1 /* 2 * Copyright (c) 2003, 2017, Oracle and/or its affiliates. All rights reserved. 3 * Copyright (c) 2014, 2015, 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 "interpreter/interpreter.hpp" 33 #include "interpreter/interp_masm.hpp" 34 #include "logging/log.hpp" 35 #include "memory/resourceArea.hpp" 36 #include "oops/compiledICHolder.hpp" 37 #include "runtime/sharedRuntime.hpp" 38 #include "runtime/vframeArray.hpp" 39 #include "vmreg_aarch64.inline.hpp" 40 #ifdef COMPILER1 41 #include "c1/c1_Runtime1.hpp" 42 #endif 43 #if defined(COMPILER2) || INCLUDE_JVMCI 44 #include "adfiles/ad_aarch64.hpp" 45 #include "opto/runtime.hpp" 46 #endif 47 #if INCLUDE_JVMCI 48 #include "jvmci/jvmciJavaClasses.hpp" 49 #endif 50 51 #ifdef BUILTIN_SIM 52 #include "../../../../../../simulator/simulator.hpp" 53 #endif 54 55 #define __ masm-> 56 57 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size; 58 59 class SimpleRuntimeFrame { 60 61 public: 62 63 // Most of the runtime stubs have this simple frame layout. 64 // This class exists to make the layout shared in one place. 65 // Offsets are for compiler stack slots, which are jints. 66 enum layout { 67 // The frame sender code expects that rbp will be in the "natural" place and 68 // will override any oopMap setting for it. We must therefore force the layout 69 // so that it agrees with the frame sender code. 70 // we don't expect any arg reg save area so aarch64 asserts that 71 // frame::arg_reg_save_area_bytes == 0 72 rbp_off = 0, 73 rbp_off2, 74 return_off, return_off2, 75 framesize 76 }; 77 }; 78 79 // FIXME -- this is used by C1 80 class RegisterSaver { 81 public: 82 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words, bool save_vectors = false); 83 static void restore_live_registers(MacroAssembler* masm, bool restore_vectors = false); 84 85 // Offsets into the register save area 86 // Used by deoptimization when it is managing result register 87 // values on its own 88 89 static int r0_offset_in_bytes(void) { return (32 + r0->encoding()) * wordSize; } 90 static int reg_offset_in_bytes(Register r) { return r0_offset_in_bytes() + r->encoding() * wordSize; } 91 static int rmethod_offset_in_bytes(void) { return reg_offset_in_bytes(rmethod); } 92 static int rscratch1_offset_in_bytes(void) { return (32 + rscratch1->encoding()) * wordSize; } 93 static int v0_offset_in_bytes(void) { return 0; } 94 static int return_offset_in_bytes(void) { return (32 /* floats*/ + 31 /* gregs*/) * wordSize; } 95 96 // During deoptimization only the result registers need to be restored, 97 // all the other values have already been extracted. 98 static void restore_result_registers(MacroAssembler* masm); 99 100 // Capture info about frame layout 101 enum layout { 102 fpu_state_off = 0, 103 fpu_state_end = fpu_state_off+FPUStateSizeInWords-1, 104 // The frame sender code expects that rfp will be in 105 // the "natural" place and will override any oopMap 106 // setting for it. We must therefore force the layout 107 // so that it agrees with the frame sender code. 108 r0_off = fpu_state_off+FPUStateSizeInWords, 109 rfp_off = r0_off + 30 * 2, 110 return_off = rfp_off + 2, // slot for return address 111 reg_save_size = return_off + 2}; 112 113 }; 114 115 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words, bool save_vectors) { 116 #if defined(COMPILER2) || INCLUDE_JVMCI 117 if (save_vectors) { 118 // Save upper half of vector registers 119 int vect_words = 32 * 8 / wordSize; 120 additional_frame_words += vect_words; 121 } 122 #else 123 assert(!save_vectors, "vectors are generated only by C2 and JVMCI"); 124 #endif 125 126 int frame_size_in_bytes = align_up(additional_frame_words*wordSize + 127 reg_save_size*BytesPerInt, 16); 128 // OopMap frame size is in compiler stack slots (jint's) not bytes or words 129 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt; 130 // The caller will allocate additional_frame_words 131 int additional_frame_slots = additional_frame_words*wordSize / BytesPerInt; 132 // CodeBlob frame size is in words. 133 int frame_size_in_words = frame_size_in_bytes / wordSize; 134 *total_frame_words = frame_size_in_words; 135 136 // Save registers, fpu state, and flags. 137 138 __ enter(); 139 __ push_CPU_state(save_vectors); 140 141 // Set an oopmap for the call site. This oopmap will map all 142 // oop-registers and debug-info registers as callee-saved. This 143 // will allow deoptimization at this safepoint to find all possible 144 // debug-info recordings, as well as let GC find all oops. 145 146 OopMapSet *oop_maps = new OopMapSet(); 147 OopMap* oop_map = new OopMap(frame_size_in_slots, 0); 148 149 for (int i = 0; i < RegisterImpl::number_of_registers; i++) { 150 Register r = as_Register(i); 151 if (r < rheapbase && r != rscratch1 && r != rscratch2) { 152 int sp_offset = 2 * (i + 32); // SP offsets are in 4-byte words, 153 // register slots are 8 bytes 154 // wide, 32 floating-point 155 // registers 156 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset + additional_frame_slots), 157 r->as_VMReg()); 158 } 159 } 160 161 for (int i = 0; i < FloatRegisterImpl::number_of_registers; i++) { 162 FloatRegister r = as_FloatRegister(i); 163 int sp_offset = save_vectors ? (4 * i) : (2 * i); 164 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset), 165 r->as_VMReg()); 166 } 167 168 return oop_map; 169 } 170 171 void RegisterSaver::restore_live_registers(MacroAssembler* masm, bool restore_vectors) { 172 #ifndef COMPILER2 173 assert(!restore_vectors, "vectors are generated only by C2 and JVMCI"); 174 #endif 175 __ pop_CPU_state(restore_vectors); 176 __ leave(); 177 } 178 179 void RegisterSaver::restore_result_registers(MacroAssembler* masm) { 180 181 // Just restore result register. Only used by deoptimization. By 182 // now any callee save register that needs to be restored to a c2 183 // caller of the deoptee has been extracted into the vframeArray 184 // and will be stuffed into the c2i adapter we create for later 185 // restoration so only result registers need to be restored here. 186 187 // Restore fp result register 188 __ ldrd(v0, Address(sp, v0_offset_in_bytes())); 189 // Restore integer result register 190 __ ldr(r0, Address(sp, r0_offset_in_bytes())); 191 192 // Pop all of the register save are off the stack 193 __ add(sp, sp, align_up(return_offset_in_bytes(), 16)); 194 } 195 196 // Is vector's size (in bytes) bigger than a size saved by default? 197 // 8 bytes vector registers are saved by default on AArch64. 198 bool SharedRuntime::is_wide_vector(int size) { 199 return size > 8; 200 } 201 202 size_t SharedRuntime::trampoline_size() { 203 return 16; 204 } 205 206 void SharedRuntime::generate_trampoline(MacroAssembler *masm, address destination) { 207 __ mov(rscratch1, destination); 208 __ br(rscratch1); 209 } 210 211 // The java_calling_convention describes stack locations as ideal slots on 212 // a frame with no abi restrictions. Since we must observe abi restrictions 213 // (like the placement of the register window) the slots must be biased by 214 // the following value. 215 static int reg2offset_in(VMReg r) { 216 // Account for saved rfp and lr 217 // This should really be in_preserve_stack_slots 218 return (r->reg2stack() + 4) * VMRegImpl::stack_slot_size; 219 } 220 221 static int reg2offset_out(VMReg r) { 222 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 223 } 224 225 // --------------------------------------------------------------------------- 226 // Read the array of BasicTypes from a signature, and compute where the 227 // arguments should go. Values in the VMRegPair regs array refer to 4-byte 228 // quantities. Values less than VMRegImpl::stack0 are registers, those above 229 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer 230 // as framesizes are fixed. 231 // VMRegImpl::stack0 refers to the first slot 0(sp). 232 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register 233 // up to RegisterImpl::number_of_registers) are the 64-bit 234 // integer registers. 235 236 // Note: the INPUTS in sig_bt are in units of Java argument words, 237 // which are 64-bit. The OUTPUTS are in 32-bit units. 238 239 // The Java calling convention is a "shifted" version of the C ABI. 240 // By skipping the first C ABI register we can call non-static jni 241 // methods with small numbers of arguments without having to shuffle 242 // the arguments at all. Since we control the java ABI we ought to at 243 // least get some advantage out of it. 244 245 int SharedRuntime::java_calling_convention(const BasicType *sig_bt, 246 VMRegPair *regs, 247 int total_args_passed, 248 int is_outgoing) { 249 250 // Create the mapping between argument positions and 251 // registers. 252 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = { 253 j_rarg0, j_rarg1, j_rarg2, j_rarg3, j_rarg4, j_rarg5, j_rarg6, j_rarg7 254 }; 255 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = { 256 j_farg0, j_farg1, j_farg2, j_farg3, 257 j_farg4, j_farg5, j_farg6, j_farg7 258 }; 259 260 261 uint int_args = 0; 262 uint fp_args = 0; 263 uint stk_args = 0; // inc by 2 each time 264 265 for (int i = 0; i < total_args_passed; i++) { 266 switch (sig_bt[i]) { 267 case T_BOOLEAN: 268 case T_CHAR: 269 case T_BYTE: 270 case T_SHORT: 271 case T_INT: 272 if (int_args < Argument::n_int_register_parameters_j) { 273 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 274 } else { 275 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 276 stk_args += 2; 277 } 278 break; 279 case T_VOID: 280 // halves of T_LONG or T_DOUBLE 281 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 282 regs[i].set_bad(); 283 break; 284 case T_LONG: 285 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 286 // fall through 287 case T_OBJECT: 288 case T_ARRAY: 289 case T_ADDRESS: 290 if (int_args < Argument::n_int_register_parameters_j) { 291 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg()); 292 } else { 293 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 294 stk_args += 2; 295 } 296 break; 297 case T_FLOAT: 298 if (fp_args < Argument::n_float_register_parameters_j) { 299 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg()); 300 } else { 301 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 302 stk_args += 2; 303 } 304 break; 305 case T_DOUBLE: 306 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 307 if (fp_args < Argument::n_float_register_parameters_j) { 308 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg()); 309 } else { 310 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 311 stk_args += 2; 312 } 313 break; 314 default: 315 ShouldNotReachHere(); 316 break; 317 } 318 } 319 320 return align_up(stk_args, 2); 321 } 322 323 // Patch the callers callsite with entry to compiled code if it exists. 324 static void patch_callers_callsite(MacroAssembler *masm) { 325 Label L; 326 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset()))); 327 __ cbz(rscratch1, L); 328 329 __ enter(); 330 __ push_CPU_state(); 331 332 // VM needs caller's callsite 333 // VM needs target method 334 // This needs to be a long call since we will relocate this adapter to 335 // the codeBuffer and it may not reach 336 337 #ifndef PRODUCT 338 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 339 #endif 340 341 __ mov(c_rarg0, rmethod); 342 __ mov(c_rarg1, lr); 343 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite))); 344 __ blrt(rscratch1, 2, 0, 0); 345 __ maybe_isb(); 346 347 __ pop_CPU_state(); 348 // restore sp 349 __ leave(); 350 __ bind(L); 351 } 352 353 static void gen_c2i_adapter(MacroAssembler *masm, 354 int total_args_passed, 355 int comp_args_on_stack, 356 const BasicType *sig_bt, 357 const VMRegPair *regs, 358 Label& skip_fixup) { 359 // Before we get into the guts of the C2I adapter, see if we should be here 360 // at all. We've come from compiled code and are attempting to jump to the 361 // interpreter, which means the caller made a static call to get here 362 // (vcalls always get a compiled target if there is one). Check for a 363 // compiled target. If there is one, we need to patch the caller's call. 364 patch_callers_callsite(masm); 365 366 __ bind(skip_fixup); 367 368 int words_pushed = 0; 369 370 // Since all args are passed on the stack, total_args_passed * 371 // Interpreter::stackElementSize is the space we need. 372 373 int extraspace = total_args_passed * Interpreter::stackElementSize; 374 375 __ mov(r13, sp); 376 377 // stack is aligned, keep it that way 378 extraspace = align_up(extraspace, 2*wordSize); 379 380 if (extraspace) 381 __ sub(sp, sp, extraspace); 382 383 // Now write the args into the outgoing interpreter space 384 for (int i = 0; i < total_args_passed; i++) { 385 if (sig_bt[i] == T_VOID) { 386 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 387 continue; 388 } 389 390 // offset to start parameters 391 int st_off = (total_args_passed - i - 1) * Interpreter::stackElementSize; 392 int next_off = st_off - Interpreter::stackElementSize; 393 394 // Say 4 args: 395 // i st_off 396 // 0 32 T_LONG 397 // 1 24 T_VOID 398 // 2 16 T_OBJECT 399 // 3 8 T_BOOL 400 // - 0 return address 401 // 402 // However to make thing extra confusing. Because we can fit a long/double in 403 // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter 404 // leaves one slot empty and only stores to a single slot. In this case the 405 // slot that is occupied is the T_VOID slot. See I said it was confusing. 406 407 VMReg r_1 = regs[i].first(); 408 VMReg r_2 = regs[i].second(); 409 if (!r_1->is_valid()) { 410 assert(!r_2->is_valid(), ""); 411 continue; 412 } 413 if (r_1->is_stack()) { 414 // memory to memory use rscratch1 415 int ld_off = (r_1->reg2stack() * VMRegImpl::stack_slot_size 416 + extraspace 417 + words_pushed * wordSize); 418 if (!r_2->is_valid()) { 419 // sign extend?? 420 __ ldrw(rscratch1, Address(sp, ld_off)); 421 __ str(rscratch1, Address(sp, st_off)); 422 423 } else { 424 425 __ ldr(rscratch1, Address(sp, ld_off)); 426 427 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG 428 // T_DOUBLE and T_LONG use two slots in the interpreter 429 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 430 // ld_off == LSW, ld_off+wordSize == MSW 431 // st_off == MSW, next_off == LSW 432 __ str(rscratch1, Address(sp, next_off)); 433 #ifdef ASSERT 434 // Overwrite the unused slot with known junk 435 __ mov(rscratch1, 0xdeadffffdeadaaaaul); 436 __ str(rscratch1, Address(sp, st_off)); 437 #endif /* ASSERT */ 438 } else { 439 __ str(rscratch1, Address(sp, st_off)); 440 } 441 } 442 } else if (r_1->is_Register()) { 443 Register r = r_1->as_Register(); 444 if (!r_2->is_valid()) { 445 // must be only an int (or less ) so move only 32bits to slot 446 // why not sign extend?? 447 __ str(r, Address(sp, st_off)); 448 } else { 449 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG 450 // T_DOUBLE and T_LONG use two slots in the interpreter 451 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 452 // long/double in gpr 453 #ifdef ASSERT 454 // Overwrite the unused slot with known junk 455 __ mov(rscratch1, 0xdeadffffdeadaaabul); 456 __ str(rscratch1, Address(sp, st_off)); 457 #endif /* ASSERT */ 458 __ str(r, Address(sp, next_off)); 459 } else { 460 __ str(r, Address(sp, st_off)); 461 } 462 } 463 } else { 464 assert(r_1->is_FloatRegister(), ""); 465 if (!r_2->is_valid()) { 466 // only a float use just part of the slot 467 __ strs(r_1->as_FloatRegister(), Address(sp, st_off)); 468 } else { 469 #ifdef ASSERT 470 // Overwrite the unused slot with known junk 471 __ mov(rscratch1, 0xdeadffffdeadaaacul); 472 __ str(rscratch1, Address(sp, st_off)); 473 #endif /* ASSERT */ 474 __ strd(r_1->as_FloatRegister(), Address(sp, next_off)); 475 } 476 } 477 } 478 479 __ mov(esp, sp); // Interp expects args on caller's expression stack 480 481 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::interpreter_entry_offset()))); 482 __ br(rscratch1); 483 } 484 485 486 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, 487 int total_args_passed, 488 int comp_args_on_stack, 489 const BasicType *sig_bt, 490 const VMRegPair *regs) { 491 492 // Note: r13 contains the senderSP on entry. We must preserve it since 493 // we may do a i2c -> c2i transition if we lose a race where compiled 494 // code goes non-entrant while we get args ready. 495 496 // In addition we use r13 to locate all the interpreter args because 497 // we must align the stack to 16 bytes. 498 499 // Adapters are frameless. 500 501 // An i2c adapter is frameless because the *caller* frame, which is 502 // interpreted, routinely repairs its own esp (from 503 // interpreter_frame_last_sp), even if a callee has modified the 504 // stack pointer. It also recalculates and aligns sp. 505 506 // A c2i adapter is frameless because the *callee* frame, which is 507 // interpreted, routinely repairs its caller's sp (from sender_sp, 508 // which is set up via the senderSP register). 509 510 // In other words, if *either* the caller or callee is interpreted, we can 511 // get the stack pointer repaired after a call. 512 513 // This is why c2i and i2c adapters cannot be indefinitely composed. 514 // In particular, if a c2i adapter were to somehow call an i2c adapter, 515 // both caller and callee would be compiled methods, and neither would 516 // clean up the stack pointer changes performed by the two adapters. 517 // If this happens, control eventually transfers back to the compiled 518 // caller, but with an uncorrected stack, causing delayed havoc. 519 520 if (VerifyAdapterCalls && 521 (Interpreter::code() != NULL || StubRoutines::code1() != NULL)) { 522 #if 0 523 // So, let's test for cascading c2i/i2c adapters right now. 524 // assert(Interpreter::contains($return_addr) || 525 // StubRoutines::contains($return_addr), 526 // "i2c adapter must return to an interpreter frame"); 527 __ block_comment("verify_i2c { "); 528 Label L_ok; 529 if (Interpreter::code() != NULL) 530 range_check(masm, rax, r11, 531 Interpreter::code()->code_start(), Interpreter::code()->code_end(), 532 L_ok); 533 if (StubRoutines::code1() != NULL) 534 range_check(masm, rax, r11, 535 StubRoutines::code1()->code_begin(), StubRoutines::code1()->code_end(), 536 L_ok); 537 if (StubRoutines::code2() != NULL) 538 range_check(masm, rax, r11, 539 StubRoutines::code2()->code_begin(), StubRoutines::code2()->code_end(), 540 L_ok); 541 const char* msg = "i2c adapter must return to an interpreter frame"; 542 __ block_comment(msg); 543 __ stop(msg); 544 __ bind(L_ok); 545 __ block_comment("} verify_i2ce "); 546 #endif 547 } 548 549 // Cut-out for having no stack args. 550 int comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord; 551 if (comp_args_on_stack) { 552 __ sub(rscratch1, sp, comp_words_on_stack * wordSize); 553 __ andr(sp, rscratch1, -16); 554 } 555 556 // Will jump to the compiled code just as if compiled code was doing it. 557 // Pre-load the register-jump target early, to schedule it better. 558 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::from_compiled_offset()))); 559 560 #if INCLUDE_JVMCI 561 if (EnableJVMCI) { 562 // check if this call should be routed towards a specific entry point 563 __ ldr(rscratch2, Address(rthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset()))); 564 Label no_alternative_target; 565 __ cbz(rscratch2, no_alternative_target); 566 __ mov(rscratch1, rscratch2); 567 __ str(zr, Address(rthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset()))); 568 __ bind(no_alternative_target); 569 } 570 #endif // INCLUDE_JVMCI 571 572 // Now generate the shuffle code. 573 for (int i = 0; i < total_args_passed; i++) { 574 if (sig_bt[i] == T_VOID) { 575 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 576 continue; 577 } 578 579 // Pick up 0, 1 or 2 words from SP+offset. 580 581 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), 582 "scrambled load targets?"); 583 // Load in argument order going down. 584 int ld_off = (total_args_passed - i - 1)*Interpreter::stackElementSize; 585 // Point to interpreter value (vs. tag) 586 int next_off = ld_off - Interpreter::stackElementSize; 587 // 588 // 589 // 590 VMReg r_1 = regs[i].first(); 591 VMReg r_2 = regs[i].second(); 592 if (!r_1->is_valid()) { 593 assert(!r_2->is_valid(), ""); 594 continue; 595 } 596 if (r_1->is_stack()) { 597 // Convert stack slot to an SP offset (+ wordSize to account for return address ) 598 int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size; 599 if (!r_2->is_valid()) { 600 // sign extend??? 601 __ ldrsw(rscratch2, Address(esp, ld_off)); 602 __ str(rscratch2, Address(sp, st_off)); 603 } else { 604 // 605 // We are using two optoregs. This can be either T_OBJECT, 606 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates 607 // two slots but only uses one for thr T_LONG or T_DOUBLE case 608 // So we must adjust where to pick up the data to match the 609 // interpreter. 610 // 611 // Interpreter local[n] == MSW, local[n+1] == LSW however locals 612 // are accessed as negative so LSW is at LOW address 613 614 // ld_off is MSW so get LSW 615 const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)? 616 next_off : ld_off; 617 __ ldr(rscratch2, Address(esp, offset)); 618 // st_off is LSW (i.e. reg.first()) 619 __ str(rscratch2, Address(sp, st_off)); 620 } 621 } else if (r_1->is_Register()) { // Register argument 622 Register r = r_1->as_Register(); 623 if (r_2->is_valid()) { 624 // 625 // We are using two VMRegs. This can be either T_OBJECT, 626 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates 627 // two slots but only uses one for thr T_LONG or T_DOUBLE case 628 // So we must adjust where to pick up the data to match the 629 // interpreter. 630 631 const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)? 632 next_off : ld_off; 633 634 // this can be a misaligned move 635 __ ldr(r, Address(esp, offset)); 636 } else { 637 // sign extend and use a full word? 638 __ ldrw(r, Address(esp, ld_off)); 639 } 640 } else { 641 if (!r_2->is_valid()) { 642 __ ldrs(r_1->as_FloatRegister(), Address(esp, ld_off)); 643 } else { 644 __ ldrd(r_1->as_FloatRegister(), Address(esp, next_off)); 645 } 646 } 647 } 648 649 // 6243940 We might end up in handle_wrong_method if 650 // the callee is deoptimized as we race thru here. If that 651 // happens we don't want to take a safepoint because the 652 // caller frame will look interpreted and arguments are now 653 // "compiled" so it is much better to make this transition 654 // invisible to the stack walking code. Unfortunately if 655 // we try and find the callee by normal means a safepoint 656 // is possible. So we stash the desired callee in the thread 657 // and the vm will find there should this case occur. 658 659 __ str(rmethod, Address(rthread, JavaThread::callee_target_offset())); 660 661 __ br(rscratch1); 662 } 663 664 #ifdef BUILTIN_SIM 665 static void generate_i2c_adapter_name(char *result, int total_args_passed, const BasicType *sig_bt) 666 { 667 strcpy(result, "i2c("); 668 int idx = 4; 669 for (int i = 0; i < total_args_passed; i++) { 670 switch(sig_bt[i]) { 671 case T_BOOLEAN: 672 result[idx++] = 'Z'; 673 break; 674 case T_CHAR: 675 result[idx++] = 'C'; 676 break; 677 case T_FLOAT: 678 result[idx++] = 'F'; 679 break; 680 case T_DOUBLE: 681 assert((i < (total_args_passed - 1)) && (sig_bt[i+1] == T_VOID), 682 "double must be followed by void"); 683 i++; 684 result[idx++] = 'D'; 685 break; 686 case T_BYTE: 687 result[idx++] = 'B'; 688 break; 689 case T_SHORT: 690 result[idx++] = 'S'; 691 break; 692 case T_INT: 693 result[idx++] = 'I'; 694 break; 695 case T_LONG: 696 assert((i < (total_args_passed - 1)) && (sig_bt[i+1] == T_VOID), 697 "long must be followed by void"); 698 i++; 699 result[idx++] = 'L'; 700 break; 701 case T_OBJECT: 702 result[idx++] = 'O'; 703 break; 704 case T_ARRAY: 705 result[idx++] = '['; 706 break; 707 case T_ADDRESS: 708 result[idx++] = 'P'; 709 break; 710 case T_NARROWOOP: 711 result[idx++] = 'N'; 712 break; 713 case T_METADATA: 714 result[idx++] = 'M'; 715 break; 716 case T_NARROWKLASS: 717 result[idx++] = 'K'; 718 break; 719 default: 720 result[idx++] = '?'; 721 break; 722 } 723 } 724 result[idx++] = ')'; 725 result[idx] = '\0'; 726 } 727 #endif 728 729 // --------------------------------------------------------------- 730 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, 731 int total_args_passed, 732 int comp_args_on_stack, 733 const BasicType *sig_bt, 734 const VMRegPair *regs, 735 AdapterFingerPrint* fingerprint) { 736 address i2c_entry = __ pc(); 737 #ifdef BUILTIN_SIM 738 char *name = NULL; 739 AArch64Simulator *sim = NULL; 740 size_t len = 65536; 741 if (NotifySimulator) { 742 name = NEW_C_HEAP_ARRAY(char, len, mtInternal); 743 } 744 745 if (name) { 746 generate_i2c_adapter_name(name, total_args_passed, sig_bt); 747 sim = AArch64Simulator::get_current(UseSimulatorCache, DisableBCCheck); 748 sim->notifyCompile(name, i2c_entry); 749 } 750 #endif 751 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs); 752 753 address c2i_unverified_entry = __ pc(); 754 Label skip_fixup; 755 756 Label ok; 757 758 Register holder = rscratch2; 759 Register receiver = j_rarg0; 760 Register tmp = r10; // A call-clobbered register not used for arg passing 761 762 // ------------------------------------------------------------------------- 763 // Generate a C2I adapter. On entry we know rmethod holds the Method* during calls 764 // to the interpreter. The args start out packed in the compiled layout. They 765 // need to be unpacked into the interpreter layout. This will almost always 766 // require some stack space. We grow the current (compiled) stack, then repack 767 // the args. We finally end in a jump to the generic interpreter entry point. 768 // On exit from the interpreter, the interpreter will restore our SP (lest the 769 // compiled code, which relys solely on SP and not FP, get sick). 770 771 { 772 __ block_comment("c2i_unverified_entry {"); 773 __ load_klass(rscratch1, receiver); 774 __ ldr(tmp, Address(holder, CompiledICHolder::holder_klass_offset())); 775 __ cmp(rscratch1, tmp); 776 __ ldr(rmethod, Address(holder, CompiledICHolder::holder_method_offset())); 777 __ br(Assembler::EQ, ok); 778 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 779 780 __ bind(ok); 781 // Method might have been compiled since the call site was patched to 782 // interpreted; if that is the case treat it as a miss so we can get 783 // the call site corrected. 784 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset()))); 785 __ cbz(rscratch1, skip_fixup); 786 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 787 __ block_comment("} c2i_unverified_entry"); 788 } 789 790 address c2i_entry = __ pc(); 791 792 #ifdef BUILTIN_SIM 793 if (name) { 794 name[0] = 'c'; 795 name[2] = 'i'; 796 sim->notifyCompile(name, c2i_entry); 797 FREE_C_HEAP_ARRAY(char, name, mtInternal); 798 } 799 #endif 800 801 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup); 802 803 __ flush(); 804 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry); 805 } 806 807 int SharedRuntime::c_calling_convention(const BasicType *sig_bt, 808 VMRegPair *regs, 809 VMRegPair *regs2, 810 int total_args_passed) { 811 assert(regs2 == NULL, "not needed on AArch64"); 812 813 // We return the amount of VMRegImpl stack slots we need to reserve for all 814 // the arguments NOT counting out_preserve_stack_slots. 815 816 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = { 817 c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5, c_rarg6, c_rarg7 818 }; 819 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = { 820 c_farg0, c_farg1, c_farg2, c_farg3, 821 c_farg4, c_farg5, c_farg6, c_farg7 822 }; 823 824 uint int_args = 0; 825 uint fp_args = 0; 826 uint stk_args = 0; // inc by 2 each time 827 828 for (int i = 0; i < total_args_passed; i++) { 829 switch (sig_bt[i]) { 830 case T_BOOLEAN: 831 case T_CHAR: 832 case T_BYTE: 833 case T_SHORT: 834 case T_INT: 835 if (int_args < Argument::n_int_register_parameters_c) { 836 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 837 } else { 838 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 839 stk_args += 2; 840 } 841 break; 842 case T_LONG: 843 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 844 // fall through 845 case T_OBJECT: 846 case T_ARRAY: 847 case T_ADDRESS: 848 case T_METADATA: 849 if (int_args < Argument::n_int_register_parameters_c) { 850 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg()); 851 } else { 852 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 853 stk_args += 2; 854 } 855 break; 856 case T_FLOAT: 857 if (fp_args < Argument::n_float_register_parameters_c) { 858 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg()); 859 } else { 860 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 861 stk_args += 2; 862 } 863 break; 864 case T_DOUBLE: 865 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 866 if (fp_args < Argument::n_float_register_parameters_c) { 867 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg()); 868 } else { 869 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 870 stk_args += 2; 871 } 872 break; 873 case T_VOID: // Halves of longs and doubles 874 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 875 regs[i].set_bad(); 876 break; 877 default: 878 ShouldNotReachHere(); 879 break; 880 } 881 } 882 883 return stk_args; 884 } 885 886 // On 64 bit we will store integer like items to the stack as 887 // 64 bits items (sparc abi) even though java would only store 888 // 32bits for a parameter. On 32bit it will simply be 32 bits 889 // So this routine will do 32->32 on 32bit and 32->64 on 64bit 890 static void move32_64(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 891 if (src.first()->is_stack()) { 892 if (dst.first()->is_stack()) { 893 // stack to stack 894 __ ldr(rscratch1, Address(rfp, reg2offset_in(src.first()))); 895 __ str(rscratch1, Address(sp, reg2offset_out(dst.first()))); 896 } else { 897 // stack to reg 898 __ ldrsw(dst.first()->as_Register(), Address(rfp, reg2offset_in(src.first()))); 899 } 900 } else if (dst.first()->is_stack()) { 901 // reg to stack 902 // Do we really have to sign extend??? 903 // __ movslq(src.first()->as_Register(), src.first()->as_Register()); 904 __ str(src.first()->as_Register(), Address(sp, reg2offset_out(dst.first()))); 905 } else { 906 if (dst.first() != src.first()) { 907 __ sxtw(dst.first()->as_Register(), src.first()->as_Register()); 908 } 909 } 910 } 911 912 // An oop arg. Must pass a handle not the oop itself 913 static void object_move(MacroAssembler* masm, 914 OopMap* map, 915 int oop_handle_offset, 916 int framesize_in_slots, 917 VMRegPair src, 918 VMRegPair dst, 919 bool is_receiver, 920 int* receiver_offset) { 921 922 // must pass a handle. First figure out the location we use as a handle 923 924 Register rHandle = dst.first()->is_stack() ? rscratch2 : dst.first()->as_Register(); 925 926 // See if oop is NULL if it is we need no handle 927 928 if (src.first()->is_stack()) { 929 930 // Oop is already on the stack as an argument 931 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 932 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots)); 933 if (is_receiver) { 934 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size; 935 } 936 937 __ ldr(rscratch1, Address(rfp, reg2offset_in(src.first()))); 938 __ lea(rHandle, Address(rfp, reg2offset_in(src.first()))); 939 // conditionally move a NULL 940 __ cmp(rscratch1, zr); 941 __ csel(rHandle, zr, rHandle, Assembler::EQ); 942 } else { 943 944 // Oop is in an a register we must store it to the space we reserve 945 // on the stack for oop_handles and pass a handle if oop is non-NULL 946 947 const Register rOop = src.first()->as_Register(); 948 int oop_slot; 949 if (rOop == j_rarg0) 950 oop_slot = 0; 951 else if (rOop == j_rarg1) 952 oop_slot = 1; 953 else if (rOop == j_rarg2) 954 oop_slot = 2; 955 else if (rOop == j_rarg3) 956 oop_slot = 3; 957 else if (rOop == j_rarg4) 958 oop_slot = 4; 959 else if (rOop == j_rarg5) 960 oop_slot = 5; 961 else if (rOop == j_rarg6) 962 oop_slot = 6; 963 else { 964 assert(rOop == j_rarg7, "wrong register"); 965 oop_slot = 7; 966 } 967 968 oop_slot = oop_slot * VMRegImpl::slots_per_word + oop_handle_offset; 969 int offset = oop_slot*VMRegImpl::stack_slot_size; 970 971 map->set_oop(VMRegImpl::stack2reg(oop_slot)); 972 // Store oop in handle area, may be NULL 973 __ str(rOop, Address(sp, offset)); 974 if (is_receiver) { 975 *receiver_offset = offset; 976 } 977 978 __ cmp(rOop, zr); 979 __ lea(rHandle, Address(sp, offset)); 980 // conditionally move a NULL 981 __ csel(rHandle, zr, rHandle, Assembler::EQ); 982 } 983 984 // If arg is on the stack then place it otherwise it is already in correct reg. 985 if (dst.first()->is_stack()) { 986 __ str(rHandle, Address(sp, reg2offset_out(dst.first()))); 987 } 988 } 989 990 // A float arg may have to do float reg int reg conversion 991 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 992 assert(src.first()->is_stack() && dst.first()->is_stack() || 993 src.first()->is_reg() && dst.first()->is_reg(), "Unexpected error"); 994 if (src.first()->is_stack()) { 995 if (dst.first()->is_stack()) { 996 __ ldrw(rscratch1, Address(rfp, reg2offset_in(src.first()))); 997 __ strw(rscratch1, Address(sp, reg2offset_out(dst.first()))); 998 } else { 999 ShouldNotReachHere(); 1000 } 1001 } else if (src.first() != dst.first()) { 1002 if (src.is_single_phys_reg() && dst.is_single_phys_reg()) 1003 __ fmovs(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 1004 else 1005 ShouldNotReachHere(); 1006 } 1007 } 1008 1009 // A long move 1010 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1011 if (src.first()->is_stack()) { 1012 if (dst.first()->is_stack()) { 1013 // stack to stack 1014 __ ldr(rscratch1, Address(rfp, reg2offset_in(src.first()))); 1015 __ str(rscratch1, Address(sp, reg2offset_out(dst.first()))); 1016 } else { 1017 // stack to reg 1018 __ ldr(dst.first()->as_Register(), Address(rfp, reg2offset_in(src.first()))); 1019 } 1020 } else if (dst.first()->is_stack()) { 1021 // reg to stack 1022 // Do we really have to sign extend??? 1023 // __ movslq(src.first()->as_Register(), src.first()->as_Register()); 1024 __ str(src.first()->as_Register(), Address(sp, reg2offset_out(dst.first()))); 1025 } else { 1026 if (dst.first() != src.first()) { 1027 __ mov(dst.first()->as_Register(), src.first()->as_Register()); 1028 } 1029 } 1030 } 1031 1032 1033 // A double move 1034 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1035 assert(src.first()->is_stack() && dst.first()->is_stack() || 1036 src.first()->is_reg() && dst.first()->is_reg(), "Unexpected error"); 1037 if (src.first()->is_stack()) { 1038 if (dst.first()->is_stack()) { 1039 __ ldr(rscratch1, Address(rfp, reg2offset_in(src.first()))); 1040 __ str(rscratch1, Address(sp, reg2offset_out(dst.first()))); 1041 } else { 1042 ShouldNotReachHere(); 1043 } 1044 } else if (src.first() != dst.first()) { 1045 if (src.is_single_phys_reg() && dst.is_single_phys_reg()) 1046 __ fmovd(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 1047 else 1048 ShouldNotReachHere(); 1049 } 1050 } 1051 1052 1053 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1054 // We always ignore the frame_slots arg and just use the space just below frame pointer 1055 // which by this time is free to use 1056 switch (ret_type) { 1057 case T_FLOAT: 1058 __ strs(v0, Address(rfp, -wordSize)); 1059 break; 1060 case T_DOUBLE: 1061 __ strd(v0, Address(rfp, -wordSize)); 1062 break; 1063 case T_VOID: break; 1064 default: { 1065 __ str(r0, Address(rfp, -wordSize)); 1066 } 1067 } 1068 } 1069 1070 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1071 // We always ignore the frame_slots arg and just use the space just below frame pointer 1072 // which by this time is free to use 1073 switch (ret_type) { 1074 case T_FLOAT: 1075 __ ldrs(v0, Address(rfp, -wordSize)); 1076 break; 1077 case T_DOUBLE: 1078 __ ldrd(v0, Address(rfp, -wordSize)); 1079 break; 1080 case T_VOID: break; 1081 default: { 1082 __ ldr(r0, Address(rfp, -wordSize)); 1083 } 1084 } 1085 } 1086 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) { 1087 RegSet x; 1088 for ( int i = first_arg ; i < arg_count ; i++ ) { 1089 if (args[i].first()->is_Register()) { 1090 x = x + args[i].first()->as_Register(); 1091 } else if (args[i].first()->is_FloatRegister()) { 1092 __ strd(args[i].first()->as_FloatRegister(), Address(__ pre(sp, -2 * wordSize))); 1093 } 1094 } 1095 __ push(x, sp); 1096 } 1097 1098 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) { 1099 RegSet x; 1100 for ( int i = first_arg ; i < arg_count ; i++ ) { 1101 if (args[i].first()->is_Register()) { 1102 x = x + args[i].first()->as_Register(); 1103 } else { 1104 ; 1105 } 1106 } 1107 __ pop(x, sp); 1108 for ( int i = first_arg ; i < arg_count ; i++ ) { 1109 if (args[i].first()->is_Register()) { 1110 ; 1111 } else if (args[i].first()->is_FloatRegister()) { 1112 __ ldrd(args[i].first()->as_FloatRegister(), Address(__ post(sp, 2 * wordSize))); 1113 } 1114 } 1115 } 1116 1117 1118 // Check GCLocker::needs_gc and enter the runtime if it's true. This 1119 // keeps a new JNI critical region from starting until a GC has been 1120 // forced. Save down any oops in registers and describe them in an 1121 // OopMap. 1122 static void check_needs_gc_for_critical_native(MacroAssembler* masm, 1123 int stack_slots, 1124 int total_c_args, 1125 int total_in_args, 1126 int arg_save_area, 1127 OopMapSet* oop_maps, 1128 VMRegPair* in_regs, 1129 BasicType* in_sig_bt) { Unimplemented(); } 1130 1131 // Unpack an array argument into a pointer to the body and the length 1132 // if the array is non-null, otherwise pass 0 for both. 1133 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, VMRegPair body_arg, VMRegPair length_arg) { Unimplemented(); } 1134 1135 1136 class ComputeMoveOrder: public StackObj { 1137 class MoveOperation: public ResourceObj { 1138 friend class ComputeMoveOrder; 1139 private: 1140 VMRegPair _src; 1141 VMRegPair _dst; 1142 int _src_index; 1143 int _dst_index; 1144 bool _processed; 1145 MoveOperation* _next; 1146 MoveOperation* _prev; 1147 1148 static int get_id(VMRegPair r) { Unimplemented(); return 0; } 1149 1150 public: 1151 MoveOperation(int src_index, VMRegPair src, int dst_index, VMRegPair dst): 1152 _src(src) 1153 , _src_index(src_index) 1154 , _dst(dst) 1155 , _dst_index(dst_index) 1156 , _next(NULL) 1157 , _prev(NULL) 1158 , _processed(false) { Unimplemented(); } 1159 1160 VMRegPair src() const { Unimplemented(); return _src; } 1161 int src_id() const { Unimplemented(); return 0; } 1162 int src_index() const { Unimplemented(); return 0; } 1163 VMRegPair dst() const { Unimplemented(); return _src; } 1164 void set_dst(int i, VMRegPair dst) { Unimplemented(); } 1165 int dst_index() const { Unimplemented(); return 0; } 1166 int dst_id() const { Unimplemented(); return 0; } 1167 MoveOperation* next() const { Unimplemented(); return 0; } 1168 MoveOperation* prev() const { Unimplemented(); return 0; } 1169 void set_processed() { Unimplemented(); } 1170 bool is_processed() const { Unimplemented(); return 0; } 1171 1172 // insert 1173 void break_cycle(VMRegPair temp_register) { Unimplemented(); } 1174 1175 void link(GrowableArray<MoveOperation*>& killer) { Unimplemented(); } 1176 }; 1177 1178 private: 1179 GrowableArray<MoveOperation*> edges; 1180 1181 public: 1182 ComputeMoveOrder(int total_in_args, VMRegPair* in_regs, int total_c_args, VMRegPair* out_regs, 1183 BasicType* in_sig_bt, GrowableArray<int>& arg_order, VMRegPair tmp_vmreg) { Unimplemented(); } 1184 1185 // Collected all the move operations 1186 void add_edge(int src_index, VMRegPair src, int dst_index, VMRegPair dst) { Unimplemented(); } 1187 1188 // Walk the edges breaking cycles between moves. The result list 1189 // can be walked in order to produce the proper set of loads 1190 GrowableArray<MoveOperation*>* get_store_order(VMRegPair temp_register) { Unimplemented(); return 0; } 1191 }; 1192 1193 1194 static void rt_call(MacroAssembler* masm, address dest, int gpargs, int fpargs, int type) { 1195 CodeBlob *cb = CodeCache::find_blob(dest); 1196 if (cb) { 1197 __ far_call(RuntimeAddress(dest)); 1198 } else { 1199 assert((unsigned)gpargs < 256, "eek!"); 1200 assert((unsigned)fpargs < 32, "eek!"); 1201 __ lea(rscratch1, RuntimeAddress(dest)); 1202 if (UseBuiltinSim) __ mov(rscratch2, (gpargs << 6) | (fpargs << 2) | type); 1203 __ blrt(rscratch1, rscratch2); 1204 __ maybe_isb(); 1205 } 1206 } 1207 1208 static void verify_oop_args(MacroAssembler* masm, 1209 methodHandle method, 1210 const BasicType* sig_bt, 1211 const VMRegPair* regs) { 1212 Register temp_reg = r19; // not part of any compiled calling seq 1213 if (VerifyOops) { 1214 for (int i = 0; i < method->size_of_parameters(); i++) { 1215 if (sig_bt[i] == T_OBJECT || 1216 sig_bt[i] == T_ARRAY) { 1217 VMReg r = regs[i].first(); 1218 assert(r->is_valid(), "bad oop arg"); 1219 if (r->is_stack()) { 1220 __ ldr(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size)); 1221 __ verify_oop(temp_reg); 1222 } else { 1223 __ verify_oop(r->as_Register()); 1224 } 1225 } 1226 } 1227 } 1228 } 1229 1230 static void gen_special_dispatch(MacroAssembler* masm, 1231 methodHandle method, 1232 const BasicType* sig_bt, 1233 const VMRegPair* regs) { 1234 verify_oop_args(masm, method, sig_bt, regs); 1235 vmIntrinsics::ID iid = method->intrinsic_id(); 1236 1237 // Now write the args into the outgoing interpreter space 1238 bool has_receiver = false; 1239 Register receiver_reg = noreg; 1240 int member_arg_pos = -1; 1241 Register member_reg = noreg; 1242 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid); 1243 if (ref_kind != 0) { 1244 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument 1245 member_reg = r19; // known to be free at this point 1246 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind); 1247 } else if (iid == vmIntrinsics::_invokeBasic) { 1248 has_receiver = true; 1249 } else { 1250 fatal("unexpected intrinsic id %d", iid); 1251 } 1252 1253 if (member_reg != noreg) { 1254 // Load the member_arg into register, if necessary. 1255 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs); 1256 VMReg r = regs[member_arg_pos].first(); 1257 if (r->is_stack()) { 1258 __ ldr(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size)); 1259 } else { 1260 // no data motion is needed 1261 member_reg = r->as_Register(); 1262 } 1263 } 1264 1265 if (has_receiver) { 1266 // Make sure the receiver is loaded into a register. 1267 assert(method->size_of_parameters() > 0, "oob"); 1268 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object"); 1269 VMReg r = regs[0].first(); 1270 assert(r->is_valid(), "bad receiver arg"); 1271 if (r->is_stack()) { 1272 // Porting note: This assumes that compiled calling conventions always 1273 // pass the receiver oop in a register. If this is not true on some 1274 // platform, pick a temp and load the receiver from stack. 1275 fatal("receiver always in a register"); 1276 receiver_reg = r2; // known to be free at this point 1277 __ ldr(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size)); 1278 } else { 1279 // no data motion is needed 1280 receiver_reg = r->as_Register(); 1281 } 1282 } 1283 1284 // Figure out which address we are really jumping to: 1285 MethodHandles::generate_method_handle_dispatch(masm, iid, 1286 receiver_reg, member_reg, /*for_compiler_entry:*/ true); 1287 } 1288 1289 // --------------------------------------------------------------------------- 1290 // Generate a native wrapper for a given method. The method takes arguments 1291 // in the Java compiled code convention, marshals them to the native 1292 // convention (handlizes oops, etc), transitions to native, makes the call, 1293 // returns to java state (possibly blocking), unhandlizes any result and 1294 // returns. 1295 // 1296 // Critical native functions are a shorthand for the use of 1297 // GetPrimtiveArrayCritical and disallow the use of any other JNI 1298 // functions. The wrapper is expected to unpack the arguments before 1299 // passing them to the callee and perform checks before and after the 1300 // native call to ensure that they GCLocker 1301 // lock_critical/unlock_critical semantics are followed. Some other 1302 // parts of JNI setup are skipped like the tear down of the JNI handle 1303 // block and the check for pending exceptions it's impossible for them 1304 // to be thrown. 1305 // 1306 // They are roughly structured like this: 1307 // if (GCLocker::needs_gc()) 1308 // SharedRuntime::block_for_jni_critical(); 1309 // tranistion to thread_in_native 1310 // unpack arrray arguments and call native entry point 1311 // check for safepoint in progress 1312 // check if any thread suspend flags are set 1313 // call into JVM and possible unlock the JNI critical 1314 // if a GC was suppressed while in the critical native. 1315 // transition back to thread_in_Java 1316 // return to caller 1317 // 1318 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm, 1319 const methodHandle& method, 1320 int compile_id, 1321 BasicType* in_sig_bt, 1322 VMRegPair* in_regs, 1323 BasicType ret_type) { 1324 #ifdef BUILTIN_SIM 1325 if (NotifySimulator) { 1326 // Names are up to 65536 chars long. UTF8-coded strings are up to 1327 // 3 bytes per character. We concatenate three such strings. 1328 // Yes, I know this is ridiculous, but it's debug code and glibc 1329 // allocates large arrays very efficiently. 1330 size_t len = (65536 * 3) * 3; 1331 char *name = new char[len]; 1332 1333 strncpy(name, method()->method_holder()->name()->as_utf8(), len); 1334 strncat(name, ".", len); 1335 strncat(name, method()->name()->as_utf8(), len); 1336 strncat(name, method()->signature()->as_utf8(), len); 1337 AArch64Simulator::get_current(UseSimulatorCache, DisableBCCheck)->notifyCompile(name, __ pc()); 1338 delete[] name; 1339 } 1340 #endif 1341 1342 if (method->is_method_handle_intrinsic()) { 1343 vmIntrinsics::ID iid = method->intrinsic_id(); 1344 intptr_t start = (intptr_t)__ pc(); 1345 int vep_offset = ((intptr_t)__ pc()) - start; 1346 1347 // First instruction must be a nop as it may need to be patched on deoptimisation 1348 __ nop(); 1349 gen_special_dispatch(masm, 1350 method, 1351 in_sig_bt, 1352 in_regs); 1353 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period 1354 __ flush(); 1355 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually 1356 return nmethod::new_native_nmethod(method, 1357 compile_id, 1358 masm->code(), 1359 vep_offset, 1360 frame_complete, 1361 stack_slots / VMRegImpl::slots_per_word, 1362 in_ByteSize(-1), 1363 in_ByteSize(-1), 1364 (OopMapSet*)NULL); 1365 } 1366 bool is_critical_native = true; 1367 address native_func = method->critical_native_function(); 1368 if (native_func == NULL) { 1369 native_func = method->native_function(); 1370 is_critical_native = false; 1371 } 1372 assert(native_func != NULL, "must have function"); 1373 1374 // An OopMap for lock (and class if static) 1375 OopMapSet *oop_maps = new OopMapSet(); 1376 intptr_t start = (intptr_t)__ pc(); 1377 1378 // We have received a description of where all the java arg are located 1379 // on entry to the wrapper. We need to convert these args to where 1380 // the jni function will expect them. To figure out where they go 1381 // we convert the java signature to a C signature by inserting 1382 // the hidden arguments as arg[0] and possibly arg[1] (static method) 1383 1384 const int total_in_args = method->size_of_parameters(); 1385 int total_c_args = total_in_args; 1386 if (!is_critical_native) { 1387 total_c_args += 1; 1388 if (method->is_static()) { 1389 total_c_args++; 1390 } 1391 } else { 1392 for (int i = 0; i < total_in_args; i++) { 1393 if (in_sig_bt[i] == T_ARRAY) { 1394 total_c_args++; 1395 } 1396 } 1397 } 1398 1399 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1400 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1401 BasicType* in_elem_bt = NULL; 1402 1403 int argc = 0; 1404 if (!is_critical_native) { 1405 out_sig_bt[argc++] = T_ADDRESS; 1406 if (method->is_static()) { 1407 out_sig_bt[argc++] = T_OBJECT; 1408 } 1409 1410 for (int i = 0; i < total_in_args ; i++ ) { 1411 out_sig_bt[argc++] = in_sig_bt[i]; 1412 } 1413 } else { 1414 Thread* THREAD = Thread::current(); 1415 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, total_in_args); 1416 SignatureStream ss(method->signature()); 1417 for (int i = 0; i < total_in_args ; i++ ) { 1418 if (in_sig_bt[i] == T_ARRAY) { 1419 // Arrays are passed as int, elem* pair 1420 out_sig_bt[argc++] = T_INT; 1421 out_sig_bt[argc++] = T_ADDRESS; 1422 Symbol* atype = ss.as_symbol(CHECK_NULL); 1423 const char* at = atype->as_C_string(); 1424 if (strlen(at) == 2) { 1425 assert(at[0] == '[', "must be"); 1426 switch (at[1]) { 1427 case 'B': in_elem_bt[i] = T_BYTE; break; 1428 case 'C': in_elem_bt[i] = T_CHAR; break; 1429 case 'D': in_elem_bt[i] = T_DOUBLE; break; 1430 case 'F': in_elem_bt[i] = T_FLOAT; break; 1431 case 'I': in_elem_bt[i] = T_INT; break; 1432 case 'J': in_elem_bt[i] = T_LONG; break; 1433 case 'S': in_elem_bt[i] = T_SHORT; break; 1434 case 'Z': in_elem_bt[i] = T_BOOLEAN; break; 1435 default: ShouldNotReachHere(); 1436 } 1437 } 1438 } else { 1439 out_sig_bt[argc++] = in_sig_bt[i]; 1440 in_elem_bt[i] = T_VOID; 1441 } 1442 if (in_sig_bt[i] != T_VOID) { 1443 assert(in_sig_bt[i] == ss.type(), "must match"); 1444 ss.next(); 1445 } 1446 } 1447 } 1448 1449 // Now figure out where the args must be stored and how much stack space 1450 // they require. 1451 int out_arg_slots; 1452 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, NULL, total_c_args); 1453 1454 // Compute framesize for the wrapper. We need to handlize all oops in 1455 // incoming registers 1456 1457 // Calculate the total number of stack slots we will need. 1458 1459 // First count the abi requirement plus all of the outgoing args 1460 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots; 1461 1462 // Now the space for the inbound oop handle area 1463 int total_save_slots = 8 * VMRegImpl::slots_per_word; // 8 arguments passed in registers 1464 if (is_critical_native) { 1465 // Critical natives may have to call out so they need a save area 1466 // for register arguments. 1467 int double_slots = 0; 1468 int single_slots = 0; 1469 for ( int i = 0; i < total_in_args; i++) { 1470 if (in_regs[i].first()->is_Register()) { 1471 const Register reg = in_regs[i].first()->as_Register(); 1472 switch (in_sig_bt[i]) { 1473 case T_BOOLEAN: 1474 case T_BYTE: 1475 case T_SHORT: 1476 case T_CHAR: 1477 case T_INT: single_slots++; break; 1478 case T_ARRAY: // specific to LP64 (7145024) 1479 case T_LONG: double_slots++; break; 1480 default: ShouldNotReachHere(); 1481 } 1482 } else if (in_regs[i].first()->is_FloatRegister()) { 1483 ShouldNotReachHere(); 1484 } 1485 } 1486 total_save_slots = double_slots * 2 + single_slots; 1487 // align the save area 1488 if (double_slots != 0) { 1489 stack_slots = align_up(stack_slots, 2); 1490 } 1491 } 1492 1493 int oop_handle_offset = stack_slots; 1494 stack_slots += total_save_slots; 1495 1496 // Now any space we need for handlizing a klass if static method 1497 1498 int klass_slot_offset = 0; 1499 int klass_offset = -1; 1500 int lock_slot_offset = 0; 1501 bool is_static = false; 1502 1503 if (method->is_static()) { 1504 klass_slot_offset = stack_slots; 1505 stack_slots += VMRegImpl::slots_per_word; 1506 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; 1507 is_static = true; 1508 } 1509 1510 // Plus a lock if needed 1511 1512 if (method->is_synchronized()) { 1513 lock_slot_offset = stack_slots; 1514 stack_slots += VMRegImpl::slots_per_word; 1515 } 1516 1517 // Now a place (+2) to save return values or temp during shuffling 1518 // + 4 for return address (which we own) and saved rfp 1519 stack_slots += 6; 1520 1521 // Ok The space we have allocated will look like: 1522 // 1523 // 1524 // FP-> | | 1525 // |---------------------| 1526 // | 2 slots for moves | 1527 // |---------------------| 1528 // | lock box (if sync) | 1529 // |---------------------| <- lock_slot_offset 1530 // | klass (if static) | 1531 // |---------------------| <- klass_slot_offset 1532 // | oopHandle area | 1533 // |---------------------| <- oop_handle_offset (8 java arg registers) 1534 // | outbound memory | 1535 // | based arguments | 1536 // | | 1537 // |---------------------| 1538 // | | 1539 // SP-> | out_preserved_slots | 1540 // 1541 // 1542 1543 1544 // Now compute actual number of stack words we need rounding to make 1545 // stack properly aligned. 1546 stack_slots = align_up(stack_slots, StackAlignmentInSlots); 1547 1548 int stack_size = stack_slots * VMRegImpl::stack_slot_size; 1549 1550 // First thing make an ic check to see if we should even be here 1551 1552 // We are free to use all registers as temps without saving them and 1553 // restoring them except rfp. rfp is the only callee save register 1554 // as far as the interpreter and the compiler(s) are concerned. 1555 1556 1557 const Register ic_reg = rscratch2; 1558 const Register receiver = j_rarg0; 1559 1560 Label hit; 1561 Label exception_pending; 1562 1563 assert_different_registers(ic_reg, receiver, rscratch1); 1564 __ verify_oop(receiver); 1565 __ cmp_klass(receiver, ic_reg, rscratch1); 1566 __ br(Assembler::EQ, hit); 1567 1568 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 1569 1570 // Verified entry point must be aligned 1571 __ align(8); 1572 1573 __ bind(hit); 1574 1575 int vep_offset = ((intptr_t)__ pc()) - start; 1576 1577 // If we have to make this method not-entrant we'll overwrite its 1578 // first instruction with a jump. For this action to be legal we 1579 // must ensure that this first instruction is a B, BL, NOP, BKPT, 1580 // SVC, HVC, or SMC. Make it a NOP. 1581 __ nop(); 1582 1583 // Generate stack overflow check 1584 if (UseStackBanging) { 1585 __ bang_stack_with_offset(JavaThread::stack_shadow_zone_size()); 1586 } else { 1587 Unimplemented(); 1588 } 1589 1590 // Generate a new frame for the wrapper. 1591 __ enter(); 1592 // -2 because return address is already present and so is saved rfp 1593 __ sub(sp, sp, stack_size - 2*wordSize); 1594 1595 // Frame is now completed as far as size and linkage. 1596 int frame_complete = ((intptr_t)__ pc()) - start; 1597 1598 // record entry into native wrapper code 1599 if (NotifySimulator) { 1600 __ notify(Assembler::method_entry); 1601 } 1602 1603 // We use r20 as the oop handle for the receiver/klass 1604 // It is callee save so it survives the call to native 1605 1606 const Register oop_handle_reg = r20; 1607 1608 if (is_critical_native) { 1609 check_needs_gc_for_critical_native(masm, stack_slots, total_c_args, total_in_args, 1610 oop_handle_offset, oop_maps, in_regs, in_sig_bt); 1611 } 1612 1613 // 1614 // We immediately shuffle the arguments so that any vm call we have to 1615 // make from here on out (sync slow path, jvmti, etc.) we will have 1616 // captured the oops from our caller and have a valid oopMap for 1617 // them. 1618 1619 // ----------------- 1620 // The Grand Shuffle 1621 1622 // The Java calling convention is either equal (linux) or denser (win64) than the 1623 // c calling convention. However the because of the jni_env argument the c calling 1624 // convention always has at least one more (and two for static) arguments than Java. 1625 // Therefore if we move the args from java -> c backwards then we will never have 1626 // a register->register conflict and we don't have to build a dependency graph 1627 // and figure out how to break any cycles. 1628 // 1629 1630 // Record esp-based slot for receiver on stack for non-static methods 1631 int receiver_offset = -1; 1632 1633 // This is a trick. We double the stack slots so we can claim 1634 // the oops in the caller's frame. Since we are sure to have 1635 // more args than the caller doubling is enough to make 1636 // sure we can capture all the incoming oop args from the 1637 // caller. 1638 // 1639 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1640 1641 // Mark location of rfp (someday) 1642 // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, vmreg(rfp)); 1643 1644 1645 int float_args = 0; 1646 int int_args = 0; 1647 1648 #ifdef ASSERT 1649 bool reg_destroyed[RegisterImpl::number_of_registers]; 1650 bool freg_destroyed[FloatRegisterImpl::number_of_registers]; 1651 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) { 1652 reg_destroyed[r] = false; 1653 } 1654 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) { 1655 freg_destroyed[f] = false; 1656 } 1657 1658 #endif /* ASSERT */ 1659 1660 // This may iterate in two different directions depending on the 1661 // kind of native it is. The reason is that for regular JNI natives 1662 // the incoming and outgoing registers are offset upwards and for 1663 // critical natives they are offset down. 1664 GrowableArray<int> arg_order(2 * total_in_args); 1665 VMRegPair tmp_vmreg; 1666 tmp_vmreg.set1(r19->as_VMReg()); 1667 1668 if (!is_critical_native) { 1669 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) { 1670 arg_order.push(i); 1671 arg_order.push(c_arg); 1672 } 1673 } else { 1674 // Compute a valid move order, using tmp_vmreg to break any cycles 1675 ComputeMoveOrder cmo(total_in_args, in_regs, total_c_args, out_regs, in_sig_bt, arg_order, tmp_vmreg); 1676 } 1677 1678 int temploc = -1; 1679 for (int ai = 0; ai < arg_order.length(); ai += 2) { 1680 int i = arg_order.at(ai); 1681 int c_arg = arg_order.at(ai + 1); 1682 __ block_comment(err_msg("move %d -> %d", i, c_arg)); 1683 if (c_arg == -1) { 1684 assert(is_critical_native, "should only be required for critical natives"); 1685 // This arg needs to be moved to a temporary 1686 __ mov(tmp_vmreg.first()->as_Register(), in_regs[i].first()->as_Register()); 1687 in_regs[i] = tmp_vmreg; 1688 temploc = i; 1689 continue; 1690 } else if (i == -1) { 1691 assert(is_critical_native, "should only be required for critical natives"); 1692 // Read from the temporary location 1693 assert(temploc != -1, "must be valid"); 1694 i = temploc; 1695 temploc = -1; 1696 } 1697 #ifdef ASSERT 1698 if (in_regs[i].first()->is_Register()) { 1699 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!"); 1700 } else if (in_regs[i].first()->is_FloatRegister()) { 1701 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!"); 1702 } 1703 if (out_regs[c_arg].first()->is_Register()) { 1704 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true; 1705 } else if (out_regs[c_arg].first()->is_FloatRegister()) { 1706 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true; 1707 } 1708 #endif /* ASSERT */ 1709 switch (in_sig_bt[i]) { 1710 case T_ARRAY: 1711 if (is_critical_native) { 1712 unpack_array_argument(masm, in_regs[i], in_elem_bt[i], out_regs[c_arg + 1], out_regs[c_arg]); 1713 c_arg++; 1714 #ifdef ASSERT 1715 if (out_regs[c_arg].first()->is_Register()) { 1716 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true; 1717 } else if (out_regs[c_arg].first()->is_FloatRegister()) { 1718 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true; 1719 } 1720 #endif 1721 int_args++; 1722 break; 1723 } 1724 case T_OBJECT: 1725 assert(!is_critical_native, "no oop arguments"); 1726 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg], 1727 ((i == 0) && (!is_static)), 1728 &receiver_offset); 1729 int_args++; 1730 break; 1731 case T_VOID: 1732 break; 1733 1734 case T_FLOAT: 1735 float_move(masm, in_regs[i], out_regs[c_arg]); 1736 float_args++; 1737 break; 1738 1739 case T_DOUBLE: 1740 assert( i + 1 < total_in_args && 1741 in_sig_bt[i + 1] == T_VOID && 1742 out_sig_bt[c_arg+1] == T_VOID, "bad arg list"); 1743 double_move(masm, in_regs[i], out_regs[c_arg]); 1744 float_args++; 1745 break; 1746 1747 case T_LONG : 1748 long_move(masm, in_regs[i], out_regs[c_arg]); 1749 int_args++; 1750 break; 1751 1752 case T_ADDRESS: assert(false, "found T_ADDRESS in java args"); 1753 1754 default: 1755 move32_64(masm, in_regs[i], out_regs[c_arg]); 1756 int_args++; 1757 } 1758 } 1759 1760 // point c_arg at the first arg that is already loaded in case we 1761 // need to spill before we call out 1762 int c_arg = total_c_args - total_in_args; 1763 1764 // Pre-load a static method's oop into c_rarg1. 1765 if (method->is_static() && !is_critical_native) { 1766 1767 // load oop into a register 1768 __ movoop(c_rarg1, 1769 JNIHandles::make_local(method->method_holder()->java_mirror()), 1770 /*immediate*/true); 1771 1772 // Now handlize the static class mirror it's known not-null. 1773 __ str(c_rarg1, Address(sp, klass_offset)); 1774 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); 1775 1776 // Now get the handle 1777 __ lea(c_rarg1, Address(sp, klass_offset)); 1778 // and protect the arg if we must spill 1779 c_arg--; 1780 } 1781 1782 // Change state to native (we save the return address in the thread, since it might not 1783 // be pushed on the stack when we do a a stack traversal). It is enough that the pc() 1784 // points into the right code segment. It does not have to be the correct return pc. 1785 // We use the same pc/oopMap repeatedly when we call out 1786 1787 intptr_t the_pc = (intptr_t) __ pc(); 1788 oop_maps->add_gc_map(the_pc - start, map); 1789 1790 __ set_last_Java_frame(sp, noreg, (address)the_pc, rscratch1); 1791 1792 Label dtrace_method_entry, dtrace_method_entry_done; 1793 { 1794 unsigned long offset; 1795 __ adrp(rscratch1, ExternalAddress((address)&DTraceMethodProbes), offset); 1796 __ ldrb(rscratch1, Address(rscratch1, offset)); 1797 __ cbnzw(rscratch1, dtrace_method_entry); 1798 __ bind(dtrace_method_entry_done); 1799 } 1800 1801 // RedefineClasses() tracing support for obsolete method entry 1802 if (log_is_enabled(Trace, redefine, class, obsolete)) { 1803 // protect the args we've loaded 1804 save_args(masm, total_c_args, c_arg, out_regs); 1805 __ mov_metadata(c_rarg1, method()); 1806 __ call_VM_leaf( 1807 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 1808 rthread, c_rarg1); 1809 restore_args(masm, total_c_args, c_arg, out_regs); 1810 } 1811 1812 // Lock a synchronized method 1813 1814 // Register definitions used by locking and unlocking 1815 1816 const Register swap_reg = r0; 1817 const Register obj_reg = r19; // Will contain the oop 1818 const Register lock_reg = r13; // Address of compiler lock object (BasicLock) 1819 const Register old_hdr = r13; // value of old header at unlock time 1820 const Register tmp = lr; 1821 1822 Label slow_path_lock; 1823 Label lock_done; 1824 1825 if (method->is_synchronized()) { 1826 assert(!is_critical_native, "unhandled"); 1827 1828 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes(); 1829 1830 // Get the handle (the 2nd argument) 1831 __ mov(oop_handle_reg, c_rarg1); 1832 1833 // Get address of the box 1834 1835 __ lea(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 1836 1837 // Load the oop from the handle 1838 __ ldr(obj_reg, Address(oop_handle_reg, 0)); 1839 1840 if (UseBiasedLocking) { 1841 __ biased_locking_enter(lock_reg, obj_reg, swap_reg, tmp, false, lock_done, &slow_path_lock); 1842 } 1843 1844 // Load (object->mark() | 1) into swap_reg %r0 1845 __ ldr(rscratch1, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 1846 __ orr(swap_reg, rscratch1, 1); 1847 1848 // Save (object->mark() | 1) into BasicLock's displaced header 1849 __ str(swap_reg, Address(lock_reg, mark_word_offset)); 1850 1851 // src -> dest iff dest == r0 else r0 <- dest 1852 { Label here; 1853 __ cmpxchg_obj_header(r0, lock_reg, obj_reg, rscratch1, lock_done, /*fallthrough*/NULL); 1854 } 1855 1856 // Hmm should this move to the slow path code area??? 1857 1858 // Test if the oopMark is an obvious stack pointer, i.e., 1859 // 1) (mark & 3) == 0, and 1860 // 2) sp <= mark < mark + os::pagesize() 1861 // These 3 tests can be done by evaluating the following 1862 // expression: ((mark - sp) & (3 - os::vm_page_size())), 1863 // assuming both stack pointer and pagesize have their 1864 // least significant 2 bits clear. 1865 // NOTE: the oopMark is in swap_reg %r0 as the result of cmpxchg 1866 1867 __ sub(swap_reg, sp, swap_reg); 1868 __ neg(swap_reg, swap_reg); 1869 __ ands(swap_reg, swap_reg, 3 - os::vm_page_size()); 1870 1871 // Save the test result, for recursive case, the result is zero 1872 __ str(swap_reg, Address(lock_reg, mark_word_offset)); 1873 __ br(Assembler::NE, slow_path_lock); 1874 1875 // Slow path will re-enter here 1876 1877 __ bind(lock_done); 1878 } 1879 1880 1881 // Finally just about ready to make the JNI call 1882 1883 // get JNIEnv* which is first argument to native 1884 if (!is_critical_native) { 1885 __ lea(c_rarg0, Address(rthread, in_bytes(JavaThread::jni_environment_offset()))); 1886 } 1887 1888 // Now set thread in native 1889 __ mov(rscratch1, _thread_in_native); 1890 __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset())); 1891 __ stlrw(rscratch1, rscratch2); 1892 1893 { 1894 int return_type = 0; 1895 switch (ret_type) { 1896 case T_VOID: break; 1897 return_type = 0; break; 1898 case T_CHAR: 1899 case T_BYTE: 1900 case T_SHORT: 1901 case T_INT: 1902 case T_BOOLEAN: 1903 case T_LONG: 1904 return_type = 1; break; 1905 case T_ARRAY: 1906 case T_OBJECT: 1907 return_type = 1; break; 1908 case T_FLOAT: 1909 return_type = 2; break; 1910 case T_DOUBLE: 1911 return_type = 3; break; 1912 default: 1913 ShouldNotReachHere(); 1914 } 1915 rt_call(masm, native_func, 1916 int_args + 2, // AArch64 passes up to 8 args in int registers 1917 float_args, // and up to 8 float args 1918 return_type); 1919 } 1920 1921 // Unpack native results. 1922 switch (ret_type) { 1923 case T_BOOLEAN: __ ubfx(r0, r0, 0, 8); break; 1924 case T_CHAR : __ ubfx(r0, r0, 0, 16); break; 1925 case T_BYTE : __ sbfx(r0, r0, 0, 8); break; 1926 case T_SHORT : __ sbfx(r0, r0, 0, 16); break; 1927 case T_INT : __ sbfx(r0, r0, 0, 32); break; 1928 case T_DOUBLE : 1929 case T_FLOAT : 1930 // Result is in v0 we'll save as needed 1931 break; 1932 case T_ARRAY: // Really a handle 1933 case T_OBJECT: // Really a handle 1934 break; // can't de-handlize until after safepoint check 1935 case T_VOID: break; 1936 case T_LONG: break; 1937 default : ShouldNotReachHere(); 1938 } 1939 1940 // Switch thread to "native transition" state before reading the synchronization state. 1941 // This additional state is necessary because reading and testing the synchronization 1942 // state is not atomic w.r.t. GC, as this scenario demonstrates: 1943 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 1944 // VM thread changes sync state to synchronizing and suspends threads for GC. 1945 // Thread A is resumed to finish this native method, but doesn't block here since it 1946 // didn't see any synchronization is progress, and escapes. 1947 __ mov(rscratch1, _thread_in_native_trans); 1948 1949 if(os::is_MP()) { 1950 if (UseMembar) { 1951 __ strw(rscratch1, Address(rthread, JavaThread::thread_state_offset())); 1952 1953 // Force this write out before the read below 1954 __ dmb(Assembler::SY); 1955 } else { 1956 __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset())); 1957 __ stlrw(rscratch1, rscratch2); 1958 1959 // Write serialization page so VM thread can do a pseudo remote membar. 1960 // We use the current thread pointer to calculate a thread specific 1961 // offset to write to within the page. This minimizes bus traffic 1962 // due to cache line collision. 1963 __ serialize_memory(rthread, r2); 1964 } 1965 } else { 1966 __ strw(rscratch1, Address(rthread, JavaThread::thread_state_offset())); 1967 } 1968 1969 // check for safepoint operation in progress and/or pending suspend requests 1970 Label safepoint_in_progress, safepoint_in_progress_done; 1971 { 1972 assert(SafepointSynchronize::_not_synchronized == 0, "fix this code"); 1973 unsigned long offset; 1974 __ adrp(rscratch1, 1975 ExternalAddress((address)SafepointSynchronize::address_of_state()), 1976 offset); 1977 __ ldrw(rscratch1, Address(rscratch1, offset)); 1978 __ cbnzw(rscratch1, safepoint_in_progress); 1979 __ ldrw(rscratch1, Address(rthread, JavaThread::suspend_flags_offset())); 1980 __ cbnzw(rscratch1, safepoint_in_progress); 1981 __ bind(safepoint_in_progress_done); 1982 } 1983 1984 // change thread state 1985 Label after_transition; 1986 __ mov(rscratch1, _thread_in_Java); 1987 __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset())); 1988 __ stlrw(rscratch1, rscratch2); 1989 __ bind(after_transition); 1990 1991 Label reguard; 1992 Label reguard_done; 1993 __ ldrb(rscratch1, Address(rthread, JavaThread::stack_guard_state_offset())); 1994 __ cmpw(rscratch1, JavaThread::stack_guard_yellow_reserved_disabled); 1995 __ br(Assembler::EQ, reguard); 1996 __ bind(reguard_done); 1997 1998 // native result if any is live 1999 2000 // Unlock 2001 Label unlock_done; 2002 Label slow_path_unlock; 2003 if (method->is_synchronized()) { 2004 2005 // Get locked oop from the handle we passed to jni 2006 __ ldr(obj_reg, Address(oop_handle_reg, 0)); 2007 2008 Label done; 2009 2010 if (UseBiasedLocking) { 2011 __ biased_locking_exit(obj_reg, old_hdr, done); 2012 } 2013 2014 // Simple recursive lock? 2015 2016 __ ldr(rscratch1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 2017 __ cbz(rscratch1, done); 2018 2019 // Must save r0 if if it is live now because cmpxchg must use it 2020 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 2021 save_native_result(masm, ret_type, stack_slots); 2022 } 2023 2024 2025 // get address of the stack lock 2026 __ lea(r0, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 2027 // get old displaced header 2028 __ ldr(old_hdr, Address(r0, 0)); 2029 2030 // Atomic swap old header if oop still contains the stack lock 2031 Label succeed; 2032 __ cmpxchg_obj_header(r0, old_hdr, obj_reg, rscratch1, succeed, &slow_path_unlock); 2033 __ bind(succeed); 2034 2035 // slow path re-enters here 2036 __ bind(unlock_done); 2037 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 2038 restore_native_result(masm, ret_type, stack_slots); 2039 } 2040 2041 __ bind(done); 2042 } 2043 2044 Label dtrace_method_exit, dtrace_method_exit_done; 2045 { 2046 unsigned long offset; 2047 __ adrp(rscratch1, ExternalAddress((address)&DTraceMethodProbes), offset); 2048 __ ldrb(rscratch1, Address(rscratch1, offset)); 2049 __ cbnzw(rscratch1, dtrace_method_exit); 2050 __ bind(dtrace_method_exit_done); 2051 } 2052 2053 __ reset_last_Java_frame(false); 2054 2055 // Unbox oop result, e.g. JNIHandles::resolve result. 2056 if (ret_type == T_OBJECT || ret_type == T_ARRAY) { 2057 Label done, not_weak; 2058 __ cbz(r0, done); // Use NULL as-is. 2059 STATIC_ASSERT(JNIHandles::weak_tag_mask == 1u); 2060 __ tbz(r0, 0, not_weak); // Test for jweak tag. 2061 // Resolve jweak. 2062 __ ldr(r0, Address(r0, -JNIHandles::weak_tag_value)); 2063 __ verify_oop(r0); 2064 #if INCLUDE_ALL_GCS 2065 if (UseG1GC) { 2066 __ g1_write_barrier_pre(noreg /* obj */, 2067 r0 /* pre_val */, 2068 rthread /* thread */, 2069 rscratch1 /* tmp */, 2070 true /* tosca_live */, 2071 true /* expand_call */); 2072 } 2073 #endif // INCLUDE_ALL_GCS 2074 __ b(done); 2075 __ bind(not_weak); 2076 // Resolve (untagged) jobject. 2077 __ ldr(r0, Address(r0, 0)); 2078 __ verify_oop(r0); 2079 __ bind(done); 2080 } 2081 2082 if (CheckJNICalls) { 2083 // clear_pending_jni_exception_check 2084 __ str(zr, Address(rthread, JavaThread::pending_jni_exception_check_fn_offset())); 2085 } 2086 2087 if (!is_critical_native) { 2088 // reset handle block 2089 __ ldr(r2, Address(rthread, JavaThread::active_handles_offset())); 2090 __ str(zr, Address(r2, JNIHandleBlock::top_offset_in_bytes())); 2091 } 2092 2093 __ leave(); 2094 2095 if (!is_critical_native) { 2096 // Any exception pending? 2097 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2098 __ cbnz(rscratch1, exception_pending); 2099 } 2100 2101 // record exit from native wrapper code 2102 if (NotifySimulator) { 2103 __ notify(Assembler::method_reentry); 2104 } 2105 2106 // We're done 2107 __ ret(lr); 2108 2109 // Unexpected paths are out of line and go here 2110 2111 if (!is_critical_native) { 2112 // forward the exception 2113 __ bind(exception_pending); 2114 2115 // and forward the exception 2116 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2117 } 2118 2119 // Slow path locking & unlocking 2120 if (method->is_synchronized()) { 2121 2122 __ block_comment("Slow path lock {"); 2123 __ bind(slow_path_lock); 2124 2125 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM 2126 // args are (oop obj, BasicLock* lock, JavaThread* thread) 2127 2128 // protect the args we've loaded 2129 save_args(masm, total_c_args, c_arg, out_regs); 2130 2131 __ mov(c_rarg0, obj_reg); 2132 __ mov(c_rarg1, lock_reg); 2133 __ mov(c_rarg2, rthread); 2134 2135 // Not a leaf but we have last_Java_frame setup as we want 2136 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3); 2137 restore_args(masm, total_c_args, c_arg, out_regs); 2138 2139 #ifdef ASSERT 2140 { Label L; 2141 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2142 __ cbz(rscratch1, L); 2143 __ stop("no pending exception allowed on exit from monitorenter"); 2144 __ bind(L); 2145 } 2146 #endif 2147 __ b(lock_done); 2148 2149 __ block_comment("} Slow path lock"); 2150 2151 __ block_comment("Slow path unlock {"); 2152 __ bind(slow_path_unlock); 2153 2154 // If we haven't already saved the native result we must save it now as xmm registers 2155 // are still exposed. 2156 2157 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) { 2158 save_native_result(masm, ret_type, stack_slots); 2159 } 2160 2161 __ mov(c_rarg2, rthread); 2162 __ lea(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 2163 __ mov(c_rarg0, obj_reg); 2164 2165 // Save pending exception around call to VM (which contains an EXCEPTION_MARK) 2166 // NOTE that obj_reg == r19 currently 2167 __ ldr(r19, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2168 __ str(zr, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2169 2170 rt_call(masm, CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 3, 0, 1); 2171 2172 #ifdef ASSERT 2173 { 2174 Label L; 2175 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2176 __ cbz(rscratch1, L); 2177 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C"); 2178 __ bind(L); 2179 } 2180 #endif /* ASSERT */ 2181 2182 __ str(r19, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2183 2184 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) { 2185 restore_native_result(masm, ret_type, stack_slots); 2186 } 2187 __ b(unlock_done); 2188 2189 __ block_comment("} Slow path unlock"); 2190 2191 } // synchronized 2192 2193 // SLOW PATH Reguard the stack if needed 2194 2195 __ bind(reguard); 2196 save_native_result(masm, ret_type, stack_slots); 2197 rt_call(masm, CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages), 0, 0, 0); 2198 restore_native_result(masm, ret_type, stack_slots); 2199 // and continue 2200 __ b(reguard_done); 2201 2202 // SLOW PATH safepoint 2203 { 2204 __ block_comment("safepoint {"); 2205 __ bind(safepoint_in_progress); 2206 2207 // Don't use call_VM as it will see a possible pending exception and forward it 2208 // and never return here preventing us from clearing _last_native_pc down below. 2209 // 2210 save_native_result(masm, ret_type, stack_slots); 2211 __ mov(c_rarg0, rthread); 2212 #ifndef PRODUCT 2213 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 2214 #endif 2215 if (!is_critical_native) { 2216 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans))); 2217 } else { 2218 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans_and_transition))); 2219 } 2220 __ blrt(rscratch1, 1, 0, 1); 2221 __ maybe_isb(); 2222 // Restore any method result value 2223 restore_native_result(masm, ret_type, stack_slots); 2224 2225 if (is_critical_native) { 2226 // The call above performed the transition to thread_in_Java so 2227 // skip the transition logic above. 2228 __ b(after_transition); 2229 } 2230 2231 __ b(safepoint_in_progress_done); 2232 __ block_comment("} safepoint"); 2233 } 2234 2235 // SLOW PATH dtrace support 2236 { 2237 __ block_comment("dtrace entry {"); 2238 __ bind(dtrace_method_entry); 2239 2240 // We have all of the arguments setup at this point. We must not touch any register 2241 // argument registers at this point (what if we save/restore them there are no oop? 2242 2243 save_args(masm, total_c_args, c_arg, out_regs); 2244 __ mov_metadata(c_rarg1, method()); 2245 __ call_VM_leaf( 2246 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 2247 rthread, c_rarg1); 2248 restore_args(masm, total_c_args, c_arg, out_regs); 2249 __ b(dtrace_method_entry_done); 2250 __ block_comment("} dtrace entry"); 2251 } 2252 2253 { 2254 __ block_comment("dtrace exit {"); 2255 __ bind(dtrace_method_exit); 2256 save_native_result(masm, ret_type, stack_slots); 2257 __ mov_metadata(c_rarg1, method()); 2258 __ call_VM_leaf( 2259 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 2260 rthread, c_rarg1); 2261 restore_native_result(masm, ret_type, stack_slots); 2262 __ b(dtrace_method_exit_done); 2263 __ block_comment("} dtrace exit"); 2264 } 2265 2266 2267 __ flush(); 2268 2269 nmethod *nm = nmethod::new_native_nmethod(method, 2270 compile_id, 2271 masm->code(), 2272 vep_offset, 2273 frame_complete, 2274 stack_slots / VMRegImpl::slots_per_word, 2275 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), 2276 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size), 2277 oop_maps); 2278 2279 if (is_critical_native) { 2280 nm->set_lazy_critical_native(true); 2281 } 2282 2283 return nm; 2284 2285 } 2286 2287 // this function returns the adjust size (in number of words) to a c2i adapter 2288 // activation for use during deoptimization 2289 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) { 2290 assert(callee_locals >= callee_parameters, 2291 "test and remove; got more parms than locals"); 2292 if (callee_locals < callee_parameters) 2293 return 0; // No adjustment for negative locals 2294 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords; 2295 // diff is counted in stack words 2296 return align_up(diff, 2); 2297 } 2298 2299 2300 //------------------------------generate_deopt_blob---------------------------- 2301 void SharedRuntime::generate_deopt_blob() { 2302 // Allocate space for the code 2303 ResourceMark rm; 2304 // Setup code generation tools 2305 int pad = 0; 2306 #if INCLUDE_JVMCI 2307 if (EnableJVMCI) { 2308 pad += 512; // Increase the buffer size when compiling for JVMCI 2309 } 2310 #endif 2311 CodeBuffer buffer("deopt_blob", 2048+pad, 1024); 2312 MacroAssembler* masm = new MacroAssembler(&buffer); 2313 int frame_size_in_words; 2314 OopMap* map = NULL; 2315 OopMapSet *oop_maps = new OopMapSet(); 2316 2317 #ifdef BUILTIN_SIM 2318 AArch64Simulator *simulator; 2319 if (NotifySimulator) { 2320 simulator = AArch64Simulator::get_current(UseSimulatorCache, DisableBCCheck); 2321 simulator->notifyCompile(const_cast<char*>("SharedRuntime::deopt_blob"), __ pc()); 2322 } 2323 #endif 2324 2325 // ------------- 2326 // This code enters when returning to a de-optimized nmethod. A return 2327 // address has been pushed on the the stack, and return values are in 2328 // registers. 2329 // If we are doing a normal deopt then we were called from the patched 2330 // nmethod from the point we returned to the nmethod. So the return 2331 // address on the stack is wrong by NativeCall::instruction_size 2332 // We will adjust the value so it looks like we have the original return 2333 // address on the stack (like when we eagerly deoptimized). 2334 // In the case of an exception pending when deoptimizing, we enter 2335 // with a return address on the stack that points after the call we patched 2336 // into the exception handler. We have the following register state from, 2337 // e.g., the forward exception stub (see stubGenerator_x86_64.cpp). 2338 // r0: exception oop 2339 // r19: exception handler 2340 // r3: throwing pc 2341 // So in this case we simply jam r3 into the useless return address and 2342 // the stack looks just like we want. 2343 // 2344 // At this point we need to de-opt. We save the argument return 2345 // registers. We call the first C routine, fetch_unroll_info(). This 2346 // routine captures the return values and returns a structure which 2347 // describes the current frame size and the sizes of all replacement frames. 2348 // The current frame is compiled code and may contain many inlined 2349 // functions, each with their own JVM state. We pop the current frame, then 2350 // push all the new frames. Then we call the C routine unpack_frames() to 2351 // populate these frames. Finally unpack_frames() returns us the new target 2352 // address. Notice that callee-save registers are BLOWN here; they have 2353 // already been captured in the vframeArray at the time the return PC was 2354 // patched. 2355 address start = __ pc(); 2356 Label cont; 2357 2358 // Prolog for non exception case! 2359 2360 // Save everything in sight. 2361 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2362 2363 // Normal deoptimization. Save exec mode for unpack_frames. 2364 __ movw(rcpool, Deoptimization::Unpack_deopt); // callee-saved 2365 __ b(cont); 2366 2367 int reexecute_offset = __ pc() - start; 2368 #if defined(INCLUDE_JVMCI) && !defined(COMPILER1) 2369 if (EnableJVMCI && UseJVMCICompiler) { 2370 // JVMCI does not use this kind of deoptimization 2371 __ should_not_reach_here(); 2372 } 2373 #endif 2374 2375 // Reexecute case 2376 // return address is the pc describes what bci to do re-execute at 2377 2378 // No need to update map as each call to save_live_registers will produce identical oopmap 2379 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2380 2381 __ movw(rcpool, Deoptimization::Unpack_reexecute); // callee-saved 2382 __ b(cont); 2383 2384 #if INCLUDE_JVMCI 2385 Label after_fetch_unroll_info_call; 2386 int implicit_exception_uncommon_trap_offset = 0; 2387 int uncommon_trap_offset = 0; 2388 2389 if (EnableJVMCI) { 2390 implicit_exception_uncommon_trap_offset = __ pc() - start; 2391 2392 __ ldr(lr, Address(rthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset()))); 2393 __ str(zr, Address(rthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset()))); 2394 2395 uncommon_trap_offset = __ pc() - start; 2396 2397 // Save everything in sight. 2398 RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2399 // fetch_unroll_info needs to call last_java_frame() 2400 Label retaddr; 2401 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2402 2403 __ ldrw(c_rarg1, Address(rthread, in_bytes(JavaThread::pending_deoptimization_offset()))); 2404 __ movw(rscratch1, -1); 2405 __ strw(rscratch1, Address(rthread, in_bytes(JavaThread::pending_deoptimization_offset()))); 2406 2407 __ movw(rcpool, (int32_t)Deoptimization::Unpack_reexecute); 2408 __ mov(c_rarg0, rthread); 2409 __ movw(c_rarg2, rcpool); // exec mode 2410 __ lea(rscratch1, 2411 RuntimeAddress(CAST_FROM_FN_PTR(address, 2412 Deoptimization::uncommon_trap))); 2413 __ blrt(rscratch1, 2, 0, MacroAssembler::ret_type_integral); 2414 __ bind(retaddr); 2415 oop_maps->add_gc_map( __ pc()-start, map->deep_copy()); 2416 2417 __ reset_last_Java_frame(false); 2418 2419 __ b(after_fetch_unroll_info_call); 2420 } // EnableJVMCI 2421 #endif // INCLUDE_JVMCI 2422 2423 int exception_offset = __ pc() - start; 2424 2425 // Prolog for exception case 2426 2427 // all registers are dead at this entry point, except for r0, and 2428 // r3 which contain the exception oop and exception pc 2429 // respectively. Set them in TLS and fall thru to the 2430 // unpack_with_exception_in_tls entry point. 2431 2432 __ str(r3, Address(rthread, JavaThread::exception_pc_offset())); 2433 __ str(r0, Address(rthread, JavaThread::exception_oop_offset())); 2434 2435 int exception_in_tls_offset = __ pc() - start; 2436 2437 // new implementation because exception oop is now passed in JavaThread 2438 2439 // Prolog for exception case 2440 // All registers must be preserved because they might be used by LinearScan 2441 // Exceptiop oop and throwing PC are passed in JavaThread 2442 // tos: stack at point of call to method that threw the exception (i.e. only 2443 // args are on the stack, no return address) 2444 2445 // The return address pushed by save_live_registers will be patched 2446 // later with the throwing pc. The correct value is not available 2447 // now because loading it from memory would destroy registers. 2448 2449 // NB: The SP at this point must be the SP of the method that is 2450 // being deoptimized. Deoptimization assumes that the frame created 2451 // here by save_live_registers is immediately below the method's SP. 2452 // This is a somewhat fragile mechanism. 2453 2454 // Save everything in sight. 2455 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2456 2457 // Now it is safe to overwrite any register 2458 2459 // Deopt during an exception. Save exec mode for unpack_frames. 2460 __ mov(rcpool, Deoptimization::Unpack_exception); // callee-saved 2461 2462 // load throwing pc from JavaThread and patch it as the return address 2463 // of the current frame. Then clear the field in JavaThread 2464 2465 __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset())); 2466 __ str(r3, Address(rfp, wordSize)); 2467 __ str(zr, Address(rthread, JavaThread::exception_pc_offset())); 2468 2469 #ifdef ASSERT 2470 // verify that there is really an exception oop in JavaThread 2471 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset())); 2472 __ verify_oop(r0); 2473 2474 // verify that there is no pending exception 2475 Label no_pending_exception; 2476 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset())); 2477 __ cbz(rscratch1, no_pending_exception); 2478 __ stop("must not have pending exception here"); 2479 __ bind(no_pending_exception); 2480 #endif 2481 2482 __ bind(cont); 2483 2484 // Call C code. Need thread and this frame, but NOT official VM entry 2485 // crud. We cannot block on this call, no GC can happen. 2486 // 2487 // UnrollBlock* fetch_unroll_info(JavaThread* thread) 2488 2489 // fetch_unroll_info needs to call last_java_frame(). 2490 2491 Label retaddr; 2492 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2493 #ifdef ASSERT0 2494 { Label L; 2495 __ ldr(rscratch1, Address(rthread, 2496 JavaThread::last_Java_fp_offset())); 2497 __ cbz(rscratch1, L); 2498 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared"); 2499 __ bind(L); 2500 } 2501 #endif // ASSERT 2502 __ mov(c_rarg0, rthread); 2503 __ mov(c_rarg1, rcpool); 2504 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info))); 2505 __ blrt(rscratch1, 1, 0, 1); 2506 __ bind(retaddr); 2507 2508 // Need to have an oopmap that tells fetch_unroll_info where to 2509 // find any register it might need. 2510 oop_maps->add_gc_map(__ pc() - start, map); 2511 2512 __ reset_last_Java_frame(false); 2513 2514 #if INCLUDE_JVMCI 2515 if (EnableJVMCI) { 2516 __ bind(after_fetch_unroll_info_call); 2517 } 2518 #endif 2519 2520 // Load UnrollBlock* into r5 2521 __ mov(r5, r0); 2522 2523 __ ldrw(rcpool, Address(r5, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes())); 2524 Label noException; 2525 __ cmpw(rcpool, Deoptimization::Unpack_exception); // Was exception pending? 2526 __ br(Assembler::NE, noException); 2527 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset())); 2528 // QQQ this is useless it was NULL above 2529 __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset())); 2530 __ str(zr, Address(rthread, JavaThread::exception_oop_offset())); 2531 __ str(zr, Address(rthread, JavaThread::exception_pc_offset())); 2532 2533 __ verify_oop(r0); 2534 2535 // Overwrite the result registers with the exception results. 2536 __ str(r0, Address(sp, RegisterSaver::r0_offset_in_bytes())); 2537 // I think this is useless 2538 // __ str(r3, Address(sp, RegisterSaver::r3_offset_in_bytes())); 2539 2540 __ bind(noException); 2541 2542 // Only register save data is on the stack. 2543 // Now restore the result registers. Everything else is either dead 2544 // or captured in the vframeArray. 2545 RegisterSaver::restore_result_registers(masm); 2546 2547 // All of the register save area has been popped of the stack. Only the 2548 // return address remains. 2549 2550 // Pop all the frames we must move/replace. 2551 // 2552 // Frame picture (youngest to oldest) 2553 // 1: self-frame (no frame link) 2554 // 2: deopting frame (no frame link) 2555 // 3: caller of deopting frame (could be compiled/interpreted). 2556 // 2557 // Note: by leaving the return address of self-frame on the stack 2558 // and using the size of frame 2 to adjust the stack 2559 // when we are done the return to frame 3 will still be on the stack. 2560 2561 // Pop deoptimized frame 2562 __ ldrw(r2, Address(r5, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes())); 2563 __ sub(r2, r2, 2 * wordSize); 2564 __ add(sp, sp, r2); 2565 __ ldp(rfp, lr, __ post(sp, 2 * wordSize)); 2566 // LR should now be the return address to the caller (3) 2567 2568 #ifdef ASSERT 2569 // Compilers generate code that bang the stack by as much as the 2570 // interpreter would need. So this stack banging should never 2571 // trigger a fault. Verify that it does not on non product builds. 2572 if (UseStackBanging) { 2573 __ ldrw(r19, Address(r5, Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes())); 2574 __ bang_stack_size(r19, r2); 2575 } 2576 #endif 2577 // Load address of array of frame pcs into r2 2578 __ ldr(r2, Address(r5, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 2579 2580 // Trash the old pc 2581 // __ addptr(sp, wordSize); FIXME ???? 2582 2583 // Load address of array of frame sizes into r4 2584 __ ldr(r4, Address(r5, Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes())); 2585 2586 // Load counter into r3 2587 __ ldrw(r3, Address(r5, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes())); 2588 2589 // Now adjust the caller's stack to make up for the extra locals 2590 // but record the original sp so that we can save it in the skeletal interpreter 2591 // frame and the stack walking of interpreter_sender will get the unextended sp 2592 // value and not the "real" sp value. 2593 2594 const Register sender_sp = r6; 2595 2596 __ mov(sender_sp, sp); 2597 __ ldrw(r19, Address(r5, 2598 Deoptimization::UnrollBlock:: 2599 caller_adjustment_offset_in_bytes())); 2600 __ sub(sp, sp, r19); 2601 2602 // Push interpreter frames in a loop 2603 __ mov(rscratch1, (address)0xDEADDEAD); // Make a recognizable pattern 2604 __ mov(rscratch2, rscratch1); 2605 Label loop; 2606 __ bind(loop); 2607 __ ldr(r19, Address(__ post(r4, wordSize))); // Load frame size 2608 __ sub(r19, r19, 2*wordSize); // We'll push pc and fp by hand 2609 __ ldr(lr, Address(__ post(r2, wordSize))); // Load pc 2610 __ enter(); // Save old & set new fp 2611 __ sub(sp, sp, r19); // Prolog 2612 // This value is corrected by layout_activation_impl 2613 __ str(zr, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 2614 __ str(sender_sp, Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable 2615 __ mov(sender_sp, sp); // Pass sender_sp to next frame 2616 __ sub(r3, r3, 1); // Decrement counter 2617 __ cbnz(r3, loop); 2618 2619 // Re-push self-frame 2620 __ ldr(lr, Address(r2)); 2621 __ enter(); 2622 2623 // Allocate a full sized register save area. We subtract 2 because 2624 // enter() just pushed 2 words 2625 __ sub(sp, sp, (frame_size_in_words - 2) * wordSize); 2626 2627 // Restore frame locals after moving the frame 2628 __ strd(v0, Address(sp, RegisterSaver::v0_offset_in_bytes())); 2629 __ str(r0, Address(sp, RegisterSaver::r0_offset_in_bytes())); 2630 2631 // Call C code. Need thread but NOT official VM entry 2632 // crud. We cannot block on this call, no GC can happen. Call should 2633 // restore return values to their stack-slots with the new SP. 2634 // 2635 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode) 2636 2637 // Use rfp because the frames look interpreted now 2638 // Don't need the precise return PC here, just precise enough to point into this code blob. 2639 address the_pc = __ pc(); 2640 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1); 2641 2642 __ mov(c_rarg0, rthread); 2643 __ movw(c_rarg1, rcpool); // second arg: exec_mode 2644 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 2645 __ blrt(rscratch1, 2, 0, 0); 2646 2647 // Set an oopmap for the call site 2648 // Use the same PC we used for the last java frame 2649 oop_maps->add_gc_map(the_pc - start, 2650 new OopMap( frame_size_in_words, 0 )); 2651 2652 // Clear fp AND pc 2653 __ reset_last_Java_frame(true); 2654 2655 // Collect return values 2656 __ ldrd(v0, Address(sp, RegisterSaver::v0_offset_in_bytes())); 2657 __ ldr(r0, Address(sp, RegisterSaver::r0_offset_in_bytes())); 2658 // I think this is useless (throwing pc?) 2659 // __ ldr(r3, Address(sp, RegisterSaver::r3_offset_in_bytes())); 2660 2661 // Pop self-frame. 2662 __ leave(); // Epilog 2663 2664 // Jump to interpreter 2665 __ ret(lr); 2666 2667 // Make sure all code is generated 2668 masm->flush(); 2669 2670 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words); 2671 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset); 2672 #if INCLUDE_JVMCI 2673 if (EnableJVMCI) { 2674 _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset); 2675 _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset); 2676 } 2677 #endif 2678 #ifdef BUILTIN_SIM 2679 if (NotifySimulator) { 2680 unsigned char *base = _deopt_blob->code_begin(); 2681 simulator->notifyRelocate(start, base - start); 2682 } 2683 #endif 2684 } 2685 2686 uint SharedRuntime::out_preserve_stack_slots() { 2687 return 0; 2688 } 2689 2690 #if defined(COMPILER2) || INCLUDE_JVMCI 2691 //------------------------------generate_uncommon_trap_blob-------------------- 2692 void SharedRuntime::generate_uncommon_trap_blob() { 2693 // Allocate space for the code 2694 ResourceMark rm; 2695 // Setup code generation tools 2696 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024); 2697 MacroAssembler* masm = new MacroAssembler(&buffer); 2698 2699 #ifdef BUILTIN_SIM 2700 AArch64Simulator *simulator; 2701 if (NotifySimulator) { 2702 simulator = AArch64Simulator::get_current(UseSimulatorCache, DisableBCCheck); 2703 simulator->notifyCompile(const_cast<char*>("SharedRuntime:uncommon_trap_blob"), __ pc()); 2704 } 2705 #endif 2706 2707 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned"); 2708 2709 address start = __ pc(); 2710 2711 // Push self-frame. We get here with a return address in LR 2712 // and sp should be 16 byte aligned 2713 // push rfp and retaddr by hand 2714 __ stp(rfp, lr, Address(__ pre(sp, -2 * wordSize))); 2715 // we don't expect an arg reg save area 2716 #ifndef PRODUCT 2717 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 2718 #endif 2719 // compiler left unloaded_class_index in j_rarg0 move to where the 2720 // runtime expects it. 2721 if (c_rarg1 != j_rarg0) { 2722 __ movw(c_rarg1, j_rarg0); 2723 } 2724 2725 // we need to set the past SP to the stack pointer of the stub frame 2726 // and the pc to the address where this runtime call will return 2727 // although actually any pc in this code blob will do). 2728 Label retaddr; 2729 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2730 2731 // Call C code. Need thread but NOT official VM entry 2732 // crud. We cannot block on this call, no GC can happen. Call should 2733 // capture callee-saved registers as well as return values. 2734 // Thread is in rdi already. 2735 // 2736 // UnrollBlock* uncommon_trap(JavaThread* thread, jint unloaded_class_index); 2737 // 2738 // n.b. 2 gp args, 0 fp args, integral return type 2739 2740 __ mov(c_rarg0, rthread); 2741 __ movw(c_rarg2, (unsigned)Deoptimization::Unpack_uncommon_trap); 2742 __ lea(rscratch1, 2743 RuntimeAddress(CAST_FROM_FN_PTR(address, 2744 Deoptimization::uncommon_trap))); 2745 __ blrt(rscratch1, 2, 0, MacroAssembler::ret_type_integral); 2746 __ bind(retaddr); 2747 2748 // Set an oopmap for the call site 2749 OopMapSet* oop_maps = new OopMapSet(); 2750 OopMap* map = new OopMap(SimpleRuntimeFrame::framesize, 0); 2751 2752 // location of rfp is known implicitly by the frame sender code 2753 2754 oop_maps->add_gc_map(__ pc() - start, map); 2755 2756 __ reset_last_Java_frame(false); 2757 2758 // move UnrollBlock* into r4 2759 __ mov(r4, r0); 2760 2761 #ifdef ASSERT 2762 { Label L; 2763 __ ldrw(rscratch1, Address(r4, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes())); 2764 __ cmpw(rscratch1, (unsigned)Deoptimization::Unpack_uncommon_trap); 2765 __ br(Assembler::EQ, L); 2766 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared"); 2767 __ bind(L); 2768 } 2769 #endif 2770 2771 // Pop all the frames we must move/replace. 2772 // 2773 // Frame picture (youngest to oldest) 2774 // 1: self-frame (no frame link) 2775 // 2: deopting frame (no frame link) 2776 // 3: caller of deopting frame (could be compiled/interpreted). 2777 2778 // Pop self-frame. We have no frame, and must rely only on r0 and sp. 2779 __ add(sp, sp, (SimpleRuntimeFrame::framesize) << LogBytesPerInt); // Epilog! 2780 2781 // Pop deoptimized frame (int) 2782 __ ldrw(r2, Address(r4, 2783 Deoptimization::UnrollBlock:: 2784 size_of_deoptimized_frame_offset_in_bytes())); 2785 __ sub(r2, r2, 2 * wordSize); 2786 __ add(sp, sp, r2); 2787 __ ldp(rfp, lr, __ post(sp, 2 * wordSize)); 2788 // LR should now be the return address to the caller (3) frame 2789 2790 #ifdef ASSERT 2791 // Compilers generate code that bang the stack by as much as the 2792 // interpreter would need. So this stack banging should never 2793 // trigger a fault. Verify that it does not on non product builds. 2794 if (UseStackBanging) { 2795 __ ldrw(r1, Address(r4, 2796 Deoptimization::UnrollBlock:: 2797 total_frame_sizes_offset_in_bytes())); 2798 __ bang_stack_size(r1, r2); 2799 } 2800 #endif 2801 2802 // Load address of array of frame pcs into r2 (address*) 2803 __ ldr(r2, Address(r4, 2804 Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 2805 2806 // Load address of array of frame sizes into r5 (intptr_t*) 2807 __ ldr(r5, Address(r4, 2808 Deoptimization::UnrollBlock:: 2809 frame_sizes_offset_in_bytes())); 2810 2811 // Counter 2812 __ ldrw(r3, Address(r4, 2813 Deoptimization::UnrollBlock:: 2814 number_of_frames_offset_in_bytes())); // (int) 2815 2816 // Now adjust the caller's stack to make up for the extra locals but 2817 // record the original sp so that we can save it in the skeletal 2818 // interpreter frame and the stack walking of interpreter_sender 2819 // will get the unextended sp value and not the "real" sp value. 2820 2821 const Register sender_sp = r8; 2822 2823 __ mov(sender_sp, sp); 2824 __ ldrw(r1, Address(r4, 2825 Deoptimization::UnrollBlock:: 2826 caller_adjustment_offset_in_bytes())); // (int) 2827 __ sub(sp, sp, r1); 2828 2829 // Push interpreter frames in a loop 2830 Label loop; 2831 __ bind(loop); 2832 __ ldr(r1, Address(r5, 0)); // Load frame size 2833 __ sub(r1, r1, 2 * wordSize); // We'll push pc and rfp by hand 2834 __ ldr(lr, Address(r2, 0)); // Save return address 2835 __ enter(); // and old rfp & set new rfp 2836 __ sub(sp, sp, r1); // Prolog 2837 __ str(sender_sp, Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable 2838 // This value is corrected by layout_activation_impl 2839 __ str(zr, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize)); 2840 __ mov(sender_sp, sp); // Pass sender_sp to next frame 2841 __ add(r5, r5, wordSize); // Bump array pointer (sizes) 2842 __ add(r2, r2, wordSize); // Bump array pointer (pcs) 2843 __ subsw(r3, r3, 1); // Decrement counter 2844 __ br(Assembler::GT, loop); 2845 __ ldr(lr, Address(r2, 0)); // save final return address 2846 // Re-push self-frame 2847 __ enter(); // & old rfp & set new rfp 2848 2849 // Use rfp because the frames look interpreted now 2850 // Save "the_pc" since it cannot easily be retrieved using the last_java_SP after we aligned SP. 2851 // Don't need the precise return PC here, just precise enough to point into this code blob. 2852 address the_pc = __ pc(); 2853 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1); 2854 2855 // Call C code. Need thread but NOT official VM entry 2856 // crud. We cannot block on this call, no GC can happen. Call should 2857 // restore return values to their stack-slots with the new SP. 2858 // Thread is in rdi already. 2859 // 2860 // BasicType unpack_frames(JavaThread* thread, int exec_mode); 2861 // 2862 // n.b. 2 gp args, 0 fp args, integral return type 2863 2864 // sp should already be aligned 2865 __ mov(c_rarg0, rthread); 2866 __ movw(c_rarg1, (unsigned)Deoptimization::Unpack_uncommon_trap); 2867 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 2868 __ blrt(rscratch1, 2, 0, MacroAssembler::ret_type_integral); 2869 2870 // Set an oopmap for the call site 2871 // Use the same PC we used for the last java frame 2872 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0)); 2873 2874 // Clear fp AND pc 2875 __ reset_last_Java_frame(true); 2876 2877 // Pop self-frame. 2878 __ leave(); // Epilog 2879 2880 // Jump to interpreter 2881 __ ret(lr); 2882 2883 // Make sure all code is generated 2884 masm->flush(); 2885 2886 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, 2887 SimpleRuntimeFrame::framesize >> 1); 2888 2889 #ifdef BUILTIN_SIM 2890 if (NotifySimulator) { 2891 unsigned char *base = _deopt_blob->code_begin(); 2892 simulator->notifyRelocate(start, base - start); 2893 } 2894 #endif 2895 } 2896 #endif // COMPILER2 2897 2898 2899 //------------------------------generate_handler_blob------ 2900 // 2901 // Generate a special Compile2Runtime blob that saves all registers, 2902 // and setup oopmap. 2903 // 2904 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) { 2905 ResourceMark rm; 2906 OopMapSet *oop_maps = new OopMapSet(); 2907 OopMap* map; 2908 2909 // Allocate space for the code. Setup code generation tools. 2910 CodeBuffer buffer("handler_blob", 2048, 1024); 2911 MacroAssembler* masm = new MacroAssembler(&buffer); 2912 2913 address start = __ pc(); 2914 address call_pc = NULL; 2915 int frame_size_in_words; 2916 bool cause_return = (poll_type == POLL_AT_RETURN); 2917 bool save_vectors = (poll_type == POLL_AT_VECTOR_LOOP); 2918 2919 // Save registers, fpu state, and flags 2920 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, save_vectors); 2921 2922 // The following is basically a call_VM. However, we need the precise 2923 // address of the call in order to generate an oopmap. Hence, we do all the 2924 // work outselves. 2925 2926 Label retaddr; 2927 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2928 2929 // The return address must always be correct so that frame constructor never 2930 // sees an invalid pc. 2931 2932 if (!cause_return) { 2933 // overwrite the return address pushed by save_live_registers 2934 __ ldr(c_rarg0, Address(rthread, JavaThread::saved_exception_pc_offset())); 2935 __ str(c_rarg0, Address(rfp, wordSize)); 2936 } 2937 2938 // Do the call 2939 __ mov(c_rarg0, rthread); 2940 __ lea(rscratch1, RuntimeAddress(call_ptr)); 2941 __ blrt(rscratch1, 1, 0, 1); 2942 __ bind(retaddr); 2943 2944 // Set an oopmap for the call site. This oopmap will map all 2945 // oop-registers and debug-info registers as callee-saved. This 2946 // will allow deoptimization at this safepoint to find all possible 2947 // debug-info recordings, as well as let GC find all oops. 2948 2949 oop_maps->add_gc_map( __ pc() - start, map); 2950 2951 Label noException; 2952 2953 __ reset_last_Java_frame(false); 2954 2955 __ maybe_isb(); 2956 __ membar(Assembler::LoadLoad | Assembler::LoadStore); 2957 2958 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset())); 2959 __ cbz(rscratch1, noException); 2960 2961 // Exception pending 2962 2963 RegisterSaver::restore_live_registers(masm); 2964 2965 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2966 2967 // No exception case 2968 __ bind(noException); 2969 2970 // Normal exit, restore registers and exit. 2971 RegisterSaver::restore_live_registers(masm, save_vectors); 2972 2973 __ ret(lr); 2974 2975 // Make sure all code is generated 2976 masm->flush(); 2977 2978 // Fill-out other meta info 2979 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words); 2980 } 2981 2982 // 2983 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss 2984 // 2985 // Generate a stub that calls into vm to find out the proper destination 2986 // of a java call. All the argument registers are live at this point 2987 // but since this is generic code we don't know what they are and the caller 2988 // must do any gc of the args. 2989 // 2990 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) { 2991 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before"); 2992 2993 // allocate space for the code 2994 ResourceMark rm; 2995 2996 CodeBuffer buffer(name, 1000, 512); 2997 MacroAssembler* masm = new MacroAssembler(&buffer); 2998 2999 int frame_size_in_words; 3000 3001 OopMapSet *oop_maps = new OopMapSet(); 3002 OopMap* map = NULL; 3003 3004 int start = __ offset(); 3005 3006 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 3007 3008 int frame_complete = __ offset(); 3009 3010 { 3011 Label retaddr; 3012 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 3013 3014 __ mov(c_rarg0, rthread); 3015 __ lea(rscratch1, RuntimeAddress(destination)); 3016 3017 __ blrt(rscratch1, 1, 0, 1); 3018 __ bind(retaddr); 3019 } 3020 3021 // Set an oopmap for the call site. 3022 // We need this not only for callee-saved registers, but also for volatile 3023 // registers that the compiler might be keeping live across a safepoint. 3024 3025 oop_maps->add_gc_map( __ offset() - start, map); 3026 3027 __ maybe_isb(); 3028 3029 // r0 contains the address we are going to jump to assuming no exception got installed 3030 3031 // clear last_Java_sp 3032 __ reset_last_Java_frame(false); 3033 // check for pending exceptions 3034 Label pending; 3035 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset())); 3036 __ cbnz(rscratch1, pending); 3037 3038 // get the returned Method* 3039 __ get_vm_result_2(rmethod, rthread); 3040 __ str(rmethod, Address(sp, RegisterSaver::reg_offset_in_bytes(rmethod))); 3041 3042 // r0 is where we want to jump, overwrite rscratch1 which is saved and scratch 3043 __ str(r0, Address(sp, RegisterSaver::rscratch1_offset_in_bytes())); 3044 RegisterSaver::restore_live_registers(masm); 3045 3046 // We are back the the original state on entry and ready to go. 3047 3048 __ br(rscratch1); 3049 3050 // Pending exception after the safepoint 3051 3052 __ bind(pending); 3053 3054 RegisterSaver::restore_live_registers(masm); 3055 3056 // exception pending => remove activation and forward to exception handler 3057 3058 __ str(zr, Address(rthread, JavaThread::vm_result_offset())); 3059 3060 __ ldr(r0, Address(rthread, Thread::pending_exception_offset())); 3061 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 3062 3063 // ------------- 3064 // make sure all code is generated 3065 masm->flush(); 3066 3067 // return the blob 3068 // frame_size_words or bytes?? 3069 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true); 3070 } 3071 3072 3073 #if defined(COMPILER2) || INCLUDE_JVMCI 3074 // This is here instead of runtime_x86_64.cpp because it uses SimpleRuntimeFrame 3075 // 3076 //------------------------------generate_exception_blob--------------------------- 3077 // creates exception blob at the end 3078 // Using exception blob, this code is jumped from a compiled method. 3079 // (see emit_exception_handler in x86_64.ad file) 3080 // 3081 // Given an exception pc at a call we call into the runtime for the 3082 // handler in this method. This handler might merely restore state 3083 // (i.e. callee save registers) unwind the frame and jump to the 3084 // exception handler for the nmethod if there is no Java level handler 3085 // for the nmethod. 3086 // 3087 // This code is entered with a jmp. 3088 // 3089 // Arguments: 3090 // r0: exception oop 3091 // r3: exception pc 3092 // 3093 // Results: 3094 // r0: exception oop 3095 // r3: exception pc in caller or ??? 3096 // destination: exception handler of caller 3097 // 3098 // Note: the exception pc MUST be at a call (precise debug information) 3099 // Registers r0, r3, r2, r4, r5, r8-r11 are not callee saved. 3100 // 3101 3102 void OptoRuntime::generate_exception_blob() { 3103 assert(!OptoRuntime::is_callee_saved_register(R3_num), ""); 3104 assert(!OptoRuntime::is_callee_saved_register(R0_num), ""); 3105 assert(!OptoRuntime::is_callee_saved_register(R2_num), ""); 3106 3107 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned"); 3108 3109 // Allocate space for the code 3110 ResourceMark rm; 3111 // Setup code generation tools 3112 CodeBuffer buffer("exception_blob", 2048, 1024); 3113 MacroAssembler* masm = new MacroAssembler(&buffer); 3114 3115 // TODO check various assumptions made here 3116 // 3117 // make sure we do so before running this 3118 3119 address start = __ pc(); 3120 3121 // push rfp and retaddr by hand 3122 // Exception pc is 'return address' for stack walker 3123 __ stp(rfp, lr, Address(__ pre(sp, -2 * wordSize))); 3124 // there are no callee save registers and we don't expect an 3125 // arg reg save area 3126 #ifndef PRODUCT 3127 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 3128 #endif 3129 // Store exception in Thread object. We cannot pass any arguments to the 3130 // handle_exception call, since we do not want to make any assumption 3131 // about the size of the frame where the exception happened in. 3132 __ str(r0, Address(rthread, JavaThread::exception_oop_offset())); 3133 __ str(r3, Address(rthread, JavaThread::exception_pc_offset())); 3134 3135 // This call does all the hard work. It checks if an exception handler 3136 // exists in the method. 3137 // If so, it returns the handler address. 3138 // If not, it prepares for stack-unwinding, restoring the callee-save 3139 // registers of the frame being removed. 3140 // 3141 // address OptoRuntime::handle_exception_C(JavaThread* thread) 3142 // 3143 // n.b. 1 gp arg, 0 fp args, integral return type 3144 3145 // the stack should always be aligned 3146 address the_pc = __ pc(); 3147 __ set_last_Java_frame(sp, noreg, the_pc, rscratch1); 3148 __ mov(c_rarg0, rthread); 3149 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, OptoRuntime::handle_exception_C))); 3150 __ blrt(rscratch1, 1, 0, MacroAssembler::ret_type_integral); 3151 __ maybe_isb(); 3152 3153 // Set an oopmap for the call site. This oopmap will only be used if we 3154 // are unwinding the stack. Hence, all locations will be dead. 3155 // Callee-saved registers will be the same as the frame above (i.e., 3156 // handle_exception_stub), since they were restored when we got the 3157 // exception. 3158 3159 OopMapSet* oop_maps = new OopMapSet(); 3160 3161 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0)); 3162 3163 __ reset_last_Java_frame(false); 3164 3165 // Restore callee-saved registers 3166 3167 // rfp is an implicitly saved callee saved register (i.e. the calling 3168 // convention will save restore it in prolog/epilog) Other than that 3169 // there are no callee save registers now that adapter frames are gone. 3170 // and we dont' expect an arg reg save area 3171 __ ldp(rfp, r3, Address(__ post(sp, 2 * wordSize))); 3172 3173 // r0: exception handler 3174 3175 // We have a handler in r0 (could be deopt blob). 3176 __ mov(r8, r0); 3177 3178 // Get the exception oop 3179 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset())); 3180 // Get the exception pc in case we are deoptimized 3181 __ ldr(r4, Address(rthread, JavaThread::exception_pc_offset())); 3182 #ifdef ASSERT 3183 __ str(zr, Address(rthread, JavaThread::exception_handler_pc_offset())); 3184 __ str(zr, Address(rthread, JavaThread::exception_pc_offset())); 3185 #endif 3186 // Clear the exception oop so GC no longer processes it as a root. 3187 __ str(zr, Address(rthread, JavaThread::exception_oop_offset())); 3188 3189 // r0: exception oop 3190 // r8: exception handler 3191 // r4: exception pc 3192 // Jump to handler 3193 3194 __ br(r8); 3195 3196 // Make sure all code is generated 3197 masm->flush(); 3198 3199 // Set exception blob 3200 _exception_blob = ExceptionBlob::create(&buffer, oop_maps, SimpleRuntimeFrame::framesize >> 1); 3201 } 3202 #endif // COMPILER2