drlvm/modules/vm/src/main/native/jitrino/shared/jet/cg_instr.cpp

/*
 *  Licensed to the Apache Software Foundation (ASF) under one or more
 *  contributor license agreements.  See the NOTICE file distributed with
 *  this work for additional information regarding copyright ownership.
 *  The ASF licenses this file to You under the Apache License, Version 2.0
 *  (the "License"); you may not use this file except in compliance with
 *  the License.  You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 *  Unless required by applicable law or agreed to in writing, software
 *  distributed under the License is distributed on an "AS IS" BASIS,
 *  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 *  See the License for the specific language governing permissions and
 *  limitations under the License.
 */
/**
 * @author Alexander Astapchuk
 */
/**
 * @file
 * @brief CodeGen's routines for instrumentation and profiling.
 */
 
#include "cg.h"
#include "trace.h"
#include "open/vm_field_access.h"
#include "jit_runtime_support.h"
#include "open/vm_class_loading.h"
#include "open/vm_class_info.h"
#include "open/vm_ee.h"

namespace Jitrino {
namespace Jet {

//Number of nops to put before counter instruction to be able to atomically
//replace beginning of the instruction with jump
#define NOPS_PER_COUNTER 1

void CodeGen::gen_prof_be(void)
{
    if (!is_set(JMF_PROF_ENTRY_BE)) {
        return;
    }

#ifdef HYX86
    AR addr = ar_x;
    int off = (int)m_p_backedge_counter;
#else
    AR addr = valloc(jobj);
    movp(addr, m_p_backedge_counter);
    int off = 0;
#endif
    U_32 offset = (U_32)(m_codeStream.ip() - m_codeStream.data() - m_bbinfo->ipoff); //store offsets inside of BB now. Fix it to method's offset after code layout
    //put number of nops to align counter instruction
    nop(NOPS_PER_COUNTER);
    alu(alu_add, Opnd(i32, addr, off), 1);

   //store information about profiling counters
    U_32 new_offset = (U_32)(m_codeStream.ip() - m_codeStream.data() - m_bbinfo->ipoff);
    ProfileCounterInfo info;
    info.bb = m_bbinfo;
    info.offsetInfo = ProfileCounterInfo::createOffsetInfo(new_offset - offset, offset);
    m_profileCountersMap.push_back(info);
}

void CodeGen::gen_gc_safe_point()
{
    if (!is_set(JMF_BBPOLLING)) {
        return;
    }
    if (m_bbstate->seen_gcpt) {
        if (is_set(DBG_TRACE_CG)) {
            dbg(";;>gc.safepoint - skipped\n");
        }
        return;
    }
    m_bbstate->seen_gcpt = true;
    // On Windows we could use a bit, but tricky way - we know about VM 
    // internals and we know how Windows manages TIB, and thus we can get a 
    // direct access to the flag, without need to call VM:
    //      mov eax, fs:14
    //      test [eax+rt_suspend_req_flag_offset], 0
    // I don't believe this will gain any improvements for .jet, so using 
    // portable and 'official' way:
    gen_call_vm(platform_v, rt_helper_get_tls_base_ptr, 0);
    // The address of flag is now in gr_ret
    AR gr_ret = platform_v.ret_reg(0);
    Opnd mem(i32, gr_ret, rt_suspend_req_flag_offset);
    alu(alu_cmp, mem, Opnd(0));
    unsigned br_off = br(z, 0, 0, taken);
    gen_call_vm_restore(false, helper_v, rt_helper_gc_safepoint, 0);
    patch(br_off, ip());
}

void CodeGen::gen_modification_watchpoint(JavaByteCodes opcode, jtype jt, Field_Handle fld) {

    unsigned ref_depth = is_wide(jt) ? 2 : 1;
    bool field_op = (opcode == OPCODE_PUTFIELD) ? true : false;

    // Check whether VM need modification notifications

    char* fld_tr_add;
    char fld_tr_mask;
    field_get_track_modification_flag(fld, &fld_tr_add, &fld_tr_mask);

    Val fld_track_mask((int)fld_tr_mask);

    AR fld_trackAr = valloc(jobj);
    movp(fld_trackAr, (void*)fld_tr_add);
    Opnd fld_track_opnd(i32, fld_trackAr, 0);
    //mov(fld_trackAr, Opnd(0xFFFFFFFF)); // Emulation to check access flag enabled
    //Opnd fld_track_opnd(fld_trackAr);
    
    rlock(fld_track_opnd);

    alu(alu_test, fld_track_opnd, fld_track_mask.as_opnd());
    runlock(fld_track_opnd);

    unsigned br_off = br(z, 0, 0, taken);
    
    // Store all scratch registers and operand stack state
    BBState saveBB;
    push_all_state(&saveBB);

    //JVMTI helper takes field handle, method handle, byte code location, pointer
    //to reference for fields or NULL for statics, pointer to field value

#ifndef HYX86_64
    // Workaround since do_mov do not put jlong on stack in gen_args on ia32
    SYNC_FIRST(static const CallSig cs_ti_fmodif(CCONV_HELPERS, jvoid, jobj, jobj, i32, i32, jobj, jobj));
    Val vlocation((jlong)m_pc);
    Val vlocationHi((jlong)0);
#else
    SYNC_FIRST(static const CallSig cs_ti_fmodif(CCONV_HELPERS, jvoid, jobj, jobj, i64, jobj, jobj));
    Val vlocation((jlong)m_pc);
#endif

    rlock(cs_ti_fmodif);

    AR fieldValBaseAr = valloc(jobj);
    Val fieldValPtr = Val(jobj, fieldValBaseAr);
    rlock(fieldValPtr);
    

    if (jt != jvoid) {
        // Make sure the value item is on the memory
        vswap(0);
        if (is_big(jt)) {
            vswap(1);
        }
        const Val& s = vstack(0);
        assert(s.is_mem());
        lea(fieldValPtr.as_opnd(), s.as_opnd());
    } else {
        Opnd stackTop(jobj, m_base, voff(m_stack.unused()));
        lea(fieldValPtr.as_opnd(), stackTop);
    }

    Val vfield(jobj, fld);
    Val vmeth(jobj, m_method);
    Val vobject =  Val(jobj, NULL);

    if (field_op) {
        vobject = vstack(ref_depth);
    }
#ifndef HYX86_64
    // Workaround since do_mov do not put jlong on stack in gen_args on ia32
    gen_args(cs_ti_fmodif, 0, &vfield, &vmeth, &vlocationHi, &vlocation, &vobject, &fieldValPtr);
#else
    gen_args(cs_ti_fmodif, 0, &vfield, &vmeth, &vlocation, &vobject, &fieldValPtr);
#endif

    // 2. Park all locals and operand stack
    vpark();
    // Store gc info
    gen_gc_stack(-1, true);

    // 3. Call VM
    AR gr = valloc(jobj);
    call( is_set(DBG_CHECK_STACK), gr, rt_helper_ti_field_modification, cs_ti_fmodif, cs_ti_fmodif.count());

    runlock(fieldValPtr);
    runlock(cs_ti_fmodif);
    
    //Restore operand stack state and scratch registers
    pop_all_state(&saveBB);

    patch(br_off, ip());
}


void CodeGen::gen_access_watchpoint(JavaByteCodes opcode, jtype jt, Field_Handle fld) {

    bool field_op = (opcode == OPCODE_GETFIELD) ? true : false;

    // Check whether VM need access notifications
    
    char* fld_tr_add;
    char fld_tr_mask;
    field_get_track_access_flag(fld, &fld_tr_add, &fld_tr_mask);

    Val fld_track_mask((int)fld_tr_mask);

    AR fld_trackAr = valloc(jobj);
    rlock(fld_trackAr);
    
    //mov(fld_trackAr, Opnd(0xFFFFFFFF)); // Emulation to check access flag enabled
    //Opnd fld_track_opnd(fld_trackAr);
    movp(fld_trackAr, (void*)fld_tr_add);
    Opnd fld_track_opnd(i32, fld_trackAr, 0);
    alu(alu_test, fld_track_opnd, fld_track_mask.as_opnd());

    runlock(fld_trackAr);

    unsigned br_off = br(z, 0, 0, taken);

    // Store all scratch registers and operand stack state
    BBState saveBB;
    push_all_state(&saveBB);


    //JVMTI helper takes field handle, method handle, byte code location, pointer
    //to reference for fields or NULL for statics

#ifndef HYX86_64
    // Workaround since do_mov do not put jlong on stack in gen_args on ia32
    SYNC_FIRST(static const CallSig cs_ti_faccess(CCONV_HELPERS, jvoid, jobj, jobj, i32, i32, jobj));
    Val vlocation((jlong)m_pc);
    Val vlocationHi((jlong)0);
#else
    SYNC_FIRST(static const CallSig cs_ti_faccess(CCONV_HELPERS, jvoid, jobj, jobj, i64, jobj));
    Val vlocation((jlong)m_pc);
#endif
    rlock(cs_ti_faccess);


    Val vfield(jobj, fld);
    Val vmeth(jobj, m_method);
    Val vobject =  Val(jobj, NULL);

    if (field_op) {
        vobject = vstack(0);
    }
    
#ifndef HYX86_64
    // Workaround since do_mov do not put jlong on stack in gen_args on ia32
    gen_args(cs_ti_faccess, 0, &vfield, &vmeth, &vlocationHi, &vlocation, &vobject);
#else
    gen_args(cs_ti_faccess, 0, &vfield, &vmeth, &vlocation, &vobject);
#endif

    // 2. Park all locals and operand stack
    vpark();
    // Store gc info
    gen_gc_stack(-1, true);

    // 3. Call VM
    rlock(cs_ti_faccess);
    AR gr = valloc(jobj);
    call( is_set(DBG_CHECK_STACK), gr, rt_helper_ti_field_access, cs_ti_faccess, cs_ti_faccess.count());
    runlock(cs_ti_faccess);

    //Restore operand stack state and scratch registers
    pop_all_state(&saveBB);


    patch(br_off, ip());
}


void CodeGen::push_all_state(BBState *saveBB){
    *saveBB = *m_bbstate;
    // 1. store scratch registers in a secret place
    // 2. park everything
    // 3. call whatever
    // 4. restore scratch regs from the secret place
    // 5. restore the state for callee-save registers
    //-----------------------------------------------
    // 1. 
    bool saveScratch = true;
    for (unsigned i=0; i<ar_num; i++) {
        AR ar = _ar(i);
        if (is_callee_save(ar)) continue;
        if (saveScratch && rrefs(ar) != 0) {
            jtype jt = is_f(ar) ? dbl64 : jobj;
            Opnd mem(jt, m_base, voff(m_stack.spill(ar)));
            Opnd reg(jt, ar);
            mov(mem, reg);
        }
        if (rlocks(ar) != 0) {
            runlock(ar, true);
        }
    }
}

void CodeGen::pop_all_state(BBState* saveBB) {
    bool saveScratch = true;
    // 4.
    *m_bbstate = *saveBB;
    // restore the registers state
    for (unsigned i=0; saveScratch && i<ar_num; i++) {
        AR ar = _ar(i);
        if (is_callee_save(ar)) continue;
        if (rrefs(ar) != 0) {
            jtype jt = is_f(ar) ? dbl64 : jobj;
            Opnd mem(jt, m_base, voff(m_stack.spill(ar)));
            Opnd reg(jt, ar);
            mov(reg, mem);
            }
        }
// 5. 
// Actually nothing to do here.
// If we had a local var on register before, then it's still on the reg
// If we had the var with static assignment which was in memory, before,
// then the memory was not corrupted.
// So, just nothing to do with callee-save regs
//
}

void CodeGen::gen_write_barrier(JavaByteCodes opcode, Field_Handle fieldHandle, Opnd fieldSlotAddress)
{
    //
    // TODO: the WB implementation is expected to perform the write
    // so, we can skip the code e.g. AASTORE after the succesfull
    // write barrier.
    //
    if (m_jframe->top() != jobj) {
        // Item on top is not Object - GC does not care, nothing to do.
        return;
    }

    const bool doGenWB4J = get_bool_arg("wb4j", false);
    const bool doGenWB4C = 
        (get_bool_arg("wb4c", false) || 
        // .exe_insert_write_barriers presumed to be WB4C
        compilation_params.exe_insert_write_barriers) &&
        !doGenWB4J;
    //  ^^^^^
    // When GC requires WBs, then VM always sets .exe_insert_write_barriers
    // May have conflict here - both doGenWB4J and doGenWB4C set. To avoid
    // such conflict, here's the rule: 
    // explicitly requested wb4j has a higher priority 
    //
    
    // No request to generate WBs - nothing to do
    if (!doGenWB4J && !doGenWB4C) {
        return;
    }
    
    const bool wb4c_skip_statics = get_bool_arg("wb4c.skip_statics", true);
    if (doGenWB4C && (opcode == OPCODE_PUTSTATIC) && wb4c_skip_statics) {
        // Currently, in DRLVM, statics are allocated outside of GC heap, 
        // no need to report them
        return;        
    }

    // WB4C has the following signature:
    //(object written to, slot written to, value written to slot)
    SYNC_FIRST(static const CallSig wb4c_sig(CCONV_CDECL, jvoid, jobj, jobj, jobj));
    //static char* wb4c_helper = xxx_gc_heap_slot_write_ref
    static char* wb4c_helper = (char*)vm_helper_get_addr(VM_RT_GC_HEAP_WRITE_REF);

    if (doGenWB4C && (NULL == wb4c_helper)) {
        // WB4C requested, but VM knows nothing about such helper
        assert(false);
        return;
    }
    // WB4J has the following signature:
    //(object written to, slot written to, value written to slot, metaA, metaB, mode)
    SYNC_FIRST(static const CallSig wb4j_sig(CCONV_CDECL, jvoid, jobj, jobj, jobj, i32, i32, i32));
    static char* wb4j_helper = NULL;
    

    static const char wbKlassName[] = "org/mmtk/vm/Barriers";
    static const char wbMethName[] = "performWriteInBarrier";
    static const char wbMethSig[] = "("
        "Lorg/vmmagic/unboxed/ObjectReference;"
        "Lorg/vmmagic/unboxed/Address;"
        "Lorg/vmmagic/unboxed/ObjectReference;"
        "Lorg/vmmagic/unboxed/Offset;"
        "II"
        ")V";
    static Class_Handle wbKlass = NULL;
    static Method_Handle wbMeth = NULL;
    static char* wbMethAddr = NULL;

    if (doGenWB4J && NULL == wbMethAddr) {
        wbKlass = vm_load_class_with_bootstrap(wbKlassName);
        if (wbKlass != NULL) {
            wbMeth = class_lookup_method_recursively(wbKlass, wbMethName, wbMethSig);
        }
        if (wbMeth != NULL) {
            if (wbMeth != NULL) {
                // Must pre-compile the method before a first usage.
                // Otherwise, if the first usage is WB for PUTSTATIC, then
                // address of the static field goes as 'slot' argument.
                // When VM tries to compile the method, then it tries to 
                // GC-protect args of the method, including this 'slot'.
                // Statics lies outside of the GC heap, so we'll get an 
                // assert that the reference is invalid.
                vm_compile_method(m_hjit, wbMeth);
                wbMethAddr = *(char**)method_get_indirect_address(wbMeth);
            }
        }
        if (NULL == wbMethAddr) {
            // Something terrible happened
#ifdef _DEBUG
            static bool printWarning = true;
            if (printWarning) {
                printWarning = false;
                fprintf(stderr, 
                    "*** WARNING: ***\n"
                    "WB4J requested, but unable to find\n"
                    "\t%s::%s(%s).\n"
                    "(No WB code is generated at all)\n"
                    "Did you put %s into *bootclasspath*?\n\n",
                    wbKlassName, wbMethName, wbMethSig, wbKlassName);
            }
#endif
            return;
        }
        //FIXME: This stores address statically - will not work in case of 
        // writeBarrier method get recompiled!
        wb4j_helper = wbMethAddr;
    }

    const CallSig& csig = doGenWB4C ? wb4c_sig : wb4j_sig;
    void* wb_helper = doGenWB4C ? wb4c_helper : wb4j_helper;
    
    Val baseObject, slotAddress, value;
    // operand stack for ASTORE:    arr, idx, ref
    // operand stack for PUTFIELD:  base, ref
    // operand stack for PUTSTATIC: ref
    rlock(csig);
    if (is_set(DBG_TRACE_CG)) { dbg(";;> write.barrier\n"); }
    
    int mode = -1;
    
    if (opcode == OPCODE_PUTFIELD) {
        mode = 0;
        baseObject = vstack(1, true);
        rlock(baseObject);
    }
    else if (opcode == OPCODE_PUTSTATIC) {
        mode = 1;
        baseObject = Val(jobj, NULL_REF);
        rlock(baseObject);
    }
    else if (opcode == OPCODE_AASTORE) {
        mode = 2;

        baseObject = vstack(2, true);
        rlock(baseObject);
        
        const Val& idx = vstack(1, vis_mem(1));
        
        AR base = baseObject.reg();
        jtype jt = jobj;
        int disp = jtypes[jt].rt_offset + (idx.is_imm() ? jtypes[jt].size*idx.ival() : 0);
        AR index = idx.is_imm() ? ar_x : idx.reg();
        unsigned scale = idx.is_imm() ? 0 : jtypes[jt].size;
        slotAddress = Val(jobj, valloc(jobj));
        rlock(slotAddress);
        Opnd address(jobj, base, disp, index, scale);
        lea(slotAddress.as_opnd(), address);
    }
    else {
        // must not happen
        assert(false);
        return;
    }
    if (mode == 0 || mode == 1) {
        assert(fieldSlotAddress.is_mem());
        slotAddress = Val(jobj, valloc(jobj));
        rlock(slotAddress);
        lea(slotAddress.as_opnd(), fieldSlotAddress);
    }

    value = vstack(0);
    runlock(slotAddress);
    runlock(baseObject);
    
    Val modeArg((int)mode);
    Val metaAArg((int)0);
    Val metaBArg((int)0);
   
    gen_args(csig, 0, &baseObject, &slotAddress, &value,  &metaAArg, &metaBArg, &modeArg);
    // according to contract with CG guys, the WB code neither 
    // throws an exception nor may lead to GC - may use gen_call_novm.
    runlock(csig);
//gen_brk();
    gen_call_novm(csig, wb_helper, csig.count());
    if (is_set(DBG_TRACE_CG)) { dbg(";;> ~write.barrier\n"); }
}
}}; // ~namespace Jitrino::Jet