On Fri, Jan 19, 2018 at 10:18 AM, Will Deacon <will.deacon@xxxxxxx> wrote:
>
> On Fri, Jan 19, 2018 at 10:12:47AM -0800, Dan Williams wrote:
> > [ adding Alexei back to the cc ]
> >
> > On Fri, Jan 19, 2018 at 9:48 AM, Adam Sampson <ats@xxxxxxxxx> wrote:
> > > Jann Horn <jannh@xxxxxxxxxx> writes:
> > >
> > >>> +/*
> > >>> + * If idx is negative or if idx > size then bit 63 is set in the mask,
> > >>> + * and the value of ~(-1L) is zero. When the mask is zero, bounds check
> > >>> + * failed, array_ptr will return NULL.
> > >>> + */
> > >>> +#ifndef array_ptr_mask
> > >>> +static inline unsigned long array_ptr_mask(unsigned long idx,
> > >>> unsigned long sz)
> > >>> +{
> > >>> + return ~(long)(idx | (sz - 1 - idx)) >> (BITS_PER_LONG - 1);
> > >>> +}
> > >>> +#endif
> > >>
> > >> Nit: Maybe add a comment saying that this is equivalent to
> > >> "return ((long)idx >= 0 && idx < sz) ? ULONG_MAX : 0"?
> > >
> > > That's only true when sz < LONG_MAX, which is documented below but not
> > > here; it's also different from the asm version, which doesn't do the idx
> > > <= LONG_MAX check. So making the constraint explicit would be a good idea.
> > >
> > > From a bit of experimentation, when the top bit of sz is set, this
> > > expression, the C version and the assembler version all have different
> > > behaviour. For example, with 32-bit unsigned long:
> > >
> > > index=00000000 size=80000001: expr=ffffffff c=00000000 asm=ffffffff
> > > index=80000000 size=80000001: expr=00000000 c=00000000 asm=ffffffff
> > > index=00000000 size=a0000000: expr=ffffffff c=00000000 asm=ffffffff
> > > index=00000001 size=a0000000: expr=ffffffff c=00000000 asm=ffffffff
> > > index=fffffffe size=ffffffff: expr=00000000 c=00000000 asm=ffffffff
> > >
> > > It may be worth noting that:
> > >
> > > return 0 - ((long) (idx < sz));
> > >
> > > causes GCC, on ia32 and amd64, to generate exactly the same cmp/sbb
> > > sequence as in Linus's asm. Are there architectures where this form
> > > would allow speculation?
> >
> > We're operating on the assumption that compilers will not try to
> > introduce branches where they don't exist in the code, so if this is
> > producing identical assembly I think we should go with it and drop the
> > x86 array_ptr_mask.
>
> Branches, perhaps, but this could easily be compiled to a conditional
> select (CSEL) instruction on arm64 and that wouldn't be safe without a
> CSDB. Of course, we can do our own thing in assembly to prevent that, but
> it would mean that the generic C implementation would not be robust for us.
>
Ah, ok good to know. Likely if the current version doesn't produce a
conditional instruction on ARM perhaps it also won't do that on other
architectures, so it is safer.
