Re: sequences of length 1 in inline-threaded
"Grzegorz B. Prokopski" <[email protected]>
| Newsgroups | gmane.comp.java.vm.sablevm.devel |
|---|---|
| Organization | SableVM - LGPL'ed Free Java VM http://sablevm.org |
| Message-ID | <[email protected]> |
On Fri, 2004-12-03 at 17:04, David Bélanger wrote: > I don't think it will change much performance. > > Slightly less memory used. Also, at code preparation, a lookup in the > splay tree is not needed for some of these 1-inst sequence. That's > about it. The patch makes sense to me. However, if you wanted to make some difference, you could look ie. at the conditional instructions. Many (most?) of them are technically inlinable or rather half-inlinable. They can appear as the *last* instruction of a superinstruction, but you have to break up the superinstruction right _after_ such bytecode. Currently we always break up construction of superinstruction _after_ *and* _before_ it, which is clearly suboptimal. These bytecodes may also include GC-check. I already once coded all the needed modifications to make testing of inlinability of such bytecodes possible, so I should be able to bring it back, *if* the problem outlined above was solved first. HTH, Grzegorz B. Prokopski BTW: I don't know whether you've seen this one (optimization of method preparation). Sounds like a good idea to me. http://sablevm.org/bugs/13 -- Grzegorz B. Prokopski <[email protected]> SableVM - Free, LGPL'ed Java VM http://sablevm.org Why SableVM ?!? http://sablevm.org/wiki/Features Debian GNU/Linux - the Free OS http://www.debian.org