The SIL optimizer has fundamental bugs that result in dropping non-Copyable
struct & enum the deinitializers.
Fix this by
1. correctly representing the ownership of struct & enum values that are
initialized from trivial values.
2. checking move-only types before deleting forwarding instructions.
These bugs block other bug fixes. They are exposed by other unrelated SIL
optimizations to SIL. I'm sure its possible to expose the bugs with source-level
tests, but the current order of inlining and deinit devirtualization has been
hiding the bugs and complicates reproduction.
Check if an operand's instruction has been deleted in `UseList.next()`.
This allows to delete an instruction, which has two uses of a value, during use-list iteration.
The `maxBitfieldID` was defined as `size_t` and used to assert
bitfield ID values. The `currentBitfieldID` of `SILFunction` is
defined as `uint64_t`, so we should consistently use `uint64_t`
for `maxBitfieldID` as well for 32-bit platforms.
This is necessary to fix a recent OSSA bug that breaks common occurrences on
mark_dependence [nonescaping]. Rather than reverting that change above, we make
forward progress toward implicit borrows scopes, as was the original intention.
In the near future, all InteriorPointer instructions will create an implicit
borrow scope. This means we have the option of not emitting extraneous
begin/end_borrow instructions around intructions like ref_element_addr,
open_existential, and project_box. After that, we can also migrate
GuaranteedForwarding instructions like tuple_extract and struct_extract.
`enum` is a signed integer type on some platforms. The return value of
`std::numeric_limits<T>::max()` is a `size_t` which is unsigned. Further
manipulation retains the unsigned-ness until the assignment which
results in truncation. Change the type to a `constexpr` constant
definition and remove the `enum` type along with a type conversion to an
explicitly unsigned integral type.
Compute, update and handle borrowed-from instruction in various utilities and passes.
Also, used borrowed-from to simplify `gatherBorrowIntroducers` and `gatherEnclosingValues`.
Replace those utilities by `Value.getBorrowIntroducers` and `Value.getEnclosingValues`, which return a lazily computed Sequence of borrowed/enclosing values.
* Let the customBits and lastInitializedBitfieldID share a single uint64_t. This increases the number of available bits in SILNode and Operand from 8 to 20. Also, it simplifies the Operand class because no PointerIntPairs are used anymore to store the operand pointer fields.
* Instead make the "deleted" flag a separate bool field in SILNode (instead of encoding it with the sign of lastInitializedBitfieldID). Another simplification
* Enable important invariant checks also in release builds by using `require` instead of `assert`. Not catching such errors in release builds would be a disaster.
* Let the Swift optimization passes use all the available bits and not only a fixed amount of 8 (SILNode) and 16 (SILBasicBlock).
The specific semantics is if we assign into a transferring parameter's field,
then we "merge" src's value into the transferring parameter, so we
conservatively leave the region of the transferring parameter alone. If we
assign over the entire transferring parameter, we perform an assign fresh since
any value that used to be in the transferring parameter cannot reference
anything in its new value since they are all gone.
To do this I used 8 spare bits in the pointers in Operand for the custom flags.
The reason I did this is just like one sometimes wants to iterate and use
sets/worklists with Nodes/Blocks, one often wants to do it with operands
especially in situations where one wants to know a using instruction and the
value on the instruction that was used.
In a downstream project we are using the Swift code as a library and
compiling with C++20. This mostly works except for two backward
incompatibilities introduced by C++20:
* [P1008] disallows aggregate initialization when defaulted or deleted
constructors are specified. This lead to a compilation error for
aggregate initialization of `swift::Compilation::Result`. The backward
and forward compatible fix for that is to provide a constructor
turning aggregate initializations into a normal constructor call.
* [P1185] introduces more candidates for overload resolution of
comparison operator calls. As explained in [P1630], this leads in some
cases to ambiguity compilation errors in C++20, which is exactly the
case with `swift::ValueOwnershipKind`. The fix in this case is to
remove some redundant operator calls conditionally on the
`__cpp_impl_three_way_comparison` feature test macro, which [includes
the papers mentioned above][feature_test].
[P1008]: https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1008r1.pdf
[P1185]: https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1185r2.html
[P1630]: https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1630r1.html
[feature_test]: https://en.cppreference.com/w/cpp/feature_test
This is phase-1 of switching from llvm::Optional to std::optional in the
next rebranch. llvm::Optional was removed from upstream LLVM, so we need
to migrate off rather soon. On Darwin, std::optional, and llvm::Optional
have the same layout, so we don't need to be as concerned about ABI
beyond the name mangling. `llvm::Optional` is only returned from one
function in
```
getStandardTypeSubst(StringRef TypeName,
bool allowConcurrencyManglings);
```
It's the return value, so it should not impact the mangling of the
function, and the layout is the same as `std::optional`, so it should be
mostly okay. This function doesn't appear to have users, and the ABI was
already broken 2 years ago for concurrency and no one seemed to notice
so this should be "okay".
I'm doing the migration incrementally so that folks working on main can
cherry-pick back to the release/5.9 branch. Once 5.9 is done and locked
away, then we can go through and finish the replacement. Since `None`
and `Optional` show up in contexts where they are not `llvm::None` and
`llvm::Optional`, I'm preparing the work now by going through and
removing the namespace unwrapping and making the `llvm` namespace
explicit. This should make it fairly mechanical to go through and
replace llvm::Optional with std::optional, and llvm::None with
std::nullopt. It's also a change that can be brought onto the
release/5.9 with minimal impact. This should be an NFC change.
"reborrow" flag on the SILArgument avoids transitive walk over the phi operandsi
to determine if it is a reborrow in multiple utilities.
SIL transforms must keep the flag up-to-date by calling SILArgument::setReborrow.
SILVerifier checks to ensure the flag is not invalidated.
Currently "escaping" is not used anywhere.
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.
This allows code to handle terminator results similar to any other
instruction result. Data flow should generally handle terminator
results like any other instruction that may forward operand ownership
to its results. The fact that it is represented as a block argument is
an implementation detail that gets in the way of conceptual
simplicity.
This lets us write optimizer unit tests and selectively debug the
optimizer in general. We'll be able trace analyses and control
optimization selectively for certain values.
Adding a trace flag to debug_value is the easiest way to start using
it experimentally and develop the rest of the infrastructure. If this
takes off, then we can consider a new `trace_value`
instruction. For now, reusing debug_value is the least intrusive way to
start writing liveness unit tests.
Andy some time ago already created the new API but didn't go through and update
the old occurences. I did that in this PR and then deprecated the old API. The
tree is clean, so I could just remove it, but I decided to be nicer to
downstream people by deprecating it first.
Since I am beginning to prepare for adding real move only types to the language,
I am renaming everything that has to do with copyable types "move only wrapped"
values instead of move only. The hope is this reduces/prevents any confusion in
between the two.
This pass lowers moveonly-ness from the IR after we have finished move only
checking. The transform can be configured in two ways: such that it only handles
trivial types and such that it does trivial and non-trivial types. For ease of
use, I created two top level transforms (TrivialMoveOnlyTypeEliminator and
MoveOnlyTypeElimintor) that invoke the two behaviors by configuring the
underlying transform slightly differently.
For now, I am running first the trivial-only and then the all of the above
lowering. The trivial only pass will remain at this part of the pipeline
forever, but with time we are going to move the lower everything pass later into
the pipeline once I have audited the optimizer pipeline to just not perform any
work on move only types. That being said, currently we do not have this
guarantee and this patch at least improves the world and lets us codegen no
implicit copy code again.
It just trampolines to calling ValueBase::dump(). The reason I added this is
sometimes one wants to do this in the debugger and confuses if one has a
ValueBase or a SILValue causing lldb to be unhappy. By mirroring this API, one
can do the same operation on either type and just move on with ones life.
I also wrapped this in LLVM_ATTRIBUTE_DEPRECATED with a msg saying only for use
in the debugger so people do not call dump() in the codebase by mistake.
