emulator · Radiant RISC-V emulator

riscv/ .clang-format 570 B .gitignore 5 B .gitsigners 112 B LICENSE 1.1 KiB Makefile 1.2 KiB README 3.7 KiB color.h 567 B emulator.c 87.6 KiB io.c 1.1 KiB io.h 444 B jit.c 32.6 KiB jit.h 5.0 KiB riscv.c 12.0 KiB riscv.h 12.0 KiB types.h 1.0 KiB
emulator.c 87.6 KiB raw
#include <errno.h>
#include <fcntl.h>
#include <limits.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/ioctl.h>
#include <termios.h>
#include <unistd.h>

#include "color.h"
#include "io.h"
#include "jit.h"
#include "riscv.h"
#include "riscv/debug.h"
#include "types.h"

#ifndef PATH_MAX
#define PATH_MAX 4096
#endif

/* Define BINARY for `bail` and `assert` functions. */
#undef BINARY
#define BINARY "emulator"

#undef assert
#define assert(condition)                                                      \
    ((condition) ? (void)0 : _assert_failed(#condition, __FILE__, __LINE__))

static inline __attribute__((noreturn)) void _assert_failed(
    const char *condition, const char *file, int line
) {
    fprintf(stderr, "%s:%d: assertion `%s` failed\n", file, line, condition);
    abort();
}

/* Maximum physical memory size (384MB). The runtime can choose any size up to
 * this value with `-memory-size=...`. */
#define MEMORY_SIZE              (384 * 1024 * 1024)
/* Default physical memory size (128MB). */
#define DEFAULT_MEMORY_SIZE      (128 * 1024 * 1024)
/* Program memory size (4MB), reserved at start of memory for program code. */
#define PROGRAM_SIZE             (4 * 1024 * 1024)
/* Writable data region base. */
#define DATA_RW_OFFSET           0xFFFFF0
/* Data memory starts at writable data region. */
#define DATA_MEMORY_START        DATA_RW_OFFSET
/* Default data memory size. */
#define DEFAULT_DATA_MEMORY_SIZE (DEFAULT_MEMORY_SIZE - DATA_MEMORY_START)
/* Default stack size (256KB), allocated at the end of memory. */
#define DEFAULT_STACK_SIZE       (256 * 1024)
/* Maximum instructions to show in the TUI. */
#define MAX_INSTR_DISPLAY        40
/* Stack words to display in the TUI. */
#define STACK_DISPLAY_WORDS      32
/* Maximum number of CPU state snapshots to store for undo. */
#define MAX_SNAPSHOTS            64
/* Maximum open files in guest runtime. */
#define MAX_OPEN_FILES           32
/* Number of history entries to print when reporting faults. */
#define FAULT_TRACE_DEPTH        8
/* Instruction trace depth for headless tracing. */
#define TRACE_HISTORY            64
/* Maximum number of steps executed in headless mode before timing out. */
#define HEADLESS_MAX_STEPS       ((u64)1000000000000)
/* Height of header and footer in rows. */
#define HEADER_HEIGHT            2
#define FOOTER_HEIGHT            3
/* Read-only data offset. */
#define DATA_RO_OFFSET           0x10000
/* TTY escape codes. */
#define TTY_CLEAR                "\033[2J\033[H"
#define TTY_GOTO_RC              "\033[%d;%dH"
/* Exit code returned on EBREAK. */
#define EBREAK_EXIT_CODE         133
/* Self-contained Radiance image header. */
#define IMAGE_MAGIC              0x30444152U
#define IMAGE_VERSION            1U
#define IMAGE_HEADER_SIZE        20U

/* Registers displayed in the TUI, in order. */
static const reg_t registers_displayed[] = {
    SP, FP, RA, A0, A1, A2, A3, A4, A5, A6, A7, T0, T1, T2, T3, T4, T5, T6
};

/* Display mode for immediates and values. */
enum display { DISPLAY_HEX, DISPLAY_DEC };

/* Debug info entry mapping PC to source location. */
struct debug_entry {
    u32  pc;
    u32  offset;
    char file[PATH_MAX];
};

/* Debug info table. */
struct debug_info {
    struct debug_entry *entries;
    size_t              count;
    size_t              capacity;
};

/* Global debug info. */
static struct debug_info g_debug = { 0 };

/* CPU state. */
struct cpu {
    u64      regs[REGISTERS];
    u32      pc;          /* Program counter. */
    u32      programsize; /* Size of loaded program. */
    instr_t *program;     /* Program instructions. */
    bool     running;     /* Execution status. */
    bool     faulted;     /* There was a fault in execution. */
    bool     ebreak;      /* Program terminated via EBREAK. */
    reg_t    modified;    /* Index of the last modified register. */
    u64      mstatus;
    u64      mie;
    u64      mtvec;
    u64      mscratch;
    u64      mepc;
    u64      mcause;
    u64      mtval;
    u64      mip;
    u64      mhartid;
    u8       privilege;   /* RISC-V privilege level: U=0, M=3. */
};

/* Snapshot of CPU and memory state for reversing execution. */
struct snapshot {
    struct cpu cpu;                 /* Copy of CPU state. */
    u8         memory[MEMORY_SIZE]; /* Copy of memory. */
};

/* Circular buffer for snapshots. */
struct snapshot_buffer {
    struct snapshot snapshots[MAX_SNAPSHOTS];
    int             head; /* Index of most recent snapshot. */
    int             count;
};

/* CPU memory. */
static u8 memory[MEMORY_SIZE];

/* Loaded section sizes, used for bounds checking and diagnostics. */
static u32 program_base  = 0;
static u32 program_bytes = 0;
static u32 rodata_bytes  = 0;
static u32 data_bytes    = 0;

/* Snapshot buffer. */
static struct snapshot_buffer snapshots;

/* File descriptor table for guest file operations. */
static int guest_fds[MAX_OPEN_FILES];

/* Initialize the guest file descriptor table. */
static void guest_fd_table_init(void) {
    for (int i = 0; i < MAX_OPEN_FILES; i++) {
        guest_fds[i] = -1;
    }
}

/* Add a host file descriptor to the guest table. */
static int guest_fd_table_add(int host_fd) {
    /* Start at 3 to skip stdin/stdout/stderr. */
    for (int i = 3; i < MAX_OPEN_FILES; i++) {
        if (guest_fds[i] == -1) {
            guest_fds[i] = host_fd;
            return i;
        }
    }
    return -1;
}

/* Get the host fd for a guest file descriptor. */
static int guest_fd_table_get(int guest_fd) {
    if (guest_fd < 0 || guest_fd >= MAX_OPEN_FILES) {
        return -1;
    }
    /* Standard streams map directly. */
    if (guest_fd < 3) {
        return guest_fd;
    }
    return guest_fds[guest_fd];
}

/* Remove a file descriptor from the guest table. */
static void guest_fd_table_remove(int guest_fd) {
    if (guest_fd >= 3 && guest_fd < MAX_OPEN_FILES) {
        guest_fds[guest_fd] = -1;
    }
}

/* Single entry in the instruction trace ring buffer. */
struct trace_entry {
    u32     pc;
    instr_t instr;
    u64     regs[REGISTERS];
};

/* Circular buffer of recent instruction traces. */
struct trace_ring {
    struct trace_entry entries[TRACE_HISTORY];
    int                head;
    int                count;
};

/* Headless-mode instruction trace buffer. */
static struct trace_ring headless_trace = { .head = -1, .count = 0 };

/* Terminal dimensions in rows and columns. */
struct termsize {
    int rows;
    int cols;
};

/* Forward declarations. */
static void ui_render_instructions(
    struct cpu *, int col, int width, int height
);
static void ui_render_registers(
    struct cpu *, enum display, int col, int height
);
static void ui_render_stack(struct cpu *, enum display, int col, int height);
static void ui_render(struct cpu *, enum display);
static void cpu_execute(struct cpu *, enum display, bool headless);
static void emit_fault_diagnostics(struct cpu *, u32 pc);

/* Emulator runtime options, populated from CLI flags. */
struct emulator_options {
    bool stack_guard;
    u32  stack_size;
    bool debug_enabled;
    bool trace_headless;
    bool trace_enabled;
    bool trace_print_instructions;
    u32  trace_depth;
    u64  headless_max_steps;
    u32  memory_size;
    u32  data_memory_size;
    bool watch_enabled;
    u32  watch_addr;
    u32  watch_size;
    u32  watch_arm_pc;
    bool watch_zero_only;
    u32  watch_skip;
    bool watch_backtrace;
    u32  watch_backtrace_depth;
    bool validate_memory;
    bool count_instructions;
    bool jit_disabled;
    bool machine_mode;
};

/* Global emulator options. */
static struct emulator_options g_opts = {
    .stack_guard              = true,
    .stack_size               = DEFAULT_STACK_SIZE,
    .debug_enabled            = false,
    .trace_headless           = false,
    .trace_enabled            = false,
    .trace_print_instructions = false,
    .trace_depth              = 32,
    .headless_max_steps       = HEADLESS_MAX_STEPS,
    .memory_size              = DEFAULT_MEMORY_SIZE,
    .data_memory_size         = DEFAULT_DATA_MEMORY_SIZE,
    .watch_enabled            = false,
    .watch_addr               = 0,
    .watch_size               = 0,
    .watch_arm_pc             = 0,
    .watch_zero_only          = false,
    .watch_skip               = 0,
    .watch_backtrace          = false,
    .watch_backtrace_depth    = 8,
    .validate_memory          = true,
    .count_instructions       = false,
    .jit_disabled             = false,
    .machine_mode             = false,
};

static void dump_watch_context(struct cpu *, u32 addr, u32 size, u32 value);

/* Return true if the given address range overlaps the watched region. */
static inline bool watch_hit(u32 addr, u32 size) {
    if (!g_opts.watch_enabled)
        return false;
    u32 start = g_opts.watch_addr;
    u32 end   = start + (g_opts.watch_size ? g_opts.watch_size : 1);
    return addr < end && (addr + size) > start;
}

/* Check a store against the memory watchpoint and halt on a hit. */
static inline void watch_store(struct cpu *cpu, u32 addr, u32 size, u32 value) {
    if (!watch_hit(addr, size))
        return;
    if (g_opts.watch_arm_pc && cpu && cpu->pc < g_opts.watch_arm_pc)
        return;
    if (g_opts.watch_zero_only && value != 0)
        return;
    if (g_opts.watch_skip > 0) {
        g_opts.watch_skip--;
        return;
    }
    fprintf(
        stderr,
        "[WATCH] pc=%08x addr=%08x size=%u value=%08x\n",
        cpu ? cpu->pc : 0,
        addr,
        size,
        value
    );
    dump_watch_context(cpu, addr, size, value);
    if (cpu) {
        cpu->running = false;
        cpu->faulted = true;
        cpu->ebreak  = true;
    }
}

/* Fixed stack guard zone size. */
#define STACK_GUARD_BYTES 16

/* Clamp and align stack size to a valid range. */
static inline u32 sanitize_stack_bytes(u32 bytes) {
    if (bytes < WORD_SIZE)
        bytes = WORD_SIZE;
    bytes = (u32)align((i32)bytes, WORD_SIZE);

    /* Keep at least one word for guard computations. */
    if (bytes >= g_opts.memory_size)
        bytes = g_opts.memory_size - WORD_SIZE;

    return bytes;
}

/* Return the active stack guard size, or 0 if guards are disabled. */
static inline u32 stack_guard_bytes(void) {
    if (!g_opts.stack_guard)
        return 0;
    return STACK_GUARD_BYTES;
}

/* Return the configured stack size. */
static inline u32 stack_size(void) {
    return g_opts.stack_size;
}

/* Return the highest addressable word-aligned memory address. */
static inline u32 memory_top(void) {
    return g_opts.memory_size - WORD_SIZE;
}

/* Return the lowest address in the stack region. */
static inline u32 stack_bottom(void) {
    return memory_top() - stack_size() + WORD_SIZE;
}

/* Return the highest usable stack address (inside the guard zone). */
static inline u32 stack_usable_top(void) {
    u32 guard = stack_guard_bytes();
    u32 size  = stack_size();
    if (guard >= size)
        guard = size - WORD_SIZE;
    return memory_top() - guard;
}

/* Return the lowest usable stack address (inside the guard zone). */
static inline u32 stack_usable_bottom(void) {
    u32 guard = stack_guard_bytes();
    u32 size  = stack_size();
    if (guard >= size)
        guard = size - WORD_SIZE;
    return stack_bottom() + guard;
}

/* Return true if addr falls within the stack region. */
static inline bool stack_contains(u32 addr) {
    return addr >= stack_bottom() && addr <= memory_top();
}

/* Return true if the range [start, end] overlaps a stack guard zone. */
static inline bool stack_guard_overlaps(u32 guard, u32 start, u32 end) {
    if (guard == 0)
        return false;

    u32 low_guard_end    = stack_bottom() + guard - 1;
    u32 high_guard_start = memory_top() - guard + 1;

    return (start <= low_guard_end && end >= stack_bottom()) ||
           (end >= high_guard_start && start <= memory_top());
}

/* Return true if addr falls inside a stack guard zone. */
static inline bool stack_guard_contains(u32 guard, u32 addr) {
    return stack_guard_overlaps(guard, addr, addr);
}

/* Load a 16-bit value from memory in little-endian byte order. */
static inline u16 memory_load_u16(u32 addr) {
    return (u16)(memory[addr] | (memory[addr + 1] << 8));
}

/* Load a 32-bit value from memory in little-endian byte order. */
static inline u32 memory_load_u32(u32 addr) {
    return memory[addr] | (memory[addr + 1] << 8) | (memory[addr + 2] << 16) |
           (memory[addr + 3] << 24);
}

/* Load a 64-bit value from memory in little-endian byte order. */
static inline u64 memory_load_u64(u32 addr) {
    return (u64)memory[addr] | ((u64)memory[addr + 1] << 8) |
           ((u64)memory[addr + 2] << 16) | ((u64)memory[addr + 3] << 24) |
           ((u64)memory[addr + 4] << 32) | ((u64)memory[addr + 5] << 40) |
           ((u64)memory[addr + 6] << 48) | ((u64)memory[addr + 7] << 56);
}

/* Store a byte to memory. */
static inline void memory_store_u8(u32 addr, u8 value) {
    memory[addr] = value;
}

/* Store a 16-bit value to memory in little-endian byte order. */
static inline void memory_store_u16(u32 addr, u16 value) {
    memory[addr]     = (u8)(value & 0xFF);
    memory[addr + 1] = (u8)((value >> 8) & 0xFF);
}

/* Store a 32-bit value to memory in little-endian byte order. */
static inline void memory_store_u32(u32 addr, u32 value) {
    assert(addr + 3 < g_opts.memory_size);
    memory[addr]     = (u8)(value & 0xFF);
    memory[addr + 1] = (u8)((value >> 8) & 0xFF);
    memory[addr + 2] = (u8)((value >> 16) & 0xFF);
    memory[addr + 3] = (u8)((value >> 24) & 0xFF);
}

/* Store a 64-bit value to memory in little-endian byte order. */
static inline void memory_store_u64(u32 addr, u64 value) {
    assert(addr + 7 < g_opts.memory_size);
    memory_store_u32(addr, (u32)(value & 0xFFFFFFFF));
    memory_store_u32(addr + 4, (u32)(value >> 32));
}

/* Load a 32-bit word from memory, returning false if out of bounds. */
static inline bool load_word_safe(u32 addr, u32 *out) {
    if (addr > g_opts.memory_size - WORD_SIZE)
        return false;
    *out = memory_load_u32(addr);
    return true;
}

/* Dump register state and backtrace when a watchpoint fires. */
static void dump_watch_context(struct cpu *cpu, u32 addr, u32 size, u32 value) {
    if (!g_opts.watch_backtrace || !cpu)
        return;

    (void)addr;
    (void)size;
    (void)value;

    fprintf(
        stderr,
        "         regs: SP=%08x FP=%08x RA=%08x A0=%08x A1=%08x A2=%08x "
        "A3=%08x\n",
        (u32)cpu->regs[SP],
        (u32)cpu->regs[FP],
        (u32)cpu->regs[RA],
        (u32)cpu->regs[A0],
        (u32)cpu->regs[A1],
        (u32)cpu->regs[A2],
        (u32)cpu->regs[A3]
    );

    u64 fp64 = cpu->regs[FP];
    u32 fp   = (fp64 <= (u64)UINT32_MAX) ? (u32)fp64 : 0;
    u32 pc   = cpu->pc;

    fprintf(
        stderr, "         backtrace (depth %u):\n", g_opts.watch_backtrace_depth
    );

    for (u32 depth = 0; depth < g_opts.watch_backtrace_depth; depth++) {
        bool has_frame = stack_contains(fp) && fp >= (2 * WORD_SIZE);
        u32  saved_ra  = 0;
        u32  prev_fp   = 0;

        if (has_frame) {
            has_frame = load_word_safe(fp - WORD_SIZE, &saved_ra) &&
                        load_word_safe(fp - 2 * WORD_SIZE, &prev_fp) &&
                        stack_contains(prev_fp);
        }

        fprintf(
            stderr,
            "           #%u pc=%08x fp=%08x ra=%08x%s\n",
            depth,
            pc,
            fp,
            has_frame ? saved_ra : 0,
            has_frame ? "" : " (?)"
        );

        if (!has_frame || prev_fp == fp || prev_fp == 0)
            break;

        pc = saved_ra;
        fp = prev_fp;
    }
}

/* Print usage information and return 1. */
static int usage(const char *prog) {
    fprintf(
        stderr,
        "usage: %s [-run] [-machine] [-no-guard-stack]"
        " [-stack-size=KB] [-no-validate] [-debug]"
        " [-trace|-trace-headless] [-trace-depth=n] [-trace-instructions]"
        " [-max-steps=n] [-memory-size=KB] [-data-size=KB]"
        " [-watch=addr] [-watch-size=bytes] [-watch-arm-pc=addr]"
        " [-watch-zero-only] [-watch-skip=n]"
        " [-watch-backtrace] [-watch-bt-depth=n]"
        " [-count-instructions] [-no-jit]"
        " <file.bin> [program args...]\n",
        prog
    );
    return 1;
}

/* Configuration parsed from CLI flags prior to launching the emulator. */
struct cli_config {
    bool        headless;
    const char *program_path;
    int         arg_index;
};

/* Parse a string as an unsigned 32-bit integer.  Returns false on error. */
static bool parse_u32(const char *str, const char *label, int base, u32 *out) {
    char *end         = NULL;
    errno             = 0;
    unsigned long val = strtoul(str, &end, base);
    if (errno != 0 || end == str || *end != '\0') {
        fprintf(stderr, "invalid %s '%s'; expected integer\n", label, str);
        return false;
    }
    if (val > UINT32_MAX)
        val = UINT32_MAX;
    *out = (u32)val;
    return true;
}

/* Parse a string as an unsigned 64-bit integer.  Returns false on error. */
static bool parse_u64(const char *str, const char *label, u64 *out) {
    char *end              = NULL;
    errno                  = 0;
    unsigned long long val = strtoull(str, &end, 10);
    if (errno != 0 || end == str || *end != '\0') {
        fprintf(stderr, "invalid %s '%s'; expected integer\n", label, str);
        return false;
    }
    *out = (u64)val;
    return true;
}

/* Parse and validate the physical memory size passed to -memory-size=. */
static bool parse_memory_size_value(const char *value) {
    u64 parsed;
    if (!parse_u64(value, "memory size", &parsed))
        return false;
    u64 bytes = parsed * 1024;
    if (bytes <= (u64)(DATA_MEMORY_START + WORD_SIZE)) {
        fprintf(
            stderr,
            "memory size too small; minimum is %u KB\n",
            (DATA_MEMORY_START + WORD_SIZE + 1024) / 1024
        );
        return false;
    }
    if (bytes > (u64)MEMORY_SIZE) {
        fprintf(
            stderr,
            "memory size too large; maximum is %u KB (recompile emulator "
            "to increase)\n",
            MEMORY_SIZE / 1024
        );
        return false;
    }
    g_opts.memory_size = (u32)bytes;
    return true;
}

/* Parse and validate the depth passed to -trace-depth=. */
static bool parse_trace_depth_value(const char *value) {
    u32 parsed;
    if (!parse_u32(value, "trace depth", 10, &parsed))
        return false;
    if (parsed == 0) {
        fprintf(stderr, "trace depth must be greater than zero\n");
        return false;
    }
    if (parsed > TRACE_HISTORY)
        parsed = TRACE_HISTORY;
    g_opts.trace_depth = parsed;
    return true;
}

/* Parse and validate the step limit passed to -max-steps=. */
static bool parse_max_steps_value(const char *value) {
    u64 parsed;
    if (!parse_u64(value, "max steps", &parsed))
        return false;
    if (parsed == 0) {
        fprintf(stderr, "max steps must be greater than zero\n");
        return false;
    }
    g_opts.headless_max_steps = parsed;
    return true;
}

/* Parse and validate the stack size passed to -stack-size=. */
static bool parse_stack_size_value(const char *value) {
    u64 parsed;
    if (!parse_u64(value, "stack size", &parsed))
        return false;
    if (parsed == 0) {
        fprintf(stderr, "stack size must be greater than zero\n");
        return false;
    }
    u64 bytes = parsed * 1024;
    if (bytes >= MEMORY_SIZE) {
        fprintf(
            stderr,
            "stack size too large; maximum is %u KB\n",
            MEMORY_SIZE / 1024
        );
        return false;
    }
    g_opts.stack_size = sanitize_stack_bytes((u32)bytes);
    return true;
}

/* Parse and validate the data size passed to -data-size=. */
static bool parse_data_size_value(const char *value) {
    u64 parsed;
    if (!parse_u64(value, "data size", &parsed))
        return false;
    if (parsed == 0) {
        fprintf(stderr, "data size must be greater than zero\n");
        return false;
    }
    u64 bytes = parsed * 1024;
    if (bytes > (u64)MEMORY_SIZE) {
        fprintf(
            stderr,
            "data size too large; maximum is %u KB\n",
            MEMORY_SIZE / 1024
        );
        return false;
    }
    g_opts.data_memory_size = (u32)bytes;
    return true;
}

/* Validate that the stack fits within available memory. */
static bool validate_memory_layout(void) {
    if (g_opts.stack_size >= g_opts.memory_size) {
        fprintf(
            stderr,
            "stack size (%u) must be smaller than memory size (%u)\n",
            g_opts.stack_size,
            g_opts.memory_size
        );
        return false;
    }
    return true;
}

/* Parse emulator CLI arguments, returning the selected mode and file path. */
static bool parse_cli_args(int argc, char *argv[], struct cli_config *cfg) {
    bool headless = false;
    int  argi     = 1;

    while (argi < argc) {
        const char *arg = argv[argi];

        if (strcmp(arg, "--") == 0) {
            argi++;
            break;
        }
        if (arg[0] != '-')
            break;

        if (strcmp(arg, "-run") == 0) {
            headless = true;
            argi++;
            continue;
        }
        if (strcmp(arg, "-machine") == 0) {
            g_opts.machine_mode = true;
            g_opts.jit_disabled = true;
            argi++;
            continue;
        }

        if (strncmp(arg, "-stack-size=", 12) == 0) {
            if (!parse_stack_size_value(arg + 12))
                return false;
            argi++;
            continue;
        }
        if (strcmp(arg, "-no-guard-stack") == 0) {
            g_opts.stack_guard = false;
            argi++;
            continue;
        }
        if (strcmp(arg, "-no-validate") == 0) {
            g_opts.validate_memory = false;
            argi++;
            continue;
        }
        if (strcmp(arg, "-debug") == 0) {
            g_opts.debug_enabled = true;
            argi++;
            continue;
        }
        if (strcmp(arg, "-trace") == 0 || strcmp(arg, "-trace-headless") == 0) {
            g_opts.trace_enabled = true;
            argi++;
            continue;
        }
        if (strcmp(arg, "-trace-instructions") == 0) {
            g_opts.trace_print_instructions = true;
            argi++;
            continue;
        }
        if (strncmp(arg, "-trace-depth=", 13) == 0) {
            if (!parse_trace_depth_value(arg + 13))
                return false;
            argi++;
            continue;
        }
        if (strncmp(arg, "-max-steps=", 11) == 0) {
            if (!parse_max_steps_value(arg + 11))
                return false;
            argi++;
            continue;
        }
        if (strncmp(arg, "-memory-size=", 13) == 0) {
            if (!parse_memory_size_value(arg + 13))
                return false;
            argi++;
            continue;
        }
        if (strncmp(arg, "-data-size=", 11) == 0) {
            if (!parse_data_size_value(arg + 11))
                return false;
            argi++;
            continue;
        }
        if (strncmp(arg, "-watch=", 7) == 0) {
            if (!parse_u32(arg + 7, "watch address", 0, &g_opts.watch_addr))
                return false;
            g_opts.watch_enabled = true;
            argi++;
            continue;
        }
        if (strncmp(arg, "-watch-size=", 12) == 0) {
            if (!parse_u32(arg + 12, "watch size", 0, &g_opts.watch_size))
                return false;
            if (g_opts.watch_size == 0) {
                fprintf(stderr, "watch size must be greater than zero\n");
                return false;
            }
            argi++;
            continue;
        }
        if (strcmp(arg, "-watch-zero-only") == 0) {
            g_opts.watch_zero_only = true;
            argi++;
            continue;
        }
        if (strncmp(arg, "-watch-skip=", 12) == 0) {
            if (!parse_u32(arg + 12, "watch skip", 0, &g_opts.watch_skip))
                return false;
            argi++;
            continue;
        }
        if (strcmp(arg, "-watch-disable") == 0) {
            g_opts.watch_enabled = false;
            argi++;
            continue;
        }
        if (strncmp(arg, "-watch-arm-pc=", 14) == 0) {
            if (!parse_u32(arg + 14, "watch arm pc", 0, &g_opts.watch_arm_pc))
                return false;
            argi++;
            continue;
        }
        if (strcmp(arg, "-watch-backtrace") == 0) {
            g_opts.watch_backtrace = true;
            argi++;
            continue;
        }
        if (strncmp(arg, "-watch-bt-depth=", 16) == 0) {
            u32 depth;
            if (!parse_u32(arg + 16, "watch backtrace depth", 0, &depth))
                return false;
            if (depth == 0) {
                fprintf(
                    stderr, "watch backtrace depth must be greater than zero\n"
                );
                return false;
            }
            g_opts.watch_backtrace       = true;
            g_opts.watch_backtrace_depth = depth;
            argi++;
            continue;
        }
        if (strcmp(arg, "-count-instructions") == 0) {
            g_opts.count_instructions = true;
            argi++;
            continue;
        }
        if (strcmp(arg, "-no-jit") == 0) {
            g_opts.jit_disabled = true;
            argi++;
            continue;
        }
        usage(argv[0]);

        return false;
    }
    if (argi >= argc) {
        usage(argv[0]);
        return false;
    }
    cfg->program_path = argv[argi++];
    cfg->arg_index    = argi;
    cfg->headless     = headless;
    if (g_opts.watch_enabled && g_opts.watch_size == 0)
        g_opts.watch_size = 4;
    if (!validate_memory_layout())
        return false;

    g_opts.stack_size = sanitize_stack_bytes(g_opts.stack_size);

    return true;
}

/* Validate a load or store against memory bounds and stack guards. */
static bool validate_memory_access(
    struct cpu *cpu,
    u64         addr,
    u32         size,
    reg_t       base_reg,
    const char *op,
    bool        is_store
) {
    /* Skip validation for performance if disabled. */
    if (!g_opts.validate_memory)
        return true;

    if (size == 0)
        size = 1;

    const char *kind = is_store ? "store" : "load";

    u64 span_end = addr + (u64)size;
    if (addr > (u64)g_opts.memory_size || span_end > (u64)g_opts.memory_size) {
        printf(
            "Memory %s out of bounds at PC=%08x: addr=%016llx size=%u (%s)\n",
            kind,
            cpu->pc,
            (unsigned long long)addr,
            size,
            op
        );
        cpu->running = false;
        emit_fault_diagnostics(cpu, cpu->pc);
        return false;
    }

    u32 addr32 = (u32)addr;
    u32 end    = (u32)(span_end - 1);

    if (addr32 < DATA_MEMORY_START) {
        if (is_store) {
            printf(
                "Read-only memory store at PC=%08x: addr=%08x size=%u (%s)\n",
                cpu->pc,
                addr32,
                size,
                op
            );
            cpu->running = false;
            emit_fault_diagnostics(cpu, cpu->pc);
            return false;
        }
        return true;
    }

    u32  guard    = stack_guard_bytes();
    u64  base_val = cpu->regs[base_reg];
    bool base_in_stack =
        base_val <= (u64)UINT32_MAX && stack_contains((u32)base_val);
    bool start_in_stack = stack_contains(addr32);
    bool end_in_stack   = stack_contains(end);

    if (base_in_stack || start_in_stack || end_in_stack) {
        u32 bottom = stack_bottom();
        if (addr32 < bottom || end > memory_top()) {
            printf(
                "Stack %s out of bounds at PC=%08x: base=%s (0x%08x) addr=%08x "
                "size=%u (%s)\n",
                kind,
                cpu->pc,
                reg_names[base_reg],
                (u32)cpu->regs[base_reg],
                addr32,
                size,
                op
            );
            cpu->running = false;
            emit_fault_diagnostics(cpu, cpu->pc);
            return false;
        }
        if (stack_guard_overlaps(guard, addr32, end)) {
            printf(
                "Stack guard %s violation at PC=%08x: base=%s (0x%08x) "
                "addr=%08x "
                "size=%u guard=%u (%s)\n",
                kind,
                cpu->pc,
                reg_names[base_reg],
                (u32)cpu->regs[base_reg],
                addr32,
                size,
                guard,
                op
            );
            cpu->running = false;
            emit_fault_diagnostics(cpu, cpu->pc);
            return false;
        }
    }
    return true;
}

/* Validate that a register holds a valid stack address. */
static bool validate_stack_register(
    struct cpu *cpu, reg_t reg, const char *label, u32 pc, bool optional
) {
    /* Skip validation for performance if disabled. */
    if (!g_opts.validate_memory)
        return true;

    u64 value = cpu->regs[reg];

    if (optional && value == 0)
        return true;

    /* Detect addresses with upper bits set -- these can never be valid
     * stack addresses in the emulator's physical memory. */
    if (value > (u64)UINT32_MAX || (u32)value < stack_bottom() ||
        (u32)value > memory_top()) {
        printf(
            "%s (%s) out of stack bounds at PC=%08x: value=%016llx\n",
            label,
            reg_names[reg],
            pc,
            (unsigned long long)value
        );
        cpu->running = false;
        emit_fault_diagnostics(cpu, pc);
        return false;
    }

    u32 guard = stack_guard_bytes();

    if (stack_guard_contains(guard, (u32)value)) {
        printf(
            "Stack guard triggered by %s (%s) at PC=%08x: value=%08x "
            "guard=%u\n",
            label,
            reg_names[reg],
            pc,
            (u32)value,
            guard
        );
        cpu->running = false;
        emit_fault_diagnostics(cpu, pc);
        return false;
    }
    return true;
}

/* Toggle stack guarding in the TUI and re-validate live stack registers. */
static void toggle_stack_guard(struct cpu *cpu) {
    g_opts.stack_guard = !g_opts.stack_guard;

    printf(
        "\nStack guard %s (%u bytes)\n",
        g_opts.stack_guard ? "enabled" : "disabled",
        STACK_GUARD_BYTES
    );

    if (g_opts.stack_guard && cpu->running) {
        validate_stack_register(cpu, SP, "SP", cpu->pc, false);
        if (cpu->running) {
            validate_stack_register(cpu, FP, "FP", cpu->pc, true);
        }
    }
}

/* Get terminal dimensions. */
static struct termsize termsize(void) {
    struct winsize  w;
    struct termsize size = { 24, 80 }; /* Default fallback. */

    if (ioctl(STDOUT_FILENO, TIOCGWINSZ, &w) != -1) {
        size.rows = w.ws_row;
        size.cols = w.ws_col;
    }
    return size;
}

/* Take a snapshot of the current CPU and memory state. */
static void snapshot_save(struct cpu *cpu) {
    int nexti = (snapshots.head + 1 + MAX_SNAPSHOTS) % MAX_SNAPSHOTS;

    memcpy(&snapshots.snapshots[nexti].cpu, cpu, sizeof(struct cpu));
    memcpy(snapshots.snapshots[nexti].memory, memory, g_opts.memory_size);

    /* Fix the program pointer to reference the snapshot's own memory. */
    snapshots.snapshots[nexti].cpu.program =
        (instr_t *)snapshots.snapshots[nexti].memory;

    snapshots.head = nexti;
    if (snapshots.count < MAX_SNAPSHOTS)
        snapshots.count++;
}

/* Restore the most recent snapshot, returning false if none remain. */
static bool snapshot_restore(struct cpu *cpu) {
    if (snapshots.count <= 1)
        return false;

    snapshots.head = (snapshots.head + MAX_SNAPSHOTS - 1) % MAX_SNAPSHOTS;
    snapshots.count--;

    int previ = snapshots.head;
    memcpy(cpu, &snapshots.snapshots[previ].cpu, sizeof(struct cpu));
    memcpy(memory, snapshots.snapshots[previ].memory, g_opts.memory_size);

    /* Fix the program pointer to reference the live memory buffer. */
    cpu->program = (instr_t *)memory;

    return true;
}

/* Initialize the snapshot buffer with an initial snapshot. */
static void snapshot_init(struct cpu *cpu) {
    snapshots.head  = -1;
    snapshots.count = 0;
    snapshot_save(cpu);
}

/* Reset the headless instruction trace buffer. */
static void trace_reset(void) {
    headless_trace.head  = -1;
    headless_trace.count = 0;
}

/* Record the current instruction into the trace ring buffer. */
static void trace_record(struct cpu *cpu, instr_t ins) {
    if (!g_opts.trace_enabled || !g_opts.trace_headless)
        return;

    int next = (headless_trace.head + 1 + TRACE_HISTORY) % TRACE_HISTORY;

    headless_trace.head                = next;
    headless_trace.entries[next].pc    = cpu->pc;
    headless_trace.entries[next].instr = ins;
    memcpy(headless_trace.entries[next].regs, cpu->regs, sizeof(cpu->regs));
    if (headless_trace.count < TRACE_HISTORY)
        headless_trace.count++;
}

/* Dump the headless instruction trace to stdout. */
static bool trace_dump(u32 fault_pc) {
    if (!g_opts.trace_enabled || !g_opts.trace_headless ||
        headless_trace.count == 0)
        return false;

    int limit = (int)g_opts.trace_depth;
    if (limit <= 0)
        limit = FAULT_TRACE_DEPTH;
    if (limit > TRACE_HISTORY)
        limit = TRACE_HISTORY;
    if (limit > headless_trace.count)
        limit = headless_trace.count;

    printf("Headless trace (newest first):\n");
    for (int i = 0; i < limit; i++) {
        int idx = (headless_trace.head - i + TRACE_HISTORY) % TRACE_HISTORY;
        struct trace_entry *entry = &headless_trace.entries[idx];
        char                istr[MAX_INSTR_STR_LEN] = { 0 };

        sprint_instr(entry->instr, istr, true);

        printf(
            "  [%d] PC=%08x %s%s\n",
            i,
            entry->pc,
            istr,
            (entry->pc == fault_pc) ? "  <-- fault" : ""
        );
        printf(
            "       SP=%08x FP=%08x RA=%08x A0=%08x A1=%08x A2=%08x\n",
            (u32)entry->regs[SP],
            (u32)entry->regs[FP],
            (u32)entry->regs[RA],
            (u32)entry->regs[A0],
            (u32)entry->regs[A1],
            (u32)entry->regs[A2]
        );
    }
    return true;
}

/* Dump recent snapshot history to stdout for fault diagnostics. */
static bool snapshot_dump_history(struct cpu *cpu, u32 fault_pc) {
    (void)cpu;
    if (snapshots.count == 0)
        return false;

    int limit = FAULT_TRACE_DEPTH;
    if (limit > snapshots.count)
        limit = snapshots.count;

    printf("Snapshot history (newest first):\n");
    for (int i = 0; i < limit; i++) {
        int idx = (snapshots.head - i + MAX_SNAPSHOTS) % MAX_SNAPSHOTS;
        struct snapshot *snap    = &snapshots.snapshots[idx];
        u32              next_pc = snap->cpu.pc;
        u32  exec_pc = next_pc >= INSTR_SIZE ? next_pc - INSTR_SIZE : next_pc;
        char istr[MAX_INSTR_STR_LEN] = { 0 };
        u32  instr_index             = exec_pc / INSTR_SIZE;

        if (instr_index < snap->cpu.programsize) {
            sprint_instr(snap->cpu.program[instr_index], istr, true);
        } else {
            snprintf(istr, sizeof(istr), "<pc %08x>", exec_pc);
        }

        printf(
            "  [%d] PC next=%08x prev=%08x %s%s\n",
            i,
            next_pc,
            exec_pc,
            istr,
            (exec_pc == fault_pc || next_pc == fault_pc) ? "  <-- fault" : ""
        );
        printf(
            "       SP=%08x FP=%08x RA=%08x A0=%08x\n",
            (u32)snap->cpu.regs[SP],
            (u32)snap->cpu.regs[FP],
            (u32)snap->cpu.regs[RA],
            (u32)snap->cpu.regs[A0]
        );
    }
    return true;
}

/* Emit runtime fault diagnostics including trace and snapshot history. */
static void emit_fault_diagnostics(struct cpu *cpu, u32 pc) {
    if (cpu->faulted)
        return;

    cpu->faulted = true;

    printf("\n--- runtime fault diagnostics ---\n");
    bool printed = false;

    printed |= trace_dump(pc);
    printed |= snapshot_dump_history(cpu, pc);

    if (!printed) {
        printf("No trace data available.\n");
    }
    printf("--- end diagnostics ---\n");
    fflush(stdout);
}

/* Return true if the CPU's PC is outside the loaded program bounds. */
static inline bool cpu_out_of_bounds(struct cpu *cpu) {
    if (!g_opts.validate_memory)
        return false;
    if (program_bytes == 0)
        return true;
    if (cpu->pc < program_base)
        return true;
    return (cpu->pc - program_base) >= program_bytes;
}

/* Last executed PC, used for detecting branches/jumps in trace mode. */
static u32 last_executed_pc = 0;

/* Reset CPU state (keeping program loaded). */
static void cpu_reset(struct cpu *cpu) {
    trace_reset();
    memset(cpu->regs, 0, sizeof(cpu->regs));
    cpu->mstatus   = 0;
    cpu->mie       = 0;
    cpu->mtvec     = 0;
    cpu->mscratch  = 0;
    cpu->mepc      = 0;
    cpu->mcause    = 0;
    cpu->mtval     = 0;
    cpu->mip       = 0;
    cpu->mhartid   = 0;
    cpu->privilege = 3;

    /* Set SP to the top of the usable stack, aligned to 16 bytes
     * as required by the RISC-V ABI. */
    cpu->regs[SP]    = stack_usable_top() & ~0xF;
    cpu->pc          = program_base;
    cpu->running     = true;
    cpu->faulted     = false;
    cpu->ebreak      = false;
    cpu->modified    = ZERO;
    last_executed_pc = 0;
}

/* Initialize CPU and memory to a clean state. */
static void cpu_init(struct cpu *cpu) {
    memset(cpu, 0, sizeof(*cpu));
    memset(memory, 0, g_opts.memory_size);
    cpu->program     = (instr_t *)memory;
    cpu->programsize = 0;
    trace_reset();
    cpu->privilege = 3;
    cpu->mhartid   = 0;
    guest_fd_table_init();
    cpu_reset(cpu);
}

/* Open a file via the openat syscall (56). */
static i32 ecall_openat(u32 pathname_addr, i32 flags) {
    if (pathname_addr >= g_opts.memory_size)
        return -1;

    /* Find the null terminator to validate the string is in bounds. */
    u32 path_end = pathname_addr;
    while (path_end < g_opts.memory_size && memory[path_end] != 0)
        path_end++;
    if (path_end >= g_opts.memory_size)
        return -1;

    i32 host_fd = open((const char *)&memory[pathname_addr], flags, 0644);
    if (host_fd < 0)
        return -1;

    i32 guest_fd = guest_fd_table_add(host_fd);
    if (guest_fd < 0) {
        close(host_fd);
        return -1;
    }
    return guest_fd;
}

/* Close a file descriptor via the close syscall (57). */
static i32 ecall_close(i32 guest_fd) {
    /* Don't close standard streams. */
    if (guest_fd < 3)
        return 0;

    i32 host_fd = guest_fd_table_get(guest_fd);
    if (host_fd >= 0) {
        i32 result = close(host_fd);
        guest_fd_table_remove(guest_fd);
        return result;
    }
    return -1;
}

/* Load a binary section from disk into emulator memory at the given offset. */
static u32 load_section(
    const char *filepath,
    const char *suffix,
    u32         offset,
    u32         limit,
    const char *label
) {
    char path[PATH_MAX];
    snprintf(path, sizeof(path), "%s.%s", filepath, suffix);

    FILE *file = fopen(path, "rb");
    if (!file)
        return 0;

    if (fseek(file, 0, SEEK_END) != 0) {
        fclose(file);
        bail("failed to seek %s section", label);
    }
    long size = ftell(file);
    if (size < 0) {
        fclose(file);
        bail("failed to determine size of %s section", label);
    }
    if (fseek(file, 0, SEEK_SET) != 0) {
        fclose(file);
        bail("failed to rewind %s section", label);
    }
    if (size == 0) {
        fclose(file);
        return 0;
    }

    u32 u_size = (u32)size;
    u64 end    = (u64)offset + (u64)u_size;
    if (end > (u64)limit) {
        fclose(file);
        u32 max_size = limit - offset;
        bail(
            "%s section too large for emulator memory: required %u bytes, max "
            "%u bytes",
            label,
            u_size,
            max_size
        );
    }
    if (end > (u64)g_opts.memory_size) {
        fclose(file);
        u32 max_size = g_opts.memory_size - offset;
        bail(
            "%s section exceeds physical memory: required %u bytes at offset "
            "%u, "
            "but only %u bytes available (total memory=%u, use "
            "-memory-size=... or recompile emulator with a larger "
            "MEMORY_SIZE)",
            label,
            u_size,
            offset,
            max_size,
            g_opts.memory_size
        );
    }
    size_t read = fread(&memory[offset], 1, u_size, file);
    fclose(file);

    if (read != u_size) {
        bail(
            "could not read entire %s section: read %zu bytes, expected %u "
            "bytes (offset=%u, limit=%u)",
            label,
            read,
            u_size,
            offset,
            limit
        );
    }
    return u_size;
}

/* Decode a little-endian word without relying on host alignment. */
static u32 decode_u32_le(const u8 *bytes) {
    return (u32)bytes[0] | ((u32)bytes[1] << 8) | ((u32)bytes[2] << 16) |
           ((u32)bytes[3] << 24);
}

/* Read exactly size bytes from file into guest memory at offset. */
static void load_image_section(
    FILE *file, u32 offset, u32 size, const char *label
) {
    if (size == 0)
        return;

    size_t read = fread(&memory[offset], 1, size, file);
    if (read != size)
        bail(
            "could not read entire %s section: read %zu bytes, expected %u "
            "bytes",
            label,
            read,
            size
        );
}

/* Prepare the environment block (argv) on the guest stack. */
static void prepare_env(struct cpu *cpu, int argc, char **argv) {
    if (argc < 0 || argv == NULL)
        argc = 0;

    usize bytes = 0;
    for (int i = 0; i < argc; i++)
        bytes += strlen(argv[i]) + 1; /* Include terminating NUL. */

    /* In RV64, slices are 16 bytes (8-byte ptr + 4-byte len + 4 padding). */
    u32 slice_size       = 16;
    u32 slice_array_size = (u32)argc * slice_size;
    u32 base_size        = slice_size + slice_array_size;
    u32 total_size       = align(base_size + (u32)bytes, 16);

    /* Place the env block below the current stack pointer so it doesn't
     * overlap with uninitialized static data.  The .rw.data file only
     * contains initialized statics; undefined statics occupy memory after
     * the loaded data but are not in the file.  Placing the env block in
     * the data region would clobber those zero-initialized areas. */
    u32 sp       = (u32)cpu->regs[SP];
    u32 env_addr = (sp - total_size) & ~0xFu;

    if (env_addr <= DATA_MEMORY_START + data_bytes)
        bail("not enough memory to prepare environment block");

    /* Move SP below the env block so the program's stack doesn't overwrite it.
     */
    cpu->regs[SP] = env_addr;

    u32 slices_addr  = env_addr + slice_size;
    u32 strings_addr = slices_addr + (argc > 0 ? slice_array_size : 0);

    /* Write the Env slice header. */
    memory_store_u64(env_addr, argc > 0 ? slices_addr : 0);
    memory_store_u32(env_addr + 8, (u32)argc);

    /* Copy argument strings and populate slices. */
    u32 curr = strings_addr;
    for (int i = 0; i < argc; i++) {
        size_t len = strlen(argv[i]);
        if (curr + len >= g_opts.memory_size)
            bail("environment string does not fit in emulator memory");

        memcpy(&memory[curr], argv[i], len);
        memory[curr + len] = 0; /* Null-terminate for syscall compatibility. */

        u32 slice_entry = slices_addr + (u32)i * slice_size;
        memory_store_u64(slice_entry, curr);
        memory_store_u32(slice_entry + 8, (u32)len);

        curr += (u32)len + 1;
    }
    cpu->regs[A0] = env_addr;
    cpu->regs[A1] = env_addr;
}

/* Load debug information from the .debug file. */
static void debug_load(const char *program_path) {
    char debugpath[PATH_MAX];
    snprintf(debugpath, sizeof(debugpath), "%s.debug", program_path);

    FILE *file = fopen(debugpath, "rb");
    if (!file)
        return; /* Debug file is optional. */

    g_debug.capacity = 64;
    g_debug.entries  = malloc(sizeof(struct debug_entry) * g_debug.capacity);
    g_debug.count    = 0;

    if (!g_debug.entries) {
        fclose(file);
        return;
    }
    while (!feof(file)) {
        struct debug_entry entry;

        if (fread(&entry.pc, sizeof(u32), 1, file) != 1)
            break;
        if (fread(&entry.offset, sizeof(u32), 1, file) != 1)
            break;

        /* Read null-terminated file path. */
        size_t i = 0;
        int    c;
        while (i < PATH_MAX - 1 && (c = fgetc(file)) != EOF && c != '\0') {
            entry.file[i++] = (char)c;
        }
        entry.file[i] = '\0';

        if (c == EOF && i == 0)
            break;

        /* Grow array if needed. */
        if (g_debug.count >= g_debug.capacity) {
            g_debug.capacity *= 2;
            g_debug.entries   = realloc(
                g_debug.entries, sizeof(struct debug_entry) * g_debug.capacity
            );
            if (!g_debug.entries) {
                fclose(file);
                return;
            }
        }
        g_debug.entries[g_debug.count++] = entry;
    }
    fclose(file);
}

/* Look up source location for a given PC. */
static struct debug_entry *debug_lookup(u32 pc) {
    struct debug_entry *best = NULL;

    for (size_t i = 0; i < g_debug.count; i++) {
        if (g_debug.entries[i].pc == pc) {
            return &g_debug.entries[i];
        }
        /* Track the closest entry at or before this PC. */
        if (g_debug.entries[i].pc <= pc) {
            best = &g_debug.entries[i];
        }
    }
    return best;
}

/* Compute the line number from a file path and byte offset. */
static int line_from_offset(const char *filepath, u32 offset) {
    FILE *file = fopen(filepath, "r");
    if (!file)
        return 0;

    u32 line = 1;
    for (u32 i = 0; i < offset; i++) {
        int c = fgetc(file);
        if (c == EOF)
            break;
        if (c == '\n')
            line++;
    }
    fclose(file);

    return line;
}

/* Load the program binary and data sections into memory. */
static void program_init(struct cpu *cpu, const char *filepath) {
    program_bytes = 0;
    data_bytes    = 0;
    if (g_opts.debug_enabled)
        debug_load(filepath);

    FILE *file = fopen(filepath, "rb");
    if (!file)
        bail("failed to open file '%s'", filepath);
    if (fseek(file, 0, SEEK_END) != 0) {
        fclose(file);
        bail("failed to seek program '%s'", filepath);
    }
    long size = ftell(file);
    if (size <= 0) {
        fclose(file);
        bail("invalid file size: %ld", size);
    }
    if (fseek(file, 0, SEEK_SET) != 0) {
        fclose(file);
        bail("failed to rewind program '%s'", filepath);
    }

    u32 data_limit = DATA_MEMORY_START + g_opts.data_memory_size;
    if (data_limit > g_opts.memory_size)
        data_limit = g_opts.memory_size;

    u8     header[IMAGE_HEADER_SIZE];
    size_t header_read = fread(header, 1, sizeof(header), file);
    bool   is_image =
        header_read >= sizeof(u32) && decode_u32_le(header) == IMAGE_MAGIC;

    if (is_image) {
        if (header_read != sizeof(header)) {
            fclose(file);
            bail("truncated Radiance image header");
        }

        u32 version   = decode_u32_le(&header[4]);
        program_bytes = decode_u32_le(&header[8]);
        rodata_bytes  = decode_u32_le(&header[12]);
        data_bytes    = decode_u32_le(&header[16]);

        if (version != IMAGE_VERSION) {
            fclose(file);
            bail("unsupported Radiance image version: %u", version);
        }
        if (program_bytes == 0 || program_bytes % sizeof(instr_t) != 0) {
            fclose(file);
            bail(
                "invalid text section size in Radiance image: %u", program_bytes
            );
        }
        u64 image_size =
            (u64)IMAGE_HEADER_SIZE + program_bytes + rodata_bytes + data_bytes;
        if (image_size != (u64)size) {
            fclose(file);
            bail(
                "Radiance image size does not match header: file has %ld "
                "bytes, header requires %llu",
                size,
                (unsigned long long)image_size
            );
        }
        if (program_bytes > PROGRAM_SIZE) {
            fclose(file);
            bail(
                "text section too large: %u bytes; maximum is %u bytes",
                program_bytes,
                PROGRAM_SIZE
            );
        }
        if ((u64)DATA_RO_OFFSET + rodata_bytes > DATA_MEMORY_START) {
            fclose(file);
            bail("read-only data section exceeds available program memory");
        }
        program_base = align(DATA_RO_OFFSET + rodata_bytes, WORD_SIZE);
        if ((u64)program_base + program_bytes > DATA_MEMORY_START) {
            fclose(file);
            bail("text section exceeds available program memory");
        }
        if ((u64)DATA_MEMORY_START + data_bytes > data_limit) {
            fclose(file);
            bail("read-write data section exceeds available data memory");
        }

        load_image_section(file, program_base, program_bytes, "text");
        load_image_section(
            file, DATA_RO_OFFSET, rodata_bytes, "read-only data"
        );
        load_image_section(
            file, DATA_MEMORY_START, data_bytes, "read-write data"
        );
    } else {
        /* Legacy flat binaries use optional sidecar data files. */
        rewind(file);
        rodata_bytes = load_section(
            filepath, "ro.data", DATA_RO_OFFSET, DATA_MEMORY_START, "ro.data"
        );
        program_base = align(DATA_RO_OFFSET + rodata_bytes, WORD_SIZE);
        if (size > PROGRAM_SIZE) {
            fclose(file);
            bail(
                "invalid file size: %ld; maximum program size is %d bytes",
                size,
                PROGRAM_SIZE
            );
        }
        if ((u64)program_base + (u32)size > DATA_MEMORY_START) {
            fclose(file);
            bail("text section exceeds available program memory");
        }
        program_bytes = (u32)size;
        load_image_section(file, program_base, program_bytes, "program");
        data_bytes = load_section(
            filepath, "rw.data", DATA_MEMORY_START, data_limit, "rw.data"
        );
    }
    fclose(file);

    cpu->programsize = program_bytes / sizeof(instr_t);
    cpu->pc          = program_base;
}

static u64 *cpu_csr(struct cpu *cpu, u32 csr) {
    switch (csr) {
    case 0x300: return &cpu->mstatus;
    case 0x304: return &cpu->mie;
    case 0x305: return &cpu->mtvec;
    case 0x340: return &cpu->mscratch;
    case 0x341: return &cpu->mepc;
    case 0x342: return &cpu->mcause;
    case 0x343: return &cpu->mtval;
    case 0x344: return &cpu->mip;
    case 0xF14: return &cpu->mhartid;
    default: return NULL;
    }
}

static void cpu_machine_trap(
    struct cpu *cpu, u32 executed_pc, u64 cause, u64 value, u32 *pc_next
) {
    u64 mie = (cpu->mstatus >> 3) & 1;
    cpu->mstatus &= ~((u64)3 << 11);
    cpu->mstatus |= (u64)cpu->privilege << 11;
    cpu->mstatus = (cpu->mstatus & ~((u64)1 << 7)) | (mie << 7);
    cpu->mstatus &= ~((u64)1 << 3);
    cpu->mepc      = executed_pc;
    cpu->mcause    = cause;
    cpu->mtval     = value;
    cpu->privilege = 3;
    *pc_next       = (u32)(cpu->mtvec & ~(u64)3);
}

/* Execute a single instruction. */
static void cpu_execute(struct cpu *cpu, enum display display, bool headless) {
    if (cpu_out_of_bounds(cpu)) {
        cpu->running = false;
        emit_fault_diagnostics(cpu, cpu->pc);
        if (headless) {
            fprintf(stderr, "program is out of bounds\n");
            return;
        }
        bail("program is out of bounds");
    }

    u32     executed_pc = cpu->pc;
    instr_t ins         = cpu->program[cpu->pc / sizeof(instr_t)];
    u32     pc_next     = cpu->pc + INSTR_SIZE;
    u32     opcode      = ins.r.opcode;

    cpu->modified = ZERO;
    trace_record(cpu, ins);

    /* Print instruction if tracing is enabled in headless mode.
     * Skip NOPs (addi x0, x0, 0 = 0x00000013). */
    if (headless && g_opts.trace_print_instructions && ins.raw != 0x00000013) {
        /* Print ellipsis if we jumped to a non-sequential instruction. */
        if (last_executed_pc != 0 &&
            executed_pc != last_executed_pc + INSTR_SIZE) {
            printf("%s  :%s\n", COLOR_GREY, COLOR_RESET);
        }

        char istr[MAX_INSTR_STR_LEN] = { 0 };
        int  len                     = sprint_instr(ins, istr, true);
        int  padding                 = INSTR_STR_LEN - len;
        if (padding < 0)
            padding = 0;
        printf(
            "%s%08x%s %s%-*s%s",
            COLOR_GREY,
            executed_pc,
            COLOR_RESET,
            istr,
            padding,
            "",
            COLOR_GREY
        );

        /* Print all non-zero registers. */
        bool first = true;
        for (int i = 0; i < REGISTERS; i++) {
            if (cpu->regs[i] != 0) {
                if (!first)
                    printf(" ");
                printf("%s=%08x", reg_names[i], (u32)cpu->regs[i]);
                first = false;
            }
        }
        printf("%s\n", COLOR_RESET);
    }

    switch (opcode) {
    case OP_LUI:
        if (ins.u.rd != 0) {
            u32 lui_val = ins.u.imm_31_12 << 12;
            cpu->regs[ins.u.rd] =
                (u64)(i64)(i32)lui_val; /* RV64: sign-extend to 64 bits. */
            cpu->modified = ins.u.rd;
        }
        break;

    case OP_AUIPC:
        if (ins.u.rd != 0) {
            u32 auipc_val = ins.u.imm_31_12 << 12;
            cpu->regs[ins.u.rd] =
                cpu->pc +
                (u64)(i64)(i32)auipc_val; /* RV64: sign-extend offset. */
            cpu->modified = (reg_t)ins.u.rd;
        }
        break;

    case OP_JAL: {
        i32 imm = get_j_imm(ins);
        if (ins.j.rd != 0) {
            cpu->regs[ins.j.rd] = pc_next;
            cpu->modified       = (reg_t)ins.j.rd;
        }
        pc_next = cpu->pc + imm;
        break;
    }

    case OP_JALR: {
        i32 imm = get_i_imm(ins);
        if (ins.i.rd != 0) {
            cpu->regs[ins.i.rd] = pc_next;
            cpu->modified       = (reg_t)ins.i.rd;
        }
        /* Calculate target address in full 64-bit precision. */
        u64 jalr_target = (cpu->regs[ins.i.rs1] + (i64)imm) & ~(u64)1;
        /* Check if this is a RET instruction (jalr x0, ra, 0). */
        if (ins.i.rd == 0 && ins.i.rs1 == 1 && imm == 0 && jalr_target == 0) {
            cpu->running = false;
            if (!headless) {
                ui_render(cpu, display);

                printf(
                    "\n%sProgram terminated with return value %d (0x%08x)%s ",
                    COLOR_BOLD_GREEN,
                    (i32)cpu->regs[A0],
                    (u32)cpu->regs[A0],
                    COLOR_RESET
                );
            }
        } else {
            pc_next = (u32)jalr_target;
        }
        break;
    }

    case OP_BRANCH: {
        bool jump = false;
        i32  imm  = get_b_imm(ins);

        switch (ins.b.funct3) {
        case FUNCT3_BYTE: /* beq.  */
            jump = (cpu->regs[ins.b.rs1] == cpu->regs[ins.b.rs2]);
            break;
        case FUNCT3_HALF: /* bne.  */
            jump = (cpu->regs[ins.b.rs1] != cpu->regs[ins.b.rs2]);
            break;
        case FUNCT3_BYTE_U: /* blt.  */
            jump = ((i64)cpu->regs[ins.b.rs1] < (i64)cpu->regs[ins.b.rs2]);
            break;
        case FUNCT3_HALF_U: /* bge.  */
            jump = ((i64)cpu->regs[ins.b.rs1] >= (i64)cpu->regs[ins.b.rs2]);
            break;
        case FUNCT3_OR: /* bltu. */
            jump = (cpu->regs[ins.b.rs1] < cpu->regs[ins.b.rs2]);
            break;
        case FUNCT3_AND: /* bgeu. */
            jump = (cpu->regs[ins.b.rs1] >= cpu->regs[ins.b.rs2]);
            break;
        }
        if (jump) {
            pc_next = cpu->pc + imm;
        }
        break;
    }

    case OP_LOAD: {
        i32 imm  = get_i_imm(ins);
        u64 addr = cpu->regs[ins.i.rs1] + (i64)imm;

        if (ins.i.rd == ZERO)
            break;

        cpu->modified = (reg_t)ins.i.rd;
        bool fault    = false;

        switch (ins.i.funct3) {
        case FUNCT3_BYTE: /* lb. */
            if (!validate_memory_access(cpu, addr, 1, ins.i.rs1, "lb", false)) {
                fault = true;
                break;
            }
            /* sign_extend returns i32; on RV64 we sign-extend to 64 bits. */
            cpu->regs[ins.i.rd] = (u64)(i64)sign_extend(memory[addr], 8);
            break;
        case FUNCT3_HALF: /* lh. */
            if (!validate_memory_access(cpu, addr, 2, ins.i.rs1, "lh", false)) {
                fault = true;
                break;
            }
            cpu->regs[ins.i.rd] =
                (u64)(i64)sign_extend(memory_load_u16(addr), 16);
            break;
        case FUNCT3_WORD: /* lw. */
            if (!validate_memory_access(cpu, addr, 4, ins.i.rs1, "lw", false)) {
                fault = true;
                break;
            }
            /* RV64: lw sign-extends the 32-bit value to 64 bits. */
            cpu->regs[ins.i.rd] = (u64)(i64)(i32)memory_load_u32(addr);
            break;
        case 0x6: /* lwu (RV64). */
            if (!validate_memory_access(
                    cpu, addr, 4, ins.i.rs1, "lwu", false
                )) {
                fault = true;
                break;
            }
            cpu->regs[ins.i.rd] = (u64)memory_load_u32(addr);
            break;
        case FUNCT3_BYTE_U: /* lbu. */
            if (!validate_memory_access(
                    cpu, addr, 1, ins.i.rs1, "lbu", false
                )) {
                fault = true;
                break;
            }
            cpu->regs[ins.i.rd] = memory[addr];
            break;
        case FUNCT3_HALF_U: /* lhu. */
            if (!validate_memory_access(
                    cpu, addr, 2, ins.i.rs1, "lhu", false
                )) {
                fault = true;
                break;
            }
            cpu->regs[ins.i.rd] = memory_load_u16(addr);
            break;
        case 0x3: /* ld (RV64). */
            if (!validate_memory_access(cpu, addr, 8, ins.i.rs1, "ld", false)) {
                fault = true;
                break;
            }
            cpu->regs[ins.i.rd] = memory_load_u64(addr);
            break;
        }
        if (fault || !cpu->running)
            break;
        break;
    }

    case OP_STORE: {
        i32 imm  = get_s_imm(ins);
        u64 addr = cpu->regs[ins.s.rs1] + (i64)imm;

        switch (ins.s.funct3) {
        case FUNCT3_BYTE: /* sb. */
            if (!validate_memory_access(cpu, addr, 1, ins.s.rs1, "sb", true))
                break;
            watch_store(cpu, (u32)addr, 1, (u32)cpu->regs[ins.s.rs2]);
            memory_store_u8(addr, (u8)cpu->regs[ins.s.rs2]);
            break;
        case FUNCT3_HALF: /* sh. */
            if (!validate_memory_access(cpu, addr, 2, ins.s.rs1, "sh", true))
                break;
            watch_store(cpu, (u32)addr, 2, (u32)cpu->regs[ins.s.rs2]);
            memory_store_u16(addr, (u16)cpu->regs[ins.s.rs2]);
            break;
        case FUNCT3_WORD: /* sw. */
            if (!validate_memory_access(cpu, addr, 4, ins.s.rs1, "sw", true))
                break;
            watch_store(cpu, (u32)addr, 4, (u32)cpu->regs[ins.s.rs2]);
            memory_store_u32(addr, (u32)cpu->regs[ins.s.rs2]);
            break;
        case 0x3: /* sd (RV64). */
            if (!validate_memory_access(cpu, addr, 8, ins.s.rs1, "sd", true))
                break;
            watch_store(cpu, (u32)addr, 8, (u32)cpu->regs[ins.s.rs2]);
            memory_store_u64(addr, cpu->regs[ins.s.rs2]);
            break;
        }
        break;
    }

    case OP_IMM: {
        i32 imm        = get_i_imm(ins);
        u32 shamt_mask = 0x3F; /* RV64: 6-bit shift amounts. */

        if (ins.i.rd == ZERO)
            break;

        cpu->modified = (reg_t)ins.i.rd;

        switch (ins.i.funct3) {
        case FUNCT3_ADD: /* addi.  */
            cpu->regs[ins.i.rd] = cpu->regs[ins.i.rs1] + imm;
            break;
        case FUNCT3_SLL: /* slli.  */
            cpu->regs[ins.i.rd] = cpu->regs[ins.i.rs1] << (imm & shamt_mask);
            break;
        case FUNCT3_SLT: /* slti.  */
            cpu->regs[ins.i.rd] =
                ((i64)cpu->regs[ins.i.rs1] < (i64)imm) ? 1 : 0;
            break;
        case FUNCT3_SLTU: /* sltiu. */
            cpu->regs[ins.i.rd] =
                (cpu->regs[ins.i.rs1] < (u64)(i64)imm) ? 1 : 0;
            break;
        case FUNCT3_XOR: /* xori.  */
            cpu->regs[ins.i.rd] = cpu->regs[ins.i.rs1] ^ imm;
            break;
        case FUNCT3_SRL: /* srli/srai. */
            if ((imm & 0x400) == 0) {
                /* srli -- logical right shift. */
                cpu->regs[ins.i.rd] =
                    cpu->regs[ins.i.rs1] >> (imm & shamt_mask);
            } else {
                /* srai -- arithmetic right shift. */
                cpu->regs[ins.i.rd] =
                    (u64)((i64)cpu->regs[ins.i.rs1] >> (imm & shamt_mask));
            }
            break;
        case FUNCT3_OR: /* ori.   */
            cpu->regs[ins.i.rd] = cpu->regs[ins.i.rs1] | imm;
            break;
        case FUNCT3_AND: /* andi.  */
            cpu->regs[ins.i.rd] = cpu->regs[ins.i.rs1] & imm;
            break;
        }
        break;
    }

    case OP_IMM_32: {
        /* RV64I: 32-bit immediate operations (ADDIW, SLLIW, SRLIW, SRAIW).
         * These operate on the lower 32 bits and sign-extend the result. */
        i32 imm = get_i_imm(ins);

        if (ins.i.rd == ZERO)
            break;

        cpu->modified = (reg_t)ins.i.rd;

        switch (ins.i.funct3) {
        case FUNCT3_ADD: { /* addiw. */
            i32 result          = (i32)cpu->regs[ins.i.rs1] + imm;
            cpu->regs[ins.i.rd] = (u64)(i64)result;
            break;
        }
        case FUNCT3_SLL: { /* slliw. */
            i32 result = (i32)((u32)cpu->regs[ins.i.rs1] << (imm & 0x1F));
            cpu->regs[ins.i.rd] = (u64)(i64)result;
            break;
        }
        case FUNCT3_SRL: { /* srliw/sraiw. */
            if ((imm & 0x400) == 0) {
                /* srliw -- logical right shift, then sign-extend. */
                i32 result = (i32)((u32)cpu->regs[ins.i.rs1] >> (imm & 0x1F));
                cpu->regs[ins.i.rd] = (u64)(i64)result;
            } else {
                /* sraiw -- arithmetic right shift, then sign-extend. */
                i32 result          = (i32)cpu->regs[ins.i.rs1] >> (imm & 0x1F);
                cpu->regs[ins.i.rd] = (u64)(i64)result;
            }
            break;
        }
        }
        break;
    }

    case OP_OP: {
        if (ins.r.rd == ZERO)
            break;

        cpu->modified = (reg_t)ins.r.rd;

        switch (ins.r.funct7) {
        case FUNCT7_NORMAL: {
            u32 shamt_mask = 0x3F;
            switch (ins.r.funct3) {
            case FUNCT3_ADD: /* add.  */
                cpu->regs[ins.r.rd] =
                    cpu->regs[ins.r.rs1] + cpu->regs[ins.r.rs2];
                break;
            case FUNCT3_SLL: /* sll.  */
                cpu->regs[ins.r.rd] = cpu->regs[ins.r.rs1]
                                      << (cpu->regs[ins.r.rs2] & shamt_mask);
                break;
            case FUNCT3_SLT: /* slt.  */
                cpu->regs[ins.r.rd] =
                    ((i64)cpu->regs[ins.r.rs1] < (i64)cpu->regs[ins.r.rs2]) ? 1
                                                                            : 0;
                break;
            case FUNCT3_SLTU: /* sltu. */
                cpu->regs[ins.r.rd] =
                    (cpu->regs[ins.r.rs1] < cpu->regs[ins.r.rs2]) ? 1 : 0;
                break;
            case FUNCT3_XOR: /* xor.  */
                cpu->regs[ins.r.rd] =
                    cpu->regs[ins.r.rs1] ^ cpu->regs[ins.r.rs2];
                break;
            case FUNCT3_SRL: /* srl.  */
                cpu->regs[ins.r.rd] =
                    cpu->regs[ins.r.rs1] >> (cpu->regs[ins.r.rs2] & shamt_mask);
                break;
            case FUNCT3_OR: /* or.   */
                cpu->regs[ins.r.rd] =
                    cpu->regs[ins.r.rs1] | cpu->regs[ins.r.rs2];
                break;
            case FUNCT3_AND: /* and.  */
                cpu->regs[ins.r.rd] =
                    cpu->regs[ins.r.rs1] & cpu->regs[ins.r.rs2];
                break;
            }
            break;
        }

        case FUNCT7_SUB:
            switch (ins.r.funct3) {
            case FUNCT3_ADD: /* sub. */
                cpu->regs[ins.r.rd] =
                    cpu->regs[ins.r.rs1] - cpu->regs[ins.r.rs2];
                break;
            case FUNCT3_SRL: /* sra. */
                cpu->regs[ins.r.rd] = (u64)((i64)cpu->regs[ins.r.rs1] >>
                                            (cpu->regs[ins.r.rs2] & 0x3F));
                break;
            }
            break;

        case FUNCT7_MUL:
            switch (ins.r.funct3) {
            case FUNCT3_ADD: /* mul.  */
                cpu->regs[ins.r.rd] =
                    cpu->regs[ins.r.rs1] * cpu->regs[ins.r.rs2];
                break;
            case FUNCT3_XOR: /* div.  */
                if (cpu->regs[ins.r.rs2] != 0) {
                    cpu->regs[ins.r.rd] = (u64)((i64)cpu->regs[ins.r.rs1] /
                                                (i64)cpu->regs[ins.r.rs2]);
                } else {
                    cpu->regs[ins.r.rd] = (u64)-1; /* Division by zero. */
                }
                break;
            case FUNCT3_SRL: /* divu. */
                if (cpu->regs[ins.r.rs2] != 0) {
                    cpu->regs[ins.r.rd] =
                        cpu->regs[ins.r.rs1] / cpu->regs[ins.r.rs2];
                } else {
                    cpu->regs[ins.r.rd] = (u64)-1; /* Division by zero. */
                }
                break;
            case FUNCT3_OR: /* rem.  */
                if (cpu->regs[ins.r.rs2] != 0) {
                    cpu->regs[ins.r.rd] = (u64)((i64)cpu->regs[ins.r.rs1] %
                                                (i64)cpu->regs[ins.r.rs2]);
                } else {
                    cpu->regs[ins.r.rd] = cpu->regs[ins.r.rs1];
                }
                break;
            case FUNCT3_AND: /* remu. */
                if (cpu->regs[ins.r.rs2] != 0) {
                    cpu->regs[ins.r.rd] =
                        cpu->regs[ins.r.rs1] % cpu->regs[ins.r.rs2];
                } else {
                    cpu->regs[ins.r.rd] = cpu->regs[ins.r.rs1];
                }
                break;
            }
            break;
        }
        break;
    }

    case OP_OP_32: {
        /* RV64I: 32-bit register-register operations (ADDW, SUBW, SLLW, SRLW,
         * SRAW, MULW, DIVW, DIVUW, REMW, REMUW). These operate on the lower 32
         * bits and sign-extend the result to 64 bits. */
        if (ins.r.rd == ZERO)
            break;

        cpu->modified = (reg_t)ins.r.rd;
        u32 rs1_32    = (u32)cpu->regs[ins.r.rs1];
        u32 rs2_32    = (u32)cpu->regs[ins.r.rs2];

        switch (ins.r.funct7) {
        case FUNCT7_NORMAL:
            switch (ins.r.funct3) {
            case FUNCT3_ADD: { /* addw. */
                i32 result          = (i32)(rs1_32 + rs2_32);
                cpu->regs[ins.r.rd] = (u64)(i64)result;
                break;
            }
            case FUNCT3_SLL: { /* sllw. */
                i32 result          = (i32)(rs1_32 << (rs2_32 & 0x1F));
                cpu->regs[ins.r.rd] = (u64)(i64)result;
                break;
            }
            case FUNCT3_SRL: { /* srlw. */
                i32 result          = (i32)(rs1_32 >> (rs2_32 & 0x1F));
                cpu->regs[ins.r.rd] = (u64)(i64)result;
                break;
            }
            }
            break;

        case FUNCT7_SUB:
            switch (ins.r.funct3) {
            case FUNCT3_ADD: { /* subw. */
                i32 result          = (i32)(rs1_32 - rs2_32);
                cpu->regs[ins.r.rd] = (u64)(i64)result;
                break;
            }
            case FUNCT3_SRL: { /* sraw. */
                i32 result          = (i32)rs1_32 >> (rs2_32 & 0x1F);
                cpu->regs[ins.r.rd] = (u64)(i64)result;
                break;
            }
            }
            break;

        case FUNCT7_MUL:
            switch (ins.r.funct3) {
            case FUNCT3_ADD: { /* mulw.  */
                i32 result          = (i32)(rs1_32 * rs2_32);
                cpu->regs[ins.r.rd] = (u64)(i64)result;
                break;
            }
            case FUNCT3_XOR: { /* divw.  */
                if (rs2_32 != 0) {
                    i32 result          = (i32)rs1_32 / (i32)rs2_32;
                    cpu->regs[ins.r.rd] = (u64)(i64)result;
                } else {
                    cpu->regs[ins.r.rd] = (u64)(i64)(i32)-1;
                }
                break;
            }
            case FUNCT3_SRL: { /* divuw. */
                if (rs2_32 != 0) {
                    i32 result          = (i32)(rs1_32 / rs2_32);
                    cpu->regs[ins.r.rd] = (u64)(i64)result;
                } else {
                    cpu->regs[ins.r.rd] = (u64)(i64)(i32)-1;
                }
                break;
            }
            case FUNCT3_OR: { /* remw.  */
                if (rs2_32 != 0) {
                    i32 result          = (i32)rs1_32 % (i32)rs2_32;
                    cpu->regs[ins.r.rd] = (u64)(i64)result;
                } else {
                    cpu->regs[ins.r.rd] = (u64)(i64)(i32)rs1_32;
                }
                break;
            }
            case FUNCT3_AND: { /* remuw. */
                if (rs2_32 != 0) {
                    i32 result          = (i32)(rs1_32 % rs2_32);
                    cpu->regs[ins.r.rd] = (u64)(i64)result;
                } else {
                    cpu->regs[ins.r.rd] = (u64)(i64)(i32)rs1_32;
                }
                break;
            }
            }
            break;
        }
        break;
    }

    case OP_SYSTEM: {
        u32 funct12 = ins.i.imm_11_0;

        if (g_opts.machine_mode) {
            u32 funct3 = (ins.raw >> 12) & 7;
            if (funct3 != 0) {
                u32 csr_num = (ins.raw >> 20) & 0xFFF;
                u32 rd      = (ins.raw >> 7) & 0x1F;
                u32 rs1     = (ins.raw >> 15) & 0x1F;
                u64 src     = funct3 >= 5 ? rs1 : cpu->regs[rs1];
                bool writes = funct3 == 1 || funct3 == 5
                    || ((funct3 == 2 || funct3 == 3
                         || funct3 == 6 || funct3 == 7) && src != 0);
                u64 *csr = cpu_csr(cpu, csr_num);
                if (funct3 == 4 || !csr || cpu->privilege < ((csr_num >> 8) & 3)
                    || (writes && ((csr_num >> 10) & 3) == 3)) {
                    cpu_machine_trap(cpu, executed_pc, 2, ins.raw, &pc_next);
                    break;
                }
                u64 old = *csr;
                switch (funct3) {
                case 1: case 5: *csr = src; break;
                case 2: case 6: if (src != 0) *csr = old | src; break;
                case 3: case 7: if (src != 0) *csr = old & ~src; break;
                default: break;
                }
                if (writes && csr_num == 0x300) {
                    u64 mpp = (*csr >> 11) & 3;
                    if (mpp != 0 && mpp != 3)
                        *csr &= ~((u64)3 << 11);
                } else if (writes && (csr_num == 0x305 || csr_num == 0x341)) {
                    *csr &= ~(u64)3;
                }
                if (rd != 0)
                    cpu->regs[rd] = old;
                break;
            }
            if (funct12 == 0) {
                cpu_machine_trap(
                    cpu, executed_pc, cpu->privilege == 0 ? 8 : 11, 0, &pc_next
                );
            } else if (funct12 == 1) {
                cpu_machine_trap(cpu, executed_pc, 3, executed_pc, &pc_next);
            } else if (funct12 == 0x105) {
                cpu->running = false;
            } else if (funct12 == 0x302) {
                if (cpu->privilege != 3) {
                    cpu_machine_trap(cpu, executed_pc, 2, ins.raw, &pc_next);
                    break;
                }
                u8 next_priv = (cpu->mstatus >> 11) & 3;
                u64 mpie     = (cpu->mstatus >> 7) & 1;
                if (next_priv != 0 && next_priv != 3)
                    next_priv = 0;
                cpu->mstatus = (cpu->mstatus & ~((u64)1 << 3)) | (mpie << 3);
                cpu->mstatus |= (u64)1 << 7;
                cpu->mstatus &= ~((u64)3 << 11);
                cpu->privilege = next_priv;
                pc_next        = (u32)(cpu->mepc & ~(u64)3);
            } else {
                cpu_machine_trap(cpu, executed_pc, 2, ins.raw, &pc_next);
            }
            break;
        }

        if (funct12 == 0) {
            u32 syscall_num = (u32)cpu->regs[A7];

            switch (syscall_num) {
            case 64: { /* write. */
                int guest_fd = (int)cpu->regs[A0];
                u64 addr     = cpu->regs[A1];
                u64 count    = cpu->regs[A2];

                if (addr + count > g_opts.memory_size ||
                    addr > (u64)g_opts.memory_size) {
                    printf(
                        "sys_write out of bounds: addr=%016llx len=%llu\n",
                        (unsigned long long)addr,
                        (unsigned long long)count
                    );
                    cpu->running = false;
                    emit_fault_diagnostics(cpu, executed_pc);
                    break;
                }
                ssize_t written = 0;
                int     host_fd = guest_fd_table_get(guest_fd);

                if (host_fd >= 0 && count > 0) {
                    written = write(host_fd, &memory[(u32)addr], (u32)count);
                    if (written < 0) {
                        written = 0;
                    }
                }
                cpu->regs[A0] = (u64)written;
                break;
            }
            case 63: { /* read. */
                int guest_fd = (int)cpu->regs[A0];
                u64 addr     = cpu->regs[A1];
                u64 count    = cpu->regs[A2];

                if (addr + count > g_opts.memory_size ||
                    addr > (u64)g_opts.memory_size) {
                    printf(
                        "sys_read out of bounds: addr=%016llx len=%llu\n",
                        (unsigned long long)addr,
                        (unsigned long long)count
                    );
                    cpu->running = false;
                    emit_fault_diagnostics(cpu, executed_pc);
                    break;
                }
                ssize_t read_bytes = 0;
                int     host_fd    = guest_fd_table_get(guest_fd);

                if (host_fd >= 0 && count > 0) {
                    read_bytes = read(host_fd, &memory[(u32)addr], (u32)count);
                    if (read_bytes < 0) {
                        read_bytes = 0;
                    }
                }
                cpu->regs[A0] = (u64)read_bytes;
                break;
            }
            case 93: { /* exit. */
                cpu->running = false;
                break;
            }
            case 56: { /* openat. */
                u64 pathname_addr = cpu->regs[A1];
                i32 flags         = (i32)cpu->regs[A2];
                if (pathname_addr > (u64)g_opts.memory_size) {
                    cpu->regs[A0] = (u64)(i64)(i32)-1;
                    break;
                }
                cpu->regs[A0] =
                    (u64)(i64)(i32)ecall_openat((u32)pathname_addr, flags);
                break;
            }
            case 57: { /* close. */
                i32 guest_fd  = (i32)cpu->regs[A0];
                cpu->regs[A0] = (u64)(i64)ecall_close(guest_fd);
                break;
            }
            default:
                cpu->regs[A0] = (u32)syscall_num;
                break;
            }
        } else if (funct12 == 1) {
            /* Look up source location for this EBREAK.  PC in the debug
             * file is relative to program start, so subtract base. */
            u32                 relative_pc = executed_pc - program_base;
            struct debug_entry *entry       = debug_lookup(relative_pc);

            printf("\n%sRuntime error (EBREAK)%s", COLOR_BOLD_RED, COLOR_RESET);
            if (entry) {
                u32 line = line_from_offset(entry->file, entry->offset);
                printf(
                    " at %s%s:%d%s", COLOR_CYAN, entry->file, line, COLOR_RESET
                );
            }
            printf("\n");

            cpu->running  = false;
            cpu->regs[A0] = EBREAK_EXIT_CODE;
            cpu->ebreak   = true;
            emit_fault_diagnostics(cpu, executed_pc);
            cpu->faulted = false;
        } else {
            printf(
                "\n%sUnknown system instruction (imm=%08x)%s\n",
                COLOR_BOLD_RED,
                funct12,
                COLOR_RESET
            );
            cpu->running = false;
            emit_fault_diagnostics(cpu, executed_pc);
        }
        break;
    }

    case OP_FENCE:
        /* Memory barriers are not implemented. */
        break;

    default:
        printf("Unknown opcode %02x at PC=%08x\n", opcode, cpu->pc);
        cpu->running = false;
        emit_fault_diagnostics(cpu, executed_pc);
        break;
    }
    /* Register x0 is hardwired to zero. */
    cpu->regs[ZERO] = 0;
    cpu->pc         = pc_next;

    if (cpu->running) {
        validate_stack_register(cpu, SP, "SP", executed_pc, false);
        if (cpu->running)
            validate_stack_register(cpu, FP, "FP", executed_pc, true);
    }

    /* Track last executed PC for trace mode. */
    if (headless && g_opts.trace_print_instructions)
        last_executed_pc = executed_pc;
}

/* Render the instructions column. */
static void ui_render_instructions(
    struct cpu *cpu, int col, int width, int height
) {
    int row = 1;

    u32 program_start_idx = program_base / INSTR_SIZE;
    u32 program_end_idx   = program_start_idx + cpu->programsize;

    /* Calculate PC index in program. */
    u32 pc_idx = cpu->pc / sizeof(instr_t);
    if (pc_idx < program_start_idx || pc_idx >= program_end_idx)
        pc_idx = program_start_idx;

    /* Calculate first instruction to display, centering PC if possible. */
    i32 progstart = (i32)pc_idx - height / 2;
    i32 min_start = (i32)program_start_idx;
    i32 max_start = (i32)program_end_idx - height;

    if (max_start < min_start)
        max_start = min_start;
    if (progstart < min_start)
        progstart = min_start;
    if (progstart > max_start)
        progstart = max_start;

    printf(TTY_GOTO_RC, row++, col);
    printf("  INSTRUCTIONS");

    for (int i = 0; i < height; i++) {
        u32 idx = (u32)progstart + i;
        if (idx >= program_end_idx)
            break;

        printf(TTY_GOTO_RC, row + i + 1, col);

        char istr[MAX_INSTR_STR_LEN] = { 0 };
        int  len = sprint_instr(cpu->program[idx], istr, true);

        if (idx == pc_idx) { /* Highlight current instruction. */
            printf("%s>%s %04x: ", COLOR_GREEN, COLOR_RESET, idx * INSTR_SIZE);
        } else {
            printf("  %s%04x:%s ", COLOR_GREY, idx * INSTR_SIZE, COLOR_RESET);
        }
        printf("%s", istr);
        printf("%-*s", width - len - 8, "");
    }
}

/* Render the registers column. */
static void ui_render_registers(
    struct cpu *cpu, enum display display, int col, int height
) {
    int row = 1;

    printf(TTY_GOTO_RC, row++, col);
    printf("REGISTERS");

    int reg_count =
        sizeof(registers_displayed) / sizeof(registers_displayed[0]);
    if (reg_count > height)
        reg_count = height;

    for (int i = 0; i < reg_count; i++) {
        printf(TTY_GOTO_RC, row + i + 1, col);

        reg_t       r = registers_displayed[i];
        const char *reg_color =
            (r == cpu->modified) ? COLOR_BOLD_BLUE : COLOR_BLUE;
        u64  reg_value = cpu->regs[r];
        bool is_stack_addr =
            reg_value <= (u64)UINT32_MAX && stack_contains((u32)reg_value);

        /* Always show registers that contain stack addresses in hex. */
        printf("%s%-2s%s = ", COLOR_GREEN, reg_names[r], COLOR_RESET);
        if (display == DISPLAY_HEX || is_stack_addr) {
            printf("%s0x%08x%s", reg_color, (u32)reg_value, COLOR_RESET);
        } else {
            printf("%s%-10d%s", reg_color, (i32)reg_value, COLOR_RESET);
        }
    }
}

/* Render the stack column. */
static void ui_render_stack(
    struct cpu *cpu, enum display display, int col, int height
) {
    int row = 1;

    printf(TTY_GOTO_RC, row++, col);
    printf("     STACK FRAME");

    assert(cpu->regs[SP] <= memory_top() && cpu->regs[FP] <= memory_top());

    u32 fp   = (u32)cpu->regs[FP];
    u32 sp   = (u32)cpu->regs[SP];
    u32 rows = (u32)height;
    if (rows > STACK_DISPLAY_WORDS)
        rows = STACK_DISPLAY_WORDS;
    if (rows == 0)
        return;

    u32 top    = stack_usable_top();
    u32 bottom = stack_usable_bottom();
    if (sp > top)
        sp = top;
    if (fp > top)
        fp = top;

    u32 used_bytes  = (top >= sp) ? (top - sp) : 0;
    u32 total_words = (used_bytes / WORD_SIZE) + 1;
    u32 frame_words = total_words;
    if (frame_words > rows)
        frame_words = rows;
    if (frame_words == 0)
        return;

    u32 start;
    if (frame_words == total_words) {
        start = top;
    } else {
        start = sp + (frame_words - 1) * WORD_SIZE;
        if (start > top)
            start = top;
    }

    if (start < bottom)
        start = bottom;

    u32 addr   = start;
    i32 offset = (i32)(start - sp);

    for (u32 i = 0; i < frame_words; i++) {
        if (addr < bottom)
            break;

        assert(addr <= memory_top());
        printf(TTY_GOTO_RC, row + i + 1, col);

        /* Mark SP and FP positions. */
        const char *marker = "  ";

        if (addr == sp) {
            marker = "sp";
        } else if (addr == fp) {
            marker = "fp";
        }
        u32 word = memory_load_u32(addr);

        char offset_buf[6];
        if (addr == sp) {
            memcpy(offset_buf, "    ", 5);
        } else {
            snprintf(offset_buf, sizeof(offset_buf), "%+4d", offset);
        }

        printf(
            "%s%s %s%s%s %08x: ",
            COLOR_GREEN,
            marker,
            COLOR_GREY,
            offset_buf,
            COLOR_RESET,
            addr
        );
        bool is_stack_addr = stack_contains(word);

        if (display == DISPLAY_HEX || is_stack_addr) {
            printf("%s0x%08x%s", COLOR_BLUE, word, COLOR_RESET);
        } else {
            printf("%s%-10d%s", COLOR_BLUE, (i32)word, COLOR_RESET);
        }
        if (addr < WORD_SIZE)
            break;

        addr   -= WORD_SIZE;
        offset -= WORD_SIZE;
    }
}

/* Render the full debugger TUI. */
static void ui_render(struct cpu *cpu, enum display display) {
    printf(TTY_CLEAR);

    struct termsize tsize = termsize();

    /* Enforce a minimum display size. */
    if (tsize.cols < 60)
        tsize.cols = 60;
    if (tsize.rows < 15)
        tsize.rows = 15;

    /* Column layout: 40% instructions, 20% registers, rest for stack. */
    int instr_width = (tsize.cols * 2) / 5;
    int reg_width   = tsize.cols / 5;

    int instr_col = 1;
    int reg_col   = instr_col + instr_width + 2;
    int stack_col = reg_col + reg_width + 2;

    int display_height = tsize.rows - FOOTER_HEIGHT - HEADER_HEIGHT;
    if (display_height > MAX_INSTR_DISPLAY)
        display_height = MAX_INSTR_DISPLAY;
    if (display_height <= 0)
        display_height = 1;

    ui_render_instructions(cpu, instr_col, instr_width, display_height);
    ui_render_registers(cpu, display, reg_col, display_height);
    ui_render_stack(cpu, display, stack_col, display_height);

    printf(TTY_GOTO_RC, display_height + FOOTER_HEIGHT, 1);
    printf(
        "%sPress `j` to step forward, `k` to step backward, `q` to quit,\n"
        "`d` to toggle decimal display, `r` to reset program.%s ",
        COLOR_GREY,
        COLOR_RESET
    );
}

/* Set up the terminal for interactive mode, saving the original settings. */
static void term_init(struct termios *oldterm) {
    struct termios term;

    tcgetattr(STDIN_FILENO, oldterm);
    term          = *oldterm;
    term.c_lflag &= ~(ICANON | ECHO);
    tcsetattr(STDIN_FILENO, TCSANOW, &term);
}

/* Restore terminal settings. */
static void term_restore(struct termios *old) {
    tcsetattr(STDIN_FILENO, TCSANOW, old);
}

int main(int argc, char *argv[]) {
    struct cpu        cpu;
    enum display      display = DISPLAY_DEC;
    struct cli_config cli     = { 0 };

    if (!parse_cli_args(argc, argv, &cli))
        return 1;

    bool headless = cli.headless;

    g_opts.trace_headless = headless;

    cpu_init(&cpu);
    program_init(&cpu, cli.program_path);
    int    prog_argc = argc - cli.arg_index;
    char **prog_argv = &argv[cli.arg_index];
    prepare_env(&cpu, prog_argc, prog_argv);

    if (headless) {
        u64 max_steps = g_opts.headless_max_steps;
        u64 steps     = 0;

        /* Try to initialise the JIT for headless mode.  Falls back to the
         * interpreter automatically when JIT is disabled, unavailable,
         * or when the code cache fills up. */
        static struct jit_state jit;
        bool                    use_jit = false;

        if (!g_opts.jit_disabled && !g_opts.trace_enabled &&
            !g_opts.trace_print_instructions && !g_opts.watch_enabled) {
            use_jit = jit_init(&jit);
        }

        if (use_jit) {
            /* ---- JIT execution loop ---- */
            while (cpu.running && steps < max_steps) {
                struct jit_block *block = jit_get_block(
                    &jit, cpu.pc, memory, program_base, program_bytes
                );
                if (!block) {
                    /* Cache full or compilation error -- fall back to
                     * interpreter for remainder. */
                    while (cpu.running && steps++ < max_steps) {
                        cpu_execute(&cpu, display, true);
                    }
                    break;
                }
                u32 next_pc = 0;
                int exit_reason =
                    jit_exec_block(block, cpu.regs, memory, &next_pc);
                steps += block->insn_count;
                jit.blocks_executed++;
                jit.insns_executed += block->insn_count;

                switch (exit_reason) {
                case JIT_EXIT_BRANCH:
                case JIT_EXIT_CHAIN:
                    cpu.pc = next_pc;
                    break;

                case JIT_EXIT_RET:
                    cpu.running = false;
                    break;

                default:
                    /* ECALL, EBREAK, FAULT -- interpreter handles it.
                     * The instruction is already counted above,
                     * so don't increment steps again. */
                    cpu.pc = next_pc;
                    cpu_execute(&cpu, display, true);
                    break;
                }
            }
            jit_destroy(&jit);
        } else {
            /* ---- Interpreter-only loop ---- */
            while (cpu.running && steps++ < max_steps) {
                cpu_execute(&cpu, display, true);
            }
        }

        if (cpu.running) {
            fprintf(
                stderr,
                "program did not terminate within %zu steps\n",
                (size_t)max_steps
            );
            return -1;
        }
        if (cpu.faulted) {
            fprintf(stderr, "program terminated due to runtime fault\n");
            return -1;
        }
        if (g_opts.count_instructions) {
            fprintf(
                stderr,
                "Processed %llu instructions\n",
                (unsigned long long)steps
            );
        }
        return (int)cpu.regs[A0];
    }
    struct termios oldterm;
    term_init(&oldterm);
    snapshot_init(&cpu);

    for (;;) {
        if (cpu.running)
            ui_render(&cpu, display);

        int ch = getchar();

        if (ch == 'q' || ch == 'Q') {
            printf("\n");
            break;
        } else if (ch == 'd' || ch == 'D') { /* Toggle display mode. */
            display = (display == DISPLAY_HEX) ? DISPLAY_DEC : DISPLAY_HEX;
        } else if (ch == 'r' || ch == 'R') { /* Reset program and state. */
            cpu_reset(&cpu);
            snapshot_init(&cpu);
        } else if (ch == 'g' || ch == 'G') { /* Toggle stack guard. */
            toggle_stack_guard(&cpu);
        } else if (ch == 'j' && cpu.running) { /* Step forward. */
            cpu_execute(&cpu, display, false);
            snapshot_save(&cpu);
        } else if (ch == 'k' && cpu.pc > 0) { /* Step backward. */
            bool restored = snapshot_restore(&cpu);

            if (restored) {
                cpu.running = true;
            } else {
                printf(
                    "\n%sNo more history to go back to.%s\n",
                    COLOR_BOLD_RED,
                    COLOR_RESET
                );
            }
        }
    }
    term_restore(&oldterm);

    return 0;
}