News for hackers. Stuff that matters.
Re:Smoking fast Haskell code using GHC's new LLVM codegen by jmillikin on Monday February 22, 2010 @12:55AM (#1142047)

  Isn't that a bit like saying that it "can outperform equivalent assembly"?
No; as shown in the link, this work allows idiomatic Haskell to outperform idiomatic C. Obviously it's possible to write C which is a direct translation of the generated assembly, but such code would be nearly unmaintainable, and certainly very difficult to reason about.

  And why can you not prove the equivalent C correct?
Proving the correctness of C code is possible, but (in my experience) significantly more difficult than for Haskell.

  After all, you're doing that proof by hand either way. A Haskell compiler will only help you prove that your program is well-typed. If you're mutating buffers then you're in a monad like IO or ST and the the compiler will throw it's hands up and tell you that correctness is entirely your problem.
I don't mean that the compiler will automatically generate a proof for me, but that authoring a proof will be possible in practice -- as opposed to C, where it's possible in theory but nobody actually can.

I'm a bit confused about why you think operating within a monad causes problems for proving correctness; to my knowledge, verifying code which uses monads is no more difficult than code using any other data types.

Re:Smoking fast Haskell code using GHC's new LLVM codegen by jrockway on Monday February 22, 2010 @01:43PM (#1143382)
as shown in the link, this work allows idiomatic Haskell to outperform idiomatic C

Interesting that you mention this. I know when I am writing C, I usually "malloc" dynamic data structures, whereas in Haskell, I usually "alloca" dynamic data structures (when talking to C libraries, obviously). Same end result, but alloca is actually faster, which makes allocation-heavy libraries faster when used from Haskell. (I guess I should use alloca in C, too. But I've never seen this anywhere.)

I noticed that this was especially helpful in a library I recently used that never allocated any of its own memory; it was always the responsibility of the caller. Haskell idioms worked much better than C idioms here.

I'm a bit confused about why you think operating within a monad causes problems for proving correctness; to my knowledge, verifying code which uses monads is no more difficult than code using any other data types.

Wait, are you saying that calling a function called "(>>=)" is no different than calling any other function? Shocking. :)