Issues resulting from inlining build

[Simon Peyton Jones]

Think of this like worker/wrapper. Given ‘crazy’, with some large body, you want to turn it into

crazy x = build (\cn. craxy’ c n x)
{-# INLINE crazy #-}

Now you can inline crazy everywhere, which will allow a consuming foldr to cancel with the build.  And without duplicating all the code.

But if there *is* no foldr, you’ll end up with calls like (crazy’  (:) []) , which will be considerably less efficient than the original function.  So you really want *both* crazy’ and crazy’ specialised to (:) and [].

GHC doesn’t do this automatically, because it really works well if the body of crazy is a good producer.  But you could imagine automating it.

Simon

From: Libraries [mailto:libraries-bounces at haskell.org] On Behalf Of David Feuer
Sent: 27 July 2014 20:29
To: Haskell Libraries
Subject: Issues resulting from inlining build


I think I finally figured out the deeper problem behind an issue I was having getting some things to fuse fully. I'm hoping someone has a brilliant idea for getting around it. Of course, it may be that someone else has thoroughly studied the matter and determined that there's no good solution. Suppose we write

crazy :: T -> [U]
crazy x = some $ long $ fusion $ pipeline

oops f = map f $ crazy x
uhOh c n = foldr c n $ crazy Pig
ohMy = take bignum $ crazy amigo

When GHC compiles crazy, it rewrites the pieces of the pipeline to build/foldr forms, fuses them, and produces

crazy x = build someBigFunction

In then inlines build, producing

crazy x = anotherBiggy

So far, this looks reasonably sensible, but it's likely bad. The problem is that GHC will (rightly) conclude that `build someBigFunction` is too big to inline, and therefore the fusion will break at that boundary and we'll produce intermediate lists in the functions that use crazy. Now if we were playing the part of the compiler by hand, we would likely factor out someBigFunction and then *refrain from inlining build*. That is, we would get

{-# NOINLINE crazyB #-}
crazyB x = someBigFunction

crazy = nonInliningBuild crazyB

Since we've factored out someBigFunction into crazyB, we can now safely inline crazy itself, allowing the pipeline to continue beyond it. The problem, of course, is that when we *don't* fuse beyond, there is some performance penalty (I have not tried to measure it yet) to passing in (:) and [] at runtime instead of fixing them at compile time.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.haskell.org/pipermail/libraries/attachments/20140728/ebb38582/attachment.html>