[LLVMdev] Whole-function isel

[Dietmar Ebner]

Evan Cheng schrieb:
> On Mar 31, 2008, at 2:37 AM, Dietmar Ebner wrote:
>> i've talked to our supporting company and they agreed to release the
>> code to interested parties. it's not a copyleft license but the code
>> can be used freely for private and research purposes.
>>
>> be warned that the code is merely a prototype implementation and not
>> ready for inclusion in LLVM. it also requires a cost augmented graph
>> grammar. for our ARM implementation, we could derive a basic set of
>> rules directly from the LLVM tablegen description, but additional
>> handwritten rules (e.g., replacing the c++ written address mode
>> selection) were necessary. the implementation has been ported from
>> another private backend which is still evident in some places.
>>
>> if you're still interested in the code, i could prepare a small  
>> package
>> to play with along with some useful instructions. if you think that a
>> PBQP based approach would be useful for LLVM and has a chance for
>> inclusion, i would be willing to bring the code in shape and make
>> changes where necessary.
> 
> Sounds like it's probably not going to be part of llvm then. :-(
> 
> Are there parts that might be introduced back?
i guess my last message was slightly misleading. the license just allows 
interested people to play with the code and do some experiments. if you 
like the method and decide that that's what you want for LLVM, then i'm 
confident that licensing issues will not be an obstacle and the code can 
be released under the usual llvm release license.

i'll try to do some minor cleanups and send you the code in the next few 
days.

cheers,

-
dietmar