Nominal type descriptors are not always unique, so testing them via pointer
equality is not correct. Introduce an "isEqual()" operation for
nominal type descriptors that performs the appropriate equality check,
using pointer equality when possible, and falling back to string
comparisons of the mangled type name when it is not possible.
Introduce a "nonunique" flag into nominal type descriptors to describe
when they are, in fact, not unique. The only nonunique nominal type
descriptors currently come from Clang-imported types; all
Swift-defined types have unique nominal type descriptors. Use this
flag to make the aforementioned operation efficient in the "unique"
case.
Use the new isEqual() operation for protocol conformance lookup, and
make sure we're caching results based on the known-canonical nominal
type descriptor.
When emitting foreign class metadata (e.g., for an imported CF type), fill
in the superclass when we have one. The superclass will itself be a foreign
metadata candidate, so also register an initialization function that uniques
the superclass metadata once we've picked the canonical foreign class
metadata.
If a generic class has non-resilient ancestry, we cannot use
the template size so far as the constant value of the base offset
variable, because the template will not contain entries for all
metadata members; instead, use metadata layout as with everything
else.
Swift class metadata has a bit to distinguish it from non-Swift Objective-C
classes. The stable ABI will use a different bit so that stable Swift and
pre-stable Swift can be distinguished from each other.
No bits are actually changed yet. Enabling the new bit needs to wait for
other coordination such as libobjc.
rdar://35767811
- Create the value witness table as a separate global object instead
of concatenating it to the metadata pattern.
- Always pass the metadata to the runtime and let the runtime handle
instantiating or modifying the value witness table.
- Pass the right layout algorithm version to the runtime; currently
this is always "Swift 5".
- Create a runtime function to instantiate single-case enums.
Among other things, this makes the copying of the VWT, and any
modifications of it, explicit and in the runtime, which is more
future-proof.
We can reduce the uniquing header from 3–4 pointer-sized words down to 1–2 32-bit words + one pointer:
- The initialization function (when present) and name are always emitted into the same binary image, so we can use relative references to shrink these down to 32-bit fields.
- We don't ever simultaneously need the initialization flags and the initialized uniqued pointer. (Keeping the "initialization function" flag bit theoretically lets us turn a "consume" load into a "relaxed" load, but that makes no practical difference on most contemporary architectures.) 12 flag bits Ought To Be Enough For Anyone and lets us reliably tell a valid pointer from a flag set, so overlap the initialization flags with the eventual invasive cache value.
The invasive cache is left inline, since we've decided we're not going to make the rest of type metadata records ever be true-const, so they'll already be sitting on a dirty page. A dynamic linker that was sufficiently Swift-optimized to precalculate the other load-time-initialized entries in metadata could likely precompute the invasive cache value as well.
rdar://problem/22527141
It's not readnone, because it reads the metatype from an object.
Readnone would let the llvm ARC optimizer reschedule the call with a release-call for the object.
fixes SR-6560.
Don't emit placeholders for field offsets and vtable entries,
since they were always null. Instead, calculate the final size
of class metadata at runtime using the size of the superclass
metadata and the number of immediate members, and only copy
this prefix from the template to the instantiated metadata,
zero-filling the rest.
For this to work with non-generic resilient classes and
non-generic subclasses of generic classes, we need a new
runtime entry point to relocate non-generic class metadata,
calculating its size at runtime using the same strategy.
Instead of recording the current offset in the template, just
collect the fill ops into a vector and emit them at the
generic argument offset stored in the MetadataLayout.
For now this is NFC, but this allows the template to omit
entries for generic arguments entirely once some more
machinery is in place.
Adjust the IRGen for ObjC interop to ensure that the section that
metadata is emitted into the correct section for non-MachO targets.
This also adds a more comprehensive test for ensuring that the IRGen can
now be tested on all targets. Since the ObjC interop is now
controllable via the driver, this test does not require that the
objc_interop feature is present as it is a IRGen test.
This is the first step to remove the `REQUIRES: objc_interop` from the
IRGen tests.
Following up on the fixes for rdar://problem/35330067. If a class inherits a class from another module with missing fields, we need to treat its size and alignment as opaque, just like the base class itself. We also need to lay out such class at runtime, since we don't know the size, alignment, or field offsets at compile time; relying on the ObjC runtime alone will slide the ivar offsets, but not the Swift instance size and alignment. Fixes rdar://problem/35747485.
We would miscompile in mixed-language-version projects when a Swift class was compiled for one language version, while using Objective-C-imported types that are only available to that version, and then imported into a Swift module with a different language version that wasn't able to see all of the properties because of incompatible imported types. This manifested in a number of ways:
- We assumed we could re-derive the constant field offsets of the class's ivars from the layout, which is wrong if properties are missing, causing accesses to final properties or subclass properties to go to the wrong offsets.
- We assumed we could re-derive the instance size and alignment of a class instance in total, causing code to allocate the wrong amount of memory.
- We neglected to account for the space that stored properties take up in the field offset vector of the class object, causing us to load vtable entries for following subclass methods from the wrong offsets.
Eventually, resilience should reduce our exposure to these kinds of problems. As an incremental step in the right direction, when we look at a class from another module in IRGen, treat it as always variably-sized, so we don't try to hardcode offsets, size, or alignment of its instances. When we import a class, and we're unable to import a stored property, leave behind a new kind of MissingMemberDecl that records the number of field offset vector slots it will take up, so that we lay out subclass objects and compute vtable offsets correctly. Fixes rdar://problem/35330067.
A side effect of this is that the RemoteAST library is no longer able to provide fixed field offsets for class ivars. This doesn't appear to impact the lldb test suite, and they will ultimately need to use more abstract access patterns to get ivar offsets from resilient classes (if they aren't already), so I just removed the RemoteAST test cases that tested for class field offsets for now.
This commit is mostly refactoring.
*) Introduce a new OptimizationMode enum and use that in SILOptions and IRGenOptions
*) Allow the optimization mode also be specified for specific SILFunctions. This is not used in this commit yet and thus still a NFC.
Also, fixes a minor bug: we didn’t run mandatory IRGen passes for functions with @_semantics("optimize.sil.never")
Switch most general endpoint to be `flags, parameters, parameterFlags, result`,
instead of opaque `void **`, more specialized ones to use follow argument scheme:
`flags, param0, [flags0], ..., paramN, [flagsN], result` and store parameter/flags
information separately in `FunctionCacheEntry::{Key, Data}` as well.
Currently only single 'inout' flag has been encoded into function
metadata, these changes extend function metadata to support up to
32 flags per parameter.
