[llvm] r228265 - Implement new heuristic for complete loop unrolling.

[Chandler Carruth]

On Fri, Feb 13, 2015 at 11:52 AM, Michael Zolotukhin <mzolotukhin at apple.com>
wrote:

> Thank you for the work and pointing out to the problems! It’s good that we
> found them early.
>
> I think that we would solve these issues if we do the following:
> 1. When looking for simplified instructions, visit each instruction only
> once. That’ll be achieved by visiting blocks in pre-order.
> 2. When looking for dead-instructions, also visit them only once - here we
> would need to visit blocks in reverse-order.
> 3. SCEV indeed might be too slow for this, but we actually only interested
> in expressions that become constants. That means, we need SCEV for
> expressions that either feed a load, or have constant Start and Step
> SCEV-expressions. We can scan all insns once and cache results for such
> potentially-constant instructions, other instructions don’t need
> SCEV-expression. So, we won’t need to talk to SCEV after this point.
> 4. As for using threshold for early-exit - that’s a really good idea, and
> it should integrate here naturally.
>
> What do you think about this? Am I missing something, or does it sound
> right?
>

Generally yes, but I would start without trying to handle dead
instructions, and following the sketch of how to build it I gave. I would
then layer dead instructions on as a follow-up. There are several different
approaches to handling dead instructions, and I think it will be useful to
see how things look without handling them first.

Note that the outline I gave naturally will skip dead *control flow*,
because we can simplify the condition to a constant and only traverse one
of the successors.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150213/89bba9c5/attachment.html>