[llvm] [msan] Fix false negatives from partially undefined constant fixed-length vectors (PR #143837)

[Florian Mayer via llvm-commits]

fmayer wrote:

> > > > LGTM but please wait until we resolve the current regression
> > > 
> > > 
> > > Can I land it earlier if I change the fix to be off by default (controlled by the new flag)? (And then eventually flip the flag on by default in a separate CL.)
> > > The advantages are:
> > > 
> > > * once the fix lands (disabled by default) in toolchains, it's easier to let owners of buggy code re-run MSan with the flag enabled, rather than telling them they need to custom patch the compiler
> > > * the fix will be lightly tested in CI by the new test case
> > 
> > 
> > IMO, ship it
> 
> To clarify, ship it off by default?

Yes. That should sufficiently address Vitaly's concern and doesn't hurt.

https://github.com/llvm/llvm-project/pull/143837