From: Josh Poimboeuf
> Sent: 30 August 2019 16:02
> On Thu, Aug 29, 2019 at 03:26:30PM -0700, Linus Torvalds wrote:
> > On Thu, Aug 29, 2019 at 1:22 PM Arnd Bergmann <a...@arndb.de> wrote:
> > >
> > > Maybe we can just pass -fno-builtin-memcpy -fno-builtin-memset
> > > for clang when CONFIG_KASAN is set and hope for the best?
> >
> > I really hate how that disables conversions both ways, which is kind
> > of pointless and wrong. It's really just "we don't want surprising
> > memcpy calls for single writes".
> >
> > Disabling all the *good* "optimize memset/memcpy" cases is really sad.
> >
> > We actually have a lot of small structures in the kernel on purpose
> > (often for type safety), and I bet we use memcpy on them on purpose at
> > times. I'd hate to see that become a function call rather than "copy
> > two words by hand".
> >
> > Even for KASAN.
> >
> > And I guess that when the compiler sees 20+ "set to zero" it's quite
> > reasonable to say "just turn it into a memset".
>
> For KASAN, the Clang threshold for inserting memset() is *2* consecutive
> writes instead of 17. Isn't that likely to cause tearing-related
> surprises?
Hmmm... I don't think I'd ever want a compiler to convert a sequence
of zero writes into a memset.
It is as bad as a certain compiler converting:
for (i = 0; i < n; i++)
tgt[i] = src[i];
into a 'rep movs' instruction.
Inlining memcpy() as 'rep movs' is one thing, the opposite is wrong.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT,
UK
Registration No: 1397386 (Wales)
