[commit: packages/binary] master: Get: Avoid needless copies of input (7532daa)

[git at git.haskell.org]

Repository : ssh://git@git.haskell.org/binary

On branch  : master
Link       : http://git.haskell.org/packages/binary.git/commitdiff/7532daa8789e5199109bb1fcde367d71effb07e2

>---------------------------------------------------------------

commit 7532daa8789e5199109bb1fcde367d71effb07e2
Author: Ben Gamari <ben at smart-cactus.org>
Date:   Sun May 15 23:55:43 2016 +0200

    Get: Avoid needless copies of input
    
    My `b-tree` library seems to tickle a rather pathological behavior in
    `binary`'s decoding logic, where `binary` will create many needless
    copies of the input buffer by evaluating things of the form `B.concat
    [B.empty, leftovers]`, where `leftovers` is large.
    
    This resulted in runtimes of over two minutes when parsing a 50 MByte
    file. With this fix run drops to less than 100 milliseconds.


>---------------------------------------------------------------

7532daa8789e5199109bb1fcde367d71effb07e2
 src/Data/Binary/Get/Internal.hs | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/src/Data/Binary/Get/Internal.hs b/src/Data/Binary/Get/Internal.hs
index cf2d012..b9a0818 100644
--- a/src/Data/Binary/Get/Internal.hs
+++ b/src/Data/Binary/Get/Internal.hs
@@ -404,7 +404,11 @@ ensureN !n0 = C $ \inp ks -> do
     enoughChunks n str
       | B.length str >= n = Right (str,B.empty)
       | otherwise = Left (n - B.length str)
-    onSucc = B.concat
+    -- Sometimes we will produce leftovers lists of the form [B.empty, nonempty]
+    -- where `nonempty` is a non-empty ByteString. In this case we can avoid a copy
+    -- by simply dropping the empty prefix. In principle ByteString might want
+    -- to gain this optimization as well
+    onSucc = B.concat . dropWhile B.null
     onFail bss = C $ \_ _ -> Fail (B.concat bss) "not enough bytes"
 {-# INLINE ensureN #-}