[GHC] #16044: Transition to C11 memory model

[GHC]

#16044: Transition to C11 memory model
-------------------------------------+-------------------------------------
           Reporter:  ulan           |             Owner:  (none)
               Type:  task           |            Status:  new
           Priority:  normal         |         Milestone:
          Component:  Runtime        |           Version:  8.6.3
  System                             |
           Keywords:                 |  Operating System:  Unknown/Multiple
       Architecture:                 |   Type of failure:  None/Unknown
  Unknown/Multiple                   |
          Test Case:                 |        Blocked By:
           Blocking:                 |   Related Tickets:
Differential Rev(s):                 |         Wiki Page:
-------------------------------------+-------------------------------------
 Following offline discussion with Ben Gamari and
 [https://mail.haskell.org/pipermail/ghc-devs/2018-November/016581.html Cmm
 Memory Model thread on ghc-devs@], I would like to propose to gradually
 transition RTS and Cmm to C11 memory model.

 There are multiple benefits:

 1) Correctness: C11 atomics are properly specified and are easier to
 reason about than the existing SMP.h barriers.

 2) Performance: compilers can optimize and generate better code for C11
 atomics. For example, on ARMv8 acquire/release operations use specialized
 LDA/STL instructions.

 3) Tooling support: once all code transitions to C11 atomics, tools like
 ThreadSanitizer can be used to find subtle data races.

 To make sure that code continues to compile with old compilers (pre C11),
 I plan to introduce wrapper functions (acquire_load, release_store, etc.)
 that are implemented either using C11 atomics or SMP.h barriers depending
 on the STDC_VERSION. Here is a
 [https://github.com/ulan/ghc/commit/0e8822a8038c3d0c770f4582c8f65b20d7823769
 prototype implementation].

 The idea is that newly written code should use the wrappers and we can
 gradually update the existing code.

 I think updating C code in RTS would work with this approach, but I am not
 sure about Cmm.

 Please let me know if you see potential issues, a better way to implement
 it, or oppose moving to C11 memory model.

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/16044>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler