1 /* 2 * Copyright (c) 2003, 2024, Oracle and/or its affiliates. All rights reserved. 3 * Copyright (c) 2014, 2021, Red Hat Inc. All rights reserved. 4 * Copyright (c) 2021, Azul Systems, Inc. All rights reserved. 5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 6 * 7 * This code is free software; you can redistribute it and/or modify it 8 * under the terms of the GNU General Public License version 2 only, as 9 * published by the Free Software Foundation. 10 * 11 * This code is distributed in the hope that it will be useful, but WITHOUT 12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 14 * version 2 for more details (a copy is included in the LICENSE file that 15 * accompanied this code). 16 * 17 * You should have received a copy of the GNU General Public License version 18 * 2 along with this work; if not, write to the Free Software Foundation, 19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 20 * 21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 22 * or visit www.oracle.com if you need additional information or have any 23 * questions. 24 * 25 */ 26 27 #include "precompiled.hpp" 28 #include "asm/macroAssembler.hpp" 29 #include "asm/macroAssembler.inline.hpp" 30 #include "code/codeCache.hpp" 31 #include "code/compiledIC.hpp" 32 #include "code/debugInfoRec.hpp" 33 #include "code/vtableStubs.hpp" 34 #include "compiler/oopMap.hpp" 35 #include "gc/shared/barrierSetAssembler.hpp" 36 #include "interpreter/interpreter.hpp" 37 #include "interpreter/interp_masm.hpp" 38 #include "logging/log.hpp" 39 #include "memory/resourceArea.hpp" 40 #include "nativeInst_aarch64.hpp" 41 #include "oops/klass.inline.hpp" 42 #include "oops/method.inline.hpp" 43 #include "prims/methodHandles.hpp" 44 #include "runtime/continuation.hpp" 45 #include "runtime/continuationEntry.inline.hpp" 46 #include "runtime/globals.hpp" 47 #include "runtime/jniHandles.hpp" 48 #include "runtime/safepointMechanism.hpp" 49 #include "runtime/sharedRuntime.hpp" 50 #include "runtime/signature.hpp" 51 #include "runtime/stubRoutines.hpp" 52 #include "runtime/vframeArray.hpp" 53 #include "utilities/align.hpp" 54 #include "utilities/formatBuffer.hpp" 55 #include "vmreg_aarch64.inline.hpp" 56 #ifdef COMPILER1 57 #include "c1/c1_Runtime1.hpp" 58 #endif 59 #ifdef COMPILER2 60 #include "adfiles/ad_aarch64.hpp" 61 #include "opto/runtime.hpp" 62 #endif 63 #if INCLUDE_JVMCI 64 #include "jvmci/jvmciJavaClasses.hpp" 65 #endif 66 67 #define __ masm-> 68 69 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size; 70 71 class SimpleRuntimeFrame { 72 73 public: 74 75 // Most of the runtime stubs have this simple frame layout. 76 // This class exists to make the layout shared in one place. 77 // Offsets are for compiler stack slots, which are jints. 78 enum layout { 79 // The frame sender code expects that rbp will be in the "natural" place and 80 // will override any oopMap setting for it. We must therefore force the layout 81 // so that it agrees with the frame sender code. 82 // we don't expect any arg reg save area so aarch64 asserts that 83 // frame::arg_reg_save_area_bytes == 0 84 rfp_off = 0, 85 rfp_off2, 86 return_off, return_off2, 87 framesize 88 }; 89 }; 90 91 // FIXME -- this is used by C1 92 class RegisterSaver { 93 const bool _save_vectors; 94 public: 95 RegisterSaver(bool save_vectors) : _save_vectors(save_vectors) {} 96 97 OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words); 98 void restore_live_registers(MacroAssembler* masm); 99 100 // Offsets into the register save area 101 // Used by deoptimization when it is managing result register 102 // values on its own 103 104 int reg_offset_in_bytes(Register r); 105 int r0_offset_in_bytes() { return reg_offset_in_bytes(r0); } 106 int rscratch1_offset_in_bytes() { return reg_offset_in_bytes(rscratch1); } 107 int v0_offset_in_bytes(); 108 109 // Total stack size in bytes for saving sve predicate registers. 110 int total_sve_predicate_in_bytes(); 111 112 // Capture info about frame layout 113 // Note this is only correct when not saving full vectors. 114 enum layout { 115 fpu_state_off = 0, 116 fpu_state_end = fpu_state_off + FPUStateSizeInWords - 1, 117 // The frame sender code expects that rfp will be in 118 // the "natural" place and will override any oopMap 119 // setting for it. We must therefore force the layout 120 // so that it agrees with the frame sender code. 121 r0_off = fpu_state_off + FPUStateSizeInWords, 122 rfp_off = r0_off + (Register::number_of_registers - 2) * Register::max_slots_per_register, 123 return_off = rfp_off + Register::max_slots_per_register, // slot for return address 124 reg_save_size = return_off + Register::max_slots_per_register}; 125 126 }; 127 128 int RegisterSaver::reg_offset_in_bytes(Register r) { 129 // The integer registers are located above the floating point 130 // registers in the stack frame pushed by save_live_registers() so the 131 // offset depends on whether we are saving full vectors, and whether 132 // those vectors are NEON or SVE. 133 134 int slots_per_vect = FloatRegister::save_slots_per_register; 135 136 #if COMPILER2_OR_JVMCI 137 if (_save_vectors) { 138 slots_per_vect = FloatRegister::slots_per_neon_register; 139 140 #ifdef COMPILER2 141 if (Matcher::supports_scalable_vector()) { 142 slots_per_vect = Matcher::scalable_vector_reg_size(T_FLOAT); 143 } 144 #endif 145 } 146 #endif 147 148 int r0_offset = v0_offset_in_bytes() + (slots_per_vect * FloatRegister::number_of_registers) * BytesPerInt; 149 return r0_offset + r->encoding() * wordSize; 150 } 151 152 int RegisterSaver::v0_offset_in_bytes() { 153 // The floating point registers are located above the predicate registers if 154 // they are present in the stack frame pushed by save_live_registers(). So the 155 // offset depends on the saved total predicate vectors in the stack frame. 156 return (total_sve_predicate_in_bytes() / VMRegImpl::stack_slot_size) * BytesPerInt; 157 } 158 159 int RegisterSaver::total_sve_predicate_in_bytes() { 160 #ifdef COMPILER2 161 if (_save_vectors && Matcher::supports_scalable_vector()) { 162 return (Matcher::scalable_vector_reg_size(T_BYTE) >> LogBitsPerByte) * 163 PRegister::number_of_registers; 164 } 165 #endif 166 return 0; 167 } 168 169 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) { 170 bool use_sve = false; 171 int sve_vector_size_in_bytes = 0; 172 int sve_vector_size_in_slots = 0; 173 int sve_predicate_size_in_slots = 0; 174 int total_predicate_in_bytes = total_sve_predicate_in_bytes(); 175 int total_predicate_in_slots = total_predicate_in_bytes / VMRegImpl::stack_slot_size; 176 177 #ifdef COMPILER2 178 use_sve = Matcher::supports_scalable_vector(); 179 if (use_sve) { 180 sve_vector_size_in_bytes = Matcher::scalable_vector_reg_size(T_BYTE); 181 sve_vector_size_in_slots = Matcher::scalable_vector_reg_size(T_FLOAT); 182 sve_predicate_size_in_slots = Matcher::scalable_predicate_reg_slots(); 183 } 184 #endif 185 186 #if COMPILER2_OR_JVMCI 187 if (_save_vectors) { 188 int extra_save_slots_per_register = 0; 189 // Save upper half of vector registers 190 if (use_sve) { 191 extra_save_slots_per_register = sve_vector_size_in_slots - FloatRegister::save_slots_per_register; 192 } else { 193 extra_save_slots_per_register = FloatRegister::extra_save_slots_per_neon_register; 194 } 195 int extra_vector_bytes = extra_save_slots_per_register * 196 VMRegImpl::stack_slot_size * 197 FloatRegister::number_of_registers; 198 additional_frame_words += ((extra_vector_bytes + total_predicate_in_bytes) / wordSize); 199 } 200 #else 201 assert(!_save_vectors, "vectors are generated only by C2 and JVMCI"); 202 #endif 203 204 int frame_size_in_bytes = align_up(additional_frame_words * wordSize + 205 reg_save_size * BytesPerInt, 16); 206 // OopMap frame size is in compiler stack slots (jint's) not bytes or words 207 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt; 208 // The caller will allocate additional_frame_words 209 int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt; 210 // CodeBlob frame size is in words. 211 int frame_size_in_words = frame_size_in_bytes / wordSize; 212 *total_frame_words = frame_size_in_words; 213 214 // Save Integer and Float registers. 215 __ enter(); 216 __ push_CPU_state(_save_vectors, use_sve, sve_vector_size_in_bytes, total_predicate_in_bytes); 217 218 // Set an oopmap for the call site. This oopmap will map all 219 // oop-registers and debug-info registers as callee-saved. This 220 // will allow deoptimization at this safepoint to find all possible 221 // debug-info recordings, as well as let GC find all oops. 222 223 OopMapSet *oop_maps = new OopMapSet(); 224 OopMap* oop_map = new OopMap(frame_size_in_slots, 0); 225 226 for (int i = 0; i < Register::number_of_registers; i++) { 227 Register r = as_Register(i); 228 if (i <= rfp->encoding() && r != rscratch1 && r != rscratch2) { 229 // SP offsets are in 4-byte words. 230 // Register slots are 8 bytes wide, 32 floating-point registers. 231 int sp_offset = Register::max_slots_per_register * i + 232 FloatRegister::save_slots_per_register * FloatRegister::number_of_registers; 233 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset + additional_frame_slots), r->as_VMReg()); 234 } 235 } 236 237 for (int i = 0; i < FloatRegister::number_of_registers; i++) { 238 FloatRegister r = as_FloatRegister(i); 239 int sp_offset = 0; 240 if (_save_vectors) { 241 sp_offset = use_sve ? (total_predicate_in_slots + sve_vector_size_in_slots * i) : 242 (FloatRegister::slots_per_neon_register * i); 243 } else { 244 sp_offset = FloatRegister::save_slots_per_register * i; 245 } 246 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset), r->as_VMReg()); 247 } 248 249 return oop_map; 250 } 251 252 void RegisterSaver::restore_live_registers(MacroAssembler* masm) { 253 #ifdef COMPILER2 254 __ pop_CPU_state(_save_vectors, Matcher::supports_scalable_vector(), 255 Matcher::scalable_vector_reg_size(T_BYTE), total_sve_predicate_in_bytes()); 256 #else 257 #if !INCLUDE_JVMCI 258 assert(!_save_vectors, "vectors are generated only by C2 and JVMCI"); 259 #endif 260 __ pop_CPU_state(_save_vectors); 261 #endif 262 __ ldp(rfp, lr, Address(__ post(sp, 2 * wordSize))); 263 __ authenticate_return_address(); 264 } 265 266 // Is vector's size (in bytes) bigger than a size saved by default? 267 // 8 bytes vector registers are saved by default on AArch64. 268 // The SVE supported min vector size is 8 bytes and we need to save 269 // predicate registers when the vector size is 8 bytes as well. 270 bool SharedRuntime::is_wide_vector(int size) { 271 return size > 8 || (UseSVE > 0 && size >= 8); 272 } 273 274 // --------------------------------------------------------------------------- 275 // Read the array of BasicTypes from a signature, and compute where the 276 // arguments should go. Values in the VMRegPair regs array refer to 4-byte 277 // quantities. Values less than VMRegImpl::stack0 are registers, those above 278 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer 279 // as framesizes are fixed. 280 // VMRegImpl::stack0 refers to the first slot 0(sp). 281 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. 282 // Register up to Register::number_of_registers are the 64-bit 283 // integer registers. 284 285 // Note: the INPUTS in sig_bt are in units of Java argument words, 286 // which are 64-bit. The OUTPUTS are in 32-bit units. 287 288 // The Java calling convention is a "shifted" version of the C ABI. 289 // By skipping the first C ABI register we can call non-static jni 290 // methods with small numbers of arguments without having to shuffle 291 // the arguments at all. Since we control the java ABI we ought to at 292 // least get some advantage out of it. 293 294 int SharedRuntime::java_calling_convention(const BasicType *sig_bt, 295 VMRegPair *regs, 296 int total_args_passed) { 297 298 // Create the mapping between argument positions and 299 // registers. 300 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = { 301 j_rarg0, j_rarg1, j_rarg2, j_rarg3, j_rarg4, j_rarg5, j_rarg6, j_rarg7 302 }; 303 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = { 304 j_farg0, j_farg1, j_farg2, j_farg3, 305 j_farg4, j_farg5, j_farg6, j_farg7 306 }; 307 308 309 uint int_args = 0; 310 uint fp_args = 0; 311 uint stk_args = 0; 312 313 for (int i = 0; i < total_args_passed; i++) { 314 switch (sig_bt[i]) { 315 case T_BOOLEAN: 316 case T_CHAR: 317 case T_BYTE: 318 case T_SHORT: 319 case T_INT: 320 if (int_args < Argument::n_int_register_parameters_j) { 321 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 322 } else { 323 stk_args = align_up(stk_args, 2); 324 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 325 stk_args += 1; 326 } 327 break; 328 case T_VOID: 329 // halves of T_LONG or T_DOUBLE 330 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 331 regs[i].set_bad(); 332 break; 333 case T_LONG: 334 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 335 // fall through 336 case T_OBJECT: 337 case T_ARRAY: 338 case T_ADDRESS: 339 if (int_args < Argument::n_int_register_parameters_j) { 340 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg()); 341 } else { 342 stk_args = align_up(stk_args, 2); 343 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 344 stk_args += 2; 345 } 346 break; 347 case T_FLOAT: 348 if (fp_args < Argument::n_float_register_parameters_j) { 349 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg()); 350 } else { 351 stk_args = align_up(stk_args, 2); 352 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 353 stk_args += 1; 354 } 355 break; 356 case T_DOUBLE: 357 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 358 if (fp_args < Argument::n_float_register_parameters_j) { 359 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg()); 360 } else { 361 stk_args = align_up(stk_args, 2); 362 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 363 stk_args += 2; 364 } 365 break; 366 default: 367 ShouldNotReachHere(); 368 break; 369 } 370 } 371 372 return stk_args; 373 } 374 375 // Patch the callers callsite with entry to compiled code if it exists. 376 static void patch_callers_callsite(MacroAssembler *masm) { 377 Label L; 378 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset()))); 379 __ cbz(rscratch1, L); 380 381 __ enter(); 382 __ push_CPU_state(); 383 384 // VM needs caller's callsite 385 // VM needs target method 386 // This needs to be a long call since we will relocate this adapter to 387 // the codeBuffer and it may not reach 388 389 #ifndef PRODUCT 390 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 391 #endif 392 393 __ mov(c_rarg0, rmethod); 394 __ mov(c_rarg1, lr); 395 __ authenticate_return_address(c_rarg1); 396 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite))); 397 __ blr(rscratch1); 398 399 // Explicit isb required because fixup_callers_callsite may change the code 400 // stream. 401 __ safepoint_isb(); 402 403 __ pop_CPU_state(); 404 // restore sp 405 __ leave(); 406 __ bind(L); 407 } 408 409 static void gen_c2i_adapter(MacroAssembler *masm, 410 int total_args_passed, 411 int comp_args_on_stack, 412 const BasicType *sig_bt, 413 const VMRegPair *regs, 414 Label& skip_fixup) { 415 // Before we get into the guts of the C2I adapter, see if we should be here 416 // at all. We've come from compiled code and are attempting to jump to the 417 // interpreter, which means the caller made a static call to get here 418 // (vcalls always get a compiled target if there is one). Check for a 419 // compiled target. If there is one, we need to patch the caller's call. 420 patch_callers_callsite(masm); 421 422 __ bind(skip_fixup); 423 424 int words_pushed = 0; 425 426 // Since all args are passed on the stack, total_args_passed * 427 // Interpreter::stackElementSize is the space we need. 428 429 int extraspace = total_args_passed * Interpreter::stackElementSize; 430 431 __ mov(r19_sender_sp, sp); 432 433 // stack is aligned, keep it that way 434 extraspace = align_up(extraspace, 2*wordSize); 435 436 if (extraspace) 437 __ sub(sp, sp, extraspace); 438 439 // Now write the args into the outgoing interpreter space 440 for (int i = 0; i < total_args_passed; i++) { 441 if (sig_bt[i] == T_VOID) { 442 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 443 continue; 444 } 445 446 // offset to start parameters 447 int st_off = (total_args_passed - i - 1) * Interpreter::stackElementSize; 448 int next_off = st_off - Interpreter::stackElementSize; 449 450 // Say 4 args: 451 // i st_off 452 // 0 32 T_LONG 453 // 1 24 T_VOID 454 // 2 16 T_OBJECT 455 // 3 8 T_BOOL 456 // - 0 return address 457 // 458 // However to make thing extra confusing. Because we can fit a Java long/double in 459 // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter 460 // leaves one slot empty and only stores to a single slot. In this case the 461 // slot that is occupied is the T_VOID slot. See I said it was confusing. 462 463 VMReg r_1 = regs[i].first(); 464 VMReg r_2 = regs[i].second(); 465 if (!r_1->is_valid()) { 466 assert(!r_2->is_valid(), ""); 467 continue; 468 } 469 if (r_1->is_stack()) { 470 // memory to memory use rscratch1 471 int ld_off = (r_1->reg2stack() * VMRegImpl::stack_slot_size 472 + extraspace 473 + words_pushed * wordSize); 474 if (!r_2->is_valid()) { 475 // sign extend?? 476 __ ldrw(rscratch1, Address(sp, ld_off)); 477 __ str(rscratch1, Address(sp, st_off)); 478 479 } else { 480 481 __ ldr(rscratch1, Address(sp, ld_off)); 482 483 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG 484 // T_DOUBLE and T_LONG use two slots in the interpreter 485 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 486 // ld_off == LSW, ld_off+wordSize == MSW 487 // st_off == MSW, next_off == LSW 488 __ str(rscratch1, Address(sp, next_off)); 489 #ifdef ASSERT 490 // Overwrite the unused slot with known junk 491 __ mov(rscratch1, (uint64_t)0xdeadffffdeadaaaaull); 492 __ str(rscratch1, Address(sp, st_off)); 493 #endif /* ASSERT */ 494 } else { 495 __ str(rscratch1, Address(sp, st_off)); 496 } 497 } 498 } else if (r_1->is_Register()) { 499 Register r = r_1->as_Register(); 500 if (!r_2->is_valid()) { 501 // must be only an int (or less ) so move only 32bits to slot 502 // why not sign extend?? 503 __ str(r, Address(sp, st_off)); 504 } else { 505 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG 506 // T_DOUBLE and T_LONG use two slots in the interpreter 507 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 508 // jlong/double in gpr 509 #ifdef ASSERT 510 // Overwrite the unused slot with known junk 511 __ mov(rscratch1, (uint64_t)0xdeadffffdeadaaabull); 512 __ str(rscratch1, Address(sp, st_off)); 513 #endif /* ASSERT */ 514 __ str(r, Address(sp, next_off)); 515 } else { 516 __ str(r, Address(sp, st_off)); 517 } 518 } 519 } else { 520 assert(r_1->is_FloatRegister(), ""); 521 if (!r_2->is_valid()) { 522 // only a float use just part of the slot 523 __ strs(r_1->as_FloatRegister(), Address(sp, st_off)); 524 } else { 525 #ifdef ASSERT 526 // Overwrite the unused slot with known junk 527 __ mov(rscratch1, (uint64_t)0xdeadffffdeadaaacull); 528 __ str(rscratch1, Address(sp, st_off)); 529 #endif /* ASSERT */ 530 __ strd(r_1->as_FloatRegister(), Address(sp, next_off)); 531 } 532 } 533 } 534 535 __ mov(esp, sp); // Interp expects args on caller's expression stack 536 537 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::interpreter_entry_offset()))); 538 __ br(rscratch1); 539 } 540 541 542 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, 543 int total_args_passed, 544 int comp_args_on_stack, 545 const BasicType *sig_bt, 546 const VMRegPair *regs) { 547 548 // Note: r19_sender_sp contains the senderSP on entry. We must 549 // preserve it since we may do a i2c -> c2i transition if we lose a 550 // race where compiled code goes non-entrant while we get args 551 // ready. 552 553 // Adapters are frameless. 554 555 // An i2c adapter is frameless because the *caller* frame, which is 556 // interpreted, routinely repairs its own esp (from 557 // interpreter_frame_last_sp), even if a callee has modified the 558 // stack pointer. It also recalculates and aligns sp. 559 560 // A c2i adapter is frameless because the *callee* frame, which is 561 // interpreted, routinely repairs its caller's sp (from sender_sp, 562 // which is set up via the senderSP register). 563 564 // In other words, if *either* the caller or callee is interpreted, we can 565 // get the stack pointer repaired after a call. 566 567 // This is why c2i and i2c adapters cannot be indefinitely composed. 568 // In particular, if a c2i adapter were to somehow call an i2c adapter, 569 // both caller and callee would be compiled methods, and neither would 570 // clean up the stack pointer changes performed by the two adapters. 571 // If this happens, control eventually transfers back to the compiled 572 // caller, but with an uncorrected stack, causing delayed havoc. 573 574 if (VerifyAdapterCalls && 575 (Interpreter::code() != nullptr || StubRoutines::final_stubs_code() != nullptr)) { 576 #if 0 577 // So, let's test for cascading c2i/i2c adapters right now. 578 // assert(Interpreter::contains($return_addr) || 579 // StubRoutines::contains($return_addr), 580 // "i2c adapter must return to an interpreter frame"); 581 __ block_comment("verify_i2c { "); 582 Label L_ok; 583 if (Interpreter::code() != nullptr) { 584 range_check(masm, rax, r11, 585 Interpreter::code()->code_start(), Interpreter::code()->code_end(), 586 L_ok); 587 } 588 if (StubRoutines::initial_stubs_code() != nullptr) { 589 range_check(masm, rax, r11, 590 StubRoutines::initial_stubs_code()->code_begin(), 591 StubRoutines::initial_stubs_code()->code_end(), 592 L_ok); 593 } 594 if (StubRoutines::final_stubs_code() != nullptr) { 595 range_check(masm, rax, r11, 596 StubRoutines::final_stubs_code()->code_begin(), 597 StubRoutines::final_stubs_code()->code_end(), 598 L_ok); 599 } 600 const char* msg = "i2c adapter must return to an interpreter frame"; 601 __ block_comment(msg); 602 __ stop(msg); 603 __ bind(L_ok); 604 __ block_comment("} verify_i2ce "); 605 #endif 606 } 607 608 // Cut-out for having no stack args. 609 int comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord; 610 if (comp_args_on_stack) { 611 __ sub(rscratch1, sp, comp_words_on_stack * wordSize); 612 __ andr(sp, rscratch1, -16); 613 } 614 615 // Will jump to the compiled code just as if compiled code was doing it. 616 // Pre-load the register-jump target early, to schedule it better. 617 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::from_compiled_offset()))); 618 619 #if INCLUDE_JVMCI 620 if (EnableJVMCI) { 621 // check if this call should be routed towards a specific entry point 622 __ ldr(rscratch2, Address(rthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset()))); 623 Label no_alternative_target; 624 __ cbz(rscratch2, no_alternative_target); 625 __ mov(rscratch1, rscratch2); 626 __ str(zr, Address(rthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset()))); 627 __ bind(no_alternative_target); 628 } 629 #endif // INCLUDE_JVMCI 630 631 // Now generate the shuffle code. 632 for (int i = 0; i < total_args_passed; i++) { 633 if (sig_bt[i] == T_VOID) { 634 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 635 continue; 636 } 637 638 // Pick up 0, 1 or 2 words from SP+offset. 639 640 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), 641 "scrambled load targets?"); 642 // Load in argument order going down. 643 int ld_off = (total_args_passed - i - 1)*Interpreter::stackElementSize; 644 // Point to interpreter value (vs. tag) 645 int next_off = ld_off - Interpreter::stackElementSize; 646 // 647 // 648 // 649 VMReg r_1 = regs[i].first(); 650 VMReg r_2 = regs[i].second(); 651 if (!r_1->is_valid()) { 652 assert(!r_2->is_valid(), ""); 653 continue; 654 } 655 if (r_1->is_stack()) { 656 // Convert stack slot to an SP offset (+ wordSize to account for return address ) 657 int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size; 658 if (!r_2->is_valid()) { 659 // sign extend??? 660 __ ldrsw(rscratch2, Address(esp, ld_off)); 661 __ str(rscratch2, Address(sp, st_off)); 662 } else { 663 // 664 // We are using two optoregs. This can be either T_OBJECT, 665 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates 666 // two slots but only uses one for thr T_LONG or T_DOUBLE case 667 // So we must adjust where to pick up the data to match the 668 // interpreter. 669 // 670 // Interpreter local[n] == MSW, local[n+1] == LSW however locals 671 // are accessed as negative so LSW is at LOW address 672 673 // ld_off is MSW so get LSW 674 const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)? 675 next_off : ld_off; 676 __ ldr(rscratch2, Address(esp, offset)); 677 // st_off is LSW (i.e. reg.first()) 678 __ str(rscratch2, Address(sp, st_off)); 679 } 680 } else if (r_1->is_Register()) { // Register argument 681 Register r = r_1->as_Register(); 682 if (r_2->is_valid()) { 683 // 684 // We are using two VMRegs. This can be either T_OBJECT, 685 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates 686 // two slots but only uses one for thr T_LONG or T_DOUBLE case 687 // So we must adjust where to pick up the data to match the 688 // interpreter. 689 690 const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)? 691 next_off : ld_off; 692 693 // this can be a misaligned move 694 __ ldr(r, Address(esp, offset)); 695 } else { 696 // sign extend and use a full word? 697 __ ldrw(r, Address(esp, ld_off)); 698 } 699 } else { 700 if (!r_2->is_valid()) { 701 __ ldrs(r_1->as_FloatRegister(), Address(esp, ld_off)); 702 } else { 703 __ ldrd(r_1->as_FloatRegister(), Address(esp, next_off)); 704 } 705 } 706 } 707 708 __ mov(rscratch2, rscratch1); 709 __ push_cont_fastpath(rthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about; kills rscratch1 710 __ mov(rscratch1, rscratch2); 711 712 // 6243940 We might end up in handle_wrong_method if 713 // the callee is deoptimized as we race thru here. If that 714 // happens we don't want to take a safepoint because the 715 // caller frame will look interpreted and arguments are now 716 // "compiled" so it is much better to make this transition 717 // invisible to the stack walking code. Unfortunately if 718 // we try and find the callee by normal means a safepoint 719 // is possible. So we stash the desired callee in the thread 720 // and the vm will find there should this case occur. 721 722 __ str(rmethod, Address(rthread, JavaThread::callee_target_offset())); 723 724 __ br(rscratch1); 725 } 726 727 // --------------------------------------------------------------- 728 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, 729 int total_args_passed, 730 int comp_args_on_stack, 731 const BasicType *sig_bt, 732 const VMRegPair *regs, 733 AdapterFingerPrint* fingerprint) { 734 address i2c_entry = __ pc(); 735 736 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs); 737 738 address c2i_unverified_entry = __ pc(); 739 Label skip_fixup; 740 741 Register data = rscratch2; 742 Register receiver = j_rarg0; 743 Register tmp = r10; // A call-clobbered register not used for arg passing 744 745 // ------------------------------------------------------------------------- 746 // Generate a C2I adapter. On entry we know rmethod holds the Method* during calls 747 // to the interpreter. The args start out packed in the compiled layout. They 748 // need to be unpacked into the interpreter layout. This will almost always 749 // require some stack space. We grow the current (compiled) stack, then repack 750 // the args. We finally end in a jump to the generic interpreter entry point. 751 // On exit from the interpreter, the interpreter will restore our SP (lest the 752 // compiled code, which relies solely on SP and not FP, get sick). 753 754 { 755 __ block_comment("c2i_unverified_entry {"); 756 // Method might have been compiled since the call site was patched to 757 // interpreted; if that is the case treat it as a miss so we can get 758 // the call site corrected. 759 __ ic_check(1 /* end_alignment */); 760 __ ldr(rmethod, Address(data, CompiledICData::speculated_method_offset())); 761 762 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset()))); 763 __ cbz(rscratch1, skip_fixup); 764 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 765 __ block_comment("} c2i_unverified_entry"); 766 } 767 768 address c2i_entry = __ pc(); 769 770 // Class initialization barrier for static methods 771 address c2i_no_clinit_check_entry = nullptr; 772 if (VM_Version::supports_fast_class_init_checks()) { 773 Label L_skip_barrier; 774 775 { // Bypass the barrier for non-static methods 776 __ ldrw(rscratch1, Address(rmethod, Method::access_flags_offset())); 777 __ andsw(zr, rscratch1, JVM_ACC_STATIC); 778 __ br(Assembler::EQ, L_skip_barrier); // non-static 779 } 780 781 __ load_method_holder(rscratch2, rmethod); 782 __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier); 783 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub())); 784 785 __ bind(L_skip_barrier); 786 c2i_no_clinit_check_entry = __ pc(); 787 } 788 789 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 790 bs->c2i_entry_barrier(masm); 791 792 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup); 793 794 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry, c2i_no_clinit_check_entry); 795 } 796 797 static int c_calling_convention_priv(const BasicType *sig_bt, 798 VMRegPair *regs, 799 int total_args_passed) { 800 801 // We return the amount of VMRegImpl stack slots we need to reserve for all 802 // the arguments NOT counting out_preserve_stack_slots. 803 804 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = { 805 c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5, c_rarg6, c_rarg7 806 }; 807 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = { 808 c_farg0, c_farg1, c_farg2, c_farg3, 809 c_farg4, c_farg5, c_farg6, c_farg7 810 }; 811 812 uint int_args = 0; 813 uint fp_args = 0; 814 uint stk_args = 0; // inc by 2 each time 815 816 for (int i = 0; i < total_args_passed; i++) { 817 switch (sig_bt[i]) { 818 case T_BOOLEAN: 819 case T_CHAR: 820 case T_BYTE: 821 case T_SHORT: 822 case T_INT: 823 if (int_args < Argument::n_int_register_parameters_c) { 824 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 825 } else { 826 #ifdef __APPLE__ 827 // Less-than word types are stored one after another. 828 // The code is unable to handle this so bailout. 829 return -1; 830 #endif 831 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 832 stk_args += 2; 833 } 834 break; 835 case T_LONG: 836 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 837 // fall through 838 case T_OBJECT: 839 case T_ARRAY: 840 case T_ADDRESS: 841 case T_METADATA: 842 if (int_args < Argument::n_int_register_parameters_c) { 843 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg()); 844 } else { 845 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 846 stk_args += 2; 847 } 848 break; 849 case T_FLOAT: 850 if (fp_args < Argument::n_float_register_parameters_c) { 851 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg()); 852 } else { 853 #ifdef __APPLE__ 854 // Less-than word types are stored one after another. 855 // The code is unable to handle this so bailout. 856 return -1; 857 #endif 858 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 859 stk_args += 2; 860 } 861 break; 862 case T_DOUBLE: 863 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 864 if (fp_args < Argument::n_float_register_parameters_c) { 865 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg()); 866 } else { 867 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 868 stk_args += 2; 869 } 870 break; 871 case T_VOID: // Halves of longs and doubles 872 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 873 regs[i].set_bad(); 874 break; 875 default: 876 ShouldNotReachHere(); 877 break; 878 } 879 } 880 881 return stk_args; 882 } 883 884 int SharedRuntime::vector_calling_convention(VMRegPair *regs, 885 uint num_bits, 886 uint total_args_passed) { 887 Unimplemented(); 888 return 0; 889 } 890 891 int SharedRuntime::c_calling_convention(const BasicType *sig_bt, 892 VMRegPair *regs, 893 int total_args_passed) 894 { 895 int result = c_calling_convention_priv(sig_bt, regs, total_args_passed); 896 guarantee(result >= 0, "Unsupported arguments configuration"); 897 return result; 898 } 899 900 901 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 902 // We always ignore the frame_slots arg and just use the space just below frame pointer 903 // which by this time is free to use 904 switch (ret_type) { 905 case T_FLOAT: 906 __ strs(v0, Address(rfp, -wordSize)); 907 break; 908 case T_DOUBLE: 909 __ strd(v0, Address(rfp, -wordSize)); 910 break; 911 case T_VOID: break; 912 default: { 913 __ str(r0, Address(rfp, -wordSize)); 914 } 915 } 916 } 917 918 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 919 // We always ignore the frame_slots arg and just use the space just below frame pointer 920 // which by this time is free to use 921 switch (ret_type) { 922 case T_FLOAT: 923 __ ldrs(v0, Address(rfp, -wordSize)); 924 break; 925 case T_DOUBLE: 926 __ ldrd(v0, Address(rfp, -wordSize)); 927 break; 928 case T_VOID: break; 929 default: { 930 __ ldr(r0, Address(rfp, -wordSize)); 931 } 932 } 933 } 934 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) { 935 RegSet x; 936 for ( int i = first_arg ; i < arg_count ; i++ ) { 937 if (args[i].first()->is_Register()) { 938 x = x + args[i].first()->as_Register(); 939 } else if (args[i].first()->is_FloatRegister()) { 940 __ strd(args[i].first()->as_FloatRegister(), Address(__ pre(sp, -2 * wordSize))); 941 } 942 } 943 __ push(x, sp); 944 } 945 946 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) { 947 RegSet x; 948 for ( int i = first_arg ; i < arg_count ; i++ ) { 949 if (args[i].first()->is_Register()) { 950 x = x + args[i].first()->as_Register(); 951 } else { 952 ; 953 } 954 } 955 __ pop(x, sp); 956 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) { 957 if (args[i].first()->is_Register()) { 958 ; 959 } else if (args[i].first()->is_FloatRegister()) { 960 __ ldrd(args[i].first()->as_FloatRegister(), Address(__ post(sp, 2 * wordSize))); 961 } 962 } 963 } 964 965 static void verify_oop_args(MacroAssembler* masm, 966 const methodHandle& method, 967 const BasicType* sig_bt, 968 const VMRegPair* regs) { 969 Register temp_reg = r19; // not part of any compiled calling seq 970 if (VerifyOops) { 971 for (int i = 0; i < method->size_of_parameters(); i++) { 972 if (sig_bt[i] == T_OBJECT || 973 sig_bt[i] == T_ARRAY) { 974 VMReg r = regs[i].first(); 975 assert(r->is_valid(), "bad oop arg"); 976 if (r->is_stack()) { 977 __ ldr(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size)); 978 __ verify_oop(temp_reg); 979 } else { 980 __ verify_oop(r->as_Register()); 981 } 982 } 983 } 984 } 985 } 986 987 // on exit, sp points to the ContinuationEntry 988 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) { 989 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, ""); 990 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, ""); 991 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, ""); 992 993 stack_slots += (int)ContinuationEntry::size()/wordSize; 994 __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata 995 996 OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize)/ VMRegImpl::stack_slot_size, 0 /* arg_slots*/); 997 998 __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset())); 999 __ str(rscratch1, Address(sp, ContinuationEntry::parent_offset())); 1000 __ mov(rscratch1, sp); // we can't use sp as the source in str 1001 __ str(rscratch1, Address(rthread, JavaThread::cont_entry_offset())); 1002 1003 return map; 1004 } 1005 1006 // on entry c_rarg1 points to the continuation 1007 // sp points to ContinuationEntry 1008 // c_rarg3 -- isVirtualThread 1009 static void fill_continuation_entry(MacroAssembler* masm) { 1010 #ifdef ASSERT 1011 __ movw(rscratch1, ContinuationEntry::cookie_value()); 1012 __ strw(rscratch1, Address(sp, ContinuationEntry::cookie_offset())); 1013 #endif 1014 1015 __ str (c_rarg1, Address(sp, ContinuationEntry::cont_offset())); 1016 __ strw(c_rarg3, Address(sp, ContinuationEntry::flags_offset())); 1017 __ str (zr, Address(sp, ContinuationEntry::chunk_offset())); 1018 __ strw(zr, Address(sp, ContinuationEntry::argsize_offset())); 1019 __ strw(zr, Address(sp, ContinuationEntry::pin_count_offset())); 1020 1021 __ ldr(rscratch1, Address(rthread, JavaThread::cont_fastpath_offset())); 1022 __ str(rscratch1, Address(sp, ContinuationEntry::parent_cont_fastpath_offset())); 1023 __ ldr(rscratch1, Address(rthread, JavaThread::held_monitor_count_offset())); 1024 __ str(rscratch1, Address(sp, ContinuationEntry::parent_held_monitor_count_offset())); 1025 1026 __ str(zr, Address(rthread, JavaThread::cont_fastpath_offset())); 1027 __ str(zr, Address(rthread, JavaThread::held_monitor_count_offset())); 1028 } 1029 1030 // on entry, sp points to the ContinuationEntry 1031 // on exit, rfp points to the spilled rfp in the entry frame 1032 static void continuation_enter_cleanup(MacroAssembler* masm) { 1033 #ifndef PRODUCT 1034 Label OK; 1035 __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset())); 1036 __ cmp(sp, rscratch1); 1037 __ br(Assembler::EQ, OK); 1038 __ stop("incorrect sp1"); 1039 __ bind(OK); 1040 #endif 1041 __ ldr(rscratch1, Address(sp, ContinuationEntry::parent_cont_fastpath_offset())); 1042 __ str(rscratch1, Address(rthread, JavaThread::cont_fastpath_offset())); 1043 1044 if (CheckJNICalls) { 1045 // Check if this is a virtual thread continuation 1046 Label L_skip_vthread_code; 1047 __ ldrw(rscratch1, Address(sp, ContinuationEntry::flags_offset())); 1048 __ cbzw(rscratch1, L_skip_vthread_code); 1049 1050 // If the held monitor count is > 0 and this vthread is terminating then 1051 // it failed to release a JNI monitor. So we issue the same log message 1052 // that JavaThread::exit does. 1053 __ ldr(rscratch1, Address(rthread, JavaThread::jni_monitor_count_offset())); 1054 __ cbz(rscratch1, L_skip_vthread_code); 1055 1056 // Save return value potentially containing the exception oop in callee-saved R19. 1057 __ mov(r19, r0); 1058 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::log_jni_monitor_still_held)); 1059 // Restore potential return value. 1060 __ mov(r0, r19); 1061 1062 // For vthreads we have to explicitly zero the JNI monitor count of the carrier 1063 // on termination. The held count is implicitly zeroed below when we restore from 1064 // the parent held count (which has to be zero). 1065 __ str(zr, Address(rthread, JavaThread::jni_monitor_count_offset())); 1066 1067 __ bind(L_skip_vthread_code); 1068 } 1069 #ifdef ASSERT 1070 else { 1071 // Check if this is a virtual thread continuation 1072 Label L_skip_vthread_code; 1073 __ ldrw(rscratch1, Address(sp, ContinuationEntry::flags_offset())); 1074 __ cbzw(rscratch1, L_skip_vthread_code); 1075 1076 // See comment just above. If not checking JNI calls the JNI count is only 1077 // needed for assertion checking. 1078 __ str(zr, Address(rthread, JavaThread::jni_monitor_count_offset())); 1079 1080 __ bind(L_skip_vthread_code); 1081 } 1082 #endif 1083 1084 __ ldr(rscratch1, Address(sp, ContinuationEntry::parent_held_monitor_count_offset())); 1085 __ str(rscratch1, Address(rthread, JavaThread::held_monitor_count_offset())); 1086 1087 __ ldr(rscratch2, Address(sp, ContinuationEntry::parent_offset())); 1088 __ str(rscratch2, Address(rthread, JavaThread::cont_entry_offset())); 1089 __ add(rfp, sp, (int)ContinuationEntry::size()); 1090 } 1091 1092 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread) 1093 // On entry: c_rarg1 -- the continuation object 1094 // c_rarg2 -- isContinue 1095 // c_rarg3 -- isVirtualThread 1096 static void gen_continuation_enter(MacroAssembler* masm, 1097 const methodHandle& method, 1098 const BasicType* sig_bt, 1099 const VMRegPair* regs, 1100 int& exception_offset, 1101 OopMapSet*oop_maps, 1102 int& frame_complete, 1103 int& stack_slots, 1104 int& interpreted_entry_offset, 1105 int& compiled_entry_offset) { 1106 //verify_oop_args(masm, method, sig_bt, regs); 1107 Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type); 1108 1109 address start = __ pc(); 1110 1111 Label call_thaw, exit; 1112 1113 // i2i entry used at interp_only_mode only 1114 interpreted_entry_offset = __ pc() - start; 1115 { 1116 1117 #ifdef ASSERT 1118 Label is_interp_only; 1119 __ ldrw(rscratch1, Address(rthread, JavaThread::interp_only_mode_offset())); 1120 __ cbnzw(rscratch1, is_interp_only); 1121 __ stop("enterSpecial interpreter entry called when not in interp_only_mode"); 1122 __ bind(is_interp_only); 1123 #endif 1124 1125 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter) 1126 __ ldr(c_rarg1, Address(esp, Interpreter::stackElementSize*2)); 1127 __ ldr(c_rarg2, Address(esp, Interpreter::stackElementSize*1)); 1128 __ ldr(c_rarg3, Address(esp, Interpreter::stackElementSize*0)); 1129 __ push_cont_fastpath(rthread); 1130 1131 __ enter(); 1132 stack_slots = 2; // will be adjusted in setup 1133 OopMap* map = continuation_enter_setup(masm, stack_slots); 1134 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe, 1135 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway. 1136 1137 fill_continuation_entry(masm); 1138 1139 __ cbnz(c_rarg2, call_thaw); 1140 1141 const address tr_call = __ trampoline_call(resolve); 1142 if (tr_call == nullptr) { 1143 fatal("CodeCache is full at gen_continuation_enter"); 1144 } 1145 1146 oop_maps->add_gc_map(__ pc() - start, map); 1147 __ post_call_nop(); 1148 1149 __ b(exit); 1150 1151 address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call); 1152 if (stub == nullptr) { 1153 fatal("CodeCache is full at gen_continuation_enter"); 1154 } 1155 } 1156 1157 // compiled entry 1158 __ align(CodeEntryAlignment); 1159 compiled_entry_offset = __ pc() - start; 1160 1161 __ enter(); 1162 stack_slots = 2; // will be adjusted in setup 1163 OopMap* map = continuation_enter_setup(masm, stack_slots); 1164 frame_complete = __ pc() - start; 1165 1166 fill_continuation_entry(masm); 1167 1168 __ cbnz(c_rarg2, call_thaw); 1169 1170 const address tr_call = __ trampoline_call(resolve); 1171 if (tr_call == nullptr) { 1172 fatal("CodeCache is full at gen_continuation_enter"); 1173 } 1174 1175 oop_maps->add_gc_map(__ pc() - start, map); 1176 __ post_call_nop(); 1177 1178 __ b(exit); 1179 1180 __ bind(call_thaw); 1181 1182 ContinuationEntry::_thaw_call_pc_offset = __ pc() - start; 1183 __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw())); 1184 oop_maps->add_gc_map(__ pc() - start, map->deep_copy()); 1185 ContinuationEntry::_return_pc_offset = __ pc() - start; 1186 __ post_call_nop(); 1187 1188 __ bind(exit); 1189 continuation_enter_cleanup(masm); 1190 __ leave(); 1191 __ ret(lr); 1192 1193 /// exception handling 1194 1195 exception_offset = __ pc() - start; 1196 { 1197 __ mov(r19, r0); // save return value contaning the exception oop in callee-saved R19 1198 1199 continuation_enter_cleanup(masm); 1200 1201 __ ldr(c_rarg1, Address(rfp, wordSize)); // return address 1202 __ authenticate_return_address(c_rarg1); 1203 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), rthread, c_rarg1); 1204 1205 // see OptoRuntime::generate_exception_blob: r0 -- exception oop, r3 -- exception pc 1206 1207 __ mov(r1, r0); // the exception handler 1208 __ mov(r0, r19); // restore return value contaning the exception oop 1209 __ verify_oop(r0); 1210 1211 __ leave(); 1212 __ mov(r3, lr); 1213 __ br(r1); // the exception handler 1214 } 1215 1216 address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call); 1217 if (stub == nullptr) { 1218 fatal("CodeCache is full at gen_continuation_enter"); 1219 } 1220 } 1221 1222 static void gen_continuation_yield(MacroAssembler* masm, 1223 const methodHandle& method, 1224 const BasicType* sig_bt, 1225 const VMRegPair* regs, 1226 OopMapSet* oop_maps, 1227 int& frame_complete, 1228 int& stack_slots, 1229 int& compiled_entry_offset) { 1230 enum layout { 1231 rfp_off1, 1232 rfp_off2, 1233 lr_off, 1234 lr_off2, 1235 framesize // inclusive of return address 1236 }; 1237 // assert(is_even(framesize/2), "sp not 16-byte aligned"); 1238 stack_slots = framesize / VMRegImpl::slots_per_word; 1239 assert(stack_slots == 2, "recheck layout"); 1240 1241 address start = __ pc(); 1242 1243 compiled_entry_offset = __ pc() - start; 1244 __ enter(); 1245 1246 __ mov(c_rarg1, sp); 1247 1248 frame_complete = __ pc() - start; 1249 address the_pc = __ pc(); 1250 1251 __ post_call_nop(); // this must be exactly after the pc value that is pushed into the frame info, we use this nop for fast CodeBlob lookup 1252 1253 __ mov(c_rarg0, rthread); 1254 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1); 1255 __ call_VM_leaf(Continuation::freeze_entry(), 2); 1256 __ reset_last_Java_frame(true); 1257 1258 Label pinned; 1259 1260 __ cbnz(r0, pinned); 1261 1262 // We've succeeded, set sp to the ContinuationEntry 1263 __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset())); 1264 __ mov(sp, rscratch1); 1265 continuation_enter_cleanup(masm); 1266 1267 __ bind(pinned); // pinned -- return to caller 1268 1269 // handle pending exception thrown by freeze 1270 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 1271 Label ok; 1272 __ cbz(rscratch1, ok); 1273 __ leave(); 1274 __ lea(rscratch1, RuntimeAddress(StubRoutines::forward_exception_entry())); 1275 __ br(rscratch1); 1276 __ bind(ok); 1277 1278 __ leave(); 1279 __ ret(lr); 1280 1281 OopMap* map = new OopMap(framesize, 1); 1282 oop_maps->add_gc_map(the_pc - start, map); 1283 } 1284 1285 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) { 1286 ::continuation_enter_cleanup(masm); 1287 } 1288 1289 static void gen_special_dispatch(MacroAssembler* masm, 1290 const methodHandle& method, 1291 const BasicType* sig_bt, 1292 const VMRegPair* regs) { 1293 verify_oop_args(masm, method, sig_bt, regs); 1294 vmIntrinsics::ID iid = method->intrinsic_id(); 1295 1296 // Now write the args into the outgoing interpreter space 1297 bool has_receiver = false; 1298 Register receiver_reg = noreg; 1299 int member_arg_pos = -1; 1300 Register member_reg = noreg; 1301 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid); 1302 if (ref_kind != 0) { 1303 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument 1304 member_reg = r19; // known to be free at this point 1305 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind); 1306 } else if (iid == vmIntrinsics::_invokeBasic) { 1307 has_receiver = true; 1308 } else if (iid == vmIntrinsics::_linkToNative) { 1309 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument 1310 member_reg = r19; // known to be free at this point 1311 } else { 1312 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid)); 1313 } 1314 1315 if (member_reg != noreg) { 1316 // Load the member_arg into register, if necessary. 1317 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs); 1318 VMReg r = regs[member_arg_pos].first(); 1319 if (r->is_stack()) { 1320 __ ldr(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size)); 1321 } else { 1322 // no data motion is needed 1323 member_reg = r->as_Register(); 1324 } 1325 } 1326 1327 if (has_receiver) { 1328 // Make sure the receiver is loaded into a register. 1329 assert(method->size_of_parameters() > 0, "oob"); 1330 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object"); 1331 VMReg r = regs[0].first(); 1332 assert(r->is_valid(), "bad receiver arg"); 1333 if (r->is_stack()) { 1334 // Porting note: This assumes that compiled calling conventions always 1335 // pass the receiver oop in a register. If this is not true on some 1336 // platform, pick a temp and load the receiver from stack. 1337 fatal("receiver always in a register"); 1338 receiver_reg = r2; // known to be free at this point 1339 __ ldr(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size)); 1340 } else { 1341 // no data motion is needed 1342 receiver_reg = r->as_Register(); 1343 } 1344 } 1345 1346 // Figure out which address we are really jumping to: 1347 MethodHandles::generate_method_handle_dispatch(masm, iid, 1348 receiver_reg, member_reg, /*for_compiler_entry:*/ true); 1349 } 1350 1351 // --------------------------------------------------------------------------- 1352 // Generate a native wrapper for a given method. The method takes arguments 1353 // in the Java compiled code convention, marshals them to the native 1354 // convention (handlizes oops, etc), transitions to native, makes the call, 1355 // returns to java state (possibly blocking), unhandlizes any result and 1356 // returns. 1357 // 1358 // Critical native functions are a shorthand for the use of 1359 // GetPrimtiveArrayCritical and disallow the use of any other JNI 1360 // functions. The wrapper is expected to unpack the arguments before 1361 // passing them to the callee. Critical native functions leave the state _in_Java, 1362 // since they block out GC. 1363 // Some other parts of JNI setup are skipped like the tear down of the JNI handle 1364 // block and the check for pending exceptions it's impossible for them 1365 // to be thrown. 1366 // 1367 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm, 1368 const methodHandle& method, 1369 int compile_id, 1370 BasicType* in_sig_bt, 1371 VMRegPair* in_regs, 1372 BasicType ret_type) { 1373 if (method->is_continuation_native_intrinsic()) { 1374 int exception_offset = -1; 1375 OopMapSet* oop_maps = new OopMapSet(); 1376 int frame_complete = -1; 1377 int stack_slots = -1; 1378 int interpreted_entry_offset = -1; 1379 int vep_offset = -1; 1380 if (method->is_continuation_enter_intrinsic()) { 1381 gen_continuation_enter(masm, 1382 method, 1383 in_sig_bt, 1384 in_regs, 1385 exception_offset, 1386 oop_maps, 1387 frame_complete, 1388 stack_slots, 1389 interpreted_entry_offset, 1390 vep_offset); 1391 } else if (method->is_continuation_yield_intrinsic()) { 1392 gen_continuation_yield(masm, 1393 method, 1394 in_sig_bt, 1395 in_regs, 1396 oop_maps, 1397 frame_complete, 1398 stack_slots, 1399 vep_offset); 1400 } else { 1401 guarantee(false, "Unknown Continuation native intrinsic"); 1402 } 1403 1404 #ifdef ASSERT 1405 if (method->is_continuation_enter_intrinsic()) { 1406 assert(interpreted_entry_offset != -1, "Must be set"); 1407 assert(exception_offset != -1, "Must be set"); 1408 } else { 1409 assert(interpreted_entry_offset == -1, "Must be unset"); 1410 assert(exception_offset == -1, "Must be unset"); 1411 } 1412 assert(frame_complete != -1, "Must be set"); 1413 assert(stack_slots != -1, "Must be set"); 1414 assert(vep_offset != -1, "Must be set"); 1415 #endif 1416 1417 __ flush(); 1418 nmethod* nm = nmethod::new_native_nmethod(method, 1419 compile_id, 1420 masm->code(), 1421 vep_offset, 1422 frame_complete, 1423 stack_slots, 1424 in_ByteSize(-1), 1425 in_ByteSize(-1), 1426 oop_maps, 1427 exception_offset); 1428 if (nm == nullptr) return nm; 1429 if (method->is_continuation_enter_intrinsic()) { 1430 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset); 1431 } else if (method->is_continuation_yield_intrinsic()) { 1432 _cont_doYield_stub = nm; 1433 } else { 1434 guarantee(false, "Unknown Continuation native intrinsic"); 1435 } 1436 return nm; 1437 } 1438 1439 if (method->is_method_handle_intrinsic()) { 1440 vmIntrinsics::ID iid = method->intrinsic_id(); 1441 intptr_t start = (intptr_t)__ pc(); 1442 int vep_offset = ((intptr_t)__ pc()) - start; 1443 1444 // First instruction must be a nop as it may need to be patched on deoptimisation 1445 __ nop(); 1446 gen_special_dispatch(masm, 1447 method, 1448 in_sig_bt, 1449 in_regs); 1450 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period 1451 __ flush(); 1452 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually 1453 return nmethod::new_native_nmethod(method, 1454 compile_id, 1455 masm->code(), 1456 vep_offset, 1457 frame_complete, 1458 stack_slots / VMRegImpl::slots_per_word, 1459 in_ByteSize(-1), 1460 in_ByteSize(-1), 1461 nullptr); 1462 } 1463 address native_func = method->native_function(); 1464 assert(native_func != nullptr, "must have function"); 1465 1466 // An OopMap for lock (and class if static) 1467 OopMapSet *oop_maps = new OopMapSet(); 1468 intptr_t start = (intptr_t)__ pc(); 1469 1470 // We have received a description of where all the java arg are located 1471 // on entry to the wrapper. We need to convert these args to where 1472 // the jni function will expect them. To figure out where they go 1473 // we convert the java signature to a C signature by inserting 1474 // the hidden arguments as arg[0] and possibly arg[1] (static method) 1475 1476 const int total_in_args = method->size_of_parameters(); 1477 int total_c_args = total_in_args + (method->is_static() ? 2 : 1); 1478 1479 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1480 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1481 BasicType* in_elem_bt = nullptr; 1482 1483 int argc = 0; 1484 out_sig_bt[argc++] = T_ADDRESS; 1485 if (method->is_static()) { 1486 out_sig_bt[argc++] = T_OBJECT; 1487 } 1488 1489 for (int i = 0; i < total_in_args ; i++ ) { 1490 out_sig_bt[argc++] = in_sig_bt[i]; 1491 } 1492 1493 // Now figure out where the args must be stored and how much stack space 1494 // they require. 1495 int out_arg_slots; 1496 out_arg_slots = c_calling_convention_priv(out_sig_bt, out_regs, total_c_args); 1497 1498 if (out_arg_slots < 0) { 1499 return nullptr; 1500 } 1501 1502 // Compute framesize for the wrapper. We need to handlize all oops in 1503 // incoming registers 1504 1505 // Calculate the total number of stack slots we will need. 1506 1507 // First count the abi requirement plus all of the outgoing args 1508 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots; 1509 1510 // Now the space for the inbound oop handle area 1511 int total_save_slots = 8 * VMRegImpl::slots_per_word; // 8 arguments passed in registers 1512 1513 int oop_handle_offset = stack_slots; 1514 stack_slots += total_save_slots; 1515 1516 // Now any space we need for handlizing a klass if static method 1517 1518 int klass_slot_offset = 0; 1519 int klass_offset = -1; 1520 int lock_slot_offset = 0; 1521 bool is_static = false; 1522 1523 if (method->is_static()) { 1524 klass_slot_offset = stack_slots; 1525 stack_slots += VMRegImpl::slots_per_word; 1526 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; 1527 is_static = true; 1528 } 1529 1530 // Plus a lock if needed 1531 1532 if (method->is_synchronized()) { 1533 lock_slot_offset = stack_slots; 1534 stack_slots += VMRegImpl::slots_per_word; 1535 } 1536 1537 // Now a place (+2) to save return values or temp during shuffling 1538 // + 4 for return address (which we own) and saved rfp 1539 stack_slots += 6; 1540 1541 // Ok The space we have allocated will look like: 1542 // 1543 // 1544 // FP-> | | 1545 // |---------------------| 1546 // | 2 slots for moves | 1547 // |---------------------| 1548 // | lock box (if sync) | 1549 // |---------------------| <- lock_slot_offset 1550 // | klass (if static) | 1551 // |---------------------| <- klass_slot_offset 1552 // | oopHandle area | 1553 // |---------------------| <- oop_handle_offset (8 java arg registers) 1554 // | outbound memory | 1555 // | based arguments | 1556 // | | 1557 // |---------------------| 1558 // | | 1559 // SP-> | out_preserved_slots | 1560 // 1561 // 1562 1563 1564 // Now compute actual number of stack words we need rounding to make 1565 // stack properly aligned. 1566 stack_slots = align_up(stack_slots, StackAlignmentInSlots); 1567 1568 int stack_size = stack_slots * VMRegImpl::stack_slot_size; 1569 1570 // First thing make an ic check to see if we should even be here 1571 1572 // We are free to use all registers as temps without saving them and 1573 // restoring them except rfp. rfp is the only callee save register 1574 // as far as the interpreter and the compiler(s) are concerned. 1575 1576 const Register receiver = j_rarg0; 1577 1578 Label exception_pending; 1579 1580 assert_different_registers(receiver, rscratch1); 1581 __ verify_oop(receiver); 1582 __ ic_check(8 /* end_alignment */); 1583 1584 // Verified entry point must be aligned 1585 int vep_offset = ((intptr_t)__ pc()) - start; 1586 1587 // If we have to make this method not-entrant we'll overwrite its 1588 // first instruction with a jump. For this action to be legal we 1589 // must ensure that this first instruction is a B, BL, NOP, BKPT, 1590 // SVC, HVC, or SMC. Make it a NOP. 1591 __ nop(); 1592 1593 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) { 1594 Label L_skip_barrier; 1595 __ mov_metadata(rscratch2, method->method_holder()); // InstanceKlass* 1596 __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier); 1597 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub())); 1598 1599 __ bind(L_skip_barrier); 1600 } 1601 1602 // Generate stack overflow check 1603 __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size())); 1604 1605 // Generate a new frame for the wrapper. 1606 __ enter(); 1607 // -2 because return address is already present and so is saved rfp 1608 __ sub(sp, sp, stack_size - 2*wordSize); 1609 1610 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 1611 bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */); 1612 1613 // Frame is now completed as far as size and linkage. 1614 int frame_complete = ((intptr_t)__ pc()) - start; 1615 1616 // We use r20 as the oop handle for the receiver/klass 1617 // It is callee save so it survives the call to native 1618 1619 const Register oop_handle_reg = r20; 1620 1621 // 1622 // We immediately shuffle the arguments so that any vm call we have to 1623 // make from here on out (sync slow path, jvmti, etc.) we will have 1624 // captured the oops from our caller and have a valid oopMap for 1625 // them. 1626 1627 // ----------------- 1628 // The Grand Shuffle 1629 1630 // The Java calling convention is either equal (linux) or denser (win64) than the 1631 // c calling convention. However the because of the jni_env argument the c calling 1632 // convention always has at least one more (and two for static) arguments than Java. 1633 // Therefore if we move the args from java -> c backwards then we will never have 1634 // a register->register conflict and we don't have to build a dependency graph 1635 // and figure out how to break any cycles. 1636 // 1637 1638 // Record esp-based slot for receiver on stack for non-static methods 1639 int receiver_offset = -1; 1640 1641 // This is a trick. We double the stack slots so we can claim 1642 // the oops in the caller's frame. Since we are sure to have 1643 // more args than the caller doubling is enough to make 1644 // sure we can capture all the incoming oop args from the 1645 // caller. 1646 // 1647 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1648 1649 // Mark location of rfp (someday) 1650 // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, vmreg(rfp)); 1651 1652 1653 int float_args = 0; 1654 int int_args = 0; 1655 1656 #ifdef ASSERT 1657 bool reg_destroyed[Register::number_of_registers]; 1658 bool freg_destroyed[FloatRegister::number_of_registers]; 1659 for ( int r = 0 ; r < Register::number_of_registers ; r++ ) { 1660 reg_destroyed[r] = false; 1661 } 1662 for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) { 1663 freg_destroyed[f] = false; 1664 } 1665 1666 #endif /* ASSERT */ 1667 1668 // For JNI natives the incoming and outgoing registers are offset upwards. 1669 GrowableArray<int> arg_order(2 * total_in_args); 1670 VMRegPair tmp_vmreg; 1671 tmp_vmreg.set2(r19->as_VMReg()); 1672 1673 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) { 1674 arg_order.push(i); 1675 arg_order.push(c_arg); 1676 } 1677 1678 int temploc = -1; 1679 for (int ai = 0; ai < arg_order.length(); ai += 2) { 1680 int i = arg_order.at(ai); 1681 int c_arg = arg_order.at(ai + 1); 1682 __ block_comment(err_msg("move %d -> %d", i, c_arg)); 1683 assert(c_arg != -1 && i != -1, "wrong order"); 1684 #ifdef ASSERT 1685 if (in_regs[i].first()->is_Register()) { 1686 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!"); 1687 } else if (in_regs[i].first()->is_FloatRegister()) { 1688 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!"); 1689 } 1690 if (out_regs[c_arg].first()->is_Register()) { 1691 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true; 1692 } else if (out_regs[c_arg].first()->is_FloatRegister()) { 1693 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true; 1694 } 1695 #endif /* ASSERT */ 1696 switch (in_sig_bt[i]) { 1697 case T_ARRAY: 1698 case T_OBJECT: 1699 __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg], 1700 ((i == 0) && (!is_static)), 1701 &receiver_offset); 1702 int_args++; 1703 break; 1704 case T_VOID: 1705 break; 1706 1707 case T_FLOAT: 1708 __ float_move(in_regs[i], out_regs[c_arg]); 1709 float_args++; 1710 break; 1711 1712 case T_DOUBLE: 1713 assert( i + 1 < total_in_args && 1714 in_sig_bt[i + 1] == T_VOID && 1715 out_sig_bt[c_arg+1] == T_VOID, "bad arg list"); 1716 __ double_move(in_regs[i], out_regs[c_arg]); 1717 float_args++; 1718 break; 1719 1720 case T_LONG : 1721 __ long_move(in_regs[i], out_regs[c_arg]); 1722 int_args++; 1723 break; 1724 1725 case T_ADDRESS: assert(false, "found T_ADDRESS in java args"); 1726 1727 default: 1728 __ move32_64(in_regs[i], out_regs[c_arg]); 1729 int_args++; 1730 } 1731 } 1732 1733 // point c_arg at the first arg that is already loaded in case we 1734 // need to spill before we call out 1735 int c_arg = total_c_args - total_in_args; 1736 1737 // Pre-load a static method's oop into c_rarg1. 1738 if (method->is_static()) { 1739 1740 // load oop into a register 1741 __ movoop(c_rarg1, 1742 JNIHandles::make_local(method->method_holder()->java_mirror())); 1743 1744 // Now handlize the static class mirror it's known not-null. 1745 __ str(c_rarg1, Address(sp, klass_offset)); 1746 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); 1747 1748 // Now get the handle 1749 __ lea(c_rarg1, Address(sp, klass_offset)); 1750 // and protect the arg if we must spill 1751 c_arg--; 1752 } 1753 1754 // Change state to native (we save the return address in the thread, since it might not 1755 // be pushed on the stack when we do a stack traversal). 1756 // We use the same pc/oopMap repeatedly when we call out 1757 Label resume_pc; 1758 __ set_last_Java_frame(sp, noreg, resume_pc, rscratch1); 1759 1760 Label dtrace_method_entry, dtrace_method_entry_done; 1761 if (DTraceMethodProbes) { 1762 __ b(dtrace_method_entry); 1763 __ bind(dtrace_method_entry_done); 1764 } 1765 1766 // RedefineClasses() tracing support for obsolete method entry 1767 if (log_is_enabled(Trace, redefine, class, obsolete)) { 1768 // protect the args we've loaded 1769 save_args(masm, total_c_args, c_arg, out_regs); 1770 __ mov_metadata(c_rarg1, method()); 1771 __ call_VM_leaf( 1772 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 1773 rthread, c_rarg1); 1774 restore_args(masm, total_c_args, c_arg, out_regs); 1775 } 1776 1777 // Lock a synchronized method 1778 1779 // Register definitions used by locking and unlocking 1780 1781 const Register swap_reg = r0; 1782 const Register obj_reg = r19; // Will contain the oop 1783 const Register lock_reg = r13; // Address of compiler lock object (BasicLock) 1784 const Register old_hdr = r13; // value of old header at unlock time 1785 const Register lock_tmp = r14; // Temporary used by lightweight_lock/unlock 1786 const Register tmp = lr; 1787 1788 Label slow_path_lock; 1789 Label lock_done; 1790 1791 if (method->is_synchronized()) { 1792 Label count; 1793 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes(); 1794 1795 // Get the handle (the 2nd argument) 1796 __ mov(oop_handle_reg, c_rarg1); 1797 1798 // Get address of the box 1799 1800 __ lea(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 1801 1802 // Load the oop from the handle 1803 __ ldr(obj_reg, Address(oop_handle_reg, 0)); 1804 1805 if (LockingMode == LM_MONITOR) { 1806 __ b(slow_path_lock); 1807 } else if (LockingMode == LM_LEGACY) { 1808 // Load (object->mark() | 1) into swap_reg %r0 1809 __ ldr(rscratch1, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 1810 __ orr(swap_reg, rscratch1, 1); 1811 1812 // Save (object->mark() | 1) into BasicLock's displaced header 1813 __ str(swap_reg, Address(lock_reg, mark_word_offset)); 1814 1815 // src -> dest iff dest == r0 else r0 <- dest 1816 __ cmpxchg_obj_header(r0, lock_reg, obj_reg, rscratch1, count, /*fallthrough*/nullptr); 1817 1818 // Hmm should this move to the slow path code area??? 1819 1820 // Test if the oopMark is an obvious stack pointer, i.e., 1821 // 1) (mark & 3) == 0, and 1822 // 2) sp <= mark < mark + os::pagesize() 1823 // These 3 tests can be done by evaluating the following 1824 // expression: ((mark - sp) & (3 - os::vm_page_size())), 1825 // assuming both stack pointer and pagesize have their 1826 // least significant 2 bits clear. 1827 // NOTE: the oopMark is in swap_reg %r0 as the result of cmpxchg 1828 1829 __ sub(swap_reg, sp, swap_reg); 1830 __ neg(swap_reg, swap_reg); 1831 __ ands(swap_reg, swap_reg, 3 - (int)os::vm_page_size()); 1832 1833 // Save the test result, for recursive case, the result is zero 1834 __ str(swap_reg, Address(lock_reg, mark_word_offset)); 1835 __ br(Assembler::NE, slow_path_lock); 1836 __ b(lock_done); 1837 } else { 1838 assert(LockingMode == LM_LIGHTWEIGHT, "must be"); 1839 __ lightweight_lock(obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock); 1840 __ b(lock_done); 1841 } 1842 __ bind(count); 1843 __ inc_held_monitor_count(); 1844 1845 // Slow path will re-enter here 1846 __ bind(lock_done); 1847 } 1848 1849 1850 // Finally just about ready to make the JNI call 1851 1852 // get JNIEnv* which is first argument to native 1853 __ lea(c_rarg0, Address(rthread, in_bytes(JavaThread::jni_environment_offset()))); 1854 1855 // Now set thread in native 1856 __ mov(rscratch1, _thread_in_native); 1857 __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset())); 1858 __ stlrw(rscratch1, rscratch2); 1859 1860 __ rt_call(native_func); 1861 1862 // Verify or restore cpu control state after JNI call 1863 __ restore_cpu_control_state_after_jni(rscratch1, rscratch2); 1864 1865 // Unpack native results. 1866 switch (ret_type) { 1867 case T_BOOLEAN: __ c2bool(r0); break; 1868 case T_CHAR : __ ubfx(r0, r0, 0, 16); break; 1869 case T_BYTE : __ sbfx(r0, r0, 0, 8); break; 1870 case T_SHORT : __ sbfx(r0, r0, 0, 16); break; 1871 case T_INT : __ sbfx(r0, r0, 0, 32); break; 1872 case T_DOUBLE : 1873 case T_FLOAT : 1874 // Result is in v0 we'll save as needed 1875 break; 1876 case T_ARRAY: // Really a handle 1877 case T_OBJECT: // Really a handle 1878 break; // can't de-handlize until after safepoint check 1879 case T_VOID: break; 1880 case T_LONG: break; 1881 default : ShouldNotReachHere(); 1882 } 1883 1884 Label safepoint_in_progress, safepoint_in_progress_done; 1885 Label after_transition; 1886 1887 // Switch thread to "native transition" state before reading the synchronization state. 1888 // This additional state is necessary because reading and testing the synchronization 1889 // state is not atomic w.r.t. GC, as this scenario demonstrates: 1890 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 1891 // VM thread changes sync state to synchronizing and suspends threads for GC. 1892 // Thread A is resumed to finish this native method, but doesn't block here since it 1893 // didn't see any synchronization is progress, and escapes. 1894 __ mov(rscratch1, _thread_in_native_trans); 1895 1896 __ strw(rscratch1, Address(rthread, JavaThread::thread_state_offset())); 1897 1898 // Force this write out before the read below 1899 if (!UseSystemMemoryBarrier) { 1900 __ dmb(Assembler::ISH); 1901 } 1902 1903 __ verify_sve_vector_length(); 1904 1905 // Check for safepoint operation in progress and/or pending suspend requests. 1906 { 1907 // We need an acquire here to ensure that any subsequent load of the 1908 // global SafepointSynchronize::_state flag is ordered after this load 1909 // of the thread-local polling word. We don't want this poll to 1910 // return false (i.e. not safepointing) and a later poll of the global 1911 // SafepointSynchronize::_state spuriously to return true. 1912 // 1913 // This is to avoid a race when we're in a native->Java transition 1914 // racing the code which wakes up from a safepoint. 1915 1916 __ safepoint_poll(safepoint_in_progress, true /* at_return */, true /* acquire */, false /* in_nmethod */); 1917 __ ldrw(rscratch1, Address(rthread, JavaThread::suspend_flags_offset())); 1918 __ cbnzw(rscratch1, safepoint_in_progress); 1919 __ bind(safepoint_in_progress_done); 1920 } 1921 1922 // change thread state 1923 __ mov(rscratch1, _thread_in_Java); 1924 __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset())); 1925 __ stlrw(rscratch1, rscratch2); 1926 __ bind(after_transition); 1927 1928 // Check preemption for Object.wait() 1929 if (method->is_object_wait0()) { 1930 Label not_preempted; 1931 __ ldr(rscratch1, Address(rthread, JavaThread::preempt_alternate_return_offset())); 1932 __ cbz(rscratch1, not_preempted); 1933 __ str(zr, Address(rthread, JavaThread::preempt_alternate_return_offset())); 1934 __ br(rscratch1); 1935 __ bind(not_preempted); 1936 } 1937 __ bind(resume_pc); 1938 1939 intptr_t the_pc = (intptr_t) __ pc(); 1940 oop_maps->add_gc_map(the_pc - start, map); 1941 1942 Label reguard; 1943 Label reguard_done; 1944 __ ldrb(rscratch1, Address(rthread, JavaThread::stack_guard_state_offset())); 1945 __ cmpw(rscratch1, StackOverflow::stack_guard_yellow_reserved_disabled); 1946 __ br(Assembler::EQ, reguard); 1947 __ bind(reguard_done); 1948 1949 // native result if any is live 1950 1951 // Unlock 1952 Label unlock_done; 1953 Label slow_path_unlock; 1954 if (method->is_synchronized()) { 1955 1956 // Get locked oop from the handle we passed to jni 1957 __ ldr(obj_reg, Address(oop_handle_reg, 0)); 1958 1959 Label done, not_recursive; 1960 1961 if (LockingMode == LM_LEGACY) { 1962 // Simple recursive lock? 1963 __ ldr(rscratch1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 1964 __ cbnz(rscratch1, not_recursive); 1965 __ b(done); 1966 } 1967 1968 __ bind(not_recursive); 1969 1970 // Must save r0 if if it is live now because cmpxchg must use it 1971 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 1972 save_native_result(masm, ret_type, stack_slots); 1973 } 1974 1975 if (LockingMode == LM_MONITOR) { 1976 __ b(slow_path_unlock); 1977 } else if (LockingMode == LM_LEGACY) { 1978 // get address of the stack lock 1979 __ lea(r0, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 1980 // get old displaced header 1981 __ ldr(old_hdr, Address(r0, 0)); 1982 1983 // Atomic swap old header if oop still contains the stack lock 1984 Label count; 1985 __ cmpxchg_obj_header(r0, old_hdr, obj_reg, rscratch1, count, &slow_path_unlock); 1986 __ bind(count); 1987 __ dec_held_monitor_count(); 1988 } else { 1989 assert(LockingMode == LM_LIGHTWEIGHT, ""); 1990 __ lightweight_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock); 1991 } 1992 1993 // slow path re-enters here 1994 __ bind(unlock_done); 1995 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 1996 restore_native_result(masm, ret_type, stack_slots); 1997 } 1998 1999 __ bind(done); 2000 } 2001 2002 Label dtrace_method_exit, dtrace_method_exit_done; 2003 if (DTraceMethodProbes) { 2004 __ b(dtrace_method_exit); 2005 __ bind(dtrace_method_exit_done); 2006 } 2007 2008 __ reset_last_Java_frame(false); 2009 2010 // Unbox oop result, e.g. JNIHandles::resolve result. 2011 if (is_reference_type(ret_type)) { 2012 __ resolve_jobject(r0, r1, r2); 2013 } 2014 2015 if (CheckJNICalls) { 2016 // clear_pending_jni_exception_check 2017 __ str(zr, Address(rthread, JavaThread::pending_jni_exception_check_fn_offset())); 2018 } 2019 2020 // reset handle block 2021 __ ldr(r2, Address(rthread, JavaThread::active_handles_offset())); 2022 __ str(zr, Address(r2, JNIHandleBlock::top_offset())); 2023 2024 __ leave(); 2025 2026 // Any exception pending? 2027 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2028 __ cbnz(rscratch1, exception_pending); 2029 2030 // We're done 2031 __ ret(lr); 2032 2033 // Unexpected paths are out of line and go here 2034 2035 // forward the exception 2036 __ bind(exception_pending); 2037 2038 // and forward the exception 2039 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2040 2041 // Slow path locking & unlocking 2042 if (method->is_synchronized()) { 2043 2044 __ block_comment("Slow path lock {"); 2045 __ bind(slow_path_lock); 2046 2047 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM 2048 // args are (oop obj, BasicLock* lock, JavaThread* thread) 2049 2050 // protect the args we've loaded 2051 save_args(masm, total_c_args, c_arg, out_regs); 2052 2053 __ mov(c_rarg0, obj_reg); 2054 __ mov(c_rarg1, lock_reg); 2055 __ mov(c_rarg2, rthread); 2056 2057 // Not a leaf but we have last_Java_frame setup as we want 2058 __ push_cont_fastpath(); 2059 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3); 2060 __ pop_cont_fastpath(); 2061 restore_args(masm, total_c_args, c_arg, out_regs); 2062 2063 #ifdef ASSERT 2064 { Label L; 2065 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2066 __ cbz(rscratch1, L); 2067 __ stop("no pending exception allowed on exit from monitorenter"); 2068 __ bind(L); 2069 } 2070 #endif 2071 __ b(lock_done); 2072 2073 __ block_comment("} Slow path lock"); 2074 2075 __ block_comment("Slow path unlock {"); 2076 __ bind(slow_path_unlock); 2077 2078 // If we haven't already saved the native result we must save it now as xmm registers 2079 // are still exposed. 2080 2081 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) { 2082 save_native_result(masm, ret_type, stack_slots); 2083 } 2084 2085 __ mov(c_rarg2, rthread); 2086 __ lea(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 2087 __ mov(c_rarg0, obj_reg); 2088 2089 // Save pending exception around call to VM (which contains an EXCEPTION_MARK) 2090 // NOTE that obj_reg == r19 currently 2091 __ ldr(r19, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2092 __ str(zr, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2093 2094 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C)); 2095 2096 #ifdef ASSERT 2097 { 2098 Label L; 2099 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2100 __ cbz(rscratch1, L); 2101 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C"); 2102 __ bind(L); 2103 } 2104 #endif /* ASSERT */ 2105 2106 __ str(r19, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2107 2108 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) { 2109 restore_native_result(masm, ret_type, stack_slots); 2110 } 2111 __ b(unlock_done); 2112 2113 __ block_comment("} Slow path unlock"); 2114 2115 } // synchronized 2116 2117 // SLOW PATH Reguard the stack if needed 2118 2119 __ bind(reguard); 2120 save_native_result(masm, ret_type, stack_slots); 2121 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages)); 2122 restore_native_result(masm, ret_type, stack_slots); 2123 // and continue 2124 __ b(reguard_done); 2125 2126 // SLOW PATH safepoint 2127 { 2128 __ block_comment("safepoint {"); 2129 __ bind(safepoint_in_progress); 2130 2131 // Don't use call_VM as it will see a possible pending exception and forward it 2132 // and never return here preventing us from clearing _last_native_pc down below. 2133 // 2134 save_native_result(masm, ret_type, stack_slots); 2135 __ mov(c_rarg0, rthread); 2136 #ifndef PRODUCT 2137 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 2138 #endif 2139 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans))); 2140 __ blr(rscratch1); 2141 2142 // Restore any method result value 2143 restore_native_result(masm, ret_type, stack_slots); 2144 2145 __ b(safepoint_in_progress_done); 2146 __ block_comment("} safepoint"); 2147 } 2148 2149 // SLOW PATH dtrace support 2150 if (DTraceMethodProbes) { 2151 { 2152 __ block_comment("dtrace entry {"); 2153 __ bind(dtrace_method_entry); 2154 2155 // We have all of the arguments setup at this point. We must not touch any register 2156 // argument registers at this point (what if we save/restore them there are no oop? 2157 2158 save_args(masm, total_c_args, c_arg, out_regs); 2159 __ mov_metadata(c_rarg1, method()); 2160 __ call_VM_leaf( 2161 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 2162 rthread, c_rarg1); 2163 restore_args(masm, total_c_args, c_arg, out_regs); 2164 __ b(dtrace_method_entry_done); 2165 __ block_comment("} dtrace entry"); 2166 } 2167 2168 { 2169 __ block_comment("dtrace exit {"); 2170 __ bind(dtrace_method_exit); 2171 save_native_result(masm, ret_type, stack_slots); 2172 __ mov_metadata(c_rarg1, method()); 2173 __ call_VM_leaf( 2174 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 2175 rthread, c_rarg1); 2176 restore_native_result(masm, ret_type, stack_slots); 2177 __ b(dtrace_method_exit_done); 2178 __ block_comment("} dtrace exit"); 2179 } 2180 } 2181 2182 __ flush(); 2183 2184 nmethod *nm = nmethod::new_native_nmethod(method, 2185 compile_id, 2186 masm->code(), 2187 vep_offset, 2188 frame_complete, 2189 stack_slots / VMRegImpl::slots_per_word, 2190 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), 2191 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size), 2192 oop_maps); 2193 2194 return nm; 2195 } 2196 2197 // this function returns the adjust size (in number of words) to a c2i adapter 2198 // activation for use during deoptimization 2199 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) { 2200 assert(callee_locals >= callee_parameters, 2201 "test and remove; got more parms than locals"); 2202 if (callee_locals < callee_parameters) 2203 return 0; // No adjustment for negative locals 2204 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords; 2205 // diff is counted in stack words 2206 return align_up(diff, 2); 2207 } 2208 2209 2210 //------------------------------generate_deopt_blob---------------------------- 2211 void SharedRuntime::generate_deopt_blob() { 2212 // Allocate space for the code 2213 ResourceMark rm; 2214 // Setup code generation tools 2215 int pad = 0; 2216 #if INCLUDE_JVMCI 2217 if (EnableJVMCI) { 2218 pad += 512; // Increase the buffer size when compiling for JVMCI 2219 } 2220 #endif 2221 CodeBuffer buffer("deopt_blob", 2048+pad, 1024); 2222 MacroAssembler* masm = new MacroAssembler(&buffer); 2223 int frame_size_in_words; 2224 OopMap* map = nullptr; 2225 OopMapSet *oop_maps = new OopMapSet(); 2226 RegisterSaver reg_save(COMPILER2_OR_JVMCI != 0); 2227 2228 // ------------- 2229 // This code enters when returning to a de-optimized nmethod. A return 2230 // address has been pushed on the stack, and return values are in 2231 // registers. 2232 // If we are doing a normal deopt then we were called from the patched 2233 // nmethod from the point we returned to the nmethod. So the return 2234 // address on the stack is wrong by NativeCall::instruction_size 2235 // We will adjust the value so it looks like we have the original return 2236 // address on the stack (like when we eagerly deoptimized). 2237 // In the case of an exception pending when deoptimizing, we enter 2238 // with a return address on the stack that points after the call we patched 2239 // into the exception handler. We have the following register state from, 2240 // e.g., the forward exception stub (see stubGenerator_x86_64.cpp). 2241 // r0: exception oop 2242 // r19: exception handler 2243 // r3: throwing pc 2244 // So in this case we simply jam r3 into the useless return address and 2245 // the stack looks just like we want. 2246 // 2247 // At this point we need to de-opt. We save the argument return 2248 // registers. We call the first C routine, fetch_unroll_info(). This 2249 // routine captures the return values and returns a structure which 2250 // describes the current frame size and the sizes of all replacement frames. 2251 // The current frame is compiled code and may contain many inlined 2252 // functions, each with their own JVM state. We pop the current frame, then 2253 // push all the new frames. Then we call the C routine unpack_frames() to 2254 // populate these frames. Finally unpack_frames() returns us the new target 2255 // address. Notice that callee-save registers are BLOWN here; they have 2256 // already been captured in the vframeArray at the time the return PC was 2257 // patched. 2258 address start = __ pc(); 2259 Label cont; 2260 2261 // Prolog for non exception case! 2262 2263 // Save everything in sight. 2264 map = reg_save.save_live_registers(masm, 0, &frame_size_in_words); 2265 2266 // Normal deoptimization. Save exec mode for unpack_frames. 2267 __ movw(rcpool, Deoptimization::Unpack_deopt); // callee-saved 2268 __ b(cont); 2269 2270 int reexecute_offset = __ pc() - start; 2271 #if INCLUDE_JVMCI && !defined(COMPILER1) 2272 if (EnableJVMCI && UseJVMCICompiler) { 2273 // JVMCI does not use this kind of deoptimization 2274 __ should_not_reach_here(); 2275 } 2276 #endif 2277 2278 // Reexecute case 2279 // return address is the pc describes what bci to do re-execute at 2280 2281 // No need to update map as each call to save_live_registers will produce identical oopmap 2282 (void) reg_save.save_live_registers(masm, 0, &frame_size_in_words); 2283 2284 __ movw(rcpool, Deoptimization::Unpack_reexecute); // callee-saved 2285 __ b(cont); 2286 2287 #if INCLUDE_JVMCI 2288 Label after_fetch_unroll_info_call; 2289 int implicit_exception_uncommon_trap_offset = 0; 2290 int uncommon_trap_offset = 0; 2291 2292 if (EnableJVMCI) { 2293 implicit_exception_uncommon_trap_offset = __ pc() - start; 2294 2295 __ ldr(lr, Address(rthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset()))); 2296 __ str(zr, Address(rthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset()))); 2297 2298 uncommon_trap_offset = __ pc() - start; 2299 2300 // Save everything in sight. 2301 reg_save.save_live_registers(masm, 0, &frame_size_in_words); 2302 // fetch_unroll_info needs to call last_java_frame() 2303 Label retaddr; 2304 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2305 2306 __ ldrw(c_rarg1, Address(rthread, in_bytes(JavaThread::pending_deoptimization_offset()))); 2307 __ movw(rscratch1, -1); 2308 __ strw(rscratch1, Address(rthread, in_bytes(JavaThread::pending_deoptimization_offset()))); 2309 2310 __ movw(rcpool, (int32_t)Deoptimization::Unpack_reexecute); 2311 __ mov(c_rarg0, rthread); 2312 __ movw(c_rarg2, rcpool); // exec mode 2313 __ lea(rscratch1, 2314 RuntimeAddress(CAST_FROM_FN_PTR(address, 2315 Deoptimization::uncommon_trap))); 2316 __ blr(rscratch1); 2317 __ bind(retaddr); 2318 oop_maps->add_gc_map( __ pc()-start, map->deep_copy()); 2319 2320 __ reset_last_Java_frame(false); 2321 2322 __ b(after_fetch_unroll_info_call); 2323 } // EnableJVMCI 2324 #endif // INCLUDE_JVMCI 2325 2326 int exception_offset = __ pc() - start; 2327 2328 // Prolog for exception case 2329 2330 // all registers are dead at this entry point, except for r0, and 2331 // r3 which contain the exception oop and exception pc 2332 // respectively. Set them in TLS and fall thru to the 2333 // unpack_with_exception_in_tls entry point. 2334 2335 __ str(r3, Address(rthread, JavaThread::exception_pc_offset())); 2336 __ str(r0, Address(rthread, JavaThread::exception_oop_offset())); 2337 2338 int exception_in_tls_offset = __ pc() - start; 2339 2340 // new implementation because exception oop is now passed in JavaThread 2341 2342 // Prolog for exception case 2343 // All registers must be preserved because they might be used by LinearScan 2344 // Exceptiop oop and throwing PC are passed in JavaThread 2345 // tos: stack at point of call to method that threw the exception (i.e. only 2346 // args are on the stack, no return address) 2347 2348 // The return address pushed by save_live_registers will be patched 2349 // later with the throwing pc. The correct value is not available 2350 // now because loading it from memory would destroy registers. 2351 2352 // NB: The SP at this point must be the SP of the method that is 2353 // being deoptimized. Deoptimization assumes that the frame created 2354 // here by save_live_registers is immediately below the method's SP. 2355 // This is a somewhat fragile mechanism. 2356 2357 // Save everything in sight. 2358 map = reg_save.save_live_registers(masm, 0, &frame_size_in_words); 2359 2360 // Now it is safe to overwrite any register 2361 2362 // Deopt during an exception. Save exec mode for unpack_frames. 2363 __ mov(rcpool, Deoptimization::Unpack_exception); // callee-saved 2364 2365 // load throwing pc from JavaThread and patch it as the return address 2366 // of the current frame. Then clear the field in JavaThread 2367 __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset())); 2368 __ protect_return_address(r3); 2369 __ str(r3, Address(rfp, wordSize)); 2370 __ str(zr, Address(rthread, JavaThread::exception_pc_offset())); 2371 2372 #ifdef ASSERT 2373 // verify that there is really an exception oop in JavaThread 2374 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset())); 2375 __ verify_oop(r0); 2376 2377 // verify that there is no pending exception 2378 Label no_pending_exception; 2379 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset())); 2380 __ cbz(rscratch1, no_pending_exception); 2381 __ stop("must not have pending exception here"); 2382 __ bind(no_pending_exception); 2383 #endif 2384 2385 __ bind(cont); 2386 2387 // Call C code. Need thread and this frame, but NOT official VM entry 2388 // crud. We cannot block on this call, no GC can happen. 2389 // 2390 // UnrollBlock* fetch_unroll_info(JavaThread* thread) 2391 2392 // fetch_unroll_info needs to call last_java_frame(). 2393 2394 Label retaddr; 2395 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2396 #ifdef ASSERT 2397 { Label L; 2398 __ ldr(rscratch1, Address(rthread, JavaThread::last_Java_fp_offset())); 2399 __ cbz(rscratch1, L); 2400 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared"); 2401 __ bind(L); 2402 } 2403 #endif // ASSERT 2404 __ mov(c_rarg0, rthread); 2405 __ mov(c_rarg1, rcpool); 2406 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info))); 2407 __ blr(rscratch1); 2408 __ bind(retaddr); 2409 2410 // Need to have an oopmap that tells fetch_unroll_info where to 2411 // find any register it might need. 2412 oop_maps->add_gc_map(__ pc() - start, map); 2413 2414 __ reset_last_Java_frame(false); 2415 2416 #if INCLUDE_JVMCI 2417 if (EnableJVMCI) { 2418 __ bind(after_fetch_unroll_info_call); 2419 } 2420 #endif 2421 2422 // Load UnrollBlock* into r5 2423 __ mov(r5, r0); 2424 2425 __ ldrw(rcpool, Address(r5, Deoptimization::UnrollBlock::unpack_kind_offset())); 2426 Label noException; 2427 __ cmpw(rcpool, Deoptimization::Unpack_exception); // Was exception pending? 2428 __ br(Assembler::NE, noException); 2429 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset())); 2430 // QQQ this is useless it was null above 2431 __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset())); 2432 __ str(zr, Address(rthread, JavaThread::exception_oop_offset())); 2433 __ str(zr, Address(rthread, JavaThread::exception_pc_offset())); 2434 2435 __ verify_oop(r0); 2436 2437 // Overwrite the result registers with the exception results. 2438 __ str(r0, Address(sp, reg_save.r0_offset_in_bytes())); 2439 // I think this is useless 2440 // __ str(r3, Address(sp, RegisterSaver::r3_offset_in_bytes())); 2441 2442 __ bind(noException); 2443 2444 // Only register save data is on the stack. 2445 // Now restore the result registers. Everything else is either dead 2446 // or captured in the vframeArray. 2447 2448 // Restore fp result register 2449 __ ldrd(v0, Address(sp, reg_save.v0_offset_in_bytes())); 2450 // Restore integer result register 2451 __ ldr(r0, Address(sp, reg_save.r0_offset_in_bytes())); 2452 2453 // Pop all of the register save area off the stack 2454 __ add(sp, sp, frame_size_in_words * wordSize); 2455 2456 // All of the register save area has been popped of the stack. Only the 2457 // return address remains. 2458 2459 // Pop all the frames we must move/replace. 2460 // 2461 // Frame picture (youngest to oldest) 2462 // 1: self-frame (no frame link) 2463 // 2: deopting frame (no frame link) 2464 // 3: caller of deopting frame (could be compiled/interpreted). 2465 // 2466 // Note: by leaving the return address of self-frame on the stack 2467 // and using the size of frame 2 to adjust the stack 2468 // when we are done the return to frame 3 will still be on the stack. 2469 2470 // Pop deoptimized frame 2471 __ ldrw(r2, Address(r5, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset())); 2472 __ sub(r2, r2, 2 * wordSize); 2473 __ add(sp, sp, r2); 2474 __ ldp(rfp, zr, __ post(sp, 2 * wordSize)); 2475 2476 #ifdef ASSERT 2477 // Compilers generate code that bang the stack by as much as the 2478 // interpreter would need. So this stack banging should never 2479 // trigger a fault. Verify that it does not on non product builds. 2480 __ ldrw(r19, Address(r5, Deoptimization::UnrollBlock::total_frame_sizes_offset())); 2481 __ bang_stack_size(r19, r2); 2482 #endif 2483 // Load address of array of frame pcs into r2 2484 __ ldr(r2, Address(r5, Deoptimization::UnrollBlock::frame_pcs_offset())); 2485 2486 // Trash the old pc 2487 // __ addptr(sp, wordSize); FIXME ???? 2488 2489 // Load address of array of frame sizes into r4 2490 __ ldr(r4, Address(r5, Deoptimization::UnrollBlock::frame_sizes_offset())); 2491 2492 // Load counter into r3 2493 __ ldrw(r3, Address(r5, Deoptimization::UnrollBlock::number_of_frames_offset())); 2494 2495 // Now adjust the caller's stack to make up for the extra locals 2496 // but record the original sp so that we can save it in the skeletal interpreter 2497 // frame and the stack walking of interpreter_sender will get the unextended sp 2498 // value and not the "real" sp value. 2499 2500 const Register sender_sp = r6; 2501 2502 __ mov(sender_sp, sp); 2503 __ ldrw(r19, Address(r5, 2504 Deoptimization::UnrollBlock:: 2505 caller_adjustment_offset())); 2506 __ sub(sp, sp, r19); 2507 2508 // Push interpreter frames in a loop 2509 __ mov(rscratch1, (uint64_t)0xDEADDEAD); // Make a recognizable pattern 2510 __ mov(rscratch2, rscratch1); 2511 Label loop; 2512 __ bind(loop); 2513 __ ldr(r19, Address(__ post(r4, wordSize))); // Load frame size 2514 __ sub(r19, r19, 2*wordSize); // We'll push pc and fp by hand 2515 __ ldr(lr, Address(__ post(r2, wordSize))); // Load pc 2516 __ enter(); // Save old & set new fp 2517 __ sub(sp, sp, r19); // Prolog 2518 // This value is corrected by layout_activation_impl 2519 __ str(zr, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 2520 __ str(sender_sp, Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable 2521 __ mov(sender_sp, sp); // Pass sender_sp to next frame 2522 __ sub(r3, r3, 1); // Decrement counter 2523 __ cbnz(r3, loop); 2524 2525 // Re-push self-frame 2526 __ ldr(lr, Address(r2)); 2527 __ enter(); 2528 2529 // Allocate a full sized register save area. We subtract 2 because 2530 // enter() just pushed 2 words 2531 __ sub(sp, sp, (frame_size_in_words - 2) * wordSize); 2532 2533 // Restore frame locals after moving the frame 2534 __ strd(v0, Address(sp, reg_save.v0_offset_in_bytes())); 2535 __ str(r0, Address(sp, reg_save.r0_offset_in_bytes())); 2536 2537 // Call C code. Need thread but NOT official VM entry 2538 // crud. We cannot block on this call, no GC can happen. Call should 2539 // restore return values to their stack-slots with the new SP. 2540 // 2541 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode) 2542 2543 // Use rfp because the frames look interpreted now 2544 // Don't need the precise return PC here, just precise enough to point into this code blob. 2545 address the_pc = __ pc(); 2546 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1); 2547 2548 __ mov(c_rarg0, rthread); 2549 __ movw(c_rarg1, rcpool); // second arg: exec_mode 2550 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 2551 __ blr(rscratch1); 2552 2553 // Set an oopmap for the call site 2554 // Use the same PC we used for the last java frame 2555 oop_maps->add_gc_map(the_pc - start, 2556 new OopMap( frame_size_in_words, 0 )); 2557 2558 // Clear fp AND pc 2559 __ reset_last_Java_frame(true); 2560 2561 // Collect return values 2562 __ ldrd(v0, Address(sp, reg_save.v0_offset_in_bytes())); 2563 __ ldr(r0, Address(sp, reg_save.r0_offset_in_bytes())); 2564 // I think this is useless (throwing pc?) 2565 // __ ldr(r3, Address(sp, RegisterSaver::r3_offset_in_bytes())); 2566 2567 // Pop self-frame. 2568 __ leave(); // Epilog 2569 2570 // Jump to interpreter 2571 __ ret(lr); 2572 2573 // Make sure all code is generated 2574 masm->flush(); 2575 2576 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words); 2577 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset); 2578 #if INCLUDE_JVMCI 2579 if (EnableJVMCI) { 2580 _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset); 2581 _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset); 2582 } 2583 #endif 2584 } 2585 2586 // Number of stack slots between incoming argument block and the start of 2587 // a new frame. The PROLOG must add this many slots to the stack. The 2588 // EPILOG must remove this many slots. aarch64 needs two slots for 2589 // return address and fp. 2590 // TODO think this is correct but check 2591 uint SharedRuntime::in_preserve_stack_slots() { 2592 return 4; 2593 } 2594 2595 uint SharedRuntime::out_preserve_stack_slots() { 2596 return 0; 2597 } 2598 2599 VMReg SharedRuntime::thread_register() { 2600 return rthread->as_VMReg(); 2601 } 2602 2603 #ifdef COMPILER2 2604 //------------------------------generate_uncommon_trap_blob-------------------- 2605 void SharedRuntime::generate_uncommon_trap_blob() { 2606 // Allocate space for the code 2607 ResourceMark rm; 2608 // Setup code generation tools 2609 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024); 2610 MacroAssembler* masm = new MacroAssembler(&buffer); 2611 2612 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned"); 2613 2614 address start = __ pc(); 2615 2616 // Push self-frame. We get here with a return address in LR 2617 // and sp should be 16 byte aligned 2618 // push rfp and retaddr by hand 2619 __ protect_return_address(); 2620 __ stp(rfp, lr, Address(__ pre(sp, -2 * wordSize))); 2621 // we don't expect an arg reg save area 2622 #ifndef PRODUCT 2623 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 2624 #endif 2625 // compiler left unloaded_class_index in j_rarg0 move to where the 2626 // runtime expects it. 2627 if (c_rarg1 != j_rarg0) { 2628 __ movw(c_rarg1, j_rarg0); 2629 } 2630 2631 // we need to set the past SP to the stack pointer of the stub frame 2632 // and the pc to the address where this runtime call will return 2633 // although actually any pc in this code blob will do). 2634 Label retaddr; 2635 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2636 2637 // Call C code. Need thread but NOT official VM entry 2638 // crud. We cannot block on this call, no GC can happen. Call should 2639 // capture callee-saved registers as well as return values. 2640 // Thread is in rdi already. 2641 // 2642 // UnrollBlock* uncommon_trap(JavaThread* thread, jint unloaded_class_index); 2643 // 2644 // n.b. 2 gp args, 0 fp args, integral return type 2645 2646 __ mov(c_rarg0, rthread); 2647 __ movw(c_rarg2, (unsigned)Deoptimization::Unpack_uncommon_trap); 2648 __ lea(rscratch1, 2649 RuntimeAddress(CAST_FROM_FN_PTR(address, 2650 Deoptimization::uncommon_trap))); 2651 __ blr(rscratch1); 2652 __ bind(retaddr); 2653 2654 // Set an oopmap for the call site 2655 OopMapSet* oop_maps = new OopMapSet(); 2656 OopMap* map = new OopMap(SimpleRuntimeFrame::framesize, 0); 2657 2658 // location of rfp is known implicitly by the frame sender code 2659 2660 oop_maps->add_gc_map(__ pc() - start, map); 2661 2662 __ reset_last_Java_frame(false); 2663 2664 // move UnrollBlock* into r4 2665 __ mov(r4, r0); 2666 2667 #ifdef ASSERT 2668 { Label L; 2669 __ ldrw(rscratch1, Address(r4, Deoptimization::UnrollBlock::unpack_kind_offset())); 2670 __ cmpw(rscratch1, (unsigned)Deoptimization::Unpack_uncommon_trap); 2671 __ br(Assembler::EQ, L); 2672 __ stop("SharedRuntime::generate_uncommon_trap_blob: expected Unpack_uncommon_trap"); 2673 __ bind(L); 2674 } 2675 #endif 2676 2677 // Pop all the frames we must move/replace. 2678 // 2679 // Frame picture (youngest to oldest) 2680 // 1: self-frame (no frame link) 2681 // 2: deopting frame (no frame link) 2682 // 3: caller of deopting frame (could be compiled/interpreted). 2683 2684 // Pop self-frame. We have no frame, and must rely only on r0 and sp. 2685 __ add(sp, sp, (SimpleRuntimeFrame::framesize) << LogBytesPerInt); // Epilog! 2686 2687 // Pop deoptimized frame (int) 2688 __ ldrw(r2, Address(r4, 2689 Deoptimization::UnrollBlock:: 2690 size_of_deoptimized_frame_offset())); 2691 __ sub(r2, r2, 2 * wordSize); 2692 __ add(sp, sp, r2); 2693 __ ldp(rfp, zr, __ post(sp, 2 * wordSize)); 2694 2695 #ifdef ASSERT 2696 // Compilers generate code that bang the stack by as much as the 2697 // interpreter would need. So this stack banging should never 2698 // trigger a fault. Verify that it does not on non product builds. 2699 __ ldrw(r1, Address(r4, 2700 Deoptimization::UnrollBlock:: 2701 total_frame_sizes_offset())); 2702 __ bang_stack_size(r1, r2); 2703 #endif 2704 2705 // Load address of array of frame pcs into r2 (address*) 2706 __ ldr(r2, Address(r4, 2707 Deoptimization::UnrollBlock::frame_pcs_offset())); 2708 2709 // Load address of array of frame sizes into r5 (intptr_t*) 2710 __ ldr(r5, Address(r4, 2711 Deoptimization::UnrollBlock:: 2712 frame_sizes_offset())); 2713 2714 // Counter 2715 __ ldrw(r3, Address(r4, 2716 Deoptimization::UnrollBlock:: 2717 number_of_frames_offset())); // (int) 2718 2719 // Now adjust the caller's stack to make up for the extra locals but 2720 // record the original sp so that we can save it in the skeletal 2721 // interpreter frame and the stack walking of interpreter_sender 2722 // will get the unextended sp value and not the "real" sp value. 2723 2724 const Register sender_sp = r8; 2725 2726 __ mov(sender_sp, sp); 2727 __ ldrw(r1, Address(r4, 2728 Deoptimization::UnrollBlock:: 2729 caller_adjustment_offset())); // (int) 2730 __ sub(sp, sp, r1); 2731 2732 // Push interpreter frames in a loop 2733 Label loop; 2734 __ bind(loop); 2735 __ ldr(r1, Address(r5, 0)); // Load frame size 2736 __ sub(r1, r1, 2 * wordSize); // We'll push pc and rfp by hand 2737 __ ldr(lr, Address(r2, 0)); // Save return address 2738 __ enter(); // and old rfp & set new rfp 2739 __ sub(sp, sp, r1); // Prolog 2740 __ str(sender_sp, Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable 2741 // This value is corrected by layout_activation_impl 2742 __ str(zr, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 2743 __ mov(sender_sp, sp); // Pass sender_sp to next frame 2744 __ add(r5, r5, wordSize); // Bump array pointer (sizes) 2745 __ add(r2, r2, wordSize); // Bump array pointer (pcs) 2746 __ subsw(r3, r3, 1); // Decrement counter 2747 __ br(Assembler::GT, loop); 2748 __ ldr(lr, Address(r2, 0)); // save final return address 2749 // Re-push self-frame 2750 __ enter(); // & old rfp & set new rfp 2751 2752 // Use rfp because the frames look interpreted now 2753 // Save "the_pc" since it cannot easily be retrieved using the last_java_SP after we aligned SP. 2754 // Don't need the precise return PC here, just precise enough to point into this code blob. 2755 address the_pc = __ pc(); 2756 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1); 2757 2758 // Call C code. Need thread but NOT official VM entry 2759 // crud. We cannot block on this call, no GC can happen. Call should 2760 // restore return values to their stack-slots with the new SP. 2761 // Thread is in rdi already. 2762 // 2763 // BasicType unpack_frames(JavaThread* thread, int exec_mode); 2764 // 2765 // n.b. 2 gp args, 0 fp args, integral return type 2766 2767 // sp should already be aligned 2768 __ mov(c_rarg0, rthread); 2769 __ movw(c_rarg1, (unsigned)Deoptimization::Unpack_uncommon_trap); 2770 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 2771 __ blr(rscratch1); 2772 2773 // Set an oopmap for the call site 2774 // Use the same PC we used for the last java frame 2775 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0)); 2776 2777 // Clear fp AND pc 2778 __ reset_last_Java_frame(true); 2779 2780 // Pop self-frame. 2781 __ leave(); // Epilog 2782 2783 // Jump to interpreter 2784 __ ret(lr); 2785 2786 // Make sure all code is generated 2787 masm->flush(); 2788 2789 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, 2790 SimpleRuntimeFrame::framesize >> 1); 2791 } 2792 #endif // COMPILER2 2793 2794 2795 //------------------------------generate_handler_blob------ 2796 // 2797 // Generate a special Compile2Runtime blob that saves all registers, 2798 // and setup oopmap. 2799 // 2800 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) { 2801 ResourceMark rm; 2802 OopMapSet *oop_maps = new OopMapSet(); 2803 OopMap* map; 2804 2805 // Allocate space for the code. Setup code generation tools. 2806 CodeBuffer buffer("handler_blob", 2048, 1024); 2807 MacroAssembler* masm = new MacroAssembler(&buffer); 2808 2809 address start = __ pc(); 2810 address call_pc = nullptr; 2811 int frame_size_in_words; 2812 bool cause_return = (poll_type == POLL_AT_RETURN); 2813 RegisterSaver reg_save(poll_type == POLL_AT_VECTOR_LOOP /* save_vectors */); 2814 2815 // When the signal occurred, the LR was either signed and stored on the stack (in which 2816 // case it will be restored from the stack before being used) or unsigned and not stored 2817 // on the stack. Stipping ensures we get the right value. 2818 __ strip_return_address(); 2819 2820 // Save Integer and Float registers. 2821 map = reg_save.save_live_registers(masm, 0, &frame_size_in_words); 2822 2823 // The following is basically a call_VM. However, we need the precise 2824 // address of the call in order to generate an oopmap. Hence, we do all the 2825 // work ourselves. 2826 2827 Label retaddr; 2828 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2829 2830 // The return address must always be correct so that frame constructor never 2831 // sees an invalid pc. 2832 2833 if (!cause_return) { 2834 // overwrite the return address pushed by save_live_registers 2835 // Additionally, r20 is a callee-saved register so we can look at 2836 // it later to determine if someone changed the return address for 2837 // us! 2838 __ ldr(r20, Address(rthread, JavaThread::saved_exception_pc_offset())); 2839 __ protect_return_address(r20); 2840 __ str(r20, Address(rfp, wordSize)); 2841 } 2842 2843 // Do the call 2844 __ mov(c_rarg0, rthread); 2845 __ lea(rscratch1, RuntimeAddress(call_ptr)); 2846 __ blr(rscratch1); 2847 __ bind(retaddr); 2848 2849 // Set an oopmap for the call site. This oopmap will map all 2850 // oop-registers and debug-info registers as callee-saved. This 2851 // will allow deoptimization at this safepoint to find all possible 2852 // debug-info recordings, as well as let GC find all oops. 2853 2854 oop_maps->add_gc_map( __ pc() - start, map); 2855 2856 Label noException; 2857 2858 __ reset_last_Java_frame(false); 2859 2860 __ membar(Assembler::LoadLoad | Assembler::LoadStore); 2861 2862 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset())); 2863 __ cbz(rscratch1, noException); 2864 2865 // Exception pending 2866 2867 reg_save.restore_live_registers(masm); 2868 2869 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2870 2871 // No exception case 2872 __ bind(noException); 2873 2874 Label no_adjust, bail; 2875 if (!cause_return) { 2876 // If our stashed return pc was modified by the runtime we avoid touching it 2877 __ ldr(rscratch1, Address(rfp, wordSize)); 2878 __ cmp(r20, rscratch1); 2879 __ br(Assembler::NE, no_adjust); 2880 __ authenticate_return_address(r20); 2881 2882 #ifdef ASSERT 2883 // Verify the correct encoding of the poll we're about to skip. 2884 // See NativeInstruction::is_ldrw_to_zr() 2885 __ ldrw(rscratch1, Address(r20)); 2886 __ ubfx(rscratch2, rscratch1, 22, 10); 2887 __ cmpw(rscratch2, 0b1011100101); 2888 __ br(Assembler::NE, bail); 2889 __ ubfx(rscratch2, rscratch1, 0, 5); 2890 __ cmpw(rscratch2, 0b11111); 2891 __ br(Assembler::NE, bail); 2892 #endif 2893 // Adjust return pc forward to step over the safepoint poll instruction 2894 __ add(r20, r20, NativeInstruction::instruction_size); 2895 __ protect_return_address(r20); 2896 __ str(r20, Address(rfp, wordSize)); 2897 } 2898 2899 __ bind(no_adjust); 2900 // Normal exit, restore registers and exit. 2901 reg_save.restore_live_registers(masm); 2902 2903 __ ret(lr); 2904 2905 #ifdef ASSERT 2906 __ bind(bail); 2907 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected"); 2908 #endif 2909 2910 // Make sure all code is generated 2911 masm->flush(); 2912 2913 // Fill-out other meta info 2914 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words); 2915 } 2916 2917 // 2918 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss 2919 // 2920 // Generate a stub that calls into vm to find out the proper destination 2921 // of a java call. All the argument registers are live at this point 2922 // but since this is generic code we don't know what they are and the caller 2923 // must do any gc of the args. 2924 // 2925 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) { 2926 assert (StubRoutines::forward_exception_entry() != nullptr, "must be generated before"); 2927 2928 // allocate space for the code 2929 ResourceMark rm; 2930 2931 CodeBuffer buffer(name, 1000, 512); 2932 MacroAssembler* masm = new MacroAssembler(&buffer); 2933 2934 int frame_size_in_words; 2935 RegisterSaver reg_save(false /* save_vectors */); 2936 2937 OopMapSet *oop_maps = new OopMapSet(); 2938 OopMap* map = nullptr; 2939 2940 int start = __ offset(); 2941 2942 map = reg_save.save_live_registers(masm, 0, &frame_size_in_words); 2943 2944 int frame_complete = __ offset(); 2945 2946 { 2947 Label retaddr; 2948 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2949 2950 __ mov(c_rarg0, rthread); 2951 __ lea(rscratch1, RuntimeAddress(destination)); 2952 2953 __ blr(rscratch1); 2954 __ bind(retaddr); 2955 } 2956 2957 // Set an oopmap for the call site. 2958 // We need this not only for callee-saved registers, but also for volatile 2959 // registers that the compiler might be keeping live across a safepoint. 2960 2961 oop_maps->add_gc_map( __ offset() - start, map); 2962 2963 // r0 contains the address we are going to jump to assuming no exception got installed 2964 2965 // clear last_Java_sp 2966 __ reset_last_Java_frame(false); 2967 // check for pending exceptions 2968 Label pending; 2969 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset())); 2970 __ cbnz(rscratch1, pending); 2971 2972 // get the returned Method* 2973 __ get_vm_result_2(rmethod, rthread); 2974 __ str(rmethod, Address(sp, reg_save.reg_offset_in_bytes(rmethod))); 2975 2976 // r0 is where we want to jump, overwrite rscratch1 which is saved and scratch 2977 __ str(r0, Address(sp, reg_save.rscratch1_offset_in_bytes())); 2978 reg_save.restore_live_registers(masm); 2979 2980 // We are back to the original state on entry and ready to go. 2981 2982 __ br(rscratch1); 2983 2984 // Pending exception after the safepoint 2985 2986 __ bind(pending); 2987 2988 reg_save.restore_live_registers(masm); 2989 2990 // exception pending => remove activation and forward to exception handler 2991 2992 __ str(zr, Address(rthread, JavaThread::vm_result_offset())); 2993 2994 __ ldr(r0, Address(rthread, Thread::pending_exception_offset())); 2995 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2996 2997 // ------------- 2998 // make sure all code is generated 2999 masm->flush(); 3000 3001 // return the blob 3002 // frame_size_words or bytes?? 3003 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true); 3004 } 3005 3006 #ifdef COMPILER2 3007 // This is here instead of runtime_aarch64_64.cpp because it uses SimpleRuntimeFrame 3008 // 3009 //------------------------------generate_exception_blob--------------------------- 3010 // creates exception blob at the end 3011 // Using exception blob, this code is jumped from a compiled method. 3012 // (see emit_exception_handler in x86_64.ad file) 3013 // 3014 // Given an exception pc at a call we call into the runtime for the 3015 // handler in this method. This handler might merely restore state 3016 // (i.e. callee save registers) unwind the frame and jump to the 3017 // exception handler for the nmethod if there is no Java level handler 3018 // for the nmethod. 3019 // 3020 // This code is entered with a jmp. 3021 // 3022 // Arguments: 3023 // r0: exception oop 3024 // r3: exception pc 3025 // 3026 // Results: 3027 // r0: exception oop 3028 // r3: exception pc in caller or ??? 3029 // destination: exception handler of caller 3030 // 3031 // Note: the exception pc MUST be at a call (precise debug information) 3032 // Registers r0, r3, r2, r4, r5, r8-r11 are not callee saved. 3033 // 3034 3035 void OptoRuntime::generate_exception_blob() { 3036 assert(!OptoRuntime::is_callee_saved_register(R3_num), ""); 3037 assert(!OptoRuntime::is_callee_saved_register(R0_num), ""); 3038 assert(!OptoRuntime::is_callee_saved_register(R2_num), ""); 3039 3040 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned"); 3041 3042 // Allocate space for the code 3043 ResourceMark rm; 3044 // Setup code generation tools 3045 CodeBuffer buffer("exception_blob", 2048, 1024); 3046 MacroAssembler* masm = new MacroAssembler(&buffer); 3047 3048 // TODO check various assumptions made here 3049 // 3050 // make sure we do so before running this 3051 3052 address start = __ pc(); 3053 3054 // push rfp and retaddr by hand 3055 // Exception pc is 'return address' for stack walker 3056 __ protect_return_address(); 3057 __ stp(rfp, lr, Address(__ pre(sp, -2 * wordSize))); 3058 // there are no callee save registers and we don't expect an 3059 // arg reg save area 3060 #ifndef PRODUCT 3061 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 3062 #endif 3063 // Store exception in Thread object. We cannot pass any arguments to the 3064 // handle_exception call, since we do not want to make any assumption 3065 // about the size of the frame where the exception happened in. 3066 __ str(r0, Address(rthread, JavaThread::exception_oop_offset())); 3067 __ str(r3, Address(rthread, JavaThread::exception_pc_offset())); 3068 3069 // This call does all the hard work. It checks if an exception handler 3070 // exists in the method. 3071 // If so, it returns the handler address. 3072 // If not, it prepares for stack-unwinding, restoring the callee-save 3073 // registers of the frame being removed. 3074 // 3075 // address OptoRuntime::handle_exception_C(JavaThread* thread) 3076 // 3077 // n.b. 1 gp arg, 0 fp args, integral return type 3078 3079 // the stack should always be aligned 3080 address the_pc = __ pc(); 3081 __ set_last_Java_frame(sp, noreg, the_pc, rscratch1); 3082 __ mov(c_rarg0, rthread); 3083 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, OptoRuntime::handle_exception_C))); 3084 __ blr(rscratch1); 3085 // handle_exception_C is a special VM call which does not require an explicit 3086 // instruction sync afterwards. 3087 3088 // May jump to SVE compiled code 3089 __ reinitialize_ptrue(); 3090 3091 // Set an oopmap for the call site. This oopmap will only be used if we 3092 // are unwinding the stack. Hence, all locations will be dead. 3093 // Callee-saved registers will be the same as the frame above (i.e., 3094 // handle_exception_stub), since they were restored when we got the 3095 // exception. 3096 3097 OopMapSet* oop_maps = new OopMapSet(); 3098 3099 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0)); 3100 3101 __ reset_last_Java_frame(false); 3102 3103 // Restore callee-saved registers 3104 3105 // rfp is an implicitly saved callee saved register (i.e. the calling 3106 // convention will save restore it in prolog/epilog) Other than that 3107 // there are no callee save registers now that adapter frames are gone. 3108 // and we dont' expect an arg reg save area 3109 __ ldp(rfp, r3, Address(__ post(sp, 2 * wordSize))); 3110 __ authenticate_return_address(r3); 3111 3112 // r0: exception handler 3113 3114 // We have a handler in r0 (could be deopt blob). 3115 __ mov(r8, r0); 3116 3117 // Get the exception oop 3118 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset())); 3119 // Get the exception pc in case we are deoptimized 3120 __ ldr(r4, Address(rthread, JavaThread::exception_pc_offset())); 3121 #ifdef ASSERT 3122 __ str(zr, Address(rthread, JavaThread::exception_handler_pc_offset())); 3123 __ str(zr, Address(rthread, JavaThread::exception_pc_offset())); 3124 #endif 3125 // Clear the exception oop so GC no longer processes it as a root. 3126 __ str(zr, Address(rthread, JavaThread::exception_oop_offset())); 3127 3128 // r0: exception oop 3129 // r8: exception handler 3130 // r4: exception pc 3131 // Jump to handler 3132 3133 __ br(r8); 3134 3135 // Make sure all code is generated 3136 masm->flush(); 3137 3138 // Set exception blob 3139 _exception_blob = ExceptionBlob::create(&buffer, oop_maps, SimpleRuntimeFrame::framesize >> 1); 3140 } 3141 3142 #endif // COMPILER2