1619206 - error: initializer for thread-local variable must be a constant expression

Status: RESOLVED FIXED

(Core :: Networking: JAR, defect, P3)

Description

[Sylvestre Ledru [:Sylvestre]]

with clang trunk on a Debian:

In file included from Unified_cpp_modules_libjar0.cpp:2:
/var/lib/jenkins/workspace/firefox-clang-last/modules/libjar/MmapFaultHandler.cpp:18:43: error: initializer for thread-local variable must be a constant expression
static MOZ_THREAD_LOCAL(MmapAccessScope*) sMmapAccessScope;
                                          ^~~~~~~~~~~~~~~~
/var/lib/jenkins/workspace/firefox-clang-last/modules/libjar/MmapFaultHandler.cpp:18:43: note: use 'thread_local' to allow this

Comment 1

[(Away)]

This may be a bug in clang, the implicit = nullptr initializer for this pointer seems like an acceptable constant expression.

Comment 2

[(Away)]

It may or may not be a clang bug, but I think this was tickled by something that changed on the m-c side. I'm seeing this failure now even with 10.0.0 RC revisions that had previously given successful builds.

Comment 3

[(Away)]

Bisection points to bug 1618165, which seems unrelated to libjar. Perhaps that commit might have brought to light a latent bug in clang-10 (was it gecko's first use of deduction guides?)

Comment 4

[(Away)]

The worst part is that I can't reproduce this locally, not even with the same clang from CI!

Comment 5

[Nathan Froyd [:froydnj]]

(In reply to :dmajor from comment #3)

Bisection points to bug 1618165, which seems unrelated to libjar. Perhaps that commit might have brought to light a latent bug in clang-10 (was it gecko's first use of deduction guides?)

AFAIK, this was Gecko's first use of deduction guides (maybe Skia or ANGLE or something that's super-aggressive about C++ support might have had them first).

Comment 6

[(Away)]

Ok, by turning the diagnostic into a crash, and saving the crash reproducer from CI, I can finally reproduce this locally. I'll continue looking tomorrow.

Comment 7

[Simon Giesecke [:sg] [he/him]]

Interesting. Does MOZ_THREAD_LOCAL use a BaseAutoLock? I had a brief look, but at least couldn't find any use of it.

Comment 8

[(Away)]

Filed as https://bugs.llvm.org/show_bug.cgi?id=45124

Comment 9

[Dragana Damjanovic [:dragana]]

Michal, can you take a look?

Comment 11

[Michal Novotny [:michal]]

Looks like a bug in clang. I'm not sure what we can do in libjar.

Comment 12

[Sylvestre Ledru [:Sylvestre]]

Fixed in 11. I asked for the clang patch to be fixed in 10.

Comment 13

[Sylvestre Ledru [:Sylvestre]]

Landed in
https://github.com/llvm/llvm-project/commit/5b5a111c6794a0bb0385d04721ea854dd56da357
for the branch 10
I think we are good now