Peeling of once rolling loops limited?

[Zdenek Dvorak]

Hello,

> Why is peeling of once rolling loops limited?

performance reasons -- to avoid analysing overly large loops.  Probably
does not really matter, but also hardly spoils anything (the large
non-rolling loops are definitely rare and even if such a loop occur, we
do not gain much by optimizing it).

> And why it is limited by 
>  a parameter that is not even documented?

I probably forgot to document it :-)

> Shouldn't peeling a once 
> rolling loop not always be an improvement?

Yes.

> the above test should probably be removed together with all traces of 
> the PARAM_MAX_ONCE_PEELED_INSNS param.

Possibly; feel free to submit the patch (unless it causes compile time
regressions, but it is IMHO unlikely).

Zdenek