On 2020-06-08 23:25, Sergey Bylokhov wrote:
On 6/8/20 2:17 pm, Claes Redestad wrote:
Not too similar, since that involves a compiler bug that basically
forced applying the -O1 flag to a critical piece of the runtime.>
Obviously there needs to be a pragmatic approach when hitting a compiler
bug like the one discussed in that thread, with the goal of removing the
special-case as soon as possible.
I meant the point raised by the Gerard
https://mail.openjdk.java.net/pipermail/hotspot-dev/2020-June/042021.html
The "-Os (the default for most hotspot files) takes 158s"
is not true anymore.
Right, -O3 is now likely the default for most files.
I'm not sure you suggest the reported ~2% difference between -Os and
-O2/-O3 as an argument against changing to -O3 by default. It might be
statistically significant - or just noise - but we should at any rate
be wary of overfitting to a particular (micro-)benchmark.
It'd also be good to re-run with the new default: when most compilation
units change to -O3, which optimization level is ideal for an individual
compilation unit (such as unsafe.cpp) might very well change, since
what optimizations are available through the build chain can shift
considerably.
/Claes
