search

rrnewton / haskell-lockfree

A collection of different packages for CAS based data structures.
106 stars 25 forks source link

No more load_load_barrier in GHC HEAD #88

Closed wavewave closed 5 months ago

wavewave commented 10 months ago

therefore, atomic-primops should be changed correspondingly.

https://gitlab.haskell.org/ghc/ghc/-/commits/master/rts/include/stg/SMP.h?ref_type=heads

https://gitlab.haskell.org/ghc/ghc/-/commit/7a7d326cdd5b640c3d40266b88f66b2a2dfeccbe

and store_load_barrier and write_barrier are also removed.

RyanGlScott commented 10 months ago

For context, https://gitlab.haskell.org/ghc/ghc/-/merge_requests/10628 removed these RTS functions.

What is the recommended alternative to using these functions?

bgamari commented 10 months ago

In general we have been trying to move away from explicit barriers entirely. {ACQUIRE,RELEASE,SEQ_CST}_FENCE exist if you really need a barrier but I would strongly recommend using a ordered atomic operation instead (e.g. ACQUIRE_LOAD, RELEASE_STORE, or similar).

RyanGlScott commented 10 months ago

For even further context, https://github.com/rrnewton/haskell-lockfree/commit/4f259b5072a0b1d9701139a1221cca4e7eeef975 is the commit that introduced the use of explicit memory barriers in atomic-primops. Judging from this, it appears that the main consumer is chaselev-deque. If push came to shove, we could just remove the memory barriers from atomic-primops. It would break chaselev-deque, but chaselev-deque is already broken.

RyanGlScott commented 9 months ago

Fixed in #89. I've uploaded these changes to Hackage as atomic-primops-0.8.5.

bgamari commented 8 months ago

It turns out that this fix wasn't quite right as the foreign imports were marked as ccall whereas they should have been prim, leading to https://gitlab.haskell.org/ghc/ghc/-/issues/24595. I will fix.

RyanGlScott commented 7 months ago

@bgamari addressed the issue in https://github.com/rrnewton/haskell-lockfree/issues/88#issuecomment-2021769236 as part of #90. I've released these changes to Hackage as part of atomic-primops-0.8.6.

hsyl20 commented 5 months ago

There is still an issue with these primops. See: https://gitlab.haskell.org/ghc/ghc/-/issues/25010

RyanGlScott commented 5 months ago

Reopening in light of https://gitlab.haskell.org/ghc/ghc/-/issues/25010.

hsyl20 commented 5 months ago

PR: https://github.com/rrnewton/haskell-lockfree/pull/91

RyanGlScott commented 5 months ago

I've uploaded atomic-primops-0.8.8 to Hackage, which contains the changes from #91. Thanks for the patch!