[GHC] #14399: NCG: dead code stripping prevention wastes space

[GHC]

#14399: NCG: dead code stripping prevention wastes space
-------------------------------------+-------------------------------------
           Reporter:  heisenbug      |             Owner:  (none)
               Type:  bug            |            Status:  new
           Priority:  normal         |         Milestone:
          Component:  Compiler       |           Version:  8.2.1
           Keywords:  low-hanging    |  Operating System:  Unknown/Multiple
       Architecture:                 |   Type of failure:  None/Unknown
  Unknown/Multiple                   |
          Test Case:                 |        Blocked By:
           Blocking:                 |   Related Tickets:
Differential Rev(s):                 |         Wiki Page:
-------------------------------------+-------------------------------------
 = Symptom

 Consider this NCG-generated code:
 {{{
         .long  _T5001a_zdp1Comorphism_info -
 _T5001a_zdp1Comorphism_info_dsp
 .data
 .align 3
 .align 0
 .globl _T5001a_mapzusentence_closure
 _T5001a_mapzusentence_closure:
         .quad   _T5001a_mapzusentence_info
 .text
 .align 3
 _T5001a_mapzusentence_info_dsp:
 .align 3
         .quad   4294967301
         .quad   0
         .quad   14
 .globl _T5001a_mapzusentence_info
 _T5001a_mapzusentence_info:
 }}}
 There is an inert `.long` at the end of each Cmm procedure. This fills up
 precious instruction cache and wastes space.

 = Potential solutions

 Some ideas to improve the situation follow:

 1. Move all those longs to a "table" at the end of the compilation unit
 (idea by Henning Thielemann),
 2. coalesce the relocation into the previous instruction by explicitly
 calculating the opcode (craziness by me),
 3. coalesce the relocation into the following info table like this:
 {{{
         .quad   4294967301 - 24 + _T5001a_zdp1Comorphism_info -
 _T5001a_zdp1Comorphism_info_dsp
 }}}
 4. Instead of exporting `_T5001a_mapzusentence_info` export
 `_T5001a_mapzusentence_info_dsp` and jump/call
 `_T5001a_mapzusentence_info_dsp + 24` instead.  (Idea by Henning
 Thielemann.)

 == Discussion

 1. Improves instruction cache only
 2. hard to implement, leads to ugly assembly
 3. wastes a long per compilation unit (at the end), needs a new state in
 the assembly writer
 4. leads to unintuitive (ugly?) assembly.

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/14399>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler