< prev index next >

src/hotspot/cpu/aarch64/sharedRuntime_aarch64.cpp

Print this page

   9  * published by the Free Software Foundation.
  10  *
  11  * This code is distributed in the hope that it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  14  * version 2 for more details (a copy is included in the LICENSE file that
  15  * accompanied this code).
  16  *
  17  * You should have received a copy of the GNU General Public License version
  18  * 2 along with this work; if not, write to the Free Software Foundation,
  19  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  20  *
  21  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  22  * or visit www.oracle.com if you need additional information or have any
  23  * questions.
  24  *
  25  */
  26 
  27 #include "asm/macroAssembler.hpp"
  28 #include "asm/macroAssembler.inline.hpp"

  29 #include "code/aotCodeCache.hpp"
  30 #include "code/codeCache.hpp"
  31 #include "code/compiledIC.hpp"
  32 #include "code/debugInfoRec.hpp"
  33 #include "code/vtableStubs.hpp"
  34 #include "compiler/oopMap.hpp"
  35 #include "gc/shared/barrierSetAssembler.hpp"
  36 #include "interpreter/interpreter.hpp"
  37 #include "interpreter/interp_masm.hpp"
  38 #include "logging/log.hpp"
  39 #include "memory/resourceArea.hpp"
  40 #include "nativeInst_aarch64.hpp"
  41 #include "oops/klass.inline.hpp"
  42 #include "oops/method.inline.hpp"
  43 #include "prims/methodHandles.hpp"
  44 #include "runtime/continuation.hpp"
  45 #include "runtime/continuationEntry.inline.hpp"
  46 #include "runtime/globals.hpp"
  47 #include "runtime/jniHandles.hpp"
  48 #include "runtime/safepointMechanism.hpp"

 184 
 185   int frame_size_in_bytes = align_up(additional_frame_words * wordSize +
 186                                      reg_save_size * BytesPerInt, 16);
 187   // OopMap frame size is in compiler stack slots (jint's) not bytes or words
 188   int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
 189   // The caller will allocate additional_frame_words
 190   int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt;
 191   // CodeBlob frame size is in words.
 192   int frame_size_in_words = frame_size_in_bytes / wordSize;
 193   *total_frame_words = frame_size_in_words;
 194 
 195   // Save Integer and Float registers.
 196   __ enter();
 197   __ push_CPU_state(_save_vectors, use_sve, sve_vector_size_in_bytes, total_predicate_in_bytes);
 198 
 199   // Set an oopmap for the call site.  This oopmap will map all
 200   // oop-registers and debug-info registers as callee-saved.  This
 201   // will allow deoptimization at this safepoint to find all possible
 202   // debug-info recordings, as well as let GC find all oops.
 203 
 204   OopMapSet *oop_maps = new OopMapSet();
 205   OopMap* oop_map = new OopMap(frame_size_in_slots, 0);
 206 
 207   for (int i = 0; i < Register::number_of_registers; i++) {
 208     Register r = as_Register(i);
 209     if (i <= rfp->encoding() && r != rscratch1 && r != rscratch2) {
 210       // SP offsets are in 4-byte words.
 211       // Register slots are 8 bytes wide, 32 floating-point registers.
 212       int sp_offset = Register::max_slots_per_register * i +
 213                       FloatRegister::save_slots_per_register * FloatRegister::number_of_registers;
 214       oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset + additional_frame_slots), r->as_VMReg());
 215     }
 216   }
 217 
 218   for (int i = 0; i < FloatRegister::number_of_registers; i++) {
 219     FloatRegister r = as_FloatRegister(i);
 220     int sp_offset = 0;
 221     if (_save_vectors) {
 222       sp_offset = use_sve ? (total_predicate_in_slots + sve_vector_size_in_slots * i) :
 223                             (FloatRegister::slots_per_neon_register * i);
 224     } else {

 334       break;
 335     case T_DOUBLE:
 336       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 337       if (fp_args < Argument::n_float_register_parameters_j) {
 338         regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
 339       } else {
 340         stk_args = align_up(stk_args, 2);
 341         regs[i].set2(VMRegImpl::stack2reg(stk_args));
 342         stk_args += 2;
 343       }
 344       break;
 345     default:
 346       ShouldNotReachHere();
 347       break;
 348     }
 349   }
 350 
 351   return stk_args;
 352 }
 353 















































































 354 // Patch the callers callsite with entry to compiled code if it exists.
 355 static void patch_callers_callsite(MacroAssembler *masm) {
 356   Label L;
 357   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset())));
 358   __ cbz(rscratch1, L);
 359 
 360   __ enter();
 361   __ push_CPU_state();
 362 
 363   // VM needs caller's callsite
 364   // VM needs target method
 365   // This needs to be a long call since we will relocate this adapter to
 366   // the codeBuffer and it may not reach
 367 
 368 #ifndef PRODUCT
 369   assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
 370 #endif
 371 
 372   __ mov(c_rarg0, rmethod);
 373   __ mov(c_rarg1, lr);
 374   __ authenticate_return_address(c_rarg1);
 375   __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)));
 376   __ blr(rscratch1);
 377 
 378   // Explicit isb required because fixup_callers_callsite may change the code
 379   // stream.
 380   __ safepoint_isb();
 381 
 382   __ pop_CPU_state();
 383   // restore sp
 384   __ leave();
 385   __ bind(L);
 386 }
 387 














































































































 388 static void gen_c2i_adapter(MacroAssembler *masm,
 389                             int total_args_passed,
 390                             int comp_args_on_stack,
 391                             const BasicType *sig_bt,
 392                             const VMRegPair *regs,
 393                             Label& skip_fixup) {




























 394   // Before we get into the guts of the C2I adapter, see if we should be here
 395   // at all.  We've come from compiled code and are attempting to jump to the
 396   // interpreter, which means the caller made a static call to get here
 397   // (vcalls always get a compiled target if there is one).  Check for a
 398   // compiled target.  If there is one, we need to patch the caller's call.
 399   patch_callers_callsite(masm);
 400 
 401   __ bind(skip_fixup);
 402 
 403   int words_pushed = 0;








 404 
 405   // Since all args are passed on the stack, total_args_passed *
 406   // Interpreter::stackElementSize is the space we need.






 407 
 408   int extraspace = total_args_passed * Interpreter::stackElementSize;












 409 
 410   __ mov(r19_sender_sp, sp);

 411 
 412   // stack is aligned, keep it that way
 413   extraspace = align_up(extraspace, 2*wordSize);
 414 
 415   if (extraspace)
 416     __ sub(sp, sp, extraspace);

 417 
 418   // Now write the args into the outgoing interpreter space
 419   for (int i = 0; i < total_args_passed; i++) {
 420     if (sig_bt[i] == T_VOID) {
 421       assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
 422       continue;
 423     }
 424 
 425     // offset to start parameters
 426     int st_off   = (total_args_passed - i - 1) * Interpreter::stackElementSize;
 427     int next_off = st_off - Interpreter::stackElementSize;
 428 
 429     // Say 4 args:
 430     // i   st_off
 431     // 0   32 T_LONG
 432     // 1   24 T_VOID
 433     // 2   16 T_OBJECT
 434     // 3    8 T_BOOL
 435     // -    0 return address
 436     //
 437     // However to make thing extra confusing. Because we can fit a Java long/double in
 438     // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
 439     // leaves one slot empty and only stores to a single slot. In this case the
 440     // slot that is occupied is the T_VOID slot. See I said it was confusing.
 441 
 442     VMReg r_1 = regs[i].first();
 443     VMReg r_2 = regs[i].second();
 444     if (!r_1->is_valid()) {
 445       assert(!r_2->is_valid(), "");
 446       continue;









 447     }
 448     if (r_1->is_stack()) {
 449       // memory to memory use rscratch1
 450       int ld_off = (r_1->reg2stack() * VMRegImpl::stack_slot_size
 451                     + extraspace
 452                     + words_pushed * wordSize);
 453       if (!r_2->is_valid()) {
 454         // sign extend??
 455         __ ldrw(rscratch1, Address(sp, ld_off));
 456         __ str(rscratch1, Address(sp, st_off));
 457 
 458       } else {




 459 
 460         __ ldr(rscratch1, Address(sp, ld_off));

 461 
 462         // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
 463         // T_DOUBLE and T_LONG use two slots in the interpreter
 464         if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
 465           // ld_off == LSW, ld_off+wordSize == MSW
 466           // st_off == MSW, next_off == LSW
 467           __ str(rscratch1, Address(sp, next_off));

























 468 #ifdef ASSERT
 469           // Overwrite the unused slot with known junk
 470           __ mov(rscratch1, (uint64_t)0xdeadffffdeadaaaaull);
 471           __ str(rscratch1, Address(sp, st_off));
 472 #endif /* ASSERT */
 473         } else {
 474           __ str(rscratch1, Address(sp, st_off));
 475         }
 476       }
 477     } else if (r_1->is_Register()) {
 478       Register r = r_1->as_Register();
 479       if (!r_2->is_valid()) {
 480         // must be only an int (or less ) so move only 32bits to slot
 481         // why not sign extend??
 482         __ str(r, Address(sp, st_off));
 483       } else {
 484         // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
 485         // T_DOUBLE and T_LONG use two slots in the interpreter
 486         if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
 487           // jlong/double in gpr
 488 #ifdef ASSERT
 489           // Overwrite the unused slot with known junk
 490           __ mov(rscratch1, (uint64_t)0xdeadffffdeadaaabull);
 491           __ str(rscratch1, Address(sp, st_off));
 492 #endif /* ASSERT */
 493           __ str(r, Address(sp, next_off));


































 494         } else {
 495           __ str(r, Address(sp, st_off));





















 496         }
 497       }
 498     } else {
 499       assert(r_1->is_FloatRegister(), "");
 500       if (!r_2->is_valid()) {
 501         // only a float use just part of the slot
 502         __ strs(r_1->as_FloatRegister(), Address(sp, st_off));
 503       } else {
 504 #ifdef ASSERT
 505         // Overwrite the unused slot with known junk
 506         __ mov(rscratch1, (uint64_t)0xdeadffffdeadaaacull);
 507         __ str(rscratch1, Address(sp, st_off));
 508 #endif /* ASSERT */
 509         __ strd(r_1->as_FloatRegister(), Address(sp, next_off));
 510       }
 511     }
 512   }
 513 
 514   __ mov(esp, sp); // Interp expects args on caller's expression stack
 515 
 516   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::interpreter_entry_offset())));
 517   __ br(rscratch1);
 518 }
 519 

 520 
 521 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
 522                                     int total_args_passed,
 523                                     int comp_args_on_stack,
 524                                     const BasicType *sig_bt,
 525                                     const VMRegPair *regs) {
 526 
 527   // Note: r19_sender_sp contains the senderSP on entry. We must
 528   // preserve it since we may do a i2c -> c2i transition if we lose a
 529   // race where compiled code goes non-entrant while we get args
 530   // ready.
 531 
 532   // Adapters are frameless.
 533 
 534   // An i2c adapter is frameless because the *caller* frame, which is
 535   // interpreted, routinely repairs its own esp (from
 536   // interpreter_frame_last_sp), even if a callee has modified the
 537   // stack pointer.  It also recalculates and aligns sp.
 538 
 539   // A c2i adapter is frameless because the *callee* frame, which is
 540   // interpreted, routinely repairs its caller's sp (from sender_sp,
 541   // which is set up via the senderSP register).
 542 
 543   // In other words, if *either* the caller or callee is interpreted, we can
 544   // get the stack pointer repaired after a call.
 545 
 546   // This is why c2i and i2c adapters cannot be indefinitely composed.
 547   // In particular, if a c2i adapter were to somehow call an i2c adapter,
 548   // both caller and callee would be compiled methods, and neither would
 549   // clean up the stack pointer changes performed by the two adapters.
 550   // If this happens, control eventually transfers back to the compiled
 551   // caller, but with an uncorrected stack, causing delayed havoc.
 552 
 553   // Cut-out for having no stack args.
 554   int comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
 555   if (comp_args_on_stack) {
 556     __ sub(rscratch1, sp, comp_words_on_stack * wordSize);
 557     __ andr(sp, rscratch1, -16);

 558   }
 559 
 560   // Will jump to the compiled code just as if compiled code was doing it.
 561   // Pre-load the register-jump target early, to schedule it better.
 562   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::from_compiled_offset())));


 563 
 564   // Now generate the shuffle code.
 565   for (int i = 0; i < total_args_passed; i++) {
 566     if (sig_bt[i] == T_VOID) {
 567       assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");

 568       continue;
 569     }
 570 
 571     // Pick up 0, 1 or 2 words from SP+offset.

 572 
 573     assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
 574             "scrambled load targets?");
 575     // Load in argument order going down.
 576     int ld_off = (total_args_passed - i - 1)*Interpreter::stackElementSize;
 577     // Point to interpreter value (vs. tag)
 578     int next_off = ld_off - Interpreter::stackElementSize;
 579     //
 580     //
 581     //
 582     VMReg r_1 = regs[i].first();
 583     VMReg r_2 = regs[i].second();
 584     if (!r_1->is_valid()) {
 585       assert(!r_2->is_valid(), "");
 586       continue;
 587     }
 588     if (r_1->is_stack()) {
 589       // Convert stack slot to an SP offset (+ wordSize to account for return address )
 590       int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size;
 591       if (!r_2->is_valid()) {
 592         // sign extend???
 593         __ ldrsw(rscratch2, Address(esp, ld_off));
 594         __ str(rscratch2, Address(sp, st_off));
 595       } else {
 596         //
 597         // We are using two optoregs. This can be either T_OBJECT,
 598         // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
 599         // two slots but only uses one for thr T_LONG or T_DOUBLE case
 600         // So we must adjust where to pick up the data to match the
 601         // interpreter.
 602         //
 603         // Interpreter local[n] == MSW, local[n+1] == LSW however locals
 604         // are accessed as negative so LSW is at LOW address
 605 
 606         // ld_off is MSW so get LSW
 607         const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)?
 608                            next_off : ld_off;
 609         __ ldr(rscratch2, Address(esp, offset));
 610         // st_off is LSW (i.e. reg.first())
 611         __ str(rscratch2, Address(sp, st_off));
 612       }
 613     } else if (r_1->is_Register()) {  // Register argument
 614       Register r = r_1->as_Register();
 615       if (r_2->is_valid()) {
 616         //
 617         // We are using two VMRegs. This can be either T_OBJECT,
 618         // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
 619         // two slots but only uses one for thr T_LONG or T_DOUBLE case
 620         // So we must adjust where to pick up the data to match the
 621         // interpreter.
 622 
 623         const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)?
 624                            next_off : ld_off;
 625 
 626         // this can be a misaligned move
 627         __ ldr(r, Address(esp, offset));
 628       } else {
 629         // sign extend and use a full word?
 630         __ ldrw(r, Address(esp, ld_off));
 631       }
 632     } else {
 633       if (!r_2->is_valid()) {
 634         __ ldrs(r_1->as_FloatRegister(), Address(esp, ld_off));
 635       } else {
 636         __ ldrd(r_1->as_FloatRegister(), Address(esp, next_off));
 637       }
 638     }
 639   }
 640 
 641   __ mov(rscratch2, rscratch1);
 642   __ push_cont_fastpath(rthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about; kills rscratch1
 643   __ mov(rscratch1, rscratch2);
 644 
 645   // 6243940 We might end up in handle_wrong_method if
 646   // the callee is deoptimized as we race thru here. If that
 647   // happens we don't want to take a safepoint because the
 648   // caller frame will look interpreted and arguments are now
 649   // "compiled" so it is much better to make this transition
 650   // invisible to the stack walking code. Unfortunately if
 651   // we try and find the callee by normal means a safepoint
 652   // is possible. So we stash the desired callee in the thread
 653   // and the vm will find there should this case occur.
 654 
 655   __ str(rmethod, Address(rthread, JavaThread::callee_target_offset()));
 656 
 657   __ br(rscratch1);
 658 }
 659 
 660 // ---------------------------------------------------------------
 661 void SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
 662                                             int total_args_passed,
 663                                             int comp_args_on_stack,
 664                                             const BasicType *sig_bt,
 665                                             const VMRegPair *regs,
 666                                             address entry_address[AdapterBlob::ENTRY_COUNT]) {
 667   entry_address[AdapterBlob::I2C] = __ pc();
 668 
 669   gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);






 670 
 671   entry_address[AdapterBlob::C2I_Unverified] = __ pc();
 672   Label skip_fixup;










 673 
 674   Register data = rscratch2;
 675   Register receiver = j_rarg0;
 676   Register tmp = r10;  // A call-clobbered register not used for arg passing
 677 
 678   // -------------------------------------------------------------------------
 679   // Generate a C2I adapter.  On entry we know rmethod holds the Method* during calls
 680   // to the interpreter.  The args start out packed in the compiled layout.  They
 681   // need to be unpacked into the interpreter layout.  This will almost always
 682   // require some stack space.  We grow the current (compiled) stack, then repack
 683   // the args.  We  finally end in a jump to the generic interpreter entry point.
 684   // On exit from the interpreter, the interpreter will restore our SP (lest the
 685   // compiled code, which relies solely on SP and not FP, get sick).
 686 
 687   {
 688     __ block_comment("c2i_unverified_entry {");
 689     // Method might have been compiled since the call site was patched to
 690     // interpreted; if that is the case treat it as a miss so we can get
 691     // the call site corrected.
 692     __ ic_check(1 /* end_alignment */);
 693     __ ldr(rmethod, Address(data, CompiledICData::speculated_method_offset()));
 694 
 695     __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset())));
 696     __ cbz(rscratch1, skip_fixup);
 697     __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
 698     __ block_comment("} c2i_unverified_entry");
 699   }
 700 
 701   entry_address[AdapterBlob::C2I] = __ pc();


 702 
 703   // Class initialization barrier for static methods
 704   entry_address[AdapterBlob::C2I_No_Clinit_Check] = nullptr;
 705   assert(VM_Version::supports_fast_class_init_checks(), "sanity");
 706   Label L_skip_barrier;
 707 
 708   // Bypass the barrier for non-static methods
 709   __ ldrh(rscratch1, Address(rmethod, Method::access_flags_offset()));
 710   __ andsw(zr, rscratch1, JVM_ACC_STATIC);
 711   __ br(Assembler::EQ, L_skip_barrier); // non-static
 712 
 713   __ load_method_holder(rscratch2, rmethod);
 714   __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier);
 715   __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
 716 
 717   __ bind(L_skip_barrier);
 718   entry_address[AdapterBlob::C2I_No_Clinit_Check] = __ pc();
 719 
 720   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
 721   bs->c2i_entry_barrier(masm);














 722 
 723   gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
 724   return;







 725 }
 726 
 727 static int c_calling_convention_priv(const BasicType *sig_bt,
 728                                          VMRegPair *regs,
 729                                          int total_args_passed) {
 730 
 731 // We return the amount of VMRegImpl stack slots we need to reserve for all
 732 // the arguments NOT counting out_preserve_stack_slots.
 733 
 734     static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
 735       c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5,  c_rarg6,  c_rarg7
 736     };
 737     static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
 738       c_farg0, c_farg1, c_farg2, c_farg3,
 739       c_farg4, c_farg5, c_farg6, c_farg7
 740     };
 741 
 742     uint int_args = 0;
 743     uint fp_args = 0;
 744     uint stk_args = 0; // inc by 2 each time

2603 
2604   // exception pending => remove activation and forward to exception handler
2605 
2606   __ str(zr, Address(rthread, JavaThread::vm_result_oop_offset()));
2607 
2608   __ ldr(r0, Address(rthread, Thread::pending_exception_offset()));
2609   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2610 
2611   // -------------
2612   // make sure all code is generated
2613   masm->flush();
2614 
2615   // return the  blob
2616   // frame_size_words or bytes??
2617   RuntimeStub* rs_blob = RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
2618 
2619   AOTCodeCache::store_code_blob(*rs_blob, AOTCodeEntry::SharedBlob, StubInfo::blob(id));
2620   return rs_blob;
2621 }
2622 






















































































































































2623 // Continuation point for throwing of implicit exceptions that are
2624 // not handled in the current activation. Fabricates an exception
2625 // oop and initiates normal exception dispatching in this
2626 // frame. Since we need to preserve callee-saved values (currently
2627 // only for C2, but done for C1 as well) we need a callee-saved oop
2628 // map and therefore have to make these stubs into RuntimeStubs
2629 // rather than BufferBlobs.  If the compiler needs all registers to
2630 // be preserved between the fault point and the exception handler
2631 // then it must assume responsibility for that in
2632 // AbstractCompiler::continuation_for_implicit_null_exception or
2633 // continuation_for_implicit_division_by_zero_exception. All other
2634 // implicit exceptions (e.g., NullPointerException or
2635 // AbstractMethodError on entry) are either at call sites or
2636 // otherwise assume that stack unwinding will be initiated, so
2637 // caller saved registers were assumed volatile in the compiler.
2638 
2639 RuntimeStub* SharedRuntime::generate_throw_exception(StubId id, address runtime_entry) {
2640   assert(is_throw_id(id), "expected a throw stub id");
2641 
2642   const char* name = SharedRuntime::stub_name(id);

   9  * published by the Free Software Foundation.
  10  *
  11  * This code is distributed in the hope that it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  14  * version 2 for more details (a copy is included in the LICENSE file that
  15  * accompanied this code).
  16  *
  17  * You should have received a copy of the GNU General Public License version
  18  * 2 along with this work; if not, write to the Free Software Foundation,
  19  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  20  *
  21  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  22  * or visit www.oracle.com if you need additional information or have any
  23  * questions.
  24  *
  25  */
  26 
  27 #include "asm/macroAssembler.hpp"
  28 #include "asm/macroAssembler.inline.hpp"
  29 #include "classfile/symbolTable.hpp"
  30 #include "code/aotCodeCache.hpp"
  31 #include "code/codeCache.hpp"
  32 #include "code/compiledIC.hpp"
  33 #include "code/debugInfoRec.hpp"
  34 #include "code/vtableStubs.hpp"
  35 #include "compiler/oopMap.hpp"
  36 #include "gc/shared/barrierSetAssembler.hpp"
  37 #include "interpreter/interpreter.hpp"
  38 #include "interpreter/interp_masm.hpp"
  39 #include "logging/log.hpp"
  40 #include "memory/resourceArea.hpp"
  41 #include "nativeInst_aarch64.hpp"
  42 #include "oops/klass.inline.hpp"
  43 #include "oops/method.inline.hpp"
  44 #include "prims/methodHandles.hpp"
  45 #include "runtime/continuation.hpp"
  46 #include "runtime/continuationEntry.inline.hpp"
  47 #include "runtime/globals.hpp"
  48 #include "runtime/jniHandles.hpp"
  49 #include "runtime/safepointMechanism.hpp"

 185 
 186   int frame_size_in_bytes = align_up(additional_frame_words * wordSize +
 187                                      reg_save_size * BytesPerInt, 16);
 188   // OopMap frame size is in compiler stack slots (jint's) not bytes or words
 189   int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
 190   // The caller will allocate additional_frame_words
 191   int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt;
 192   // CodeBlob frame size is in words.
 193   int frame_size_in_words = frame_size_in_bytes / wordSize;
 194   *total_frame_words = frame_size_in_words;
 195 
 196   // Save Integer and Float registers.
 197   __ enter();
 198   __ push_CPU_state(_save_vectors, use_sve, sve_vector_size_in_bytes, total_predicate_in_bytes);
 199 
 200   // Set an oopmap for the call site.  This oopmap will map all
 201   // oop-registers and debug-info registers as callee-saved.  This
 202   // will allow deoptimization at this safepoint to find all possible
 203   // debug-info recordings, as well as let GC find all oops.
 204 

 205   OopMap* oop_map = new OopMap(frame_size_in_slots, 0);
 206 
 207   for (int i = 0; i < Register::number_of_registers; i++) {
 208     Register r = as_Register(i);
 209     if (i <= rfp->encoding() && r != rscratch1 && r != rscratch2) {
 210       // SP offsets are in 4-byte words.
 211       // Register slots are 8 bytes wide, 32 floating-point registers.
 212       int sp_offset = Register::max_slots_per_register * i +
 213                       FloatRegister::save_slots_per_register * FloatRegister::number_of_registers;
 214       oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset + additional_frame_slots), r->as_VMReg());
 215     }
 216   }
 217 
 218   for (int i = 0; i < FloatRegister::number_of_registers; i++) {
 219     FloatRegister r = as_FloatRegister(i);
 220     int sp_offset = 0;
 221     if (_save_vectors) {
 222       sp_offset = use_sve ? (total_predicate_in_slots + sve_vector_size_in_slots * i) :
 223                             (FloatRegister::slots_per_neon_register * i);
 224     } else {

 334       break;
 335     case T_DOUBLE:
 336       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 337       if (fp_args < Argument::n_float_register_parameters_j) {
 338         regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
 339       } else {
 340         stk_args = align_up(stk_args, 2);
 341         regs[i].set2(VMRegImpl::stack2reg(stk_args));
 342         stk_args += 2;
 343       }
 344       break;
 345     default:
 346       ShouldNotReachHere();
 347       break;
 348     }
 349   }
 350 
 351   return stk_args;
 352 }
 353 
 354 
 355 const uint SharedRuntime::java_return_convention_max_int = Argument::n_int_register_parameters_j;
 356 const uint SharedRuntime::java_return_convention_max_float = Argument::n_float_register_parameters_j;
 357 
 358 int SharedRuntime::java_return_convention(const BasicType *sig_bt, VMRegPair *regs, int total_args_passed) {
 359 
 360   // Create the mapping between argument positions and registers.
 361 
 362   static const Register INT_ArgReg[java_return_convention_max_int] = {
 363     r0 /* j_rarg7 */, j_rarg6, j_rarg5, j_rarg4, j_rarg3, j_rarg2, j_rarg1, j_rarg0
 364   };
 365 
 366   static const FloatRegister FP_ArgReg[java_return_convention_max_float] = {
 367     j_farg0, j_farg1, j_farg2, j_farg3, j_farg4, j_farg5, j_farg6, j_farg7
 368   };
 369 
 370   uint int_args = 0;
 371   uint fp_args = 0;
 372 
 373   for (int i = 0; i < total_args_passed; i++) {
 374     switch (sig_bt[i]) {
 375     case T_BOOLEAN:
 376     case T_CHAR:
 377     case T_BYTE:
 378     case T_SHORT:
 379     case T_INT:
 380       if (int_args < SharedRuntime::java_return_convention_max_int) {
 381         regs[i].set1(INT_ArgReg[int_args]->as_VMReg());
 382         int_args ++;
 383       } else {
 384         return -1;
 385       }
 386       break;
 387     case T_VOID:
 388       // halves of T_LONG or T_DOUBLE
 389       assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
 390       regs[i].set_bad();
 391       break;
 392     case T_LONG:
 393       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 394       // fall through
 395     case T_OBJECT:
 396     case T_ARRAY:
 397     case T_ADDRESS:
 398       // Should T_METADATA be added to java_calling_convention as well ?
 399     case T_METADATA:
 400       if (int_args < SharedRuntime::java_return_convention_max_int) {
 401         regs[i].set2(INT_ArgReg[int_args]->as_VMReg());
 402         int_args ++;
 403       } else {
 404         return -1;
 405       }
 406       break;
 407     case T_FLOAT:
 408       if (fp_args < SharedRuntime::java_return_convention_max_float) {
 409         regs[i].set1(FP_ArgReg[fp_args]->as_VMReg());
 410         fp_args ++;
 411       } else {
 412         return -1;
 413       }
 414       break;
 415     case T_DOUBLE:
 416       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 417       if (fp_args < SharedRuntime::java_return_convention_max_float) {
 418         regs[i].set2(FP_ArgReg[fp_args]->as_VMReg());
 419         fp_args ++;
 420       } else {
 421         return -1;
 422       }
 423       break;
 424     default:
 425       ShouldNotReachHere();
 426       break;
 427     }
 428   }
 429 
 430   return int_args + fp_args;
 431 }
 432 
 433 // Patch the callers callsite with entry to compiled code if it exists.
 434 static void patch_callers_callsite(MacroAssembler *masm) {
 435   Label L;
 436   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset())));
 437   __ cbz(rscratch1, L);
 438 
 439   __ enter();
 440   __ push_CPU_state();
 441 
 442   // VM needs caller's callsite
 443   // VM needs target method
 444   // This needs to be a long call since we will relocate this adapter to
 445   // the codeBuffer and it may not reach
 446 
 447 #ifndef PRODUCT
 448   assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
 449 #endif
 450 
 451   __ mov(c_rarg0, rmethod);
 452   __ mov(c_rarg1, lr);
 453   __ authenticate_return_address(c_rarg1);
 454   __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)));
 455   __ blr(rscratch1);
 456 
 457   // Explicit isb required because fixup_callers_callsite may change the code
 458   // stream.
 459   __ safepoint_isb();
 460 
 461   __ pop_CPU_state();
 462   // restore sp
 463   __ leave();
 464   __ bind(L);
 465 }
 466 
 467 // For each inline type argument, sig includes the list of fields of
 468 // the inline type. This utility function computes the number of
 469 // arguments for the call if inline types are passed by reference (the
 470 // calling convention the interpreter expects).
 471 static int compute_total_args_passed_int(const GrowableArray<SigEntry>* sig_extended) {
 472   int total_args_passed = 0;
 473   if (InlineTypePassFieldsAsArgs) {
 474     for (int i = 0; i < sig_extended->length(); i++) {
 475       BasicType bt = sig_extended->at(i)._bt;
 476       if (bt == T_METADATA) {
 477         // In sig_extended, an inline type argument starts with:
 478         // T_METADATA, followed by the types of the fields of the
 479         // inline type and T_VOID to mark the end of the value
 480         // type. Inline types are flattened so, for instance, in the
 481         // case of an inline type with an int field and an inline type
 482         // field that itself has 2 fields, an int and a long:
 483         // T_METADATA T_INT T_METADATA T_INT T_LONG T_VOID (second
 484         // slot for the T_LONG) T_VOID (inner inline type) T_VOID
 485         // (outer inline type)
 486         total_args_passed++;
 487         int vt = 1;
 488         do {
 489           i++;
 490           BasicType bt = sig_extended->at(i)._bt;
 491           BasicType prev_bt = sig_extended->at(i-1)._bt;
 492           if (bt == T_METADATA) {
 493             vt++;
 494           } else if (bt == T_VOID &&
 495                      prev_bt != T_LONG &&
 496                      prev_bt != T_DOUBLE) {
 497             vt--;
 498           }
 499         } while (vt != 0);
 500       } else {
 501         total_args_passed++;
 502       }
 503     }
 504   } else {
 505     total_args_passed = sig_extended->length();
 506   }
 507   return total_args_passed;
 508 }
 509 
 510 
 511 static void gen_c2i_adapter_helper(MacroAssembler* masm,
 512                                    BasicType bt,
 513                                    BasicType prev_bt,
 514                                    size_t size_in_bytes,
 515                                    const VMRegPair& reg_pair,
 516                                    const Address& to,
 517                                    Register tmp1,
 518                                    Register tmp2,
 519                                    Register tmp3,
 520                                    int extraspace,
 521                                    bool is_oop) {
 522   if (bt == T_VOID) {
 523     assert(prev_bt == T_LONG || prev_bt == T_DOUBLE, "missing half");
 524     return;
 525   }
 526 
 527   // Say 4 args:
 528   // i   st_off
 529   // 0   32 T_LONG
 530   // 1   24 T_VOID
 531   // 2   16 T_OBJECT
 532   // 3    8 T_BOOL
 533   // -    0 return address
 534   //
 535   // However to make thing extra confusing. Because we can fit a Java long/double in
 536   // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
 537   // leaves one slot empty and only stores to a single slot. In this case the
 538   // slot that is occupied is the T_VOID slot. See I said it was confusing.
 539 
 540   bool wide = (size_in_bytes == wordSize);
 541   VMReg r_1 = reg_pair.first();
 542   VMReg r_2 = reg_pair.second();
 543   assert(r_2->is_valid() == wide, "invalid size");
 544   if (!r_1->is_valid()) {
 545     assert(!r_2->is_valid(), "");
 546     return;
 547   }
 548 
 549   if (!r_1->is_FloatRegister()) {
 550     Register val = r25;
 551     if (r_1->is_stack()) {
 552       // memory to memory use r25 (scratch registers is used by store_heap_oop)
 553       int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
 554       __ load_sized_value(val, Address(sp, ld_off), size_in_bytes, /* is_signed */ false);
 555     } else {
 556       val = r_1->as_Register();
 557     }
 558     assert_different_registers(to.base(), val, tmp1, tmp2, tmp3);
 559     if (is_oop) {
 560       // store_heap_oop transitively calls oop_store_at which corrupts to.base(). We need to keep it valid.
 561       __ push(to.base(), sp);
 562       __ store_heap_oop(to, val, tmp1, tmp2, tmp3, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
 563       __ pop(to.base(), sp);
 564     } else {
 565       __ store_sized_value(to, val, size_in_bytes);
 566     }
 567   } else {
 568     if (wide) {
 569       __ strd(r_1->as_FloatRegister(), to);
 570     } else {
 571       // only a float use just part of the slot
 572       __ strs(r_1->as_FloatRegister(), to);
 573     }
 574   }
 575 }
 576 
 577 static void gen_c2i_adapter(MacroAssembler *masm,
 578                             const GrowableArray<SigEntry>* sig_extended,


 579                             const VMRegPair *regs,
 580                             bool requires_clinit_barrier,
 581                             address& c2i_no_clinit_check_entry,
 582                             Label& skip_fixup,
 583                             address start,
 584                             OopMapSet* oop_maps,
 585                             int& frame_complete,
 586                             int& frame_size_in_words,
 587                             bool alloc_inline_receiver) {
 588   if (requires_clinit_barrier) {
 589     assert(VM_Version::supports_fast_class_init_checks(), "sanity");
 590     Label L_skip_barrier;
 591 
 592     { // Bypass the barrier for non-static methods
 593       __ ldrh(rscratch1, Address(rmethod, Method::access_flags_offset()));
 594       __ andsw(zr, rscratch1, JVM_ACC_STATIC);
 595       __ br(Assembler::EQ, L_skip_barrier); // non-static
 596     }
 597 
 598     __ load_method_holder(rscratch2, rmethod);
 599     __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier);
 600     __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
 601 
 602     __ bind(L_skip_barrier);
 603     c2i_no_clinit_check_entry = __ pc();
 604   }
 605 
 606   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
 607   bs->c2i_entry_barrier(masm);
 608 
 609   // Before we get into the guts of the C2I adapter, see if we should be here
 610   // at all.  We've come from compiled code and are attempting to jump to the
 611   // interpreter, which means the caller made a static call to get here
 612   // (vcalls always get a compiled target if there is one).  Check for a
 613   // compiled target.  If there is one, we need to patch the caller's call.
 614   patch_callers_callsite(masm);
 615 
 616   __ bind(skip_fixup);
 617 
 618   // Name some registers to be used in the following code. We can use
 619   // anything except r0-r7 which are arguments in the Java calling
 620   // convention, rmethod (r12), and r19 which holds the outgoing sender
 621   // SP for the interpreter.
 622   Register buf_array = r10;   // Array of buffered inline types
 623   Register buf_oop = r11;     // Buffered inline type oop
 624   Register tmp1 = r15;
 625   Register tmp2 = r16;
 626   Register tmp3 = r17;
 627 
 628 #ifdef ASSERT
 629   RegSet clobbered_gp_regs = MacroAssembler::call_clobbered_gp_registers();
 630   assert(clobbered_gp_regs.contains(buf_array), "buf_array must be saved explicitly if it's not a clobber");
 631   assert(clobbered_gp_regs.contains(buf_oop), "buf_oop must be saved explicitly if it's not a clobber");
 632   assert(clobbered_gp_regs.contains(tmp1), "tmp1 must be saved explicitly if it's not a clobber");
 633   assert(clobbered_gp_regs.contains(tmp2), "tmp2 must be saved explicitly if it's not a clobber");
 634   assert(clobbered_gp_regs.contains(tmp3), "tmp3 must be saved explicitly if it's not a clobber");
 635 #endif
 636 
 637   if (InlineTypePassFieldsAsArgs) {
 638     // Is there an inline type argument?
 639     bool has_inline_argument = false;
 640     for (int i = 0; i < sig_extended->length() && !has_inline_argument; i++) {
 641       has_inline_argument = (sig_extended->at(i)._bt == T_METADATA);
 642     }
 643     if (has_inline_argument) {
 644       // There is at least a value type argument: we're coming from
 645       // compiled code so we may not have buffers to back the value
 646       // objects. Allocate the buffers here with a runtime call for
 647       // the value arguments that needs a buffer.
 648       RegisterSaver reg_save(true /* save_vectors */);
 649       OopMap* map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
 650 
 651       frame_complete = __ offset();
 652       address the_pc = __ pc();
 653 
 654       Label retaddr;
 655       __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
 656 
 657       __ mov(c_rarg0, rthread);
 658       __ mov(c_rarg1, rmethod);
 659       __ mov(c_rarg2, (int64_t)alloc_inline_receiver);
 660 
 661       __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::allocate_inline_types)));
 662       __ blr(rscratch1);
 663       __ bind(retaddr);



 664 
 665       oop_maps->add_gc_map(__ pc() - start, map);
 666       __ reset_last_Java_frame(false);














 667 
 668       reg_save.restore_live_registers(masm);
 669 
 670       Label no_exception;
 671       __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
 672       __ cbz(rscratch1, no_exception);
 673 
 674       __ str(zr, Address(rthread, JavaThread::vm_result_oop_offset()));
 675       __ ldr(r0, Address(rthread, Thread::pending_exception_offset()));
 676       __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
 677 
 678       __ bind(no_exception);
 679 
 680       // We get an array of objects from the runtime call
 681       __ get_vm_result_oop(buf_array, rthread);
 682     }
 683   }








 684 
 685   // Since all args are passed on the stack, total_args_passed *
 686   // Interpreter::stackElementSize is the space we need.
 687 
 688   int total_args_passed = compute_total_args_passed_int(sig_extended);
 689   int extraspace = total_args_passed * Interpreter::stackElementSize;
 690 
 691   // stack is aligned, keep it that way
 692   extraspace = align_up(extraspace, StackAlignmentInBytes);
 693 
 694   // set senderSP value
 695   __ mov(r19_sender_sp, sp);
 696 
 697   __ sub(sp, sp, extraspace);
 698 
 699   // Now write the args into the outgoing interpreter space
 700 
 701   // next_arg_comp is the next argument from the compiler point of
 702   // view (inline type fields are passed in registers/on the stack). In
 703   // sig_extended, an inline type argument starts with: T_METADATA,
 704   // followed by the types of the fields of the inline type and T_VOID
 705   // to mark the end of the inline type. ignored counts the number of
 706   // T_METADATA/T_VOID. next_vt_arg is the next inline type argument:
 707   // used to get the buffer for that argument from the pool of buffers
 708   // we allocated above and want to pass to the
 709   // interpreter. next_arg_int is the next argument from the
 710   // interpreter point of view (inline types are passed by reference).
 711   for (int next_arg_comp = 0, ignored = 0, next_vt_arg = 0, next_arg_int = 0;
 712        next_arg_comp < sig_extended->length(); next_arg_comp++) {
 713     assert(ignored <= next_arg_comp, "shouldn't skip over more slots than there are arguments");
 714     assert(next_arg_int <= total_args_passed, "more arguments for the interpreter than expected?");
 715     BasicType bt = sig_extended->at(next_arg_comp)._bt;
 716     int st_off = (total_args_passed - next_arg_int - 1) * Interpreter::stackElementSize;
 717     if (!InlineTypePassFieldsAsArgs || bt != T_METADATA) {
 718       int next_off = st_off - Interpreter::stackElementSize;
 719       const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : st_off;
 720       const VMRegPair reg_pair = regs[next_arg_comp-ignored];
 721       size_t size_in_bytes = reg_pair.second()->is_valid() ? 8 : 4;
 722       gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
 723                              size_in_bytes, reg_pair, Address(sp, offset), tmp1, tmp2, tmp3, extraspace, false);
 724       next_arg_int++;
 725 #ifdef ASSERT
 726       if (bt == T_LONG || bt == T_DOUBLE) {
 727         // Overwrite the unused slot with known junk
 728         __ mov(rscratch1, CONST64(0xdeadffffdeadaaaa));
 729         __ str(rscratch1, Address(sp, st_off));



 730       }















 731 #endif /* ASSERT */
 732     } else {
 733       ignored++;
 734       next_arg_int++;
 735       int vt = 1;
 736       // write fields we get from compiled code in registers/stack
 737       // slots to the buffer: we know we are done with that inline type
 738       // argument when we hit the T_VOID that acts as an end of inline
 739       // type delimiter for this inline type. Inline types are flattened
 740       // so we might encounter embedded inline types. Each entry in
 741       // sig_extended contains a field offset in the buffer.
 742       Label L_null;
 743       Label not_null_buffer;
 744       do {
 745         next_arg_comp++;
 746         BasicType bt = sig_extended->at(next_arg_comp)._bt;
 747         BasicType prev_bt = sig_extended->at(next_arg_comp - 1)._bt;
 748         if (bt == T_METADATA) {
 749           vt++;
 750           ignored++;
 751         } else if (bt == T_VOID && prev_bt != T_LONG && prev_bt != T_DOUBLE) {
 752           vt--;
 753           ignored++;
 754         } else if (sig_extended->at(next_arg_comp)._vt_oop) {
 755           VMReg buffer = regs[next_arg_comp-ignored].first();
 756           if (buffer->is_stack()) {
 757             int ld_off = buffer->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
 758             __ ldr(buf_oop, Address(sp, ld_off));
 759           } else {
 760             __ mov(buf_oop, buffer->as_Register());
 761           }
 762           __ cbnz(buf_oop, not_null_buffer);
 763           // get the buffer from the just allocated pool of buffers
 764           int index = arrayOopDesc::base_offset_in_bytes(T_OBJECT) + next_vt_arg * type2aelembytes(T_OBJECT);
 765           __ load_heap_oop(buf_oop, Address(buf_array, index), rscratch1, tmp2);
 766           next_vt_arg++;
 767         } else {
 768           int off = sig_extended->at(next_arg_comp)._offset;
 769           if (off == -1) {
 770             // Nullable inline type argument, emit null check
 771             VMReg reg = regs[next_arg_comp-ignored].first();
 772             Label L_notNull;
 773             if (reg->is_stack()) {
 774               int ld_off = reg->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
 775               __ ldrb(tmp1, Address(sp, ld_off));
 776               __ cbnz(tmp1, L_notNull);
 777             } else {
 778               __ cbnz(reg->as_Register(), L_notNull);
 779             }
 780             __ str(zr, Address(sp, st_off));
 781             __ b(L_null);
 782             __ bind(L_notNull);
 783             continue;
 784           }
 785           assert(off > 0, "offset in object should be positive");
 786           size_t size_in_bytes = is_java_primitive(bt) ? type2aelembytes(bt) : wordSize;
 787           bool is_oop = is_reference_type(bt);
 788           gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
 789                                  size_in_bytes, regs[next_arg_comp-ignored], Address(buf_oop, off), tmp1, tmp2, tmp3, extraspace, is_oop);
 790         }
 791       } while (vt != 0);
 792       // pass the buffer to the interpreter
 793       __ bind(not_null_buffer);
 794       __ str(buf_oop, Address(sp, st_off));
 795       __ bind(L_null);









 796     }
 797   }
 798 
 799   __ mov(esp, sp); // Interp expects args on caller's expression stack
 800 
 801   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::interpreter_entry_offset())));
 802   __ br(rscratch1);
 803 }
 804 
 805 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, int comp_args_on_stack, const GrowableArray<SigEntry>* sig, const VMRegPair *regs) {
 806 





 807 
 808   // Note: r19_sender_sp contains the senderSP on entry. We must
 809   // preserve it since we may do a i2c -> c2i transition if we lose a
 810   // race where compiled code goes non-entrant while we get args
 811   // ready.
 812 
 813   // Adapters are frameless.
 814 
 815   // An i2c adapter is frameless because the *caller* frame, which is
 816   // interpreted, routinely repairs its own esp (from
 817   // interpreter_frame_last_sp), even if a callee has modified the
 818   // stack pointer.  It also recalculates and aligns sp.
 819 
 820   // A c2i adapter is frameless because the *callee* frame, which is
 821   // interpreted, routinely repairs its caller's sp (from sender_sp,
 822   // which is set up via the senderSP register).
 823 
 824   // In other words, if *either* the caller or callee is interpreted, we can
 825   // get the stack pointer repaired after a call.
 826 
 827   // This is why c2i and i2c adapters cannot be indefinitely composed.
 828   // In particular, if a c2i adapter were to somehow call an i2c adapter,
 829   // both caller and callee would be compiled methods, and neither would
 830   // clean up the stack pointer changes performed by the two adapters.
 831   // If this happens, control eventually transfers back to the compiled
 832   // caller, but with an uncorrected stack, causing delayed havoc.
 833 
 834   // Cut-out for having no stack args.
 835   int comp_words_on_stack = 0;
 836   if (comp_args_on_stack) {
 837      comp_words_on_stack = align_up(comp_args_on_stack * VMRegImpl::stack_slot_size, wordSize) >> LogBytesPerWord;
 838      __ sub(rscratch1, sp, comp_words_on_stack * wordSize);
 839      __ andr(sp, rscratch1, -16);
 840   }
 841 
 842   // Will jump to the compiled code just as if compiled code was doing it.
 843   // Pre-load the register-jump target early, to schedule it better.
 844   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::from_compiled_inline_offset())));
 845 
 846   int total_args_passed = sig->length();
 847 
 848   // Now generate the shuffle code.
 849   for (int i = 0; i < total_args_passed; i++) {
 850     BasicType bt = sig->at(i)._bt;
 851     if (bt == T_VOID) {
 852       assert(i > 0 && (sig->at(i - 1)._bt == T_LONG || sig->at(i - 1)._bt == T_DOUBLE), "missing half");
 853       continue;
 854     }
 855 
 856     // Pick up 0, 1 or 2 words from SP+offset.
 857     assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), "scrambled load targets?");
 858 


 859     // Load in argument order going down.
 860     int ld_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
 861     // Point to interpreter value (vs. tag)
 862     int next_off = ld_off - Interpreter::stackElementSize;
 863     //
 864     //
 865     //
 866     VMReg r_1 = regs[i].first();
 867     VMReg r_2 = regs[i].second();
 868     if (!r_1->is_valid()) {
 869       assert(!r_2->is_valid(), "");
 870       continue;
 871     }
 872     if (r_1->is_stack()) {
 873       // Convert stack slot to an SP offset (+ wordSize to account for return address )
 874       int st_off = regs[i].first()->reg2stack() * VMRegImpl::stack_slot_size;
 875       if (!r_2->is_valid()) {
 876         // sign extend???
 877         __ ldrsw(rscratch2, Address(esp, ld_off));
 878         __ str(rscratch2, Address(sp, st_off));
 879       } else {
 880         //
 881         // We are using two optoregs. This can be either T_OBJECT,
 882         // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
 883         // two slots but only uses one for thr T_LONG or T_DOUBLE case
 884         // So we must adjust where to pick up the data to match the
 885         // interpreter.
 886         //
 887         // Interpreter local[n] == MSW, local[n+1] == LSW however locals
 888         // are accessed as negative so LSW is at LOW address
 889 
 890         // ld_off is MSW so get LSW
 891         const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : ld_off;

 892         __ ldr(rscratch2, Address(esp, offset));
 893         // st_off is LSW (i.e. reg.first())
 894          __ str(rscratch2, Address(sp, st_off));
 895        }
 896      } else if (r_1->is_Register()) {  // Register argument
 897        Register r = r_1->as_Register();
 898        if (r_2->is_valid()) {
 899          //
 900          // We are using two VMRegs. This can be either T_OBJECT,
 901          // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
 902          // two slots but only uses one for thr T_LONG or T_DOUBLE case
 903          // So we must adjust where to pick up the data to match the
 904          // interpreter.
 905 
 906         const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : ld_off;
 907 
 908          // this can be a misaligned move
 909          __ ldr(r, Address(esp, offset));
 910        } else {
 911          // sign extend and use a full word?
 912          __ ldrw(r, Address(esp, ld_off));
 913        }
 914      } else {
 915        if (!r_2->is_valid()) {
 916          __ ldrs(r_1->as_FloatRegister(), Address(esp, ld_off));
 917        } else {
 918          __ ldrd(r_1->as_FloatRegister(), Address(esp, next_off));
 919        }
 920      }
 921    }

 922 
 923   __ mov(rscratch2, rscratch1);
 924   __ push_cont_fastpath(rthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about; kills rscratch1
 925   __ mov(rscratch1, rscratch2);
 926 
 927   // 6243940 We might end up in handle_wrong_method if
 928   // the callee is deoptimized as we race thru here. If that
 929   // happens we don't want to take a safepoint because the
 930   // caller frame will look interpreted and arguments are now
 931   // "compiled" so it is much better to make this transition
 932   // invisible to the stack walking code. Unfortunately if
 933   // we try and find the callee by normal means a safepoint
 934   // is possible. So we stash the desired callee in the thread
 935   // and the vm will find there should this case occur.
 936 
 937   __ str(rmethod, Address(rthread, JavaThread::callee_target_offset()));
 938 
 939   __ br(rscratch1);
 940 }
 941 
 942 static void gen_inline_cache_check(MacroAssembler *masm, Label& skip_fixup) {
 943   Register data = rscratch2;
 944   __ ic_check(1 /* end_alignment */);
 945   __ ldr(rmethod, Address(data, CompiledICData::speculated_method_offset()));




 946 
 947   // Method might have been compiled since the call site was patched to
 948   // interpreted; if that is the case treat it as a miss so we can get
 949   // the call site corrected.
 950   __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset())));
 951   __ cbz(rscratch1, skip_fixup);
 952   __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
 953 }
 954 
 955 // ---------------------------------------------------------------
 956 void SharedRuntime::generate_i2c2i_adapters(MacroAssembler* masm,
 957                                             int comp_args_on_stack,
 958                                             const GrowableArray<SigEntry>* sig,
 959                                             const VMRegPair* regs,
 960                                             const GrowableArray<SigEntry>* sig_cc,
 961                                             const VMRegPair* regs_cc,
 962                                             const GrowableArray<SigEntry>* sig_cc_ro,
 963                                             const VMRegPair* regs_cc_ro,
 964                                             address entry_address[AdapterBlob::ENTRY_COUNT],
 965                                             AdapterBlob*& new_adapter,
 966                                             bool allocate_code_blob) {
 967 
 968   entry_address[AdapterBlob::I2C] = __ pc();
 969   gen_i2c_adapter(masm, comp_args_on_stack, sig, regs);

 970 
 971   // -------------------------------------------------------------------------
 972   // Generate a C2I adapter.  On entry we know rmethod holds the Method* during calls
 973   // to the interpreter.  The args start out packed in the compiled layout.  They
 974   // need to be unpacked into the interpreter layout.  This will almost always
 975   // require some stack space.  We grow the current (compiled) stack, then repack
 976   // the args.  We  finally end in a jump to the generic interpreter entry point.
 977   // On exit from the interpreter, the interpreter will restore our SP (lest the
 978   // compiled code, which relies solely on SP and not FP, get sick).
 979 
 980   entry_address[AdapterBlob::C2I_Unverified] = __ pc();
 981   entry_address[AdapterBlob::C2I_Unverified_Inline] = __ pc();
 982   Label skip_fixup;




 983 
 984   gen_inline_cache_check(masm, skip_fixup);




 985 
 986   OopMapSet* oop_maps = new OopMapSet();
 987   int frame_complete = CodeOffsets::frame_never_safe;
 988   int frame_size_in_words = 0;
 989 
 990   // Scalarized c2i adapter with non-scalarized receiver (i.e., don't pack receiver)
 991   entry_address[AdapterBlob::C2I_No_Clinit_Check] = nullptr;
 992   entry_address[AdapterBlob::C2I_Inline_RO] = __ pc();
 993   if (regs_cc != regs_cc_ro) {
 994     // No class init barrier needed because method is guaranteed to be non-static
 995     gen_c2i_adapter(masm, sig_cc_ro, regs_cc_ro, /* requires_clinit_barrier = */ false, entry_address[AdapterBlob::C2I_No_Clinit_Check],
 996                     skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
 997     skip_fixup.reset();
 998   }







 999 
1000   // Scalarized c2i adapter
1001   entry_address[AdapterBlob::C2I]        = __ pc();
1002   entry_address[AdapterBlob::C2I_Inline] = __ pc();
1003   gen_c2i_adapter(masm, sig_cc, regs_cc, /* requires_clinit_barrier = */ true, entry_address[AdapterBlob::C2I_No_Clinit_Check],
1004                   skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ true);
1005 
1006   // Non-scalarized c2i adapter
1007   if (regs != regs_cc) {
1008     entry_address[AdapterBlob::C2I_Unverified_Inline] = __ pc();
1009     Label inline_entry_skip_fixup;
1010     gen_inline_cache_check(masm, inline_entry_skip_fixup);
1011 
1012     entry_address[AdapterBlob::C2I_Inline] = __ pc();
1013     gen_c2i_adapter(masm, sig, regs, /* requires_clinit_barrier = */ true, entry_address[AdapterBlob::C2I_No_Clinit_Check],
1014                     inline_entry_skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
1015   }
1016 
1017   // The c2i adapters might safepoint and trigger a GC. The caller must make sure that
1018   // the GC knows about the location of oop argument locations passed to the c2i adapter.
1019   if (allocate_code_blob) {
1020     bool caller_must_gc_arguments = (regs != regs_cc);
1021     int entry_offset[AdapterHandlerEntry::ENTRIES_COUNT];
1022     assert(AdapterHandlerEntry::ENTRIES_COUNT == 7, "sanity");
1023     AdapterHandlerLibrary::address_to_offset(entry_address, entry_offset);
1024     new_adapter = AdapterBlob::create(masm->code(), entry_offset, frame_complete, frame_size_in_words, oop_maps, caller_must_gc_arguments);
1025   }
1026 }
1027 
1028 static int c_calling_convention_priv(const BasicType *sig_bt,
1029                                          VMRegPair *regs,
1030                                          int total_args_passed) {
1031 
1032 // We return the amount of VMRegImpl stack slots we need to reserve for all
1033 // the arguments NOT counting out_preserve_stack_slots.
1034 
1035     static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1036       c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5,  c_rarg6,  c_rarg7
1037     };
1038     static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
1039       c_farg0, c_farg1, c_farg2, c_farg3,
1040       c_farg4, c_farg5, c_farg6, c_farg7
1041     };
1042 
1043     uint int_args = 0;
1044     uint fp_args = 0;
1045     uint stk_args = 0; // inc by 2 each time

2904 
2905   // exception pending => remove activation and forward to exception handler
2906 
2907   __ str(zr, Address(rthread, JavaThread::vm_result_oop_offset()));
2908 
2909   __ ldr(r0, Address(rthread, Thread::pending_exception_offset()));
2910   __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2911 
2912   // -------------
2913   // make sure all code is generated
2914   masm->flush();
2915 
2916   // return the  blob
2917   // frame_size_words or bytes??
2918   RuntimeStub* rs_blob = RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
2919 
2920   AOTCodeCache::store_code_blob(*rs_blob, AOTCodeEntry::SharedBlob, StubInfo::blob(id));
2921   return rs_blob;
2922 }
2923 
2924 BufferedInlineTypeBlob* SharedRuntime::generate_buffered_inline_type_adapter(const InlineKlass* vk) {
2925   BufferBlob* buf = BufferBlob::create("inline types pack/unpack", 16 * K);
2926   if (buf == nullptr) {
2927     return nullptr;
2928   }
2929   CodeBuffer buffer(buf);
2930   short buffer_locs[20];
2931   buffer.insts()->initialize_shared_locs((relocInfo*)buffer_locs,
2932                                          sizeof(buffer_locs)/sizeof(relocInfo));
2933 
2934   MacroAssembler _masm(&buffer);
2935   MacroAssembler* masm = &_masm;
2936 
2937   const Array<SigEntry>* sig_vk = vk->extended_sig();
2938   const Array<VMRegPair>* regs = vk->return_regs();
2939 
2940   int pack_fields_jobject_off = __ offset();
2941   // Resolve pre-allocated buffer from JNI handle.
2942   // We cannot do this in generate_call_stub() because it requires GC code to be initialized.
2943   Register Rresult = r14;  // See StubGenerator::generate_call_stub().
2944   __ ldr(r0, Address(Rresult));
2945   __ resolve_jobject(r0 /* value */,
2946                      rthread /* thread */,
2947                      r12 /* tmp */);
2948   __ str(r0, Address(Rresult));
2949 
2950   int pack_fields_off = __ offset();
2951 
2952   int j = 1;
2953   for (int i = 0; i < sig_vk->length(); i++) {
2954     BasicType bt = sig_vk->at(i)._bt;
2955     if (bt == T_METADATA) {
2956       continue;
2957     }
2958     if (bt == T_VOID) {
2959       if (sig_vk->at(i-1)._bt == T_LONG ||
2960           sig_vk->at(i-1)._bt == T_DOUBLE) {
2961         j++;
2962       }
2963       continue;
2964     }
2965     int off = sig_vk->at(i)._offset;
2966     VMRegPair pair = regs->at(j);
2967     VMReg r_1 = pair.first();
2968     VMReg r_2 = pair.second();
2969     Address to(r0, off);
2970     if (bt == T_FLOAT) {
2971       __ strs(r_1->as_FloatRegister(), to);
2972     } else if (bt == T_DOUBLE) {
2973       __ strd(r_1->as_FloatRegister(), to);
2974     } else {
2975       Register val = r_1->as_Register();
2976       assert_different_registers(to.base(), val, r15, r16, r17);
2977       if (is_reference_type(bt)) {
2978         // store_heap_oop transitively calls oop_store_at which corrupts to.base(). We need to keep r0 valid.
2979         __ mov(r17, r0);
2980         Address to_with_r17(r17, off);
2981         __ store_heap_oop(to_with_r17, val, r15, r16, r17, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
2982       } else {
2983         __ store_sized_value(to, r_1->as_Register(), type2aelembytes(bt));
2984       }
2985     }
2986     j++;
2987   }
2988   assert(j == regs->length(), "missed a field?");
2989   if (vk->supports_nullable_layouts()) {
2990     // Zero the null marker (setting it to 1 would be better but would require an additional register)
2991     __ strb(zr, Address(r0, vk->null_marker_offset()));
2992   }
2993   __ ret(lr);
2994 
2995   int unpack_fields_off = __ offset();
2996 
2997   Label skip;
2998   Label not_null;
2999   __ cbnz(r0, not_null);
3000 
3001   // Return value is null. Zero all registers because the runtime requires a canonical
3002   // representation of a flat null.
3003   j = 1;
3004   for (int i = 0; i < sig_vk->length(); i++) {
3005     BasicType bt = sig_vk->at(i)._bt;
3006     if (bt == T_METADATA) {
3007       continue;
3008     }
3009     if (bt == T_VOID) {
3010       if (sig_vk->at(i-1)._bt == T_LONG ||
3011           sig_vk->at(i-1)._bt == T_DOUBLE) {
3012         j++;
3013       }
3014       continue;
3015     }
3016 
3017     VMRegPair pair = regs->at(j);
3018     VMReg r_1 = pair.first();
3019     if (r_1->is_FloatRegister()) {
3020       __ mov(r_1->as_FloatRegister(), Assembler::T2S, 0);
3021     } else {
3022       __ mov(r_1->as_Register(), zr);
3023     }
3024     j++;
3025   }
3026   __ b(skip);
3027   __ bind(not_null);
3028 
3029   j = 1;
3030   for (int i = 0; i < sig_vk->length(); i++) {
3031     BasicType bt = sig_vk->at(i)._bt;
3032     if (bt == T_METADATA) {
3033       continue;
3034     }
3035     if (bt == T_VOID) {
3036       if (sig_vk->at(i-1)._bt == T_LONG ||
3037           sig_vk->at(i-1)._bt == T_DOUBLE) {
3038         j++;
3039       }
3040       continue;
3041     }
3042     int off = sig_vk->at(i)._offset;
3043     assert(off > 0, "offset in object should be positive");
3044     VMRegPair pair = regs->at(j);
3045     VMReg r_1 = pair.first();
3046     VMReg r_2 = pair.second();
3047     Address from(r0, off);
3048     if (bt == T_FLOAT) {
3049       __ ldrs(r_1->as_FloatRegister(), from);
3050     } else if (bt == T_DOUBLE) {
3051       __ ldrd(r_1->as_FloatRegister(), from);
3052     } else if (bt == T_OBJECT || bt == T_ARRAY) {
3053       assert_different_registers(r0, r_1->as_Register());
3054       __ load_heap_oop(r_1->as_Register(), from, rscratch1, rscratch2);
3055     } else {
3056       assert(is_java_primitive(bt), "unexpected basic type");
3057       assert_different_registers(r0, r_1->as_Register());
3058       size_t size_in_bytes = type2aelembytes(bt);
3059       __ load_sized_value(r_1->as_Register(), from, size_in_bytes, bt != T_CHAR && bt != T_BOOLEAN);
3060     }
3061     j++;
3062   }
3063   assert(j == regs->length(), "missed a field?");
3064 
3065   __ bind(skip);
3066 
3067   __ ret(lr);
3068 
3069   __ flush();
3070 
3071   return BufferedInlineTypeBlob::create(&buffer, pack_fields_off, pack_fields_jobject_off, unpack_fields_off);
3072 }
3073 
3074 // Continuation point for throwing of implicit exceptions that are
3075 // not handled in the current activation. Fabricates an exception
3076 // oop and initiates normal exception dispatching in this
3077 // frame. Since we need to preserve callee-saved values (currently
3078 // only for C2, but done for C1 as well) we need a callee-saved oop
3079 // map and therefore have to make these stubs into RuntimeStubs
3080 // rather than BufferBlobs.  If the compiler needs all registers to
3081 // be preserved between the fault point and the exception handler
3082 // then it must assume responsibility for that in
3083 // AbstractCompiler::continuation_for_implicit_null_exception or
3084 // continuation_for_implicit_division_by_zero_exception. All other
3085 // implicit exceptions (e.g., NullPointerException or
3086 // AbstractMethodError on entry) are either at call sites or
3087 // otherwise assume that stack unwinding will be initiated, so
3088 // caller saved registers were assumed volatile in the compiler.
3089 
3090 RuntimeStub* SharedRuntime::generate_throw_exception(StubId id, address runtime_entry) {
3091   assert(is_throw_id(id), "expected a throw stub id");
3092 
3093   const char* name = SharedRuntime::stub_name(id);
< prev index next >