'ModuleDependencyScanner' maintains a Thread Pool along with a pool of workers
which are capable of executing a filesystem lookup of a named module dependency.
When resolving imports of a given Swift module, each import's resolution
operation can be issued asunchronously.
From being a scattered collection of 'static' methods in ScanDependencies.cpp
and member methods of ASTContext. This makes 'ScanDependencies.cpp' much easier
to read, and abstracts the actual scanning logic away to a place with common
state which will make it easier to reason about in the future.
dependencies
It is valuable for clients to be able to distinguish which dependencies of a
Swift module originated from 'import' statements, and which ones are implicit
dependency Swift overlays of imported Clang modules.
Instead of the code querying the compiler's built-in Clang instance, refactor the
dependency scanner to explicitly keep track of module output path. It is still
set according to '-module-cache-path' as it has been prior to this change, but
now the scanner can use a different module cache for scanning PCMs, as specified
with '-clang-scanner-module-cache-path', without affecting module output path.
Resolves rdar://113222853
Add a flag `finalized` to indicate that a module entry in the dependency
cache is finalized and no longer needs to be updated. This prevents the
command-line flags from dependency inputs get added multiple times on
re-scan with the same service.
While during normal compilation, adding the same command-line flags
multiple times are fine, it is bad for caching builds as a new
compilation cache key needs to be computed every time.
Unlike `swift-frontend -scan-dependencies` option, when dependency
scanner is used as a library by swift driver, the SwiftScanningService
is shared for multiple driver invocations. It can't keep states (like
common file dependencies) that can change from one invocation to
another.
Instead, the clang/SDK file dependencies are computed from each driver
invocations to avoid out-of-date information when scanning service is
reused.
The test case for a shared Service will be added to swift-driver repo
since there is no tool to test it within swift compiler.
Reformatting everything now that we have `llvm` namespaces. I've
separated this from the main commit to help manage merge-conflicts and
for making it a bit easier to read the mega-patch.
This is phase-1 of switching from llvm::Optional to std::optional in the
next rebranch. llvm::Optional was removed from upstream LLVM, so we need
to migrate off rather soon. On Darwin, std::optional, and llvm::Optional
have the same layout, so we don't need to be as concerned about ABI
beyond the name mangling. `llvm::Optional` is only returned from one
function in
```
getStandardTypeSubst(StringRef TypeName,
bool allowConcurrencyManglings);
```
It's the return value, so it should not impact the mangling of the
function, and the layout is the same as `std::optional`, so it should be
mostly okay. This function doesn't appear to have users, and the ABI was
already broken 2 years ago for concurrency and no one seemed to notice
so this should be "okay".
I'm doing the migration incrementally so that folks working on main can
cherry-pick back to the release/5.9 branch. Once 5.9 is done and locked
away, then we can go through and finish the replacement. Since `None`
and `Optional` show up in contexts where they are not `llvm::None` and
`llvm::Optional`, I'm preparing the work now by going through and
removing the namespace unwrapping and making the `llvm` namespace
explicit. This should make it fairly mechanical to go through and
replace llvm::Optional with std::optional, and llvm::None with
std::nullopt. It's also a change that can be brought onto the
release/5.9 with minimal impact. This should be an NFC change.
Teach swift dependency scanner to use CAS to capture the full dependencies for a build and construct build commands with immutable inputs from CAS.
This allows swift compilation caching using CAS.
There is a special case that already exists in 'ClangImporter' for implicit module loading:
Import of a "submodule" named "Foo.Private" is treated as a top-level module named "Foo_Private".
Clang has special support for this.
Resolves rdar://108287140
Instead of being a part of 'directDependencies' on a module dependency info, make them a separate array of dependency IDs for Swift Source and Textual modules.
This will allow clients to still distinguish direct module dependencies imported from a given module, versus dependencies added because direct/transitive Clang module dependencies have Swift overlays.
This change does *not* remove overlay dependencies from 'directDependencies' yet, just adds them as a separate field on the module details info. A followup change will remove overlay and bridging header dependencies from 'directDependencies' once the clients have had a chance to adopt to this change.
Instead, treat them like any other module that is specific to the scanning context hash of the scan it originates from.
Otherwise we may actually have simultaneous scans happening for the same source module but with different context hashes, and the current scheme leads to collisions.
Using mutual exclusion, ensuring that multiple threads executing dependency scans do not encounter data races on shared mutable state.
There are two layers with shared state where we need to be careful:
- `DependencyScanningTool`, as the main entity that scanning clients interact with. This tool instantiates compiler instances for individual scans, computing a scanning invocation hash. It needs to remember those instances for future use, and when creating instances it needs to reset LLVM argument processor's global state, meaning all uses of argument processing must be in a critical section.
- `SwiftDependencyScanningService`, as the main cache where dependency scanning results are stored. Each individual scan instantiates a `ModuleDependenciesCache`, which uses the scanning service as the underlying storage. The services' storage is segmented to storing dependencies discovered in a scan with a given context hash, which means two different scanning invocations running at the same time will be accessing different locations in its storage, thus not requiring synchronization. But the service still has some shared state that must be protected, such as the collection of discovered source modules, and the map used to query context-hash-specific underlying cache storage.
Do this by computing a transitive closure on the computed dependency graph, relying on the fact that it is a DAG.
The used algorithm is:
```
for each v ∈ V {
T(v) = { v }
}
for v ∈ V in reverse topological order {
for each (v, w) ∈ E {
T(v) = T(v) ∪ T(w)
}
}
```
This changes the scanner's behavior to "resolve" a discovered module's dependencies to a set of Module IDs: module name + module kind (swift textual, swift binary, clang, etc.).
The 'ModuleDependencyInfo' objects that are stored in the dependency scanner's cache now carry a set of kind-qualified ModuleIDs for their dependencies, in addition to unqualified imported module names of their dependencies.
Previously, the scanner's internal state would cache a module dependnecy as having its own set of dependencies which were stored as names of imported modules. This led to a design where any time we needed to process the dependency downstream from its discovery (e.g. cycle detection, graph construction), we had to query the ASTContext to resolve this dependency's imports, which shouldn't be necessary. Now, upon discovery, we "resolve" a discovered dependency by executing a lookup for each of its imported module names (this operation happens regardless of this patch) and store a fully-resolved set of dependencies in the dependency module info.
Moreover, looking up a given module dependency by name (via `ASTContext`'s `getModuleDependencies`) would result in iterating over the scanner's module "loaders" and querying each for the module name. The corresponding modules would then check the scanner's cache for a respective discovered module, and if no such module is found the "loader" would search the filesystem.
This meant that in practice, we searched the filesystem on many occasions where we actually had cached the required dependency, as follows:
Suppose we had previously discovered a Clang module "foo" and cached its dependency info.
-> ASTContext.getModuleDependencies("foo")
--> (1) Swift Module "Loader" checks caches for a Swift module "foo" and doesn't find one, so it searches the filesystem for "foo" and fails to find one.
--> (2) Clang Module "Loader" checks caches for a Clang module "foo", finds one and returns it to the client.
This means that we were always searching the filesystem in (1) even if we knew that to be futile.
With this change, queries to `ASTContext`'s `getModuleDependencies` will always check all the caches first, and only delegate to the scanner "loaders" if no cached dependency is found. The loaders are then no longer in the business of checking the cached contents.
To handle cases in the scanner where we must only lookup either a Swift-only module or a Clang-only module, this patch splits 'getModuleDependencies' into an alrady-existing 'getSwiftModuleDependencies' and a newly-added 'getClangModuleDependencies'.
Adopts Clang's 'DependencyScanningWorkerFilesystem' for use by the scanner, with the persistent
scanner instance keeping a 'DependencyScanningFilesystemSharedCache'.
Introduces a concept of a dependency scanning action context hash, which is used to select an instance of a global dependency scanning cache which gets re-used across dependency scanning actions.
It was recently moved to the ModuleDependenciesCache. This is undesireable, instead this should live in the GlobalModuleDependenciesCache so that we benefit from the filesystem caching it performs across diferent scanning actions.
Move clangScanningTool and clangScanningService to be parts of 'ModuleDependenciesCache' state, getting rid of the 'ClangModuleDependenciesCacheImpl', which is no-longer needed since we moved moved to by-name lookup of Clang modules.
This change tweaks the 'GlobalModuleDependenciesCache', which persists across scanner invocations with the same 'DependencyScanningTool' to no longer cache discovered Clang modules.
Doing so felt like a premature optimization, and we should instead attempt to share as much state as possible by keeping around the actual Clang scanner's state, which performs its own caching. Caching discovered dependencies both in the Clang scanner instance, and in our own cache is much more error-prone - the Clang scanner has a richer context for what is okay and not okay to cache/re-use.
Instead, we still cache discovered Clang dependencies *within* a given scan, since those are discovered using a common Clang scanner instance and should be safe to keep for the duration of the scan.
This change should make it simpler to pin down the core functionality and correctness of the scanner.
Once we turn our attention to the scanner's performance, we can revisit this strategy and optimize the caching behaviour.
When we are building a Swift module which has an underlying Clang module, and which generates an ObjC interface ('-Swift.h'), the mechanism for building the latter involves a VFS redirect of its modulemap to one that does not yet have the generated Swift code, because it must be built before the Swift portion is built because the Swift portion depends on it. This means that the invocation to build this module is different to one used by the clients which depend on this module.
To avoid the subsequent client scans from re-using the partial (VFS-redirected) module, ensure that we do not store dependency info of the underlying Clang module into the global scanner cache. This will cause subsequent client scans to re-scan for this module, and find the fully-resolved modulemap without a VFS redirect.
Resolves rdar://88309064
Having it be in the potentially-persistent global cache state seems to be causing issues with search paths of subsequent scans. While I investigate the cause, moving this state to the local cache works around the problem.
Doing so will allow clients to know which Swift-specific PCM arguments are already captured from the scan that first discovered this module.
SwiftDriver, in particular, will be able to use this information to avoid re-scanning a given Clang module if the initial scan was sufficient for all possible sets of PCM arguments on Swift modules that depend on said Clang module.
And only resolve cached dependencies that came from scanning actions with the same target triple.
This change means that the `GlobalModuleDependenciesCache` must be configured with a specific target triple for every scannig action, and it will only resolve previously-found dependencies from previous scannig actions using the exact same triple.
Furthermore, the `GlobalModuleDependenciesCache` separately tracks source-file-based module dependencies as those represent main Swift modules of previous scanning actions, and we must be able to resolve those regardless of the target triple.
Resolves rdar://83105455
These kinds of modules differ from `SwiftTextual` modules in that they do not have an interface and have source-files.
It is cleaner to enforce this distinction with types, instead of checking for interface optionality everywhere.
This change causes the cache to be layered with a local "cache" that wraps the global cache, which will serve as the source of truth. The local cache persists only for the duration of a given scanning action, and has a store of references to dependencies resolved as a part of the current scanning action only, while the global cache is the one that persists across scanning actions (e.g. in `DependencyScanningTool`) and stores actual module dependency info values.
Only the local cache can answer dependency lookup queries, checking current scanning action results first, before falling back to querying the global cache, with queries disambiguated by the current scannning action's search paths, ensuring we never resolve a dependency lookup query with a module info that could not be found in the current action's search paths.
This change is required because search-path disambiguation can lead to false-negatives: for example, the Clang dependency scanner may find modules relative to the compiler's path that are not on the compiler's direct search paths. While such false-negative query responses should be functionally safe, we rely on the current scanning action's results being always-present-in-the-cache for the scanner's functionality. This layering ensures that the cache use-sites remain unchanged and that we get both: preserved global state which can be queried disambiguated with the search path details, and an always-consistent local (current action) cache state.
The dependency scanner's cache persists across different queries and answering a subsequent query's module lookup with a module not in the query's search path is not correct.
For example, suppose we are looking for a Swift module `Foo` with a set of search paths `SP`.
And dependency scanner cache already contains a module `Foo`, for which we found an interface file at location `L`. If `L`∉`SP`, then we cannot re-use the cached entry because we’d be resolving the scanning query to a filesystem location that the current scanning context is not aware of.
Resolves rdar://81175942
This matches the behavior of the current client (`swift-driver`) and reduces ambiguity in how the nodes in the graph are to be treated. Swift dependencies with a textual interface, for example, must be built into a binary module by clients. Swift dependencies without a textual interface, with only a binary module, are to be used directly, without any up-to-date checks.
Note, this is distinct from Swift dependencies that have a textual interface, for which we also detect potential pre-build binary module candidates. Those are still reported in the `details` field of textual Swift dependencies as `prebuiltModuleCandidates`.
In the fast dependency scanner, depending on whether a module intrface was found via the import search path or framework search path, encode into the dependency graph Swift module details, whether a given module is a framework.
