We can encounter these when the compiler modifies an inlinable
function to break apart a struct and the struct uses a private
type for one of its fields. It's questionable whether we /should/
handle this, but meanwhile this /is/ a non-intrusive fix that
preserves the performance of non-resilient libraries.
(That is, it appears this worked in Swift 4.0, though perhaps
not all of the same optimizations kicked in.)
https://bugs.swift.org/browse/SR-6874
@noescape function types will eventually be trivial. A
convert_escape_to_noescape instruction does not take ownership of its
operand. It is a projection to the trivial value carried by the closure
-- both context and implementation function viewed as a trivial value.
A safe SIL program must ensure that the object that the project value is based
on is live beyond the last use of the trivial value. This will be
achieve by means of making the lifetimes dependent.
For example:
%e = partial_apply [callee_guaranteed] %f(%z) : $@convention(thin) (Builtin.Int64) -> ()
%n = convert_escape_to_noescape %e : $@callee_guaranteed () -> () to $@noescape @callee_guaranteed () -> ()
%n2 = mark_dependence %n : $@noescape @callee_guaranteed () -> () on %e : $@callee_guaranteed () -> ()
%f2 = function_ref @use : $@convention(thin) (@noescape @callee_guaranteed () -> ()) -> ()
apply %f2(%n2) : $@convention(thin) (@noescape @callee_guaranteed () -> ()) -> ()
release_value %e : $@callee_guaranteed () -> ()
Note: This is not yet actually used.
Part of:
SR-5441
rdar://36116691
* Implement #warning and #error
* Fix #warning/#error in switch statements
* Fix AST printing for #warning/#error
* Add to test case
* Add extra handling to ParseDeclPoundDiagnostic
* fix dumping
* Consume the right paren even in the failure case
* Diagnose extra tokens on the same line after a diagnostic directive
Also remove the decl from the known decls and remove a
bunch of code referencing that decl as well as a bunch of other
random things including deserialization support.
This includes removing some specialized diagnostics code that
matched the identifier ImplicitlyUnwrappedOptional, and tweaking
diagnostics for various modes and various issues.
Fixes most of rdar://problem/37121121, among other things.
...rather than the ad hoc CustomTypeNameManglingAttr I was using
before. As John pointed out, the AST should be semantic wherever
possible.
We may someday want to get out of this being an attribute altogether,
or duplicating information that's available in the original Clang
node, by actually storing a reference to that node somewhere. This is
tricky and mixed up with deciding what hasClangNode() or
getClangDecl() would mean, though, so for now the attribute just
carries the information we need.
When importing a C enum with the ns_error_domain attribute, we
synthesize a struct containing an NSError object to represent errors
in that domain. That synthesized struct should have a mangled name
that ties it to the original C enum, if we want it to be stable, and
now it does.
Before: $SSC7MyErrorV (a normal struct, which is a lie)
After: $SSC11MyErrorCode13ns_error_enumLLV
kind=Global
kind=Structure
kind=Module, text="__C_Synthesized"
kind=PrivateDeclName
kind=Identifier, text="ns_error_enum"
kind=Identifier, text="MyErrorCode"
Using the "private discriminator" feature allows us to pack in extra
information about the declaration without changing the mangling
grammar, and without stepping on anything the importer is using.
More rdar://problem/24688918
- Clear the 'serialized' flag on witness tables and vtables
after serialization, not just functions. This fixes SIL
verifier failures if post-serialization SIL is printed
out and parsed back in.
- Clear the 'serialized' flag when deserializing functions,
witness tables and vtables in a module that has already
been serialized. This fixes SIL verifier failures if
we deserialize more declarations after serializing SIL.
We were seeing SIL verifier failures on bots that run the
tests with the stdlib built with non-standard flags.
Unfortunately I don't have a reduced test case that would
fail in PR testing without these fixes.
Fixes <rdar://problem/36682929>.
Consider a module with two files, one of which references a
function in the other file. If the function had PublicNonABI
linkage, the forward reference has HiddenExternal linkage.
If we saw the forward reference first in the merge modules
process, we would not update the linkage of the function
when we later deserialized its body.
This would result in it being dropped from the final merged
SIL module. Previously this did not cause any problems
because all inlineable functions had public linkage, but now
that default argument generators no longer have public entry
points, this could cause linker errors.
For the majority of artificial helper functions the filename is
actively misleading since it usually represents the file of the caller
that triggered the helper to be generated. Instead, this patch creates
a virtual filname `<compiler-generated>` to make it very obvious that
the function has not correspondence to any source code.
<rdar://problem/33809560>
This is going to be used for "always emit into client" functions,
such as default argument generators and stored property
initializers.
- In dead function elimination, these functions behave identically to
public functions, serving as "anchors" for the mark-and-sweep
analysis.
- There is no external variant of this linkage, because external
declarations can use HiddenExternal linkage -- the definition should
always be emitted by another translation unit in the same Swift
module.
- When deserialized, they receive shared linkage, because we want the
linker to coalesce multiple copies of the same deserialized
definition if it was deserialized from multiple translation units
in the same Swift module.
- When IRGen emits a definition with this linkage, it receives the
same LLVM-level linkage as a hidden definition, ensuring it does not
have a public entry point.
This has three principal advantages:
- It gives some additional type-safety when working
with known accessors.
- It makes it significantly easier to test whether a declaration
is an accessor and encourages the use of a common idiom.
- It saves a small amount of memory in both FuncDecl and its
serialized form.
* Reduce array abstraction on apple platforms dealing with literals
Part of the ongoing quest to reduce swift array literal abstraction
penalties: make the SIL optimizer able to eliminate bridging overhead
when dealing with array literals.
Introduce a new classify_bridge_object SIL instruction to handle the
logic of extracting platform specific bits from a Builtin.BridgeObject
value that indicate whether it contains a ObjC tagged pointer object,
or a normal ObjC object. This allows the SIL optimizer to eliminate
these, which allows constant folding a ton of code. On the example
added to test/SILOptimizer/static_arrays.swift, this results in 4x
less SIL code, and also leads to a lot more commonality between linux
and apple platform codegen when passing an array literal.
This also introduces a couple of SIL combines for patterns that occur
in the array literal passing case.
When loading the named members for a given name, we want to load all
of the members with that base name... not only the ones that match the
full name, because the lookup table is indexed by base name and
filtering too early drops candidates.
Fixes rdar://problem/36085994.
The deserialization of the type witnesses for a normal protocol conformance
is crucial to the usability of said conformance. Deserializing the
value witnesses first can fail if they somehow rely on the type
witnesses (e.g., through a recursive conformance).
As a stop-gap, deserialize and record type witnesses *first*, then
deserialize value witnesses afterward. A longer-term solution would
make deserialization of the normal protocol conformance far more
lazy.
Fixes SR-6522 / rdar://problem/35830641, a merge-modules crasher in a
nontrivial project.
- Outlaw duplicate input files, fix driver, fix tests, and add test.
- Reflect that no buffer is present without a (possibly pseudo) named file.
- Reflect fact that every input has a (possible pseudo) name.
- Break up CompilerInstance::setup.
Don't bail on dups.
Follow-up to ac6fd7214 that adds a similar feature for enums. This
probably doesn't come up much, but the infrastructure's already there.
(The previous commit showed that the other places we do recovery
already handle this from their normal collection of "dependency
types".)
Refactoring suggested by Slava on ac6fd7214. In cases where multiple
requirements mention the same nominal type, this will also result in
doing slightly less work because we'll only record it once. Other than
that, no user-visible change expected.
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.
We could handle a typealias itself disappearing, but not if the
typealias was okay but the underlying type wasn't. This came up in
real Swift 3/4 mix-and-match code.
rdar://problem/34940079
Rather than storing contextual types in the type witnesses and associated
conformances of NormalProtocolConformance, store only interface types.
@huonw did most of the work here, and @DougGregor patched things up to
complete the change.
If we can't resolve a cross-reference unambiguously, we're supposed to
produce an llvm::Error and let the calling code handle it. However, if
we couldn't even resolve the /type/ of the cross-reference, we would
just crash. Follow the supported error path in that case too -- in
many cases the error can just propagate upwards to something that can
handle it.
rdar://problem/34821187, plus an extra test case from
rdar://problem/35157494. (The latter will be fixed better later, but
meanwhile let's not regress on the crashing part.)
For now these are underscored attributes, i.e. compiler internal attributes:
@_optimize(speed)
@_optimize(size)
@_optimize(none)
Those attributes override the command-line specified optimization mode for a specific function.
The @_optimize(none) attribute is equivalent to the already existing @_semantics("optimize.sil.never") attribute
