This reverts commit 2262bd579a.
This information isn't necessary for field descriptor lookup,
after all. It's only the fields that need to have generic information,
which is already in the field descriptor.
Previously, the mangling didn't include generics, but these are
needed to key off of the new field descriptor metadata, as well
as to construct type references for the nominal type.
The size of a protocol's metadata was not a multiple of 8 bytes, so
on 64-bit platforms, the runtime would copy default witnesses from
the wrong address, because IRGen metadata does not add alignment padding,
whereas the in-memory structure does.
Fix this by adding a 32-bit padding field at the end of the protocol
descriptor. Technically this is not necessary on 32-bit, but this keeps
things simpler for now.
The test case for this is a library evolution test exercising resilient
default protocol requirements, but it is not quite ready to go in yet.
Previously SILDefaultWitnessTables only included "resilient" default
implementations, which are currently defined as those that appear at the
end of a protocol, after any requirements without defaults.
However, this was too inflexible. Instead, include all entries in the
SILDefaultWitnessTable, with invalid entries standing in for requirements
without defaults.
Previously, the minimum witness table size was a separate parameter, also
appearing in SIL syntax; now it can be calculated by looking at the entries
themselves. The getResilientDefaultEntries() method of SILDefaultWitnessTable
returns the same result as getEntries() did previously.
"minimal" is defined as the set of requirements that would be
passed to a function with the type's generic signature that
takes the thick metadata of the parent type as its only argument.
Each runtime function definition in RuntimeFunctions.def states which calling convention
should be used for this runtime function. But IRGen and LLVMPasses were not always
properly propagating this declared calling convention all the way down to llvm's Call instructions.
In many cases, the standard C convention was set for the call irrespective of the actual calling
convention defined for a given runtime function. As a result, incorrect code was generated.
This commit tries to fix all those places, where such a mismatch was found. In many cases this is
achieved by defining a helper function CreateCall in such a way that makes sure that the call instruction
gets the same calling convention as the one used by its callee operand.
from the witness tables for their associations rather than passing
them separately.
This drastically reduces the number of physical arguments required
to invoke a generic function with a complex protocol hierarchy. It's
also an important step towards allowing recursive protocol
constraints. However, it may cause some performance problems in
generic code that we'll have to figure out ways to remediate.
There are still a few places in IRGen that rely on recursive eager
expansion of associated types and protocol witnesses. For example,
passing generic arguments requires us to map from a dependent type
back to an index into the all-dependent-types list in order to
find the right Substitution; that's something we'll need to fix
more generally. Specific to IRGen, there are still a few abstractions
like NecessaryBindings that use recursive expansion and are therefore
probably extremely expensive under this patch; I intend to fix those
up in follow-ups to the greatest extent possible.
There are also still a few things that could be made lazier about
type fulfillment; for example, we eagerly project the dynamic type
metadata of class parameters rather than waiting for the first place
we actually need to do so. We should be able to be lazier about
that, at least when the parameter is @guaranteed.
Technical notes follow. Most of the basic infrastructure I set up
for this over the last few months stood up, although there were
some unanticipated complexities:
The first is that the all-dependent-types list still does not
reliably contain all the dependent types in the minimized signature,
even with my last patch, because the primary type parameters aren't
necessarily representatives. It is, unfortunately, important to
give the witness marker to the primary type parameter because
otherwise substitution won't be able to replace that parameter at all.
There are better representations for all of that, but it's not
something I wanted to condition this patch on; therefore, we have to
do a significantly more expensive check in order to figure out a
dependent type's index in the all-dependent-types list.
The second is that the ability to add requirements to associated
types in protocol refinements means that we have to find the *right*
associatedtype declaration in order to find the associated witness
table. There seems to be relatively poor AST support for this
operation; maybe I just missed it.
The third complexity (so far) is that the association between an
archetype and its parent isn't particularly more important than
any other association it has. We need to be able to recover
witness tables linked with *all* of the associations that lead
to an archetype. This is, again, not particularly well-supported
by the AST, and we may run into problems here when we eliminate
recursive associated type expansion in signatures.
Finally, it's a known fault that this potentially leaves debug
info in a bit of a mess, since we won't have any informaton for
a type parameter unless we actually needed it somewhere.
Similarly to how we've always handled parameter types, we
now recursively expand tuples in result types and separately
determine a result convention for each result.
The most important code-generation change here is that
indirect results are now returned separately from each
other and from any direct results. It is generally far
better, when receiving an indirect result, to receive it
as an independent result; the caller is much more likely
to be able to directly receive the result in the address
they want to initialize, rather than having to receive it
in temporary memory and then copy parts of it into the
target.
The most important conceptual change here that clients and
producers of SIL must be aware of is the new distinction
between a SILFunctionType's *parameters* and its *argument
list*. The former is just the formal parameters, derived
purely from the parameter types of the original function;
indirect results are no longer in this list. The latter
includes the indirect result arguments; as always, all
the indirect results strictly precede the parameters.
Apply instructions and entry block arguments follow the
argument list, not the parameter list.
A relatively minor change is that there can now be multiple
direct results, each with its own result convention.
This is a minor change because I've chosen to leave
return instructions as taking a single operand and
apply instructions as producing a single result; when
the type describes multiple results, they are implicitly
bound up in a tuple. It might make sense to split these
up and allow e.g. return instructions to take a list
of operands; however, it's not clear what to do on the
caller side, and this would be a major change that can
be separated out from this already over-large patch.
Unsurprisingly, the most invasive changes here are in
SILGen; this requires substantial reworking of both call
emission and reabstraction. It also proved important
to switch several SILGen operations over to work with
RValue instead of ManagedValue, since otherwise they
would be forced to spuriously "implode" buffers.
This function points the ivar offsets at per-metadata field offsets
before passing off the class to the Objective-C runtime, ensuring
we don't slide the global ivar offsets multiple times.
The swift_initializeSuperclass() function does not do this ivar
cloning, so it cannot be used for generic classes even if they do
not have generically-sized fields.
An alternative would be to refactor the runtime to clone ivar offset
vectors in swift_initializeSuperclass(), but this doesn't seem to be
worth it for now.
IRGen now uses a ConstantBuilder to build protocol metadata, which may
now have additional fields at the end for default witnesses.
For now, the default implementations in the test have to external
because IRGen cannot emit a witness_method body where Self is
abstract. I will fix this by passing in the witness table as part
of the witness_method calling convention.
On the IRGen side, other than the calling convention change, the only
remaining piece here is emitting GenericWitnessTables and accessor
functions for conformances where the conformance is defined in
a different module than the protocol, and the protocol is resilient.
Sema still needs to infer default witnesses and store them in the
ProtocolDecl, so that SILGen can emit default witness thunks for
them.
- Implement emission of type references for nominal type field
reflection, using a small custom encoder resulting in packed
structs, not strings. This will let us embed 7-bit encoded
32-bit relative offsets directly in the structure (not yet
hooked in).
- Use the AST Mangler for encoding type references
Archetypes and internal references were complicating this before, so we
can take the opportunity to reuse this machinery and avoid unique code
and new ABI.
Next up: Tests for reading the reflection sections and converting the
demangle tree into a tree of type references.
Todo: For concrete types, serialize the types for associated types of
their conformances to bootstrap the typeref substitution process.
rdar://problem/15617914
This comes with a fix for a null pointer dereference in _typeByName()
that would pop with foreign classes that do not have a
NominalTypeDescriptor.
Also, I decided to back out part of the change for now, where the
NominalTypeDescriptor references an accessor function instead of a
pattern, since this broke LLDB, which reaches into the pattern to
get the generic cache.
Soon we will split off the generic cache from the pattern, and at
that time we can change the NominalTypeDescriptor to point at the
cache. But for now, let's avoid needless churn in LLDB by keeping
that part of the setup unchanged.
Change conformance records to reference NominalTypeDescriptors instead of
metadata patterns for resilient or generic types.
For a resilient type, we don't know if the metadata is constant or not,
so we can't directly reference either constant metadata or the metadata
template.
Also, whereas previously NominalTypeDescriptors would point to the
metadata pattern, they now point to the metadata accessor function.
This allows the recently-added logic for instantiating concrete types
by name to continue working.
In turn, swift_initClassMetadata_UniversalStrategy() would reach into
the NominalTypeDescriptor to get the pattern out, so that its bump
allocator could be used to allocate ivar tables. Since the pattern is
no longer available this way, we have to pass it in as a parameter.
In the future, we will split off the read-write metadata cache entry
from the pattern; then swift_initClassMetadata_UniversalStrategy() can
just take a pointer to that, since it doesn't actually need anything
else from the pattern.
Since Clang doesn't guarantee alignment for function pointers, I had
to kill the cute trick that packed the NominalTypeKind into the low
bits of the relative pointer to the pattern; instead the kind is now
stored out of line. We could fix this by packing it with some other
field, or keep it this way in case we add new flags later.
Now that generic metadata is instantiated by calling accessor functions,
this change removes the last remaining place that metadata patterns were
referenced from outside the module they were defined in. Now, the layout
of the metadata pattern and the behavior of swift_getGenericMetadata()
is purely an implementation detail of generic metadata accessors.
This patch allows two previously-XFAIL'd tests to pass.
Instead of directly emitting calls to swift_getGenericMetadata*() and
referencing metadata templates, call a metadata accessor function
corresponding to the UnboundGenericType of the NominalTypeDecl.
The body of this accessor forwards arguments to a runtime metadata
instantiation function, together with the template.
Also, move some code around, so that metadata accesses which are
only done as part of the body of a metadata accessor function are
handled separately in emitTypeMetadataAccessFunction().
Apart from protocol conformances, this means metadata templates are
no longer referenced from outside the module where they were defined.
This is a preliminary refactoring toward emitting generic metadata
instantiation accessors. This will let us stop exporting metadata
template symbols and reference accessors from conformance tables
instead. The latter is required to enable conformances where the
conforming type is resilient.
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.
