On Wed, 18 Oct 2023 14:47:58 GMT, Jorn Vernee <jver...@openjdk.org> wrote:
> > I wonder if the native_invoker_size_per_arg thing still works good enough. > > We may exceed the computed size, now, right? > > Good point. I'll have a look at enhancing the test we have for this. > > Intuitively, I think it will be okay. It's true that we generate more code to > add the oops and offsets together, but at the same time, we don't have any > code to shuffle the offsets. Looks like we use 2 input regs per obj, so we reserve 2x native_invoker_size_per_arg. However, the case in which `reg_oop` and `reg_offset` are on stack together with `arg_shuffle` can produce more than this size (depending on platform). If we pass lots of objects on stack, we may exceed the computed size. Correct? ------------- PR Comment: https://git.openjdk.org/jdk/pull/16201#issuecomment-1770938705
