Re: [ANN] LuaJIT Roadmap 2012/2013
- From: Mike Pall <mike-1206@xxxxxxxxxx>
- To: luajit@xxxxxxxxxxxxx
- Date: Thu, 14 Jun 2012 12:31:46 +0200
Geoff Leyland wrote:
> On 14/06/2012, at 1:20 PM, Mike Pall wrote:
> > Yes, more or less. Inside the loop everything will be computed in
> > registers. There's still a 'point' allocation at the start of the
> > loop. And another 'point' object will be materialized from the
> > registers when the loop exits (it'll be eliminated if/when that
> > side trace is compiled eventually).
>
> Cool! Are the two allocations because the trace starts most
> likely starts on the "for i" rather than the line before and
> ends at the end of the loop?
The trace for a 'for' loop is initiated by a hot 'FORL'
instruction. Trace recording starts at the _next_ instruction,
which is the first instruction in the loop body. The trace ends
after it hits the 'FORL' instruction again.
The code before a 'for' loop, including the 'FORI' setup, is only
compiled when a second trace ends up joining to the inner loop.
But then it's too late to change the logic of the first trace,
which expects a table (or a boxed object) and not separate values.
--Mike
Other related posts:
