Decrease the size of nominal type descriptors and make them true-const by relative-addressing the other metadata they need to reference, which should all be included in the same image as the descriptor itself. Relative-referencing string constants exposes a bug in the Apple linker, which crashes when resolving relative relocations to coalesceable symbols (rdar://problem/22674524); work around this for now by revoking the `unnamed_addr`-ness of string constants that we take relative references to. (I haven't tested whether GNU ld or gold also have this problem on Linux; it may be possible to conditionalize the workaround to only apply to Darwin targets for now.)
Since that's somewhat expensive, allow the generation of meaningful
IR value names to be efficiently controlled in IRGen. By default,
enable meaningful value names only when generating .ll output.
I considered giving protocol witness tables the name T:Protocol
instead of T.Protocol, but decided that I didn't want to update that
many test cases.
Most notably, the source caches did not respect dominance. The
simplest solution was just to drop them in favor of the ordinary
caching system; this is unfortunate because it requires walking
over the path twice instead of exploiting the trie, but it's much
easier to make this work, especially in combination with the other
caching mechanisms at play.
This will be tested by later commits that enable lazy-loading of
local type data in various contexts.
There are several interesting new features here.
The first is that, when emitting a SILFunction, we're now able to
cache type data according to the full dominance structure of the
original function. For example, if we ask for type metadata, and
we've already computed it in a dominating position, we're now able
to re-use that value; previously, we were limited to only doing this
if the value was from the entry block or the LLVM basic block
matched exactly. Since this tracks the SIL dominance relationship,
things in IRGen which add their own control flow must be careful
to suppress caching within blocks that may not dominate the
fallthrough; this mechanism is currently very crude, but could be
made to allow a limited amount of caching within the
conditionally-executed blocks.
This query is done using a proper dominator tree analysis, even at -O0.
I do not expect that we will frequently need to actually build the
tree, and I expect that the code-size benefits of doing a real
analysis will be significant, especially as we move towards making
more metadata lazily computed.
The second feature is that this adds support for "abstract"
cache entries, which indicate that we know how to derive the metadata
but haven't actually done so. This code isn't yet tested, but
it's going to be the basis of making a lot of things much lazier.
of associated types in protocol witness tables.
We use the global access functions when the result isn't
dependent, and a simple accessor when the result can be cheaply
recovered from the conforming metadata. Otherwise, we add a
cache slot to a private section of the witness table, forcing
an instantiation per conformance. Like generic type metadata,
concrete instantiations of generic conformances are memoized.
There's a fair amount of code in this patch that can't be
dynamically tested at the moment because of the widespread
reliance on recursive expansion of archetypes / dependent
types. That's something we're now theoretically in a position
to change, and as we do so, we'll test more of this code.
This speculatively re-applies 7576a91009,
i.e. reverts commit 11ab3d537f.
We have not been able to duplicate the build failure in
independent testing; it might have been spurious or unrelated.
This should cover most temporary buffers, except for those used by indirected value arguments, which need some cooperation with CallEmission to properly mark lifetime end after the call's completed.
of associated types in protocol witness tables.
We use the global access functions when the result isn't
dependent, and a simple accessor when the result can be cheaply
recovered from the conforming metadata. Otherwise, we add a
cache slot to a private section of the witness table, forcing
an instantiation per conformance. Like generic type metadata,
concrete instantiations of generic conformances are memoized.
There's a fair amount of code in this patch that can't be
dynamically tested at the moment because of the widespread
reliance on recursive expansion of archetypes / dependent
types. That's something we're now theoretically in a position
to change, and as we do so, we'll test more of this code.
This reverts commit 6528ec2887, i.e.
it reapplies b1e3120a28, with a fix
to unbreak release builds.
Now that all the machinery is in place, the ClassMetadataBuilder
can (more accurately) query the ClassLayout instead of trying to
re-derive whether the field offset vector is dependent, etc.
Apart from performing dynamic layout for resiliently-sized fields
in concrete classes, this also lets us *skip* dynamic layout
if we have a generic class without any dependent fields.
I haven't tested subclassing with resilient field layout yet, but
getting that working is the next step and should not be too much
work.
Also, swift_initClassMetadata_UniversalStrategy() only stores
the computed field offsets in the field offset globals when the
Objective-C runtime is available, because it gets the offset
pointers from the Objective-C class rodata. On Linux, we will
need to emit code to copy from the field offset vector into
field offset globals in IRGen. This is pretty easy, but I'll do
it in a follow-up patch so for now the new execution test is
XFAIL'd on Linux.
It is questionable if emitting globals for constant offsets is of
any use; the only scenario where one could not compute the offset
directly in IRGen is if the class has a fixed-size layout inside
its own resilience domain, but has a non-fixed layout outside,
*and* we're doing direct stored property access.
Right now we don't plan on doing this, but it might come up in
the future with more exotic resilience domain configurations or
perhaps optimizations.
The runtime entry doesn't just report the error, unlike the other report* functions, it also does the crashing.
Reapplying independent of unrelated reverted patches.
This reverts commit b1e3120a28.
Reverting because this patch uses WitnessTableBuilder::PI in NDEBUG code.
That field only exists when NDEBUG is not defined, but now NextCacheIndex, a
field that exists regardless, is being updated based on information from PI.
This problem means that Release builds do not work.
ClassLayoutBuilder computes a bunch of stuff in addition to the
StructLayout, which is then stashed in ClassTypeInfo. Extract this
into a new ClassLayout type. It probably should not exist at all,
if we only generalized StructLayout a bit.
of associated types in protocol witness tables.
We use the global access functions when the result isn't
dependent, and a simple accessor when the result can be cheaply
recovered from the conforming metadata. Otherwise, we add a
cache slot to a private section of the witness table, forcing
an instantiation per conformance. Like generic type metadata,
concrete instantiations of generic conformances are memoized.
There's a fair amount of code in this patch that can't be
dynamically tested at the moment because of the widespread
reliance on recursive expansion of archetypes / dependent
types. That's something we're now theoretically in a position
to change, and as we do so, we'll test more of this code.
Now, such classes will emit a metadata pattern and use the
generic metadata instantiation logic.
This was all wired up to handle the case of no generic
parameters previously, to support resilient struct layout
in the runtime.
The swift_initializeSuperclass() entry point still exists,
providing a fast path for when there's no field layout to
do, which is currently always true if we have a concrete
class.
This entry point no longer needs the global lock, since
now we get a per-class lock from the metadata cache.
Also, previously we would call the superclass accessor
function on every access of class metadata for a concrete
subclass of a generic class. Now that we re-use the
existing metadata cache logic, this extra call only occurs
during initialization.
Both swift_initializeSuperclass() and
swift_initClassMetadata_UniversalStrategy() used to take
the superclass as a parameter, but this isn't really
necessary, since it was loaded out of the class metadata
immediately prior to the call by the caller. Removing
this parameter makes the ABI a little simpler.
Once class layout supports resilient types, we will also
use swift_initClassMetadata_UniversalStrategy() to lay
out classes with resilient types as fields.
Singleton metadata caches will still allocate a copy of
the template, which is a slight performance regression
from the previous implementation of concrete subclasses
of generic classes. This will be optimized soon.
Right now, the template can always be modified in place;
in the future, it will be possible to modify in place as
long as the superclass is fixed-layout; a resilient superclass
might add or remove fields, thus we cannot leave room for
it in the metadata of the subclass, and will need to grow
the metadata and slide field offsets at runtime using a
new entry point.
Also, the representation of the cache itself could be
optimized to handle the singleton case, since all we
really need here is a lock without any kind of mapping
table.
Move the following from IRGen to runtime:
- Copying generic parameters from superclass to subclass
- Copying field offsets from superclass to subclass
- Initializing the Objective-C runtime name of the subclass
This eliminates some duplication between the generic subclass and
concrete subclass of a generic class cases.
Also this should reduce generated code size and have no impact on
performance (the instantiation logic only runs once per substituted
type).
class B<T> : NSFoo {}
class A : B<Int> {}
IRGen computes the ivar layout starting from offset zero, since
the size of the 'NSFoo' is unknown and we rely on the Objective-C
runtime to slide the ivar offsets.
The instantiated metadata for B<Int> would contain a field offset
vector with the correct offsets, because of how
swift_initClassMetadata_UniversalStrategy() works.
However, A's metadata is emitted statically, and this includes a
copy of the field offset vector from the superclass. A's metadata
was initialized by swift_initializeSuperclass(), which did not
copy the field offset vector over from A<Int>. And since the
Objective-C runtime only slides the immediate ivars of a class,
the field offsets corresponding to A<Int>'s fields in B's type
metadata were never slid, resulting in problems when an instance
of B was passed to a function operating on an A<T> generically.
Fixes <rdar://problem/23200051>.
If an enum is fixed-layout in our resilience domain but not
universally fixed-layout, other resilience domains will use
runtime functions to project and inject payloads.
These expect to find the payload size in the metadata, so
emit it if the enum is not universally fixed-layout.
Note that we do know the payload size, so it is constant
for us; there's no runtime call required to initialize
the metadata.
We've already got -enable-resilience and @_fixed_layout to compare
IRGen differences when lowering super_method instructions, so nuke
the -force-resilient-super-dispatch testing flag. While
we're at it, consolidate the super tests a bit.
We only need to indirectly load the superclass's metadata when the
superclass is resilient. For example, we may be calling a super method
inside an extension of a class we don't own. Otherwise, we can
immediately load the statically known superclass.
Adds a testing flag to force resilient super dispatch, since the
IRGen module resilience checks aren't fully implemented yet.
SIL -> IR tests added.
rdar://problem/22749732
Also fixes rdar://problem/23884670
Decrease the size of nominal type descriptors and make them true-const by relative-addressing the other metadata they need to reference, which should all be included in the same image as the descriptor itself. Relative-referencing string constants exposes a bug in the Apple linker, which crashes when resolving relative relocations to coalesceable symbols (rdar://problem/22674524); work around this for now by revoking the `unnamed_addr`-ness of string constants that we take relative references to. (I haven't tested whether GNU ld or gold also have this problem on Linux; it may be possible to conditionalize the workaround to only apply to Darwin targets for now.)
