57ddc6f · reflex · git.iko.soy

git.iko.soy/reflex

Add NOINLINE pragma to mkFastWeakTicket

(Also move what I think is the corresponding comment) Without this NOINLINE, the FastWeakTicket can be fused away (in cacheSubscription in particular), causing early unsubscriptions. This manifested itself in reflex-dom-core as items in a nested list never being added to the DOM, because runWithReplace in the immediate builder was being unsubscribed from the lifted runWithReplace result events.

commit57ddc6f878e9ce44452852abc7e4e790cfb008bf authorTom Smalley <tfsmalley@gmail.com> dateMon 30 Mar 2020, 19:44 (7 years ago) parent3e81151

2 files changed+10−10

compare with develop

unifiedsplit

src/Reflex/FastWeak.hs+10 src/Reflex/Spider/Internal.hs−10

src/Reflex/FastWeak.hs+10

@@ -151,6 +151,16 @@ getFastWeakTicket w = do

151151

152152-- | Create a 'FastWeakTicket' directly from a value, creating a 'FastWeak' in the process

153153-- which can be obtained with 'getFastWeakTicketValue'.

154+--

155+-- This function is marked NOINLINE so it is opaque to GHC.

156+-- If we do not do this, then GHC will sometimes fuse the constructor away

157+-- so any weak references that are attached to the ticket will have their

158+-- finalizer run. Using the opaque constructor, GHC does not see the

159+-- constructor application, so it behaves like an IORef and cannot be fused away.

160+--

161+-- The result is also evaluated to WHNF, since forcing a thunk invalidates

162+-- the weak pointer to it in some cases.

163+{-# NOINLINE mkFastWeakTicket #-}

154164mkFastWeakTicket :: a -> IO (FastWeakTicket a)

155165-- I think it's fine if this is lazy - it'll retain the 'a', but so would the output; we just need to make sure it's forced before we start relying on the

156166-- associated FastWeak to actually be weak

src/Reflex/Spider/Internal.hs−10

@@ -450,16 +450,6 @@ data Subscriber x a = Subscriber

450450 , subscriberRecalculateHeight :: !(Height -> IO ())

451451 }

452452

453−--TODO: Move this comment to WeakBag

454−-- These function are constructor functions that are marked NOINLINE so they are

455−-- opaque to GHC. If we do not do this, then GHC will sometimes fuse the constructor away

456−-- so any weak references that are attached to the constructors will have their

457−-- finalizer run. Using the opaque constructor, does not see the

458−-- constructor application, so it behaves like an IORef and cannot be fused away.

459−--

460−-- The result is also evaluated to WHNF, since forcing a thunk invalidates

461−-- the weak pointer to it in some cases.

462−

463453newSubscriberHold :: (HasSpiderTimeline x, Patch p) => Hold x p -> IO (Subscriber x p)

464454newSubscriberHold h = return $ Subscriber

465455 { subscriberPropagate = {-# SCC "traverseHold" #-} propagateSubscriberHold h