Why write out-of-line PrimOps in Cmm?

[Simon Marlow]

On 14/02/13 11:55, Alexander Kjeldaas wrote:
>
> I was looking at the eventlog code, and I wanted to move processing of a
> full eventlog buffer into Haskell, instead of the now fixed behavior of
> writing the data to a file.
>
> To do this, I though that having a haskell process blocked on an MVar,
> or Chan would be nice, and then some way to signal the process.
>
> But the low-level PrimOps for MVars, takeMVar, tryTakeMVar etc are in
> PrimOps.cmm and I don't know how to call them from the RTS.

You can't call them directly.  The problem is that to call Haskell code 
you need a Haskell thread to run it in.

There are a couple of ways that we call Haskell code from the RTS:

  - rts_lock()/rts_evalIO()/rts_unlock().  This is what calling a
    foreign export does, and it's quite heavyweight.

  - create a new Haskell thread and add it to the run queue, like we do
    for finalizers.

> This lead me to question what the point of out-of-line PrimOps in
> PrimOps.cmm is.  I don't think the commentary covers this.
>
> http://hackage.haskell.org/trac/ghc/wiki/Commentary/PrimOps
>
> So why aren't all the stuff in PrimOps.cmm just "ccall" wrappers around
> C implementations?  Wouldn't that in general be more flexible for the RTS?

They need to do things like block, which you can't do from a ccall. 
There's also the overhead - a primop can allocate memory directly from 
the nursery, whereas a ccall would have to call allocate() (and couldn't 
GC).

Cheers,
	Simon