[GHC] #16064: Improving Placement of Heap Checks - Avoiding Slowdowns in Hot Code

[GHC]

#16064: Improving Placement of Heap Checks - Avoiding Slowdowns in Hot Code
-------------------------------------+-------------------------------------
           Reporter:  _recursion     |             Owner:  (none)
               Type:  task           |            Status:  new
           Priority:  normal         |         Milestone:  8.9
          Component:  Compiler       |           Version:
           Keywords:  CodeGen,       |  Operating System:  Unknown/Multiple
  Performance                        |
       Architecture:                 |   Type of failure:  Runtime
  Unknown/Multiple                   |  performance bug
          Test Case:                 |        Blocked By:
           Blocking:                 |   Related Tickets:  #16040, #14791,
                                     |  #8326, #12231, #1498
Differential Rev(s):                 |         Wiki Page:
-------------------------------------+-------------------------------------
 When generating both heap and stack checks in Cmm, GHC can sometimes be
 fairly obtuse about their placement. As a result, they can be generated
 for paths in hot code that actually do not perform any allocation.
 Examples can be found in the related tickets, particularly #14791 and
 #16040.

 This ticket exists to track the work
 [https://ghc.haskell.org/trac/ghc/ticket/16040#comment:5 proposed] by
 Simon to detect such code patterns. They are described as follows:

 1. A primitive case with no allocation upstream from it.
 2. One alternative that performs no allocation.

 Under such circumstances the stack/heap checks can be moved from being
 performed on every iteration to being performed just prior to the
 allocation itself.

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/16064>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler