fu/c/jsonfmt.c

static void fujson_fmt(pTHX_ fustr *, SV *);

static void fujson_fmt_str(pTHX_ fustr *out, const char *stri, size_t len, int utf8) {
    size_t off = 0, loff;
    const unsigned char *str = (const unsigned char *)stri;
    unsigned char *buf;
    unsigned char x = 0;

    /* Validate entire string for conformance if this is flagged as a utf8
     * string, this lets us be lazy further on.
     * Commenting this out doubles the performance for formatting unicode
     * strings, I suspect there's room for optimizations in
     * is_c9strict_utf8_string().  */
    if (utf8 && !is_c9strict_utf8_string(str, len)) {
        return; /* TODO: Throw error. */
    }

    fustr_write(out, "\"", 1);
    fustr_reserve(out, len);

    while (off < len) {
        /* Fast path: no escaping needed */
        loff = off;
        if (utf8) {
            /* we already validated everything >=0x80 */
            while (off < len) {
                x = str[off];
                if (x <= 0x1f || x == '"' || x == '\\' || x == 0x7f) break;
                off++;
            }
        } else {
            /* binary strings need special handling for >=0x80 */
            while (off < len) {
                x = str[off];
                if (x <= 0x1f || x == '"' || x == '\\' || x >= 0x7f) break;
                off++;
            }
        }
        fustr_write(out, (char *)str+loff, off-loff);

        if (off < len) { /* early break, which means current byte needs special processing */
            switch (x) {
                case '"':  fustr_write(out, "\\\"", 2); break;
                case '\\': fustr_write(out, "\\\\", 2); break;
                case 0x08: fustr_write(out, "\\b", 2); break;
                case 0x09: fustr_write(out, "\\t", 2); break;
                case 0x0a: fustr_write(out, "\\n", 2); break;
                case 0x0c: fustr_write(out, "\\f", 2); break;
                case 0x0d: fustr_write(out, "\\r", 2); break;
                default:
                    if (x < 0x80) {
                        buf = (unsigned char *)fustr_write_buf(out, 6);
                        memcpy(buf, "\\u00", 4);
                        buf[4] = PL_hexdigit[(x >> 4) & 0x0f];
                        buf[5] = PL_hexdigit[x & 0x0f];
                    } else { /* x >= 0x80, !utf8, so encode as 2-byte UTF-8 */
                        buf = (unsigned char *)fustr_write_buf(out, 2);
                        buf[0] = 0xc0 | (x >> 6);
                        buf[1] = 0x80 | (x & 0x3f);
                    }
            }
            off++;
        }
    }

    fustr_write(out, "\"", 1);
}

/* All digits between 0 and 100, a trick I borrowed from the Zig stdlib. */
static const char fujson_digits[] =
    "00010203040506070809"
    "10111213141516171819"
    "20212223242526272829"
    "30313233343536373839"
    "40414243444546474849"
    "50515253545556575859"
    "60616263646566676869"
    "70717273747576777879"
    "80818283848586878889"
    "90919293949596979899";

static void fujson_fmt_int(pTHX_ fustr *out, SV *val) {
    char buf[32];
    char *r = buf+31;
    int neg = 0;
    IV iv;
    UV uv;

    if (SvIsUV(val)) { /* Why is this macro not documented? */
        uv = SvUV_nomg(val);
    } else {
        iv = SvIV_nomg(val);
        neg = iv < 0;
        uv = neg ? -iv : iv;
    }

    if (uv == 0) {
        fustr_write(out, "0", 1);
        return;
    }

    while (uv >= 10) {
        r -= 2;
        memcpy(r, fujson_digits + ((uv % 100)<<1), 2);
        uv /= 100;
    }
    if (uv > 0) *(--r) = '0' + (uv % 10);
    if (neg) *(--r) = '-';
    uv = 31 - (r - buf);
    fustr_write(out, r, uv);
}

static void fujson_fmt_av(pTHX_ fustr *out, AV *av) {
    int i, len = av_count(av);
    fustr_write(out, "[", 1);
    for (i=0; i<len; i++) {
        if (i) fustr_write(out, ",", 1);
        SV **sv = av_fetch(av, i, 0);
        if (sv) fujson_fmt(aTHX_ out, *sv); /* sv will have magic if av is tied, but fujson_fmt() handles that. */
        else fustr_write(out, "null", 4);
    }
    fustr_write(out, "]", 1);
}

static void fujson_fmt_hv(pTHX_ fustr *out, HV *hv) {
    HE *he;
    STRLEN helen;
    char *hestr = NULL;

    hv_iterinit(hv);
    fustr_write(out, "{", 1);
    while ((he = hv_iternext(hv))) {
        if (hestr) fustr_write(out, ",", 1);
        hestr = HePV(he, helen);
        fujson_fmt_str(aTHX_ out, hestr, helen, HeUTF8(he));
        fustr_write(out, ":", 1);
        fujson_fmt(aTHX_ out, UNLIKELY(SvMAGICAL(hv)) ? hv_iterval(hv, he) : HeVAL(he));
    }
    fustr_write(out, "}", 1);
}


static void fujson_fmt(pTHX_ fustr *out, SV *val) {
    SvGETMAGIC(val);

    /* XXX: &PL_sv_yes and &PL_sv_no are proper booleans under 5.40, so no need
     * to explicitly check for those; does this work in 5.36 as well? */
    if (SvIsBOOL(val)) { /* Must check before IOKp & POKp, because bool implies both flags */
        if (BOOL_INTERNALS_sv_isbool_true(val)) fustr_write(out, "true", 4);
        else fustr_write(out, "false", 5);
    } else if (SvPOKp(val)) {
        fujson_fmt_str(aTHX_ out, SvPVX(val), SvCUR(val), SvUTF8(val));
    } else if (SvNOKp(val)) { /* Must check before IOKp, because integer conversion might have been lossy */
        /* TODO: quadmath? */
        NV nv = SvNV_nomg(val);
        if (isinfnan(nv)) return; /* TODO: error */
        fustr_reserve(out, NV_DIG+1);
        Gconvert(nv, NV_DIG, 0, out->cur);
        out->cur += strlen(out->cur);
    } else if (SvIOKp(val)) {
        fujson_fmt_int(aTHX_ out, val);
    } else if (SvROK(val)) {
        SV *rv = SvRV(val);
        SvGETMAGIC(rv);
        if (UNLIKELY(SvOBJECT(rv))) { /* TODO: Check for TO_JSON */ }
        else if (SvTYPE(rv) == SVt_PVHV) fujson_fmt_hv(aTHX_ out, (HV *)rv);
        else if (SvTYPE(rv) == SVt_PVAV) fujson_fmt_av(aTHX_ out, (AV *)rv);
        else return; /* TODO: error */
    } else if (!SvOK(val)) {
        fustr_write(out, "null", 4);
    } else {
        /* TODO: error */
    }
}

/* TODO: canonical */
/* TODO: pretty */
/* TODO: max depth? */