RFC: are duplications / float outs acceptable for prefetches or not?

[Carter Schonwald]

just to clarify:

the issue lies with the pure operations.
 Strictly speaking, the one on mutable byte arrays doesn't even need the
effects annotation, because the ST s / IO ordering will be enough to make
sure it will always be run when it should.  the pure ones, would almost
always need to be in some sort of monadic wrapper to make sure they're
sequences ahead of time at the right time, i think

the concern lies with prefetches being for performance:  they only make
sense as part of a thoughtful approach to leveraging the memory bandwidth
of a system, and GHC does make use of system memory buses pretty
aggressively as is!


to be clear: I'd very much like to be able to write code like

prefetchPtrM :: Monad m =>  Ptr a -> m ()
prefetchPtrM (Ptr adr) = do (return $ prefetchAddrOp addr) ; return ()


this is reasonably motivated use case: the Vector library has monadic
operations on its Pure vectors to enforce evaluation order.

naively, it seems like the right approach requires at least the "can fail"
annotation, and perhaps maybe the "has side effects annotation", so that
the above prefetchM operation is definable.

or maybe im not understanding something, and this is simpler than I fear?

thanks!






On Mon, Sep 16, 2013 at 1:35 AM, Carter Schonwald <
carter.schonwald at gmail.com> wrote:

> just for reference, the  current ops are
>
>
> primop PrefetchByteArrayOp2 "prefetchByteArray2#" GenPrimOp
>    ByteArray# -> Int# -> ByteArray#
>    with llvm_only = True
>
> primop PrefetchMutableByteArrayOp2 "prefetchMutableByteArray2#" GenPrimOp
>    MutableByteArray# s -> Int# -> State# s -> State# s
>    with has_side_effects = True
>         llvm_only = True
>
> primop PrefetchAddrOp2 "prefetchAddr2#" GenPrimOp
>    Addr# -> Int# -> Addr#
>    with llvm_only = True
>
> one of the things my patch (for 7.8) will do is add preliminary native
> code gen support for prefetch, admittedly in the form of Nops. (later work
> will fix that to be a bit more actionable)
>
>
> any opinions that help us suss this out would be appreciated
>
>
>
>
> On Mon, Sep 16, 2013 at 1:08 AM, Carter Schonwald <
> carter.schonwald at gmail.com> wrote:
>
>> Hey all,
>> I"ve been doing some work to finish up my prefetch patch that  I hope to
>> land in 7.8, and one design matter thats come up is what the optimization
>> semantics should be!
>>
>> is it appropriate to duplicate prefetches? (probably yes)
>> is it appropriate to float OUT prefetched? (probably not?)
>>
>> operations that are dupable which shouldn't float out are marked with the
>> can_fail= true
>> property,
>> operations which cannot be safely floated out and can't be duped are
>> marked as having side effects.
>>
>> on the GHC irc channel, Duncan was patient enough to walk my through
>> understanding this, and this makes me lean towards the conclusion that
>> duplication is ok, but floating out isnt. Thus all the prefetch ops (pure
>> or not) should have the attribute "can fail", rather than no attribut or
>> "has effects".
>>
>> this is exploiting that "can fail"  really just means "don't float out".
>>
>> thoughts and counter points?
>>
>> thanks
>>
>> -Carter
>>
>>
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.haskell.org/pipermail/ghc-devs/attachments/20130916/9a38b536/attachment.htm>