< prev index next >

src/hotspot/cpu/x86/sharedRuntime_x86_64.cpp

Print this page

   1 /*
   2  * Copyright (c) 2003, 2025, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #ifndef _WINDOWS
  26 #include "alloca.h"
  27 #endif
  28 #include "asm/macroAssembler.hpp"
  29 #include "asm/macroAssembler.inline.hpp"

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

 617       break;
 618     case T_DOUBLE:
 619       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 620       if (fp_args < Argument::n_float_register_parameters_j) {
 621         regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
 622       } else {
 623         stk_args = align_up(stk_args, 2);
 624         regs[i].set2(VMRegImpl::stack2reg(stk_args));
 625         stk_args += 2;
 626       }
 627       break;
 628     default:
 629       ShouldNotReachHere();
 630       break;
 631     }
 632   }
 633 
 634   return stk_args;
 635 }
 636 

















































































 637 // Patch the callers callsite with entry to compiled code if it exists.
 638 static void patch_callers_callsite(MacroAssembler *masm) {
 639   Label L;
 640   __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), NULL_WORD);
 641   __ jcc(Assembler::equal, L);
 642 
 643   // Save the current stack pointer
 644   __ mov(r13, rsp);
 645   // Schedule the branch target address early.
 646   // Call into the VM to patch the caller, then jump to compiled callee
 647   // rax isn't live so capture return address while we easily can
 648   __ movptr(rax, Address(rsp, 0));
 649 
 650   // align stack so push_CPU_state doesn't fault
 651   __ andptr(rsp, -(StackAlignmentInBytes));
 652   __ push_CPU_state();
 653   __ vzeroupper();
 654   // VM needs caller's callsite
 655   // VM needs target method
 656   // This needs to be a long call since we will relocate this adapter to

 659   // Allocate argument register save area
 660   if (frame::arg_reg_save_area_bytes != 0) {
 661     __ subptr(rsp, frame::arg_reg_save_area_bytes);
 662   }
 663   __ mov(c_rarg0, rbx);
 664   __ mov(c_rarg1, rax);
 665   __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)));
 666 
 667   // De-allocate argument register save area
 668   if (frame::arg_reg_save_area_bytes != 0) {
 669     __ addptr(rsp, frame::arg_reg_save_area_bytes);
 670   }
 671 
 672   __ vzeroupper();
 673   __ pop_CPU_state();
 674   // restore sp
 675   __ mov(rsp, r13);
 676   __ bind(L);
 677 }
 678 













































































































 679 static void gen_c2i_adapter(MacroAssembler *masm,
 680                             int total_args_passed,
 681                             int comp_args_on_stack,
 682                             const BasicType *sig_bt,
 683                             const VMRegPair *regs,
 684                             Label& skip_fixup) {































 685   // Before we get into the guts of the C2I adapter, see if we should be here
 686   // at all.  We've come from compiled code and are attempting to jump to the
 687   // interpreter, which means the caller made a static call to get here
 688   // (vcalls always get a compiled target if there is one).  Check for a
 689   // compiled target.  If there is one, we need to patch the caller's call.
 690   patch_callers_callsite(masm);
 691 
 692   __ bind(skip_fixup);
 693 










































 694   // Since all args are passed on the stack, total_args_passed *
 695   // Interpreter::stackElementSize is the space we need.
 696 
 697   assert(total_args_passed >= 0, "total_args_passed is %d", total_args_passed);
 698 
 699   int extraspace = (total_args_passed * Interpreter::stackElementSize);
 700 
 701   // stack is aligned, keep it that way
 702   // This is not currently needed or enforced by the interpreter, but
 703   // we might as well conform to the ABI.
 704   extraspace = align_up(extraspace, 2*wordSize);
 705 
 706   // set senderSP value
 707   __ lea(r13, Address(rsp, wordSize));
 708 
 709 #ifdef ASSERT
 710   __ check_stack_alignment(r13, "sender stack not aligned");
 711 #endif
 712   if (extraspace > 0) {
 713     // Pop the return address
 714     __ pop(rax);
 715 
 716     __ subptr(rsp, extraspace);
 717 
 718     // Push the return address
 719     __ push(rax);
 720 
 721     // Account for the return address location since we store it first rather
 722     // than hold it in a register across all the shuffling
 723     extraspace += wordSize;
 724   }
 725 
 726 #ifdef ASSERT
 727   __ check_stack_alignment(rsp, "callee stack not aligned", wordSize, rax);
 728 #endif
 729 
 730   // Now write the args into the outgoing interpreter space
 731   for (int i = 0; i < total_args_passed; i++) {
 732     if (sig_bt[i] == T_VOID) {
 733       assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
 734       continue;
 735     }
 736 
 737     // offset to start parameters
 738     int st_off   = (total_args_passed - i) * Interpreter::stackElementSize;
 739     int next_off = st_off - Interpreter::stackElementSize;
 740 
 741     // Say 4 args:
 742     // i   st_off
 743     // 0   32 T_LONG
 744     // 1   24 T_VOID
 745     // 2   16 T_OBJECT
 746     // 3    8 T_BOOL
 747     // -    0 return address
 748     //
 749     // However to make thing extra confusing. Because we can fit a long/double in
 750     // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
 751     // leaves one slot empty and only stores to a single slot. In this case the
 752     // slot that is occupied is the T_VOID slot. See I said it was confusing.
 753 
 754     VMReg r_1 = regs[i].first();
 755     VMReg r_2 = regs[i].second();
 756     if (!r_1->is_valid()) {
 757       assert(!r_2->is_valid(), "");
 758       continue;
 759     }
 760     if (r_1->is_stack()) {
 761       // memory to memory use rax
 762       int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
 763       if (!r_2->is_valid()) {
 764         // sign extend??
 765         __ movl(rax, Address(rsp, ld_off));
 766         __ movptr(Address(rsp, st_off), rax);
 767 
 768       } else {
 769 
 770         __ movq(rax, Address(rsp, ld_off));
 771 
 772         // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
 773         // T_DOUBLE and T_LONG use two slots in the interpreter
 774         if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
 775           // ld_off == LSW, ld_off+wordSize == MSW
 776           // st_off == MSW, next_off == LSW
 777           __ movq(Address(rsp, next_off), rax);














 778 #ifdef ASSERT
 779           // Overwrite the unused slot with known junk
 780           __ mov64(rax, CONST64(0xdeadffffdeadaaaa));
 781           __ movptr(Address(rsp, st_off), rax);
 782 #endif /* ASSERT */
 783         } else {
 784           __ movq(Address(rsp, st_off), rax);
 785         }
 786       }
 787     } else if (r_1->is_Register()) {
 788       Register r = r_1->as_Register();
 789       if (!r_2->is_valid()) {
 790         // must be only an int (or less ) so move only 32bits to slot
 791         // why not sign extend??
 792         __ movl(Address(rsp, st_off), r);
 793       } else {
 794         // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
 795         // T_DOUBLE and T_LONG use two slots in the interpreter
 796         if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
 797           // long/double in gpr
 798 #ifdef ASSERT
 799           // Overwrite the unused slot with known junk
 800           __ mov64(rax, CONST64(0xdeadffffdeadaaab));
 801           __ movptr(Address(rsp, st_off), rax);
 802 #endif /* ASSERT */
 803           __ movq(Address(rsp, next_off), r);

























 804         } else {
 805           __ movptr(Address(rsp, st_off), r);





















 806         }
 807       }
 808     } else {
 809       assert(r_1->is_XMMRegister(), "");
 810       if (!r_2->is_valid()) {
 811         // only a float use just part of the slot
 812         __ movflt(Address(rsp, st_off), r_1->as_XMMRegister());
 813       } else {
 814 #ifdef ASSERT
 815         // Overwrite the unused slot with known junk
 816         __ mov64(rax, CONST64(0xdeadffffdeadaaac));
 817         __ movptr(Address(rsp, st_off), rax);
 818 #endif /* ASSERT */
 819         __ movdbl(Address(rsp, next_off), r_1->as_XMMRegister());
 820       }
 821     }
 822   }
 823 
 824   // Schedule the branch target address early.
 825   __ movptr(rcx, Address(rbx, in_bytes(Method::interpreter_entry_offset())));
 826   __ jmp(rcx);
 827 }
 828 
 829 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
 830                                     int total_args_passed,
 831                                     int comp_args_on_stack,
 832                                     const BasicType *sig_bt,
 833                                     const VMRegPair *regs) {
 834 
 835   // Note: r13 contains the senderSP on entry. We must preserve it since
 836   // we may do a i2c -> c2i transition if we lose a race where compiled
 837   // code goes non-entrant while we get args ready.
 838   // In addition we use r13 to locate all the interpreter args as
 839   // we must align the stack to 16 bytes on an i2c entry else we
 840   // lose alignment we expect in all compiled code and register
 841   // save code can segv when fxsave instructions find improperly
 842   // aligned stack pointer.
 843 
 844   // Adapters can be frameless because they do not require the caller
 845   // to perform additional cleanup work, such as correcting the stack pointer.
 846   // An i2c adapter is frameless because the *caller* frame, which is interpreted,
 847   // routinely repairs its own stack pointer (from interpreter_frame_last_sp),
 848   // even if a callee has modified the stack pointer.
 849   // A c2i adapter is frameless because the *callee* frame, which is interpreted,
 850   // routinely repairs its caller's stack pointer (from sender_sp, which is set
 851   // up via the senderSP register).
 852   // In other words, if *either* the caller or callee is interpreted, we can

 868   // Convert 4-byte c2 stack slots to words.
 869   int comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
 870 
 871   if (comp_args_on_stack) {
 872     __ subptr(rsp, comp_words_on_stack * wordSize);
 873   }
 874 
 875   // Ensure compiled code always sees stack at proper alignment
 876   __ andptr(rsp, -16);
 877 
 878   // push the return address and misalign the stack that youngest frame always sees
 879   // as far as the placement of the call instruction
 880   __ push(rax);
 881 
 882   // Put saved SP in another register
 883   const Register saved_sp = rax;
 884   __ movptr(saved_sp, r11);
 885 
 886   // Will jump to the compiled code just as if compiled code was doing it.
 887   // Pre-load the register-jump target early, to schedule it better.
 888   __ movptr(r11, Address(rbx, in_bytes(Method::from_compiled_offset())));
 889 
 890 #if INCLUDE_JVMCI
 891   if (EnableJVMCI) {
 892     // check if this call should be routed towards a specific entry point
 893     __ cmpptr(Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())), 0);
 894     Label no_alternative_target;
 895     __ jcc(Assembler::equal, no_alternative_target);
 896     __ movptr(r11, Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
 897     __ movptr(Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())), 0);
 898     __ bind(no_alternative_target);
 899   }
 900 #endif // INCLUDE_JVMCI
 901 


 902   // Now generate the shuffle code.  Pick up all register args and move the
 903   // rest through the floating point stack top.
 904   for (int i = 0; i < total_args_passed; i++) {
 905     if (sig_bt[i] == T_VOID) {

 906       // Longs and doubles are passed in native word order, but misaligned
 907       // in the 32-bit build.
 908       assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");

 909       continue;
 910     }
 911 
 912     // Pick up 0, 1 or 2 words from SP+offset.
 913 
 914     assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
 915             "scrambled load targets?");
 916     // Load in argument order going down.
 917     int ld_off = (total_args_passed - i)*Interpreter::stackElementSize;
 918     // Point to interpreter value (vs. tag)
 919     int next_off = ld_off - Interpreter::stackElementSize;
 920     //
 921     //
 922     //
 923     VMReg r_1 = regs[i].first();
 924     VMReg r_2 = regs[i].second();
 925     if (!r_1->is_valid()) {
 926       assert(!r_2->is_valid(), "");
 927       continue;
 928     }

 930       // Convert stack slot to an SP offset (+ wordSize to account for return address )
 931       int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size + wordSize;
 932 
 933       // We can use r13 as a temp here because compiled code doesn't need r13 as an input
 934       // and if we end up going thru a c2i because of a miss a reasonable value of r13
 935       // will be generated.
 936       if (!r_2->is_valid()) {
 937         // sign extend???
 938         __ movl(r13, Address(saved_sp, ld_off));
 939         __ movptr(Address(rsp, st_off), r13);
 940       } else {
 941         //
 942         // We are using two optoregs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
 943         // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
 944         // So we must adjust where to pick up the data to match the interpreter.
 945         //
 946         // Interpreter local[n] == MSW, local[n+1] == LSW however locals
 947         // are accessed as negative so LSW is at LOW address
 948 
 949         // ld_off is MSW so get LSW
 950         const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)?
 951                            next_off : ld_off;
 952         __ movq(r13, Address(saved_sp, offset));
 953         // st_off is LSW (i.e. reg.first())
 954         __ movq(Address(rsp, st_off), r13);
 955       }
 956     } else if (r_1->is_Register()) {  // Register argument
 957       Register r = r_1->as_Register();
 958       assert(r != rax, "must be different");
 959       if (r_2->is_valid()) {
 960         //
 961         // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
 962         // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
 963         // So we must adjust where to pick up the data to match the interpreter.
 964 
 965         const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)?
 966                            next_off : ld_off;
 967 
 968         // this can be a misaligned move
 969         __ movq(r, Address(saved_sp, offset));
 970       } else {
 971         // sign extend and use a full word?
 972         __ movl(r, Address(saved_sp, ld_off));
 973       }
 974     } else {
 975       if (!r_2->is_valid()) {
 976         __ movflt(r_1->as_XMMRegister(), Address(saved_sp, ld_off));
 977       } else {
 978         __ movdbl(r_1->as_XMMRegister(), Address(saved_sp, next_off));
 979       }
 980     }
 981   }
 982 
 983   __ push_cont_fastpath(); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
 984 
 985   // 6243940 We might end up in handle_wrong_method if
 986   // the callee is deoptimized as we race thru here. If that
 987   // happens we don't want to take a safepoint because the
 988   // caller frame will look interpreted and arguments are now
 989   // "compiled" so it is much better to make this transition
 990   // invisible to the stack walking code. Unfortunately if
 991   // we try and find the callee by normal means a safepoint
 992   // is possible. So we stash the desired callee in the thread
 993   // and the vm will find there should this case occur.
 994 
 995   __ movptr(Address(r15_thread, JavaThread::callee_target_offset()), rbx);
 996 
 997   // put Method* where a c2i would expect should we end up there
 998   // only needed because eof c2 resolve stubs return Method* as a result in
 999   // rax
1000   __ mov(rax, rbx);
1001   __ jmp(r11);
1002 }
1003 













1004 // ---------------------------------------------------------------
1005 void SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
1006                                             int total_args_passed,
1007                                             int comp_args_on_stack,
1008                                             const BasicType *sig_bt,
1009                                             const VMRegPair *regs,
1010                                             address entry_address[AdapterBlob::ENTRY_COUNT]) {






1011   entry_address[AdapterBlob::I2C] = __ pc();
1012 
1013   gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
1014 
1015   // -------------------------------------------------------------------------
1016   // Generate a C2I adapter.  On entry we know rbx holds the Method* during calls
1017   // to the interpreter.  The args start out packed in the compiled layout.  They
1018   // need to be unpacked into the interpreter layout.  This will almost always
1019   // require some stack space.  We grow the current (compiled) stack, then repack
1020   // the args.  We  finally end in a jump to the generic interpreter entry point.
1021   // On exit from the interpreter, the interpreter will restore our SP (lest the
1022   // compiled code, which relies solely on SP and not RBP, get sick).
1023 
1024   entry_address[AdapterBlob::C2I_Unverified] = __ pc();

1025   Label skip_fixup;
1026 
1027   Register data = rax;
1028   Register receiver = j_rarg0;
1029   Register temp = rbx;
1030 
1031   {
1032     __ ic_check(1 /* end_alignment */);
1033     __ movptr(rbx, Address(data, CompiledICData::speculated_method_offset()));
1034     // Method might have been compiled since the call site was patched to
1035     // interpreted if that is the case treat it as a miss so we can get
1036     // the call site corrected.
1037     __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), NULL_WORD);
1038     __ jcc(Assembler::equal, skip_fixup);
1039     __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1040   }
1041 
1042   entry_address[AdapterBlob::C2I] = __ pc();
1043 
1044   // Class initialization barrier for static methods
1045   entry_address[AdapterBlob::C2I_No_Clinit_Check] = nullptr;
1046   if (VM_Version::supports_fast_class_init_checks()) {
1047     Label L_skip_barrier;
1048     Register method = rbx;
1049 
1050     { // Bypass the barrier for non-static methods
1051       Register flags = rscratch1;
1052       __ load_unsigned_short(flags, Address(method, Method::access_flags_offset()));
1053       __ testl(flags, JVM_ACC_STATIC);
1054       __ jcc(Assembler::zero, L_skip_barrier); // non-static
1055     }
1056 
1057     Register klass = rscratch1;
1058     __ load_method_holder(klass, method);
1059     __ clinit_barrier(klass, &L_skip_barrier /*L_fast_path*/);
1060 
1061     __ jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub())); // slow path
1062 
1063     __ bind(L_skip_barrier);
1064     entry_address[AdapterBlob::C2I_No_Clinit_Check] = __ pc();














1065   }
1066 
1067   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1068   bs->c2i_entry_barrier(masm);
1069 
1070   gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
1071   return;
1072 }
1073 
1074 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
1075                                          VMRegPair *regs,
1076                                          int total_args_passed) {
1077 
1078 // We return the amount of VMRegImpl stack slots we need to reserve for all
1079 // the arguments NOT counting out_preserve_stack_slots.
1080 
1081 // NOTE: These arrays will have to change when c1 is ported
1082 #ifdef _WIN64
1083     static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1084       c_rarg0, c_rarg1, c_rarg2, c_rarg3
1085     };
1086     static const XMMRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
1087       c_farg0, c_farg1, c_farg2, c_farg3
1088     };
1089 #else
1090     static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1091       c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5

3493   julong *scratch = (julong *)alloca(total_allocation);
3494 
3495   // Local scratch arrays
3496   julong
3497     *a = scratch + 0 * longwords,
3498     *n = scratch + 1 * longwords,
3499     *m = scratch + 2 * longwords;
3500 
3501   reverse_words((julong *)a_ints, a, longwords);
3502   reverse_words((julong *)n_ints, n, longwords);
3503 
3504   if (len >= MONTGOMERY_SQUARING_THRESHOLD) {
3505     ::montgomery_square(a, n, m, (julong)inv, longwords);
3506   } else {
3507     ::montgomery_multiply(a, a, n, m, (julong)inv, longwords);
3508   }
3509 
3510   reverse_words(m, (julong *)m_ints, longwords);
3511 }
3512 















































































































































3513 #if INCLUDE_JFR
3514 
3515 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3516 // It returns a jobject handle to the event writer.
3517 // The handle is dereferenced and the return value is the event writer oop.
3518 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
3519   enum layout {
3520     rbp_off,
3521     rbpH_off,
3522     return_off,
3523     return_off2,
3524     framesize // inclusive of return address
3525   };
3526 
3527   const char* name = SharedRuntime::stub_name(StubId::shared_jfr_write_checkpoint_id);
3528   CodeBuffer code(name, 1024, 64);
3529   MacroAssembler* masm = new MacroAssembler(&code);
3530   address start = __ pc();
3531 
3532   __ enter();

3585   __ reset_last_Java_frame(true);
3586 
3587   __ leave();
3588   __ ret(0);
3589 
3590   OopMapSet* oop_maps = new OopMapSet();
3591   OopMap* map = new OopMap(framesize, 1);
3592   oop_maps->add_gc_map(frame_complete, map);
3593 
3594   RuntimeStub* stub =
3595     RuntimeStub::new_runtime_stub(name,
3596                                   &code,
3597                                   frame_complete,
3598                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3599                                   oop_maps,
3600                                   false);
3601   return stub;
3602 }
3603 
3604 #endif // INCLUDE_JFR
3605 

   1 /*
   2  * Copyright (c) 2003, 2026, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #ifndef _WINDOWS
  26 #include "alloca.h"
  27 #endif
  28 #include "asm/macroAssembler.hpp"
  29 #include "asm/macroAssembler.inline.hpp"
  30 #include "classfile/symbolTable.hpp"
  31 #include "code/aotCodeCache.hpp"
  32 #include "code/compiledIC.hpp"
  33 #include "code/debugInfoRec.hpp"
  34 #include "code/nativeInst.hpp"
  35 #include "code/vtableStubs.hpp"
  36 #include "compiler/oopMap.hpp"
  37 #include "gc/shared/collectedHeap.hpp"
  38 #include "gc/shared/gcLocker.hpp"
  39 #include "gc/shared/barrierSet.hpp"
  40 #include "gc/shared/barrierSetAssembler.hpp"
  41 #include "interpreter/interpreter.hpp"
  42 #include "logging/log.hpp"
  43 #include "memory/resourceArea.hpp"
  44 #include "memory/universe.hpp"
  45 #include "oops/klass.inline.hpp"
  46 #include "oops/method.inline.hpp"
  47 #include "prims/methodHandles.hpp"
  48 #include "runtime/continuation.hpp"
  49 #include "runtime/continuationEntry.inline.hpp"
  50 #include "runtime/globals.hpp"

 618       break;
 619     case T_DOUBLE:
 620       assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
 621       if (fp_args < Argument::n_float_register_parameters_j) {
 622         regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
 623       } else {
 624         stk_args = align_up(stk_args, 2);
 625         regs[i].set2(VMRegImpl::stack2reg(stk_args));
 626         stk_args += 2;
 627       }
 628       break;
 629     default:
 630       ShouldNotReachHere();
 631       break;
 632     }
 633   }
 634 
 635   return stk_args;
 636 }
 637 
 638 // Same as java_calling_convention() but for multiple return
 639 // values. There's no way to store them on the stack so if we don't
 640 // have enough registers, multiple values can't be returned.
 641 const uint SharedRuntime::java_return_convention_max_int = Argument::n_int_register_parameters_j+1;
 642 const uint SharedRuntime::java_return_convention_max_float = Argument::n_float_register_parameters_j;
 643 int SharedRuntime::java_return_convention(const BasicType *sig_bt,
 644                                           VMRegPair *regs,
 645                                           int total_args_passed) {
 646   // Create the mapping between argument positions and
 647   // registers.
 648   static const Register INT_ArgReg[java_return_convention_max_int] = {
 649     rax, j_rarg5, j_rarg4, j_rarg3, j_rarg2, j_rarg1, j_rarg0
 650   };
 651   static const XMMRegister FP_ArgReg[java_return_convention_max_float] = {
 652     j_farg0, j_farg1, j_farg2, j_farg3,
 653     j_farg4, j_farg5, j_farg6, j_farg7
 654   };
 655 
 656 
 657   uint int_args = 0;
 658   uint fp_args = 0;
 659 
 660   for (int i = 0; i < total_args_passed; i++) {
 661     switch (sig_bt[i]) {
 662     case T_BOOLEAN:
 663     case T_CHAR:
 664     case T_BYTE:
 665     case T_SHORT:
 666     case T_INT:
 667       if (int_args < Argument::n_int_register_parameters_j+1) {
 668         regs[i].set1(INT_ArgReg[int_args]->as_VMReg());
 669         int_args++;
 670       } else {
 671         return -1;
 672       }
 673       break;
 674     case T_VOID:
 675       // halves of T_LONG or T_DOUBLE
 676       assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
 677       regs[i].set_bad();
 678       break;
 679     case T_LONG:
 680       assert(sig_bt[i + 1] == T_VOID, "expecting half");
 681       // fall through
 682     case T_OBJECT:
 683     case T_ARRAY:
 684     case T_ADDRESS:
 685     case T_METADATA:
 686       if (int_args < Argument::n_int_register_parameters_j+1) {
 687         regs[i].set2(INT_ArgReg[int_args]->as_VMReg());
 688         int_args++;
 689       } else {
 690         return -1;
 691       }
 692       break;
 693     case T_FLOAT:
 694       if (fp_args < Argument::n_float_register_parameters_j) {
 695         regs[i].set1(FP_ArgReg[fp_args]->as_VMReg());
 696         fp_args++;
 697       } else {
 698         return -1;
 699       }
 700       break;
 701     case T_DOUBLE:
 702       assert(sig_bt[i + 1] == T_VOID, "expecting half");
 703       if (fp_args < Argument::n_float_register_parameters_j) {
 704         regs[i].set2(FP_ArgReg[fp_args]->as_VMReg());
 705         fp_args++;
 706       } else {
 707         return -1;
 708       }
 709       break;
 710     default:
 711       ShouldNotReachHere();
 712       break;
 713     }
 714   }
 715 
 716   return int_args + fp_args;
 717 }
 718 
 719 // Patch the callers callsite with entry to compiled code if it exists.
 720 static void patch_callers_callsite(MacroAssembler *masm) {
 721   Label L;
 722   __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), NULL_WORD);
 723   __ jcc(Assembler::equal, L);
 724 
 725   // Save the current stack pointer
 726   __ mov(r13, rsp);
 727   // Schedule the branch target address early.
 728   // Call into the VM to patch the caller, then jump to compiled callee
 729   // rax isn't live so capture return address while we easily can
 730   __ movptr(rax, Address(rsp, 0));
 731 
 732   // align stack so push_CPU_state doesn't fault
 733   __ andptr(rsp, -(StackAlignmentInBytes));
 734   __ push_CPU_state();
 735   __ vzeroupper();
 736   // VM needs caller's callsite
 737   // VM needs target method
 738   // This needs to be a long call since we will relocate this adapter to

 741   // Allocate argument register save area
 742   if (frame::arg_reg_save_area_bytes != 0) {
 743     __ subptr(rsp, frame::arg_reg_save_area_bytes);
 744   }
 745   __ mov(c_rarg0, rbx);
 746   __ mov(c_rarg1, rax);
 747   __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)));
 748 
 749   // De-allocate argument register save area
 750   if (frame::arg_reg_save_area_bytes != 0) {
 751     __ addptr(rsp, frame::arg_reg_save_area_bytes);
 752   }
 753 
 754   __ vzeroupper();
 755   __ pop_CPU_state();
 756   // restore sp
 757   __ mov(rsp, r13);
 758   __ bind(L);
 759 }
 760 
 761 // For each inline type argument, sig includes the list of fields of
 762 // the inline type. This utility function computes the number of
 763 // arguments for the call if inline types are passed by reference (the
 764 // calling convention the interpreter expects).
 765 static int compute_total_args_passed_int(const GrowableArray<SigEntry>* sig_extended) {
 766   int total_args_passed = 0;
 767   if (InlineTypePassFieldsAsArgs) {
 768     for (int i = 0; i < sig_extended->length(); i++) {
 769       BasicType bt = sig_extended->at(i)._bt;
 770       if (bt == T_METADATA) {
 771         // In sig_extended, an inline type argument starts with:
 772         // T_METADATA, followed by the types of the fields of the
 773         // inline type and T_VOID to mark the end of the value
 774         // type. Inline types are flattened so, for instance, in the
 775         // case of an inline type with an int field and an inline type
 776         // field that itself has 2 fields, an int and a long:
 777         // T_METADATA T_INT T_METADATA T_INT T_LONG T_VOID (second
 778         // slot for the T_LONG) T_VOID (inner inline type) T_VOID
 779         // (outer inline type)
 780         total_args_passed++;
 781         int vt = 1;
 782         do {
 783           i++;
 784           BasicType bt = sig_extended->at(i)._bt;
 785           BasicType prev_bt = sig_extended->at(i-1)._bt;
 786           if (bt == T_METADATA) {
 787             vt++;
 788           } else if (bt == T_VOID &&
 789                      prev_bt != T_LONG &&
 790                      prev_bt != T_DOUBLE) {
 791             vt--;
 792           }
 793         } while (vt != 0);
 794       } else {
 795         total_args_passed++;
 796       }
 797     }
 798   } else {
 799     total_args_passed = sig_extended->length();
 800   }
 801   return total_args_passed;
 802 }
 803 
 804 
 805 static void gen_c2i_adapter_helper(MacroAssembler* masm,
 806                                    BasicType bt,
 807                                    BasicType prev_bt,
 808                                    size_t size_in_bytes,
 809                                    const VMRegPair& reg_pair,
 810                                    const Address& to,
 811                                    int extraspace,
 812                                    bool is_oop) {
 813   if (bt == T_VOID) {
 814     assert(prev_bt == T_LONG || prev_bt == T_DOUBLE, "missing half");
 815     return;
 816   }
 817 
 818   // Say 4 args:
 819   // i   st_off
 820   // 0   32 T_LONG
 821   // 1   24 T_VOID
 822   // 2   16 T_OBJECT
 823   // 3    8 T_BOOL
 824   // -    0 return address
 825   //
 826   // However to make thing extra confusing. Because we can fit a long/double in
 827   // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
 828   // leaves one slot empty and only stores to a single slot. In this case the
 829   // slot that is occupied is the T_VOID slot. See I said it was confusing.
 830 
 831   bool wide = (size_in_bytes == wordSize);
 832   VMReg r_1 = reg_pair.first();
 833   VMReg r_2 = reg_pair.second();
 834   assert(r_2->is_valid() == wide, "invalid size");
 835   if (!r_1->is_valid()) {
 836     assert(!r_2->is_valid(), "must be invalid");
 837     return;
 838   }
 839 
 840   if (!r_1->is_XMMRegister()) {
 841     Register val = rax;
 842     if (r_1->is_stack()) {
 843       int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
 844       __ load_sized_value(val, Address(rsp, ld_off), size_in_bytes, /* is_signed */ false);
 845     } else {
 846       val = r_1->as_Register();
 847     }
 848     assert_different_registers(to.base(), val, rscratch1);
 849     if (is_oop) {
 850       __ push(r13);
 851       __ push(rbx);
 852       // store_heap_oop transitively calls oop_store_at which corrupts to.base(). We need to keep it valid.
 853       __ push(to.base());
 854       __ store_heap_oop(to, val, rscratch1, r13, rbx, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
 855       __ pop(to.base());
 856       __ pop(rbx);
 857       __ pop(r13);
 858     } else {
 859       __ store_sized_value(to, val, size_in_bytes);
 860     }
 861   } else {
 862     if (wide) {
 863       __ movdbl(to, r_1->as_XMMRegister());
 864     } else {
 865       __ movflt(to, r_1->as_XMMRegister());
 866     }
 867   }
 868 }
 869 
 870 static void gen_c2i_adapter(MacroAssembler *masm,
 871                             const GrowableArray<SigEntry>* sig_extended,


 872                             const VMRegPair *regs,
 873                             bool requires_clinit_barrier,
 874                             address& c2i_no_clinit_check_entry,
 875                             Label& skip_fixup,
 876                             address start,
 877                             OopMapSet* oop_maps,
 878                             int& frame_complete,
 879                             int& frame_size_in_words,
 880                             bool alloc_inline_receiver) {
 881   if (requires_clinit_barrier && VM_Version::supports_fast_class_init_checks()) {
 882     Label L_skip_barrier;
 883     Register method = rbx;
 884 
 885     { // Bypass the barrier for non-static methods
 886       Register flags = rscratch1;
 887       __ load_unsigned_short(flags, Address(method, Method::access_flags_offset()));
 888       __ testl(flags, JVM_ACC_STATIC);
 889       __ jcc(Assembler::zero, L_skip_barrier); // non-static
 890     }
 891 
 892     Register klass = rscratch1;
 893     __ load_method_holder(klass, method);
 894     __ clinit_barrier(klass, &L_skip_barrier /*L_fast_path*/);
 895 
 896     __ jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub())); // slow path
 897 
 898     __ bind(L_skip_barrier);
 899     c2i_no_clinit_check_entry = __ pc();
 900   }
 901 
 902   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
 903   bs->c2i_entry_barrier(masm);
 904 
 905   // Before we get into the guts of the C2I adapter, see if we should be here
 906   // at all.  We've come from compiled code and are attempting to jump to the
 907   // interpreter, which means the caller made a static call to get here
 908   // (vcalls always get a compiled target if there is one).  Check for a
 909   // compiled target.  If there is one, we need to patch the caller's call.
 910   patch_callers_callsite(masm);
 911 
 912   __ bind(skip_fixup);
 913 
 914   if (InlineTypePassFieldsAsArgs) {
 915     // Is there an inline type argument?
 916     bool has_inline_argument = false;
 917     for (int i = 0; i < sig_extended->length() && !has_inline_argument; i++) {
 918       has_inline_argument = (sig_extended->at(i)._bt == T_METADATA);
 919     }
 920     if (has_inline_argument) {
 921       // There is at least an inline type argument: we're coming from
 922       // compiled code so we have no buffers to back the inline types.
 923       // Allocate the buffers here with a runtime call.
 924       OopMap* map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, /*save_wide_vectors*/ false);
 925 
 926       frame_complete = __ offset();
 927 
 928       __ set_last_Java_frame(noreg, noreg, nullptr, rscratch1);
 929 
 930       __ mov(c_rarg0, r15_thread);
 931       __ mov(c_rarg1, rbx);
 932       __ mov64(c_rarg2, (int64_t)alloc_inline_receiver);
 933       __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::allocate_inline_types)));
 934 
 935       oop_maps->add_gc_map((int)(__ pc() - start), map);
 936       __ reset_last_Java_frame(false);
 937 
 938       RegisterSaver::restore_live_registers(masm);
 939 
 940       Label no_exception;
 941       __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
 942       __ jcc(Assembler::equal, no_exception);
 943 
 944       __ movptr(Address(r15_thread, JavaThread::vm_result_oop_offset()), NULL_WORD);
 945       __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
 946       __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
 947 
 948       __ bind(no_exception);
 949 
 950       // We get an array of objects from the runtime call
 951       __ get_vm_result_oop(rscratch2); // Use rscratch2 (r11) as temporary because rscratch1 (r10) is trashed by movptr()
 952       __ get_vm_result_metadata(rbx); // TODO: required to keep the callee Method live?
 953     }
 954   }
 955 
 956   // Since all args are passed on the stack, total_args_passed *
 957   // Interpreter::stackElementSize is the space we need.
 958   int total_args_passed = compute_total_args_passed_int(sig_extended);
 959   assert(total_args_passed >= 0, "total_args_passed is %d", total_args_passed);
 960 
 961   int extraspace = (total_args_passed * Interpreter::stackElementSize);
 962 
 963   // stack is aligned, keep it that way
 964   // This is not currently needed or enforced by the interpreter, but
 965   // we might as well conform to the ABI.
 966   extraspace = align_up(extraspace, 2*wordSize);
 967 
 968   // set senderSP value
 969   __ lea(r13, Address(rsp, wordSize));
 970 
 971 #ifdef ASSERT
 972   __ check_stack_alignment(r13, "sender stack not aligned");
 973 #endif
 974   if (extraspace > 0) {
 975     // Pop the return address
 976     __ pop(rax);
 977 
 978     __ subptr(rsp, extraspace);
 979 
 980     // Push the return address
 981     __ push(rax);
 982 
 983     // Account for the return address location since we store it first rather
 984     // than hold it in a register across all the shuffling
 985     extraspace += wordSize;
 986   }
 987 
 988 #ifdef ASSERT
 989   __ check_stack_alignment(rsp, "callee stack not aligned", wordSize, rax);
 990 #endif
 991 
 992   // Now write the args into the outgoing interpreter space




































 993 
 994   // next_arg_comp is the next argument from the compiler point of
 995   // view (inline type fields are passed in registers/on the stack). In
 996   // sig_extended, an inline type argument starts with: T_METADATA,
 997   // followed by the types of the fields of the inline type and T_VOID
 998   // to mark the end of the inline type. ignored counts the number of
 999   // T_METADATA/T_VOID. next_vt_arg is the next inline type argument:
1000   // used to get the buffer for that argument from the pool of buffers
1001   // we allocated above and want to pass to the
1002   // interpreter. next_arg_int is the next argument from the
1003   // interpreter point of view (inline types are passed by reference).
1004   for (int next_arg_comp = 0, ignored = 0, next_vt_arg = 0, next_arg_int = 0;
1005        next_arg_comp < sig_extended->length(); next_arg_comp++) {
1006     assert(ignored <= next_arg_comp, "shouldn't skip over more slots than there are arguments");
1007     assert(next_arg_int <= total_args_passed, "more arguments for the interpreter than expected?");
1008     BasicType bt = sig_extended->at(next_arg_comp)._bt;
1009     int st_off = (total_args_passed - next_arg_int) * Interpreter::stackElementSize;
1010     if (!InlineTypePassFieldsAsArgs || bt != T_METADATA) {
1011       int next_off = st_off - Interpreter::stackElementSize;
1012       const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : st_off;
1013       const VMRegPair reg_pair = regs[next_arg_comp-ignored];
1014       size_t size_in_bytes = reg_pair.second()->is_valid() ? 8 : 4;
1015       gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
1016                              size_in_bytes, reg_pair, Address(rsp, offset), extraspace, false);
1017       next_arg_int++;
1018 #ifdef ASSERT
1019       if (bt == T_LONG || bt == T_DOUBLE) {
1020         // Overwrite the unused slot with known junk
1021         __ mov64(rax, CONST64(0xdeadffffdeadaaaa));
1022         __ movptr(Address(rsp, st_off), rax);



1023       }















1024 #endif /* ASSERT */
1025     } else {
1026       ignored++;
1027       // get the buffer from the just allocated pool of buffers
1028       int index = arrayOopDesc::base_offset_in_bytes(T_OBJECT) + next_vt_arg * type2aelembytes(T_OBJECT);
1029       __ load_heap_oop(r14, Address(rscratch2, index));
1030       next_vt_arg++; next_arg_int++;
1031       int vt = 1;
1032       // write fields we get from compiled code in registers/stack
1033       // slots to the buffer: we know we are done with that inline type
1034       // argument when we hit the T_VOID that acts as an end of inline
1035       // type delimiter for this inline type. Inline types are flattened
1036       // so we might encounter embedded inline types. Each entry in
1037       // sig_extended contains a field offset in the buffer.
1038       Label L_null;
1039       do {
1040         next_arg_comp++;
1041         BasicType bt = sig_extended->at(next_arg_comp)._bt;
1042         BasicType prev_bt = sig_extended->at(next_arg_comp-1)._bt;
1043         if (bt == T_METADATA) {
1044           vt++;
1045           ignored++;
1046         } else if (bt == T_VOID &&
1047                    prev_bt != T_LONG &&
1048                    prev_bt != T_DOUBLE) {
1049           vt--;
1050           ignored++;
1051         } else {
1052           int off = sig_extended->at(next_arg_comp)._offset;
1053           if (off == -1) {
1054             // Nullable inline type argument, emit null check
1055             VMReg reg = regs[next_arg_comp-ignored].first();
1056             Label L_notNull;
1057             if (reg->is_stack()) {
1058               int ld_off = reg->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
1059               __ testb(Address(rsp, ld_off), 1);
1060             } else {
1061               __ testb(reg->as_Register(), 1);
1062             }
1063             __ jcc(Assembler::notZero, L_notNull);
1064             __ movptr(Address(rsp, st_off), 0);
1065             __ jmp(L_null);
1066             __ bind(L_notNull);
1067             continue;
1068           }
1069           assert(off > 0, "offset in object should be positive");
1070           size_t size_in_bytes = is_java_primitive(bt) ? type2aelembytes(bt) : wordSize;
1071           bool is_oop = is_reference_type(bt);
1072           gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
1073                                  size_in_bytes, regs[next_arg_comp-ignored], Address(r14, off), extraspace, is_oop);
1074         }
1075       } while (vt != 0);
1076       // pass the buffer to the interpreter
1077       __ movptr(Address(rsp, st_off), r14);
1078       __ bind(L_null);










1079     }
1080   }
1081 
1082   // Schedule the branch target address early.
1083   __ movptr(rcx, Address(rbx, in_bytes(Method::interpreter_entry_offset())));
1084   __ jmp(rcx);
1085 }
1086 
1087 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,

1088                                     int comp_args_on_stack,
1089                                     const GrowableArray<SigEntry>* sig,
1090                                     const VMRegPair *regs) {
1091 
1092   // Note: r13 contains the senderSP on entry. We must preserve it since
1093   // we may do a i2c -> c2i transition if we lose a race where compiled
1094   // code goes non-entrant while we get args ready.
1095   // In addition we use r13 to locate all the interpreter args as
1096   // we must align the stack to 16 bytes on an i2c entry else we
1097   // lose alignment we expect in all compiled code and register
1098   // save code can segv when fxsave instructions find improperly
1099   // aligned stack pointer.
1100 
1101   // Adapters can be frameless because they do not require the caller
1102   // to perform additional cleanup work, such as correcting the stack pointer.
1103   // An i2c adapter is frameless because the *caller* frame, which is interpreted,
1104   // routinely repairs its own stack pointer (from interpreter_frame_last_sp),
1105   // even if a callee has modified the stack pointer.
1106   // A c2i adapter is frameless because the *callee* frame, which is interpreted,
1107   // routinely repairs its caller's stack pointer (from sender_sp, which is set
1108   // up via the senderSP register).
1109   // In other words, if *either* the caller or callee is interpreted, we can

1125   // Convert 4-byte c2 stack slots to words.
1126   int comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
1127 
1128   if (comp_args_on_stack) {
1129     __ subptr(rsp, comp_words_on_stack * wordSize);
1130   }
1131 
1132   // Ensure compiled code always sees stack at proper alignment
1133   __ andptr(rsp, -16);
1134 
1135   // push the return address and misalign the stack that youngest frame always sees
1136   // as far as the placement of the call instruction
1137   __ push(rax);
1138 
1139   // Put saved SP in another register
1140   const Register saved_sp = rax;
1141   __ movptr(saved_sp, r11);
1142 
1143   // Will jump to the compiled code just as if compiled code was doing it.
1144   // Pre-load the register-jump target early, to schedule it better.
1145   __ movptr(r11, Address(rbx, in_bytes(Method::from_compiled_inline_offset())));
1146 
1147 #if INCLUDE_JVMCI
1148   if (EnableJVMCI) {
1149     // check if this call should be routed towards a specific entry point
1150     __ cmpptr(Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())), 0);
1151     Label no_alternative_target;
1152     __ jcc(Assembler::equal, no_alternative_target);
1153     __ movptr(r11, Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
1154     __ movptr(Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())), 0);
1155     __ bind(no_alternative_target);
1156   }
1157 #endif // INCLUDE_JVMCI
1158 
1159   int total_args_passed = sig->length();
1160 
1161   // Now generate the shuffle code.  Pick up all register args and move the
1162   // rest through the floating point stack top.
1163   for (int i = 0; i < total_args_passed; i++) {
1164     BasicType bt = sig->at(i)._bt;
1165     if (bt == T_VOID) {
1166       // Longs and doubles are passed in native word order, but misaligned
1167       // in the 32-bit build.
1168       BasicType prev_bt = (i > 0) ? sig->at(i-1)._bt : T_ILLEGAL;
1169       assert(i > 0 && (prev_bt == T_LONG || prev_bt == T_DOUBLE), "missing half");
1170       continue;
1171     }
1172 
1173     // Pick up 0, 1 or 2 words from SP+offset.
1174 
1175     assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
1176             "scrambled load targets?");
1177     // Load in argument order going down.
1178     int ld_off = (total_args_passed - i)*Interpreter::stackElementSize;
1179     // Point to interpreter value (vs. tag)
1180     int next_off = ld_off - Interpreter::stackElementSize;
1181     //
1182     //
1183     //
1184     VMReg r_1 = regs[i].first();
1185     VMReg r_2 = regs[i].second();
1186     if (!r_1->is_valid()) {
1187       assert(!r_2->is_valid(), "");
1188       continue;
1189     }

1191       // Convert stack slot to an SP offset (+ wordSize to account for return address )
1192       int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size + wordSize;
1193 
1194       // We can use r13 as a temp here because compiled code doesn't need r13 as an input
1195       // and if we end up going thru a c2i because of a miss a reasonable value of r13
1196       // will be generated.
1197       if (!r_2->is_valid()) {
1198         // sign extend???
1199         __ movl(r13, Address(saved_sp, ld_off));
1200         __ movptr(Address(rsp, st_off), r13);
1201       } else {
1202         //
1203         // We are using two optoregs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
1204         // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
1205         // So we must adjust where to pick up the data to match the interpreter.
1206         //
1207         // Interpreter local[n] == MSW, local[n+1] == LSW however locals
1208         // are accessed as negative so LSW is at LOW address
1209 
1210         // ld_off is MSW so get LSW
1211         const int offset = (bt==T_LONG||bt==T_DOUBLE)?
1212                            next_off : ld_off;
1213         __ movq(r13, Address(saved_sp, offset));
1214         // st_off is LSW (i.e. reg.first())
1215         __ movq(Address(rsp, st_off), r13);
1216       }
1217     } else if (r_1->is_Register()) {  // Register argument
1218       Register r = r_1->as_Register();
1219       assert(r != rax, "must be different");
1220       if (r_2->is_valid()) {
1221         //
1222         // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
1223         // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
1224         // So we must adjust where to pick up the data to match the interpreter.
1225 
1226         const int offset = (bt==T_LONG||bt==T_DOUBLE)?
1227                            next_off : ld_off;
1228 
1229         // this can be a misaligned move
1230         __ movq(r, Address(saved_sp, offset));
1231       } else {
1232         // sign extend and use a full word?
1233         __ movl(r, Address(saved_sp, ld_off));
1234       }
1235     } else {
1236       if (!r_2->is_valid()) {
1237         __ movflt(r_1->as_XMMRegister(), Address(saved_sp, ld_off));
1238       } else {
1239         __ movdbl(r_1->as_XMMRegister(), Address(saved_sp, next_off));
1240       }
1241     }
1242   }
1243 
1244   __ push_cont_fastpath(); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
1245 
1246   // 6243940 We might end up in handle_wrong_method if
1247   // the callee is deoptimized as we race thru here. If that
1248   // happens we don't want to take a safepoint because the
1249   // caller frame will look interpreted and arguments are now
1250   // "compiled" so it is much better to make this transition
1251   // invisible to the stack walking code. Unfortunately if
1252   // we try and find the callee by normal means a safepoint
1253   // is possible. So we stash the desired callee in the thread
1254   // and the vm will find there should this case occur.
1255 
1256   __ movptr(Address(r15_thread, JavaThread::callee_target_offset()), rbx);
1257 
1258   // put Method* where a c2i would expect should we end up there
1259   // only needed because of c2 resolve stubs return Method* as a result in
1260   // rax
1261   __ mov(rax, rbx);
1262   __ jmp(r11);
1263 }
1264 
1265 static void gen_inline_cache_check(MacroAssembler *masm, Label& skip_fixup) {
1266   Register data = rax;
1267   __ ic_check(1 /* end_alignment */);
1268   __ movptr(rbx, Address(data, CompiledICData::speculated_method_offset()));
1269 
1270   // Method might have been compiled since the call site was patched to
1271   // interpreted if that is the case treat it as a miss so we can get
1272   // the call site corrected.
1273   __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), NULL_WORD);
1274   __ jcc(Assembler::equal, skip_fixup);
1275   __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1276 }
1277 
1278 // ---------------------------------------------------------------
1279 void SharedRuntime::generate_i2c2i_adapters(MacroAssembler* masm,

1280                                             int comp_args_on_stack,
1281                                             const GrowableArray<SigEntry>* sig,
1282                                             const VMRegPair* regs,
1283                                             const GrowableArray<SigEntry>* sig_cc,
1284                                             const VMRegPair* regs_cc,
1285                                             const GrowableArray<SigEntry>* sig_cc_ro,
1286                                             const VMRegPair* regs_cc_ro,
1287                                             address entry_address[AdapterBlob::ENTRY_COUNT],
1288                                             AdapterBlob*& new_adapter,
1289                                             bool allocate_code_blob) {
1290   entry_address[AdapterBlob::I2C] = __ pc();
1291   gen_i2c_adapter(masm, comp_args_on_stack, sig, regs);

1292 
1293   // -------------------------------------------------------------------------
1294   // Generate a C2I adapter.  On entry we know rbx holds the Method* during calls
1295   // to the interpreter.  The args start out packed in the compiled layout.  They
1296   // need to be unpacked into the interpreter layout.  This will almost always
1297   // require some stack space.  We grow the current (compiled) stack, then repack
1298   // the args.  We  finally end in a jump to the generic interpreter entry point.
1299   // On exit from the interpreter, the interpreter will restore our SP (lest the
1300   // compiled code, which relies solely on SP and not RBP, get sick).
1301 
1302   entry_address[AdapterBlob::C2I_Unverified] = __ pc();
1303   entry_address[AdapterBlob::C2I_Unverified_Inline] = __ pc();
1304   Label skip_fixup;
1305 
1306   gen_inline_cache_check(masm, skip_fixup);


1307 
1308   OopMapSet* oop_maps = new OopMapSet();
1309   int frame_complete = CodeOffsets::frame_never_safe;
1310   int frame_size_in_words = 0;









1311 
1312   // Scalarized c2i adapter with non-scalarized receiver (i.e., don't pack receiver)
1313   entry_address[AdapterBlob::C2I_No_Clinit_Check] = nullptr;
1314   entry_address[AdapterBlob::C2I_Inline_RO] = __ pc();
1315   if (regs_cc != regs_cc_ro) {
1316     // No class init barrier needed because method is guaranteed to be non-static
1317     gen_c2i_adapter(masm, sig_cc_ro, regs_cc_ro, /* requires_clinit_barrier = */ false, entry_address[AdapterBlob::C2I_No_Clinit_Check],
1318                     skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
1319     skip_fixup.reset();
1320   }
1321 
1322   // Scalarized c2i adapter
1323   entry_address[AdapterBlob::C2I]        = __ pc();
1324   entry_address[AdapterBlob::C2I_Inline] = __ pc();
1325   gen_c2i_adapter(masm, sig_cc, regs_cc, /* requires_clinit_barrier = */ true, entry_address[AdapterBlob::C2I_No_Clinit_Check],
1326                   skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ true);
1327 
1328   // Non-scalarized c2i adapter
1329   if (regs != regs_cc) {
1330     entry_address[AdapterBlob::C2I_Unverified_Inline] = __ pc();
1331     Label inline_entry_skip_fixup;
1332     gen_inline_cache_check(masm, inline_entry_skip_fixup);
1333 
1334     entry_address[AdapterBlob::C2I_Inline] = __ pc();
1335     gen_c2i_adapter(masm, sig, regs, /* requires_clinit_barrier = */ true, entry_address[AdapterBlob::C2I_No_Clinit_Check],
1336                     inline_entry_skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
1337   }
1338 
1339   // The c2i adapters might safepoint and trigger a GC. The caller must make sure that
1340   // the GC knows about the location of oop argument locations passed to the c2i adapter.
1341   if (allocate_code_blob) {
1342     bool caller_must_gc_arguments = (regs != regs_cc);
1343     int entry_offset[AdapterHandlerEntry::ENTRIES_COUNT];
1344     assert(AdapterHandlerEntry::ENTRIES_COUNT == 7, "sanity");
1345     AdapterHandlerLibrary::address_to_offset(entry_address, entry_offset);
1346     new_adapter = AdapterBlob::create(masm->code(), entry_offset, frame_complete, frame_size_in_words, oop_maps, caller_must_gc_arguments);
1347   }






1348 }
1349 
1350 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
1351                                          VMRegPair *regs,
1352                                          int total_args_passed) {
1353 
1354 // We return the amount of VMRegImpl stack slots we need to reserve for all
1355 // the arguments NOT counting out_preserve_stack_slots.
1356 
1357 // NOTE: These arrays will have to change when c1 is ported
1358 #ifdef _WIN64
1359     static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1360       c_rarg0, c_rarg1, c_rarg2, c_rarg3
1361     };
1362     static const XMMRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
1363       c_farg0, c_farg1, c_farg2, c_farg3
1364     };
1365 #else
1366     static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1367       c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5

3769   julong *scratch = (julong *)alloca(total_allocation);
3770 
3771   // Local scratch arrays
3772   julong
3773     *a = scratch + 0 * longwords,
3774     *n = scratch + 1 * longwords,
3775     *m = scratch + 2 * longwords;
3776 
3777   reverse_words((julong *)a_ints, a, longwords);
3778   reverse_words((julong *)n_ints, n, longwords);
3779 
3780   if (len >= MONTGOMERY_SQUARING_THRESHOLD) {
3781     ::montgomery_square(a, n, m, (julong)inv, longwords);
3782   } else {
3783     ::montgomery_multiply(a, a, n, m, (julong)inv, longwords);
3784   }
3785 
3786   reverse_words(m, (julong *)m_ints, longwords);
3787 }
3788 
3789 BufferedInlineTypeBlob* SharedRuntime::generate_buffered_inline_type_adapter(const InlineKlass* vk) {
3790   BufferBlob* buf = BufferBlob::create("inline types pack/unpack", 16 * K);
3791   if (buf == nullptr) {
3792     return nullptr;
3793   }
3794   CodeBuffer buffer(buf);
3795   short buffer_locs[20];
3796   buffer.insts()->initialize_shared_locs((relocInfo*)buffer_locs,
3797                                          sizeof(buffer_locs)/sizeof(relocInfo));
3798 
3799   MacroAssembler* masm = new MacroAssembler(&buffer);
3800 
3801   const Array<SigEntry>* sig_vk = vk->extended_sig();
3802   const Array<VMRegPair>* regs = vk->return_regs();
3803 
3804   int pack_fields_jobject_off = __ offset();
3805   // Resolve pre-allocated buffer from JNI handle.
3806   // We cannot do this in generate_call_stub() because it requires GC code to be initialized.
3807   __ movptr(rax, Address(r13, 0));
3808   __ resolve_jobject(rax /* value */,
3809                      r12 /* tmp */);
3810   __ movptr(Address(r13, 0), rax);
3811 
3812   int pack_fields_off = __ offset();
3813 
3814   int j = 1;
3815   for (int i = 0; i < sig_vk->length(); i++) {
3816     BasicType bt = sig_vk->at(i)._bt;
3817     if (bt == T_METADATA) {
3818       continue;
3819     }
3820     if (bt == T_VOID) {
3821       if (sig_vk->at(i-1)._bt == T_LONG ||
3822           sig_vk->at(i-1)._bt == T_DOUBLE) {
3823         j++;
3824       }
3825       continue;
3826     }
3827     int off = sig_vk->at(i)._offset;
3828     assert(off > 0, "offset in object should be positive");
3829     VMRegPair pair = regs->at(j);
3830     VMReg r_1 = pair.first();
3831     Address to(rax, off);
3832     if (bt == T_FLOAT) {
3833       __ movflt(to, r_1->as_XMMRegister());
3834     } else if (bt == T_DOUBLE) {
3835       __ movdbl(to, r_1->as_XMMRegister());
3836     } else {
3837       Register val = r_1->as_Register();
3838       assert_different_registers(to.base(), val, r14, r13, rbx, rscratch1);
3839       if (is_reference_type(bt)) {
3840         // store_heap_oop transitively calls oop_store_at which corrupts to.base(). We need to keep rax valid.
3841         __ mov(rbx, rax);
3842         Address to_with_rbx(rbx, off);
3843         __ store_heap_oop(to_with_rbx, val, r14, r13, rbx, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
3844       } else {
3845         __ store_sized_value(to, r_1->as_Register(), type2aelembytes(bt));
3846       }
3847     }
3848     j++;
3849   }
3850   assert(j == regs->length(), "missed a field?");
3851   if (vk->has_nullable_atomic_layout()) {
3852     // Set the null marker
3853     __ movb(Address(rax, vk->null_marker_offset()), 1);
3854   }
3855   __ ret(0);
3856 
3857   int unpack_fields_off = __ offset();
3858 
3859   Label skip;
3860   Label not_null;
3861   __ testptr(rax, rax);
3862   __ jcc(Assembler::notZero, not_null);
3863 
3864   // Return value is null. Zero oop registers to make the GC happy.
3865   j = 1;
3866   for (int i = 0; i < sig_vk->length(); i++) {
3867     BasicType bt = sig_vk->at(i)._bt;
3868     if (bt == T_METADATA) {
3869       continue;
3870     }
3871     if (bt == T_VOID) {
3872       if (sig_vk->at(i-1)._bt == T_LONG ||
3873           sig_vk->at(i-1)._bt == T_DOUBLE) {
3874         j++;
3875       }
3876       continue;
3877     }
3878     if (bt == T_OBJECT || bt == T_ARRAY) {
3879       VMRegPair pair = regs->at(j);
3880       VMReg r_1 = pair.first();
3881       __ xorq(r_1->as_Register(), r_1->as_Register());
3882     }
3883     j++;
3884   }
3885   __ jmp(skip);
3886   __ bind(not_null);
3887 
3888   j = 1;
3889   for (int i = 0; i < sig_vk->length(); i++) {
3890     BasicType bt = sig_vk->at(i)._bt;
3891     if (bt == T_METADATA) {
3892       continue;
3893     }
3894     if (bt == T_VOID) {
3895       if (sig_vk->at(i-1)._bt == T_LONG ||
3896           sig_vk->at(i-1)._bt == T_DOUBLE) {
3897         j++;
3898       }
3899       continue;
3900     }
3901     int off = sig_vk->at(i)._offset;
3902     assert(off > 0, "offset in object should be positive");
3903     VMRegPair pair = regs->at(j);
3904     VMReg r_1 = pair.first();
3905     VMReg r_2 = pair.second();
3906     Address from(rax, off);
3907     if (bt == T_FLOAT) {
3908       __ movflt(r_1->as_XMMRegister(), from);
3909     } else if (bt == T_DOUBLE) {
3910       __ movdbl(r_1->as_XMMRegister(), from);
3911     } else if (bt == T_OBJECT || bt == T_ARRAY) {
3912       assert_different_registers(rax, r_1->as_Register());
3913       __ load_heap_oop(r_1->as_Register(), from);
3914     } else {
3915       assert(is_java_primitive(bt), "unexpected basic type");
3916       assert_different_registers(rax, r_1->as_Register());
3917       size_t size_in_bytes = type2aelembytes(bt);
3918       __ load_sized_value(r_1->as_Register(), from, size_in_bytes, bt != T_CHAR && bt != T_BOOLEAN);
3919     }
3920     j++;
3921   }
3922   assert(j == regs->length(), "missed a field?");
3923 
3924   __ bind(skip);
3925   __ ret(0);
3926 
3927   __ flush();
3928 
3929   return BufferedInlineTypeBlob::create(&buffer, pack_fields_off, pack_fields_jobject_off, unpack_fields_off);
3930 }
3931 
3932 #if INCLUDE_JFR
3933 
3934 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3935 // It returns a jobject handle to the event writer.
3936 // The handle is dereferenced and the return value is the event writer oop.
3937 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
3938   enum layout {
3939     rbp_off,
3940     rbpH_off,
3941     return_off,
3942     return_off2,
3943     framesize // inclusive of return address
3944   };
3945 
3946   const char* name = SharedRuntime::stub_name(StubId::shared_jfr_write_checkpoint_id);
3947   CodeBuffer code(name, 1024, 64);
3948   MacroAssembler* masm = new MacroAssembler(&code);
3949   address start = __ pc();
3950 
3951   __ enter();

4004   __ reset_last_Java_frame(true);
4005 
4006   __ leave();
4007   __ ret(0);
4008 
4009   OopMapSet* oop_maps = new OopMapSet();
4010   OopMap* map = new OopMap(framesize, 1);
4011   oop_maps->add_gc_map(frame_complete, map);
4012 
4013   RuntimeStub* stub =
4014     RuntimeStub::new_runtime_stub(name,
4015                                   &code,
4016                                   frame_complete,
4017                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
4018                                   oop_maps,
4019                                   false);
4020   return stub;
4021 }
4022 
4023 #endif // INCLUDE_JFR

< prev index next >