Azure & Shippable: Upgrade LDC-LLVM to v9.0.0

[kinke]

No description provided.

[kinke]

@Hardcode84: jit-rt hits deprecations such as:

jit_context.cpp:109:7: warning: 'LegacyRTDyldObjectLinkingLayer' is deprecated: ORCv1 layers (layers with the 'Legacy' prefix) are deprecated. Please use ORCv2 (see docs/ORCv2.rst)

[Hardcode84]

jit-rt hits deprecations such as

ok, I'll check it after merge, thanks

[kinke]

@Hardcode84: Thanks - there's a transitioning guide in the release notes.

The PGO counter linkage changes on Windows probably render our LLVM workaround obsolete - will check. Edit: Yes, workaround removed.

[kinke]

Okay, only remaining bullet point is how to handle libfuzzer's dropped support for trace-pc-guard (see https://reviews.llvm.org/rL352564).

[JohanEngelen]

Okay, only remaining bullet point is how to handle libfuzzer's dropped support for trace-pc-guard (see https://reviews.llvm.org/rL352564).

Let's not stress about this. We can simply drop it as well if Clang dropped support.

[kinke]

I already dropped it (as libFuzzer would just terminate the program after printing an error). The question was more wrt. how to adapt the tests - duplicate the 4 failing ones without trace-pc-guard stuff for LLVM 9+ or just remove the related checks for all LLVM versions?

[JohanEngelen]

For the trace-pc-guard stuff, this is an important Clang commit https://www.mail-archive.com/cfe-commits@lists.llvm.org/msg67181.html
So we can drop testing -fsanitize-coverage=trace-pc-guard, and change the code for -fsanitize=fuzzer to mean sancovOpts.Inline8bitCounters = true; instead of sancovOpts.TracePCGuard = true; (and add sancovOpts.PCTable =true; too).

Edit: I've used the github interface to modify the code, hope it works.
Edit2: I now realize that my code change is going to break a bunch of tests, because they all test for calls to _sanitizer_cov_trace_pc_guard. I think that is (no longer?) necessary, and fuzzing should work without it. So if we just remove all the filecheck lines checking those calls, the rest of fuzz functionality should be preserved.

[JohanEngelen]

Tested the fuzz changes locally, and fuzzing seems to work as before.

[kinke]

Thx Johan!

[JohanEngelen]

Ah you already fixed the test :)
I want to add a little bit of extra testing. I'm having trouble pushing to this PR, I thought one can now push to other people's branches if it is a PR?

[kinke]

Yes, pushing should work, the only time I wasn't able to was when Stefanos used his default branch as PR branch.

[kinke]

sanitizers/link_fuzzer.d apparently fails to fail with -link-no-cpp with LLVM 9 and shared druntime/Phobos (new Travis job)...

[JohanEngelen]

sanitizers/link_fuzzer.d apparently fails to fail with -link-no-cpp with LLVM 9 and shared druntime/Phobos (new Travis job)...

Hmm, so that simply means that c++ libs are no longer necessary to fuzz?

[kinke]

With static druntime/Phobos (Azure: building both shared/static, lit-tests only testing default static libs), the C++ lib is apparently still required, for both Linux and Mac...

[JohanEngelen]

I think the test is meant to only test that -fsanitize=fuzzer -link-no-cpp means no C++ libs (i.e. that you can override what -fsanitize=fuzzer is adding to linker line). We need something like a dry-run -v output, without actually invoking the linker.

[kinke]

Johan, you're welcome to follow-up with additional tests.