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.
Previously, two conditions were necessary to enable differentiable programming:
- Using the `-enable-experimental-differentiable-programming` frontend flag.
- Importing the `_Differentiation` module.
Importing the `_Differentiation` module is the true condition because it
contains the required compiler-known `Differentiable` protocol. The frontend
flag is redundant and cumbersome.
Now, the frontend flag is removed.
Importing `_Differentiation` is the only condition.
Request-based incremental dependencies are enabled by default. For the time being, add a flag that will turn them off and switch back to manual dependency tracking.
* [Diagnostics] Turn educational notes on-by-default
* [Diagnostics] Only include educational notes in printed output if -print-educational-notes is passed
* Make -print-educational-notes a driver option
* [Diagnostics] Issue a printed remark if educational notes are available, but disabled
* [docs] Update educational notes documentation and add a contributing guide
* [Diagnostics] Cleanup PrintingDiagnosticConsumer handling of edu notes
* Revert "[Diagnostics] Issue a printed remark if educational notes are available, but disabled"
For now, don't notify users if edu notes are available but disabled. This decision can be reevaluated later.
Add `AdditiveArithmetic` derived conformances for structs, gated by the
`-enable-experimential-additive-arithmetic-derivation` flag.
Structs whose stored properties all conform to `AdditiveArithmetic` can derive
`AdditiveArithmetic`:
- `static var zero: Self`
- `static func +(lhs: Self, rhs: Self) -> Self`
- `static func -(lhs: Self, rhs: Self) -> Self`
- An "effective memberwise initializer":
- Either a synthesized memberwise initializer or a user-defined initializer
with the same type.
Effective memberwise initializers are used only by derived conformances for
`Self`-returning protocol requirements like `AdditiveArithmetic.+`, which
require memberwise initialization.
Resolves TF-844.
Unblocks TF-845: upstream `Differentiable` derived conformances.
This allows the usage of the whole remark infrastructure developed in
LLVM, which includes a new binary format, metadata in object files, etc.
This gets rid of the YAMLTraits-based remark serialization and does the
plumbing for hooking to LLVM's main remark streamer.
For more about the idea behind LLVM's main remark streamer, see the
docs/Remarks.rst changes in https://reviews.llvm.org/D73676.
The flags are now:
* -save-optimization-record: enable remarks, defaults to YAML
* -save-optimization-record=<format>: enable remarks, use <format> for
serialization
* -save-optimization-record-passes <regex>: only serialize passes that
match <regex>.
The YAMLTraits in swift had a different `flow` setting for the debug
location, resulting in some test changes.
When enabled at the driver level, the frontends will inherit the flag. For each frontend that recieves this option, all primaries will have their reference dependencies validated.
* Stage in #filePath
To give users of #file time to transition, we are first adding #filePath without changing #file’s behavior. This commit makes that change.
Fixes <rdar://problem/58586626>.
* Correct swiftinterface test line
Static-linked libraries could add symbols to the final tbd file. We need
this flag to specify additional module names to collect symbols from.
rdar://59399684
This adds an argument to allow negating `-whole-module-optimization`.
This is useful for cases where it's easier to add an extra flag to your
swiftc invocation rather than removing the original one.
Add support in the driver and frontend for macCatalyst target
targets and library search paths.
The compiler now adds two library search paths for overlays when compiling
for macCatalyst: one for macCatalyst libraries and one for zippered macOS
libraries. The macCatalyst path must take priority over the normal macOS path
so that in the case of 'unzippered twins' the macCatalyst library is
found instead of the macOS library.
To support 'zippered' builds, also add support for a new -target-variant
flag. For zippered libraries, the driver invocation takes both a -target and a
-target-variant flag passes them along to the frontend. We support builds both
when the target is a macOS triple and the target variant is macCatalyst and
also the 'reverse zippered' configuration where the target is macCatalyst and the
target-variant is macOS.
Restructure fine-grained-dependencies to enable unit testing
Get frontend to emit correct swiftdeps file (fine-grained when needed) and only emit dot file for -emit-fine-grained-dependency-sourcefile-dot-files
Use deterministic order for more information outputs.
Set EnableFineGrainedDependencies consistently in frontend.
Tolerate errors that result in null getExtendedNominal()
Fix memory issue by removing node everywhere.
Break up print routine
Be more verbose so it will compile on Linux.
Sort batchable jobs, too.
Restructure fine-grained-dependencies to enable unit testing
Get frontend to emit correct swiftdeps file (fine-grained when needed) and only emit dot file for -emit-fine-grained-dependency-sourcefile-dot-files
Use deterministic order for more information outputs.
Set EnableFineGrainedDependencies consistently in frontend.
Tolerate errors that result in null getExtendedNominal()
Fix memory issue by removing node everywhere.
Break up print routine
Be more verbose so it will compile on Linux.
Sort batchable jobs, too.
Rather than only emitting the target triple, provide additional
information about that particular target, including the module triple
(i.e., what file names will be used for Swift modules for that
triple), the runtime compatibility version if there is one, and
whether linking with rpaths is required for the standard library and
other libraries shipped with Swift. Encode this as JSON so we can
extend it in the future. For now, it looks like this:
```
{
"target": {
"triple": "arm64-apple-ios12.0",
"moduleTriple": "arm64-apple-ios",
"swiftRuntimeCompatibilityVersion": "5.0",
"librariesRequireRPath": true
}
}
```
Which you can deserialize into a TargetInfo instance as defined below:
```
struct Target: Codable {
/// The target triple.
var triple: String
/// The triple used for module file names.
var moduleTriple: String
/// If this platform provides the Swift runtime, the Swift language
version
/// with which that runtime is compatible.
var swiftRuntimeCompatibilityVersion: String?
/// Whether linking against the Swift libraries requires the use of
rpaths.
var librariesRequireRPath: Bool
}
struct TargetInfo: Codable {
var target: Target
}
```
Implements rdar://problem/47095159.
Add a -print-target-triple command line option to the Swift frontend
and driver to allow other tools (e.g., SwiftPM) to query the host
triple as it is understood by the Swift compiler. This follows the
precedent set by Clang. Implements rdar://problem/57434967.
This is a first version of cross module optimization (CMO).
The basic idea for CMO is to use the existing library evolution compiler features, but in an automated way. A new SIL module pass "annotates" functions and types with @inlinable and @usableFromInline. This results in functions being serialized into the swiftmodule file and thus available for optimizations in client modules.
The annotation is done with a worklist-algorithm, starting from public functions and continuing with entities which are used from already selected functions. A heuristic performs a preselection on which functions to consider - currently just generic functions are selected.
The serializer then writes annotated functions (including function bodies) into the swiftmodule file of the compiled module. Client modules are able to de-serialize such functions from their imported modules and use them for optimiations, like generic specialization.
The optimization is gated by a new compiler option -cross-module-optimization (also available in the swift driver).
By default this option is off. Without turning the option on, this change is (almost) a NFC.
rdar://problem/22591518
Frontend outputs source-as-compiled, and source-ranges file with function body ranges and ranges that were unparsed in secondaries.
Driver computes diffs for each source file. If diffs are in function bodies, only recompiles that one file. Else if diffs are in what another file did not parse, then the other file need not be rebuilt.
The new option `-sanitize-recover=` takes a list of sanitizers that
recovery instrumentation should be enabled for. Currently we only
support it for Address Sanitizer.
If the option is not specified then the generated instrumentation does
not allow error recovery.
This option mirrors the `-fsanitize-recover=` option of Clang.
We don't enable recoverable instrumentation by default because it may
lead to code size blow up (control flow has to be resumable).
The motivation behind this change is that today, setting
`ASAN_OPTIONS=halt_on_error=0` at runtime doesn't always work. If you
compile without the `-sanitize-recover=address` option (equivalent to
the current behavior of the swift compiler) then the generated
instrumentation doesn't allow for error recovery. What this means is
that if you set `ASAN_OPTIONS=halt_on_error=0` at runtime and if an ASan
issue is caught via instrumentation then the process will always halt
regardless of how `halt_on_error` is set. However, if ASan catches an
issue via one of its interceptors (e.g. memcpy) then `the halt_on_error`
runtime option is respected.
With `-sanitize-recover=address` the generated instrumentation allows
for error recovery which means that the `halt_on_error` runtime option
is also respected when the ASan issue is caught by instrumentation.
ASan's default for `halt_on_error` is true which means this issue only
effects people who choose to not use the default behavior.
rdar://problem/56346688
This flag will enable all experimental differentiable programming features.
The default will be `true` on tensorflow branch but `false` on master branch.
Features will first be updated on tensorflow branch to use this flag, before
being upstreamed to master. The goal is to achieve a minimal code diff between
the two branches.
The [TF-820](https://bugs.swift.org/browse/TF-820) master issue tracks upstreaming differentiable programming.
---
Rationale: we chose to add a frontend flag rather than a `build-script`/CMake flag for easier testing. Differentiable programming `lit` tests can be run by specifying this additional flag without recompiling the compiler and standard library.
[Forum discussion on upstreaming differentiable programming.](https://forums.swift.org/t/upstreaming-differentiable-attribute-and-differentiable-protocol/26821)
This flag, currently staged in as `-experimental-skip-non-inlinable-function-bodies`, will cause the typechecker to skip typechecking bodies of functions that will not be serialized in the resulting `.swiftmodule`. This patch also includes a SIL verifier that ensures that we don’t accidentally include a body that we should have skipped.
There is still some work left to make sure the emitted .swiftmodule is exactly the same as what’s emitted without the flag, which is what’s causing the benchmark noise above. I’ll be committing follow-up patches to address those, but for now I’m going to land the implementation behind a flag.
Add `-no-toolchain-stdlib-rpath` flag: the negative version of
`-toolchain-stdlib-rpath`.
Make `-no-toolchain-stdlib-rpath` be the default: use `/usr/lib/swift` as
default RPATH on Darwin platforms instead of toolchain standard library.
Adapted from https://github.com/apple/swift/pull/27206.
tensorflow branch requires the opposite default (use toolchain standard
library as RPATH) because some stdlib modules like TensorFlow do not exist in
`/usr/lib/swift`.
This flag adds diagnostic names to the end of their messages, e.g. 'error: cannot convert value of type '[Any]' to specified type '[Int]' [cannot_convert_initializer_value]'. It's intended to be used for debugging purposes when working on the compiler.
