[new-regalloc] Dataflow too slow...
Geert Bosch
bosch@gnat.com
Sat Feb 3 15:13:00 GMT 2001
On Sun, 04 Feb 2001 12:03:34 +1300 (NZDT), Michael Hayes wrote:
I used a simple worklist algorithm for this to get the ball rolling.
The usual solution to impove the performance is to order the blocks
with a depth first search. This was on my todo list.
Yes, I see. It seems that indeed in this case there could be O(N)
passes, even though using reverse post-order this would probably
be at most 5 passes. This is the big gain...
As for the esoteric algorithms, we first need to see what speed we
have with the dfs approach.
-Geert
More information about the Gcc
mailing list