[Date Prev][Date Next][Thread Prev][Thread Next]
[Date Index]
[Thread Index]
- Subject: Re: [ANN] LuaJIT-2.0.0-beta1
- From: Mike Pall <mikelu-0911@...>
- Date: Sun, 1 Nov 2009 20:17:32 +0100
David Given wrote:
> At this point the illustrious reader may wish to insert my standard rant
> on the evils of languages with no 'goto' (which I'm sure you've all
> heard before).
Actually LJ2 makes extensive use of the fact that Lua doesn't have
a goto. This enables the use of a new region-selection algorithm,
which I termed NLF (natural-loop-first). IMHO it outperforms NET
(next-executed tail) and LEI (last-executed iteration) which has
previously been used for trace compilers.
ObRef:
Improving region selection in dynamic optimization systems,
D. Hiniker, Intl. Symp. on Microarchitecture, 2005
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.102.9615
> I'm assuming that all the argument passing would usually turn into no
> code in LuaJIT due to the way the trace compiler works.
Yes.
> Actually, Clue uses some of the shootout benchmarks as unit tests.
> Results follow:
Since others may be misinterpreting this: these are the results
for a C->Lua translation (and, as we have discussed, a rather
suboptimal one). The native Lua variants of these benchmarks run
*much* faster.
--Mike