RFR: 8310299: C2: 8275201 broke constant folding of array store check in some cases

[Sandhya Viswanathan]

On Mon, 19 Jun 2023 08:56:26 GMT, Roland Westrelin <roland at openjdk.org> wrote:

> Before 8275201, loading the element klass of an array returned:
> 
> 
> TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
> 
> 
> that is exact if the array type is exact. I changed it to:
> 
> 
> tkls->is_aryklassptr()->elem();
> 
> 
> When the array type is exact (newly allocated array for instance) but
> the element class has subclasses, this doesn't return an exact class
> (so the logic is different from the one that was there before). That
> affects array store checks that no longer constant fold.

@rwestrel Thanks a lot for this fix. It fully recovers the performance drop that we have been observing since JDK 19.

-------------

PR Comment: https://git.openjdk.org/jdk/pull/14536#issuecomment-1601848258