We can finally get rid of -sil-serialize-all when building the standard library! This option will be completely eliminated in the future commits.
Instead of serializing just everything as we did before, we now serialize only functions annotated with @_inlineable. This way we can selectively control what needs to be available to the clients. This is an important step towards building a resilient standard library.
While this is a huge change for the serialization of the stdlib, it should be virtually invisible to the clients. For example, there are no noticeable performance regressions on any of the benchmarks.
When moving to the swift-4.0-branch of LLVM, we started using the
LLVM_DISABLE_ABI_BREAKING_CHECKS_ENFORCING cmake option to turn off those
checks. This was necessary for the stdlib build, where it uses some LLVM
headers without linking libSupport.a, but the cmake option forced us to
disable the checks across the board. It also caused a lot of churn for
people who needed to remove their cmake caches when switching to the new
version of LLVM. This change switches to use a preprocessor macro to
disable the checks only in the context of the stdlib build. It requires
LLVM r295090. rdar://problem/30098953
for sib/sibgen files for libraries without building them by default.
This will let engineers just cd into the build directory and type:
ninja swift-stdlib-sib
ninja swift-stdlib-sibgen
To generate sib and sibgen files respectively.
There are still some dependency issues in between the sib targets, so to get
this to work well, I would suggest doing a full build and then using these
targets.
My recent change to enable SWIFT_BUILD_RUNTIME_WITH_HOST_COMPILER by default
exposed this problem with ASan. There are some places, e.g., the
swift-reflection-dump tool, where we link together Swift runtime code with
other code built by the host compiler. If we try to use the sanitizers
with both compilers mixed together, this can cause link failures.
To get the same coverage, we probably need to sanitize the runtime
separately. rdar://problem/29567862
This splits the `--build-swift-stdlib` and `--build-swift-sdk-overlay`
arguments into `dynamic` and `static` variants, which makes the
following build command possible:
```
utils/build-script -- \
--build-swift-dynamic-stdlib=0 --build-swift-dynamic-sdk-overlay=0 \
--build-swift-static-stdlib=1 --build-swift-static-sdk-overlay=0
```
This command produces *only* static libraries for the stdlib, and no
SDK overlay libraries at all. Many other finely-grained build options
are now possible.
As a first step to allowing the build script to build *only*
static library versions of the stdlib, change `add_swift_library`
such that callers must pass in `SHARED`, `STATIC`, or `OBJECT_LIBRARY`.
Ideally, only these flags would be used to determine whether to
build shared, static, or object libraries, but that is not currently
the case -- `add_swift_library` also checks whether the library
`IS_STDLIB` before performing certain additional actions. This will be
cleaned up in a future commit.
This commit makes the swift build scripts use the clang that we build as part of
the build process in order to build the Swift runtime, and not the host
compiler. This change is necessary for using new clang features immediately after they are introduced.
One example for such feature is new function calling conventions, or new c-level attributes.
We want to have the option of building the standard library without
-sil-serialize-all.
For now, the build crashes in the SIL verifier due to various issues;
upcoming patches will address these.
This is part of the larger goal of building the stdlib with
-enable-resilience.
Part of https://bugs.swift.org/browse/SR-267.
This brings down StdlibUnittest build time to 90 seconds with either
a DebugAssert or a ReleaseAssert compiler.
The new library, StdlibCollectionTests, is only built when running
validation tests.
