[LLVMdev] Live Intervals vs. Live Variables

[Roman Levenstein]

--- Evan Cheng <evan.cheng at apple.com> wrote:
> I do want to comment on the "conservative" nature of  
> LiveIntervalAnalysis. I think the comment is misleading and is  
> probably just a relic of early implementation. The biggest problem  
> with the current implementation is the overly aggressive copy  
> coalescer. Right now the register allocator cannot split live ranges.
>  
> So if the coalescer has formed a long live interval it will causes  
> lots of interferences (and thus lots of spills).
> 
> I would encourage those who are interested in implementing new  
> register allocators for LLVM to think about this issue. I do not  
> believe it's possible to build a noticeably "better" allocator if it 
> is relying on the existing live interval information.

You are right. Therefore, for my implementation of Wimmer's linear scan
algorithm I have rewritten the LiveIntervalAnalysis pass and
LiveInterval class. Basically, it does not do coalescing during the
LiveIntervalAnalysis stage. Instead I use register hints as indication
that two registers can be probably coalesced, or better said - can be
assigned to the same physical register. This avoids building long live
coalesced intervals that have a lot of interferences. It also speeds up
the analysis. BTW, I also do not use live interval weights computed by
the LiveIntervalAnalysis, though I might cosidering using them again
for better determination of registers to be split or spilled. 

Roman


 
____________________________________________________________________________________
It's here! Your new message!  
Get new email alerts with the free Yahoo! Toolbar.
http://tools.search.yahoo.com/toolbar/features/mail/