duplicate feature ft21

[JeremySinger]

Hi Grigori,

I have been playing with the static program features in featlstn.P - I
think your setup is really neat, thanks.

However I notice that you have two different definitions for ft21. At
first it is "number of unconditional branches" and later it becomes
"number of assignments". ft21 is also repeated (presumably because of
a clever shell script expansion?) in the wiki page at
http://ctuning.org/wiki/index.php/CTools:MilepostGCC:StaticFeatures
and also in your papers e.g. fmtp2008.pdf - "MILEPOST GCC: machine
learning based research compiler", page 13, Table 1.

I guess the later feature definition overrides the earlier one in the
actual databases?

Cheers,
Jeremy

[Grigori Fursin]

Hi Jeremy,

> I have been playing with the static program features in featlstn.P - I
> think your setup is really neat, thanks.

Just to mention again that the feature extractor pass including featlstn.P
has been written fully by Mircea Namolaru (in CC) from IBM Haifa, so thanks and credit
for that should go to him ;) !..

> However I notice that you have two different definitions for ft21. At
> first it is "number of unconditional branches" and later it becomes
> "number of assignments". ft21 is also repeated (presumably because of
> a clever shell script expansion?) in the wiki page at
> http://ctuning.org/wiki/index.php/CTools:MilepostGCC:StaticFeatures
> and also in your papers e.g. fmtp2008.pdf - "MILEPOST GCC: machine
> learning based research compiler", page 13, Table 1.

Ups, I didn't notice that. We will need to check it ...

> I guess the later feature definition overrides the earlier one in the
> actual databases?

Well, we record the feature vector to the database as it comes out of the
MILEPOST GCC so we just need to check what ft21 means inside MILEPOST GCC...
I just checked the http://cTuning.org/cpredict prediction page
and it seems that ft21=0 in the test example which means that it is likely
unconditional branches and not assignments.

We are extremely busy at the moment preparing a few papers, etc but me and
maybe Mircea will try to find time to check it out ...

Thanks again a lot for this feedback!
Grigori