ReflectionContext is getting too large, and having to thread a
<Remote> template parameter through all code that wants to construct
typerefs is getting tricky. This is the first patch in a refactoring
to move some stuff out of ReflectionContext.
The compiler is generally free to not include pointers to metadata in
heap boxes, which are used for closure captures, if it knows you can get
to metadata through some other path. These MetadataSource classes will
describe a sequence of steps to get to metadata at runtime.
In the short term, this will be useful for describing the layout of
function/closure capture contexts, which can vary depending on what is
captured.
Previously we would pre-process the same input files in the ObjC
and non-ObjC tests. This made the tests difficult to update because
the output of one was a subset of the other, and only one of the
two tests would run on any given platform.
Instead, let's just put the Objective-C tests in their own test
and input files.
In order to perform layout, the remote mirrors library needs to know
about the size, alignment and extra inhabitants of builtin types.
Ideally we would emit a reflection info section in libswiftRuntime.o,
but in the meantime just duplicate builtin type metadata for all
builtin types referenced from the current module instead.
In practice only the stdlib and a handful of overlays like the SIMD
overlay use builtin types, and only a few at a time.
Tested manually by running swift-reflection-tool on the standard
library -- I'll add automated tests by using -parse-stdlib to
reference Builtin types in a subsequent patch that adds more layout
logic.
NFC if -enable-reflection-metadata is off.
When creating a TypeRef from metadata, we have a parent pointer
handy, and construct the TypeRef directly, so there's no need
to mutate the TypeRef after the fact.
When demangling a TypeRef from a string, the mangling encodes
the parent module or type context, so we can set it when
constructing the TypeRef there too.
We will be handing pointers to typerefs over the SwiftRemoteMirrors C
API boundary, at which point it is unclear who will hold onto a shared
pointer. The useful lifetime of a typeref is tied to the
ReflectionContext for which they were created anyway so, when it goes
away, all of those typerefs can go away anyway.
We can't use LLVM's bump-pointer allocator here because we only build
the Support library for the host. As a compromise, stuff new typeref
pointers into a vector pool, where they will be taken down during
ReflectionContext's destructor.
The most important change is a fix for dependent member lookup in
bound generic TypeRefs.
The TypeRef logic here is a similified version of what Sema does,
and roughly works as follows:
When looking up a member type O.M for an original type O and
member type M, we first substitute O, to yield a concrete type S.
Then, we look up M in the serialized conformance info section
to yield the unsubstituted member type T', and apply substitutions
to produce the substituted member type T.
Trouble is, we were applying the wrong substitutions the second
time around.
If S is a bound generic type, the "base type substitutions" of S
are the substitutions required to turn the fully-abstracted
bound generic type into S. This might not be the same substitution
map that took us from O to S.
Apart from that, add some assertions that the result of a
substitution is always concrete and that the substitution map
must not contain all generic parameters referenced in the
original type.
Just drop labels when demangling TypeRefs. This is OK for now
since labels do not affect layout.
Vararg tuple types cannot appear directly as the type of
storage, however they can appear in function input types, and
therefore must be minimally supported. Since we don't plan on
doing function call reflection just yet, this doesn't matter
for now, but again we need to not crash.
With this patch, all TypeRefs in the Swift standard library now
successfully demangle and print.
swift-reflection-test is now the test that forks a swift executable
and performs remote reflection, making it runnable on other targets,
such as the iOS simulator.
swift-reflection-dump is now a host-side tool that dumps the remote
reflection sections for any platform binary and will continue to
link in LLVM object file support.
This necessitates finally moving lib/Refleciton into stdlib/public,
since we're linking target-specific versions of the test tool and
we would eventually like to adopt some of this functionality in
the runtime anyway.
We'll need to drill into nested structs to get their field typerefs and
so on, without metadata necessarily available. Decouple the lookup from
the address.
TODO: Cache associated type descriptors based on typeref or mangled
name.
This is more amenable to cross-platform remote reflection tests.
Also add a new callback to the memory reader: getSymbolAddress,
which will be used for getting the addresses of nominal type
descriptors for concrete and fixed generic types.
- Read the Parent pointer out of Class/Value metadata and create
typerefs for them.
- Add Parent fields to NominalTypeRef and BoundGenericTypeRef.
- Add TypeRef::getSubstMap(), which creates a new generic argument
map after substitution has taken place on it. This is used to
continue to burrow into nested value types, where generic type
parameters may have a different index.
- Use a DenseMap as that generic argument map.
- Unconditionally key the generic argument map with (Depth, Index)
- Clean up ordering and presentation of Index and Depth. In the rest
of the compiler, Depth comes before Index.
Once an unsubstituted typeref for a field is built, we substitute
`GenericTypeParameterTypeRef`s with concrete ones built from the generic
arguments of concrete bound generic metadata.
During that process, if we run into a `DependentMemberTypeRef` (e.g.
something of type T.Index), we substitute the base (T) using the current
list of substitutions, and then resolve what `Index` is for the base
using the associated type metadata in the 'assocty' data section.
Nominal type descriptors use declared types for their mangled names,
so we need to use them when scanning the fieldmd section for a
matching record. This is fine because the descriptor can tell us
about the type's generics. Individual field records continue to use
the interface type.
