quick_entrypoints_arm64.S revision d72945c6b18831630af8755fcf041f5a59c82cae
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 /* 23 * Macro that sets up the callee save frame to conform with 24 * Runtime::CreateCalleeSaveMethod(kSaveAll) 25 */ 26.macro SETUP_SAVE_ALL_CALLEE_SAVE_FRAME 27 adrp xIP0, :got:_ZN3art7Runtime9instance_E 28 ldr xIP0, [xIP0, #:got_lo12:_ZN3art7Runtime9instance_E] 29 30 // Our registers aren't intermixed - just spill in order. 31 ldr xIP0, [xIP0] // xIP0 = & (art::Runtime * art::Runtime.instance_) . 32 33 // xIP0 = (ArtMethod*) Runtime.instance_.callee_save_methods[kRefAndArgs] . 34 // Loads appropriate callee-save-method. 35 ldr xIP0, [xIP0, RUNTIME_SAVE_ALL_CALLEE_SAVE_FRAME_OFFSET ] 36 37 sub sp, sp, #176 38 .cfi_adjust_cfa_offset 176 39 40 // Ugly compile-time check, but we only have the preprocessor. 41#if (FRAME_SIZE_SAVE_ALL_CALLEE_SAVE != 176) 42#error "SAVE_ALL_CALLEE_SAVE_FRAME(ARM64) size not as expected." 43#endif 44 45 // Stack alignment filler [sp, #8]. 46 // FP callee-saves. 47 stp d8, d9, [sp, #16] 48 stp d10, d11, [sp, #32] 49 stp d12, d13, [sp, #48] 50 stp d14, d15, [sp, #64] 51 52 // GP callee-saves 53 stp x19, x20, [sp, #80] 54 .cfi_rel_offset x19, 80 55 .cfi_rel_offset x20, 88 56 57 stp x21, x22, [sp, #96] 58 .cfi_rel_offset x21, 96 59 .cfi_rel_offset x22, 104 60 61 stp x23, x24, [sp, #112] 62 .cfi_rel_offset x23, 112 63 .cfi_rel_offset x24, 120 64 65 stp x25, x26, [sp, #128] 66 .cfi_rel_offset x25, 128 67 .cfi_rel_offset x26, 136 68 69 stp x27, x28, [sp, #144] 70 .cfi_rel_offset x27, 144 71 .cfi_rel_offset x28, 152 72 73 stp x29, xLR, [sp, #160] 74 .cfi_rel_offset x29, 160 75 .cfi_rel_offset x30, 168 76 77 // Store ArtMethod* Runtime::callee_save_methods_[kRefsAndArgs]. 78 str xIP0, [sp] 79 // Place sp in Thread::Current()->top_quick_frame. 80 mov xIP0, sp 81 str xIP0, [xSELF, # THREAD_TOP_QUICK_FRAME_OFFSET] 82.endm 83 84 /* 85 * Macro that sets up the callee save frame to conform with 86 * Runtime::CreateCalleeSaveMethod(kRefsOnly). 87 */ 88.macro SETUP_REFS_ONLY_CALLEE_SAVE_FRAME 89 adrp xIP0, :got:_ZN3art7Runtime9instance_E 90 ldr xIP0, [xIP0, #:got_lo12:_ZN3art7Runtime9instance_E] 91 92 // Our registers aren't intermixed - just spill in order. 93 ldr xIP0, [xIP0] // xIP0 = & (art::Runtime * art::Runtime.instance_) . 94 95 // xIP0 = (ArtMethod*) Runtime.instance_.callee_save_methods[kRefOnly] . 96 // Loads appropriate callee-save-method. 97 ldr xIP0, [xIP0, RUNTIME_REFS_ONLY_CALLEE_SAVE_FRAME_OFFSET ] 98 99 sub sp, sp, #96 100 .cfi_adjust_cfa_offset 96 101 102 // Ugly compile-time check, but we only have the preprocessor. 103#if (FRAME_SIZE_REFS_ONLY_CALLEE_SAVE != 96) 104#error "REFS_ONLY_CALLEE_SAVE_FRAME(ARM64) size not as expected." 105#endif 106 107 // GP callee-saves. 108 // x20 paired with ArtMethod* - see below. 109 stp x21, x22, [sp, #16] 110 .cfi_rel_offset x21, 16 111 .cfi_rel_offset x22, 24 112 113 stp x23, x24, [sp, #32] 114 .cfi_rel_offset x23, 32 115 .cfi_rel_offset x24, 40 116 117 stp x25, x26, [sp, #48] 118 .cfi_rel_offset x25, 48 119 .cfi_rel_offset x26, 56 120 121 stp x27, x28, [sp, #64] 122 .cfi_rel_offset x27, 64 123 .cfi_rel_offset x28, 72 124 125 stp x29, xLR, [sp, #80] 126 .cfi_rel_offset x29, 80 127 .cfi_rel_offset x30, 88 128 129 // Store ArtMethod* Runtime::callee_save_methods_[kRefsOnly]. 130 stp xIP0, x20, [sp] 131 .cfi_rel_offset x20, 8 132 133 // Place sp in Thread::Current()->top_quick_frame. 134 mov xIP0, sp 135 str xIP0, [xSELF, # THREAD_TOP_QUICK_FRAME_OFFSET] 136.endm 137 138// TODO: Probably no need to restore registers preserved by aapcs64. 139.macro RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME 140 // Callee-saves. 141 ldr x20, [sp, #8] 142 .cfi_restore x20 143 144 ldp x21, x22, [sp, #16] 145 .cfi_restore x21 146 .cfi_restore x22 147 148 ldp x23, x24, [sp, #32] 149 .cfi_restore x23 150 .cfi_restore x24 151 152 ldp x25, x26, [sp, #48] 153 .cfi_restore x25 154 .cfi_restore x26 155 156 ldp x27, x28, [sp, #64] 157 .cfi_restore x27 158 .cfi_restore x28 159 160 ldp x29, xLR, [sp, #80] 161 .cfi_restore x29 162 .cfi_restore x30 163 164 add sp, sp, #96 165 .cfi_adjust_cfa_offset -96 166.endm 167 168.macro POP_REFS_ONLY_CALLEE_SAVE_FRAME 169 add sp, sp, #96 170 .cfi_adjust_cfa_offset - 96 171.endm 172 173.macro RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME_AND_RETURN 174 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME 175 ret 176.endm 177 178 179.macro SETUP_REFS_AND_ARGS_CALLEE_SAVE_FRAME_INTERNAL 180 sub sp, sp, #224 181 .cfi_adjust_cfa_offset 224 182 183 // Ugly compile-time check, but we only have the preprocessor. 184#if (FRAME_SIZE_REFS_AND_ARGS_CALLEE_SAVE != 224) 185#error "REFS_AND_ARGS_CALLEE_SAVE_FRAME(ARM64) size not as expected." 186#endif 187 188 // Stack alignment filler [sp, #8]. 189 // FP args. 190 stp d0, d1, [sp, #16] 191 stp d2, d3, [sp, #32] 192 stp d4, d5, [sp, #48] 193 stp d6, d7, [sp, #64] 194 195 // Core args. 196 stp x1, x2, [sp, #80] 197 .cfi_rel_offset x1, 80 198 .cfi_rel_offset x2, 88 199 200 stp x3, x4, [sp, #96] 201 .cfi_rel_offset x3, 96 202 .cfi_rel_offset x4, 104 203 204 stp x5, x6, [sp, #112] 205 .cfi_rel_offset x5, 112 206 .cfi_rel_offset x6, 120 207 208 // x7, Callee-saves. 209 stp x7, x20, [sp, #128] 210 .cfi_rel_offset x7, 128 211 .cfi_rel_offset x20, 136 212 213 stp x21, x22, [sp, #144] 214 .cfi_rel_offset x21, 144 215 .cfi_rel_offset x22, 152 216 217 stp x23, x24, [sp, #160] 218 .cfi_rel_offset x23, 160 219 .cfi_rel_offset x24, 168 220 221 stp x25, x26, [sp, #176] 222 .cfi_rel_offset x25, 176 223 .cfi_rel_offset x26, 184 224 225 stp x27, x28, [sp, #192] 226 .cfi_rel_offset x27, 192 227 .cfi_rel_offset x28, 200 228 229 // x29(callee-save) and LR. 230 stp x29, xLR, [sp, #208] 231 .cfi_rel_offset x29, 208 232 .cfi_rel_offset x30, 216 233 234.endm 235 236 /* 237 * Macro that sets up the callee save frame to conform with 238 * Runtime::CreateCalleeSaveMethod(kRefsAndArgs). 239 * 240 * TODO This is probably too conservative - saving FP & LR. 241 */ 242.macro SETUP_REFS_AND_ARGS_CALLEE_SAVE_FRAME 243 adrp xIP0, :got:_ZN3art7Runtime9instance_E 244 ldr xIP0, [xIP0, #:got_lo12:_ZN3art7Runtime9instance_E] 245 246 // Our registers aren't intermixed - just spill in order. 247 ldr xIP0, [xIP0] // xIP0 = & (art::Runtime * art::Runtime.instance_) . 248 249 // xIP0 = (ArtMethod*) Runtime.instance_.callee_save_methods[kRefAndArgs] . 250 ldr xIP0, [xIP0, RUNTIME_REFS_AND_ARGS_CALLEE_SAVE_FRAME_OFFSET ] 251 252 SETUP_REFS_AND_ARGS_CALLEE_SAVE_FRAME_INTERNAL 253 254 str xIP0, [sp] // Store ArtMethod* Runtime::callee_save_methods_[kRefsAndArgs] 255 // Place sp in Thread::Current()->top_quick_frame. 256 mov xIP0, sp 257 str xIP0, [xSELF, # THREAD_TOP_QUICK_FRAME_OFFSET] 258.endm 259 260.macro SETUP_REFS_AND_ARGS_CALLEE_SAVE_FRAME_WITH_METHOD_IN_X0 261 SETUP_REFS_AND_ARGS_CALLEE_SAVE_FRAME_INTERNAL 262 str x0, [sp, #0] // Store ArtMethod* to bottom of stack. 263 // Place sp in Thread::Current()->top_quick_frame. 264 mov xIP0, sp 265 str xIP0, [xSELF, # THREAD_TOP_QUICK_FRAME_OFFSET] 266.endm 267 268// TODO: Probably no need to restore registers preserved by aapcs64. 269.macro RESTORE_REFS_AND_ARGS_CALLEE_SAVE_FRAME 270 // FP args. 271 ldp d0, d1, [sp, #16] 272 ldp d2, d3, [sp, #32] 273 ldp d4, d5, [sp, #48] 274 ldp d6, d7, [sp, #64] 275 276 // Core args. 277 ldp x1, x2, [sp, #80] 278 .cfi_restore x1 279 .cfi_restore x2 280 281 ldp x3, x4, [sp, #96] 282 .cfi_restore x3 283 .cfi_restore x4 284 285 ldp x5, x6, [sp, #112] 286 .cfi_restore x5 287 .cfi_restore x6 288 289 // x7, Callee-saves. 290 ldp x7, x20, [sp, #128] 291 .cfi_restore x7 292 .cfi_restore x20 293 294 ldp x21, x22, [sp, #144] 295 .cfi_restore x21 296 .cfi_restore x22 297 298 ldp x23, x24, [sp, #160] 299 .cfi_restore x23 300 .cfi_restore x24 301 302 ldp x25, x26, [sp, #176] 303 .cfi_restore x25 304 .cfi_restore x26 305 306 ldp x27, x28, [sp, #192] 307 .cfi_restore x27 308 .cfi_restore x28 309 310 // x29(callee-save) and LR. 311 ldp x29, xLR, [sp, #208] 312 .cfi_restore x29 313 .cfi_restore x30 314 315 add sp, sp, #224 316 .cfi_adjust_cfa_offset -224 317.endm 318 319.macro RETURN_IF_RESULT_IS_ZERO 320 cbnz x0, 1f // result non-zero branch over 321 ret // return 3221: 323.endm 324 325.macro RETURN_IF_RESULT_IS_NON_ZERO 326 cbz x0, 1f // result zero branch over 327 ret // return 3281: 329.endm 330 331 /* 332 * Macro that set calls through to artDeliverPendingExceptionFromCode, where the pending 333 * exception is Thread::Current()->exception_ 334 */ 335.macro DELIVER_PENDING_EXCEPTION 336 SETUP_SAVE_ALL_CALLEE_SAVE_FRAME 337 mov x0, xSELF 338 339 // Point of no return. 340 b artDeliverPendingExceptionFromCode // artDeliverPendingExceptionFromCode(Thread*) 341 brk 0 // Unreached 342.endm 343 344.macro RETURN_OR_DELIVER_PENDING_EXCEPTION_REG reg 345 ldr \reg, [xSELF, # THREAD_EXCEPTION_OFFSET] // Get exception field. 346 cbnz \reg, 1f 347 ret 3481: 349 DELIVER_PENDING_EXCEPTION 350.endm 351 352.macro RETURN_OR_DELIVER_PENDING_EXCEPTION 353 RETURN_OR_DELIVER_PENDING_EXCEPTION_REG xIP0 354.endm 355 356// Same as above with x1. This is helpful in stubs that want to avoid clobbering another register. 357.macro RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 358 RETURN_OR_DELIVER_PENDING_EXCEPTION_REG x1 359.endm 360 361.macro RETURN_IF_W0_IS_ZERO_OR_DELIVER 362 cbnz w0, 1f // result non-zero branch over 363 ret // return 3641: 365 DELIVER_PENDING_EXCEPTION 366.endm 367 368.macro NO_ARG_RUNTIME_EXCEPTION c_name, cxx_name 369 .extern \cxx_name 370ENTRY \c_name 371 SETUP_SAVE_ALL_CALLEE_SAVE_FRAME // save all registers as basis for long jump context 372 mov x0, xSELF // pass Thread::Current 373 b \cxx_name // \cxx_name(Thread*) 374END \c_name 375.endm 376 377.macro ONE_ARG_RUNTIME_EXCEPTION c_name, cxx_name 378 .extern \cxx_name 379ENTRY \c_name 380 SETUP_SAVE_ALL_CALLEE_SAVE_FRAME // save all registers as basis for long jump context. 381 mov x1, xSELF // pass Thread::Current. 382 b \cxx_name // \cxx_name(arg, Thread*). 383 brk 0 384END \c_name 385.endm 386 387.macro TWO_ARG_RUNTIME_EXCEPTION c_name, cxx_name 388 .extern \cxx_name 389ENTRY \c_name 390 SETUP_SAVE_ALL_CALLEE_SAVE_FRAME // save all registers as basis for long jump context 391 mov x2, xSELF // pass Thread::Current 392 b \cxx_name // \cxx_name(arg1, arg2, Thread*) 393 brk 0 394END \c_name 395.endm 396 397 /* 398 * Called by managed code, saves callee saves and then calls artThrowException 399 * that will place a mock Method* at the bottom of the stack. Arg1 holds the exception. 400 */ 401ONE_ARG_RUNTIME_EXCEPTION art_quick_deliver_exception, artDeliverExceptionFromCode 402 403 /* 404 * Called by managed code to create and deliver a NullPointerException. 405 */ 406NO_ARG_RUNTIME_EXCEPTION art_quick_throw_null_pointer_exception, artThrowNullPointerExceptionFromCode 407 408 /* 409 * Called by managed code to create and deliver an ArithmeticException. 410 */ 411NO_ARG_RUNTIME_EXCEPTION art_quick_throw_div_zero, artThrowDivZeroFromCode 412 413 /* 414 * Called by managed code to create and deliver an ArrayIndexOutOfBoundsException. Arg1 holds 415 * index, arg2 holds limit. 416 */ 417TWO_ARG_RUNTIME_EXCEPTION art_quick_throw_array_bounds, artThrowArrayBoundsFromCode 418 419 /* 420 * Called by managed code to create and deliver a StackOverflowError. 421 */ 422NO_ARG_RUNTIME_EXCEPTION art_quick_throw_stack_overflow, artThrowStackOverflowFromCode 423 424 /* 425 * Called by managed code to create and deliver a NoSuchMethodError. 426 */ 427ONE_ARG_RUNTIME_EXCEPTION art_quick_throw_no_such_method, artThrowNoSuchMethodFromCode 428 429 /* 430 * All generated callsites for interface invokes and invocation slow paths will load arguments 431 * as usual - except instead of loading arg0/x0 with the target Method*, arg0/x0 will contain 432 * the method_idx. This wrapper will save arg1-arg3, and call the appropriate C helper. 433 * NOTE: "this" is first visible argument of the target, and so can be found in arg1/x1. 434 * 435 * The helper will attempt to locate the target and return a 128-bit result in x0/x1 consisting 436 * of the target Method* in x0 and method->code_ in x1. 437 * 438 * If unsuccessful, the helper will return null/????. There will be a pending exception in the 439 * thread and we branch to another stub to deliver it. 440 * 441 * On success this wrapper will restore arguments and *jump* to the target, leaving the lr 442 * pointing back to the original caller. 443 * 444 * Adapted from ARM32 code. 445 * 446 * Clobbers xIP0. 447 */ 448.macro INVOKE_TRAMPOLINE_BODY cxx_name 449 .extern \cxx_name 450 SETUP_REFS_AND_ARGS_CALLEE_SAVE_FRAME // save callee saves in case allocation triggers GC 451 // Helper signature is always 452 // (method_idx, *this_object, *caller_method, *self, sp) 453 454 mov x2, xSELF // pass Thread::Current 455 mov x3, sp 456 bl \cxx_name // (method_idx, this, Thread*, SP) 457 mov xIP0, x1 // save Method*->code_ 458 RESTORE_REFS_AND_ARGS_CALLEE_SAVE_FRAME 459 cbz x0, 1f // did we find the target? if not go to exception delivery 460 br xIP0 // tail call to target 4611: 462 DELIVER_PENDING_EXCEPTION 463.endm 464.macro INVOKE_TRAMPOLINE c_name, cxx_name 465ENTRY \c_name 466 INVOKE_TRAMPOLINE_BODY \cxx_name 467END \c_name 468.endm 469 470INVOKE_TRAMPOLINE art_quick_invoke_interface_trampoline_with_access_check, artInvokeInterfaceTrampolineWithAccessCheck 471 472INVOKE_TRAMPOLINE art_quick_invoke_static_trampoline_with_access_check, artInvokeStaticTrampolineWithAccessCheck 473INVOKE_TRAMPOLINE art_quick_invoke_direct_trampoline_with_access_check, artInvokeDirectTrampolineWithAccessCheck 474INVOKE_TRAMPOLINE art_quick_invoke_super_trampoline_with_access_check, artInvokeSuperTrampolineWithAccessCheck 475INVOKE_TRAMPOLINE art_quick_invoke_virtual_trampoline_with_access_check, artInvokeVirtualTrampolineWithAccessCheck 476 477 478.macro INVOKE_STUB_CREATE_FRAME 479 480SAVE_SIZE=15*8 // x4, x5, x19, x20, x21, x22, x23, x24, x25, x26, x27, x28, SP, LR, FP saved. 481SAVE_SIZE_AND_METHOD=SAVE_SIZE+8 482 483 484 mov x9, sp // Save stack pointer. 485 .cfi_register sp,x9 486 487 add x10, x2, # SAVE_SIZE_AND_METHOD // calculate size of frame. 488 sub x10, sp, x10 // Calculate SP position - saves + ArtMethod* + args 489 and x10, x10, # ~0xf // Enforce 16 byte stack alignment. 490 mov sp, x10 // Set new SP. 491 492 sub x10, x9, #SAVE_SIZE // Calculate new FP (later). Done here as we must move SP 493 .cfi_def_cfa_register x10 // before this. 494 .cfi_adjust_cfa_offset SAVE_SIZE 495 496 str x28, [x10, #112] 497 .cfi_rel_offset x28, 112 498 499 stp x26, x27, [x10, #96] 500 .cfi_rel_offset x26, 96 501 .cfi_rel_offset x27, 104 502 503 stp x24, x25, [x10, #80] 504 .cfi_rel_offset x24, 80 505 .cfi_rel_offset x25, 88 506 507 stp x22, x23, [x10, #64] 508 .cfi_rel_offset x22, 64 509 .cfi_rel_offset x23, 72 510 511 stp x20, x21, [x10, #48] 512 .cfi_rel_offset x20, 48 513 .cfi_rel_offset x21, 56 514 515 stp x9, x19, [x10, #32] // Save old stack pointer and x19. 516 .cfi_rel_offset sp, 32 517 .cfi_rel_offset x19, 40 518 519 stp x4, x5, [x10, #16] // Save result and shorty addresses. 520 .cfi_rel_offset x4, 16 521 .cfi_rel_offset x5, 24 522 523 stp xFP, xLR, [x10] // Store LR & FP. 524 .cfi_rel_offset x29, 0 525 .cfi_rel_offset x30, 8 526 527 mov xFP, x10 // Use xFP now, as it's callee-saved. 528 .cfi_def_cfa_register x29 529 mov xSELF, x3 // Move thread pointer into SELF register. 530 531 // Copy arguments into stack frame. 532 // Use simple copy routine for now. 533 // 4 bytes per slot. 534 // X1 - source address 535 // W2 - args length 536 // X9 - destination address. 537 // W10 - temporary 538 add x9, sp, #8 // Destination address is bottom of stack + null. 539 540 // Copy parameters into the stack. Use numeric label as this is a macro and Clang's assembler 541 // does not have unique-id variables. 5421: 543 cmp w2, #0 544 beq 2f 545 sub w2, w2, #4 // Need 65536 bytes of range. 546 ldr w10, [x1, x2] 547 str w10, [x9, x2] 548 549 b 1b 550 5512: 552 // Store null into ArtMethod* at bottom of frame. 553 str xzr, [sp] 554.endm 555 556.macro INVOKE_STUB_CALL_AND_RETURN 557 558 // load method-> METHOD_QUICK_CODE_OFFSET 559 ldr x9, [x0, #ART_METHOD_QUICK_CODE_OFFSET_64] 560 // Branch to method. 561 blr x9 562 563 // Restore return value address and shorty address. 564 ldp x4,x5, [xFP, #16] 565 .cfi_restore x4 566 .cfi_restore x5 567 568 ldr x28, [xFP, #112] 569 .cfi_restore x28 570 571 ldp x26, x27, [xFP, #96] 572 .cfi_restore x26 573 .cfi_restore x27 574 575 ldp x24, x25, [xFP, #80] 576 .cfi_restore x24 577 .cfi_restore x25 578 579 ldp x22, x23, [xFP, #64] 580 .cfi_restore x22 581 .cfi_restore x23 582 583 ldp x20, x21, [xFP, #48] 584 .cfi_restore x20 585 .cfi_restore x21 586 587 // Store result (w0/x0/s0/d0) appropriately, depending on resultType. 588 ldrb w10, [x5] 589 590 // Check the return type and store the correct register into the jvalue in memory. 591 // Use numeric label as this is a macro and Clang's assembler does not have unique-id variables. 592 593 // Don't set anything for a void type. 594 cmp w10, #'V' 595 beq 3f 596 597 // Is it a double? 598 cmp w10, #'D' 599 bne 1f 600 str d0, [x4] 601 b 3f 602 6031: // Is it a float? 604 cmp w10, #'F' 605 bne 2f 606 str s0, [x4] 607 b 3f 608 6092: // Just store x0. Doesn't matter if it is 64 or 32 bits. 610 str x0, [x4] 611 6123: // Finish up. 613 ldp x2, x19, [xFP, #32] // Restore stack pointer and x19. 614 .cfi_restore x19 615 mov sp, x2 616 .cfi_restore sp 617 618 ldp xFP, xLR, [xFP] // Restore old frame pointer and link register. 619 .cfi_restore x29 620 .cfi_restore x30 621 622 ret 623 624.endm 625 626 627/* 628 * extern"C" void art_quick_invoke_stub(ArtMethod *method, x0 629 * uint32_t *args, x1 630 * uint32_t argsize, w2 631 * Thread *self, x3 632 * JValue *result, x4 633 * char *shorty); x5 634 * +----------------------+ 635 * | | 636 * | C/C++ frame | 637 * | LR'' | 638 * | FP'' | <- SP' 639 * +----------------------+ 640 * +----------------------+ 641 * | x28 | <- TODO: Remove callee-saves. 642 * | : | 643 * | x19 | 644 * | SP' | 645 * | X5 | 646 * | X4 | Saved registers 647 * | LR' | 648 * | FP' | <- FP 649 * +----------------------+ 650 * | uint32_t out[n-1] | 651 * | : : | Outs 652 * | uint32_t out[0] | 653 * | ArtMethod* | <- SP value=null 654 * +----------------------+ 655 * 656 * Outgoing registers: 657 * x0 - Method* 658 * x1-x7 - integer parameters. 659 * d0-d7 - Floating point parameters. 660 * xSELF = self 661 * SP = & of ArtMethod* 662 * x1 = "this" pointer. 663 * 664 */ 665ENTRY art_quick_invoke_stub 666 // Spill registers as per AACPS64 calling convention. 667 INVOKE_STUB_CREATE_FRAME 668 669 // Fill registers x/w1 to x/w7 and s/d0 to s/d7 with parameters. 670 // Parse the passed shorty to determine which register to load. 671 // Load addresses for routines that load WXSD registers. 672 adr x11, .LstoreW2 673 adr x12, .LstoreX2 674 adr x13, .LstoreS0 675 adr x14, .LstoreD0 676 677 // Initialize routine offsets to 0 for integers and floats. 678 // x8 for integers, x15 for floating point. 679 mov x8, #0 680 mov x15, #0 681 682 add x10, x5, #1 // Load shorty address, plus one to skip return value. 683 ldr w1, [x9],#4 // Load "this" parameter, and increment arg pointer. 684 685 // Loop to fill registers. 686.LfillRegisters: 687 ldrb w17, [x10], #1 // Load next character in signature, and increment. 688 cbz w17, .LcallFunction // Exit at end of signature. Shorty 0 terminated. 689 690 cmp w17, #'F' // is this a float? 691 bne .LisDouble 692 693 cmp x15, # 8*12 // Skip this load if all registers full. 694 beq .Ladvance4 695 696 add x17, x13, x15 // Calculate subroutine to jump to. 697 br x17 698 699.LisDouble: 700 cmp w17, #'D' // is this a double? 701 bne .LisLong 702 703 cmp x15, # 8*12 // Skip this load if all registers full. 704 beq .Ladvance8 705 706 add x17, x14, x15 // Calculate subroutine to jump to. 707 br x17 708 709.LisLong: 710 cmp w17, #'J' // is this a long? 711 bne .LisOther 712 713 cmp x8, # 6*12 // Skip this load if all registers full. 714 beq .Ladvance8 715 716 add x17, x12, x8 // Calculate subroutine to jump to. 717 br x17 718 719.LisOther: // Everything else takes one vReg. 720 cmp x8, # 6*12 // Skip this load if all registers full. 721 beq .Ladvance4 722 723 add x17, x11, x8 // Calculate subroutine to jump to. 724 br x17 725 726.Ladvance4: 727 add x9, x9, #4 728 b .LfillRegisters 729 730.Ladvance8: 731 add x9, x9, #8 732 b .LfillRegisters 733 734// Macro for loading a parameter into a register. 735// counter - the register with offset into these tables 736// size - the size of the register - 4 or 8 bytes. 737// register - the name of the register to be loaded. 738.macro LOADREG counter size register return 739 ldr \register , [x9], #\size 740 add \counter, \counter, 12 741 b \return 742.endm 743 744// Store ints. 745.LstoreW2: 746 LOADREG x8 4 w2 .LfillRegisters 747 LOADREG x8 4 w3 .LfillRegisters 748 LOADREG x8 4 w4 .LfillRegisters 749 LOADREG x8 4 w5 .LfillRegisters 750 LOADREG x8 4 w6 .LfillRegisters 751 LOADREG x8 4 w7 .LfillRegisters 752 753// Store longs. 754.LstoreX2: 755 LOADREG x8 8 x2 .LfillRegisters 756 LOADREG x8 8 x3 .LfillRegisters 757 LOADREG x8 8 x4 .LfillRegisters 758 LOADREG x8 8 x5 .LfillRegisters 759 LOADREG x8 8 x6 .LfillRegisters 760 LOADREG x8 8 x7 .LfillRegisters 761 762// Store singles. 763.LstoreS0: 764 LOADREG x15 4 s0 .LfillRegisters 765 LOADREG x15 4 s1 .LfillRegisters 766 LOADREG x15 4 s2 .LfillRegisters 767 LOADREG x15 4 s3 .LfillRegisters 768 LOADREG x15 4 s4 .LfillRegisters 769 LOADREG x15 4 s5 .LfillRegisters 770 LOADREG x15 4 s6 .LfillRegisters 771 LOADREG x15 4 s7 .LfillRegisters 772 773// Store doubles. 774.LstoreD0: 775 LOADREG x15 8 d0 .LfillRegisters 776 LOADREG x15 8 d1 .LfillRegisters 777 LOADREG x15 8 d2 .LfillRegisters 778 LOADREG x15 8 d3 .LfillRegisters 779 LOADREG x15 8 d4 .LfillRegisters 780 LOADREG x15 8 d5 .LfillRegisters 781 LOADREG x15 8 d6 .LfillRegisters 782 LOADREG x15 8 d7 .LfillRegisters 783 784 785.LcallFunction: 786 787 INVOKE_STUB_CALL_AND_RETURN 788 789END art_quick_invoke_stub 790 791/* extern"C" 792 * void art_quick_invoke_static_stub(ArtMethod *method, x0 793 * uint32_t *args, x1 794 * uint32_t argsize, w2 795 * Thread *self, x3 796 * JValue *result, x4 797 * char *shorty); x5 798 */ 799ENTRY art_quick_invoke_static_stub 800 // Spill registers as per AACPS64 calling convention. 801 INVOKE_STUB_CREATE_FRAME 802 803 // Fill registers x/w1 to x/w7 and s/d0 to s/d7 with parameters. 804 // Parse the passed shorty to determine which register to load. 805 // Load addresses for routines that load WXSD registers. 806 adr x11, .LstoreW1_2 807 adr x12, .LstoreX1_2 808 adr x13, .LstoreS0_2 809 adr x14, .LstoreD0_2 810 811 // Initialize routine offsets to 0 for integers and floats. 812 // x8 for integers, x15 for floating point. 813 mov x8, #0 814 mov x15, #0 815 816 add x10, x5, #1 // Load shorty address, plus one to skip return value. 817 818 // Loop to fill registers. 819.LfillRegisters2: 820 ldrb w17, [x10], #1 // Load next character in signature, and increment. 821 cbz w17, .LcallFunction2 // Exit at end of signature. Shorty 0 terminated. 822 823 cmp w17, #'F' // is this a float? 824 bne .LisDouble2 825 826 cmp x15, # 8*12 // Skip this load if all registers full. 827 beq .Ladvance4_2 828 829 add x17, x13, x15 // Calculate subroutine to jump to. 830 br x17 831 832.LisDouble2: 833 cmp w17, #'D' // is this a double? 834 bne .LisLong2 835 836 cmp x15, # 8*12 // Skip this load if all registers full. 837 beq .Ladvance8_2 838 839 add x17, x14, x15 // Calculate subroutine to jump to. 840 br x17 841 842.LisLong2: 843 cmp w17, #'J' // is this a long? 844 bne .LisOther2 845 846 cmp x8, # 7*12 // Skip this load if all registers full. 847 beq .Ladvance8_2 848 849 add x17, x12, x8 // Calculate subroutine to jump to. 850 br x17 851 852.LisOther2: // Everything else takes one vReg. 853 cmp x8, # 7*12 // Skip this load if all registers full. 854 beq .Ladvance4_2 855 856 add x17, x11, x8 // Calculate subroutine to jump to. 857 br x17 858 859.Ladvance4_2: 860 add x9, x9, #4 861 b .LfillRegisters2 862 863.Ladvance8_2: 864 add x9, x9, #8 865 b .LfillRegisters2 866 867// Store ints. 868.LstoreW1_2: 869 LOADREG x8 4 w1 .LfillRegisters2 870 LOADREG x8 4 w2 .LfillRegisters2 871 LOADREG x8 4 w3 .LfillRegisters2 872 LOADREG x8 4 w4 .LfillRegisters2 873 LOADREG x8 4 w5 .LfillRegisters2 874 LOADREG x8 4 w6 .LfillRegisters2 875 LOADREG x8 4 w7 .LfillRegisters2 876 877// Store longs. 878.LstoreX1_2: 879 LOADREG x8 8 x1 .LfillRegisters2 880 LOADREG x8 8 x2 .LfillRegisters2 881 LOADREG x8 8 x3 .LfillRegisters2 882 LOADREG x8 8 x4 .LfillRegisters2 883 LOADREG x8 8 x5 .LfillRegisters2 884 LOADREG x8 8 x6 .LfillRegisters2 885 LOADREG x8 8 x7 .LfillRegisters2 886 887// Store singles. 888.LstoreS0_2: 889 LOADREG x15 4 s0 .LfillRegisters2 890 LOADREG x15 4 s1 .LfillRegisters2 891 LOADREG x15 4 s2 .LfillRegisters2 892 LOADREG x15 4 s3 .LfillRegisters2 893 LOADREG x15 4 s4 .LfillRegisters2 894 LOADREG x15 4 s5 .LfillRegisters2 895 LOADREG x15 4 s6 .LfillRegisters2 896 LOADREG x15 4 s7 .LfillRegisters2 897 898// Store doubles. 899.LstoreD0_2: 900 LOADREG x15 8 d0 .LfillRegisters2 901 LOADREG x15 8 d1 .LfillRegisters2 902 LOADREG x15 8 d2 .LfillRegisters2 903 LOADREG x15 8 d3 .LfillRegisters2 904 LOADREG x15 8 d4 .LfillRegisters2 905 LOADREG x15 8 d5 .LfillRegisters2 906 LOADREG x15 8 d6 .LfillRegisters2 907 LOADREG x15 8 d7 .LfillRegisters2 908 909 910.LcallFunction2: 911 912 INVOKE_STUB_CALL_AND_RETURN 913 914END art_quick_invoke_static_stub 915 916 917 918/* extern"C" void art_quick_osr_stub(void** stack, x0 919 * size_t stack_size_in_bytes, x1 920 * const uin8_t* native_pc, x2 921 * JValue *result, x3 922 * char *shorty, x4 923 * Thread *self) x5 924 */ 925ENTRY art_quick_osr_stub 926SAVE_SIZE=15*8 // x3, x4, x19, x20, x21, x22, x23, x24, x25, x26, x27, x28, SP, LR, FP saved. 927 mov x9, sp // Save stack pointer. 928 .cfi_register sp,x9 929 930 sub x10, sp, # SAVE_SIZE 931 and x10, x10, # ~0xf // Enforce 16 byte stack alignment. 932 mov sp, x10 // Set new SP. 933 934 str x28, [sp, #112] 935 stp x26, x27, [sp, #96] 936 stp x24, x25, [sp, #80] 937 stp x22, x23, [sp, #64] 938 stp x20, x21, [sp, #48] 939 stp x9, x19, [sp, #32] // Save old stack pointer and x19. 940 stp x3, x4, [sp, #16] // Save result and shorty addresses. 941 stp xFP, xLR, [sp] // Store LR & FP. 942 mov xSELF, x5 // Move thread pointer into SELF register. 943 944 sub sp, sp, #16 945 str xzr, [sp] // Store null for ArtMethod* slot 946 // Branch to stub. 947 bl .Losr_entry 948 add sp, sp, #16 949 950 // Restore return value address and shorty address. 951 ldp x3,x4, [sp, #16] 952 ldr x28, [sp, #112] 953 ldp x26, x27, [sp, #96] 954 ldp x24, x25, [sp, #80] 955 ldp x22, x23, [sp, #64] 956 ldp x20, x21, [sp, #48] 957 958 // Store result (w0/x0/s0/d0) appropriately, depending on resultType. 959 ldrb w10, [x4] 960 961 // Check the return type and store the correct register into the jvalue in memory. 962 963 // Don't set anything for a void type. 964 cmp w10, #'V' 965 beq .Losr_exit 966 967 // Is it a double? 968 cmp w10, #'D' 969 bne .Lno_double 970 str d0, [x3] 971 b .Losr_exit 972 973.Lno_double: // Is it a float? 974 cmp w10, #'F' 975 bne .Lno_float 976 str s0, [x3] 977 b .Losr_exit 978 979.Lno_float: // Just store x0. Doesn't matter if it is 64 or 32 bits. 980 str x0, [x3] 981 982.Losr_exit: // Finish up. 983 ldp x2, x19, [sp, #32] // Restore stack pointer and x19. 984 ldp xFP, xLR, [sp] // Restore old frame pointer and link register. 985 mov sp, x2 986 ret 987 988.Losr_entry: 989 // Update stack pointer for the callee 990 sub sp, sp, x1 991 992 // Update link register slot expected by the callee. 993 sub w1, w1, #8 994 str lr, [sp, x1] 995 996 // Copy arguments into stack frame. 997 // Use simple copy routine for now. 998 // 4 bytes per slot. 999 // X0 - source address 1000 // W1 - args length 1001 // SP - destination address. 1002 // W10 - temporary 1003.Losr_loop_entry: 1004 cmp w1, #0 1005 beq .Losr_loop_exit 1006 sub w1, w1, #4 1007 ldr w10, [x0, x1] 1008 str w10, [sp, x1] 1009 b .Losr_loop_entry 1010 1011.Losr_loop_exit: 1012 // Branch to the OSR entry point. 1013 br x2 1014 1015END art_quick_osr_stub 1016 1017 /* 1018 * On entry x0 is uintptr_t* gprs_ and x1 is uint64_t* fprs_ 1019 */ 1020 1021ENTRY art_quick_do_long_jump 1022 // Load FPRs 1023 ldp d0, d1, [x1], #16 1024 ldp d2, d3, [x1], #16 1025 ldp d4, d5, [x1], #16 1026 ldp d6, d7, [x1], #16 1027 ldp d8, d9, [x1], #16 1028 ldp d10, d11, [x1], #16 1029 ldp d12, d13, [x1], #16 1030 ldp d14, d15, [x1], #16 1031 ldp d16, d17, [x1], #16 1032 ldp d18, d19, [x1], #16 1033 ldp d20, d21, [x1], #16 1034 ldp d22, d23, [x1], #16 1035 ldp d24, d25, [x1], #16 1036 ldp d26, d27, [x1], #16 1037 ldp d28, d29, [x1], #16 1038 ldp d30, d31, [x1] 1039 1040 // Load GPRs 1041 // TODO: lots of those are smashed, could optimize. 1042 add x0, x0, #30*8 1043 ldp x30, x1, [x0], #-16 // LR & SP 1044 ldp x28, x29, [x0], #-16 1045 ldp x26, x27, [x0], #-16 1046 ldp x24, x25, [x0], #-16 1047 ldp x22, x23, [x0], #-16 1048 ldp x20, x21, [x0], #-16 1049 ldp x18, x19, [x0], #-16 1050 ldp x16, x17, [x0], #-16 1051 ldp x14, x15, [x0], #-16 1052 ldp x12, x13, [x0], #-16 1053 ldp x10, x11, [x0], #-16 1054 ldp x8, x9, [x0], #-16 1055 ldp x6, x7, [x0], #-16 1056 ldp x4, x5, [x0], #-16 1057 ldp x2, x3, [x0], #-16 1058 mov sp, x1 1059 1060 // Need to load PC, it's at the end (after the space for the unused XZR). Use x1. 1061 ldr x1, [x0, #33*8] 1062 // And the value of x0. 1063 ldr x0, [x0] 1064 1065 br x1 1066END art_quick_do_long_jump 1067 1068 /* 1069 * Entry from managed code that calls artLockObjectFromCode, may block for GC. x0 holds the 1070 * possibly null object to lock. 1071 * 1072 * Derived from arm32 code. 1073 */ 1074 .extern artLockObjectFromCode 1075ENTRY art_quick_lock_object 1076 cbz w0, .Lslow_lock 1077 add x4, x0, #MIRROR_OBJECT_LOCK_WORD_OFFSET // exclusive load/store has no immediate anymore 1078.Lretry_lock: 1079 ldr w2, [xSELF, #THREAD_ID_OFFSET] // TODO: Can the thread ID really change during the loop? 1080 ldxr w1, [x4] 1081 mov x3, x1 1082 and w3, w3, #LOCK_WORD_READ_BARRIER_STATE_MASK_TOGGLED // zero the read barrier bits 1083 cbnz w3, .Lnot_unlocked // already thin locked 1084 // unlocked case - x1: original lock word that's zero except for the read barrier bits. 1085 orr x2, x1, x2 // x2 holds thread id with count of 0 with preserved read barrier bits 1086 stxr w3, w2, [x4] 1087 cbnz w3, .Llock_stxr_fail // store failed, retry 1088 dmb ishld // full (LoadLoad|LoadStore) memory barrier 1089 ret 1090.Lnot_unlocked: // x1: original lock word 1091 lsr w3, w1, LOCK_WORD_STATE_SHIFT 1092 cbnz w3, .Lslow_lock // if either of the top two bits are set, go slow path 1093 eor w2, w1, w2 // lock_word.ThreadId() ^ self->ThreadId() 1094 uxth w2, w2 // zero top 16 bits 1095 cbnz w2, .Lslow_lock // lock word and self thread id's match -> recursive lock 1096 // else contention, go to slow path 1097 mov x3, x1 // copy the lock word to check count overflow. 1098 and w3, w3, #LOCK_WORD_READ_BARRIER_STATE_MASK_TOGGLED // zero the read barrier bits. 1099 add w2, w3, #LOCK_WORD_THIN_LOCK_COUNT_ONE // increment count in lock word placing in w2 to check overflow 1100 lsr w3, w2, LOCK_WORD_READ_BARRIER_STATE_SHIFT // if either of the upper two bits (28-29) are set, we overflowed. 1101 cbnz w3, .Lslow_lock // if we overflow the count go slow path 1102 add w2, w1, #LOCK_WORD_THIN_LOCK_COUNT_ONE // increment count for real 1103 stxr w3, w2, [x4] 1104 cbnz w3, .Llock_stxr_fail // store failed, retry 1105 ret 1106.Llock_stxr_fail: 1107 b .Lretry_lock // retry 1108.Lslow_lock: 1109 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // save callee saves in case we block 1110 mov x1, xSELF // pass Thread::Current 1111 bl artLockObjectFromCode // (Object* obj, Thread*) 1112 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME 1113 RETURN_IF_W0_IS_ZERO_OR_DELIVER 1114END art_quick_lock_object 1115 1116 /* 1117 * Entry from managed code that calls artUnlockObjectFromCode and delivers exception on failure. 1118 * x0 holds the possibly null object to lock. 1119 * 1120 * Derived from arm32 code. 1121 */ 1122 .extern artUnlockObjectFromCode 1123ENTRY art_quick_unlock_object 1124 cbz x0, .Lslow_unlock 1125 add x4, x0, #MIRROR_OBJECT_LOCK_WORD_OFFSET // exclusive load/store has no immediate anymore 1126.Lretry_unlock: 1127#ifndef USE_READ_BARRIER 1128 ldr w1, [x4] 1129#else 1130 ldxr w1, [x4] // Need to use atomic instructions for read barrier 1131#endif 1132 lsr w2, w1, LOCK_WORD_STATE_SHIFT 1133 cbnz w2, .Lslow_unlock // if either of the top two bits are set, go slow path 1134 ldr w2, [xSELF, #THREAD_ID_OFFSET] 1135 mov x3, x1 // copy lock word to check thread id equality 1136 and w3, w3, #LOCK_WORD_READ_BARRIER_STATE_MASK_TOGGLED // zero the read barrier bits 1137 eor w3, w3, w2 // lock_word.ThreadId() ^ self->ThreadId() 1138 uxth w3, w3 // zero top 16 bits 1139 cbnz w3, .Lslow_unlock // do lock word and self thread id's match? 1140 mov x3, x1 // copy lock word to detect transition to unlocked 1141 and w3, w3, #LOCK_WORD_READ_BARRIER_STATE_MASK_TOGGLED // zero the read barrier bits 1142 cmp w3, #LOCK_WORD_THIN_LOCK_COUNT_ONE 1143 bpl .Lrecursive_thin_unlock 1144 // transition to unlocked 1145 mov x3, x1 1146 and w3, w3, #LOCK_WORD_READ_BARRIER_STATE_MASK // w3: zero except for the preserved read barrier bits 1147 dmb ish // full (LoadStore|StoreStore) memory barrier 1148#ifndef USE_READ_BARRIER 1149 str w3, [x4] 1150#else 1151 stxr w2, w3, [x4] // Need to use atomic instructions for read barrier 1152 cbnz w2, .Lunlock_stxr_fail // store failed, retry 1153#endif 1154 ret 1155.Lrecursive_thin_unlock: // w1: original lock word 1156 sub w1, w1, #LOCK_WORD_THIN_LOCK_COUNT_ONE // decrement count 1157#ifndef USE_READ_BARRIER 1158 str w1, [x4] 1159#else 1160 stxr w2, w1, [x4] // Need to use atomic instructions for read barrier 1161 cbnz w2, .Lunlock_stxr_fail // store failed, retry 1162#endif 1163 ret 1164.Lunlock_stxr_fail: 1165 b .Lretry_unlock // retry 1166.Lslow_unlock: 1167 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // save callee saves in case exception allocation triggers GC 1168 mov x1, xSELF // pass Thread::Current 1169 bl artUnlockObjectFromCode // (Object* obj, Thread*) 1170 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME 1171 RETURN_IF_W0_IS_ZERO_OR_DELIVER 1172END art_quick_unlock_object 1173 1174 /* 1175 * Entry from managed code that calls artIsAssignableFromCode and on failure calls 1176 * artThrowClassCastException. 1177 */ 1178 .extern artThrowClassCastException 1179ENTRY art_quick_check_cast 1180 // Store arguments and link register 1181 // Stack needs to be 16B aligned on calls. 1182 stp x0, x1, [sp,#-32]! 1183 .cfi_adjust_cfa_offset 32 1184 .cfi_rel_offset x0, 0 1185 .cfi_rel_offset x1, 8 1186 str xLR, [sp, #24] 1187 .cfi_rel_offset x30, 24 1188 1189 // Call runtime code 1190 bl artIsAssignableFromCode 1191 1192 // Check for exception 1193 cbz x0, .Lthrow_class_cast_exception 1194 1195 // Restore and return 1196 ldr xLR, [sp, #24] 1197 .cfi_restore x30 1198 ldp x0, x1, [sp], #32 1199 .cfi_restore x0 1200 .cfi_restore x1 1201 .cfi_adjust_cfa_offset -32 1202 ret 1203 1204 .cfi_adjust_cfa_offset 32 // Reset unwind info so following code unwinds. 1205 1206.Lthrow_class_cast_exception: 1207 // Restore 1208 ldr xLR, [sp, #24] 1209 .cfi_restore x30 1210 ldp x0, x1, [sp], #32 1211 .cfi_restore x0 1212 .cfi_restore x1 1213 .cfi_adjust_cfa_offset -32 1214 1215 SETUP_SAVE_ALL_CALLEE_SAVE_FRAME // save all registers as basis for long jump context 1216 mov x2, xSELF // pass Thread::Current 1217 b artThrowClassCastException // (Class*, Class*, Thread*) 1218 brk 0 // We should not return here... 1219END art_quick_check_cast 1220 1221// Restore xReg's value from [sp, #offset] if xReg is not the same as xExclude. 1222.macro POP_REG_NE xReg, offset, xExclude 1223 .ifnc \xReg, \xExclude 1224 ldr \xReg, [sp, #\offset] // restore xReg 1225 .cfi_restore \xReg 1226 .endif 1227.endm 1228 1229 /* 1230 * Macro to insert read barrier, only used in art_quick_aput_obj. 1231 * xDest, wDest and xObj are registers, offset is a defined literal such as 1232 * MIRROR_OBJECT_CLASS_OFFSET. Dest needs both x and w versions of the same register to handle 1233 * name mismatch between instructions. This macro uses the lower 32b of register when possible. 1234 * TODO: When read barrier has a fast path, add heap unpoisoning support for the fast path. 1235 */ 1236.macro READ_BARRIER xDest, wDest, xObj, offset 1237#ifdef USE_READ_BARRIER 1238 // Store registers used in art_quick_aput_obj (x0-x4, LR), stack is 16B aligned. 1239 stp x0, x1, [sp, #-48]! 1240 .cfi_adjust_cfa_offset 48 1241 .cfi_rel_offset x0, 0 1242 .cfi_rel_offset x1, 8 1243 stp x2, x3, [sp, #16] 1244 .cfi_rel_offset x2, 16 1245 .cfi_rel_offset x3, 24 1246 stp x4, xLR, [sp, #32] 1247 .cfi_rel_offset x4, 32 1248 .cfi_rel_offset x30, 40 1249 1250 // mov x0, \xRef // pass ref in x0 (no-op for now since parameter ref is unused) 1251 .ifnc \xObj, x1 1252 mov x1, \xObj // pass xObj 1253 .endif 1254 mov w2, #\offset // pass offset 1255 bl artReadBarrierSlow // artReadBarrierSlow(ref, xObj, offset) 1256 // No need to unpoison return value in w0, artReadBarrierSlow() would do the unpoisoning. 1257 .ifnc \wDest, w0 1258 mov \wDest, w0 // save return value in wDest 1259 .endif 1260 1261 // Conditionally restore saved registers 1262 POP_REG_NE x0, 0, \xDest 1263 POP_REG_NE x1, 8, \xDest 1264 POP_REG_NE x2, 16, \xDest 1265 POP_REG_NE x3, 24, \xDest 1266 POP_REG_NE x4, 32, \xDest 1267 ldr xLR, [sp, #40] 1268 .cfi_restore x30 1269 add sp, sp, #48 1270 .cfi_adjust_cfa_offset -48 1271#else 1272 ldr \wDest, [\xObj, #\offset] // Heap reference = 32b. This also zero-extends to \xDest. 1273 UNPOISON_HEAP_REF \wDest 1274#endif // USE_READ_BARRIER 1275.endm 1276 1277 /* 1278 * Entry from managed code for array put operations of objects where the value being stored 1279 * needs to be checked for compatibility. 1280 * x0 = array, x1 = index, x2 = value 1281 * 1282 * Currently all values should fit into w0/w1/w2, and w1 always will as indices are 32b. We 1283 * assume, though, that the upper 32b are zeroed out. At least for x1/w1 we can do better by 1284 * using index-zero-extension in load/stores. 1285 * 1286 * Temporaries: x3, x4 1287 * TODO: x4 OK? ip seems wrong here. 1288 */ 1289ENTRY art_quick_aput_obj_with_null_and_bound_check 1290 tst x0, x0 1291 bne art_quick_aput_obj_with_bound_check 1292 b art_quick_throw_null_pointer_exception 1293END art_quick_aput_obj_with_null_and_bound_check 1294 1295ENTRY art_quick_aput_obj_with_bound_check 1296 ldr w3, [x0, #MIRROR_ARRAY_LENGTH_OFFSET] 1297 cmp w3, w1 1298 bhi art_quick_aput_obj 1299 mov x0, x1 1300 mov x1, x3 1301 b art_quick_throw_array_bounds 1302END art_quick_aput_obj_with_bound_check 1303 1304#ifdef USE_READ_BARRIER 1305 .extern artReadBarrierSlow 1306#endif 1307ENTRY art_quick_aput_obj 1308 cbz x2, .Ldo_aput_null 1309 READ_BARRIER x3, w3, x0, MIRROR_OBJECT_CLASS_OFFSET // Heap reference = 32b 1310 // This also zero-extends to x3 1311 READ_BARRIER x4, w4, x2, MIRROR_OBJECT_CLASS_OFFSET // Heap reference = 32b 1312 // This also zero-extends to x4 1313 READ_BARRIER x3, w3, x3, MIRROR_CLASS_COMPONENT_TYPE_OFFSET // Heap reference = 32b 1314 // This also zero-extends to x3 1315 cmp w3, w4 // value's type == array's component type - trivial assignability 1316 bne .Lcheck_assignability 1317.Ldo_aput: 1318 add x3, x0, #MIRROR_OBJECT_ARRAY_DATA_OFFSET 1319 // "Compress" = do nothing 1320 POISON_HEAP_REF w2 1321 str w2, [x3, x1, lsl #2] // Heap reference = 32b 1322 ldr x3, [xSELF, #THREAD_CARD_TABLE_OFFSET] 1323 lsr x0, x0, #7 1324 strb w3, [x3, x0] 1325 ret 1326.Ldo_aput_null: 1327 add x3, x0, #MIRROR_OBJECT_ARRAY_DATA_OFFSET 1328 // "Compress" = do nothing 1329 str w2, [x3, x1, lsl #2] // Heap reference = 32b 1330 ret 1331.Lcheck_assignability: 1332 // Store arguments and link register 1333 stp x0, x1, [sp,#-32]! 1334 .cfi_adjust_cfa_offset 32 1335 .cfi_rel_offset x0, 0 1336 .cfi_rel_offset x1, 8 1337 stp x2, xLR, [sp, #16] 1338 .cfi_rel_offset x2, 16 1339 .cfi_rel_offset x30, 24 1340 1341 // Call runtime code 1342 mov x0, x3 // Heap reference, 32b, "uncompress" = do nothing, already zero-extended 1343 mov x1, x4 // Heap reference, 32b, "uncompress" = do nothing, already zero-extended 1344 bl artIsAssignableFromCode 1345 1346 // Check for exception 1347 cbz x0, .Lthrow_array_store_exception 1348 1349 // Restore 1350 ldp x2, x30, [sp, #16] 1351 .cfi_restore x2 1352 .cfi_restore x30 1353 ldp x0, x1, [sp], #32 1354 .cfi_restore x0 1355 .cfi_restore x1 1356 .cfi_adjust_cfa_offset -32 1357 1358 add x3, x0, #MIRROR_OBJECT_ARRAY_DATA_OFFSET 1359 // "Compress" = do nothing 1360 POISON_HEAP_REF w2 1361 str w2, [x3, x1, lsl #2] // Heap reference = 32b 1362 ldr x3, [xSELF, #THREAD_CARD_TABLE_OFFSET] 1363 lsr x0, x0, #7 1364 strb w3, [x3, x0] 1365 ret 1366 .cfi_adjust_cfa_offset 32 // 4 restores after cbz for unwinding. 1367.Lthrow_array_store_exception: 1368 ldp x2, x30, [sp, #16] 1369 .cfi_restore x2 1370 .cfi_restore x30 1371 ldp x0, x1, [sp], #32 1372 .cfi_restore x0 1373 .cfi_restore x1 1374 .cfi_adjust_cfa_offset -32 1375 1376 SETUP_SAVE_ALL_CALLEE_SAVE_FRAME 1377 mov x1, x2 // Pass value. 1378 mov x2, xSELF // Pass Thread::Current. 1379 b artThrowArrayStoreException // (Object*, Object*, Thread*). 1380 brk 0 // Unreached. 1381END art_quick_aput_obj 1382 1383// Macro to facilitate adding new allocation entrypoints. 1384.macro ONE_ARG_DOWNCALL name, entrypoint, return 1385 .extern \entrypoint 1386ENTRY \name 1387 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // save callee saves in case of GC 1388 mov x1, xSELF // pass Thread::Current 1389 bl \entrypoint // (uint32_t type_idx, Method* method, Thread*) 1390 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME 1391 \return 1392END \name 1393.endm 1394 1395// Macro to facilitate adding new allocation entrypoints. 1396.macro TWO_ARG_DOWNCALL name, entrypoint, return 1397 .extern \entrypoint 1398ENTRY \name 1399 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // save callee saves in case of GC 1400 mov x2, xSELF // pass Thread::Current 1401 bl \entrypoint // (uint32_t type_idx, Method* method, Thread*) 1402 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME 1403 \return 1404END \name 1405.endm 1406 1407// Macro to facilitate adding new allocation entrypoints. 1408.macro THREE_ARG_DOWNCALL name, entrypoint, return 1409 .extern \entrypoint 1410ENTRY \name 1411 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // save callee saves in case of GC 1412 mov x3, xSELF // pass Thread::Current 1413 bl \entrypoint 1414 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME 1415 \return 1416END \name 1417.endm 1418 1419// Macro to facilitate adding new allocation entrypoints. 1420.macro FOUR_ARG_DOWNCALL name, entrypoint, return 1421 .extern \entrypoint 1422ENTRY \name 1423 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // save callee saves in case of GC 1424 mov x4, xSELF // pass Thread::Current 1425 bl \entrypoint // 1426 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME 1427 \return 1428 DELIVER_PENDING_EXCEPTION 1429END \name 1430.endm 1431 1432// Macros taking opportunity of code similarities for downcalls with referrer. 1433.macro ONE_ARG_REF_DOWNCALL name, entrypoint, return 1434 .extern \entrypoint 1435ENTRY \name 1436 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // save callee saves in case of GC 1437 ldr x1, [sp, #FRAME_SIZE_REFS_ONLY_CALLEE_SAVE] // Load referrer 1438 mov x2, xSELF // pass Thread::Current 1439 bl \entrypoint // (uint32_t type_idx, Method* method, Thread*, SP) 1440 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME 1441 \return 1442END \name 1443.endm 1444 1445.macro TWO_ARG_REF_DOWNCALL name, entrypoint, return 1446 .extern \entrypoint 1447ENTRY \name 1448 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // save callee saves in case of GC 1449 ldr x2, [sp, #FRAME_SIZE_REFS_ONLY_CALLEE_SAVE] // Load referrer 1450 mov x3, xSELF // pass Thread::Current 1451 bl \entrypoint 1452 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME 1453 \return 1454END \name 1455.endm 1456 1457.macro THREE_ARG_REF_DOWNCALL name, entrypoint, return 1458 .extern \entrypoint 1459ENTRY \name 1460 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // save callee saves in case of GC 1461 ldr x3, [sp, #FRAME_SIZE_REFS_ONLY_CALLEE_SAVE] // Load referrer 1462 mov x4, xSELF // pass Thread::Current 1463 bl \entrypoint 1464 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME 1465 \return 1466END \name 1467.endm 1468 1469.macro RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1470 cbz w0, 1f // result zero branch over 1471 ret // return 14721: 1473 DELIVER_PENDING_EXCEPTION 1474.endm 1475 1476 /* 1477 * Entry from managed code that calls artHandleFillArrayDataFromCode and delivers exception on 1478 * failure. 1479 */ 1480TWO_ARG_REF_DOWNCALL art_quick_handle_fill_data, artHandleFillArrayDataFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1481 1482 /* 1483 * Entry from managed code when uninitialized static storage, this stub will run the class 1484 * initializer and deliver the exception on error. On success the static storage base is 1485 * returned. 1486 */ 1487ONE_ARG_DOWNCALL art_quick_initialize_static_storage, artInitializeStaticStorageFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1488 1489ONE_ARG_DOWNCALL art_quick_initialize_type, artInitializeTypeFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1490ONE_ARG_DOWNCALL art_quick_initialize_type_and_verify_access, artInitializeTypeAndVerifyAccessFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1491 1492ONE_ARG_REF_DOWNCALL art_quick_get_boolean_static, artGetBooleanStaticFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1493ONE_ARG_REF_DOWNCALL art_quick_get_byte_static, artGetByteStaticFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1494ONE_ARG_REF_DOWNCALL art_quick_get_char_static, artGetCharStaticFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1495ONE_ARG_REF_DOWNCALL art_quick_get_short_static, artGetShortStaticFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1496ONE_ARG_REF_DOWNCALL art_quick_get32_static, artGet32StaticFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1497ONE_ARG_REF_DOWNCALL art_quick_get64_static, artGet64StaticFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1498ONE_ARG_REF_DOWNCALL art_quick_get_obj_static, artGetObjStaticFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1499 1500TWO_ARG_REF_DOWNCALL art_quick_get_boolean_instance, artGetBooleanInstanceFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1501TWO_ARG_REF_DOWNCALL art_quick_get_byte_instance, artGetByteInstanceFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1502TWO_ARG_REF_DOWNCALL art_quick_get_char_instance, artGetCharInstanceFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1503TWO_ARG_REF_DOWNCALL art_quick_get_short_instance, artGetShortInstanceFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1504TWO_ARG_REF_DOWNCALL art_quick_get32_instance, artGet32InstanceFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1505TWO_ARG_REF_DOWNCALL art_quick_get64_instance, artGet64InstanceFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1506TWO_ARG_REF_DOWNCALL art_quick_get_obj_instance, artGetObjInstanceFromCode, RETURN_OR_DELIVER_PENDING_EXCEPTION_X1 1507 1508TWO_ARG_REF_DOWNCALL art_quick_set8_static, artSet8StaticFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1509TWO_ARG_REF_DOWNCALL art_quick_set16_static, artSet16StaticFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1510TWO_ARG_REF_DOWNCALL art_quick_set32_static, artSet32StaticFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1511TWO_ARG_REF_DOWNCALL art_quick_set_obj_static, artSetObjStaticFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1512 1513THREE_ARG_REF_DOWNCALL art_quick_set8_instance, artSet8InstanceFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1514THREE_ARG_REF_DOWNCALL art_quick_set16_instance, artSet16InstanceFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1515THREE_ARG_REF_DOWNCALL art_quick_set32_instance, artSet32InstanceFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1516THREE_ARG_REF_DOWNCALL art_quick_set64_instance, artSet64InstanceFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1517THREE_ARG_REF_DOWNCALL art_quick_set_obj_instance, artSetObjInstanceFromCode, RETURN_IF_W0_IS_ZERO_OR_DELIVER 1518 1519// This is separated out as the argument order is different. 1520 .extern artSet64StaticFromCode 1521ENTRY art_quick_set64_static 1522 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // save callee saves in case of GC 1523 ldr x1, [sp, #FRAME_SIZE_REFS_ONLY_CALLEE_SAVE] // Load referrer 1524 // x2 contains the parameter 1525 mov x3, xSELF // pass Thread::Current 1526 bl artSet64StaticFromCode 1527 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME 1528 RETURN_IF_W0_IS_ZERO_OR_DELIVER 1529END art_quick_set64_static 1530 1531 /* 1532 * Entry from managed code to resolve a string, this stub will allocate a String and deliver an 1533 * exception on error. On success the String is returned. w0 holds the string index. The fast 1534 * path check for hit in strings cache has already been performed. 1535 */ 1536ONE_ARG_DOWNCALL art_quick_resolve_string, artResolveStringFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1537 1538// Generate the allocation entrypoints for each allocator. 1539GENERATE_ALLOC_ENTRYPOINTS_FOR_EACH_ALLOCATOR 1540 1541// A hand-written override for GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT(_rosalloc, RosAlloc). 1542ENTRY art_quick_alloc_object_rosalloc 1543 // Fast path rosalloc allocation. 1544 // x0: type_idx/return value, x1: ArtMethod*, xSELF(x19): Thread::Current 1545 // x2-x7: free. 1546 ldr x2, [x1, #ART_METHOD_DEX_CACHE_TYPES_OFFSET_64] // Load dex cache resolved types array 1547 // Load the class (x2) 1548 ldr w2, [x2, x0, lsl #COMPRESSED_REFERENCE_SIZE_SHIFT] 1549 cbz x2, .Lart_quick_alloc_object_rosalloc_slow_path // Check null class 1550 // Check class status. 1551 ldr w3, [x2, #MIRROR_CLASS_STATUS_OFFSET] 1552 cmp x3, #MIRROR_CLASS_STATUS_INITIALIZED 1553 bne .Lart_quick_alloc_object_rosalloc_slow_path 1554 // Add a fake dependence from the 1555 // following access flag and size 1556 // loads to the status load. 1557 // This is to prevent those loads 1558 // from being reordered above the 1559 // status load and reading wrong 1560 // values (an alternative is to use 1561 // a load-acquire for the status). 1562 eor x3, x3, x3 1563 add x2, x2, x3 1564 // Check access flags has 1565 // kAccClassIsFinalizable 1566 ldr w3, [x2, #MIRROR_CLASS_ACCESS_FLAGS_OFFSET] 1567 tst x3, #ACCESS_FLAGS_CLASS_IS_FINALIZABLE 1568 bne .Lart_quick_alloc_object_rosalloc_slow_path 1569 ldr x3, [xSELF, #THREAD_LOCAL_ALLOC_STACK_TOP_OFFSET] // Check if the thread local 1570 // allocation stack has room. 1571 // ldp won't work due to large offset. 1572 ldr x4, [xSELF, #THREAD_LOCAL_ALLOC_STACK_END_OFFSET] 1573 cmp x3, x4 1574 bhs .Lart_quick_alloc_object_rosalloc_slow_path 1575 ldr w3, [x2, #MIRROR_CLASS_OBJECT_SIZE_OFFSET] // Load the object size (x3) 1576 cmp x3, #ROSALLOC_MAX_THREAD_LOCAL_BRACKET_SIZE // Check if the size is for a thread 1577 // local allocation 1578 bhs .Lart_quick_alloc_object_rosalloc_slow_path 1579 // Compute the rosalloc bracket index 1580 // from the size. 1581 // Align up the size by the rosalloc 1582 // bracket quantum size and divide 1583 // by the quantum size and subtract 1584 // by 1. This code is a shorter but 1585 // equivalent version. 1586 sub x3, x3, #1 1587 lsr x3, x3, #ROSALLOC_BRACKET_QUANTUM_SIZE_SHIFT 1588 // Load the rosalloc run (x4) 1589 add x4, xSELF, x3, lsl #POINTER_SIZE_SHIFT 1590 ldr x4, [x4, #THREAD_ROSALLOC_RUNS_OFFSET] 1591 // Load the free list head (x3). This 1592 // will be the return val. 1593 ldr x3, [x4, #(ROSALLOC_RUN_FREE_LIST_OFFSET + ROSALLOC_RUN_FREE_LIST_HEAD_OFFSET)] 1594 cbz x3, .Lart_quick_alloc_object_rosalloc_slow_path 1595 // "Point of no slow path". Won't go to the slow path from here on. OK to clobber x0 and x1. 1596 ldr x1, [x3, #ROSALLOC_SLOT_NEXT_OFFSET] // Load the next pointer of the head 1597 // and update the list head with the 1598 // next pointer. 1599 str x1, [x4, #(ROSALLOC_RUN_FREE_LIST_OFFSET + ROSALLOC_RUN_FREE_LIST_HEAD_OFFSET)] 1600 // Store the class pointer in the 1601 // header. This also overwrites the 1602 // next pointer. The offsets are 1603 // asserted to match. 1604#if ROSALLOC_SLOT_NEXT_OFFSET != MIRROR_OBJECT_CLASS_OFFSET 1605#error "Class pointer needs to overwrite next pointer." 1606#endif 1607 POISON_HEAP_REF w2 1608 str w2, [x3, #MIRROR_OBJECT_CLASS_OFFSET] 1609 // Push the new object onto the thread 1610 // local allocation stack and 1611 // increment the thread local 1612 // allocation stack top. 1613 ldr x1, [xSELF, #THREAD_LOCAL_ALLOC_STACK_TOP_OFFSET] 1614 str w3, [x1], #COMPRESSED_REFERENCE_SIZE // (Increment x1 as a side effect.) 1615 str x1, [xSELF, #THREAD_LOCAL_ALLOC_STACK_TOP_OFFSET] 1616 // Decrement the size of the free list 1617 ldr w1, [x4, #(ROSALLOC_RUN_FREE_LIST_OFFSET + ROSALLOC_RUN_FREE_LIST_SIZE_OFFSET)] 1618 sub x1, x1, #1 1619 // TODO: consider combining this store 1620 // and the list head store above using 1621 // strd. 1622 str w1, [x4, #(ROSALLOC_RUN_FREE_LIST_OFFSET + ROSALLOC_RUN_FREE_LIST_SIZE_OFFSET)] 1623 // Fence. This is "ish" not "ishst" so 1624 // that the code after this allocation 1625 // site will see the right values in 1626 // the fields of the class. 1627 // Alternatively we could use "ishst" 1628 // if we use load-acquire for the 1629 // class status load.) 1630 dmb ish 1631 mov x0, x3 // Set the return value and return. 1632 ret 1633.Lart_quick_alloc_object_rosalloc_slow_path: 1634 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // save callee saves in case of GC 1635 mov x2, xSELF // pass Thread::Current 1636 bl artAllocObjectFromCodeRosAlloc // (uint32_t type_idx, Method* method, Thread*) 1637 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME 1638 RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1639END art_quick_alloc_object_rosalloc 1640 1641// A hand-written override for GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT(_tlab, TLAB). 1642ENTRY art_quick_alloc_object_tlab 1643 // Fast path tlab allocation. 1644 // x0: type_idx/return value, x1: ArtMethod*, xSELF(x19): Thread::Current 1645 // x2-x7: free. 1646#if defined(USE_READ_BARRIER) 1647 mvn x0, xzr // Read barrier not supported here. 1648 ret // Return -1. 1649#endif 1650 ldr x2, [x1, #ART_METHOD_DEX_CACHE_TYPES_OFFSET_64] // Load dex cache resolved types array 1651 // Load the class (x2) 1652 ldr w2, [x2, x0, lsl #COMPRESSED_REFERENCE_SIZE_SHIFT] 1653 cbz x2, .Lart_quick_alloc_object_tlab_slow_path // Check null class 1654 // Check class status. 1655 ldr w3, [x2, #MIRROR_CLASS_STATUS_OFFSET] 1656 cmp x3, #MIRROR_CLASS_STATUS_INITIALIZED 1657 bne .Lart_quick_alloc_object_tlab_slow_path 1658 // Add a fake dependence from the 1659 // following access flag and size 1660 // loads to the status load. 1661 // This is to prevent those loads 1662 // from being reordered above the 1663 // status load and reading wrong 1664 // values (an alternative is to use 1665 // a load-acquire for the status). 1666 eor x3, x3, x3 1667 add x2, x2, x3 1668 // Check access flags has 1669 // kAccClassIsFinalizable. 1670 ldr w3, [x2, #MIRROR_CLASS_ACCESS_FLAGS_OFFSET] 1671 tbnz x3, #ACCESS_FLAGS_CLASS_IS_FINALIZABLE_BIT, .Lart_quick_alloc_object_tlab_slow_path 1672 // Load thread_local_pos (x4) and 1673 // thread_local_end (x5). 1674 ldr x4, [xSELF, #THREAD_LOCAL_POS_OFFSET] 1675 ldr x5, [xSELF, #THREAD_LOCAL_END_OFFSET] 1676 sub x6, x5, x4 // Compute the remaining buf size. 1677 ldr w7, [x2, #MIRROR_CLASS_OBJECT_SIZE_OFFSET] // Load the object size (x7). 1678 cmp x7, x6 // Check if it fits. OK to do this 1679 // before rounding up the object size 1680 // assuming the buf size alignment. 1681 bhi .Lart_quick_alloc_object_tlab_slow_path 1682 // "Point of no slow path". Won't go to the slow path from here on. OK to clobber x0 and x1. 1683 // Round up the object size by the 1684 // object alignment. (addr + 7) & ~7. 1685 add x7, x7, #OBJECT_ALIGNMENT_MASK 1686 and x7, x7, #OBJECT_ALIGNMENT_MASK_TOGGLED 1687 // Move old thread_local_pos to x0 1688 // for the return value. 1689 mov x0, x4 1690 add x5, x0, x7 1691 str x5, [xSELF, #THREAD_LOCAL_POS_OFFSET] // Store new thread_local_pos. 1692 ldr x5, [xSELF, #THREAD_LOCAL_OBJECTS_OFFSET] // Increment thread_local_objects. 1693 add x5, x5, #1 1694 str x5, [xSELF, #THREAD_LOCAL_OBJECTS_OFFSET] 1695 POISON_HEAP_REF w2 1696 str w2, [x0, #MIRROR_OBJECT_CLASS_OFFSET] // Store the class pointer. 1697 // Fence. This is "ish" not "ishst" so 1698 // that the code after this allocation 1699 // site will see the right values in 1700 // the fields of the class. 1701 // Alternatively we could use "ishst" 1702 // if we use load-acquire for the 1703 // class status load.) 1704 dmb ish 1705 ret 1706.Lart_quick_alloc_object_tlab_slow_path: 1707 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // Save callee saves in case of GC. 1708 mov x2, xSELF // Pass Thread::Current. 1709 bl artAllocObjectFromCodeTLAB // (uint32_t type_idx, Method* method, Thread*) 1710 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME 1711 RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER 1712END art_quick_alloc_object_tlab 1713 1714GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT(_region_tlab, RegionTLAB) 1715 1716 /* 1717 * Called by managed code when the thread has been asked to suspend. 1718 */ 1719 .extern artTestSuspendFromCode 1720ENTRY art_quick_test_suspend 1721 ldrh w0, [xSELF, #THREAD_FLAGS_OFFSET] // get xSELF->state_and_flags.as_struct.flags 1722 cbnz w0, .Lneed_suspend // check flags == 0 1723 ret // return if flags == 0 1724.Lneed_suspend: 1725 mov x0, xSELF 1726 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // save callee saves for stack crawl 1727 bl artTestSuspendFromCode // (Thread*) 1728 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME_AND_RETURN 1729END art_quick_test_suspend 1730 1731ENTRY art_quick_implicit_suspend 1732 mov x0, xSELF 1733 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // save callee saves for stack crawl 1734 bl artTestSuspendFromCode // (Thread*) 1735 RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME_AND_RETURN 1736END art_quick_implicit_suspend 1737 1738 /* 1739 * Called by managed code that is attempting to call a method on a proxy class. On entry 1740 * x0 holds the proxy method and x1 holds the receiver; The frame size of the invoked proxy 1741 * method agrees with a ref and args callee save frame. 1742 */ 1743 .extern artQuickProxyInvokeHandler 1744ENTRY art_quick_proxy_invoke_handler 1745 SETUP_REFS_AND_ARGS_CALLEE_SAVE_FRAME_WITH_METHOD_IN_X0 1746 mov x2, xSELF // pass Thread::Current 1747 mov x3, sp // pass SP 1748 bl artQuickProxyInvokeHandler // (Method* proxy method, receiver, Thread*, SP) 1749 ldr x2, [xSELF, THREAD_EXCEPTION_OFFSET] 1750 cbnz x2, .Lexception_in_proxy // success if no exception is pending 1751 RESTORE_REFS_AND_ARGS_CALLEE_SAVE_FRAME // Restore frame 1752 fmov d0, x0 // Store result in d0 in case it was float or double 1753 ret // return on success 1754.Lexception_in_proxy: 1755 RESTORE_REFS_AND_ARGS_CALLEE_SAVE_FRAME 1756 DELIVER_PENDING_EXCEPTION 1757END art_quick_proxy_invoke_handler 1758 1759 /* 1760 * Called to resolve an imt conflict. xIP1 is a hidden argument that holds the target method's 1761 * dex method index. 1762 */ 1763 .extern artInvokeInterfaceTrampoline 1764ENTRY art_quick_imt_conflict_trampoline 1765 mov x0, xIP1 1766 INVOKE_TRAMPOLINE_BODY artInvokeInterfaceTrampoline 1767END art_quick_imt_conflict_trampoline 1768 1769ENTRY art_quick_resolution_trampoline 1770 SETUP_REFS_AND_ARGS_CALLEE_SAVE_FRAME 1771 mov x2, xSELF 1772 mov x3, sp 1773 bl artQuickResolutionTrampoline // (called, receiver, Thread*, SP) 1774 cbz x0, 1f 1775 mov xIP0, x0 // Remember returned code pointer in xIP0. 1776 ldr x0, [sp, #0] // artQuickResolutionTrampoline puts called method in *SP. 1777 RESTORE_REFS_AND_ARGS_CALLEE_SAVE_FRAME 1778 br xIP0 17791: 1780 RESTORE_REFS_AND_ARGS_CALLEE_SAVE_FRAME 1781 DELIVER_PENDING_EXCEPTION 1782END art_quick_resolution_trampoline 1783 1784/* 1785 * Generic JNI frame layout: 1786 * 1787 * #-------------------# 1788 * | | 1789 * | caller method... | 1790 * #-------------------# <--- SP on entry 1791 * | Return X30/LR | 1792 * | X29/FP | callee save 1793 * | X28 | callee save 1794 * | X27 | callee save 1795 * | X26 | callee save 1796 * | X25 | callee save 1797 * | X24 | callee save 1798 * | X23 | callee save 1799 * | X22 | callee save 1800 * | X21 | callee save 1801 * | X20 | callee save 1802 * | X19 | callee save 1803 * | X7 | arg7 1804 * | X6 | arg6 1805 * | X5 | arg5 1806 * | X4 | arg4 1807 * | X3 | arg3 1808 * | X2 | arg2 1809 * | X1 | arg1 1810 * | D7 | float arg 8 1811 * | D6 | float arg 7 1812 * | D5 | float arg 6 1813 * | D4 | float arg 5 1814 * | D3 | float arg 4 1815 * | D2 | float arg 3 1816 * | D1 | float arg 2 1817 * | D0 | float arg 1 1818 * | Method* | <- X0 1819 * #-------------------# 1820 * | local ref cookie | // 4B 1821 * | handle scope size | // 4B 1822 * #-------------------# 1823 * | JNI Call Stack | 1824 * #-------------------# <--- SP on native call 1825 * | | 1826 * | Stack for Regs | The trampoline assembly will pop these values 1827 * | | into registers for native call 1828 * #-------------------# 1829 * | Native code ptr | 1830 * #-------------------# 1831 * | Free scratch | 1832 * #-------------------# 1833 * | Ptr to (1) | <--- SP 1834 * #-------------------# 1835 */ 1836 /* 1837 * Called to do a generic JNI down-call 1838 */ 1839ENTRY art_quick_generic_jni_trampoline 1840 SETUP_REFS_AND_ARGS_CALLEE_SAVE_FRAME_WITH_METHOD_IN_X0 1841 1842 // Save SP , so we can have static CFI info. 1843 mov x28, sp 1844 .cfi_def_cfa_register x28 1845 1846 // This looks the same, but is different: this will be updated to point to the bottom 1847 // of the frame when the handle scope is inserted. 1848 mov xFP, sp 1849 1850 mov xIP0, #5120 1851 sub sp, sp, xIP0 1852 1853 // prepare for artQuickGenericJniTrampoline call 1854 // (Thread*, SP) 1855 // x0 x1 <= C calling convention 1856 // xSELF xFP <= where they are 1857 1858 mov x0, xSELF // Thread* 1859 mov x1, xFP 1860 bl artQuickGenericJniTrampoline // (Thread*, sp) 1861 1862 // The C call will have registered the complete save-frame on success. 1863 // The result of the call is: 1864 // x0: pointer to native code, 0 on error. 1865 // x1: pointer to the bottom of the used area of the alloca, can restore stack till there. 1866 1867 // Check for error = 0. 1868 cbz x0, .Lexception_in_native 1869 1870 // Release part of the alloca. 1871 mov sp, x1 1872 1873 // Save the code pointer 1874 mov xIP0, x0 1875 1876 // Load parameters from frame into registers. 1877 // TODO Check with artQuickGenericJniTrampoline. 1878 // Also, check again APPCS64 - the stack arguments are interleaved. 1879 ldp x0, x1, [sp] 1880 ldp x2, x3, [sp, #16] 1881 ldp x4, x5, [sp, #32] 1882 ldp x6, x7, [sp, #48] 1883 1884 ldp d0, d1, [sp, #64] 1885 ldp d2, d3, [sp, #80] 1886 ldp d4, d5, [sp, #96] 1887 ldp d6, d7, [sp, #112] 1888 1889 add sp, sp, #128 1890 1891 blr xIP0 // native call. 1892 1893 // result sign extension is handled in C code 1894 // prepare for artQuickGenericJniEndTrampoline call 1895 // (Thread*, result, result_f) 1896 // x0 x1 x2 <= C calling convention 1897 mov x1, x0 // Result (from saved). 1898 mov x0, xSELF // Thread register. 1899 fmov x2, d0 // d0 will contain floating point result, but needs to go into x2 1900 1901 bl artQuickGenericJniEndTrampoline 1902 1903 // Pending exceptions possible. 1904 ldr x2, [xSELF, THREAD_EXCEPTION_OFFSET] 1905 cbnz x2, .Lexception_in_native 1906 1907 // Tear down the alloca. 1908 mov sp, x28 1909 .cfi_def_cfa_register sp 1910 1911 // Tear down the callee-save frame. 1912 RESTORE_REFS_AND_ARGS_CALLEE_SAVE_FRAME 1913 1914 // store into fpr, for when it's a fpr return... 1915 fmov d0, x0 1916 ret 1917 1918.Lexception_in_native: 1919 // Move to x1 then sp to please assembler. 1920 ldr x1, [xSELF, # THREAD_TOP_QUICK_FRAME_OFFSET] 1921 mov sp, x1 1922 .cfi_def_cfa_register sp 1923 # This will create a new save-all frame, required by the runtime. 1924 DELIVER_PENDING_EXCEPTION 1925END art_quick_generic_jni_trampoline 1926 1927/* 1928 * Called to bridge from the quick to interpreter ABI. On entry the arguments match those 1929 * of a quick call: 1930 * x0 = method being called/to bridge to. 1931 * x1..x7, d0..d7 = arguments to that method. 1932 */ 1933ENTRY art_quick_to_interpreter_bridge 1934 SETUP_REFS_AND_ARGS_CALLEE_SAVE_FRAME // Set up frame and save arguments. 1935 1936 // x0 will contain mirror::ArtMethod* method. 1937 mov x1, xSELF // How to get Thread::Current() ??? 1938 mov x2, sp 1939 1940 // uint64_t artQuickToInterpreterBridge(mirror::ArtMethod* method, Thread* self, 1941 // mirror::ArtMethod** sp) 1942 bl artQuickToInterpreterBridge 1943 1944 RESTORE_REFS_AND_ARGS_CALLEE_SAVE_FRAME // TODO: no need to restore arguments in this case. 1945 1946 fmov d0, x0 1947 1948 RETURN_OR_DELIVER_PENDING_EXCEPTION 1949END art_quick_to_interpreter_bridge 1950 1951 1952// 1953// Instrumentation-related stubs 1954// 1955 .extern artInstrumentationMethodEntryFromCode 1956ENTRY art_quick_instrumentation_entry 1957 SETUP_REFS_AND_ARGS_CALLEE_SAVE_FRAME 1958 1959 mov x20, x0 // Preserve method reference in a callee-save. 1960 1961 mov x2, xSELF 1962 mov x3, xLR 1963 bl artInstrumentationMethodEntryFromCode // (Method*, Object*, Thread*, LR) 1964 1965 mov xIP0, x0 // x0 = result of call. 1966 mov x0, x20 // Reload method reference. 1967 1968 RESTORE_REFS_AND_ARGS_CALLEE_SAVE_FRAME // Note: will restore xSELF 1969 adr xLR, art_quick_instrumentation_exit 1970 br xIP0 // Tail-call method with lr set to art_quick_instrumentation_exit. 1971END art_quick_instrumentation_entry 1972 1973 .extern artInstrumentationMethodExitFromCode 1974ENTRY art_quick_instrumentation_exit 1975 mov xLR, #0 // Clobber LR for later checks. 1976 1977 SETUP_REFS_ONLY_CALLEE_SAVE_FRAME 1978 1979 // We need to save x0 and d0. We could use a callee-save from SETUP_REF_ONLY, but then 1980 // we would need to fully restore it. As there are a lot of callee-save registers, it seems 1981 // easier to have an extra small stack area. 1982 1983 str x0, [sp, #-16]! // Save integer result. 1984 .cfi_adjust_cfa_offset 16 1985 str d0, [sp, #8] // Save floating-point result. 1986 1987 add x1, sp, #16 // Pass SP. 1988 mov x2, x0 // Pass integer result. 1989 fmov x3, d0 // Pass floating-point result. 1990 mov x0, xSELF // Pass Thread. 1991 bl artInstrumentationMethodExitFromCode // (Thread*, SP, gpr_res, fpr_res) 1992 1993 mov xIP0, x0 // Return address from instrumentation call. 1994 mov xLR, x1 // r1 is holding link register if we're to bounce to deoptimize 1995 1996 ldr d0, [sp, #8] // Restore floating-point result. 1997 ldr x0, [sp], 16 // Restore integer result, and drop stack area. 1998 .cfi_adjust_cfa_offset 16 1999 2000 POP_REFS_ONLY_CALLEE_SAVE_FRAME 2001 2002 br xIP0 // Tail-call out. 2003END art_quick_instrumentation_exit 2004 2005 /* 2006 * Instrumentation has requested that we deoptimize into the interpreter. The deoptimization 2007 * will long jump to the upcall with a special exception of -1. 2008 */ 2009 .extern artDeoptimize 2010ENTRY art_quick_deoptimize 2011 SETUP_SAVE_ALL_CALLEE_SAVE_FRAME 2012 mov x0, xSELF // Pass thread. 2013 bl artDeoptimize // artDeoptimize(Thread*) 2014 brk 0 2015END art_quick_deoptimize 2016 2017 /* 2018 * Compiled code has requested that we deoptimize into the interpreter. The deoptimization 2019 * will long jump to the upcall with a special exception of -1. 2020 */ 2021 .extern artDeoptimizeFromCompiledCode 2022ENTRY art_quick_deoptimize_from_compiled_code 2023 SETUP_SAVE_ALL_CALLEE_SAVE_FRAME 2024 mov x0, xSELF // Pass thread. 2025 bl artDeoptimizeFromCompiledCode // artDeoptimizeFromCompiledCode(Thread*) 2026 brk 0 2027END art_quick_deoptimize_from_compiled_code 2028 2029 2030 /* 2031 * String's indexOf. 2032 * 2033 * TODO: Not very optimized. 2034 * On entry: 2035 * x0: string object (known non-null) 2036 * w1: char to match (known <= 0xFFFF) 2037 * w2: Starting offset in string data 2038 */ 2039ENTRY art_quick_indexof 2040 ldr w3, [x0, #MIRROR_STRING_COUNT_OFFSET] 2041 add x0, x0, #MIRROR_STRING_VALUE_OFFSET 2042 2043 /* Clamp start to [0..count] */ 2044 cmp w2, #0 2045 csel w2, wzr, w2, lt 2046 cmp w2, w3 2047 csel w2, w3, w2, gt 2048 2049 /* Save a copy to compute result */ 2050 mov x5, x0 2051 2052 /* Build pointer to start of data to compare and pre-bias */ 2053 add x0, x0, x2, lsl #1 2054 sub x0, x0, #2 2055 2056 /* Compute iteration count */ 2057 sub w2, w3, w2 2058 2059 /* 2060 * At this point we have: 2061 * x0: start of the data to test 2062 * w1: char to compare 2063 * w2: iteration count 2064 * x5: original start of string data 2065 */ 2066 2067 subs w2, w2, #4 2068 b.lt .Lindexof_remainder 2069 2070.Lindexof_loop4: 2071 ldrh w6, [x0, #2]! 2072 ldrh w7, [x0, #2]! 2073 ldrh wIP0, [x0, #2]! 2074 ldrh wIP1, [x0, #2]! 2075 cmp w6, w1 2076 b.eq .Lmatch_0 2077 cmp w7, w1 2078 b.eq .Lmatch_1 2079 cmp wIP0, w1 2080 b.eq .Lmatch_2 2081 cmp wIP1, w1 2082 b.eq .Lmatch_3 2083 subs w2, w2, #4 2084 b.ge .Lindexof_loop4 2085 2086.Lindexof_remainder: 2087 adds w2, w2, #4 2088 b.eq .Lindexof_nomatch 2089 2090.Lindexof_loop1: 2091 ldrh w6, [x0, #2]! 2092 cmp w6, w1 2093 b.eq .Lmatch_3 2094 subs w2, w2, #1 2095 b.ne .Lindexof_loop1 2096 2097.Lindexof_nomatch: 2098 mov x0, #-1 2099 ret 2100 2101.Lmatch_0: 2102 sub x0, x0, #6 2103 sub x0, x0, x5 2104 asr x0, x0, #1 2105 ret 2106.Lmatch_1: 2107 sub x0, x0, #4 2108 sub x0, x0, x5 2109 asr x0, x0, #1 2110 ret 2111.Lmatch_2: 2112 sub x0, x0, #2 2113 sub x0, x0, x5 2114 asr x0, x0, #1 2115 ret 2116.Lmatch_3: 2117 sub x0, x0, x5 2118 asr x0, x0, #1 2119 ret 2120END art_quick_indexof 2121 2122 /* 2123 * String's compareTo. 2124 * 2125 * TODO: Not very optimized. 2126 * 2127 * On entry: 2128 * x0: this object pointer 2129 * x1: comp object pointer 2130 * 2131 */ 2132 .extern __memcmp16 2133ENTRY art_quick_string_compareto 2134 mov x2, x0 // x0 is return, use x2 for first input. 2135 sub x0, x2, x1 // Same string object? 2136 cbnz x0,1f 2137 ret 21381: // Different string objects. 2139 2140 ldr w4, [x2, #MIRROR_STRING_COUNT_OFFSET] 2141 ldr w3, [x1, #MIRROR_STRING_COUNT_OFFSET] 2142 add x2, x2, #MIRROR_STRING_VALUE_OFFSET 2143 add x1, x1, #MIRROR_STRING_VALUE_OFFSET 2144 2145 /* 2146 * Now: Data* Count 2147 * first arg x2 w4 2148 * second arg x1 w3 2149 */ 2150 2151 // x0 := str1.length(w4) - str2.length(w3). ldr zero-extended w3/w4 into x3/x4. 2152 subs x0, x4, x3 2153 // Min(count1, count2) into w3. 2154 csel x3, x3, x4, ge 2155 2156 // TODO: Tune this value. 2157 // Check for long string, do memcmp16 for them. 2158 cmp w3, #28 // Constant from arm32. 2159 bgt .Ldo_memcmp16 2160 2161 /* 2162 * Now: 2163 * x2: *first string data 2164 * x1: *second string data 2165 * w3: iteration count 2166 * x0: return value if comparison equal 2167 * x4, x5, x6, x7: free 2168 */ 2169 2170 // Do a simple unrolled loop. 2171.Lloop: 2172 // At least two more elements? 2173 subs w3, w3, #2 2174 b.lt .Lremainder_or_done 2175 2176 ldrh w4, [x2], #2 2177 ldrh w5, [x1], #2 2178 2179 ldrh w6, [x2], #2 2180 ldrh w7, [x1], #2 2181 2182 subs w4, w4, w5 2183 b.ne .Lw4_result 2184 2185 subs w6, w6, w7 2186 b.ne .Lw6_result 2187 2188 b .Lloop 2189 2190.Lremainder_or_done: 2191 adds w3, w3, #1 2192 b.eq .Lremainder 2193 ret 2194 2195.Lremainder: 2196 ldrh w4, [x2], #2 2197 ldrh w5, [x1], #2 2198 subs w4, w4, w5 2199 b.ne .Lw4_result 2200 ret 2201 2202// Result is in w4 2203.Lw4_result: 2204 sxtw x0, w4 2205 ret 2206 2207// Result is in w6 2208.Lw6_result: 2209 sxtw x0, w6 2210 ret 2211 2212.Ldo_memcmp16: 2213 mov x14, x0 // Save x0 and LR. __memcmp16 does not use these temps. 2214 mov x15, xLR // TODO: Codify and check that? 2215 2216 mov x0, x2 2217 uxtw x2, w3 2218 bl __memcmp16 2219 2220 mov xLR, x15 // Restore LR. 2221 2222 cmp x0, #0 // Check the memcmp difference. 2223 csel x0, x0, x14, ne // x0 := x0 != 0 ? x14(prev x0=length diff) : x1. 2224 ret 2225END art_quick_string_compareto 2226