals address. 866 // 867 // Our signature handlers copy required arguments to the C stack 868 // (outgoing C args), R3_ARG1 to R10_ARG8, and FARG1 to FARG13. 869 __ mr(R3_ARG1, R18_locals); 870 #if !defined(ABI_ELFv2) 871 __ ld(signature_handler_fd, 0, signature_handler_fd); 872 #endif 873 874 __ call_stub(signature_handler_fd); 875 876 // Remove the register parameter varargs slots we allocated in 877 // compute_interpreter_state. SP+16 ends up pointing to the ABI 878 // outgoing argument area. 879 // 880 // Not needed on PPC64. 881 //__ add(SP, SP, Argument::n_register_parameters*BytesPerWord); 882 883 assert(result_handler_addr->is_nonvolatile(), "result_handler_addr must be in a non-volatile register"); 884 // Save across call to native method. 885 __ mr(result_handler_addr, R3_RET); 886 887 __ isync(); // Acquire signature handler before trying to fetch the native entry point and klass mirror. 888 889 // Set up fixed parameters and call the native method. 890 // If the method is static, get mirror into R4_ARG2. 891 { 892 Label method_is_not_static; 893 // Access_flags is non-volatile and still, no need to restore it. 894 895 // Restore access flags. 896 __ testbitdi(CCR0, R0, access_flags, JVM_ACC_STATIC_BIT); 897 __ bfalse(CCR0, method_is_not_static); 898 899 // constants = method->constants(); 900 __ ld(R11_scratch1, in_bytes(Method::const_offset()), R19_method); 901 __ ld(R11_scratch1, in_bytes(ConstMethod::constants_offset()), R11_scratch1); 902 // pool_holder = method->constants()->pool_holder(); 903 __ ld(R11_scratch1/*pool_holder*/, ConstantPool::pool_holder_offset_in_bytes(), 904 R11_scratch1/*constants*/); 905 906 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 907 908 // mirror = pool_holder->klass_part()->java_mirror(); 909 __ ld(R0/*mirror*/, mirror_offset, R11_scratch1/*pool_holder*/); 910 // state->_native_mirror = mirror; 911 912 __ ld(R11_scratch1, 0, R1_SP); 913 __ std(R0/*mirror*/, _ijava_state_neg(oop_tmp), R11_scratch1); 914 // R4_ARG2 = &state->_oop_temp; 915 __ addi(R4_ARG2, R11_scratch1, _ijava_state_neg(oop_tmp)); 916 BIND(method_is_not_static); 917 } 918 919 // At this point, arguments have been copied off the stack into 920 // their JNI positions. Oops are boxed in-place on the stack, with 921 // handles copied to arguments. The result handler address is in a 922 // register. 923 924 // Pass JNIEnv address as first parameter. 925 __ addir(R3_ARG1, thread_(jni_environment)); 926 927 // Load the native_method entry before we change the thread state. 928 __ ld(native_method_fd, method_(native_function)); 929 930 //============================================================================= 931 // Transition from _thread_in_Java to _thread_in_native. As soon as 932 // we make this change the safepoint code needs to be certain that 933 // the last Java frame we established is good. The pc in that frame 934 // just needs to be near here not an actual return address. 935 936 // We use release_store_fence to update values like the thread state, where 937 // we don't want the current thread to continue until all our prior memory 938 // accesses (including the new thread state) are visible to other threads. 939 __ li(R0, _thread_in_native); 940 __ release(); 941 942 // TODO PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size"); 943 __ stw(R0, thread_(thread_state)); 944 945 if (UseMembar) { 946 __ fence(); 947 } 948 949 //============================================================================= 950 // Call the native method. Argument registers must not have been 951 // overwritten since "__ call_stub(signature_handler);" (except for 952 // ARG1 and ARG2 for static methods). 953 __ call_c(native_method_fd); 954 955 __ li(R0, 0); 956 __ ld(R11_scratch1, 0, R1_SP); 957 __ std(R3_RET, _ijava_state_neg(lresult), R11_scratch1); 958 __ stfd(F1_RET, _ijava_state_neg(fresult), R11_scratch1); 959 __ std(R0/*mirror*/, _ijava_state_neg(oop_tmp), R11_scratch1); // reset 960 961 // Note: C++ interpreter needs the following here: 962 // The frame_manager_lr field, which we use for setting the last 963 // java frame, gets overwritten by the signature handler. Restore 964 // it now. 965 //__ get_PC_trash_LR(R11_scratch1); 966 //__ std(R11_scratch1, _top_ijava_frame_abi(frame_manager_lr), R1_SP); 967 968 // Because of GC R19_method may no longer be valid. 969 970 // Block, if necessary, before resuming in _thread_in_Java state. 971 // In order for GC to work, don't clear the last_Java_sp until after 972 // blocking. 973 974 //============================================================================= 975 // Switch thread to "native transition" state before reading the 976 // synchronization state. This additional state is necessary 977 // because reading and testing the synchronization state is not 978 // atomic w.r.t. GC, as this scenario demonstrates: Java thread A, 979 // in _thread_in_native state, loads _not_synchronized and is 980 // preempted. VM thread changes sync state to synchronizing and 981 // suspends threads for GC. Thread A is resumed to finish this 982 // native method, but doesn't block here since it didn't see any 983 // synchronization in progress, and escapes. 984 985 // We use release_store_fence to update values like the thread state, where 986 // we don't want the current thread to continue until all our prior memory 987 // accesses (including the new thread state) are visible to other threads. 988 __ li(R0/*thread_state*/, _thread_in_native_trans); 989 __ release(); 990 __ stw(R0/*thread_state*/, thread_(thread_state)); 991 if (UseMembar) { 992 __ fence(); 993 } 994 // Write serialization page so that the VM thread can do a pseudo remote 995 // membar. We use the current thread pointer to calculate a thread 996 // specific offset to write to within the page. This minimizes bus 997 // traffic due to cache line collision. 998 else { 999 __ serialize_memory(R16_thread, R11_scratch1, R12_scratch2); 1000 } 1001 1002 // Now before we return to java we must look for a current safepoint 1003 // (a new safepoint can not start since we entered native_trans). 1004 // We must check here because a current safepoint could be modifying 1005 // the callers registers right this moment. 1006 1007 // Acquire isn't strictly necessary here because of the fence, but 1008 // sync_state is declared to be volatile, so we do it anyway 1009 // (cmp-br-isync on one path, release (same as acquire on PPC64) on the other path). 1010 int sync_state_offs = __ load_const_optimized(sync_state_addr, SafepointSynchronize::address_of_state(), /*temp*/R0, true); 1011 1012 // TODO PPC port assert(4 == SafepointSynchronize::sz_state(), "unexpected field size"); 1013 __ lwz(sync_state, sync_state_offs, sync_state_addr); 1014 1015 // TODO PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size"); 1016 __ lwz(suspend_flags, thread_(suspend_flags)); 1017 1018 Label sync_check_done; 1019 Label do_safepoint; 1020 // No synchronization in progress nor yet synchronized. 1021 __ cmpwi(CCR0, sync_state, SafepointSynchronize::_not_synchronized); 1022 // Not suspended. 1023 __ cmpwi(CCR1, suspend_flags, 0); 1024 1025 __ bne(CCR0, do_safepoint); 1026 __ beq(CCR1, sync_check_done); 1027 __ bind(do_safepoint); 1028 __ isync(); 1029 // Block. We do the call directly and leave the current 1030 // last_Java_frame setup undisturbed. We must save any possible 1031 // native result across the call. No oop is present. 1032 1033 __ mr(R3_ARG1, R16_thread); 1034 #if defined(ABI_ELFv2) 1035 __ call_c(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans), 1036 relocInfo::none); 1037 #else 1038 __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, JavaThread::check_special_condition_for_native_trans), 1039 relocInfo::none); 1040 #endif 1041 1042 __ bind(sync_check_done); 1043 1044 //============================================================================= 1045 // <<<<<< Back in Interpreter Frame >>>>> 1046 1047 // We are in thread_in_native_trans here and back in the normal 1048 // interpreter frame. We don't have to do anything special about 1049 // safepoints and we can switch to Java mode anytime we are ready. 1050 1051 // Note: frame::interpreter_frame_result has a dependency on how the 1052 // method result is saved across the call to post_method_exit. For 1053 // native methods it assumes that the non-FPU/non-void result is 1054 // saved in _native_lresult and a FPU result in _native_fresult. If 1055 // this changes then the interpreter_frame_result implementation 1056 // will need to be updated too. 1057 1058 // On PPC64, we have stored the result directly after the native call. 1059 1060 //============================================================================= 1061 // Back in Java 1062 1063 // We use release_store_fence to update values like the thread state, where 1064 // we don't want the current thread to continue until all our prior memory 1065 // accesses (including the new thread state) are visible to other threads. 1066 __ li(R0/*thread_state*/, _thread_in_Java); 1067 __ release(); 1068 __ stw(R0/*thread_state*/, thread_(thread_state)); 1069 if (UseMembar) { 1070 __ fence(); 1071 } 1072 1073 __ reset_last_Java_frame(); 1074 1075 // Jvmdi/jvmpi support. Whether we've got an exception pending or 1076 // not, and whether unlocking throws an exception or not, we notify 1077 // on native method exit. If we do have an exception, we'll end up 1078 // in the caller's context to handle it, so if we don't do the 1079 // notify here, we'll drop it on the floor. 1080 __ notify_method_exit(true/*native method*/, 1081 ilgl /*illegal state (not used for native methods)*/, 1082 InterpreterMacroAssembler::NotifyJVMTI, 1083 false /*check_exceptions*/); 1084 1085 //============================================================================= 1086 // Handle exceptions 1087 1088 if (synchronized) { 1089 // Don't check for exceptions since we're still in the i2n frame. Do that 1090 // manually afterwards. 1091 unlock_method(false); 1092 } 1093 1094 // Reset active handles after returning from native. 1095 // thread->active_handles()->clear(); 1096 __ ld(active_handles, thread_(active_handles)); 1097 // TODO PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size"); 1098 __ li(R0, 0); 1099 __ stw(R0, JNIHandleBlock::top_offset_in_bytes(), active_handles); 1100 1101 Label exception_return_sync_check_already_unlocked; 1102 __ ld(R0/*pending_exception*/, thread_(pending_exception)); 1103 __ cmpdi(CCR0, R0/*pending_exception*/, 0); 1104 __ bne(CCR0, exception_return_sync_check_already_unlocked); 1105 1106 //----------------------------------------------------------------------------- 1107 // No exception pending. 1108 1109 // Move native method result back into proper registers and return. 1110 // Invoke result handler (may unbox/promote). 1111 __ ld(R11_scratch1, 0, R1_SP); 1112 __ ld(R3_RET, _ijava_state_neg(lresult), R11_scratch1); 1113 __ lfd(F1_RET, _ijava_state_neg(fresult), R11_scratch1); 1114 __ call_stub(result_handler_addr); 1115 1116 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ R0, R11_scratch1, R12_scratch2); 1117 1118 // Must use the return pc which was loaded from the caller's frame 1119 // as the VM uses return-pc-patching for deoptimization. 1120 __ mtlr(R0); 1121 __ blr(); 1122 1123 //----------------------------------------------------------------------------- 1124 // An exception is pending. We call into the runtime only if the 1125 // caller was not interpreted. If it was interpreted the 1126 // interpreter will do the correct thing. If it isn't interpreted 1127 // (call stub/compiled code) we will change our return and continue. 1128 1129 BIND(exception_return_sync_check); 1130 1131 if (synchronized) { 1132 // Don't check for exceptions since we're still in the i2n frame. Do that 1133 // manually afterwards. 1134 unlock_method(false); 1135 } 1136 BIND(exception_return_sync_check_already_unlocked); 1137 1138 const Register return_pc = R31; 1139 1140 __ ld(return_pc, 0, R1_SP); 1141 __ ld(return_pc, _abi(lr), return_pc); 1142 1143 // Get the address of the exception handler. 1144 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), 1145 R16_thread, 1146 return_pc /* return pc */); 1147 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, noreg, R11_scratch1, R12_scratch2); 1148 1149 // Load the PC of the the exception handler into LR. 1150 __ mtlr(R3_RET); 1151 1152 // Load exception into R3_ARG1 and clear pending exception in thread. 1153 __ ld(R3_ARG1/*exception*/, thread_(pending_exception)); 1154 __ li(R4_ARG2, 0); 1155 __ std(R4_ARG2, thread_(pending_exception)); 1156 1157 // Load the original return pc into R4_ARG2. 1158 __ mr(R4_ARG2/*issuing_pc*/, return_pc); 1159 1160 // Return to exception handler. 1161 __ blr(); 1162 1163 //============================================================================= 1164 // Counter overflow. 1165 1166 if (inc_counter) { 1167 // Handle invocation counter overflow. 1168 __ bind(invocation_counter_overflow); 1169 1170 generate_counter_overflow(continue_after_compile); 1171 } 1172 1173 return entry; 1174 } 1175 1176 // Generic interpreted method entry to (asm) interpreter. 1177 // 1178 address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) { 1179 bool inc_counter = UseCompiler || CountCompiledCalls; 1180 address entry = __ pc(); 1181 // Generate the code to allocate the interpreter stack frame. 1182 Register Rsize_of_parameters = R4_ARG2, // Written by generate_fixed_frame. 1183 Rsize_of_locals = R5_ARG3; // Written by generate_fixed_frame. 1184 1185 generate_fixed_frame(false, Rsize_of_parameters, Rsize_of_locals); 1186 1187 #ifdef FAST_DISPATCH 1188 __ unimplemented("Fast dispatch in generate_normal_entry"); 1189 #if 0 1190 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables); 1191 // Set bytecode dispatch table base. 1192 #endif 1193 #endif 1194 1195 // -------------------------------------------------------------------------- 1196 // Zero out non-parameter locals. 1197 // Note: *Always* zero out non-parameter locals as Sparc does. It's not 1198 // worth to ask the flag, just do it. 1199 Register Rslot_addr = R6_ARG4, 1200 Rnum = R7_ARG5; 1201 Label Lno_locals, Lzero_loop; 1202 1203 // Set up the zeroing loop. 1204 __ subf(Rnum, Rsize_of_parameters, Rsize_of_locals); 1205 __ subf(Rslot_addr, Rsize_of_parameters, R18_locals); 1206 __ srdi_(Rnum, Rnum, Interpreter::logStackElementSize); 1207 __ beq(CCR0, Lno_locals); 1208 __ li(R0, 0); 1209 __ mtctr(Rnum); 1210 1211 // The zero locals loop. 1212 __ bind(Lzero_loop); 1213 __ std(R0, 0, Rslot_addr); 1214 __ addi(Rslot_addr, Rslot_addr, -Interpreter::stackElementSize); 1215 __ bdnz(Lzero_loop); 1216 1217 __ bind(Lno_locals); 1218 1219 // -------------------------------------------------------------------------- 1220 // Counter increment and overflow check. 1221 Label invocation_counter_overflow, 1222 profile_method, 1223 profile_method_continue; 1224 if (inc_counter || ProfileInterpreter) { 1225 1226 Register Rdo_not_unlock_if_synchronized_addr = R11_scratch1; 1227 if (synchronized) { 1228 // Since at this point in the method invocation the exception handler 1229 // would try to exit the monitor of synchronized methods which hasn't 1230 // been entered yet, we set the thread local variable 1231 // _do_not_unlock_if_synchronized to true. If any exception was thrown by 1232 // runtime, exception handling i.e. unlock_if_synchronized_method will 1233 // check this thread local flag. 1234 // This flag has two effects, one is to force an unwind in the topmost 1235 // interpreter frame and not perform an unlock while doing so. 1236 __ li(R0, 1); 1237 __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread); 1238 } 1239 1240 // Argument and return type profiling. 1241 __ profile_parameters_type(R3_ARG1, R4_ARG2, R5_ARG3, R6_ARG4); 1242 1243 // Increment invocation counter and check for overflow. 1244 if (inc_counter) { 1245 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue); 1246 } 1247 1248 __ bind(profile_method_continue); 1249 1250 // Reset the _do_not_unlock_if_synchronized flag. 1251 if (synchronized) { 1252 __ li(R0, 0); 1253 __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread); 1254 } 1255 } 1256 1257 // -------------------------------------------------------------------------- 1258 // Locking of synchronized methods. Must happen AFTER invocation_counter 1259 // check and stack overflow check, so method is not locked if overflows. 1260 if (synchronized) { 1261 lock_method(R3_ARG1, R4_ARG2, R5_ARG3); 1262 } 1263 #ifdef ASSERT 1264 else { 1265 Label Lok; 1266 __ lwz(R0, in_bytes(Method::access_flags_offset()), R19_method); 1267 __ andi_(R0, R0, JVM_ACC_SYNCHRONIZED); 1268 __ asm_assert_eq("method needs synchronization", 0x8521); 1269 __ bind(Lok); 1270 } 1271 #endif // ASSERT 1272 1273 __ verify_thread(); 1274 1275 // -------------------------------------------------------------------------- 1276 // JVMTI support 1277 __ notify_method_entry(); 1278 1279 // -------------------------------------------------------------------------- 1280 // Start executing instructions. 1281 __ dispatch_next(vtos); 1282 1283 // -------------------------------------------------------------------------- 1284 // Out of line counter overflow and MDO creation code. 1285 if (ProfileInterpreter) { 1286 // We have decided to profile this method in the interpreter. 1287 __ bind(profile_method); 1288 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method)); 1289 __ set_method_data_pointer_for_bcp(); 1290 __ b(profile_method_continue); 1291 } 1292 1293 if (inc_counter) { 1294 // Handle invocation counter overflow. 1295 __ bind(invocation_counter_overflow); 1296 generate_counter_overflow(profile_method_continue); 1297 } 1298 return entry; 1299 } 1300 1301 // ============================================================================= 1302 // Entry points 1303 1304 address AbstractInterpreterGenerator::generate_method_entry( 1305 AbstractInterpreter::MethodKind kind) { 1306 // Determine code generation flags. 1307 bool synchronized = false; 1308 address entry_point = NULL; 1309 1310 switch (kind) { 1311 case Interpreter::zerolocals : break; 1312 case Interpreter::zerolocals_synchronized: synchronized = true; break; 1313 case Interpreter::native : entry_point = ((InterpreterGenerator*) this)->generate_native_entry(false); break; 1314 case Interpreter::native_synchronized : entry_point = ((InterpreterGenerator*) this)->generate_native_entry(true); break; 1315 case Interpreter::empty : entry_point = ((InterpreterGenerator*) this)->generate_empty_entry(); break; 1316 case Interpreter::accessor : entry_point = ((InterpreterGenerator*) this)->generate_accessor_entry(); break; 1317 case Interpreter::abstract : entry_point = ((InterpreterGenerator*) this)->generate_abstract_entry(); break; 1318 1319 case Interpreter::java_lang_math_sin : // fall thru 1320 case Interpreter::java_lang_math_cos : // fall thru 1321 case Interpreter::java_lang_math_tan : // fall thru 1322 case Interpreter::java_lang_math_abs : // fall thru 1323 case Interpreter::java_lang_math_log : // fall thru 1324 case Interpreter::java_lang_math_log10 : // fall thru 1325 case Interpreter::java_lang_math_sqrt : // fall thru 1326 case Interpreter::java_lang_math_pow : // fall thru 1327 case Interpreter::java_lang_math_exp : entry_point = ((InterpreterGenerator*) this)->generate_math_entry(kind); break; 1328 case Interpreter::java_lang_ref_reference_get 1329 : entry_point = ((InterpreterGenerator*)this)->generate_Reference_get_entry(); break; 1330 case Interpreter::java_util_zip_CRC32_update : entry_point = ((InterpreterGenerator*)this)->generate_CRC32_update_entry(); break; 1331 case Interpreter::java_util_zip_CRC32_updateBytes : entry_point = ((InterpreterGenerator*)this)->generate_CRC32_updateBytes_entry(kind); break; 1332 case Interpreter::java_util_zip_CRC32_updateByteBuffer : break; 1333 default : ShouldNotReachHere(); break; 1334 } 1335 1336 if (entry_point) { 1337 return entry_point; 1338 } 1339 1340 return ((InterpreterGenerator*) this)->generate_normal_entry(synchronized); 1341 } 1342 1343 // CRC32 Intrinsics. 1344 // 1345 // Contract on scratch and work registers. 1346 // ======================================= 1347 // 1348 // On ppc, the register set {R2..R12} is available in the interpreter as scratch/work registers. 1349 // You should, however, keep in mind that {R3_ARG1..R10_ARG8} is the C-ABI argument register set. 1350 // You can't rely on these registers across calls. 1351 // 1352 // The generators for CRC32_update and for CRC32_updateBytes use the 1353 // scratch/work register set internally, passing the work registers 1354 // as arguments to the MacroAssembler emitters as required. 1355 // 1356 // R3_ARG1..R6_ARG4 are preset to hold the incoming java arguments. 1357 // Their contents is not constant but may change according to the requirements 1358 // of the emitted code. 1359 // 1360 // All other registers from the scratch/work register set are used "internally" 1361 // and contain garbage (i.e. unpredictable values) once blr() is reached. 1362 // Basically, only R3_RET contains a defined value which is the function result. 1363 // 1364 /** 1365 * Method entry for static native methods: 1366 * int java.util.zip.CRC32.update(int crc, int b) 1367 */ 1368 address InterpreterGenerator::generate_CRC32_update_entry() { 1369 address start = __ pc(); // Remember stub start address (is rtn value). 1370 1371 if (UseCRC32Intrinsics) { 1372 Label slow_path; 1373 1374 // Safepoint check 1375 const Register sync_state = R11_scratch1; 1376 int sync_state_offs = __ load_const_optimized(sync_state, SafepointSynchronize::address_of_state(), /*temp*/R0, true); 1377 __ lwz(sync_state, sync_state_offs, sync_state); 1378 __ cmpwi(CCR0, sync_state, SafepointSynchronize::_not_synchronized); 1379 __ bne(CCR0, slow_path); 1380 1381 // We don't generate local frame and don't align stack because 1382 // we not even call stub code (we generate the code inline) 1383 // and there is no safepoint on this path. 1384 1385 // Load java parameters. 1386 // R15_esp is callers operand stack pointer, i.e. it points to the parameters. 1387 const Register argP = R15_esp; 1388 const Register crc = R3_ARG1; // crc value 1389 const Register data = R4_ARG2; // address of java byte value (kernel_crc32 needs address) 1390 const Register dataLen = R5_ARG3; // source data len (1 byte). Not used because calling the single-byte emitter. 1391 const Register table = R6_ARG4; // address of crc32 table 1392 const Register tmp = dataLen; // Reuse unused len register to show we don't actually need a separate tmp here. 1393 1394 BLOCK_COMMENT("CRC32_update {"); 1395 1396 // Arguments are reversed on java expression stack 1397 #ifdef VM_LITTLE_ENDIAN 1398 __ addi(data, argP, 0+1*wordSize); // (stack) address of byte value. Emitter expects address, not value. 1399 // Being passed as an int, the single byte is at offset +0. 1400 #else 1401 __ addi(data, argP, 3+1*wordSize); // (stack) address of byte value. Emitter expects address, not value. 1402 // Being passed from java as an int, the single byte is at offset +3. 1403 #endif 1404 __ lwz(crc, 2*wordSize, argP); // Current crc state, zero extend to 64 bit to have a clean register. 1405 1406 StubRoutines::ppc64::generate_load_crc_table_addr(_masm, table); 1407 __ kernel_crc32_singleByte(crc, data, dataLen, table, tmp); 1408 1409 // Restore caller sp for c2i case and return. 1410 __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started. 1411 __ blr(); 1412 1413 // Generate a vanilla native entry as the slow path. 1414 BLOCK_COMMENT("} CRC32_update"); 1415 BIND(slow_path); 1416 } 1417 1418 (void) generate_native_entry(false); 1419 1420 return start; 1421 } 1422 1423 // CRC32 Intrinsics. 1424 /** 1425 * Method entry for static native methods: 1426 * int java.util.zip.CRC32.updateBytes( int crc, byte[] b, int off, int len) 1427 * int java.util.zip.CRC32.updateByteBuffer(int crc, long* buf, int off, int len) 1428 */ 1429 address InterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractInterpreter::MethodKind kind) { 1430 address start = __ pc(); // Remember stub start address (is rtn value). 1431 1432 if (UseCRC32Intrinsics) { 1433 Label slow_path; 1434 1435 // Safepoint check 1436 const Register sync_state = R11_scratch1; 1437 int sync_state_offs = __ load_const_optimized(sync_state, SafepointSynchronize::address_of_state(), /*temp*/R0, true); 1438 __ lwz(sync_state, sync_state_offs, sync_state); 1439 __ cmpwi(CCR0, sync_state, SafepointSynchronize::_not_synchronized); 1440 __ bne(CCR0, slow_path); 1441 1442 // We don't generate local frame and don't align stack because 1443 // we not even call stub code (we generate the code inline) 1444 // and there is no safepoint on this path. 1445 1446 // Load parameters. 1447 // Z_esp is callers operand stack pointer, i.e. it points to the parameters. 1448 const Register argP = R15_esp; 1449 const Register crc = R3_ARG1; // crc value 1450 const Register data = R4_ARG2; // address of java byte array 1451 const Register dataLen = R5_ARG3; // source data len 1452 const Register table = R6_ARG4; // address of crc32 table 1453 1454 const Register t0 = R9; // scratch registers for crc calculation 1455 const Register t1 = R10; 1456 const Register t2 = R11; 1457 const Register t3 = R12; 1458 1459 const Register tc0 = R2; // registers to hold pre-calculated column addresses 1460 const Register tc1 = R7; 1461 const Register tc2 = R8; 1462 const Register tc3 = table; // table address is reconstructed at the end of kernel_crc32_* emitters 1463 1464 const Register tmp = t0; // Only used very locally to calculate byte buffer address. 1465 1466 // Arguments are reversed on java expression stack. 1467 // Calculate address of start element. 1468 if (kind == Interpreter::java_util_zip_CRC32_updateByteBuffer) { // Used for "updateByteBuffer direct". 1469 BLOCK_COMMENT("CRC32_updateByteBuffer {"); 1470 // crc @ (SP + 5W) (32bit) 1471 // buf @ (SP + 3W) (64bit ptr to long array) 1472 // off @ (SP + 2W) (32bit) 1473 // dataLen @ (SP + 1W) (32bit) 1474 // data = buf + off 1475 __ ld( data, 3*wordSize, argP); // start of byte buffer 1476 __ lwa( tmp, 2*wordSize, argP); // byte buffer offset 1477 __ lwa( dataLen, 1*wordSize, argP); // #bytes to process 1478 __ lwz( crc, 5*wordSize, argP); // current crc state 1479 __ add( data, data, tmp); // Add byte buffer offset. 1480 } else { // Used for "updateBytes update". 1481 BLOCK_COMMENT("CRC32_updateBytes {"); 1482 // crc @ (SP + 4W) (32bit) 1483 // buf @ (SP + 3W) (64bit ptr to byte array) 1484 // off @ (SP + 2W) (32bit) 1485 // dataLen @ (SP + 1W) (32bit) 1486 // data = buf + off + base_offset 1487 __ ld( data, 3*wordSize, argP); // start of byte buffer 1488 __ lwa( tmp, 2*wordSize, argP); // byte buffer offset 1489 __ lwa( dataLen, 1*wordSize, argP); // #bytes to process 1490 __ add( data, data, tmp); // add byte buffer offset 1491 __ lwz( crc, 4*wordSize, argP); // current crc state 1492 __ addi(data, data, arrayOopDesc::base_offset_in_bytes(T_BYTE)); 1493 } 1494 1495 StubRoutines::ppc64::generate_load_crc_table_addr(_masm, table); 1496 1497 // Performance measurements show the 1word and 2word variants to be almost equivalent, 1498 // with very light advantages for the 1word variant. We chose the 1word variant for 1499 // code compactness. 1500 __ kernel_crc32_1word(crc, data, dataLen, table, t0, t1, t2, t3, tc0, tc1, tc2, tc3); 1501 1502 // Restore caller sp for c2i case and return. 1503 __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started. 1504 __ blr(); 1505 1506 // Generate a vanilla native entry as the slow path. 1507 BLOCK_COMMENT("} CRC32_updateBytes(Buffer)"); 1508 BIND(slow_path); 1509 } 1510 1511 (void) generate_native_entry(false); 1512 1513 return start; 1514 } 1515 1516 // These should never be compiled since the interpreter will prefer 1517 // the compiled version to the intrinsic version. 1518 bool AbstractInterpreter::can_be_compiled(methodHandle m) { 1519 return !math_entry_available(method_kind(m)); 1520 } 1521 1522 // How much stack a method activation needs in stack slots. 1523 // We must calc this exactly like in generate_fixed_frame. 1524 // Note: This returns the conservative size assuming maximum alignment. 1525 int AbstractInterpreter::size_top_interpreter_activation(Method* method) { 1526 const int max_alignment_size = 2; 1527 const int abi_scratch = frame::abi_reg_args_size; 1528 return method->max_locals() + method->max_stack() + 1529 frame::interpreter_frame_monitor_size() + max_alignment_size + abi_scratch; 1530 } 1531 1532 // Returns number of stackElementWords needed for the interpreter frame with the 1533 // given sections. 1534 // This overestimates the stack by one slot in case of alignments. 1535 int AbstractInterpreter::size_activation(int max_stack, 1536 int temps, 1537 int extra_args, 1538 int monitors, 1539 int callee_params, 1540 int callee_locals, 1541 bool is_top_frame) { 1542 // Note: This calculation must exactly parallel the frame setup 1543 // in AbstractInterpreterGenerator::generate_method_entry. 1544 assert(Interpreter::stackElementWords == 1, "sanity"); 1545 const int max_alignment_space = StackAlignmentInBytes / Interpreter::stackElementSize; 1546 const int abi_scratch = is_top_frame ? (frame::abi_reg_args_size / Interpreter::stackElementSize) : 1547 (frame::abi_minframe_size / Interpreter::stackElementSize); 1548 const int size = 1549 max_stack + 1550 (callee_locals - callee_params) + 1551 monitors * frame::interpreter_frame_monitor_size() + 1552 max_alignment_space + 1553 abi_scratch + 1554 frame::ijava_state_size / Interpreter::stackElementSize; 1555 1556 // Fixed size of an interpreter frame, align to 16-byte. 1557 return (size & -2); 1558 } 1559 1560 // Fills a sceletal interpreter frame generated during deoptimizations. 1561 // 1562 // Parameters: 1563 // 1564 // interpreter_frame != NULL: 1565 // set up the method, locals, and monitors. 1566 // The frame interpreter_frame, if not NULL, is guaranteed to be the 1567 // right size, as determined by a previous call to this method. 1568 // It is also guaranteed to be walkable even though it is in a skeletal state 1569 // 1570 // is_top_frame == true: 1571 // We're processing the *oldest* interpreter frame! 1572 // 1573 // pop_frame_extra_args: 1574 // If this is != 0 we are returning to a deoptimized frame by popping 1575 // off the callee frame. We want to re-execute the call that called the 1576 // callee interpreted, but since the return to the interpreter would pop 1577 // the arguments off advance the esp by dummy popframe_extra_args slots. 1578 // Popping off those will establish the stack layout as it was before the call. 1579 // 1580 void AbstractInterpreter::layout_activation(Method* method, 1581 int tempcount, 1582 int popframe_extra_args, 1583 int moncount, 1584 int caller_actual_parameters, 1585 int callee_param_count, 1586 int callee_locals_count, 1587 frame* caller, 1588 frame* interpreter_frame, 1589 bool is_top_frame, 1590 bool is_bottom_frame) { 1591 1592 const int abi_scratch = is_top_frame ? (frame::abi_reg_args_size / Interpreter::stackElementSize) : 1593 (frame::abi_minframe_size / Interpreter::stackElementSize); 1594 1595 intptr_t* locals_base = (caller->is_interpreted_frame()) ? 1596 caller->interpreter_frame_esp() + caller_actual_parameters : 1597 caller->sp() + method->max_locals() - 1 + (frame::abi_minframe_size / Interpreter::stackElementSize) ; 1598 1599 intptr_t* monitor_base = caller->sp() - frame::ijava_state_size / Interpreter::stackElementSize ; 1600 intptr_t* monitor = monitor_base - (moncount * frame::interpreter_frame_monitor_size()); 1601 intptr_t* esp_base = monitor - 1; 1602 intptr_t* esp = esp_base - tempcount - popframe_extra_args; 1603 intptr_t* sp = (intptr_t *) (((intptr_t) (esp_base - callee_locals_count + callee_param_count - method->max_stack()- abi_scratch)) & -StackAlignmentInBytes); 1604 intptr_t* sender_sp = caller->sp() + (frame::abi_minframe_size - frame::abi_reg_args_size) / Interpreter::stackElementSize; 1605 intptr_t* top_frame_sp = is_top_frame ? sp : sp + (frame::abi_minframe_size - frame::abi_reg_args_size) / Interpreter::stackElementSize; 1606 1607 interpreter_frame->interpreter_frame_set_method(method); 1608 interpreter_frame->interpreter_frame_set_locals(locals_base); 1609 interpreter_frame->interpreter_frame_set_cpcache(method->constants()->cache()); 1610 interpreter_frame->interpreter_frame_set_esp(esp); 1611 interpreter_frame->interpreter_frame_set_monitor_end((BasicObjectLock *)monitor); 1612 interpreter_frame->interpreter_frame_set_top_frame_sp(top_frame_sp); 1613 if (!is_bottom_frame) { 1614 interpreter_frame->interpreter_frame_set_sender_sp(sender_sp); 1615 } 1616 } 1617 1618 // ============================================================================= 1619 // Exceptions 1620 1621 void TemplateInterpreterGenerator::generate_throw_exception() { 1622 Register Rexception = R17_tos, 1623 Rcontinuation = R3_RET; 1624 1625 // -------------------------------------------------------------------------- 1626 // Entry point if an method returns with a pending exception (rethrow). 1627 Interpreter::_rethrow_exception_entry = __ pc(); 1628 { 1629 __ restore_interpreter_state(R11_scratch1); // Sets R11_scratch1 = fp. 1630 __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1); 1631 __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0); 1632 1633 // Compiled code destroys templateTableBase, reload. 1634 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1); 1635 } 1636 1637 // Entry point if a interpreted method throws an exception (throw). 1638 Interpreter::_throw_exception_entry = __ pc(); 1639 { 1640 __ mr(Rexception, R3_RET); 1641 1642 __ verify_thread(); 1643 __ verify_oop(Rexception); 1644 1645 // Expression stack must be empty before entering the VM in case of an exception. 1646 __ empty_expression_stack(); 1647 // Find exception handler address and preserve exception oop. 1648 // Call C routine to find handler and jump to it. 1649 __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Rexception); 1650 __ mtctr(Rcontinuation); 1651 // Push exception for exception handler bytecodes. 1652 __ push_ptr(Rexception); 1653 1654 // Jump to exception handler (may be remove activation entry!). 1655 __ bctr(); 1656 } 1657 1658 // If the exception is not handled in the current frame the frame is 1659 // removed and the exception is rethrown (i.e. exception 1660 // continuation is _rethrow_exception). 1661 // 1662 // Note: At this point the bci is still the bxi for the instruction 1663 // which caused the exception and the expression stack is 1664 // empty. Thus, for any VM calls at this point, GC will find a legal 1665 // oop map (with empty expression stack). 1666 1667 // In current activation 1668 // tos: exception 1669 // bcp: exception bcp 1670 1671 // -------------------------------------------------------------------------- 1672 // JVMTI PopFrame support 1673 1674 Interpreter::_remove_activation_preserving_args_entry = __ pc(); 1675 { 1676 // Set the popframe_processing bit in popframe_condition indicating that we are 1677 // currently handling popframe, so that call_VMs that may happen later do not 1678 // trigger new popframe handling cycles. 1679 __ lwz(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread); 1680 __ ori(R11_scratch1, R11_scratch1, JavaThread::popframe_processing_bit); 1681 __ stw(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread); 1682 1683 // Empty the expression stack, as in normal exception handling. 1684 __ empty_expression_stack(); 1685 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false); 1686 1687 // Check to see whether we are returning to a deoptimized frame. 1688 // (The PopFrame call ensures that the caller of the popped frame is 1689 // either interpreted or compiled and deoptimizes it if compiled.) 1690 // Note that we don't compare the return PC against the 1691 // deoptimization blob's unpack entry because of the presence of 1692 // adapter frames in C2. 1693 Label Lcaller_not_deoptimized; 1694 Register return_pc = R3_ARG1; 1695 __ ld(return_pc, 0, R1_SP); 1696 __ ld(return_pc, _abi(lr), return_pc); 1697 __ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), return_pc); 1698 __ cmpdi(CCR0, R3_RET, 0); 1699 __ bne(CCR0, Lcaller_not_deoptimized); 1700 1701 // The deoptimized case. 1702 // In this case, we can't call dispatch_next() after the frame is 1703 // popped, but instead must save the incoming arguments and restore 1704 // them after deoptimization has occurred. 1705 __ ld(R4_ARG2, in_bytes(Method::const_offset()), R19_method); 1706 __ lhz(R4_ARG2 /* number of params */, in_bytes(ConstMethod::size_of_parameters_offset()), R4_ARG2); 1707 __ slwi(R4_ARG2, R4_ARG2, Interpreter::logStackElementSize); 1708 __ addi(R5_ARG3, R18_locals, Interpreter::stackElementSize); 1709 __ subf(R5_ARG3, R4_ARG2, R5_ARG3); 1710 // Save these arguments. 1711 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), R16_thread, R4_ARG2, R5_ARG3); 1712 1713 // Inform deoptimization that it is responsible for restoring these arguments. 1714 __ load_const_optimized(R11_scratch1, JavaThread::popframe_force_deopt_reexecution_bit); 1715 __ stw(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread); 1716 1717 // Return from the current method into the deoptimization blob. Will eventually 1718 // end up in the deopt interpeter entry, deoptimization prepared everything that 1719 // we will reexecute the call that called us. 1720 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*reload return_pc*/ return_pc, R11_scratch1, R12_scratch2); 1721 __ mtlr(return_pc); 1722 __ blr(); 1723 1724 // The non-deoptimized case. 1725 __ bind(Lcaller_not_deoptimized); 1726 1727 // Clear the popframe condition flag. 1728 __ li(R0, 0); 1729 __ stw(R0, in_bytes(JavaThread::popframe_condition_offset()), R16_thread); 1730 1731 // Get out of the current method and re-execute the call that called us. 1732 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ noreg, R11_scratch1, R12_scratch2); 1733 __ restore_interpreter_state(R11_scratch1); 1734 __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1); 1735 __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0); 1736 if (ProfileInterpreter) { 1737 __ set_method_data_pointer_for_bcp(); 1738 __ ld(R11_scratch1, 0, R1_SP); 1739 __ std(R28_mdx, _ijava_state_neg(mdx), R11_scratch1); 1740 } 1741 #if INCLUDE_JVMTI 1742 Label L_done; 1743 1744 __ lbz(R11_scratch1, 0, R14_bcp); 1745 __ cmpwi(CCR0, R11_scratch1, Bytecodes::_invokestatic); 1746 __ bne(CCR0, L_done); 1747 1748 // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call. 1749 // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL. 1750 __ ld(R4_ARG2, 0, R18_locals); 1751 __ MacroAssembler::call_VM(R4_ARG2, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), R4_ARG2, R19_method, R14_bcp, false); 1752 __ restore_interpreter_state(R11_scratch1, /*bcp_and_mdx_only*/ true); 1753 __ cmpdi(CCR0, R4_ARG2, 0); 1754 __ beq(CCR0, L_done); 1755 __ std(R4_ARG2, wordSize, R15_esp); 1756 __ bind(L_done); 1757 #endif // INCLUDE_JVMTI 1758 __ dispatch_next(vtos); 1759 } 1760 // end of JVMTI PopFrame support 1761 1762 // -------------------------------------------------------------------------- 1763 // Remove activation exception entry. 1764 // This is jumped to if an interpreted method can't handle an exception itself 1765 // (we come from the throw/rethrow exception entry above). We're going to call 1766 // into the VM to find the exception handler in the caller, pop the current 1767 // frame and return the handler we calculated. 1768 Interpreter::_remove_activation_entry = __ pc(); 1769 { 1770 __ pop_ptr(Rexception); 1771 __ verify_thread(); 1772 __ verify_oop(Rexception); 1773 __ std(Rexception, in_bytes(JavaThread::vm_result_offset()), R16_thread); 1774 1775 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, true); 1776 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI, false); 1777 1778 __ get_vm_result(Rexception); 1779 1780 // We are done with this activation frame; find out where to go next. 1781 // The continuation point will be an exception handler, which expects 1782 // the following registers set up: 1783 // 1784 // RET: exception oop 1785 // ARG2: Issuing PC (see generate_exception_blob()), only used if the caller is compiled. 1786 1787 Register return_pc = R31; // Needs to survive the runtime call. 1788 __ ld(return_pc, 0, R1_SP); 1789 __ ld(return_pc, _abi(lr), return_pc); 1790 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), R16_thread, return_pc); 1791 1792 // Remove the current activation. 1793 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ noreg, R11_scratch1, R12_scratch2); 1794 1795 __ mr(R4_ARG2, return_pc); 1796 __ mtlr(R3_RET); 1797 __ mr(R3_RET, Rexception); 1798 __ blr(); 1799 } 1800 } 1801 1802 // JVMTI ForceEarlyReturn support. 1803 // Returns "in the middle" of a method with a "fake" return value. 1804 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) { 1805 1806 Register Rscratch1 = R11_scratch1, 1807 Rscratch2 = R12_scratch2; 1808 1809 address entry = __ pc(); 1810 __ empty_expression_stack(); 1811 1812 __ load_earlyret_value(state, Rscratch1); 1813 1814 __ ld(Rscratch1, in_bytes(JavaThread::jvmti_thread_state_offset()), R16_thread); 1815 // Clear the earlyret state. 1816 __ li(R0, 0); 1817 __ stw(R0, in_bytes(JvmtiThreadState::earlyret_state_offset()), Rscratch1); 1818 1819 __ remove_activation(state, false, false); 1820 // Copied from TemplateTable::_return. 1821 // Restoration of lr done by remove_activation. 1822 switch (state) { 1823 // Narrow result if state is itos but result type is smaller. 1824 case itos: __ narrow(R17_tos); /* fall through */ 1825 case ltos: 1826 case btos: 1827 case ztos: 1828 case ctos: 1829 case stos: 1830 case atos: __ mr(R3_RET, R17_tos); break; 1831 case ftos: 1832 case dtos: __ fmr(F1_RET, F15_ftos); break; 1833 case vtos: // This might be a constructor. Final fields (and volatile fields on PPC64) need 1834 // to get visible before the reference to the object gets stored anywhere. 1835 __ membar(Assembler::StoreStore); break; 1836 default : ShouldNotReachHere(); 1837 } 1838 __ blr(); 1839 1840 return entry; 1841 } // end of ForceEarlyReturn support 1842 1843 //----------------------------------------------------------------------------- 1844 // Helper for vtos entry point generation 1845 1846 void TemplateInterpreterGenerator::set_vtos_entry_points(Template* t, 1847 address& bep, 1848 address& cep, 1849 address& sep, 1850 address& aep, 1851 address& iep, 1852 address& lep, 1853 address& fep, 1854 address& dep, 1855 address& vep) { 1856 assert(t->is_valid() && t->tos_in() == vtos, "illegal template"); 1857 Label L; 1858 1859 aep = __ pc(); __ push_ptr(); __ b(L); 1860 fep = __ pc(); __ push_f(); __ b(L); 1861 dep = __ pc(); __ push_d(); __ b(L); 1862 lep = __ pc(); __ push_l(); __ b(L); 1863 __ align(32, 12, 24); // align L 1864 bep = cep = sep = 1865 iep = __ pc(); __ push_i(); 1866 vep = __ pc(); 1867 __ bind(L); 1868 generate_and_dispatch(t); 1869 } 1870 1871 //----------------------------------------------------------------------------- 1872 // Generation of individual instructions 1873 1874 // helpers for generate_and_dispatch 1875 1876 InterpreterGenerator::InterpreterGenerator(StubQueue* code) 1877 : TemplateInterpreterGenerator(code) { 1878 generate_all(); // Down here so it can be "virtual". 1879 } 1880 1881 //----------------------------------------------------------------------------- 1882 1883 // Non-product code 1884 #ifndef PRODUCT 1885 address TemplateInterpreterGenerator::generate_trace_code(TosState state) { 1886 //__ flush_bundle(); 1887 address entry = __ pc(); 1888 1889 const char *bname = NULL; 1890 uint tsize = 0; 1891 switch(state) { 1892 case ftos: 1893 bname = "trace_code_ftos {"; 1894 tsize = 2; 1895 break; 1896 case btos: 1897 bname = "trace_code_btos {"; 1898 tsize = 2; 1899 break; 1900 case ztos: 1901 bname = "trace_code_ztos {"; 1902 tsize = 2; 1903 break; 1904 case ctos: 1905 bname = "trace_code_ctos {"; 1906 tsize = 2; 1907 break; 1908 case stos: 1909 bname = "trace_code_stos {"; 1910 tsize = 2; 1911 break; 1912 case itos: 1913 bname = "trace_code_itos {"; 1914 tsize = 2; 1915 break; 1916 case ltos: 1917 bname = "trace_code_ltos {"; 1918 tsize = 3; 1919 break; 1920 case atos: 1921 bname = "trace_code_atos {"; 1922 tsize = 2; 1923 break; 1924 case vtos: 1925 // Note: In case of vtos, the topmost of stack value could be a int or doubl 1926 // In case of a double (2 slots) we won't see the 2nd stack value. 1927 // Maybe we simply should print the topmost 3 stack slots to cope with the problem. 1928 bname = "trace_code_vtos {"; 1929 tsize = 2; 1930 1931 break; 1932 case dtos: 1933 bname = "trace_code_dtos {"; 1934 tsize = 3; 1935 break; 1936 default: 1937 ShouldNotReachHere(); 1938 } 1939 BLOCK_COMMENT(bname); 1940 1941 // Support short-cut for TraceBytecodesAt. 1942 // Don't call into the VM if we don't want to trace to speed up things. 1943 Label Lskip_vm_call; 1944 if (TraceBytecodesAt > 0 && TraceBytecodesAt < max_intx) { 1945 int offs1 = __ load_const_optimized(R11_scratch1, (address) &TraceBytecodesAt, R0, true); 1946 int offs2 = __ load_const_optimized(R12_scratch2, (address) &BytecodeCounter::_counter_value, R0, true); 1947 __ ld(R11_scratch1, offs1, R11_scratch1); 1948 __ lwa(R12_scratch2, offs2, R12_scratch2); 1949 __ cmpd(CCR0, R12_scratch2, R11_scratch1); 1950 __ blt(CCR0, Lskip_vm_call); 1951 } 1952 1953 __ push(state); 1954 // Load 2 topmost expression stack values. 1955 __ ld(R6_ARG4, tsize*Interpreter::stackElementSize, R15_esp); 1956 __ ld(R5_ARG3, Interpreter::stackElementSize, R15_esp); 1957 __ mflr(R31); 1958 __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), /* unused */ R4_ARG2, R5_ARG3, R6_ARG4, false); 1959 __ mtlr(R31); 1960 __ pop(state); 1961 1962 if (TraceBytecodesAt > 0 && TraceBytecodesAt < max_intx) { 1963 __ bind(Lskip_vm_call); 1964 } 1965 __ blr(); 1966 BLOCK_COMMENT("} trace_code"); 1967 return entry; 1968 } 1969 1970 void TemplateInterpreterGenerator::count_bytecode() { 1971 int offs = __ load_const_optimized(R11_scratch1, (address) &BytecodeCounter::_counter_value, R12_scratch2, true); 1972 __ lwz(R12_scratch2, offs, R11_scratch1); 1973 __ addi(R12_scratch2, R12_scratch2, 1); 1974 __ stw(R12_scratch2, offs, R11_scratch1); 1975 } 1976 1977 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) { 1978 int offs = __ load_const_optimized(R11_scratch1, (address) &BytecodeHistogram::_counters[t->bytecode()], R12_scratch2, true); 1979 __ lwz(R12_scratch2, offs, R11_scratch1); 1980 __ addi(R12_scratch2, R12_scratch2, 1); 1981 __ stw(R12_scratch2, offs, R11_scratch1); 1982 } 1983 1984 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) { 1985 const Register addr = R11_scratch1, 1986 tmp = R12_scratch2; 1987 // Get index, shift out old bytecode, bring in new bytecode, and store it. 1988 // _index = (_index >> log2_number_of_codes) | 1989 // (bytecode << log2_number_of_codes); 1990 int offs1 = __ load_const_optimized(addr, (address)&BytecodePairHistogram::_index, tmp, true); 1991 __ lwz(tmp, offs1, addr); 1992 __ srwi(tmp, tmp, BytecodePairHistogram::log2_number_of_codes); 1993 __ ori(tmp, tmp, ((int) t->bytecode()) << BytecodePairHistogram::log2_number_of_codes); 1994 __ stw(tmp, offs1, addr); 1995 1996 // Bump bucket contents. 1997 // _counters[_index] ++; 1998 int offs2 = __ load_const_optimized(addr, (address)&BytecodePairHistogram::_counters, R0, true); 1999 __ sldi(tmp, tmp, LogBytesPerInt); 2000 __ add(addr, tmp, addr); 2001 __ lwz(tmp, offs2, addr); 2002 __ addi(tmp, tmp, 1); 2003 __ stw(tmp, offs2, addr); 2004 } 2005 2006 void TemplateInterpreterGenerator::trace_bytecode(Template* t) { 2007 // Call a little run-time stub to avoid blow-up for each bytecode. 2008 // The run-time runtime saves the right registers, depending on 2009 // the tosca in-state for the given template. 2010 2011 assert(Interpreter::trace_code(t->tos_in()) != NULL, 2012 "entry must have been generated"); 2013 2014 // Note: we destroy LR here. 2015 __ bl(Interpreter::trace_code(t->tos_in())); 2016 } 2017 2018 void TemplateInterpreterGenerator::stop_interpreter_at() { 2019 Label L; 2020 int offs1 = __ load_const_optimized(R11_scratch1, (address) &StopInterpreterAt, R0, true); 2021 int offs2 = __ load_const_optimized(R12_scratch2, (address) &BytecodeCounter::_counter_value, R0, true); 2022 __ ld(R11_scratch1, offs1, R11_scratch1); 2023 __ lwa(R12_scratch2, offs2, R12_scratch2); 2024 __ cmpd(CCR0, R12_scratch2, R11_scratch1); 2025 __ bne(CCR0, L); 2026 __ illtrap(); 2027 __ bind(L); 2028 } 2029 2030 #endif // !PRODUCT 2031 #endif // !CC_INTERP