Re: [PATCH] compiler.h: Fix barrier_data() on clang

[Nick Desaulniers <ndesaulniers@google.com>]

On Thu, Oct 15, 2020 at 8:39 AM Arvind Sankar <nivedita@alum.mit.edu> wrote:
>
> On Thu, Oct 15, 2020 at 03:24:09PM +0000, David Laight wrote:
> > > I think the comment is unclear now that you bring it up, but the problem
> > > it actually addresses is not that the data is held in registers: in the
> > > sha256_transform() case mentioned in the commit message, for example,
> > > the data is in fact in memory even before this change (it's a 256-byte
> > > array), and that together with the memory clobber is enough for gcc to
> > > assume that the asm might use it. But with clang, if the address of that
> > > data has never escaped -- in this case the data is a local variable
> > > whose address was never passed out of the function -- the compiler
> > > assumes that the asm cannot possibly depend on that memory, because it
> > > has no way of getting its address.
> >
> > Ok, slightly different from what i thought.
> > But the current comment is just wrong.
>
> Should I fix up the comment in the same commit, or do a second one after
> moving the macro?

I would prefer undoing the mistake from 815f0ddb346c and getting that
backported to stable, then rewriting comments or the trick to retain
the memset separately.

>
> > > i.e. something like:
> > >     static inline void barrier_data(void *ptr, size_t size)
> > >     {
> > >             asm volatile("" : "+m"(*(char (*)[size])ptr));
> >
> > I think it has to be a struct with an array member?
>
> I don't think so, this is actually an example in gcc's documentation:
>
>   An x86 example where the string memory argument is of unknown length.
>
>         asm("repne scasb"
>         : "=c" (count), "+D" (p)
>         : "m" (*(const char (*)[]) p), "0" (-1), "a" (0));
>
>   If you know the above will only be reading a ten byte array then you
>   could instead use a memory input like: "m" (*(const char (*)[10]) p).
>
> >
> > >     }
> > > plus some magic to disable the VLA warning, otherwise it causes a build
> > > error.
> >
> > It shouldn't if the size is a compile time constant.
> > And given this is an instruction to the compiler it better be.
>
> Ah right. I saw the warning when playing with something else where size
> could be constant or variable depending on the call.
>
> > >
> > > With a memory clobber, the compiler has to keep x and y at different
> > > addresses, since the first barrier_data() might have saved the address
> > > of x.
> >
> > Maybe "+m"(*ptr) : "r"(ptr) would work.
>
> Nothing that can only modify what ptr points to could avoid this, since
> that storage is dead after the barrier.
>
> > OTOH a "memory" clobber at the bottom of a function isn't going to
> > cause bloat.
> >
> > The explicit ranged memory access (without "memory") probably has its
> > uses - but only if the full "memory" clobber causes grief.
> >
> >       David
> >
> > -
> > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> > Registration No: 1397386 (Wales)



-- 
Thanks,
~Nick Desaulniers