Luca Dariz, le jeu. 02 mars 2023 09:20:15 +0100, a ecrit:
> Il 02/03/23 09:00, Samuel Thibault ha scritto:
> > Luca Dariz, le jeu. 02 mars 2023 08:55:38 +0100, a ecrit:
> > > Il 01/03/23 21:18, Samuel Thibault ha scritto:
> > > > Luca Dariz, le mer. 01 mars 2023 18:40:37 +0100, a ecrit:
> > > > > + asm volatile("wrmsr"
> > > > > + :
> > > > > + : "c" (regaddr), "a" (low), "d" (high)
> > > > > + : "memory" /* wrmsr is a serializing instruction */
> > > >
> > > > The comment could be misleading.
> > > >
> > > > The fact that it's a serialization instruction does not *require* to
> > > > express it to the compiler.
> > > >
> > > > But the fact that wrmsr needs to be a serialization instruction (because
> > > > it may depend on other writes etc.) means that one *also* wants to make
> > > > the asm snippet serialized by the compiler thanks to the memory clobber.
> > > >
> > > > So I'd rather see:
> > > >
> > > > : "memory" /* wrmsr usage needs serialization */
> > >
> > > The comment comes directly from the intel doc about WRMSR:
> > >
> > > The WRMSR instruction is a serializing instruction
> >
> > I'm not saying it's not a serializing instruction.
> >
> > I'm saying that the compiler does not have to *care* about the
> > instruction being serializing.
> >
> > But I'm also saying that the very reason why the instruction is
> > serialized is also the reason why which should give a memory clobber to
> > the compiler.
>
> I think we agree on this
>
> > > The Intel 64 and IA-32 architectures define several serializing
> > > instructions. These instructions force the processor to complete all
> > > modifications to flags, registers, and memory by previous instructions and
> > > to drain all buffered writes to memory before the next instruction is
> > > fetched and executed.
> >
> > Yes, and that is *completely* fine with the compiler not flushing
> > variables etc. to buffers before that.
> >
> > But the very reason why all these flushes are done is that wrmsr can
> > have side effect which *requires* that to be done (e.g. structure
> > preparation and whatnot), and thus we should *also* tell the compiler to
> > do the same.
>
> and on this as well.
Ok, then perhaps
"memory" /* wrmsr usage needs serialization from the compiler too */
?
Samuel
