Re: [Haskell-cafe] ANN: ieee version 0.7

On Tuesday 21 September 2010 07:11:24, Conrad Parker wrote:

On 21 September 2010 12:18, John Millikin wrote:

...

On Mon, Sep 20, 2010 at 03:22, Daniel Fischer wrote:

...

unsafeCoerce is not supposed to work for casts between Integral and Floating types. If you try to unsafeCoerce# between unboxed types, say Double# and Word64#, you're likely to get a compile failure (ghc panic). If you unsafeCoerce between the boxed types, it will probably work, but there are no guarantees.

There's a feature request for unboxed coercion (i.e. reinterpretation of the bit-pattern):

http://hackage.haskell.org/trac/ghc/ticket/4092

Interesting -- in that bug report, Simon Mar says that converting the value using pointers will work correctly. I've changed d-b-ieee754 over to use this method (v 0.4.2); the tests are still passing, so I'll call it success.

And I'd expect it to be a heck of a lot faster than the previous implementation. Have you done any benchmarks?

I've been using unsafeCoerce:

getFloat64be :: Get Double getFloat64be = do n <- getWord64be return (unsafeCoerce n :: Double)

putFloat64be :: Double -> Put putFloat64be n = putWord64be (unsafeCoerce n :: Word64)

but only tested it with quickcheck -- it passes about 10^7 checks, comparing roundtrips in combinatrion with the previous data-binary-ieee754 versions. However could that sometimes behave incorrectly?

Should the d-b-iee754-0.4.2 versions with castPtr etc. be even faster?

No, Simon says "Not terribly efficient, but better than using the FFI." I would expect unsafeCoerce to be the fastest you can get (so far). From the docs of unsafeCoerce#, I get the impression that will probably work, but it's not listed among the cases where it's *supposed* to work, hence it's probably in fact a little unsafe. One problem I see with both, unsafeCoerce and poke/peek is endianness. Will the bit-pattern of a double be interpreted as the same uint64_t on little-endian and on big-endian machines? In other words, is the byte order for doubles endianness-dependent too? If yes, that's fine, if no, it would break between machines of different endianness.

Conrad.

Cheers, Daniel