This indicates that the "self" argument to the current function is always dynamically of the exact
static base class type, allowing metadata accesses in IRGen to use the local self metadata to answer
metadata requests for the class type. Set this attribute on allocating entry points of designated
inits, which is one of the most common places where we emit redundant metadata accesses.
Recent Swift uses 2 as the is-Swift bit when running on newer versions, and 1 on older versions. Since it's difficult or impossible to know what we'll be running on at build time, make the selection at runtime.
Field offset vectors are always filled out with either zero or the static layout's offset, depending on the metadata initialization strategy. This change means that the static layout's offset will only be non-zero for properties with a statically-known layout. Existing runtimes doing dynamic class layout assign class properties a zero offset if the field offset vector entry is zero and the property is zero-sized. So this effectively brings the compiler into accord with the runtime (for all newly-compiled Swift code, which will eventually be all Swift code because the current public releases of Swift 5 are not yet considered ABI-stable) and guarantees a zero value for the offset everywhere.
Since the runtime will agree with the compiler about the zero value of the offset, the compiler can continue to emit such offset variables as constant. The exception to this rule is if the class has non-fragile ObjC ancestry, in which case the ObjC runtime (which is not aware of this special rule for empty fields) will attempt to slide it along with everything else.
Fixes rdar://48031465, in which the `FixedClassMetadataBuilder` for a class with a legacy-fixed layout was writing a non-zero offset for an empty field into the field offset vector, causing the runtime to not apply the special case and thus to compute a non-zero offset, which it then attempted to copy into the global field offset variable, which the compiler had emitted as a true-constant zero.
In our initial approach for resolving metadata dependency cycles with classes, non-transitively complete superclass metadata was fetched by the subclass's metadata completion function and passed to `swift_initClassMetadata`. That could mean generating quite a lot of code in the completion function, and so we fairly recently changed it so that `swift_initClassMetadata` instead fetched the superclass metadata via a demangling. Unfortunately, the metadata demangler only fetches _abstract_ metadata by default, and class metadata cannot be considered even non-transitively complete when its superclass reference not at that stage. If the superclass metadata is being completed on one thread, and a subclass is being completed on another, and the subclass installs the incomplete superclass metadata in its superclass field and attempts to register the subclass with the Objective-C runtime, the runtime may crash reading the incompletely-initialized superclass.
The proper fix is to make `swift_initClassMetadata` fetch non-transitively complete metadata for the superclass, delaying completion if that metadata is unavailable. Unfortunately, that can't actually be implemented on top of `swift_initClassMetadata` because that function has no means of reporting an unsatisfied dependency to its caller, and we can no longer simply change its signature without worrying about a small of internal code that might still be using it. We cannot simply perform a blocking metadata request in `swift_initClassMetadata` because it is deeply problematic to block within a metadata completion function. The solution is therefore to add a `swift_initClassMetadata2` which has the ability to report unsatisfied dependencies. That was done in #22386; this patch builds on that by teaching the compiler to generate code to actually use it. It is therefore not safe to use this patch if you might be running on an OS that only provides the old runtime function, but that should be a temporary Apple-internal problem.
Fixes rdar://47549859.
Update the IRGen tests for Windows IRGen. This is an improvement over
what we have currently, but is not sufficient to get all the IRGen tests
passing yet.
We want to be able to define classes with a fixed storage layout,
but a resilient (opaque) vtable. If the class is also generic,
we still have to load field offsets from the metadata if they
are dependent.
So put the field offsets after the generic arguments and before
the vtable.
This is an ABI break for @_fixed_layout classes, which are
defined by the stdlib.
IRGen always just emits a simple implementation that immediately
calls swift_relocateClassMetadata(); so allow the function to be
null in this case to save on code size.
Rather than rely on the metadata initialization function to compute and
fill in the superclass, use the mangled superclass name to construct the
superclass metadata.
Previously we would emit class metadata for classes with resilient
ancestry, and relocate it at runtime once the correct size was known.
However most of the fields were blank, so it makes more sense to
construct the metadata from scratch, and store the few bits that we
do need in a true-const pattern where we can use relative pointers.
They were, already, but remove the isConstant parameter to
getAddrOfTypeMetadataPattern(), and just assert that its true for
patterns in defineTypeMetadata() instead.
Also, metadata patterns are i8*, not i8**. In fact they don't contain any
absolute pointers at all.
Should be NFC other than the LLVM type change.
Similar to the non-resilient case, except we also emit a 'relocation
function'. The class descriptor now contains this relocation function
if the class has resilient ancestry, and the relocation function
calls the runtime's swift_relocateClassMetadata() entry point.
The metadata completion function calls swift_initClassMetadata() and
does layout, just like the non-resilient case.
Fixes <rdar://problem/40810002>.
Now that we don't need the superclass before calling
swift_relocateClassMetadata(), it seems simpler to set it
here instead of doing it in various places in IRGen.
Using the superclass metadata here no longer makes sense with two-phase
init, in case the superclass metadata depends on the class being
instantiated.
It would also be nice to rework the resilient class metadata 'pattern'
to be its own data structure that's true const, instead of just the
prefix of a real class metadata, but for now let's keep the existing
crappy design.
If a class has generic ancestry or resiliently-sized fields, but is
itself not generic and does not have resilient ancestry, we must
perform runtime metadata initialization, but we can initialize
the metadata in-place.
As with generic classes or classes with resilient ancestry, we
copy generic requirements and field offset vectors from the
superclass. We also calculate the layout of the class's direct
fields at runtime.
Unlike the fully resilient case, we don't copy vtable entries
from the superclass, or install the direct class's vtable
entries from the type context descriptor. Instead, we statically
emit the vtable as with fixed-size class metadata.
Both the in-place and resilient case call the same runtime
entry point to initialize class metadata; the new HasStaticVTable
flag in ClassLayoutFlags is used to select between the two
behaviors concerning the vtable.
This saves us some expensive cross-referencing and caching in the runtime, and lets us reclaim the `isReflectable` bit from the context descriptor flags (since a null field descriptor is a suitable and more accurate indicator of whether a type is reflectable).
As part of this, rename TypeMetadataRecordKind to TypeReferenceKind
and consistently give it three bits of storage.
The better modelling of these type references appears to have been
sufficient to make dynamic conformance checks succeed, which is good
but unexpected.
It hass been a longstanding principle in LLVM that the presence of
debug info shall not affect code generation. This patch brings the
Swift frontend closer to this ideal:
- unconditionally emit shadow copies
- unconditionally bind type metadata
The extra allocas, bitcasts, geps, and stores being emitted get
optimized away when compiling at anything but -Onone. There are few
use-cases for compiling at -Onone without -g, so this shouldn't affect
performance for any real-world use-cases.
This adds the dllstorage annotations on the tests. This first pass gets
most of the IRGen tests passing on Windows (though has dependencies on
other changes). However, this allows for the changes to be merged more
easily as we cannot regress other platforms here.
Ensure that we use the correct python to run the python based tools.
This also allows these tools to run on Windows which will not
necessarily associate the python script with an interpreter (python).
Most of the work of this patch is just propagating metadata states
throughout the system, especially local-type-data caching and
metadata-path resolution. It took a few design revisions to get both
DynamicMetadataRequest and MetadataResponse to a shape that felt
right and seemed to make everything easier.
The design is laid out pretty clearly (I hope) in the comments on
DynamicMetadataRequest and MetadataResponse, so I'm not going to
belabor it again here. Instead, I'll list out the work that's still
outstanding:
- I'm sure there are places we're asking for complete metadata where
we could be asking for something weaker.
- I need to actually test the runtime behavior to verify that it's
breaking the cycles it's supposed to, instead of just not regressing
anything else.
- I need to add something to the runtime to actually force all the
generic arguments of a generic type to be complete before reporting
completion. I think we can get away with this for now because all
existing types construct themselves completely on the first request,
but there might be a race condition there if another asks for the
type argument, gets an abstract metadata, and constructs a type with
it without ever needing it to be completed.
- Non-generic resilient types need to be switched over to an IRGen
pattern that supports initialization suspension.
- We should probably space out the MetadataStates so that there's some
space between Abstract and Complete.
- The runtime just calmly sits there, never making progress and
permanently blocking any waiting threads, if you actually form an
unresolvable metadata dependency cycle. It is possible to set up such
a thing in a way that Sema can't diagnose, and we should detect it at
runtime. I've set up some infrastructure so that it should be
straightforward to diagnose this, but I haven't actually implemented
the diagnostic yet.
- It's not clear to me that swift_checkMetadataState is really cheap
enough that it doesn't make sense to use a cache for type-fulfilled
metadata in associated type access functions. Fortunately this is not
ABI-affecting, so we can evaluate it anytime.
- Type layout really seems like a lot of code now that we sometimes
need to call swift_checkMetadataState for generic arguments. Maybe
we can have the runtime do this by marking low bits or something, so
that a TypeLayoutRef is actually either (1) a TypeLayout, (2) a known
layout-complete metadata, or (3) a metadata of unknown state. We could
do that later with a flag, but we'll need to at least future-proof by
allowing the runtime functions to return a MetadataDependency.
Rename it to swift_initClassMetadata() just like we recently did
swift_initStructMetadata(), and add a StructLayoutFlags parameter
so we can version calls to this function in the future.
Maybe at some point this will become a separate ClassLayoutFlags
type, but at this point it doesn't matter because IRGen always
passes a value of 0.
This includes global generic and non-generic global access
functions, protocol associated type access functions,
swift_getGenericMetadata, and generic type completion functions.
The main part of this change is that the functions now need to take
a MetadataRequest and return a MetadataResponse, which is capable
of expressing that the request can fail. The state of the returned
metadata is reported as an second, independent return value; this
allows the caller to easily check the possibility of failure without
having to mask it out from the returned metadata pointer, as well
as allowing it to be easily ignored.
Also, change metadata access functions to use swiftcc to ensure that
this return value is indeed returned in two separate registers.
Also, change protocol associated conformance access functions to use
swiftcc. This isn't really related, but for some reason it snuck in.
Since it's clearly the right thing to do, and since I really didn't
want to retroactively tease that back out from all the rest of the
test changes, I've left it in.
Also, change generic metadata access functions to either pass all
the generic arguments directly or pass them all indirectly. I don't
know how we ended up with the hybrid approach. I needed to change all
the code-generation and calls here anyway in order to pass the request
parameter, and I figured I might as well change the ABI to something
sensible.
The allocation phase is guaranteed to succeed and just puts enough
of the structure together to make things work.
The completion phase does any component metadata lookups that are
necessary (for the superclass, fields, etc.) and performs layout;
it can fail and require restart.
Next up is to support this in the runtime; then we can start the
process of making metadata accessors actually allow incomplete
metadata to be fetched.
