Anton Korobeynikov
Anton Korobeynikov
> Somewhat in line with [#152 (comment)](https://github.com/ablab/spades/issues/152#issuecomment-766323916), I was wondering, @asl, is the error correction step deterministic? If so, can it be performed on individual reads of a pair separately?...
@kmkappa Please do not hijack unrelated issues, open a new one
The SIL for generated pullback looks a bit suspicious to me: ``` // pullback of B.a(_:) sil private [ossa] @$s4main1BV1ay23DifferentiationUnittest7TrackedVySfGAHFTJpSUpSr : $@convention(thin) (@in_guaranteed Tracked, @owned _AD__$s4main1BV1ay23DifferentiationUnittest7TrackedVySfGAHF_bb0__PB__src_0_wrt_0) -> @out Tracked {...
> > So it looks like we're changing calling convention here from @callee_guaranteed to @callee_owned here. As a result no release is generated after function call and as callee assumes...
> Weird... If there's nothing I missed from applying the pullback bitcasting workaround, I'd expect any leaks to trigger an OSSA verification failure. Something apparently slipped through. We can check...
So, yes. On `main` at the end of `b(g: A)` we're having just one unowned refcount for `3`, we release the object and therefore deallocate it. On branch we're having...
So, the value is captured in the pullback of `*` here: ``` extension ${Self} where T: Differentiable & SignedNumeric, T == T.Magnitude, T == T.TangentVector { @usableFromInline @derivative(of: *) internal...
> Could this be an IRGen issue when its emitting a non -[callee_guaranteed]` `partial_apply`? The partial apply is `[callee_guaranteed]` here, yes. And it's pretty same in `main`, the values are...
> > > Could this be an IRGen issue when its emitting a non -[callee_guaranteed]` `partial_apply`? > > > > > > The partial apply is `[callee_guaranteed]` here, yes. And...
@rxwei I rebased this PR into `main` and ran into bunch of ABI-compatibility assertions. Looks like instead of reabstraction thunks (removed here: https://github.com/apple/swift/pull/34935/files#diff-01cf87f81a8c47d84be8508fac2cb0f1a4ba15919ad86d3c8904bccee60151b5L915) this PR emits just unchecked bitcasts. However,...