Re: [RFC v5 00/57] objtool: Add support for arm64

[Nick Desaulniers]

On Tue, Jan 21, 2020 at 2:31 AM Will Deacon <will@xxxxxxxxxx> wrote:
>
> On Mon, Jan 13, 2020 at 07:57:48AM +0000, Julien Thierry wrote:
> > On 1/12/20 8:42 AM, Nathan Chancellor wrote:
> > > The 0day bot reported a couple of issues with clang with this series;
> > > the full report is available here (clang reports are only sent to our
> > > mailing lists for manual triage for the time being):
> > >
> > > https://groups.google.com/d/msg/clang-built-linux/MJbl_xPxawg/mWjgDgZgBwAJ
> > >
> >
> > Thanks, I'll have a look at those.
> >
> > > The first obvious issue is that this series appears to depend on a GCC
> > > plugin? I'll be quite honest, objtool and everything it does is rather
> > > over my head but I see this warning during configuration (allyesconfig):
> > >
> > > WARNING: unmet direct dependencies detected for GCC_PLUGIN_SWITCH_TABLES
> > >    Depends on [n]: GCC_PLUGINS [=n] && ARM64 [=y]
> > >      Selected by [y]:
> > >        - ARM64 [=y] && STACK_VALIDATION [=y]
> > >
> > > Followed by the actual error:
> > >
> > > error: unable to load plugin
> > > './scripts/gcc-plugins/arm64_switch_table_detection_plugin.so':
> > > './scripts/gcc-plugins/arm64_switch_table_detection_plugin.so: cannot
> > > open shared object file: No such file or directory'
> > >
> > > If this plugin is absolutely necessary and can't be implemented in
> > > another way so that clang can be used, seems like STACK_VALIDATION
> > > should only be selected on ARM64 when CONFIG_CC_IS_GCC is not zero.
> > >
> >
> > So currently the plugin is necessary for proper validation. One option can
> > be to just let objtool output false positives on files containing jump
> > tables when the plugin cannot be used. But overall I guess it makes more
> > sense to disable stack validation for non-gcc builds, for now.
>
> Alternatively, could we add '-fno-jump-tables' to the KBUILD_CFLAGS if
> STACK_VALIDATION is selected but we're not using GCC? Is that sufficient
> to prevent generation of these things?

Surely we wouldn't want to replace jump tables with long chains of
comparisons just because objtool couldn't validate jump tables without
a GCC plugin for aarch64 for some reason, right?  objtool validation
is valuable, but tying runtime performance to a GCC plugin used for
validation seems bad.
-- 
Thanks,
~Nick Desaulniers