Re: Performance of records with variant parts

[John Perry <john.perry@usm.edu>]

On Friday, March 26, 2021 at 10:38:22 AM UTC-5, Stephen Leake wrote:
> As a workaround, you can suppress checks in that block of code: 
> 
> pragma Suppress (Discriminant_Check); 

Sorry, I should have updated earlier. I discovered tried this a couple of nights ago, and it did nothing to help. I checked the IR and that did change, so it certainly did something; it just didn't improve the performance as I expected. It has made me think that perhaps the problem is elsewhere, but I have no idea where.

So, thank you very much for remembering this and getting back to me! But by now I've concluded that gprof misled me. Wouldn't be the first time. The difference in performance is real, but must be caused elsewhere. I don't even feel confidence submitting an enhancement request to AdaCore because by now I have no idea what the problem is.

For what it's worth: I reimplemented it a fourth time as tagged records, and while the code is quite a bit wordier, it was more efficient than variant records.

(I also tried examining the generated assembly code, but I'm not strong enough in assembly to work that out.)