[LLVMdev] fix for loop scale limiting in BFI
chandlerc at google.com
Sun Mar 29 12:32:51 PDT 2015
On Sun, Mar 29, 2015 at 12:22 PM, Xinliang David Li <xinliangli at gmail.com>
> Special casing infinite loops sounds reasonable for now. Maybe just stick
> to the current 4096 scale limit for them. Infinite loops do exist (e.g. in
> server code), but I don't yet have data to show how important it is to
> handle them more precisely.
My preference remains to handle infinite loops as a symbolic entity, and
then pick a value for it only when mapping to the fixed range, and then we
should pick an arbitrary scale higher than the next-highest non-infinite
frequency. I agree that 4096 seems as good of a choice as any other.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-dev