How to avoid excessive GC

[wismill]

So I have the following situation in unicode-data:

• Bench function on chars.
• Lots of codepoints are unassigned or private use area, so we want to filter these out.
• We do not want the filtering be part of the benchmark, so we use env to create a list of chars to bench.
• This big list seems to trigger the GC anyway, giving every unstable results.

Lately I have experimented a hack that creates a pinned “list” using the foreign interface. This is ugly but seems to work, thanks to pinned memory being excluded from GC.

@Bodigrim What do you think? Is there a better way? Note that we would like to keep the dependencies as small as possible.

[Bodigrim]

@wismill maybe pack "interesting" Chars into a ByteArray instead of list?

[wismill]

You mean using e.g. newPinnedByteArray? But then I guess I can just reach for newPinnedByteArray# and further functions from GHC.Exts.

Anyway, you confirmed that pinned memory a good direction. Thanks for your insights. It may be worth noting it in the README.

[Bodigrim]

Not necessarily even pinned, copying continuous blocks of memory is extremely fast on modern processors. And any arrays over 3-4K are automatically pinned by RTS.