Canonicalizes `differentiable_function` instructions by filling in missing
derivative function operands.
Derivative function emission rules, based on the original function value:
- `function_ref`: look up differentiability witness with the exact or a minimal
superset derivative configuration. Emit a `differentiability_witness_function`
for the derivative function.
- `witness_method`: emit a `witness_method` with the minimal superset derivative
configuration for the derivative function.
- `class_method`: emit a `class_method` with the minimal superset derivative
configuration for the derivative function.
If an *actual* emitted derivative function has a superset derivative
configuration versus the *desired* derivative configuration, create a "subset
parameters thunk" to thunk the actual derivative to the desired type.
For `differentiable_function` instructions formed from curry thunk applications:
clone the curry thunk (with type `(Self) -> (T, ...) -> U`) and create a new
version with type `(Self) -> @differentiable (T, ...) -> U`.
Progress towards TF-1211.
wrapper original wrapped value expression inside of CSApply.
This prevents type checking the synthesized backing storage initializer
twice - once with the original expression and again with the placeholder.
Like switch cases, a catch clause may now include a comma-
separated list of patterns. The body will be executed if any
one of those patterns is matched.
This patch replaces `CatchStmt` with `CaseStmt` as the children
of `DoCatchStmt` in the AST. This necessitates a number of changes
throughout the compiler, including:
- Parser & libsyntax support for the new syntax and AST structure
- Typechecking of multi-pattern catches, including those which
contain bindings.
- SILGen support
- Code completion updates
- Profiler updates
- Name lookup changes
Short overview of new TBD-v4 format changes:
* special section for reexported symbols (which is not seen any
differently to the linker)
* target based slices as opposed to just architecture
more information in: rdar://problem/60586390
Unfortunately we still need this performance hack because otherwise
e.g. if initializer returns a tuple its type is going to be connected
to a type variable representing a pattern type, which means all of the
tuple element types are going to form a single constraint system component.
Resolves: rdar://problem/60961087
In each of the following situations `getTypeForPattern` would
add a new pattern element to the path:
- Element of a tuple pattern
- Sub-pattern of a typed pattern
- Sub-pattern of optional .some
Without whole module optimization, the metadata accessors are emitted on
a per-file basis. The result is that if the file containing a generic
type is processed before the file containing a usage of that type that
would result in that prespecialization, the metadata accessor would have
already been emitted by the time that the usage is noted, making it
impossible for the newly created prespecialization to be returned from
the already-emitted metadata accessor.
Here, require that either whole module optimization is enabled so that
the metadata accessors are all emitted at once at the end, or else that
the usage of the prespecialization is in the same file as the type is
declared.
We used to take all the captures of a local function and treat them all
as read and write usages of vars from an outer scope. Instead, let's
refactor the analysis to walk into local functions.
Type erasure requires a circular construction by its very nature:
@_typeEraser(AnyProto)
protocol Proto { /**/ }
public struct AnyProto : Proto {}
If we eagerly resolve AnyProto, the chain of resolution steps that
deserialization must make goes a little something like this:
Lookup(Proto)
-> Deserialize(@_typeEraser(AnyProto))
-> Lookup(AnyProto)
-> DeserializeInheritedStuff(AnyProto)
-> Lookup(Proto)
This cycle could be broken if the order of incremental inputs was
such that we had already cached the lookup of Proto.
Resolve this cycle in any case by suspending the deserialization of the
type eraser until the point it's demanded by adding
ResolveTypeEraserTypeRequest.
rdar://61270195
If an import-as-member property was used in a key path, we'd try to identify the component by its
foreign-to-native thunk, which isn't normally generated (causing a crash from the missing symbol)
and wouldn't be globally unique even if it were. Fixes rdar://problem/60519829.
Before attempting to get the superclass of a
self parameter type, check to see if we have a
metatype, and perform the necessary unwrapping and
re-wrapping if needed.
This loop optimization hoists and sinks a group of loads and stores to
the same address.
Consider this SIL...
PRELOOP:
%stackAddr = alloc_stack $Index
%outerAddr1 = struct_element_addr %stackAddr : $*Index, #Index.value
%innerAddr1 = struct_element_addr %outerAddr1 : $*Int, #Int._value
%outerAddr2 = struct_element_addr %stackAddr : $*Index, #Index.value
%innerAddr2 = struct_element_addr %outerAddr2 : $*Int, #Int._value
LOOP:
%_ = load %innerAddr2 : $*Builtin.Int64
store %_ to %outerAddr2 : $*Int
%_ = load %innerAddr1 : $*Builtin.Int64
There are two bugs:
1) LICM miscompiles code during combined load/store hoisting and sinking.
When the loop contains an aliasing load from a difference projection
value, the optimization sinks the store but never replaces the
load. At runtime, the load reads a stale value.
FIX: isOnlyLoadedAndStored needs to check for other load instructions
before hoisting/sinking a seemingly unrelated set of
loads/stores. Checking side effect instructions is insufficient. The
same bug could happen with stores, which also do not produce side
effects.
Fixes <rdar://61246061> LICM miscompile:
Combined load/store hoisting/sinking with aliases
2) The LICM algorithm is not robust with respect to address projection
because it identifies a projected address by its SILValue. This
should never be done! It is trivial to represent a project path
using an IndexTrieNode (there is also an abstraction called
"ProjectionPath", but it should _never_ actually be stored by an
analysis because of the time and space complexity of doing so).
The second bug is not necessary to fix for correctness, so will be
fixed in a follow-up commit.
This imports const members of C++ structs/classes stored properties with
an inaccessible setter.
Note that in C++ there are ways to change the values of const members,
so we don't use `WriteImplKind::Immutable` storage.
Resolves: [SR-12463](https://bugs.swift.org/browse/SR-12463)
In case an unchecked_addr_cast is used to convert between pointer and non-pointer, we missed some escaping values.
This can be the case when using C unions.
https://bugs.swift.org/browse/SR-12427
rdar://problem/60983997
Recent-ish SDKs for Darwin platforms include an SDKSettings.json
file with version information and Catalyst SDK version mappings. Read
these (when available) and use them to pass the appropriate SDK
version down to the Darwin linker via `-platform_version`.
Finishes rdar://problem/55972144.
Mac Catalyst was introduced with an iOS deployment target of 13.0.
If given a deployment target before that, adjust the deployment target
to 13.0 for the linker.
Standardize the way in which we pass platform version information to
the Darwin linker, using the `-platform_version` option. In the case
of Mac Catalyst, there may be two such platform arguments.
The eventual point of this refactoring is to also pass information
about the SDK version, which `-platform_version` supports but the
mix of `-*_version_min` parameters do not. For now, the SDK
version is stubbed out to 0.0.0, which is this option's "unknown"
value.
Part of rdar://problem/55972144.
The API was accidentally undefined, presumably because I checked in
the wrong code or there was a bad merge. The API will be used by
upcoming commits.
Meanwhile, getAccessedAddress was not stripping access markers, which
means some analysis may have been too conservative.
This fix could expose issues by making existing analyses more effective.
The differentiation transform does the following:
- Canonicalizes differentiability witnesses by filling in missing derivative
function entries.
- Canonicalizes `differentiable_function` instructions by filling in missing
derivative function operands.
- If necessary, performs automatic differentiation: generating derivative
functions for original functions.
- When encountering non-differentiability code, produces a diagnostic and
errors out.
Partially resolves TF-1211: add the main canonicalization loop.
To incrementally stage changes, derivative functions are currently created
with empty bodies that fatal error with a nice message.
Derivative emitters will be upstreamed separately.
