#include <stdio.h>
#include <stdint.h>
#include <string.h>
#include <stdlib.h>

#include "../include/gl.h"
#include "../include/glext.h"
#include "private.h"
#include "profiler.h"

typedef struct {
    const void* ptr;
    GLenum type;
    GLsizei stride;
    GLint size;
} AttribPointer;


static AttribPointer VERTEX_POINTER;
static AttribPointer UV_POINTER;
static AttribPointer ST_POINTER;
static AttribPointer NORMAL_POINTER;
static AttribPointer DIFFUSE_POINTER;

#define VERTEX_ENABLED_FLAG     (1 << 0)
#define UV_ENABLED_FLAG         (1 << 1)
#define ST_ENABLED_FLAG         (1 << 2)
#define DIFFUSE_ENABLED_FLAG    (1 << 3)
#define NORMAL_ENABLED_FLAG     (1 << 4)

static GLuint ENABLED_VERTEX_ATTRIBUTES = 0;
static GLubyte ACTIVE_CLIENT_TEXTURE = 0;

void initAttributePointers() {
    TRACE();

    VERTEX_POINTER.ptr = NULL;
    VERTEX_POINTER.stride = 0;
    VERTEX_POINTER.type = GL_FLOAT;
    VERTEX_POINTER.size = 4;

    DIFFUSE_POINTER.ptr = NULL;
    DIFFUSE_POINTER.stride = 0;
    DIFFUSE_POINTER.type = GL_FLOAT;
    DIFFUSE_POINTER.size = 4;

    UV_POINTER.ptr = NULL;
    UV_POINTER.stride = 0;
    UV_POINTER.type = GL_FLOAT;
    UV_POINTER.size = 4;

    ST_POINTER.ptr = NULL;
    ST_POINTER.stride = 0;
    ST_POINTER.type = GL_FLOAT;
    ST_POINTER.size = 4;

    NORMAL_POINTER.ptr = NULL;
    NORMAL_POINTER.stride = 0;
    NORMAL_POINTER.type = GL_FLOAT;
    NORMAL_POINTER.size = 3;
}

static inline GLuint byte_size(GLenum type) {
    switch(type) {
    case GL_BYTE: return sizeof(GLbyte);
    case GL_UNSIGNED_BYTE: return sizeof(GLubyte);
    case GL_SHORT: return sizeof(GLshort);
    case GL_UNSIGNED_SHORT: return sizeof(GLushort);
    case GL_INT: return sizeof(GLint);
    case GL_UNSIGNED_INT: return sizeof(GLuint);
    case GL_DOUBLE: return sizeof(GLdouble);
    case GL_FLOAT:
    default: return sizeof(GLfloat);
    }
}

typedef void (*FloatParseFunc)(GLfloat* out, const GLubyte* in);
typedef void (*PolyBuildFunc)(ClipVertex* first, ClipVertex* previous, ClipVertex* vertex, ClipVertex* next, const GLsizei i);

static inline void _parseVec3FromShort3(GLfloat* out, const GLubyte* in) {
    GLshort* ptr = (GLshort*) in;

    out[0] = (GLfloat) ptr[0];
    out[1] = (GLfloat) ptr[1];
    out[2] = (GLfloat) ptr[2];
}

static inline void _parseVec3FromInt3(GLfloat* out, const GLubyte* in) {
    GLint* ptr = (GLint*) in;

    out[0] = (GLfloat) ptr[0];
    out[1] = (GLfloat) ptr[1];
    out[2] = (GLfloat) ptr[2];
}

static inline void _parseVec3FromFloat3(GLfloat* out, const GLubyte* in) {
    GLfloat* ptr = (GLfloat*) in;

    out[0] = ptr[0];
    out[1] = ptr[1];
    out[2] = ptr[2];
}

static inline void _parseVec2FromFloat2(GLfloat* out, const GLubyte* in) {
    GLfloat* ptr = (GLfloat*) in;

    out[0] = ptr[0];
    out[1] = ptr[1];
}

static inline void _parseVec3FromFloat2(GLfloat* out, const GLubyte* in) {
    GLfloat* ptr = (GLfloat*) in;

    out[0] = ptr[0];
    out[1] = ptr[1];
    out[2] = 0.0f;
}

static inline void _parseVec4FromFloat3(GLfloat* out, const GLubyte* in) {
    GLfloat* ptr = (GLfloat*) in;

    out[0] = ptr[0];
    out[1] = ptr[1];
    out[2] = ptr[2];
    out[3] = 1.0;
}

static inline void _parseVec4FromFloat4(GLfloat* out, const GLubyte* in) {
    GLfloat* ptr = (GLfloat*) in;

    out[0] = ptr[0];
    out[1] = ptr[1];
    out[2] = ptr[2];
    out[3] = ptr[3];
}

static inline void _parseColourFromUByte4(GLfloat* out, const GLubyte* in) {
    const float ONE_OVER_255 = 1.0f / 255.0f;
    out[0] = ((GLfloat) in[0]) * ONE_OVER_255;
    out[1] = ((GLfloat) in[1]) * ONE_OVER_255;
    out[2] = ((GLfloat) in[2]) * ONE_OVER_255;
    out[3] = ((GLfloat) in[3]) * ONE_OVER_255;
}

static inline void _constVec2Zero(GLfloat* out, const GLubyte* in) {
    out[0] = 0.0f;
    out[1] = 0.0f;
}

static inline void _constVec3NegZ(GLfloat* out, const GLubyte* in) {
    out[0] = 0.0f;
    out[1] = 0.0f;
    out[2] = -1.0f;
}

static inline void _constVec4One(GLfloat* out, const GLubyte* in) {
    out[0] = 1.0f;
    out[1] = 1.0f;
    out[2] = 1.0f;
    out[3] = 1.0f;
}

typedef GLuint (*IndexParseFunc)(const GLubyte* in);

static inline GLuint _parseUByteIndex(const GLubyte* in) {
    return (GLuint) *in;
}

static inline GLuint _parseUIntIndex(const GLubyte* in) {
    return *((GLuint*) in);
}

static inline GLuint _parseUShortIndex(const GLubyte* in) {
    return *((GLshort*) in);
}


static inline IndexParseFunc _calcParseIndexFunc(GLenum type) {
    switch(type) {
    case GL_UNSIGNED_BYTE:
        return &_parseUByteIndex;
    break;
    case GL_UNSIGNED_INT:
        return &_parseUIntIndex;
    break;
    case GL_UNSIGNED_SHORT:
    default:
        break;
    }

    return &_parseUShortIndex;
}


/* There was a bug in this macro that shipped with Kos
 * which has now been fixed. But just in case...
 */
#undef mat_trans_single3_nodiv
#define mat_trans_single3_nodiv(x, y, z) { \
    register float __x __asm__("fr12") = (x); \
    register float __y __asm__("fr13") = (y); \
    register float __z __asm__("fr14") = (z); \
    __asm__ __volatile__( \
                          "fldi1 fr15\n" \
                          "ftrv  xmtrx, fv12\n" \
                          : "=f" (__x), "=f" (__y), "=f" (__z) \
                          : "0" (__x), "1" (__y), "2" (__z) \
                          : "fr15"); \
    x = __x; y = __y; z = __z; \
}


/* FIXME: Is this right? Shouldn't it be fr12->15? */
#undef mat_trans_normal3
#define mat_trans_normal3(x, y, z) { \
    register float __x __asm__("fr8") = (x); \
    register float __y __asm__("fr9") = (y); \
    register float __z __asm__("fr10") = (z); \
    __asm__ __volatile__( \
                          "fldi0 fr11\n" \
                          "ftrv  xmtrx, fv8\n" \
                          : "=f" (__x), "=f" (__y), "=f" (__z) \
                          : "0" (__x), "1" (__y), "2" (__z) \
                          : "fr11"); \
    x = __x; y = __y; z = __z; \
}


static inline void transformToEyeSpace(GLfloat* point) {
    _matrixLoadModelView();
    mat_trans_single3_nodiv(point[0], point[1], point[2]);
}

static inline void transformNormalToEyeSpace(GLfloat* normal) {
    _matrixLoadNormal();
    mat_trans_normal3(normal[0], normal[1], normal[2]);
}

static inline void swapVertex(ClipVertex* v1, ClipVertex* v2) {
    static ClipVertex tmp;

    tmp = *v1;
    *v1 = *v2;
    *v2 = tmp;
}

static inline FloatParseFunc _calcVertexParseFunc() {
    switch(VERTEX_POINTER.type) {
    case GL_SHORT: {
        if(VERTEX_POINTER.size == 3) {
            return &_parseVec3FromShort3;
        }
    } break;
    case GL_INT: {
        if(VERTEX_POINTER.size == 3) {
            return &_parseVec3FromInt3;
        }
    } break;
    case GL_FLOAT: {
        if(VERTEX_POINTER.size == 3) {
            return &_parseVec3FromFloat3;
        } else if(VERTEX_POINTER.size == 2) {
            return &_parseVec3FromFloat2;
        }
    } break;
    default:
        break;
    }

    return NULL;
}

static inline FloatParseFunc _calcDiffuseParseFunc() {
    if((ENABLED_VERTEX_ATTRIBUTES & DIFFUSE_ENABLED_FLAG) != DIFFUSE_ENABLED_FLAG) {
        return &_constVec4One;
    }

    switch(DIFFUSE_POINTER.type) {
    case GL_BYTE:
    case GL_UNSIGNED_BYTE: {
        if(DIFFUSE_POINTER.size == 4) {
            return &_parseColourFromUByte4;
        }
    } break;
    case GL_INT: {
        if(DIFFUSE_POINTER.size == 3) {
            return &_parseVec3FromInt3;
        }
    } break;
    case GL_FLOAT: {
        if(DIFFUSE_POINTER.size == 3) {
            return &_parseVec4FromFloat3;
        } else if(DIFFUSE_POINTER.size == 4) {
            return &_parseVec4FromFloat4;
        }
    } break;
    default:
        break;
    }

    return &_constVec4One;
}

static inline FloatParseFunc _calcUVParseFunc() {
    if((ENABLED_VERTEX_ATTRIBUTES & UV_ENABLED_FLAG) != UV_ENABLED_FLAG) {
        return &_constVec2Zero;
    }

    switch(UV_POINTER.type) {
    case GL_FLOAT: {
        if(UV_POINTER.size == 2) {
            return &_parseVec2FromFloat2;
        }
    } break;
    default:
        break;
    }

    return &_constVec2Zero;
}

static inline FloatParseFunc _calcSTParseFunc() {
    if((ENABLED_VERTEX_ATTRIBUTES & ST_ENABLED_FLAG) != ST_ENABLED_FLAG) {
        return &_constVec2Zero;
    }

    switch(ST_POINTER.type) {
    case GL_FLOAT: {
        if(ST_POINTER.size == 2) {
            return &_parseVec2FromFloat2;
        }
    } break;
    default:
        break;
    }

    return &_constVec2Zero;
}

static inline FloatParseFunc _calcNormalParseFunc() {
    if((ENABLED_VERTEX_ATTRIBUTES & NORMAL_ENABLED_FLAG) != NORMAL_ENABLED_FLAG) {
        return &_constVec3NegZ;
    }

    switch(NORMAL_POINTER.type) {
    case GL_SHORT: {
        if(NORMAL_POINTER.size == 3) {
            return &_parseVec3FromShort3;
        }
    } break;
    case GL_INT: {
        if(NORMAL_POINTER.size == 3) {
            return &_parseVec3FromInt3;
        }
    } break;
    case GL_FLOAT: {
        if(NORMAL_POINTER.size == 3) {
            return &_parseVec3FromFloat3;
        } else if(NORMAL_POINTER.size == 2) {
            return &_parseVec3FromFloat2;
        }
    } break;
    default:
        break;
    }

    return &_constVec3NegZ;
}


static void _buildTriangle(ClipVertex* first, ClipVertex* previous, ClipVertex* vertex, ClipVertex* next, const GLsizei i) {
    if(((i + 1) % 3) == 0) {
        vertex->flags = PVR_CMD_VERTEX_EOL;
    }
}

static inline GLsizei fast_mod(const GLsizei input, const GLsizei ceil) {
    return input >= ceil ? input % ceil : input;
}

static void _buildQuad(ClipVertex* first, ClipVertex* previous, ClipVertex* vertex, ClipVertex* next, const GLsizei i) {
    if((i + 1) % 4 == 0) {
        previous->flags = PVR_CMD_VERTEX_EOL;
        swapVertex(previous, vertex);
    }
}

static void _buildTriangleFan(ClipVertex* first, ClipVertex* previous, ClipVertex* vertex, ClipVertex* next, const GLsizei i) {
    if(i == 2) {
        swapVertex(previous, vertex);
        vertex->flags = PVR_CMD_VERTEX_EOL;
    } else if(i > 2) {
        ClipVertex* next = vertex + 1;

        *next = *first;

        swapVertex(next, vertex);

        vertex = next + 1;
        *vertex = *previous;

        vertex->flags = PVR_CMD_VERTEX_EOL;
    }
}

static void _buildStrip(ClipVertex* first, ClipVertex* previous, ClipVertex* vertex, ClipVertex* next, const GLsizei i) {
    if(!next) {
        /* If the mode was triangle strip, then the last vertex is the last vertex */
        vertex->flags = PVR_CMD_VERTEX_EOL;
    }
}

static inline PolyBuildFunc _calcBuildFunc(const GLenum type) {
    switch(type) {
    case GL_TRIANGLES:
        return &_buildTriangle;
    break;
    case GL_QUADS:
        return &_buildQuad;
    break;
    case GL_TRIANGLE_FAN:
    case GL_POLYGON:
        return &_buildTriangleFan;
    break;
    default:
        break;
    }

    return &_buildStrip;
}

typedef struct {
    const GLubyte* vptr;
    const GLuint vstride;
    const GLubyte* cptr;
    const GLuint cstride;
    const GLubyte* uvptr;
    const GLuint uvstride;
    const GLubyte* stptr;
    const GLuint ststride;
    const GLubyte* nptr;
    const GLuint nstride;
} GenerateParams;

static void generate(AlignedVector* output, const GLenum mode, const GLsizei first, const GLsizei count,
        const GLubyte* indices, const GLenum type, const GenerateParams* pointers) {
    /* Read from the client buffers and generate an array of ClipVertices */

    const GLsizei max = first + count;
    const GLsizei spaceNeeded = (mode == GL_POLYGON || mode == GL_TRIANGLE_FAN) ? ((count - 2) * 3) : count;

    /* Make sure we have room for the output */
    ClipVertex* vertex = aligned_vector_resize(output, spaceNeeded);

    const FloatParseFunc vertexFunc = _calcVertexParseFunc();
    const FloatParseFunc diffuseFunc = _calcDiffuseParseFunc();
    const FloatParseFunc uvFunc = _calcUVParseFunc();
    const FloatParseFunc stFunc = _calcSTParseFunc();
    const FloatParseFunc normalFunc = _calcNormalParseFunc();

    const PolyBuildFunc buildFunc = _calcBuildFunc(mode);
    const IndexParseFunc indexFunc = _calcParseIndexFunc(type);

    const GLsizei type_byte_size = byte_size(type);

    ClipVertex* previous = NULL;
    ClipVertex* firstV = vertex;
    ClipVertex* next = NULL;

    GLsizei i;

    for(i = first; i < max; ++i, ++vertex) {
        vertex->flags = PVR_CMD_VERTEX;

        const GLuint idx = (indices) ?
            indexFunc(&indices[type_byte_size * i]) : i;

        const GLubyte* vin = pointers->vptr + (idx * pointers->vstride);
        const GLubyte* din = pointers->cptr + (idx * pointers->cstride);
        const GLubyte* uin = pointers->uvptr + (idx * pointers->uvstride);
        const GLubyte* sin = pointers->stptr + (idx * pointers->ststride);
        const GLubyte* nin = pointers->nptr + (idx * pointers->nstride);

        vertexFunc(vertex->xyz, vin);
        diffuseFunc(vertex->diffuse, din);
        uvFunc(vertex->uv, uin);
        stFunc(vertex->st, sin);
        normalFunc(vertex->nxyz, nin);
    }

    vertex = firstV;

    for(i = 0; i < count; ++i, ++vertex) {
        next = (i < count - 1) ? vertex + 1 : NULL;
        previous = (i > 0) ? vertex - 1 : NULL;
        buildFunc(firstV, previous, vertex, next, i);
    }
}

static void transform(AlignedVector* vertices) {
    /* Perform modelview transform, storing W */

    ClipVertex* vertex = (ClipVertex*) vertices->data;

    _applyRenderMatrix(); /* Apply the Render Matrix Stack */

    GLsizei i;
    for(i = 0; i < vertices->size; ++i, ++vertex) {
        register float __x __asm__("fr12") = (vertex->xyz[0]);
        register float __y __asm__("fr13") = (vertex->xyz[1]);
        register float __z __asm__("fr14") = (vertex->xyz[2]);
        register float __w __asm__("fr15");

        __asm__ __volatile__(
            "fldi1 fr15\n"
            "ftrv   xmtrx,fv12\n"
            : "=f" (__x), "=f" (__y), "=f" (__z), "=f" (__w)
            : "0" (__x), "1" (__y), "2" (__z), "3" (__w)
        );

        vertex->xyz[0] = __x;
        vertex->xyz[1] = __y;
        vertex->xyz[2] = __z;
        vertex->w = __w;
    }
}

static void clip(AlignedVector* vertices) {
    /* Perform clipping, generating new vertices as necessary */

    static AlignedVector* CLIP_BUFFER = NULL;

    /* First entry into this, allocate the clip buffer */
    if(!CLIP_BUFFER) {
        CLIP_BUFFER = (AlignedVector*) malloc(sizeof(AlignedVector));
        aligned_vector_init(CLIP_BUFFER, sizeof(ClipVertex));
    }

    /* Make sure we allocate roughly enough space */
    aligned_vector_reserve(CLIP_BUFFER, vertices->size * 1.5);

    /* Start from empty */
    aligned_vector_resize(CLIP_BUFFER, 0);

    /* Now perform clipping! */
    clipTriangleStrip(vertices, CLIP_BUFFER);

    /* Copy the clip buffer over the vertices */
    aligned_vector_resize(vertices, CLIP_BUFFER->size);
    memcpy(vertices->data, CLIP_BUFFER->data, CLIP_BUFFER->size * CLIP_BUFFER->element_size);
}

static void mat_transform3(const float* xyz, const float* xyzOut, const uint32_t count, const uint32_t stride) {
    uint8_t* dataIn = (uint8_t*) xyz;
    uint8_t* dataOut = (uint8_t*) xyzOut;
    uint32_t i = count;

    while(i--) {
        float* in = (float*) dataIn;
        float* out = (float*) dataOut;

        mat_trans_single3_nodiv_nomod(in[0], in[1], in[2], out[0], out[1], out[2]);

        dataIn += stride;
        dataOut += stride;
    }
}

static void mat_transform_normal3(const float* xyz, const float* xyzOut, const uint32_t count, const uint32_t stride) {
    uint8_t* dataIn = (uint8_t*) xyz;
    uint8_t* dataOut = (uint8_t*) xyzOut;
    uint32_t i = count;

    while(i--) {
        float* in = (float*) dataIn;
        float* out = (float*) dataOut;

        mat_trans_normal3_nomod(in[0], in[1], in[2], out[0], out[1], out[2]);

        dataIn += stride;
        dataOut += stride;
    }
}

static void light(AlignedVector* vertices) {
    if(!isLightingEnabled()) {
        return;
    }

    /* Perform lighting calculations and manipulate the colour */
    ClipVertex* vertex = (ClipVertex*) vertices->data;

    _matrixLoadModelView();
    mat_transform3(vertex->xyz, vertex->xyzES, vertices->size, sizeof(ClipVertex));

    _matrixLoadNormal();
    mat_transform_normal3(vertex->nxyz, vertex->nES, vertices->size, sizeof(ClipVertex));

    GLsizei i;
    for(i = 0; i < vertices->size; ++i, ++vertex) {
        /* We ignore diffuse colour when lighting is enabled. If GL_COLOR_MATERIAL is enabled
         * then the lighting calculation should possibly take it into account */
        memset(vertex->diffuse, 0, sizeof(float) * 4);

        GLfloat to_add [] = {0.0f, 0.0f, 0.0f, 0.0f};

        GLubyte j;
        for(j = 0; j < MAX_LIGHTS; ++j) {
            if(isLightEnabled(j)) {
                calculateLightingContribution(j, vertex->xyzES, vertex->nES, to_add);

                vertex->diffuse[0] += to_add[0];
                vertex->diffuse[1] += to_add[1];
                vertex->diffuse[2] += to_add[2];
                vertex->diffuse[3] += to_add[3];
            }
        }
    }
}

static void divide(AlignedVector* vertices) {
    /* Perform perspective divide on each vertex */
    ClipVertex* vertex = (ClipVertex*) vertices->data;

    GLsizei i;
    for(i = 0; i < vertices->size; ++i, ++vertex) {
        vertex->xyz[2] = 1.0f / vertex->w;
        vertex->xyz[0] *= vertex->xyz[2];
        vertex->xyz[1] *= vertex->xyz[2];
    }
}

typedef struct {
    unsigned int list_type;
    pvr_poly_hdr_t hdr;
} ListToHeader;


#define MAX_LISTS 5

static void push(const AlignedVector* vertices, PolyList* activePolyList, GLshort textureUnit) {
    /* Copy the vertices to the active poly list */

    static GLuint LIST_COUNTER = 0;
    static ListToHeader LAST_HEADERS[MAX_LISTS];

    /* If the list was empty, this is the first submission this frame so we
     * always send the header in this case */
    GLboolean listWasEmpty = activePolyList->vector.size > 0;

    // Make room for the element + the header
    PVRCommand* dst = (PVRCommand*) aligned_vector_extend(&activePolyList->vector, vertices->size);

    // Store a pointer to the header
    pvr_poly_hdr_t* hdr = (pvr_poly_hdr_t*) dst;

    // Compile
    pvr_poly_cxt_t cxt = *getPVRContext();
    cxt.list_type = activePolyList->list_type;

    _glUpdatePVRTextureContext(&cxt, textureUnit);

    pvr_poly_compile(hdr, &cxt);

    /* We store a list of the last "hdr" to be submitted for a list, and then compare before
     * specifying another one. Apparently it's quite slow to change header
     */

    GLuint c;
    GLboolean sendHeader = GL_FALSE;
    GLboolean listFound = GL_FALSE;

    for(c = 0; c < LIST_COUNTER; ++c) {
        if(LAST_HEADERS[c].list_type == activePolyList->list_type) {
            /* Send the header if this was the first submission to this list, or the header has changed since
             * the last sent */
            sendHeader = listWasEmpty || memcmp(&LAST_HEADERS[c].hdr, &hdr, sizeof(pvr_poly_hdr_t)) != 0;
            listFound = GL_TRUE;
            break;
        }
    }

    if(!listFound) {
        if(LIST_COUNTER == MAX_LISTS) {
            fprintf(stderr, "Ran out of space!\n");
        }
        /* First time we've seen this list, add it to the array */
        LAST_HEADERS[LIST_COUNTER].list_type = activePolyList->list_type;
        LAST_HEADERS[LIST_COUNTER++].hdr = *hdr;

        /* Send the header the first time */
        sendHeader = GL_TRUE;
    }

    if(sendHeader) {
        // Point dest at the first new vertex to populate, if we're not sending a header
        // we won't increment and instead overwrite the header we just created with the
        // first vertex
        dst++;

        // Add one more to the list
        aligned_vector_extend(&activePolyList->vector, 1);
    }

    GLsizei i;
    ClipVertex* vin = aligned_vector_at(vertices, 0);

    for(i = 0; i < vertices->size; ++i, dst++) {
        pvr_vertex_t* vout = (pvr_vertex_t*) dst;
        vout->flags = vin->flags;
        vout->x = vin->xyz[0];
        vout->y = vin->xyz[1];
        vout->z = vin->xyz[2];
        vout->u = vin->uv[0];
        vout->v = vin->uv[1];
        vout->argb = PVR_PACK_COLOR(vin->diffuse[3], vin->diffuse[0], vin->diffuse[1], vin->diffuse[2]);
        vout->oargb = 0;

        vin++;
    }
}

static void submitVertices(GLenum mode, GLsizei first, GLsizei count, GLenum type, const GLvoid* indices) {
    static AlignedVector* buffer = NULL;

    /* Do nothing if vertices aren't enabled */
    if(!(ENABLED_VERTEX_ATTRIBUTES & VERTEX_ENABLED_FLAG)) {
        return;
    }

    /* Initialize the buffer on first call */
    if(!buffer) {
        buffer = (AlignedVector*) malloc(sizeof(AlignedVector));
        aligned_vector_init(buffer, sizeof(ClipVertex));

        /* Reserve 64k up-front */
        aligned_vector_reserve(buffer, 64 * 1024);
    } else {
        /* Else, resize to zero (this will retain the allocated memory) */
        aligned_vector_resize(buffer, 0);
    }

    const GLuint vstride = (VERTEX_POINTER.stride) ? VERTEX_POINTER.stride : VERTEX_POINTER.size * byte_size(VERTEX_POINTER.type);
    const GLubyte* vptr = VERTEX_POINTER.ptr;

    const GLuint cstride = (DIFFUSE_POINTER.stride) ? DIFFUSE_POINTER.stride : DIFFUSE_POINTER.size * byte_size(DIFFUSE_POINTER.type);
    const GLubyte* cptr = DIFFUSE_POINTER.ptr;

    const GLuint uvstride = (UV_POINTER.stride) ? UV_POINTER.stride : UV_POINTER.size * byte_size(UV_POINTER.type);
    const GLubyte* uvptr = UV_POINTER.ptr;

    const GLuint ststride = (ST_POINTER.stride) ? ST_POINTER.stride : ST_POINTER.size * byte_size(ST_POINTER.type);
    const GLubyte* stptr = ST_POINTER.ptr;

    const GLuint nstride = (NORMAL_POINTER.stride) ? NORMAL_POINTER.stride : NORMAL_POINTER.size * byte_size(NORMAL_POINTER.type);
    const GLubyte* nptr = NORMAL_POINTER.ptr;

    GenerateParams params = {
        .vptr = vptr,
        .vstride = vstride,
        .cptr = cptr,
        .cstride = cstride,
        .uvptr = uvptr,
        .uvstride = uvstride,
        .stptr = stptr,
        .ststride = ststride,
        .nptr = nptr,
        .nstride = nstride
    };

    generate(buffer, mode, first, count, (GLubyte*) indices, type, &params);

    light(buffer);

    transform(buffer);

    if(isClippingEnabled()) {
        clip(buffer);
    }

    divide(buffer);

    push(buffer, activePolyList(), 0);

    /*
       Now, if multitexturing is enabled, we want to send exactly the same vertices again, except:
       - We want to enable blending, and send them to the TR list
       - We want to set the depth func to GL_EQUAL
       - We want to set the second texture ID
       - We want to set the uv coordinates to the passed st ones
    */

    GLboolean doMultitexture;
    glGetBooleanv(GL_TEXTURE_2D, &doMultitexture);

    if(!doMultitexture) {
        /* Multitexture actively disabled */
        return;
    }

    TextureObject* texture1 = getTexture1();

    if(!texture1 || ((ENABLED_VERTEX_ATTRIBUTES & ST_ENABLED_FLAG) != ST_ENABLED_FLAG)) {
        /* Multitexture implicitly disabled */
        return;
    }

    ClipVertex* vertex = (ClipVertex*) aligned_vector_at(buffer, 0);

    /* Copy ST coordinates to UV ones */
    GLsizei i = 0;
    for(; i < buffer->size; ++i, ++vertex) {
        vertex->uv[0] = vertex->st[0];
        vertex->uv[1] = vertex->st[1];
    }

    /* Store state, as we're about to mess around with it */
    GLint depthFunc, blendSrc, blendDst;
    glGetIntegerv(GL_DEPTH_FUNC, &depthFunc);
    glGetIntegerv(GL_BLEND_SRC, &blendSrc);
    glGetIntegerv(GL_BLEND_DST, &blendDst);

    GLboolean blendEnabled = glIsEnabled(GL_BLEND);
    GLboolean depthEnabled = glIsEnabled(GL_DEPTH_TEST);

    glDepthFunc(GL_EQUAL);
    glEnable(GL_BLEND);
    glBlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);

    /* Send the buffer again to the transparent list */
    push(buffer, transparentPolyList(), 1);

    /* Reset state */
    glDepthFunc(depthFunc);
    glBlendFunc(blendSrc, blendDst);
    (blendEnabled) ? glEnable(GL_BLEND) : glDisable(GL_BLEND);
    (depthEnabled) ? glEnable(GL_DEPTH_TEST) : glDisable(GL_DEPTH_TEST);
}

void APIENTRY glDrawElements(GLenum mode, GLsizei count, GLenum type, const GLvoid* indices) {
    TRACE();

    if(checkImmediateModeInactive(__func__)) {
        return;
    }

    submitVertices(mode, 0, count, type, indices);
}

void APIENTRY glDrawArrays(GLenum mode, GLint first, GLsizei count) {
    TRACE();

    if(checkImmediateModeInactive(__func__)) {
        return;
    }

    submitVertices(mode, first, count, GL_UNSIGNED_SHORT, NULL);
}

void APIENTRY glEnableClientState(GLenum cap) {
    TRACE();

    switch(cap) {
    case GL_VERTEX_ARRAY:
        ENABLED_VERTEX_ATTRIBUTES |= VERTEX_ENABLED_FLAG;
    break;
    case GL_COLOR_ARRAY:
        ENABLED_VERTEX_ATTRIBUTES |= DIFFUSE_ENABLED_FLAG;
    break;
    case GL_NORMAL_ARRAY:
        ENABLED_VERTEX_ATTRIBUTES |= NORMAL_ENABLED_FLAG;
    break;
    case GL_TEXTURE_COORD_ARRAY:
        (ACTIVE_CLIENT_TEXTURE) ?
            (ENABLED_VERTEX_ATTRIBUTES |= ST_ENABLED_FLAG):
            (ENABLED_VERTEX_ATTRIBUTES |= UV_ENABLED_FLAG);
    break;
    default:
        _glKosThrowError(GL_INVALID_ENUM, "glEnableClientState");
    }
}

void APIENTRY glDisableClientState(GLenum cap) {
    TRACE();

    switch(cap) {
    case GL_VERTEX_ARRAY:
        ENABLED_VERTEX_ATTRIBUTES &= ~VERTEX_ENABLED_FLAG;
    break;
    case GL_COLOR_ARRAY:
        ENABLED_VERTEX_ATTRIBUTES &= ~DIFFUSE_ENABLED_FLAG;
    break;
    case GL_NORMAL_ARRAY:
        ENABLED_VERTEX_ATTRIBUTES &= ~NORMAL_ENABLED_FLAG;
    break;
    case GL_TEXTURE_COORD_ARRAY:
        (ACTIVE_CLIENT_TEXTURE) ?
            (ENABLED_VERTEX_ATTRIBUTES &= ~ST_ENABLED_FLAG):
            (ENABLED_VERTEX_ATTRIBUTES &= ~UV_ENABLED_FLAG);
    break;
    default:
        _glKosThrowError(GL_INVALID_ENUM, "glDisableClientState");
    }
}

void APIENTRY glClientActiveTextureARB(GLenum texture) {
    TRACE();

    if(texture < GL_TEXTURE0_ARB || texture > GL_TEXTURE0_ARB + MAX_TEXTURE_UNITS) {
        _glKosThrowError(GL_INVALID_ENUM, "glClientActiveTextureARB");
    }

    if(_glKosHasError()) {
        _glKosPrintError();
        return;
    }

    ACTIVE_CLIENT_TEXTURE = (texture == GL_TEXTURE1_ARB) ? 1 : 0;
}

void APIENTRY glTexCoordPointer(GLint size,  GLenum type,  GLsizei stride,  const GLvoid * pointer) {
    TRACE();

    AttribPointer* tointer = (ACTIVE_CLIENT_TEXTURE == 0) ? &UV_POINTER : &ST_POINTER;

    tointer->ptr = pointer;
    tointer->stride = stride;
    tointer->type = type;
    tointer->size = size;
}

void APIENTRY glVertexPointer(GLint size,  GLenum type,  GLsizei stride,  const GLvoid * pointer) {
    TRACE();

    VERTEX_POINTER.ptr = pointer;
    VERTEX_POINTER.stride = stride;
    VERTEX_POINTER.type = type;
    VERTEX_POINTER.size = size;
}

void APIENTRY glColorPointer(GLint size,  GLenum type,  GLsizei stride,  const GLvoid * pointer) {
    TRACE();

    DIFFUSE_POINTER.ptr = pointer;
    DIFFUSE_POINTER.stride = stride;
    DIFFUSE_POINTER.type = type;
    DIFFUSE_POINTER.size = size;
}

void APIENTRY glNormalPointer(GLenum type,  GLsizei stride,  const GLvoid * pointer) {
    TRACE();

    NORMAL_POINTER.ptr = pointer;
    NORMAL_POINTER.stride = stride;
    NORMAL_POINTER.type = type;
    NORMAL_POINTER.size = 3;
}