[LLVMdev] some superoptimizer results

[John Regehr]

> Another option would be exclude extensions of i1 and non cmov-like selects 
> from the candidate set.

Sorry, can you explain what a non-cmov-like select looks like?

OK, you and Sean have convinced me about the condition codes-- I'll assign 
cost >1 to sext/zext of i1 and we'll see how that works out.

> - Give each entry a unique ID (result set 3 #25 would be fine). That would 
> make them much easier to discuss.

Oops-- I had intended to give permalinks to the optimizations but somehow 
dropped that, will do it next time.

> - Exclude examples where only a prefix changes.  There were several instances 
> where the last instruction in the sequence was duplicated exactly.  This is 
> essentially an instance of synthesizing the subproblem and results in 
> duplicates in the output.

Weird, I read the entire list and totally overlooked this, will filter 
next time.

> - Group all inputs at the beginning.  Follow with an instructions which are 
> shared between input and output.  This'll make it easier to scan the output.

Yeah. Alternately, we already translate each Souper LHS into LLVM-- I 
could put the LLVM code on the web instead of the Souper code, then you'd 
have inputs as function arguments, if that's easier on the eyes?

Thanks!

John