Enable linker dead stripping on Darwin #22839
Conversation
See r://48283130. This shaves 6MB+ off of swift and SourceKitService, 26.8MB off of swift-llvm-opt, 29.4MB off sil-passpipeline-dumper, etc.
|
@swift-ci smoke test |
|
@akyrtzi Do you know of any build/link issues with plugins that could be triggered by dead stripping? Or are all those problems resolved at this point? |
cmake/modules/AddSwift.cmake
Outdated
| if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin") | ||
| # See rdar://48283130: This gives 6MB+ size reductions for swift and | ||
| # SourceKitService, and much larger size reductions for sil-opt etc. | ||
| list(APPEND result "-Wl,-dead_strip") |
There was a problem hiding this comment.
add_link_opts only adds the option for a release build. It may be useful to avoid stripping in debug builds because stripping can hide linker errors caused by dependency cycles.
Or try to just delegate to add_link_opts ?
There was a problem hiding this comment.
I'll add a Debug check. It'll take extensive refactoring to pull in the definition of add_link_opts here (just including AddLLVM breaks the swift build due to conflicting variables).
There was a problem hiding this comment.
Hmm, I'm certainly no CMake expert but just want to point out that I've added a call to add_link_opts in here:
https://github.com/apple/swift/blob/34f8670d2a8d1ce05c6da32da8e484e543e8ec5e/tools/libSwiftSyntaxParser/CMakeLists.txt#L21
without issues, but maybe this particular file is problematic for it.
There was a problem hiding this comment.
Yeah, add_link_opts requires that a target exists (it sets properties on the target directly). Most of swift's cmake logic collects properties, and then creates a target.
Maybe we should just factor out the core logic in add_link_opts into a helper which returns a list (in llvm), then use that here?
I'm not aware of any such problems, AFAIK we don't support loading plugins from swift or SourceKitService. |
It can hide linker errors caused by dependency cycles.
|
@swift-ci smoke test |
| # `add_link_opts` function (which, perhaps, should have been used here in the | ||
| # first place, but at this point it's hard to say whether that's feasible). | ||
| # | ||
| # TODO: Evaluate/enable -f{function,data}-sections --gc-sections for bfd, |
There was a problem hiding this comment.
BFD is an unsupported linker (it cannot handle the protected semantics needed for swift)
|
@swift-ci smoke test Linux platform |
See r://48283130. This shaves 6MB+ off of swift and SourceKitService,
26.8MB off of swift-llvm-opt, 29.4MB off sil-passpipeline-dumper, etc.