[Haskell-cafe] transformers appears to benefit from more inline

[Oliver Charles]

Hi everyone,

I've recently been playing with a little mtl-like approach to streaming
data, whereby I have a single type class - MonadYield - which can yield
data, and then I use various implementations of this type class to
implement operations. My usual approach to this is to implement this by
newtyping around appropriate "off-the-shelf" monad transformers from the
transformers library, but I found that this incurs a significant
performance penalty.

I've tried to put some fairly extensive benchmarks in place, which you can
find at https://github.com/ocharles/monad-yield. In that repository is a
README.md file that describes how I have been performing these benchmarks.
The benchmarks are defined over a common interface that each implementation
of MonadYield exports. The benchmarks are defined in "Benchmarks.hs", and
the three implementations are "Transformers.hs" (using transformers from
GHC), "TransformersInline.hs" (using transformers-ocharles from that
repository, which has many more INLINE pragmas) and "Inline.hs" (which
doesn't depend on anything other than base).

There are three main benchmarks that are ran - one is benchmarking
essentially the cost of ReaderT, the next the cost of StateT, and the last
a composition of ReaderT over StateT over ReaderT. The results of the
benchmark can be found here: https://ocharles.github.io/monad-yield/.

It seems that the current darcs release of transformers loses every time,
but if I sprinkle {-# INLINE #-} across the definition of lazy state, I get
identical performance to just writing out the lazy state monad by hand.

I was very surprised to see that I have to pay when I use transformers, and
it seems like this cost can be removed at the cost of slightly larger
interface files.

Before I submit a patch, I'd love to hear others thoughts. Should {-#
INLINE #-} be necessary? Is there any reason not to add it to every symbol
in transformers?

-- ocharles
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/haskell-cafe/attachments/20160124/857413a4/attachment.html>