llvm dev - Nov 2012 - [LLVMdev] Removing nuw and nsw

Hello everybody,

currently I'm writing a thesis about compiler optimization techniques  
and I've choosen LLVM as a framework for implementing and testing my  
ideas.

While reading the LLVM Language Reference Manual I noticed that the  
nuw and nsw keywords for the instructions add, sub and mul complicate  
reasoning about integer arithmetic for the purpose of my thesis  
somewhat. Things would be easier for me if there would be just the  
plain variants of these instructions so I could formally treat them  
modular arithmetic operations. Thus, I'm wondering if it's a valid  
transformation to simply drop these keywords during an optimisation  
pass ...

As I see it, the effect of nuw resp. nsw is to create poison values in  
case of an overflow, which are essentially undefined values with the  
additional effect that they can cause other operation that depend on  
them to also produce poison values. So, as I understand it, removing  
these keywords would basically turn undefined values into defined  
ones, thus dropping some information that could be used for further  
optimizations, but it would not strictly speaking be an incorrect  
transformation - is this correct or do I miss or misunderstand  
something?

Thanks in advance for any comments and insights about this matter!

Raffael Bild

Hi Raffael,
> currently I'm writing a thesis about compiler optimization techniques
and I've
> choosen LLVM as a framework for implementing and testing my ideas.
>
> While reading the LLVM Language Reference Manual I noticed that the nuw and
nsw
> keywords for the instructions add, sub and mul complicate reasoning about
> integer arithmetic for the purpose of my thesis somewhat. Things would be
easier
> for me if there would be just the plain variants of these instructions so I
> could formally treat them modular arithmetic operations. Thus, I'm
wondering if
> it's a valid transformation to simply drop these keywords during an
optimisation
> pass ...
it is valid.  Dropping them decreases the effectiveness of the optimizers, but
it doesn't create any correctness issues.
> As I see it, the effect of nuw resp. nsw is to create poison values in case
of
> an overflow, which are essentially undefined values with the additional
effect
> that they can cause other operation that depend on them to also produce
poison
> values. So, as I understand it, removing these keywords would basically
turn
> undefined values into defined ones, thus dropping some information that
could be
> used for further optimizations, but it would not strictly speaking be an
> incorrect transformation - is this correct or do I miss or misunderstand
something?
That's right, you got it.

Ciao, Duncan.
>
> Thanks in advance for any comments and insights about this matter!
>
> Raffael Bild
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev