js_atod returns NaN for incomplete exponents instead of backtracking

[davesnx]

Description

According to the JavaScript specification, parseFloat("1e"), parseFloat("1e+"), and parseFloat("1e-") should all return 1 by backtracking when no valid digit follows the exponent marker. However, QuickJS returns NaN instead.

Expected Behavior (per ECMA-262 / TC39 Test262)

parseFloat("1e")   // should return 1
parseFloat("1e+")  // should return 1
parseFloat("1e-")  // should return 1
parseFloat("1E")   // should return 1
parseFloat("42e")  // should return 42
parseFloat("3.14e+") // should return 3.14

Current Behavior

All of the above return NaN.

Root Cause

In dtoa.c, the js_atod function has this code (around line 1534):

c = to_digit(*p);
if (c >= 10)
    goto fail; /* XXX: could stop before the exponent part */

The comment even acknowledges the issue. When there's no valid digit after the exponent marker (and optional sign), the parser should backtrack to before the e character and return the parsed number, not fail entirely.

Proposed Fix

Save the position before entering the exponent parsing block and restore it if no valid digit is found:

const char *p_before_exp = p; /* save position before exponent */
is_bin_exp = (*p == 'p' || *p == 'P');
p++;
exp_is_neg = 0;
if (*p == '+') {
    p++;
} else if (*p == '-') {
    exp_is_neg = true;
    p++;
}
c = to_digit(*p);
if (c >= 10) {
    /* No valid digit after exponent marker: backtrack to before 'e'.
       JavaScript spec: parseFloat("1e") === 1, parseFloat("1e+") === 1 */
    p = p_before_exp;
    goto after_exp;  /* skip exponent processing, continue with result */
}

Then add after_exp: label before the if (p == p_start) check.

References

• ECMA-262 Section 21.1.1.1 (ToNumber Applied to the String Type)
• TC39 Test262 test/built-ins/parseFloat/
• Found while running Test262 tests via quickjs.ml

[bnoordhuis]

With what version did you test? parseFloat already works that way but it's a fairly recent change, late September.

[davesnx]

Ah, I was testing against an older version (the og quickjs) in my project, and I thought the atod didn't change). Confirmed that recent quickjs-ng (binding to master) has this fixed. Thanks for the quick response! Closing this issue and the PR