[llvm] r306529 - [X86] Correct dwarf unwind information in function epilogue

[Petar Jovanovic via llvm-commits]

+ Violeta

> So, this says "[X86]", but is it really X86-specific? there is a pretty
> huge set of changes to common CFI infrastructure that is used by PPC and
> AArch64 at least.

It is x86-specific in the way that improvements made in this patch should
affect/improve x86 target only. Other arches can benefit if they make
similar changes.

> It seems like these late passes (that run on all targets from what I can
> tell) *require* the new information to be set up in the basic block...

Yes, but if the new information is not set (i.e. has default values), the
passes would not make any changes. An alternative is to run these passes for
selected target(s) only.

> How does this work? Is this not a (serious) bug?

Violeta has more description in https://reviews.llvm.org/D18046
It should not be a bug, more like a place where things can be improved for
other architectures too.

So the passes should not be the problem. The only thing I am concerned is
that other common code changes may have side effects elsewhere. There is
a failure on PPC64 sanitizer buildbot and this change is on the blame list,
so that is being investigated along with the case Eli has reported.