[cfe-dev] Enable -funwind-tables by default on Power PC (Bugzilla Bug 32611)

[Reid Kleckner via cfe-dev]

I think not generating unwind tables for nounwind functions with EH
personalities is a bug.

It's true that the program never unwinds out of func, but it internally
catches all possibly unwinding exceptions and calls __cxa_call_unexpected.
That requires .cfi directives.

It looks like this bug applies to both ppc64 and x86_64. It must be in
common prologue emission code, or all the targets have the same bug. We
shouldn't need the uwtable attribute to get this case right.

On Fri, Aug 17, 2018 at 10:47 AM Stefan Pintilie via cfe-dev <
cfe-dev at lists.llvm.org> wrote:

> Hi all,
>
> I've been looking at this bug:
> https://bugs.llvm.org/show_bug.cgi?id=32611
>
> The bug involves a function that is specified as not throwing anything but
> in reality it does throw.
>
> void func() throw() {
>   throw 100;
> }
>
> When this code is run we should see a call to __cxa_call_unexpected but we
> do not. The reason is because the function does not generate the unwind
> tables.
> func() is marked with the nounwind attribute as is expected due to the
> fact the user has told us it does not throw anything. However, that
> attribute also blocks the generation of the unwind tables which are
> required for the proper functioning of the call to unexpected. The solution
> is to also use the attribute uwtable which forces the generation of the
> unwind tables.
>
> Checking how this works on x86 I realised that this works because
> -funwind-tables is on by default for x86 and therefore adds the uwtable
> attribute to almost all of the functions being compiled.
> I've added a patch
> https://reviews.llvm.org/D50908
> that does the same thing for PowerPC as is done on x86 and it fixes the
> bug.
>
> However, I feel like this solution adds a function attribute to a whole
> set of functions where it is not needed. The vast majority of functions
> that will now receive the uwtable attribute don't actually need it.
> Has anyone encountered an issue where a function marked as nounwind
> actually requires unwind tables? Does anyone know of a better solution for
> this?
>
> Thank you,
> Stefan
>
>
> _______________________________________________
> cfe-dev mailing list
> cfe-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20180817/960d663d/attachment.html>