Re: [PATCH] x86: work around clang IAS bug referencing __force_order

[Nick Desaulniers]

On Fri, Aug 21, 2020 at 4:04 PM Arvind Sankar <nivedita@xxxxxxxxxxxx> wrote:
>
> On Fri, Aug 21, 2020 at 02:37:48AM +0200, Thomas Gleixner wrote:
> > On Thu, Aug 20 2020 at 09:06, Arvind Sankar wrote:
> > > I don't think that's an issue, or at least, not one where force_order
> > > helps.
> > >
> > > If the source for foo() is not visible to the compiler, the only reason
> > > force_order prevents the reordering is because foo() might have
> > > references to it, but equally foo() might have volatile asm, so the
> > > reordering isn't possible anyway.
> > >
> > > If the source is visible, force_order won't prevent any reordering
> > > except across references to force_order, but the only references are
> > > from the volatile asm's which already prevent reordering.
> > >
> > > I think force_order can only help with buggy compilers, and for those it
> > > should really have been an input-output operand -- it wouldn't currently
> > > do anything to prevent cr writes from being reordered.

I agree 100%.  From the link to GCC docs, the code in question doesn't
even follow the pattern from the doc from informing the compiler of
any dependency, it just looks like !@#$.

> >
> > Fair enough. Care to provide a patch which has the collected wisdom of
> > this thread in the changelog?
> >
> > Thanks,
> >
> >         tglx
>
> The gcc bug I linked to earlier is only fixed in gcc-6 onwards. Is that

(based on https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82602#c14)

> good enough to remove force_order? I can test gcc-4.9 and gcc-5 to check
> if it would currently have any impact.

I think checking the disassemblies with a pre-gcc-6 would be good
enough then; that bug isn't specific to this particular case.

> CBL guys, can you confirm that clang also will not reorder volatile asm?

Full disassemblies of vmlinux pre vs post __force_order removal are
the same.  That's pretty good actually; I was worried for a code base
of this size whether two compiles would produce the exact same
disassemblies; I know the version strings are timestamped, for
instance, but I didn't compare data, just .text.  I should triple
check i386, and some of the ko's that use modified functions.  I'd be
happy to help provide a tested by tag for numerous configurations with
Clang.

Attaching the diff I was testing, feel free to add a commit message.
--
Thanks,
~Nick Desaulniers

Attachment: force_order.patch
Description: Binary data