[LLVMdev] [Polly][GSOC2013] FastPolly -- SCOP Detection Pass

Tobias Grosser tobias at grosser.es
Mon Jul 1 08:58:06 PDT 2013

On 07/01/2013 08:18 AM, Star Tan wrote:
>>> (3) About detecting scop regions in bottom-up order.
>>> Detecting scop regions in bottom-up order can significantly speed up the scop detection pass. However, as I have discussed with Sebastian, detecting scops in bottom-up order and up-bottom order will lead to different results. As a result, we should not change the detection order.
>> Sebastian had a patch for this. Does his patch improve the scop
>> detection time.
> LNT testing results for Sebastian's patch file can be seen on (Run Order: ScopDetect130615). You can compare ScopDetect130615 (Polly with Sebastian's patch) to pOpt130615 (Polly without Sebastian's patch). The result seems not show significant performance improvements with the bottom-up patch for LLVM test-suite benchmarks.
> You are right. I think I should better firstly focus on the some simple examples, such as oggenc, to understand where scop detection pass spend its time. After that, we can then investigate the scop detection order.

This is actually surprising. I expected his patch to make a difference 
in the scop detect time.

Btw, if you rerun the experiments, it probably makes more sense to not 
modify the commit number, but instead use different lnt tester names
to report the changes. Changing the commit numbers causes crashes e.g.
when you click on the links for individual benchmarks.


More information about the llvm-dev mailing list