deepseq-magic-1.0.0.3: Control/DeepSeq/Magic.hs
{-# LANGUAGE BangPatterns, MagicHash, UnboxedTuples #-}
module Control.DeepSeq.Magic where
-----------------------------------------------------------------------------
-- |
-- Module : Control.DeepSeq.Magic
-- Copyright : Public Domain
-- License : Public Domain
--
-- Maintainer : ezyang@cs.stanford.edu
-- Stability : stable
-- Portability : portable
--
-- The functions in this module are exactly the same as
-- 'Control.DeepSeq', but without the need for an 'NFData' instance.
--
-- The semantics are not terribly well defined, since they are based on
-- what 'unpackClosure#' returns in its pointer list, but it appears
-- that we NEVER recurse into AP or PAP objects.
import GHC.Exts
-- NOTE: This is a bit slower than a normal NFData instance
-- because unpackClosure# has to allocate arrays to report
-- the results. I tried writing a hand-rolled loop but it
-- was tricky. If this is a bottleneck I'll try again.
--
-- NOTE: Thankfully, unpackClosure# doesn't look through
-- thunks/function closures, so this behavior is actually
-- somewhat reasonable.
rnf :: a -> ()
rnf a = a `seq` rest `seq` () where
rest = case unpackClosure# a of
(# _, ptrs#, _ #)->
let s = I# (sizeofArray# ptrs#)
-- Despite the mucking about with unboxed integers,
-- use normal equality because the API changed
-- backwards incompatibly in 7.10. The optimizer
-- can figure out how to get the unboxed version.
go n@(I# n#) | n == s = ()
| otherwise =
case indexArray# ptrs# n# of
(# a #) -> rnf a `seq` go (n + 1)
in go 0
infixr 0 $!!
deepseq :: a -> b -> b
deepseq a b = rnf a `seq` b
($!!) :: (a -> b) -> a -> b
f $!! x = x `deepseq` f x
force :: a -> a
force x = x `deepseq` x