Re: [GHC] #8326: Place heap checks common in case alternatives before the case

5 Aug 2015

      #8326: Place heap checks common in case alternatives before the case
-------------------------------------+-------------------------------------
        Reporter:  jstolarek         |                   Owner:
            Type:  task              |                  Status:  new
        Priority:  normal            |               Milestone:
       Component:  Compiler          |                 Version:  7.7
      Resolution:                    |                Keywords:
Operating System:  Unknown/Multiple  |            Architecture:
 Type of failure:  Runtime           |  Unknown/Multiple
  performance bug                    |               Test Case:
      Blocked By:                    |                Blocking:  8317
 Related Tickets:  #1498             |  Differential Revisions:  Phab:D343
-------------------------------------+-------------------------------------

Comment (by bgamari):

 I think the issue here is that while on paper the code we emit is
 independent of our `GcPlan`, in practice we emit the code in most cases in
 `heapCheck` and **how** we end up in `heapCheck` is very much dependent on
 the `GcPlan`.

 I believe this almost does what we want,
 {{{#!hs
 maybeAltHeapCheck :: (GcPlan,ReturnKind) -> FCode a -> FCode a
 maybeAltHeapCheck gc_plan code = do
     codeOnly $ case gc_plan of
       (NoGcInAlts,_)                       -> code
       (GcInAlts regs, AssignedDirectly)    -> altHeapCheck regs code
       (GcInAlts regs, ReturnedTo lret off) -> altHeapCheckReturnsTo regs
 regs lret off code
 }}}
 The trouble is you have now thrown away the heap usage information from
 the code you emitted in the `NoGcInAlts` case.

 Alternatively, you can lift the code emission out of `heapCheck`, but this
 breaks its nice interface,
 {{{#!hs
 maybeAltHeapCheck :: (GcPlan,ReturnKind) -> FCode a -> FCode a
 maybeAltHeapCheck gc_plan code = do
     codeOnly $ case gc_plan of
       (NoGcInAlts,_)                       -> return ()
       -- These now only compute the heap usage of 'code' and do not emit
 it
       (GcInAlts regs, AssignedDirectly)    -> altHeapCheck regs code
       (GcInAlts regs, ReturnedTo lret off) -> altHeapCheckReturnsTo regs
 regs lret off code
     code
 }}}

 Anyways, let's see how Reid's approach performs; perhaps this is all
 irrelevant.

--
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/8326#comment:32
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler