From: | David Christensen <david(dot)christensen(at)crunchydata(dot)com> |
---|---|
To: | John Naylor <john(dot)naylor(at)enterprisedb(dot)com> |
Cc: | PostgreSQL-development <pgsql-hackers(at)postgresql(dot)org>, Andres Freund <andres(at)anarazel(dot)de>, Tomas Vondra <tomas(dot)vondra(at)enterprisedb(dot)com>, Stephen Frost <sfrost(at)snowman(dot)net> |
Subject: | Re: Initdb-time block size specification |
Date: | 2023-08-31 16:01:02 |
Message-ID: | CAOxo6X+OVBedhNMo_9TMWidj94zoBHOaV-BZYVf7jvAMg+3c1w@mail.gmail.com |
Views: | Raw Message | Whole Thread | Download mbox | Resend email |
Thread: | |
Lists: | pgsql-hackers |
> + * pg_fastmod - calculates the modulus of a 32-bit number against a constant
> + * divisor without using the division operator
> + */
> +static inline uint32 pg_fastmod(uint32 n, uint32 divisor, uint64 fastinv)
> +{
> +#ifdef HAVE_INT128
> + uint64_t lowbits = fastinv * n;
> + return ((uint128)lowbits * divisor) >> 64;
> +#else
> + return n % divisor;
> +#endif
> +}
>
> Requiring 128-bit arithmetic to avoid serious regression is a non-starter as written. Software that relies on fast 128-bit multiplication has to do backflips to get that working on multiple platforms. But I'm not sure it's necessary -- if the max block number is UINT32_MAX and max block size is UINT16_MAX, can't we just use 64-bit multiplication?
I was definitely hand-waving additional implementation here for
non-native 128 bit support; the modulus algorithm as presented
requires 4 times the space as the divisor, so a uint16 implementation
should work for all 64-bit machines. Certainly open to other ideas or
implementations, this was the one I was able to find initially. If
the 16bit approach is all that is needed in practice we can also see
about narrowing the domain and not worry about making this a
general-purpose function.
Best,
David
From | Date | Subject | |
---|---|---|---|
Next Message | David Christensen | 2023-08-31 16:13:18 | Re: Initdb-time block size specification |
Previous Message | John Naylor | 2023-08-31 15:54:02 | Re: Initdb-time block size specification |