[GHC] #8256: adding locality levels to prefetch# and friends

8 Sep 2013

      #8256: adding locality levels to prefetch# and friends
------------------------------------+-------------------------------------
       Reporter:  carter            |             Owner:  carter
           Type:  task              |            Status:  new
       Priority:  normal            |         Milestone:  7.8.1
      Component:  Compiler          |           Version:  7.7
       Keywords:                    |  Operating System:  Unknown/Multiple
   Architecture:  Unknown/Multiple  |   Type of failure:  None/Unknown
     Difficulty:  Unknown           |         Test Case:
     Blocked By:                    |          Blocking:
Related Tickets:  #8252 ,#8107      |
------------------------------------+-------------------------------------
 currently in HEAD / 7.7, the prefetch primop only does the equivalent of
 __builtin_prefetch(ptr,0,3)
 (0 here denoting a read prefetch, 3 denoting "high locality, keep this in
 cache for a while")

 On modern hardware, we can do better!  For many natural use case of
 prefetch in  haskell, we have a streaming workload, so we want a prefetch
 to also hint that "once we've worked on a piece of memory, no need to keep
 it around for any period of time, don't pollute our memory with it"

 the attached patch takes each prefetchBlah# operation (where
 blah=ByteArray,MutableByteArray, or Addr) and replaces it with
 prefetchBlah0# through prefetchBlah3#,  and passes the integer information
 through into the code generators.

 To make the engineering reasonable, I enriched MO_Prefetch_Data with an
 Int parameter (which must be a value between ranging 0-3). Theres probably
 a better way to model the locality paramter, but that maps directly to how
 its used in llvm code gen.

 This patch does not include a test case yet. (interestingly, currently the
 test suite doesn't have any tests for the prefetch primops as yet!). So
 that needs to be added.

 Also, theres no good reason for the prefetch ops to be LLVM only.

 worst case we could just treat them as noop's and drop them. But at least
 for x86, should be easy to add the support (though if anyone wants to add
 support for the ppc and sparc stuff, or help me do that, that'd be awesome
 too!) . I"ll look into doing that in a few days.

 theres probably some other things i'm overlooking.

 anyways, i'll attach a preliminary patch for feedback now

-- 
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/8256
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler

tags

participants (1)