[LLVMdev] Predicated Vector Operations

[Chandler Carruth]

On Thu, May 9, 2013 at 6:16 AM, Nadav Rotem <nrotem at apple.com> wrote:

>
>
> I'm not sure I understand the full impact of this example, and I would
> like to.
>
> What are the desired memory model semantics for a masked store?
> Specifically, let me suppose a simplified vector model of <2 x i64> on an
> i64-word-size platform.
>
>
> Hi Chandler,
>
> I brought the example in this email thread to show that the optimizations
> that we currently have won't work on masked load/store operations because
> they don't take the mask into consideration. The memory model interesting
> question but I am not sure how it is related.
>

I understand why you originally brought up the example, but I'm trying to
dig deeper. The reason why the memory model is related is because it
drastically changes the set of options available for representing masked
loads and stores. Masked loads and stores must participate in the memory
model (just as all other loads and stores do), and we will need to update
it to reflect them if they are added to the IR. I'm trying to understand
the potential requirements of the actual code in question.


> In our example you can see the problem with a single thread.
>

Sure, but any solution is going to have to solve both single thread
problems, and concurrent problems. Also, the memory model isn't *only*
about concurrent issues. It's a fundamental description of how loads and
stores interact. That's why I presented the examples I did, and I would
very much appreciate you responding to the entirety of my email.


> Both MIC and AVX[1] have masked stores operations and they have a
> different memory model.
>

Ok, but you've only given me a link to one set of instructions, and it
doesn't really have a proper description of the memory model. If I assume
that this instruction's specification is the complete spec for AVX masked
stores, then they are atomic stores (!!!) which means that they do not form
a data race on *any* of the memory locations. This is, to say the least, a
deeply surprising and restrictive model. Certainly, it seem too strong of a
guarantee to enshrine as the *only* form of masked store in the IR.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20130512/6ebdaf84/attachment.html>