1798789 - Differing results for 10**23 between v8 and SpiderMonkey

Status: NEW

(Core :: JavaScript Engine, task, P3)

Description

[Steve Fink [:sfink] [:s:]]

Migrated from bug 1316557 comment 1:

I'm having an issue with Math.pow too.
Math.pow(10, 23) and 10 ** 23 returns 1.0000000000000001e+23
On Chrome it's 1e+23, so I'm using Number("1e" + 23) to get consistent results.

This seems like a potential interoperability/fingerprinting issue.

Comment 1

[André Bargull [:anba]]

We seem to do better in general (at least under Linux, not sure about the precision of std::pow under Windows/MacOS). The following test case shows more imprecise results for Chrome, cf. https://bugzilla.mozilla.org/show_bug.cgi?id=1775254#c4. As part of bug 1775254, I've also added "pow-approx-pow10.js" in https://phabricator.services.mozilla.com/D150791 to ensure the error is less than 2 ULP for Math.pow(10, i).

for (let i = 0; i <= Math.trunc(Math.log10(Number.MAX_VALUE)); ++i) {
  let n = 10 ** i;
  let k = Number("1e"+i);

  let abs = n => n < 0 ? -n : n;

  let e = 10n ** BigInt(i);
  let d1 = abs(e - BigInt(n));
  let d2 = abs(e - BigInt(k));

  // Check the results are consistent.
  if ((d1 < d2 && n !== k) || (d2 > d1 && n === k)) {
    console.log("inconsistent results");
    continue;
  }

  if (n !== k) {
    console.log(i, n, k);

    if (d1 > d2) {
      console.log("choose wrong");
    } else {
      // The delta between both points is equal.
      console.log("middle point");
    }

    let ulp = new BigUint64Array(new Float64Array([n]).buffer)[0];
    let down = new Float64Array(new BigUint64Array([ulp - 1n]).buffer)[0];
    let up = new Float64Array(new BigUint64Array([ulp + 1n]).buffer)[0];

    if (down === k) {
      console.log("down is better");
    } else if (up === k) {
      console.log("up is better");
    } else {
      // Something is wrong if we end up here.
      console.log("???");
    }
  }
}

Fingerprinting is also discussed in bug 1775254, which also links to bug 1722181. Bug 1775254 also introduced a regression due to fingerprinting -> bug 1784111.