[llvm-commits] [llvm] r123105 - in /llvm/trunk: lib/Analysis/ScalarEvolution.cpp test/Analysis/ScalarEvolution/nsw.ll

[Dan Gohman]

On Jan 8, 2011, at 6:28 PM, Chris Lattner wrote:

> Author: lattner
> Date: Sat Jan  8 20:28:48 2011
> New Revision: 123105
> 
> URL: http://llvm.org/viewvc/llvm-project?rev=123105&view=rev
> Log:
> teach SCEV analysis of PHI nodes that PHI recurences formed
> with GEP instructions are always NUW, because PHIs cannot wrap
> the end of the address space.
> 
> Modified:
>    llvm/trunk/lib/Analysis/ScalarEvolution.cpp
>    llvm/trunk/test/Analysis/ScalarEvolution/nsw.ll
> 
> Modified: llvm/trunk/lib/Analysis/ScalarEvolution.cpp
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/ScalarEvolution.cpp?rev=123105&r1=123104&r2=123105&view=diff
> ==============================================================================
> --- llvm/trunk/lib/Analysis/ScalarEvolution.cpp (original)
> +++ llvm/trunk/lib/Analysis/ScalarEvolution.cpp Sat Jan  8 20:28:48 2011
> @@ -2719,6 +2719,11 @@
>                   HasNUW = true;
>                 if (OBO->hasNoSignedWrap())
>                   HasNSW = true;
> +              } else if (isa<GEPOperator>(BEValueV)) {
> +                // If the increment is a GEP, then we know it won't perform an
> +                // unsigned overflow, because the address space cannot be
> +                // wrapped around.
> +                HasNUW = true;


Hi Chris, this isn't safe unless the GEP has the inbounds flag.
inbounds is to GEP very nearly what nuw and nsw are to add and sub.

Dan