Re: [PATCH] ARM: mtd-xip: work around clang/llvm bug

[Arnd Bergmann]

On Wed, Jul 10, 2019 at 12:25 AM Russell King - ARM Linux admin
<linux@xxxxxxxxxxxxxxx> wrote:
>
> On Tue, Jul 09, 2019 at 08:40:51PM +0200, Arnd Bergmann wrote:
> > On Tue, Jul 9, 2019 at 8:08 PM 'Nick Desaulniers' via Clang Built
> > Linux <clang-built-linux@xxxxxxxxxxxxxxxx> wrote:
> > > On Tue, Jul 9, 2019 at 1:41 AM Linus Walleij <linus.walleij@xxxxxxxxxx> wrote:
> > >
> > > > I guess this brings up the old question whether the compiler should
> > > > be worked around or just considered immature, but as it happens this
> > >
> > > Definitely a balancing act; we prioritize work based on what's
> > > feasible to work around vs must be implemented.  A lot of my time is
> > > going into validation of asm goto right now, but others are ramping up
> > > on the integrated assembler (clang itself can be invoked as a
> > > substitute for GNU AS; but there's not enough support to do `make
> > > AS=clang` for the kernel just yet).
> >
> > Note that this bug is the same with both gas and AS=clang, which also
> > indicates that clang implemented the undocumented .rep directive
> > for compatibility.
> >
> > Overall I think we are at the point where building the kernel with clang-8
> > is mature enough that we should work around bugs like this where it is
> > easy enough rather than waiting for clang-9.
>
> While both assemblers seem to support both .rept and .rep, might it
> be an idea to check what the clang-8 situation is with .rept ?

Good idea. I tried this patch now:

--- a/arch/arm/include/asm/mtd-xip.h
+++ b/arch/arm/include/asm/mtd-xip.h
@@ -15,6 +15,6 @@
 #include <mach/mtd-xip.h>

 /* fill instruction prefetch */
-#define xip_iprefetch()        do { asm volatile (".rep 8; nop;
.endr"); } while (0)
+#define xip_iprefetch()        do { asm volatile (".rept 8; nop;
.endr"); } while (0)

 #endif /* __ARM_MTD_XIP_H__ */

Unfortunately that has no effect, clang treats them both the same way.

      Arnd