[llvm-commits] [llvm] r139517 - in /llvm/trunk: lib/Transforms/Scalar/ test/Transforms/IndVarSimplify/ test/Transforms/LoopUnroll/

[Tobias Grosser]

On 09/12/2011 11:55 PM, Andrew Trick wrote:
> On Sep 12, 2011, at 1:31 PM, Tobias Grosser wrote:
>>> I don't know much about Polly, but there's no question that dependence analysis should make use of SCEV.
>> I fully agree. We do not use it for any analysis, but for final code generation. Here it seemed a lot simpler to rewrite based on canonical induction variables. I will have another look and see if we can find an approach that does not rely on the induction variables being
>> canonical.
>>
>> Thanks for your explanation
>> Tobi
>
> It sounds like you want to stick with the default mode for llvm optimization passes. You can either call the SCEV rewriter in your own backend, analogous to CodeGen prepare for machine targets, our you can take a look at the SCEV rewriter logic and adapt your backend make use of SCEV expressions.
Yes, I actually based my first implementation on the scev rewriter, but 
had some issues with it. I will reevaluate and ask for help when I find 
problems again.

> I'm just about to change the default indvars mode. We need to do this to begin making progress on long standing issues in LSR and gep rewriting. In the short term, I'll try not to break the -enable-iv-rewrite mode, but I'm hoping you won't have much trouble switching over.
I hope so too. Please go ahead. We should be able to temporarily 
reenable it for our work.

Cheers
Tobi