Re: Frequency of integer division in source code?

meissner@osf.org writes:
>[re modulus for hash tables]


tfj@apusapus.demon.co.uk (Trevor Jenkins) writes:
|> Better yet is not to be mis-lead by Maurer's paper and to ignore
|> his advice by using a table whose size is a power of two. Then the
|> integer divide becomes an AND operation which is faster that
|> division or multiplication on almost all architectures.


Personal experience: my hash-table impementation uses power-of-two table
sizes. It doubles the table size when it gets too full. I get good
spread and density: for identitifier + keyword strings from a large
collection of C programs I had 85% of my bins holding 1 key, and no more
than 3 keys in any bin.


I would be more worried about code like:


    int offset;
    struct foo *a, *b;
    ...
    offset = a - b;


where sizeof(struct foo) is not a power of two.
--
cliffc@cs.rice.edu -- Massively Scalar Compiler Group, Rice University
--

Return to the comp.compilers page.
Search the comp.compilers archives again.