inline void MacroAssembler::fmr_if_needed(FloatRegister rd, FloatRegister rs) { if (rs != rd) fmr(rd, rs); }
address InterpreterGenerator::generate_native_entry(bool synchronized) { const Register handler = r14; const Register function = r15; assert_different_registers(Rmethod, Rlocals, Rthread, Rstate, Rmonitor, handler, function); // We use the same code for synchronized and not if (native_entry) return native_entry; address start = __ pc(); // Allocate and initialize our stack frame. __ load (Rstate, 0); generate_compute_interpreter_state(true); // Make sure method is native and not abstract #ifdef ASSERT { Label ok; __ lwz (r0, Address(Rmethod, methodOopDesc::access_flags_offset())); __ andi_ (r0, r0, JVM_ACC_NATIVE | JVM_ACC_ABSTRACT); __ compare (r0, JVM_ACC_NATIVE); __ beq (ok); __ should_not_reach_here (__FILE__, __LINE__); __ bind (ok); } #endif // Lock if necessary Label not_synchronized_1; __ bne (CRsync, not_synchronized_1); __ lock_object (Rmonitor); __ bind (not_synchronized_1); // Get signature handler const Address signature_handler_addr( Rmethod, methodOopDesc::signature_handler_offset()); Label return_to_caller, got_signature_handler; __ load (handler, signature_handler_addr); __ compare (handler, 0); __ bne (got_signature_handler); __ call_VM (noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Rmethod, CALL_VM_NO_EXCEPTION_CHECKS); __ load (r0, Address(Rthread, Thread::pending_exception_offset())); __ compare (r0, 0); __ bne (return_to_caller); __ load (handler, signature_handler_addr); __ bind (got_signature_handler); // Get the native function entry point const Address native_function_addr( Rmethod, methodOopDesc::native_function_offset()); Label got_function; __ load (function, native_function_addr); #ifdef ASSERT { // InterpreterRuntime::prepare_native_call() sets the mirror // handle and native function address first and the signature // handler last, so function should always be set here. Label ok; __ compare (function, 0); __ bne (ok); __ should_not_reach_here (__FILE__, __LINE__); __ bind (ok); } #endif // Call signature handler __ mtctr (handler); __ bctrl (); __ mr (handler, r0); // Pass JNIEnv __ la (r3, Address(Rthread, JavaThread::jni_environment_offset())); // Pass mirror handle if static const Address oop_temp_addr = STATE(_oop_temp); Label not_static; __ bne (CRstatic, not_static); __ get_mirror_handle (r4); __ store (r4, oop_temp_addr); __ la (r4, oop_temp_addr); __ bind (not_static); // Set up the Java frame anchor __ set_last_Java_frame (); // Change the thread state to native const Address thread_state_addr(Rthread, JavaThread::thread_state_offset()); #ifdef ASSERT { Label ok; __ lwz (r0, thread_state_addr); __ compare (r0, _thread_in_Java); __ beq (ok); __ should_not_reach_here (__FILE__, __LINE__); __ bind (ok); } #endif __ load (r0, _thread_in_native); __ stw (r0, thread_state_addr); // Make the call __ call (function); __ fixup_after_potential_safepoint (); // The result will be in r3 (and maybe r4 on 32-bit) or f1. // Wherever it is, we need to store it before calling anything const Register r3_save = r16; #ifdef PPC32 const Register r4_save = r17; #endif const FloatRegister f1_save = f14; __ mr (r3_save, r3); #ifdef PPC32 __ mr (r4_save, r4); #endif __ fmr (f1_save, f1); // Switch thread to "native transition" state before reading the // synchronization state. This additional state is necessary // because reading and testing the synchronization state is not // atomic with respect to garbage collection. __ load (r0, _thread_in_native_trans); __ stw (r0, thread_state_addr); // Ensure the new state is visible to the VM thread. if(os::is_MP()) { if (UseMembar) __ sync (); else __ serialize_memory (r3, r4); } // Check for safepoint operation in progress and/or pending // suspend requests. We use a leaf call in order to leave // the last_Java_frame setup undisturbed. Label block, no_block; __ load (r3, (intptr_t) SafepointSynchronize::address_of_state()); __ lwz (r0, Address(r3, 0)); __ compare (r0, SafepointSynchronize::_not_synchronized); __ bne (block); __ lwz (r0, Address(Rthread, JavaThread::suspend_flags_offset())); __ compare (r0, 0); __ beq (no_block); __ bind (block); __ call_VM_leaf ( CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans)); __ fixup_after_potential_safepoint (); __ bind (no_block); // Change the thread state __ load (r0, _thread_in_Java); __ stw (r0, thread_state_addr); // Reset the frame anchor __ reset_last_Java_frame (); // If the result was an OOP then unbox it and store it in the frame // (where it will be safe from garbage collection) before we release // the handle it might be protected by Label non_oop, store_oop; __ load (r0, (intptr_t) AbstractInterpreter::result_handler(T_OBJECT)); __ compare (r0, handler); __ bne (non_oop); __ compare (r3_save, 0); __ beq (store_oop); __ load (r3_save, Address(r3_save, 0)); __ bind (store_oop); __ store (r3_save, STATE(_oop_temp)); __ bind (non_oop); // Reset handle block __ load (r3, Address(Rthread, JavaThread::active_handles_offset())); __ load (r0, 0); __ stw (r0, Address(r3, JNIHandleBlock::top_offset_in_bytes())); // If there is an exception we skip the result handler and return. // Note that this also skips unlocking which seems totally wrong, // but apparently this is what the asm interpreter does so we do // too. __ load (r0, Address(Rthread, Thread::pending_exception_offset())); __ compare (r0, 0); __ bne (return_to_caller); // Unlock if necessary Label not_synchronized_2; __ bne (CRsync, not_synchronized_2); __ unlock_object (Rmonitor); __ bind (not_synchronized_2); // Restore saved result and call the result handler __ mr (r3, r3_save); #ifdef PPC32 __ mr (r4, r4_save); #endif __ fmr (f1, f1_save); __ mtctr (handler); __ bctrl (); // Unwind the current activation and return __ bind (return_to_caller); generate_unwind_interpreter_state(); __ blr (); native_entry = start; return start; }
address AbstractInterpreterGenerator::generate_slow_signature_handler() { // Slow_signature handler that respects the PPC C calling conventions. // // We get called by the native entry code with our output register // area == 8. First we call InterpreterRuntime::get_result_handler // to copy the pointer to the signature string temporarily to the // first C-argument and to return the result_handler in // R3_RET. Since native_entry will copy the jni-pointer to the // first C-argument slot later on, it is OK to occupy this slot // temporarilly. Then we copy the argument list on the java // expression stack into native varargs format on the native stack // and load arguments into argument registers. Integer arguments in // the varargs vector will be sign-extended to 8 bytes. // // On entry: // R3_ARG1 - intptr_t* Address of java argument list in memory. // R15_prev_state - BytecodeInterpreter* Address of interpreter state for // this method // R19_method // // On exit (just before return instruction): // R3_RET - contains the address of the result_handler. // R4_ARG2 - is not updated for static methods and contains "this" otherwise. // R5_ARG3-R10_ARG8: - When the (i-2)th Java argument is not of type float or double, // ARGi contains this argument. Otherwise, ARGi is not updated. // F1_ARG1-F13_ARG13 - contain the first 13 arguments of type float or double. const int LogSizeOfTwoInstructions = 3; // FIXME: use Argument:: GL: Argument names different numbers! const int max_fp_register_arguments = 13; const int max_int_register_arguments = 6; // first 2 are reserved const Register arg_java = R21_tmp1; const Register arg_c = R22_tmp2; const Register signature = R23_tmp3; // is string const Register sig_byte = R24_tmp4; const Register fpcnt = R25_tmp5; const Register argcnt = R26_tmp6; const Register intSlot = R27_tmp7; const Register target_sp = R28_tmp8; const FloatRegister floatSlot = F0; address entry = __ function_entry(); __ save_LR_CR(R0); __ save_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14)); // We use target_sp for storing arguments in the C frame. __ mr(target_sp, R1_SP); __ push_frame_reg_args_nonvolatiles(0, R11_scratch1); __ mr(arg_java, R3_ARG1); __ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::get_signature), R16_thread, R19_method); // Signature is in R3_RET. Signature is callee saved. __ mr(signature, R3_RET); // Get the result handler. __ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::get_result_handler), R16_thread, R19_method); { Label L; // test if static // _access_flags._flags must be at offset 0. // TODO PPC port: requires change in shared code. //assert(in_bytes(AccessFlags::flags_offset()) == 0, // "MethodDesc._access_flags == MethodDesc._access_flags._flags"); // _access_flags must be a 32 bit value. assert(sizeof(AccessFlags) == 4, "wrong size"); __ lwa(R11_scratch1/*access_flags*/, method_(access_flags)); // testbit with condition register. __ testbitdi(CCR0, R0, R11_scratch1/*access_flags*/, JVM_ACC_STATIC_BIT); __ btrue(CCR0, L); // For non-static functions, pass "this" in R4_ARG2 and copy it // to 2nd C-arg slot. // We need to box the Java object here, so we use arg_java // (address of current Java stack slot) as argument and don't // dereference it as in case of ints, floats, etc. __ mr(R4_ARG2, arg_java); __ addi(arg_java, arg_java, -BytesPerWord); __ std(R4_ARG2, _abi(carg_2), target_sp); __ bind(L); } // Will be incremented directly after loop_start. argcnt=0 // corresponds to 3rd C argument. __ li(argcnt, -1); // arg_c points to 3rd C argument __ addi(arg_c, target_sp, _abi(carg_3)); // no floating-point args parsed so far __ li(fpcnt, 0); Label move_intSlot_to_ARG, move_floatSlot_to_FARG; Label loop_start, loop_end; Label do_int, do_long, do_float, do_double, do_dontreachhere, do_object, do_array, do_boxed; // signature points to '(' at entry #ifdef ASSERT __ lbz(sig_byte, 0, signature); __ cmplwi(CCR0, sig_byte, '('); __ bne(CCR0, do_dontreachhere); #endif __ bind(loop_start); __ addi(argcnt, argcnt, 1); __ lbzu(sig_byte, 1, signature); __ cmplwi(CCR0, sig_byte, ')'); // end of signature __ beq(CCR0, loop_end); __ cmplwi(CCR0, sig_byte, 'B'); // byte __ beq(CCR0, do_int); __ cmplwi(CCR0, sig_byte, 'C'); // char __ beq(CCR0, do_int); __ cmplwi(CCR0, sig_byte, 'D'); // double __ beq(CCR0, do_double); __ cmplwi(CCR0, sig_byte, 'F'); // float __ beq(CCR0, do_float); __ cmplwi(CCR0, sig_byte, 'I'); // int __ beq(CCR0, do_int); __ cmplwi(CCR0, sig_byte, 'J'); // long __ beq(CCR0, do_long); __ cmplwi(CCR0, sig_byte, 'S'); // short __ beq(CCR0, do_int); __ cmplwi(CCR0, sig_byte, 'Z'); // boolean __ beq(CCR0, do_int); __ cmplwi(CCR0, sig_byte, 'L'); // object __ beq(CCR0, do_object); __ cmplwi(CCR0, sig_byte, '['); // array __ beq(CCR0, do_array); // __ cmplwi(CCR0, sig_byte, 'V'); // void cannot appear since we do not parse the return type // __ beq(CCR0, do_void); __ bind(do_dontreachhere); __ unimplemented("ShouldNotReachHere in slow_signature_handler", 120); __ bind(do_array); { Label start_skip, end_skip; __ bind(start_skip); __ lbzu(sig_byte, 1, signature); __ cmplwi(CCR0, sig_byte, '['); __ beq(CCR0, start_skip); // skip further brackets __ cmplwi(CCR0, sig_byte, '9'); __ bgt(CCR0, end_skip); // no optional size __ cmplwi(CCR0, sig_byte, '0'); __ bge(CCR0, start_skip); // skip optional size __ bind(end_skip); __ cmplwi(CCR0, sig_byte, 'L'); __ beq(CCR0, do_object); // for arrays of objects, the name of the object must be skipped __ b(do_boxed); // otherwise, go directly to do_boxed } __ bind(do_object); { Label L; __ bind(L); __ lbzu(sig_byte, 1, signature); __ cmplwi(CCR0, sig_byte, ';'); __ bne(CCR0, L); } // Need to box the Java object here, so we use arg_java (address of // current Java stack slot) as argument and don't dereference it as // in case of ints, floats, etc. Label do_null; __ bind(do_boxed); __ ld(R0,0, arg_java); __ cmpdi(CCR0, R0, 0); __ li(intSlot,0); __ beq(CCR0, do_null); __ mr(intSlot, arg_java); __ bind(do_null); __ std(intSlot, 0, arg_c); __ addi(arg_java, arg_java, -BytesPerWord); __ addi(arg_c, arg_c, BytesPerWord); __ cmplwi(CCR0, argcnt, max_int_register_arguments); __ blt(CCR0, move_intSlot_to_ARG); __ b(loop_start); __ bind(do_int); __ lwa(intSlot, 0, arg_java); __ std(intSlot, 0, arg_c); __ addi(arg_java, arg_java, -BytesPerWord); __ addi(arg_c, arg_c, BytesPerWord); __ cmplwi(CCR0, argcnt, max_int_register_arguments); __ blt(CCR0, move_intSlot_to_ARG); __ b(loop_start); __ bind(do_long); __ ld(intSlot, -BytesPerWord, arg_java); __ std(intSlot, 0, arg_c); __ addi(arg_java, arg_java, - 2 * BytesPerWord); __ addi(arg_c, arg_c, BytesPerWord); __ cmplwi(CCR0, argcnt, max_int_register_arguments); __ blt(CCR0, move_intSlot_to_ARG); __ b(loop_start); __ bind(do_float); __ lfs(floatSlot, 0, arg_java); #if defined(LINUX) // Linux uses ELF ABI. Both original ELF and ELFv2 ABIs have float // in the least significant word of an argument slot. #if defined(VM_LITTLE_ENDIAN) __ stfs(floatSlot, 0, arg_c); #else __ stfs(floatSlot, 4, arg_c); #endif #elif defined(AIX) // Although AIX runs on big endian CPU, float is in most significant // word of an argument slot. __ stfs(floatSlot, 0, arg_c); #else #error "unknown OS" #endif __ addi(arg_java, arg_java, -BytesPerWord); __ addi(arg_c, arg_c, BytesPerWord); __ cmplwi(CCR0, fpcnt, max_fp_register_arguments); __ blt(CCR0, move_floatSlot_to_FARG); __ b(loop_start); __ bind(do_double); __ lfd(floatSlot, - BytesPerWord, arg_java); __ stfd(floatSlot, 0, arg_c); __ addi(arg_java, arg_java, - 2 * BytesPerWord); __ addi(arg_c, arg_c, BytesPerWord); __ cmplwi(CCR0, fpcnt, max_fp_register_arguments); __ blt(CCR0, move_floatSlot_to_FARG); __ b(loop_start); __ bind(loop_end); __ pop_frame(); __ restore_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14)); __ restore_LR_CR(R0); __ blr(); Label move_int_arg, move_float_arg; __ bind(move_int_arg); // each case must consist of 2 instructions (otherwise adapt LogSizeOfTwoInstructions) __ mr(R5_ARG3, intSlot); __ b(loop_start); __ mr(R6_ARG4, intSlot); __ b(loop_start); __ mr(R7_ARG5, intSlot); __ b(loop_start); __ mr(R8_ARG6, intSlot); __ b(loop_start); __ mr(R9_ARG7, intSlot); __ b(loop_start); __ mr(R10_ARG8, intSlot); __ b(loop_start); __ bind(move_float_arg); // each case must consist of 2 instructions (otherwise adapt LogSizeOfTwoInstructions) __ fmr(F1_ARG1, floatSlot); __ b(loop_start); __ fmr(F2_ARG2, floatSlot); __ b(loop_start); __ fmr(F3_ARG3, floatSlot); __ b(loop_start); __ fmr(F4_ARG4, floatSlot); __ b(loop_start); __ fmr(F5_ARG5, floatSlot); __ b(loop_start); __ fmr(F6_ARG6, floatSlot); __ b(loop_start); __ fmr(F7_ARG7, floatSlot); __ b(loop_start); __ fmr(F8_ARG8, floatSlot); __ b(loop_start); __ fmr(F9_ARG9, floatSlot); __ b(loop_start); __ fmr(F10_ARG10, floatSlot); __ b(loop_start); __ fmr(F11_ARG11, floatSlot); __ b(loop_start); __ fmr(F12_ARG12, floatSlot); __ b(loop_start); __ fmr(F13_ARG13, floatSlot); __ b(loop_start); __ bind(move_intSlot_to_ARG); __ sldi(R0, argcnt, LogSizeOfTwoInstructions); __ load_const(R11_scratch1, move_int_arg); // Label must be bound here. __ add(R11_scratch1, R0, R11_scratch1); __ mtctr(R11_scratch1/*branch_target*/); __ bctr(); __ bind(move_floatSlot_to_FARG); __ sldi(R0, fpcnt, LogSizeOfTwoInstructions); __ addi(fpcnt, fpcnt, 1); __ load_const(R11_scratch1, move_float_arg); // Label must be bound here. __ add(R11_scratch1, R0, R11_scratch1); __ mtctr(R11_scratch1/*branch_target*/); __ bctr(); return entry; }