On Fri, Feb 09, 2024, Linus Torvalds wrote:
> Sean? Does this work for the case you noticed?
Yep. You can quite literally see the effect of the asm(""). A "good" sequence
directly propagates the result from the VMREAD's destination register to its
final destination
<+1756>: mov $0x280e,%r13d
<+1762>: vmread %r13,%r13
<+1766>: jbe 0x209fa <sync_vmcs02_to_vmcs12+1834>
<+1768>: mov %r13,0xe8(%rbx)
whereas the "bad" sequence bounces through a different register.
<+1780>: mov $0x2810,%eax
<+1785>: vmread %rax,%rax
<+1788>: jbe 0x209e4 <sync_vmcs02_to_vmcs12+1812>
<+1790>: mov %rax,%r12
<+1793>: mov %r12,0xf0(%rbx)
> That (b) is very much voodoo programming, but it matches the old magic
> barrier thing that Jakub Jelinek suggested for the really *old* gcc
> bug wrt plain (non-output) "asm goto". The underlying bug for _that_
> was fixed long ago:
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58670
>
> We removed that for plain "asm goto" workaround a couple of years ago,
> so "asm_volatile_goto()" has been a no-op since June 2022, but this
> now resurrects that hack for the output case.
>
> I'm not loving it, but Sean seemed to confirm that it fixes the code
> generation problem, so ...
Yeah, I'm in the same boat.
