William Mahoney <wmahoney@xxxxxxxxxxxxxxxx> writes:
> I have moved the pass so that it is immediately after pass_ira - no
> difference, I still see register 75 (and 74). Also, df.c doesn't appear
to
> be in my source - I assume it is now df-core.c which does have quite the
> documentation and I will take a look. Are you basically saying that
after
> the hard assignments the live sets are not necessarily valid any more?
> That would be .. oh, the word that comes to mind is "bad"!
As far as I know the live sets should be valid provided you call
df_live_add_problem and df_analyze.
Ian
Resolved; and Ian, FYI you're right but the issue is that in the new IRA
code, it does not do the df_live... and df_analyze steps unless you are
optimizing. This makes sense because there's no use wasting the time -
except that in my case I need the live-in sets to be accurate even if
you're not optimizing. Interestingly, it does do at least a minimal amount
even with no optimization because the original problem was that rax was
live but not marked as such. Regardless, this does the trick in ira.c:
df_finish_pass (true);
/* if (optimize > 1) -- Bill was here */
df_live_add_problem ();
df_scan_alloc (NULL);
df_scan_blocks ();
/* if (optimize) -- Bill was here */
df_analyze ();
Kudos to Vladimir for pointing that out.
Bill
