SIL differentiability witnesses are a new top-level SIL construct mapping
an "original" SIL function and derivative configuration to derivative SIL
functions.
This patch adds `SILDifferentiabilityWitness` IRGen.
`SILDifferentiabilityWitness` has a fixed `{ i8*, i8* }` layout:
JVP and VJP derivative function pointers.
Resolves TF-1146.
This reverts commit 8247525471. While
correct, it has uncovered several issues in existing code bases that
need to be sorted out before we can land it again.
Fixes rdar://problem/57846390.
Prespecialized metadata records must refer to value witness tables that
have correct values for size and stride. In order to do that, a
prespecialized value witness table must emitted and referred to. Here,
a fully specialized value witness table is emitted.
For the moment, the value witness table is fully specialized. Having it
be specialized, though, will likely cause serious code size problems,
since it results in specialized value witness functions being generated.
rdar://problem/58088270
Metadata accessors are dependent on prespecializations of the metadata
of generic, in-module types. Those prespecializations are themselves
dependent on usages of the types in functions. Consequently, the
accessors must be emitted after all the functions are emitted.
When possible, directly reference metadata prespecializations. Doing so
is possible when the type is defined in the same module, because in
those cases the metadata accessor can be modified to ensure that the
prespecialized metadata is canonical.
rdar://problem/56994171
Added worklist of prespecializations awaiting lazy emission to
IRGenModule. Added map from type decl to list of bound types for which
prespecializations will be emitted.
For now, no specializations are emitted.
WASM currently is treated identically to the ELF paths. Collocate the
types to make it easier to ensure that all the paths are correctly
handling the emission. This adds the missed case for the module hash.
We've changed *what* is serialized by changing the way
@_dynamicReplacement is type checked, but not *how* it's
serialized. Bump the format so there aren't strange incompatibilities
because of this.
Complete the refactoring by splitting the semantic callers for the original decl of a dynamically replaced declaration.
There's also a change to the way this attribute is validated and placed. The old model visited the attribute on any functions and variable declarations it encountered in the primary. Once there, it would strip the attribute off of variables and attach the corresponding attribute to each parsed accessor, then perform some additional ObjC-related validation.
The new approach instead leaves the attribute alone. The request exists specifically to perform the lookups and type matching required to find replaced decls, and the attribute visitor no longer needs to worry about revisiting decls it has just grafted attributes onto. This also means that a bunch of parts of IRGen and SILGen that needed to fan out to the accessors to ask for the @_dynamicReplacement attribute to undo the work the type checker had done can just look at the storage itself. Further, syntactic requests for the attribute will now consistently succeed, where before they would fail dependending on whether or not the type checker had run - which was generally not an issue by the time we hit SIL.
This is a first version of cross module optimization (CMO).
The basic idea for CMO is to use the existing library evolution compiler features, but in an automated way. A new SIL module pass "annotates" functions and types with @inlinable and @usableFromInline. This results in functions being serialized into the swiftmodule file and thus available for optimizations in client modules.
The annotation is done with a worklist-algorithm, starting from public functions and continuing with entities which are used from already selected functions. A heuristic performs a preselection on which functions to consider - currently just generic functions are selected.
The serializer then writes annotated functions (including function bodies) into the swiftmodule file of the compiled module. Client modules are able to de-serialize such functions from their imported modules and use them for optimiations, like generic specialization.
The optimization is gated by a new compiler option -cross-module-optimization (also available in the swift driver).
By default this option is off. Without turning the option on, this change is (almost) a NFC.
rdar://problem/22591518
Today in far more cases we are using mangled strings to look up metadata at
runtime. If we do this for an objc class but for whatever reason we do not have
any other references to the class, the static linker will fail to link in the
relevant framework. The reason why this happens is that autolinking is treated
by the static linker as a hint that a framework may be needed rather than as a
"one must link against the framework". If there aren't any undefined symbols
needed by the app from that framework, the linker just will ignore the hint. Of
course this then causes the class lookup to fail at runtime when we use our
mangled name to try to lookup the class.
I included an Interpreter test as well as IRGen tests to make sure that we do
not regress here in the future.
NOTE: The test modifications here are due to my moving the ObjCClasses framework
out of ./test/Interpreters/Inputs => test/Inputs since I am using it in the
IRGen test along side the interpreter test.
rdar://56136123
Otherwise one TU could only require the type descriptor without metadata
and another TU could require metadata and type descriptor. Whether the
metadata access function is available would then depend on the linking
order of the two TUs.
rdar://56929811
First, remove the AvailabilityContext parameter; it was confusing because
we actually always want to use the deployment target here.
Then, split this method up into three methods:
- isAlwaysWeakImported(): simply checks for a @_weakLinked attribute, either
on the declaration itself or one of its parent contexts.
- getAvailabilityForLinkage(): returns the OS version availability when
this declaration was introduced, or if the declaration does not have
explicit availability, check it's storage (if its an accessor), or its
parent contexts.
- isWeakImported(ModuleDecl *fromModule): combines these two checks to
determine if the declaration should be weak linked when referenced from
the given module, or if it might be weak referenced from some module
(if the module parameter is null).
When we generate code that asks for complete metadata for a fully concrete specific type that
doesn't have trivial metadata access, like `(Int, String)` or `[String: [Any]]`,
generate a cache variable that points to a mangled name, and use a common accessor function
that turns that cache variable into a pointer to the instantiated metadata. This saves a bunch
of code size, and should have minimal runtime impact, since the demangling of any string only
has to happen once.
This mostly just works, though it exposed a couple of issues:
- Mangling a type ref including objc protocols didn't cause the objc protocol record to get
instantiated. Fixed as part of this patch.
- The runtime type demangler doesn't correctly handle retroactive conformances. If there are
multiple retroactive conformances in a process at runtime, then even though the mangled string
refers to a specific conformance, the runtime still just picks one without listening to the
mangler. This is left to fix later, rdar://problem/53828345.
There is some more follow-up work that we can do to further improve the gains:
- We could improve the runtime-provided entry points, adding versions that don't require size
to be cached, and which can handle arbitrary metadata requests. This would allow for mangled
names to also be used for incomplete metadata accesses and improve code size of some generic
type accessors. However, we'd only be able to take advantage of the new entry points in
OSes that ship a new runtime.
- We could choose to always symbolic reference all type references, which would generally reduce
the size of mangled strings, as well as make runtime demangling more efficient, since it wouldn't
need to hit the runtime caches. This would however require that we be able to handle symbolic
references across files in the MetadataReader in order to avoid regressing remote mirror
functionality.
Correct the COMDAT handling for the type metadata emission. In
particular, declarations should not be emitted with COMDATs. The
pattern used in the compiler for the metadata emission is different than
the other structures. In particular, `getAddrOfTypeMetadata` emits the
declaration for the type metadata symbol. The definition for the type
metadata will be emitted via `defineTypeMetadata`. This change adjusts
the IRGen to remove the COMDAT on the declaration permitting the
definition from `defineTypeMetadata` being emitted strongly.
This exposes a bug in the LLVM MC layer which is a separate change that
will enable this to function properly.
