Avoid string instructions in x86_64 memset

Avoid string instructions in x86_64 memset

We want to avoid revealing the sandbox base address in newlib code; memset
currently uses stos, (which leaves the full 64 bits in rdi) and does not clear
it before returning to user code. Instead of just clearing it, we remove
the stos instructions because if we want to implement a validator
check in the future, it will be easier to just ban them.

Also:
* Optimize the unrolled loop for large aligned buffers by using rbp
  as a base register.
* Remove non-temporal store instructions (and the fence that separated them
  from the subsequent stores).

R=jvoung@chromium.org, mseaborn@chromium.org
BUG=  https://code.google.com/p/nativeclient/issues/detail?id=4088

Committed: https://git.chromium.org/gitweb?p=native_client/nacl-newlib.git;a=commit;h=8c4da477c5348743d900307ce8443da4cc2fcdb8

[jvoung (off chromium), 2015-02-27 20:33:52]

jvoung@chromium.org changed reviewers:
+ jvoung@chromium.org

[jvoung (off chromium), 2015-02-27 20:33:53]

lgtm

https://codereview.chromium.org/957063002/diff/1/newlib/libc/machine/x86_64/m...
File newlib/libc/machine/x86_64/memset.S (right):

https://codereview.chromium.org/957063002/diff/1/newlib/libc/machine/x86_64/m...
newlib/libc/machine/x86_64/memset.S:74: leal    128 (rdi), edi
Might have been able to add to ebp and then restore r15 together w/ a custom
bundle_lock/unlock, rather than go through edi (if you start the loop after
setting rbp to edi + 15)?

[Derek Schuff, 2015-02-27 20:56:53]

Committed patchset #2 (id:20001) manually as
8c4da477c5348743d900307ce8443da4cc2fcdb8 (tree was closed).

[Derek Schuff, 2015-02-27 21:03:19]

https://codereview.chromium.org/957063002/diff/1/newlib/libc/machine/x86_64/m...
File newlib/libc/machine/x86_64/memset.S (right):

https://codereview.chromium.org/957063002/diff/1/newlib/libc/machine/x86_64/m...
newlib/libc/machine/x86_64/memset.S:74: leal    128 (rdi), edi
On 2015/02/27 20:33:52, jvoung wrote:
> Might have been able to add to ebp and then restore r15 together w/ a custom
> bundle_lock/unlock, rather than go through edi (if you start the loop after
> setting rbp to edi + 15)?

I like it. Done.

[Mark Seaborn, 2015-03-09 17:57:28]

For context: As with https://codereview.chromium.org/917463003/, the reason for
removing the non-temporal memory accesses here is just in case they can be used
for row hammering.