[Cryptography] Speculation re Intel HW cockup; reqs. OS rewrites & slow execution

[Tom Mitchell]

On Thu, Jan 4, 2018 at 11:33 PM <jamesd at echeque.com> wrote:

> On 1/5/2018 2:38 PM, Bill Frantz wrote:
> > I can imagine speculative execution which does not proceed if the needed
> > data is not already in the cache.
>
> Indeed loading stuff into cache on speculative execution increases your
> working set, so arguably there are efficiency grounds for not
> speculatively hitting the cache.
>
> If you need to load stuff into cache, your speculative execution is
> going to be delayed, does not buy you much.  Worst case, your
> speculative execution is delayed by a cache miss, and then turns out to
> be wrong, then you take the the correct path - and it has a cache miss
> again...


Speculation hides loading cache and that is in part the design win.   First
level cache
is crazy fast today and actions there are an order of magnitude quicker
than main memory.
Constraining speculation to cached data and instructions is something
compilers try to do
With loop optimization including loop unrolling and memory footprints to
not trigger cache line sharing and read modify write actions to memory.

> --
Tinny keyboard.. Mobile ... I am
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.metzdowd.com/pipermail/cryptography/attachments/20180105/86195f35/attachment.html>