[llvm-commits] [llvm] r169728 - in /llvm/trunk: include/llvm/Analysis/PtrUseVisitor.h lib/Analysis/CMakeLists.txt lib/Analysis/PtrUseVisitor.cpp lib/Transforms/Scalar/SROA.cpp

[Duncan Sands]

Hi Chandler,

On 10/12/12 10:49, Chandler Carruth wrote:
> On Mon, Dec 10, 2012 at 1:45 AM, Duncan Sands <baldrick at free.fr
> <mailto:baldrick at free.fr>> wrote:
>
>     Hi Chandler,
>
>
>                  +    /// \brief Mark the visit as aborted. Intended for use in
>         a void
>                  return.
>
>
>              Intended for use in a void return?  What does this mean?  What
>         happens if I is
>              used for an instruction that is not a void return?
>
>
>         The 'set' methods are intended to work cleanly with code like:
>
>         void visitMyInstruction(__MyInstruction &I) {
>             // .. some code ..
>
>             if (detectErrorCondition())
>               return PI.setAborted(&I);
>
>             // .. rest of my code ..
>         }
>
>         Thus, they are designed to be called within a void-type return statement
>         by the
>         implementor of the visitor. Does that clarify? It has nothing to do with the
>         return of the LLVM IR under analysis.
>
>
>     are you saying that the user should not continue to use the visitor after having
>     called setAborted, and should bail out at once?
>
>
> Nothing breaks if the user continues to use it. Setting abort just halts the
> visit across the uses. Often this is done when the user is no longer interested
> in the results, including the visit currently taking place, so this is expected
> to be a common pattern but not necessarily a required pattern.

in that case, why not drop the "they are designed to be called within a
void-type return statement" comments altogether?  At worst someone will write

   if (detectErrorCondition()) {
     PI.setAborted(&I);
     return;
   }

rather than your neater version, but this is harmless.

Ciao, Duncan.