[llvm] r240787 - [DAGCombine] fold (X >>?, exact C1) << C2 --> X << (C2-C1)

[Benjamin Kramer]

> On 26.06.2015, at 17:27, Tobias Grosser <tobias at grosser.es> wrote:
> 
> On 06/26/2015 04:51 PM, Benjamin Kramer wrote:
>> Author: d0k
>> Date: Fri Jun 26 09:51:36 2015
>> New Revision: 240787
>> 
>> URL: http://llvm.org/viewvc/llvm-project?rev=240787&view=rev
>> Log:
>> [DAGCombine] fold (X >>?,exact C1) << C2 --> X << (C2-C1)
>> 
>> Instcombine also does this but many opportunities only become visible
>> after GEPs are lowered.
> 
> The Polly bots report a recent miscompile, with this and the following commit looking to be the most likely ones to cause this.
> 
> http://lab.llvm.org:8011/builders/perf-x86_64-penryn-O3-polly-fast/builds/8707
> 
> The plain clang -O3 bot did not reach this state yet, but I will have a look at it later today:
> 
> http://lab.llvm.org:8011/builders/perf-x86_64-penryn-O3/builds/5315
> 
> (It runs 10 iterations and should be finished in about 5 hours. If we just want to check for correctness, within one hour the first iteration should be done and we can check the logs for a failure).

Fixed in r240796, thanks for watching the bots.

- Ben