package pdf

import "core:math"
import "core:mem"

// Anti-aliased scanline fill.
//
// WHICH ALGORITHM, AND WHY
//
// This is a signed-area cell accumulator in the style of FreeType's `smooth`
// renderer, not a sub-scanline sampler. Every line segment is chopped at the
// scanline boundaries it crosses and then at the pixel boundaries it crosses
// within a scanline, and each resulting piece deposits two numbers into the
// cell it lies in:
//
//     cover[i] += dy                  the signed height the edge spans
//     area [i] += dy * (fx1 + fx2)    twice the trapezoid between the edge
//                                     and the LEFT wall of the cell
//
// A left-to-right prefix sum over `cover` then gives the winding-weighted
// coverage of every cell exactly:
//
//     coverage(i) = sum(cover[j] for j < i) + cover[i] - area[i]/2
//
// which is the integral of the winding number over the pixel -- exact in both
// axes for a polygon, with no sampling error at all. Sixteen sub-scanlines
// would quantise vertical coverage to 1/16 and cost sixteen passes over every
// edge; this costs one, and a near-horizontal glyph stem comes out smooth
// instead of banded. The one place it is an approximation rather than an
// identity is a pixel that a path enters with winding +1 and leaves with
// winding -1: the integral cancels where the true nonzero rule would fill.
// That is a sub-pixel artefact at a self-intersection, it is what FreeType,
// stb_truetype and every other accumulator-based rasteriser do, and it is
// invisible next to the cost of the alternative.
//
// The prefix sum is kept in floating point rather than 26.6 fixed point
// because the input is already f64 device coordinates: rounding to a fixed
// grid first would throw away precision that the analytic area then pretends
// to have.

// Span_Proc receives the coverage of one horizontal run on scanline `y`,
// beginning at device pixel `x0`. Values are 0..255. The slice is scratch
// owned by the rasteriser and is only valid for the duration of the call.
//
// Runs arrive in increasing `y`, and in increasing `x` within a row, and never
// overlap, so a consumer can write straight into a row without sorting.
@(private)
Span_Proc :: #type proc(user: rawptr, y: int, x0: int, coverage: []u8)

// Raster_Clip is the device-pixel box outside which nothing is drawn, given as
// a half-open range: x0 <= x < x1, y0 <= y < y1.
@(private)
Raster_Clip :: struct {
    x0, y0, x1, y1: int,
}

// MAX_RASTER_EDGES caps the edge table. A page-sized fill needs a handful of
// edges and the most complex real content -- a vector map -- a few hundred
// thousand; a million is past anything legitimate and stops a hostile content
// stream from turning a 200-byte path into gigabytes of edges. Past the cap
// the remaining segments are dropped, which distorts the shape but keeps the
// page rendering, per the degrade-don't-fail rule.
@(private)
MAX_RASTER_EDGES :: 1 << 20

// RASTER_COORD_MAX is where coordinates stop being geometry and start being
// noise. Everything is clipped to the clip box anyway, so the only job of this
// clamp is to keep infinities out of the arithmetic that computes the clip
// intersection.
@(private)
RASTER_COORD_MAX :: 1e15

// Raster_Edge is one non-horizontal segment, already clipped to the raster
// box in both axes, oriented downwards (y0 < y1) with the original direction
// kept in `dir` -- the winding number needs the direction, the sweep needs the
// order, and carrying both separately is cheaper than re-deriving either.
@(private)
Raster_Edge :: struct {
    x0, y0: f64,
    x1, y1: f64,
    // 1/(y1-y0), for interpolating x at a scanline boundary.
    inv_dy: f64,
    // +1 when the segment ran down the page, -1 when it ran up it.
    dir:    f64,
    // Next edge starting on the same scanline, or -1.
    next:   i32,
}

@(private)
Raster_State :: struct {
    // The raster box: the clip intersected with the path's bounds. Cells run
    // rx0 ..= rx1 inclusive -- one past the right edge is a guard cell that
    // collects the winding of everything clipped off to the right and is never
    // emitted.
    rx0, rx1, ry0, ry1: int,
    fx0, fx1:           f64,

    edges:    [dynamic]Raster_Edge,
    // Head of the list of edges that start on each scanline, indexed from ry0.
    buckets:  []i32,
    active:   []i32,
    n_active: int,

    cover:  []f64,
    area:   []f64,
    covbuf: []u8,

    // Cells touched on the row being built, so the next row only has to clear
    // what was written. A page of text is thousands of narrow rows; clearing
    // the full clip width for each would dominate the whole fill.
    min_x, max_x: int,
}

// raster_fill scan-converts `fp` and reports its coverage through `span`.
//
// Subpaths are closed implicitly whether or not `Flat_Subpath.closed` is set:
// that is what the fill operators mean (ISO 32000-2, 8.5.3.3.1). Horizontal
// segments are dropped, since they carry no winding.
//
// Degenerate input -- no path, no points, a single point, a subpath that never
// leaves one horizontal line, a path wholly outside `clip` -- produces no
// spans and no error. `allocator` is used for scratch only; everything it
// allocates is released before this returns.
@(private)
raster_fill :: proc(fp:       ^Flat_Path,
                    even_odd: bool,
                    clip:     Raster_Clip,
                    span:     Span_Proc,
                    user:     rawptr,
                    allocator := context.allocator) -> Error {
    if span == nil { return .Invalid_Argument }
    if fp == nil || len(fp.points) == 0 || len(fp.subpaths) == 0 { return .None }
    if clip.x1 <= clip.x0 || clip.y1 <= clip.y0 { return .None }

    // Narrowing the raster box to the path is not just an optimisation: it is
    // what keeps a glyph in the corner of a 300 dpi page from allocating and
    // clearing a 2550-cell row, and what bounds a path whose bounds are 1e9
    // wide to the size of the clip.
    b := flat_path_bounds(fp)
    r := Raster_State{
        rx0 = raster_floor_clamped(b.x0, clip.x0, clip.x1),
        rx1 = raster_ceil_clamped (b.x1, clip.x0, clip.x1),
        ry0 = raster_floor_clamped(b.y0, clip.y0, clip.y1),
        ry1 = raster_ceil_clamped (b.y1, clip.y0, clip.y1),
    }
    if r.rx1 <= r.rx0 || r.ry1 <= r.ry0 { return .None }
    r.fx0 = f64(r.rx0)
    r.fx1 = f64(r.rx1)

    width  := r.rx1 - r.rx0
    height := r.ry1 - r.ry0

    // One allocation each, up front. Nothing below this line allocates, which
    // is what makes the per-scanline cost independent of the path.
    segments := 0
    for sp in fp.subpaths {
        if sp.count >= 2 { segments += sp.count }
    }
    if segments <= 0 { return .None }

    aerr: mem.Allocator_Error
    r.edges, aerr = make([dynamic]Raster_Edge, 0, min(segments, 1024), allocator)
    if aerr != nil { return .Out_Of_Memory }
    defer delete(r.edges)

    r.buckets, aerr = make([]i32, height, allocator)
    if aerr != nil { return .Out_Of_Memory }
    defer delete(r.buckets, allocator)
    for &h in r.buckets { h = -1 }

    r.cover, aerr = make([]f64, width + 1, allocator)
    if aerr != nil { return .Out_Of_Memory }
    defer delete(r.cover, allocator)

    r.area, aerr = make([]f64, width + 1, allocator)
    if aerr != nil { return .Out_Of_Memory }
    defer delete(r.area, allocator)

    r.covbuf, aerr = make([]u8, width, allocator)
    if aerr != nil { return .Out_Of_Memory }
    defer delete(r.covbuf, allocator)

    for i in 0 ..< len(fp.subpaths) {
        pts := flat_path_slice(fp, i)
        if len(pts) < 2 { continue }
        for j in 0 ..< len(pts) - 1 {
            raster_add_edge(&r, pts[j], pts[j + 1])
        }
        // The closing segment. A subpath whose last point already coincides
        // with its first would only add a zero-length edge, which is dropped
        // by the horizontal test anyway, but skipping it keeps the edge table
        // honest for the common case of an explicitly closed rectangle.
        if pts[len(pts) - 1] != pts[0] {
            raster_add_edge(&r, pts[len(pts) - 1], pts[0])
        }
    }
    if len(r.edges) == 0 { return .None }

    r.active, aerr = make([]i32, len(r.edges), allocator)
    if aerr != nil { return .Out_Of_Memory }
    defer delete(r.active, allocator)

    for y in r.ry0 ..< r.ry1 {
        for e := r.buckets[y - r.ry0]; e >= 0; {
            next := r.edges[e].next
            if r.n_active < len(r.active) {
                r.active[r.n_active] = e
                r.n_active += 1
            }
            e = next
        }
        if r.n_active == 0 { continue }

        row_top := f64(y)
        row_bot := f64(y + 1)
        r.min_x = r.rx1 + 1
        r.max_x = r.rx0 - 1

        for k in 0 ..< r.n_active {
            e := &r.edges[r.active[k]]
            ya := max(row_top, e.y0)
            yb := min(row_bot, e.y1)
            if yb <= ya { continue }
            xa := ya <= e.y0 ? e.x0 : raster_edge_x(&r, e, ya)
            xb := yb >= e.y1 ? e.x1 : raster_edge_x(&r, e, yb)
            raster_accumulate(&r, xa, ya, xb, yb, e.dir)
        }

        raster_sweep(&r, y, even_odd, span, user)

        if r.min_x <= r.max_x {
            for i in r.min_x ..= r.max_x {
                r.cover[i - r.rx0] = 0
                r.area [i - r.rx0] = 0
            }
        }

        w := 0
        for k in 0 ..< r.n_active {
            if r.edges[r.active[k]].y1 > row_bot {
                r.active[w] = r.active[k]
                w += 1
            }
        }
        r.n_active = w
    }
    return .None
}

// ---------------------------------------------------------------------------
// Building the edge table
// ---------------------------------------------------------------------------

// raster_add_edge clips one segment to the raster box and appends the one to
// three edges that survive.
//
// Clipping in y is a plain truncation: coverage on a scanline depends only on
// the parts of edges inside that scanline, so anything above or below the box
// cannot affect it. Clipping in x is NOT a truncation -- a segment to the left
// of the box still winds every pixel to its right -- so the parts outside are
// replaced by vertical edges standing on the box's walls, which wind
// identically. That substitution is what makes a path 10^9 units wide cost the
// same as one the width of the clip.
@(private)
raster_add_edge :: proc(r: ^Raster_State, p0, p1: Point) {
    if len(r.edges) >= MAX_RASTER_EDGES { return }

    x0 := raster_sane(p0.x)
    y0 := raster_sane(p0.y)
    x1 := raster_sane(p1.x)
    y1 := raster_sane(p1.y)

    if y0 == y1 { return }
    dir := 1.0
    if y0 > y1 {
        x0, x1 = x1, x0
        y0, y1 = y1, y0
        dir = -1
    }

    top := f64(r.ry0)
    bot := f64(r.ry1)
    cy0 := max(y0, top)
    cy1 := min(y1, bot)
    if cy1 <= cy0 { return }
    cx0 := cy0 == y0 ? x0 : raster_lerp_x(x0, y0, x1, y1, cy0)
    cx1 := cy1 == y1 ? x1 : raster_lerp_x(x0, y0, x1, y1, cy1)

    // Split where the segment crosses the left and right walls, so that each
    // piece is wholly inside or wholly outside and clamping its endpoints is
    // exact rather than a bend.
    ys: [4]f64
    n := 0
    ys[n] = cy0; n += 1
    dx := cx1 - cx0
    if dx != 0 {
        for wall in ([2]f64{r.fx0, r.fx1}) {
            if (cx0 < wall) == (cx1 < wall) { continue }
            t := (wall - cx0) / dx
            if !(t > 0 && t < 1) { continue }
            yy := cy0 + (cy1 - cy0) * t
            if yy > cy0 && yy < cy1 {
                ys[n] = yy; n += 1
            }
        }
    }
    ys[n] = cy1; n += 1
    if n == 4 && ys[1] > ys[2] {
        ys[1], ys[2] = ys[2], ys[1]
    }

    for i in 0 ..< n - 1 {
        ya := ys[i]
        yb := ys[i + 1]
        if yb <= ya { continue }
        xa := clamp(ya == cy0 ? cx0 : raster_lerp_x(cx0, cy0, cx1, cy1, ya), r.fx0, r.fx1)
        xb := clamp(yb == cy1 ? cx1 : raster_lerp_x(cx0, cy0, cx1, cy1, yb), r.fx0, r.fx1)
        raster_push_edge(r, xa, ya, xb, yb, dir)
    }
}

@(private)
raster_push_edge :: proc(r: ^Raster_State, x0, y0, x1, y1, dir: f64) {
    if len(r.edges) >= MAX_RASTER_EDGES { return }
    dy := y1 - y0
    if !(dy > 0) { return }

    row := int(math.floor(y0))
    row = clamp(row, r.ry0, r.ry1 - 1)

    idx := i32(len(r.edges))
    append(&r.edges, Raster_Edge{
        x0     = x0,
        y0     = y0,
        x1     = x1,
        y1     = y1,
        inv_dy = 1 / dy,
        dir    = dir,
        next   = r.buckets[row - r.ry0],
    })
    r.buckets[row - r.ry0] = idx
}

// raster_edge_x interpolates the edge's x at a scanline boundary. The guards
// are not decoration: an edge whose y extent is denormally small has an
// infinite 1/dy, and 0 * infinity is a NaN that would otherwise reach the cell
// walk and index an array with it.
@(private)
raster_edge_x :: proc(r: ^Raster_State, e: ^Raster_Edge, y: f64) -> f64 {
    t := (y - e.y0) * e.inv_dy
    if !is_finite(t) { t = 0 }
    t = clamp(t, 0, 1)
    x := e.x0 + (e.x1 - e.x0) * t
    if !is_finite(x) { x = e.x0 }
    return clamp(x, r.fx0, r.fx1)
}

@(private)
raster_lerp_x :: proc "contextless" (x0, y0, x1, y1, y: f64) -> f64 {
    dy := y1 - y0
    t := dy != 0 ? (y - y0) / dy : 0
    if !is_finite(t) { t = 0 }
    t = clamp(t, 0, 1)
    x := x0 + (x1 - x0) * t
    return is_finite(x) ? x : x0
}

@(private)
raster_sane :: proc "contextless" (v: f64) -> f64 {
    if v != v { return 0 }
    return clamp(v, -RASTER_COORD_MAX, RASTER_COORD_MAX)
}

// raster_floor_clamped and raster_ceil_clamped convert a bound to a pixel
// index without ever letting a huge or non-finite f64 reach the integer
// conversion, which is undefined when the value does not fit. A NaN bound
// widens to the clip rather than narrowing to nothing: dropping content is a
// worse failure than drawing too much.
@(private)
raster_floor_clamped :: proc "contextless" (v: f64, lo, hi: int) -> int {
    if v != v || v <= f64(lo) { return lo }
    if v >= f64(hi) { return hi }
    return int(math.floor(v))
}

@(private)
raster_ceil_clamped :: proc "contextless" (v: f64, lo, hi: int) -> int {
    if v != v || v >= f64(hi) { return hi }
    if v <= f64(lo) { return lo }
    return int(math.ceil(v))
}

// ---------------------------------------------------------------------------
// Accumulation
// ---------------------------------------------------------------------------

@(private)
raster_cell :: proc "contextless" (r: ^Raster_State, x: f64) -> int {
    if !(x > r.fx0) { return r.rx0 }
    if x >= r.fx1 { return r.rx1 }
    return int(math.floor(x))
}

// raster_accumulate deposits one piece of one edge -- already confined to a
// single scanline -- into the cells it crosses.
@(private)
raster_accumulate :: proc(r: ^Raster_State, xa, ya, xb, yb, dir: f64) {
    if !(yb > ya) { return }

    ex1 := raster_cell(r, xa)
    ex2 := raster_cell(r, xb)

    if ex1 == ex2 {
        raster_cell_add(r, ex1, (yb - ya) * dir, xa, xb)
        return
    }

    // dy per unit x. Both endpoints sit inside the raster box, so the number
    // of cells crossed is bounded by its width.
    dydx := (yb - ya) / (xb - xa)
    step := xb > xa ? 1 : -1
    ex := ex1
    xc := xa
    yc := ya
    for ex != ex2 {
        xn := f64(step > 0 ? ex + 1 : ex)
        yn := ya + (xn - xa) * dydx
        if !(yn > yc) { yn = yc }
        if yn > yb    { yn = yb }
        raster_cell_add(r, ex, (yn - yc) * dir, xc, xn)
        xc = xn
        yc = yn
        ex += step
    }
    raster_cell_add(r, ex, (yb - yc) * dir, xc, xb)
}

// raster_cell_add records `dy` of signed height crossing cell `ex` between the
// x offsets `x_from` and `x_to`. `area` holds twice the trapezoid on the left
// of the edge, which the sweep subtracts from the running cover.
@(private)
raster_cell_add :: proc "contextless" (r: ^Raster_State, ex: int, dy, x_from, x_to: f64) {
    if dy == 0 { return }
    i := ex - r.rx0
    if i < 0 || i >= len(r.cover) { return }
    fx := f64(ex)
    r.cover[i] += dy
    r.area [i] += dy * ((x_from - fx) + (x_to - fx))
    r.min_x = min(r.min_x, ex)
    r.max_x = max(r.max_x, ex)
}

// ---------------------------------------------------------------------------
// Sweep
// ---------------------------------------------------------------------------

// raster_coverage folds a winding-weighted coverage into a 0..255 alpha.
//
// Nonzero clamps the magnitude; even-odd folds it, which is the floating-point
// spelling of the usual integer trick of masking the accumulator to nine bits
// and mirroring everything above the halfway point. Even-odd is not a corner
// case: it is what `f*` and `B*` mean, and Type1 glyph counters depend on it.
@(private)
raster_coverage :: proc "contextless" (cov: f64, even_odd: bool) -> u8 {
    a := abs(cov)
    if even_odd {
        a -= 2 * math.floor(a * 0.5)
        if a > 1 { a = 2 - a }
    } else if a > 1 {
        a = 1
    }
    // Written as a positive test so a NaN that survived every other guard
    // becomes transparent rather than an out-of-range conversion.
    if !(a > 0) { return 0 }
    v := int(a * 255 + 0.5)
    return u8(clamp(v, 0, 255))
}

@(private)
raster_sweep :: proc(r: ^Raster_State, y: int, even_odd: bool, span: Span_Proc, user: rawptr) {
    if r.min_x > r.max_x { return }

    lo := r.min_x
    // The sweep stops at the last cell that any edge touched, and never needs
    // a tail: a closed path's windings sum to zero at the far right, and the
    // ones that would have closed beyond the clip were moved onto the guard
    // cell rather than dropped, so `running` is back at zero by the time the
    // touched range ends. The guard cell itself is one past the last emittable
    // pixel, which is why `hi` is capped rather than clamped away.
    hi := min(r.max_x, r.rx1 - 1)
    if lo > hi { return }

    end := hi + 1
    running := 0.0
    for i in lo ..= hi {
        j := i - r.rx0
        c := running + r.cover[j] - r.area[j] * 0.5
        running += r.cover[j]
        r.covbuf[j] = raster_coverage(c, even_odd)
    }

    i := lo
    for i < end {
        if r.covbuf[i - r.rx0] == 0 {
            i += 1
            continue
        }
        j := i
        for j < end && r.covbuf[j - r.rx0] != 0 { j += 1 }
        span(user, y, i, r.covbuf[i - r.rx0 : j - r.rx0])
        i = j
    }
}