OSSA lifetime canonicalization can take a very long time in certain
cases in which there are large basic blocks. to mitigate this, add logic
to skip walking the liveness boundary for extending liveness to dead
ends when there aren't any dead ends in the function.
Updates `DeadEndBlocks` with a new `isEmpty` method and cache to
determine if there are any dead-end blocks in a given function.
When the utility is used by the ConsumeOperatorCopyableValuesChecker,
the checker guarantees that the lifetime can end at the consumes, that
there are no uses after those consumes. In that circumstance, the
utility maintains liveness to those consumes and as far as possible
without introducing a copy everywhere else.
The lack of complete lifetimes has forced the utility to extend liveness
of values to dead-ends. That extension, however, is in tension with the
use that the checker is putting the utility to. If there is a dead-end
after a consume, liveness must not be maintained to that dead-end.
rdar://147586673
Regardless of consumes of copies, if the original lexical value is not
consumed on a dead-end path, it must remain live to that dead-end.
rdar://145226757
It's sufficient just to have a struct with a kind and a value. There
aren't any cases where the payload's original type benefits from being
statically preserved--they're only ever obtained as `SILValue`s. Keep
the type safety by way of constructors.
In preparation for only recording the defs once, replace the
GraphNodeWorklist of defs with a SetVector. Preserve the current
visitation order by creating a worklist of indices to be visited.
Add a type which distinguishes among the types of defs that are pushed
onto the "def-use worklist". Note that it's not possible to rely on the
kind of value because the root may itself be a copy_value. For now, the
distinction is discarded as soon as the def is visited.
Use the more precise areUsesWithinBoundary API (which takes dead-end
blocks into account). This requires first updating liveness with the
newly created destroys.
Just clear all structures in a single method which is called wherever
clearing is done. Fixes a failure to clear discoveredBlocks under
certain circumstances.
Enhance the utility with the ability to end lifetimes of lexical values
at indicated instructions, overriding the usual behavior of maintaining
such lifetimes' previous endpoints (modulo non-deinit-barrier
instructions).
Parameterized `extendUnconsumedLiveness` on the ends of interest and the
action to take when visiting the extended boundary and named the
resulting function `visitExtendedUnconsumedBoundary`.
Not every block in a region which begins with the non-lifetime-ending
boundary of a value and ending with unreachable-terminated blocks has
the value available. If the unreachable-terminated blocks in this
boundary are not available, it is incorrect to insert destroys of the
value in them: it is an overconsume on some paths. Previously,
however, destroys were simply being inserted at the unreachable.
Here, this is fixed by finding the boundary of availability within that
region and inserting destroys before the terminators of the blocks on
that boundary.
rdar://116255254
OSSALifetimeCompletion needs to insert not at unreachable instructions
that appear after the non-lifetime-ending boundary of a value but rather
at the terminators of the availability boundary of the value within that
region. Once it does so, it will no longer be sufficient to check
whether the insertion point is an unreachable because such terminators
may be another terminator that appears on the availability boundary.
Prepare for that by recording the instructions that were found and
checking whether the destroy insertion point is such an instruction
before bailing rather than specifically checking for `unreachable`.
When canonicalizing the lifetime of a lexical value, deinit barriers are
respected. This is done by walking backwards from lifetime ends and
adding encountered deinit barriers to liveness.
Only destroy lifetime ends were walked back from under the assumption
that lifetimes would be complete. Without complete OSSA lifetimes,
however, it's necessary to also necessary to consider lifetimes that end
with unreachables. Unfortunately, we can't simply walk back from those
unreachables because there may be instructions which are secretly users
of the value being canonicalized (e.g. destroys of `partial_apply`s to
which a `begin_borrow` of the value was passed). Such uses don't appear
in the use list because lifetime canonicalization expects complete
lifetimes and only visits lifetime ends of `begin_borrow`s.
Here, instead, the instructions before the relevant unreachables are
added to liveness. In order to determine which unreachables are
relevant, it's necessary to have a liveness that includes the original
destroys. So a copy of liveness is created and those destroys are added
to it.
rdar://115468707
llvm::SmallSetVector changed semantics
(https://reviews.llvm.org/D152497) resulting in build failures in Swift.
The old semantics allowed usage of types that did not have an
`operator==` because `SmallDenseSet` uses `DenseSetInfo<T>::isEqual` to
determine equality. The new implementation switched to using
`std::find`, which internally uses `operator==`. This type is used
pretty frequently with `swift::Type`, which intentionally deletes
`operator==` as it is not the canonical type and therefore cannot be
compared in normal circumstances.
This patch adds a new type-alias to the Swift namespace that provides
the old semantic behavior for `SmallSetVector`. I've also gone through
and replaced usages of `llvm::SmallSetVector` with the
`Swift::SmallSetVector` in places where we're storing a type that
doesn't implement or explicitly deletes `operator==`. The changes to
`llvm::SmallSetVector` should improve compile-time performance, so I
left the `llvm::SmallSetVector` where possible.
Previously, the utility bailed out on lexical lifetimes because it
didn't respect deinit barriers. Here, deinit barriers are found and
added to liveness if the value is lexical. This enables copies to be
propagated without hoisting destroys over deinit barriers.
rdar://104630103
Use BasicBlockBitfield to record per-block liveness state. This has
been the intention since BasicBlockBitfield was first introduced.
Remove the per-field bitfield from PrunedLiveBlocks. This
(re)specializes the data structure for scalar liveness and drastically
simplifies the implementation.
This utility is fundamental to all ownership utilities. It will be on
the critical path in many areas of the compiler, including at
-Onone. It needs to be minimal and as easy as possible for compiler
engineers to understand, investigate, and debug.
This is in preparation for fixing bugs related to multi-def liveness
as used by the move checker.
Specifically, previously if we emitted an error we just dumped all of the
consuming uses. Now instead for each consuming use that needs a copy, we perform
a search for a specific boundary use (consuming or non-consuming) that is
reachable from the former and emit a specialized error for it. Thus we emit for
the two consuming case the normal consumed twice error, and now for
non-consuming errors we emit the "use after consume" error.
For those who are unaware, CanonicalizeOSSALifetime::canonicalizeValueLifetime()
is really a high level driver routine for the functionality of
CanonicalizeOSSALifetime that computes liveness and then rewrites copies using
boundary information. This change introduces splits the implementation of
canonicalizeValueLifetime into two parts: a first part called computeLiveness
and a second part called rewriteLifetimes. Internally canonicalizeValueLifetime
still just calls these two methods.
The reason why I am doing this is that it lets the move only object checker use
the raw liveness information computed before the rewriting mucks with the
analysis information. This information is used by the checker to compute the raw
liveness boundary of a value and use that information to determine the list of
consuming uses not on the boundary, consuming uses on the boundary, and
non-consuming uses on the boundary. This is then used by later parts of the
checker to emit our errors.
Some additional benefits of doing this are:
1. I was able to eliminate callbacks in the rewriting stage of
CanonicalOSSALifetimes which previously gave the checker this information.
2. Previously the move checker did not have access to the non-consuming boundary
uses causing us to always fail appropriately, but sadly not emit a note showing
the non-consuming use. I am going to wire this up in a subsequent commit.
The other change to the implementation of the move checker that this caused is
that I needed to add an extra diagnostic check for instructions that consume the
value twice or consume the value and use the value. The reason why this must be
done is that liveness does not distinguish in between different operands on the
same instruction meaning such an error would be lost.
For most uses, some access scopes must be "respected"--if an extended
value's original lifetime originally extends beyond an access scope, its
canonicalized lifetime must not end _within_ such scopes (although
ending before them is fine). Currently, to be conservative, the utility
applies this behavior to all access scopes.
For move-only values, however, lifetimes end at final consumes without
regard to access scopes.
Allow this behavior to be controlled by whether or not a
NonLocalAccessBlockAnalysis is provided to the utility in its
constructor.
rdar://104635319
When encountering inside a borrow scope a non-lexical move_value or a
move_value [lexical] where the borrowed value is itself already lexical,
delete the move_value and regard its uses as uses of the moved-from
value.
To improve the debugging experience of values whose lifetimes are
canonicalized without compromising the semantics expressed in the source
language, when canonicalizing OSSA lifetimes at Onone, lengthen
lifetimes as much as possible without incurring copies that would be
eliminated at O.
rdar://99618502
Rather than having finding the boundary be a single combined step,
separate finding the original boundary from extending that boundary.
This enables inserting an optional step between those steps, namely to
extend unconsumed liveness to its original extent at Onone.
It is possible for phis to be marked live. With guaranteed phis, they
will be the last uses and be non-consuming. In this case, the
merge block will have multiple predecessors whose terminators are on the
boundary. When inserting destroys, track whether a merge point has been
visited previously.
To facilitate this, restructure the boundary extension and destroy
insertion code.
Previously, the extender was building up a list of places at which to
insert destroys. In particular it was using the "boundary edge"
collection for all blocks at the beginning of which a destroy should be
created. In particular, it would add merge blocks. Such blocks are not
boundary blocks.
Here, the extender produces a PrunedLivenessBoundary which doesn't
violate that invariant.
This required some changes to the destroy insertion code to find where
to insert destroys. It is now similar to
PrunedLivenessBoundary::visitInsertionPoints and could be used as a
template for a PrunedLivenessBoundary::visitBoundaryPoints through which
::visitInsertionPoints might be factored.
Previously, the destroys set (now set vector) wasn't ever being cleared.
The result was that users could get overly pessimistic behavior if they
had previously used the utility with a destroy that came after the
destroys relevant for its current run. Here, it is cleared when the
utility is initialized with a new def. Addresses a TODO in the
copy_propagation test.
