Re: [Xen-devel] [PATCH] mm: fix LLVM code-generation issue

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Thu, Nov 22, 2018 at 12:23:22PM +0000, Andrew Cooper wrote:
> On 22/11/2018 12:03, Roger Pau Monne wrote:
> > LLVM code generation can attempt to perform a load from a variable in
> > the next condition of an expression under certain circumstances, thus
> > turning the following condition:
> >
> > if ( system_state < SYS_STATE_active && opt_bootscrub == BOOTSCRUB_IDLE )
> >
> > Into:
> >
> > 0xffff82d080223967 <+103>: cmpl   $0x3,0x37b032(%rip) # 0xffff82d08059e9a0 
> > <system_state>
> > 0xffff82d08022396e <+110>: setb   -0x29(%rbp)
> > 0xffff82d080223972 <+114>: cmpl   $0x2,0x228a8b(%rip) # 0xffff82d08044c404 
> > <opt_bootscrub>
> 
> You're actually missing two pieces here.  If I recall the disassembly
> correctly, there was a `sete %al`, and an `and %al, -0x29(%rbp)` which
> had the cumulative effect of calculating `idle_scrub` branchlessly
> (which is no doubt the intended optimisation).

Right, my point here was to merely show that opt_bootscrub is accessed
regardless of the value of system_state. I can expand the asm chunk,
but the bug reference below already has the disassembly of the whole
function.

> >
> > Such code will trigger a page fault if system_state >=
> > SYS_STATE_active.
> >
> > In order to prevent such optimization signal to the compiler that
> > accessing opt_bootscrub can have side effects by using ACCESS_ONCE.
> > This has been reported and discussed with upstream LLVM:
> >
> > https://bugs.llvm.org/show_bug.cgi?id=39707

Thanks, Roger.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel