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

[Linus Walleij]

On Mon, Jul 8, 2019 at 10:31 PM Arnd Bergmann <arnd@xxxxxxxx> wrote:

> llvm gets confused by inline asm with .rep directives,

Are the LLVM developers aware of the bug?
It seems like something we can work around but should
eventually be fixed properly in LLVM, right?

> which
> can lead to miscalculating the number of instructions inside it,
> and in turn lead to an overflow for relative address calculation:
>
> /tmp/cfi_cmdset_0002-539a47.s: Assembler messages:
> /tmp/cfi_cmdset_0002-539a47.s:11288: Error: bad immediate value for offset (4100)
> /tmp/cfi_cmdset_0002-539a47.s:11289: Error: bad immediate value for offset (4100)
>
> This might be fixed in future clang versions, but is not hard
> to work around by just replacing the .rep with a series of
> eight unrolled nop instructions.
>
> Link: https://bugs.llvm.org/show_bug.cgi?id=42539
> https://godbolt.org/z/DSM2Jy
> Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>

I guess this brings up the old question whether the compiler should
be worked around or just considered immature, but as it happens this
other day I was grep:ing around to find "the 8 NOP" that is so
compulsively inserted in ARM executables (like at the very start of
the kernel execution) and I couldn't find them and now I see why.
Spelling them out makes it easier to find so:

Reviewed-by: Linus Walleij <linus.walleij@xxxxxxxxxx>

Yours,
Linus Walleij