[LLVMdev] [RFC] Heuristic for complete loop unrolling

[Owen Anderson]

> On Jan 23, 2015, at 2:26 PM, Michael Zolotukhin <mzolotukhin at apple.com> wrote:
> 
> 
>> On Jan 23, 2015, at 1:38 PM, Owen Anderson <resistor at mac.com <mailto:resistor at mac.com>> wrote:
>> 
>> Hi Michael,
>> 
>> I’m unsure about other targets, but for the ones I work on converting register-indexed loads into constant-indexed ones is a huge win, even in situations where the load cannot be constant folded.  Do you think there is any reasonable way to incorporate that into this change?  Maybe a TTI hook?
> Hi Owen,
> 
> It’s not difficult to incorporate such customization into the patch. Actually, right now there is a magic constant in the patch which gives a weight to every potentially removed load instruction. It might make sense to make this constant target-dependent. And if we do it, we can also add a weight for loads that will become constant-indexed (not necessarily constant-folded).

That would definitely be very useful for me!

—Owen

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150124/6dae075d/attachment.html>