[LLVMdev] New Register Allocation Algorithm
Fernando Magno Quintao Pereira
fernando at CS.UCLA.EDU
Wed Jul 18 22:22:56 PDT 2007
> From the web page, it appears it's not currently passing 100% of the
> llvm tests. What's your plan to get it there?
There are some benchmarks that I cannot compile, but they are very few. I
am still improving the implementation, and each day I get some more
compiled. Some parts of the register allocator are very complex, like
implementing swaps for registers that may alias, reusing colors across the
dominance tree, etc, and bugs are still likely to linger in the code for a
while, but I will be working on the allocator, and if someone else is
interested, I can help in understanding the code and the concepts that
underline it.
> Do you have a plan to improve the MachineBasicBlock dominator pass?
> This is something we can use right away. :-)
Yes, I really want to improve that pass, but not right away. Quadratic
structures are very bad, you know. Currently, to compute the dominator
tree takes longer than to perform register allocation. I will fix that as
soon as I have time.
Thanks for looking into it,
Fernando
More information about the llvm-dev
mailing list