The G1 post-barrier is quite large, containing several if's.

E.g. for

x.a = y;

if (region(@x.a) == region(y)) exit
if (y == nullptr) exit
if (card(@x.a) == Young) exit
[...]
if (card(@x.a) == Dirty) exit
[...]
if (dcqs is full) goto veryslowpath
[...]

All of these conditions are inlined into the C2 node tree with fixed static probabilities; some old investigation however showed that they differ quite a bit depending on the application (well, benchmarks).

Investigate whether profiling this branches helps performance by inserting appropriate counters into interpreter/c1 and use them for c2 compilation.

This is the same request as JDK-8130918. JDK-8225776 shows some manual branch-order changes make sense, but it would be interesting to see if automatic adaptation gives better results.