[GHC] #15124: Improve block layout for the NCG

[GHC]

#15124: Improve block layout for the NCG
-------------------------------------+-------------------------------------
        Reporter:  AndreasK          |                Owner:  AndreasK
            Type:  task              |               Status:  new
        Priority:  normal            |            Milestone:  8.6.1
       Component:  Compiler (NCG)    |              Version:  8.2.2
      Resolution:                    |             Keywords:  CodeGen
Operating System:  Unknown/Multiple  |         Architecture:
                                     |  Unknown/Multiple
 Type of failure:  None/Unknown      |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:  #8082 #14672      |  Differential Rev(s):  Phab:D4726
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by AndreasK):

 Replying to [comment:5 AndreasK]:
 > I've wrote up a patch implementing an experimental code layout
 algorithm, which hopefully is easy to adjust to take advantage of
 additional information about control flow.
 >
 > Code is at Phab:D4726. As it stands:
 >
 > * It works (only) on x64.
 > * Performance is around the same. (Better but within the margin of
 error).
 > * Compiler performance is slightly worse.
 >
 > There are a lot of knobs to tune this approach further but verifying any
 results is tedious so I stopped eventually.
 >
 > For now I hope to use this eventually as a substrate for the work on
 #14672.
 >
 > I've also created a wiki page here
 https://ghc.haskell.org/trac/ghc/wiki/Commentary/Compiler/CodeLayout with
 more details.

 I've came over a comment here:
 https://ghc.haskell.org/trac/ghc/wiki/Commentary/Compiler/Backends/NCG#Redundantreconstructionofthecontrolflowgraph

 > The allocator goes to considerable computational expense to construct
 all the flow edges in the group of instructions it's allocating for, by
 using the insnFuture function in the Instr pseudo-abstract type.

 Since my patch already constructs and maintains a CFG maybe we can reuse
 this. Although currently it's on the basic block not instruction level and
 I haven't checked if the comment is still relevant either.

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/15124#comment:7>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler