1 /* 2 * Copyright (c) 2003, 2020, Oracle and/or its affiliates. All rights reserved. 3 * Copyright (c) 2014, 2020, Red Hat Inc. All rights reserved. 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 5 * 6 * This code is free software; you can redistribute it and/or modify it 7 * under the terms of the GNU General Public License version 2 only, as 8 * published by the Free Software Foundation. 9 * 10 * This code is distributed in the hope that it will be useful, but WITHOUT 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 13 * version 2 for more details (a copy is included in the LICENSE file that 14 * accompanied this code). 15 * 16 * You should have received a copy of the GNU General Public License version 17 * 2 along with this work; if not, write to the Free Software Foundation, 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 19 * 20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 21 * or visit www.oracle.com if you need additional information or have any 22 * questions. 23 * 24 */ 25 26 #include "precompiled.hpp" 27 #include "asm/macroAssembler.hpp" 28 #include "asm/macroAssembler.inline.hpp" 29 #include "code/debugInfoRec.hpp" 30 #include "code/icBuffer.hpp" 31 #include "code/vtableStubs.hpp" 32 #include "gc/shared/barrierSetAssembler.hpp" 33 #include "interpreter/interpreter.hpp" 34 #include "interpreter/interp_masm.hpp" 35 #include "logging/log.hpp" 36 #include "memory/resourceArea.hpp" 37 #include "nativeInst_aarch64.hpp" 38 #include "oops/compiledICHolder.hpp" 39 #include "oops/klass.inline.hpp" 40 #include "runtime/safepointMechanism.hpp" 41 #include "runtime/sharedRuntime.hpp" 42 #include "runtime/vframeArray.hpp" 43 #include "utilities/align.hpp" 44 #include "vmreg_aarch64.inline.hpp" 45 #ifdef COMPILER1 46 #include "c1/c1_Runtime1.hpp" 47 #endif 48 #ifdef COMPILER2 49 #include "adfiles/ad_aarch64.hpp" 50 #include "opto/runtime.hpp" 51 #endif 52 #if INCLUDE_JVMCI 53 #include "jvmci/jvmciJavaClasses.hpp" 54 #endif 55 56 #define __ masm-> 57 58 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size; 59 60 class SimpleRuntimeFrame { 61 62 public: 63 64 // Most of the runtime stubs have this simple frame layout. 65 // This class exists to make the layout shared in one place. 66 // Offsets are for compiler stack slots, which are jints. 67 enum layout { 68 // The frame sender code expects that rbp will be in the "natural" place and 69 // will override any oopMap setting for it. We must therefore force the layout 70 // so that it agrees with the frame sender code. 71 // we don't expect any arg reg save area so aarch64 asserts that 72 // frame::arg_reg_save_area_bytes == 0 73 rbp_off = 0, 74 rbp_off2, 75 return_off, return_off2, 76 framesize 77 }; 78 }; 79 80 // FIXME -- this is used by C1 81 class RegisterSaver { 82 public: 83 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words, bool save_vectors = false); 84 static void restore_live_registers(MacroAssembler* masm, bool restore_vectors = false); 85 86 // Offsets into the register save area 87 // Used by deoptimization when it is managing result register 88 // values on its own 89 90 static int r0_offset_in_bytes(void) { return (32 + r0->encoding()) * wordSize; } 91 static int reg_offset_in_bytes(Register r) { return r0_offset_in_bytes() + r->encoding() * wordSize; } 92 static int rmethod_offset_in_bytes(void) { return reg_offset_in_bytes(rmethod); } 93 static int rscratch1_offset_in_bytes(void) { return (32 + rscratch1->encoding()) * wordSize; } 94 static int v0_offset_in_bytes(void) { return 0; } 95 static int return_offset_in_bytes(void) { return (32 /* floats*/ + 31 /* gregs*/) * wordSize; } 96 97 // During deoptimization only the result registers need to be restored, 98 // all the other values have already been extracted. 99 static void restore_result_registers(MacroAssembler* masm); 100 101 // Capture info about frame layout 102 enum layout { 103 fpu_state_off = 0, 104 fpu_state_end = fpu_state_off + FPUStateSizeInWords - 1, 105 // The frame sender code expects that rfp will be in 106 // the "natural" place and will override any oopMap 107 // setting for it. We must therefore force the layout 108 // so that it agrees with the frame sender code. 109 r0_off = fpu_state_off + FPUStateSizeInWords, 110 rfp_off = r0_off + (RegisterImpl::number_of_registers - 2) * RegisterImpl::max_slots_per_register, 111 return_off = rfp_off + RegisterImpl::max_slots_per_register, // slot for return address 112 reg_save_size = return_off + RegisterImpl::max_slots_per_register}; 113 114 }; 115 116 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words, bool save_vectors) { 117 #if COMPILER2_OR_JVMCI 118 if (save_vectors) { 119 // Save upper half of vector registers 120 int vect_words = FloatRegisterImpl::number_of_registers * FloatRegisterImpl::extra_save_slots_per_register / 121 VMRegImpl::slots_per_word; 122 additional_frame_words += vect_words; 123 } 124 #else 125 assert(!save_vectors, "vectors are generated only by C2 and JVMCI"); 126 #endif 127 128 int frame_size_in_bytes = align_up(additional_frame_words * wordSize + 129 reg_save_size * BytesPerInt, 16); 130 // OopMap frame size is in compiler stack slots (jint's) not bytes or words 131 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt; 132 // The caller will allocate additional_frame_words 133 int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt; 134 // CodeBlob frame size is in words. 135 int frame_size_in_words = frame_size_in_bytes / wordSize; 136 *total_frame_words = frame_size_in_words; 137 138 // Save Integer and Float registers. 139 __ enter(); 140 __ push_CPU_state(save_vectors); 141 142 // Set an oopmap for the call site. This oopmap will map all 143 // oop-registers and debug-info registers as callee-saved. This 144 // will allow deoptimization at this safepoint to find all possible 145 // debug-info recordings, as well as let GC find all oops. 146 147 OopMapSet *oop_maps = new OopMapSet(); 148 OopMap* oop_map = new OopMap(frame_size_in_slots, 0); 149 150 for (int i = 0; i < RegisterImpl::number_of_registers; i++) { 151 Register r = as_Register(i); 152 if (r <= rfp && r != rscratch1 && r != rscratch2) { 153 // SP offsets are in 4-byte words. 154 // Register slots are 8 bytes wide, 32 floating-point registers. 155 int sp_offset = RegisterImpl::max_slots_per_register * i + 156 FloatRegisterImpl::save_slots_per_register * FloatRegisterImpl::number_of_registers; 157 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset + additional_frame_slots), 158 r->as_VMReg()); 159 } 160 } 161 162 for (int i = 0; i < FloatRegisterImpl::number_of_registers; i++) { 163 FloatRegister r = as_FloatRegister(i); 164 int sp_offset = save_vectors ? (FloatRegisterImpl::max_slots_per_register * i) : 165 (FloatRegisterImpl::save_slots_per_register * i); 166 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset), 167 r->as_VMReg()); 168 } 169 170 return oop_map; 171 } 172 173 void RegisterSaver::restore_live_registers(MacroAssembler* masm, bool restore_vectors) { 174 #if !COMPILER2_OR_JVMCI 175 assert(!restore_vectors, "vectors are generated only by C2 and JVMCI"); 176 #endif 177 __ pop_CPU_state(restore_vectors); 178 __ leave(); 179 180 } 181 182 void RegisterSaver::restore_result_registers(MacroAssembler* masm) { 183 184 // Just restore result register. Only used by deoptimization. By 185 // now any callee save register that needs to be restored to a c2 186 // caller of the deoptee has been extracted into the vframeArray 187 // and will be stuffed into the c2i adapter we create for later 188 // restoration so only result registers need to be restored here. 189 190 // Restore fp result register 191 __ ldrd(v0, Address(sp, v0_offset_in_bytes())); 192 // Restore integer result register 193 __ ldr(r0, Address(sp, r0_offset_in_bytes())); 194 195 // Pop all of the register save are off the stack 196 __ add(sp, sp, align_up(return_offset_in_bytes(), 16)); 197 } 198 199 // Is vector's size (in bytes) bigger than a size saved by default? 200 // 8 bytes vector registers are saved by default on AArch64. 201 bool SharedRuntime::is_wide_vector(int size) { 202 return size > 8; 203 } 204 205 size_t SharedRuntime::trampoline_size() { 206 return 16; 207 } 208 209 void SharedRuntime::generate_trampoline(MacroAssembler *masm, address destination) { 210 __ mov(rscratch1, destination); 211 __ br(rscratch1); 212 } 213 214 // The java_calling_convention describes stack locations as ideal slots on 215 // a frame with no abi restrictions. Since we must observe abi restrictions 216 // (like the placement of the register window) the slots must be biased by 217 // the following value. 218 static int reg2offset_in(VMReg r) { 219 // Account for saved rfp and lr 220 // This should really be in_preserve_stack_slots 221 return (r->reg2stack() + 4) * VMRegImpl::stack_slot_size; 222 } 223 224 static int reg2offset_out(VMReg r) { 225 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 226 } 227 228 // --------------------------------------------------------------------------- 229 // Read the array of BasicTypes from a signature, and compute where the 230 // arguments should go. Values in the VMRegPair regs array refer to 4-byte 231 // quantities. Values less than VMRegImpl::stack0 are registers, those above 232 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer 233 // as framesizes are fixed. 234 // VMRegImpl::stack0 refers to the first slot 0(sp). 235 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register 236 // up to RegisterImpl::number_of_registers) are the 64-bit 237 // integer registers. 238 239 // Note: the INPUTS in sig_bt are in units of Java argument words, 240 // which are 64-bit. The OUTPUTS are in 32-bit units. 241 242 // The Java calling convention is a "shifted" version of the C ABI. 243 // By skipping the first C ABI register we can call non-static jni 244 // methods with small numbers of arguments without having to shuffle 245 // the arguments at all. Since we control the java ABI we ought to at 246 // least get some advantage out of it. 247 248 int SharedRuntime::java_calling_convention(const BasicType *sig_bt, 249 VMRegPair *regs, 250 int total_args_passed, 251 int is_outgoing) { 252 253 // Create the mapping between argument positions and 254 // registers. 255 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = { 256 j_rarg0, j_rarg1, j_rarg2, j_rarg3, j_rarg4, j_rarg5, j_rarg6, j_rarg7 257 }; 258 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = { 259 j_farg0, j_farg1, j_farg2, j_farg3, 260 j_farg4, j_farg5, j_farg6, j_farg7 261 }; 262 263 264 uint int_args = 0; 265 uint fp_args = 0; 266 uint stk_args = 0; // inc by 2 each time 267 268 for (int i = 0; i < total_args_passed; i++) { 269 switch (sig_bt[i]) { 270 case T_BOOLEAN: 271 case T_CHAR: 272 case T_BYTE: 273 case T_SHORT: 274 case T_INT: 275 if (int_args < Argument::n_int_register_parameters_j) { 276 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 277 } else { 278 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 279 stk_args += 2; 280 } 281 break; 282 case T_VOID: 283 // halves of T_LONG or T_DOUBLE 284 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 285 regs[i].set_bad(); 286 break; 287 case T_LONG: 288 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 289 // fall through 290 case T_OBJECT: 291 case T_ARRAY: 292 case T_ADDRESS: 293 if (int_args < Argument::n_int_register_parameters_j) { 294 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg()); 295 } else { 296 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 297 stk_args += 2; 298 } 299 break; 300 case T_FLOAT: 301 if (fp_args < Argument::n_float_register_parameters_j) { 302 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg()); 303 } else { 304 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 305 stk_args += 2; 306 } 307 break; 308 case T_DOUBLE: 309 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 310 if (fp_args < Argument::n_float_register_parameters_j) { 311 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg()); 312 } else { 313 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 314 stk_args += 2; 315 } 316 break; 317 default: 318 ShouldNotReachHere(); 319 break; 320 } 321 } 322 323 return align_up(stk_args, 2); 324 } 325 326 // Patch the callers callsite with entry to compiled code if it exists. 327 static void patch_callers_callsite(MacroAssembler *masm) { 328 Label L; 329 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset()))); 330 __ cbz(rscratch1, L); 331 332 __ enter(); 333 __ push_CPU_state(); 334 335 // VM needs caller's callsite 336 // VM needs target method 337 // This needs to be a long call since we will relocate this adapter to 338 // the codeBuffer and it may not reach 339 340 #ifndef PRODUCT 341 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 342 #endif 343 344 __ mov(c_rarg0, rmethod); 345 __ mov(c_rarg1, lr); 346 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite))); 347 __ blr(rscratch1); 348 __ maybe_isb(); 349 350 __ pop_CPU_state(); 351 // restore sp 352 __ leave(); 353 __ bind(L); 354 } 355 356 static void gen_c2i_adapter(MacroAssembler *masm, 357 int total_args_passed, 358 int comp_args_on_stack, 359 const BasicType *sig_bt, 360 const VMRegPair *regs, 361 Label& skip_fixup) { 362 // Before we get into the guts of the C2I adapter, see if we should be here 363 // at all. We've come from compiled code and are attempting to jump to the 364 // interpreter, which means the caller made a static call to get here 365 // (vcalls always get a compiled target if there is one). Check for a 366 // compiled target. If there is one, we need to patch the caller's call. 367 patch_callers_callsite(masm); 368 369 __ bind(skip_fixup); 370 371 int words_pushed = 0; 372 373 // Since all args are passed on the stack, total_args_passed * 374 // Interpreter::stackElementSize is the space we need. 375 376 int extraspace = total_args_passed * Interpreter::stackElementSize; 377 378 __ mov(r13, sp); 379 380 // stack is aligned, keep it that way 381 extraspace = align_up(extraspace, 2*wordSize); 382 383 if (extraspace) 384 __ sub(sp, sp, extraspace); 385 386 // Now write the args into the outgoing interpreter space 387 for (int i = 0; i < total_args_passed; i++) { 388 if (sig_bt[i] == T_VOID) { 389 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 390 continue; 391 } 392 393 // offset to start parameters 394 int st_off = (total_args_passed - i - 1) * Interpreter::stackElementSize; 395 int next_off = st_off - Interpreter::stackElementSize; 396 397 // Say 4 args: 398 // i st_off 399 // 0 32 T_LONG 400 // 1 24 T_VOID 401 // 2 16 T_OBJECT 402 // 3 8 T_BOOL 403 // - 0 return address 404 // 405 // However to make thing extra confusing. Because we can fit a long/double in 406 // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter 407 // leaves one slot empty and only stores to a single slot. In this case the 408 // slot that is occupied is the T_VOID slot. See I said it was confusing. 409 410 VMReg r_1 = regs[i].first(); 411 VMReg r_2 = regs[i].second(); 412 if (!r_1->is_valid()) { 413 assert(!r_2->is_valid(), ""); 414 continue; 415 } 416 if (r_1->is_stack()) { 417 // memory to memory use rscratch1 418 int ld_off = (r_1->reg2stack() * VMRegImpl::stack_slot_size 419 + extraspace 420 + words_pushed * wordSize); 421 if (!r_2->is_valid()) { 422 // sign extend?? 423 __ ldrw(rscratch1, Address(sp, ld_off)); 424 __ str(rscratch1, Address(sp, st_off)); 425 426 } else { 427 428 __ ldr(rscratch1, Address(sp, ld_off)); 429 430 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG 431 // T_DOUBLE and T_LONG use two slots in the interpreter 432 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 433 // ld_off == LSW, ld_off+wordSize == MSW 434 // st_off == MSW, next_off == LSW 435 __ str(rscratch1, Address(sp, next_off)); 436 #ifdef ASSERT 437 // Overwrite the unused slot with known junk 438 __ mov(rscratch1, 0xdeadffffdeadaaaaul); 439 __ str(rscratch1, Address(sp, st_off)); 440 #endif /* ASSERT */ 441 } else { 442 __ str(rscratch1, Address(sp, st_off)); 443 } 444 } 445 } else if (r_1->is_Register()) { 446 Register r = r_1->as_Register(); 447 if (!r_2->is_valid()) { 448 // must be only an int (or less ) so move only 32bits to slot 449 // why not sign extend?? 450 __ str(r, Address(sp, st_off)); 451 } else { 452 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG 453 // T_DOUBLE and T_LONG use two slots in the interpreter 454 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 455 // long/double in gpr 456 #ifdef ASSERT 457 // Overwrite the unused slot with known junk 458 __ mov(rscratch1, 0xdeadffffdeadaaabul); 459 __ str(rscratch1, Address(sp, st_off)); 460 #endif /* ASSERT */ 461 __ str(r, Address(sp, next_off)); 462 } else { 463 __ str(r, Address(sp, st_off)); 464 } 465 } 466 } else { 467 assert(r_1->is_FloatRegister(), ""); 468 if (!r_2->is_valid()) { 469 // only a float use just part of the slot 470 __ strs(r_1->as_FloatRegister(), Address(sp, st_off)); 471 } else { 472 #ifdef ASSERT 473 // Overwrite the unused slot with known junk 474 __ mov(rscratch1, 0xdeadffffdeadaaacul); 475 __ str(rscratch1, Address(sp, st_off)); 476 #endif /* ASSERT */ 477 __ strd(r_1->as_FloatRegister(), Address(sp, next_off)); 478 } 479 } 480 } 481 482 __ mov(esp, sp); // Interp expects args on caller's expression stack 483 484 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::interpreter_entry_offset()))); 485 __ br(rscratch1); 486 } 487 488 489 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, 490 int total_args_passed, 491 int comp_args_on_stack, 492 const BasicType *sig_bt, 493 const VMRegPair *regs) { 494 495 // Note: r13 contains the senderSP on entry. We must preserve it since 496 // we may do a i2c -> c2i transition if we lose a race where compiled 497 // code goes non-entrant while we get args ready. 498 499 // In addition we use r13 to locate all the interpreter args because 500 // we must align the stack to 16 bytes. 501 502 // Adapters are frameless. 503 504 // An i2c adapter is frameless because the *caller* frame, which is 505 // interpreted, routinely repairs its own esp (from 506 // interpreter_frame_last_sp), even if a callee has modified the 507 // stack pointer. It also recalculates and aligns sp. 508 509 // A c2i adapter is frameless because the *callee* frame, which is 510 // interpreted, routinely repairs its caller's sp (from sender_sp, 511 // which is set up via the senderSP register). 512 513 // In other words, if *either* the caller or callee is interpreted, we can 514 // get the stack pointer repaired after a call. 515 516 // This is why c2i and i2c adapters cannot be indefinitely composed. 517 // In particular, if a c2i adapter were to somehow call an i2c adapter, 518 // both caller and callee would be compiled methods, and neither would 519 // clean up the stack pointer changes performed by the two adapters. 520 // If this happens, control eventually transfers back to the compiled 521 // caller, but with an uncorrected stack, causing delayed havoc. 522 523 if (VerifyAdapterCalls && 524 (Interpreter::code() != NULL || StubRoutines::code1() != NULL)) { 525 #if 0 526 // So, let's test for cascading c2i/i2c adapters right now. 527 // assert(Interpreter::contains($return_addr) || 528 // StubRoutines::contains($return_addr), 529 // "i2c adapter must return to an interpreter frame"); 530 __ block_comment("verify_i2c { "); 531 Label L_ok; 532 if (Interpreter::code() != NULL) 533 range_check(masm, rax, r11, 534 Interpreter::code()->code_start(), Interpreter::code()->code_end(), 535 L_ok); 536 if (StubRoutines::code1() != NULL) 537 range_check(masm, rax, r11, 538 StubRoutines::code1()->code_begin(), StubRoutines::code1()->code_end(), 539 L_ok); 540 if (StubRoutines::code2() != NULL) 541 range_check(masm, rax, r11, 542 StubRoutines::code2()->code_begin(), StubRoutines::code2()->code_end(), 543 L_ok); 544 const char* msg = "i2c adapter must return to an interpreter frame"; 545 __ block_comment(msg); 546 __ stop(msg); 547 __ bind(L_ok); 548 __ block_comment("} verify_i2ce "); 549 #endif 550 } 551 552 // Cut-out for having no stack args. 553 int comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord; 554 if (comp_args_on_stack) { 555 __ sub(rscratch1, sp, comp_words_on_stack * wordSize); 556 __ andr(sp, rscratch1, -16); 557 } 558 559 // Will jump to the compiled code just as if compiled code was doing it. 560 // Pre-load the register-jump target early, to schedule it better. 561 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::from_compiled_offset()))); 562 563 #if INCLUDE_JVMCI 564 if (EnableJVMCI || UseAOT) { 565 // check if this call should be routed towards a specific entry point 566 __ ldr(rscratch2, Address(rthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset()))); 567 Label no_alternative_target; 568 __ cbz(rscratch2, no_alternative_target); 569 __ mov(rscratch1, rscratch2); 570 __ str(zr, Address(rthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset()))); 571 __ bind(no_alternative_target); 572 } 573 #endif // INCLUDE_JVMCI 574 575 // Now generate the shuffle code. 576 for (int i = 0; i < total_args_passed; i++) { 577 if (sig_bt[i] == T_VOID) { 578 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 579 continue; 580 } 581 582 // Pick up 0, 1 or 2 words from SP+offset. 583 584 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), 585 "scrambled load targets?"); 586 // Load in argument order going down. 587 int ld_off = (total_args_passed - i - 1)*Interpreter::stackElementSize; 588 // Point to interpreter value (vs. tag) 589 int next_off = ld_off - Interpreter::stackElementSize; 590 // 591 // 592 // 593 VMReg r_1 = regs[i].first(); 594 VMReg r_2 = regs[i].second(); 595 if (!r_1->is_valid()) { 596 assert(!r_2->is_valid(), ""); 597 continue; 598 } 599 if (r_1->is_stack()) { 600 // Convert stack slot to an SP offset (+ wordSize to account for return address ) 601 int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size; 602 if (!r_2->is_valid()) { 603 // sign extend??? 604 __ ldrsw(rscratch2, Address(esp, ld_off)); 605 __ str(rscratch2, Address(sp, st_off)); 606 } else { 607 // 608 // We are using two optoregs. This can be either T_OBJECT, 609 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates 610 // two slots but only uses one for thr T_LONG or T_DOUBLE case 611 // So we must adjust where to pick up the data to match the 612 // interpreter. 613 // 614 // Interpreter local[n] == MSW, local[n+1] == LSW however locals 615 // are accessed as negative so LSW is at LOW address 616 617 // ld_off is MSW so get LSW 618 const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)? 619 next_off : ld_off; 620 __ ldr(rscratch2, Address(esp, offset)); 621 // st_off is LSW (i.e. reg.first()) 622 __ str(rscratch2, Address(sp, st_off)); 623 } 624 } else if (r_1->is_Register()) { // Register argument 625 Register r = r_1->as_Register(); 626 if (r_2->is_valid()) { 627 // 628 // We are using two VMRegs. This can be either T_OBJECT, 629 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates 630 // two slots but only uses one for thr T_LONG or T_DOUBLE case 631 // So we must adjust where to pick up the data to match the 632 // interpreter. 633 634 const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)? 635 next_off : ld_off; 636 637 // this can be a misaligned move 638 __ ldr(r, Address(esp, offset)); 639 } else { 640 // sign extend and use a full word? 641 __ ldrw(r, Address(esp, ld_off)); 642 } 643 } else { 644 if (!r_2->is_valid()) { 645 __ ldrs(r_1->as_FloatRegister(), Address(esp, ld_off)); 646 } else { 647 __ ldrd(r_1->as_FloatRegister(), Address(esp, next_off)); 648 } 649 } 650 } 651 652 // 6243940 We might end up in handle_wrong_method if 653 // the callee is deoptimized as we race thru here. If that 654 // happens we don't want to take a safepoint because the 655 // caller frame will look interpreted and arguments are now 656 // "compiled" so it is much better to make this transition 657 // invisible to the stack walking code. Unfortunately if 658 // we try and find the callee by normal means a safepoint 659 // is possible. So we stash the desired callee in the thread 660 // and the vm will find there should this case occur. 661 662 __ str(rmethod, Address(rthread, JavaThread::callee_target_offset())); 663 664 __ br(rscratch1); 665 } 666 667 // --------------------------------------------------------------- 668 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, 669 int total_args_passed, 670 int comp_args_on_stack, 671 const BasicType *sig_bt, 672 const VMRegPair *regs, 673 AdapterFingerPrint* fingerprint) { 674 address i2c_entry = __ pc(); 675 676 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs); 677 678 address c2i_unverified_entry = __ pc(); 679 Label skip_fixup; 680 681 Label ok; 682 683 Register holder = rscratch2; 684 Register receiver = j_rarg0; 685 Register tmp = r10; // A call-clobbered register not used for arg passing 686 687 // ------------------------------------------------------------------------- 688 // Generate a C2I adapter. On entry we know rmethod holds the Method* during calls 689 // to the interpreter. The args start out packed in the compiled layout. They 690 // need to be unpacked into the interpreter layout. This will almost always 691 // require some stack space. We grow the current (compiled) stack, then repack 692 // the args. We finally end in a jump to the generic interpreter entry point. 693 // On exit from the interpreter, the interpreter will restore our SP (lest the 694 // compiled code, which relys solely on SP and not FP, get sick). 695 696 { 697 __ block_comment("c2i_unverified_entry {"); 698 __ load_klass(rscratch1, receiver); 699 __ ldr(tmp, Address(holder, CompiledICHolder::holder_klass_offset())); 700 __ cmp(rscratch1, tmp); 701 __ ldr(rmethod, Address(holder, CompiledICHolder::holder_metadata_offset())); 702 __ br(Assembler::EQ, ok); 703 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 704 705 __ bind(ok); 706 // Method might have been compiled since the call site was patched to 707 // interpreted; if that is the case treat it as a miss so we can get 708 // the call site corrected. 709 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset()))); 710 __ cbz(rscratch1, skip_fixup); 711 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 712 __ block_comment("} c2i_unverified_entry"); 713 } 714 715 address c2i_entry = __ pc(); 716 717 // Class initialization barrier for static methods 718 address c2i_no_clinit_check_entry = NULL; 719 if (VM_Version::supports_fast_class_init_checks()) { 720 Label L_skip_barrier; 721 722 { // Bypass the barrier for non-static methods 723 __ ldrw(rscratch1, Address(rmethod, Method::access_flags_offset())); 724 __ andsw(zr, rscratch1, JVM_ACC_STATIC); 725 __ br(Assembler::EQ, L_skip_barrier); // non-static 726 } 727 728 __ load_method_holder(rscratch2, rmethod); 729 __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier); 730 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub())); 731 732 __ bind(L_skip_barrier); 733 c2i_no_clinit_check_entry = __ pc(); 734 } 735 736 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 737 bs->c2i_entry_barrier(masm); 738 739 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup); 740 741 __ flush(); 742 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry, c2i_no_clinit_check_entry); 743 } 744 745 int SharedRuntime::c_calling_convention(const BasicType *sig_bt, 746 VMRegPair *regs, 747 VMRegPair *regs2, 748 int total_args_passed) { 749 assert(regs2 == NULL, "not needed on AArch64"); 750 751 // We return the amount of VMRegImpl stack slots we need to reserve for all 752 // the arguments NOT counting out_preserve_stack_slots. 753 754 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = { 755 c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5, c_rarg6, c_rarg7 756 }; 757 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = { 758 c_farg0, c_farg1, c_farg2, c_farg3, 759 c_farg4, c_farg5, c_farg6, c_farg7 760 }; 761 762 uint int_args = 0; 763 uint fp_args = 0; 764 uint stk_args = 0; // inc by 2 each time 765 766 for (int i = 0; i < total_args_passed; i++) { 767 switch (sig_bt[i]) { 768 case T_BOOLEAN: 769 case T_CHAR: 770 case T_BYTE: 771 case T_SHORT: 772 case T_INT: 773 if (int_args < Argument::n_int_register_parameters_c) { 774 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 775 } else { 776 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 777 stk_args += 2; 778 } 779 break; 780 case T_LONG: 781 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 782 // fall through 783 case T_OBJECT: 784 case T_ARRAY: 785 case T_ADDRESS: 786 case T_METADATA: 787 if (int_args < Argument::n_int_register_parameters_c) { 788 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg()); 789 } else { 790 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 791 stk_args += 2; 792 } 793 break; 794 case T_FLOAT: 795 if (fp_args < Argument::n_float_register_parameters_c) { 796 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg()); 797 } else { 798 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 799 stk_args += 2; 800 } 801 break; 802 case T_DOUBLE: 803 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 804 if (fp_args < Argument::n_float_register_parameters_c) { 805 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg()); 806 } else { 807 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 808 stk_args += 2; 809 } 810 break; 811 case T_VOID: // Halves of longs and doubles 812 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 813 regs[i].set_bad(); 814 break; 815 default: 816 ShouldNotReachHere(); 817 break; 818 } 819 } 820 821 return stk_args; 822 } 823 824 int SharedRuntime::vector_calling_convention(VMRegPair *regs, 825 uint num_bits, 826 uint total_args_passed) { 827 assert(!Matcher::supports_vector_calling_convention(), "not implemented"); 828 Unimplemented(); 829 return 0; 830 } 831 832 // On 64 bit we will store integer like items to the stack as 833 // 64 bits items (sparc abi) even though java would only store 834 // 32bits for a parameter. On 32bit it will simply be 32 bits 835 // So this routine will do 32->32 on 32bit and 32->64 on 64bit 836 static void move32_64(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 837 if (src.first()->is_stack()) { 838 if (dst.first()->is_stack()) { 839 // stack to stack 840 __ ldr(rscratch1, Address(rfp, reg2offset_in(src.first()))); 841 __ str(rscratch1, Address(sp, reg2offset_out(dst.first()))); 842 } else { 843 // stack to reg 844 __ ldrsw(dst.first()->as_Register(), Address(rfp, reg2offset_in(src.first()))); 845 } 846 } else if (dst.first()->is_stack()) { 847 // reg to stack 848 // Do we really have to sign extend??? 849 // __ movslq(src.first()->as_Register(), src.first()->as_Register()); 850 __ str(src.first()->as_Register(), Address(sp, reg2offset_out(dst.first()))); 851 } else { 852 if (dst.first() != src.first()) { 853 __ sxtw(dst.first()->as_Register(), src.first()->as_Register()); 854 } 855 } 856 } 857 858 // An oop arg. Must pass a handle not the oop itself 859 static void object_move(MacroAssembler* masm, 860 OopMap* map, 861 int oop_handle_offset, 862 int framesize_in_slots, 863 VMRegPair src, 864 VMRegPair dst, 865 bool is_receiver, 866 int* receiver_offset) { 867 868 // must pass a handle. First figure out the location we use as a handle 869 870 Register rHandle = dst.first()->is_stack() ? rscratch2 : dst.first()->as_Register(); 871 872 // See if oop is NULL if it is we need no handle 873 874 if (src.first()->is_stack()) { 875 876 // Oop is already on the stack as an argument 877 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 878 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots)); 879 if (is_receiver) { 880 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size; 881 } 882 883 __ ldr(rscratch1, Address(rfp, reg2offset_in(src.first()))); 884 __ lea(rHandle, Address(rfp, reg2offset_in(src.first()))); 885 // conditionally move a NULL 886 __ cmp(rscratch1, zr); 887 __ csel(rHandle, zr, rHandle, Assembler::EQ); 888 } else { 889 890 // Oop is in an a register we must store it to the space we reserve 891 // on the stack for oop_handles and pass a handle if oop is non-NULL 892 893 const Register rOop = src.first()->as_Register(); 894 int oop_slot; 895 if (rOop == j_rarg0) 896 oop_slot = 0; 897 else if (rOop == j_rarg1) 898 oop_slot = 1; 899 else if (rOop == j_rarg2) 900 oop_slot = 2; 901 else if (rOop == j_rarg3) 902 oop_slot = 3; 903 else if (rOop == j_rarg4) 904 oop_slot = 4; 905 else if (rOop == j_rarg5) 906 oop_slot = 5; 907 else if (rOop == j_rarg6) 908 oop_slot = 6; 909 else { 910 assert(rOop == j_rarg7, "wrong register"); 911 oop_slot = 7; 912 } 913 914 oop_slot = oop_slot * VMRegImpl::slots_per_word + oop_handle_offset; 915 int offset = oop_slot*VMRegImpl::stack_slot_size; 916 917 map->set_oop(VMRegImpl::stack2reg(oop_slot)); 918 // Store oop in handle area, may be NULL 919 __ str(rOop, Address(sp, offset)); 920 if (is_receiver) { 921 *receiver_offset = offset; 922 } 923 924 __ cmp(rOop, zr); 925 __ lea(rHandle, Address(sp, offset)); 926 // conditionally move a NULL 927 __ csel(rHandle, zr, rHandle, Assembler::EQ); 928 } 929 930 // If arg is on the stack then place it otherwise it is already in correct reg. 931 if (dst.first()->is_stack()) { 932 __ str(rHandle, Address(sp, reg2offset_out(dst.first()))); 933 } 934 } 935 936 // A float arg may have to do float reg int reg conversion 937 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 938 assert(src.first()->is_stack() && dst.first()->is_stack() || 939 src.first()->is_reg() && dst.first()->is_reg(), "Unexpected error"); 940 if (src.first()->is_stack()) { 941 if (dst.first()->is_stack()) { 942 __ ldrw(rscratch1, Address(rfp, reg2offset_in(src.first()))); 943 __ strw(rscratch1, Address(sp, reg2offset_out(dst.first()))); 944 } else { 945 ShouldNotReachHere(); 946 } 947 } else if (src.first() != dst.first()) { 948 if (src.is_single_phys_reg() && dst.is_single_phys_reg()) 949 __ fmovs(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 950 else 951 ShouldNotReachHere(); 952 } 953 } 954 955 // A long move 956 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 957 if (src.first()->is_stack()) { 958 if (dst.first()->is_stack()) { 959 // stack to stack 960 __ ldr(rscratch1, Address(rfp, reg2offset_in(src.first()))); 961 __ str(rscratch1, Address(sp, reg2offset_out(dst.first()))); 962 } else { 963 // stack to reg 964 __ ldr(dst.first()->as_Register(), Address(rfp, reg2offset_in(src.first()))); 965 } 966 } else if (dst.first()->is_stack()) { 967 // reg to stack 968 // Do we really have to sign extend??? 969 // __ movslq(src.first()->as_Register(), src.first()->as_Register()); 970 __ str(src.first()->as_Register(), Address(sp, reg2offset_out(dst.first()))); 971 } else { 972 if (dst.first() != src.first()) { 973 __ mov(dst.first()->as_Register(), src.first()->as_Register()); 974 } 975 } 976 } 977 978 979 // A double move 980 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 981 assert(src.first()->is_stack() && dst.first()->is_stack() || 982 src.first()->is_reg() && dst.first()->is_reg(), "Unexpected error"); 983 if (src.first()->is_stack()) { 984 if (dst.first()->is_stack()) { 985 __ ldr(rscratch1, Address(rfp, reg2offset_in(src.first()))); 986 __ str(rscratch1, Address(sp, reg2offset_out(dst.first()))); 987 } else { 988 ShouldNotReachHere(); 989 } 990 } else if (src.first() != dst.first()) { 991 if (src.is_single_phys_reg() && dst.is_single_phys_reg()) 992 __ fmovd(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 993 else 994 ShouldNotReachHere(); 995 } 996 } 997 998 999 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1000 // We always ignore the frame_slots arg and just use the space just below frame pointer 1001 // which by this time is free to use 1002 switch (ret_type) { 1003 case T_FLOAT: 1004 __ strs(v0, Address(rfp, -wordSize)); 1005 break; 1006 case T_DOUBLE: 1007 __ strd(v0, Address(rfp, -wordSize)); 1008 break; 1009 case T_VOID: break; 1010 default: { 1011 __ str(r0, Address(rfp, -wordSize)); 1012 } 1013 } 1014 } 1015 1016 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1017 // We always ignore the frame_slots arg and just use the space just below frame pointer 1018 // which by this time is free to use 1019 switch (ret_type) { 1020 case T_FLOAT: 1021 __ ldrs(v0, Address(rfp, -wordSize)); 1022 break; 1023 case T_DOUBLE: 1024 __ ldrd(v0, Address(rfp, -wordSize)); 1025 break; 1026 case T_VOID: break; 1027 default: { 1028 __ ldr(r0, Address(rfp, -wordSize)); 1029 } 1030 } 1031 } 1032 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) { 1033 RegSet x; 1034 for ( int i = first_arg ; i < arg_count ; i++ ) { 1035 if (args[i].first()->is_Register()) { 1036 x = x + args[i].first()->as_Register(); 1037 } else if (args[i].first()->is_FloatRegister()) { 1038 __ strd(args[i].first()->as_FloatRegister(), Address(__ pre(sp, -2 * wordSize))); 1039 } 1040 } 1041 __ push(x, sp); 1042 } 1043 1044 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) { 1045 RegSet x; 1046 for ( int i = first_arg ; i < arg_count ; i++ ) { 1047 if (args[i].first()->is_Register()) { 1048 x = x + args[i].first()->as_Register(); 1049 } else { 1050 ; 1051 } 1052 } 1053 __ pop(x, sp); 1054 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) { 1055 if (args[i].first()->is_Register()) { 1056 ; 1057 } else if (args[i].first()->is_FloatRegister()) { 1058 __ ldrd(args[i].first()->as_FloatRegister(), Address(__ post(sp, 2 * wordSize))); 1059 } 1060 } 1061 } 1062 1063 1064 // Check GCLocker::needs_gc and enter the runtime if it's true. This 1065 // keeps a new JNI critical region from starting until a GC has been 1066 // forced. Save down any oops in registers and describe them in an 1067 // OopMap. 1068 static void check_needs_gc_for_critical_native(MacroAssembler* masm, 1069 int stack_slots, 1070 int total_c_args, 1071 int total_in_args, 1072 int arg_save_area, 1073 OopMapSet* oop_maps, 1074 VMRegPair* in_regs, 1075 BasicType* in_sig_bt) { Unimplemented(); } 1076 1077 // Unpack an array argument into a pointer to the body and the length 1078 // if the array is non-null, otherwise pass 0 for both. 1079 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, VMRegPair body_arg, VMRegPair length_arg) { Unimplemented(); } 1080 1081 1082 class ComputeMoveOrder: public StackObj { 1083 class MoveOperation: public ResourceObj { 1084 friend class ComputeMoveOrder; 1085 private: 1086 VMRegPair _src; 1087 VMRegPair _dst; 1088 int _src_index; 1089 int _dst_index; 1090 bool _processed; 1091 MoveOperation* _next; 1092 MoveOperation* _prev; 1093 1094 static int get_id(VMRegPair r) { Unimplemented(); return 0; } 1095 1096 public: 1097 MoveOperation(int src_index, VMRegPair src, int dst_index, VMRegPair dst): 1098 _src(src) 1099 , _dst(dst) 1100 , _src_index(src_index) 1101 , _dst_index(dst_index) 1102 , _processed(false) 1103 , _next(NULL) 1104 , _prev(NULL) { Unimplemented(); } 1105 1106 VMRegPair src() const { Unimplemented(); return _src; } 1107 int src_id() const { Unimplemented(); return 0; } 1108 int src_index() const { Unimplemented(); return 0; } 1109 VMRegPair dst() const { Unimplemented(); return _src; } 1110 void set_dst(int i, VMRegPair dst) { Unimplemented(); } 1111 int dst_index() const { Unimplemented(); return 0; } 1112 int dst_id() const { Unimplemented(); return 0; } 1113 MoveOperation* next() const { Unimplemented(); return 0; } 1114 MoveOperation* prev() const { Unimplemented(); return 0; } 1115 void set_processed() { Unimplemented(); } 1116 bool is_processed() const { Unimplemented(); return 0; } 1117 1118 // insert 1119 void break_cycle(VMRegPair temp_register) { Unimplemented(); } 1120 1121 void link(GrowableArray<MoveOperation*>& killer) { Unimplemented(); } 1122 }; 1123 1124 private: 1125 GrowableArray<MoveOperation*> edges; 1126 1127 public: 1128 ComputeMoveOrder(int total_in_args, VMRegPair* in_regs, int total_c_args, VMRegPair* out_regs, 1129 BasicType* in_sig_bt, GrowableArray<int>& arg_order, VMRegPair tmp_vmreg) { Unimplemented(); } 1130 1131 // Collected all the move operations 1132 void add_edge(int src_index, VMRegPair src, int dst_index, VMRegPair dst) { Unimplemented(); } 1133 1134 // Walk the edges breaking cycles between moves. The result list 1135 // can be walked in order to produce the proper set of loads 1136 GrowableArray<MoveOperation*>* get_store_order(VMRegPair temp_register) { Unimplemented(); return 0; } 1137 }; 1138 1139 1140 static void rt_call(MacroAssembler* masm, address dest, int gpargs, int fpargs, int type) { 1141 CodeBlob *cb = CodeCache::find_blob(dest); 1142 if (cb) { 1143 __ far_call(RuntimeAddress(dest)); 1144 } else { 1145 assert((unsigned)gpargs < 256, "eek!"); 1146 assert((unsigned)fpargs < 32, "eek!"); 1147 __ lea(rscratch1, RuntimeAddress(dest)); 1148 __ blr(rscratch1); 1149 __ maybe_isb(); 1150 } 1151 } 1152 1153 static void verify_oop_args(MacroAssembler* masm, 1154 const methodHandle& method, 1155 const BasicType* sig_bt, 1156 const VMRegPair* regs) { 1157 Register temp_reg = r19; // not part of any compiled calling seq 1158 if (VerifyOops) { 1159 for (int i = 0; i < method->size_of_parameters(); i++) { 1160 if (sig_bt[i] == T_OBJECT || 1161 sig_bt[i] == T_ARRAY) { 1162 VMReg r = regs[i].first(); 1163 assert(r->is_valid(), "bad oop arg"); 1164 if (r->is_stack()) { 1165 __ ldr(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size)); 1166 __ verify_oop(temp_reg); 1167 } else { 1168 __ verify_oop(r->as_Register()); 1169 } 1170 } 1171 } 1172 } 1173 } 1174 1175 static void gen_special_dispatch(MacroAssembler* masm, 1176 const methodHandle& method, 1177 const BasicType* sig_bt, 1178 const VMRegPair* regs) { 1179 verify_oop_args(masm, method, sig_bt, regs); 1180 vmIntrinsics::ID iid = method->intrinsic_id(); 1181 1182 // Now write the args into the outgoing interpreter space 1183 bool has_receiver = false; 1184 Register receiver_reg = noreg; 1185 int member_arg_pos = -1; 1186 Register member_reg = noreg; 1187 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid); 1188 if (ref_kind != 0) { 1189 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument 1190 member_reg = r19; // known to be free at this point 1191 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind); 1192 } else if (iid == vmIntrinsics::_invokeBasic) { 1193 has_receiver = true; 1194 } else { 1195 fatal("unexpected intrinsic id %d", iid); 1196 } 1197 1198 if (member_reg != noreg) { 1199 // Load the member_arg into register, if necessary. 1200 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs); 1201 VMReg r = regs[member_arg_pos].first(); 1202 if (r->is_stack()) { 1203 __ ldr(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size)); 1204 } else { 1205 // no data motion is needed 1206 member_reg = r->as_Register(); 1207 } 1208 } 1209 1210 if (has_receiver) { 1211 // Make sure the receiver is loaded into a register. 1212 assert(method->size_of_parameters() > 0, "oob"); 1213 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object"); 1214 VMReg r = regs[0].first(); 1215 assert(r->is_valid(), "bad receiver arg"); 1216 if (r->is_stack()) { 1217 // Porting note: This assumes that compiled calling conventions always 1218 // pass the receiver oop in a register. If this is not true on some 1219 // platform, pick a temp and load the receiver from stack. 1220 fatal("receiver always in a register"); 1221 receiver_reg = r2; // known to be free at this point 1222 __ ldr(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size)); 1223 } else { 1224 // no data motion is needed 1225 receiver_reg = r->as_Register(); 1226 } 1227 } 1228 1229 // Figure out which address we are really jumping to: 1230 MethodHandles::generate_method_handle_dispatch(masm, iid, 1231 receiver_reg, member_reg, /*for_compiler_entry:*/ true); 1232 } 1233 1234 // --------------------------------------------------------------------------- 1235 // Generate a native wrapper for a given method. The method takes arguments 1236 // in the Java compiled code convention, marshals them to the native 1237 // convention (handlizes oops, etc), transitions to native, makes the call, 1238 // returns to java state (possibly blocking), unhandlizes any result and 1239 // returns. 1240 // 1241 // Critical native functions are a shorthand for the use of 1242 // GetPrimtiveArrayCritical and disallow the use of any other JNI 1243 // functions. The wrapper is expected to unpack the arguments before 1244 // passing them to the callee and perform checks before and after the 1245 // native call to ensure that they GCLocker 1246 // lock_critical/unlock_critical semantics are followed. Some other 1247 // parts of JNI setup are skipped like the tear down of the JNI handle 1248 // block and the check for pending exceptions it's impossible for them 1249 // to be thrown. 1250 // 1251 // They are roughly structured like this: 1252 // if (GCLocker::needs_gc()) 1253 // SharedRuntime::block_for_jni_critical(); 1254 // tranistion to thread_in_native 1255 // unpack arrray arguments and call native entry point 1256 // check for safepoint in progress 1257 // check if any thread suspend flags are set 1258 // call into JVM and possible unlock the JNI critical 1259 // if a GC was suppressed while in the critical native. 1260 // transition back to thread_in_Java 1261 // return to caller 1262 // 1263 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm, 1264 const methodHandle& method, 1265 int compile_id, 1266 BasicType* in_sig_bt, 1267 VMRegPair* in_regs, 1268 BasicType ret_type, 1269 address critical_entry) { 1270 if (method->is_method_handle_intrinsic()) { 1271 vmIntrinsics::ID iid = method->intrinsic_id(); 1272 intptr_t start = (intptr_t)__ pc(); 1273 int vep_offset = ((intptr_t)__ pc()) - start; 1274 1275 // First instruction must be a nop as it may need to be patched on deoptimisation 1276 __ nop(); 1277 gen_special_dispatch(masm, 1278 method, 1279 in_sig_bt, 1280 in_regs); 1281 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period 1282 __ flush(); 1283 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually 1284 return nmethod::new_native_nmethod(method, 1285 compile_id, 1286 masm->code(), 1287 vep_offset, 1288 frame_complete, 1289 stack_slots / VMRegImpl::slots_per_word, 1290 in_ByteSize(-1), 1291 in_ByteSize(-1), 1292 (OopMapSet*)NULL); 1293 } 1294 bool is_critical_native = true; 1295 address native_func = critical_entry; 1296 if (native_func == NULL) { 1297 native_func = method->native_function(); 1298 is_critical_native = false; 1299 } 1300 assert(native_func != NULL, "must have function"); 1301 1302 // An OopMap for lock (and class if static) 1303 OopMapSet *oop_maps = new OopMapSet(); 1304 intptr_t start = (intptr_t)__ pc(); 1305 1306 // We have received a description of where all the java arg are located 1307 // on entry to the wrapper. We need to convert these args to where 1308 // the jni function will expect them. To figure out where they go 1309 // we convert the java signature to a C signature by inserting 1310 // the hidden arguments as arg[0] and possibly arg[1] (static method) 1311 1312 const int total_in_args = method->size_of_parameters(); 1313 int total_c_args = total_in_args; 1314 if (!is_critical_native) { 1315 total_c_args += 1; 1316 if (method->is_static()) { 1317 total_c_args++; 1318 } 1319 } else { 1320 for (int i = 0; i < total_in_args; i++) { 1321 if (in_sig_bt[i] == T_ARRAY) { 1322 total_c_args++; 1323 } 1324 } 1325 } 1326 1327 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1328 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1329 BasicType* in_elem_bt = NULL; 1330 1331 int argc = 0; 1332 if (!is_critical_native) { 1333 out_sig_bt[argc++] = T_ADDRESS; 1334 if (method->is_static()) { 1335 out_sig_bt[argc++] = T_OBJECT; 1336 } 1337 1338 for (int i = 0; i < total_in_args ; i++ ) { 1339 out_sig_bt[argc++] = in_sig_bt[i]; 1340 } 1341 } else { 1342 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, total_in_args); 1343 SignatureStream ss(method->signature()); 1344 for (int i = 0; i < total_in_args ; i++ ) { 1345 if (in_sig_bt[i] == T_ARRAY) { 1346 // Arrays are passed as int, elem* pair 1347 out_sig_bt[argc++] = T_INT; 1348 out_sig_bt[argc++] = T_ADDRESS; 1349 ss.skip_array_prefix(1); // skip one '[' 1350 assert(ss.is_primitive(), "primitive type expected"); 1351 in_elem_bt[i] = ss.type(); 1352 } else { 1353 out_sig_bt[argc++] = in_sig_bt[i]; 1354 in_elem_bt[i] = T_VOID; 1355 } 1356 if (in_sig_bt[i] != T_VOID) { 1357 assert(in_sig_bt[i] == ss.type() || 1358 in_sig_bt[i] == T_ARRAY, "must match"); 1359 ss.next(); 1360 } 1361 } 1362 } 1363 1364 // Now figure out where the args must be stored and how much stack space 1365 // they require. 1366 int out_arg_slots; 1367 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, NULL, total_c_args); 1368 1369 // Compute framesize for the wrapper. We need to handlize all oops in 1370 // incoming registers 1371 1372 // Calculate the total number of stack slots we will need. 1373 1374 // First count the abi requirement plus all of the outgoing args 1375 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots; 1376 1377 // Now the space for the inbound oop handle area 1378 int total_save_slots = 8 * VMRegImpl::slots_per_word; // 8 arguments passed in registers 1379 if (is_critical_native) { 1380 // Critical natives may have to call out so they need a save area 1381 // for register arguments. 1382 int double_slots = 0; 1383 int single_slots = 0; 1384 for ( int i = 0; i < total_in_args; i++) { 1385 if (in_regs[i].first()->is_Register()) { 1386 const Register reg = in_regs[i].first()->as_Register(); 1387 switch (in_sig_bt[i]) { 1388 case T_BOOLEAN: 1389 case T_BYTE: 1390 case T_SHORT: 1391 case T_CHAR: 1392 case T_INT: single_slots++; break; 1393 case T_ARRAY: // specific to LP64 (7145024) 1394 case T_LONG: double_slots++; break; 1395 default: ShouldNotReachHere(); 1396 } 1397 } else if (in_regs[i].first()->is_FloatRegister()) { 1398 ShouldNotReachHere(); 1399 } 1400 } 1401 total_save_slots = double_slots * 2 + single_slots; 1402 // align the save area 1403 if (double_slots != 0) { 1404 stack_slots = align_up(stack_slots, 2); 1405 } 1406 } 1407 1408 int oop_handle_offset = stack_slots; 1409 stack_slots += total_save_slots; 1410 1411 // Now any space we need for handlizing a klass if static method 1412 1413 int klass_slot_offset = 0; 1414 int klass_offset = -1; 1415 int lock_slot_offset = 0; 1416 bool is_static = false; 1417 1418 if (method->is_static()) { 1419 klass_slot_offset = stack_slots; 1420 stack_slots += VMRegImpl::slots_per_word; 1421 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; 1422 is_static = true; 1423 } 1424 1425 // Plus a lock if needed 1426 1427 if (method->is_synchronized()) { 1428 lock_slot_offset = stack_slots; 1429 stack_slots += VMRegImpl::slots_per_word; 1430 } 1431 1432 // Now a place (+2) to save return values or temp during shuffling 1433 // + 4 for return address (which we own) and saved rfp 1434 stack_slots += 6; 1435 1436 // Ok The space we have allocated will look like: 1437 // 1438 // 1439 // FP-> | | 1440 // |---------------------| 1441 // | 2 slots for moves | 1442 // |---------------------| 1443 // | lock box (if sync) | 1444 // |---------------------| <- lock_slot_offset 1445 // | klass (if static) | 1446 // |---------------------| <- klass_slot_offset 1447 // | oopHandle area | 1448 // |---------------------| <- oop_handle_offset (8 java arg registers) 1449 // | outbound memory | 1450 // | based arguments | 1451 // | | 1452 // |---------------------| 1453 // | | 1454 // SP-> | out_preserved_slots | 1455 // 1456 // 1457 1458 1459 // Now compute actual number of stack words we need rounding to make 1460 // stack properly aligned. 1461 stack_slots = align_up(stack_slots, StackAlignmentInSlots); 1462 1463 int stack_size = stack_slots * VMRegImpl::stack_slot_size; 1464 1465 // First thing make an ic check to see if we should even be here 1466 1467 // We are free to use all registers as temps without saving them and 1468 // restoring them except rfp. rfp is the only callee save register 1469 // as far as the interpreter and the compiler(s) are concerned. 1470 1471 1472 const Register ic_reg = rscratch2; 1473 const Register receiver = j_rarg0; 1474 1475 Label hit; 1476 Label exception_pending; 1477 1478 assert_different_registers(ic_reg, receiver, rscratch1); 1479 __ verify_oop(receiver); 1480 __ cmp_klass(receiver, ic_reg, rscratch1); 1481 __ br(Assembler::EQ, hit); 1482 1483 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 1484 1485 // Verified entry point must be aligned 1486 __ align(8); 1487 1488 __ bind(hit); 1489 1490 int vep_offset = ((intptr_t)__ pc()) - start; 1491 1492 // If we have to make this method not-entrant we'll overwrite its 1493 // first instruction with a jump. For this action to be legal we 1494 // must ensure that this first instruction is a B, BL, NOP, BKPT, 1495 // SVC, HVC, or SMC. Make it a NOP. 1496 __ nop(); 1497 1498 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) { 1499 Label L_skip_barrier; 1500 __ mov_metadata(rscratch2, method->method_holder()); // InstanceKlass* 1501 __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier); 1502 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub())); 1503 1504 __ bind(L_skip_barrier); 1505 } 1506 1507 // Generate stack overflow check 1508 if (UseStackBanging) { 1509 __ bang_stack_with_offset(JavaThread::stack_shadow_zone_size()); 1510 } else { 1511 Unimplemented(); 1512 } 1513 1514 // Generate a new frame for the wrapper. 1515 __ enter(); 1516 // -2 because return address is already present and so is saved rfp 1517 __ sub(sp, sp, stack_size - 2*wordSize); 1518 1519 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler(); 1520 bs->nmethod_entry_barrier(masm); 1521 1522 // Frame is now completed as far as size and linkage. 1523 int frame_complete = ((intptr_t)__ pc()) - start; 1524 1525 // We use r20 as the oop handle for the receiver/klass 1526 // It is callee save so it survives the call to native 1527 1528 const Register oop_handle_reg = r20; 1529 1530 if (is_critical_native) { 1531 check_needs_gc_for_critical_native(masm, stack_slots, total_c_args, total_in_args, 1532 oop_handle_offset, oop_maps, in_regs, in_sig_bt); 1533 } 1534 1535 // 1536 // We immediately shuffle the arguments so that any vm call we have to 1537 // make from here on out (sync slow path, jvmti, etc.) we will have 1538 // captured the oops from our caller and have a valid oopMap for 1539 // them. 1540 1541 // ----------------- 1542 // The Grand Shuffle 1543 1544 // The Java calling convention is either equal (linux) or denser (win64) than the 1545 // c calling convention. However the because of the jni_env argument the c calling 1546 // convention always has at least one more (and two for static) arguments than Java. 1547 // Therefore if we move the args from java -> c backwards then we will never have 1548 // a register->register conflict and we don't have to build a dependency graph 1549 // and figure out how to break any cycles. 1550 // 1551 1552 // Record esp-based slot for receiver on stack for non-static methods 1553 int receiver_offset = -1; 1554 1555 // This is a trick. We double the stack slots so we can claim 1556 // the oops in the caller's frame. Since we are sure to have 1557 // more args than the caller doubling is enough to make 1558 // sure we can capture all the incoming oop args from the 1559 // caller. 1560 // 1561 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1562 1563 // Mark location of rfp (someday) 1564 // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, vmreg(rfp)); 1565 1566 1567 int float_args = 0; 1568 int int_args = 0; 1569 1570 #ifdef ASSERT 1571 bool reg_destroyed[RegisterImpl::number_of_registers]; 1572 bool freg_destroyed[FloatRegisterImpl::number_of_registers]; 1573 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) { 1574 reg_destroyed[r] = false; 1575 } 1576 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) { 1577 freg_destroyed[f] = false; 1578 } 1579 1580 #endif /* ASSERT */ 1581 1582 // This may iterate in two different directions depending on the 1583 // kind of native it is. The reason is that for regular JNI natives 1584 // the incoming and outgoing registers are offset upwards and for 1585 // critical natives they are offset down. 1586 GrowableArray<int> arg_order(2 * total_in_args); 1587 VMRegPair tmp_vmreg; 1588 tmp_vmreg.set2(r19->as_VMReg()); 1589 1590 if (!is_critical_native) { 1591 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) { 1592 arg_order.push(i); 1593 arg_order.push(c_arg); 1594 } 1595 } else { 1596 // Compute a valid move order, using tmp_vmreg to break any cycles 1597 ComputeMoveOrder cmo(total_in_args, in_regs, total_c_args, out_regs, in_sig_bt, arg_order, tmp_vmreg); 1598 } 1599 1600 int temploc = -1; 1601 for (int ai = 0; ai < arg_order.length(); ai += 2) { 1602 int i = arg_order.at(ai); 1603 int c_arg = arg_order.at(ai + 1); 1604 __ block_comment(err_msg("move %d -> %d", i, c_arg)); 1605 if (c_arg == -1) { 1606 assert(is_critical_native, "should only be required for critical natives"); 1607 // This arg needs to be moved to a temporary 1608 __ mov(tmp_vmreg.first()->as_Register(), in_regs[i].first()->as_Register()); 1609 in_regs[i] = tmp_vmreg; 1610 temploc = i; 1611 continue; 1612 } else if (i == -1) { 1613 assert(is_critical_native, "should only be required for critical natives"); 1614 // Read from the temporary location 1615 assert(temploc != -1, "must be valid"); 1616 i = temploc; 1617 temploc = -1; 1618 } 1619 #ifdef ASSERT 1620 if (in_regs[i].first()->is_Register()) { 1621 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!"); 1622 } else if (in_regs[i].first()->is_FloatRegister()) { 1623 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!"); 1624 } 1625 if (out_regs[c_arg].first()->is_Register()) { 1626 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true; 1627 } else if (out_regs[c_arg].first()->is_FloatRegister()) { 1628 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true; 1629 } 1630 #endif /* ASSERT */ 1631 switch (in_sig_bt[i]) { 1632 case T_ARRAY: 1633 if (is_critical_native) { 1634 unpack_array_argument(masm, in_regs[i], in_elem_bt[i], out_regs[c_arg + 1], out_regs[c_arg]); 1635 c_arg++; 1636 #ifdef ASSERT 1637 if (out_regs[c_arg].first()->is_Register()) { 1638 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true; 1639 } else if (out_regs[c_arg].first()->is_FloatRegister()) { 1640 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true; 1641 } 1642 #endif 1643 int_args++; 1644 break; 1645 } 1646 case T_OBJECT: 1647 assert(!is_critical_native, "no oop arguments"); 1648 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg], 1649 ((i == 0) && (!is_static)), 1650 &receiver_offset); 1651 int_args++; 1652 break; 1653 case T_VOID: 1654 break; 1655 1656 case T_FLOAT: 1657 float_move(masm, in_regs[i], out_regs[c_arg]); 1658 float_args++; 1659 break; 1660 1661 case T_DOUBLE: 1662 assert( i + 1 < total_in_args && 1663 in_sig_bt[i + 1] == T_VOID && 1664 out_sig_bt[c_arg+1] == T_VOID, "bad arg list"); 1665 double_move(masm, in_regs[i], out_regs[c_arg]); 1666 float_args++; 1667 break; 1668 1669 case T_LONG : 1670 long_move(masm, in_regs[i], out_regs[c_arg]); 1671 int_args++; 1672 break; 1673 1674 case T_ADDRESS: assert(false, "found T_ADDRESS in java args"); 1675 1676 default: 1677 move32_64(masm, in_regs[i], out_regs[c_arg]); 1678 int_args++; 1679 } 1680 } 1681 1682 // point c_arg at the first arg that is already loaded in case we 1683 // need to spill before we call out 1684 int c_arg = total_c_args - total_in_args; 1685 1686 // Pre-load a static method's oop into c_rarg1. 1687 if (method->is_static() && !is_critical_native) { 1688 1689 // load oop into a register 1690 __ movoop(c_rarg1, 1691 JNIHandles::make_local(method->method_holder()->java_mirror()), 1692 /*immediate*/true); 1693 1694 // Now handlize the static class mirror it's known not-null. 1695 __ str(c_rarg1, Address(sp, klass_offset)); 1696 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); 1697 1698 // Now get the handle 1699 __ lea(c_rarg1, Address(sp, klass_offset)); 1700 // and protect the arg if we must spill 1701 c_arg--; 1702 } 1703 1704 // Change state to native (we save the return address in the thread, since it might not 1705 // be pushed on the stack when we do a stack traversal). 1706 // We use the same pc/oopMap repeatedly when we call out 1707 1708 Label native_return; 1709 __ set_last_Java_frame(sp, noreg, native_return, rscratch1); 1710 1711 Label dtrace_method_entry, dtrace_method_entry_done; 1712 { 1713 unsigned long offset; 1714 __ adrp(rscratch1, ExternalAddress((address)&DTraceMethodProbes), offset); 1715 __ ldrb(rscratch1, Address(rscratch1, offset)); 1716 __ cbnzw(rscratch1, dtrace_method_entry); 1717 __ bind(dtrace_method_entry_done); 1718 } 1719 1720 // RedefineClasses() tracing support for obsolete method entry 1721 if (log_is_enabled(Trace, redefine, class, obsolete)) { 1722 // protect the args we've loaded 1723 save_args(masm, total_c_args, c_arg, out_regs); 1724 __ mov_metadata(c_rarg1, method()); 1725 __ call_VM_leaf( 1726 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 1727 rthread, c_rarg1); 1728 restore_args(masm, total_c_args, c_arg, out_regs); 1729 } 1730 1731 // Lock a synchronized method 1732 1733 // Register definitions used by locking and unlocking 1734 1735 const Register swap_reg = r0; 1736 const Register obj_reg = r19; // Will contain the oop 1737 const Register lock_reg = r13; // Address of compiler lock object (BasicLock) 1738 const Register old_hdr = r13; // value of old header at unlock time 1739 const Register tmp = lr; 1740 1741 Label slow_path_lock; 1742 Label lock_done; 1743 1744 if (method->is_synchronized()) { 1745 assert(!is_critical_native, "unhandled"); 1746 1747 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes(); 1748 1749 // Get the handle (the 2nd argument) 1750 __ mov(oop_handle_reg, c_rarg1); 1751 1752 // Get address of the box 1753 1754 __ lea(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 1755 1756 // Load the oop from the handle 1757 __ ldr(obj_reg, Address(oop_handle_reg, 0)); 1758 1759 __ resolve(IS_NOT_NULL, obj_reg); 1760 1761 if (UseBiasedLocking) { 1762 __ biased_locking_enter(lock_reg, obj_reg, swap_reg, tmp, false, lock_done, &slow_path_lock); 1763 } 1764 1765 // Load (object->mark() | 1) into swap_reg %r0 1766 __ ldr(rscratch1, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 1767 __ orr(swap_reg, rscratch1, 1); 1768 1769 // Save (object->mark() | 1) into BasicLock's displaced header 1770 __ str(swap_reg, Address(lock_reg, mark_word_offset)); 1771 1772 // src -> dest iff dest == r0 else r0 <- dest 1773 { Label here; 1774 __ cmpxchg_obj_header(r0, lock_reg, obj_reg, rscratch1, lock_done, /*fallthrough*/NULL); 1775 } 1776 1777 // Hmm should this move to the slow path code area??? 1778 1779 // Test if the oopMark is an obvious stack pointer, i.e., 1780 // 1) (mark & 3) == 0, and 1781 // 2) sp <= mark < mark + os::pagesize() 1782 // These 3 tests can be done by evaluating the following 1783 // expression: ((mark - sp) & (3 - os::vm_page_size())), 1784 // assuming both stack pointer and pagesize have their 1785 // least significant 2 bits clear. 1786 // NOTE: the oopMark is in swap_reg %r0 as the result of cmpxchg 1787 1788 __ sub(swap_reg, sp, swap_reg); 1789 __ neg(swap_reg, swap_reg); 1790 __ ands(swap_reg, swap_reg, 3 - os::vm_page_size()); 1791 1792 // Save the test result, for recursive case, the result is zero 1793 __ str(swap_reg, Address(lock_reg, mark_word_offset)); 1794 __ br(Assembler::NE, slow_path_lock); 1795 1796 // Slow path will re-enter here 1797 1798 __ bind(lock_done); 1799 } 1800 1801 1802 // Finally just about ready to make the JNI call 1803 1804 // get JNIEnv* which is first argument to native 1805 if (!is_critical_native) { 1806 __ lea(c_rarg0, Address(rthread, in_bytes(JavaThread::jni_environment_offset()))); 1807 } 1808 1809 // Now set thread in native 1810 __ mov(rscratch1, _thread_in_native); 1811 __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset())); 1812 __ stlrw(rscratch1, rscratch2); 1813 1814 { 1815 int return_type = 0; 1816 switch (ret_type) { 1817 case T_VOID: break; 1818 return_type = 0; break; 1819 case T_CHAR: 1820 case T_BYTE: 1821 case T_SHORT: 1822 case T_INT: 1823 case T_BOOLEAN: 1824 case T_LONG: 1825 return_type = 1; break; 1826 case T_ARRAY: 1827 case T_OBJECT: 1828 return_type = 1; break; 1829 case T_FLOAT: 1830 return_type = 2; break; 1831 case T_DOUBLE: 1832 return_type = 3; break; 1833 default: 1834 ShouldNotReachHere(); 1835 } 1836 rt_call(masm, native_func, 1837 int_args + 2, // AArch64 passes up to 8 args in int registers 1838 float_args, // and up to 8 float args 1839 return_type); 1840 } 1841 1842 __ bind(native_return); 1843 1844 intptr_t return_pc = (intptr_t) __ pc(); 1845 oop_maps->add_gc_map(return_pc - start, map); 1846 1847 // Unpack native results. 1848 switch (ret_type) { 1849 case T_BOOLEAN: __ c2bool(r0); break; 1850 case T_CHAR : __ ubfx(r0, r0, 0, 16); break; 1851 case T_BYTE : __ sbfx(r0, r0, 0, 8); break; 1852 case T_SHORT : __ sbfx(r0, r0, 0, 16); break; 1853 case T_INT : __ sbfx(r0, r0, 0, 32); break; 1854 case T_DOUBLE : 1855 case T_FLOAT : 1856 // Result is in v0 we'll save as needed 1857 break; 1858 case T_ARRAY: // Really a handle 1859 case T_OBJECT: // Really a handle 1860 break; // can't de-handlize until after safepoint check 1861 case T_VOID: break; 1862 case T_LONG: break; 1863 default : ShouldNotReachHere(); 1864 } 1865 1866 // Switch thread to "native transition" state before reading the synchronization state. 1867 // This additional state is necessary because reading and testing the synchronization 1868 // state is not atomic w.r.t. GC, as this scenario demonstrates: 1869 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 1870 // VM thread changes sync state to synchronizing and suspends threads for GC. 1871 // Thread A is resumed to finish this native method, but doesn't block here since it 1872 // didn't see any synchronization is progress, and escapes. 1873 __ mov(rscratch1, _thread_in_native_trans); 1874 1875 __ strw(rscratch1, Address(rthread, JavaThread::thread_state_offset())); 1876 1877 // Force this write out before the read below 1878 __ dmb(Assembler::ISH); 1879 1880 // check for safepoint operation in progress and/or pending suspend requests 1881 Label safepoint_in_progress, safepoint_in_progress_done; 1882 { 1883 __ safepoint_poll_acquire(safepoint_in_progress); 1884 __ ldrw(rscratch1, Address(rthread, JavaThread::suspend_flags_offset())); 1885 __ cbnzw(rscratch1, safepoint_in_progress); 1886 __ bind(safepoint_in_progress_done); 1887 } 1888 1889 // change thread state 1890 Label after_transition; 1891 __ mov(rscratch1, _thread_in_Java); 1892 __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset())); 1893 __ stlrw(rscratch1, rscratch2); 1894 __ bind(after_transition); 1895 1896 Label reguard; 1897 Label reguard_done; 1898 __ ldrb(rscratch1, Address(rthread, JavaThread::stack_guard_state_offset())); 1899 __ cmpw(rscratch1, JavaThread::stack_guard_yellow_reserved_disabled); 1900 __ br(Assembler::EQ, reguard); 1901 __ bind(reguard_done); 1902 1903 // native result if any is live 1904 1905 // Unlock 1906 Label unlock_done; 1907 Label slow_path_unlock; 1908 if (method->is_synchronized()) { 1909 1910 // Get locked oop from the handle we passed to jni 1911 __ ldr(obj_reg, Address(oop_handle_reg, 0)); 1912 1913 __ resolve(IS_NOT_NULL, obj_reg); 1914 1915 Label done; 1916 1917 if (UseBiasedLocking) { 1918 __ biased_locking_exit(obj_reg, old_hdr, done); 1919 } 1920 1921 // Simple recursive lock? 1922 1923 __ ldr(rscratch1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 1924 __ cbz(rscratch1, done); 1925 1926 // Must save r0 if if it is live now because cmpxchg must use it 1927 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 1928 save_native_result(masm, ret_type, stack_slots); 1929 } 1930 1931 1932 // get address of the stack lock 1933 __ lea(r0, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 1934 // get old displaced header 1935 __ ldr(old_hdr, Address(r0, 0)); 1936 1937 // Atomic swap old header if oop still contains the stack lock 1938 Label succeed; 1939 __ cmpxchg_obj_header(r0, old_hdr, obj_reg, rscratch1, succeed, &slow_path_unlock); 1940 __ bind(succeed); 1941 1942 // slow path re-enters here 1943 __ bind(unlock_done); 1944 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 1945 restore_native_result(masm, ret_type, stack_slots); 1946 } 1947 1948 __ bind(done); 1949 } 1950 1951 Label dtrace_method_exit, dtrace_method_exit_done; 1952 { 1953 unsigned long offset; 1954 __ adrp(rscratch1, ExternalAddress((address)&DTraceMethodProbes), offset); 1955 __ ldrb(rscratch1, Address(rscratch1, offset)); 1956 __ cbnzw(rscratch1, dtrace_method_exit); 1957 __ bind(dtrace_method_exit_done); 1958 } 1959 1960 __ reset_last_Java_frame(false); 1961 1962 // Unbox oop result, e.g. JNIHandles::resolve result. 1963 if (is_reference_type(ret_type)) { 1964 __ resolve_jobject(r0, rthread, rscratch2); 1965 } 1966 1967 if (CheckJNICalls) { 1968 // clear_pending_jni_exception_check 1969 __ str(zr, Address(rthread, JavaThread::pending_jni_exception_check_fn_offset())); 1970 } 1971 1972 if (!is_critical_native) { 1973 // reset handle block 1974 __ ldr(r2, Address(rthread, JavaThread::active_handles_offset())); 1975 __ str(zr, Address(r2, JNIHandleBlock::top_offset_in_bytes())); 1976 } 1977 1978 __ leave(); 1979 1980 if (!is_critical_native) { 1981 // Any exception pending? 1982 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 1983 __ cbnz(rscratch1, exception_pending); 1984 } 1985 1986 // We're done 1987 __ ret(lr); 1988 1989 // Unexpected paths are out of line and go here 1990 1991 if (!is_critical_native) { 1992 // forward the exception 1993 __ bind(exception_pending); 1994 1995 // and forward the exception 1996 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 1997 } 1998 1999 // Slow path locking & unlocking 2000 if (method->is_synchronized()) { 2001 2002 __ block_comment("Slow path lock {"); 2003 __ bind(slow_path_lock); 2004 2005 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM 2006 // args are (oop obj, BasicLock* lock, JavaThread* thread) 2007 2008 // protect the args we've loaded 2009 save_args(masm, total_c_args, c_arg, out_regs); 2010 2011 __ mov(c_rarg0, obj_reg); 2012 __ mov(c_rarg1, lock_reg); 2013 __ mov(c_rarg2, rthread); 2014 2015 // Not a leaf but we have last_Java_frame setup as we want 2016 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3); 2017 restore_args(masm, total_c_args, c_arg, out_regs); 2018 2019 #ifdef ASSERT 2020 { Label L; 2021 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2022 __ cbz(rscratch1, L); 2023 __ stop("no pending exception allowed on exit from monitorenter"); 2024 __ bind(L); 2025 } 2026 #endif 2027 __ b(lock_done); 2028 2029 __ block_comment("} Slow path lock"); 2030 2031 __ block_comment("Slow path unlock {"); 2032 __ bind(slow_path_unlock); 2033 2034 // If we haven't already saved the native result we must save it now as xmm registers 2035 // are still exposed. 2036 2037 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) { 2038 save_native_result(masm, ret_type, stack_slots); 2039 } 2040 2041 __ mov(c_rarg2, rthread); 2042 __ lea(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 2043 __ mov(c_rarg0, obj_reg); 2044 2045 // Save pending exception around call to VM (which contains an EXCEPTION_MARK) 2046 // NOTE that obj_reg == r19 currently 2047 __ ldr(r19, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2048 __ str(zr, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2049 2050 rt_call(masm, CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 3, 0, 1); 2051 2052 #ifdef ASSERT 2053 { 2054 Label L; 2055 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2056 __ cbz(rscratch1, L); 2057 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C"); 2058 __ bind(L); 2059 } 2060 #endif /* ASSERT */ 2061 2062 __ str(r19, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2063 2064 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) { 2065 restore_native_result(masm, ret_type, stack_slots); 2066 } 2067 __ b(unlock_done); 2068 2069 __ block_comment("} Slow path unlock"); 2070 2071 } // synchronized 2072 2073 // SLOW PATH Reguard the stack if needed 2074 2075 __ bind(reguard); 2076 save_native_result(masm, ret_type, stack_slots); 2077 rt_call(masm, CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages), 0, 0, 0); 2078 restore_native_result(masm, ret_type, stack_slots); 2079 // and continue 2080 __ b(reguard_done); 2081 2082 // SLOW PATH safepoint 2083 { 2084 __ block_comment("safepoint {"); 2085 __ bind(safepoint_in_progress); 2086 2087 // Don't use call_VM as it will see a possible pending exception and forward it 2088 // and never return here preventing us from clearing _last_native_pc down below. 2089 // 2090 save_native_result(masm, ret_type, stack_slots); 2091 __ mov(c_rarg0, rthread); 2092 #ifndef PRODUCT 2093 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 2094 #endif 2095 if (!is_critical_native) { 2096 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans))); 2097 } else { 2098 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans_and_transition))); 2099 } 2100 __ blr(rscratch1); 2101 __ maybe_isb(); 2102 // Restore any method result value 2103 restore_native_result(masm, ret_type, stack_slots); 2104 2105 if (is_critical_native) { 2106 // The call above performed the transition to thread_in_Java so 2107 // skip the transition logic above. 2108 __ b(after_transition); 2109 } 2110 2111 __ b(safepoint_in_progress_done); 2112 __ block_comment("} safepoint"); 2113 } 2114 2115 // SLOW PATH dtrace support 2116 { 2117 __ block_comment("dtrace entry {"); 2118 __ bind(dtrace_method_entry); 2119 2120 // We have all of the arguments setup at this point. We must not touch any register 2121 // argument registers at this point (what if we save/restore them there are no oop? 2122 2123 save_args(masm, total_c_args, c_arg, out_regs); 2124 __ mov_metadata(c_rarg1, method()); 2125 __ call_VM_leaf( 2126 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 2127 rthread, c_rarg1); 2128 restore_args(masm, total_c_args, c_arg, out_regs); 2129 __ b(dtrace_method_entry_done); 2130 __ block_comment("} dtrace entry"); 2131 } 2132 2133 { 2134 __ block_comment("dtrace exit {"); 2135 __ bind(dtrace_method_exit); 2136 save_native_result(masm, ret_type, stack_slots); 2137 __ mov_metadata(c_rarg1, method()); 2138 __ call_VM_leaf( 2139 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 2140 rthread, c_rarg1); 2141 restore_native_result(masm, ret_type, stack_slots); 2142 __ b(dtrace_method_exit_done); 2143 __ block_comment("} dtrace exit"); 2144 } 2145 2146 2147 __ flush(); 2148 2149 nmethod *nm = nmethod::new_native_nmethod(method, 2150 compile_id, 2151 masm->code(), 2152 vep_offset, 2153 frame_complete, 2154 stack_slots / VMRegImpl::slots_per_word, 2155 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), 2156 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size), 2157 oop_maps); 2158 2159 if (is_critical_native) { 2160 nm->set_lazy_critical_native(true); 2161 } 2162 2163 return nm; 2164 2165 } 2166 2167 // this function returns the adjust size (in number of words) to a c2i adapter 2168 // activation for use during deoptimization 2169 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) { 2170 assert(callee_locals >= callee_parameters, 2171 "test and remove; got more parms than locals"); 2172 if (callee_locals < callee_parameters) 2173 return 0; // No adjustment for negative locals 2174 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords; 2175 // diff is counted in stack words 2176 return align_up(diff, 2); 2177 } 2178 2179 2180 //------------------------------generate_deopt_blob---------------------------- 2181 void SharedRuntime::generate_deopt_blob() { 2182 // Allocate space for the code 2183 ResourceMark rm; 2184 // Setup code generation tools 2185 int pad = 0; 2186 #if INCLUDE_JVMCI 2187 if (EnableJVMCI || UseAOT) { 2188 pad += 512; // Increase the buffer size when compiling for JVMCI 2189 } 2190 #endif 2191 CodeBuffer buffer("deopt_blob", 2048+pad, 1024); 2192 MacroAssembler* masm = new MacroAssembler(&buffer); 2193 int frame_size_in_words; 2194 OopMap* map = NULL; 2195 OopMapSet *oop_maps = new OopMapSet(); 2196 2197 // ------------- 2198 // This code enters when returning to a de-optimized nmethod. A return 2199 // address has been pushed on the the stack, and return values are in 2200 // registers. 2201 // If we are doing a normal deopt then we were called from the patched 2202 // nmethod from the point we returned to the nmethod. So the return 2203 // address on the stack is wrong by NativeCall::instruction_size 2204 // We will adjust the value so it looks like we have the original return 2205 // address on the stack (like when we eagerly deoptimized). 2206 // In the case of an exception pending when deoptimizing, we enter 2207 // with a return address on the stack that points after the call we patched 2208 // into the exception handler. We have the following register state from, 2209 // e.g., the forward exception stub (see stubGenerator_x86_64.cpp). 2210 // r0: exception oop 2211 // r19: exception handler 2212 // r3: throwing pc 2213 // So in this case we simply jam r3 into the useless return address and 2214 // the stack looks just like we want. 2215 // 2216 // At this point we need to de-opt. We save the argument return 2217 // registers. We call the first C routine, fetch_unroll_info(). This 2218 // routine captures the return values and returns a structure which 2219 // describes the current frame size and the sizes of all replacement frames. 2220 // The current frame is compiled code and may contain many inlined 2221 // functions, each with their own JVM state. We pop the current frame, then 2222 // push all the new frames. Then we call the C routine unpack_frames() to 2223 // populate these frames. Finally unpack_frames() returns us the new target 2224 // address. Notice that callee-save registers are BLOWN here; they have 2225 // already been captured in the vframeArray at the time the return PC was 2226 // patched. 2227 address start = __ pc(); 2228 Label cont; 2229 2230 // Prolog for non exception case! 2231 2232 // Save everything in sight. 2233 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2234 2235 // Normal deoptimization. Save exec mode for unpack_frames. 2236 __ movw(rcpool, Deoptimization::Unpack_deopt); // callee-saved 2237 __ b(cont); 2238 2239 int reexecute_offset = __ pc() - start; 2240 #if INCLUDE_JVMCI && !defined(COMPILER1) 2241 if (EnableJVMCI && UseJVMCICompiler) { 2242 // JVMCI does not use this kind of deoptimization 2243 __ should_not_reach_here(); 2244 } 2245 #endif 2246 2247 // Reexecute case 2248 // return address is the pc describes what bci to do re-execute at 2249 2250 // No need to update map as each call to save_live_registers will produce identical oopmap 2251 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2252 2253 __ movw(rcpool, Deoptimization::Unpack_reexecute); // callee-saved 2254 __ b(cont); 2255 2256 #if INCLUDE_JVMCI 2257 Label after_fetch_unroll_info_call; 2258 int implicit_exception_uncommon_trap_offset = 0; 2259 int uncommon_trap_offset = 0; 2260 2261 if (EnableJVMCI || UseAOT) { 2262 implicit_exception_uncommon_trap_offset = __ pc() - start; 2263 2264 __ ldr(lr, Address(rthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset()))); 2265 __ str(zr, Address(rthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset()))); 2266 2267 uncommon_trap_offset = __ pc() - start; 2268 2269 // Save everything in sight. 2270 RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2271 // fetch_unroll_info needs to call last_java_frame() 2272 Label retaddr; 2273 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2274 2275 __ ldrw(c_rarg1, Address(rthread, in_bytes(JavaThread::pending_deoptimization_offset()))); 2276 __ movw(rscratch1, -1); 2277 __ strw(rscratch1, Address(rthread, in_bytes(JavaThread::pending_deoptimization_offset()))); 2278 2279 __ movw(rcpool, (int32_t)Deoptimization::Unpack_reexecute); 2280 __ mov(c_rarg0, rthread); 2281 __ movw(c_rarg2, rcpool); // exec mode 2282 __ lea(rscratch1, 2283 RuntimeAddress(CAST_FROM_FN_PTR(address, 2284 Deoptimization::uncommon_trap))); 2285 __ blr(rscratch1); 2286 __ bind(retaddr); 2287 oop_maps->add_gc_map( __ pc()-start, map->deep_copy()); 2288 2289 __ reset_last_Java_frame(false); 2290 2291 __ b(after_fetch_unroll_info_call); 2292 } // EnableJVMCI 2293 #endif // INCLUDE_JVMCI 2294 2295 int exception_offset = __ pc() - start; 2296 2297 // Prolog for exception case 2298 2299 // all registers are dead at this entry point, except for r0, and 2300 // r3 which contain the exception oop and exception pc 2301 // respectively. Set them in TLS and fall thru to the 2302 // unpack_with_exception_in_tls entry point. 2303 2304 __ str(r3, Address(rthread, JavaThread::exception_pc_offset())); 2305 __ str(r0, Address(rthread, JavaThread::exception_oop_offset())); 2306 2307 int exception_in_tls_offset = __ pc() - start; 2308 2309 // new implementation because exception oop is now passed in JavaThread 2310 2311 // Prolog for exception case 2312 // All registers must be preserved because they might be used by LinearScan 2313 // Exceptiop oop and throwing PC are passed in JavaThread 2314 // tos: stack at point of call to method that threw the exception (i.e. only 2315 // args are on the stack, no return address) 2316 2317 // The return address pushed by save_live_registers will be patched 2318 // later with the throwing pc. The correct value is not available 2319 // now because loading it from memory would destroy registers. 2320 2321 // NB: The SP at this point must be the SP of the method that is 2322 // being deoptimized. Deoptimization assumes that the frame created 2323 // here by save_live_registers is immediately below the method's SP. 2324 // This is a somewhat fragile mechanism. 2325 2326 // Save everything in sight. 2327 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2328 2329 // Now it is safe to overwrite any register 2330 2331 // Deopt during an exception. Save exec mode for unpack_frames. 2332 __ mov(rcpool, Deoptimization::Unpack_exception); // callee-saved 2333 2334 // load throwing pc from JavaThread and patch it as the return address 2335 // of the current frame. Then clear the field in JavaThread 2336 2337 __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset())); 2338 __ str(r3, Address(rfp, wordSize)); 2339 __ str(zr, Address(rthread, JavaThread::exception_pc_offset())); 2340 2341 #ifdef ASSERT 2342 // verify that there is really an exception oop in JavaThread 2343 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset())); 2344 __ verify_oop(r0); 2345 2346 // verify that there is no pending exception 2347 Label no_pending_exception; 2348 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset())); 2349 __ cbz(rscratch1, no_pending_exception); 2350 __ stop("must not have pending exception here"); 2351 __ bind(no_pending_exception); 2352 #endif 2353 2354 __ bind(cont); 2355 2356 // Call C code. Need thread and this frame, but NOT official VM entry 2357 // crud. We cannot block on this call, no GC can happen. 2358 // 2359 // UnrollBlock* fetch_unroll_info(JavaThread* thread) 2360 2361 // fetch_unroll_info needs to call last_java_frame(). 2362 2363 Label retaddr; 2364 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2365 #ifdef ASSERT0 2366 { Label L; 2367 __ ldr(rscratch1, Address(rthread, 2368 JavaThread::last_Java_fp_offset())); 2369 __ cbz(rscratch1, L); 2370 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared"); 2371 __ bind(L); 2372 } 2373 #endif // ASSERT 2374 __ mov(c_rarg0, rthread); 2375 __ mov(c_rarg1, rcpool); 2376 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info))); 2377 __ blr(rscratch1); 2378 __ bind(retaddr); 2379 2380 // Need to have an oopmap that tells fetch_unroll_info where to 2381 // find any register it might need. 2382 oop_maps->add_gc_map(__ pc() - start, map); 2383 2384 __ reset_last_Java_frame(false); 2385 2386 #if INCLUDE_JVMCI 2387 if (EnableJVMCI || UseAOT) { 2388 __ bind(after_fetch_unroll_info_call); 2389 } 2390 #endif 2391 2392 // Load UnrollBlock* into r5 2393 __ mov(r5, r0); 2394 2395 __ ldrw(rcpool, Address(r5, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes())); 2396 Label noException; 2397 __ cmpw(rcpool, Deoptimization::Unpack_exception); // Was exception pending? 2398 __ br(Assembler::NE, noException); 2399 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset())); 2400 // QQQ this is useless it was NULL above 2401 __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset())); 2402 __ str(zr, Address(rthread, JavaThread::exception_oop_offset())); 2403 __ str(zr, Address(rthread, JavaThread::exception_pc_offset())); 2404 2405 __ verify_oop(r0); 2406 2407 // Overwrite the result registers with the exception results. 2408 __ str(r0, Address(sp, RegisterSaver::r0_offset_in_bytes())); 2409 // I think this is useless 2410 // __ str(r3, Address(sp, RegisterSaver::r3_offset_in_bytes())); 2411 2412 __ bind(noException); 2413 2414 // Only register save data is on the stack. 2415 // Now restore the result registers. Everything else is either dead 2416 // or captured in the vframeArray. 2417 RegisterSaver::restore_result_registers(masm); 2418 2419 // All of the register save area has been popped of the stack. Only the 2420 // return address remains. 2421 2422 // Pop all the frames we must move/replace. 2423 // 2424 // Frame picture (youngest to oldest) 2425 // 1: self-frame (no frame link) 2426 // 2: deopting frame (no frame link) 2427 // 3: caller of deopting frame (could be compiled/interpreted). 2428 // 2429 // Note: by leaving the return address of self-frame on the stack 2430 // and using the size of frame 2 to adjust the stack 2431 // when we are done the return to frame 3 will still be on the stack. 2432 2433 // Pop deoptimized frame 2434 __ ldrw(r2, Address(r5, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes())); 2435 __ sub(r2, r2, 2 * wordSize); 2436 __ add(sp, sp, r2); 2437 __ ldp(rfp, lr, __ post(sp, 2 * wordSize)); 2438 // LR should now be the return address to the caller (3) 2439 2440 #ifdef ASSERT 2441 // Compilers generate code that bang the stack by as much as the 2442 // interpreter would need. So this stack banging should never 2443 // trigger a fault. Verify that it does not on non product builds. 2444 if (UseStackBanging) { 2445 __ ldrw(r19, Address(r5, Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes())); 2446 __ bang_stack_size(r19, r2); 2447 } 2448 #endif 2449 // Load address of array of frame pcs into r2 2450 __ ldr(r2, Address(r5, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 2451 2452 // Trash the old pc 2453 // __ addptr(sp, wordSize); FIXME ???? 2454 2455 // Load address of array of frame sizes into r4 2456 __ ldr(r4, Address(r5, Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes())); 2457 2458 // Load counter into r3 2459 __ ldrw(r3, Address(r5, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes())); 2460 2461 // Now adjust the caller's stack to make up for the extra locals 2462 // but record the original sp so that we can save it in the skeletal interpreter 2463 // frame and the stack walking of interpreter_sender will get the unextended sp 2464 // value and not the "real" sp value. 2465 2466 const Register sender_sp = r6; 2467 2468 __ mov(sender_sp, sp); 2469 __ ldrw(r19, Address(r5, 2470 Deoptimization::UnrollBlock:: 2471 caller_adjustment_offset_in_bytes())); 2472 __ sub(sp, sp, r19); 2473 2474 // Push interpreter frames in a loop 2475 __ mov(rscratch1, (address)0xDEADDEAD); // Make a recognizable pattern 2476 __ mov(rscratch2, rscratch1); 2477 Label loop; 2478 __ bind(loop); 2479 __ ldr(r19, Address(__ post(r4, wordSize))); // Load frame size 2480 __ sub(r19, r19, 2*wordSize); // We'll push pc and fp by hand 2481 __ ldr(lr, Address(__ post(r2, wordSize))); // Load pc 2482 __ enter(); // Save old & set new fp 2483 __ sub(sp, sp, r19); // Prolog 2484 // This value is corrected by layout_activation_impl 2485 __ str(zr, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 2486 __ str(sender_sp, Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable 2487 __ mov(sender_sp, sp); // Pass sender_sp to next frame 2488 __ sub(r3, r3, 1); // Decrement counter 2489 __ cbnz(r3, loop); 2490 2491 // Re-push self-frame 2492 __ ldr(lr, Address(r2)); 2493 __ enter(); 2494 2495 // Allocate a full sized register save area. We subtract 2 because 2496 // enter() just pushed 2 words 2497 __ sub(sp, sp, (frame_size_in_words - 2) * wordSize); 2498 2499 // Restore frame locals after moving the frame 2500 __ strd(v0, Address(sp, RegisterSaver::v0_offset_in_bytes())); 2501 __ str(r0, Address(sp, RegisterSaver::r0_offset_in_bytes())); 2502 2503 // Call C code. Need thread but NOT official VM entry 2504 // crud. We cannot block on this call, no GC can happen. Call should 2505 // restore return values to their stack-slots with the new SP. 2506 // 2507 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode) 2508 2509 // Use rfp because the frames look interpreted now 2510 // Don't need the precise return PC here, just precise enough to point into this code blob. 2511 address the_pc = __ pc(); 2512 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1); 2513 2514 __ mov(c_rarg0, rthread); 2515 __ movw(c_rarg1, rcpool); // second arg: exec_mode 2516 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 2517 __ blr(rscratch1); 2518 2519 // Set an oopmap for the call site 2520 // Use the same PC we used for the last java frame 2521 oop_maps->add_gc_map(the_pc - start, 2522 new OopMap( frame_size_in_words, 0 )); 2523 2524 // Clear fp AND pc 2525 __ reset_last_Java_frame(true); 2526 2527 // Collect return values 2528 __ ldrd(v0, Address(sp, RegisterSaver::v0_offset_in_bytes())); 2529 __ ldr(r0, Address(sp, RegisterSaver::r0_offset_in_bytes())); 2530 // I think this is useless (throwing pc?) 2531 // __ ldr(r3, Address(sp, RegisterSaver::r3_offset_in_bytes())); 2532 2533 // Pop self-frame. 2534 __ leave(); // Epilog 2535 2536 // Jump to interpreter 2537 __ ret(lr); 2538 2539 // Make sure all code is generated 2540 masm->flush(); 2541 2542 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words); 2543 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset); 2544 #if INCLUDE_JVMCI 2545 if (EnableJVMCI || UseAOT) { 2546 _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset); 2547 _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset); 2548 } 2549 #endif 2550 } 2551 2552 uint SharedRuntime::out_preserve_stack_slots() { 2553 return 0; 2554 } 2555 2556 #ifdef COMPILER2 2557 //------------------------------generate_uncommon_trap_blob-------------------- 2558 void SharedRuntime::generate_uncommon_trap_blob() { 2559 // Allocate space for the code 2560 ResourceMark rm; 2561 // Setup code generation tools 2562 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024); 2563 MacroAssembler* masm = new MacroAssembler(&buffer); 2564 2565 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned"); 2566 2567 address start = __ pc(); 2568 2569 // Push self-frame. We get here with a return address in LR 2570 // and sp should be 16 byte aligned 2571 // push rfp and retaddr by hand 2572 __ stp(rfp, lr, Address(__ pre(sp, -2 * wordSize))); 2573 // we don't expect an arg reg save area 2574 #ifndef PRODUCT 2575 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 2576 #endif 2577 // compiler left unloaded_class_index in j_rarg0 move to where the 2578 // runtime expects it. 2579 if (c_rarg1 != j_rarg0) { 2580 __ movw(c_rarg1, j_rarg0); 2581 } 2582 2583 // we need to set the past SP to the stack pointer of the stub frame 2584 // and the pc to the address where this runtime call will return 2585 // although actually any pc in this code blob will do). 2586 Label retaddr; 2587 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2588 2589 // Call C code. Need thread but NOT official VM entry 2590 // crud. We cannot block on this call, no GC can happen. Call should 2591 // capture callee-saved registers as well as return values. 2592 // Thread is in rdi already. 2593 // 2594 // UnrollBlock* uncommon_trap(JavaThread* thread, jint unloaded_class_index); 2595 // 2596 // n.b. 2 gp args, 0 fp args, integral return type 2597 2598 __ mov(c_rarg0, rthread); 2599 __ movw(c_rarg2, (unsigned)Deoptimization::Unpack_uncommon_trap); 2600 __ lea(rscratch1, 2601 RuntimeAddress(CAST_FROM_FN_PTR(address, 2602 Deoptimization::uncommon_trap))); 2603 __ blr(rscratch1); 2604 __ bind(retaddr); 2605 2606 // Set an oopmap for the call site 2607 OopMapSet* oop_maps = new OopMapSet(); 2608 OopMap* map = new OopMap(SimpleRuntimeFrame::framesize, 0); 2609 2610 // location of rfp is known implicitly by the frame sender code 2611 2612 oop_maps->add_gc_map(__ pc() - start, map); 2613 2614 __ reset_last_Java_frame(false); 2615 2616 // move UnrollBlock* into r4 2617 __ mov(r4, r0); 2618 2619 #ifdef ASSERT 2620 { Label L; 2621 __ ldrw(rscratch1, Address(r4, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes())); 2622 __ cmpw(rscratch1, (unsigned)Deoptimization::Unpack_uncommon_trap); 2623 __ br(Assembler::EQ, L); 2624 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared"); 2625 __ bind(L); 2626 } 2627 #endif 2628 2629 // Pop all the frames we must move/replace. 2630 // 2631 // Frame picture (youngest to oldest) 2632 // 1: self-frame (no frame link) 2633 // 2: deopting frame (no frame link) 2634 // 3: caller of deopting frame (could be compiled/interpreted). 2635 2636 // Pop self-frame. We have no frame, and must rely only on r0 and sp. 2637 __ add(sp, sp, (SimpleRuntimeFrame::framesize) << LogBytesPerInt); // Epilog! 2638 2639 // Pop deoptimized frame (int) 2640 __ ldrw(r2, Address(r4, 2641 Deoptimization::UnrollBlock:: 2642 size_of_deoptimized_frame_offset_in_bytes())); 2643 __ sub(r2, r2, 2 * wordSize); 2644 __ add(sp, sp, r2); 2645 __ ldp(rfp, lr, __ post(sp, 2 * wordSize)); 2646 // LR should now be the return address to the caller (3) frame 2647 2648 #ifdef ASSERT 2649 // Compilers generate code that bang the stack by as much as the 2650 // interpreter would need. So this stack banging should never 2651 // trigger a fault. Verify that it does not on non product builds. 2652 if (UseStackBanging) { 2653 __ ldrw(r1, Address(r4, 2654 Deoptimization::UnrollBlock:: 2655 total_frame_sizes_offset_in_bytes())); 2656 __ bang_stack_size(r1, r2); 2657 } 2658 #endif 2659 2660 // Load address of array of frame pcs into r2 (address*) 2661 __ ldr(r2, Address(r4, 2662 Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 2663 2664 // Load address of array of frame sizes into r5 (intptr_t*) 2665 __ ldr(r5, Address(r4, 2666 Deoptimization::UnrollBlock:: 2667 frame_sizes_offset_in_bytes())); 2668 2669 // Counter 2670 __ ldrw(r3, Address(r4, 2671 Deoptimization::UnrollBlock:: 2672 number_of_frames_offset_in_bytes())); // (int) 2673 2674 // Now adjust the caller's stack to make up for the extra locals but 2675 // record the original sp so that we can save it in the skeletal 2676 // interpreter frame and the stack walking of interpreter_sender 2677 // will get the unextended sp value and not the "real" sp value. 2678 2679 const Register sender_sp = r8; 2680 2681 __ mov(sender_sp, sp); 2682 __ ldrw(r1, Address(r4, 2683 Deoptimization::UnrollBlock:: 2684 caller_adjustment_offset_in_bytes())); // (int) 2685 __ sub(sp, sp, r1); 2686 2687 // Push interpreter frames in a loop 2688 Label loop; 2689 __ bind(loop); 2690 __ ldr(r1, Address(r5, 0)); // Load frame size 2691 __ sub(r1, r1, 2 * wordSize); // We'll push pc and rfp by hand 2692 __ ldr(lr, Address(r2, 0)); // Save return address 2693 __ enter(); // and old rfp & set new rfp 2694 __ sub(sp, sp, r1); // Prolog 2695 __ str(sender_sp, Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable 2696 // This value is corrected by layout_activation_impl 2697 __ str(zr, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 2698 __ mov(sender_sp, sp); // Pass sender_sp to next frame 2699 __ add(r5, r5, wordSize); // Bump array pointer (sizes) 2700 __ add(r2, r2, wordSize); // Bump array pointer (pcs) 2701 __ subsw(r3, r3, 1); // Decrement counter 2702 __ br(Assembler::GT, loop); 2703 __ ldr(lr, Address(r2, 0)); // save final return address 2704 // Re-push self-frame 2705 __ enter(); // & old rfp & set new rfp 2706 2707 // Use rfp because the frames look interpreted now 2708 // Save "the_pc" since it cannot easily be retrieved using the last_java_SP after we aligned SP. 2709 // Don't need the precise return PC here, just precise enough to point into this code blob. 2710 address the_pc = __ pc(); 2711 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1); 2712 2713 // Call C code. Need thread but NOT official VM entry 2714 // crud. We cannot block on this call, no GC can happen. Call should 2715 // restore return values to their stack-slots with the new SP. 2716 // Thread is in rdi already. 2717 // 2718 // BasicType unpack_frames(JavaThread* thread, int exec_mode); 2719 // 2720 // n.b. 2 gp args, 0 fp args, integral return type 2721 2722 // sp should already be aligned 2723 __ mov(c_rarg0, rthread); 2724 __ movw(c_rarg1, (unsigned)Deoptimization::Unpack_uncommon_trap); 2725 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 2726 __ blr(rscratch1); 2727 2728 // Set an oopmap for the call site 2729 // Use the same PC we used for the last java frame 2730 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0)); 2731 2732 // Clear fp AND pc 2733 __ reset_last_Java_frame(true); 2734 2735 // Pop self-frame. 2736 __ leave(); // Epilog 2737 2738 // Jump to interpreter 2739 __ ret(lr); 2740 2741 // Make sure all code is generated 2742 masm->flush(); 2743 2744 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, 2745 SimpleRuntimeFrame::framesize >> 1); 2746 } 2747 #endif // COMPILER2 2748 2749 2750 //------------------------------generate_handler_blob------ 2751 // 2752 // Generate a special Compile2Runtime blob that saves all registers, 2753 // and setup oopmap. 2754 // 2755 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) { 2756 ResourceMark rm; 2757 OopMapSet *oop_maps = new OopMapSet(); 2758 OopMap* map; 2759 2760 // Allocate space for the code. Setup code generation tools. 2761 CodeBuffer buffer("handler_blob", 2048, 1024); 2762 MacroAssembler* masm = new MacroAssembler(&buffer); 2763 2764 address start = __ pc(); 2765 address call_pc = NULL; 2766 int frame_size_in_words; 2767 bool cause_return = (poll_type == POLL_AT_RETURN); 2768 bool save_vectors = (poll_type == POLL_AT_VECTOR_LOOP); 2769 2770 // Save Integer and Float registers. 2771 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, save_vectors); 2772 2773 // The following is basically a call_VM. However, we need the precise 2774 // address of the call in order to generate an oopmap. Hence, we do all the 2775 // work outselves. 2776 2777 Label retaddr; 2778 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2779 2780 // The return address must always be correct so that frame constructor never 2781 // sees an invalid pc. 2782 2783 if (!cause_return) { 2784 // overwrite the return address pushed by save_live_registers 2785 // Additionally, r20 is a callee-saved register so we can look at 2786 // it later to determine if someone changed the return address for 2787 // us! 2788 __ ldr(r20, Address(rthread, JavaThread::saved_exception_pc_offset())); 2789 __ str(r20, Address(rfp, wordSize)); 2790 } 2791 2792 // Do the call 2793 __ mov(c_rarg0, rthread); 2794 __ lea(rscratch1, RuntimeAddress(call_ptr)); 2795 __ blr(rscratch1); 2796 __ bind(retaddr); 2797 2798 // Set an oopmap for the call site. This oopmap will map all 2799 // oop-registers and debug-info registers as callee-saved. This 2800 // will allow deoptimization at this safepoint to find all possible 2801 // debug-info recordings, as well as let GC find all oops. 2802 2803 oop_maps->add_gc_map( __ pc() - start, map); 2804 2805 Label noException; 2806 2807 __ reset_last_Java_frame(false); 2808 2809 __ maybe_isb(); 2810 __ membar(Assembler::LoadLoad | Assembler::LoadStore); 2811 2812 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset())); 2813 __ cbz(rscratch1, noException); 2814 2815 // Exception pending 2816 2817 RegisterSaver::restore_live_registers(masm, save_vectors); 2818 2819 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2820 2821 // No exception case 2822 __ bind(noException); 2823 2824 Label no_adjust, bail; 2825 if (!cause_return) { 2826 // If our stashed return pc was modified by the runtime we avoid touching it 2827 __ ldr(rscratch1, Address(rfp, wordSize)); 2828 __ cmp(r20, rscratch1); 2829 __ br(Assembler::NE, no_adjust); 2830 2831 #ifdef ASSERT 2832 // Verify the correct encoding of the poll we're about to skip. 2833 // See NativeInstruction::is_ldrw_to_zr() 2834 __ ldrw(rscratch1, Address(r20)); 2835 __ ubfx(rscratch2, rscratch1, 22, 10); 2836 __ cmpw(rscratch2, 0b1011100101); 2837 __ br(Assembler::NE, bail); 2838 __ ubfx(rscratch2, rscratch1, 0, 5); 2839 __ cmpw(rscratch2, 0b11111); 2840 __ br(Assembler::NE, bail); 2841 #endif 2842 // Adjust return pc forward to step over the safepoint poll instruction 2843 __ add(r20, r20, NativeInstruction::instruction_size); 2844 __ str(r20, Address(rfp, wordSize)); 2845 } 2846 2847 __ bind(no_adjust); 2848 // Normal exit, restore registers and exit. 2849 RegisterSaver::restore_live_registers(masm, save_vectors); 2850 2851 __ ret(lr); 2852 2853 #ifdef ASSERT 2854 __ bind(bail); 2855 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected"); 2856 #endif 2857 2858 // Make sure all code is generated 2859 masm->flush(); 2860 2861 // Fill-out other meta info 2862 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words); 2863 } 2864 2865 // 2866 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss 2867 // 2868 // Generate a stub that calls into vm to find out the proper destination 2869 // of a java call. All the argument registers are live at this point 2870 // but since this is generic code we don't know what they are and the caller 2871 // must do any gc of the args. 2872 // 2873 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) { 2874 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before"); 2875 2876 // allocate space for the code 2877 ResourceMark rm; 2878 2879 CodeBuffer buffer(name, 1000, 512); 2880 MacroAssembler* masm = new MacroAssembler(&buffer); 2881 2882 int frame_size_in_words; 2883 2884 OopMapSet *oop_maps = new OopMapSet(); 2885 OopMap* map = NULL; 2886 2887 int start = __ offset(); 2888 2889 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2890 2891 int frame_complete = __ offset(); 2892 2893 { 2894 Label retaddr; 2895 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2896 2897 __ mov(c_rarg0, rthread); 2898 __ lea(rscratch1, RuntimeAddress(destination)); 2899 2900 __ blr(rscratch1); 2901 __ bind(retaddr); 2902 } 2903 2904 // Set an oopmap for the call site. 2905 // We need this not only for callee-saved registers, but also for volatile 2906 // registers that the compiler might be keeping live across a safepoint. 2907 2908 oop_maps->add_gc_map( __ offset() - start, map); 2909 2910 __ maybe_isb(); 2911 2912 // r0 contains the address we are going to jump to assuming no exception got installed 2913 2914 // clear last_Java_sp 2915 __ reset_last_Java_frame(false); 2916 // check for pending exceptions 2917 Label pending; 2918 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset())); 2919 __ cbnz(rscratch1, pending); 2920 2921 // get the returned Method* 2922 __ get_vm_result_2(rmethod, rthread); 2923 __ str(rmethod, Address(sp, RegisterSaver::reg_offset_in_bytes(rmethod))); 2924 2925 // r0 is where we want to jump, overwrite rscratch1 which is saved and scratch 2926 __ str(r0, Address(sp, RegisterSaver::rscratch1_offset_in_bytes())); 2927 RegisterSaver::restore_live_registers(masm); 2928 2929 // We are back the the original state on entry and ready to go. 2930 2931 __ br(rscratch1); 2932 2933 // Pending exception after the safepoint 2934 2935 __ bind(pending); 2936 2937 RegisterSaver::restore_live_registers(masm); 2938 2939 // exception pending => remove activation and forward to exception handler 2940 2941 __ str(zr, Address(rthread, JavaThread::vm_result_offset())); 2942 2943 __ ldr(r0, Address(rthread, Thread::pending_exception_offset())); 2944 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2945 2946 // ------------- 2947 // make sure all code is generated 2948 masm->flush(); 2949 2950 // return the blob 2951 // frame_size_words or bytes?? 2952 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true); 2953 } 2954 2955 #ifdef COMPILER2 2956 // This is here instead of runtime_x86_64.cpp because it uses SimpleRuntimeFrame 2957 // 2958 //------------------------------generate_exception_blob--------------------------- 2959 // creates exception blob at the end 2960 // Using exception blob, this code is jumped from a compiled method. 2961 // (see emit_exception_handler in x86_64.ad file) 2962 // 2963 // Given an exception pc at a call we call into the runtime for the 2964 // handler in this method. This handler might merely restore state 2965 // (i.e. callee save registers) unwind the frame and jump to the 2966 // exception handler for the nmethod if there is no Java level handler 2967 // for the nmethod. 2968 // 2969 // This code is entered with a jmp. 2970 // 2971 // Arguments: 2972 // r0: exception oop 2973 // r3: exception pc 2974 // 2975 // Results: 2976 // r0: exception oop 2977 // r3: exception pc in caller or ??? 2978 // destination: exception handler of caller 2979 // 2980 // Note: the exception pc MUST be at a call (precise debug information) 2981 // Registers r0, r3, r2, r4, r5, r8-r11 are not callee saved. 2982 // 2983 2984 void OptoRuntime::generate_exception_blob() { 2985 assert(!OptoRuntime::is_callee_saved_register(R3_num), ""); 2986 assert(!OptoRuntime::is_callee_saved_register(R0_num), ""); 2987 assert(!OptoRuntime::is_callee_saved_register(R2_num), ""); 2988 2989 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned"); 2990 2991 // Allocate space for the code 2992 ResourceMark rm; 2993 // Setup code generation tools 2994 CodeBuffer buffer("exception_blob", 2048, 1024); 2995 MacroAssembler* masm = new MacroAssembler(&buffer); 2996 2997 // TODO check various assumptions made here 2998 // 2999 // make sure we do so before running this 3000 3001 address start = __ pc(); 3002 3003 // push rfp and retaddr by hand 3004 // Exception pc is 'return address' for stack walker 3005 __ stp(rfp, lr, Address(__ pre(sp, -2 * wordSize))); 3006 // there are no callee save registers and we don't expect an 3007 // arg reg save area 3008 #ifndef PRODUCT 3009 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 3010 #endif 3011 // Store exception in Thread object. We cannot pass any arguments to the 3012 // handle_exception call, since we do not want to make any assumption 3013 // about the size of the frame where the exception happened in. 3014 __ str(r0, Address(rthread, JavaThread::exception_oop_offset())); 3015 __ str(r3, Address(rthread, JavaThread::exception_pc_offset())); 3016 3017 // This call does all the hard work. It checks if an exception handler 3018 // exists in the method. 3019 // If so, it returns the handler address. 3020 // If not, it prepares for stack-unwinding, restoring the callee-save 3021 // registers of the frame being removed. 3022 // 3023 // address OptoRuntime::handle_exception_C(JavaThread* thread) 3024 // 3025 // n.b. 1 gp arg, 0 fp args, integral return type 3026 3027 // the stack should always be aligned 3028 address the_pc = __ pc(); 3029 __ set_last_Java_frame(sp, noreg, the_pc, rscratch1); 3030 __ mov(c_rarg0, rthread); 3031 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, OptoRuntime::handle_exception_C))); 3032 __ blr(rscratch1); 3033 __ maybe_isb(); 3034 3035 // Set an oopmap for the call site. This oopmap will only be used if we 3036 // are unwinding the stack. Hence, all locations will be dead. 3037 // Callee-saved registers will be the same as the frame above (i.e., 3038 // handle_exception_stub), since they were restored when we got the 3039 // exception. 3040 3041 OopMapSet* oop_maps = new OopMapSet(); 3042 3043 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0)); 3044 3045 __ reset_last_Java_frame(false); 3046 3047 // Restore callee-saved registers 3048 3049 // rfp is an implicitly saved callee saved register (i.e. the calling 3050 // convention will save restore it in prolog/epilog) Other than that 3051 // there are no callee save registers now that adapter frames are gone. 3052 // and we dont' expect an arg reg save area 3053 __ ldp(rfp, r3, Address(__ post(sp, 2 * wordSize))); 3054 3055 // r0: exception handler 3056 3057 // We have a handler in r0 (could be deopt blob). 3058 __ mov(r8, r0); 3059 3060 // Get the exception oop 3061 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset())); 3062 // Get the exception pc in case we are deoptimized 3063 __ ldr(r4, Address(rthread, JavaThread::exception_pc_offset())); 3064 #ifdef ASSERT 3065 __ str(zr, Address(rthread, JavaThread::exception_handler_pc_offset())); 3066 __ str(zr, Address(rthread, JavaThread::exception_pc_offset())); 3067 #endif 3068 // Clear the exception oop so GC no longer processes it as a root. 3069 __ str(zr, Address(rthread, JavaThread::exception_oop_offset())); 3070 3071 // r0: exception oop 3072 // r8: exception handler 3073 // r4: exception pc 3074 // Jump to handler 3075 3076 __ br(r8); 3077 3078 // Make sure all code is generated 3079 masm->flush(); 3080 3081 // Set exception blob 3082 _exception_blob = ExceptionBlob::create(&buffer, oop_maps, SimpleRuntimeFrame::framesize >> 1); 3083 } 3084 #endif // COMPILER2