On 06/07/2011 07:39 PM, Alexander Monakov wrote:
>
>
> On Fri, 3 Jun 2011, Bernd Schmidt wrote:
>
>>>> Ok. Although I wonder how sel-sched can end up reusing an entry in
>>>> h_d_i_d?
>
> Unlike Haifa scheduler, we recompute INSN_LUIDs for each region. However, we
> call sched_deps_{init,finish} once per function (like Haifa) and that makes us
> reuse entries in h_d_i_d.
>
> The proposed patch clears h_d_i_d when finishing a region in sel-sched.
> Bootstrapped and regtested on ia64-linux, OK for trunk?
You probably know best, so I'll approve it. Thanks for working on this.
> Normal forward scan is only done to compute insn priorities; now, it will also
> reset INSN_CONDs for instructions followed by an assignment to their
> predicate (for exposed-pipeline targets). After that, sel-sched's dependency
> analysis will treat such instructions as if they had no predicate, which is
> very conservative. Unfortunately, correctness still may be broken after a
> predicate assignment is moved into a different BB (or copied to one as a
> bookkeeping copy). I'll try to think of a way to fix it when preparing the
> predication patch.
It's probably not urgent. We can't really use sel-sched on c6x, as I'd
have to add the backtracking machinery to sel-sched for scheduling delay
slots, and I don't understand it well enough to do so.
Bernd
