RFR: 8334755: Asymptotically faster implementation of square root algorithm [v30]

[Raffaello Giulietti]

On Thu, 18 Jul 2024 15:09:41 GMT, fabioromano1 <duke at openjdk.org> wrote:

>>> I mean only restricted to unsigned `long` perfect squares, something like the following, but written as a proper test
>>> 
>>> ```
>>>         long i = 0;
>>>         for (; i < 1L << 32; ++i) {
>>>             long x = i * i;
>>>             long s = (long) Math.sqrt(x >= 0 ? x : x + 0x1p64);
>>>             if (!(s + 1 == i || s == i)) {
>>>                 System.out.format("oops... i=%d, but s=%d%n", i, s);
>>>                 System.exit(1);
>>>             }
>>>         }
>>> ```
>> 
>> So, it is okay although the code does not test directly BigInteger.sqrt()?
>
> I found the way: I can test directly the code through `java.math.Accessor.java`

I think there's a misunderstanding here.

What I'd like to see is a test that verifies the claim

        /* For every long value s in [0, 2^32) such that x == s * s,
         * it is true that s - 1 <= (long) Math.sqrt(x >= 0 ? x : x + 0x1p64) <= s,

A test similar to my code above should be more than enough.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/19710#discussion_r1683045737