86Box/src/cpu/386_dynarec.c

#include <stdarg.h>
#include <stdint.h>
#include <stdio.h>
#include <string.h>
#include <stdlib.h>
#if defined(__APPLE__) && defined(__aarch64__)
#    include <pthread.h>
#endif
#include <wchar.h>
#include <math.h>
#ifndef INFINITY
#    define INFINITY (__builtin_inff())
#endif

#define HAVE_STDARG_H
#include <86box/86box.h>
#include "cpu.h"
#include "x86.h"
#include "x86_ops.h"
#include "x87.h"
#include <86box/io.h>
#include <86box/mem.h>
#include <86box/nmi.h>
#include <86box/pic.h>
#include <86box/timer.h>
#include <86box/fdd.h>
#include <86box/fdc.h>
#include <86box/machine.h>
#include <86box/gdbstub.h>
#ifdef USE_DYNAREC
#    include "codegen.h"
#    ifdef USE_NEW_DYNAREC
#        include "codegen_backend.h"
#    endif
#endif

#ifdef IS_DYNAREC
#    undef IS_DYNAREC
#endif

#include "386_common.h"

#if defined(__APPLE__) && defined(__aarch64__)
#    include <pthread.h>
#endif

#define CPU_BLOCK_END() cpu_block_end = 1

int inrecomp = 0, cpu_block_end = 0;
int cpu_end_block_after_ins = 0;

#ifdef ENABLE_386_DYNAREC_LOG
int x386_dynarec_do_log = ENABLE_386_DYNAREC_LOG;

void
x386_dynarec_log(const char *fmt, ...)
{
    va_list ap;

    if (x386_dynarec_do_log) {
        va_start(ap, fmt);
        pclog_ex(fmt, ap);
        va_end(ap);
    }
}
#else
#    define x386_dynarec_log(fmt, ...)
#endif

static __inline void
fetch_ea_32_long(uint32_t rmdat)
{
    eal_r = eal_w = NULL;
    easeg         = cpu_state.ea_seg->base;
    if (cpu_rm == 4) {
        uint8_t sib = rmdat >> 8;

        switch (cpu_mod) {
            case 0:
                cpu_state.eaaddr = cpu_state.regs[sib & 7].l;
                cpu_state.pc++;
                break;
            case 1:
                cpu_state.pc++;
                cpu_state.eaaddr = ((uint32_t) (int8_t) getbyte()) + cpu_state.regs[sib & 7].l;
                break;
            case 2:
                cpu_state.eaaddr = (fastreadl(cs + cpu_state.pc + 1)) + cpu_state.regs[sib & 7].l;
                cpu_state.pc += 5;
                break;
        }
        /*SIB byte present*/
        if ((sib & 7) == 5 && !cpu_mod)
            cpu_state.eaaddr = getlong();
        else if ((sib & 6) == 4 && !cpu_state.ssegs) {
            easeg            = ss;
            cpu_state.ea_seg = &cpu_state.seg_ss;
        }
        if (((sib >> 3) & 7) != 4)
            cpu_state.eaaddr += cpu_state.regs[(sib >> 3) & 7].l << (sib >> 6);
    } else {
        cpu_state.eaaddr = cpu_state.regs[cpu_rm].l;
        if (cpu_mod) {
            if (cpu_rm == 5 && !cpu_state.ssegs) {
                easeg            = ss;
                cpu_state.ea_seg = &cpu_state.seg_ss;
            }
            if (cpu_mod == 1) {
                cpu_state.eaaddr += ((uint32_t) (int8_t) (rmdat >> 8));
                cpu_state.pc++;
            } else {
                cpu_state.eaaddr += getlong();
            }
        } else if (cpu_rm == 5) {
            cpu_state.eaaddr = getlong();
        }
    }
    if (easeg != 0xFFFFFFFF && ((easeg + cpu_state.eaaddr) & 0xFFF) <= 0xFFC) {
        uint32_t addr = easeg + cpu_state.eaaddr;
        if (readlookup2[addr >> 12] != (uintptr_t) -1)
            eal_r = (uint32_t *) (readlookup2[addr >> 12] + addr);
        if (writelookup2[addr >> 12] != (uintptr_t) -1)
            eal_w = (uint32_t *) (writelookup2[addr >> 12] + addr);
    }
}

static __inline void
fetch_ea_16_long(uint32_t rmdat)
{
    eal_r = eal_w = NULL;
    easeg         = cpu_state.ea_seg->base;
    if (!cpu_mod && cpu_rm == 6) {
        cpu_state.eaaddr = getword();
    } else {
        switch (cpu_mod) {
            case 0:
                cpu_state.eaaddr = 0;
                break;
            case 1:
                cpu_state.eaaddr = (uint16_t) (int8_t) (rmdat >> 8);
                cpu_state.pc++;
                break;
            case 2:
                cpu_state.eaaddr = getword();
                break;
        }
        cpu_state.eaaddr += (*mod1add[0][cpu_rm]) + (*mod1add[1][cpu_rm]);
        if (mod1seg[cpu_rm] == &ss && !cpu_state.ssegs) {
            easeg            = ss;
            cpu_state.ea_seg = &cpu_state.seg_ss;
        }
        cpu_state.eaaddr &= 0xFFFF;
    }
    if (easeg != 0xFFFFFFFF && ((easeg + cpu_state.eaaddr) & 0xFFF) <= 0xFFC) {
        uint32_t addr = easeg + cpu_state.eaaddr;
        if (readlookup2[addr >> 12] != (uintptr_t) -1)
            eal_r = (uint32_t *) (readlookup2[addr >> 12] + addr);
        if (writelookup2[addr >> 12] != (uintptr_t) -1)
            eal_w = (uint32_t *) (writelookup2[addr >> 12] + addr);
    }
}

#define fetch_ea_16(rmdat)       \
    cpu_state.pc++;              \
    cpu_mod = (rmdat >> 6) & 3;  \
    cpu_reg = (rmdat >> 3) & 7;  \
    cpu_rm  = rmdat & 7;         \
    if (cpu_mod != 3) {          \
        fetch_ea_16_long(rmdat); \
        if (cpu_state.abrt)      \
            return 1;            \
    }
#define fetch_ea_32(rmdat)       \
    cpu_state.pc++;              \
    cpu_mod = (rmdat >> 6) & 3;  \
    cpu_reg = (rmdat >> 3) & 7;  \
    cpu_rm  = rmdat & 7;         \
    if (cpu_mod != 3) {          \
        fetch_ea_32_long(rmdat); \
    }                            \
    if (cpu_state.abrt)          \
    return 1

#include "x86_flags.h"

/*Prefetch emulation is a fairly simplistic model:
  - All instruction bytes must be fetched before it starts.
  - Cycles used for non-instruction memory accesses are counted and subtracted
    from the total cycles taken
  - Any remaining cycles are used to refill the prefetch queue.

  Note that this is only used for 286 / 386 systems. It is disabled when the
  internal cache on 486+ CPUs is enabled.
*/
static int prefetch_bytes    = 0;
static int prefetch_prefixes = 0;

static void
prefetch_run(int instr_cycles, int bytes, int modrm, int reads, int reads_l, int writes, int writes_l, int ea32)
{
    int mem_cycles = reads * cpu_cycles_read + reads_l * cpu_cycles_read_l + writes * cpu_cycles_write + writes_l * cpu_cycles_write_l;

    if (instr_cycles < mem_cycles)
        instr_cycles = mem_cycles;

    prefetch_bytes -= prefetch_prefixes;
    prefetch_bytes -= bytes;
    if (modrm != -1) {
        if (ea32) {
            if ((modrm & 7) == 4) {
                if ((modrm & 0x700) == 0x500)
                    prefetch_bytes -= 5;
                else if ((modrm & 0xc0) == 0x40)
                    prefetch_bytes -= 2;
                else if ((modrm & 0xc0) == 0x80)
                    prefetch_bytes -= 5;
            } else {
                if ((modrm & 0xc7) == 0x05)
                    prefetch_bytes -= 4;
                else if ((modrm & 0xc0) == 0x40)
                    prefetch_bytes--;
                else if ((modrm & 0xc0) == 0x80)
                    prefetch_bytes -= 4;
            }
        } else {
            if ((modrm & 0xc7) == 0x06)
                prefetch_bytes -= 2;
            else if ((modrm & 0xc0) != 0xc0)
                prefetch_bytes -= ((modrm & 0xc0) >> 6);
        }
    }

    /* Fill up prefetch queue */
    while (prefetch_bytes < 0) {
        prefetch_bytes += cpu_prefetch_width;
        cycles -= cpu_prefetch_cycles;
    }

    /* Subtract cycles used for memory access by instruction */
    instr_cycles -= mem_cycles;

    while (instr_cycles >= cpu_prefetch_cycles) {
        prefetch_bytes += cpu_prefetch_width;
        instr_cycles -= cpu_prefetch_cycles;
    }

    prefetch_prefixes = 0;
    if (prefetch_bytes > 16)
        prefetch_bytes = 16;
}

static void
prefetch_flush(void)
{
    prefetch_bytes = 0;
}

#define PREFETCH_RUN(instr_cycles, bytes, modrm, reads, reads_l, writes, writes_l, ea32)      \
    do {                                                                                      \
        if (cpu_prefetch_cycles)                                                              \
            prefetch_run(instr_cycles, bytes, modrm, reads, reads_l, writes, writes_l, ea32); \
    } while (0)

#define PREFETCH_PREFIX()        \
    do {                         \
        if (cpu_prefetch_cycles) \
            prefetch_prefixes++; \
    } while (0)
#define PREFETCH_FLUSH() prefetch_flush()

#define OP_TABLE(name)   ops_##name
#if 0
#    define CLOCK_CYCLES(c)               \
        {                                 \
            if (fpu_cycles > 0) {         \
                fpu_cycles -= (c);        \
                if (fpu_cycles < 0) {     \
                    cycles += fpu_cycles; \
                }                         \
            } else {                      \
                cycles -= (c);            \
            }                             \
        }
#    define CLOCK_CYCLES_FPU(c)   cycles -= (c)
#    define CONCURRENCY_CYCLES(c) fpu_cycles = (c)
#else
#    define CLOCK_CYCLES(c)     cycles -= (c)
#    define CLOCK_CYCLES_FPU(c) cycles -= (c)
#    define CONCURRENCY_CYCLES(c)
#endif
#define CLOCK_CYCLES_ALWAYS(c) cycles -= (c)

#include "386_ops.h"

#define CACHE_ON() (!(cr0 & (1 << 30)) && !(cpu_state.flags & T_FLAG))

#ifdef USE_DYNAREC
int             cycles_main = 0;
static int      cycles_old  = 0;
static uint64_t tsc_old     = 0;

#    ifdef USE_ACYCS
int acycs = 0;
#    endif

void
update_tsc(void)
{
    int      cycdiff;
    uint64_t delta;

    cycdiff = cycles_old - cycles;
#    ifdef USE_ACYCS
    if (inrecomp)
        cycdiff += acycs;
#    endif

    delta = tsc - tsc_old;
    if (delta > 0) {
        /* TSC has changed, this means interim timer processing has happened,
           see how much we still need to add. */
        cycdiff -= delta;
    }

    if (cycdiff > 0)
        tsc += cycdiff;

    if (cycdiff > 0) {
        if (TIMER_VAL_LESS_THAN_VAL(timer_target, (uint32_t) tsc))
            timer_process_inline();
    }
}

static __inline void
exec386_dynarec_int(void)
{
    cpu_block_end = 0;
    x86_was_reset = 0;

    while (!cpu_block_end) {
#    ifndef USE_NEW_DYNAREC
        oldcs  = CS;
        oldcpl = CPL;
#    endif
        cpu_state.oldpc = cpu_state.pc;
        cpu_state.op32  = use32;

        cpu_state.ea_seg = &cpu_state.seg_ds;
        cpu_state.ssegs  = 0;

        fetchdat = fastreadl_fetch(cs + cpu_state.pc);
#    ifdef ENABLE_386_DYNAREC_LOG
        if (in_smm)
            x386_dynarec_log("[%04X:%08X] fetchdat = %08X\n", CS, cpu_state.pc, fetchdat);
#    endif

        if (!cpu_state.abrt) {
            opcode = fetchdat & 0xFF;
            fetchdat >>= 8;

            trap = cpu_state.flags & T_FLAG;

            cpu_state.pc++;
            x86_opcodes[(opcode | cpu_state.op32) & 0x3ff](fetchdat);
        }

#    ifndef USE_NEW_DYNAREC
        if (!use32)
            cpu_state.pc &= 0xffff;
#    endif

        if (((cs + cpu_state.pc) >> 12) != pccache)
            CPU_BLOCK_END();

        if (cpu_end_block_after_ins) {
            cpu_end_block_after_ins--;
            if (!cpu_end_block_after_ins)
                CPU_BLOCK_END();
        }

        if (cpu_state.abrt)
            CPU_BLOCK_END();
        if (smi_line)
            CPU_BLOCK_END();
        else if (trap)
            CPU_BLOCK_END();
        else if (nmi && nmi_enable && nmi_mask)
            CPU_BLOCK_END();
        else if ((cpu_state.flags & I_FLAG) && pic.int_pending && !cpu_end_block_after_ins)
            CPU_BLOCK_END();
    }

    if (!cpu_state.abrt && trap) {
        trap = 0;
#    ifndef USE_NEW_DYNAREC
        oldcs = CS;
#    endif
        cpu_state.oldpc = cpu_state.pc;
        dr[6] |= 0x4000;
        x86_int(1);
    }

    cpu_end_block_after_ins = 0;
}

static __inline void
exec386_dynarec_dyn(void)
{
    uint32_t start_pc = 0, phys_addr = get_phys(cs + cpu_state.pc);
    int      hash = HASH(phys_addr);
#    ifdef USE_NEW_DYNAREC
    codeblock_t *block = &codeblock[codeblock_hash[hash]];
#    else
    codeblock_t *block = codeblock_hash[hash];
#    endif
    int valid_block = 0;

#    ifdef USE_NEW_DYNAREC
    if (!cpu_state.abrt)
#    else
    if (block && !cpu_state.abrt)
#    endif
    {
        page_t *page = &pages[phys_addr >> 12];

        /* Block must match current CS, PC, code segment size,
           and physical address. The physical address check will
           also catch any page faults at this stage */
        valid_block = (block->pc == cs + cpu_state.pc) && (block->_cs == cs) && (block->phys == phys_addr) && !((block->status ^ cpu_cur_status) & CPU_STATUS_FLAGS) && ((block->status & cpu_cur_status & CPU_STATUS_MASK) == (cpu_cur_status & CPU_STATUS_MASK));
        if (!valid_block) {
            uint64_t mask = (uint64_t) 1 << ((phys_addr >> PAGE_MASK_SHIFT) & PAGE_MASK_MASK);
#    ifdef USE_NEW_DYNAREC
            int      byte_offset = (phys_addr >> PAGE_BYTE_MASK_SHIFT) & PAGE_BYTE_MASK_OFFSET_MASK;
            uint64_t byte_mask   = 1ull << (PAGE_BYTE_MASK_MASK & 0x3f);

            if ((page->code_present_mask & mask) || (page->byte_code_present_mask[byte_offset] & byte_mask))
#    else
            if (page->code_present_mask[(phys_addr >> PAGE_MASK_INDEX_SHIFT) & PAGE_MASK_INDEX_MASK] & mask)
#    endif
            {
                /* Walk page tree to see if we find the correct block */
                codeblock_t *new_block = codeblock_tree_find(phys_addr, cs);
                if (new_block) {
                    valid_block = (new_block->pc == cs + cpu_state.pc) && (new_block->_cs == cs) && (new_block->phys == phys_addr) && !((new_block->status ^ cpu_cur_status) & CPU_STATUS_FLAGS) && ((new_block->status & cpu_cur_status & CPU_STATUS_MASK) == (cpu_cur_status & CPU_STATUS_MASK));
                    if (valid_block) {
                        block = new_block;
#    ifdef USE_NEW_DYNAREC
                        codeblock_hash[hash] = get_block_nr(block);
#    endif
                    }
                }
            }
        }

        if (valid_block && (block->page_mask & *block->dirty_mask)) {
#    ifdef USE_NEW_DYNAREC
            codegen_check_flush(page, page->dirty_mask, phys_addr);
            if (block->pc == BLOCK_PC_INVALID)
                valid_block = 0;
            else if (block->flags & CODEBLOCK_IN_DIRTY_LIST)
                block->flags &= ~CODEBLOCK_WAS_RECOMPILED;
#    else
            codegen_check_flush(page, page->dirty_mask[(phys_addr >> 10) & 3], phys_addr);
            page->dirty_mask[(phys_addr >> 10) & 3] = 0;
            if (!block->valid)
                valid_block = 0;
#    endif
        }
        if (valid_block && block->page_mask2) {
            /* We don't want the second page to cause a page
               fault at this stage - that would break any
               code crossing a page boundary where the first
               page is present but the second isn't. Instead
               allow the first page to be interpreted and for
               the page fault to occur when the page boundary
               is actually crossed.*/
#    ifdef USE_NEW_DYNAREC
            uint32_t phys_addr_2 = get_phys_noabrt(block->pc + ((block->flags & CODEBLOCK_BYTE_MASK) ? 0x40 : 0x400));
#    else
            uint32_t phys_addr_2 = get_phys_noabrt(block->endpc);
#    endif
            page_t *page_2 = &pages[phys_addr_2 >> 12];

            if ((block->phys_2 ^ phys_addr_2) & ~0xfff)
                valid_block = 0;
            else if (block->page_mask2 & *block->dirty_mask2) {
#    ifdef USE_NEW_DYNAREC
                codegen_check_flush(page_2, page_2->dirty_mask, phys_addr_2);
                if (block->pc == BLOCK_PC_INVALID)
                    valid_block = 0;
                else if (block->flags & CODEBLOCK_IN_DIRTY_LIST)
                    block->flags &= ~CODEBLOCK_WAS_RECOMPILED;
#    else
                codegen_check_flush(page_2, page_2->dirty_mask[(phys_addr_2 >> 10) & 3], phys_addr_2);
                page_2->dirty_mask[(phys_addr_2 >> 10) & 3] = 0;
                if (!block->valid)
                    valid_block = 0;
#    endif
            }
        }
#    ifdef USE_NEW_DYNAREC
        if (valid_block && (block->flags & CODEBLOCK_IN_DIRTY_LIST)) {
            block->flags &= ~CODEBLOCK_WAS_RECOMPILED;
            if (block->flags & CODEBLOCK_BYTE_MASK)
                block->flags |= CODEBLOCK_NO_IMMEDIATES;
            else
                block->flags |= CODEBLOCK_BYTE_MASK;
        }
        if (valid_block && (block->flags & CODEBLOCK_WAS_RECOMPILED) && (block->flags & CODEBLOCK_STATIC_TOP) && block->TOP != (cpu_state.TOP & 7))
#    else
        if (valid_block && block->was_recompiled && (block->flags & CODEBLOCK_STATIC_TOP) && block->TOP != cpu_state.TOP)
#    endif
        {
            /* FPU top-of-stack does not match the value this block was compiled
               with, re-compile using dynamic top-of-stack*/
#    ifdef USE_NEW_DYNAREC
            block->flags &= ~(CODEBLOCK_STATIC_TOP | CODEBLOCK_WAS_RECOMPILED);
#    else
            block->flags &= ~CODEBLOCK_STATIC_TOP;
            block->was_recompiled = 0;
#    endif
        }
    }

#    ifdef USE_NEW_DYNAREC
    if (valid_block && (block->flags & CODEBLOCK_WAS_RECOMPILED))
#    else
    if (valid_block && block->was_recompiled)
#    endif
    {
        void (*code)(void) = (void *) &block->data[BLOCK_START];

#    ifndef USE_NEW_DYNAREC
        codeblock_hash[hash] = block;
#    endif
        inrecomp = 1;
        code();
#    ifdef USE_ACYCS
        acycs = 0;
#    endif
        inrecomp = 0;

#    ifndef USE_NEW_DYNAREC
        if (!use32)
            cpu_state.pc &= 0xffff;
#    endif
    } else if (valid_block && !cpu_state.abrt) {
#    ifdef USE_NEW_DYNAREC
        start_pc                 = cs + cpu_state.pc;
        const int max_block_size = (block->flags & CODEBLOCK_BYTE_MASK) ? ((128 - 25) - (start_pc & 0x3f)) : 1000;
#    else
        start_pc = cpu_state.pc;
#    endif

        cpu_block_end = 0;
        x86_was_reset = 0;

#    if defined(__APPLE__) && defined(__aarch64__)
        pthread_jit_write_protect_np(0);
#    endif
        codegen_block_start_recompile(block);
        codegen_in_recompile = 1;

        while (!cpu_block_end) {
#    ifndef USE_NEW_DYNAREC
            oldcs  = CS;
            oldcpl = CPL;
#    endif
            cpu_state.oldpc = cpu_state.pc;
            cpu_state.op32  = use32;

            cpu_state.ea_seg = &cpu_state.seg_ds;
            cpu_state.ssegs  = 0;

            fetchdat = fastreadl_fetch(cs + cpu_state.pc);
#    ifdef ENABLE_386_DYNAREC_LOG
            if (in_smm)
                x386_dynarec_log("[%04X:%08X] fetchdat = %08X\n", CS, cpu_state.pc, fetchdat);
#    endif

            if (!cpu_state.abrt) {
                opcode = fetchdat & 0xFF;
                fetchdat >>= 8;

                cpu_state.pc++;

                codegen_generate_call(opcode, x86_opcodes[(opcode | cpu_state.op32) & 0x3ff], fetchdat, cpu_state.pc, cpu_state.pc - 1);

                x86_opcodes[(opcode | cpu_state.op32) & 0x3ff](fetchdat);

                if (x86_was_reset)
                    break;
            }

#    ifndef USE_NEW_DYNAREC
            if (!use32)
                cpu_state.pc &= 0xffff;
#    endif

                /* Cap source code at 4000 bytes per block; this
                   will prevent any block from spanning more than
                   2 pages. In practice this limit will never be
                   hit, as host block size is only 2kB*/
#    ifdef USE_NEW_DYNAREC
            if (((cs + cpu_state.pc) - start_pc) >= max_block_size)
#    else
            if ((cpu_state.pc - start_pc) > 1000)
#    endif
                CPU_BLOCK_END();

            if (cpu_state.flags & T_FLAG)
                CPU_BLOCK_END();
            if (smi_line)
                CPU_BLOCK_END();
            if (nmi && nmi_enable && nmi_mask)
                CPU_BLOCK_END();
            if ((cpu_state.flags & I_FLAG) && pic.int_pending && !cpu_end_block_after_ins)
                CPU_BLOCK_END();

            if (cpu_end_block_after_ins) {
                cpu_end_block_after_ins--;
                if (!cpu_end_block_after_ins)
                    CPU_BLOCK_END();
            }

            if (cpu_state.abrt) {
                if (!(cpu_state.abrt & ABRT_EXPECTED))
                    codegen_block_remove();
                CPU_BLOCK_END();
            }
        }

        cpu_end_block_after_ins = 0;

        if ((!cpu_state.abrt || (cpu_state.abrt & ABRT_EXPECTED)) && !x86_was_reset)
            codegen_block_end_recompile(block);

        if (x86_was_reset)
            codegen_reset();

        codegen_in_recompile = 0;
#    if defined(__APPLE__) && defined(__aarch64__)
        pthread_jit_write_protect_np(1);
#    endif
    } else if (!cpu_state.abrt) {
        /* Mark block but do not recompile */
#    ifdef USE_NEW_DYNAREC
        start_pc                 = cs + cpu_state.pc;
        const int max_block_size = (block->flags & CODEBLOCK_BYTE_MASK) ? ((128 - 25) - (start_pc & 0x3f)) : 1000;
#    else
        start_pc = cpu_state.pc;
#    endif

        cpu_block_end = 0;
        x86_was_reset = 0;

        codegen_block_init(phys_addr);

        while (!cpu_block_end) {
#    ifndef USE_NEW_DYNAREC
            oldcs  = CS;
            oldcpl = CPL;
#    endif
            cpu_state.oldpc = cpu_state.pc;
            cpu_state.op32  = use32;

            cpu_state.ea_seg = &cpu_state.seg_ds;
            cpu_state.ssegs  = 0;

            codegen_endpc = (cs + cpu_state.pc) + 8;
            fetchdat      = fastreadl_fetch(cs + cpu_state.pc);

#    ifdef ENABLE_386_DYNAREC_LOG
            if (in_smm)
                x386_dynarec_log("[%04X:%08X] fetchdat = %08X\n", CS, cpu_state.pc, fetchdat);
#    endif

            if (!cpu_state.abrt) {
                opcode = fetchdat & 0xFF;
                fetchdat >>= 8;

                cpu_state.pc++;

                x86_opcodes[(opcode | cpu_state.op32) & 0x3ff](fetchdat);

                if (x86_was_reset)
                    break;
            }

#    ifndef USE_NEW_DYNAREC
            if (!use32)
                cpu_state.pc &= 0xffff;
#    endif

                /* Cap source code at 4000 bytes per block; this
                   will prevent any block from spanning more than
                   2 pages. In practice this limit will never be
                   hit, as host block size is only 2kB */
#    ifdef USE_NEW_DYNAREC
            if (((cs + cpu_state.pc) - start_pc) >= max_block_size)
#    else
            if ((cpu_state.pc - start_pc) > 1000)
#    endif
                CPU_BLOCK_END();

            if (cpu_state.flags & T_FLAG)
                CPU_BLOCK_END();
            if (smi_line)
                CPU_BLOCK_END();
            if (nmi && nmi_enable && nmi_mask)
                CPU_BLOCK_END();
            if ((cpu_state.flags & I_FLAG) && pic.int_pending && !cpu_end_block_after_ins)
                CPU_BLOCK_END();

            if (cpu_end_block_after_ins) {
                cpu_end_block_after_ins--;
                if (!cpu_end_block_after_ins)
                    CPU_BLOCK_END();
            }

            if (cpu_state.abrt) {
                if (!(cpu_state.abrt & ABRT_EXPECTED))
                    codegen_block_remove();
                CPU_BLOCK_END();
            }
        }

        cpu_end_block_after_ins = 0;

        if ((!cpu_state.abrt || (cpu_state.abrt & ABRT_EXPECTED)) && !x86_was_reset)
            codegen_block_end();

        if (x86_was_reset)
            codegen_reset();
    }
#    ifdef USE_NEW_DYNAREC
    else
        cpu_state.oldpc = cpu_state.pc;
#    endif
}

void
exec386_dynarec(int cycs)
{
    int      vector, tempi;
    int      cycdiff;
    int      oldcyc, oldcyc2;
    uint64_t oldtsc, delta;

    int cyc_period = cycs / 2000; /*5us*/

#    ifdef USE_ACYCS
    acycs = 0;
#    endif
    cycles_main += cycs;
    while (cycles_main > 0) {
        int cycles_start;

        cycles += cyc_period;
        cycles_start = cycles;

        while (cycles > 0) {
#    ifndef USE_NEW_DYNAREC
            oldcs           = CS;
            cpu_state.oldpc = cpu_state.pc;
            oldcpl          = CPL;
            cpu_state.op32  = use32;

            cycdiff = 0;
#    endif
            oldcyc = oldcyc2 = cycles;
            cycles_old       = cycles;
            oldtsc           = tsc;
            tsc_old          = tsc;
            if (!CACHE_ON()) /*Interpret block*/
            {
                exec386_dynarec_int();
            } else {
                exec386_dynarec_dyn();
            }

            if (cpu_state.abrt) {
                flags_rebuild();
                tempi          = cpu_state.abrt & ABRT_MASK;
                cpu_state.abrt = 0;
                x86_doabrt(tempi);
                if (cpu_state.abrt) {
                    cpu_state.abrt = 0;
                    cpu_state.pc   = cpu_state.oldpc;
#    ifndef USE_NEW_DYNAREC
                    CS = oldcs;
#    endif
                    pmodeint(8, 0);
                    if (cpu_state.abrt) {
                        cpu_state.abrt = 0;
                        softresetx86();
                        cpu_set_edx();
#    ifdef ENABLE_386_DYNAREC_LOG
                        x386_dynarec_log("Triple fault - reset\n");
#    endif
                    }
                }
            }

            if (smi_line)
                enter_smm_check(0);
            else if (nmi && nmi_enable && nmi_mask) {
#    ifndef USE_NEW_DYNAREC
                oldcs = CS;
#    endif
                cpu_state.oldpc = cpu_state.pc;
                x86_int(2);
                nmi_enable = 0;
#    ifdef OLD_NMI_BEHAVIOR
                if (nmi_auto_clear) {
                    nmi_auto_clear = 0;
                    nmi            = 0;
                }
#    else
                nmi = 0;
#    endif
            } else if ((cpu_state.flags & I_FLAG) && pic.int_pending) {
                vector = picinterrupt();
                if (vector != -1) {
#    ifndef USE_NEW_DYNAREC
                    oldcs = CS;
#    endif
                    cpu_state.oldpc = cpu_state.pc;
                    x86_int(vector);
                }
            }

            cycdiff = oldcyc - cycles;
            delta   = tsc - oldtsc;
            if (delta > 0) {
                /* TSC has changed, this means interim timer processing has happened,
                   see how much we still need to add. */
                cycdiff -= delta;
                if (cycdiff > 0)
                    tsc += cycdiff;
            } else {
                /* TSC has not changed. */
                tsc += cycdiff;
            }

            if (cycdiff > 0) {
                if (TIMER_VAL_LESS_THAN_VAL(timer_target, (uint32_t) tsc))
                    timer_process_inline();
            }

#    ifdef USE_GDBSTUB
            if (gdbstub_instruction())
                return;
#    endif
        }

        cycles_main -= (cycles_start - cycles);
    }
}
#endif