1 /* 2 * Copyright (c) 2003, 2024, Oracle and/or its affiliates. All rights reserved. 3 * Copyright (c) 2014, 2020, Red Hat Inc. All rights reserved. 4 * Copyright (c) 2020, 2023, Huawei Technologies Co., Ltd. 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/compiledIC.hpp" 31 #include "code/debugInfoRec.hpp" 32 #include "code/vtableStubs.hpp" 33 #include "compiler/oopMap.hpp" 34 #include "gc/shared/barrierSetAssembler.hpp" 35 #include "interpreter/interp_masm.hpp" 36 #include "interpreter/interpreter.hpp" 37 #include "logging/log.hpp" 38 #include "memory/resourceArea.hpp" 39 #include "nativeInst_riscv.hpp" 40 #include "oops/klass.inline.hpp" 41 #include "oops/method.inline.hpp" 42 #include "prims/methodHandles.hpp" 43 #include "runtime/continuation.hpp" 44 #include "runtime/continuationEntry.inline.hpp" 45 #include "runtime/globals.hpp" 46 #include "runtime/jniHandles.hpp" 47 #include "runtime/safepointMechanism.hpp" 48 #include "runtime/sharedRuntime.hpp" 49 #include "runtime/signature.hpp" 50 #include "runtime/stubRoutines.hpp" 51 #include "runtime/vframeArray.hpp" 52 #include "utilities/align.hpp" 53 #include "utilities/formatBuffer.hpp" 54 #include "vmreg_riscv.inline.hpp" 55 #ifdef COMPILER1 56 #include "c1/c1_Runtime1.hpp" 57 #endif 58 #ifdef COMPILER2 59 #include "adfiles/ad_riscv.hpp" 60 #include "opto/runtime.hpp" 61 #endif 62 #if INCLUDE_JVMCI 63 #include "jvmci/jvmciJavaClasses.hpp" 64 #endif 65 66 #define __ masm-> 67 68 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size; 69 70 class SimpleRuntimeFrame { 71 public: 72 73 // Most of the runtime stubs have this simple frame layout. 74 // This class exists to make the layout shared in one place. 75 // Offsets are for compiler stack slots, which are jints. 76 enum layout { 77 // The frame sender code expects that fp will be in the "natural" place and 78 // will override any oopMap setting for it. We must therefore force the layout 79 // so that it agrees with the frame sender code. 80 // we don't expect any arg reg save area so riscv asserts that 81 // frame::arg_reg_save_area_bytes == 0 82 fp_off = 0, fp_off2, 83 return_off, return_off2, 84 framesize 85 }; 86 }; 87 88 class RegisterSaver { 89 const bool _save_vectors; 90 public: 91 RegisterSaver(bool save_vectors) : _save_vectors(UseRVV && save_vectors) {} 92 ~RegisterSaver() {} 93 OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words); 94 void restore_live_registers(MacroAssembler* masm); 95 96 // Offsets into the register save area 97 // Used by deoptimization when it is managing result register 98 // values on its own 99 // gregs:28, float_register:32; except: x1(ra) & x2(sp) & gp(x3) & tp(x4) 100 // |---v0---|<---SP 101 // |---v1---|save vectors only in generate_handler_blob 102 // |-- .. --| 103 // |---v31--|----- 104 // |---f0---| 105 // |---f1---| 106 // | .. | 107 // |---f31--| 108 // |---reserved slot for stack alignment---| 109 // |---x5---| 110 // | x6 | 111 // |---.. --| 112 // |---x31--| 113 // |---fp---| 114 // |---ra---| 115 int v0_offset_in_bytes(void) { return 0; } 116 int f0_offset_in_bytes(void) { 117 int f0_offset = 0; 118 #ifdef COMPILER2 119 if (_save_vectors) { 120 f0_offset += Matcher::scalable_vector_reg_size(T_INT) * VectorRegister::number_of_registers * 121 BytesPerInt; 122 } 123 #endif 124 return f0_offset; 125 } 126 int reserved_slot_offset_in_bytes(void) { 127 return f0_offset_in_bytes() + 128 FloatRegister::max_slots_per_register * 129 FloatRegister::number_of_registers * 130 BytesPerInt; 131 } 132 133 int reg_offset_in_bytes(Register r) { 134 assert (r->encoding() > 4, "ra, sp, gp and tp not saved"); 135 return reserved_slot_offset_in_bytes() + (r->encoding() - 4 /* x1, x2, x3, x4 */) * wordSize; 136 } 137 138 int freg_offset_in_bytes(FloatRegister f) { 139 return f0_offset_in_bytes() + f->encoding() * wordSize; 140 } 141 142 int ra_offset_in_bytes(void) { 143 return reserved_slot_offset_in_bytes() + 144 (Register::number_of_registers - 3) * 145 Register::max_slots_per_register * 146 BytesPerInt; 147 } 148 }; 149 150 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) { 151 int vector_size_in_bytes = 0; 152 int vector_size_in_slots = 0; 153 #ifdef COMPILER2 154 if (_save_vectors) { 155 vector_size_in_bytes += Matcher::scalable_vector_reg_size(T_BYTE); 156 vector_size_in_slots += Matcher::scalable_vector_reg_size(T_INT); 157 } 158 #endif 159 160 int frame_size_in_bytes = align_up(additional_frame_words * wordSize + ra_offset_in_bytes() + wordSize, 16); 161 // OopMap frame size is in compiler stack slots (jint's) not bytes or words 162 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt; 163 // The caller will allocate additional_frame_words 164 int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt; 165 // CodeBlob frame size is in words. 166 int frame_size_in_words = frame_size_in_bytes / wordSize; 167 *total_frame_words = frame_size_in_words; 168 169 // Save Integer, Float and Vector registers. 170 __ enter(); 171 __ push_CPU_state(_save_vectors, vector_size_in_bytes); 172 173 // Set an oopmap for the call site. This oopmap will map all 174 // oop-registers and debug-info registers as callee-saved. This 175 // will allow deoptimization at this safepoint to find all possible 176 // debug-info recordings, as well as let GC find all oops. 177 178 OopMapSet *oop_maps = new OopMapSet(); 179 OopMap* oop_map = new OopMap(frame_size_in_slots, 0); 180 assert_cond(oop_maps != nullptr && oop_map != nullptr); 181 182 int sp_offset_in_slots = 0; 183 int step_in_slots = 0; 184 if (_save_vectors) { 185 step_in_slots = vector_size_in_slots; 186 for (int i = 0; i < VectorRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) { 187 VectorRegister r = as_VectorRegister(i); 188 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg()); 189 } 190 } 191 192 step_in_slots = FloatRegister::max_slots_per_register; 193 for (int i = 0; i < FloatRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) { 194 FloatRegister r = as_FloatRegister(i); 195 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg()); 196 } 197 198 step_in_slots = Register::max_slots_per_register; 199 // skip the slot reserved for alignment, see MacroAssembler::push_reg; 200 // also skip x5 ~ x6 on the stack because they are caller-saved registers. 201 sp_offset_in_slots += Register::max_slots_per_register * 3; 202 // besides, we ignore x0 ~ x4 because push_CPU_state won't push them on the stack. 203 for (int i = 7; i < Register::number_of_registers; i++, sp_offset_in_slots += step_in_slots) { 204 Register r = as_Register(i); 205 if (r != xthread) { 206 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots + additional_frame_slots), r->as_VMReg()); 207 } 208 } 209 210 return oop_map; 211 } 212 213 void RegisterSaver::restore_live_registers(MacroAssembler* masm) { 214 #ifdef COMPILER2 215 __ pop_CPU_state(_save_vectors, Matcher::scalable_vector_reg_size(T_BYTE)); 216 #else 217 #if !INCLUDE_JVMCI 218 assert(!_save_vectors, "vectors are generated only by C2 and JVMCI"); 219 #endif 220 __ pop_CPU_state(_save_vectors); 221 #endif 222 __ leave(); 223 } 224 225 // Is vector's size (in bytes) bigger than a size saved by default? 226 // riscv does not ovlerlay the floating-point registers on vector registers like aarch64. 227 bool SharedRuntime::is_wide_vector(int size) { 228 return UseRVV; 229 } 230 231 // --------------------------------------------------------------------------- 232 // Read the array of BasicTypes from a signature, and compute where the 233 // arguments should go. Values in the VMRegPair regs array refer to 4-byte 234 // quantities. Values less than VMRegImpl::stack0 are registers, those above 235 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer 236 // as framesizes are fixed. 237 // VMRegImpl::stack0 refers to the first slot 0(sp). 238 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. 239 // Register up to Register::number_of_registers) are the 64-bit 240 // integer registers. 241 242 // Note: the INPUTS in sig_bt are in units of Java argument words, 243 // which are 64-bit. The OUTPUTS are in 32-bit units. 244 245 // The Java calling convention is a "shifted" version of the C ABI. 246 // By skipping the first C ABI register we can call non-static jni 247 // methods with small numbers of arguments without having to shuffle 248 // the arguments at all. Since we control the java ABI we ought to at 249 // least get some advantage out of it. 250 251 int SharedRuntime::java_calling_convention(const BasicType *sig_bt, 252 VMRegPair *regs, 253 int total_args_passed) { 254 // Create the mapping between argument positions and 255 // registers. 256 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = { 257 j_rarg0, j_rarg1, j_rarg2, j_rarg3, 258 j_rarg4, j_rarg5, j_rarg6, j_rarg7 259 }; 260 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = { 261 j_farg0, j_farg1, j_farg2, j_farg3, 262 j_farg4, j_farg5, j_farg6, j_farg7 263 }; 264 265 uint int_args = 0; 266 uint fp_args = 0; 267 uint stk_args = 0; 268 269 for (int i = 0; i < total_args_passed; i++) { 270 switch (sig_bt[i]) { 271 case T_BOOLEAN: // fall through 272 case T_CHAR: // fall through 273 case T_BYTE: // fall through 274 case T_SHORT: // fall through 275 case T_INT: 276 if (int_args < Argument::n_int_register_parameters_j) { 277 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 278 } else { 279 stk_args = align_up(stk_args, 2); 280 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 281 stk_args += 1; 282 } 283 break; 284 case T_VOID: 285 // halves of T_LONG or T_DOUBLE 286 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 287 regs[i].set_bad(); 288 break; 289 case T_LONG: // fall through 290 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 291 case T_OBJECT: // fall through 292 case T_ARRAY: // fall through 293 case T_ADDRESS: 294 if (int_args < Argument::n_int_register_parameters_j) { 295 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg()); 296 } else { 297 stk_args = align_up(stk_args, 2); 298 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 299 stk_args += 2; 300 } 301 break; 302 case T_FLOAT: 303 if (fp_args < Argument::n_float_register_parameters_j) { 304 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg()); 305 } else { 306 stk_args = align_up(stk_args, 2); 307 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 308 stk_args += 1; 309 } 310 break; 311 case T_DOUBLE: 312 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 313 if (fp_args < Argument::n_float_register_parameters_j) { 314 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg()); 315 } else { 316 stk_args = align_up(stk_args, 2); 317 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 318 stk_args += 2; 319 } 320 break; 321 default: 322 ShouldNotReachHere(); 323 } 324 } 325 326 return stk_args; 327 } 328 329 // Patch the callers callsite with entry to compiled code if it exists. 330 static void patch_callers_callsite(MacroAssembler *masm) { 331 Label L; 332 __ ld(t0, Address(xmethod, in_bytes(Method::code_offset()))); 333 __ beqz(t0, L); 334 335 __ enter(); 336 __ push_CPU_state(); 337 338 // VM needs caller's callsite 339 // VM needs target method 340 // This needs to be a long call since we will relocate this adapter to 341 // the codeBuffer and it may not reach 342 343 #ifndef PRODUCT 344 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 345 #endif 346 347 __ mv(c_rarg0, xmethod); 348 __ mv(c_rarg1, ra); 349 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)); 350 351 __ pop_CPU_state(); 352 // restore sp 353 __ leave(); 354 __ bind(L); 355 } 356 357 static void gen_c2i_adapter(MacroAssembler *masm, 358 int total_args_passed, 359 int comp_args_on_stack, 360 const BasicType *sig_bt, 361 const VMRegPair *regs, 362 Label& skip_fixup) { 363 // Before we get into the guts of the C2I adapter, see if we should be here 364 // at all. We've come from compiled code and are attempting to jump to the 365 // interpreter, which means the caller made a static call to get here 366 // (vcalls always get a compiled target if there is one). Check for a 367 // compiled target. If there is one, we need to patch the caller's call. 368 patch_callers_callsite(masm); 369 370 __ bind(skip_fixup); 371 372 int words_pushed = 0; 373 374 // Since all args are passed on the stack, total_args_passed * 375 // Interpreter::stackElementSize is the space we need. 376 377 int extraspace = total_args_passed * Interpreter::stackElementSize; 378 379 __ mv(x19_sender_sp, sp); 380 381 // stack is aligned, keep it that way 382 extraspace = align_up(extraspace, 2 * wordSize); 383 384 if (extraspace) { 385 __ sub(sp, sp, extraspace); 386 } 387 388 // Now write the args into the outgoing interpreter space 389 for (int i = 0; i < total_args_passed; i++) { 390 if (sig_bt[i] == T_VOID) { 391 assert(i > 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "missing half"); 392 continue; 393 } 394 395 // offset to start parameters 396 int st_off = (total_args_passed - i - 1) * Interpreter::stackElementSize; 397 int next_off = st_off - Interpreter::stackElementSize; 398 399 // Say 4 args: 400 // i st_off 401 // 0 32 T_LONG 402 // 1 24 T_VOID 403 // 2 16 T_OBJECT 404 // 3 8 T_BOOL 405 // - 0 return address 406 // 407 // However to make thing extra confusing. Because we can fit a Java long/double in 408 // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter 409 // leaves one slot empty and only stores to a single slot. In this case the 410 // slot that is occupied is the T_VOID slot. See I said it was confusing. 411 412 VMReg r_1 = regs[i].first(); 413 VMReg r_2 = regs[i].second(); 414 if (!r_1->is_valid()) { 415 assert(!r_2->is_valid(), ""); 416 continue; 417 } 418 if (r_1->is_stack()) { 419 // memory to memory use t0 420 int ld_off = (r_1->reg2stack() * VMRegImpl::stack_slot_size 421 + extraspace 422 + words_pushed * wordSize); 423 if (!r_2->is_valid()) { 424 __ lwu(t0, Address(sp, ld_off)); 425 __ sd(t0, Address(sp, st_off), /*temp register*/esp); 426 } else { 427 __ ld(t0, Address(sp, ld_off), /*temp register*/esp); 428 429 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG 430 // T_DOUBLE and T_LONG use two slots in the interpreter 431 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 432 // ld_off == LSW, ld_off+wordSize == MSW 433 // st_off == MSW, next_off == LSW 434 __ sd(t0, Address(sp, next_off), /*temp register*/esp); 435 #ifdef ASSERT 436 // Overwrite the unused slot with known junk 437 __ mv(t0, 0xdeadffffdeadaaaaul); 438 __ sd(t0, Address(sp, st_off), /*temp register*/esp); 439 #endif /* ASSERT */ 440 } else { 441 __ sd(t0, Address(sp, st_off), /*temp register*/esp); 442 } 443 } 444 } else if (r_1->is_Register()) { 445 Register r = r_1->as_Register(); 446 if (!r_2->is_valid()) { 447 // must be only an int (or less ) so move only 32bits to slot 448 __ sd(r, Address(sp, st_off)); 449 } else { 450 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG 451 // T_DOUBLE and T_LONG use two slots in the interpreter 452 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 453 // long/double in gpr 454 #ifdef ASSERT 455 // Overwrite the unused slot with known junk 456 __ mv(t0, 0xdeadffffdeadaaabul); 457 __ sd(t0, Address(sp, st_off), /*temp register*/esp); 458 #endif /* ASSERT */ 459 __ sd(r, Address(sp, next_off)); 460 } else { 461 __ sd(r, Address(sp, st_off)); 462 } 463 } 464 } else { 465 assert(r_1->is_FloatRegister(), ""); 466 if (!r_2->is_valid()) { 467 // only a float use just part of the slot 468 __ fsw(r_1->as_FloatRegister(), Address(sp, st_off)); 469 } else { 470 #ifdef ASSERT 471 // Overwrite the unused slot with known junk 472 __ mv(t0, 0xdeadffffdeadaaacul); 473 __ sd(t0, Address(sp, st_off), /*temp register*/esp); 474 #endif /* ASSERT */ 475 __ fsd(r_1->as_FloatRegister(), Address(sp, next_off)); 476 } 477 } 478 } 479 480 __ mv(esp, sp); // Interp expects args on caller's expression stack 481 482 __ ld(t0, Address(xmethod, in_bytes(Method::interpreter_entry_offset()))); 483 __ jr(t0); 484 } 485 486 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, 487 int total_args_passed, 488 int comp_args_on_stack, 489 const BasicType *sig_bt, 490 const VMRegPair *regs) { 491 // Note: x19_sender_sp contains the senderSP on entry. We must 492 // preserve it since we may do a i2c -> c2i transition if we lose a 493 // race where compiled code goes non-entrant while we get args 494 // ready. 495 496 // Cut-out for having no stack args. 497 int comp_words_on_stack = align_up(comp_args_on_stack * VMRegImpl::stack_slot_size, wordSize) >> LogBytesPerWord; 498 if (comp_args_on_stack != 0) { 499 __ sub(t0, sp, comp_words_on_stack * wordSize); 500 __ andi(sp, t0, -16); 501 } 502 503 // Will jump to the compiled code just as if compiled code was doing it. 504 // Pre-load the register-jump target early, to schedule it better. 505 __ ld(t1, Address(xmethod, in_bytes(Method::from_compiled_offset()))); 506 507 #if INCLUDE_JVMCI 508 if (EnableJVMCI) { 509 // check if this call should be routed towards a specific entry point 510 __ ld(t0, Address(xthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset()))); 511 Label no_alternative_target; 512 __ beqz(t0, no_alternative_target); 513 __ mv(t1, t0); 514 __ sd(zr, Address(xthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset()))); 515 __ bind(no_alternative_target); 516 } 517 #endif // INCLUDE_JVMCI 518 519 // Now generate the shuffle code. 520 for (int i = 0; i < total_args_passed; i++) { 521 if (sig_bt[i] == T_VOID) { 522 assert(i > 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "missing half"); 523 continue; 524 } 525 526 // Pick up 0, 1 or 2 words from SP+offset. 527 528 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), 529 "scrambled load targets?"); 530 // Load in argument order going down. 531 int ld_off = (total_args_passed - i - 1) * Interpreter::stackElementSize; 532 // Point to interpreter value (vs. tag) 533 int next_off = ld_off - Interpreter::stackElementSize; 534 535 VMReg r_1 = regs[i].first(); 536 VMReg r_2 = regs[i].second(); 537 if (!r_1->is_valid()) { 538 assert(!r_2->is_valid(), ""); 539 continue; 540 } 541 if (r_1->is_stack()) { 542 // Convert stack slot to an SP offset (+ wordSize to account for return address ) 543 int st_off = regs[i].first()->reg2stack() * VMRegImpl::stack_slot_size; 544 if (!r_2->is_valid()) { 545 __ lw(t0, Address(esp, ld_off)); 546 __ sd(t0, Address(sp, st_off), /*temp register*/t2); 547 } else { 548 // 549 // We are using two optoregs. This can be either T_OBJECT, 550 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates 551 // two slots but only uses one for thr T_LONG or T_DOUBLE case 552 // So we must adjust where to pick up the data to match the 553 // interpreter. 554 // 555 // Interpreter local[n] == MSW, local[n+1] == LSW however locals 556 // are accessed as negative so LSW is at LOW address 557 558 // ld_off is MSW so get LSW 559 const int offset = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ? 560 next_off : ld_off; 561 __ ld(t0, Address(esp, offset)); 562 // st_off is LSW (i.e. reg.first()) 563 __ sd(t0, Address(sp, st_off), /*temp register*/t2); 564 } 565 } else if (r_1->is_Register()) { // Register argument 566 Register r = r_1->as_Register(); 567 if (r_2->is_valid()) { 568 // 569 // We are using two VMRegs. This can be either T_OBJECT, 570 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates 571 // two slots but only uses one for thr T_LONG or T_DOUBLE case 572 // So we must adjust where to pick up the data to match the 573 // interpreter. 574 575 const int offset = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ? 576 next_off : ld_off; 577 578 // this can be a misaligned move 579 __ ld(r, Address(esp, offset)); 580 } else { 581 // sign extend and use a full word? 582 __ lw(r, Address(esp, ld_off)); 583 } 584 } else { 585 if (!r_2->is_valid()) { 586 __ flw(r_1->as_FloatRegister(), Address(esp, ld_off)); 587 } else { 588 __ fld(r_1->as_FloatRegister(), Address(esp, next_off)); 589 } 590 } 591 } 592 593 __ push_cont_fastpath(xthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about 594 595 // 6243940 We might end up in handle_wrong_method if 596 // the callee is deoptimized as we race thru here. If that 597 // happens we don't want to take a safepoint because the 598 // caller frame will look interpreted and arguments are now 599 // "compiled" so it is much better to make this transition 600 // invisible to the stack walking code. Unfortunately if 601 // we try and find the callee by normal means a safepoint 602 // is possible. So we stash the desired callee in the thread 603 // and the vm will find there should this case occur. 604 605 __ sd(xmethod, Address(xthread, JavaThread::callee_target_offset())); 606 607 __ jr(t1); 608 } 609 610 // --------------------------------------------------------------- 611 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, 612 int total_args_passed, 613 int comp_args_on_stack, 614 const BasicType *sig_bt, 615 const VMRegPair *regs, 616 AdapterFingerPrint* fingerprint) { 617 address i2c_entry = __ pc(); 618 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs); 619 620 address c2i_unverified_entry = __ pc(); 621 Label skip_fixup; 622 623 const Register receiver = j_rarg0; 624 const Register data = t1; 625 const Register tmp = t2; // A call-clobbered register not used for arg passing 626 627 // ------------------------------------------------------------------------- 628 // Generate a C2I adapter. On entry we know xmethod holds the Method* during calls 629 // to the interpreter. The args start out packed in the compiled layout. They 630 // need to be unpacked into the interpreter layout. This will almost always 631 // require some stack space. We grow the current (compiled) stack, then repack 632 // the args. We finally end in a jump to the generic interpreter entry point. 633 // On exit from the interpreter, the interpreter will restore our SP (lest the 634 // compiled code, which relies solely on SP and not FP, get sick). 635 636 { 637 __ block_comment("c2i_unverified_entry {"); 638 639 __ ic_check(); 640 __ ld(xmethod, Address(data, CompiledICData::speculated_method_offset())); 641 642 __ ld(t0, Address(xmethod, in_bytes(Method::code_offset()))); 643 __ beqz(t0, skip_fixup); 644 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 645 __ block_comment("} c2i_unverified_entry"); 646 } 647 648 address c2i_entry = __ pc(); 649 650 // Class initialization barrier for static methods 651 address c2i_no_clinit_check_entry = nullptr; 652 if (VM_Version::supports_fast_class_init_checks()) { 653 Label L_skip_barrier; 654 655 { // Bypass the barrier for non-static methods 656 __ lwu(t0, Address(xmethod, Method::access_flags_offset())); 657 __ test_bit(t1, t0, exact_log2(JVM_ACC_STATIC)); 658 __ beqz(t1, L_skip_barrier); // non-static 659 } 660 661 __ load_method_holder(t1, xmethod); 662 __ clinit_barrier(t1, t0, &L_skip_barrier); 663 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub())); 664 665 __ bind(L_skip_barrier); 666 c2i_no_clinit_check_entry = __ pc(); 667 } 668 669 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 670 bs->c2i_entry_barrier(masm); 671 672 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup); 673 674 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry, c2i_no_clinit_check_entry); 675 } 676 677 int SharedRuntime::vector_calling_convention(VMRegPair *regs, 678 uint num_bits, 679 uint total_args_passed) { 680 Unimplemented(); 681 return 0; 682 } 683 684 int SharedRuntime::c_calling_convention(const BasicType *sig_bt, 685 VMRegPair *regs, 686 int total_args_passed) { 687 688 // We return the amount of VMRegImpl stack slots we need to reserve for all 689 // the arguments NOT counting out_preserve_stack_slots. 690 691 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = { 692 c_rarg0, c_rarg1, c_rarg2, c_rarg3, 693 c_rarg4, c_rarg5, c_rarg6, c_rarg7 694 }; 695 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = { 696 c_farg0, c_farg1, c_farg2, c_farg3, 697 c_farg4, c_farg5, c_farg6, c_farg7 698 }; 699 700 uint int_args = 0; 701 uint fp_args = 0; 702 uint stk_args = 0; // inc by 2 each time 703 704 for (int i = 0; i < total_args_passed; i++) { 705 switch (sig_bt[i]) { 706 case T_BOOLEAN: // fall through 707 case T_CHAR: // fall through 708 case T_BYTE: // fall through 709 case T_SHORT: // fall through 710 case T_INT: 711 if (int_args < Argument::n_int_register_parameters_c) { 712 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 713 } else { 714 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 715 stk_args += 2; 716 } 717 break; 718 case T_LONG: // fall through 719 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 720 case T_OBJECT: // fall through 721 case T_ARRAY: // fall through 722 case T_ADDRESS: // fall through 723 case T_METADATA: 724 if (int_args < Argument::n_int_register_parameters_c) { 725 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg()); 726 } else { 727 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 728 stk_args += 2; 729 } 730 break; 731 case T_FLOAT: 732 if (fp_args < Argument::n_float_register_parameters_c) { 733 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg()); 734 } else if (int_args < Argument::n_int_register_parameters_c) { 735 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 736 } else { 737 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 738 stk_args += 2; 739 } 740 break; 741 case T_DOUBLE: 742 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 743 if (fp_args < Argument::n_float_register_parameters_c) { 744 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg()); 745 } else if (int_args < Argument::n_int_register_parameters_c) { 746 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg()); 747 } else { 748 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 749 stk_args += 2; 750 } 751 break; 752 case T_VOID: // Halves of longs and doubles 753 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 754 regs[i].set_bad(); 755 break; 756 default: 757 ShouldNotReachHere(); 758 } 759 } 760 761 return stk_args; 762 } 763 764 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 765 // We always ignore the frame_slots arg and just use the space just below frame pointer 766 // which by this time is free to use 767 switch (ret_type) { 768 case T_FLOAT: 769 __ fsw(f10, Address(fp, -3 * wordSize)); 770 break; 771 case T_DOUBLE: 772 __ fsd(f10, Address(fp, -3 * wordSize)); 773 break; 774 case T_VOID: break; 775 default: { 776 __ sd(x10, Address(fp, -3 * wordSize)); 777 } 778 } 779 } 780 781 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 782 // We always ignore the frame_slots arg and just use the space just below frame pointer 783 // which by this time is free to use 784 switch (ret_type) { 785 case T_FLOAT: 786 __ flw(f10, Address(fp, -3 * wordSize)); 787 break; 788 case T_DOUBLE: 789 __ fld(f10, Address(fp, -3 * wordSize)); 790 break; 791 case T_VOID: break; 792 default: { 793 __ ld(x10, Address(fp, -3 * wordSize)); 794 } 795 } 796 } 797 798 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) { 799 RegSet x; 800 for ( int i = first_arg ; i < arg_count ; i++ ) { 801 if (args[i].first()->is_Register()) { 802 x = x + args[i].first()->as_Register(); 803 } else if (args[i].first()->is_FloatRegister()) { 804 __ addi(sp, sp, -2 * wordSize); 805 __ fsd(args[i].first()->as_FloatRegister(), Address(sp, 0)); 806 } 807 } 808 __ push_reg(x, sp); 809 } 810 811 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) { 812 RegSet x; 813 for ( int i = first_arg ; i < arg_count ; i++ ) { 814 if (args[i].first()->is_Register()) { 815 x = x + args[i].first()->as_Register(); 816 } else { 817 ; 818 } 819 } 820 __ pop_reg(x, sp); 821 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) { 822 if (args[i].first()->is_Register()) { 823 ; 824 } else if (args[i].first()->is_FloatRegister()) { 825 __ fld(args[i].first()->as_FloatRegister(), Address(sp, 0)); 826 __ add(sp, sp, 2 * wordSize); 827 } 828 } 829 } 830 831 static void verify_oop_args(MacroAssembler* masm, 832 const methodHandle& method, 833 const BasicType* sig_bt, 834 const VMRegPair* regs) { 835 const Register temp_reg = x9; // not part of any compiled calling seq 836 if (VerifyOops) { 837 for (int i = 0; i < method->size_of_parameters(); i++) { 838 if (sig_bt[i] == T_OBJECT || 839 sig_bt[i] == T_ARRAY) { 840 VMReg r = regs[i].first(); 841 assert(r->is_valid(), "bad oop arg"); 842 if (r->is_stack()) { 843 __ ld(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size)); 844 __ verify_oop(temp_reg); 845 } else { 846 __ verify_oop(r->as_Register()); 847 } 848 } 849 } 850 } 851 } 852 853 // on exit, sp points to the ContinuationEntry 854 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) { 855 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, ""); 856 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, ""); 857 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, ""); 858 859 stack_slots += (int)ContinuationEntry::size() / wordSize; 860 __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata 861 862 OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize) / VMRegImpl::stack_slot_size, 0 /* arg_slots*/); 863 864 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset())); 865 __ sd(t0, Address(sp, ContinuationEntry::parent_offset())); 866 __ sd(sp, Address(xthread, JavaThread::cont_entry_offset())); 867 868 return map; 869 } 870 871 // on entry c_rarg1 points to the continuation 872 // sp points to ContinuationEntry 873 // c_rarg3 -- isVirtualThread 874 static void fill_continuation_entry(MacroAssembler* masm) { 875 #ifdef ASSERT 876 __ mv(t0, ContinuationEntry::cookie_value()); 877 __ sw(t0, Address(sp, ContinuationEntry::cookie_offset())); 878 #endif 879 880 __ sd(c_rarg1, Address(sp, ContinuationEntry::cont_offset())); 881 __ sw(c_rarg3, Address(sp, ContinuationEntry::flags_offset())); 882 __ sd(zr, Address(sp, ContinuationEntry::chunk_offset())); 883 __ sw(zr, Address(sp, ContinuationEntry::argsize_offset())); 884 __ sw(zr, Address(sp, ContinuationEntry::pin_count_offset())); 885 886 __ ld(t0, Address(xthread, JavaThread::cont_fastpath_offset())); 887 __ sd(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset())); 888 __ ld(t0, Address(xthread, JavaThread::held_monitor_count_offset())); 889 __ sd(t0, Address(sp, ContinuationEntry::parent_held_monitor_count_offset())); 890 891 __ sd(zr, Address(xthread, JavaThread::cont_fastpath_offset())); 892 __ sd(zr, Address(xthread, JavaThread::held_monitor_count_offset())); 893 } 894 895 // on entry, sp points to the ContinuationEntry 896 // on exit, fp points to the spilled fp + 2 * wordSize in the entry frame 897 static void continuation_enter_cleanup(MacroAssembler* masm) { 898 #ifndef PRODUCT 899 Label OK; 900 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset())); 901 __ beq(sp, t0, OK); 902 __ stop("incorrect sp"); 903 __ bind(OK); 904 #endif 905 906 __ ld(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset())); 907 __ sd(t0, Address(xthread, JavaThread::cont_fastpath_offset())); 908 909 if (CheckJNICalls) { 910 // Check if this is a virtual thread continuation 911 Label L_skip_vthread_code; 912 __ lwu(t0, Address(sp, ContinuationEntry::flags_offset())); 913 __ beqz(t0, L_skip_vthread_code); 914 915 // If the held monitor count is > 0 and this vthread is terminating then 916 // it failed to release a JNI monitor. So we issue the same log message 917 // that JavaThread::exit does. 918 __ ld(t0, Address(xthread, JavaThread::jni_monitor_count_offset())); 919 __ beqz(t0, L_skip_vthread_code); 920 921 // Save return value potentially containing the exception oop in callee-saved x9 922 __ mv(x9, x10); 923 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::log_jni_monitor_still_held)); 924 // Restore potential return value 925 __ mv(x10, x9); 926 927 // For vthreads we have to explicitly zero the JNI monitor count of the carrier 928 // on termination. The held count is implicitly zeroed below when we restore from 929 // the parent held count (which has to be zero). 930 __ sd(zr, Address(xthread, JavaThread::jni_monitor_count_offset())); 931 932 __ bind(L_skip_vthread_code); 933 } 934 #ifdef ASSERT 935 else { 936 // Check if this is a virtual thread continuation 937 Label L_skip_vthread_code; 938 __ lwu(t0, Address(sp, ContinuationEntry::flags_offset())); 939 __ beqz(t0, L_skip_vthread_code); 940 941 // See comment just above. If not checking JNI calls the JNI count is only 942 // needed for assertion checking. 943 __ sd(zr, Address(xthread, JavaThread::jni_monitor_count_offset())); 944 945 __ bind(L_skip_vthread_code); 946 } 947 #endif 948 949 __ ld(t0, Address(sp, ContinuationEntry::parent_held_monitor_count_offset())); 950 __ sd(t0, Address(xthread, JavaThread::held_monitor_count_offset())); 951 952 __ ld(t0, Address(sp, ContinuationEntry::parent_offset())); 953 __ sd(t0, Address(xthread, JavaThread::cont_entry_offset())); 954 __ add(fp, sp, (int)ContinuationEntry::size() + 2 * wordSize /* 2 extra words to match up with leave() */); 955 } 956 957 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread) 958 // On entry: c_rarg1 -- the continuation object 959 // c_rarg2 -- isContinue 960 // c_rarg3 -- isVirtualThread 961 static void gen_continuation_enter(MacroAssembler* masm, 962 const methodHandle& method, 963 const BasicType* sig_bt, 964 const VMRegPair* regs, 965 int& exception_offset, 966 OopMapSet*oop_maps, 967 int& frame_complete, 968 int& stack_slots, 969 int& interpreted_entry_offset, 970 int& compiled_entry_offset) { 971 // verify_oop_args(masm, method, sig_bt, regs); 972 Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type); 973 974 address start = __ pc(); 975 976 Label call_thaw, exit; 977 978 // i2i entry used at interp_only_mode only 979 interpreted_entry_offset = __ pc() - start; 980 { 981 #ifdef ASSERT 982 Label is_interp_only; 983 __ lw(t0, Address(xthread, JavaThread::interp_only_mode_offset())); 984 __ bnez(t0, is_interp_only); 985 __ stop("enterSpecial interpreter entry called when not in interp_only_mode"); 986 __ bind(is_interp_only); 987 #endif 988 989 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter) 990 __ ld(c_rarg1, Address(esp, Interpreter::stackElementSize * 2)); 991 __ ld(c_rarg2, Address(esp, Interpreter::stackElementSize * 1)); 992 __ ld(c_rarg3, Address(esp, Interpreter::stackElementSize * 0)); 993 __ push_cont_fastpath(xthread); 994 995 __ enter(); 996 stack_slots = 2; // will be adjusted in setup 997 OopMap* map = continuation_enter_setup(masm, stack_slots); 998 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe, 999 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway. 1000 1001 fill_continuation_entry(masm); 1002 1003 __ bnez(c_rarg2, call_thaw); 1004 1005 // Make sure the call is patchable 1006 __ align(NativeInstruction::instruction_size); 1007 1008 const address tr_call = __ trampoline_call(resolve); 1009 if (tr_call == nullptr) { 1010 fatal("CodeCache is full at gen_continuation_enter"); 1011 } 1012 1013 oop_maps->add_gc_map(__ pc() - start, map); 1014 __ post_call_nop(); 1015 1016 __ j(exit); 1017 1018 address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call); 1019 if (stub == nullptr) { 1020 fatal("CodeCache is full at gen_continuation_enter"); 1021 } 1022 } 1023 1024 // compiled entry 1025 __ align(CodeEntryAlignment); 1026 compiled_entry_offset = __ pc() - start; 1027 1028 __ enter(); 1029 stack_slots = 2; // will be adjusted in setup 1030 OopMap* map = continuation_enter_setup(masm, stack_slots); 1031 frame_complete = __ pc() - start; 1032 1033 fill_continuation_entry(masm); 1034 1035 __ bnez(c_rarg2, call_thaw); 1036 1037 // Make sure the call is patchable 1038 __ align(NativeInstruction::instruction_size); 1039 1040 const address tr_call = __ trampoline_call(resolve); 1041 if (tr_call == nullptr) { 1042 fatal("CodeCache is full at gen_continuation_enter"); 1043 } 1044 1045 oop_maps->add_gc_map(__ pc() - start, map); 1046 __ post_call_nop(); 1047 1048 __ j(exit); 1049 1050 __ bind(call_thaw); 1051 1052 __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw())); 1053 oop_maps->add_gc_map(__ pc() - start, map->deep_copy()); 1054 ContinuationEntry::_return_pc_offset = __ pc() - start; 1055 __ post_call_nop(); 1056 1057 __ bind(exit); 1058 continuation_enter_cleanup(masm); 1059 __ leave(); 1060 __ ret(); 1061 1062 // exception handling 1063 exception_offset = __ pc() - start; 1064 { 1065 __ mv(x9, x10); // save return value contaning the exception oop in callee-saved x9 1066 1067 continuation_enter_cleanup(masm); 1068 1069 __ ld(c_rarg1, Address(fp, -1 * wordSize)); // return address 1070 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), xthread, c_rarg1); 1071 1072 // see OptoRuntime::generate_exception_blob: x10 -- exception oop, x13 -- exception pc 1073 1074 __ mv(x11, x10); // the exception handler 1075 __ mv(x10, x9); // restore return value contaning the exception oop 1076 __ verify_oop(x10); 1077 1078 __ leave(); 1079 __ mv(x13, ra); 1080 __ jr(x11); // the exception handler 1081 } 1082 1083 address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call); 1084 if (stub == nullptr) { 1085 fatal("CodeCache is full at gen_continuation_enter"); 1086 } 1087 } 1088 1089 static void gen_continuation_yield(MacroAssembler* masm, 1090 const methodHandle& method, 1091 const BasicType* sig_bt, 1092 const VMRegPair* regs, 1093 OopMapSet* oop_maps, 1094 int& frame_complete, 1095 int& stack_slots, 1096 int& compiled_entry_offset) { 1097 enum layout { 1098 fp_off, 1099 fp_off2, 1100 return_off, 1101 return_off2, 1102 framesize // inclusive of return address 1103 }; 1104 // assert(is_even(framesize/2), "sp not 16-byte aligned"); 1105 1106 stack_slots = framesize / VMRegImpl::slots_per_word; 1107 assert(stack_slots == 2, "recheck layout"); 1108 1109 address start = __ pc(); 1110 1111 compiled_entry_offset = __ pc() - start; 1112 __ enter(); 1113 1114 __ mv(c_rarg1, sp); 1115 1116 frame_complete = __ pc() - start; 1117 address the_pc = __ pc(); 1118 1119 __ 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 1120 1121 __ mv(c_rarg0, xthread); 1122 __ set_last_Java_frame(sp, fp, the_pc, t0); 1123 __ call_VM_leaf(Continuation::freeze_entry(), 2); 1124 __ reset_last_Java_frame(true); 1125 1126 Label pinned; 1127 1128 __ bnez(x10, pinned); 1129 1130 // We've succeeded, set sp to the ContinuationEntry 1131 __ ld(sp, Address(xthread, JavaThread::cont_entry_offset())); 1132 continuation_enter_cleanup(masm); 1133 1134 __ bind(pinned); // pinned -- return to caller 1135 1136 // handle pending exception thrown by freeze 1137 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset()))); 1138 Label ok; 1139 __ beqz(t0, ok); 1140 __ leave(); 1141 __ la(t0, RuntimeAddress(StubRoutines::forward_exception_entry())); 1142 __ jr(t0); 1143 __ bind(ok); 1144 1145 __ leave(); 1146 __ ret(); 1147 1148 OopMap* map = new OopMap(framesize, 1); 1149 oop_maps->add_gc_map(the_pc - start, map); 1150 } 1151 1152 static void gen_special_dispatch(MacroAssembler* masm, 1153 const methodHandle& method, 1154 const BasicType* sig_bt, 1155 const VMRegPair* regs) { 1156 verify_oop_args(masm, method, sig_bt, regs); 1157 vmIntrinsics::ID iid = method->intrinsic_id(); 1158 1159 // Now write the args into the outgoing interpreter space 1160 bool has_receiver = false; 1161 Register receiver_reg = noreg; 1162 int member_arg_pos = -1; 1163 Register member_reg = noreg; 1164 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid); 1165 if (ref_kind != 0) { 1166 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument 1167 member_reg = x9; // known to be free at this point 1168 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind); 1169 } else if (iid == vmIntrinsics::_invokeBasic) { 1170 has_receiver = true; 1171 } else if (iid == vmIntrinsics::_linkToNative) { 1172 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument 1173 member_reg = x9; // known to be free at this point 1174 } else { 1175 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid)); 1176 } 1177 1178 if (member_reg != noreg) { 1179 // Load the member_arg into register, if necessary. 1180 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs); 1181 VMReg r = regs[member_arg_pos].first(); 1182 if (r->is_stack()) { 1183 __ ld(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size)); 1184 } else { 1185 // no data motion is needed 1186 member_reg = r->as_Register(); 1187 } 1188 } 1189 1190 if (has_receiver) { 1191 // Make sure the receiver is loaded into a register. 1192 assert(method->size_of_parameters() > 0, "oob"); 1193 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object"); 1194 VMReg r = regs[0].first(); 1195 assert(r->is_valid(), "bad receiver arg"); 1196 if (r->is_stack()) { 1197 // Porting note: This assumes that compiled calling conventions always 1198 // pass the receiver oop in a register. If this is not true on some 1199 // platform, pick a temp and load the receiver from stack. 1200 fatal("receiver always in a register"); 1201 receiver_reg = x12; // known to be free at this point 1202 __ ld(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size)); 1203 } else { 1204 // no data motion is needed 1205 receiver_reg = r->as_Register(); 1206 } 1207 } 1208 1209 // Figure out which address we are really jumping to: 1210 MethodHandles::generate_method_handle_dispatch(masm, iid, 1211 receiver_reg, member_reg, /*for_compiler_entry:*/ true); 1212 } 1213 1214 // --------------------------------------------------------------------------- 1215 // Generate a native wrapper for a given method. The method takes arguments 1216 // in the Java compiled code convention, marshals them to the native 1217 // convention (handlizes oops, etc), transitions to native, makes the call, 1218 // returns to java state (possibly blocking), unhandlizes any result and 1219 // returns. 1220 // 1221 // Critical native functions are a shorthand for the use of 1222 // GetPrimtiveArrayCritical and disallow the use of any other JNI 1223 // functions. The wrapper is expected to unpack the arguments before 1224 // passing them to the callee and perform checks before and after the 1225 // native call to ensure that they GCLocker 1226 // lock_critical/unlock_critical semantics are followed. Some other 1227 // parts of JNI setup are skipped like the tear down of the JNI handle 1228 // block and the check for pending exceptions it's impossible for them 1229 // to be thrown. 1230 // 1231 // They are roughly structured like this: 1232 // if (GCLocker::needs_gc()) SharedRuntime::block_for_jni_critical() 1233 // tranistion to thread_in_native 1234 // unpack array arguments and call native entry point 1235 // check for safepoint in progress 1236 // check if any thread suspend flags are set 1237 // call into JVM and possible unlock the JNI critical 1238 // if a GC was suppressed while in the critical native. 1239 // transition back to thread_in_Java 1240 // return to caller 1241 // 1242 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm, 1243 const methodHandle& method, 1244 int compile_id, 1245 BasicType* in_sig_bt, 1246 VMRegPair* in_regs, 1247 BasicType ret_type) { 1248 if (method->is_continuation_native_intrinsic()) { 1249 int exception_offset = -1; 1250 OopMapSet* oop_maps = new OopMapSet(); 1251 int frame_complete = -1; 1252 int stack_slots = -1; 1253 int interpreted_entry_offset = -1; 1254 int vep_offset = -1; 1255 if (method->is_continuation_enter_intrinsic()) { 1256 gen_continuation_enter(masm, 1257 method, 1258 in_sig_bt, 1259 in_regs, 1260 exception_offset, 1261 oop_maps, 1262 frame_complete, 1263 stack_slots, 1264 interpreted_entry_offset, 1265 vep_offset); 1266 } else if (method->is_continuation_yield_intrinsic()) { 1267 gen_continuation_yield(masm, 1268 method, 1269 in_sig_bt, 1270 in_regs, 1271 oop_maps, 1272 frame_complete, 1273 stack_slots, 1274 vep_offset); 1275 } else { 1276 guarantee(false, "Unknown Continuation native intrinsic"); 1277 } 1278 1279 #ifdef ASSERT 1280 if (method->is_continuation_enter_intrinsic()) { 1281 assert(interpreted_entry_offset != -1, "Must be set"); 1282 assert(exception_offset != -1, "Must be set"); 1283 } else { 1284 assert(interpreted_entry_offset == -1, "Must be unset"); 1285 assert(exception_offset == -1, "Must be unset"); 1286 } 1287 assert(frame_complete != -1, "Must be set"); 1288 assert(stack_slots != -1, "Must be set"); 1289 assert(vep_offset != -1, "Must be set"); 1290 #endif 1291 1292 __ flush(); 1293 nmethod* nm = nmethod::new_native_nmethod(method, 1294 compile_id, 1295 masm->code(), 1296 vep_offset, 1297 frame_complete, 1298 stack_slots, 1299 in_ByteSize(-1), 1300 in_ByteSize(-1), 1301 oop_maps, 1302 exception_offset); 1303 if (nm == nullptr) return nm; 1304 if (method->is_continuation_enter_intrinsic()) { 1305 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset); 1306 } else if (method->is_continuation_yield_intrinsic()) { 1307 _cont_doYield_stub = nm; 1308 } else { 1309 guarantee(false, "Unknown Continuation native intrinsic"); 1310 } 1311 return nm; 1312 } 1313 1314 if (method->is_method_handle_intrinsic()) { 1315 vmIntrinsics::ID iid = method->intrinsic_id(); 1316 intptr_t start = (intptr_t)__ pc(); 1317 int vep_offset = ((intptr_t)__ pc()) - start; 1318 1319 // First instruction must be a nop as it may need to be patched on deoptimisation 1320 { 1321 Assembler::IncompressibleRegion ir(masm); // keep the nop as 4 bytes for patching. 1322 MacroAssembler::assert_alignment(__ pc()); 1323 __ nop(); // 4 bytes 1324 } 1325 gen_special_dispatch(masm, 1326 method, 1327 in_sig_bt, 1328 in_regs); 1329 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period 1330 __ flush(); 1331 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually 1332 return nmethod::new_native_nmethod(method, 1333 compile_id, 1334 masm->code(), 1335 vep_offset, 1336 frame_complete, 1337 stack_slots / VMRegImpl::slots_per_word, 1338 in_ByteSize(-1), 1339 in_ByteSize(-1), 1340 (OopMapSet*)nullptr); 1341 } 1342 address native_func = method->native_function(); 1343 assert(native_func != nullptr, "must have function"); 1344 1345 // An OopMap for lock (and class if static) 1346 OopMapSet *oop_maps = new OopMapSet(); 1347 assert_cond(oop_maps != nullptr); 1348 intptr_t start = (intptr_t)__ pc(); 1349 1350 // We have received a description of where all the java arg are located 1351 // on entry to the wrapper. We need to convert these args to where 1352 // the jni function will expect them. To figure out where they go 1353 // we convert the java signature to a C signature by inserting 1354 // the hidden arguments as arg[0] and possibly arg[1] (static method) 1355 1356 const int total_in_args = method->size_of_parameters(); 1357 int total_c_args = total_in_args + (method->is_static() ? 2 : 1); 1358 1359 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1360 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1361 BasicType* in_elem_bt = nullptr; 1362 1363 int argc = 0; 1364 out_sig_bt[argc++] = T_ADDRESS; 1365 if (method->is_static()) { 1366 out_sig_bt[argc++] = T_OBJECT; 1367 } 1368 1369 for (int i = 0; i < total_in_args ; i++) { 1370 out_sig_bt[argc++] = in_sig_bt[i]; 1371 } 1372 1373 // Now figure out where the args must be stored and how much stack space 1374 // they require. 1375 int out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args); 1376 1377 // Compute framesize for the wrapper. We need to handlize all oops in 1378 // incoming registers 1379 1380 // Calculate the total number of stack slots we will need. 1381 1382 // First count the abi requirement plus all of the outgoing args 1383 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots; 1384 1385 // Now the space for the inbound oop handle area 1386 int total_save_slots = 8 * VMRegImpl::slots_per_word; // 8 arguments passed in registers 1387 1388 int oop_handle_offset = stack_slots; 1389 stack_slots += total_save_slots; 1390 1391 // Now any space we need for handlizing a klass if static method 1392 1393 int klass_slot_offset = 0; 1394 int klass_offset = -1; 1395 int lock_slot_offset = 0; 1396 bool is_static = false; 1397 1398 if (method->is_static()) { 1399 klass_slot_offset = stack_slots; 1400 stack_slots += VMRegImpl::slots_per_word; 1401 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; 1402 is_static = true; 1403 } 1404 1405 // Plus a lock if needed 1406 1407 if (method->is_synchronized()) { 1408 lock_slot_offset = stack_slots; 1409 stack_slots += VMRegImpl::slots_per_word; 1410 } 1411 1412 // Now a place (+2) to save return values or temp during shuffling 1413 // + 4 for return address (which we own) and saved fp 1414 stack_slots += 6; 1415 1416 // Ok The space we have allocated will look like: 1417 // 1418 // 1419 // FP-> | | 1420 // | 2 slots (ra) | 1421 // | 2 slots (fp) | 1422 // |---------------------| 1423 // | 2 slots for moves | 1424 // |---------------------| 1425 // | lock box (if sync) | 1426 // |---------------------| <- lock_slot_offset 1427 // | klass (if static) | 1428 // |---------------------| <- klass_slot_offset 1429 // | oopHandle area | 1430 // |---------------------| <- oop_handle_offset (8 java arg registers) 1431 // | outbound memory | 1432 // | based arguments | 1433 // | | 1434 // |---------------------| 1435 // | | 1436 // SP-> | out_preserved_slots | 1437 // 1438 // 1439 1440 1441 // Now compute actual number of stack words we need rounding to make 1442 // stack properly aligned. 1443 stack_slots = align_up(stack_slots, StackAlignmentInSlots); 1444 1445 int stack_size = stack_slots * VMRegImpl::stack_slot_size; 1446 1447 // First thing make an ic check to see if we should even be here 1448 1449 // We are free to use all registers as temps without saving them and 1450 // restoring them except fp. fp is the only callee save register 1451 // as far as the interpreter and the compiler(s) are concerned. 1452 1453 1454 const Register ic_reg = t1; 1455 const Register receiver = j_rarg0; 1456 1457 __ verify_oop(receiver); 1458 assert_different_registers(receiver, t0, t1); 1459 1460 __ ic_check(); 1461 1462 int vep_offset = ((intptr_t)__ pc()) - start; 1463 1464 // If we have to make this method not-entrant we'll overwrite its 1465 // first instruction with a jump. 1466 { 1467 Assembler::IncompressibleRegion ir(masm); // keep the nop as 4 bytes for patching. 1468 MacroAssembler::assert_alignment(__ pc()); 1469 __ nop(); // 4 bytes 1470 } 1471 1472 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) { 1473 Label L_skip_barrier; 1474 __ mov_metadata(t1, method->method_holder()); // InstanceKlass* 1475 __ clinit_barrier(t1, t0, &L_skip_barrier); 1476 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub())); 1477 1478 __ bind(L_skip_barrier); 1479 } 1480 1481 // Generate stack overflow check 1482 __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size())); 1483 1484 // Generate a new frame for the wrapper. 1485 __ enter(); 1486 // -2 because return address is already present and so is saved fp 1487 __ sub(sp, sp, stack_size - 2 * wordSize); 1488 1489 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 1490 assert_cond(bs != nullptr); 1491 bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */); 1492 1493 // Frame is now completed as far as size and linkage. 1494 int frame_complete = ((intptr_t)__ pc()) - start; 1495 1496 // We use x18 as the oop handle for the receiver/klass 1497 // It is callee save so it survives the call to native 1498 1499 const Register oop_handle_reg = x18; 1500 1501 // 1502 // We immediately shuffle the arguments so that any vm call we have to 1503 // make from here on out (sync slow path, jvmti, etc.) we will have 1504 // captured the oops from our caller and have a valid oopMap for 1505 // them. 1506 1507 // ----------------- 1508 // The Grand Shuffle 1509 1510 // The Java calling convention is either equal (linux) or denser (win64) than the 1511 // c calling convention. However the because of the jni_env argument the c calling 1512 // convention always has at least one more (and two for static) arguments than Java. 1513 // Therefore if we move the args from java -> c backwards then we will never have 1514 // a register->register conflict and we don't have to build a dependency graph 1515 // and figure out how to break any cycles. 1516 // 1517 1518 // Record esp-based slot for receiver on stack for non-static methods 1519 int receiver_offset = -1; 1520 1521 // This is a trick. We double the stack slots so we can claim 1522 // the oops in the caller's frame. Since we are sure to have 1523 // more args than the caller doubling is enough to make 1524 // sure we can capture all the incoming oop args from the 1525 // caller. 1526 // 1527 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1528 assert_cond(map != nullptr); 1529 1530 int float_args = 0; 1531 int int_args = 0; 1532 1533 #ifdef ASSERT 1534 bool reg_destroyed[Register::number_of_registers]; 1535 bool freg_destroyed[FloatRegister::number_of_registers]; 1536 for ( int r = 0 ; r < Register::number_of_registers ; r++ ) { 1537 reg_destroyed[r] = false; 1538 } 1539 for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) { 1540 freg_destroyed[f] = false; 1541 } 1542 1543 #endif /* ASSERT */ 1544 1545 // For JNI natives the incoming and outgoing registers are offset upwards. 1546 GrowableArray<int> arg_order(2 * total_in_args); 1547 VMRegPair tmp_vmreg; 1548 tmp_vmreg.set2(x9->as_VMReg()); 1549 1550 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) { 1551 arg_order.push(i); 1552 arg_order.push(c_arg); 1553 } 1554 1555 int temploc = -1; 1556 for (int ai = 0; ai < arg_order.length(); ai += 2) { 1557 int i = arg_order.at(ai); 1558 int c_arg = arg_order.at(ai + 1); 1559 __ block_comment(err_msg("mv %d -> %d", i, c_arg)); 1560 assert(c_arg != -1 && i != -1, "wrong order"); 1561 #ifdef ASSERT 1562 if (in_regs[i].first()->is_Register()) { 1563 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!"); 1564 } else if (in_regs[i].first()->is_FloatRegister()) { 1565 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!"); 1566 } 1567 if (out_regs[c_arg].first()->is_Register()) { 1568 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true; 1569 } else if (out_regs[c_arg].first()->is_FloatRegister()) { 1570 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true; 1571 } 1572 #endif /* ASSERT */ 1573 switch (in_sig_bt[i]) { 1574 case T_ARRAY: 1575 case T_OBJECT: 1576 __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg], 1577 ((i == 0) && (!is_static)), 1578 &receiver_offset); 1579 int_args++; 1580 break; 1581 case T_VOID: 1582 break; 1583 1584 case T_FLOAT: 1585 __ float_move(in_regs[i], out_regs[c_arg]); 1586 float_args++; 1587 break; 1588 1589 case T_DOUBLE: 1590 assert( i + 1 < total_in_args && 1591 in_sig_bt[i + 1] == T_VOID && 1592 out_sig_bt[c_arg + 1] == T_VOID, "bad arg list"); 1593 __ double_move(in_regs[i], out_regs[c_arg]); 1594 float_args++; 1595 break; 1596 1597 case T_LONG : 1598 __ long_move(in_regs[i], out_regs[c_arg]); 1599 int_args++; 1600 break; 1601 1602 case T_ADDRESS: 1603 assert(false, "found T_ADDRESS in java args"); 1604 break; 1605 1606 default: 1607 __ move32_64(in_regs[i], out_regs[c_arg]); 1608 int_args++; 1609 } 1610 } 1611 1612 // point c_arg at the first arg that is already loaded in case we 1613 // need to spill before we call out 1614 int c_arg = total_c_args - total_in_args; 1615 1616 // Pre-load a static method's oop into c_rarg1. 1617 if (method->is_static()) { 1618 1619 // load oop into a register 1620 __ movoop(c_rarg1, 1621 JNIHandles::make_local(method->method_holder()->java_mirror())); 1622 1623 // Now handlize the static class mirror it's known not-null. 1624 __ sd(c_rarg1, Address(sp, klass_offset)); 1625 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); 1626 1627 // Now get the handle 1628 __ la(c_rarg1, Address(sp, klass_offset)); 1629 // and protect the arg if we must spill 1630 c_arg--; 1631 } 1632 1633 // Change state to native (we save the return address in the thread, since it might not 1634 // be pushed on the stack when we do a stack traversal). 1635 // We use the same pc/oopMap repeatedly when we call out 1636 1637 Label native_return; 1638 __ set_last_Java_frame(sp, noreg, native_return, t0); 1639 1640 Label dtrace_method_entry, dtrace_method_entry_done; 1641 { 1642 ExternalAddress target((address)&DTraceMethodProbes); 1643 __ relocate(target.rspec(), [&] { 1644 int32_t offset; 1645 __ la(t0, target.target(), offset); 1646 __ lbu(t0, Address(t0, offset)); 1647 }); 1648 __ bnez(t0, dtrace_method_entry); 1649 __ bind(dtrace_method_entry_done); 1650 } 1651 1652 // RedefineClasses() tracing support for obsolete method entry 1653 if (log_is_enabled(Trace, redefine, class, obsolete)) { 1654 // protect the args we've loaded 1655 save_args(masm, total_c_args, c_arg, out_regs); 1656 __ mov_metadata(c_rarg1, method()); 1657 __ call_VM_leaf( 1658 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 1659 xthread, c_rarg1); 1660 restore_args(masm, total_c_args, c_arg, out_regs); 1661 } 1662 1663 // Lock a synchronized method 1664 1665 // Register definitions used by locking and unlocking 1666 1667 const Register swap_reg = x10; 1668 const Register obj_reg = x9; // Will contain the oop 1669 const Register lock_reg = x30; // Address of compiler lock object (BasicLock) 1670 const Register old_hdr = x30; // value of old header at unlock time 1671 const Register lock_tmp = x31; // Temporary used by lightweight_lock/unlock 1672 const Register tmp = ra; 1673 1674 Label slow_path_lock; 1675 Label lock_done; 1676 1677 if (method->is_synchronized()) { 1678 Label count; 1679 1680 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes(); 1681 1682 // Get the handle (the 2nd argument) 1683 __ mv(oop_handle_reg, c_rarg1); 1684 1685 // Get address of the box 1686 1687 __ la(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 1688 1689 // Load the oop from the handle 1690 __ ld(obj_reg, Address(oop_handle_reg, 0)); 1691 1692 if (LockingMode == LM_MONITOR) { 1693 __ j(slow_path_lock); 1694 } else if (LockingMode == LM_LEGACY) { 1695 // Load (object->mark() | 1) into swap_reg % x10 1696 __ ld(t0, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 1697 __ ori(swap_reg, t0, 1); 1698 1699 // Save (object->mark() | 1) into BasicLock's displaced header 1700 __ sd(swap_reg, Address(lock_reg, mark_word_offset)); 1701 1702 // src -> dest if dest == x10 else x10 <- dest 1703 __ cmpxchg_obj_header(x10, lock_reg, obj_reg, lock_tmp, count, /*fallthrough*/nullptr); 1704 1705 // Test if the oopMark is an obvious stack pointer, i.e., 1706 // 1) (mark & 3) == 0, and 1707 // 2) sp <= mark < mark + os::pagesize() 1708 // These 3 tests can be done by evaluating the following 1709 // expression: ((mark - sp) & (3 - os::vm_page_size())), 1710 // assuming both stack pointer and pagesize have their 1711 // least significant 2 bits clear. 1712 // NOTE: the oopMark is in swap_reg % 10 as the result of cmpxchg 1713 1714 __ sub(swap_reg, swap_reg, sp); 1715 __ andi(swap_reg, swap_reg, 3 - (int)os::vm_page_size()); 1716 1717 // Save the test result, for recursive case, the result is zero 1718 __ sd(swap_reg, Address(lock_reg, mark_word_offset)); 1719 __ bnez(swap_reg, slow_path_lock); 1720 } else { 1721 assert(LockingMode == LM_LIGHTWEIGHT, "must be"); 1722 __ lightweight_lock(obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock); 1723 } 1724 1725 __ bind(count); 1726 __ increment(Address(xthread, JavaThread::held_monitor_count_offset())); 1727 1728 // Slow path will re-enter here 1729 __ bind(lock_done); 1730 } 1731 1732 1733 // Finally just about ready to make the JNI call 1734 1735 // get JNIEnv* which is first argument to native 1736 __ la(c_rarg0, Address(xthread, in_bytes(JavaThread::jni_environment_offset()))); 1737 1738 // Now set thread in native 1739 __ la(t1, Address(xthread, JavaThread::thread_state_offset())); 1740 __ mv(t0, _thread_in_native); 1741 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore); 1742 __ sw(t0, Address(t1)); 1743 1744 __ rt_call(native_func); 1745 1746 __ bind(native_return); 1747 1748 intptr_t return_pc = (intptr_t) __ pc(); 1749 oop_maps->add_gc_map(return_pc - start, map); 1750 1751 // Verify or restore cpu control state after JNI call 1752 __ restore_cpu_control_state_after_jni(t0); 1753 1754 // Unpack native results. 1755 if (ret_type != T_OBJECT && ret_type != T_ARRAY) { 1756 __ cast_primitive_type(ret_type, x10); 1757 } 1758 1759 Label safepoint_in_progress, safepoint_in_progress_done; 1760 Label after_transition; 1761 1762 // Switch thread to "native transition" state before reading the synchronization state. 1763 // This additional state is necessary because reading and testing the synchronization 1764 // state is not atomic w.r.t. GC, as this scenario demonstrates: 1765 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 1766 // VM thread changes sync state to synchronizing and suspends threads for GC. 1767 // Thread A is resumed to finish this native method, but doesn't block here since it 1768 // didn't see any synchronization is progress, and escapes. 1769 __ mv(t0, _thread_in_native_trans); 1770 1771 __ sw(t0, Address(xthread, JavaThread::thread_state_offset())); 1772 1773 // Force this write out before the read below 1774 if (!UseSystemMemoryBarrier) { 1775 __ membar(MacroAssembler::AnyAny); 1776 } 1777 1778 // check for safepoint operation in progress and/or pending suspend requests 1779 { 1780 // We need an acquire here to ensure that any subsequent load of the 1781 // global SafepointSynchronize::_state flag is ordered after this load 1782 // of the thread-local polling word. We don't want this poll to 1783 // return false (i.e. not safepointing) and a later poll of the global 1784 // SafepointSynchronize::_state spuriously to return true. 1785 // This is to avoid a race when we're in a native->Java transition 1786 // racing the code which wakes up from a safepoint. 1787 1788 __ safepoint_poll(safepoint_in_progress, true /* at_return */, true /* acquire */, false /* in_nmethod */); 1789 __ lwu(t0, Address(xthread, JavaThread::suspend_flags_offset())); 1790 __ bnez(t0, safepoint_in_progress); 1791 __ bind(safepoint_in_progress_done); 1792 } 1793 1794 // change thread state 1795 __ la(t1, Address(xthread, JavaThread::thread_state_offset())); 1796 __ mv(t0, _thread_in_Java); 1797 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore); 1798 __ sw(t0, Address(t1)); 1799 __ bind(after_transition); 1800 1801 Label reguard; 1802 Label reguard_done; 1803 __ lbu(t0, Address(xthread, JavaThread::stack_guard_state_offset())); 1804 __ mv(t1, StackOverflow::stack_guard_yellow_reserved_disabled); 1805 __ beq(t0, t1, reguard); 1806 __ bind(reguard_done); 1807 1808 // native result if any is live 1809 1810 // Unlock 1811 Label unlock_done; 1812 Label slow_path_unlock; 1813 if (method->is_synchronized()) { 1814 1815 // Get locked oop from the handle we passed to jni 1816 __ ld(obj_reg, Address(oop_handle_reg, 0)); 1817 1818 Label done, not_recursive; 1819 1820 if (LockingMode == LM_LEGACY) { 1821 // Simple recursive lock? 1822 __ ld(t0, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 1823 __ bnez(t0, not_recursive); 1824 __ decrement(Address(xthread, JavaThread::held_monitor_count_offset())); 1825 __ j(done); 1826 } 1827 1828 __ bind(not_recursive); 1829 1830 // Must save x10 if if it is live now because cmpxchg must use it 1831 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 1832 save_native_result(masm, ret_type, stack_slots); 1833 } 1834 1835 if (LockingMode == LM_MONITOR) { 1836 __ j(slow_path_unlock); 1837 } else if (LockingMode == LM_LEGACY) { 1838 // get address of the stack lock 1839 __ la(x10, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 1840 // get old displaced header 1841 __ ld(old_hdr, Address(x10, 0)); 1842 1843 // Atomic swap old header if oop still contains the stack lock 1844 Label count; 1845 __ cmpxchg_obj_header(x10, old_hdr, obj_reg, lock_tmp, count, &slow_path_unlock); 1846 __ bind(count); 1847 __ decrement(Address(xthread, JavaThread::held_monitor_count_offset())); 1848 } else { 1849 assert(LockingMode == LM_LIGHTWEIGHT, ""); 1850 __ lightweight_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock); 1851 __ decrement(Address(xthread, JavaThread::held_monitor_count_offset())); 1852 } 1853 1854 // slow path re-enters here 1855 __ bind(unlock_done); 1856 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 1857 restore_native_result(masm, ret_type, stack_slots); 1858 } 1859 1860 __ bind(done); 1861 } 1862 1863 Label dtrace_method_exit, dtrace_method_exit_done; 1864 { 1865 ExternalAddress target((address)&DTraceMethodProbes); 1866 __ relocate(target.rspec(), [&] { 1867 int32_t offset; 1868 __ la(t0, target.target(), offset); 1869 __ lbu(t0, Address(t0, offset)); 1870 }); 1871 __ bnez(t0, dtrace_method_exit); 1872 __ bind(dtrace_method_exit_done); 1873 } 1874 1875 __ reset_last_Java_frame(false); 1876 1877 // Unbox oop result, e.g. JNIHandles::resolve result. 1878 if (is_reference_type(ret_type)) { 1879 __ resolve_jobject(x10, x11, x12); 1880 } 1881 1882 if (CheckJNICalls) { 1883 // clear_pending_jni_exception_check 1884 __ sd(zr, Address(xthread, JavaThread::pending_jni_exception_check_fn_offset())); 1885 } 1886 1887 // reset handle block 1888 __ ld(x12, Address(xthread, JavaThread::active_handles_offset())); 1889 __ sd(zr, Address(x12, JNIHandleBlock::top_offset())); 1890 1891 __ leave(); 1892 1893 // Any exception pending? 1894 Label exception_pending; 1895 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset()))); 1896 __ bnez(t0, exception_pending); 1897 1898 // We're done 1899 __ ret(); 1900 1901 // Unexpected paths are out of line and go here 1902 1903 // forward the exception 1904 __ bind(exception_pending); 1905 1906 // and forward the exception 1907 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 1908 1909 // Slow path locking & unlocking 1910 if (method->is_synchronized()) { 1911 1912 __ block_comment("Slow path lock {"); 1913 __ bind(slow_path_lock); 1914 1915 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM 1916 // args are (oop obj, BasicLock* lock, JavaThread* thread) 1917 1918 // protect the args we've loaded 1919 save_args(masm, total_c_args, c_arg, out_regs); 1920 1921 __ mv(c_rarg0, obj_reg); 1922 __ mv(c_rarg1, lock_reg); 1923 __ mv(c_rarg2, xthread); 1924 1925 // Not a leaf but we have last_Java_frame setup as we want 1926 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3); 1927 restore_args(masm, total_c_args, c_arg, out_regs); 1928 1929 #ifdef ASSERT 1930 { Label L; 1931 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset()))); 1932 __ beqz(t0, L); 1933 __ stop("no pending exception allowed on exit from monitorenter"); 1934 __ bind(L); 1935 } 1936 #endif 1937 __ j(lock_done); 1938 1939 __ block_comment("} Slow path lock"); 1940 1941 __ block_comment("Slow path unlock {"); 1942 __ bind(slow_path_unlock); 1943 1944 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) { 1945 save_native_result(masm, ret_type, stack_slots); 1946 } 1947 1948 __ mv(c_rarg2, xthread); 1949 __ la(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 1950 __ mv(c_rarg0, obj_reg); 1951 1952 // Save pending exception around call to VM (which contains an EXCEPTION_MARK) 1953 // NOTE that obj_reg == x9 currently 1954 __ ld(x9, Address(xthread, in_bytes(Thread::pending_exception_offset()))); 1955 __ sd(zr, Address(xthread, in_bytes(Thread::pending_exception_offset()))); 1956 1957 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C)); 1958 1959 #ifdef ASSERT 1960 { 1961 Label L; 1962 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset()))); 1963 __ beqz(t0, L); 1964 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C"); 1965 __ bind(L); 1966 } 1967 #endif /* ASSERT */ 1968 1969 __ sd(x9, Address(xthread, in_bytes(Thread::pending_exception_offset()))); 1970 1971 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) { 1972 restore_native_result(masm, ret_type, stack_slots); 1973 } 1974 __ j(unlock_done); 1975 1976 __ block_comment("} Slow path unlock"); 1977 1978 } // synchronized 1979 1980 // SLOW PATH Reguard the stack if needed 1981 1982 __ bind(reguard); 1983 save_native_result(masm, ret_type, stack_slots); 1984 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages)); 1985 restore_native_result(masm, ret_type, stack_slots); 1986 // and continue 1987 __ j(reguard_done); 1988 1989 // SLOW PATH safepoint 1990 { 1991 __ block_comment("safepoint {"); 1992 __ bind(safepoint_in_progress); 1993 1994 // Don't use call_VM as it will see a possible pending exception and forward it 1995 // and never return here preventing us from clearing _last_native_pc down below. 1996 // 1997 save_native_result(masm, ret_type, stack_slots); 1998 __ mv(c_rarg0, xthread); 1999 #ifndef PRODUCT 2000 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 2001 #endif 2002 __ rt_call(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans)); 2003 2004 // Restore any method result value 2005 restore_native_result(masm, ret_type, stack_slots); 2006 2007 __ j(safepoint_in_progress_done); 2008 __ block_comment("} safepoint"); 2009 } 2010 2011 // SLOW PATH dtrace support 2012 { 2013 __ block_comment("dtrace entry {"); 2014 __ bind(dtrace_method_entry); 2015 2016 // We have all of the arguments setup at this point. We must not touch any register 2017 // argument registers at this point (what if we save/restore them there are no oop? 2018 2019 save_args(masm, total_c_args, c_arg, out_regs); 2020 __ mov_metadata(c_rarg1, method()); 2021 __ call_VM_leaf( 2022 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 2023 xthread, c_rarg1); 2024 restore_args(masm, total_c_args, c_arg, out_regs); 2025 __ j(dtrace_method_entry_done); 2026 __ block_comment("} dtrace entry"); 2027 } 2028 2029 { 2030 __ block_comment("dtrace exit {"); 2031 __ bind(dtrace_method_exit); 2032 save_native_result(masm, ret_type, stack_slots); 2033 __ mov_metadata(c_rarg1, method()); 2034 __ call_VM_leaf( 2035 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 2036 xthread, c_rarg1); 2037 restore_native_result(masm, ret_type, stack_slots); 2038 __ j(dtrace_method_exit_done); 2039 __ block_comment("} dtrace exit"); 2040 } 2041 2042 __ flush(); 2043 2044 nmethod *nm = nmethod::new_native_nmethod(method, 2045 compile_id, 2046 masm->code(), 2047 vep_offset, 2048 frame_complete, 2049 stack_slots / VMRegImpl::slots_per_word, 2050 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), 2051 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size), 2052 oop_maps); 2053 assert(nm != nullptr, "create native nmethod fail!"); 2054 return nm; 2055 } 2056 2057 // this function returns the adjust size (in number of words) to a c2i adapter 2058 // activation for use during deoptimization 2059 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) { 2060 assert(callee_locals >= callee_parameters, 2061 "test and remove; got more parms than locals"); 2062 if (callee_locals < callee_parameters) { 2063 return 0; // No adjustment for negative locals 2064 } 2065 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords; 2066 // diff is counted in stack words 2067 return align_up(diff, 2); 2068 } 2069 2070 //------------------------------generate_deopt_blob---------------------------- 2071 void SharedRuntime::generate_deopt_blob() { 2072 // Allocate space for the code 2073 ResourceMark rm; 2074 // Setup code generation tools 2075 int pad = 0; 2076 #if INCLUDE_JVMCI 2077 if (EnableJVMCI) { 2078 pad += 512; // Increase the buffer size when compiling for JVMCI 2079 } 2080 #endif 2081 CodeBuffer buffer("deopt_blob", 2048 + pad, 1024); 2082 MacroAssembler* masm = new MacroAssembler(&buffer); 2083 int frame_size_in_words = -1; 2084 OopMap* map = nullptr; 2085 OopMapSet *oop_maps = new OopMapSet(); 2086 assert_cond(masm != nullptr && oop_maps != nullptr); 2087 RegisterSaver reg_saver(COMPILER2_OR_JVMCI != 0); 2088 2089 // ------------- 2090 // This code enters when returning to a de-optimized nmethod. A return 2091 // address has been pushed on the stack, and return values are in 2092 // registers. 2093 // If we are doing a normal deopt then we were called from the patched 2094 // nmethod from the point we returned to the nmethod. So the return 2095 // address on the stack is wrong by NativeCall::instruction_size 2096 // We will adjust the value so it looks like we have the original return 2097 // address on the stack (like when we eagerly deoptimized). 2098 // In the case of an exception pending when deoptimizing, we enter 2099 // with a return address on the stack that points after the call we patched 2100 // into the exception handler. We have the following register state from, 2101 // e.g., the forward exception stub (see stubGenerator_riscv.cpp). 2102 // x10: exception oop 2103 // x9: exception handler 2104 // x13: throwing pc 2105 // So in this case we simply jam x13 into the useless return address and 2106 // the stack looks just like we want. 2107 // 2108 // At this point we need to de-opt. We save the argument return 2109 // registers. We call the first C routine, fetch_unroll_info(). This 2110 // routine captures the return values and returns a structure which 2111 // describes the current frame size and the sizes of all replacement frames. 2112 // The current frame is compiled code and may contain many inlined 2113 // functions, each with their own JVM state. We pop the current frame, then 2114 // push all the new frames. Then we call the C routine unpack_frames() to 2115 // populate these frames. Finally unpack_frames() returns us the new target 2116 // address. Notice that callee-save registers are BLOWN here; they have 2117 // already been captured in the vframeArray at the time the return PC was 2118 // patched. 2119 address start = __ pc(); 2120 Label cont; 2121 2122 // Prolog for non exception case! 2123 2124 // Save everything in sight. 2125 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words); 2126 2127 // Normal deoptimization. Save exec mode for unpack_frames. 2128 __ mv(xcpool, Deoptimization::Unpack_deopt); // callee-saved 2129 __ j(cont); 2130 2131 int reexecute_offset = __ pc() - start; 2132 #if INCLUDE_JVMCI && !defined(COMPILER1) 2133 if (EnableJVMCI && UseJVMCICompiler) { 2134 // JVMCI does not use this kind of deoptimization 2135 __ should_not_reach_here(); 2136 } 2137 #endif 2138 2139 // Reexecute case 2140 // return address is the pc describes what bci to do re-execute at 2141 2142 // No need to update map as each call to save_live_registers will produce identical oopmap 2143 (void) reg_saver.save_live_registers(masm, 0, &frame_size_in_words); 2144 2145 __ mv(xcpool, Deoptimization::Unpack_reexecute); // callee-saved 2146 __ j(cont); 2147 2148 #if INCLUDE_JVMCI 2149 Label after_fetch_unroll_info_call; 2150 int implicit_exception_uncommon_trap_offset = 0; 2151 int uncommon_trap_offset = 0; 2152 2153 if (EnableJVMCI) { 2154 implicit_exception_uncommon_trap_offset = __ pc() - start; 2155 2156 __ ld(ra, Address(xthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset()))); 2157 __ sd(zr, Address(xthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset()))); 2158 2159 uncommon_trap_offset = __ pc() - start; 2160 2161 // Save everything in sight. 2162 reg_saver.save_live_registers(masm, 0, &frame_size_in_words); 2163 // fetch_unroll_info needs to call last_java_frame() 2164 Label retaddr; 2165 __ set_last_Java_frame(sp, noreg, retaddr, t0); 2166 2167 __ lw(c_rarg1, Address(xthread, in_bytes(JavaThread::pending_deoptimization_offset()))); 2168 __ mv(t0, -1); 2169 __ sw(t0, Address(xthread, in_bytes(JavaThread::pending_deoptimization_offset()))); 2170 2171 __ mv(xcpool, Deoptimization::Unpack_reexecute); 2172 __ mv(c_rarg0, xthread); 2173 __ orrw(c_rarg2, zr, xcpool); // exec mode 2174 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap)); 2175 __ bind(retaddr); 2176 oop_maps->add_gc_map( __ pc()-start, map->deep_copy()); 2177 2178 __ reset_last_Java_frame(false); 2179 2180 __ j(after_fetch_unroll_info_call); 2181 } // EnableJVMCI 2182 #endif // INCLUDE_JVMCI 2183 2184 int exception_offset = __ pc() - start; 2185 2186 // Prolog for exception case 2187 2188 // all registers are dead at this entry point, except for x10, and 2189 // x13 which contain the exception oop and exception pc 2190 // respectively. Set them in TLS and fall thru to the 2191 // unpack_with_exception_in_tls entry point. 2192 2193 __ sd(x13, Address(xthread, JavaThread::exception_pc_offset())); 2194 __ sd(x10, Address(xthread, JavaThread::exception_oop_offset())); 2195 2196 int exception_in_tls_offset = __ pc() - start; 2197 2198 // new implementation because exception oop is now passed in JavaThread 2199 2200 // Prolog for exception case 2201 // All registers must be preserved because they might be used by LinearScan 2202 // Exceptiop oop and throwing PC are passed in JavaThread 2203 // tos: stack at point of call to method that threw the exception (i.e. only 2204 // args are on the stack, no return address) 2205 2206 // The return address pushed by save_live_registers will be patched 2207 // later with the throwing pc. The correct value is not available 2208 // now because loading it from memory would destroy registers. 2209 2210 // NB: The SP at this point must be the SP of the method that is 2211 // being deoptimized. Deoptimization assumes that the frame created 2212 // here by save_live_registers is immediately below the method's SP. 2213 // This is a somewhat fragile mechanism. 2214 2215 // Save everything in sight. 2216 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words); 2217 2218 // Now it is safe to overwrite any register 2219 2220 // Deopt during an exception. Save exec mode for unpack_frames. 2221 __ mv(xcpool, Deoptimization::Unpack_exception); // callee-saved 2222 2223 // load throwing pc from JavaThread and patch it as the return address 2224 // of the current frame. Then clear the field in JavaThread 2225 2226 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset())); 2227 __ sd(x13, Address(fp, frame::return_addr_offset * wordSize)); 2228 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset())); 2229 2230 #ifdef ASSERT 2231 // verify that there is really an exception oop in JavaThread 2232 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset())); 2233 __ verify_oop(x10); 2234 2235 // verify that there is no pending exception 2236 Label no_pending_exception; 2237 __ ld(t0, Address(xthread, Thread::pending_exception_offset())); 2238 __ beqz(t0, no_pending_exception); 2239 __ stop("must not have pending exception here"); 2240 __ bind(no_pending_exception); 2241 #endif 2242 2243 __ bind(cont); 2244 2245 // Call C code. Need thread and this frame, but NOT official VM entry 2246 // crud. We cannot block on this call, no GC can happen. 2247 // 2248 // UnrollBlock* fetch_unroll_info(JavaThread* thread) 2249 2250 // fetch_unroll_info needs to call last_java_frame(). 2251 2252 Label retaddr; 2253 __ set_last_Java_frame(sp, noreg, retaddr, t0); 2254 #ifdef ASSERT 2255 { 2256 Label L; 2257 __ ld(t0, Address(xthread, 2258 JavaThread::last_Java_fp_offset())); 2259 __ beqz(t0, L); 2260 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared"); 2261 __ bind(L); 2262 } 2263 #endif // ASSERT 2264 __ mv(c_rarg0, xthread); 2265 __ mv(c_rarg1, xcpool); 2266 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info)); 2267 __ bind(retaddr); 2268 2269 // Need to have an oopmap that tells fetch_unroll_info where to 2270 // find any register it might need. 2271 oop_maps->add_gc_map(__ pc() - start, map); 2272 2273 __ reset_last_Java_frame(false); 2274 2275 #if INCLUDE_JVMCI 2276 if (EnableJVMCI) { 2277 __ bind(after_fetch_unroll_info_call); 2278 } 2279 #endif 2280 2281 // Load UnrollBlock* into x15 2282 __ mv(x15, x10); 2283 2284 __ lwu(xcpool, Address(x15, Deoptimization::UnrollBlock::unpack_kind_offset())); 2285 Label noException; 2286 __ mv(t0, Deoptimization::Unpack_exception); 2287 __ bne(xcpool, t0, noException); // Was exception pending? 2288 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset())); 2289 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset())); 2290 __ sd(zr, Address(xthread, JavaThread::exception_oop_offset())); 2291 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset())); 2292 2293 __ verify_oop(x10); 2294 2295 // Overwrite the result registers with the exception results. 2296 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10))); 2297 2298 __ bind(noException); 2299 2300 // Only register save data is on the stack. 2301 // Now restore the result registers. Everything else is either dead 2302 // or captured in the vframeArray. 2303 2304 // Restore fp result register 2305 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10))); 2306 // Restore integer result register 2307 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10))); 2308 2309 // Pop all of the register save area off the stack 2310 __ add(sp, sp, frame_size_in_words * wordSize); 2311 2312 // All of the register save area has been popped of the stack. Only the 2313 // return address remains. 2314 2315 // Pop all the frames we must move/replace. 2316 // 2317 // Frame picture (youngest to oldest) 2318 // 1: self-frame (no frame link) 2319 // 2: deopting frame (no frame link) 2320 // 3: caller of deopting frame (could be compiled/interpreted). 2321 // 2322 // Note: by leaving the return address of self-frame on the stack 2323 // and using the size of frame 2 to adjust the stack 2324 // when we are done the return to frame 3 will still be on the stack. 2325 2326 // Pop deoptimized frame 2327 __ lwu(x12, Address(x15, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset())); 2328 __ sub(x12, x12, 2 * wordSize); 2329 __ add(sp, sp, x12); 2330 __ ld(fp, Address(sp, 0)); 2331 __ ld(ra, Address(sp, wordSize)); 2332 __ addi(sp, sp, 2 * wordSize); 2333 // RA should now be the return address to the caller (3) 2334 2335 #ifdef ASSERT 2336 // Compilers generate code that bang the stack by as much as the 2337 // interpreter would need. So this stack banging should never 2338 // trigger a fault. Verify that it does not on non product builds. 2339 __ lwu(x9, Address(x15, Deoptimization::UnrollBlock::total_frame_sizes_offset())); 2340 __ bang_stack_size(x9, x12); 2341 #endif 2342 // Load address of array of frame pcs into x12 2343 __ ld(x12, Address(x15, Deoptimization::UnrollBlock::frame_pcs_offset())); 2344 2345 // Load address of array of frame sizes into x14 2346 __ ld(x14, Address(x15, Deoptimization::UnrollBlock::frame_sizes_offset())); 2347 2348 // Load counter into x13 2349 __ lwu(x13, Address(x15, Deoptimization::UnrollBlock::number_of_frames_offset())); 2350 2351 // Now adjust the caller's stack to make up for the extra locals 2352 // but record the original sp so that we can save it in the skeletal interpreter 2353 // frame and the stack walking of interpreter_sender will get the unextended sp 2354 // value and not the "real" sp value. 2355 2356 const Register sender_sp = x16; 2357 2358 __ mv(sender_sp, sp); 2359 __ lwu(x9, Address(x15, 2360 Deoptimization::UnrollBlock:: 2361 caller_adjustment_offset())); 2362 __ sub(sp, sp, x9); 2363 2364 // Push interpreter frames in a loop 2365 __ mv(t0, 0xDEADDEAD); // Make a recognizable pattern 2366 __ mv(t1, t0); 2367 Label loop; 2368 __ bind(loop); 2369 __ ld(x9, Address(x14, 0)); // Load frame size 2370 __ addi(x14, x14, wordSize); 2371 __ sub(x9, x9, 2 * wordSize); // We'll push pc and fp by hand 2372 __ ld(ra, Address(x12, 0)); // Load pc 2373 __ addi(x12, x12, wordSize); 2374 __ enter(); // Save old & set new fp 2375 __ sub(sp, sp, x9); // Prolog 2376 // This value is corrected by layout_activation_impl 2377 __ sd(zr, Address(fp, frame::interpreter_frame_last_sp_offset * wordSize)); 2378 __ sd(sender_sp, Address(fp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable 2379 __ mv(sender_sp, sp); // Pass sender_sp to next frame 2380 __ addi(x13, x13, -1); // Decrement counter 2381 __ bnez(x13, loop); 2382 2383 // Re-push self-frame 2384 __ ld(ra, Address(x12)); 2385 __ enter(); 2386 2387 // Allocate a full sized register save area. We subtract 2 because 2388 // enter() just pushed 2 words 2389 __ sub(sp, sp, (frame_size_in_words - 2) * wordSize); 2390 2391 // Restore frame locals after moving the frame 2392 __ fsd(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10))); 2393 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10))); 2394 2395 // Call C code. Need thread but NOT official VM entry 2396 // crud. We cannot block on this call, no GC can happen. Call should 2397 // restore return values to their stack-slots with the new SP. 2398 // 2399 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode) 2400 2401 // Use fp because the frames look interpreted now 2402 // Don't need the precise return PC here, just precise enough to point into this code blob. 2403 address the_pc = __ pc(); 2404 __ set_last_Java_frame(sp, fp, the_pc, t0); 2405 2406 __ mv(c_rarg0, xthread); 2407 __ mv(c_rarg1, xcpool); // second arg: exec_mode 2408 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)); 2409 2410 // Set an oopmap for the call site 2411 // Use the same PC we used for the last java frame 2412 oop_maps->add_gc_map(the_pc - start, 2413 new OopMap(frame_size_in_words, 0)); 2414 2415 // Clear fp AND pc 2416 __ reset_last_Java_frame(true); 2417 2418 // Collect return values 2419 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10))); 2420 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10))); 2421 2422 // Pop self-frame. 2423 __ leave(); // Epilog 2424 2425 // Jump to interpreter 2426 __ ret(); 2427 2428 // Make sure all code is generated 2429 masm->flush(); 2430 2431 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words); 2432 assert(_deopt_blob != nullptr, "create deoptimization blob fail!"); 2433 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset); 2434 #if INCLUDE_JVMCI 2435 if (EnableJVMCI) { 2436 _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset); 2437 _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset); 2438 } 2439 #endif 2440 } 2441 2442 // Number of stack slots between incoming argument block and the start of 2443 // a new frame. The PROLOG must add this many slots to the stack. The 2444 // EPILOG must remove this many slots. 2445 // RISCV needs two words for RA (return address) and FP (frame pointer). 2446 uint SharedRuntime::in_preserve_stack_slots() { 2447 return 2 * VMRegImpl::slots_per_word; 2448 } 2449 2450 uint SharedRuntime::out_preserve_stack_slots() { 2451 return 0; 2452 } 2453 2454 #ifdef COMPILER2 2455 //------------------------------generate_uncommon_trap_blob-------------------- 2456 void SharedRuntime::generate_uncommon_trap_blob() { 2457 // Allocate space for the code 2458 ResourceMark rm; 2459 // Setup code generation tools 2460 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024); 2461 MacroAssembler* masm = new MacroAssembler(&buffer); 2462 assert_cond(masm != nullptr); 2463 2464 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned"); 2465 2466 address start = __ pc(); 2467 2468 // Push self-frame. We get here with a return address in RA 2469 // and sp should be 16 byte aligned 2470 // push fp and retaddr by hand 2471 __ addi(sp, sp, -2 * wordSize); 2472 __ sd(ra, Address(sp, wordSize)); 2473 __ sd(fp, Address(sp, 0)); 2474 // we don't expect an arg reg save area 2475 #ifndef PRODUCT 2476 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 2477 #endif 2478 // compiler left unloaded_class_index in j_rarg0 move to where the 2479 // runtime expects it. 2480 __ sign_extend(c_rarg1, j_rarg0, 32); 2481 2482 // we need to set the past SP to the stack pointer of the stub frame 2483 // and the pc to the address where this runtime call will return 2484 // although actually any pc in this code blob will do). 2485 Label retaddr; 2486 __ set_last_Java_frame(sp, noreg, retaddr, t0); 2487 2488 // Call C code. Need thread but NOT official VM entry 2489 // crud. We cannot block on this call, no GC can happen. Call should 2490 // capture callee-saved registers as well as return values. 2491 // 2492 // UnrollBlock* uncommon_trap(JavaThread* thread, jint unloaded_class_index, jint exec_mode) 2493 // 2494 // n.b. 3 gp args, 0 fp args, integral return type 2495 2496 __ mv(c_rarg0, xthread); 2497 __ mv(c_rarg2, Deoptimization::Unpack_uncommon_trap); 2498 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap)); 2499 __ bind(retaddr); 2500 2501 // Set an oopmap for the call site 2502 OopMapSet* oop_maps = new OopMapSet(); 2503 OopMap* map = new OopMap(SimpleRuntimeFrame::framesize, 0); 2504 assert_cond(oop_maps != nullptr && map != nullptr); 2505 2506 // location of fp is known implicitly by the frame sender code 2507 2508 oop_maps->add_gc_map(__ pc() - start, map); 2509 2510 __ reset_last_Java_frame(false); 2511 2512 // move UnrollBlock* into x14 2513 __ mv(x14, x10); 2514 2515 #ifdef ASSERT 2516 { Label L; 2517 __ lwu(t0, Address(x14, Deoptimization::UnrollBlock::unpack_kind_offset())); 2518 __ mv(t1, Deoptimization::Unpack_uncommon_trap); 2519 __ beq(t0, t1, L); 2520 __ stop("SharedRuntime::generate_uncommon_trap_blob: expected Unpack_uncommon_trap"); 2521 __ bind(L); 2522 } 2523 #endif 2524 2525 // Pop all the frames we must move/replace. 2526 // 2527 // Frame picture (youngest to oldest) 2528 // 1: self-frame (no frame link) 2529 // 2: deopting frame (no frame link) 2530 // 3: caller of deopting frame (could be compiled/interpreted). 2531 2532 __ add(sp, sp, (SimpleRuntimeFrame::framesize) << LogBytesPerInt); // Epilog! 2533 2534 // Pop deoptimized frame (int) 2535 __ lwu(x12, Address(x14, 2536 Deoptimization::UnrollBlock:: 2537 size_of_deoptimized_frame_offset())); 2538 __ sub(x12, x12, 2 * wordSize); 2539 __ add(sp, sp, x12); 2540 __ ld(fp, Address(sp, 0)); 2541 __ ld(ra, Address(sp, wordSize)); 2542 __ addi(sp, sp, 2 * wordSize); 2543 // RA should now be the return address to the caller (3) frame 2544 2545 #ifdef ASSERT 2546 // Compilers generate code that bang the stack by as much as the 2547 // interpreter would need. So this stack banging should never 2548 // trigger a fault. Verify that it does not on non product builds. 2549 __ lwu(x11, Address(x14, 2550 Deoptimization::UnrollBlock:: 2551 total_frame_sizes_offset())); 2552 __ bang_stack_size(x11, x12); 2553 #endif 2554 2555 // Load address of array of frame pcs into x12 (address*) 2556 __ ld(x12, Address(x14, 2557 Deoptimization::UnrollBlock::frame_pcs_offset())); 2558 2559 // Load address of array of frame sizes into x15 (intptr_t*) 2560 __ ld(x15, Address(x14, 2561 Deoptimization::UnrollBlock:: 2562 frame_sizes_offset())); 2563 2564 // Counter 2565 __ lwu(x13, Address(x14, 2566 Deoptimization::UnrollBlock:: 2567 number_of_frames_offset())); // (int) 2568 2569 // Now adjust the caller's stack to make up for the extra locals but 2570 // record the original sp so that we can save it in the skeletal 2571 // interpreter frame and the stack walking of interpreter_sender 2572 // will get the unextended sp value and not the "real" sp value. 2573 2574 const Register sender_sp = t1; // temporary register 2575 2576 __ lwu(x11, Address(x14, 2577 Deoptimization::UnrollBlock:: 2578 caller_adjustment_offset())); // (int) 2579 __ mv(sender_sp, sp); 2580 __ sub(sp, sp, x11); 2581 2582 // Push interpreter frames in a loop 2583 Label loop; 2584 __ bind(loop); 2585 __ ld(x11, Address(x15, 0)); // Load frame size 2586 __ sub(x11, x11, 2 * wordSize); // We'll push pc and fp by hand 2587 __ ld(ra, Address(x12, 0)); // Save return address 2588 __ enter(); // and old fp & set new fp 2589 __ sub(sp, sp, x11); // Prolog 2590 __ sd(sender_sp, Address(fp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable 2591 // This value is corrected by layout_activation_impl 2592 __ sd(zr, Address(fp, frame::interpreter_frame_last_sp_offset * wordSize)); 2593 __ mv(sender_sp, sp); // Pass sender_sp to next frame 2594 __ add(x15, x15, wordSize); // Bump array pointer (sizes) 2595 __ add(x12, x12, wordSize); // Bump array pointer (pcs) 2596 __ subw(x13, x13, 1); // Decrement counter 2597 __ bgtz(x13, loop); 2598 __ ld(ra, Address(x12, 0)); // save final return address 2599 // Re-push self-frame 2600 __ enter(); // & old fp & set new fp 2601 2602 // Use fp because the frames look interpreted now 2603 // Save "the_pc" since it cannot easily be retrieved using the last_java_SP after we aligned SP. 2604 // Don't need the precise return PC here, just precise enough to point into this code blob. 2605 address the_pc = __ pc(); 2606 __ set_last_Java_frame(sp, fp, the_pc, t0); 2607 2608 // Call C code. Need thread but NOT official VM entry 2609 // crud. We cannot block on this call, no GC can happen. Call should 2610 // restore return values to their stack-slots with the new SP. 2611 // 2612 // BasicType unpack_frames(JavaThread* thread, int exec_mode) 2613 // 2614 2615 // n.b. 2 gp args, 0 fp args, integral return type 2616 2617 // sp should already be aligned 2618 __ mv(c_rarg0, xthread); 2619 __ mv(c_rarg1, Deoptimization::Unpack_uncommon_trap); 2620 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)); 2621 2622 // Set an oopmap for the call site 2623 // Use the same PC we used for the last java frame 2624 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0)); 2625 2626 // Clear fp AND pc 2627 __ reset_last_Java_frame(true); 2628 2629 // Pop self-frame. 2630 __ leave(); // Epilog 2631 2632 // Jump to interpreter 2633 __ ret(); 2634 2635 // Make sure all code is generated 2636 masm->flush(); 2637 2638 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, 2639 SimpleRuntimeFrame::framesize >> 1); 2640 } 2641 #endif // COMPILER2 2642 2643 //------------------------------generate_handler_blob------ 2644 // 2645 // Generate a special Compile2Runtime blob that saves all registers, 2646 // and setup oopmap. 2647 // 2648 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) { 2649 ResourceMark rm; 2650 OopMapSet *oop_maps = new OopMapSet(); 2651 assert_cond(oop_maps != nullptr); 2652 OopMap* map = nullptr; 2653 2654 // Allocate space for the code. Setup code generation tools. 2655 CodeBuffer buffer("handler_blob", 2048, 1024); 2656 MacroAssembler* masm = new MacroAssembler(&buffer); 2657 assert_cond(masm != nullptr); 2658 2659 address start = __ pc(); 2660 address call_pc = nullptr; 2661 int frame_size_in_words = -1; 2662 bool cause_return = (poll_type == POLL_AT_RETURN); 2663 RegisterSaver reg_saver(poll_type == POLL_AT_VECTOR_LOOP /* save_vectors */); 2664 2665 // Save Integer and Float registers. 2666 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words); 2667 2668 // The following is basically a call_VM. However, we need the precise 2669 // address of the call in order to generate an oopmap. Hence, we do all the 2670 // work ourselves. 2671 2672 Label retaddr; 2673 __ set_last_Java_frame(sp, noreg, retaddr, t0); 2674 2675 // The return address must always be correct so that frame constructor never 2676 // sees an invalid pc. 2677 2678 if (!cause_return) { 2679 // overwrite the return address pushed by save_live_registers 2680 // Additionally, x18 is a callee-saved register so we can look at 2681 // it later to determine if someone changed the return address for 2682 // us! 2683 __ ld(x18, Address(xthread, JavaThread::saved_exception_pc_offset())); 2684 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize)); 2685 } 2686 2687 // Do the call 2688 __ mv(c_rarg0, xthread); 2689 __ rt_call(call_ptr); 2690 __ bind(retaddr); 2691 2692 // Set an oopmap for the call site. This oopmap will map all 2693 // oop-registers and debug-info registers as callee-saved. This 2694 // will allow deoptimization at this safepoint to find all possible 2695 // debug-info recordings, as well as let GC find all oops. 2696 2697 oop_maps->add_gc_map( __ pc() - start, map); 2698 2699 Label noException; 2700 2701 __ reset_last_Java_frame(false); 2702 2703 __ membar(MacroAssembler::LoadLoad | MacroAssembler::LoadStore); 2704 2705 __ ld(t0, Address(xthread, Thread::pending_exception_offset())); 2706 __ beqz(t0, noException); 2707 2708 // Exception pending 2709 2710 reg_saver.restore_live_registers(masm); 2711 2712 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2713 2714 // No exception case 2715 __ bind(noException); 2716 2717 Label no_adjust, bail; 2718 if (!cause_return) { 2719 // If our stashed return pc was modified by the runtime we avoid touching it 2720 __ ld(t0, Address(fp, frame::return_addr_offset * wordSize)); 2721 __ bne(x18, t0, no_adjust); 2722 2723 #ifdef ASSERT 2724 // Verify the correct encoding of the poll we're about to skip. 2725 // See NativeInstruction::is_lwu_to_zr() 2726 __ lwu(t0, Address(x18)); 2727 __ andi(t1, t0, 0b0000011); 2728 __ mv(t2, 0b0000011); 2729 __ bne(t1, t2, bail); // 0-6:0b0000011 2730 __ srli(t1, t0, 7); 2731 __ andi(t1, t1, 0b00000); 2732 __ bnez(t1, bail); // 7-11:0b00000 2733 __ srli(t1, t0, 12); 2734 __ andi(t1, t1, 0b110); 2735 __ mv(t2, 0b110); 2736 __ bne(t1, t2, bail); // 12-14:0b110 2737 #endif 2738 // Adjust return pc forward to step over the safepoint poll instruction 2739 __ add(x18, x18, NativeInstruction::instruction_size); 2740 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize)); 2741 } 2742 2743 __ bind(no_adjust); 2744 // Normal exit, restore registers and exit. 2745 2746 reg_saver.restore_live_registers(masm); 2747 __ ret(); 2748 2749 #ifdef ASSERT 2750 __ bind(bail); 2751 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected"); 2752 #endif 2753 2754 // Make sure all code is generated 2755 masm->flush(); 2756 2757 // Fill-out other meta info 2758 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words); 2759 } 2760 2761 // 2762 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss 2763 // 2764 // Generate a stub that calls into vm to find out the proper destination 2765 // of a java call. All the argument registers are live at this point 2766 // but since this is generic code we don't know what they are and the caller 2767 // must do any gc of the args. 2768 // 2769 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) { 2770 assert(StubRoutines::forward_exception_entry() != nullptr, "must be generated before"); 2771 2772 // allocate space for the code 2773 ResourceMark rm; 2774 2775 CodeBuffer buffer(name, 1000, 512); 2776 MacroAssembler* masm = new MacroAssembler(&buffer); 2777 assert_cond(masm != nullptr); 2778 2779 int frame_size_in_words = -1; 2780 RegisterSaver reg_saver(false /* save_vectors */); 2781 2782 OopMapSet *oop_maps = new OopMapSet(); 2783 assert_cond(oop_maps != nullptr); 2784 OopMap* map = nullptr; 2785 2786 int start = __ offset(); 2787 2788 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words); 2789 2790 int frame_complete = __ offset(); 2791 2792 { 2793 Label retaddr; 2794 __ set_last_Java_frame(sp, noreg, retaddr, t0); 2795 2796 __ mv(c_rarg0, xthread); 2797 __ rt_call(destination); 2798 __ bind(retaddr); 2799 } 2800 2801 // Set an oopmap for the call site. 2802 // We need this not only for callee-saved registers, but also for volatile 2803 // registers that the compiler might be keeping live across a safepoint. 2804 2805 oop_maps->add_gc_map( __ offset() - start, map); 2806 2807 // x10 contains the address we are going to jump to assuming no exception got installed 2808 2809 // clear last_Java_sp 2810 __ reset_last_Java_frame(false); 2811 // check for pending exceptions 2812 Label pending; 2813 __ ld(t0, Address(xthread, Thread::pending_exception_offset())); 2814 __ bnez(t0, pending); 2815 2816 // get the returned Method* 2817 __ get_vm_result_2(xmethod, xthread); 2818 __ sd(xmethod, Address(sp, reg_saver.reg_offset_in_bytes(xmethod))); 2819 2820 // x10 is where we want to jump, overwrite t0 which is saved and temporary 2821 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(t0))); 2822 reg_saver.restore_live_registers(masm); 2823 2824 // We are back to the original state on entry and ready to go. 2825 2826 __ jr(t0); 2827 2828 // Pending exception after the safepoint 2829 2830 __ bind(pending); 2831 2832 reg_saver.restore_live_registers(masm); 2833 2834 // exception pending => remove activation and forward to exception handler 2835 2836 __ sd(zr, Address(xthread, JavaThread::vm_result_offset())); 2837 2838 __ ld(x10, Address(xthread, Thread::pending_exception_offset())); 2839 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2840 2841 // ------------- 2842 // make sure all code is generated 2843 masm->flush(); 2844 2845 // return the blob 2846 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true); 2847 } 2848 2849 #ifdef COMPILER2 2850 //------------------------------generate_exception_blob--------------------------- 2851 // creates exception blob at the end 2852 // Using exception blob, this code is jumped from a compiled method. 2853 // (see emit_exception_handler in riscv.ad file) 2854 // 2855 // Given an exception pc at a call we call into the runtime for the 2856 // handler in this method. This handler might merely restore state 2857 // (i.e. callee save registers) unwind the frame and jump to the 2858 // exception handler for the nmethod if there is no Java level handler 2859 // for the nmethod. 2860 // 2861 // This code is entered with a jmp. 2862 // 2863 // Arguments: 2864 // x10: exception oop 2865 // x13: exception pc 2866 // 2867 // Results: 2868 // x10: exception oop 2869 // x13: exception pc in caller 2870 // destination: exception handler of caller 2871 // 2872 // Note: the exception pc MUST be at a call (precise debug information) 2873 // Registers x10, x13, x12, x14, x15, t0 are not callee saved. 2874 // 2875 2876 void OptoRuntime::generate_exception_blob() { 2877 assert(!OptoRuntime::is_callee_saved_register(R13_num), ""); 2878 assert(!OptoRuntime::is_callee_saved_register(R10_num), ""); 2879 assert(!OptoRuntime::is_callee_saved_register(R12_num), ""); 2880 2881 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned"); 2882 2883 // Allocate space for the code 2884 ResourceMark rm; 2885 // Setup code generation tools 2886 CodeBuffer buffer("exception_blob", 2048, 1024); 2887 MacroAssembler* masm = new MacroAssembler(&buffer); 2888 assert_cond(masm != nullptr); 2889 2890 // TODO check various assumptions made here 2891 // 2892 // make sure we do so before running this 2893 2894 address start = __ pc(); 2895 2896 // push fp and retaddr by hand 2897 // Exception pc is 'return address' for stack walker 2898 __ addi(sp, sp, -2 * wordSize); 2899 __ sd(ra, Address(sp, wordSize)); 2900 __ sd(fp, Address(sp)); 2901 // there are no callee save registers and we don't expect an 2902 // arg reg save area 2903 #ifndef PRODUCT 2904 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 2905 #endif 2906 // Store exception in Thread object. We cannot pass any arguments to the 2907 // handle_exception call, since we do not want to make any assumption 2908 // about the size of the frame where the exception happened in. 2909 __ sd(x10, Address(xthread, JavaThread::exception_oop_offset())); 2910 __ sd(x13, Address(xthread, JavaThread::exception_pc_offset())); 2911 2912 // This call does all the hard work. It checks if an exception handler 2913 // exists in the method. 2914 // If so, it returns the handler address. 2915 // If not, it prepares for stack-unwinding, restoring the callee-save 2916 // registers of the frame being removed. 2917 // 2918 // address OptoRuntime::handle_exception_C(JavaThread* thread) 2919 // 2920 // n.b. 1 gp arg, 0 fp args, integral return type 2921 2922 // the stack should always be aligned 2923 address the_pc = __ pc(); 2924 __ set_last_Java_frame(sp, noreg, the_pc, t0); 2925 __ mv(c_rarg0, xthread); 2926 __ rt_call(CAST_FROM_FN_PTR(address, OptoRuntime::handle_exception_C)); 2927 2928 // handle_exception_C is a special VM call which does not require an explicit 2929 // instruction sync afterwards. 2930 2931 // Set an oopmap for the call site. This oopmap will only be used if we 2932 // are unwinding the stack. Hence, all locations will be dead. 2933 // Callee-saved registers will be the same as the frame above (i.e., 2934 // handle_exception_stub), since they were restored when we got the 2935 // exception. 2936 2937 OopMapSet* oop_maps = new OopMapSet(); 2938 assert_cond(oop_maps != nullptr); 2939 2940 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0)); 2941 2942 __ reset_last_Java_frame(false); 2943 2944 // Restore callee-saved registers 2945 2946 // fp is an implicitly saved callee saved register (i.e. the calling 2947 // convention will save restore it in prolog/epilog) Other than that 2948 // there are no callee save registers now that adapter frames are gone. 2949 // and we dont' expect an arg reg save area 2950 __ ld(fp, Address(sp)); 2951 __ ld(x13, Address(sp, wordSize)); 2952 __ addi(sp, sp , 2 * wordSize); 2953 2954 // x10: exception handler 2955 2956 // We have a handler in x10 (could be deopt blob). 2957 __ mv(t0, x10); 2958 2959 // Get the exception oop 2960 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset())); 2961 // Get the exception pc in case we are deoptimized 2962 __ ld(x14, Address(xthread, JavaThread::exception_pc_offset())); 2963 #ifdef ASSERT 2964 __ sd(zr, Address(xthread, JavaThread::exception_handler_pc_offset())); 2965 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset())); 2966 #endif 2967 // Clear the exception oop so GC no longer processes it as a root. 2968 __ sd(zr, Address(xthread, JavaThread::exception_oop_offset())); 2969 2970 // x10: exception oop 2971 // t0: exception handler 2972 // x14: exception pc 2973 // Jump to handler 2974 2975 __ jr(t0); 2976 2977 // Make sure all code is generated 2978 masm->flush(); 2979 2980 // Set exception blob 2981 _exception_blob = ExceptionBlob::create(&buffer, oop_maps, SimpleRuntimeFrame::framesize >> 1); 2982 } 2983 #endif // COMPILER2