We still only parse transferring... but this sets us up for adding the new
'sending' syntax by first validating that this internal change does not mess up
the current transferring impl since we want both to keep working for now.
rdar://128216574
Add a new demangler option which excludes a closure's type signature.
This will be used in lldb.
Closures are not subject to overloading, and so the signature will never be used to
disambiguate. A demangled closure is uniquely identifiable by its index(s) and parent.
Where opaque types are involved, the concrete type signature can be quite complex. This
demangling option allows callers to avoid printing the underlying complex nested
concrete types.
Example:
before: `closure #1 (Swift.Int) -> () in closure #1 (Swift.Int) -> () in main`
after: `closure #1 in closure #1 in main`
Replace use of `snprintf()` with some custom code (this is around 10x
faster on my machine).
Move a few of the `Node` member functions to the header so they're
inlined.
Optimize the `deepEquals()` function by adding an `isSimilarTo()`
method on `Node`; the checks that were happening in the `deepEquals()`
function could be implemented more efficiently by making use of
details of the internal representation of `Node`.
rdar://125739630
This PR implements first set of changes required to support autodiff for coroutines. It mostly targeted to `_modify` accessors in standard library (and beyond), but overall implementation is quite generic.
There are some specifics of implementation and known limitations:
- Only `@yield_once` coroutines are naturally supported
- VJP is a coroutine itself: it yields the results *and* returns a pullback closure as a normal return. This allows us to capture values produced in resume part of a coroutine (this is required for defers and other cleanups / commits)
- Pullback is a coroutine, we assume that coroutine cannot abort and therefore we execute the original coroutine in reverse from return via yield and then back to the entry
- It seems there is no semantically sane way to support `_read` coroutines (as we will need to "accept" adjoints via yields), therefore only coroutines with inout yields are supported (`_modify` accessors). Pullbacks of such coroutines take adjoint buffer as input argument, yield this buffer (to accumulate adjoint values in the caller) and finally return the adjoints indirectly.
- Coroutines (as opposed to normal functions) are not first-class values: there is no AST type for them, one cannot e.g. store them into tuples, etc. So, everywhere where AST type is required, we have to hack around.
- As there is no AST type for coroutines, there is no way one could register custom derivative for coroutines. So far only compiler-produced derivatives are supported
- There are lots of common things wrt normal function apply's, but still there are subtle but important differences. I tried to organize the code to enable code reuse, still it was not always possible, so some code duplication could be seen
- The order of how pullback closures are produced in VJP is a bit different: for normal apply's VJP produces both value and pullback closure via a single nested VJP apply. This is not so anymore with coroutine VJP's: yielded values are produced at `begin_apply` site and pullback closure is available only from `end_apply`, so we need to track the order in which pullbacks are produced (and arrange consumption of the values accordingly – effectively delay them)
- On the way some complementary changes were required in e.g. mangler / demangler
This patch covers the generation of derivatives up to SIL level, however, it is not enough as codegen of `partial_apply` of a coroutine is completely broken. The fix for this will be submitted separately as it is not directly autodiff-related.
---------
Co-authored-by: Andrew Savonichev <andrew.savonichev@gmail.com>
Co-authored-by: Richard Wei <rxwei@apple.com>
The names of the private witness table accessor thunks we generate for
an opaque return type mangle the concrete conformance of the underlying
type.
If a conformance requirement of the opaque return type was witnessed by
a conditional conformance of a variadic generic type, we would crash
because of an unimplemented case in the mangler.
Fixes rdar://problem/125668798.
Invertible protocols are currently always mangled with `Ri`, followed by
a single letter for each invertible protocol (e.g., `c` and `e` for
`Copyable` and `Escapable`, respectively), followed by the generic
parameter index. However, this requires that we extend the mangling
for any future invertible protocols, which mean they won't be
backward compatible.
Replace this mangling with one that mangles the bit # for the
invertible protocol, e.g., `Ri_` (followed by the generic parameter
index) is bit 0, which is `Copyable`. `Ri0_` (then generic parameter
index) is bit 1, which is `Escapable`. This allows us to round-trip
through mangled names for any invertible protocol, without any
knowledge of what the invertible protocol is, providing forward
compatibility. The same forward compatibility is present in all
metadata and the runtime, allowing us to add more invertible
protocols in the future without updating any of them, and also
allowing backward compatibility.
Only the demangling to human-readable strings maps the bit numbers
back to their names, and there's a fallback printing with just the bit
number when appropriate.
Also generalize the mangling a bit to allow for mangling of invertible
requirements on associated types, e.g., `S.Sequence: ~Copyable`. This
is currently unsupported by the compiler or runtime, but that may
change, and it was easy enough to finish off the mangling work for it.
Extend TypeDecoder with support for inverse requirements, passing them
along to the type builder. Then implement support for inverse
requirements within the AST demangler, which addresses the round-trip
demangling failures we've been seeing.
The runtime and remote inspection facilities still need metadata to
deal with inverse requirements.
Fixes rdar://124564447.
LLVM is presumably moving towards `std::string_view` -
`StringRef::startswith` is deprecated on tip. `SmallString::startswith`
was just renamed there (maybe with some small deprecation inbetween, but
if so, we've missed it).
The `SmallString::startswith` references were moved to
`.str().starts_with()`, rather than adding the `starts_with` on
`stable/20230725` as we only had a few of them. Open to switching that
over if anyone feels strongly though.
We preserve the current semantics that we have today by requiring that either all SILResultInfo are transferring or none are transferring. This also let me swap to @sil_transferring representation.
I did both of these things to fix SIL issues around transferring.
It also ensures that we now properly emit
The `ABI` headers had accidentally grown an `#include` into compiler headers,
allowing the enum constant values of the `ValueOwnership` enum to leak into
the runtime ABI. Sever this inappropriate relationship by declaring a separate
`ParameterOwnership` enum with ABI-stable values in the ABI headers, and
explicitly converting between the AST and ABI representation where needed.
Fixes rdar://122435628.
The main piece that's still missing here is support for closures;
they actually mostly work, but they infer the wrong isolation for
actor-isolated closures (it's not expressed in the type, so obviously
they're non-isolated), so it's not really functional. We also have
a significant problem where reabstraction thunks collide incorrectly
because we don't mangle (or represent!) formal isolation into
SILFunctionType; that's another follow-up. Otherwise, I think SILGen
is working.
This includes runtime support for instantiating transferring param/result in
function types. This is especially important since that is how we instantiate
function types like: typealias Fn = (transferring X) -> ().
rdar://123118061
Not quite NFC because apparently the representation bleeds into what's
accepted in some situations where we're supposed to be warning about
conflicts and then making an arbitrary choice. But what we're doing
is nonsense, so we definitely need to break behavior here.
This is setting up for isolated(any) and isolated(caller). I tried
to keep that out of the patch as much as possible, though.
This library uses GenericMetadataBuilder with a ReaderWriter that can read data and resolve pointers from MachO files, and emit a JSON representation of a dylib containing the built metadata.
We use LLVM's binary file readers to parse the MachO files and resolve fixups so we can follow pointers. This code is somewhat MachO specific, but could be generalized to other formats that LLVM supports.
rdar://116592577
Create a version of the metadata specialization code which is abstracted so that it can work in different contexts, such as building specialized metadata from dylibs on disk rather than from inside a running process.
The GenericMetadataBuilder class is templatized on a ReaderWriter. The ReaderWriter abstracts out everything that's different between in-process and external construction of this data. Instead of reading and writing pointers directly, the builder calls the ReaderWriter to resolve and write pointers. The ReaderWriter also handles symbol lookups and looking up other Swift types by name.
This is accompanied by a simple implementation of the ReaderWriter which works in-process. The abstracted calls to resolve and write pointers are implemented using standard pointer dereferencing.
A new SWIFT_DEBUG_VALIDATE_EXTERNAL_GENERIC_METADATA_BUILDER environment variable uses the in-process ReaderWriter to validate the builder by running it in parallel with the existing metadata builder code in the runtime. When enabled, the GenericMetadataBuilder is used to build a second copy of metadata built by the runtime, and the two are compared to ensure that they match. When this environment variable is not set, the new builder code is inactive.
The builder is incomplete, and this initial version only works on structs. Any unsupported type produces an error, and skips the validation.
rdar://116592420
I am doing this in preparation for adding options to SILParameterInfo/
SILResultInfo that state that a parameter/result is transferring. Even though I
could have just introduced a new bit here, I instead streamlined the interface
of SILParameterInfo/SILResultInfo to use an OptionSet instead of individual bits
to make it easier to add new flags here. The reason why it is easier is that
along API (e.x.: function argument) boundaries one does not have to marshal each
field or pass each field. Instead one can just pass the whole OptionSet as an
opaque thing. Using this I was able to change serialization/deserialization of
SILParameterInfo/SILResultInfo so that one does not need to update them if one
adds new fields!
The reason why I am doing this for both SILParameterInfo/SILResultInfo in the
same commit is because they share code in the demangler that I did not want to
have to duplicate in an intervening commit. By changing them both at the same
type, I didn't have to change anything without an actual need to.
I am doing this in a separate commit from adding transferring support so I can
validate correctness using the tests for the options already supported
(currently only differentiability).
