Michael Gottesman 5294e51ef1 [mandatory-inlining] Fix lifetime extension error. ...

This bug is caused by a quirk in the API of the linear lifetime
checker. Specifically, even though valueHasLinearLifetime is passed a SILValue
(the value whose lifetime one is checking), really it doesnt care about that
value (except for error diagnostics). Really it just cares about the parent
block of the value since it assumes that the value is guaranteed to dominate all
uses.

This creates a footgun when if one is writing code using "generic ossa/non-ossa"
routines on SILBuilder (the emit*Operation methods), if one in non-ossa code
calls that function, it returns the input value of the strong_retain. This
causes the linear lifetime error, to use the parent block of the argument of the
retain, instead of the parent block of the retain itself. This then causes it to
find the wrong leaking blocks and thus insert destroys in the wrong places.

I fix this problem in this commit by noting that the partial apply is our
original insertion point for the copy, so of course it is going to be in the
same block. So I changed the linear lifetime checker to check for leaks with
respect to the partial applies result.

In a subsequent commit, I am going to add a new API on top of this that is based
around the use of the value by the partial apply (maybe
extendLifetimeFromUseToInsertionPoint?). By using the use, it will express in
code more clearly what is happening here and will insert the copy for you.

rdar://54234011

2019-08-22 09:37:29 -07:00

1.8 KiB

Raw Blame History

View Raw