< prev index next >

src/hotspot/cpu/x86/sharedRuntime_x86_64.cpp

Print this page

  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 #include "precompiled.hpp"
  26 #ifndef _WINDOWS
  27 #include "alloca.h"
  28 #endif
  29 #include "asm/macroAssembler.hpp"
  30 #include "asm/macroAssembler.inline.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"
  50 #include "runtime/jniHandles.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 
 680 static void gen_c2i_adapter(MacroAssembler *masm,
 681                             int total_args_passed,
 682                             int comp_args_on_stack,
 683                             const BasicType *sig_bt,
 684                             const VMRegPair *regs,
 685                             Label& skip_fixup) {































 686   // Before we get into the guts of the C2I adapter, see if we should be here
 687   // at all.  We've come from compiled code and are attempting to jump to the
 688   // interpreter, which means the caller made a static call to get here
 689   // (vcalls always get a compiled target if there is one).  Check for a
 690   // compiled target.  If there is one, we need to patch the caller's call.
 691   patch_callers_callsite(masm);
 692 
 693   __ bind(skip_fixup);
 694 










































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
















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

























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





















 807         }
 808       }
 809     } else {
 810       assert(r_1->is_XMMRegister(), "");
 811       if (!r_2->is_valid()) {
 812         // only a float use just part of the slot
 813         __ movflt(Address(rsp, st_off), r_1->as_XMMRegister());
 814       } else {
 815 #ifdef ASSERT
 816         // Overwrite the unused slot with known junk
 817         __ mov64(rax, CONST64(0xdeadffffdeadaaac));
 818         __ movptr(Address(rsp, st_off), rax);
 819 #endif /* ASSERT */
 820         __ movdbl(Address(rsp, next_off), r_1->as_XMMRegister());
 821       }
 822     }
 823   }
 824 
 825   // Schedule the branch target address early.
 826   __ movptr(rcx, Address(rbx, in_bytes(Method::interpreter_entry_offset())));
 827   __ jmp(rcx);
 828 }
 829 
 830 static void range_check(MacroAssembler* masm, Register pc_reg, Register temp_reg,
 831                         address code_start, address code_end,
 832                         Label& L_ok) {
 833   Label L_fail;
 834   __ lea(temp_reg, AddressLiteral(code_start, relocInfo::none));
 835   __ cmpptr(pc_reg, temp_reg);
 836   __ jcc(Assembler::belowEqual, L_fail);
 837   __ lea(temp_reg, AddressLiteral(code_end, relocInfo::none));
 838   __ cmpptr(pc_reg, temp_reg);
 839   __ jcc(Assembler::below, L_ok);
 840   __ bind(L_fail);
 841 }
 842 
 843 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
 844                                     int total_args_passed,
 845                                     int comp_args_on_stack,
 846                                     const BasicType *sig_bt,
 847                                     const VMRegPair *regs) {
 848 
 849   // Note: r13 contains the senderSP on entry. We must preserve it since
 850   // we may do a i2c -> c2i transition if we lose a race where compiled
 851   // code goes non-entrant while we get args ready.
 852   // In addition we use r13 to locate all the interpreter args as
 853   // we must align the stack to 16 bytes on an i2c entry else we
 854   // lose alignment we expect in all compiled code and register
 855   // save code can segv when fxsave instructions find improperly
 856   // aligned stack pointer.
 857 
 858   // Adapters can be frameless because they do not require the caller
 859   // to perform additional cleanup work, such as correcting the stack pointer.
 860   // An i2c adapter is frameless because the *caller* frame, which is interpreted,
 861   // routinely repairs its own stack pointer (from interpreter_frame_last_sp),
 862   // even if a callee has modified the stack pointer.
 863   // A c2i adapter is frameless because the *callee* frame, which is interpreted,
 864   // routinely repairs its caller's stack pointer (from sender_sp, which is set
 865   // up via the senderSP register).
 866   // In other words, if *either* the caller or callee is interpreted, we can

 917   // Convert 4-byte c2 stack slots to words.
 918   int comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
 919 
 920   if (comp_args_on_stack) {
 921     __ subptr(rsp, comp_words_on_stack * wordSize);
 922   }
 923 
 924   // Ensure compiled code always sees stack at proper alignment
 925   __ andptr(rsp, -16);
 926 
 927   // push the return address and misalign the stack that youngest frame always sees
 928   // as far as the placement of the call instruction
 929   __ push(rax);
 930 
 931   // Put saved SP in another register
 932   const Register saved_sp = rax;
 933   __ movptr(saved_sp, r11);
 934 
 935   // Will jump to the compiled code just as if compiled code was doing it.
 936   // Pre-load the register-jump target early, to schedule it better.
 937   __ movptr(r11, Address(rbx, in_bytes(Method::from_compiled_offset())));
 938 
 939 #if INCLUDE_JVMCI
 940   if (EnableJVMCI) {
 941     // check if this call should be routed towards a specific entry point
 942     __ cmpptr(Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())), 0);
 943     Label no_alternative_target;
 944     __ jcc(Assembler::equal, no_alternative_target);
 945     __ movptr(r11, Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
 946     __ movptr(Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())), 0);
 947     __ bind(no_alternative_target);
 948   }
 949 #endif // INCLUDE_JVMCI
 950 


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

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

 958       continue;
 959     }
 960 
 961     // Pick up 0, 1 or 2 words from SP+offset.
 962 
 963     assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
 964             "scrambled load targets?");
 965     // Load in argument order going down.
 966     int ld_off = (total_args_passed - i)*Interpreter::stackElementSize;
 967     // Point to interpreter value (vs. tag)
 968     int next_off = ld_off - Interpreter::stackElementSize;
 969     //
 970     //
 971     //
 972     VMReg r_1 = regs[i].first();
 973     VMReg r_2 = regs[i].second();
 974     if (!r_1->is_valid()) {
 975       assert(!r_2->is_valid(), "");
 976       continue;
 977     }

 979       // Convert stack slot to an SP offset (+ wordSize to account for return address )
 980       int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size + wordSize;
 981 
 982       // We can use r13 as a temp here because compiled code doesn't need r13 as an input
 983       // and if we end up going thru a c2i because of a miss a reasonable value of r13
 984       // will be generated.
 985       if (!r_2->is_valid()) {
 986         // sign extend???
 987         __ movl(r13, Address(saved_sp, ld_off));
 988         __ movptr(Address(rsp, st_off), r13);
 989       } else {
 990         //
 991         // We are using two optoregs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
 992         // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
 993         // So we must adjust where to pick up the data to match the interpreter.
 994         //
 995         // Interpreter local[n] == MSW, local[n+1] == LSW however locals
 996         // are accessed as negative so LSW is at LOW address
 997 
 998         // ld_off is MSW so get LSW
 999         const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)?
1000                            next_off : ld_off;
1001         __ movq(r13, Address(saved_sp, offset));
1002         // st_off is LSW (i.e. reg.first())
1003         __ movq(Address(rsp, st_off), r13);
1004       }
1005     } else if (r_1->is_Register()) {  // Register argument
1006       Register r = r_1->as_Register();
1007       assert(r != rax, "must be different");
1008       if (r_2->is_valid()) {
1009         //
1010         // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
1011         // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
1012         // So we must adjust where to pick up the data to match the interpreter.
1013 
1014         const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)?
1015                            next_off : ld_off;
1016 
1017         // this can be a misaligned move
1018         __ movq(r, Address(saved_sp, offset));
1019       } else {
1020         // sign extend and use a full word?
1021         __ movl(r, Address(saved_sp, ld_off));
1022       }
1023     } else {
1024       if (!r_2->is_valid()) {
1025         __ movflt(r_1->as_XMMRegister(), Address(saved_sp, ld_off));
1026       } else {
1027         __ movdbl(r_1->as_XMMRegister(), Address(saved_sp, next_off));
1028       }
1029     }
1030   }
1031 
1032   __ push_cont_fastpath(); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
1033 
1034   // 6243940 We might end up in handle_wrong_method if
1035   // the callee is deoptimized as we race thru here. If that
1036   // happens we don't want to take a safepoint because the
1037   // caller frame will look interpreted and arguments are now
1038   // "compiled" so it is much better to make this transition
1039   // invisible to the stack walking code. Unfortunately if
1040   // we try and find the callee by normal means a safepoint
1041   // is possible. So we stash the desired callee in the thread
1042   // and the vm will find there should this case occur.
1043 
1044   __ movptr(Address(r15_thread, JavaThread::callee_target_offset()), rbx);
1045 
1046   // put Method* where a c2i would expect should we end up there
1047   // only needed because eof c2 resolve stubs return Method* as a result in
1048   // rax
1049   __ mov(rax, rbx);
1050   __ jmp(r11);
1051 }
1052 













1053 // ---------------------------------------------------------------
1054 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
1055                                                             int total_args_passed,
1056                                                             int comp_args_on_stack,
1057                                                             const BasicType *sig_bt,
1058                                                             const VMRegPair *regs,
1059                                                             AdapterFingerPrint* fingerprint) {






1060   address i2c_entry = __ pc();
1061 
1062   gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
1063 
1064   // -------------------------------------------------------------------------
1065   // Generate a C2I adapter.  On entry we know rbx holds the Method* during calls
1066   // to the interpreter.  The args start out packed in the compiled layout.  They
1067   // need to be unpacked into the interpreter layout.  This will almost always
1068   // require some stack space.  We grow the current (compiled) stack, then repack
1069   // the args.  We  finally end in a jump to the generic interpreter entry point.
1070   // On exit from the interpreter, the interpreter will restore our SP (lest the
1071   // compiled code, which relies solely on SP and not RBP, get sick).
1072 
1073   address c2i_unverified_entry = __ pc();

1074   Label skip_fixup;
1075 
1076   Register data = rax;
1077   Register receiver = j_rarg0;
1078   Register temp = rbx;
1079 
1080   {
1081     __ ic_check(1 /* end_alignment */);
1082     __ movptr(rbx, Address(data, CompiledICData::speculated_method_offset()));
1083     // Method might have been compiled since the call site was patched to
1084     // interpreted if that is the case treat it as a miss so we can get
1085     // the call site corrected.
1086     __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), NULL_WORD);
1087     __ jcc(Assembler::equal, skip_fixup);
1088     __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1089   }
1090 
1091   address c2i_entry = __ pc();


1092 
1093   // Class initialization barrier for static methods
1094   address c2i_no_clinit_check_entry = nullptr;
1095   if (VM_Version::supports_fast_class_init_checks()) {
1096     Label L_skip_barrier;
1097     Register method = rbx;
1098 
1099     { // Bypass the barrier for non-static methods
1100       Register flags = rscratch1;
1101       __ movl(flags, Address(method, Method::access_flags_offset()));
1102       __ testl(flags, JVM_ACC_STATIC);
1103       __ jcc(Assembler::zero, L_skip_barrier); // non-static
1104     }
1105 
1106     Register klass = rscratch1;
1107     __ load_method_holder(klass, method);
1108     __ clinit_barrier(klass, r15_thread, &L_skip_barrier /*L_fast_path*/);


1109 
1110     __ jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub())); // slow path




1111 
1112     __ bind(L_skip_barrier);
1113     c2i_no_clinit_check_entry = __ pc();

1114   }
1115 
1116   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1117   bs->c2i_entry_barrier(masm);
1118 
1119   gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);


1120 
1121   return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry, c2i_no_clinit_check_entry);
1122 }
1123 
1124 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
1125                                          VMRegPair *regs,
1126                                          int total_args_passed) {
1127 
1128 // We return the amount of VMRegImpl stack slots we need to reserve for all
1129 // the arguments NOT counting out_preserve_stack_slots.
1130 
1131 // NOTE: These arrays will have to change when c1 is ported
1132 #ifdef _WIN64
1133     static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1134       c_rarg0, c_rarg1, c_rarg2, c_rarg3
1135     };
1136     static const XMMRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
1137       c_farg0, c_farg1, c_farg2, c_farg3
1138     };
1139 #else
1140     static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1141       c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5

2228     const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes();
2229 
2230     // Get the handle (the 2nd argument)
2231     __ mov(oop_handle_reg, c_rarg1);
2232 
2233     // Get address of the box
2234 
2235     __ lea(lock_reg, Address(rsp, lock_slot_offset * VMRegImpl::stack_slot_size));
2236 
2237     // Load the oop from the handle
2238     __ movptr(obj_reg, Address(oop_handle_reg, 0));
2239 
2240     if (LockingMode == LM_MONITOR) {
2241       __ jmp(slow_path_lock);
2242     } else if (LockingMode == LM_LEGACY) {
2243       // Load immediate 1 into swap_reg %rax
2244       __ movl(swap_reg, 1);
2245 
2246       // Load (object->mark() | 1) into swap_reg %rax
2247       __ orptr(swap_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));




2248 
2249       // Save (object->mark() | 1) into BasicLock's displaced header
2250       __ movptr(Address(lock_reg, mark_word_offset), swap_reg);
2251 
2252       // src -> dest iff dest == rax else rax <- dest
2253       __ lock();
2254       __ cmpxchgptr(lock_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
2255       __ jcc(Assembler::equal, count_mon);
2256 
2257       // Hmm should this move to the slow path code area???
2258 
2259       // Test if the oopMark is an obvious stack pointer, i.e.,
2260       //  1) (mark & 3) == 0, and
2261       //  2) rsp <= mark < mark + os::pagesize()
2262       // These 3 tests can be done by evaluating the following
2263       // expression: ((mark - rsp) & (3 - os::vm_page_size())),
2264       // assuming both stack pointer and pagesize have their
2265       // least significant 2 bits clear.
2266       // NOTE: the oopMark is in swap_reg %rax as the result of cmpxchg
2267 

3569   julong *scratch = (julong *)alloca(total_allocation);
3570 
3571   // Local scratch arrays
3572   julong
3573     *a = scratch + 0 * longwords,
3574     *n = scratch + 1 * longwords,
3575     *m = scratch + 2 * longwords;
3576 
3577   reverse_words((julong *)a_ints, a, longwords);
3578   reverse_words((julong *)n_ints, n, longwords);
3579 
3580   if (len >= MONTGOMERY_SQUARING_THRESHOLD) {
3581     ::montgomery_square(a, n, m, (julong)inv, longwords);
3582   } else {
3583     ::montgomery_multiply(a, a, n, m, (julong)inv, longwords);
3584   }
3585 
3586   reverse_words(m, (julong *)m_ints, longwords);
3587 }
3588 















































































































3589 #if INCLUDE_JFR
3590 
3591 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3592 // It returns a jobject handle to the event writer.
3593 // The handle is dereferenced and the return value is the event writer oop.
3594 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
3595   enum layout {
3596     rbp_off,
3597     rbpH_off,
3598     return_off,
3599     return_off2,
3600     framesize // inclusive of return address
3601   };
3602 
3603   const char* name = SharedRuntime::stub_name(SharedStubId::jfr_write_checkpoint_id);
3604   CodeBuffer code(name, 1024, 64);
3605   MacroAssembler* masm = new MacroAssembler(&code);
3606   address start = __ pc();
3607 
3608   __ enter();

3661   __ reset_last_Java_frame(true);
3662 
3663   __ leave();
3664   __ ret(0);
3665 
3666   OopMapSet* oop_maps = new OopMapSet();
3667   OopMap* map = new OopMap(framesize, 1);
3668   oop_maps->add_gc_map(frame_complete, map);
3669 
3670   RuntimeStub* stub =
3671     RuntimeStub::new_runtime_stub(name,
3672                                   &code,
3673                                   frame_complete,
3674                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3675                                   oop_maps,
3676                                   false);
3677   return stub;
3678 }
3679 
3680 #endif // INCLUDE_JFR
3681 

  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 #include "precompiled.hpp"
  26 #ifndef _WINDOWS
  27 #include "alloca.h"
  28 #endif
  29 #include "asm/macroAssembler.hpp"
  30 #include "asm/macroAssembler.inline.hpp"
  31 #include "classfile/symbolTable.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"
  51 #include "runtime/jniHandles.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(to, val, rscratch1, r13, rbx, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
 853       __ pop(rbx);
 854       __ pop(r13);
 855     } else {
 856       __ store_sized_value(to, val, size_in_bytes);
 857     }
 858   } else {
 859     if (wide) {
 860       __ movdbl(to, r_1->as_XMMRegister());
 861     } else {
 862       __ movflt(to, r_1->as_XMMRegister());
 863     }
 864   }
 865 }
 866 
 867 static void gen_c2i_adapter(MacroAssembler *masm,
 868                             const GrowableArray<SigEntry>* sig_extended,


 869                             const VMRegPair *regs,
 870                             bool requires_clinit_barrier,
 871                             address& c2i_no_clinit_check_entry,
 872                             Label& skip_fixup,
 873                             address start,
 874                             OopMapSet* oop_maps,
 875                             int& frame_complete,
 876                             int& frame_size_in_words,
 877                             bool alloc_inline_receiver) {
 878   if (requires_clinit_barrier && VM_Version::supports_fast_class_init_checks()) {
 879     Label L_skip_barrier;
 880     Register method = rbx;
 881 
 882     { // Bypass the barrier for non-static methods
 883       Register flags = rscratch1;
 884       __ movl(flags, Address(method, Method::access_flags_offset()));
 885       __ testl(flags, JVM_ACC_STATIC);
 886       __ jcc(Assembler::zero, L_skip_barrier); // non-static
 887     }
 888 
 889     Register klass = rscratch1;
 890     __ load_method_holder(klass, method);
 891     __ clinit_barrier(klass, r15_thread, &L_skip_barrier /*L_fast_path*/);
 892 
 893     __ jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub())); // slow path
 894 
 895     __ bind(L_skip_barrier);
 896     c2i_no_clinit_check_entry = __ pc();
 897   }
 898 
 899   BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
 900   bs->c2i_entry_barrier(masm);
 901 
 902   // Before we get into the guts of the C2I adapter, see if we should be here
 903   // at all.  We've come from compiled code and are attempting to jump to the
 904   // interpreter, which means the caller made a static call to get here
 905   // (vcalls always get a compiled target if there is one).  Check for a
 906   // compiled target.  If there is one, we need to patch the caller's call.
 907   patch_callers_callsite(masm);
 908 
 909   __ bind(skip_fixup);
 910 
 911   if (InlineTypePassFieldsAsArgs) {
 912     // Is there an inline type argument?
 913     bool has_inline_argument = false;
 914     for (int i = 0; i < sig_extended->length() && !has_inline_argument; i++) {
 915       has_inline_argument = (sig_extended->at(i)._bt == T_METADATA);
 916     }
 917     if (has_inline_argument) {
 918       // There is at least an inline type argument: we're coming from
 919       // compiled code so we have no buffers to back the inline types.
 920       // Allocate the buffers here with a runtime call.
 921       OopMap* map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, /*save_vectors*/ false);
 922 
 923       frame_complete = __ offset();
 924 
 925       __ set_last_Java_frame(noreg, noreg, nullptr, rscratch1);
 926 
 927       __ mov(c_rarg0, r15_thread);
 928       __ mov(c_rarg1, rbx);
 929       __ mov64(c_rarg2, (int64_t)alloc_inline_receiver);
 930       __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::allocate_inline_types)));
 931 
 932       oop_maps->add_gc_map((int)(__ pc() - start), map);
 933       __ reset_last_Java_frame(false);
 934 
 935       RegisterSaver::restore_live_registers(masm);
 936 
 937       Label no_exception;
 938       __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
 939       __ jcc(Assembler::equal, no_exception);
 940 
 941       __ movptr(Address(r15_thread, JavaThread::vm_result_offset()), NULL_WORD);
 942       __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
 943       __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
 944 
 945       __ bind(no_exception);
 946 
 947       // We get an array of objects from the runtime call
 948       __ get_vm_result(rscratch2, r15_thread); // Use rscratch2 (r11) as temporary because rscratch1 (r10) is trashed by movptr()
 949       __ get_vm_result_2(rbx, r15_thread); // TODO: required to keep the callee Method live?
 950     }
 951   }
 952 
 953   // Since all args are passed on the stack, total_args_passed *
 954   // Interpreter::stackElementSize is the space we need.
 955   int total_args_passed = compute_total_args_passed_int(sig_extended);
 956   assert(total_args_passed >= 0, "total_args_passed is %d", total_args_passed);
 957 
 958   int extraspace = (total_args_passed * Interpreter::stackElementSize);
 959 
 960   // stack is aligned, keep it that way
 961   // This is not currently needed or enforced by the interpreter, but
 962   // we might as well conform to the ABI.
 963   extraspace = align_up(extraspace, 2*wordSize);
 964 
 965   // set senderSP value
 966   __ lea(r13, Address(rsp, wordSize));
 967 
 968 #ifdef ASSERT
 969   __ check_stack_alignment(r13, "sender stack not aligned");
 970 #endif
 971   if (extraspace > 0) {
 972     // Pop the return address
 973     __ pop(rax);
 974 
 975     __ subptr(rsp, extraspace);
 976 
 977     // Push the return address
 978     __ push(rax);
 979 
 980     // Account for the return address location since we store it first rather
 981     // than hold it in a register across all the shuffling
 982     extraspace += wordSize;
 983   }
 984 
 985 #ifdef ASSERT
 986   __ check_stack_alignment(rsp, "callee stack not aligned", wordSize, rax);
 987 #endif
 988 
 989   // Now write the args into the outgoing interpreter space






































 990 
 991   // next_arg_comp is the next argument from the compiler point of
 992   // view (inline type fields are passed in registers/on the stack). In
 993   // sig_extended, an inline type argument starts with: T_METADATA,
 994   // followed by the types of the fields of the inline type and T_VOID
 995   // to mark the end of the inline type. ignored counts the number of
 996   // T_METADATA/T_VOID. next_vt_arg is the next inline type argument:
 997   // used to get the buffer for that argument from the pool of buffers
 998   // we allocated above and want to pass to the
 999   // interpreter. next_arg_int is the next argument from the
1000   // interpreter point of view (inline types are passed by reference).
1001   for (int next_arg_comp = 0, ignored = 0, next_vt_arg = 0, next_arg_int = 0;
1002        next_arg_comp < sig_extended->length(); next_arg_comp++) {
1003     assert(ignored <= next_arg_comp, "shouldn't skip over more slots than there are arguments");
1004     assert(next_arg_int <= total_args_passed, "more arguments for the interpreter than expected?");
1005     BasicType bt = sig_extended->at(next_arg_comp)._bt;
1006     int st_off = (total_args_passed - next_arg_int) * Interpreter::stackElementSize;
1007     if (!InlineTypePassFieldsAsArgs || bt != T_METADATA) {
1008       int next_off = st_off - Interpreter::stackElementSize;
1009       const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : st_off;
1010       const VMRegPair reg_pair = regs[next_arg_comp-ignored];
1011       size_t size_in_bytes = reg_pair.second()->is_valid() ? 8 : 4;
1012       gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
1013                              size_in_bytes, reg_pair, Address(rsp, offset), extraspace, false);
1014       next_arg_int++;
1015 #ifdef ASSERT
1016       if (bt == T_LONG || bt == T_DOUBLE) {
1017         // Overwrite the unused slot with known junk
1018         __ mov64(rax, CONST64(0xdeadffffdeadaaaa));
1019         __ movptr(Address(rsp, st_off), rax);



1020       }















1021 #endif /* ASSERT */
1022     } else {
1023       ignored++;
1024       // get the buffer from the just allocated pool of buffers
1025       int index = arrayOopDesc::base_offset_in_bytes(T_OBJECT) + next_vt_arg * type2aelembytes(T_OBJECT);
1026       __ load_heap_oop(r14, Address(rscratch2, index));
1027       next_vt_arg++; next_arg_int++;
1028       int vt = 1;
1029       // write fields we get from compiled code in registers/stack
1030       // slots to the buffer: we know we are done with that inline type
1031       // argument when we hit the T_VOID that acts as an end of inline
1032       // type delimiter for this inline type. Inline types are flattened
1033       // so we might encounter embedded inline types. Each entry in
1034       // sig_extended contains a field offset in the buffer.
1035       Label L_null;
1036       do {
1037         next_arg_comp++;
1038         BasicType bt = sig_extended->at(next_arg_comp)._bt;
1039         BasicType prev_bt = sig_extended->at(next_arg_comp-1)._bt;
1040         if (bt == T_METADATA) {
1041           vt++;
1042           ignored++;
1043         } else if (bt == T_VOID &&
1044                    prev_bt != T_LONG &&
1045                    prev_bt != T_DOUBLE) {
1046           vt--;
1047           ignored++;
1048         } else {
1049           int off = sig_extended->at(next_arg_comp)._offset;
1050           if (off == -1) {
1051             // Nullable inline type argument, emit null check
1052             VMReg reg = regs[next_arg_comp-ignored].first();
1053             Label L_notNull;
1054             if (reg->is_stack()) {
1055               int ld_off = reg->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
1056               __ testb(Address(rsp, ld_off), 1);
1057             } else {
1058               __ testb(reg->as_Register(), 1);
1059             }
1060             __ jcc(Assembler::notZero, L_notNull);
1061             __ movptr(Address(rsp, st_off), 0);
1062             __ jmp(L_null);
1063             __ bind(L_notNull);
1064             continue;
1065           }
1066           assert(off > 0, "offset in object should be positive");
1067           size_t size_in_bytes = is_java_primitive(bt) ? type2aelembytes(bt) : wordSize;
1068           bool is_oop = is_reference_type(bt);
1069           gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
1070                                  size_in_bytes, regs[next_arg_comp-ignored], Address(r14, off), extraspace, is_oop);
1071         }
1072       } while (vt != 0);
1073       // pass the buffer to the interpreter
1074       __ movptr(Address(rsp, st_off), r14);
1075       __ bind(L_null);










1076     }
1077   }
1078 
1079   // Schedule the branch target address early.
1080   __ movptr(rcx, Address(rbx, in_bytes(Method::interpreter_entry_offset())));
1081   __ jmp(rcx);
1082 }
1083 
1084 static void range_check(MacroAssembler* masm, Register pc_reg, Register temp_reg,
1085                         address code_start, address code_end,
1086                         Label& L_ok) {
1087   Label L_fail;
1088   __ lea(temp_reg, AddressLiteral(code_start, relocInfo::none));
1089   __ cmpptr(pc_reg, temp_reg);
1090   __ jcc(Assembler::belowEqual, L_fail);
1091   __ lea(temp_reg, AddressLiteral(code_end, relocInfo::none));
1092   __ cmpptr(pc_reg, temp_reg);
1093   __ jcc(Assembler::below, L_ok);
1094   __ bind(L_fail);
1095 }
1096 
1097 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,

1098                                     int comp_args_on_stack,
1099                                     const GrowableArray<SigEntry>* sig,
1100                                     const VMRegPair *regs) {
1101 
1102   // Note: r13 contains the senderSP on entry. We must preserve it since
1103   // we may do a i2c -> c2i transition if we lose a race where compiled
1104   // code goes non-entrant while we get args ready.
1105   // In addition we use r13 to locate all the interpreter args as
1106   // we must align the stack to 16 bytes on an i2c entry else we
1107   // lose alignment we expect in all compiled code and register
1108   // save code can segv when fxsave instructions find improperly
1109   // aligned stack pointer.
1110 
1111   // Adapters can be frameless because they do not require the caller
1112   // to perform additional cleanup work, such as correcting the stack pointer.
1113   // An i2c adapter is frameless because the *caller* frame, which is interpreted,
1114   // routinely repairs its own stack pointer (from interpreter_frame_last_sp),
1115   // even if a callee has modified the stack pointer.
1116   // A c2i adapter is frameless because the *callee* frame, which is interpreted,
1117   // routinely repairs its caller's stack pointer (from sender_sp, which is set
1118   // up via the senderSP register).
1119   // In other words, if *either* the caller or callee is interpreted, we can

1170   // Convert 4-byte c2 stack slots to words.
1171   int comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
1172 
1173   if (comp_args_on_stack) {
1174     __ subptr(rsp, comp_words_on_stack * wordSize);
1175   }
1176 
1177   // Ensure compiled code always sees stack at proper alignment
1178   __ andptr(rsp, -16);
1179 
1180   // push the return address and misalign the stack that youngest frame always sees
1181   // as far as the placement of the call instruction
1182   __ push(rax);
1183 
1184   // Put saved SP in another register
1185   const Register saved_sp = rax;
1186   __ movptr(saved_sp, r11);
1187 
1188   // Will jump to the compiled code just as if compiled code was doing it.
1189   // Pre-load the register-jump target early, to schedule it better.
1190   __ movptr(r11, Address(rbx, in_bytes(Method::from_compiled_inline_offset())));
1191 
1192 #if INCLUDE_JVMCI
1193   if (EnableJVMCI) {
1194     // check if this call should be routed towards a specific entry point
1195     __ cmpptr(Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())), 0);
1196     Label no_alternative_target;
1197     __ jcc(Assembler::equal, no_alternative_target);
1198     __ movptr(r11, Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
1199     __ movptr(Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())), 0);
1200     __ bind(no_alternative_target);
1201   }
1202 #endif // INCLUDE_JVMCI
1203 
1204   int total_args_passed = sig->length();
1205 
1206   // Now generate the shuffle code.  Pick up all register args and move the
1207   // rest through the floating point stack top.
1208   for (int i = 0; i < total_args_passed; i++) {
1209     BasicType bt = sig->at(i)._bt;
1210     if (bt == T_VOID) {
1211       // Longs and doubles are passed in native word order, but misaligned
1212       // in the 32-bit build.
1213       BasicType prev_bt = (i > 0) ? sig->at(i-1)._bt : T_ILLEGAL;
1214       assert(i > 0 && (prev_bt == T_LONG || prev_bt == T_DOUBLE), "missing half");
1215       continue;
1216     }
1217 
1218     // Pick up 0, 1 or 2 words from SP+offset.
1219 
1220     assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
1221             "scrambled load targets?");
1222     // Load in argument order going down.
1223     int ld_off = (total_args_passed - i)*Interpreter::stackElementSize;
1224     // Point to interpreter value (vs. tag)
1225     int next_off = ld_off - Interpreter::stackElementSize;
1226     //
1227     //
1228     //
1229     VMReg r_1 = regs[i].first();
1230     VMReg r_2 = regs[i].second();
1231     if (!r_1->is_valid()) {
1232       assert(!r_2->is_valid(), "");
1233       continue;
1234     }

1236       // Convert stack slot to an SP offset (+ wordSize to account for return address )
1237       int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size + wordSize;
1238 
1239       // We can use r13 as a temp here because compiled code doesn't need r13 as an input
1240       // and if we end up going thru a c2i because of a miss a reasonable value of r13
1241       // will be generated.
1242       if (!r_2->is_valid()) {
1243         // sign extend???
1244         __ movl(r13, Address(saved_sp, ld_off));
1245         __ movptr(Address(rsp, st_off), r13);
1246       } else {
1247         //
1248         // We are using two optoregs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
1249         // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
1250         // So we must adjust where to pick up the data to match the interpreter.
1251         //
1252         // Interpreter local[n] == MSW, local[n+1] == LSW however locals
1253         // are accessed as negative so LSW is at LOW address
1254 
1255         // ld_off is MSW so get LSW
1256         const int offset = (bt==T_LONG||bt==T_DOUBLE)?
1257                            next_off : ld_off;
1258         __ movq(r13, Address(saved_sp, offset));
1259         // st_off is LSW (i.e. reg.first())
1260         __ movq(Address(rsp, st_off), r13);
1261       }
1262     } else if (r_1->is_Register()) {  // Register argument
1263       Register r = r_1->as_Register();
1264       assert(r != rax, "must be different");
1265       if (r_2->is_valid()) {
1266         //
1267         // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
1268         // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
1269         // So we must adjust where to pick up the data to match the interpreter.
1270 
1271         const int offset = (bt==T_LONG||bt==T_DOUBLE)?
1272                            next_off : ld_off;
1273 
1274         // this can be a misaligned move
1275         __ movq(r, Address(saved_sp, offset));
1276       } else {
1277         // sign extend and use a full word?
1278         __ movl(r, Address(saved_sp, ld_off));
1279       }
1280     } else {
1281       if (!r_2->is_valid()) {
1282         __ movflt(r_1->as_XMMRegister(), Address(saved_sp, ld_off));
1283       } else {
1284         __ movdbl(r_1->as_XMMRegister(), Address(saved_sp, next_off));
1285       }
1286     }
1287   }
1288 
1289   __ push_cont_fastpath(); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
1290 
1291   // 6243940 We might end up in handle_wrong_method if
1292   // the callee is deoptimized as we race thru here. If that
1293   // happens we don't want to take a safepoint because the
1294   // caller frame will look interpreted and arguments are now
1295   // "compiled" so it is much better to make this transition
1296   // invisible to the stack walking code. Unfortunately if
1297   // we try and find the callee by normal means a safepoint
1298   // is possible. So we stash the desired callee in the thread
1299   // and the vm will find there should this case occur.
1300 
1301   __ movptr(Address(r15_thread, JavaThread::callee_target_offset()), rbx);
1302 
1303   // put Method* where a c2i would expect should we end up there
1304   // only needed because of c2 resolve stubs return Method* as a result in
1305   // rax
1306   __ mov(rax, rbx);
1307   __ jmp(r11);
1308 }
1309 
1310 static void gen_inline_cache_check(MacroAssembler *masm, Label& skip_fixup) {
1311   Register data = rax;
1312   __ ic_check(1 /* end_alignment */);
1313   __ movptr(rbx, Address(data, CompiledICData::speculated_method_offset()));
1314 
1315   // Method might have been compiled since the call site was patched to
1316   // interpreted if that is the case treat it as a miss so we can get
1317   // the call site corrected.
1318   __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), NULL_WORD);
1319   __ jcc(Assembler::equal, skip_fixup);
1320   __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1321 }
1322 
1323 // ---------------------------------------------------------------
1324 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler* masm,

1325                                                             int comp_args_on_stack,
1326                                                             const GrowableArray<SigEntry>* sig,
1327                                                             const VMRegPair* regs,
1328                                                             const GrowableArray<SigEntry>* sig_cc,
1329                                                             const VMRegPair* regs_cc,
1330                                                             const GrowableArray<SigEntry>* sig_cc_ro,
1331                                                             const VMRegPair* regs_cc_ro,
1332                                                             AdapterFingerPrint* fingerprint,
1333                                                             AdapterBlob*& new_adapter,
1334                                                             bool allocate_code_blob) {
1335   address i2c_entry = __ pc();
1336   gen_i2c_adapter(masm, comp_args_on_stack, sig, regs);

1337 
1338   // -------------------------------------------------------------------------
1339   // Generate a C2I adapter.  On entry we know rbx holds the Method* during calls
1340   // to the interpreter.  The args start out packed in the compiled layout.  They
1341   // need to be unpacked into the interpreter layout.  This will almost always
1342   // require some stack space.  We grow the current (compiled) stack, then repack
1343   // the args.  We  finally end in a jump to the generic interpreter entry point.
1344   // On exit from the interpreter, the interpreter will restore our SP (lest the
1345   // compiled code, which relies solely on SP and not RBP, get sick).
1346 
1347   address c2i_unverified_entry        = __ pc();
1348   address c2i_unverified_inline_entry = __ pc();
1349   Label skip_fixup;
1350 
1351   gen_inline_cache_check(masm, skip_fixup);













1352 
1353   OopMapSet* oop_maps = new OopMapSet();
1354   int frame_complete = CodeOffsets::frame_never_safe;
1355   int frame_size_in_words = 0;
1356 
1357   // Scalarized c2i adapter with non-scalarized receiver (i.e., don't pack receiver)
1358   address c2i_no_clinit_check_entry = nullptr;
1359   address c2i_inline_ro_entry = __ pc();
1360   if (regs_cc != regs_cc_ro) {
1361     // No class init barrier needed because method is guaranteed to be non-static
1362     gen_c2i_adapter(masm, sig_cc_ro, regs_cc_ro, /* requires_clinit_barrier = */ false, c2i_no_clinit_check_entry,
1363                     skip_fixup, i2c_entry, oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
1364     skip_fixup.reset();
1365   }



1366 
1367   // Scalarized c2i adapter
1368   address c2i_entry        = __ pc();
1369   address c2i_inline_entry = __ pc();
1370   gen_c2i_adapter(masm, sig_cc, regs_cc, /* requires_clinit_barrier = */ true, c2i_no_clinit_check_entry,
1371                   skip_fixup, i2c_entry, oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ true);
1372 
1373   // Non-scalarized c2i adapter
1374   if (regs != regs_cc) {
1375     c2i_unverified_inline_entry = __ pc();
1376     Label inline_entry_skip_fixup;
1377     gen_inline_cache_check(masm, inline_entry_skip_fixup);
1378 
1379     c2i_inline_entry = __ pc();
1380     gen_c2i_adapter(masm, sig, regs, /* requires_clinit_barrier = */ true, c2i_no_clinit_check_entry,
1381                     inline_entry_skip_fixup, i2c_entry, oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
1382   }
1383 
1384   // The c2i adapters might safepoint and trigger a GC. The caller must make sure that
1385   // the GC knows about the location of oop argument locations passed to the c2i adapter.
1386   if (allocate_code_blob) {
1387     bool caller_must_gc_arguments = (regs != regs_cc);
1388     new_adapter = AdapterBlob::create(masm->code(), frame_complete, frame_size_in_words, oop_maps, caller_must_gc_arguments);
1389   }
1390 
1391   return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_inline_entry, c2i_inline_ro_entry, c2i_unverified_entry, c2i_unverified_inline_entry, c2i_no_clinit_check_entry);
1392 }
1393 
1394 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
1395                                          VMRegPair *regs,
1396                                          int total_args_passed) {
1397 
1398 // We return the amount of VMRegImpl stack slots we need to reserve for all
1399 // the arguments NOT counting out_preserve_stack_slots.
1400 
1401 // NOTE: These arrays will have to change when c1 is ported
1402 #ifdef _WIN64
1403     static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1404       c_rarg0, c_rarg1, c_rarg2, c_rarg3
1405     };
1406     static const XMMRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
1407       c_farg0, c_farg1, c_farg2, c_farg3
1408     };
1409 #else
1410     static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1411       c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5

2498     const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes();
2499 
2500     // Get the handle (the 2nd argument)
2501     __ mov(oop_handle_reg, c_rarg1);
2502 
2503     // Get address of the box
2504 
2505     __ lea(lock_reg, Address(rsp, lock_slot_offset * VMRegImpl::stack_slot_size));
2506 
2507     // Load the oop from the handle
2508     __ movptr(obj_reg, Address(oop_handle_reg, 0));
2509 
2510     if (LockingMode == LM_MONITOR) {
2511       __ jmp(slow_path_lock);
2512     } else if (LockingMode == LM_LEGACY) {
2513       // Load immediate 1 into swap_reg %rax
2514       __ movl(swap_reg, 1);
2515 
2516       // Load (object->mark() | 1) into swap_reg %rax
2517       __ orptr(swap_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
2518       if (EnableValhalla) {
2519         // Mask inline_type bit such that we go to the slow path if object is an inline type
2520         __ andptr(swap_reg, ~((int) markWord::inline_type_bit_in_place));
2521       }
2522 
2523       // Save (object->mark() | 1) into BasicLock's displaced header
2524       __ movptr(Address(lock_reg, mark_word_offset), swap_reg);
2525 
2526       // src -> dest iff dest == rax else rax <- dest
2527       __ lock();
2528       __ cmpxchgptr(lock_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
2529       __ jcc(Assembler::equal, count_mon);
2530 
2531       // Hmm should this move to the slow path code area???
2532 
2533       // Test if the oopMark is an obvious stack pointer, i.e.,
2534       //  1) (mark & 3) == 0, and
2535       //  2) rsp <= mark < mark + os::pagesize()
2536       // These 3 tests can be done by evaluating the following
2537       // expression: ((mark - rsp) & (3 - os::vm_page_size())),
2538       // assuming both stack pointer and pagesize have their
2539       // least significant 2 bits clear.
2540       // NOTE: the oopMark is in swap_reg %rax as the result of cmpxchg
2541 

3843   julong *scratch = (julong *)alloca(total_allocation);
3844 
3845   // Local scratch arrays
3846   julong
3847     *a = scratch + 0 * longwords,
3848     *n = scratch + 1 * longwords,
3849     *m = scratch + 2 * longwords;
3850 
3851   reverse_words((julong *)a_ints, a, longwords);
3852   reverse_words((julong *)n_ints, n, longwords);
3853 
3854   if (len >= MONTGOMERY_SQUARING_THRESHOLD) {
3855     ::montgomery_square(a, n, m, (julong)inv, longwords);
3856   } else {
3857     ::montgomery_multiply(a, a, n, m, (julong)inv, longwords);
3858   }
3859 
3860   reverse_words(m, (julong *)m_ints, longwords);
3861 }
3862 
3863 BufferedInlineTypeBlob* SharedRuntime::generate_buffered_inline_type_adapter(const InlineKlass* vk) {
3864   BufferBlob* buf = BufferBlob::create("inline types pack/unpack", 16 * K);
3865   CodeBuffer buffer(buf);
3866   short buffer_locs[20];
3867   buffer.insts()->initialize_shared_locs((relocInfo*)buffer_locs,
3868                                          sizeof(buffer_locs)/sizeof(relocInfo));
3869 
3870   MacroAssembler* masm = new MacroAssembler(&buffer);
3871 
3872   const Array<SigEntry>* sig_vk = vk->extended_sig();
3873   const Array<VMRegPair>* regs = vk->return_regs();
3874 
3875   int pack_fields_jobject_off = __ offset();
3876   // Resolve pre-allocated buffer from JNI handle.
3877   // We cannot do this in generate_call_stub() because it requires GC code to be initialized.
3878   __ movptr(rax, Address(r13, 0));
3879   __ resolve_jobject(rax /* value */,
3880                      r15_thread /* thread */,
3881                      r12 /* tmp */);
3882   __ movptr(Address(r13, 0), rax);
3883 
3884   int pack_fields_off = __ offset();
3885 
3886   int j = 1;
3887   for (int i = 0; i < sig_vk->length(); i++) {
3888     BasicType bt = sig_vk->at(i)._bt;
3889     if (bt == T_METADATA) {
3890       continue;
3891     }
3892     if (bt == T_VOID) {
3893       if (sig_vk->at(i-1)._bt == T_LONG ||
3894           sig_vk->at(i-1)._bt == T_DOUBLE) {
3895         j++;
3896       }
3897       continue;
3898     }
3899     int off = sig_vk->at(i)._offset;
3900     assert(off > 0, "offset in object should be positive");
3901     VMRegPair pair = regs->at(j);
3902     VMReg r_1 = pair.first();
3903     VMReg r_2 = pair.second();
3904     Address to(rax, off);
3905     if (bt == T_FLOAT) {
3906       __ movflt(to, r_1->as_XMMRegister());
3907     } else if (bt == T_DOUBLE) {
3908       __ movdbl(to, r_1->as_XMMRegister());
3909     } else {
3910       Register val = r_1->as_Register();
3911       assert_different_registers(to.base(), val, r14, r13, rbx, rscratch1);
3912       if (is_reference_type(bt)) {
3913         __ store_heap_oop(to, val, r14, r13, rbx, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
3914       } else {
3915         __ store_sized_value(to, r_1->as_Register(), type2aelembytes(bt));
3916       }
3917     }
3918     j++;
3919   }
3920   assert(j == regs->length(), "missed a field?");
3921 
3922   __ ret(0);
3923 
3924   int unpack_fields_off = __ offset();
3925 
3926   Label skip;
3927   __ testptr(rax, rax);
3928   __ jcc(Assembler::zero, skip);
3929 
3930   j = 1;
3931   for (int i = 0; i < sig_vk->length(); i++) {
3932     BasicType bt = sig_vk->at(i)._bt;
3933     if (bt == T_METADATA) {
3934       continue;
3935     }
3936     if (bt == T_VOID) {
3937       if (sig_vk->at(i-1)._bt == T_LONG ||
3938           sig_vk->at(i-1)._bt == T_DOUBLE) {
3939         j++;
3940       }
3941       continue;
3942     }
3943     int off = sig_vk->at(i)._offset;
3944     assert(off > 0, "offset in object should be positive");
3945     VMRegPair pair = regs->at(j);
3946     VMReg r_1 = pair.first();
3947     VMReg r_2 = pair.second();
3948     Address from(rax, off);
3949     if (bt == T_FLOAT) {
3950       __ movflt(r_1->as_XMMRegister(), from);
3951     } else if (bt == T_DOUBLE) {
3952       __ movdbl(r_1->as_XMMRegister(), from);
3953     } else if (bt == T_OBJECT || bt == T_ARRAY) {
3954       assert_different_registers(rax, r_1->as_Register());
3955       __ load_heap_oop(r_1->as_Register(), from);
3956     } else {
3957       assert(is_java_primitive(bt), "unexpected basic type");
3958       assert_different_registers(rax, r_1->as_Register());
3959       size_t size_in_bytes = type2aelembytes(bt);
3960       __ load_sized_value(r_1->as_Register(), from, size_in_bytes, bt != T_CHAR && bt != T_BOOLEAN);
3961     }
3962     j++;
3963   }
3964   assert(j == regs->length(), "missed a field?");
3965 
3966   __ bind(skip);
3967   __ ret(0);
3968 
3969   __ flush();
3970 
3971   return BufferedInlineTypeBlob::create(&buffer, pack_fields_off, pack_fields_jobject_off, unpack_fields_off);
3972 }
3973 
3974 #if INCLUDE_JFR
3975 
3976 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3977 // It returns a jobject handle to the event writer.
3978 // The handle is dereferenced and the return value is the event writer oop.
3979 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
3980   enum layout {
3981     rbp_off,
3982     rbpH_off,
3983     return_off,
3984     return_off2,
3985     framesize // inclusive of return address
3986   };
3987 
3988   const char* name = SharedRuntime::stub_name(SharedStubId::jfr_write_checkpoint_id);
3989   CodeBuffer code(name, 1024, 64);
3990   MacroAssembler* masm = new MacroAssembler(&code);
3991   address start = __ pc();
3992 
3993   __ enter();

4046   __ reset_last_Java_frame(true);
4047 
4048   __ leave();
4049   __ ret(0);
4050 
4051   OopMapSet* oop_maps = new OopMapSet();
4052   OopMap* map = new OopMap(framesize, 1);
4053   oop_maps->add_gc_map(frame_complete, map);
4054 
4055   RuntimeStub* stub =
4056     RuntimeStub::new_runtime_stub(name,
4057                                   &code,
4058                                   frame_complete,
4059                                   (framesize >> (LogBytesPerWord - LogBytesPerInt)),
4060                                   oop_maps,
4061                                   false);
4062   return stub;
4063 }
4064 
4065 #endif // INCLUDE_JFR

< prev index next >