1 /* 2 * Copyright (c) 2003, 2018, Oracle and/or its affiliates. All rights reserved. 3 * Copyright (c) 2014, Red Hat Inc. All rights reserved. 4 * Copyright (c) 2015, Linaro Ltd. All rights reserved. 5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 6 * 7 * This code is free software; you can redistribute it and/or modify it 8 * under the terms of the GNU General Public License version 2 only, as 9 * published by the Free Software Foundation. 10 * 11 * This code is distributed in the hope that it will be useful, but WITHOUT 12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 14 * version 2 for more details (a copy is included in the LICENSE file that 15 * accompanied this code). 16 * 17 * You should have received a copy of the GNU General Public License version 18 * 2 along with this work; if not, write to the Free Software Foundation, 19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 20 * 21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 22 * or visit www.oracle.com if you need additional information or have any 23 * questions. 24 * 25 */ 26 27 #include "precompiled.hpp" 28 #include "asm/macroAssembler.hpp" 29 #include "asm/macroAssembler.inline.hpp" 30 #include "code/debugInfoRec.hpp" 31 #include "code/icBuffer.hpp" 32 #include "code/vtableStubs.hpp" 33 #include "interp_masm_aarch32.hpp" 34 #include "interpreter/interpreter.hpp" 35 #include "logging/log.hpp" 36 #include "oops/compiledICHolder.hpp" 37 #include "runtime/safepointMechanism.hpp" 38 #include "runtime/sharedRuntime.hpp" 39 #include "runtime/vframeArray.hpp" 40 #include "utilities/align.hpp" 41 #include "utilities/formatBuffer.hpp" 42 #include "vmreg_aarch32.inline.hpp" 43 #include "register_aarch32.hpp" 44 #include "vm_version_aarch32.hpp" 45 #ifdef COMPILER1 46 #include "c1/c1_Runtime1.hpp" 47 #endif 48 #if COMPILER2_OR_JVMCI 49 #include "adfiles/ad_aarch32.hpp" 50 #include "opto/runtime.hpp" 51 #endif 52 #if INCLUDE_JVMCI 53 #include "jvmci/jvmciJavaClasses.hpp" 54 #endif 55 56 57 #define __ masm-> 58 59 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size; 60 61 class SimpleRuntimeFrame { 62 63 public: 64 65 // Most of the runtime stubs have this simple frame layout. 66 // This class exists to make the layout shared in one place. 67 // Offsets are for compiler stack slots, which are jints. 68 enum layout { 69 // The frame sender code expects that rbp will be in the "natural" place and 70 // will override any oopMap setting for it. We must therefore force the layout 71 // so that it agrees with the frame sender code. 72 // we don't expect any arg reg save area so aarch32 asserts that 73 // frame::arg_reg_save_area_bytes == 0 74 rbp_off = 0, 75 rbp_off2, 76 return_off, return_off2, 77 framesize 78 }; 79 }; 80 81 // FIXME -- this is used by C1 82 class RegisterSaver { 83 public: 84 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words, bool lr_pushed = false); 85 static void restore_live_registers(MacroAssembler* masm, bool restore_lr = true); 86 87 // Capture info about frame layout 88 enum layout { 89 fpu_state_off = 0, 90 fpu_state_end = fpu_state_off+FPUStateSizeInWords-1, 91 // The frame sender code expects that rfp will be in 92 // the "natural" place and will override any oopMap 93 // setting for it. We must therefore force the layout 94 // so that it agrees with the frame sender code. 95 // 96 // FIXME there are extra saved register (from `push_CPU_state`) note that r11 == rfp 97 r0_off, 98 r1_off, 99 r2_off, 100 r3_off, 101 r4_off, 102 r5_off, 103 r6_off, 104 r7_off, 105 r8_off, rmethod_off = r8_off, 106 r9_off, rscratch1_off = r9_off, 107 r10_off, 108 r11_off, 109 r12_off, 110 r14_off, // with C2 can hold value different to LR entry in the frame 111 reg_save_size, 112 }; 113 114 115 // Offsets into the register save area 116 // Used by deoptimization when it is managing result register 117 // values on its own 118 119 static int offset_in_bytes(int offset) { return offset * wordSize; } 120 121 // During deoptimization only the result registers need to be restored, 122 // all the other values have already been extracted. 123 static void restore_result_registers(MacroAssembler* masm); 124 125 }; 126 127 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words, bool lr_pushed) { 128 int frame_size_in_bytes = additional_frame_words*wordSize + (reg_save_size + frame::get_frame_size()) *BytesPerInt; 129 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt; 130 int additional_frame_slots = additional_frame_words*wordSize / BytesPerInt; 131 *total_frame_words = frame_size_in_bytes / wordSize;; 132 133 if (lr_pushed) { 134 if (FrameAPCS) 135 Unimplemented(); 136 137 __ push(rfp); 138 __ add(rfp, sp, wordSize); 139 } else 140 __ enter(); 141 __ push_CPU_state(); 142 143 // Set an oopmap for the call site. This oopmap will map all 144 // oop-registers and debug-info registers as callee-saved. This 145 // will allow deoptimization at this safepoint to find all possible 146 // debug-info recordings, as well as let GC find all oops. 147 148 OopMapSet *oop_maps = new OopMapSet(); 149 OopMap* oop_map = new OopMap(frame_size_in_slots, 0); 150 151 oop_map->set_callee_saved(VMRegImpl::stack2reg(r0_off + additional_frame_slots), r0->as_VMReg()); 152 oop_map->set_callee_saved(VMRegImpl::stack2reg(r1_off + additional_frame_slots), r1->as_VMReg()); 153 oop_map->set_callee_saved(VMRegImpl::stack2reg(r2_off + additional_frame_slots), r2->as_VMReg()); 154 oop_map->set_callee_saved(VMRegImpl::stack2reg(r3_off + additional_frame_slots), r3->as_VMReg()); 155 oop_map->set_callee_saved(VMRegImpl::stack2reg(r4_off + additional_frame_slots), r4->as_VMReg()); 156 oop_map->set_callee_saved(VMRegImpl::stack2reg(r5_off + additional_frame_slots), r5->as_VMReg()); 157 oop_map->set_callee_saved(VMRegImpl::stack2reg(r6_off + additional_frame_slots), r6->as_VMReg()); 158 oop_map->set_callee_saved(VMRegImpl::stack2reg(r7_off + additional_frame_slots), r7->as_VMReg()); 159 oop_map->set_callee_saved(VMRegImpl::stack2reg(r8_off + additional_frame_slots), r8->as_VMReg()); 160 oop_map->set_callee_saved(VMRegImpl::stack2reg(r9_off + additional_frame_slots), r9->as_VMReg()); 161 oop_map->set_callee_saved(VMRegImpl::stack2reg(r10_off + additional_frame_slots), r10->as_VMReg()); 162 oop_map->set_callee_saved(VMRegImpl::stack2reg(r12_off + additional_frame_slots), r12->as_VMReg()); 163 oop_map->set_callee_saved(VMRegImpl::stack2reg(r14_off + additional_frame_slots), r14->as_VMReg()); 164 if (hasFPU()) { 165 for (int i = 0; i < FPUStateSizeInWords; ++i) { 166 oop_map->set_callee_saved(VMRegImpl::stack2reg(fpu_state_off + i + additional_frame_slots), 167 as_FloatRegister(i)->as_VMReg()); 168 } 169 } 170 171 return oop_map; 172 } 173 174 void RegisterSaver::restore_live_registers(MacroAssembler* masm, bool restore_lr) { 175 __ pop_CPU_state(); 176 if (restore_lr) 177 __ leave(); 178 else { 179 if (FrameAPCS) 180 Unimplemented(); 181 182 __ sub(sp, rfp, wordSize); 183 __ pop(rfp); 184 } 185 } 186 187 void RegisterSaver::restore_result_registers(MacroAssembler* masm) { 188 189 // Just restore result register. Only used by deoptimization. By 190 // now any callee save register that needs to be restored to a c2 191 // caller of the deoptee has been extracted into the vframeArray 192 // and will be stuffed into the c2i adapter we create for later 193 // restoration so only result registers need to be restored here. 194 195 196 if(hasFPU()) { 197 // Restore fp result register 198 __ vldr_f64(d0, Address(sp, offset_in_bytes(fpu_state_off))); 199 } 200 201 // Restore integer result register 202 __ ldr(r0, Address(sp, offset_in_bytes(r0_off))); 203 __ ldr(r1, Address(sp, offset_in_bytes(r1_off))); 204 205 // Pop all of the register save are off the stack 206 __ add(sp, sp, (reg_save_size + frame::get_frame_size()) * wordSize); 207 } 208 209 // Is vector's size (in bytes) bigger than a size saved by default? 210 // 16 bytes XMM registers are saved by default using fxsave/fxrstor instructions. 211 bool SharedRuntime::is_wide_vector(int size) { 212 return size > 16; 213 } 214 215 size_t SharedRuntime::trampoline_size() { 216 return NativeCall::instruction_size; 217 } 218 219 void SharedRuntime::generate_trampoline(MacroAssembler *masm, address destination) { 220 __ mov(rscratch1, destination); 221 __ b(rscratch1); 222 } 223 // This functions returns offset from fp to java arguments on stack. 224 // 225 // The java_calling_convention describes stack locations as ideal slots on 226 // a frame with no abi restrictions. Since we must observe abi restrictions 227 // (like the placement of the register window) the slots must be biased by 228 // the following value. 229 static int reg2offset_in(VMReg r) { 230 // After stack frame created, fp points to 1 slot after previous sp value. 231 return (r->reg2stack() + 1) * VMRegImpl::stack_slot_size; 232 } 233 234 static int reg2offset_out(VMReg r) { 235 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 236 } 237 238 template <class T> static const T& min (const T& a, const T& b) { 239 return (a > b) ? b : a; 240 } 241 242 // --------------------------------------------------------------------------- 243 // Read the array of BasicTypes from a signature, and compute where the 244 // arguments should go. Values in the VMRegPair regs array refer to 4-byte 245 // quantities. Values less than VMRegImpl::stack0 are registers, those above 246 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer 247 // as framesizes are fixed. 248 // VMRegImpl::stack0 refers to the first slot 0(sp). 249 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register 250 // up to RegisterImpl::number_of_registers) are the 64-bit 251 // integer registers. 252 253 // Note: the INPUTS in sig_bt are in units of Java argument words, 254 // which are 64-bit. The OUTPUTS are in 32-bit units. 255 256 int SharedRuntime::java_calling_convention(const BasicType *sig_bt, 257 VMRegPair *regs, 258 int total_args_passed, 259 int is_outgoing) { 260 261 assert(j_rarg0 == c_rarg0, "assumed"); 262 263 #ifndef HARD_FLOAT_CC 264 if (hasFPU()) { 265 // Create the mapping between argument positions and 266 // registers. 267 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = { 268 j_rarg0, j_rarg1, j_rarg2, j_rarg3 269 }; 270 const uint FP_ArgReg_N = 16; 271 static const FloatRegister FP_ArgReg[] = { 272 f0, f1, f2, f3, 273 f4, f5, f6, f7, 274 f8, f9, f10, f11, 275 f12, f13, f14, f15, 276 }; 277 278 uint int_args = 0; 279 uint fp_args = 0; 280 uint stk_args = 0; 281 282 for (int i = 0; i < total_args_passed; i++) { 283 switch (sig_bt[i]) { 284 case T_FLOAT: 285 if (fp_args < FP_ArgReg_N) { 286 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg()); 287 } else { 288 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 289 stk_args += 1; 290 } 291 break; 292 case T_BOOLEAN: 293 case T_CHAR: 294 case T_BYTE: 295 case T_SHORT: 296 case T_INT: 297 case T_OBJECT: 298 case T_ARRAY: 299 case T_ADDRESS: 300 if (int_args < Argument::n_int_register_parameters_j) { 301 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 302 } else { 303 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 304 stk_args += 1; 305 } 306 break; 307 case T_VOID: 308 // halves of T_LONG or T_DOUBLE 309 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 310 regs[i].set_bad(); 311 break; 312 case T_DOUBLE: 313 assert(sig_bt[i + 1] == T_VOID, "expecting half"); 314 fp_args = align_up(fp_args, 2); 315 if (fp_args < FP_ArgReg_N) { 316 regs[i].set2(FP_ArgReg[fp_args]->as_VMReg()); 317 fp_args += 2; 318 } else { 319 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 320 stk_args += 2; 321 } 322 break; 323 case T_LONG: 324 assert(sig_bt[i + 1] == T_VOID, "expecting half"); 325 if (int_args + 1 < Argument::n_int_register_parameters_j) { 326 if ((int_args % 2) != 0) { 327 ++int_args; 328 } 329 regs[i].set2(INT_ArgReg[int_args]->as_VMReg()); 330 int_args += 2; 331 } else { 332 if (stk_args % 2 != 0) { 333 ++stk_args; 334 } 335 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 336 stk_args += 2; 337 int_args = Argument::n_int_register_parameters_j; 338 } 339 break; 340 default: 341 ShouldNotReachHere(); 342 break; 343 } 344 } 345 346 return align_up(stk_args, StackAlignmentInBytes/wordSize); 347 } else 348 #endif // ndef HARD_FLOAT_CC 349 { 350 // in aarch32 pure soft-float mode the java calling convention is set the same as C one 351 return c_calling_convention(sig_bt, regs, NULL, total_args_passed); 352 } 353 } 354 355 // Patch the callers callsite with entry to compiled code if it exists. 356 static void patch_callers_callsite(MacroAssembler *masm) { 357 Label L; 358 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset()))); 359 __ cbz(rscratch1, L); 360 361 __ enter(); 362 __ push_CPU_state(); 363 364 // VM needs caller's callsite 365 // VM needs target method 366 // This needs to be a long call since we will relocate this adapter to 367 // the codeBuffer and it may not reach 368 369 #ifndef PRODUCT 370 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 371 #endif 372 373 __ mov(c_rarg0, rmethod); 374 __ mov(c_rarg1, lr); 375 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite))); 376 __ bl(rscratch1); 377 __ maybe_isb(); 378 379 __ pop_CPU_state(); 380 // restore sp 381 __ leave(); 382 __ bind(L); 383 } 384 385 static void gen_c2i_adapter(MacroAssembler *masm, 386 int total_args_passed, 387 int comp_args_on_stack, 388 const BasicType *sig_bt, 389 const VMRegPair *regs, 390 Label& skip_fixup) { 391 // Before we get into the guts of the C2I adapter, see if we should be here 392 // at all. We've come from compiled code and are attempting to jump to the 393 // interpreter, which means the caller made a static call to get here 394 // (vcalls always get a compiled target if there is one). Check for a 395 // compiled target. If there is one, we need to patch the caller's call. 396 patch_callers_callsite(masm); 397 398 __ bind(skip_fixup); 399 400 // Since all args are passed on the stack, total_args_passed * 401 // Interpreter::stackElementSize is the space we need. 402 403 const int extraspace = total_args_passed * Interpreter::stackElementSize; 404 const Register compArgPos = lr; 405 int ld_shift = 0; 406 407 __ str(compArgPos, Address(sp, -(extraspace + wordSize))); 408 __ mov(compArgPos, sp); 409 410 // Now write the args into the outgoing interpreter space 411 for (int i = 0; i < total_args_passed; i++) { 412 413 if (sig_bt[i] == T_VOID) { 414 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 415 continue; 416 } 417 418 // next stack slot offset 419 const int next_off = -Interpreter::stackElementSize; 420 421 VMReg r_1 = regs[i].first(); 422 VMReg r_2 = regs[i].second(); 423 if (!r_1->is_valid()) { 424 assert(!r_2->is_valid(), ""); 425 continue; 426 } 427 428 if (r_2->is_valid()) { 429 assert(i + 1 < total_args_passed && sig_bt[i + 1] == T_VOID, "going to overrwrite reg_2 value"); 430 } 431 432 if (r_1->is_stack()) { 433 // memory to memory use rscratch1 434 int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size - ld_shift; 435 if (!r_2->is_valid()) { 436 __ ldr(rscratch1, Address(compArgPos, ld_off)); 437 __ str(rscratch1, Address(sp, next_off, Address::pre)); 438 } else { 439 int tmp_off = ld_off; 440 // ldrd accepts only imm8 441 if(abs(ld_off) > (255 << 2)) { 442 if(__ is_valid_for_imm12(ld_off)) { 443 __ add(compArgPos, compArgPos, ld_off); 444 } else { 445 // add operates encoded imm12, NOT plain 446 __ mov(rscratch1, ld_off); 447 __ add(compArgPos, compArgPos, rscratch1); 448 } 449 tmp_off = 0; 450 ld_shift += ld_off; 451 } 452 __ ldrd(rscratch1, rscratch2, Address(compArgPos, tmp_off)); 453 __ strd(rscratch1, rscratch2, Address(sp, 2* next_off, Address::pre)); 454 } 455 } else if (r_1->is_Register()) { 456 Register r = r_1->as_Register(); 457 assert(r != compArgPos, "compArgPos was modified"); 458 if (!r_2->is_valid()) { 459 __ str(r, Address(sp, next_off, Address::pre)); 460 } else { 461 assert(r_2->as_Register() != compArgPos, "compArgPos was modified"); 462 __ strd(r, r_2->as_Register(), Address(sp, 2 * next_off, Address::pre)); 463 } 464 } else { 465 assert(r_1->is_FloatRegister(), ""); 466 if (!r_2->is_valid()) { 467 // Can't do pre or post addressing for vldr, vstr 468 __ add(sp, sp, next_off); 469 __ vstr_f32(r_1->as_FloatRegister(), Address(sp)); 470 } else { 471 // TODO assert(r_2->is_FloatRegister() && r_2->as_FloatRegister() == r_1->as_FloatRegister() + 1, ""); 472 // Can't do pre or post addressing for vldr, vstr 473 __ add(sp, sp, 2 * next_off); 474 __ vstr_f64(r_1->as_FloatRegister(), Address(sp)); 475 } 476 } 477 } 478 479 // hope, sp is returned to desired value 480 __ ldr(compArgPos, Address(sp, -wordSize)); 481 482 // set sender sp 483 if(__ is_valid_for_imm12(extraspace)) { 484 __ add(r4, sp, extraspace); 485 } else { 486 __ mov(rscratch1, extraspace); 487 __ add(r4, sp, rscratch1); 488 } 489 490 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::interpreter_entry_offset()))); 491 __ b(rscratch1); 492 } 493 494 static void range_check(MacroAssembler* masm, Register pc_reg, Register temp_reg, 495 address code_start, address code_end, 496 Label& L_ok) { 497 Label L_fail; 498 __ lea(temp_reg, ExternalAddress(code_start)); 499 __ cmp(pc_reg, temp_reg); 500 __ b(L_fail, Assembler::LO); 501 __ lea(temp_reg, ExternalAddress(code_end)); 502 __ cmp(pc_reg, temp_reg); 503 __ b(L_ok, Assembler::LO); 504 __ bind(L_fail); 505 } 506 507 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, 508 int total_args_passed, 509 int comp_args_on_stack, 510 const BasicType *sig_bt, 511 const VMRegPair *regs) { 512 513 // Note: r13 contains the senderSP on entry. We must preserve it since 514 // we may do a i2c -> c2i transition if we lose a race where compiled 515 // code goes non-entrant while we get args ready. 516 517 // In addition we use r13 to locate all the interpreter args because 518 // we must align the stack to 16 bytes. 519 520 // Adapters are frameless. 521 522 // An i2c adapter is frameless because the *caller* frame, which is 523 // interpreted, routinely repairs its own sp (from 524 // interpreter_frame_last_sp), even if a callee has modified the 525 // stack pointer. It also recalculates and aligns sp. 526 527 // A c2i adapter is frameless because the *callee* frame, which is 528 // interpreted, routinely repairs its caller's sp (from sender_sp, 529 // which is set up via the senderSP register). 530 531 // In other words, if *either* the caller or callee is interpreted, we can 532 // get the stack pointer repaired after a call. 533 534 // This is why c2i and i2c adapters cannot be indefinitely composed. 535 // In particular, if a c2i adapter were to somehow call an i2c adapter, 536 // both caller and callee would be compiled methods, and neither would 537 // clean up the stack pointer changes performed by the two adapters. 538 // If this happens, control eventually transfers back to the compiled 539 // caller, but with an uncorrected stack, causing delayed havoc. 540 541 if (VerifyAdapterCalls && 542 (Interpreter::code() != NULL || StubRoutines::code1() != NULL)) { 543 // So, let's test for cascading c2i/i2c adapters right now. 544 // assert(Interpreter::contains($return_addr) || 545 // StubRoutines::contains($return_addr), 546 // "i2c adapter must return to an interpreter frame"); 547 __ block_comment("verify_i2c { "); 548 Label L_ok; 549 if (Interpreter::code() != NULL) 550 range_check(masm, lr, rscratch1, 551 Interpreter::code()->code_start(), Interpreter::code()->code_end(), 552 L_ok); 553 if (StubRoutines::code1() != NULL) 554 range_check(masm, lr, rscratch1, 555 StubRoutines::code1()->code_begin(), StubRoutines::code1()->code_end(), 556 L_ok); 557 if (StubRoutines::code2() != NULL) 558 range_check(masm, lr, rscratch1, 559 StubRoutines::code2()->code_begin(), StubRoutines::code2()->code_end(), 560 L_ok); 561 const char* msg = "i2c adapter must return to an interpreter frame"; 562 __ block_comment(msg); 563 __ stop(msg); 564 __ bind(L_ok); 565 __ block_comment("} verify_i2ce "); 566 } 567 568 const int stack_space = align_up(comp_args_on_stack * VMRegImpl::stack_slot_size, StackAlignmentInBytes); 569 const int ld_high = total_args_passed *Interpreter::stackElementSize; 570 // Point to interpreter value (vs. tag) 571 const int next_off = -Interpreter::stackElementSize; // offset from ld ptr 572 const Register loadCounter = lr; 573 574 // Align sp to StackAlignmentInBytes so compiled frame starts always aligned 575 // This is required by APCS, so all native code depends on it. The compiled 576 // Java code is not required to follow this standard however doing so 577 // simplifies the code because allows to have fixed size for compiled frames 578 __ mov(rscratch2, sp); 579 __ align_stack(); 580 if(total_args_passed) { 581 // put below reserved stack space, imm12 should be enough 582 __ str(loadCounter, Address(sp, -(stack_space + wordSize))); 583 584 if(__ is_valid_for_imm12(ld_high)) { 585 __ add(loadCounter, rscratch2, ld_high); 586 } else { 587 // add operates encoded imm12, we need plain 588 __ mov(rscratch1, ld_high); 589 __ add(loadCounter, rscratch2, rscratch1); 590 } 591 } 592 593 if(comp_args_on_stack) { 594 if(__ is_valid_for_imm12(stack_space)) { 595 __ sub(sp, sp, stack_space); 596 } else { 597 // add operates encoded imm12, we need plain 598 __ mov(rscratch1, stack_space); 599 __ sub(sp, sp, rscratch1); 600 } 601 } 602 603 // +------+ -> r4 604 // | 0 | \ 605 // | 1 | \ 606 // | 2 | - > Load in argument order going down. 607 // | x | / 608 // | N | / 609 // +------+ -> inital sp 610 // | pad | maybe 1 word to align the stack to 8 bytes 611 // | M | \ 612 // | x | \ 613 // | 2 | -> Load in argument order going up. 614 // | 1 | / 615 // | 0 | / 616 // +------+ -> 617 618 619 int sp_offset = 0; 620 621 // Now generate the shuffle code. 622 for (int i = 0; i < total_args_passed; i++) { 623 624 if (sig_bt[i] == T_VOID) { 625 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 626 continue; 627 } 628 629 // Pick up 0, 1 or 2 words from SP+offset. 630 631 // 632 // 633 // 634 VMReg r_1 = regs[i].first(); 635 VMReg r_2 = regs[i].second(); 636 if (!r_1->is_valid()) { 637 assert(!r_2->is_valid(), ""); 638 continue; 639 } 640 641 if (r_2->is_valid()) { 642 assert(i + 1 < total_args_passed && sig_bt[i + 1] == T_VOID, "going to overrwrite reg_2 value"); 643 } 644 645 if (r_1->is_stack()) { 646 // Convert stack slot to an SP offset 647 int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size - sp_offset; 648 649 if (!r_2->is_valid()) { 650 __ ldr(rscratch2, Address(loadCounter, next_off, Address::pre)); 651 __ str(rscratch2, Address(sp, st_off)); 652 } else { 653 int tmp_off = st_off; 654 if(abs(st_off) > (255 << 2)) { 655 //st_off doesn't fit imm8 required by strd 656 657 if(__ is_valid_for_imm12(st_off)) { 658 __ add(sp, sp, st_off); 659 } else { 660 // add operates encoded imm12, NOT plain 661 __ mov(rscratch1, st_off); 662 __ add(sp, sp, rscratch1); 663 } 664 tmp_off = 0; 665 sp_offset += st_off; 666 } 667 668 669 // Interpreter local[n] == MSW, local[n+1] == LSW however locals 670 // are accessed as negative so LSW is at LOW address 671 672 // this can be a misaligned move 673 __ ldrd(rscratch1, rscratch2, Address(loadCounter, 2 * next_off, Address::pre)); 674 __ strd(rscratch1, rscratch2, Address(sp, tmp_off)); 675 } 676 } else if (r_1->is_Register()) { // Register argument 677 Register r = r_1->as_Register(); 678 assert(r != loadCounter, "loadCounter is reloaded"); 679 if (r_2->is_valid()) { 680 assert(r_2->as_Register() != loadCounter, "loadCounter is reloaded"); 681 // this can be a misaligned move 682 // ldrd can handle inconsecutive registers 683 __ ldrd(r, r_2->as_Register(), Address(loadCounter, 2 * next_off, Address::pre)); 684 } else { 685 __ ldr(r, Address(loadCounter, next_off, Address::pre)); 686 } 687 } else { 688 assert(r_1->is_FloatRegister(), ""); 689 if (!r_2->is_valid()) { 690 // Can't do pre or post addressing for vldr, vstr 691 __ add(loadCounter, loadCounter, next_off); 692 __ vldr_f32(r_1->as_FloatRegister(), Address(loadCounter)); 693 } else { 694 // TODO assert(r_2->is_FloatRegister() && r_2->as_FloatRegister() == r_1->as_FloatRegister() + 1, ""); 695 // Can't do pre or post addressing for vldr, vstr 696 __ add(loadCounter, loadCounter, 2 * next_off); 697 __ vldr_f64(r_1->as_FloatRegister(), Address(loadCounter)); 698 } 699 } 700 } 701 702 // restore sp 703 if(sp_offset) { 704 if(__ is_valid_for_imm12(sp_offset)) { 705 __ sub(sp, sp, sp_offset); 706 } else { 707 // add operates encoded imm12, we need plain 708 __ mov(rscratch1, sp_offset); 709 __ sub(sp, sp, rscratch1); 710 } 711 } 712 713 if(total_args_passed) { 714 // restore loadCounter 715 __ ldr(loadCounter, Address(sp, -wordSize)); 716 } 717 718 // 6243940 We might end up in handle_wrong_method if 719 // the callee is deoptimized as we race thru here. If that 720 // happens we don't want to take a safepoint because the 721 // caller frame will look interpreted and arguments are now 722 // "compiled" so it is much better to make this transition 723 // invisible to the stack walking code. Unfortunately if 724 // we try and find the callee by normal means a safepoint 725 // is possible. So we stash the desired callee in the thread 726 // and the vm will find there should this case occur. 727 728 __ str(rmethod, Address(rthread, JavaThread::callee_target_offset())); 729 730 // Will jump to the compiled code just as if compiled code was doing it. 731 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::from_compiled_offset()))); 732 __ b(rscratch1); 733 } 734 735 // --------------------------------------------------------------- 736 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, 737 int total_args_passed, 738 int comp_args_on_stack, 739 const BasicType *sig_bt, 740 const VMRegPair *regs, 741 AdapterFingerPrint* fingerprint) { 742 address i2c_entry = __ pc(); 743 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs); 744 745 address c2i_unverified_entry = __ pc(); 746 Label skip_fixup; 747 748 Label ok; 749 750 Register holder = rscratch2; 751 Register receiver = j_rarg0; 752 Register tmp = r8; // A call-clobbered register not used for arg passing 753 754 // ------------------------------------------------------------------------- 755 // Generate a C2I adapter. On entry we know rmethod holds the Method* during calls 756 // to the interpreter. The args start out packed in the compiled layout. They 757 // need to be unpacked into the interpreter layout. This will almost always 758 // require some stack space. We grow the current (compiled) stack, then repack 759 // the args. We finally end in a jump to the generic interpreter entry point. 760 // On exit from the interpreter, the interpreter will restore our SP (lest the 761 // compiled code, which relys solely on SP and not FP, get sick). 762 763 { 764 __ block_comment("c2i_unverified_entry {"); 765 __ load_klass(rscratch1, receiver); 766 __ ldr(tmp, Address(holder, CompiledICHolder::holder_klass_offset())); 767 __ cmp(rscratch1, tmp); 768 __ ldr(rmethod, Address(holder, CompiledICHolder::holder_metadata_offset())); 769 __ b(ok, Assembler::EQ); 770 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 771 772 __ bind(ok); 773 // Method might have been compiled since the call site was patched to 774 // interpreted; if that is the case treat it as a miss so we can get 775 // the call site corrected. 776 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset()))); 777 __ cbz(rscratch1, skip_fixup); 778 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 779 __ block_comment("} c2i_unverified_entry"); 780 } 781 782 address c2i_entry = __ pc(); 783 784 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup); 785 786 __ flush(); 787 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry); 788 } 789 790 int SharedRuntime::c_calling_convention(const BasicType *sig_bt, 791 VMRegPair *regs, 792 VMRegPair *regs2, 793 int total_args_passed) { 794 assert(regs2 == NULL, "not needed on AArch32"); 795 796 // We return the amount of VMRegImpl stack slots we need to reserve for all 797 // the arguments NOT counting out_preserve_stack_slots. 798 799 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = { 800 c_rarg0, c_rarg1, c_rarg2, c_rarg3 801 }; 802 #ifdef HARD_FLOAT_CC 803 const int FP_ArgReg_N = 16; 804 static const FloatRegister FP_ArgReg[] = { 805 f0, f1, f2, f3, 806 f4, f5, f6, f7, 807 f8, f9, f10, f11, 808 f12, f13, f14, f15, 809 }; 810 unsigned long fp_free_mask = (1 << FP_ArgReg_N) - 1; 811 uint fp_args = 0; 812 #endif //HARD_FLOAT_CC 813 814 uint int_args = 0; 815 uint stk_args = 0; 816 817 for (int i = 0; i < total_args_passed; i++) { 818 switch (sig_bt[i]) { 819 case T_BOOLEAN: 820 case T_CHAR: 821 case T_BYTE: 822 case T_SHORT: 823 case T_INT: 824 case T_OBJECT: 825 case T_ARRAY: 826 case T_ADDRESS: 827 case T_METADATA: 828 #ifndef HARD_FLOAT_CC 829 // soft FP case 830 case T_FLOAT: 831 #endif 832 if (int_args < Argument::n_int_register_parameters_c) { 833 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 834 } else { 835 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 836 stk_args += 1; 837 } 838 break; 839 #ifndef HARD_FLOAT_CC 840 // soft FP case 841 case T_DOUBLE: 842 #endif 843 case T_LONG: 844 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 845 if (int_args + 1 < Argument::n_int_register_parameters_c) { 846 if ((int_args % 2) != 0) { 847 ++int_args; 848 } 849 regs[i].set2(INT_ArgReg[int_args]->as_VMReg()); 850 int_args += 2; 851 } else { 852 if (stk_args % 2 != 0) { 853 ++stk_args; 854 } 855 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 856 stk_args += 2; 857 int_args = Argument::n_int_register_parameters_c; 858 } 859 break; 860 #ifdef HARD_FLOAT_CC 861 case T_FLOAT: 862 if (fp_free_mask & ((1 << FP_ArgReg_N)-1)) { 863 unsigned index = __builtin_ctz(fp_free_mask); 864 regs[i].set1(FP_ArgReg[index]->as_VMReg()); 865 fp_free_mask &= ~(1 << index); 866 fp_args += 2 * ((~index) & 1); 867 } else { 868 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 869 stk_args += 1; 870 } 871 break; 872 case T_DOUBLE: 873 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 874 if (fp_args + 1 < FP_ArgReg_N) { 875 fp_free_mask &= ~(3 << fp_args); 876 regs[i].set2(FP_ArgReg[fp_args]->as_VMReg()); 877 fp_args += 2; 878 } else { 879 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 880 stk_args += 2; 881 } 882 break; 883 #endif //HARD_FLOAT_CC 884 case T_VOID: // Halves of longs and doubles 885 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 886 regs[i].set_bad(); 887 break; 888 default: 889 ShouldNotReachHere(); 890 break; 891 } 892 } 893 894 return align_up(stk_args, StackAlignmentInBytes/wordSize); 895 } 896 897 // On 64 bit we will store integer like items to the stack as 898 // 64 bits items (sparc abi) even though java would only store 899 // 32bits for a parameter. On 32bit it will simply be 32 bits 900 // So this routine will do 32->32 on 32bit and 32->64 on 64bit 901 902 static void move_int(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 903 if (src.first()->is_stack()) { 904 if (dst.first()->is_stack()) { 905 // stack to stack 906 __ ldr(rscratch1, Address(rfp, reg2offset_in(src.first()))); 907 __ str(rscratch1, Address(sp, reg2offset_out(dst.first()))); 908 } else { 909 // stack to reg 910 __ ldr(dst.first()->as_Register(), Address(rfp, reg2offset_in(src.first()))); 911 } 912 } else if (dst.first()->is_stack()) { 913 // reg to stack 914 __ str(src.first()->as_Register(), Address(sp, reg2offset_out(dst.first()))); 915 } else { 916 if (dst.first() != src.first()) { 917 __ mov(dst.first()->as_Register(), src.first()->as_Register()); 918 } 919 } 920 } 921 922 // An oop arg. Must pass a handle not the oop itself 923 static void object_move(MacroAssembler* masm, 924 OopMap* map, 925 int oop_handle_offset, 926 int framesize_in_slots, 927 VMRegPair src, 928 VMRegPair dst, 929 bool is_receiver, 930 int* receiver_offset) { 931 932 // must pass a handle. First figure out the location we use as a handle 933 934 Register rHandle = dst.first()->is_stack() ? rscratch2 : dst.first()->as_Register(); 935 936 // See if oop is NULL if it is we need no handle 937 938 if (src.first()->is_stack()) { 939 940 // Oop is already on the stack as an argument 941 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 942 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots)); 943 if (is_receiver) { 944 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size; 945 } 946 947 __ ldr(rscratch1, Address(rfp, reg2offset_in(src.first()))); 948 __ lea(rHandle, Address(rfp, reg2offset_in(src.first()))); 949 // conditionally move a NULL 950 __ cmp(rscratch1, 0); 951 __ mov(rHandle, 0, Assembler::EQ); 952 } else { 953 954 // Oop is in an a register we must store it to the space we reserve 955 // on the stack for oop_handles and pass a handle if oop is non-NULL 956 957 const Register rOop = src.first()->as_Register(); 958 int oop_slot; 959 if (rOop == j_rarg0) 960 oop_slot = 0; 961 else if (rOop == j_rarg1) 962 oop_slot = 1; 963 else if (rOop == j_rarg2) 964 oop_slot = 2; 965 else { 966 assert(rOop == j_rarg3, "wrong register"); 967 oop_slot = 3; 968 } 969 970 oop_slot = oop_slot * VMRegImpl::slots_per_word + oop_handle_offset; 971 int offset = oop_slot*VMRegImpl::stack_slot_size; 972 973 map->set_oop(VMRegImpl::stack2reg(oop_slot)); 974 // Store oop in handle area, may be NULL 975 __ str(rOop, Address(sp, offset)); 976 if (is_receiver) { 977 *receiver_offset = offset; 978 } 979 980 __ cmp(rOop, 0); 981 __ lea(rHandle, Address(sp, offset)); 982 // conditionally move a NULL 983 __ mov(rHandle, 0, Assembler::EQ); 984 } 985 986 // If arg is on the stack then place it otherwise it is already in correct reg. 987 if (dst.first()->is_stack()) { 988 __ str(rHandle, Address(sp, reg2offset_out(dst.first()))); 989 } 990 } 991 992 // A float arg may have to do float reg int reg conversion 993 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 994 if(hasFPU()) { 995 if (src.first()->is_stack()) { 996 if (dst.first()->is_stack()) { 997 // stack to stack 998 // Have no vfp scratch registers, so copy via gpr 999 __ ldr(rscratch1, Address(rfp, reg2offset_in(src.first()))); 1000 __ str(rscratch1, Address(sp, reg2offset_out(dst.first()))); 1001 } else { 1002 // stack to reg 1003 __ vldr_f32(dst.first()->as_FloatRegister(), Address(rfp, reg2offset_in(src.first()))); 1004 } 1005 } else if (dst.first()->is_stack()) { 1006 // reg to stack 1007 __ vstr_f32(src.first()->as_FloatRegister(), Address(sp, reg2offset_out(dst.first()))); 1008 } else { 1009 #ifndef HARD_FLOAT_CC 1010 if(dst.first()->is_Register()) { 1011 __ vmov_f32(dst.first()->as_Register(), src.first()->as_FloatRegister()); 1012 } else 1013 #endif 1014 if (dst.first() != src.first()) { 1015 __ vmov_f32(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 1016 } 1017 } 1018 } else { 1019 move_int(masm, src, dst); 1020 } 1021 } 1022 1023 // A long move 1024 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1025 if (src.first()->is_stack()) { 1026 if (dst.first()->is_stack()) { 1027 // stack to stack 1028 __ ldrd(rscratch1, rscratch2, Address(rfp, reg2offset_in(src.first()))); 1029 __ strd(rscratch1, rscratch2, Address(sp, reg2offset_out(dst.first()))); 1030 } else { 1031 // stack to reg 1032 __ ldrd(dst.first()->as_Register(), dst.second()->as_Register(), 1033 Address(rfp, reg2offset_in(src.first()))); 1034 } 1035 } else if (dst.first()->is_stack()) { 1036 // reg to stack 1037 __ strd(src.first()->as_Register(), src.second()->as_Register(), 1038 Address(sp, reg2offset_out(dst.first()))); 1039 } else { 1040 // reg to reg 1041 if (dst.first() != src.first()) { 1042 if (dst.first() != src.second()) { 1043 __ mov(dst.first()->as_Register(), src.first()->as_Register()); 1044 __ mov(dst.second()->as_Register(), src.second()->as_Register()); 1045 } else { 1046 __ mov(dst.second()->as_Register(), src.second()->as_Register()); 1047 __ mov(dst.first()->as_Register(), src.first()->as_Register()); 1048 } 1049 } 1050 } 1051 } 1052 1053 // A double move 1054 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1055 if(hasFPU()) { 1056 if (src.first()->is_stack()) { 1057 if (dst.first()->is_stack()) { 1058 // stack to stack 1059 // Have no vfp scratch registers, so copy via gpr 1060 __ ldrd(rscratch1, rscratch2, Address(rfp, reg2offset_in(src.first()))); 1061 __ strd(rscratch1, rscratch2, Address(sp, reg2offset_out(dst.first()))); 1062 } else { 1063 // stack to reg 1064 __ vldr_f64(dst.first()->as_FloatRegister(), Address(rfp, reg2offset_in(src.first()))); 1065 } 1066 } else if (dst.first()->is_stack()) { 1067 // reg to stack 1068 __ vstr_f64(src.first()->as_FloatRegister(), Address(sp, reg2offset_out(dst.first()))); 1069 } else { 1070 #ifndef HARD_FLOAT_CC 1071 if(dst.first()->is_Register()) { 1072 __ vmov_f64(dst.first()->as_Register(), dst.second()->as_Register(), src.first()->as_FloatRegister()); 1073 } else 1074 #endif 1075 if (dst.first() != src.first()) { 1076 __ vmov_f64(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 1077 } 1078 } 1079 } else { 1080 long_move(masm, src, dst); 1081 } 1082 } 1083 1084 1085 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1086 // We always ignore the frame_slots arg and just use the space just below frame pointer 1087 // which by this time is free to use 1088 switch (ret_type) { 1089 case T_DOUBLE: 1090 #ifdef HARD_FLOAT_CC 1091 __ vstr_f64(d0, Address(rfp, -(frame::get_frame_size() + 1) * wordSize)); 1092 break; 1093 #endif//fall through otherwise 1094 case T_LONG: 1095 __ strd(r0, r1, Address(rfp, -(frame::get_frame_size() + 1) * wordSize)); 1096 break; 1097 case T_VOID: 1098 break; 1099 case T_FLOAT: 1100 #ifdef HARD_FLOAT_CC 1101 __ vstr_f32(f0, Address(rfp, -frame::get_frame_size() * wordSize)); 1102 break; 1103 #endif//fall through otherwise 1104 default: 1105 __ str(r0, Address(rfp, -frame::get_frame_size() * wordSize)); 1106 break; 1107 } 1108 } 1109 1110 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1111 // We always ignore the frame_slots arg and just use the space just below frame pointer 1112 // which by this time is free to use 1113 switch (ret_type) { 1114 case T_DOUBLE: 1115 #ifdef HARD_FLOAT_CC 1116 __ vldr_f64(d0, Address(rfp, -(frame::get_frame_size() + 1) * wordSize)); 1117 break; 1118 #endif//fall through otherwise 1119 case T_LONG: 1120 __ ldrd(r0, r1, Address(rfp, -(frame::get_frame_size() + 1) * wordSize)); 1121 break; 1122 case T_VOID: 1123 break; 1124 case T_FLOAT: 1125 #ifdef HARD_FLOAT_CC 1126 __ vldr_f32(d0, Address(rfp, -frame::get_frame_size() * wordSize)); 1127 break; 1128 #endif//fall through otherwise 1129 default: 1130 __ ldr(r0, Address(rfp, -frame::get_frame_size() * wordSize)); 1131 break; 1132 } 1133 } 1134 1135 static int save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) { 1136 RegSet x; 1137 int saved_slots = 0; 1138 for ( int i = first_arg ; i < arg_count ; i++ ) { 1139 if (args[i].first()->is_Register()) { 1140 x = x + args[i].first()->as_Register(); 1141 ++saved_slots; 1142 } 1143 if (args[i].second()->is_Register()) { 1144 x = x + args[i].second()->as_Register(); 1145 ++saved_slots; 1146 } 1147 #ifdef HARD_FLOAT_CC 1148 else if (args[i].first()->is_FloatRegister()) { 1149 FloatRegister fr = args[i].first()->as_FloatRegister(); 1150 1151 if (args[i].second()->is_FloatRegister()) { 1152 assert(args[i].is_single_phys_reg(), "doubles should be 2 consequents float regs"); 1153 __ decrement(sp, 2 * wordSize); 1154 __ vstr_f64(fr, Address(sp)); 1155 saved_slots += 2; 1156 } else { 1157 __ decrement(sp, wordSize); 1158 __ vstr_f32(fr, Address(sp)); 1159 ++saved_slots; 1160 } 1161 } 1162 #endif//HARD_FLOAT_CC 1163 } 1164 __ push(x, sp); 1165 return saved_slots; 1166 } 1167 1168 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) { 1169 RegSet x; 1170 for ( int i = first_arg ; i < arg_count ; i++ ) { 1171 if (args[i].first()->is_Register()) { 1172 x = x + args[i].first()->as_Register(); 1173 } else { 1174 ; 1175 } 1176 if (args[i].second()->is_Register()) { 1177 x = x + args[i].second()->as_Register(); 1178 } 1179 } 1180 __ pop(x, sp); 1181 for ( int i = first_arg ; i < arg_count ; i++ ) { 1182 if (args[i].first()->is_Register()) { 1183 ; 1184 } 1185 #ifdef HARD_FLOAT_CC 1186 else if (args[i].first()->is_FloatRegister()) { 1187 FloatRegister fr = args[i].first()->as_FloatRegister(); 1188 1189 if (args[i].second()->is_FloatRegister()) { 1190 assert(args[i].is_single_phys_reg(), "doubles should be 2 consequents float regs"); 1191 __ vstr_f64(fr, Address(sp)); 1192 __ increment(sp, 2 * wordSize); 1193 } else { 1194 __ vstr_f32(fr, Address(sp)); 1195 __ increment(sp, wordSize); 1196 } 1197 } 1198 #endif//HARD_FLOAT_CC 1199 } 1200 } 1201 1202 1203 // Check GCLocker::needs_gc and enter the runtime if it's true. This 1204 // keeps a new JNI critical region from starting until a GC has been 1205 // forced. Save down any oops in registers and describe them in an 1206 // OopMap. 1207 static void check_needs_gc_for_critical_native(MacroAssembler* masm, 1208 int stack_slots, 1209 int total_c_args, 1210 int total_in_args, 1211 int arg_save_area, 1212 OopMapSet* oop_maps, 1213 VMRegPair* in_regs, 1214 BasicType* in_sig_bt) { Unimplemented(); } 1215 1216 // Unpack an array argument into a pointer to the body and the length 1217 // if the array is non-null, otherwise pass 0 for both. 1218 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, VMRegPair body_arg, VMRegPair length_arg) { Unimplemented(); } 1219 1220 1221 class ComputeMoveOrder: public StackObj { 1222 class MoveOperation: public ResourceObj { 1223 friend class ComputeMoveOrder; 1224 private: 1225 VMRegPair _src; 1226 VMRegPair _dst; 1227 int _src_index; 1228 int _dst_index; 1229 bool _processed; 1230 MoveOperation* _next; 1231 MoveOperation* _prev; 1232 1233 static int get_id(VMRegPair r) { Unimplemented(); return 0; } 1234 1235 public: 1236 MoveOperation(int src_index, VMRegPair src, int dst_index, VMRegPair dst): 1237 _src(src) 1238 , _src_index(src_index) 1239 , _dst(dst) 1240 , _dst_index(dst_index) 1241 , _next(NULL) 1242 , _prev(NULL) 1243 , _processed(false) { Unimplemented(); } 1244 1245 VMRegPair src() const { Unimplemented(); return _src; } 1246 int src_id() const { Unimplemented(); return 0; } 1247 int src_index() const { Unimplemented(); return 0; } 1248 VMRegPair dst() const { Unimplemented(); return _src; } 1249 void set_dst(int i, VMRegPair dst) { Unimplemented(); } 1250 int dst_index() const { Unimplemented(); return 0; } 1251 int dst_id() const { Unimplemented(); return 0; } 1252 MoveOperation* next() const { Unimplemented(); return 0; } 1253 MoveOperation* prev() const { Unimplemented(); return 0; } 1254 void set_processed() { Unimplemented(); } 1255 bool is_processed() const { Unimplemented(); return 0; } 1256 1257 // insert 1258 void break_cycle(VMRegPair temp_register) { Unimplemented(); } 1259 1260 void link(GrowableArray<MoveOperation*>& killer) { Unimplemented(); } 1261 }; 1262 1263 private: 1264 GrowableArray<MoveOperation*> edges; 1265 1266 public: 1267 ComputeMoveOrder(int total_in_args, VMRegPair* in_regs, int total_c_args, VMRegPair* out_regs, 1268 BasicType* in_sig_bt, GrowableArray<int>& arg_order, VMRegPair tmp_vmreg) { Unimplemented(); } 1269 1270 // Collected all the move operations 1271 void add_edge(int src_index, VMRegPair src, int dst_index, VMRegPair dst) { Unimplemented(); } 1272 1273 // Walk the edges breaking cycles between moves. The result list 1274 // can be walked in order to produce the proper set of loads 1275 GrowableArray<MoveOperation*>* get_store_order(VMRegPair temp_register) { Unimplemented(); return 0; } 1276 }; 1277 1278 1279 static void rt_call(MacroAssembler* masm, address dest) { 1280 CodeBlob *cb = CodeCache::find_blob(dest); 1281 if (cb) { 1282 __ far_call(RuntimeAddress(dest), NULL); 1283 } else { 1284 __ lea(rscratch2, RuntimeAddress(dest)); 1285 __ bl(rscratch2); 1286 __ maybe_isb(); 1287 } 1288 } 1289 1290 static void verify_oop_args(MacroAssembler* masm, 1291 const methodHandle &method, 1292 const BasicType* sig_bt, 1293 const VMRegPair* regs) { 1294 Register temp_reg = rscratch2; // not part of any compiled calling seq 1295 if (VerifyOops) { 1296 for (int i = 0; i < method->size_of_parameters(); i++) { 1297 if (sig_bt[i] == T_OBJECT || 1298 sig_bt[i] == T_ARRAY) { 1299 VMReg r = regs[i].first(); 1300 assert(r->is_valid(), "bad oop arg"); 1301 if (r->is_stack()) { 1302 __ ldr(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size)); 1303 __ verify_oop(temp_reg); 1304 } else { 1305 __ verify_oop(r->as_Register()); 1306 } 1307 } 1308 } 1309 } 1310 } 1311 1312 static void gen_special_dispatch(MacroAssembler* masm, 1313 const methodHandle &method, 1314 const BasicType* sig_bt, 1315 const VMRegPair* regs) { 1316 verify_oop_args(masm, method, sig_bt, regs); 1317 vmIntrinsics::ID iid = method->intrinsic_id(); 1318 1319 // Now write the args into the outgoing interpreter space 1320 bool has_receiver = false; 1321 Register receiver_reg = noreg; 1322 int member_arg_pos = -1; 1323 Register member_reg = noreg; 1324 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid); 1325 if (ref_kind != 0) { 1326 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument 1327 member_reg = r4; 1328 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind); 1329 } else if (iid == vmIntrinsics::_invokeBasic) { 1330 has_receiver = true; 1331 } else { 1332 fatal("unexpected intrinsic id %d", iid); 1333 } 1334 1335 if (member_reg != noreg) { 1336 // Load the member_arg into register, if necessary. 1337 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs); 1338 VMReg r = regs[member_arg_pos].first(); 1339 if (r->is_stack()) { 1340 __ ldr(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size)); 1341 } else { 1342 // no data motion is needed 1343 member_reg = r->as_Register(); 1344 } 1345 } 1346 1347 if (has_receiver) { 1348 // Make sure the receiver is loaded into a register. 1349 assert(method->size_of_parameters() > 0, "oob"); 1350 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object"); 1351 VMReg r = regs[0].first(); 1352 assert(r->is_valid(), "bad receiver arg"); 1353 if (r->is_stack()) { 1354 // Porting note: This assumes that compiled calling conventions always 1355 // pass the receiver oop in a register. If this is not true on some 1356 // platform, pick a temp and load the receiver from stack. 1357 fatal("receiver always in a register"); 1358 } else { 1359 // no data motion is needed 1360 receiver_reg = r->as_Register(); 1361 } 1362 } 1363 1364 // Figure out which address we are really jumping to: 1365 MethodHandles::generate_method_handle_dispatch(masm, iid, 1366 receiver_reg, member_reg, /*for_compiler_entry:*/ true); 1367 } 1368 1369 // --------------------------------------------------------------------------- 1370 // Generate a native wrapper for a given method. The method takes arguments 1371 // in the Java compiled code convention, marshals them to the native 1372 // convention (handlizes oops, etc), transitions to native, makes the call, 1373 // returns to java state (possibly blocking), unhandlizes any result and 1374 // returns. 1375 // 1376 // Critical native functions are a shorthand for the use of 1377 // GetPrimtiveArrayCritical and disallow the use of any other JNI 1378 // functions. The wrapper is expected to unpack the arguments before 1379 // passing them to the callee and perform checks before and after the 1380 // native call to ensure that they GC_locker 1381 // lock_critical/unlock_critical semantics are followed. Some other 1382 // parts of JNI setup are skipped like the tear down of the JNI handle 1383 // block and the check for pending exceptions it's impossible for them 1384 // to be thrown. 1385 // 1386 // They are roughly structured like this: 1387 // if (GC_locker::needs_gc()) 1388 // SharedRuntime::block_for_jni_critical(); 1389 // tranistion to thread_in_native 1390 // unpack arrray arguments and call native entry point 1391 // check for safepoint in progress 1392 // check if any thread suspend flags are set 1393 // call into JVM and possible unlock the JNI critical 1394 // if a GC was suppressed while in the critical native. 1395 // transition back to thread_in_Java 1396 // return to caller 1397 // 1398 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm, 1399 const methodHandle& method, 1400 int compile_id, 1401 BasicType* in_sig_bt, 1402 VMRegPair* in_regs, 1403 BasicType ret_type) { 1404 if (method->is_method_handle_intrinsic()) { 1405 vmIntrinsics::ID iid = method->intrinsic_id(); 1406 intptr_t start = (intptr_t)__ pc(); 1407 int vep_offset = ((intptr_t)__ pc()) - start; 1408 1409 // First instruction must be a nop as it may need to be patched on deoptimisation 1410 __ nop(); 1411 gen_special_dispatch(masm, 1412 method, 1413 in_sig_bt, 1414 in_regs); 1415 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period 1416 __ flush(); 1417 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually 1418 return nmethod::new_native_nmethod(method, 1419 compile_id, 1420 masm->code(), 1421 vep_offset, 1422 frame_complete, 1423 stack_slots / VMRegImpl::slots_per_word, 1424 in_ByteSize(-1), 1425 in_ByteSize(-1), 1426 (OopMapSet*)NULL); 1427 } 1428 1429 bool is_critical_native = true; 1430 address native_func = method->critical_native_function(); 1431 if (native_func == NULL) { 1432 native_func = method->native_function(); 1433 is_critical_native = false; 1434 } 1435 assert(native_func != NULL, "must have function"); 1436 1437 // An OopMap for lock (and class if static) 1438 OopMapSet *oop_maps = new OopMapSet(); 1439 intptr_t start = (intptr_t)__ pc(); 1440 1441 // We have received a description of where all the java arg are located 1442 // on entry to the wrapper. We need to convert these args to where 1443 // the jni function will expect them. To figure out where they go 1444 // we convert the java signature to a C signature by inserting 1445 // the hidden arguments as arg[0] and possibly arg[1] (static method) 1446 1447 const int total_in_args = method->size_of_parameters(); 1448 int total_c_args = total_in_args; 1449 if (!is_critical_native) { 1450 total_c_args += 1; 1451 if (method->is_static()) { 1452 total_c_args++; 1453 } 1454 } else { 1455 for (int i = 0; i < total_in_args; i++) { 1456 if (in_sig_bt[i] == T_ARRAY) { 1457 total_c_args++; 1458 } 1459 } 1460 } 1461 1462 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1463 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1464 BasicType* in_elem_bt = NULL; 1465 1466 int argc = 0; 1467 if (!is_critical_native) { 1468 out_sig_bt[argc++] = T_ADDRESS; 1469 if (method->is_static()) { 1470 out_sig_bt[argc++] = T_OBJECT; 1471 } 1472 1473 for (int i = 0; i < total_in_args ; i++ ) { 1474 out_sig_bt[argc++] = in_sig_bt[i]; 1475 } 1476 } else { 1477 Thread* THREAD = Thread::current(); 1478 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, total_in_args); 1479 SignatureStream ss(method->signature()); 1480 for (int i = 0; i < total_in_args ; i++ ) { 1481 if (in_sig_bt[i] == T_ARRAY) { 1482 // Arrays are passed as int, elem* pair 1483 out_sig_bt[argc++] = T_INT; 1484 out_sig_bt[argc++] = T_ADDRESS; 1485 Symbol* atype = ss.as_symbol(CHECK_NULL); 1486 const char* at = atype->as_C_string(); 1487 if (strlen(at) == 2) { 1488 assert(at[0] == '[', "must be"); 1489 switch (at[1]) { 1490 case 'B': in_elem_bt[i] = T_BYTE; break; 1491 case 'C': in_elem_bt[i] = T_CHAR; break; 1492 case 'D': in_elem_bt[i] = T_DOUBLE; break; 1493 case 'F': in_elem_bt[i] = T_FLOAT; break; 1494 case 'I': in_elem_bt[i] = T_INT; break; 1495 case 'J': in_elem_bt[i] = T_LONG; break; 1496 case 'S': in_elem_bt[i] = T_SHORT; break; 1497 case 'Z': in_elem_bt[i] = T_BOOLEAN; break; 1498 default: ShouldNotReachHere(); 1499 } 1500 } 1501 } else { 1502 out_sig_bt[argc++] = in_sig_bt[i]; 1503 in_elem_bt[i] = T_VOID; 1504 } 1505 if (in_sig_bt[i] != T_VOID) { 1506 assert(in_sig_bt[i] == ss.type(), "must match"); 1507 ss.next(); 1508 } 1509 } 1510 } 1511 1512 // Now figure out where the args must be stored and how much stack space 1513 // they require. 1514 int out_arg_slots; 1515 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, NULL, total_c_args); 1516 1517 // Compute framesize for the wrapper. We need to handlize all oops in 1518 // incoming registers 1519 1520 // Calculate the total number of stack slots we will need. 1521 1522 // First count the abi requirement plus all of the outgoing args 1523 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots; 1524 1525 // Now the space for the inbound oop handle area 1526 int total_save_slots = -1; 1527 if (is_critical_native) { 1528 // Critical natives may have to call out so they need a save area 1529 // for register arguments. 1530 int double_slots = 0; 1531 int single_slots = 0; 1532 for ( int i = 0; i < total_in_args; i++) { 1533 if (in_regs[i].first()->is_Register()) { 1534 const Register reg = in_regs[i].first()->as_Register(); 1535 switch (in_sig_bt[i]) { 1536 case T_ARRAY: // critical array (uses 2 slots on LP64) 1537 case T_BOOLEAN: 1538 case T_BYTE: 1539 case T_SHORT: 1540 case T_CHAR: 1541 case T_INT: single_slots++; break; 1542 case T_LONG: double_slots++; break; 1543 default: ShouldNotReachHere(); 1544 } 1545 } else 1546 #ifdef HARD_FLOAT_CC 1547 if (in_regs[i].first()->is_FloatRegister()) 1548 #endif // HARD_FLOAT_CC 1549 ShouldNotReachHere(); 1550 } 1551 total_save_slots = double_slots * 2 + single_slots; 1552 // align the save area 1553 if (double_slots != 0) { 1554 stack_slots = align_up(stack_slots, 2); 1555 } 1556 } else { 1557 total_save_slots = 4 * VMRegImpl::slots_per_word; // 4 arguments passed in registers 1558 } 1559 assert(total_save_slots != -1, "initialize total_save_slots!"); 1560 1561 int oop_handle_offset = stack_slots; 1562 stack_slots += total_save_slots; 1563 1564 // Now any space we need for handlizing a klass if static method 1565 1566 int klass_slot_offset = 0; 1567 int klass_offset = -1; 1568 int lock_slot_offset = 0; 1569 bool is_static = false; 1570 1571 if (method->is_static()) { 1572 klass_slot_offset = stack_slots; 1573 stack_slots += VMRegImpl::slots_per_word; 1574 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; 1575 is_static = true; 1576 } 1577 1578 // Plus a lock if needed 1579 1580 if (method->is_synchronized()) { 1581 lock_slot_offset = stack_slots; 1582 stack_slots += VMRegImpl::slots_per_word; 1583 } 1584 1585 const int enter_frame_size = frame::get_frame_size(); 1586 1587 // Now a place (+2) to save return values or temp during shuffling 1588 // + {2,4} words which pushed by enter() 1589 // (return address (which we own), saved rfp, ...) 1590 stack_slots += 2 + enter_frame_size; 1591 1592 // Ok The space we have allocated will look like: 1593 // 1594 // 1595 // FP-> | saved lr | 1596 // |---------------------| 1597 // | saved fp | 1598 // |---------------------| 1599 // | 2 slots for moves | 1600 // |.....................| 1601 // | 1 slot opt padding | 1602 // |---------------------| 1603 // | lock box (if sync) | 1604 // |---------------------| <- lock_slot_offset 1605 // | klass (if static) | 1606 // |---------------------| <- klass_slot_offset 1607 // | oopHandle area | 1608 // |---------------------| <- oop_handle_offset (8 java arg registers) 1609 // | outbound memory | 1610 // | based arguments | 1611 // | | 1612 // |---------------------| 1613 // | | 1614 // SP-> | out_preserved_slots | 1615 // 1616 // 1617 1618 1619 // Now compute actual number of stack words we need rounding to make 1620 // stack properly aligned. 1621 stack_slots = align_up(stack_slots, StackAlignmentInSlots); 1622 1623 int stack_size = stack_slots * VMRegImpl::stack_slot_size; 1624 1625 // First thing make an ic check to see if we should even be here 1626 1627 // We are free to use all registers as temps without saving them and 1628 // restoring them except rfp. rfp is the only callee save register 1629 // as far as the interpreter and the compiler(s) are concerned. 1630 1631 1632 const Register ic_reg = rscratch2; 1633 const Register receiver = j_rarg0; 1634 1635 Label hit; 1636 Label exception_pending; 1637 1638 assert_different_registers(ic_reg, receiver, rscratch1); 1639 __ verify_oop(receiver); 1640 __ cmp_klass(receiver, ic_reg, rscratch1); 1641 __ b(hit, Assembler::EQ); 1642 1643 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 1644 1645 // Verified entry point must be aligned 1646 __ align(8); 1647 1648 __ bind(hit); 1649 1650 #ifdef ASSERT 1651 __ mov(ic_reg, 0xdead); // trash ic_reg(rscratch2), as used as real scratch further 1652 #endif 1653 1654 int vep_offset = ((intptr_t)__ pc()) - start; 1655 1656 // Generate stack overflow check 1657 1658 // If we have to make this method not-entrant we'll overwrite its 1659 // first instruction with a jump. For this action to be legal we 1660 // must ensure that this first instruction is a B, BL, NOP, BKPT, 1661 // SVC, HVC, or SMC. Make it a NOP. 1662 __ nop(); 1663 1664 if (UseStackBanging) { 1665 __ bang_stack_with_offset(JavaThread::stack_shadow_zone_size()); 1666 } else { 1667 Unimplemented(); 1668 } 1669 1670 // Generate a new frame for the wrapper. 1671 __ enter(); 1672 // some words are pushed by enter, so adjust frame size on this value 1673 __ sub(sp, sp, stack_size - enter_frame_size * wordSize); 1674 1675 // Frame is now completed as far as size and linkage. 1676 int frame_complete = ((intptr_t)__ pc()) - start; 1677 1678 if (is_critical_native) { 1679 check_needs_gc_for_critical_native(masm, stack_slots, total_c_args, total_in_args, 1680 oop_handle_offset, oop_maps, in_regs, in_sig_bt); 1681 } 1682 1683 // 1684 // We immediately shuffle the arguments so that any vm call we have to 1685 // make from here on out (sync slow path, jvmti, etc.) we will have 1686 // captured the oops from our caller and have a valid oopMap for 1687 // them. 1688 1689 // ----------------- 1690 // The Grand Shuffle 1691 1692 // The Java calling convention is either equal (linux) or denser (win64) than the 1693 // c calling convention. However the because of the jni_env argument the c calling 1694 // convention always has at least one more (and two for static) arguments than Java. 1695 // Therefore if we move the args from java -> c backwards then we will never have 1696 // a register->register conflict and we don't have to build a dependency graph 1697 // and figure out how to break any cycles. 1698 // 1699 1700 // Record sp-based slot for receiver on stack for non-static methods 1701 int receiver_offset = -1; 1702 1703 // This is a trick. We double the stack slots so we can claim 1704 // the oops in the caller's frame. Since we are sure to have 1705 // more args than the caller doubling is enough to make 1706 // sure we can capture all the incoming oop args from the 1707 // caller. 1708 // 1709 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1710 1711 // Mark location of rfp (someday) 1712 // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, vmreg(rfp)); 1713 1714 1715 #ifdef ASSERT 1716 bool reg_destroyed[RegisterImpl::number_of_registers]; 1717 bool freg_destroyed[FloatRegisterImpl::number_of_registers]; 1718 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) { 1719 reg_destroyed[r] = false; 1720 } 1721 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) { 1722 freg_destroyed[f] = false; 1723 } 1724 1725 #endif // ASSERT 1726 1727 // This may iterate in two different directions depending on the 1728 // kind of native it is. The reason is that for regular JNI natives 1729 // the incoming and outgoing registers are offset upwards and for 1730 // critical natives they are offset down. 1731 GrowableArray<int> arg_order(2 * total_in_args); 1732 VMRegPair tmp_vmreg; 1733 tmp_vmreg.set2(rscratch2->as_VMReg()); 1734 1735 if (!is_critical_native) { 1736 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) { 1737 arg_order.push(i); 1738 arg_order.push(c_arg); 1739 } 1740 } else { 1741 // Compute a valid move order, using tmp_vmreg to break any cycles 1742 ComputeMoveOrder cmo(total_in_args, in_regs, total_c_args, out_regs, in_sig_bt, arg_order, tmp_vmreg); 1743 } 1744 1745 int temploc = -1; 1746 for (int ai = 0; ai < arg_order.length(); ai += 2) { 1747 int i = arg_order.at(ai); 1748 int c_arg = arg_order.at(ai + 1); 1749 __ block_comment(err_msg("move %d -> %d", i, c_arg)); 1750 if (c_arg == -1) { 1751 assert(is_critical_native, "should only be required for critical natives"); 1752 // This arg needs to be moved to a temporary 1753 __ mov(tmp_vmreg.first()->as_Register(), in_regs[i].first()->as_Register()); 1754 in_regs[i] = tmp_vmreg; 1755 temploc = i; 1756 continue; 1757 } else if (i == -1) { 1758 assert(is_critical_native, "should only be required for critical natives"); 1759 // Read from the temporary location 1760 assert(temploc != -1, "must be valid"); 1761 i = temploc; 1762 temploc = -1; 1763 } 1764 #ifdef ASSERT 1765 if (in_regs[i].first()->is_Register()) { 1766 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!"); 1767 } else if (in_regs[i].first()->is_FloatRegister()) { 1768 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!"); 1769 } 1770 if (out_regs[c_arg].first()->is_Register()) { 1771 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true; 1772 } else if (out_regs[c_arg].first()->is_FloatRegister()) { 1773 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true; 1774 } 1775 #endif // ASSERT 1776 switch (in_sig_bt[i]) { 1777 case T_ARRAY: 1778 if (is_critical_native) { 1779 unpack_array_argument(masm, in_regs[i], in_elem_bt[i], out_regs[c_arg + 1], out_regs[c_arg]); 1780 c_arg++; 1781 #ifdef ASSERT 1782 if (out_regs[c_arg].first()->is_Register()) { 1783 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true; 1784 } else if (out_regs[c_arg].first()->is_FloatRegister()) { 1785 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true; 1786 } 1787 #endif 1788 break; 1789 } 1790 case T_OBJECT: 1791 assert(!is_critical_native, "no oop arguments"); 1792 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg], 1793 ((i == 0) && (!is_static)), 1794 &receiver_offset); 1795 break; 1796 case T_VOID: 1797 break; 1798 1799 case T_FLOAT: 1800 float_move(masm, in_regs[i], out_regs[c_arg]); 1801 break; 1802 1803 case T_DOUBLE: 1804 assert( i + 1 < total_in_args && 1805 in_sig_bt[i + 1] == T_VOID && 1806 out_sig_bt[c_arg+1] == T_VOID, "bad arg list"); 1807 double_move(masm, in_regs[i], out_regs[c_arg]); 1808 break; 1809 1810 case T_LONG : 1811 long_move(masm, in_regs[i], out_regs[c_arg]); 1812 break; 1813 1814 case T_BOOLEAN : 1815 case T_BYTE : 1816 case T_CHAR : 1817 case T_SHORT : 1818 case T_INT : 1819 move_int(masm, in_regs[i], out_regs[c_arg]); 1820 break; 1821 1822 case T_ADDRESS: assert(false, "found T_ADDRESS in java args"); 1823 case T_NARROWOOP : 1824 case T_METADATA : 1825 case T_NARROWKLASS : 1826 default: 1827 ShouldNotReachHere(); 1828 } 1829 } 1830 1831 // point c_arg at the first arg that is already loaded in case we 1832 // need to spill before we call out 1833 int c_arg = total_c_args - total_in_args; 1834 1835 // We use r4 as the oop handle for the receiver/klass 1836 // It is callee save so it survives the call to native 1837 1838 const Register oop_handle_reg = r4; 1839 1840 // Pre-load a static method's oop. Used both by locking code and 1841 // the normal JNI call code. 1842 if (method->is_static() && !is_critical_native) { 1843 1844 // load oop into a register 1845 __ movoop(oop_handle_reg, 1846 JNIHandles::make_local(method->method_holder()->java_mirror()), 1847 /*immediate*/true); 1848 1849 // Now handlize the static class mirror it's known not-null. 1850 __ str(oop_handle_reg, Address(sp, klass_offset)); 1851 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); 1852 1853 // Now get the handle 1854 __ lea(oop_handle_reg, Address(sp, klass_offset)); 1855 // store the klass handle as second argument 1856 __ mov(c_rarg1, oop_handle_reg); 1857 // and protect the arg if we must spill 1858 c_arg--; 1859 } 1860 1861 // Change state to native (we save the return address in the thread, since it might not 1862 // be pushed on the stack when we do a a stack traversal). It is enough that the pc() 1863 // points into the right code segment. It does not have to be the correct return pc. 1864 // We use the same pc/oopMap repeatedly when we call out 1865 1866 intptr_t the_pc = (intptr_t) __ pc(); 1867 oop_maps->add_gc_map(the_pc - start, map); 1868 1869 __ set_last_Java_frame(sp, noreg, (address)the_pc, rscratch1); 1870 1871 1872 // We have all of the arguments setup at this point. We must not touch any register 1873 // argument registers at this point (what if we save/restore them there are no oop? 1874 1875 #ifdef DTRACE_ENABLED 1876 { 1877 SkipIfEqual skip(masm, &DTraceMethodProbes, false); 1878 // protect the args we've loaded 1879 (void) save_args(masm, total_c_args, c_arg, out_regs); 1880 __ mov_metadata(c_rarg1, method()); 1881 __ call_VM_leaf( 1882 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 1883 rthread, c_rarg1); 1884 restore_args(masm, total_c_args, c_arg, out_regs); 1885 } 1886 #endif 1887 1888 // RedefineClasses() tracing support for obsolete method entry 1889 if (log_is_enabled(Trace, redefine, class, obsolete)) { 1890 // protect the args we've loaded 1891 save_args(masm, total_c_args, c_arg, out_regs); 1892 __ mov_metadata(c_rarg1, method()); 1893 __ call_VM_leaf( 1894 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 1895 rthread, c_rarg1); 1896 restore_args(masm, total_c_args, c_arg, out_regs); 1897 } 1898 1899 // Lock a synchronized method 1900 1901 // Register definitions used by locking and unlocking 1902 1903 Label slow_path_lock; 1904 Label lock_done; 1905 1906 if (method->is_synchronized()) { 1907 assert(!is_critical_native, "unhandled"); 1908 1909 // registers below are not used to pass parameters 1910 // and they are caller save in C1 1911 // => safe to use as temporary here 1912 const Register swap_reg = r5; 1913 const Register obj_reg = r6; // Will contain the oop 1914 const Register lock_reg = r7; // Address of compiler lock object (BasicLock) 1915 1916 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes(); 1917 1918 // Get the handle (the 2nd argument) 1919 __ mov(oop_handle_reg, c_rarg1); 1920 1921 // Get address of the box 1922 1923 __ lea(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 1924 1925 // Load the oop from the handle 1926 __ ldr(obj_reg, Address(oop_handle_reg, 0)); 1927 1928 if (UseBiasedLocking) { 1929 __ biased_locking_enter(obj_reg, swap_reg, rscratch2, rscratch1, false, lock_done, &slow_path_lock); 1930 } 1931 1932 // Load (object->mark() | 1) into swap_reg %r0 1933 __ ldr(swap_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 1934 __ orr(swap_reg, swap_reg, 1); 1935 1936 // Save (object->mark() | 1) into BasicLock's displaced header 1937 __ str(swap_reg, Address(lock_reg, mark_word_offset)); 1938 1939 // src -> dest iff dest == r0 else r0 <- dest 1940 { Label here; 1941 __ cmpxchg_obj_header(swap_reg, lock_reg, obj_reg, rscratch1, lock_done, &slow_path_lock); 1942 } 1943 1944 // Slow path will re-enter here 1945 __ bind(lock_done); 1946 } 1947 1948 1949 // Finally just about ready to make the JNI call 1950 1951 1952 // get JNIEnv* which is first argument to native 1953 if (!is_critical_native) { 1954 __ lea(c_rarg0, Address(rthread, in_bytes(JavaThread::jni_environment_offset()))); 1955 } 1956 1957 // Now set thread in native 1958 __ mov(rscratch1, _thread_in_native); 1959 __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset())); 1960 __ dmb(Assembler::ISH); 1961 __ str(rscratch1, rscratch2); 1962 1963 // Do the call 1964 rt_call(masm, native_func); 1965 1966 // Unpack native results. 1967 switch (ret_type) { 1968 case T_BOOLEAN: __ c2bool(r0); break; 1969 case T_CHAR : __ uxth(r0, r0); break; 1970 case T_BYTE : __ sxtb(r0, r0); break; 1971 case T_SHORT : __ sxth(r0, r0); break; 1972 case T_INT : break; 1973 case T_FLOAT : 1974 #ifndef HARD_FLOAT_CC 1975 if(hasFPU()) { 1976 __ vmov_f32(d0, r0); 1977 } 1978 #endif 1979 break; 1980 case T_DOUBLE : 1981 #ifndef HARD_FLOAT_CC 1982 if(hasFPU()) { 1983 __ vmov_f64(d0, r0, r1); 1984 } 1985 #endif 1986 break; 1987 case T_ARRAY: // Really a handle 1988 case T_OBJECT: // Really a handle 1989 break; // can't de-handlize until after safepoint check 1990 case T_VOID: break; 1991 case T_LONG: break; 1992 default : ShouldNotReachHere(); 1993 } 1994 1995 // Switch thread to "native transition" state before reading the synchronization state. 1996 // This additional state is necessary because reading and testing the synchronization 1997 // state is not atomic w.r.t. GC, as this scenario demonstrates: 1998 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 1999 // VM thread changes sync state to synchronizing and suspends threads for GC. 2000 // Thread A is resumed to finish this native method, but doesn't block here since it 2001 // didn't see any synchronization is progress, and escapes. 2002 __ mov(rscratch1, _thread_in_native_trans); 2003 __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset())); 2004 __ dmb(Assembler::ISH); 2005 __ str(rscratch1, rscratch2); 2006 2007 if(os::is_MP()) { 2008 if (UseMembar) { 2009 // Force this write out before the read below 2010 __ membar(Assembler::AnyAny); 2011 } else { 2012 // Write serialization page so VM thread can do a pseudo remote membar. 2013 // We use the current thread pointer to calculate a thread specific 2014 // offset to write to within the page. This minimizes bus traffic 2015 // due to cache line collision. 2016 __ serialize_memory(rthread, rscratch1); 2017 } 2018 } 2019 2020 Label after_transition; 2021 2022 // check for safepoint operation in progress and/or pending suspend requests 2023 { 2024 Label Continue; 2025 2026 Label L; 2027 __ safepoint_poll_acquire(L); 2028 __ ldr(rscratch1, Address(rthread, JavaThread::suspend_flags_offset())); 2029 __ cbz(rscratch1, Continue); 2030 __ bind(L); 2031 2032 // Don't use call_VM as it will see a possible pending exception and forward it 2033 // and never return here preventing us from clearing _last_native_pc down below. 2034 // 2035 save_native_result(masm, ret_type, stack_slots); 2036 __ mov(c_rarg0, rthread); 2037 #ifndef PRODUCT 2038 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 2039 #endif 2040 if (!is_critical_native) { 2041 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans))); 2042 } else { 2043 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans_and_transition))); 2044 } 2045 __ bl(rscratch1); 2046 __ maybe_isb(); 2047 // Restore any method result value 2048 restore_native_result(masm, ret_type, stack_slots); 2049 2050 if (is_critical_native) { 2051 // The call above performed the transition to thread_in_Java so 2052 // skip the transition logic below. 2053 __ b(after_transition); 2054 } 2055 2056 __ bind(Continue); 2057 } 2058 2059 // change thread state 2060 __ mov(rscratch1, _thread_in_Java); 2061 __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset())); 2062 __ dmb(Assembler::ISH); 2063 __ str(rscratch1, rscratch2); 2064 __ bind(after_transition); 2065 2066 Label reguard; 2067 Label reguard_done; 2068 __ ldrb(rscratch1, Address(rthread, JavaThread::stack_guard_state_offset())); 2069 __ cmp(rscratch1, JavaThread::stack_guard_yellow_reserved_disabled); 2070 __ b(reguard, Assembler::EQ); 2071 __ bind(reguard_done); 2072 2073 // native result if any is live 2074 2075 // Unlock 2076 Label unlock_done; 2077 Label slow_path_unlock; 2078 if (method->is_synchronized()) { 2079 const Register obj_reg = r2; // Will contain the oop 2080 const Register lock_reg = rscratch1; // Address of compiler lock object (BasicLock) 2081 const Register old_hdr = r3; // value of old header at unlock time 2082 2083 // Get locked oop from the handle we passed to jni 2084 __ ldr(obj_reg, Address(oop_handle_reg, 0)); 2085 2086 if (UseBiasedLocking) { 2087 __ biased_locking_exit(obj_reg, old_hdr, unlock_done); 2088 } 2089 2090 // Simple recursive lock? 2091 // get address of the stack lock 2092 __ lea(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 2093 2094 // get old displaced header 2095 __ ldr(old_hdr, Address(lock_reg, 0)); 2096 __ cbz(old_hdr, unlock_done); 2097 2098 // Atomic swap old header if oop still contains the stack lock 2099 Label succeed; 2100 __ cmpxchg_obj_header(lock_reg, old_hdr, obj_reg, rscratch2, succeed, &slow_path_unlock); 2101 __ bind(succeed); 2102 2103 // slow path re-enters here 2104 __ bind(unlock_done); 2105 } 2106 2107 #ifdef DTRACE_ENABLED 2108 { 2109 SkipIfEqual skip(masm, &DTraceMethodProbes, false); 2110 save_native_result(masm, ret_type, stack_slots); 2111 __ mov_metadata(c_rarg1, method()); 2112 __ call_VM_leaf( 2113 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 2114 rthread, c_rarg1); 2115 restore_native_result(masm, ret_type, stack_slots); 2116 } 2117 #endif 2118 2119 __ reset_last_Java_frame(false); 2120 2121 // Unbox oop result, e.g. JNIHandles::resolve result. 2122 if (ret_type == T_OBJECT || ret_type == T_ARRAY) { 2123 __ resolve_jobject(r0, rthread, rscratch2); 2124 } 2125 2126 if (CheckJNICalls) { 2127 // clear_pending_jni_exception_check 2128 __ mov(rscratch1, 0); 2129 __ str(rscratch1, Address(rthread, JavaThread::pending_jni_exception_check_fn_offset())); 2130 } 2131 2132 if (!is_critical_native) { 2133 // reset handle block 2134 __ mov(rscratch1, 0); 2135 __ ldr(r2, Address(rthread, JavaThread::active_handles_offset())); 2136 __ str(rscratch1, Address(r2, JNIHandleBlock::top_offset_in_bytes())); 2137 } 2138 2139 __ leave(); 2140 2141 if (!is_critical_native) { 2142 // Any exception pending? 2143 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2144 __ cbnz(rscratch1, exception_pending); 2145 } 2146 2147 // We're done 2148 __ b(lr); 2149 2150 // Unexpected paths are out of line and go here 2151 2152 if (!is_critical_native) { 2153 // forward the exception 2154 __ bind(exception_pending); 2155 2156 // and forward the exception 2157 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2158 } 2159 2160 // Slow path locking & unlocking 2161 if (method->is_synchronized()) { 2162 2163 // BEGIN Slow path lock 2164 __ bind(slow_path_lock); 2165 2166 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM 2167 // args are (oop obj, BasicLock* lock, JavaThread* thread) 2168 2169 // protect the args we've loaded 2170 const int extra_words = save_args(masm, total_c_args, c_arg, out_regs); 2171 2172 __ ldr(c_rarg0, Address(oop_handle_reg)); 2173 __ lea(c_rarg1, Address(sp, (extra_words + lock_slot_offset) * VMRegImpl::stack_slot_size)); 2174 __ mov(c_rarg2, rthread); 2175 2176 // Not a leaf but we have last_Java_frame setup as we want 2177 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3); 2178 restore_args(masm, total_c_args, c_arg, out_regs); 2179 2180 #ifdef ASSERT 2181 { Label L; 2182 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2183 __ cbz(rscratch1, L); 2184 __ stop("no pending exception allowed on exit from monitorenter"); 2185 __ bind(L); 2186 } 2187 #endif 2188 __ b(lock_done); 2189 2190 // END Slow path lock 2191 2192 // BEGIN Slow path unlock 2193 __ bind(slow_path_unlock); 2194 2195 // If we haven't already saved the native result we must save it now as xmm registers 2196 // are still exposed. 2197 2198 save_native_result(masm, ret_type, stack_slots); 2199 2200 __ mov(c_rarg2, rthread); 2201 __ lea(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size)); 2202 __ ldr(c_rarg0, Address(oop_handle_reg)); 2203 2204 // Save pending exception around call to VM (which contains an EXCEPTION_MARK) 2205 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2206 __ mov(rscratch2, 0); 2207 __ str(rscratch2, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2208 2209 rt_call(masm, CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C)); 2210 2211 #ifdef ASSERT 2212 { 2213 Label L; 2214 __ ldr(rscratch2, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2215 __ cbz(rscratch2, L); 2216 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C"); 2217 __ bind(L); 2218 } 2219 #endif // ASSERT 2220 2221 __ str(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset()))); 2222 2223 restore_native_result(masm, ret_type, stack_slots); 2224 2225 __ b(unlock_done); 2226 2227 // END Slow path unlock 2228 2229 } // synchronized 2230 2231 // SLOW PATH Reguard the stack if needed 2232 2233 __ bind(reguard); 2234 save_native_result(masm, ret_type, stack_slots); 2235 rt_call(masm, CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages)); 2236 restore_native_result(masm, ret_type, stack_slots); 2237 // and continue 2238 __ b(reguard_done); 2239 2240 2241 2242 __ flush(); 2243 2244 nmethod *nm = nmethod::new_native_nmethod(method, 2245 compile_id, 2246 masm->code(), 2247 vep_offset, 2248 frame_complete, 2249 stack_slots / VMRegImpl::slots_per_word, 2250 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), 2251 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size), 2252 oop_maps); 2253 2254 if (is_critical_native) { 2255 nm->set_lazy_critical_native(true); 2256 } 2257 2258 return nm; 2259 } 2260 2261 // this function returns the adjust size (in number of words) to a c2i adapter 2262 // activation for use during deoptimization 2263 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) { 2264 assert(callee_locals >= callee_parameters, 2265 "test and remove; got more parms than locals"); 2266 if (callee_locals < callee_parameters) 2267 return 0; // No adjustment for negative locals 2268 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords; 2269 // diff is counted in stack words 2270 return align_up(diff, 2); 2271 } 2272 2273 2274 //------------------------------generate_deopt_blob---------------------------- 2275 void SharedRuntime::generate_deopt_blob() { 2276 2277 // Allocate space for the code 2278 ResourceMark rm; 2279 // Setup code generation tools 2280 CodeBuffer buffer("deopt_blob", 2048, 1024); 2281 MacroAssembler* masm = new MacroAssembler(&buffer); 2282 int frame_size_in_words; 2283 OopMap* map = NULL; 2284 OopMapSet *oop_maps = new OopMapSet(); 2285 2286 // ------------- 2287 // This code enters when returning to a de-optimized nmethod. A return 2288 // address has been pushed on the the stack, and return values are in 2289 // registers. 2290 // If we are doing a normal deopt then we were called from the patched 2291 // nmethod from the point we returned to the nmethod. So the return 2292 // address on the stack is wrong by NativeCall::instruction_size 2293 // We will adjust the value so it looks like we have the original return 2294 // address on the stack (like when we eagerly deoptimized). 2295 // In the case of an exception pending when deoptimizing, we enter 2296 // with a return address on the stack that points after the call we patched 2297 // into the exception handler. We have the following register state from, 2298 // e.g., the forward exception stub (see stubGenerator_x86_64.cpp). 2299 // r0: exception oop 2300 // r7: exception handler 2301 // r3: throwing pc 2302 // So in this case we simply jam r3 into the useless return address and 2303 // the stack looks just like we want. 2304 // 2305 // At this point we need to de-opt. We save the argument return 2306 // registers. We call the first C routine, fetch_unroll_info(). This 2307 // routine captures the return values and returns a structure which 2308 // describes the current frame size and the sizes of all replacement frames. 2309 // The current frame is compiled code and may contain many inlined 2310 // functions, each with their own JVM state. We pop the current frame, then 2311 // push all the new frames. Then we call the C routine unpack_frames() to 2312 // populate these frames. Finally unpack_frames() returns us the new target 2313 // address. Notice that callee-save registers are BLOWN here; they have 2314 // already been captured in the vframeArray at the time the return PC was 2315 // patched. 2316 address start = __ pc(); 2317 Label cont; 2318 2319 // Prolog for non exception case! 2320 2321 // Save everything in sight. 2322 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, true); 2323 2324 // Normal deoptimization. Save exec mode for unpack_frames. 2325 __ mov(r7, Deoptimization::Unpack_deopt); // callee-saved 2326 __ b(cont); 2327 2328 int reexecute_offset = __ pc() - start; 2329 2330 // Reexecute case 2331 // return address is the pc describes what bci to do re-execute at 2332 2333 // No need to update map as each call to save_live_registers will produce identical oopmap 2334 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2335 2336 __ mov(r7, Deoptimization::Unpack_reexecute); // callee-saved 2337 __ b(cont); 2338 2339 int exception_offset = __ pc() - start; 2340 2341 // Prolog for exception case 2342 2343 // all registers are dead at this entry point, except for r0, and 2344 // r3 which contain the exception oop and exception pc 2345 // respectively. Set them in TLS and fall thru to the 2346 // unpack_with_exception_in_tls entry point. 2347 2348 __ str(r3, Address(rthread, JavaThread::exception_pc_offset())); 2349 __ str(r0, Address(rthread, JavaThread::exception_oop_offset())); 2350 2351 int exception_in_tls_offset = __ pc() - start; 2352 2353 // new implementation because exception oop is now passed in JavaThread 2354 2355 // Prolog for exception case 2356 // All registers must be preserved because they might be used by LinearScan 2357 // Exceptiop oop and throwing PC are passed in JavaThread 2358 // tos: stack at point of call to method that threw the exception (i.e. only 2359 // args are on the stack, no return address) 2360 2361 // The return address pushed by save_live_registers will be patched 2362 // later with the throwing pc. The correct value is not available 2363 // now because loading it from memory would destroy registers. 2364 2365 // NB: The SP at this point must be the SP of the method that is 2366 // being deoptimized. Deoptimization assumes that the frame created 2367 // here by save_live_registers is immediately below the method's SP. 2368 // This is a somewhat fragile mechanism. 2369 2370 // Save everything in sight. 2371 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2372 2373 // Now it is safe to overwrite any register 2374 2375 // Deopt during an exception. Save exec mode for unpack_frames. 2376 __ mov(r7, Deoptimization::Unpack_exception); // callee-saved 2377 2378 // load throwing pc from JavaThread and patch it as the return address 2379 // of the current frame. Then clear the field in JavaThread 2380 2381 __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset())); 2382 __ str(r3, Address(rfp, frame::get_return_addr_offset() * wordSize)); 2383 __ mov(rscratch1, 0); 2384 __ str(rscratch1, Address(rthread, JavaThread::exception_pc_offset())); 2385 2386 #ifdef ASSERT 2387 // verify that there is really an exception oop in JavaThread 2388 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset())); 2389 __ verify_oop(r0); 2390 2391 // verify that there is no pending exception 2392 Label no_pending_exception; 2393 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset())); 2394 __ cbz(rscratch1, no_pending_exception); 2395 __ stop("must not have pending exception here"); 2396 __ bind(no_pending_exception); 2397 #endif 2398 2399 __ bind(cont); 2400 2401 // Call C code. Need thread and this frame, but NOT official VM entry 2402 // crud. We cannot block on this call, no GC can happen. 2403 // 2404 // UnrollBlock* fetch_unroll_info(JavaThread* thread) 2405 2406 // fetch_unroll_info needs to call last_java_frame(). 2407 2408 Label retaddr; 2409 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2410 #ifdef ASSERT0 2411 { Label L; 2412 __ ldr(rscratch1, Address(rthread, 2413 JavaThread::last_Java_fp_offset())); 2414 __ cbz(rscratch1, L); 2415 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared"); 2416 __ bind(L); 2417 } 2418 #endif // ASSERT 2419 __ mov(c_rarg0, rthread); 2420 __ mov(c_rarg1, r7); // rcpool 2421 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info))); 2422 __ bl(rscratch1); 2423 __ bind(retaddr); 2424 2425 // Need to have an oopmap that tells fetch_unroll_info where to 2426 // find any register it might need. 2427 oop_maps->add_gc_map(__ pc() - start, map); 2428 2429 __ reset_last_Java_frame(false); 2430 2431 // Load UnrollBlock* into r5 2432 __ mov(r5, r0); 2433 2434 Label noException; 2435 __ cmp(r7, Deoptimization::Unpack_exception); // Was exception pending? 2436 __ b(noException, Assembler::NE); 2437 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset())); 2438 // QQQ this is useless it was NULL above 2439 __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset())); 2440 __ mov(rscratch1, 0); 2441 __ str(rscratch1, Address(rthread, JavaThread::exception_oop_offset())); 2442 __ str(rscratch1, Address(rthread, JavaThread::exception_pc_offset())); 2443 2444 __ verify_oop(r0); 2445 2446 // Overwrite the result registers with the exception results. 2447 __ str(r0, Address(sp, RegisterSaver::offset_in_bytes(RegisterSaver::r0_off))); 2448 // I think this is useless 2449 // __ str(r3, Address(sp, RegisterSaver::r3_offset_in_bytes())); 2450 2451 __ bind(noException); 2452 2453 // Only register save data is on the stack. 2454 // Now restore the result registers. Everything else is either dead 2455 // or captured in the vframeArray. 2456 RegisterSaver::restore_result_registers(masm); 2457 2458 // All of the register save area has been popped of the stack. Only the 2459 // return address remains. 2460 2461 // Pop all the frames we must move/replace. 2462 // 2463 // Frame picture (youngest to oldest) 2464 // 1: self-frame (no frame link) 2465 // 2: deopting frame (no frame link) 2466 // 3: caller of deopting frame (could be compiled/interpreted). 2467 // 2468 // Note: by leaving the return address of self-frame on the stack 2469 // and using the size of frame 2 to adjust the stack 2470 // when we are done the return to frame 3 will still be on the stack. 2471 2472 // Pop deoptimized frame 2473 __ ldr(r2, Address(r5, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes())); 2474 __ sub(r2, r2, frame::get_frame_size() * wordSize); 2475 __ add(sp, sp, r2); 2476 if (FrameAPCS) { 2477 // frame constructed with 2478 // push {r11, r12, lr, pc} 2479 __ ldr(rfp, __ post(sp, 2 * wordSize)); 2480 __ ldr(lr, __ post(sp, 2 * wordSize)); 2481 } else { 2482 __ ldrd(rfp, lr, __ post(sp, 2 * wordSize)); 2483 } 2484 // LR should now be the return address to the caller (3) 2485 2486 #ifdef ASSERT 2487 // Compilers generate code that bang the stack by as much as the 2488 // interpreter would need. So this stack banging should never 2489 // trigger a fault. Verify that it does not on non product builds. 2490 if (UseStackBanging) { 2491 __ ldr(rscratch2, Address(r5, Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes())); 2492 __ bang_stack_size(rscratch2, r2); 2493 } 2494 #endif 2495 // Load address of array of frame pcs into r2 2496 __ ldr(r2, Address(r5, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 2497 2498 // Trash the old pc 2499 // __ addptr(sp, wordSize); FIXME ???? 2500 2501 // Load address of array of frame sizes into r4 2502 __ ldr(r4, Address(r5, Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes())); 2503 2504 // Load counter into r3 2505 __ ldr(r3, Address(r5, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes())); 2506 2507 // Now adjust the caller's stack to make up for the extra locals 2508 // but record the original sp so that we can save it in the skeletal interpreter 2509 // frame and the stack walking of interpreter_sender will get the unextended sp 2510 // value and not the "real" sp value. 2511 2512 const Register sender_sp = r6; 2513 2514 __ mov(sender_sp, sp); 2515 __ ldr(rscratch1, Address(r5, 2516 Deoptimization::UnrollBlock:: 2517 caller_adjustment_offset_in_bytes())); 2518 __ sub(sp, sp, rscratch1); 2519 2520 // Push interpreter frames in a loop 2521 __ mov(rscratch1, (address)0xDEADDEAD); // Make a recognizable pattern 2522 // Initially used to place 0xDEADDEAD in rscratch2 as well - why? 2523 __ mov(rscratch2, 0); 2524 Label loop; 2525 __ bind(loop); 2526 __ ldr(rscratch1, Address(__ post(r4, wordSize))); // Load frame size 2527 __ sub(rscratch1, rscratch1, frame::get_frame_size() * wordSize); // We'll push frame backtrace by hand 2528 __ ldr(lr, Address(__ post(r2, wordSize))); // Load pc 2529 __ enter(); // Save old & set new fp 2530 __ sub(sp, sp, rscratch1); // Prolog 2531 // This value is corrected by layout_activation_impl 2532 __ str(rscratch2, Address(rfp, frame::get_interpreter_frame_last_sp_offset() * wordSize)); 2533 __ str(sender_sp, Address(rfp, frame::get_interpreter_frame_sender_sp_offset() * wordSize)); // Make it walkable 2534 __ mov(sender_sp, sp); // Pass sender_sp to next frame 2535 __ sub(r3, r3, 1); // Decrement counter 2536 __ cbnz(r3, loop); 2537 2538 // Re-push self-frame 2539 __ ldr(lr, Address(r2)); 2540 __ enter(); 2541 2542 // Allocate a full sized register save area. We subtract frame::get_frame_size() words, 2543 // because enter() just pushed them. 2544 __ sub(sp, sp, (frame_size_in_words - frame::get_frame_size()) * wordSize); 2545 2546 // Restore frame locals after moving the frame 2547 if(hasFPU()) { 2548 __ vstr_f64(d0, Address(sp, RegisterSaver::offset_in_bytes(RegisterSaver::fpu_state_off))); 2549 } 2550 __ strd(r0, Address(sp, RegisterSaver::offset_in_bytes(RegisterSaver::r0_off))); 2551 2552 // Call C code. Need thread but NOT official VM entry 2553 // crud. We cannot block on this call, no GC can happen. Call should 2554 // restore return values to their stack-slots with the new SP. 2555 // 2556 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode) 2557 2558 // Use rfp because the frames look interpreted now 2559 // Don't need the precise return PC here, just precise enough to point into this code blob. 2560 address the_pc = __ pc(); 2561 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1); 2562 2563 __ mov(c_rarg0, rthread); 2564 __ mov(c_rarg1, r7); // second arg: exec_mode 2565 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 2566 __ bl(rscratch1); 2567 2568 // Set an oopmap for the call site 2569 // Use the same PC we used for the last java frame 2570 oop_maps->add_gc_map(the_pc - start, 2571 new OopMap( frame_size_in_words, 0 )); 2572 2573 // Clear fp AND pc 2574 __ reset_last_Java_frame(true); 2575 2576 // Collect return values 2577 if(hasFPU()) { 2578 __ vldr_f64(d0, Address(sp, RegisterSaver::offset_in_bytes(RegisterSaver::fpu_state_off))); 2579 } 2580 __ ldrd(r0, Address(sp, RegisterSaver::offset_in_bytes(RegisterSaver::r0_off))); 2581 // I think this is useless (throwing pc?) 2582 // __ ldr(r3, Address(sp, RegisterSaver::r3_offset_in_bytes())); 2583 2584 // Pop self-frame. 2585 __ leave(); // Epilog 2586 2587 // Jump to interpreter 2588 __ b(lr); 2589 2590 // Make sure all code is generated 2591 masm->flush(); 2592 2593 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words); 2594 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset); 2595 2596 } 2597 2598 uint SharedRuntime::out_preserve_stack_slots() { 2599 return 0; 2600 } 2601 2602 #if COMPILER2_OR_JVMCI 2603 //------------------------------generate_uncommon_trap_blob-------------------- 2604 void SharedRuntime::generate_uncommon_trap_blob() { 2605 // Allocate space for the code 2606 ResourceMark rm; 2607 // Setup code generation tools 2608 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024); 2609 MacroAssembler* masm = new MacroAssembler(&buffer); 2610 2611 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned"); 2612 2613 address start = __ pc(); 2614 2615 // Push self-frame. We get here with a return address in LR 2616 // and sp should be 16 byte aligned 2617 // push rfp and retaddr by hand 2618 __ enter(); 2619 // we don't expect an arg reg save area 2620 #ifndef PRODUCT 2621 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area"); 2622 #endif 2623 // compiler left unloaded_class_index in j_rarg0 2624 __ mov(c_rarg1, j_rarg0); 2625 2626 // we need to set the past SP to the stack pointer of the stub frame 2627 // and the pc to the address where this runtime call will return 2628 // although actually any pc in this code blob will do). 2629 Label retaddr; 2630 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2631 2632 // Call C code. Need thread but NOT official VM entry 2633 // crud. We cannot block on this call, no GC can happen. Call should 2634 // capture callee-saved registers as well as return values. 2635 // 2636 // UnrollBlock* uncommon_trap(JavaThread* thread, jint unloaded_class_index); 2637 // 2638 // n.b. 2 gp args, 0 fp args, integral return type 2639 2640 __ mov(c_rarg0, rthread); 2641 __ mov(c_rarg2, (unsigned)Deoptimization::Unpack_uncommon_trap); 2642 __ lea(rscratch1, 2643 RuntimeAddress(CAST_FROM_FN_PTR(address, 2644 Deoptimization::uncommon_trap))); 2645 __ bl(rscratch1); 2646 __ bind(retaddr); 2647 2648 // Set an oopmap for the call site 2649 OopMapSet* oop_maps = new OopMapSet(); 2650 OopMap* map = new OopMap(SimpleRuntimeFrame::framesize, 0); 2651 2652 // location of rfp is known implicitly by the frame sender code 2653 2654 oop_maps->add_gc_map(__ pc() - start, map); 2655 2656 __ reset_last_Java_frame(false); 2657 2658 // move UnrollBlock* into r4 2659 __ mov(r4, r0); 2660 2661 // Pop all the frames we must move/replace. 2662 // 2663 // Frame picture (youngest to oldest) 2664 // 1: self-frame (no frame link) 2665 // 2: deopting frame (no frame link) 2666 // 3: caller of deopting frame (could be compiled/interpreted). 2667 2668 // Pop self-frame 2669 __ leave(); 2670 2671 // Pop deoptimized frame (int) 2672 __ ldr(r2, Address(r4, 2673 Deoptimization::UnrollBlock:: 2674 size_of_deoptimized_frame_offset_in_bytes())); 2675 __ add(sp, sp, r2); 2676 2677 #ifdef ASSERT 2678 // Compilers generate code that bang the stack by as much as the 2679 // interpreter would need. So this stack banging should never 2680 // trigger a fault. Verify that it does not on non product builds. 2681 if (UseStackBanging) { 2682 // The compiled method that we are deoptimizing was popped from the stack. 2683 // If the stack bang results in a stack overflow, we don't return to the 2684 // method that is being deoptimized. The stack overflow exception is 2685 // propagated to the caller of the deoptimized method. Need to get the pc 2686 // from the caller in LR and restore FP. 2687 __ ldr(r2, Address(r4, 2688 Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 2689 __ ldr(lr, Address(r2, 0)); 2690 __ ldr(rfp, Address(r4, Deoptimization::UnrollBlock::initial_info_offset_in_bytes())); 2691 __ ldr(r1, Address(r4, 2692 Deoptimization::UnrollBlock:: 2693 total_frame_sizes_offset_in_bytes())); 2694 __ bang_stack_size(r1, r2); 2695 } 2696 #endif 2697 // Now is the time to restore frameptr. Need to take what was in the frame header 2698 // since it can be real FP if previous frame was interpreted/C1 or arbitrary value if C2 2699 __ ldr(rfp, Address(sp, -2*wordSize)/*Address(r4, 2700 Deoptimization::UnrollBlock::initial_info_offset_in_bytes())*/); 2701 2702 // Load address of array of frame pcs into r2 (address*) 2703 __ ldr(r2, Address(r4, 2704 Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 2705 2706 // Load address of array of frame sizes into r5 (intptr_t*) 2707 __ ldr(r5, Address(r4, 2708 Deoptimization::UnrollBlock:: 2709 frame_sizes_offset_in_bytes())); 2710 2711 // Counter 2712 __ ldr(r3, Address(r4, 2713 Deoptimization::UnrollBlock:: 2714 number_of_frames_offset_in_bytes())); // (int) 2715 2716 // Now adjust the caller's stack to make up for the extra locals but 2717 // record the original sp so that we can save it in the skeletal 2718 // interpreter frame and the stack walking of interpreter_sender 2719 // will get the unextended sp value and not the "real" sp value. 2720 2721 const Register sender_sp = r7; 2722 2723 __ mov(sender_sp, sp); 2724 __ ldr(r1, Address(r4, 2725 Deoptimization::UnrollBlock:: 2726 caller_adjustment_offset_in_bytes())); // (int) 2727 __ sub(sp, sp, r1); 2728 2729 __ mov(rscratch1, 0); 2730 // Push interpreter frames in a loop 2731 Label loop; 2732 __ bind(loop); 2733 __ ldr(r1, __ post(r5, wordSize)); // Load frame size 2734 __ sub(r1, r1, 2 * wordSize); // We'll push pc and rfp by hand 2735 __ ldr(lr, __ post(r2, wordSize)); // Save return address 2736 __ enter(); // and old rfp & set new rfp 2737 __ sub(sp, sp, r1); // Prolog 2738 __ str(sender_sp, Address(rfp, frame::get_interpreter_frame_sender_sp_offset() * wordSize)); // Make it walkable 2739 // This value is corrected by layout_activation_impl 2740 __ str(rscratch1, Address(rfp, frame::get_interpreter_frame_last_sp_offset() * wordSize)); //zero it 2741 __ mov(sender_sp, sp); // Pass sender_sp to next frame 2742 __ subs(r3, r3, 1); // Decrement counter 2743 __ b(loop, Assembler::GT); 2744 __ ldr(lr, Address(r2, 0)); // save final return address 2745 // Re-push self-frame 2746 __ enter(); // & old rfp & set new rfp 2747 2748 // Use rfp because the frames look interpreted now 2749 // Save "the_pc" since it cannot easily be retrieved using the last_java_SP after we aligned SP. 2750 // Don't need the precise return PC here, just precise enough to point into this code blob. 2751 address the_pc = __ pc(); 2752 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1); 2753 2754 // Call C code. Need thread but NOT official VM entry 2755 // crud. We cannot block on this call, no GC can happen. Call should 2756 // restore return values to their stack-slots with the new SP. 2757 // Thread is in rdi already. 2758 // 2759 // BasicType unpack_frames(JavaThread* thread, int exec_mode); 2760 // 2761 // n.b. 2 gp args, 0 fp args, integral return type 2762 2763 // sp should already be aligned 2764 __ mov(c_rarg0, rthread); 2765 __ mov(c_rarg1, (unsigned)Deoptimization::Unpack_uncommon_trap); 2766 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 2767 __ bl(rscratch1); 2768 2769 // Set an oopmap for the call site 2770 // Use the same PC we used for the last java frame 2771 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0)); 2772 2773 // Clear fp AND pc 2774 __ reset_last_Java_frame(true); 2775 2776 // Pop self-frame. 2777 __ leave(); // Epilog 2778 2779 // Jump to interpreter 2780 __ b(lr); 2781 2782 // Make sure all code is generated 2783 masm->flush(); 2784 2785 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, 2786 SimpleRuntimeFrame::framesize >> 1); 2787 } 2788 #endif // COMPILER2_OR_JVMCI 2789 2790 2791 //------------------------------generate_handler_blob------ 2792 // 2793 // Generate a special Compile2Runtime blob that saves all registers, 2794 // and setup oopmap. 2795 // 2796 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) { 2797 ResourceMark rm; 2798 OopMapSet *oop_maps = new OopMapSet(); 2799 OopMap* map; 2800 2801 // Allocate space for the code. Setup code generation tools. 2802 CodeBuffer buffer("handler_blob", 2048, 1024); 2803 MacroAssembler* masm = new MacroAssembler(&buffer); 2804 2805 address start = __ pc(); 2806 address call_pc = NULL; 2807 int frame_size_in_words; 2808 bool cause_return = (poll_type == POLL_AT_RETURN); 2809 2810 // If cause_return is true we are at a poll_return and there is 2811 // the return address on the stack to the caller on the nmethod 2812 // that is safepoint. We can leave this return on the stack and 2813 // effectively complete the return and safepoint in the caller. 2814 // Otherwise we push space for a return address that the safepoint 2815 // handler will install later to make the stack walking sensible. 2816 if (!cause_return) { 2817 __ sub(sp, sp, wordSize); // make room for return address 2818 } 2819 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, !cause_return); 2820 2821 // The following is basically a call_VM. However, we need the precise 2822 // address of the call in order to generate an oopmap. Hence, we do all the 2823 // work outselves. 2824 2825 Label retaddr; 2826 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2827 2828 // The return address must always be correct so that frame constructor never 2829 // sees an invalid pc. 2830 2831 if (!cause_return) { 2832 // overwrite the return address pushed by save_live_registers 2833 // Additionally, r5 is a callee-saved register so we can look at 2834 // it later to determine if someone changed the return address for 2835 // us! 2836 __ ldr(r5, Address(rthread, JavaThread::saved_exception_pc_offset())); 2837 __ str(r5, Address(rfp, frame::get_return_addr_offset() * wordSize)); 2838 } 2839 2840 // Do the call 2841 __ mov(c_rarg0, rthread); 2842 __ lea(rscratch1, RuntimeAddress(call_ptr)); 2843 __ bl(rscratch1); 2844 __ bind(retaddr); 2845 2846 // Set an oopmap for the call site. This oopmap will map all 2847 // oop-registers and debug-info registers as callee-saved. This 2848 // will allow deoptimization at this safepoint to find all possible 2849 // debug-info recordings, as well as let GC find all oops. 2850 2851 oop_maps->add_gc_map( __ pc() - start, map); 2852 2853 Label noException, no_adjust, bail; 2854 2855 __ reset_last_Java_frame(false); 2856 2857 __ maybe_isb(); 2858 __ membar(Assembler::LoadLoad | Assembler::LoadStore); 2859 2860 if (SafepointMechanism::uses_thread_local_poll() && !cause_return) { 2861 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset())); 2862 __ cbnz(rscratch1, no_adjust); 2863 2864 // If our stashed return pc was modified by the runtime we avoid touching it 2865 __ ldr(rscratch1, Address(rfp, frame::get_return_addr_offset() * wordSize)); 2866 __ cmp(r5, rscratch1); 2867 __ b(no_adjust, Assembler::NE); 2868 2869 #ifdef ASSERT 2870 // Verify the correct encoding of the poll we're about to skip. 2871 // ldr(r12, [r12, #0]); 2872 __ ldr(rscratch1, Address(r5)); 2873 __ bic(rscratch1, rscratch1, ~0xfff0ffff); 2874 __ mov(rscratch2, 0xe590c000); 2875 __ cmp(rscratch1, rscratch2); 2876 __ b(bail, Assembler::NE); 2877 #endif 2878 // Adjust return pc forward to step over the safepoint poll instruction 2879 __ add(r5, r5, NativeInstruction::arm_insn_sz); 2880 __ str(r5, Address(rfp, frame::get_return_addr_offset() * wordSize)); 2881 } 2882 2883 __ bind(no_adjust); 2884 2885 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset())); 2886 __ cmp(rscratch1, 0); 2887 2888 // does not kill flags 2889 RegisterSaver::restore_live_registers(masm, cause_return); 2890 // for !POLL_AT_RETURN the stack has return address on it 2891 2892 __ b(noException, Assembler::EQ); 2893 2894 // Exception pending 2895 if (cause_return) 2896 __ mov(r3, lr); 2897 else 2898 __ pop(r3); 2899 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2900 2901 // No exception case 2902 __ bind(noException); 2903 2904 if (cause_return) 2905 __ b(lr); 2906 else 2907 __ pop(r15_pc); 2908 2909 #ifdef ASSERT 2910 __ bind(bail); 2911 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected"); 2912 #endif 2913 2914 // Make sure all code is generated 2915 masm->flush(); 2916 2917 // Fill-out other meta info 2918 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words); 2919 } 2920 2921 // 2922 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss 2923 // 2924 // Generate a stub that calls into vm to find out the proper destination 2925 // of a java call. All the argument registers are live at this point 2926 // but since this is generic code we don't know what they are and the caller 2927 // must do any gc of the args. 2928 // 2929 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) { 2930 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before"); 2931 2932 // allocate space for the code 2933 ResourceMark rm; 2934 2935 //CodeBuffer buffer(name, 1000, 512); 2936 CodeBuffer buffer(name, 2048, 512 ); // changed as error later 2937 MacroAssembler* masm = new MacroAssembler(&buffer); 2938 2939 int frame_size_in_words; 2940 2941 OopMapSet *oop_maps = new OopMapSet(); 2942 OopMap* map = NULL; 2943 2944 int start = __ offset(); 2945 2946 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2947 2948 int frame_complete = __ offset(); 2949 2950 { 2951 Label retaddr; 2952 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1); 2953 2954 __ mov(c_rarg0, rthread); 2955 __ lea(rscratch1, RuntimeAddress(destination)); 2956 2957 __ bl(rscratch1); 2958 __ bind(retaddr); 2959 } 2960 2961 // Set an oopmap for the call site. 2962 // We need this not only for callee-saved registers, but also for volatile 2963 // registers that the compiler might be keeping live across a safepoint. 2964 2965 oop_maps->add_gc_map( __ offset() - start, map); 2966 2967 __ maybe_isb(); 2968 2969 // r0 contains the address we are going to jump to assuming no exception got installed 2970 2971 // clear last_Java_sp 2972 __ reset_last_Java_frame(false); 2973 // check for pending exceptions 2974 Label pending; 2975 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset())); 2976 __ cbnz(rscratch1, pending); 2977 2978 // get the returned Method* 2979 __ get_vm_result_2(rmethod, rthread); 2980 __ str(rmethod, Address(sp, RegisterSaver::offset_in_bytes(RegisterSaver::rmethod_off))); 2981 2982 // r0 is where we want to jump, overwrite rscratch1 which is saved and scratch 2983 __ str(r0, Address(sp, RegisterSaver::offset_in_bytes(RegisterSaver::rscratch1_off))); 2984 RegisterSaver::restore_live_registers(masm); 2985 2986 // We are back the the original state on entry and ready to go. 2987 2988 __ b(rscratch1); 2989 2990 // Pending exception after the safepoint 2991 2992 __ bind(pending); 2993 2994 RegisterSaver::restore_live_registers(masm); 2995 2996 // exception pending => remove activation and forward to exception handler 2997 __ mov(rscratch1, 0); 2998 __ str(rscratch1, Address(rthread, JavaThread::vm_result_offset())); 2999 3000 __ ldr(r0, Address(rthread, Thread::pending_exception_offset())); 3001 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 3002 3003 // ------------- 3004 // make sure all code is generated 3005 masm->flush(); 3006 3007 // return the blob 3008 // frame_size_words or bytes?? 3009 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true); 3010 } 3011 3012 3013 #if COMPILER2_OR_JVMCI 3014 // This is here instead of runtime_x86_64.cpp because it uses SimpleRuntimeFrame 3015 // 3016 //------------------------------generate_exception_blob--------------------------- 3017 // creates exception blob at the end 3018 // Using exception blob, this code is jumped from a compiled method. 3019 // (see emit_exception_handler in x86_64.ad file) 3020 // 3021 // Given an exception pc at a call we call into the runtime for the 3022 // handler in this method. This handler might merely restore state 3023 // (i.e. callee save registers) unwind the frame and jump to the 3024 // exception handler for the nmethod if there is no Java level handler 3025 // for the nmethod. 3026 // 3027 // This code is entered with a jmp. 3028 // 3029 // Arguments: 3030 // r0: exception oop 3031 // r3: exception pc 3032 // 3033 // Results: 3034 // r0: exception oop 3035 // r3: exception pc in caller or ??? 3036 // destination: exception handler of caller 3037 // 3038 // Note: the exception pc MUST be at a call (precise debug information) 3039 // Registers r0, r3, r2, r4, r5, r8-r11 are not callee saved. 3040 // 3041 3042 void OptoRuntime::generate_exception_blob() { 3043 // allocate space for code 3044 ResourceMark rm; 3045 int pad = VerifyThread ? 256 : 0;// Extra slop space for more verify code 3046 3047 // setup code generation tools 3048 // Measured 8/7/03 at 256 in 32bit debug build (no VerifyThread) 3049 // Measured 8/7/03 at 528 in 32bit debug build (VerifyThread) 3050 CodeBuffer buffer("exception_blob", 600+pad, 512); 3051 MacroAssembler* masm = new MacroAssembler(&buffer); 3052 3053 int framesize_in_words = 2; // FP + LR 3054 int framesize_in_bytes = framesize_in_words * wordSize; 3055 int framesize_in_slots = framesize_in_bytes / sizeof(jint); 3056 3057 address start = __ pc(); 3058 3059 __ str(r0, Address(rthread, JavaThread::exception_oop_offset())); 3060 __ str(r3, Address(rthread, JavaThread::exception_pc_offset())); 3061 3062 // This call does all the hard work. It checks if an exception catch 3063 // exists in the method. 3064 // If so, it returns the handler address. 3065 // If the nmethod has been deoptimized and it had a handler the handler 3066 // address is the deopt blob unpack_with_exception entry. 3067 // 3068 // If no handler exists it prepares for stack-unwinding, restoring the callee-save 3069 // registers of the frame being removed. 3070 // 3071 __ mov(lr, r3); 3072 __ enter(); 3073 address the_pc = __ pc(); 3074 __ set_last_Java_frame(sp, noreg, the_pc, rscratch1); 3075 3076 __ mov(r0, rthread); 3077 3078 // This call can block at exit and nmethod can be deoptimized at that 3079 // point. If the nmethod had a catch point we would jump to the 3080 // now deoptimized catch point and fall thru the vanilla deopt 3081 // path and lose the exception 3082 // Sure would be simpler if this call didn't block! 3083 __ call(CAST_FROM_FN_PTR(address, OptoRuntime::handle_exception_C)); 3084 3085 // Set an oopmap for the call site. This oopmap will only be used if we 3086 // are unwinding the stack. Hence, all locations will be dead. 3087 // Callee-saved registers will be the same as the frame above (i.e., 3088 // handle_exception_stub), since they were restored when we got the 3089 // exception. 3090 3091 OopMapSet* oop_maps = new OopMapSet(); 3092 3093 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0)); 3094 3095 __ reset_last_Java_frame(false); 3096 3097 __ leave(); 3098 3099 // Restore SP from its saved reg (FP) if the exception PC is a MethodHandle call site. 3100 __ ldr(rscratch1, Address(rthread, JavaThread::is_method_handle_return_offset())); 3101 __ cmp(rscratch1, 0); 3102 __ mov(sp, rfp, Assembler::NE); 3103 3104 // We have a handler in r0 (could be deopt blob). 3105 __ mov(rscratch2, r0); 3106 3107 // Since this may be the deopt blob we must set R3 to look like we returned 3108 // from the original pc that threw the exception 3109 3110 __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset())); 3111 3112 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset())); 3113 __ mov(rscratch1, 0); 3114 #ifdef ASSERT 3115 __ str(rscratch1, Address(rthread, JavaThread::exception_handler_pc_offset())); 3116 __ str(rscratch1, Address(rthread, JavaThread::exception_pc_offset())); 3117 #endif 3118 // Clear the exception oop so GC no longer processes it as a root. 3119 __ str(rscratch1, Address(rthread, JavaThread::exception_oop_offset())); 3120 __ b(rscratch2); 3121 3122 // ------------- 3123 // make sure all code is generated 3124 masm->flush(); 3125 3126 _exception_blob = ExceptionBlob::create(&buffer, oop_maps, framesize_in_words); 3127 } 3128 #endif // COMPILER2_OR_JVMCI