[LLVMdev] readonly and infinite loops

[Nuno Lopes]

>> People have been aware of this issue for a long time (as you can see
>> by
>> Nick's 2010 patch).  I guess it was not pushed further because of
>> lack of
>> practical interest.
>
> I can't comment about the rationale in 2010, but this came up again when I 
> was originally working on heap-to-stack conversion. Being able to mark 
> functions as halting is essential for doing heap-to-stack conversion. This 
> was put on hold, however, essentially awaiting the new pass manager. The 
> issue is that you'd like to be able to use SCEV in a module/CGSCC-level 
> pass to infer the attribute on functions with loops, and this cannot be 
> done directly until we have the new pass manager.

Interesting.  Could you give an example why knowing a function will halt is 
essential for the heap-to-stack conversion?  It's definitely an optimization 
we should be doing (although, correct me if I'm wrong, LLVM already has some 
form of this for very small mallocs, no?)

Thanks,
Nuno