This is currently disabled by default. Building the client library can be enabled with the CMake option SWIFT_BUILD_CLIENT_RETAIN_RELEASE, and using the library can be enabled with the flags -Xfrontend -enable-client-retain-release.
To improve retain/release performance, we build a static library containing optimized implementations of the fast paths of swift_retain, swift_release, and the corresponding bridgeObject functions. This avoids going through a stub to make a cross-library call.
IRGen gains awareness of these new functions and emits calls to them when the functionality is enabled and the target supports them. Two options are added to force use of them on or off: -enable-client-retain-release and -disable-client-retain-release. When enabled, the compiler auto-links the static library containing the implementations.
The new calls also use LLVM's preserve_most calling convention. Since retain/release doesn't need a large number of scratch registers, this is mostly harmless for the implementation, while allowing callers to improve code size and performance by spilling fewer registers around refcounting calls. (Experiments with an even more aggressive calling convention preserving x2 and up showed an insignificant savings in code size, so preserve_most seems to be a good middle ground.)
Since the implementations are embedded into client binaries, any change in the runtime's refcounting implementation needs to stay compatible with this new fast path implementation. This is ensured by having the implementation use a runtime-provided mask to check whether it can proceed into its fast path. The mask is provided as the address of the absolute symbol _swift_retainRelease_slowpath_mask_v1. If that mask ANDed with the object's current refcount field is non-zero, then we take the slow path. A future runtime that changes the refcounting implementation can adjust this mask to match, or set the mask to all 1s to disable the old embedded fast path entirely (as long as the new representation never uses 0 as a valid refcount field value).
As part of this work, the overall approach for bridgeObjectRetain is changed slightly. Previously, it would mask off the spare bits from the native pointer and then call through to swift_retain. This either lost the spare bits in the return value (when tail calling swift_retain) which is problematic since it's supposed to return its parameter, or it required pushing a stack frame which is inefficient. Now, swift_retain takes on the responsibility of masking off spare bits from the parameter and preserving them in the return value. This is a trivial addition to the fast path (just a quick mask and an extra register for saving the original value) and makes bridgeObjectRetain quite a bit more efficient when implemented correctly to return the exact value it was passed.
The runtime's implementations of swift_retain/release are now also marked as preserve_most so that they can be tail called from the client library. preserve_most is compatible with callers expecting the standard calling convention so this doesn't break any existing clients. Some ugly tricks were needed to prevent the compiler from creating unnecessary stack frames with the new calling convention. Avert your eyes.
To allow back deployment, the runtime now has aliases for these functions called swift_retain_preservemost and swift_release_preservemost. The client library brings weak definitions of these functions that save the extra registers and call through to swift_retain/release. This allows them to work correctly on older runtimes, with a small performance penalty, while still running at full speed on runtimes that have the new preservemost symbols.
Although this is only supported on Darwin at the moment, it shouldn't be too much work to adapt it to other ARM64 targets. We need to ensure the assembly plays nice with the other platforms' assemblers, and make sure the implementation is correct for the non-ObjC-interop case.
rdar://122595871
This is currently not wired up to anything. I am going to wire it up in
subsequent commits.
The reason why we are introducing this new Builtin type is to represent that we
are going to start stealing bits from the protocol witness table pointer of the
Optional<any Actor> that this type is bitwise compatible with. The type will
ensure that this value is only used in places where we know that it will be
properly masked out giving us certainty that this value will not be used in any
manner without it first being bit cleared and transformed back to Optional<any
Actor>.
Lazy emission of SIL global variables caused us to go through the
creation of two different LLVM struct types for the same
initialization, tripping an assertion in LLVM. Cache it along with
other information about the static-initialized object rather than
rebuilding it.
Also fix the lazy_globals test to account for the laziness,
generalizing it to also run on arm64 so I won't miss it locally.
Delay the emission of SIL global variables that aren't externally
visible until they are actually used. This is the same lazy emission
approach that we use for a number of other entities, such as SIL
functions.
Part of rdar://158363967.
functions to compute them directly without a TypeLowering object, and
change a lot of getTypeLowering call sites to just use that.
There is one subtle change here that I think is okay: SILBuilder used to
use different TypeExpansionContexts when inserting into a global:
- getTypeLowering() always used a minimal context when inserting into
a global
- getTypeExpansionContext() always returned a maximal context for the
module scope
The latter seems more correct, as AFAIK global initializers are never
inlinable. If they are, we probably need to configure the builder with
an actual context properly rather than making global assumptions.
This is incremental progress towards computing this for most types
without a TypeLowering, and hopefully eventually removing TL entirely.
Calling setupLLVMOptimizationRemarks overwrites the MainRemarkStreamer
in the LLVM context. This prevents LLVM from serializing the remark meta
information for the already emitted SIL remarks into the object file.
Without the meta information bitstream remarks don't work correctly.
Instead, emit SIL remarks and LLVM remarks to the same RemarkSerializer,
and keep the file stream alive until after CodeGen.
Expose the createStructType helper to clients of IRGenModule which want
to define types which won't ever be used elsewhere. This is just a
convenience--such clients could already have directly used the API on
llvm::Module directly.
Raw identifiers are backtick-delimited identifiers that can contain any
non-identifier character other than the backtick itself, CR, LF, or other
non-printable ASCII code units, and which are also not composed entirely
of operator characters.
When it's available, use an open-coded allocator function that returns
an alloca without popping if the allocator is nullptr and otherwise
calls swift_coro_alloc. When it's not available, use the malloc
allocator in the synchronous context.
The well known builtin and structural types are strongly defined in the
runtime which is compacted into the standard library. Given that the VWT
is defined in the runtime, it is not visible to the Swift compilation
process and as we do not provide a Swift definition, we would previously
compute the linkage as being module external (`dllimport` for shared
library builds). This formed incorrect references to these variables and
would require thunking to adjust the references.
One special case that we add here is the "any function" type
representation (`@escaping () -> ()`) as we do use the VWT for this type
in the standard library but do not consider it part of the well known
builtin or structural type enumeration.
These errors were previously being swallowed by the build system and
thus escaped from being fixed when the other cases of incorrect DLL
storage were.
Remove `IRGenModule::useDllStorage()` as there is a standalone version
that is available and the necessary information is public from the
`IRGenModule` type. Additionally, avoid the wrapped `isStandardLibrary`
preferring to use the same method off of the public accessors. This
works towards removing some of the standard library special casing so
that it is possible to support both static and dynamic standard
libraries on Windows.
In C++, a primary base class that is placed in the beginning of the type's memory layout isn't always the type that is the first in the list of bases – the base types might be laid out in memory in a different order.
This makes sure that IRGen handles base types of C++ structs in the correct order.
This fixes an assertion in asserts-enabled compilers, and an out-of-memory error in asserts-disabled compilers. The issue was happening for both value types and foreign reference types. This change also includes a small refactoring to reuse the logic between the two code paths.
rdar://140848603
When compiling with library evolution and a pre-Swift 6.0 deployment
target, a mismatch between the notion of resilience used for determining
whether a protocol that inherits Sendable might need to be treated as
"dependent" differed from how other parts of IR generation decided
whether to conformance should be considered as resilient. The
difference came when both the protocol and its conforming type are in
the same module as the user.
Switch over to the "is this conformance resilient?" query that takes
into account such conformances.
Fixes rdar://136586922.
The main change here is to associate a witness table with a `ProtocolConformance` instead of a `RootProtocolConformance`.
A `ProtocolConformance` is the base class and can be a `RootProtocolConformance` or a `SpecializedProtocolConformance`.
For types like `Atomic` and `Mutex`, we want to know that even though they are
technically bitwise-takable, they differ from other bitwise-takable types until
this point because they are not also "bitwise-borrowable"; while borrowed,
they are pinned in memory, so they cannot be passed by value as a borrowed
parameter, unlike copyable bitwise-takable types. Add a bit to the value witness
table flags to record this.
Note that this patch does not include any accompanying runtime support for
propagating the flag into runtime-instantiated type metadata. There isn't yet
any runtime functionality that varies based on this flag, so that can
be implemented separately.
rdar://136396806
