to work with aggregates containing unknown values. Such aggregates
can be generated when an instruction is skipped during constant
evaluation and its results are used to create a struct.
Linear maps are captured in vjp routine via callee-guaranteed partial apply and are passed as @owned references to the enclosing pullback that finally consumes them. Necessary retains are inserted by a partial apply forwarder.
However, this is not the case when the function being differentiated contains loops as heap-allocated context is used and bare pointer is captured by the pullback partial apply. As a result, partial apply forwarder does not retain the linear maps that are owned by a heap-allocated context, however, they are still treated as @owned references and therefore are released in the pullback after the first call. As a result, subsequent pullback calls release linear maps and we'd end with possible use-after-free.
Ensure we retain values when we load values from the context.
Reproducible only when:
* Loops (so, heap-allocated context)
* Pullbacks of thick functions (so context is non-zero)
* Multiple pullback calls
* Some cleanup while there
Fixes#64257
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.
Add a separate 'verifyOwnership()' entry point so it's possible
to check OSSA lifetimes at various points.
Move SILGenCleanup into a SILGen pass pipeline.
After SILGen, verify incomplete OSSA.
After SILGenCleanup, verify ownership.
The `isEscaping` function is called a lot from ARCSequenceOpt and ReleaseHoisting.
To avoid quadratic complexity for large functions, limit the amount of work what the EscapeUtils are allowed to to.
This keeps the complexity linear.
The arbitrary limit is good enough for almost all functions.
It lets the EscapeUtils do several hundred up/down walks which is much more than needed in most cases.
Fixes a compiler hang
https://github.com/apple/swift/issues/63846
rdar://105795976
Otherwise, sometimes when the object checker emits a diagnostic and cleans up
the IR, some of the cleaned up copies are copies that should have been handled
by the address checker. The end result is that the address checker does not emit
diagnostics for that IR. I found this problem was exascerbated when writing code
for escaping closures.
This commit also cleans up the passes in preparation for at a future time moving
some of the transformations into the utils folder.
Although nonescaping closures are representationally trivial pointers to their
on-stack context, it is useful to model them as borrowing their captures, which
allows for checking correct use of move-only values across the closure, and
lets us model the lifetime dependence between a closure and its captures without
an ad-hoc web of `mark_dependence` instructions.
During ownership elimination, We eliminate copy/destroy_value instructions and
end the partial_apply's lifetime with an explicit dealloc_stack as before,
for compatibility with existing IRGen and non-OSSA aware passes.
Encapsulate all the complexity of reborrows and guaranteed phi in 3
ownership liveness interfaces:
LinerLiveness, InteriorLiveness, and ExtendedLiveness.
The Swift Simplification pass can do more than the old MandatoryCombine pass: simplification of more instruction types and dead code elimination.
The result is a better -Onone performance while still keeping debug info consistent.
Currently following code patterns are simplified:
* `struct` -> `struct_extract`
* `enum` -> `unchecked_enum_data`
* `partial_apply` -> `apply`
* `br` to a 1:1 related block
* `cond_br` with a constant condition
* `isConcrete` and `is_same_metadata` builtins
More simplifications can be added in the future.
rdar://96708429
rdar://104562580
If a `debug_step` has the same debug location as a previous or succeeding instruction it is removed.
It's just important that there is at least one instruction for a certain debug location so that single stepping on that location will work.
The changes are intentionally were made close to the original implementation w/o possible simplifications to ease the review
Fixes#63207, supersedes #63379 (and fixes#63234)
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
This is a cleaner separation of concerns. The reason why I did not do this
originally is that I thought I would need to reuse this functionality in the
address checker, but this issue actually does not come up there since we project
the address and then load instead of load and then project.
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.
This enables us to emit the appropriate error for consuming uses of fields and
also causes us to eliminate copies exposed by using fields of a move only type
in a non-consuming way.
rdar://103271138
If a unit test is miswritten in the sense that the test expects an
instance of one type by an instance of some other type is specified,
print that out.
* for testing: add the option `-simplify-instruction=<instruction-name>` to only run simplification passes for that instruction type
* on the swift side, add `Options.enableSimplification`
* split the `PassContext` into multiple protocols and structs: `Context`, `MutatingContext`, `FunctionPassContext` and `SimplifyContext`
* change how instruction passes work: implement the `simplify` function in conformance to `SILCombineSimplifyable`
* add a mechanism to add a callback for inserted instructions
