<anarsoul>
I'd really like to merge it ASAP to stop answering the questions where to get this patch :)
<enunes>
anarsoul: I'm a bit away from those issues, but to me it looks good to merge
<enunes>
one thing I have in my mind about the debug flags (certainly for a separate patch) is about the 'all' flag, maybe we should eventually split it so that 'all' just enables the increased-debug-output related ones and not the disable-feature-debug ones
<enunes>
or maybe just add a 'debug' option and leave 'all' as-is
<anarsoul>
yeah, it's better to add 'debug' option and leave all as is
<anarsoul>
or drop it since it doesn't make much sense
<enunes>
btw about ppir scheduler: I keep hitting corner cases with the new liveness analysis, piglit is without regressions but deqp hit some new regressions... I'm still working on it and have redesigned it a couple of times already
<enunes>
it's very tricky because some corner cases end up generating a result with variable that is written but never read or something like that
<anarsoul>
enunes: that's why we need optimizations for ppir
<anarsoul>
written but never read variables should be eliminated with DCE
<enunes>
well DCE can basically be a byproduct of the new liveness analysis that I have
<anarsoul>
yet it should be done as a separate optimization pass
<enunes>
the algorithm is exactly the same
<anarsoul>
let's stick to class backend design unless you have very good reasons not to :)
<anarsoul>
s/class/classic
<enunes>
right, the algorithm seems to be the common one from every compiler 101
<enunes>
it's mostly tricky because of the non-ssa registers and for ppir because of registers written and read in the same instruction
<enunes>
and other things that we created ourselves like cloning varying load that end up leaving the original varying load orphan
<enunes>
anyway, working on it
<enunes>
I don't think that something like DCE should be part of the backend though, dead code doesn't really happen from nir as far I have seen unless our current optimizations create it
<enunes>
I think the next step after liveness analysis is to take a look at how to improve the scheduler, it's what I plan to do next after this
<rellla>
anarsoul: i tried to debug alpha->ref_value and friends at this place but they are always 0. may i miss anything?
monstr has quit [Remote host closed the connection]