Experimental Physics and
Industrial Control System

Ben Franksen via Core-talk <core-talk at aps.anl.gov> · Wed, 25 Mar 2020 14:20:38 -0000

After studying this for a while I think I understand the logic now.

Caching the conversion routine from the last time a request was made on this channel makes sense only if we circumvent the call to dbGet and call the conversion routine directly. This is what the last two cases are about.

I am not sure this optimization buys a lot. The code path inside dbGet up until the point where we call the conversion routine is mostly a small number of binary decisions and a few simple scalar assignments. We don't specify any options to dbGet (the MS link option is handled separately). Saving these few lines of code doesn't strike me as worth the trouble. The only noticeable benefit this could have is that we avoid the table lookup, which /may/ allow the processor to keep the conversion routine in the cache.

Has this ever been benchmarked?

Anyway, I will try to refactor this in my write-filters branch. It should be possible to make this optimization work with less complication by factoring out the code from dbGet up to the point where we know which conversion routine we want to call and then just return that pointer.
-- 
https://code.launchpad.net/~dirk.zimoch/epics-base/+git/epics-base/+merge/378968
Your team EPICS Core Developers is subscribed to branch epics-base:7.0.

Experimental Physics and Industrial Control System

Experimental Physics and
Industrial Control System