Re: [PATCH] powerpc: Use aligned stores in memset

[Steven Munroe <munroesj@linux.vnet.ibm.com>]

On Tue, 2017-09-12 at 08:18 -0400, Zack Weinberg wrote:
> On Tue, Sep 12, 2017 at 6:30 AM, Florian Weimer <fweimer@redhat.com> wrote:
> >
> > I could not find the manual which has the requirement that the mem*
> > functions do not use unaligned accesses.  Unless they are worded in a
> > very peculiar way, right now, the GCC/glibc combination does not comply
> > with a requirement that memset & Co. can be used for device memory access.
> 
> mem* are required to behave as-if they access memory as an array of
> unsigned char.  Therefore it is valid to give them arbitrarily
> (un)aligned pointers.  The C abstract machine doesn't specifically
> contemplate the possibility of a CPU that can do unaligned word reads
> but maybe not to all memory addresses, but I would argue that if there
> is such a CPU, then mem* are obliged to cope with it.
> 
> > ...the current glibc
> > implementation accesses locations which are outside the specified object
> > boundaries.
> 
> I think that's technically a defect.  Nothing in the C standard
> licenses it to do that; we just get away with it because, on the
> implementations to date, it's not observable (unless you go past the
> end of a page, which you'll note there are a bunch of tests to ensure
> we don't do).  If an over-read by a single byte is observable, then
> mem* is not allowed to do that.
> 
Also a bit of over reaction.

As long a the library routine does no cause a visible artifact (segfault
or alignment check) aligned access before or after the requested start
address and length is an optimization.

For example accessing the source at offset 3 and length 10 with an
aligned quadword load is Ok as long I clear the leading and trailing
bytes.

But attempting to store 7 bytes within a quadword by merging bytes in a
register and storing the whole quadword would violate single copy
atomicity and is not allowed.

> zw
>