This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

Re: Iterative dataflow function

> So i got tired of seeing 8 million different implementations of
> worklist algorithms, and made a function that performs iterative
> bitvector dataflow for us.
> Is something like this useful enough? It seriously cuts down on the
> amount of code in most of our dataflow algorithms (LCM, etc).
Yes, Yes, YES. (IMO:)
Avoiding any code duplication in gcc is wonderfull IMO!
We should be curefull to not introduce so much of it.

In this case the duplicated hunk is small, but yet usefull and may be
optimized if dataflow turns out to be performance bottleneck in future.

> I've already got it doing iteration in reverse completion order, using a
> fibheap based priority queue as the worklist (df_visit_next is, right
> now, O(n)).

Adding fibheaps themselves to the libiberty can be wonderfull addition too.
I have some other palces, where fibheap is best fitting datastructure, but
my plan was to avoid using it, just as I am lazy to implement it and sorting
woud work just "well enought".


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]