[llvm-commits] [llvm] r94433 - in /llvm/trunk: include/llvm/Value.h lib/Transforms/Scalar/ScalarReplAggregates.cpp lib/VMCore/Value.cpp

[Duncan Sands]

Hi Bob,

>> if sroa has to go back through a gazillion GEPS and bitcasts to get to the
>> alloca, maybe it should just give up instead?
> 
> This is happening at a point where SROA has already decided to do the transformation.  It is just using getUnderlyingObject as a convenience.  Changing SROA to have a limit like that is possible, but it would involve more extensive changes.  We could also avoid using getUnderlyingObject by passing the alloca along as a parameter as the function recurses.  I'm not sure that's worth doing.

fair enough - thanks for the explanation.

Ciao,

Duncan.