// dvxPlatformDos.c — DOS/DJGPP platform implementation for DVX GUI
//
// All BIOS calls, DPMI functions, port I/O, inline assembly, and
// DOS-specific file handling are isolated in this single file.
//
// This file is the ONLY place where DJGPP headers (dpmi.h, go32.h,
// sys/nearptr.h, etc.) appear. Every other DVX module calls through
// the dvxPlatform.h interface, so porting to a new OS (Linux/SDL,
// Win32, or bare-metal ARM) requires replacing only this file and
// nothing else. The abstraction covers five domains:
//   1. VESA VBE video init / mode set / LFB mapping
//   2. Backbuffer-to-LFB flush using rep movsl
//   3. Span fill/copy primitives using inline asm (rep stosl / rep movsl)
//   4. Mouse input via INT 33h driver
//   5. Keyboard input via BIOS INT 16h
//
// Why BIOS INT 16h for keyboard instead of direct port I/O (scancode
// reading from port 0x60): BIOS handles typematic repeat, keyboard
// translation tables, and extended key decoding. Direct port I/O would
// require reimplementing all of that, and the DPMI host already hooks
// IRQ1 to feed the BIOS buffer. The BIOS approach is simpler and more
// portable across emulators (DOSBox, 86Box, PCem all handle it correctly).
//
// Why INT 33h for mouse: same rationale — the mouse driver handles
// PS/2 and serial mice transparently, and every DOS emulator provides
// a compatible driver. Polling via function 03h avoids the complexity
// of installing a real-mode callback for mouse events.

#include "dvxPlatform.h"
#include "../dvxPalette.h"

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <ctype.h>
#include <signal.h>

// DJGPP-specific headers — this is the ONLY file that includes these
#include <dpmi.h>
#include <go32.h>
#include <pc.h>
#include <sys/nearptr.h>
#include <sys/farptr.h>

// ============================================================
// Prototypes
// ============================================================

static int32_t findBestMode(int32_t requestedW, int32_t requestedH, int32_t preferredBpp, uint16_t *outMode, DisplayT *d);
static void    getModeInfo(uint16_t mode, DisplayT *d, int32_t *score, int32_t requestedW, int32_t requestedH, int32_t preferredBpp);
static int32_t mapLfb(DisplayT *d, uint32_t physAddr);
void           platformVideoEnumModes(void (*cb)(int32_t w, int32_t h, int32_t bpp, void *userData), void *userData);
static int32_t setVesaMode(uint16_t mode);

// Alt+key scan code to ASCII lookup table (indexed by BIOS scan code).
// INT 16h returns these scan codes with ascii=0 for Alt+key combos.
// Using a 256-byte lookup table instead of a switch or if-chain because
// this is called on every keypress and the table fits in a single cache
// line cluster. The designated initializer syntax leaves all other
// entries as zero, which is the "no mapping" sentinel.
static const char sAltScanToAscii[256] = {
    // Alt+letters
    [0x10] = 'q', [0x11] = 'w', [0x12] = 'e', [0x13] = 'r',
    [0x14] = 't', [0x15] = 'y', [0x16] = 'u', [0x17] = 'i',
    [0x18] = 'o', [0x19] = 'p', [0x1E] = 'a', [0x1F] = 's',
    [0x20] = 'd', [0x21] = 'f', [0x22] = 'g', [0x23] = 'h',
    [0x24] = 'j', [0x25] = 'k', [0x26] = 'l', [0x2C] = 'z',
    [0x2D] = 'x', [0x2E] = 'c', [0x2F] = 'v', [0x30] = 'b',
    [0x31] = 'n', [0x32] = 'm',
    // Alt+digits
    [0x78] = '1', [0x79] = '2', [0x7A] = '3', [0x7B] = '4',
    [0x7C] = '5', [0x7D] = '6', [0x7E] = '7', [0x7F] = '8',
    [0x80] = '9', [0x81] = '0',
};


// ============================================================
// findBestMode
// ============================================================
//
// Enumerates all VESA VBE modes and selects the best match for the
// requested resolution and color depth using a scoring algorithm.
//
// The approach is: call VBE function 0x4F00 to get the controller
// info block (which contains a pointer to the mode list), then call
// VBE function 0x4F01 for each mode to get its attributes. Each mode
// is scored by getModeInfo() and the highest-scoring mode wins.
//
// Why scoring instead of exact-match: real VESA BIOSes vary wildly in
// what modes they expose. Some have 640x480x16 but not x32; some only
// have 800x600. The scoring heuristic picks the closest usable mode
// rather than failing outright if the exact requested mode is absent.
//
// All VBE info block reads use DJGPP's far pointer API (_farpeekb/w/l)
// to access the DPMI transfer buffer (__tb), which lives in the first
// 1MB of address space (conventional memory). VBE BIOS calls use
// real-mode interrupts via __dpmi_int(), so all data must pass through
// the transfer buffer.

static int32_t findBestMode(int32_t requestedW, int32_t requestedH, int32_t preferredBpp, uint16_t *outMode, DisplayT *d) {
    __dpmi_regs r;
    uint16_t    bestMode  = 0;
    int32_t     bestScore = -1;
    DisplayT    bestDisplay;

    memset(&bestDisplay, 0, sizeof(bestDisplay));

    // Get VBE controller info — the transfer buffer (__tb) is the DJGPP-
    // provided region in conventional memory that real-mode BIOS calls
    // can read/write. We split it into seg:off for the INT 10h call.
    uint32_t infoSeg = __tb >> 4;
    uint32_t infoOff = __tb & 0x0F;

    // Writing "VBE2" tells the BIOS we want VBE 2.0+ extended info.
    // Without this, we'd get VBE 1.x info which lacks LFB addresses.
    _farpokeb(_dos_ds, __tb + 0, 'V');
    _farpokeb(_dos_ds, __tb + 1, 'B');
    _farpokeb(_dos_ds, __tb + 2, 'E');
    _farpokeb(_dos_ds, __tb + 3, '2');

    memset(&r, 0, sizeof(r));
    r.x.ax = 0x4F00;
    r.x.es = infoSeg;
    r.x.di = infoOff;
    __dpmi_int(0x10, &r);

    // VBE functions return 0x004F in AX on success. Any other value
    // means the function failed or isn't supported.
    if (r.x.ax != 0x004F) {
        fprintf(stderr, "VBE: Function 0x4F00 failed (AX=0x%04X)\n", r.x.ax);
        return -1;
    }

    // On success the BIOS overwrites "VBE2" with "VESA" in the buffer
    char sig[5];
    for (int32_t i = 0; i < 4; i++) {
        sig[i] = _farpeekb(_dos_ds, __tb + i);
    }
    sig[4] = '\0';

    if (strcmp(sig, "VESA") != 0) {
        fprintf(stderr, "VBE: Bad signature '%s'\n", sig);
        return -1;
    }

    // VBE 2.0+ is required for LFB (Linear Frame Buffer) support.
    // VBE 1.x only supports bank switching, which we explicitly don't
    // implement — the complexity isn't worth it for 486+ targets.
    uint16_t vbeVersion = _farpeekw(_dos_ds, __tb + 4);
    if (vbeVersion < 0x0200) {
        fprintf(stderr, "VBE: Version %d.%d too old (need 2.0+)\n",
                vbeVersion >> 8, vbeVersion & 0xFF);
        return -1;
    }

    // The mode list is a far pointer (seg:off) at offset 14 in the info
    // block. It points to a null-terminated (0xFFFF) array of mode numbers
    // in conventional memory.
    uint16_t modeListOff = _farpeekw(_dos_ds, __tb + 14);
    uint16_t modeListSeg = _farpeekw(_dos_ds, __tb + 16);
    uint32_t modeListAddr = ((uint32_t)modeListSeg << 4) + modeListOff;

    // Walk the mode list. Cap at 256 to prevent runaway on corrupt BIOS
    // data (real hardware rarely has more than ~50 modes).
    for (int32_t i = 0; i < 256; i++) {
        uint16_t mode = _farpeekw(_dos_ds, modeListAddr + i * 2);

        if (mode == 0xFFFF) {
            break;
        }

        DisplayT  candidate;
        int32_t   score = 0;

        memset(&candidate, 0, sizeof(candidate));
        getModeInfo(mode, &candidate, &score, requestedW, requestedH, preferredBpp);

        if (score > bestScore) {
            bestScore   = score;
            bestMode    = mode;
            bestDisplay = candidate;
        }
    }

    if (bestScore < 0) {
        fprintf(stderr, "VBE: No suitable mode found for %ldx%ld\n", (long)requestedW, (long)requestedH);
        return -1;
    }

    *outMode = bestMode;
    *d       = bestDisplay;
    return 0;
}


// ============================================================
// platformVideoEnumModes
// ============================================================

void platformVideoEnumModes(void (*cb)(int32_t w, int32_t h, int32_t bpp, void *userData), void *userData) {
    __dpmi_regs r;

    memset(&r, 0, sizeof(r));
    r.x.ax = 0x4F00;
    r.x.es = __tb >> 4;
    r.x.di = __tb & 0x0F;

    // Write "VBE2" signature to request VBE 2.0+ info
    _farpokeb(_dos_ds, __tb + 0, 'V');
    _farpokeb(_dos_ds, __tb + 1, 'B');
    _farpokeb(_dos_ds, __tb + 2, 'E');
    _farpokeb(_dos_ds, __tb + 3, '2');

    __dpmi_int(0x10, &r);

    if (r.x.ax != 0x004F) {
        return;
    }

    uint16_t modeListOff = _farpeekw(_dos_ds, __tb + 14);
    uint16_t modeListSeg = _farpeekw(_dos_ds, __tb + 16);
    uint32_t modeListAddr = ((uint32_t)modeListSeg << 4) + modeListOff;

    for (int32_t i = 0; i < 256; i++) {
        uint16_t mode = _farpeekw(_dos_ds, modeListAddr + i * 2);

        if (mode == 0xFFFF) {
            break;
        }

        memset(&r, 0, sizeof(r));
        r.x.ax = 0x4F01;
        r.x.cx = mode;
        r.x.es = __tb >> 4;
        r.x.di = __tb & 0x0F;
        __dpmi_int(0x10, &r);

        if (r.x.ax != 0x004F) {
            continue;
        }

        uint16_t attr = _farpeekw(_dos_ds, __tb + 0);

        // Only report LFB-capable graphics modes
        if (!(attr & 0x0080) || !(attr & 0x0010)) {
            continue;
        }

        int32_t w   = _farpeekw(_dos_ds, __tb + 18);
        int32_t h   = _farpeekw(_dos_ds, __tb + 20);
        int32_t bpp = _farpeekb(_dos_ds, __tb + 25);

        cb(w, h, bpp, userData);
    }
}


// ============================================================
// getModeInfo
// ============================================================
//
// Queries VBE mode info (function 0x4F01) for a single mode and
// scores it against the requested parameters. The scoring algorithm:
//
//   Base score by bpp:  16-bit=100, 15-bit=90, 32-bit=85, 8-bit=70
//   +20 if bpp matches preferredBpp
//   +10 if exact resolution match, -10 if oversize
//   -1 (rejected) if mode lacks LFB, is text-mode, is below requested
//     resolution, or uses an unsupported bpp (e.g. 24-bit)
//
// 16-bit is preferred over 32-bit because it's twice as fast for
// span fill/copy on a 486/Pentium bus (half the bytes). 15-bit scores
// slightly below 16-bit because some VESA BIOSes report 15bpp modes
// as 16bpp with a dead high bit, causing confusion. 8-bit scores
// lowest because palette management adds complexity.
//
// 24-bit is explicitly rejected (not 8/15/16/32) because its 3-byte
// pixels can't use dword-aligned rep stosl fills without masking.
//
// The physical LFB address is temporarily stored in d->lfb as a raw
// integer cast — it will be properly mapped via DPMI in mapLfb() later.

static void getModeInfo(uint16_t mode, DisplayT *d, int32_t *score, int32_t requestedW, int32_t requestedH, int32_t preferredBpp) {
    __dpmi_regs r;

    *score = -1;

    memset(&r, 0, sizeof(r));
    r.x.ax = 0x4F01;
    r.x.cx = mode;
    r.x.es = __tb >> 4;
    r.x.di = __tb & 0x0F;
    __dpmi_int(0x10, &r);

    if (r.x.ax != 0x004F) {
        return;
    }

    // VBE mode attribute word at offset 0:
    //   bit 7 = LFB available, bit 4 = graphics mode (not text)
    // Both are required — we never bank-switch and never want text modes.
    uint16_t attr = _farpeekw(_dos_ds, __tb + 0);

    if (!(attr & 0x0080)) {
        return;
    }

    if (!(attr & 0x0010)) {
        return;
    }

    int32_t  w        = _farpeekw(_dos_ds, __tb + 18);
    int32_t  h        = _farpeekw(_dos_ds, __tb + 20);
    int32_t  bpp      = _farpeekb(_dos_ds, __tb + 25);
    int32_t  pitch    = _farpeekw(_dos_ds, __tb + 16);
    uint32_t physAddr = _farpeekl(_dos_ds, __tb + 40);

    // Must match or exceed requested resolution
    if (w < requestedW || h < requestedH) {
        return;
    }

    // Must be a supported bpp
    if (bpp != 8 && bpp != 15 && bpp != 16 && bpp != 32) {
        return;
    }

    // Score this mode
    int32_t s = 0;

    if (bpp == 16) {
        s = 100;
    } else if (bpp == 15) {
        s = 90;
    } else if (bpp == 32) {
        s = 85;
    } else if (bpp == 8) {
        s = 70;
    }

    // Prefer the user's preferred bpp
    if (bpp == preferredBpp) {
        s += 20;
    }

    // Exact resolution match is preferred
    if (w == requestedW && h == requestedH) {
        s += 10;
    } else {
        s -= 10;
    }

    *score = s;

    // Fill in display info
    d->width  = w;
    d->height = h;
    d->pitch  = pitch;

    d->format.bitsPerPixel  = bpp;
    d->format.bytesPerPixel = (bpp + 7) / 8;

    // Read the channel mask layout from the VBE mode info block.
    // These offsets (31-36) define the bit position and size of each
    // color channel. This is essential because the channel layout
    // varies: some cards use RGB565, others BGR565, etc.
    if (bpp >= 15) {
        int32_t redSize   = _farpeekb(_dos_ds, __tb + 31);
        int32_t redPos    = _farpeekb(_dos_ds, __tb + 32);
        int32_t greenSize = _farpeekb(_dos_ds, __tb + 33);
        int32_t greenPos  = _farpeekb(_dos_ds, __tb + 34);
        int32_t blueSize  = _farpeekb(_dos_ds, __tb + 35);
        int32_t bluePos   = _farpeekb(_dos_ds, __tb + 36);

        d->format.redBits    = redSize;
        d->format.redShift   = redPos;
        d->format.redMask    = ((1U << redSize) - 1) << redPos;
        d->format.greenBits  = greenSize;
        d->format.greenShift = greenPos;
        d->format.greenMask  = ((1U << greenSize) - 1) << greenPos;
        d->format.blueBits   = blueSize;
        d->format.blueShift  = bluePos;
        d->format.blueMask   = ((1U << blueSize) - 1) << bluePos;
    }

    // Store physical address in lfb field temporarily (will be remapped)
    d->lfb = (uint8_t *)(uintptr_t)physAddr;
}


// ============================================================
// mapLfb
// ============================================================
//
// Maps the video card's physical LFB address into the DPMI linear
// address space, then converts it to a near pointer for direct C
// access.
//
// The mapping process has three steps:
//   1. __dpmi_physical_address_mapping() — asks the DPMI host to
//      create a linear address mapping for the physical framebuffer.
//      This is necessary because DPMI runs in protected mode with
//      paging; physical addresses aren't directly accessible.
//   2. __dpmi_lock_linear_region() — pins the mapped pages so they
//      can't be swapped out. The LFB is memory-mapped I/O to the
//      video card; paging it would be catastrophic.
//   3. __djgpp_nearptr_enable() — disables DJGPP's default segment
//      limit checking so we can use plain C pointers to access the
//      LFB address. Without this, all LFB access would require far
//      pointer calls (_farpokeb etc.), which are much slower because
//      each one involves a segment register load.
//
// Why near pointers: the performance difference is dramatic.
// platformFlushRect() copies thousands of dwords per frame using
// rep movsl — this only works with near pointers. Far pointer access
// would add ~10 cycles per byte and make 60fps impossible on a 486.
//
// The final pointer calculation adds __djgpp_conventional_base, which
// is the offset DJGPP applies to convert linear addresses to near
// pointer addresses (compensating for the DS segment base).

static int32_t mapLfb(DisplayT *d, uint32_t physAddr) {
    __dpmi_meminfo info;
    uint32_t       fbSize = (uint32_t)d->pitch * (uint32_t)d->height;

    info.address = physAddr;
    info.size    = fbSize;

    if (__dpmi_physical_address_mapping(&info) != 0) {
        fprintf(stderr, "VBE: Failed to map LFB at 0x%08lX\n", (unsigned long)physAddr);
        return -1;
    }

    __dpmi_meminfo lockInfo;
    lockInfo.address = info.address;
    lockInfo.size    = fbSize;
    __dpmi_lock_linear_region(&lockInfo);

    if (__djgpp_nearptr_enable() == 0) {
        fprintf(stderr, "VBE: Failed to enable near pointers\n");
        return -1;
    }

    // Convert linear address to near pointer by adding the DS base offset
    d->lfb = (uint8_t *)(info.address + __djgpp_conventional_base);

    return 0;
}


// ============================================================
// platformAltScanToChar
// ============================================================

char platformAltScanToChar(int32_t scancode) {
    if (scancode < 0 || scancode > 255) {
        return 0;
    }

    return sAltScanToAscii[scancode];
}


// ============================================================
// platformFlushRect
// ============================================================
//
// Copies a dirty rectangle from the system RAM backbuffer to the LFB.
// This is the critical path for display updates — the compositor calls
// it once per dirty rect per frame.
//
// Two code paths:
//   1. Full-width: if the rect spans the entire scanline (rowBytes ==
//      pitch), collapse all rows into a single large rep movsl. This
//      avoids per-row loop overhead and is the common case for full-
//      screen redraws.
//   2. Partial-width: copy each scanline individually with rep movsl,
//      advancing src/dst by pitch (not rowBytes) between rows.
//
// rep movsl is used instead of memcpy() because on 486/Pentium, GCC's
// memcpy may not generate the optimal dword-aligned string move, and
// the DJGPP C library's memcpy isn't always tuned for large copies.
// The explicit asm guarantees exactly the instruction sequence we want.
//
// __builtin_expect hints tell GCC to generate branch-free fast paths
// for the common cases (non-zero w/h, no trailing bytes).

void platformFlushRect(const DisplayT *d, const RectT *r) {
    int32_t bpp = d->format.bytesPerPixel;

    int32_t x = r->x;
    int32_t y = r->y;
    int32_t w = r->w;
    int32_t h = r->h;

    if (__builtin_expect(w <= 0 || h <= 0, 0)) {
        return;
    }

    int32_t  rowBytes = w * bpp;
    int32_t  pitch    = d->pitch;
    uint8_t *src      = d->backBuf + y * pitch + x * bpp;
    uint8_t *dst      = d->lfb + y * pitch + x * bpp;

    // Full-width flush: single large copy
    if (rowBytes == pitch) {
        int32_t totalBytes = pitch * h;
        int32_t dwords     = totalBytes >> 2;
        int32_t remainder  = totalBytes & 3;
        __asm__ __volatile__ (
            "rep movsl"
            : "+D"(dst), "+S"(src), "+c"(dwords)
            :
            : "memory"
        );
        while (remainder-- > 0) {
            *dst++ = *src++;
        }
    } else {
        // Partial scanlines — copy row by row with rep movsd
        int32_t dwords    = rowBytes >> 2;
        int32_t remainder = rowBytes & 3;
        for (int32_t i = 0; i < h; i++) {
            int32_t  dc = dwords;
            uint8_t *s  = src;
            uint8_t *dd = dst;
            __asm__ __volatile__ (
                "rep movsl"
                : "+D"(dd), "+S"(s), "+c"(dc)
                :
                : "memory"
            );
            if (__builtin_expect(remainder > 0, 0)) {
                int32_t rem = remainder;
                while (rem-- > 0) {
                    *dd++ = *s++;
                }
            }
            src += pitch;
            dst += pitch;
        }
    }
}


// ============================================================
// platformInit
// ============================================================

void platformInit(void) {
    // Disable Ctrl+C/Break so the user can't accidentally kill the
    // GUI while in graphics mode (which would leave the display in
    // an unusable state without restoring text mode first).
    signal(SIGINT, SIG_IGN);
}


// ============================================================
// platformKeyboardGetModifiers
// ============================================================
//
// Returns the current modifier key state via INT 16h function 12h
// (enhanced get extended shift flags). The low byte contains:
//   bit 0 = right shift, bit 1 = left shift
//   bit 2 = ctrl, bit 3 = alt
// The widget system uses these bits for keyboard accelerators
// (Alt+key) and text editing shortcuts (Ctrl+C/V/X).

int32_t platformKeyboardGetModifiers(void) {
    __dpmi_regs r;

    memset(&r, 0, sizeof(r));
    r.x.ax = 0x1200;
    __dpmi_int(0x16, &r);

    return r.x.ax & 0xFF;
}


// ============================================================
// platformKeyboardRead
// ============================================================
//
// Non-blocking keyboard read using enhanced INT 16h functions.
//
// Uses the "enhanced" functions (10h/11h) rather than the original
// (00h/01h) because the originals can't distinguish between grey
// and numpad arrow keys, and they don't report F11/F12. The enhanced
// functions have been standard since AT-class machines (1984).
//
// The two-step peek-then-read is necessary because function 10h
// (read key) blocks until a key is available — there's no non-blocking
// read in the BIOS API. Function 11h (check key) peeks without
// consuming, letting us poll without blocking the event loop.

bool platformKeyboardRead(PlatformKeyEventT *evt) {
    __dpmi_regs r;

    // Peek: function 11h sets ZF if buffer is empty
    r.x.ax = 0x1100;
    __dpmi_int(0x16, &r);

    // Test the Zero Flag (bit 6 of the flags register)
    if (r.x.flags & 0x40) {
        return false;
    }

    // Consume: function 10h removes the key from the BIOS buffer.
    // AH = scan code, AL = ASCII character (0 for extended keys).
    r.x.ax = 0x1000;
    __dpmi_int(0x16, &r);

    evt->scancode = (r.x.ax >> 8) & 0xFF;
    evt->ascii    = r.x.ax & 0xFF;

    // Enhanced INT 16h uses 0xE0 as the ASCII byte for grey/extended
    // keys (arrows, Home, End, Insert, Delete on 101-key keyboards).
    // Normalize to 0 so the rest of the codebase can use a single
    // "ascii == 0 means extended key, check scancode" convention.
    if (evt->ascii == 0xE0) {
        evt->ascii = 0;
    }

    return true;
}


// ============================================================
// platformMouseInit
// ============================================================
//
// Initializes the INT 33h mouse driver. The mouse driver is a TSR
// (or emulated by the DOS environment) that tracks position and
// buttons independently of the application.
//
// We must set the movement range to match our VESA resolution,
// because the default range may be 640x200 (CGA text mode).
// Without this, mouse coordinates would be wrong or clipped.
//
// The hardware cursor is never shown — DVX composites its own
// software cursor on top of the backbuffer. We only use INT 33h
// for position/button state via polling (function 03h).

void platformMouseInit(int32_t screenW, int32_t screenH) {
    __dpmi_regs r;

    // Function 00h: reset driver, detect mouse hardware
    memset(&r, 0, sizeof(r));
    r.x.ax = 0x0000;
    __dpmi_int(0x33, &r);

    // Function 07h: set horizontal min/max range
    memset(&r, 0, sizeof(r));
    r.x.ax = 0x0007;
    r.x.cx = 0;
    r.x.dx = screenW - 1;
    __dpmi_int(0x33, &r);

    // Function 08h: set vertical min/max range
    memset(&r, 0, sizeof(r));
    r.x.ax = 0x0008;
    r.x.cx = 0;
    r.x.dx = screenH - 1;
    __dpmi_int(0x33, &r);

    // Function 04h: warp cursor to center of screen
    memset(&r, 0, sizeof(r));
    r.x.ax = 0x0004;
    r.x.cx = screenW / 2;
    r.x.dx = screenH / 2;
    __dpmi_int(0x33, &r);
}


// ============================================================
// platformMousePoll
// ============================================================
//
// Reads current mouse state via INT 33h function 03h.
// Returns: CX=X position, DX=Y position, BX=button state
// (bit 0 = left, bit 1 = right, bit 2 = middle).
//
// Polling is used instead of a callback/event model because the
// DVX event loop already runs at frame rate. Installing a real-mode
// callback for mouse events would add DPMI mode-switch overhead
// on every mickeyed movement, which is wasteful when we only sample
// once per frame anyway.

void platformMousePoll(int32_t *mx, int32_t *my, int32_t *buttons) {
    __dpmi_regs r;

    memset(&r, 0, sizeof(r));
    r.x.ax = 0x0003;
    __dpmi_int(0x33, &r);

    *mx      = r.x.cx;
    *my      = r.x.dx;
    *buttons = r.x.bx;
}


// ============================================================
// platformSpanCopy8
// ============================================================
//
// Copies 'count' 8-bit pixels from src to dst using dword-aligned
// rep movsl for the bulk transfer.
//
// All span operations (Copy8/16/32, Fill8/16/32) follow the same
// pattern: align to a dword boundary, do the bulk as rep movsl or
// rep stosl, then handle the remainder. This pattern exists because
// on 486/Pentium, misaligned dword moves incur a 3-cycle penalty per
// access. Aligning first ensures the critical rep loop runs at full
// bus speed.
//
// rep movsl moves 4 bytes per iteration with hardware loop decrement,
// which is faster than a C for-loop — the CPU string move pipeline
// optimizes sequential memory access patterns.

void platformSpanCopy8(uint8_t *dst, const uint8_t *src, int32_t count) {
    // Align dst to a dword boundary with byte copies
    while (((uintptr_t)dst & 3) && count > 0) {
        *dst++ = *src++;
        count--;
    }

    if (count >= 4) {
        int32_t dwordCount = count >> 2;
        __asm__ __volatile__ (
            "rep movsl"
            : "+D"(dst), "+S"(src), "+c"(dwordCount)
            :
            : "memory"
        );
        dst += dwordCount * 4;
        src += dwordCount * 4;
    }

    int32_t rem = count & 3;
    while (rem-- > 0) {
        *dst++ = *src++;
    }
}


// ============================================================
// platformSpanCopy16
// ============================================================
//
// Copies 'count' 16-bit pixels. Since each pixel is 2 bytes, we
// only need to check bit 1 of the address for dword alignment
// (bit 0 is always clear for 16-bit aligned data). A single
// leading pixel copy brings us to a dword boundary, then rep movsl
// copies pixel pairs as dwords.

void platformSpanCopy16(uint8_t *dst, const uint8_t *src, int32_t count) {
    // Copy one pixel to reach dword alignment if needed
    if (((uintptr_t)dst & 2) && count > 0) {
        *(uint16_t *)dst = *(const uint16_t *)src;
        dst += 2;
        src += 2;
        count--;
    }

    if (count >= 2) {
        int32_t dwordCount = count >> 1;
        __asm__ __volatile__ (
            "rep movsl"
            : "+D"(dst), "+S"(src), "+c"(dwordCount)
            :
            : "memory"
        );
        dst += dwordCount * 4;
        src += dwordCount * 4;
    }

    if (count & 1) {
        *(uint16_t *)dst = *(const uint16_t *)src;
    }
}


// ============================================================
// platformSpanCopy32
// ============================================================
//
// 32-bit pixels are inherently dword-aligned, so no alignment
// preamble is needed — straight to rep movsl.

void platformSpanCopy32(uint8_t *dst, const uint8_t *src, int32_t count) {
    __asm__ __volatile__ (
        "rep movsl"
        : "+D"(dst), "+S"(src), "+c"(count)
        :
        : "memory"
    );
}


// ============================================================
// platformSpanFill8
// ============================================================
//
// Fills 'count' 8-bit pixels with a single color value.
// The 8-bit value is replicated into all four bytes of a dword so
// that rep stosl writes 4 identical pixels per iteration. This is
// 4x faster than byte-at-a-time for large fills (window backgrounds,
// screen clears).

void platformSpanFill8(uint8_t *dst, uint32_t color, int32_t count) {
    uint8_t  c     = (uint8_t)color;
    uint32_t dword = (uint32_t)c | ((uint32_t)c << 8) | ((uint32_t)c << 16) | ((uint32_t)c << 24);

    // Align to 4 bytes — skip if already aligned
    if (__builtin_expect((uintptr_t)dst & 3, 0)) {
        while (((uintptr_t)dst & 3) && count > 0) {
            *dst++ = c;
            count--;
        }
    }

    if (count >= 4) {
        int32_t dwordCount = count >> 2;
        __asm__ __volatile__ (
            "rep stosl"
            : "+D"(dst), "+c"(dwordCount)
            : "a"(dword)
            : "memory"
        );
        dst += dwordCount * 4;
    }

    int32_t rem = count & 3;
    while (rem-- > 0) {
        *dst++ = c;
    }
}


// ============================================================
// platformSpanFill16
// ============================================================
//
// Fills 'count' 16-bit pixels. Two pixels are packed into a dword
// (low half = first pixel, high half = second pixel) so rep stosl
// writes 2 pixels per iteration.

void platformSpanFill16(uint8_t *dst, uint32_t color, int32_t count) {
    uint16_t c = (uint16_t)color;

    // Handle odd leading pixel for dword alignment
    if (((uintptr_t)dst & 2) && count > 0) {
        *(uint16_t *)dst = c;
        dst += 2;
        count--;
    }

    // Fill pairs of pixels as 32-bit dwords
    if (count >= 2) {
        uint32_t dword      = ((uint32_t)c << 16) | c;
        int32_t  dwordCount = count >> 1;
        __asm__ __volatile__ (
            "rep stosl"
            : "+D"(dst), "+c"(dwordCount)
            : "a"(dword)
            : "memory"
        );
        dst += dwordCount * 4;
    }

    // Handle trailing odd pixel
    if (count & 1) {
        *(uint16_t *)dst = c;
    }
}


// ============================================================
// platformSpanFill32
// ============================================================
//
// 32-bit fill is the simplest case — each pixel is already a dword,
// so rep stosl writes exactly one pixel per iteration with no
// alignment or packing concerns.

void platformSpanFill32(uint8_t *dst, uint32_t color, int32_t count) {
    __asm__ __volatile__ (
        "rep stosl"
        : "+D"(dst), "+c"(count)
        : "a"(color)
        : "memory"
    );
}


// ============================================================
// platformValidateFilename — DOS 8.3 filename validation
// ============================================================
//
// Validates that a filename conforms to DOS 8.3 conventions:
//   - Base name: 1-8 chars, extension: 0-3 chars, one dot max
//   - No spaces or special characters that DOS can't handle
//   - Not a reserved device name (CON, PRN, AUX, NUL, COMn, LPTn)
//
// The reserved name check compares the base name only (before the
// dot), case-insensitive, because DOS treats "CON.TXT" the same
// as the CON device — the extension is ignored for device names.
//
// Returns NULL on success, or a human-readable error string on failure.
// On non-DOS platforms, this function would be replaced with one that
// validates for that platform's filesystem rules.

const char *platformValidateFilename(const char *name) {
    static const char *reserved[] = {
        "CON", "PRN", "AUX", "NUL",
        "COM1", "COM2", "COM3", "COM4", "COM5", "COM6", "COM7", "COM8", "COM9",
        "LPT1", "LPT2", "LPT3", "LPT4", "LPT5", "LPT6", "LPT7", "LPT8", "LPT9",
        NULL
    };

    if (!name || name[0] == '\0') {
        return "Filename must not be empty.";
    }

    // Split into base and extension
    const char *dot = strrchr(name, '.');
    int32_t baseLen;
    int32_t extLen;

    if (dot) {
        baseLen = (int32_t)(dot - name);
        extLen  = (int32_t)strlen(dot + 1);
    } else {
        baseLen = (int32_t)strlen(name);
        extLen  = 0;
    }

    if (baseLen < 1 || baseLen > 8) {
        return "Filename must be 1-8 characters before the extension.";
    }

    if (extLen > 3) {
        return "Extension must be 3 characters or fewer.";
    }

    // Check for invalid characters
    for (const char *p = name; *p; p++) {
        if (*p == '.') {
            continue;
        }

        if (*p < '!' || *p > '~') {
            return "Filename contains invalid characters.";
        }

        if (strchr(" \"*+,/:;<=>?[\\]|", *p)) {
            return "Filename contains invalid characters.";
        }
    }

    // Check for multiple dots
    if (dot && strchr(name, '.') != dot) {
        return "Filename may contain only one dot.";
    }

    // Check reserved device names (compare base only, case-insensitive)
    char base[9];
    int32_t copyLen = baseLen < 8 ? baseLen : 8;

    for (int32_t i = 0; i < copyLen; i++) {
        base[i] = toupper((unsigned char)name[i]);
    }

    base[copyLen] = '\0';

    for (const char **r = reserved; *r; r++) {
        if (strcmp(base, *r) == 0) {
            return "That name is a reserved device name.";
        }
    }

    return NULL;
}


// ============================================================
// platformVideoInit
// ============================================================
//
// Complete video initialization sequence:
//   1. findBestMode() — enumerate VESA modes and pick the best match
//   2. setVesaMode() — actually switch to the chosen mode with LFB
//   3. mapLfb() — DPMI-map the physical framebuffer into linear memory
//   4. Allocate system RAM backbuffer (same size as LFB)
//   5. Set up 8-bit palette if needed
//   6. Initialize clip rect to full display
//
// The backbuffer is allocated in system RAM rather than drawing
// directly to the LFB because: (a) reads from the LFB are extremely
// slow on ISA/VLB/PCI (uncached MMIO), so any compositing that reads
// pixels would crawl; (b) the dirty rect system only flushes changed
// regions, so most of the LFB is never touched per frame.

int32_t platformVideoInit(DisplayT *d, int32_t requestedW, int32_t requestedH, int32_t preferredBpp) {
    uint16_t bestMode;
    uint32_t physAddr;

    memset(d, 0, sizeof(*d));

    // Find the best VESA mode
    if (findBestMode(requestedW, requestedH, preferredBpp, &bestMode, d) != 0) {
        return -1;
    }

    // Save the physical address before we overwrite it
    physAddr = (uint32_t)(uintptr_t)d->lfb;

    // Set the mode
    if (setVesaMode(bestMode) != 0) {
        return -1;
    }

    // Map the LFB
    if (mapLfb(d, physAddr) != 0) {
        return -1;
    }

    // Allocate backbuffer
    uint32_t fbSize = (uint32_t)d->pitch * (uint32_t)d->height;
    d->backBuf = (uint8_t *)malloc(fbSize);

    if (!d->backBuf) {
        fprintf(stderr, "VBE: Failed to allocate %lu byte backbuffer\n", (unsigned long)fbSize);
        __djgpp_nearptr_disable();
        return -1;
    }

    memset(d->backBuf, 0, fbSize);

    // Set up palette for 8-bit mode
    if (d->format.bitsPerPixel == 8) {
        d->palette = (uint8_t *)malloc(768);

        if (!d->palette) {
            fprintf(stderr, "VBE: Failed to allocate palette\n");
            free(d->backBuf);
            d->backBuf = NULL;
            __djgpp_nearptr_disable();
            return -1;
        }

        dvxGeneratePalette(d->palette);
        platformVideoSetPalette(d->palette, 0, 256);
    }

    // Initialize clip rect to full display
    d->clipX = 0;
    d->clipY = 0;
    d->clipW = d->width;
    d->clipH = d->height;

    fprintf(stderr, "VBE: Mode 0x%04X set: %ldx%ldx%ld, pitch=%ld\n",
            bestMode, (long)d->width, (long)d->height, (long)d->format.bitsPerPixel, (long)d->pitch);

    return 0;
}


// ============================================================
// platformVideoSetPalette
// ============================================================
//
// Programs the VGA DAC palette registers via direct port I/O.
// Port 0x3C8 = write index, port 0x3C9 = data (auto-increments).
//
// The VGA DAC expects 6-bit values (0-63) but our palette stores
// 8-bit values (0-255), hence the >> 2 shift. This is standard
// VGA behavior dating back to the original IBM VGA in 1987.
//
// Direct port I/O is used instead of VBE function 09h (set palette)
// because the VGA DAC ports are faster (no BIOS call overhead) and
// universally compatible — even VBE 3.0 cards still have the standard
// VGA DAC at ports 0x3C8/0x3C9.

void platformVideoSetPalette(const uint8_t *pal, int32_t firstEntry, int32_t count) {
    outportb(0x3C8, (uint8_t)firstEntry);

    for (int32_t i = 0; i < count; i++) {
        int32_t idx = (firstEntry + i) * 3;
        outportb(0x3C9, pal[idx + 0] >> 2);
        outportb(0x3C9, pal[idx + 1] >> 2);
        outportb(0x3C9, pal[idx + 2] >> 2);
    }
}


// ============================================================
// platformVideoShutdown
// ============================================================
//
// Tears down the graphics mode and restores standard 80x25 text
// mode (BIOS mode 3) so the user gets their DOS prompt back.
// Also frees the backbuffer, palette, and disables near pointers
// (re-enables DJGPP's segment limit checking for safety).

void platformVideoShutdown(DisplayT *d) {
    // INT 10h function 00h, mode 03h = 80x25 color text
    __dpmi_regs r;
    memset(&r, 0, sizeof(r));
    r.x.ax = 0x0003;
    __dpmi_int(0x10, &r);

    if (d->backBuf) {
        free(d->backBuf);
        d->backBuf = NULL;
    }

    if (d->palette) {
        free(d->palette);
        d->palette = NULL;
    }

    d->lfb = NULL;

    __djgpp_nearptr_disable();
}


// ============================================================
// platformYield
// ============================================================
//
// Cooperative yield to the DPMI host. In a multitasking DOS
// environment (Windows 3.x DOS box, OS/2 VDM, or DESQview),
// this gives other tasks a chance to run. Under a single-tasking
// DPMI server (CWSDPMI) it's essentially a no-op, but it doesn't
// hurt. Called once per event loop iteration when idle.

void platformYield(void) {
    __dpmi_yield();
}


// ============================================================
// setVesaMode
// ============================================================
//
// Sets a VBE video mode via function 0x4F02. Bit 14 of the mode
// number tells the BIOS to enable the Linear Frame Buffer instead
// of the default banked mode. This is the only mode we support.

static int32_t setVesaMode(uint16_t mode) {
    __dpmi_regs r;

    memset(&r, 0, sizeof(r));
    r.x.ax = 0x4F02;
    r.x.bx = mode | 0x4000; // bit 14 = use LFB
    __dpmi_int(0x10, &r);

    if (r.x.ax != 0x004F) {
        fprintf(stderr, "VBE: Failed to set mode 0x%04X (AX=0x%04X)\n", mode, r.x.ax);
        return -1;
    }

    return 0;
}