[polly] r256199 - Add option to assume single-loop scops with sufficient compute are profitable
Tobias Grosser via llvm-commits
llvm-commits at lists.llvm.org
Tue Dec 22 05:51:40 PST 2015
On 12/22/2015 02:31 PM, Johannes Doerfert wrote:
> On 12/22, Tobias Grosser wrote:
>> On 12/22/2015 12:09 PM, Johannes Doerfert wrote:
>>> In this spirit I would like to change the default heuristic to:
>>> #loops + #blocks > 2
>>> as that is the condition under which we can perform "by default enabled"
>>> optimizations.
>>
>> If this results in run-time benefits for at least one application we care
>> about and no run-time regressions, then I think this is fine.
>>
>> (I care a little less about compile-time regressions. The more we expose
>> them, the more visible they become and the more likely we are
>> to add improvements here).
> It will at least trigger distribution on Spec2006 hmmer. I think that's
> worth it.
Sure. Just would prefer to have a look at the performance data at least
for LNT and SPEC. Most likely we are lucky and it is purely beneficial.
Best,
Tobias
More information about the llvm-commits
mailing list