[CMake] Don't add an extraneous rpath to the swift-frontend

[finagolfin]

When checking rpaths for #63782, I just noticed these incorrect runpaths on linux:

> readelf -d swift*-DEVELOPMENT-SNAPSHOT-2023-0*-a-ubuntu20.04/usr/bin/swift-frontend | ag "File:|runpath"
File: swift-5.8-DEVELOPMENT-SNAPSHOT-2023-02-23-a-ubuntu20.04/usr/bin/swift-frontend
 0x000000000000001d (RUNPATH)            Library runpath: [$ORIGIN/../lib/swift/linux:$ORIGIN/../../lib/swift/linux]
File: swift-DEVELOPMENT-SNAPSHOT-2023-03-01-a-ubuntu20.04/usr/bin/swift-frontend
 0x000000000000001d (RUNPATH)            Library runpath: [$ORIGIN/../lib/swift/linux:$ORIGIN/../../lib/swift/linux]

It looks like this line was incorrectly added in #60943, then incorporated into #61426 and merged.

This line is unused on Darwin because the CI builds with BOOTSTRAPPING-WITH-HOSTLIBS, but the incorrect runpath should show up on macOS too if a full bootstrap were done.

@zoecarver and @DougGregor, let me know what you think.

[finagolfin]

@rintaro, you spun out _add_swift_runtime_link_flags() last year in 9f4ce61, maybe you can chime in on whether it's properly used here.

[finagolfin]

Pinging @ktoso, haven't gotten a response on this so far and it's about to ship with the 5.8 release.

[finagolfin]

Ping @eeckstein, you have some experience with these rpath issues with your bootstrapping work.

[eeckstein]

Sorry, I can't remember the details. What you should do is to test all affected bootstrapping modes and see if the RPATH in the binary is set correct and if the final swift-frontend is loading the correct libraries (with otool -L on macos)

[finagolfin]

@eeckstein, nothing wrong with your bootstrapping work, the issue is this line that I'm removing in this pull, that was added by Doug and Zoe last year. It adds the incorrect second rpath shown in my OP for linux, and should add it incorrectly for macOS too for full bootstraps.

I'd simply like your opinion on if this line should be removed.

[eeckstein]

I can't tell for sure without testing it.

[finagolfin]

@edymtt, perhaps you can review?

[finagolfin]

@CodaFi, would you take a look at this?

[finagolfin]

@etcwilde, any input on this issue?

[edymtt]

@buttaface apologies for acknowledging this late -- at the moment I am busy with other efforts, so cannot devote the time to reason on any impact of this change.

In the meantime, could you comment on the impact this has on the Linux build?

[finagolfin]

@edymtt, I believe this is just a cleanup of an extra unneeded runpath added on linux and certain macOS configs. This could be a mild security issue in certain system install layouts and cause libraries to be loaded that shouldn't be, so it is better to clean it up.

If you would run the CI on this pull, that should give us an idea if this line was unneeded in the first place.

[finagolfin]

@artemcm, would you run the CI on this?

[edymtt]

@swift-ci please test

[finagolfin]

@edymtt, passed CI, which heavily implies this line was extraneous all along, though the CI obviously doesn't test all configurations.

I'm fairly certain it is never needed for linux, I just can't speak for some obscure macOS config.

[finagolfin]

@DougGregor, you committed this line, can you review?

[edymtt]

Found some time to review this, and in fact I see that _add_swift_runtime_link_flags is called for the swift-frontend target twice -- once as part of add_swift_host_tool and once explicitly -- and that the former invocation causes the addition of a second rpath when building for Linux (that does not seem to happen for Darwin).

@finagolfin -- could you check if my understanding of your concern is correct?

(in the meantime I will check on the "obscure macOS configs")

[edymtt]

@swift-ci please test

[edymtt]

@swift-ci please build toolchain

[finagolfin]

@edymtt, yep, that is the problem, you got it.

[finagolfin]

Mac toolchain build failure is unrelated, some libcxx issue:

CMake Error at test/CMakeLists.txt:153 (message):
  Couldn't find libcxx build in
  '/Users/ec2-user/jenkins/workspace/swift-PR-toolchain-macos/branch-main/buildbot_osx/libcxx-macosx-x86_64'.
  To run the test-suite for a standalone LLDB build please build libcxx and
  point LLDB_TEST_LIBCXX_ROOT_DIR to it.

[edymtt]

Agree, that seems related to this LLVM PR

[edymtt]

Double checked that in the generated linux toolchain we have only one rpath

$ readelf -d usr/bin/swift-frontend | grep -E "File:|runpath"
 0x000000000000001d (RUNPATH)            Library runpath: [$ORIGIN/../lib/swift/linux]

[edymtt]

LGTM, thanks for noticing this!

[finagolfin]

Thanks @edymtt, I will submit this for the release branches next.