[llvm] r259675 - Revert r259662, which caused regressions on polly tests.

Tobias Grosser via llvm-commits llvm-commits at lists.llvm.org
Wed Feb 3 13:02:55 PST 2016


On 02/03/2016 07:05 PM, Wei Mi via llvm-commits wrote:
> Author: wmi
> Date: Wed Feb  3 12:05:57 2016
> New Revision: 259675
>
> URL: http://llvm.org/viewvc/llvm-project?rev=259675&view=rev
> Log:
> Revert r259662, which caused regressions on polly tests.
>

Hi Wei Mi,

thanks for taking care of the buildbots and also thank you for improving 
SCEV. I already committed one patch that makes Polly less
sensitive to your cleanups and have attached a patch that fixes the 
Polly test case to take into account your simplifications.

Two issues remain:

1) test/Isl/CodeGen/non-affine-phi-node-expansion.ll

In this test case LLVM after your patch actually produces more 
complicated IR then before. Specifically, it does not inline the 'true'
value any more. This is clearly a minor regression which will be cleaned 
up by later passes, but it might be of interest to you

2) Polly currently uses SCEV in a way that we need to invalidate this 
instruction cache at some point as otherwise we reference instructions
that are not reachable at the point of use. If you could add 
aninvalidateNewCache() function (maybe use a better name), I could work 
out a patch for Polly, such that we can then commit both together.

Best,
Tobias

PS: Thanks again for taking care of SCEV simplifications
1)

Best,
Tobias
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-Adjust-polly-to-SCEVBuilder-changes.patch
Type: text/x-patch
Size: 2972 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160203/db3fd892/attachment.bin>


More information about the llvm-commits mailing list