quick_entrypoints_arm64.S revision 0719b5b9b458cb3eb9f0823f0dacdfe1a71214dd
1/* 2 * Copyright (C) 2014 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17#include "asm_support_arm64.S" 18 19#include "arch/quick_alloc_entrypoints.S" 20 21 22.macro INCREASE_FRAME frame_adjustment 23 sub sp, sp, #(\frame_adjustment) 24 .cfi_adjust_cfa_offset (\frame_adjustment) 25.endm 26 27.macro DECREASE_FRAME frame_adjustment 28 add sp, sp, #(\frame_adjustment) 29 .cfi_adjust_cfa_offset -(\frame_adjustment) 30.endm 31 32.macro SAVE_REG reg, offset 33 str \reg, [sp, #(\offset)] 34 .cfi_rel_offset \reg, (\offset) 35.endm 36 37.macro RESTORE_REG reg, offset 38 ldr \reg, [sp, #(\offset)] 39 .cfi_restore \reg 40.endm 41 42.macro SAVE_TWO_REGS reg1, reg2, offset 43 stp \reg1, \reg2, [sp, #(\offset)] 44 .cfi_rel_offset \reg1, (\offset) 45 .cfi_rel_offset \reg2, (\offset) + 8 46.endm 47 48.macro RESTORE_TWO_REGS reg1, reg2, offset 49 ldp \reg1, \reg2, [sp, #(\offset)] 50 .cfi_restore \reg1 51 .cfi_restore \reg2 52.endm 53 54.macro SAVE_TWO_REGS_INCREASE_FRAME reg1, reg2, frame_adjustment 55 stp \reg1, \reg2, [sp, #-(\frame_adjustment)]! 56 .cfi_adjust_cfa_offset (\frame_adjustment) 57 .cfi_rel_offset \reg1, 0 58 .cfi_rel_offset \reg2, 8 59.endm 60 61.macro RESTORE_TWO_REGS_DECREASE_FRAME reg1, reg2, frame_adjustment 62 ldp \reg1, \reg2, [sp], #(\frame_adjustment) 63 .cfi_restore \reg1 64 .cfi_restore \reg2 65 .cfi_adjust_cfa_offset -(\frame_adjustment) 66.endm 67 68 /* 69 * Macro that sets up the callee save frame to conform with 70 * Runtime::CreateCalleeSaveMethod(kSaveAllCalleeSaves) 71 */ 72.macro SETUP_SAVE_ALL_CALLEE_SAVES_FRAME 73 // art::Runtime** xIP0 = &art::Runtime::instance_ 74 adrp xIP0, :got:_ZN3art7Runtime9instance_E 75 ldr xIP0, [xIP0, #:got_lo12:_ZN3art7Runtime9instance_E] 76 77 // Our registers aren't intermixed - just spill in order. 78 ldr xIP0, [xIP0] // art::Runtime* xIP0 = art::Runtime::instance_; 79 80 // ArtMethod* xIP0 = Runtime::instance_->callee_save_methods_[kSaveAllCalleeSaves]; 81 ldr xIP0, [xIP0, RUNTIME_SAVE_ALL_CALLEE_SAVES_METHOD_OFFSET] 82 83 INCREASE_FRAME 176 84 85 // Ugly compile-time check, but we only have the preprocessor. 86#if (FRAME_SIZE_SAVE_ALL_CALLEE_SAVES != 176) 87#error "FRAME_SIZE_SAVE_ALL_CALLEE_SAVES(ARM64) size not as expected." 88#endif 89 90 // Stack alignment filler [sp, #8]. 91 // FP callee-saves. 92 stp d8, d9, [sp, #16] 93 stp d10, d11, [sp, #32] 94 stp d12, d13, [sp, #48] 95 stp d14, d15, [sp, #64] 96 97 // GP callee-saves 98 SAVE_TWO_REGS x19, x20, 80 99 SAVE_TWO_REGS x21, x22, 96 100 SAVE_TWO_REGS x23, x24, 112 101 SAVE_TWO_REGS x25, x26, 128 102 SAVE_TWO_REGS x27, x28, 144 103 SAVE_TWO_REGS x29, xLR, 160 104 105 // Store ArtMethod* Runtime::callee_save_methods_[kSaveAllCalleeSaves]. 106 str xIP0, [sp] 107 // Place sp in Thread::Current()->top_quick_frame. 108 mov xIP0, sp 109 str xIP0, [xSELF, # THREAD_TOP_QUICK_FRAME_OFFSET] 110.endm 111 112 /* 113 * Macro that sets up the callee save frame to conform with 114 * Runtime::CreateCalleeSaveMethod(kSaveRefsOnly). 115 */ 116.macro SETUP_SAVE_REFS_ONLY_FRAME 117 // art::Runtime** xIP0 = &art::Runtime::instance_ 118 adrp xIP0, :got:_ZN3art7Runtime9instance_E 119 ldr xIP0, [xIP0, #:got_lo12:_ZN3art7Runtime9instance_E] 120 121 // Our registers aren't intermixed - just spill in order. 122 ldr xIP0, [xIP0] // art::Runtime* xIP0 = art::Runtime::instance_; 123 124 // ArtMethod* xIP0 = Runtime::instance_->callee_save_methods_[kSaveRefOnly]; 125 ldr xIP0, [xIP0, RUNTIME_SAVE_REFS_ONLY_METHOD_OFFSET] 126 127 INCREASE_FRAME 96 128 129 // Ugly compile-time check, but we only have the preprocessor. 130#if (FRAME_SIZE_SAVE_REFS_ONLY != 96) 131#error "FRAME_SIZE_SAVE_REFS_ONLY(ARM64) size not as expected." 132#endif 133 134 // GP callee-saves. 135 // x20 paired with ArtMethod* - see below. 136 SAVE_TWO_REGS x21, x22, 16 137 SAVE_TWO_REGS x23, x24, 32 138 SAVE_TWO_REGS x25, x26, 48 139 SAVE_TWO_REGS x27, x28, 64 140 SAVE_TWO_REGS x29, xLR, 80 141 142 // Store ArtMethod* Runtime::callee_save_methods_[kSaveRefsOnly]. 143 stp xIP0, x20, [sp] 144 .cfi_rel_offset x20, 8 145 146 // Place sp in Thread::Current()->top_quick_frame. 147 mov xIP0, sp 148 str xIP0, [xSELF, # THREAD_TOP_QUICK_FRAME_OFFSET] 149.endm 150 151// TODO: Probably no need to restore registers preserved by aapcs64. 152.macro RESTORE_SAVE_REFS_ONLY_FRAME 153 // Callee-saves. 154 RESTORE_REG x20, 8 155 RESTORE_TWO_REGS x21, x22, 16 156 RESTORE_TWO_REGS x23, x24, 32 157 RESTORE_TWO_REGS x25, x26, 48 158 RESTORE_TWO_REGS x27, x28, 64 159 RESTORE_TWO_REGS x29, xLR, 80 160 161 DECREASE_FRAME 96 162.endm 163 164.macro POP_SAVE_REFS_ONLY_FRAME 165 DECREASE_FRAME 96 166.endm 167 168.macro RESTORE_SAVE_REFS_ONLY_FRAME_AND_RETURN 169 RESTORE_SAVE_REFS_ONLY_FRAME 170 ret 171.endm 172 173 174.macro SETUP_SAVE_REFS_AND_ARGS_FRAME_INTERNAL 175 INCREASE_FRAME 224 176 177 // Ugly compile-time check, but we only have the preprocessor. 178#if (FRAME_SIZE_SAVE_REFS_AND_ARGS != 224) 179#error "FRAME_SIZE_SAVE_REFS_AND_ARGS(ARM64) size not as expected." 180#endif 181 182 // Stack alignment filler [sp, #8]. 183 // FP args. 184 stp d0, d1, [sp, #16] 185 stp d2, d3, [sp, #32] 186 stp d4, d5, [sp, #48] 187 stp d6, d7, [sp, #64] 188 189 // Core args. 190 SAVE_TWO_REGS x1, x2, 80 191 SAVE_TWO_REGS x3, x4, 96 192 SAVE_TWO_REGS x5, x6, 112 193 194 // x7, Callee-saves. 195 SAVE_TWO_REGS x7, x20, 128 196 SAVE_TWO_REGS x21, x22, 144 197 SAVE_TWO_REGS x23, x24, 160 198 SAVE_TWO_REGS x25, x26, 176 199 SAVE_TWO_REGS x27, x28, 192 200 201 // x29(callee-save) and LR. 202 SAVE_TWO_REGS x29, xLR, 208 203 204.endm 205 206 /* 207 * Macro that sets up the callee save frame to conform with 208 * Runtime::CreateCalleeSaveMethod(kSaveRefsAndArgs). 209 * 210 * TODO This is probably too conservative - saving FP & LR. 211 */ 212.macro SETUP_SAVE_REFS_AND_ARGS_FRAME 213 // art::Runtime** xIP0 = &art::Runtime::instance_ 214 adrp xIP0, :got:_ZN3art7Runtime9instance_E 215 ldr xIP0, [xIP0, #:got_lo12:_ZN3art7Runtime9instance_E] 216 217 // Our registers aren't intermixed - just spill in order. 218 ldr xIP0, [xIP0] // art::Runtime* xIP0 = art::Runtime::instance_; 219 220 // ArtMethod* xIP0 = Runtime::instance_->callee_save_methods_[kSaveRefAndArgs]; 221 ldr xIP0, [xIP0, RUNTIME_SAVE_REFS_AND_ARGS_METHOD_OFFSET] 222 223 SETUP_SAVE_REFS_AND_ARGS_FRAME_INTERNAL 224 225 str xIP0, [sp] // Store ArtMethod* Runtime::callee_save_methods_[kSaveRefsAndArgs]. 226 // Place sp in Thread::Current()->top_quick_frame. 227 mov xIP0, sp 228 str xIP0, [xSELF, # THREAD_TOP_QUICK_FRAME_OFFSET] 229.endm 230 231.macro SETUP_SAVE_REFS_AND_ARGS_FRAME_WITH_METHOD_IN_X0 232 SETUP_SAVE_REFS_AND_ARGS_FRAME_INTERNAL 233 str x0, [sp, #0] // Store ArtMethod* to bottom of stack. 234 // Place sp in Thread::Current()->top_quick_frame. 235 mov xIP0, sp 236 str xIP0, [xSELF, # THREAD_TOP_QUICK_FRAME_OFFSET] 237.endm 238 239// TODO: Probably no need to restore registers preserved by aapcs64. 240.macro RESTORE_SAVE_REFS_AND_ARGS_FRAME 241 // FP args. 242 ldp d0, d1, [sp, #16] 243 ldp d2, d3, [sp, #32] 244 ldp d4, d5, [sp, #48] 245 ldp d6, d7, [sp, #64] 246 247 // Core args. 248 RESTORE_TWO_REGS x1, x2, 80 249 RESTORE_TWO_REGS x3, x4, 96 250 RESTORE_TWO_REGS x5, x6, 112 251 252 // x7, Callee-saves. 253 RESTORE_TWO_REGS x7, x20, 128 254 RESTORE_TWO_REGS x21, x22, 144 255 RESTORE_TWO_REGS x23, x24, 160 256 RESTORE_TWO_REGS x25, x26, 176 257 RESTORE_TWO_REGS x27, x28, 192 258 259 // x29(callee-save) and LR. 260 RESTORE_TWO_REGS x29, xLR, 208 261 262 DECREASE_FRAME 224 263.endm 264 265 /* 266 * Macro that sets up the callee save frame to conform with 267 * Runtime::CreateCalleeSaveMethod(kSaveEverything) 268 */ 269.macro SETUP_SAVE_EVERYTHING_FRAME 270 INCREASE_FRAME 512 271 272 // Ugly compile-time check, but we only have the preprocessor. 273#if (FRAME_SIZE_SAVE_EVERYTHING != 512) 274#error "FRAME_SIZE_SAVE_EVERYTHING(ARM64) size not as expected." 275#endif 276 277 // Save FP registers. 278 // For better performance, store d0 and d31 separately, so that all STPs are 16-byte aligned. 279 str d0, [sp, #8] 280 stp d1, d2, [sp, #16] 281 stp d3, d4, [sp, #32] 282 stp d5, d6, [sp, #48] 283 stp d7, d8, [sp, #64] 284 stp d9, d10, [sp, #80] 285 stp d11, d12, [sp, #96] 286 stp d13, d14, [sp, #112] 287 stp d15, d16, [sp, #128] 288 stp d17, d18, [sp, #144] 289 stp d19, d20, [sp, #160] 290 stp d21, d22, [sp, #176] 291 stp d23, d24, [sp, #192] 292 stp d25, d26, [sp, #208] 293 stp d27, d28, [sp, #224] 294 stp d29, d30, [sp, #240] 295 str d31, [sp, #256] 296 297 // Save core registers. 298 SAVE_REG x0, 264 299 SAVE_TWO_REGS x1, x2, 272 300 SAVE_TWO_REGS x3, x4, 288 301 SAVE_TWO_REGS x5, x6, 304 302 SAVE_TWO_REGS x7, x8, 320 303 SAVE_TWO_REGS x9, x10, 336 304 SAVE_TWO_REGS x11, x12, 352 305 SAVE_TWO_REGS x13, x14, 368 306 SAVE_TWO_REGS x15, x16, 384 307 SAVE_TWO_REGS x17, x18, 400 308 SAVE_TWO_REGS x19, x20, 416 309 SAVE_TWO_REGS x21, x22, 432 310 SAVE_TWO_REGS x23, x24, 448 311 SAVE_TWO_REGS x25, x26, 464 312 SAVE_TWO_REGS x27, x28, 480 313 SAVE_TWO_REGS x29, xLR, 496 314 315 // art::Runtime** xIP0 = &art::Runtime::instance_ 316 adrp xIP0, :got:_ZN3art7Runtime9instance_E 317 ldr xIP0, [xIP0, #:got_lo12:_ZN3art7Runtime9instance_E] 318 319 ldr xIP0, [xIP0] // art::Runtime* xIP0 = art::Runtime::instance_; 320 321 // ArtMethod* xIP0 = Runtime::instance_->callee_save_methods_[kSaveEverything]; 322 ldr xIP0, [xIP0, RUNTIME_SAVE_EVERYTHING_METHOD_OFFSET] 323 324 // Store ArtMethod* Runtime::callee_save_methods_[kSaveEverything]. 325 str xIP0, [sp] 326 // Place sp in Thread::Current()->top_quick_frame. 327 mov xIP0, sp 328 str xIP0, [xSELF, # THREAD_TOP_QUICK_FRAME_OFFSET] 329.endm 330 331.macro RESTORE_SAVE_EVERYTHING_FRAME 332 // Restore FP registers. 333 // For better performance, load d0 and d31 separately, so that all LDPs are 16-byte aligned. 334 ldr d0, [sp, #8] 335 ldp d1, d2, [sp, #16] 336 ldp d3, d4, [sp, #32] 337 ldp d5, d6, [sp, #48] 338 ldp d7, d8, [sp, #64] 339 ldp d9, d10, [sp, #80] 340 ldp d11, d12, [sp, #96] 341 ldp d13, d14, [sp, #112] 342 ldp d15, d16, [sp, #128] 343 ldp d17, d18, [sp, #144] 344 ldp d19, d20, [sp, #160] 345 ldp d21, d22, [sp, #176] 346 ldp d23, d24, [sp, #192] 347 ldp d25, d26, [sp, #208] 348 ldp d27, d28, [sp, #224] 349 ldp d29, d30, [sp, #240] 350 ldr d31, [sp, #256] 351 352 // Restore core registers. 353 RESTORE_REG x0, 264 354 RESTORE_TWO_REGS x1, x2, 272 355 RESTORE_TWO_REGS x3, x4, 288 356 RESTORE_TWO_REGS x5, x6, 304 357 RESTORE_TWO_REGS x7, x8, 320 358 RESTORE_TWO_REGS x9, x10, 336 359 RESTORE_TWO_REGS x11, x12, 352 360 RESTORE_TWO_REGS x13, x14, 368 361 RESTORE_TWO_REGS x15, x16, 384 362 RESTORE_TWO_REGS x17, x18, 400 363 RESTORE_TWO_REGS x19, x20, 416 364 RESTORE_TWO_REGS x21, x22, 432 365 RESTORE_TWO_REGS x23, x24, 448 366 RESTORE_TWO_REGS x25, x26, 464 367 RESTORE_TWO_REGS x27, x28, 480 368 RESTORE_TWO_REGS x29, xLR, 496 369 370 DECREASE_FRAME 512 371.endm 372 373.macro RETURN_IF_RESULT_IS_ZERO 374 cbnz x0, 1f // result non-zero branch over 375 ret // return 3761: 377.endm 378 379.macro RETURN_IF_RESULT_IS_NON_ZERO 380 cbz x0, 1f // result zero branch over 381 ret // return 3821: 383.endm 384 385 /* 386 * Macro that set calls through to artDeliverPendingExceptionFromCode, where the pending 387 * exception is Thread::Current()->exception_ 388 */ 389.macro DELIVER_PENDING_EXCEPTION 390 SETUP_SAVE_ALL_CALLEE_SAVES_FRAME 391 mov x0, xSELF 392 393 // Point of no return. 394 b artDeliverPendingExceptionFromCode // artDeliverPendingExceptionFromCode(Thread*) 395 brk 0 // Unreached 396.endm 397 398.macro RETURN_OR_DELIVER_PENDING_EXCEPTION_REG reg 399 ldr \reg, [xSELF, # THREAD_EXCEPTION_OFFSET] // Get exception field. 400 cbnz \reg, 1f 401 ret 4021: 403 DELIVER_PENDING_EXCEPTION 404.endm 405 406.macro RETURN_OR_DELIVER_PENDING_EXCEPTION 407 RETURN_OR_DELIVER_PENDING_EXCEPTION_REG xIP0 408.endm 409 410// Same as above with x1. This is helpful in stubs that want to avoid clobbering another register. 411.macro RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 412 RETURN_OR_DELIVER_PENDING_EXCEPTION_REG x1 413.endm 414 415.macro RETURN_IF_W0_IS_ZERO_OR_DELIVER 416 cbnz w0, 1f // result non-zero branch over 417 ret // return 4181: 419 DELIVER_PENDING_EXCEPTION 420.endm 421 422.macro NO_ARG_RUNTIME_EXCEPTION c_name, cxx_name 423 .extern \cxx_name 424ENTRY \c_name 425 SETUP_SAVE_ALL_CALLEE_SAVES_FRAME // save all registers as basis for long jump context 426 mov x0, xSELF // pass Thread::Current 427 b \cxx_name // \cxx_name(Thread*) 428END \c_name 429.endm 430 431.macro ONE_ARG_RUNTIME_EXCEPTION c_name, cxx_name 432 .extern \cxx_name 433ENTRY \c_name 434 SETUP_SAVE_ALL_CALLEE_SAVES_FRAME // save all registers as basis for long jump context. 435 mov x1, xSELF // pass Thread::Current. 436 b \cxx_name // \cxx_name(arg, Thread*). 437 brk 0 438END \c_name 439.endm 440 441.macro TWO_ARG_RUNTIME_EXCEPTION c_name, cxx_name 442 .extern \cxx_name 443ENTRY \c_name 444 SETUP_SAVE_ALL_CALLEE_SAVES_FRAME // save all registers as basis for long jump context 445 mov x2, xSELF // pass Thread::Current 446 b \cxx_name // \cxx_name(arg1, arg2, Thread*) 447 brk 0 448END \c_name 449.endm 450 451 /* 452 * Called by managed code, saves callee saves and then calls artThrowException 453 * that will place a mock Method* at the bottom of the stack. Arg1 holds the exception. 454 */ 455ONE_ARG_RUNTIME_EXCEPTION art_quick_deliver_exception, artDeliverExceptionFromCode 456 457 /* 458 * Called by managed code to create and deliver a NullPointerException. 459 */ 460NO_ARG_RUNTIME_EXCEPTION art_quick_throw_null_pointer_exception, artThrowNullPointerExceptionFromCode 461 462 /* 463 * Call installed by a signal handler to create and deliver a NullPointerException. 464 */ 465ONE_ARG_RUNTIME_EXCEPTION art_quick_throw_null_pointer_exception_from_signal, artThrowNullPointerExceptionFromSignal 466 467 /* 468 * Called by managed code to create and deliver an ArithmeticException. 469 */ 470NO_ARG_RUNTIME_EXCEPTION art_quick_throw_div_zero, artThrowDivZeroFromCode 471 472 /* 473 * Called by managed code to create and deliver an ArrayIndexOutOfBoundsException. Arg1 holds 474 * index, arg2 holds limit. 475 */ 476TWO_ARG_RUNTIME_EXCEPTION art_quick_throw_array_bounds, artThrowArrayBoundsFromCode 477 478 /* 479 * Called by managed code to create and deliver a StringIndexOutOfBoundsException 480 * as if thrown from a call to String.charAt(). Arg1 holds index, arg2 holds limit. 481 */ 482TWO_ARG_RUNTIME_EXCEPTION art_quick_throw_string_bounds, artThrowStringBoundsFromCode 483 484 /* 485 * Called by managed code to create and deliver a StackOverflowError. 486 */ 487NO_ARG_RUNTIME_EXCEPTION art_quick_throw_stack_overflow, artThrowStackOverflowFromCode 488 489 /* 490 * Called by managed code to create and deliver a NoSuchMethodError. 491 */ 492ONE_ARG_RUNTIME_EXCEPTION art_quick_throw_no_such_method, artThrowNoSuchMethodFromCode 493 494 /* 495 * All generated callsites for interface invokes and invocation slow paths will load arguments 496 * as usual - except instead of loading arg0/x0 with the target Method*, arg0/x0 will contain 497 * the method_idx. This wrapper will save arg1-arg3, and call the appropriate C helper. 498 * NOTE: "this" is first visible argument of the target, and so can be found in arg1/x1. 499 * 500 * The helper will attempt to locate the target and return a 128-bit result in x0/x1 consisting 501 * of the target Method* in x0 and method->code_ in x1. 502 * 503 * If unsuccessful, the helper will return null/????. There will be a pending exception in the 504 * thread and we branch to another stub to deliver it. 505 * 506 * On success this wrapper will restore arguments and *jump* to the target, leaving the lr 507 * pointing back to the original caller. 508 * 509 * Adapted from ARM32 code. 510 * 511 * Clobbers xIP0. 512 */ 513.macro INVOKE_TRAMPOLINE_BODY cxx_name 514 .extern \cxx_name 515 SETUP_SAVE_REFS_AND_ARGS_FRAME // save callee saves in case allocation triggers GC 516 // Helper signature is always 517 // (method_idx, *this_object, *caller_method, *self, sp) 518 519 mov x2, xSELF // pass Thread::Current 520 mov x3, sp 521 bl \cxx_name // (method_idx, this, Thread*, SP) 522 mov xIP0, x1 // save Method*->code_ 523 RESTORE_SAVE_REFS_AND_ARGS_FRAME 524 cbz x0, 1f // did we find the target? if not go to exception delivery 525 br xIP0 // tail call to target 5261: 527 DELIVER_PENDING_EXCEPTION 528.endm 529.macro INVOKE_TRAMPOLINE c_name, cxx_name 530ENTRY \c_name 531 INVOKE_TRAMPOLINE_BODY \cxx_name 532END \c_name 533.endm 534 535INVOKE_TRAMPOLINE art_quick_invoke_interface_trampoline_with_access_check, artInvokeInterfaceTrampolineWithAccessCheck 536 537INVOKE_TRAMPOLINE art_quick_invoke_static_trampoline_with_access_check, artInvokeStaticTrampolineWithAccessCheck 538INVOKE_TRAMPOLINE art_quick_invoke_direct_trampoline_with_access_check, artInvokeDirectTrampolineWithAccessCheck 539INVOKE_TRAMPOLINE art_quick_invoke_super_trampoline_with_access_check, artInvokeSuperTrampolineWithAccessCheck 540INVOKE_TRAMPOLINE art_quick_invoke_virtual_trampoline_with_access_check, artInvokeVirtualTrampolineWithAccessCheck 541 542 543.macro INVOKE_STUB_CREATE_FRAME 544 545SAVE_SIZE=15*8 // x4, x5, x19, x20, x21, x22, x23, x24, x25, x26, x27, x28, SP, LR, FP saved. 546SAVE_SIZE_AND_METHOD=SAVE_SIZE+8 547 548 549 mov x9, sp // Save stack pointer. 550 .cfi_register sp,x9 551 552 add x10, x2, # SAVE_SIZE_AND_METHOD // calculate size of frame. 553 sub x10, sp, x10 // Calculate SP position - saves + ArtMethod* + args 554 and x10, x10, # ~0xf // Enforce 16 byte stack alignment. 555 mov sp, x10 // Set new SP. 556 557 sub x10, x9, #SAVE_SIZE // Calculate new FP (later). Done here as we must move SP 558 .cfi_def_cfa_register x10 // before this. 559 .cfi_adjust_cfa_offset SAVE_SIZE 560 561 str x28, [x10, #112] 562 .cfi_rel_offset x28, 112 563 564 stp x26, x27, [x10, #96] 565 .cfi_rel_offset x26, 96 566 .cfi_rel_offset x27, 104 567 568 stp x24, x25, [x10, #80] 569 .cfi_rel_offset x24, 80 570 .cfi_rel_offset x25, 88 571 572 stp x22, x23, [x10, #64] 573 .cfi_rel_offset x22, 64 574 .cfi_rel_offset x23, 72 575 576 stp x20, x21, [x10, #48] 577 .cfi_rel_offset x20, 48 578 .cfi_rel_offset x21, 56 579 580 stp x9, x19, [x10, #32] // Save old stack pointer and x19. 581 .cfi_rel_offset sp, 32 582 .cfi_rel_offset x19, 40 583 584 stp x4, x5, [x10, #16] // Save result and shorty addresses. 585 .cfi_rel_offset x4, 16 586 .cfi_rel_offset x5, 24 587 588 stp xFP, xLR, [x10] // Store LR & FP. 589 .cfi_rel_offset x29, 0 590 .cfi_rel_offset x30, 8 591 592 mov xFP, x10 // Use xFP now, as it's callee-saved. 593 .cfi_def_cfa_register x29 594 mov xSELF, x3 // Move thread pointer into SELF register. 595 596 // Copy arguments into stack frame. 597 // Use simple copy routine for now. 598 // 4 bytes per slot. 599 // X1 - source address 600 // W2 - args length 601 // X9 - destination address. 602 // W10 - temporary 603 add x9, sp, #8 // Destination address is bottom of stack + null. 604 605 // Copy parameters into the stack. Use numeric label as this is a macro and Clang's assembler 606 // does not have unique-id variables. 6071: 608 cmp w2, #0 609 beq 2f 610 sub w2, w2, #4 // Need 65536 bytes of range. 611 ldr w10, [x1, x2] 612 str w10, [x9, x2] 613 614 b 1b 615 6162: 617 // Store null into ArtMethod* at bottom of frame. 618 str xzr, [sp] 619.endm 620 621.macro INVOKE_STUB_CALL_AND_RETURN 622 623 // load method-> METHOD_QUICK_CODE_OFFSET 624 ldr x9, [x0, #ART_METHOD_QUICK_CODE_OFFSET_64] 625 // Branch to method. 626 blr x9 627 628 // Restore return value address and shorty address. 629 ldp x4,x5, [xFP, #16] 630 .cfi_restore x4 631 .cfi_restore x5 632 633 ldr x28, [xFP, #112] 634 .cfi_restore x28 635 636 ldp x26, x27, [xFP, #96] 637 .cfi_restore x26 638 .cfi_restore x27 639 640 ldp x24, x25, [xFP, #80] 641 .cfi_restore x24 642 .cfi_restore x25 643 644 ldp x22, x23, [xFP, #64] 645 .cfi_restore x22 646 .cfi_restore x23 647 648 ldp x20, x21, [xFP, #48] 649 .cfi_restore x20 650 .cfi_restore x21 651 652 // Store result (w0/x0/s0/d0) appropriately, depending on resultType. 653 ldrb w10, [x5] 654 655 // Check the return type and store the correct register into the jvalue in memory. 656 // Use numeric label as this is a macro and Clang's assembler does not have unique-id variables. 657 658 // Don't set anything for a void type. 659 cmp w10, #'V' 660 beq 3f 661 662 // Is it a double? 663 cmp w10, #'D' 664 bne 1f 665 str d0, [x4] 666 b 3f 667 6681: // Is it a float? 669 cmp w10, #'F' 670 bne 2f 671 str s0, [x4] 672 b 3f 673 6742: // Just store x0. Doesn't matter if it is 64 or 32 bits. 675 str x0, [x4] 676 6773: // Finish up. 678 ldp x2, x19, [xFP, #32] // Restore stack pointer and x19. 679 .cfi_restore x19 680 mov sp, x2 681 .cfi_restore sp 682 683 ldp xFP, xLR, [xFP] // Restore old frame pointer and link register. 684 .cfi_restore x29 685 .cfi_restore x30 686 687 ret 688 689.endm 690 691 692/* 693 * extern"C" void art_quick_invoke_stub(ArtMethod *method, x0 694 * uint32_t *args, x1 695 * uint32_t argsize, w2 696 * Thread *self, x3 697 * JValue *result, x4 698 * char *shorty); x5 699 * +----------------------+ 700 * | | 701 * | C/C++ frame | 702 * | LR'' | 703 * | FP'' | <- SP' 704 * +----------------------+ 705 * +----------------------+ 706 * | x28 | <- TODO: Remove callee-saves. 707 * | : | 708 * | x19 | 709 * | SP' | 710 * | X5 | 711 * | X4 | Saved registers 712 * | LR' | 713 * | FP' | <- FP 714 * +----------------------+ 715 * | uint32_t out[n-1] | 716 * | : : | Outs 717 * | uint32_t out[0] | 718 * | ArtMethod* | <- SP value=null 719 * +----------------------+ 720 * 721 * Outgoing registers: 722 * x0 - Method* 723 * x1-x7 - integer parameters. 724 * d0-d7 - Floating point parameters. 725 * xSELF = self 726 * SP = & of ArtMethod* 727 * x1 = "this" pointer. 728 * 729 */ 730ENTRY art_quick_invoke_stub 731 // Spill registers as per AACPS64 calling convention. 732 INVOKE_STUB_CREATE_FRAME 733 734 // Fill registers x/w1 to x/w7 and s/d0 to s/d7 with parameters. 735 // Parse the passed shorty to determine which register to load. 736 // Load addresses for routines that load WXSD registers. 737 adr x11, .LstoreW2 738 adr x12, .LstoreX2 739 adr x13, .LstoreS0 740 adr x14, .LstoreD0 741 742 // Initialize routine offsets to 0 for integers and floats. 743 // x8 for integers, x15 for floating point. 744 mov x8, #0 745 mov x15, #0 746 747 add x10, x5, #1 // Load shorty address, plus one to skip return value. 748 ldr w1, [x9],#4 // Load "this" parameter, and increment arg pointer. 749 750 // Loop to fill registers. 751.LfillRegisters: 752 ldrb w17, [x10], #1 // Load next character in signature, and increment. 753 cbz w17, .LcallFunction // Exit at end of signature. Shorty 0 terminated. 754 755 cmp w17, #'F' // is this a float? 756 bne .LisDouble 757 758 cmp x15, # 8*12 // Skip this load if all registers full. 759 beq .Ladvance4 760 761 add x17, x13, x15 // Calculate subroutine to jump to. 762 br x17 763 764.LisDouble: 765 cmp w17, #'D' // is this a double? 766 bne .LisLong 767 768 cmp x15, # 8*12 // Skip this load if all registers full. 769 beq .Ladvance8 770 771 add x17, x14, x15 // Calculate subroutine to jump to. 772 br x17 773 774.LisLong: 775 cmp w17, #'J' // is this a long? 776 bne .LisOther 777 778 cmp x8, # 6*12 // Skip this load if all registers full. 779 beq .Ladvance8 780 781 add x17, x12, x8 // Calculate subroutine to jump to. 782 br x17 783 784.LisOther: // Everything else takes one vReg. 785 cmp x8, # 6*12 // Skip this load if all registers full. 786 beq .Ladvance4 787 788 add x17, x11, x8 // Calculate subroutine to jump to. 789 br x17 790 791.Ladvance4: 792 add x9, x9, #4 793 b .LfillRegisters 794 795.Ladvance8: 796 add x9, x9, #8 797 b .LfillRegisters 798 799// Macro for loading a parameter into a register. 800// counter - the register with offset into these tables 801// size - the size of the register - 4 or 8 bytes. 802// register - the name of the register to be loaded. 803.macro LOADREG counter size register return 804 ldr \register , [x9], #\size 805 add \counter, \counter, 12 806 b \return 807.endm 808 809// Store ints. 810.LstoreW2: 811 LOADREG x8 4 w2 .LfillRegisters 812 LOADREG x8 4 w3 .LfillRegisters 813 LOADREG x8 4 w4 .LfillRegisters 814 LOADREG x8 4 w5 .LfillRegisters 815 LOADREG x8 4 w6 .LfillRegisters 816 LOADREG x8 4 w7 .LfillRegisters 817 818// Store longs. 819.LstoreX2: 820 LOADREG x8 8 x2 .LfillRegisters 821 LOADREG x8 8 x3 .LfillRegisters 822 LOADREG x8 8 x4 .LfillRegisters 823 LOADREG x8 8 x5 .LfillRegisters 824 LOADREG x8 8 x6 .LfillRegisters 825 LOADREG x8 8 x7 .LfillRegisters 826 827// Store singles. 828.LstoreS0: 829 LOADREG x15 4 s0 .LfillRegisters 830 LOADREG x15 4 s1 .LfillRegisters 831 LOADREG x15 4 s2 .LfillRegisters 832 LOADREG x15 4 s3 .LfillRegisters 833 LOADREG x15 4 s4 .LfillRegisters 834 LOADREG x15 4 s5 .LfillRegisters 835 LOADREG x15 4 s6 .LfillRegisters 836 LOADREG x15 4 s7 .LfillRegisters 837 838// Store doubles. 839.LstoreD0: 840 LOADREG x15 8 d0 .LfillRegisters 841 LOADREG x15 8 d1 .LfillRegisters 842 LOADREG x15 8 d2 .LfillRegisters 843 LOADREG x15 8 d3 .LfillRegisters 844 LOADREG x15 8 d4 .LfillRegisters 845 LOADREG x15 8 d5 .LfillRegisters 846 LOADREG x15 8 d6 .LfillRegisters 847 LOADREG x15 8 d7 .LfillRegisters 848 849 850.LcallFunction: 851 852 INVOKE_STUB_CALL_AND_RETURN 853 854END art_quick_invoke_stub 855 856/* extern"C" 857 * void art_quick_invoke_static_stub(ArtMethod *method, x0 858 * uint32_t *args, x1 859 * uint32_t argsize, w2 860 * Thread *self, x3 861 * JValue *result, x4 862 * char *shorty); x5 863 */ 864ENTRY art_quick_invoke_static_stub 865 // Spill registers as per AACPS64 calling convention. 866 INVOKE_STUB_CREATE_FRAME 867 868 // Fill registers x/w1 to x/w7 and s/d0 to s/d7 with parameters. 869 // Parse the passed shorty to determine which register to load. 870 // Load addresses for routines that load WXSD registers. 871 adr x11, .LstoreW1_2 872 adr x12, .LstoreX1_2 873 adr x13, .LstoreS0_2 874 adr x14, .LstoreD0_2 875 876 // Initialize routine offsets to 0 for integers and floats. 877 // x8 for integers, x15 for floating point. 878 mov x8, #0 879 mov x15, #0 880 881 add x10, x5, #1 // Load shorty address, plus one to skip return value. 882 883 // Loop to fill registers. 884.LfillRegisters2: 885 ldrb w17, [x10], #1 // Load next character in signature, and increment. 886 cbz w17, .LcallFunction2 // Exit at end of signature. Shorty 0 terminated. 887 888 cmp w17, #'F' // is this a float? 889 bne .LisDouble2 890 891 cmp x15, # 8*12 // Skip this load if all registers full. 892 beq .Ladvance4_2 893 894 add x17, x13, x15 // Calculate subroutine to jump to. 895 br x17 896 897.LisDouble2: 898 cmp w17, #'D' // is this a double? 899 bne .LisLong2 900 901 cmp x15, # 8*12 // Skip this load if all registers full. 902 beq .Ladvance8_2 903 904 add x17, x14, x15 // Calculate subroutine to jump to. 905 br x17 906 907.LisLong2: 908 cmp w17, #'J' // is this a long? 909 bne .LisOther2 910 911 cmp x8, # 7*12 // Skip this load if all registers full. 912 beq .Ladvance8_2 913 914 add x17, x12, x8 // Calculate subroutine to jump to. 915 br x17 916 917.LisOther2: // Everything else takes one vReg. 918 cmp x8, # 7*12 // Skip this load if all registers full. 919 beq .Ladvance4_2 920 921 add x17, x11, x8 // Calculate subroutine to jump to. 922 br x17 923 924.Ladvance4_2: 925 add x9, x9, #4 926 b .LfillRegisters2 927 928.Ladvance8_2: 929 add x9, x9, #8 930 b .LfillRegisters2 931 932// Store ints. 933.LstoreW1_2: 934 LOADREG x8 4 w1 .LfillRegisters2 935 LOADREG x8 4 w2 .LfillRegisters2 936 LOADREG x8 4 w3 .LfillRegisters2 937 LOADREG x8 4 w4 .LfillRegisters2 938 LOADREG x8 4 w5 .LfillRegisters2 939 LOADREG x8 4 w6 .LfillRegisters2 940 LOADREG x8 4 w7 .LfillRegisters2 941 942// Store longs. 943.LstoreX1_2: 944 LOADREG x8 8 x1 .LfillRegisters2 945 LOADREG x8 8 x2 .LfillRegisters2 946 LOADREG x8 8 x3 .LfillRegisters2 947 LOADREG x8 8 x4 .LfillRegisters2 948 LOADREG x8 8 x5 .LfillRegisters2 949 LOADREG x8 8 x6 .LfillRegisters2 950 LOADREG x8 8 x7 .LfillRegisters2 951 952// Store singles. 953.LstoreS0_2: 954 LOADREG x15 4 s0 .LfillRegisters2 955 LOADREG x15 4 s1 .LfillRegisters2 956 LOADREG x15 4 s2 .LfillRegisters2 957 LOADREG x15 4 s3 .LfillRegisters2 958 LOADREG x15 4 s4 .LfillRegisters2 959 LOADREG x15 4 s5 .LfillRegisters2 960 LOADREG x15 4 s6 .LfillRegisters2 961 LOADREG x15 4 s7 .LfillRegisters2 962 963// Store doubles. 964.LstoreD0_2: 965 LOADREG x15 8 d0 .LfillRegisters2 966 LOADREG x15 8 d1 .LfillRegisters2 967 LOADREG x15 8 d2 .LfillRegisters2 968 LOADREG x15 8 d3 .LfillRegisters2 969 LOADREG x15 8 d4 .LfillRegisters2 970 LOADREG x15 8 d5 .LfillRegisters2 971 LOADREG x15 8 d6 .LfillRegisters2 972 LOADREG x15 8 d7 .LfillRegisters2 973 974 975.LcallFunction2: 976 977 INVOKE_STUB_CALL_AND_RETURN 978 979END art_quick_invoke_static_stub 980 981 982 983/* extern"C" void art_quick_osr_stub(void** stack, x0 984 * size_t stack_size_in_bytes, x1 985 * const uin8_t* native_pc, x2 986 * JValue *result, x3 987 * char *shorty, x4 988 * Thread *self) x5 989 */ 990ENTRY art_quick_osr_stub 991SAVE_SIZE=15*8 // x3, x4, x19, x20, x21, x22, x23, x24, x25, x26, x27, x28, SP, LR, FP saved. 992 mov x9, sp // Save stack pointer. 993 .cfi_register sp,x9 994 995 sub x10, sp, # SAVE_SIZE 996 and x10, x10, # ~0xf // Enforce 16 byte stack alignment. 997 mov sp, x10 // Set new SP. 998 999 str x28, [sp, #112] 1000 stp x26, x27, [sp, #96] 1001 stp x24, x25, [sp, #80] 1002 stp x22, x23, [sp, #64] 1003 stp x20, x21, [sp, #48] 1004 stp x9, x19, [sp, #32] // Save old stack pointer and x19. 1005 stp x3, x4, [sp, #16] // Save result and shorty addresses. 1006 stp xFP, xLR, [sp] // Store LR & FP. 1007 mov xSELF, x5 // Move thread pointer into SELF register. 1008 1009 sub sp, sp, #16 1010 str xzr, [sp] // Store null for ArtMethod* slot 1011 // Branch to stub. 1012 bl .Losr_entry 1013 add sp, sp, #16 1014 1015 // Restore return value address and shorty address. 1016 ldp x3,x4, [sp, #16] 1017 ldr x28, [sp, #112] 1018 ldp x26, x27, [sp, #96] 1019 ldp x24, x25, [sp, #80] 1020 ldp x22, x23, [sp, #64] 1021 ldp x20, x21, [sp, #48] 1022 1023 // Store result (w0/x0/s0/d0) appropriately, depending on resultType. 1024 ldrb w10, [x4] 1025 1026 // Check the return type and store the correct register into the jvalue in memory. 1027 1028 // Don't set anything for a void type. 1029 cmp w10, #'V' 1030 beq .Losr_exit 1031 1032 // Is it a double? 1033 cmp w10, #'D' 1034 bne .Lno_double 1035 str d0, [x3] 1036 b .Losr_exit 1037 1038.Lno_double: // Is it a float? 1039 cmp w10, #'F' 1040 bne .Lno_float 1041 str s0, [x3] 1042 b .Losr_exit 1043 1044.Lno_float: // Just store x0. Doesn't matter if it is 64 or 32 bits. 1045 str x0, [x3] 1046 1047.Losr_exit: // Finish up. 1048 ldp x2, x19, [sp, #32] // Restore stack pointer and x19. 1049 ldp xFP, xLR, [sp] // Restore old frame pointer and link register. 1050 mov sp, x2 1051 ret 1052 1053.Losr_entry: 1054 // Update stack pointer for the callee 1055 sub sp, sp, x1 1056 1057 // Update link register slot expected by the callee. 1058 sub w1, w1, #8 1059 str lr, [sp, x1] 1060 1061 // Copy arguments into stack frame. 1062 // Use simple copy routine for now. 1063 // 4 bytes per slot. 1064 // X0 - source address 1065 // W1 - args length 1066 // SP - destination address. 1067 // W10 - temporary 1068.Losr_loop_entry: 1069 cmp w1, #0 1070 beq .Losr_loop_exit 1071 sub w1, w1, #4 1072 ldr w10, [x0, x1] 1073 str w10, [sp, x1] 1074 b .Losr_loop_entry 1075 1076.Losr_loop_exit: 1077 // Branch to the OSR entry point. 1078 br x2 1079 1080END art_quick_osr_stub 1081 1082 /* 1083 * On entry x0 is uintptr_t* gprs_ and x1 is uint64_t* fprs_ 1084 */ 1085 1086ENTRY art_quick_do_long_jump 1087 // Load FPRs 1088 ldp d0, d1, [x1], #16 1089 ldp d2, d3, [x1], #16 1090 ldp d4, d5, [x1], #16 1091 ldp d6, d7, [x1], #16 1092 ldp d8, d9, [x1], #16 1093 ldp d10, d11, [x1], #16 1094 ldp d12, d13, [x1], #16 1095 ldp d14, d15, [x1], #16 1096 ldp d16, d17, [x1], #16 1097 ldp d18, d19, [x1], #16 1098 ldp d20, d21, [x1], #16 1099 ldp d22, d23, [x1], #16 1100 ldp d24, d25, [x1], #16 1101 ldp d26, d27, [x1], #16 1102 ldp d28, d29, [x1], #16 1103 ldp d30, d31, [x1] 1104 1105 // Load GPRs 1106 // TODO: lots of those are smashed, could optimize. 1107 add x0, x0, #30*8 1108 ldp x30, x1, [x0], #-16 // LR & SP 1109 ldp x28, x29, [x0], #-16 1110 ldp x26, x27, [x0], #-16 1111 ldp x24, x25, [x0], #-16 1112 ldp x22, x23, [x0], #-16 1113 ldp x20, x21, [x0], #-16 1114 ldp x18, x19, [x0], #-16 1115 ldp x16, x17, [x0], #-16 1116 ldp x14, x15, [x0], #-16 1117 ldp x12, x13, [x0], #-16 1118 ldp x10, x11, [x0], #-16 1119 ldp x8, x9, [x0], #-16 1120 ldp x6, x7, [x0], #-16 1121 ldp x4, x5, [x0], #-16 1122 ldp x2, x3, [x0], #-16 1123 mov sp, x1 1124 1125 // Need to load PC, it's at the end (after the space for the unused XZR). Use x1. 1126 ldr x1, [x0, #33*8] 1127 // And the value of x0. 1128 ldr x0, [x0] 1129 1130 br x1 1131END art_quick_do_long_jump 1132 1133 /* 1134 * Entry from managed code that calls artLockObjectFromCode, may block for GC. x0 holds the 1135 * possibly null object to lock. 1136 * 1137 * Derived from arm32 code. 1138 */ 1139 .extern artLockObjectFromCode 1140ENTRY art_quick_lock_object 1141 cbz w0, .Lslow_lock 1142 add x4, x0, #MIRROR_OBJECT_LOCK_WORD_OFFSET // exclusive load/store has no immediate anymore 1143.Lretry_lock: 1144 ldr w2, [xSELF, #THREAD_ID_OFFSET] // TODO: Can the thread ID really change during the loop? 1145 ldxr w1, [x4] 1146 mov x3, x1 1147 and w3, w3, #LOCK_WORD_GC_STATE_MASK_SHIFTED_TOGGLED // zero the gc bits 1148 cbnz w3, .Lnot_unlocked // already thin locked 1149 // unlocked case - x1: original lock word that's zero except for the read barrier bits. 1150 orr x2, x1, x2 // x2 holds thread id with count of 0 with preserved read barrier bits 1151 stxr w3, w2, [x4] 1152 cbnz w3, .Llock_stxr_fail // store failed, retry 1153 dmb ishld // full (LoadLoad|LoadStore) memory barrier 1154 ret 1155.Lnot_unlocked: // x1: original lock word 1156 lsr w3, w1, LOCK_WORD_STATE_SHIFT 1157 cbnz w3, .Lslow_lock // if either of the top two bits are set, go slow path 1158 eor w2, w1, w2 // lock_word.ThreadId() ^ self->ThreadId() 1159 uxth w2, w2 // zero top 16 bits 1160 cbnz w2, .Lslow_lock // lock word and self thread id's match -> recursive lock 1161 // else contention, go to slow path 1162 mov x3, x1 // copy the lock word to check count overflow. 1163 and w3, w3, #LOCK_WORD_GC_STATE_MASK_SHIFTED_TOGGLED // zero the gc bits. 1164 add w2, w3, #LOCK_WORD_THIN_LOCK_COUNT_ONE // increment count in lock word placing in w2 to check overflow 1165 lsr w3, w2, #LOCK_WORD_GC_STATE_SHIFT // if the first gc state bit is set, we overflowed. 1166 cbnz w3, .Lslow_lock // if we overflow the count go slow path 1167 add w2, w1, #LOCK_WORD_THIN_LOCK_COUNT_ONE // increment count for real 1168 stxr w3, w2, [x4] 1169 cbnz w3, .Llock_stxr_fail // store failed, retry 1170 ret 1171.Llock_stxr_fail: 1172 b .Lretry_lock // retry 1173.Lslow_lock: 1174 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case we block 1175 mov x1, xSELF // pass Thread::Current 1176 bl artLockObjectFromCode // (Object* obj, Thread*) 1177 RESTORE_SAVE_REFS_ONLY_FRAME 1178 RETURN_IF_W0_IS_ZERO_OR_DELIVER 1179END art_quick_lock_object 1180 1181ENTRY art_quick_lock_object_no_inline 1182 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case we block 1183 mov x1, xSELF // pass Thread::Current 1184 bl artLockObjectFromCode // (Object* obj, Thread*) 1185 RESTORE_SAVE_REFS_ONLY_FRAME 1186 RETURN_IF_W0_IS_ZERO_OR_DELIVER 1187END art_quick_lock_object_no_inline 1188 1189 /* 1190 * Entry from managed code that calls artUnlockObjectFromCode and delivers exception on failure. 1191 * x0 holds the possibly null object to lock. 1192 * 1193 * Derived from arm32 code. 1194 */ 1195 .extern artUnlockObjectFromCode 1196ENTRY art_quick_unlock_object 1197 cbz x0, .Lslow_unlock 1198 add x4, x0, #MIRROR_OBJECT_LOCK_WORD_OFFSET // exclusive load/store has no immediate anymore 1199.Lretry_unlock: 1200#ifndef USE_READ_BARRIER 1201 ldr w1, [x4] 1202#else 1203 ldxr w1, [x4] // Need to use atomic instructions for read barrier 1204#endif 1205 lsr w2, w1, LOCK_WORD_STATE_SHIFT 1206 cbnz w2, .Lslow_unlock // if either of the top two bits are set, go slow path 1207 ldr w2, [xSELF, #THREAD_ID_OFFSET] 1208 mov x3, x1 // copy lock word to check thread id equality 1209 and w3, w3, #LOCK_WORD_GC_STATE_MASK_SHIFTED_TOGGLED // zero the gc bits 1210 eor w3, w3, w2 // lock_word.ThreadId() ^ self->ThreadId() 1211 uxth w3, w3 // zero top 16 bits 1212 cbnz w3, .Lslow_unlock // do lock word and self thread id's match? 1213 mov x3, x1 // copy lock word to detect transition to unlocked 1214 and w3, w3, #LOCK_WORD_GC_STATE_MASK_SHIFTED_TOGGLED // zero the gc bits 1215 cmp w3, #LOCK_WORD_THIN_LOCK_COUNT_ONE 1216 bpl .Lrecursive_thin_unlock 1217 // transition to unlocked 1218 mov x3, x1 1219 and w3, w3, #LOCK_WORD_GC_STATE_MASK_SHIFTED // w3: zero except for the preserved read barrier bits 1220 dmb ish // full (LoadStore|StoreStore) memory barrier 1221#ifndef USE_READ_BARRIER 1222 str w3, [x4] 1223#else 1224 stxr w2, w3, [x4] // Need to use atomic instructions for read barrier 1225 cbnz w2, .Lunlock_stxr_fail // store failed, retry 1226#endif 1227 ret 1228.Lrecursive_thin_unlock: // w1: original lock word 1229 sub w1, w1, #LOCK_WORD_THIN_LOCK_COUNT_ONE // decrement count 1230#ifndef USE_READ_BARRIER 1231 str w1, [x4] 1232#else 1233 stxr w2, w1, [x4] // Need to use atomic instructions for read barrier 1234 cbnz w2, .Lunlock_stxr_fail // store failed, retry 1235#endif 1236 ret 1237.Lunlock_stxr_fail: 1238 b .Lretry_unlock // retry 1239.Lslow_unlock: 1240 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case exception allocation triggers GC 1241 mov x1, xSELF // pass Thread::Current 1242 bl artUnlockObjectFromCode // (Object* obj, Thread*) 1243 RESTORE_SAVE_REFS_ONLY_FRAME 1244 RETURN_IF_W0_IS_ZERO_OR_DELIVER 1245END art_quick_unlock_object 1246 1247ENTRY art_quick_unlock_object_no_inline 1248 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case exception allocation triggers GC 1249 mov x1, xSELF // pass Thread::Current 1250 bl artUnlockObjectFromCode // (Object* obj, Thread*) 1251 RESTORE_SAVE_REFS_ONLY_FRAME 1252 RETURN_IF_W0_IS_ZERO_OR_DELIVER 1253END art_quick_unlock_object_no_inline 1254 1255 /* 1256 * Entry from managed code that calls artIsAssignableFromCode and on failure calls 1257 * artThrowClassCastException. 1258 */ 1259 .extern artThrowClassCastException 1260ENTRY art_quick_check_cast 1261 // Store arguments and link register 1262 // Stack needs to be 16B aligned on calls. 1263 SAVE_TWO_REGS_INCREASE_FRAME x0, x1, 32 1264 SAVE_REG xLR, 24 1265 1266 // Call runtime code 1267 bl artIsAssignableFromCode 1268 1269 // Check for exception 1270 cbz x0, .Lthrow_class_cast_exception 1271 1272 // Restore and return 1273 .cfi_remember_state 1274 RESTORE_REG xLR, 24 1275 RESTORE_TWO_REGS_DECREASE_FRAME x0, x1, 32 1276 ret 1277 .cfi_restore_state // Reset unwind info so following code unwinds. 1278 1279.Lthrow_class_cast_exception: 1280 // Restore 1281 RESTORE_REG xLR, 24 1282 RESTORE_TWO_REGS_DECREASE_FRAME x0, x1, 32 1283 1284 SETUP_SAVE_ALL_CALLEE_SAVES_FRAME // save all registers as basis for long jump context 1285 mov x2, xSELF // pass Thread::Current 1286 b artThrowClassCastException // (Class*, Class*, Thread*) 1287 brk 0 // We should not return here... 1288END art_quick_check_cast 1289 1290// Restore xReg's value from [sp, #offset] if xReg is not the same as xExclude. 1291.macro POP_REG_NE xReg, offset, xExclude 1292 .ifnc \xReg, \xExclude 1293 ldr \xReg, [sp, #\offset] // restore xReg 1294 .cfi_restore \xReg 1295 .endif 1296.endm 1297 1298// Restore xReg1's value from [sp, #offset] if xReg1 is not the same as xExclude. 1299// Restore xReg2's value from [sp, #(offset + 8)] if xReg2 is not the same as xExclude. 1300.macro POP_REGS_NE xReg1, xReg2, offset, xExclude 1301 .ifc \xReg1, \xExclude 1302 ldr \xReg2, [sp, #(\offset + 8)] // restore xReg2 1303 .else 1304 .ifc \xReg2, \xExclude 1305 ldr \xReg1, [sp, #\offset] // restore xReg1 1306 .else 1307 ldp \xReg1, \xReg2, [sp, #\offset] // restore xReg1 and xReg2 1308 .endif 1309 .endif 1310 .cfi_restore \xReg1 1311 .cfi_restore \xReg2 1312.endm 1313 1314 /* 1315 * Macro to insert read barrier, only used in art_quick_aput_obj. 1316 * xDest, wDest and xObj are registers, offset is a defined literal such as 1317 * MIRROR_OBJECT_CLASS_OFFSET. Dest needs both x and w versions of the same register to handle 1318 * name mismatch between instructions. This macro uses the lower 32b of register when possible. 1319 * TODO: When read barrier has a fast path, add heap unpoisoning support for the fast path. 1320 */ 1321.macro READ_BARRIER xDest, wDest, xObj, xTemp, wTemp, offset, number 1322#ifdef USE_READ_BARRIER 1323#ifdef USE_BAKER_READ_BARRIER 1324 ldr \wTemp, [\xObj, #MIRROR_OBJECT_LOCK_WORD_OFFSET] 1325 tbnz \wTemp, #LOCK_WORD_READ_BARRIER_STATE_SHIFT, .Lrb_slowpath\number 1326 // False dependency to avoid needing load/load fence. 1327 add \xObj, \xObj, \xTemp, lsr #32 1328 ldr \wDest, [\xObj, #\offset] // Heap reference = 32b. This also zero-extends to \xDest. 1329 UNPOISON_HEAP_REF \wDest 1330 b .Lrb_exit\number 1331#endif 1332.Lrb_slowpath\number: 1333 // Store registers used in art_quick_aput_obj (x0-x4, LR), stack is 16B aligned. 1334 SAVE_TWO_REGS_INCREASE_FRAME x0, x1, 48 1335 SAVE_TWO_REGS x2, x3, 16 1336 SAVE_TWO_REGS x4, xLR, 32 1337 1338 // mov x0, \xRef // pass ref in x0 (no-op for now since parameter ref is unused) 1339 .ifnc \xObj, x1 1340 mov x1, \xObj // pass xObj 1341 .endif 1342 mov w2, #\offset // pass offset 1343 bl artReadBarrierSlow // artReadBarrierSlow(ref, xObj, offset) 1344 // No need to unpoison return value in w0, artReadBarrierSlow() would do the unpoisoning. 1345 .ifnc \wDest, w0 1346 mov \wDest, w0 // save return value in wDest 1347 .endif 1348 1349 // Conditionally restore saved registers 1350 POP_REG_NE x0, 0, \xDest 1351 POP_REG_NE x1, 8, \xDest 1352 POP_REG_NE x2, 16, \xDest 1353 POP_REG_NE x3, 24, \xDest 1354 POP_REG_NE x4, 32, \xDest 1355 RESTORE_REG xLR, 40 1356 DECREASE_FRAME 48 1357.Lrb_exit\number: 1358#else 1359 ldr \wDest, [\xObj, #\offset] // Heap reference = 32b. This also zero-extends to \xDest. 1360 UNPOISON_HEAP_REF \wDest 1361#endif // USE_READ_BARRIER 1362.endm 1363 1364 /* 1365 * Entry from managed code for array put operations of objects where the value being stored 1366 * needs to be checked for compatibility. 1367 * x0 = array, x1 = index, x2 = value 1368 * 1369 * Currently all values should fit into w0/w1/w2, and w1 always will as indices are 32b. We 1370 * assume, though, that the upper 32b are zeroed out. At least for x1/w1 we can do better by 1371 * using index-zero-extension in load/stores. 1372 * 1373 * Temporaries: x3, x4 1374 * TODO: x4 OK? ip seems wrong here. 1375 */ 1376ENTRY art_quick_aput_obj_with_null_and_bound_check 1377 tst x0, x0 1378 bne art_quick_aput_obj_with_bound_check 1379 b art_quick_throw_null_pointer_exception 1380END art_quick_aput_obj_with_null_and_bound_check 1381 1382ENTRY art_quick_aput_obj_with_bound_check 1383 ldr w3, [x0, #MIRROR_ARRAY_LENGTH_OFFSET] 1384 cmp w3, w1 1385 bhi art_quick_aput_obj 1386 mov x0, x1 1387 mov x1, x3 1388 b art_quick_throw_array_bounds 1389END art_quick_aput_obj_with_bound_check 1390 1391#ifdef USE_READ_BARRIER 1392 .extern artReadBarrierSlow 1393#endif 1394ENTRY art_quick_aput_obj 1395 cbz x2, .Ldo_aput_null 1396 READ_BARRIER x3, w3, x0, x3, w3, MIRROR_OBJECT_CLASS_OFFSET, 0 // Heap reference = 32b 1397 // This also zero-extends to x3 1398 READ_BARRIER x3, w3, x3, x4, w4, MIRROR_CLASS_COMPONENT_TYPE_OFFSET, 1 // Heap reference = 32b 1399 // This also zero-extends to x3 1400 READ_BARRIER x4, w4, x2, x4, w4, MIRROR_OBJECT_CLASS_OFFSET, 2 // Heap reference = 32b 1401 // This also zero-extends to x4 1402 cmp w3, w4 // value's type == array's component type - trivial assignability 1403 bne .Lcheck_assignability 1404.Ldo_aput: 1405 add x3, x0, #MIRROR_OBJECT_ARRAY_DATA_OFFSET 1406 // "Compress" = do nothing 1407 POISON_HEAP_REF w2 1408 str w2, [x3, x1, lsl #2] // Heap reference = 32b 1409 ldr x3, [xSELF, #THREAD_CARD_TABLE_OFFSET] 1410 lsr x0, x0, #7 1411 strb w3, [x3, x0] 1412 ret 1413.Ldo_aput_null: 1414 add x3, x0, #MIRROR_OBJECT_ARRAY_DATA_OFFSET 1415 // "Compress" = do nothing 1416 str w2, [x3, x1, lsl #2] // Heap reference = 32b 1417 ret 1418.Lcheck_assignability: 1419 // Store arguments and link register 1420 SAVE_TWO_REGS_INCREASE_FRAME x0, x1, 32 1421 SAVE_TWO_REGS x2, xLR, 16 1422 1423 // Call runtime code 1424 mov x0, x3 // Heap reference, 32b, "uncompress" = do nothing, already zero-extended 1425 mov x1, x4 // Heap reference, 32b, "uncompress" = do nothing, already zero-extended 1426 bl artIsAssignableFromCode 1427 1428 // Check for exception 1429 cbz x0, .Lthrow_array_store_exception 1430 1431 // Restore 1432 .cfi_remember_state 1433 RESTORE_TWO_REGS x2, xLR, 16 1434 RESTORE_TWO_REGS_DECREASE_FRAME x0, x1, 32 1435 1436 add x3, x0, #MIRROR_OBJECT_ARRAY_DATA_OFFSET 1437 // "Compress" = do nothing 1438 POISON_HEAP_REF w2 1439 str w2, [x3, x1, lsl #2] // Heap reference = 32b 1440 ldr x3, [xSELF, #THREAD_CARD_TABLE_OFFSET] 1441 lsr x0, x0, #7 1442 strb w3, [x3, x0] 1443 ret 1444 .cfi_restore_state // Reset unwind info so following code unwinds. 1445.Lthrow_array_store_exception: 1446 RESTORE_TWO_REGS x2, xLR, 16 1447 RESTORE_TWO_REGS_DECREASE_FRAME x0, x1, 32 1448 1449 SETUP_SAVE_ALL_CALLEE_SAVES_FRAME 1450 mov x1, x2 // Pass value. 1451 mov x2, xSELF // Pass Thread::Current. 1452 b artThrowArrayStoreException // (Object*, Object*, Thread*). 1453 brk 0 // Unreached. 1454END art_quick_aput_obj 1455 1456// Macro to facilitate adding new allocation entrypoints. 1457.macro ONE_ARG_DOWNCALL name, entrypoint, return 1458 .extern \entrypoint 1459ENTRY \name 1460 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case of GC 1461 mov x1, xSELF // pass Thread::Current 1462 bl \entrypoint // (uint32_t type_idx, Method* method, Thread*) 1463 RESTORE_SAVE_REFS_ONLY_FRAME 1464 \return 1465END \name 1466.endm 1467 1468// Macro to facilitate adding new allocation entrypoints. 1469.macro TWO_ARG_DOWNCALL name, entrypoint, return 1470 .extern \entrypoint 1471ENTRY \name 1472 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case of GC 1473 mov x2, xSELF // pass Thread::Current 1474 bl \entrypoint // (uint32_t type_idx, Method* method, Thread*) 1475 RESTORE_SAVE_REFS_ONLY_FRAME 1476 \return 1477END \name 1478.endm 1479 1480// Macro to facilitate adding new allocation entrypoints. 1481.macro THREE_ARG_DOWNCALL name, entrypoint, return 1482 .extern \entrypoint 1483ENTRY \name 1484 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case of GC 1485 mov x3, xSELF // pass Thread::Current 1486 bl \entrypoint 1487 RESTORE_SAVE_REFS_ONLY_FRAME 1488 \return 1489END \name 1490.endm 1491 1492// Macro to facilitate adding new allocation entrypoints. 1493.macro FOUR_ARG_DOWNCALL name, entrypoint, return 1494 .extern \entrypoint 1495ENTRY \name 1496 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case of GC 1497 mov x4, xSELF // pass Thread::Current 1498 bl \entrypoint // 1499 RESTORE_SAVE_REFS_ONLY_FRAME 1500 \return 1501 DELIVER_PENDING_EXCEPTION 1502END \name 1503.endm 1504 1505// Macros taking opportunity of code similarities for downcalls with referrer. 1506.macro ONE_ARG_REF_DOWNCALL name, entrypoint, return 1507 .extern \entrypoint 1508ENTRY \name 1509 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case of GC 1510 ldr x1, [sp, #FRAME_SIZE_SAVE_REFS_ONLY] // Load referrer 1511 mov x2, xSELF // pass Thread::Current 1512 bl \entrypoint // (uint32_t type_idx, Method* method, Thread*, SP) 1513 RESTORE_SAVE_REFS_ONLY_FRAME 1514 \return 1515END \name 1516.endm 1517 1518.macro TWO_ARG_REF_DOWNCALL name, entrypoint, return 1519 .extern \entrypoint 1520ENTRY \name 1521 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case of GC 1522 ldr x2, [sp, #FRAME_SIZE_SAVE_REFS_ONLY] // Load referrer 1523 mov x3, xSELF // pass Thread::Current 1524 bl \entrypoint 1525 RESTORE_SAVE_REFS_ONLY_FRAME 1526 \return 1527END \name 1528.endm 1529 1530.macro THREE_ARG_REF_DOWNCALL name, entrypoint, return 1531 .extern \entrypoint 1532ENTRY \name 1533 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case of GC 1534 ldr x3, [sp, #FRAME_SIZE_SAVE_REFS_ONLY] // Load referrer 1535 mov x4, xSELF // pass Thread::Current 1536 bl \entrypoint 1537 RESTORE_SAVE_REFS_ONLY_FRAME 1538 \return 1539END \name 1540.endm 1541 1542.macro RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1543 cbz w0, 1f // result zero branch over 1544 ret // return 15451: 1546 DELIVER_PENDING_EXCEPTION 1547.endm 1548 1549 /* 1550 * Entry from managed code that calls artHandleFillArrayDataFromCode and delivers exception on 1551 * failure. 1552 */ 1553TWO_ARG_REF_DOWNCALL art_quick_handle_fill_data, artHandleFillArrayDataFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1554 1555 /* 1556 * Entry from managed code when uninitialized static storage, this stub will run the class 1557 * initializer and deliver the exception on error. On success the static storage base is 1558 * returned. 1559 */ 1560ONE_ARG_DOWNCALL art_quick_initialize_static_storage, artInitializeStaticStorageFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1561 1562ONE_ARG_DOWNCALL art_quick_initialize_type, artInitializeTypeFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1563ONE_ARG_DOWNCALL art_quick_initialize_type_and_verify_access, artInitializeTypeAndVerifyAccessFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1564 1565ONE_ARG_REF_DOWNCALL art_quick_get_boolean_static, artGetBooleanStaticFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1566ONE_ARG_REF_DOWNCALL art_quick_get_byte_static, artGetByteStaticFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1567ONE_ARG_REF_DOWNCALL art_quick_get_char_static, artGetCharStaticFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1568ONE_ARG_REF_DOWNCALL art_quick_get_short_static, artGetShortStaticFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1569ONE_ARG_REF_DOWNCALL art_quick_get32_static, artGet32StaticFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1570ONE_ARG_REF_DOWNCALL art_quick_get64_static, artGet64StaticFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1571ONE_ARG_REF_DOWNCALL art_quick_get_obj_static, artGetObjStaticFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1572 1573TWO_ARG_REF_DOWNCALL art_quick_get_boolean_instance, artGetBooleanInstanceFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1574TWO_ARG_REF_DOWNCALL art_quick_get_byte_instance, artGetByteInstanceFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1575TWO_ARG_REF_DOWNCALL art_quick_get_char_instance, artGetCharInstanceFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1576TWO_ARG_REF_DOWNCALL art_quick_get_short_instance, artGetShortInstanceFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1577TWO_ARG_REF_DOWNCALL art_quick_get32_instance, artGet32InstanceFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1578TWO_ARG_REF_DOWNCALL art_quick_get64_instance, artGet64InstanceFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1579TWO_ARG_REF_DOWNCALL art_quick_get_obj_instance, artGetObjInstanceFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1580 1581TWO_ARG_REF_DOWNCALL art_quick_set8_static, artSet8StaticFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1582TWO_ARG_REF_DOWNCALL art_quick_set16_static, artSet16StaticFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1583TWO_ARG_REF_DOWNCALL art_quick_set32_static, artSet32StaticFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1584TWO_ARG_REF_DOWNCALL art_quick_set_obj_static, artSetObjStaticFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1585 1586THREE_ARG_REF_DOWNCALL art_quick_set8_instance, artSet8InstanceFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1587THREE_ARG_REF_DOWNCALL art_quick_set16_instance, artSet16InstanceFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1588THREE_ARG_REF_DOWNCALL art_quick_set32_instance, artSet32InstanceFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1589THREE_ARG_REF_DOWNCALL art_quick_set64_instance, artSet64InstanceFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1590THREE_ARG_REF_DOWNCALL art_quick_set_obj_instance, artSetObjInstanceFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1591 1592// This is separated out as the argument order is different. 1593 .extern artSet64StaticFromCode 1594ENTRY art_quick_set64_static 1595 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case of GC 1596 ldr x1, [sp, #FRAME_SIZE_SAVE_REFS_ONLY] // Load referrer 1597 // x2 contains the parameter 1598 mov x3, xSELF // pass Thread::Current 1599 bl artSet64StaticFromCode 1600 RESTORE_SAVE_REFS_ONLY_FRAME 1601 RETURN_IF_W0_IS_ZERO_OR_DELIVER 1602END art_quick_set64_static 1603 1604 /* 1605 * Entry from managed code to resolve a string, this stub will 1606 * check the dex cache for a matching string (the fast path), and if not found, 1607 * it will allocate a String and deliver an exception on error. 1608 * On success the String is returned. R0 holds the string index. 1609 */ 1610 1611ENTRY art_quick_resolve_string 1612 ldr x1, [sp] // load referrer 1613 ldr w2, [x1, #ART_METHOD_DECLARING_CLASS_OFFSET] // load declaring class 1614 ldr x1, [x2, #DECLARING_CLASS_DEX_CACHE_STRINGS_OFFSET] // load string dex cache 1615 ubfx x2, x0, #0, #STRING_DEX_CACHE_HASH_BITS // get masked string index into x2 1616 ldr x2, [x1, x2, lsl #STRING_DEX_CACHE_ELEMENT_SIZE_SHIFT] // load dex cache pair into x2 1617 cmp x0, x2, lsr #32 // compare against upper 32 bits 1618 bne .Lart_quick_resolve_string_slow_path 1619 ubfx x0, x2, #0, #32 // extract lower 32 bits into x0 1620#ifdef USE_READ_BARRIER 1621 // Most common case: GC is not marking. 1622 ldr w3, [xSELF, #THREAD_IS_GC_MARKING_OFFSET] 1623 cbnz x3, .Lart_quick_resolve_string_marking 1624#endif 1625 ret 1626 1627// Slow path case, the index did not match. 1628.Lart_quick_resolve_string_slow_path: 1629 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case of GC 1630 mov x1, xSELF // pass Thread::Current 1631 bl artResolveStringFromCode // (int32_t string_idx, Thread* self) 1632 RESTORE_SAVE_REFS_ONLY_FRAME 1633 RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1634 1635// GC is marking case, need to check the mark bit. 1636.Lart_quick_resolve_string_marking: 1637 ldr x3, [x0, #MIRROR_OBJECT_LOCK_WORD_OFFSET] 1638 tbnz x3, #LOCK_WORD_MARK_BIT_SHIFT, .Lart_quick_resolve_string_no_rb 1639 // Save LR so that we can return, also x1 for alignment purposes. 1640 SAVE_TWO_REGS_INCREASE_FRAME x1, xLR, 16 // Save x1, LR. 1641 bl artReadBarrierMark // Get the marked string back. 1642 RESTORE_TWO_REGS_DECREASE_FRAME x1, xLR, 16 // Restore registers. 1643.Lart_quick_resolve_string_no_rb: 1644 ret 1645 1646END art_quick_resolve_string 1647 1648// Generate the allocation entrypoints for each allocator. 1649GENERATE_ALLOC_ENTRYPOINTS_FOR_NON_REGION_TLAB_ALLOCATORS 1650// Comment out allocators that have arm64 specific asm. 1651// GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT(_region_tlab, RegionTLAB) implemented in asm 1652// GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT_RESOLVED(_region_tlab, RegionTLAB) 1653// GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT_INITIALIZED(_region_tlab, RegionTLAB) 1654GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT_WITH_ACCESS_CHECK(_region_tlab, RegionTLAB) 1655// GENERATE_ALLOC_ENTRYPOINTS_ALLOC_ARRAY(_region_tlab, RegionTLAB) implemented in asm 1656// GENERATE_ALLOC_ENTRYPOINTS_ALLOC_ARRAY_RESOLVED(_region_tlab, RegionTLAB) 1657GENERATE_ALLOC_ENTRYPOINTS_ALLOC_ARRAY_WITH_ACCESS_CHECK(_region_tlab, RegionTLAB) 1658GENERATE_ALLOC_ENTRYPOINTS_CHECK_AND_ALLOC_ARRAY(_region_tlab, RegionTLAB) 1659GENERATE_ALLOC_ENTRYPOINTS_CHECK_AND_ALLOC_ARRAY_WITH_ACCESS_CHECK(_region_tlab, RegionTLAB) 1660GENERATE_ALLOC_ENTRYPOINTS_ALLOC_STRING_FROM_BYTES(_region_tlab, RegionTLAB) 1661GENERATE_ALLOC_ENTRYPOINTS_ALLOC_STRING_FROM_CHARS(_region_tlab, RegionTLAB) 1662GENERATE_ALLOC_ENTRYPOINTS_ALLOC_STRING_FROM_STRING(_region_tlab, RegionTLAB) 1663 1664// A hand-written override for GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT(_rosalloc, RosAlloc). 1665ENTRY art_quick_alloc_object_rosalloc 1666 // Fast path rosalloc allocation. 1667 // x0: type_idx/return value, x1: ArtMethod*, xSELF(x19): Thread::Current 1668 // x2-x7: free. 1669 ldr x2, [x1, #ART_METHOD_DEX_CACHE_TYPES_OFFSET_64] // Load dex cache resolved types array 1670 // Load the class (x2) 1671 ldr w2, [x2, x0, lsl #COMPRESSED_REFERENCE_SIZE_SHIFT] 1672 cbz x2, .Lart_quick_alloc_object_rosalloc_slow_path // Check null class 1673 ldr x3, [xSELF, #THREAD_LOCAL_ALLOC_STACK_TOP_OFFSET] // Check if the thread local 1674 // allocation stack has room. 1675 // ldp won't work due to large offset. 1676 ldr x4, [xSELF, #THREAD_LOCAL_ALLOC_STACK_END_OFFSET] 1677 cmp x3, x4 1678 bhs .Lart_quick_alloc_object_rosalloc_slow_path 1679 ldr w3, [x2, #MIRROR_CLASS_OBJECT_SIZE_ALLOC_FAST_PATH_OFFSET] // Load the object size (x3) 1680 cmp x3, #ROSALLOC_MAX_THREAD_LOCAL_BRACKET_SIZE // Check if the size is for a thread 1681 // local allocation. Also does the 1682 // finalizable and initialization 1683 // checks. 1684 bhs .Lart_quick_alloc_object_rosalloc_slow_path 1685 // Compute the rosalloc bracket index 1686 // from the size. Since the size is 1687 // already aligned we can combine the 1688 // two shifts together. 1689 add x4, xSELF, x3, lsr #(ROSALLOC_BRACKET_QUANTUM_SIZE_SHIFT - POINTER_SIZE_SHIFT) 1690 // Subtract pointer size since ther 1691 // are no runs for 0 byte allocations 1692 // and the size is already aligned. 1693 ldr x4, [x4, #(THREAD_ROSALLOC_RUNS_OFFSET - __SIZEOF_POINTER__)] 1694 // Load the free list head (x3). This 1695 // will be the return val. 1696 ldr x3, [x4, #(ROSALLOC_RUN_FREE_LIST_OFFSET + ROSALLOC_RUN_FREE_LIST_HEAD_OFFSET)] 1697 cbz x3, .Lart_quick_alloc_object_rosalloc_slow_path 1698 // "Point of no slow path". Won't go to the slow path from here on. OK to clobber x0 and x1. 1699 ldr x1, [x3, #ROSALLOC_SLOT_NEXT_OFFSET] // Load the next pointer of the head 1700 // and update the list head with the 1701 // next pointer. 1702 str x1, [x4, #(ROSALLOC_RUN_FREE_LIST_OFFSET + ROSALLOC_RUN_FREE_LIST_HEAD_OFFSET)] 1703 // Store the class pointer in the 1704 // header. This also overwrites the 1705 // next pointer. The offsets are 1706 // asserted to match. 1707#if ROSALLOC_SLOT_NEXT_OFFSET != MIRROR_OBJECT_CLASS_OFFSET 1708#error "Class pointer needs to overwrite next pointer." 1709#endif 1710 POISON_HEAP_REF w2 1711 str w2, [x3, #MIRROR_OBJECT_CLASS_OFFSET] 1712 // Fence. This is "ish" not "ishst" so 1713 // that it also ensures ordering of 1714 // the object size load with respect 1715 // to later accesses to the class 1716 // object. Alternatively we could use 1717 // "ishst" if we use load-acquire for 1718 // the class status load. 1719 // Needs to be done before pushing on 1720 // allocation since Heap::VisitObjects 1721 // relies on seeing the class pointer. 1722 // b/28790624 1723 dmb ish 1724 // Push the new object onto the thread 1725 // local allocation stack and 1726 // increment the thread local 1727 // allocation stack top. 1728 ldr x1, [xSELF, #THREAD_LOCAL_ALLOC_STACK_TOP_OFFSET] 1729 str w3, [x1], #COMPRESSED_REFERENCE_SIZE // (Increment x1 as a side effect.) 1730 str x1, [xSELF, #THREAD_LOCAL_ALLOC_STACK_TOP_OFFSET] 1731 // Decrement the size of the free list 1732 ldr w1, [x4, #(ROSALLOC_RUN_FREE_LIST_OFFSET + ROSALLOC_RUN_FREE_LIST_SIZE_OFFSET)] 1733 sub x1, x1, #1 1734 // TODO: consider combining this store 1735 // and the list head store above using 1736 // strd. 1737 str w1, [x4, #(ROSALLOC_RUN_FREE_LIST_OFFSET + ROSALLOC_RUN_FREE_LIST_SIZE_OFFSET)] 1738 1739 mov x0, x3 // Set the return value and return. 1740 ret 1741.Lart_quick_alloc_object_rosalloc_slow_path: 1742 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case of GC 1743 mov x2, xSELF // pass Thread::Current 1744 bl artAllocObjectFromCodeRosAlloc // (uint32_t type_idx, Method* method, Thread*) 1745 RESTORE_SAVE_REFS_ONLY_FRAME 1746 RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1747END art_quick_alloc_object_rosalloc 1748 1749 1750// The common fast path code for art_quick_alloc_array_region_tlab. 1751.macro ALLOC_ARRAY_TLAB_FAST_PATH slowPathLabel, xClass, wClass, xCount, wCount, xTemp0, wTemp0, xTemp1, wTemp1, xTemp2, wTemp2 1752 // Check null class 1753 cbz \wClass, \slowPathLabel 1754 ALLOC_ARRAY_TLAB_FAST_PATH_RESOLVED \slowPathLabel, \xClass, \wClass, \xCount, \wCount, \xTemp0, \wTemp0, \xTemp1, \wTemp1, \xTemp2, \wTemp2 1755.endm 1756 1757// The common fast path code for art_quick_alloc_array_region_tlab. 1758.macro ALLOC_ARRAY_TLAB_FAST_PATH_RESOLVED slowPathLabel, xClass, wClass, xCount, wCount, xTemp0, wTemp0, xTemp1, wTemp1, xTemp2, wTemp2 1759 // Array classes are never finalizable or uninitialized, no need to check. 1760 ldr \wTemp0, [\xClass, #MIRROR_CLASS_COMPONENT_TYPE_OFFSET] // Load component type 1761 UNPOISON_HEAP_REF \wTemp0 1762 ldr \wTemp0, [\xTemp0, #MIRROR_CLASS_OBJECT_PRIMITIVE_TYPE_OFFSET] 1763 lsr \xTemp0, \xTemp0, #PRIMITIVE_TYPE_SIZE_SHIFT_SHIFT // Component size shift is in high 16 1764 // bits. 1765 // xCount is holding a 32 bit value, 1766 // it can not overflow. 1767 lsl \xTemp1, \xCount, \xTemp0 // Calculate data size 1768 // Add array data offset and alignment. 1769 add \xTemp1, \xTemp1, #(MIRROR_INT_ARRAY_DATA_OFFSET + OBJECT_ALIGNMENT_MASK) 1770#if MIRROR_LONG_ARRAY_DATA_OFFSET != MIRROR_INT_ARRAY_DATA_OFFSET + 4 1771#error Long array data offset must be 4 greater than int array data offset. 1772#endif 1773 1774 add \xTemp0, \xTemp0, #1 // Add 4 to the length only if the 1775 // component size shift is 3 1776 // (for 64 bit alignment). 1777 and \xTemp0, \xTemp0, #4 1778 add \xTemp1, \xTemp1, \xTemp0 1779 and \xTemp1, \xTemp1, #OBJECT_ALIGNMENT_MASK_TOGGLED64 // Apply alignemnt mask 1780 // (addr + 7) & ~7. The mask must 1781 // be 64 bits to keep high bits in 1782 // case of overflow. 1783 // Negative sized arrays are handled here since xCount holds a zero extended 32 bit value. 1784 // Negative ints become large 64 bit unsigned ints which will always be larger than max signed 1785 // 32 bit int. Since the max shift for arrays is 3, it can not become a negative 64 bit int. 1786 cmp \xTemp1, #MIN_LARGE_OBJECT_THRESHOLD // Possibly a large object, go slow 1787 bhs \slowPathLabel // path. 1788 1789 ldr \xTemp0, [xSELF, #THREAD_LOCAL_POS_OFFSET] // Check tlab for space, note that 1790 // we use (end - begin) to handle 1791 // negative size arrays. It is 1792 // assumed that a negative size will 1793 // always be greater unsigned than 1794 // region size. 1795 ldr \xTemp2, [xSELF, #THREAD_LOCAL_END_OFFSET] 1796 sub \xTemp2, \xTemp2, \xTemp0 1797 cmp \xTemp1, \xTemp2 1798 bhi \slowPathLabel 1799 // "Point of no slow path". Won't go to the slow path from here on. OK to clobber x0 and x1. 1800 // Move old thread_local_pos to x0 1801 // for the return value. 1802 mov x0, \xTemp0 1803 add \xTemp0, \xTemp0, \xTemp1 1804 str \xTemp0, [xSELF, #THREAD_LOCAL_POS_OFFSET] // Store new thread_local_pos. 1805 ldr \xTemp0, [xSELF, #THREAD_LOCAL_OBJECTS_OFFSET] // Increment thread_local_objects. 1806 add \xTemp0, \xTemp0, #1 1807 str \xTemp0, [xSELF, #THREAD_LOCAL_OBJECTS_OFFSET] 1808 POISON_HEAP_REF \wClass 1809 str \wClass, [x0, #MIRROR_OBJECT_CLASS_OFFSET] // Store the class pointer. 1810 str \wCount, [x0, #MIRROR_ARRAY_LENGTH_OFFSET] // Store the array length. 1811 // Fence. 1812 dmb ishst 1813 ret 1814.endm 1815 1816// The common fast path code for art_quick_alloc_object_tlab and art_quick_alloc_object_region_tlab. 1817// 1818// x0: type_idx/return value, x1: ArtMethod*, x2: Class*, xSELF(x19): Thread::Current 1819// x3-x7: free. 1820// Need to preserve x0 and x1 to the slow path. 1821.macro ALLOC_OBJECT_TLAB_FAST_PATH slowPathLabel 1822 cbz x2, \slowPathLabel // Check null class 1823 ALLOC_OBJECT_TLAB_FAST_PATH_RESOLVED \slowPathLabel 1824.endm 1825 1826// TODO: delete ALLOC_OBJECT_TLAB_FAST_PATH_RESOLVED since it is the same as 1827// ALLOC_OBJECT_TLAB_FAST_PATH_INITIALIZED. 1828.macro ALLOC_OBJECT_TLAB_FAST_PATH_RESOLVED slowPathLabel 1829 ALLOC_OBJECT_TLAB_FAST_PATH_INITIALIZED \slowPathLabel 1830.endm 1831 1832.macro ALLOC_OBJECT_TLAB_FAST_PATH_INITIALIZED slowPathLabel 1833 ldr x4, [xSELF, #THREAD_LOCAL_POS_OFFSET] 1834 ldr x5, [xSELF, #THREAD_LOCAL_END_OFFSET] 1835 ldr w7, [x2, #MIRROR_CLASS_OBJECT_SIZE_ALLOC_FAST_PATH_OFFSET] // Load the object size (x7). 1836 add x6, x4, x7 // Add object size to tlab pos. 1837 cmp x6, x5 // Check if it fits, overflow works 1838 // since the tlab pos and end are 32 1839 // bit values. 1840 bhi \slowPathLabel 1841 // "Point of no slow path". Won't go to the slow path from here on. OK to clobber x0 and x1. 1842 mov x0, x4 1843 str x6, [xSELF, #THREAD_LOCAL_POS_OFFSET] // Store new thread_local_pos. 1844 ldr x5, [xSELF, #THREAD_LOCAL_OBJECTS_OFFSET] // Increment thread_local_objects. 1845 add x5, x5, #1 1846 str x5, [xSELF, #THREAD_LOCAL_OBJECTS_OFFSET] 1847 POISON_HEAP_REF w2 1848 str w2, [x0, #MIRROR_OBJECT_CLASS_OFFSET] // Store the class pointer. 1849 // Fence. This is "ish" not "ishst" so 1850 // that the code after this allocation 1851 // site will see the right values in 1852 // the fields of the class. 1853 // Alternatively we could use "ishst" 1854 // if we use load-acquire for the 1855 // object size load.) 1856 dmb ish 1857 ret 1858.endm 1859 1860// A hand-written override for GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT(_tlab, TLAB). 1861ENTRY art_quick_alloc_object_tlab 1862 // Fast path tlab allocation. 1863 // x0: type_idx/return value, x1: ArtMethod*, xSELF(x19): Thread::Current 1864 // x2-x7: free. 1865#if defined(USE_READ_BARRIER) 1866 mvn x0, xzr // Read barrier not supported here. 1867 ret // Return -1. 1868#endif 1869 ldr x2, [x1, #ART_METHOD_DEX_CACHE_TYPES_OFFSET_64] // Load dex cache resolved types array 1870 // Load the class (x2) 1871 ldr w2, [x2, x0, lsl #COMPRESSED_REFERENCE_SIZE_SHIFT] 1872 ALLOC_OBJECT_TLAB_FAST_PATH .Lart_quick_alloc_object_tlab_slow_path 1873.Lart_quick_alloc_object_tlab_slow_path: 1874 SETUP_SAVE_REFS_ONLY_FRAME // Save callee saves in case of GC. 1875 mov x2, xSELF // Pass Thread::Current. 1876 bl artAllocObjectFromCodeTLAB // (uint32_t type_idx, Method* method, Thread*) 1877 RESTORE_SAVE_REFS_ONLY_FRAME 1878 RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1879END art_quick_alloc_object_tlab 1880 1881// The common code for art_quick_alloc_object_*region_tlab 1882.macro GENERATE_ALLOC_OBJECT_REGION_TLAB name, entrypoint, fast_path, is_resolved, read_barrier 1883ENTRY \name 1884 // Fast path region tlab allocation. 1885 // x0: type_idx/resolved class/return value, x1: ArtMethod*, xSELF(x19): Thread::Current 1886 // If is_resolved is 1 then x0 is the resolved type, otherwise it is the index. 1887 // x2-x7: free. 1888#if !defined(USE_READ_BARRIER) 1889 mvn x0, xzr // Read barrier must be enabled here. 1890 ret // Return -1. 1891#endif 1892.if \is_resolved 1893 mov x2, x0 // class is actually stored in x0 already 1894.else 1895 ldr x2, [x1, #ART_METHOD_DEX_CACHE_TYPES_OFFSET_64] // Load dex cache resolved types array 1896 // Load the class (x2) 1897 ldr w2, [x2, x0, lsl #COMPRESSED_REFERENCE_SIZE_SHIFT] 1898 // If the class is null, go slow path. The check is required to read the lock word. 1899 cbz w2, .Lslow_path\name 1900.endif 1901.if \read_barrier 1902 // Most common case: GC is not marking. 1903 ldr w3, [xSELF, #THREAD_IS_GC_MARKING_OFFSET] 1904 cbnz x3, .Lmarking\name 1905.endif 1906.Ldo_allocation\name: 1907 \fast_path .Lslow_path\name 1908.Lmarking\name: 1909.if \read_barrier 1910 // GC is marking, check the lock word of the class for the mark bit. 1911 // Class is not null, check mark bit in lock word. 1912 ldr w3, [x2, #MIRROR_OBJECT_LOCK_WORD_OFFSET] 1913 // If the bit is not zero, do the allocation. 1914 tbnz w3, #LOCK_WORD_MARK_BIT_SHIFT, .Ldo_allocation\name 1915 // The read barrier slow path. Mark 1916 // the class. 1917 SAVE_TWO_REGS_INCREASE_FRAME x0, x1, 32 // Save registers (x0, x1, lr). 1918 SAVE_REG xLR, 24 // Align sp by 16 bytes. 1919 mov x0, x2 // Pass the class as the first param. 1920 bl artReadBarrierMark 1921 mov x2, x0 // Get the (marked) class back. 1922 RESTORE_REG xLR, 24 1923 RESTORE_TWO_REGS_DECREASE_FRAME x0, x1, 32 // Restore registers. 1924 b .Ldo_allocation\name 1925.endif 1926.Lslow_path\name: 1927 SETUP_SAVE_REFS_ONLY_FRAME // Save callee saves in case of GC. 1928 mov x2, xSELF // Pass Thread::Current. 1929 bl \entrypoint // (uint32_t type_idx, Method* method, Thread*) 1930 RESTORE_SAVE_REFS_ONLY_FRAME 1931 RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1932END \name 1933.endm 1934 1935// Use ALLOC_OBJECT_TLAB_FAST_PATH_RESOLVED since the null check is already done in GENERATE_ALLOC_OBJECT_TLAB. 1936GENERATE_ALLOC_OBJECT_REGION_TLAB art_quick_alloc_object_region_tlab, artAllocObjectFromCodeRegionTLAB, ALLOC_OBJECT_TLAB_FAST_PATH_RESOLVED, 0, 1 1937// No read barrier for the resolved or initialized cases since the caller is responsible for the 1938// read barrier due to the to-space invariant. 1939GENERATE_ALLOC_OBJECT_REGION_TLAB art_quick_alloc_object_resolved_region_tlab, artAllocObjectFromCodeResolvedRegionTLAB, ALLOC_OBJECT_TLAB_FAST_PATH_RESOLVED, 1, 0 1940GENERATE_ALLOC_OBJECT_REGION_TLAB art_quick_alloc_object_initialized_region_tlab, artAllocObjectFromCodeInitializedRegionTLAB, ALLOC_OBJECT_TLAB_FAST_PATH_INITIALIZED, 1, 0 1941 1942// TODO: We could use this macro for the normal tlab allocator too. 1943 1944// The common code for art_quick_alloc_array_*region_tlab 1945.macro GENERATE_ALLOC_ARRAY_REGION_TLAB name, entrypoint, fast_path, is_resolved 1946ENTRY \name 1947 // Fast path array allocation for region tlab allocation. 1948 // x0: uint32_t type_idx 1949 // x1: int32_t component_count 1950 // x2: ArtMethod* method 1951 // x3-x7: free. 1952#if !defined(USE_READ_BARRIER) 1953 mvn x0, xzr // Read barrier must be enabled here. 1954 ret // Return -1. 1955#endif 1956.if \is_resolved 1957 mov x3, x0 1958 // If already resolved, class is stored in x0 1959.else 1960 ldr x3, [x2, #ART_METHOD_DEX_CACHE_TYPES_OFFSET_64] // Load dex cache resolved types array 1961 // Load the class (x2) 1962 ldr w3, [x3, x0, lsl #COMPRESSED_REFERENCE_SIZE_SHIFT] 1963.endif 1964 // Most common case: GC is not marking. 1965 ldr w4, [xSELF, #THREAD_IS_GC_MARKING_OFFSET] 1966 cbnz x4, .Lmarking\name 1967.Ldo_allocation\name: 1968 \fast_path .Lslow_path\name, x3, w3, x1, w1, x4, w4, x5, w5, x6, w6 1969.Lmarking\name: 1970 // GC is marking, check the lock word of the class for the mark bit. 1971 // If the class is null, go slow path. The check is required to read the lock word. 1972 cbz w3, .Lslow_path\name 1973 // Class is not null, check mark bit in lock word. 1974 ldr w4, [x3, #MIRROR_OBJECT_LOCK_WORD_OFFSET] 1975 // If the bit is not zero, do the allocation. 1976 tbnz w4, #LOCK_WORD_MARK_BIT_SHIFT, .Ldo_allocation\name 1977 // The read barrier slow path. Mark 1978 // the class. 1979 stp x0, x1, [sp, #-32]! // Save registers (x0, x1, x2, lr). 1980 stp x2, xLR, [sp, #16] 1981 mov x0, x3 // Pass the class as the first param. 1982 bl artReadBarrierMark 1983 mov x3, x0 // Get the (marked) class back. 1984 ldp x2, xLR, [sp, #16] 1985 ldp x0, x1, [sp], #32 // Restore registers. 1986 b .Ldo_allocation\name 1987.Lslow_path\name: 1988 // x0: uint32_t type_idx / mirror::Class* klass (if resolved) 1989 // x1: int32_t component_count 1990 // x2: ArtMethod* method 1991 // x3: Thread* self 1992 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case of GC 1993 mov x3, xSELF // pass Thread::Current 1994 bl \entrypoint 1995 RESTORE_SAVE_REFS_ONLY_FRAME 1996 RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1997END \name 1998.endm 1999 2000GENERATE_ALLOC_ARRAY_REGION_TLAB art_quick_alloc_array_region_tlab, artAllocArrayFromCodeRegionTLAB, ALLOC_ARRAY_TLAB_FAST_PATH, 0 2001// TODO: art_quick_alloc_array_resolved_region_tlab seems to not get called. Investigate compiler. 2002GENERATE_ALLOC_ARRAY_REGION_TLAB art_quick_alloc_array_resolved_region_tlab, artAllocArrayFromCodeResolvedRegionTLAB, ALLOC_ARRAY_TLAB_FAST_PATH_RESOLVED, 1 2003 2004 /* 2005 * Called by managed code when the thread has been asked to suspend. 2006 */ 2007 .extern artTestSuspendFromCode 2008ENTRY art_quick_test_suspend 2009 SETUP_SAVE_EVERYTHING_FRAME // save callee saves for stack crawl 2010 mov x0, xSELF 2011 bl artTestSuspendFromCode // (Thread*) 2012 RESTORE_SAVE_EVERYTHING_FRAME 2013 ret 2014END art_quick_test_suspend 2015 2016ENTRY art_quick_implicit_suspend 2017 mov x0, xSELF 2018 SETUP_SAVE_REFS_ONLY_FRAME // save callee saves for stack crawl 2019 bl artTestSuspendFromCode // (Thread*) 2020 RESTORE_SAVE_REFS_ONLY_FRAME_AND_RETURN 2021END art_quick_implicit_suspend 2022 2023 /* 2024 * Called by managed code that is attempting to call a method on a proxy class. On entry 2025 * x0 holds the proxy method and x1 holds the receiver; The frame size of the invoked proxy 2026 * method agrees with a ref and args callee save frame. 2027 */ 2028 .extern artQuickProxyInvokeHandler 2029ENTRY art_quick_proxy_invoke_handler 2030 SETUP_SAVE_REFS_AND_ARGS_FRAME_WITH_METHOD_IN_X0 2031 mov x2, xSELF // pass Thread::Current 2032 mov x3, sp // pass SP 2033 bl artQuickProxyInvokeHandler // (Method* proxy method, receiver, Thread*, SP) 2034 ldr x2, [xSELF, THREAD_EXCEPTION_OFFSET] 2035 cbnz x2, .Lexception_in_proxy // success if no exception is pending 2036 RESTORE_SAVE_REFS_AND_ARGS_FRAME // Restore frame 2037 fmov d0, x0 // Store result in d0 in case it was float or double 2038 ret // return on success 2039.Lexception_in_proxy: 2040 RESTORE_SAVE_REFS_AND_ARGS_FRAME 2041 DELIVER_PENDING_EXCEPTION 2042END art_quick_proxy_invoke_handler 2043 2044 /* 2045 * Called to resolve an imt conflict. 2046 * x0 is the conflict ArtMethod. 2047 * xIP1 is a hidden argument that holds the target interface method's dex method index. 2048 * 2049 * Note that this stub writes to xIP0, xIP1, and x0. 2050 */ 2051 .extern artInvokeInterfaceTrampoline 2052ENTRY art_quick_imt_conflict_trampoline 2053 ldr xIP0, [sp, #0] // Load referrer 2054 ldr xIP0, [xIP0, #ART_METHOD_DEX_CACHE_METHODS_OFFSET_64] // Load dex cache methods array 2055 ldr xIP0, [xIP0, xIP1, lsl #POINTER_SIZE_SHIFT] // Load interface method 2056 ldr xIP1, [x0, #ART_METHOD_JNI_OFFSET_64] // Load ImtConflictTable 2057 ldr x0, [xIP1] // Load first entry in ImtConflictTable. 2058.Limt_table_iterate: 2059 cmp x0, xIP0 2060 // Branch if found. Benchmarks have shown doing a branch here is better. 2061 beq .Limt_table_found 2062 // If the entry is null, the interface method is not in the ImtConflictTable. 2063 cbz x0, .Lconflict_trampoline 2064 // Iterate over the entries of the ImtConflictTable. 2065 ldr x0, [xIP1, #(2 * __SIZEOF_POINTER__)]! 2066 b .Limt_table_iterate 2067.Limt_table_found: 2068 // We successfully hit an entry in the table. Load the target method 2069 // and jump to it. 2070 ldr x0, [xIP1, #__SIZEOF_POINTER__] 2071 ldr xIP0, [x0, #ART_METHOD_QUICK_CODE_OFFSET_64] 2072 br xIP0 2073.Lconflict_trampoline: 2074 // Call the runtime stub to populate the ImtConflictTable and jump to the 2075 // resolved method. 2076 INVOKE_TRAMPOLINE_BODY artInvokeInterfaceTrampoline 2077END art_quick_imt_conflict_trampoline 2078 2079ENTRY art_quick_resolution_trampoline 2080 SETUP_SAVE_REFS_AND_ARGS_FRAME 2081 mov x2, xSELF 2082 mov x3, sp 2083 bl artQuickResolutionTrampoline // (called, receiver, Thread*, SP) 2084 cbz x0, 1f 2085 mov xIP0, x0 // Remember returned code pointer in xIP0. 2086 ldr x0, [sp, #0] // artQuickResolutionTrampoline puts called method in *SP. 2087 RESTORE_SAVE_REFS_AND_ARGS_FRAME 2088 br xIP0 20891: 2090 RESTORE_SAVE_REFS_AND_ARGS_FRAME 2091 DELIVER_PENDING_EXCEPTION 2092END art_quick_resolution_trampoline 2093 2094/* 2095 * Generic JNI frame layout: 2096 * 2097 * #-------------------# 2098 * | | 2099 * | caller method... | 2100 * #-------------------# <--- SP on entry 2101 * | Return X30/LR | 2102 * | X29/FP | callee save 2103 * | X28 | callee save 2104 * | X27 | callee save 2105 * | X26 | callee save 2106 * | X25 | callee save 2107 * | X24 | callee save 2108 * | X23 | callee save 2109 * | X22 | callee save 2110 * | X21 | callee save 2111 * | X20 | callee save 2112 * | X19 | callee save 2113 * | X7 | arg7 2114 * | X6 | arg6 2115 * | X5 | arg5 2116 * | X4 | arg4 2117 * | X3 | arg3 2118 * | X2 | arg2 2119 * | X1 | arg1 2120 * | D7 | float arg 8 2121 * | D6 | float arg 7 2122 * | D5 | float arg 6 2123 * | D4 | float arg 5 2124 * | D3 | float arg 4 2125 * | D2 | float arg 3 2126 * | D1 | float arg 2 2127 * | D0 | float arg 1 2128 * | Method* | <- X0 2129 * #-------------------# 2130 * | local ref cookie | // 4B 2131 * | handle scope size | // 4B 2132 * #-------------------# 2133 * | JNI Call Stack | 2134 * #-------------------# <--- SP on native call 2135 * | | 2136 * | Stack for Regs | The trampoline assembly will pop these values 2137 * | | into registers for native call 2138 * #-------------------# 2139 * | Native code ptr | 2140 * #-------------------# 2141 * | Free scratch | 2142 * #-------------------# 2143 * | Ptr to (1) | <--- SP 2144 * #-------------------# 2145 */ 2146 /* 2147 * Called to do a generic JNI down-call 2148 */ 2149ENTRY art_quick_generic_jni_trampoline 2150 SETUP_SAVE_REFS_AND_ARGS_FRAME_WITH_METHOD_IN_X0 2151 2152 // Save SP , so we can have static CFI info. 2153 mov x28, sp 2154 .cfi_def_cfa_register x28 2155 2156 // This looks the same, but is different: this will be updated to point to the bottom 2157 // of the frame when the handle scope is inserted. 2158 mov xFP, sp 2159 2160 mov xIP0, #5120 2161 sub sp, sp, xIP0 2162 2163 // prepare for artQuickGenericJniTrampoline call 2164 // (Thread*, SP) 2165 // x0 x1 <= C calling convention 2166 // xSELF xFP <= where they are 2167 2168 mov x0, xSELF // Thread* 2169 mov x1, xFP 2170 bl artQuickGenericJniTrampoline // (Thread*, sp) 2171 2172 // The C call will have registered the complete save-frame on success. 2173 // The result of the call is: 2174 // x0: pointer to native code, 0 on error. 2175 // x1: pointer to the bottom of the used area of the alloca, can restore stack till there. 2176 2177 // Check for error = 0. 2178 cbz x0, .Lexception_in_native 2179 2180 // Release part of the alloca. 2181 mov sp, x1 2182 2183 // Save the code pointer 2184 mov xIP0, x0 2185 2186 // Load parameters from frame into registers. 2187 // TODO Check with artQuickGenericJniTrampoline. 2188 // Also, check again APPCS64 - the stack arguments are interleaved. 2189 ldp x0, x1, [sp] 2190 ldp x2, x3, [sp, #16] 2191 ldp x4, x5, [sp, #32] 2192 ldp x6, x7, [sp, #48] 2193 2194 ldp d0, d1, [sp, #64] 2195 ldp d2, d3, [sp, #80] 2196 ldp d4, d5, [sp, #96] 2197 ldp d6, d7, [sp, #112] 2198 2199 add sp, sp, #128 2200 2201 blr xIP0 // native call. 2202 2203 // result sign extension is handled in C code 2204 // prepare for artQuickGenericJniEndTrampoline call 2205 // (Thread*, result, result_f) 2206 // x0 x1 x2 <= C calling convention 2207 mov x1, x0 // Result (from saved). 2208 mov x0, xSELF // Thread register. 2209 fmov x2, d0 // d0 will contain floating point result, but needs to go into x2 2210 2211 bl artQuickGenericJniEndTrampoline 2212 2213 // Pending exceptions possible. 2214 ldr x2, [xSELF, THREAD_EXCEPTION_OFFSET] 2215 cbnz x2, .Lexception_in_native 2216 2217 // Tear down the alloca. 2218 mov sp, x28 2219 .cfi_def_cfa_register sp 2220 2221 // Tear down the callee-save frame. 2222 RESTORE_SAVE_REFS_AND_ARGS_FRAME 2223 2224 // store into fpr, for when it's a fpr return... 2225 fmov d0, x0 2226 ret 2227 2228.Lexception_in_native: 2229 // Move to x1 then sp to please assembler. 2230 ldr x1, [xSELF, # THREAD_TOP_QUICK_FRAME_OFFSET] 2231 mov sp, x1 2232 .cfi_def_cfa_register sp 2233 # This will create a new save-all frame, required by the runtime. 2234 DELIVER_PENDING_EXCEPTION 2235END art_quick_generic_jni_trampoline 2236 2237/* 2238 * Called to bridge from the quick to interpreter ABI. On entry the arguments match those 2239 * of a quick call: 2240 * x0 = method being called/to bridge to. 2241 * x1..x7, d0..d7 = arguments to that method. 2242 */ 2243ENTRY art_quick_to_interpreter_bridge 2244 SETUP_SAVE_REFS_AND_ARGS_FRAME // Set up frame and save arguments. 2245 2246 // x0 will contain mirror::ArtMethod* method. 2247 mov x1, xSELF // How to get Thread::Current() ??? 2248 mov x2, sp 2249 2250 // uint64_t artQuickToInterpreterBridge(mirror::ArtMethod* method, Thread* self, 2251 // mirror::ArtMethod** sp) 2252 bl artQuickToInterpreterBridge 2253 2254 RESTORE_SAVE_REFS_AND_ARGS_FRAME // TODO: no need to restore arguments in this case. 2255 2256 fmov d0, x0 2257 2258 RETURN_OR_DELIVER_PENDING_EXCEPTION 2259END art_quick_to_interpreter_bridge 2260 2261 2262// 2263// Instrumentation-related stubs 2264// 2265 .extern artInstrumentationMethodEntryFromCode 2266ENTRY art_quick_instrumentation_entry 2267 SETUP_SAVE_REFS_AND_ARGS_FRAME 2268 2269 mov x20, x0 // Preserve method reference in a callee-save. 2270 2271 mov x2, xSELF 2272 mov x3, xLR 2273 bl artInstrumentationMethodEntryFromCode // (Method*, Object*, Thread*, LR) 2274 2275 mov xIP0, x0 // x0 = result of call. 2276 mov x0, x20 // Reload method reference. 2277 2278 RESTORE_SAVE_REFS_AND_ARGS_FRAME // Note: will restore xSELF 2279 adr xLR, art_quick_instrumentation_exit 2280 br xIP0 // Tail-call method with lr set to art_quick_instrumentation_exit. 2281END art_quick_instrumentation_entry 2282 2283 .extern artInstrumentationMethodExitFromCode 2284ENTRY art_quick_instrumentation_exit 2285 mov xLR, #0 // Clobber LR for later checks. 2286 2287 SETUP_SAVE_REFS_ONLY_FRAME 2288 2289 // We need to save x0 and d0. We could use a callee-save from SETUP_REF_ONLY, but then 2290 // we would need to fully restore it. As there are a lot of callee-save registers, it seems 2291 // easier to have an extra small stack area. 2292 2293 str x0, [sp, #-16]! // Save integer result. 2294 .cfi_adjust_cfa_offset 16 2295 str d0, [sp, #8] // Save floating-point result. 2296 2297 add x1, sp, #16 // Pass SP. 2298 mov x2, x0 // Pass integer result. 2299 fmov x3, d0 // Pass floating-point result. 2300 mov x0, xSELF // Pass Thread. 2301 bl artInstrumentationMethodExitFromCode // (Thread*, SP, gpr_res, fpr_res) 2302 2303 mov xIP0, x0 // Return address from instrumentation call. 2304 mov xLR, x1 // r1 is holding link register if we're to bounce to deoptimize 2305 2306 ldr d0, [sp, #8] // Restore floating-point result. 2307 ldr x0, [sp], #16 // Restore integer result, and drop stack area. 2308 .cfi_adjust_cfa_offset 16 2309 2310 POP_SAVE_REFS_ONLY_FRAME 2311 2312 br xIP0 // Tail-call out. 2313END art_quick_instrumentation_exit 2314 2315 /* 2316 * Instrumentation has requested that we deoptimize into the interpreter. The deoptimization 2317 * will long jump to the upcall with a special exception of -1. 2318 */ 2319 .extern artDeoptimize 2320ENTRY art_quick_deoptimize 2321 SETUP_SAVE_ALL_CALLEE_SAVES_FRAME 2322 mov x0, xSELF // Pass thread. 2323 bl artDeoptimize // artDeoptimize(Thread*) 2324 brk 0 2325END art_quick_deoptimize 2326 2327 /* 2328 * Compiled code has requested that we deoptimize into the interpreter. The deoptimization 2329 * will long jump to the upcall with a special exception of -1. 2330 */ 2331 .extern artDeoptimizeFromCompiledCode 2332ENTRY art_quick_deoptimize_from_compiled_code 2333 SETUP_SAVE_EVERYTHING_FRAME 2334 mov x0, xSELF // Pass thread. 2335 bl artDeoptimizeFromCompiledCode // artDeoptimizeFromCompiledCode(Thread*) 2336 brk 0 2337END art_quick_deoptimize_from_compiled_code 2338 2339 2340 /* 2341 * String's indexOf. 2342 * 2343 * TODO: Not very optimized. 2344 * On entry: 2345 * x0: string object (known non-null) 2346 * w1: char to match (known <= 0xFFFF) 2347 * w2: Starting offset in string data 2348 */ 2349ENTRY art_quick_indexof 2350 ldr w3, [x0, #MIRROR_STRING_COUNT_OFFSET] 2351 add x0, x0, #MIRROR_STRING_VALUE_OFFSET 2352 2353 /* Clamp start to [0..count] */ 2354 cmp w2, #0 2355 csel w2, wzr, w2, lt 2356 cmp w2, w3 2357 csel w2, w3, w2, gt 2358 2359 /* Save a copy to compute result */ 2360 mov x5, x0 2361 2362 /* Build pointer to start of data to compare and pre-bias */ 2363 add x0, x0, x2, lsl #1 2364 sub x0, x0, #2 2365 2366 /* Compute iteration count */ 2367 sub w2, w3, w2 2368 2369 /* 2370 * At this point we have: 2371 * x0: start of the data to test 2372 * w1: char to compare 2373 * w2: iteration count 2374 * x5: original start of string data 2375 */ 2376 2377 subs w2, w2, #4 2378 b.lt .Lindexof_remainder 2379 2380.Lindexof_loop4: 2381 ldrh w6, [x0, #2]! 2382 ldrh w7, [x0, #2]! 2383 ldrh wIP0, [x0, #2]! 2384 ldrh wIP1, [x0, #2]! 2385 cmp w6, w1 2386 b.eq .Lmatch_0 2387 cmp w7, w1 2388 b.eq .Lmatch_1 2389 cmp wIP0, w1 2390 b.eq .Lmatch_2 2391 cmp wIP1, w1 2392 b.eq .Lmatch_3 2393 subs w2, w2, #4 2394 b.ge .Lindexof_loop4 2395 2396.Lindexof_remainder: 2397 adds w2, w2, #4 2398 b.eq .Lindexof_nomatch 2399 2400.Lindexof_loop1: 2401 ldrh w6, [x0, #2]! 2402 cmp w6, w1 2403 b.eq .Lmatch_3 2404 subs w2, w2, #1 2405 b.ne .Lindexof_loop1 2406 2407.Lindexof_nomatch: 2408 mov x0, #-1 2409 ret 2410 2411.Lmatch_0: 2412 sub x0, x0, #6 2413 sub x0, x0, x5 2414 asr x0, x0, #1 2415 ret 2416.Lmatch_1: 2417 sub x0, x0, #4 2418 sub x0, x0, x5 2419 asr x0, x0, #1 2420 ret 2421.Lmatch_2: 2422 sub x0, x0, #2 2423 sub x0, x0, x5 2424 asr x0, x0, #1 2425 ret 2426.Lmatch_3: 2427 sub x0, x0, x5 2428 asr x0, x0, #1 2429 ret 2430END art_quick_indexof 2431 2432 /* 2433 * Create a function `name` calling the ReadBarrier::Mark routine, 2434 * getting its argument and returning its result through W register 2435 * `wreg` (corresponding to X register `xreg`), saving and restoring 2436 * all caller-save registers. 2437 * 2438 * If `wreg` is different from `w0`, the generated function follows a 2439 * non-standard runtime calling convention: 2440 * - register `wreg` is used to pass the (sole) argument of this 2441 * function (instead of W0); 2442 * - register `wreg` is used to return the result of this function 2443 * (instead of W0); 2444 * - W0 is treated like a normal (non-argument) caller-save register; 2445 * - everything else is the same as in the standard runtime calling 2446 * convention (e.g. standard callee-save registers are preserved). 2447 */ 2448.macro READ_BARRIER_MARK_REG name, wreg, xreg 2449ENTRY \name 2450 // Reference is null, no work to do at all. 2451 cbz \wreg, .Lret_rb_\name 2452 /* 2453 * Allocate 46 stack slots * 8 = 368 bytes: 2454 * - 20 slots for core registers X0-X19 2455 * - 24 slots for floating-point registers D0-D7 and D16-D31 2456 * - 1 slot for return address register XLR 2457 * - 1 padding slot for 16-byte stack alignment 2458 */ 2459 // Use wIP0 as temp and check the mark bit of the reference. wIP0 is not used by the compiler. 2460 ldr wIP0, [\xreg, #MIRROR_OBJECT_LOCK_WORD_OFFSET] 2461 tbz wIP0, #LOCK_WORD_MARK_BIT_SHIFT, .Lslow_path_rb_\name 2462 ret 2463.Lslow_path_rb_\name: 2464 // Save all potentially live caller-save core registers. 2465 SAVE_TWO_REGS_INCREASE_FRAME x0, x1, 368 2466 SAVE_TWO_REGS x2, x3, 16 2467 SAVE_TWO_REGS x4, x5, 32 2468 SAVE_TWO_REGS x6, x7, 48 2469 SAVE_TWO_REGS x8, x9, 64 2470 SAVE_TWO_REGS x10, x11, 80 2471 SAVE_TWO_REGS x12, x13, 96 2472 SAVE_TWO_REGS x14, x15, 112 2473 SAVE_TWO_REGS x16, x17, 128 2474 SAVE_TWO_REGS x18, x19, 144 2475 // Save all potentially live caller-save floating-point registers. 2476 stp d0, d1, [sp, #160] 2477 stp d2, d3, [sp, #176] 2478 stp d4, d5, [sp, #192] 2479 stp d6, d7, [sp, #208] 2480 stp d16, d17, [sp, #224] 2481 stp d18, d19, [sp, #240] 2482 stp d20, d21, [sp, #256] 2483 stp d22, d23, [sp, #272] 2484 stp d24, d25, [sp, #288] 2485 stp d26, d27, [sp, #304] 2486 stp d28, d29, [sp, #320] 2487 stp d30, d31, [sp, #336] 2488 // Save return address. 2489 // (sp + #352 is a padding slot) 2490 SAVE_REG xLR, 360 2491 2492 .ifnc \wreg, w0 2493 mov w0, \wreg // Pass arg1 - obj from `wreg` 2494 .endif 2495 bl artReadBarrierMark // artReadBarrierMark(obj) 2496 .ifnc \wreg, w0 2497 mov \wreg, w0 // Return result into `wreg` 2498 .endif 2499 2500 // Restore core regs, except `xreg`, as `wreg` is used to return the 2501 // result of this function (simply remove it from the stack instead). 2502 POP_REGS_NE x0, x1, 0, \xreg 2503 POP_REGS_NE x2, x3, 16, \xreg 2504 POP_REGS_NE x4, x5, 32, \xreg 2505 POP_REGS_NE x6, x7, 48, \xreg 2506 POP_REGS_NE x8, x9, 64, \xreg 2507 POP_REGS_NE x10, x11, 80, \xreg 2508 POP_REGS_NE x12, x13, 96, \xreg 2509 POP_REGS_NE x14, x15, 112, \xreg 2510 POP_REGS_NE x16, x17, 128, \xreg 2511 POP_REGS_NE x18, x19, 144, \xreg 2512 // Restore floating-point registers. 2513 ldp d0, d1, [sp, #160] 2514 ldp d2, d3, [sp, #176] 2515 ldp d4, d5, [sp, #192] 2516 ldp d6, d7, [sp, #208] 2517 ldp d16, d17, [sp, #224] 2518 ldp d18, d19, [sp, #240] 2519 ldp d20, d21, [sp, #256] 2520 ldp d22, d23, [sp, #272] 2521 ldp d24, d25, [sp, #288] 2522 ldp d26, d27, [sp, #304] 2523 ldp d28, d29, [sp, #320] 2524 ldp d30, d31, [sp, #336] 2525 // Restore return address and remove padding. 2526 RESTORE_REG xLR, 360 2527 DECREASE_FRAME 368 2528.Lret_rb_\name: 2529 ret 2530END \name 2531.endm 2532 2533READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg00, w0, x0 2534READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg01, w1, x1 2535READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg02, w2, x2 2536READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg03, w3, x3 2537READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg04, w4, x4 2538READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg05, w5, x5 2539READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg06, w6, x6 2540READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg07, w7, x7 2541READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg08, w8, x8 2542READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg09, w9, x9 2543READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg10, w10, x10 2544READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg11, w11, x11 2545READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg12, w12, x12 2546READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg13, w13, x13 2547READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg14, w14, x14 2548READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg15, w15, x15 2549// READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg16, w16, x16 ip0 is blocked 2550READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg17, w17, x17 2551READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg18, w18, x18 2552READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg19, w19, x19 2553READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg20, w20, x20 2554READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg21, w21, x21 2555READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg22, w22, x22 2556READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg23, w23, x23 2557READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg24, w24, x24 2558READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg25, w25, x25 2559READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg26, w26, x26 2560READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg27, w27, x27 2561READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg28, w28, x28 2562READ_BARRIER_MARK_REG art_quick_read_barrier_mark_reg29, w29, x29 2563