On Tue, Apr 15, 2014 at 03:41:55PM +0200, Joerg Sonnenberger wrote:
>
> I remember a discussion about this topic from the LLVM lists and the
> reasons for the standard language on this are extremely weak. IIRC the
> *only* justification was for some platforms with broken (trapping)
> prefetch instructions.
That can only possibly matter if memcpy() is implemented with something
(equivalent to) 'rep movsb' - otherwise the broken prefetch would happen
during normal code sequences.
I'd have thought a zero length memcpy() would also be valid with (the valid)
pointer to 'just beyond' and item. A prefetch there could also fault.
David
--
David Laight: da...@l8s.co.uk
