On Fri, 17 Mar 2023 06:33:43 GMT, Roman Kennke <rken...@openjdk.org> wrote:
> > > > In my last changes I made a stupid mistake and don't set the condition > > > flags correctly to force the slow-path, on aarch64. This is only relevant > > > when we exceed the lock-stack capacity, that is why it's failing so > > > rarely. I don't see a similar problem on x86_64 - have we observed any > > > failures on x86_64? I pushed a fix for aarch64. > > > > > > I noticed this too for arm; I used cmp to clear EQ but using tst seems > > better. I also do it inside fast_lock, to give it a defined exit state wrt > > EQ|NE, since it saves me from having to think about this on every call > > site. But at least the fail case may be fiddly without conditional > > execution. > > Cmp(r,r) would not clear EQ, but set it. Unless you do cmp(r,0) on a non-null > register. Sure. I used cmp with an immediate that I knew was not the value. Clunky, I know. As I wrote, tst seems better. > Tst is better at least on x86 because it encodes smaller. _shrugs_ > > You can do it in the shared fast_lock() but it's really only needed in C2, > that's why I'm doing it there. Maybe I'm too perfectionist when it comes to > assembly code? I felt just better having it there, at least for the start. I may still move it outside to C2. Lets see. ------------- PR: https://git.openjdk.org/jdk/pull/10907
