Code breaks with '-O', any ideas.

[Donald Bruce Stewart]

haskell2:
> David Brown wrote:
> > Ian Lynagh wrote:
> >> On Mon, Mar 05, 2007 at 10:51:46AM -0800, David Brown wrote:
> >>> My program 'harchive' <http://www.davidb.org/darcs/harchive/> breaks
> >>> if compiled with '-O'.  Just wondering if anyone has suggestions on
> >>> how to figure out what might be causing this.
> >> What do you mean by "breaks"?
> >
> > I'm now rebuilding individual parts with -O to see if I can isolate
> > the problem(s).
> 
> Well, that helped.  I found the problem.  I had something like:
> 
>    import qualified Data.ByteString as B
>  
>    let hash = B.pack $ replicate (#const SHA_DIGEST_LENGTH) 0
>    ...
>    B.useAsCStringLen hash $ \ (hdata, _) -> do
>      c_sha1Final hash ctx
>    return hash
> 
> But this seems to be causing the hash to be garbage, and possibly have
> c_sha1Final write a 20-byte hash into some garbled piece of memory.
> 
> Ireplaced this with
> 
>    let hashLen = (#const SHA_DIGEST_LENGTH)
>    hashData <- mallocForeignPtrBytes hashLen
>    withForeignPtr hashData $ \hashP -> do
>       c_sha1Final hashP ctx
>       B.copyCStringLen (hashP, hashLen)
> 
> and the code works both with and without optimization.
> 
> Dave

Ah, this is my fault! The old api used 'useAsCSTringLen' unsafely. Its a
zero-copying version. The default is now copyCStringLen. See this thread
for the issues involved:

    http://www.haskell.org/pipermail/libraries/2007-January/006783.html

and for how it has been resolved:

    http://www.haskell.org/pipermail/libraries/2007-January/006801.html

So yes, using 'copyCStringlen' is the appropriate thing to do, if you
don't want to guarantee the C side of things.

-- Don