pdf-odin-llm/font_type1.odin

package pdf

import "core:mem"
import "core:slice"

// Type 1 font programs (Adobe, "Adobe Type 1 Font Format", 1990).
//
// Bytes in, outlines out. Nothing here knows what a PDF is; a Type 1 program
// is the same file whether it arrived as a /FontFile stream, a .pfb on disk or
// a .pfa pasted into a PostScript job, and keeping this file ignorant of the
// difference is what makes it testable from byte literals.
//
// SHAPE OF THE FILE
//
//   %!FontType1-1.0 ...          cleartext PostScript: /FontMatrix, /Encoding
//   currentfile eexec
//   <encrypted>                  the Private dict: /lenIV, /Subrs, /CharStrings
//   512 zeros, cleartomark       trailer
//
// The encrypted portion is enciphered with the "eexec" stream cipher keyed at
// r = 55665, whose first four plaintext bytes are random padding. It arrives
// either as raw binary (.pfb, and PDF /FontFile streams) or as hex ASCII
// (.pfa); both are accepted, told apart by whether the first four bytes are hex
// digits -- which works because Adobe's own encryptor is required to choose
// padding that keeps the two forms distinguishable. Each charstring inside is
// enciphered again with the same cipher keyed at r = 4330, with /lenIV bytes
// (4 unless the font says otherwise) of padding to drop.
//
// A .pfb wraps the three parts in segment headers -- 0x80, a type byte, then a
// little-endian length -- which are stripped and the payloads concatenated
// before anything else looks at the bytes.
//
// WHAT SUBSETTERS DO TO THESE FILES
//
// Font programs are the most reliably damaged data in a PDF, because every
// producer rewrites them. So: a /Subrs or /CharStrings count that disagrees
// with the entries that follow is ignored in favour of the entries; a
// charstring shorter than /lenIV yields an empty glyph rather than an error; an
// eexec block that stops in the middle yields the glyphs decoded up to that
// point; and a font with no eexec section at all -- some are shipped with that
// layer already stripped -- is read as though its Private dict were the
// cleartext it is.

// ---------------------------------------------------------------------------
// Limits. Every one of these bounds something an attacker controls.
// ---------------------------------------------------------------------------

// Tokens read from one cleartext or Private section. These are data, not
// programs, so the cap only ever fires on garbage -- which is exactly what the
// tail of a truncated eexec block decrypts to.
@(private)
MAX_T1_TOKENS :: 1 << 20
// Glyph indices are u16 everywhere in this layer, so the dictionary stops one
// short of where the index would stop being representable.
@(private)
MAX_T1_GLYPHS :: 0xFFFF
@(private)
MAX_T1_SUBRS :: 1 << 16
// The charstring interpreter's bounds. The format allows 24 operands and a
// call depth of 10; the margin absorbs the fonts that exceed the spec without
// letting a hostile one run away.
@(private)
T1_STACK_MAX :: 48
@(private)
T1_PS_MAX :: 32
@(private)
T1_MAX_CALL_DEPTH :: 30
// Charstring bytes executed for one glyph, subroutine calls included.
@(private)
T1_MAX_FUEL :: 1 << 17
// Points a FLEX collects: a reference point plus the six that make two curves.
@(private)
T1_FLEX_POINTS :: 7

// The two cipher keys, and the constants they share (Type 1 spec, 7.3).
@(private)
T1_EEXEC_KEY :: 55665
@(private)
T1_CHARSTRING_KEY :: 4330

// ---------------------------------------------------------------------------
// The parsed font
// ---------------------------------------------------------------------------

// T1_Name_Entry is one glyph name and the glyph index it belongs to, in a table
// sorted by name so that a lookup is a binary search. A simple Type 1 font is
// addressed entirely by name, so this is the hot path.
@(private)
T1_Name_Entry :: struct {
    name: string,
    gid:  u16,
}

// Type1_Font is the parsed form of a font program. Declared here so that Font can
// reference it; the reader that fills it in owns this file and adds the fields
// it needs. Nothing in this file may depend on Document or on any PDF type --
// a font program is bytes in and outlines out, and stays separately testable.
Type1_Font :: struct {
    // The whole program with any PFB segment headers removed. When the input
    // was not PFB-wrapped this aliases the caller's bytes rather than copying
    // them, so `data` passed to `t1_parse` must outlive the font: the built-in
    // encoding's glyph names point into it.
    data: []byte,
    // The decrypted Private dictionary, always a buffer of this font's own.
    // Glyph names, charstrings and subroutines point into it.
    priv: []byte,

    // /FontName, for diagnostics only.
    name: string,
    // /FontMatrix, defaulting to the 1000-units-per-em [0.001 0 0 0.001 0 0].
    mtx:  Matrix,

    // The program's built-in encoding: 256 glyph names, "" where the font
    // leaves a code undefined. Meaningful only when `has_encoding` is set.
    encoding:     [256]string,
    has_encoding: bool,

    // /lenIV: bytes of padding at the head of every decrypted charstring.
    // Negative means the charstrings are not enciphered at all.
    len_iv: int,

    subrs: [][]byte,
    names: []string,
    chars: [][]byte,
    // `names` sorted, for t1_gid_for_name.
    by_name: []T1_Name_Entry,

    // Advances are memoised: laying out a page asks for the same handful of
    // glyphs thousands of times, and each answer costs a charstring run.
    advances:  []f64,
    adv_known: []bool,

    // Set when the eexec block or a charstring ran off the end of the input.
    // The font is still usable; it simply holds fewer glyphs than it claims.
    truncated: bool,
}

// ---------------------------------------------------------------------------
// Parsing
// ---------------------------------------------------------------------------

// t1_parse reads a Type 1 font program: PFB-wrapped, raw binary or PFA.
//
// The returned font and everything it points at are allocated with `allocator`
// and belong to the caller; nothing is freed individually. `data` is referenced,
// not copied, so it must outlive the font.
//
// An error comes back only when there is genuinely nothing to hand over -- no
// charstring could be read at all. A font that is damaged but still holds one
// glyph parses successfully, because a page that draws that glyph should draw
// it.
@(private)
t1_parse :: proc(data: []byte, allocator := context.allocator) -> (^Type1_Font, Error) {
    if len(data) == 0 {
        return nil, .Bad_Type1
    }

    f := new(Type1_Font, allocator)
    if f == nil {
        return nil, .Out_Of_Memory
    }
    f.mtx = Matrix{0.001, 0, 0, 0.001, 0, 0}
    f.len_iv = 4

    raw, raw_owned := t1_unwrap_pfb(data, allocator)
    f.data = raw

    // The cleartext is everything before `eexec`. For a font that has none,
    // stop at the first thing that introduces binary data instead, so the
    // tokeniser never walks over charstring bytes.
    clear_end := len(raw)
    eexec_at := t1_find_token(raw, "eexec")
    if eexec_at >= 0 {
        clear_end = eexec_at
    }
    if p := t1_find_token(raw[:clear_end], "/CharStrings"); p >= 0 {
        clear_end = p
    }
    if p := t1_find_token(raw[:clear_end], "/Subrs"); p >= 0 {
        clear_end = p
    }
    t1_parse_cleartext(f, raw[:clear_end])

    // The encrypted region: everything after `eexec` and the whitespace that
    // follows it. Adobe requires the first byte of the ciphertext not to be
    // whitespace precisely so that skipping it is unambiguous (Type 1 spec,
    // 7.2).
    body := raw
    encrypted := false
    if eexec_at >= 0 {
        i := eexec_at + len("eexec")
        for i < len(raw) && t1_is_space(raw[i]) {
            i += 1
        }
        body = raw[i:]
        encrypted = true
    }
    // A font whose eexec layer has already been stripped still says `eexec`.
    // Twelve plaintext bytes surviving encipherment is not a thing that
    // happens, so finding the token means the region is not enciphered.
    if encrypted && t1_find_token(body, "/CharStrings") >= 0 {
        encrypted = false
    }

    // `f.priv` is always the whole allocation, so it can be handed back to the
    // allocator; `priv_body` is the part after the four bytes of random
    // padding at the head of the eexec plaintext.
    priv_body: []byte
    switch {
    case encrypted && t1_is_hex_eexec(body):
        buf := make([]byte, len(body) / 2 + 1, allocator)
        n := t1_hex_decode(buf, body)
        t1_decipher(buf[:n], T1_EEXEC_KEY)
        f.priv = buf
        priv_body = n >= 4 ? buf[4:n] : buf[n:n]
    case encrypted:
        f.priv = t1_copy(body, allocator)
        t1_decipher(f.priv, T1_EEXEC_KEY)
        priv_body = len(f.priv) >= 4 ? f.priv[4:] : f.priv[len(f.priv):]
    case:
        f.priv = t1_copy(body, allocator)
        priv_body = f.priv
    }

    ctx := T1_Parse_Ctx{
        f     = f,
        subrs = make([dynamic][]byte, allocator),
        names = make([dynamic]string, allocator),
        chars = make([dynamic][]byte, allocator),
    }
    t1_parse_private(&ctx, priv_body)

    if len(ctx.chars) == 0 {
        delete(ctx.subrs)
        delete(ctx.names)
        delete(ctx.chars)
        delete(f.priv, allocator)
        if raw_owned {
            delete(raw, allocator)
        }
        free(f, allocator)
        return nil, .Bad_Type1
    }

    f.subrs = ctx.subrs[:]
    f.names = ctx.names[:]
    f.chars = ctx.chars[:]

    n := len(f.names)
    f.by_name = make([]T1_Name_Entry, n, allocator)
    for i in 0 ..< n {
        f.by_name[i] = T1_Name_Entry{name = f.names[i], gid = u16(i)}
    }
    slice.sort_by(f.by_name, proc(a, b: T1_Name_Entry) -> bool {
        return a.name < b.name
    })

    f.advances = make([]f64, n, allocator)
    f.adv_known = make([]bool, n, allocator)

    return f, .None
}

// t1_glyph_count is the number of charstrings the font holds. Glyph indices
// are positions in the /CharStrings dictionary, in the order it was written.
@(private)
t1_glyph_count :: proc(f: ^Type1_Font) -> int {
    if f == nil {
        return 0
    }
    return len(f.chars)
}

// t1_font_matrix maps glyph space to text space. Outlines from t1_glyph_path
// are in glyph space and have not been through it.
@(private)
t1_font_matrix :: proc(f: ^Type1_Font) -> Matrix {
    if f == nil {
        return Matrix{0.001, 0, 0, 0.001, 0, 0}
    }
    return f.mtx
}

// t1_builtin_encoding is the font program's own encoding, or nil when it
// declares none -- which the caller must read as "use StandardEncoding",
// not as "this font has no glyphs at those codes".
@(private)
t1_builtin_encoding :: proc(f: ^Type1_Font) -> ^[256]string {
    if f == nil || !f.has_encoding {
        return nil
    }
    return &f.encoding
}

// t1_gid_for_name resolves a glyph name against the /CharStrings dictionary.
@(private)
t1_gid_for_name :: proc(f: ^Type1_Font, name: string) -> (u16, bool) {
    if f == nil || len(name) == 0 {
        return 0, false
    }
    lo, hi := 0, len(f.by_name)
    for lo < hi {
        mid := lo + (hi - lo) / 2
        if f.by_name[mid].name < name {
            lo = mid + 1
        } else {
            hi = mid
        }
    }
    if lo < len(f.by_name) && f.by_name[lo].name == name {
        return f.by_name[lo].gid, true
    }
    return 0, false
}

// t1_glyph_advance is the width `hsbw` or `sbw` declares, in glyph space.
// Zero for a glyph that declares none, which is what a viewer advances by.
@(private)
t1_glyph_advance :: proc(f: ^Type1_Font, gid: u16) -> f64 {
    if f == nil || int(gid) >= len(f.chars) {
        return 0
    }
    i := int(gid)
    if i < len(f.adv_known) && f.adv_known[i] {
        return f.advances[i]
    }

    ip := T1_Interp{width_only = true}
    t1_run(f, i, nil, Point{0, 0}, &ip)
    w := ip.have_width ? ip.width : 0
    if !is_finite(w) {
        w = 0
    }
    if i < len(f.adv_known) {
        f.advances[i] = w
        f.adv_known[i] = true
    }
    return w
}

// t1_glyph_path appends one glyph's outline, in glyph space, to `out`.
//
// A glyph that cannot be interpreted appends whatever it managed to draw and
// still returns .None: a broken glyph is a blank, never a failed page. Only an
// index that names no glyph is an error.
@(private)
t1_glyph_path :: proc(f: ^Type1_Font, gid: u16, out: ^Path) -> Error {
    if f == nil || out == nil {
        return .Invalid_Argument
    }
    if int(gid) >= len(f.chars) {
        return .No_Glyph
    }

    ip := T1_Interp{allow_seac = true}
    t1_run(f, int(gid), out, Point{0, 0}, &ip)
    if !ip.seac {
        return .None
    }

    // seac composes an accented glyph out of two others named by their
    // StandardEncoding codes -- the one place a font program refers to an
    // encoding it does not carry. The table is font_standard.odin's, which is
    // the same table the format's Annex D defines.
    asb := ip.seac_args[0]
    adx := ip.seac_args[1]
    ady := ip.seac_args[2]
    if !is_finite(asb) || !is_finite(adx) || !is_finite(ady) {
        return .None
    }
    bcode := int(ip.seac_args[3])
    acode := int(ip.seac_args[4])
    std := encoding_standard()

    if bcode >= 0 && bcode < 256 {
        if bgid, ok := t1_gid_for_name(f, std[bcode]); ok {
            base := T1_Interp{}
            t1_run(f, int(bgid), out, Point{0, 0}, &base)
        }
    }
    if acode >= 0 && acode < 256 {
        if agid, ok := t1_gid_for_name(f, std[acode]); ok {
            // adx/ady position the accent's left sidebearing point relative to
            // the COMPOSITE's left sidebearing point, not its origin -- so the
            // composite's own sidebearing, which `ip.sbx` holds, is part of the
            // offset, and `asb` (the accent's sidebearing, which the accent's
            // own hsbw is about to apply again) is taken back out.
            //
            // Getting this wrong is invisible in the many fonts whose accented
            // glyphs have a zero sidebearing and glaring in the ones that do
            // not: dropping `ip.sbx` moves the dieresis off the I by the width
            // of the sidebearing in 104 of the 448 seac composites in the
            // Bitstream and URW faces this was checked against.
            acc := T1_Interp{}
            t1_run(f, int(agid), out, Point{ip.sbx + adx - asb, ady}, &acc)
        }
    }
    return .None
}

// ---------------------------------------------------------------------------
// Container: PFB segments, eexec, and the two ciphers
// ---------------------------------------------------------------------------

@(private = "file")
t1_copy :: proc(src: []byte, allocator: mem.Allocator) -> []byte {
    if len(src) == 0 {
        return nil
    }
    out := make([]byte, len(src), allocator)
    if len(out) != len(src) {
        return nil
    }
    copy(out, src)
    return out
}

// t1_unwrap_pfb strips .pfb segment headers and concatenates the payloads.
// Returns the input untouched, and owned = false, for anything else.
@(private = "file")
t1_unwrap_pfb :: proc(data: []byte, allocator: mem.Allocator) -> (out: []byte, owned: bool) {
    if len(data) < 6 || data[0] != 0x80 {
        return data, false
    }

    // Two passes so the join is exactly sized: a declared length is a file
    // value and must not be believed as an allocation size.
    total := 0
    for pass in 0 ..< 2 {
        pos := 0
        n := 0
        for pos + 6 <= len(data) {
            if data[pos] != 0x80 {
                break
            }
            kind := data[pos + 1]
            // 3 is the end-of-file marker; anything else is not a segment.
            if kind != 1 && kind != 2 {
                break
            }
            seg := int(u32(data[pos + 2]) |
                       u32(data[pos + 3]) << 8 |
                       u32(data[pos + 4]) << 16 |
                       u32(data[pos + 5]) << 24)
            pos += 6
            if seg > len(data) - pos {
                seg = len(data) - pos
            }
            if seg <= 0 {
                continue
            }
            if pass == 0 {
                total += seg
            } else {
                copy(out[n:], data[pos:pos + seg])
                n += seg
            }
            pos += seg
        }
        if pass == 0 {
            if total == 0 {
                return data, false
            }
            out = make([]byte, total, allocator)
            if len(out) != total {
                return data, false
            }
        }
    }
    return out, true
}

// t1_decipher runs the eexec cipher over `buf` in place (Type 1 spec, 7.3).
@(private = "file")
t1_decipher :: proc(buf: []byte, key: u16) {
    r := u32(key)
    for i in 0 ..< len(buf) {
        c := u32(buf[i])
        buf[i] = u8(c) ~ u8(r >> 8)
        r = ((c + r) * 52845 + 22719) & 0xFFFF
    }
}

// t1_is_hex_eexec tells the .pfa form from the binary one. Adobe requires the
// random padding to be chosen so that the first four ciphertext bytes are not
// all hex digits, which is what makes this test safe rather than a guess.
@(private = "file")
t1_is_hex_eexec :: proc(body: []byte) -> bool {
    if len(body) < 4 {
        return false
    }
    for i in 0 ..< 4 {
        if _, ok := hex_value(body[i]); !ok {
            return false
        }
    }
    return true
}

// t1_hex_decode decodes the hex-ASCII eexec form into `dst`, returning how many
// bytes it wrote. Decoding stops at the first character that is neither hex nor
// whitespace, which in a .pfa is the `cleartomark` trailer.
@(private = "file")
t1_hex_decode :: proc(dst, src: []byte) -> int {
    n := 0
    hi: u8
    have_hi := false
    for c in src {
        if t1_is_space(c) {
            continue
        }
        v, ok := hex_value(c)
        if !ok {
            break
        }
        if !have_hi {
            hi = v
            have_hi = true
            continue
        }
        if n >= len(dst) {
            break
        }
        dst[n] = hi << 4 | v
        n += 1
        have_hi = false
    }
    return n
}

// t1_decrypt_charstring deciphers one charstring in place and drops its /lenIV
// bytes of padding. A charstring shorter than the padding yields an empty
// glyph, which draws nothing -- the alternative is a slice out of range.
@(private = "file")
t1_decrypt_charstring :: proc(buf: []byte, len_iv: int) -> []byte {
    if len_iv < 0 {
        return buf
    }
    t1_decipher(buf, T1_CHARSTRING_KEY)
    if len(buf) <= len_iv {
        return buf[len(buf):]
    }
    return buf[len_iv:]
}

// ---------------------------------------------------------------------------
// A PostScript-shaped tokeniser
//
// Only enough of the syntax to find the handful of constructs that matter.
// Strings and hex strings come back whole so that their contents can never be
// mistaken for keywords: a /Notice is allowed to contain the word "eexec".
// ---------------------------------------------------------------------------

@(private = "file")
T1_Lexer :: struct {
    src: []byte,
    pos: int,
}

@(private = "file")
t1_is_space :: proc "contextless" (c: byte) -> bool {
    return c == 0 || c == 9 || c == 10 || c == 12 || c == 13 || c == 32
}

@(private = "file")
t1_is_special :: proc "contextless" (c: byte) -> bool {
    switch c {
    case '(', ')', '<', '>', '[', ']', '{', '}', '/', '%':
        return true
    }
    return false
}

@(private = "file")
t1_next :: proc(lx: ^T1_Lexer) -> (tok: []byte, ok: bool) {
    for {
        for lx.pos < len(lx.src) && t1_is_space(lx.src[lx.pos]) {
            lx.pos += 1
        }
        if lx.pos >= len(lx.src) {
            return nil, false
        }
        if lx.src[lx.pos] != '%' {
            break
        }
        for lx.pos < len(lx.src) && lx.src[lx.pos] != '\n' && lx.src[lx.pos] != '\r' {
            lx.pos += 1
        }
    }

    start := lx.pos
    switch lx.src[lx.pos] {
    case '/':
        lx.pos += 1
        for lx.pos < len(lx.src) &&
            !t1_is_space(lx.src[lx.pos]) && !t1_is_special(lx.src[lx.pos]) {
            lx.pos += 1
        }
    case '(':
        depth := 0
        for lx.pos < len(lx.src) {
            c := lx.src[lx.pos]
            lx.pos += 1
            switch c {
            case '\\':
                lx.pos += 1
            case '(':
                depth += 1
            case ')':
                depth -= 1
            }
            if depth == 0 {
                break
            }
        }
        if lx.pos > len(lx.src) {
            lx.pos = len(lx.src)
        }
    case '<':
        lx.pos += 1
        for lx.pos < len(lx.src) && lx.src[lx.pos] != '>' {
            lx.pos += 1
        }
        if lx.pos < len(lx.src) {
            lx.pos += 1
        }
    case ')', '>', '[', ']', '{', '}':
        lx.pos += 1
    case:
        for lx.pos < len(lx.src) &&
            !t1_is_space(lx.src[lx.pos]) && !t1_is_special(lx.src[lx.pos]) {
            lx.pos += 1
        }
    }
    if lx.pos == start {
        // Nothing consumed would stall the caller's loop; a delimiter with no
        // body is still one byte of progress.
        lx.pos += 1
    }
    return lx.src[start:lx.pos], true
}

@(private = "file")
t1_eq :: proc "contextless" (tok: []byte, s: string) -> bool {
    return string(tok) == s
}

// t1_tok_int reads a plain decimal integer, which is all the constructs here
// ever need. Values far past anything meaningful are rejected rather than
// wrapped, because every one of them is a length or a count.
@(private = "file")
t1_tok_int :: proc "contextless" (tok: []byte) -> (int, bool) {
    if len(tok) == 0 {
        return 0, false
    }
    i := 0
    neg := false
    if tok[0] == '+' || tok[0] == '-' {
        neg = tok[0] == '-'
        i = 1
    }
    if i >= len(tok) {
        return 0, false
    }
    v := 0
    for ; i < len(tok); i += 1 {
        c := tok[i]
        if c < '0' || c > '9' {
            return 0, false
        }
        if v > 1 << 30 {
            return 0, false
        }
        v = v * 10 + int(c - '0')
    }
    return neg ? -v : v, true
}

@(private = "file")
t1_next_int :: proc(lx: ^T1_Lexer) -> (int, bool) {
    tok, ok := t1_next(lx)
    if !ok {
        return 0, false
    }
    return t1_tok_int(tok)
}

// t1_find_token locates a whitespace-delimited keyword. Delimiting matters:
// `/Subrs` is not to be found inside `/SubrsCount`, and a bare substring search
// finds `eexec` inside the middle of a word.
@(private = "file")
t1_find_token :: proc(hay: []byte, needle: string) -> int {
    if len(needle) == 0 || len(hay) < len(needle) {
        return -1
    }
    for i in 0 ..= len(hay) - len(needle) {
        if hay[i] != needle[0] {
            continue
        }
        if string(hay[i:i + len(needle)]) != needle {
            continue
        }
        if i > 0 {
            p := hay[i - 1]
            if !t1_is_space(p) && p != ')' && p != ']' && p != '}' && p != '>' {
                continue
            }
        }
        j := i + len(needle)
        if j < len(hay) {
            c := hay[j]
            if !t1_is_space(c) && !t1_is_special(c) {
                continue
            }
        }
        return i
    }
    return -1
}

// ---------------------------------------------------------------------------
// The cleartext portion
// ---------------------------------------------------------------------------

@(private = "file")
t1_parse_cleartext :: proc(f: ^Type1_Font, src: []byte) {
    lx := T1_Lexer{src = src}
    for fuel := MAX_T1_TOKENS; fuel > 0; fuel -= 1 {
        tok, ok := t1_next(&lx)
        if !ok {
            break
        }
        switch {
        case t1_eq(tok, "/FontMatrix"):
            t1_read_matrix(f, &lx)
        case t1_eq(tok, "/Encoding"):
            t1_read_encoding(f, &lx)
        case t1_eq(tok, "/FontName"):
            if name, nok := t1_next(&lx); nok && len(name) > 1 && name[0] == '/' {
                f.name = string(name[1:])
            }
        }
    }
}

// t1_read_matrix reads `[a b c d e f]`. A matrix that is not invertible, or
// that carries a NaN, is discarded in favour of the 1000-per-em default: it
// would otherwise collapse every glyph in the font to a point.
@(private = "file")
t1_read_matrix :: proc(f: ^Type1_Font, lx: ^T1_Lexer) {
    open, ok := t1_next(lx)
    if !ok || (!t1_eq(open, "[") && !t1_eq(open, "{")) {
        return
    }
    vals: [6]f64
    n := 0
    for n < 6 {
        tok, tok_ok := t1_next(lx)
        if !tok_ok || t1_eq(tok, "]") || t1_eq(tok, "}") {
            return
        }
        num, num_ok := parse_number(string(tok))
        v, v_ok := to_number(num)
        if !num_ok || !v_ok {
            return
        }
        vals[n] = v
        n += 1
    }
    m := Matrix{vals[0], vals[1], vals[2], vals[3], vals[4], vals[5]}
    if !matrix_is_finite(m) || matrix_determinant(m) == 0 {
        return
    }
    f.mtx = m
}

// t1_read_encoding reads both spellings the format allows: the
// `StandardEncoding` shorthand, and the `256 array ... dup <code> /<name> put`
// form that every subsetter emits.
@(private = "file")
t1_read_encoding :: proc(f: ^Type1_Font, lx: ^T1_Lexer) {
    tok, ok := t1_next(lx)
    if !ok {
        return
    }
    if t1_eq(tok, "StandardEncoding") {
        f.encoding = encoding_standard()^
        f.has_encoding = true
        return
    }

    depth := 0
    for fuel := MAX_T1_TOKENS; fuel > 0; fuel -= 1 {
        t, tok_ok := t1_next(lx)
        if !tok_ok {
            return
        }
        switch {
        case t1_eq(t, "{"):
            depth += 1
            continue
        case t1_eq(t, "}"):
            depth -= 1
            continue
        }
        if depth <= 0 && (t1_eq(t, "def") || t1_eq(t, "readonly") || t1_eq(t, "eexec")) {
            return
        }
        if depth != 0 || !t1_eq(t, "dup") {
            continue
        }

        // `dup <code> /<name> put`. The three tokens are read together, and
        // anything that does not fit the shape is stepped over rather than
        // consumed: a font is allowed to `dup` for other reasons.
        save := lx.pos
        code, code_ok := t1_next_int(lx)
        name, name_ok := t1_next(lx)
        put, put_ok := t1_next(lx)
        if !code_ok || !name_ok || !put_ok ||
           code < 0 || code > 255 || len(name) < 2 || name[0] != '/' ||
           !t1_eq(put, "put") {
            lx.pos = save
            continue
        }
        f.encoding[code] = string(name[1:])
        f.has_encoding = true
    }
}

// ---------------------------------------------------------------------------
// The Private dictionary
// ---------------------------------------------------------------------------

@(private = "file")
T1_Parse_Ctx :: struct {
    f:     ^Type1_Font,
    subrs: [dynamic][]byte,
    names: [dynamic]string,
    chars: [dynamic][]byte,
}

@(private = "file")
t1_parse_private :: proc(ctx: ^T1_Parse_Ctx, src: []byte) {
    ctx.f.len_iv = t1_scan_len_iv(src)

    lx := T1_Lexer{src = src}
    subrs_open := false
    fuel := MAX_T1_TOKENS
    for fuel > 0 {
        fuel -= 1
        tok, ok := t1_next(&lx)
        if !ok {
            return
        }
        switch {
        case t1_eq(tok, "/CharStrings"):
            t1_skip_to_begin(&lx)
            t1_read_charstrings(ctx, &lx, &fuel)
            return
        case t1_eq(tok, "/Subrs"):
            // The count that follows is not read at all: the array grows to fit
            // the entries that turn up, which is what a subsetter that dropped
            // some without rewriting the count needs, and what stops a count of
            // a billion from being an allocation.
            subrs_open = true
        case subrs_open && t1_eq(tok, "dup"):
            t1_read_subr(ctx, &lx)
        }
    }
}

// t1_scan_len_iv finds /lenIV before any binary data can be reached, so that a
// charstring byte pattern cannot be mistaken for the key.
@(private = "file")
t1_scan_len_iv :: proc(src: []byte) -> int {
    limit := len(src)
    if p := t1_find_token(src, "/Subrs"); p >= 0 {
        limit = min(limit, p)
    }
    if p := t1_find_token(src, "/CharStrings"); p >= 0 {
        limit = min(limit, p)
    }
    at := t1_find_token(src[:limit], "/lenIV")
    if at < 0 {
        return 4
    }
    lx := T1_Lexer{src = src[:limit], pos = at}
    if _, ok := t1_next(&lx); !ok {
        return 4
    }
    v, ok := t1_next_int(&lx)
    // 0..16 is the padding a font can mean. A negative value is how the tools
    // that strip the charstring cipher record that they did.
    if !ok || v > 16 || v < -1 {
        return 4
    }
    return v
}

// t1_skip_to_begin steps over the `<n> dict dup begin` that introduces the
// /CharStrings dictionary. A font that writes it differently is not followed
// off a cliff: the lexer is put back where it was and the entries are read from
// there.
@(private = "file")
t1_skip_to_begin :: proc(lx: ^T1_Lexer) {
    save := lx.pos
    for _ in 0 ..< 32 {
        tok, ok := t1_next(lx)
        if !ok {
            break
        }
        if t1_eq(tok, "begin") {
            return
        }
        if len(tok) > 0 && tok[0] == '/' {
            break
        }
    }
    lx.pos = save
}

// t1_read_binary takes the `<len>` bytes that follow an RD token. Exactly one
// delimiter separates the two, and the byte after it is data even when it looks
// like whitespace.
//
// A length that runs past the end of the input is clamped rather than rejected:
// that is a truncated eexec block, and the partial charstring it yields draws
// what it can. The clamp always leaves the lexer at the end of the input, so
// the truncation is recorded by the token read that follows, in
// t1_read_charstrings.
@(private = "file")
t1_read_binary :: proc(lx: ^T1_Lexer, n: int) -> ([]byte, bool) {
    if lx.pos >= len(lx.src) || n < 0 {
        return nil, false
    }
    lx.pos += 1
    take := min(n, len(lx.src) - lx.pos)
    out := lx.src[lx.pos:lx.pos + take]
    lx.pos += take
    return out, true
}

// t1_read_subr reads one `dup <index> <len> RD <bytes> NP`. The RD token's name
// is whatever the font defined it as -- RD, -|, or anything else -- so it is
// read and discarded rather than matched.
@(private = "file")
t1_read_subr :: proc(ctx: ^T1_Parse_Ctx, lx: ^T1_Lexer) {
    save := lx.pos
    idx, idx_ok := t1_next_int(lx)
    n, n_ok := t1_next_int(lx)
    if !idx_ok || !n_ok || idx < 0 || n < 0 || idx >= MAX_T1_SUBRS {
        lx.pos = save
        return
    }
    if _, ok := t1_next(lx); !ok {
        lx.pos = save
        return
    }
    blob, blob_ok := t1_read_binary(lx, n)
    if !blob_ok {
        lx.pos = save
        return
    }
    if idx >= len(ctx.subrs) {
        resize(&ctx.subrs, idx + 1)
    }
    ctx.subrs[idx] = t1_decrypt_charstring(blob, ctx.f.len_iv)
}

// t1_read_charstrings reads `/<name> <len> RD <bytes> ND` until `end`.
@(private = "file")
t1_read_charstrings :: proc(ctx: ^T1_Parse_Ctx, lx: ^T1_Lexer, fuel: ^int) {
    for fuel^ > 0 {
        fuel^ -= 1
        tok, ok := t1_next(lx)
        if !ok {
            ctx.f.truncated = true
            return
        }
        if t1_eq(tok, "end") {
            return
        }
        if len(tok) < 2 || tok[0] != '/' {
            continue
        }

        save := lx.pos
        n, n_ok := t1_next_int(lx)
        if !n_ok || n < 0 {
            lx.pos = save
            continue
        }
        // Both of these mean the dictionary stopped in the middle of an
        // entry, which is what a truncated eexec block looks like from here.
        if _, rd_ok := t1_next(lx); !rd_ok {
            ctx.f.truncated = true
            return
        }
        blob, blob_ok := t1_read_binary(lx, n)
        if !blob_ok {
            ctx.f.truncated = true
            return
        }
        if len(ctx.names) >= MAX_T1_GLYPHS {
            return
        }
        append(&ctx.names, string(tok[1:]))
        append(&ctx.chars, t1_decrypt_charstring(blob, ctx.f.len_iv))
    }
}

// ---------------------------------------------------------------------------
// The charstring interpreter (Type 1 spec, chapter 6)
// ---------------------------------------------------------------------------

@(private = "file")
T1_Interp :: struct {
    f:   ^Type1_Font,
    out: ^Path,

    stack: [T1_STACK_MAX]f64,
    sp:    int,
    // Values an OtherSubr left behind for the `pop`s that follow it. The
    // format models this as the PostScript operand stack, which a charstring
    // can only reach through `pop`.
    res:   [T1_PS_MAX]f64,
    nres:  int,
    ires:  int,

    // Where the glyph is being drawn. `origin` is non-zero only for the accent
    // half of a seac, and translates the whole outline because every Type 1
    // path operator is relative.
    cur:    Point,
    origin: Point,

    sbx:        f64,
    width:      f64,
    have_width: bool,
    // A subpath has been begun and not yet closed.
    open: bool,

    in_flex: bool,
    // Where the pen stood when the FLEX began; nothing is emitted while its
    // points are collected, so this is still the subpath's live end.
    flex_start: Point,
    flex:       [T1_FLEX_POINTS]Point,
    nflex:      int,

    fuel: int,
    stop: bool,

    allow_seac: bool,
    seac:       bool,
    seac_args:  [5]f64,
    width_only: bool,
}

@(private = "file")
t1_run :: proc(f: ^Type1_Font, gid: int, out: ^Path, origin: Point, ip: ^T1_Interp) {
    if gid < 0 || gid >= len(f.chars) {
        return
    }
    ip.f = f
    ip.out = out
    ip.origin = origin
    ip.cur = origin
    ip.fuel = T1_MAX_FUEL
    t1_exec(ip, f.chars[gid], 0)
    // Glyph outlines are closed shapes, so whatever the charstring left open --
    // whether it ended at `endchar`, ran out of bytes, or hit a bound -- is
    // closed here.
    t1_close_subpath(ip)
}

@(private = "file")
t1_push :: proc(ip: ^T1_Interp, v: f64) {
    if ip.sp >= len(ip.stack) {
        // The format allows 24 operands. Past our margin the charstring is
        // corrupt, and carrying on would mis-align every operator after it.
        ip.stop = true
        return
    }
    ip.stack[ip.sp] = is_finite(v) ? v : 0
    ip.sp += 1
}

@(private = "file")
t1_args :: proc(ip: ^T1_Interp, n: int) -> ([]f64, bool) {
    if ip.sp < n {
        return nil, false
    }
    return ip.stack[ip.sp - n:ip.sp], true
}

@(private = "file")
t1_ensure_open :: proc(ip: ^T1_Interp) {
    if ip.open {
        return
    }
    if ip.out != nil {
        path_move_to(ip.out, ip.cur.x, ip.cur.y)
    }
    ip.open = true
}

// t1_close_subpath ends the current subpath. Type 1 `closepath` leaves the
// current point where it was, unlike the content stream's `h`, so `cur` is
// untouched here.
@(private = "file")
t1_close_subpath :: proc(ip: ^T1_Interp) {
    if !ip.open {
        return
    }
    if ip.out != nil {
        path_close(ip.out)
    }
    ip.open = false
}

@(private = "file")
t1_rmoveto :: proc(ip: ^T1_Interp, dx, dy: f64) {
    nx := ip.cur.x + dx
    ny := ip.cur.y + dy
    if !is_finite(nx) || !is_finite(ny) {
        ip.stop = true
        return
    }
    ip.cur = Point{nx, ny}

    if ip.in_flex {
        // Inside a FLEX the seven movetos are control points being collected,
        // not seven separate subpaths. A reader that draws them as movetos
        // loses the curve entirely and the glyph goes visibly lumpy.
        if ip.nflex < len(ip.flex) {
            ip.flex[ip.nflex] = ip.cur
        }
        ip.nflex += 1
        return
    }

    // Glyph outlines are always closed, and fonts that leave the `closepath`
    // out are common enough that closing here rather than trusting them is the
    // only way a stroked text render mode looks right.
    t1_close_subpath(ip)
    if ip.out != nil {
        path_move_to(ip.out, nx, ny)
    }
    ip.open = true
}

@(private = "file")
t1_rlineto :: proc(ip: ^T1_Interp, dx, dy: f64) {
    nx := ip.cur.x + dx
    ny := ip.cur.y + dy
    if !is_finite(nx) || !is_finite(ny) {
        ip.stop = true
        return
    }
    t1_ensure_open(ip)
    if ip.out != nil {
        path_line_to(ip.out, nx, ny)
    }
    ip.cur = Point{nx, ny}
}

@(private = "file")
t1_rrcurveto :: proc(ip: ^T1_Interp, dx1, dy1, dx2, dy2, dx3, dy3: f64) {
    x1 := ip.cur.x + dx1
    y1 := ip.cur.y + dy1
    x2 := x1 + dx2
    y2 := y1 + dy2
    x3 := x2 + dx3
    y3 := y2 + dy3
    t1_curve_abs(ip, Point{x1, y1}, Point{x2, y2}, Point{x3, y3})
}

@(private = "file")
t1_curve_abs :: proc(ip: ^T1_Interp, c1, c2, e: Point) {
    if !is_finite(c1.x) || !is_finite(c1.y) || !is_finite(c2.x) ||
       !is_finite(c2.y) || !is_finite(e.x) || !is_finite(e.y) {
        ip.stop = true
        return
    }
    t1_ensure_open(ip)
    if ip.out != nil {
        path_curve_to(ip.out, c1.x, c1.y, c2.x, c2.y, e.x, e.y)
    }
    ip.cur = e
}

@(private = "file")
t1_exec :: proc(ip: ^T1_Interp, code: []byte, depth: int) {
    if depth > T1_MAX_CALL_DEPTH {
        ip.stop = true
        return
    }

    i := 0
    for i < len(code) {
        if ip.stop {
            return
        }
        ip.fuel -= 1
        if ip.fuel <= 0 {
            ip.stop = true
            return
        }

        b := code[i]
        i += 1

        // 32..255 begin a number; 0..31 are the operators (Type 1 spec, 6.2).
        if b >= 32 {
            v: f64
            switch {
            case b <= 246:
                v = f64(int(b) - 139)
            case b <= 250:
                if i >= len(code) {
                    ip.stop = true
                    return
                }
                v = f64((int(b) - 247) * 256 + int(code[i]) + 108)
                i += 1
            case b <= 254:
                if i >= len(code) {
                    ip.stop = true
                    return
                }
                v = f64(-(int(b) - 251) * 256 - int(code[i]) - 108)
                i += 1
            case:
                if i + 4 > len(code) {
                    ip.stop = true
                    return
                }
                u := u32(code[i]) << 24 | u32(code[i + 1]) << 16 |
                     u32(code[i + 2]) << 8 | u32(code[i + 3])
                v = f64(i32(u))
                i += 4
            }
            t1_push(ip, v)
            continue
        }

        switch b {
        case 1, 3:
            // hstem, vstem: hints, which this reader does not use.
            ip.sp = 0

        case 4: // vmoveto
            if a, ok := t1_args(ip, 1); ok {
                t1_rmoveto(ip, 0, a[0])
            }
            ip.sp = 0

        case 5: // rlineto
            if a, ok := t1_args(ip, 2); ok {
                t1_rlineto(ip, a[0], a[1])
            }
            ip.sp = 0

        case 6: // hlineto
            if a, ok := t1_args(ip, 1); ok {
                t1_rlineto(ip, a[0], 0)
            }
            ip.sp = 0

        case 7: // vlineto
            if a, ok := t1_args(ip, 1); ok {
                t1_rlineto(ip, 0, a[0])
            }
            ip.sp = 0

        case 8: // rrcurveto
            if a, ok := t1_args(ip, 6); ok {
                t1_rrcurveto(ip, a[0], a[1], a[2], a[3], a[4], a[5])
            }
            ip.sp = 0

        case 9: // closepath
            t1_close_subpath(ip)
            ip.sp = 0

        case 10: // callsubr
            if ip.sp < 1 {
                ip.sp = 0
                break
            }
            ip.sp -= 1
            idx := int(ip.stack[ip.sp])
            if idx < 0 || idx >= len(ip.f.subrs) || len(ip.f.subrs[idx]) == 0 {
                break
            }
            t1_exec(ip, ip.f.subrs[idx], depth + 1)

        case 11: // return
            return

        case 13: // hsbw
            if a, ok := t1_args(ip, 2); ok {
                ip.sbx = a[0]
                ip.width = a[1]
                ip.have_width = true
                // The left sidebearing is where the outline starts, so setting
                // the current point here is what translates the whole glyph.
                ip.cur = Point{ip.origin.x + a[0], ip.origin.y}
            }
            ip.sp = 0
            if ip.width_only && ip.have_width {
                ip.stop = true
                return
            }

        case 14: // endchar
            // The open subpath is closed by t1_run rather than here, because a
            // charstring that simply runs out of bytes has to be closed too.
            ip.sp = 0
            ip.stop = true
            return

        case 21: // rmoveto
            if a, ok := t1_args(ip, 2); ok {
                t1_rmoveto(ip, a[0], a[1])
            }
            ip.sp = 0

        case 22: // hmoveto
            if a, ok := t1_args(ip, 1); ok {
                t1_rmoveto(ip, a[0], 0)
            }
            ip.sp = 0

        case 30: // vhcurveto
            if a, ok := t1_args(ip, 4); ok {
                t1_rrcurveto(ip, 0, a[0], a[1], a[2], a[3], 0)
            }
            ip.sp = 0

        case 31: // hvcurveto
            if a, ok := t1_args(ip, 4); ok {
                t1_rrcurveto(ip, a[0], 0, a[1], a[2], 0, a[3])
            }
            ip.sp = 0

        case 12:
            if i >= len(code) {
                ip.stop = true
                return
            }
            b2 := code[i]
            i += 1
            if t1_escape(ip, b2) {
                return
            }

        case:
            // An operator this reader does not know is a byte of corruption,
            // not a reason to abandon the glyph: clear the operands it would
            // have taken and carry on.
            ip.sp = 0
        }
    }
}

// t1_escape runs the two-byte operators. Returns true when the caller must
// stop executing this charstring.
@(private = "file")
t1_escape :: proc(ip: ^T1_Interp, op: byte) -> bool {
    switch op {
    case 0, 1, 2:
        // dotsection, vstem3, hstem3: hints.
        ip.sp = 0

    case 6: // seac
        if a, ok := t1_args(ip, 5); ok && ip.allow_seac {
            ip.seac = true
            for k in 0 ..< 5 {
                ip.seac_args[k] = a[k]
            }
        }
        ip.sp = 0
        ip.stop = true
        return true

    case 7: // sbw
        if a, ok := t1_args(ip, 4); ok {
            ip.sbx = a[0]
            ip.width = a[2]
            ip.have_width = true
            ip.cur = Point{ip.origin.x + a[0], ip.origin.y + a[1]}
        }
        ip.sp = 0
        if ip.width_only && ip.have_width {
            ip.stop = true
            return true
        }

    case 12: // div
        if ip.sp < 2 {
            ip.sp = 0
            break
        }
        den := ip.stack[ip.sp - 1]
        num := ip.stack[ip.sp - 2]
        ip.sp -= 2
        // A zero denominator gives an infinity or a NaN, which t1_push turns
        // into a zero: this is the only arithmetic in the interpreter, so it
        // is the only place a non-finite number can enter the operand stack.
        t1_push(ip, num / den)

    case 16: // callothersubr
        t1_othersubr(ip)

    case 17: // pop
        v := 0.0
        if ip.ires < ip.nres {
            v = ip.res[ip.ires]
            ip.ires += 1
        }
        t1_push(ip, v)

    case 33: // setcurrentpoint
        // The operands restate a current point the interpreter already holds:
        // the FLEX sequence ends `pop pop setcurrentpoint`, handing back the
        // two values OtherSubr 0 left. Honouring them could only move the pen
        // somewhere the outline never went, so they are dropped.
        ip.in_flex = false
        ip.sp = 0

    case:
        ip.sp = 0
    }
    return false
}

// t1_othersubr implements `callothersubr`, which is where the font program
// calls out to PostScript procedures the interpreter is expected to know.
//
// Skipping it is not an option. OtherSubrs 0-2 are FLEX, which every serif face
// uses on the shallow curves of stems and bowls; a reader that treats the seven
// collected movetos as movetos draws those curves as a stack of disconnected
// points. OtherSubr 3 is hint replacement, whose one argument has to come back
// out through the `pop` that follows so the `callsubr` after it reaches the
// right subroutine.
@(private = "file")
t1_othersubr :: proc(ip: ^T1_Interp) {
    if ip.sp < 2 {
        ip.sp = 0
        return
    }
    othersubr := int(ip.stack[ip.sp - 1])
    n := int(ip.stack[ip.sp - 2])
    ip.sp -= 2
    if n < 0 {
        n = 0
    }
    if n > ip.sp {
        n = ip.sp
    }
    args := ip.stack[ip.sp - n:ip.sp]
    ip.sp -= n

    ip.nres = 0
    ip.ires = 0

    switch othersubr {
    case 0:
        t1_end_flex(ip)
    case 1:
        ip.in_flex = true
        ip.nflex = 0
        ip.flex_start = ip.cur
    case 2:
        // Collect one FLEX point. The rmoveto that precedes it has already
        // recorded it, so there is nothing left to do.
    case 3:
        // Hint replacement: the subroutine number is handed straight back. A
        // reader that answered 3 instead -- the value the non-interpreting
        // fallback pushes -- would run the do-nothing subroutine every font
        // ships as Subrs[3] and lose the replacement.
        ip.nres = 1
        ip.res[0] = len(args) > 0 ? args[0] : 3
    case:
        // An OtherSubr this reader does not know: leave its arguments for the
        // pops that follow, which is what the PostScript procedure it stands in
        // for would have left on the stack.
        cnt := min(len(args), len(ip.res))
        for k in 0 ..< cnt {
            ip.res[k] = args[k]
        }
        ip.nres = cnt
    }
}

// t1_end_flex turns the seven collected points into the two curves they
// describe. The first is a reference point the hinting machinery uses and the
// outline ignores; the other six are two cubics' control and end points.
@(private = "file")
t1_end_flex :: proc(ip: ^T1_Interp) {
    if !ip.in_flex {
        return
    }
    ip.in_flex = false

    // Nothing was emitted while the points were collected, so the subpath is
    // still sitting where the FLEX began.
    if !ip.open {
        if ip.out != nil {
            path_move_to(ip.out, ip.flex_start.x, ip.flex_start.y)
        }
        ip.open = true
    }

    if ip.nflex >= T1_FLEX_POINTS {
        t1_curve_abs(ip, ip.flex[1], ip.flex[2], ip.flex[3])
        t1_curve_abs(ip, ip.flex[4], ip.flex[5], ip.flex[6])
    } else if ip.nflex > 0 {
        // A FLEX that collected the wrong number of points is malformed, but
        // the pen did move: joining the ends with a straight segment keeps the
        // outline in one piece instead of leaving a gap across the glyph.
        if ip.out != nil {
            path_line_to(ip.out, ip.cur.x, ip.cur.y)
        }
    }

    // The two values `pop pop setcurrentpoint` is about to consume.
    ip.res[0] = ip.cur.x
    ip.res[1] = ip.cur.y
    ip.nres = 2
    ip.ires = 0
}