[llvm-commits] [llvm] r122678 - in /llvm/trunk: lib/Transforms/Scalar/LoopIdiomRecognize.cpp test/Transforms/LoopIdiom/basic.ll

[Dan Gohman]

On Jan 1, 2011, at 7:37 PM, Chris Lattner wrote:
> 
> +  // The # stored bytes is (BECount+1)*Size.  Expand the trip count out to
> +  // pointer size if it isn't already.
> +  const Type *IntPtr = TD->getIntPtrType(SI->getContext());
> +  unsigned BESize = SE->getTypeSizeInBits(BECount->getType());
> +  if (BESize < TD->getPointerSizeInBits())
> +    BECount = SE->getZeroExtendExpr(BECount, IntPtr);
> +  else if (BESize > TD->getPointerSizeInBits())
> +    BECount = SE->getTruncateExpr(BECount, IntPtr);

This code (and other instances of it) can be simplified using
ScalarEvolution's getTruncateOrZeroExtend.

Also, truncating is surprising here, though I guess maybe it's safe
since any loop where it would change the value would be iterating
over the entire address space.

> +  
> +  const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
> +                                         true, true /*nooverflow*/);

Setting nuw here is surprising, but as above, overflow would mean
the loop is iterating over the entire address space.

Setting nsw here seems incorrect, as there's no fundamental reason why
iterating over half of the address space is invalid.

Dan