[GHC] #13426: compile-time memory-usage regression for DynFlags between GHC 8.0 and GHC 8.2

[GHC]

#13426: compile-time memory-usage regression for DynFlags between GHC 8.0 and GHC
8.2
-------------------------------------+-------------------------------------
        Reporter:  hvr               |                Owner:  rwbarton
            Type:  bug               |               Status:  new
        Priority:  high              |            Milestone:  8.2.1
       Component:  Compiler          |              Version:  8.1
      Resolution:                    |             Keywords:
Operating System:  Unknown/Multiple  |         Architecture:
 Type of failure:  Compile-time      |  Unknown/Multiple
  performance bug                    |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):  Phab:D3399,
       Wiki Page:                    |  Phab:D3400
-------------------------------------+-------------------------------------
Changes (by rwbarton):

 * differential:   => Phab:D3399, Phab:D3400


Comment:

 Replying to [comment:14 simonpj]:
 > PS: you may then be able to get rid of the call to `coreBindsSize` in
 `SimplCore`.  If so, that would be good: faster compilation!  Worth
 trying... it might I suppose show up another leak currently patched by
 `coreBindsSize`.

 This line of investigation already paid some dividends. It turns out that
 that `coreBindsSize` is what takes care of forcing the demand and
 strictness info produced by the demand analysis pass. Otherwise, the
 demand analysis pass leaks basically a copy of the whole Core program.

 But ''late'' demand analysis runs after all simplifier passes, and there
 is no more `coreBindsSize` or `seqBinds` after that. Presumably code
 generation uses some of the information added by this pass, but it doesn't
 force all of it; and the result is that we hang on to two copies of the
 Core program during code generation. And code generation is the phase that
 uses the most memory anyways, so this makes a big difference in peak
 memory usage.

 As a crude measure I added a `seqBinds` to demand analysis (Phab:D3400)
 and it reduced peak memory usage on DynFlags by around 30% and even made
 the running time a few percent faster (I think because of reduced GC
 work).

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/13426#comment:17>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler