Re: [PATCH 2/2] x86/asm: Fix inline asm call constraints for clang

[Josh Poimboeuf]

On Tue, Sep 19, 2017 at 08:18:23PM -0500, Josh Poimboeuf wrote:
> On Tue, Sep 19, 2017 at 01:45:28PM -0500, Josh Poimboeuf wrote:
> > For inline asm statements which have a CALL instruction, we list the
> > stack pointer as a constraint to convince GCC to ensure the frame
> > pointer is set up first:
> > 
> >   static inline void foo()
> >   {
> >   	register void *__sp asm(_ASM_SP);
> >   	asm("call bar" : "+r" (__sp))
> >   }
> > 
> > Unfortunately, that pattern causes clang to corrupt the stack pointer.
> > 
> > There's actually an easier way to achieve the same goal in GCC, without
> > causing trouble for clang.  If we declare the stack pointer register
> > variable as a global variable, and remove the constraint altogether,
> > that convinces GCC to always set up the frame pointer before inserting
> > *any* inline asm.
> > 
> > It basically acts as if *every* inline asm statement has a CALL
> > instruction.  It's a bit overkill, but the performance impact should be
> > negligible.
> > 
> > Here are the vmlinux .text size differences with the following configs
> > on GCC:
> > 
> > - defconfig
> > - defconfig without frame pointers
> > - Fedora distro config
> > - Fedora distro config without frame pointers
> > 
> > 	defconfig	defconfig-nofp	distro		distro-nofp
> > before	9796300		9466764		9076191		8789745
> > after	9796941		9462859		9076381		8785325
> > 
> > With frame pointers, the text size increases slightly.  Without frame
> > pointers, the text size decreases, and a little more significantly.
> > 
> > Reported-by: Matthias Kaehlcke <mka@xxxxxxxxxxxx>
> > Signed-off-by: Josh Poimboeuf <jpoimboe@xxxxxxxxxx>
> 
> NAK - kbuild bot is reporting some cases where this patch doesn't force
> the frame pointer setup.

So it turns out that for GCC 7, it works as described above: the global
register variable results in the frame pointer getting set up before
*all* inline asm.

But for GCC 6, it doesn't work that way.  The global register variable
has no such effect.

So we need both the global register variable *and* the output constraint
after all.  Will post another patch after some more testing.

-- 
Josh