[PATCH] D13716: [IndVars] Split `WidenIV::cloneIVUser`; NFC

[Sanjoy Das via llvm-commits]

sanjoy added a comment.

In http://reviews.llvm.org/D13716#267189, @reames wrote:

> I'm a bit confused by the need/desirability of this change.  Isn't it fairly simple to make the special case for widening conditional on the operator type within a single function?  Or is there something more fundamental I'm missing?


Firstly, the structure of `cloneArithmeticIVUser` will end up being a little different from `cloneBitwiseIVUser`; and I think having an "early dispatch" will make it easier to avoid bugs.  This is the practical reason why I split the two.

There is a more fundamental difference between the two as well, which I've not drawn attention to here, because it isn't being used -- SCEV is usually to unable to analyze bitwise operations like `x & y` (unless they're just optimized arithmetic).  This means we don't widen uses of IVs like `and %iv, 42` (and insert a `trunc`) even though bitwise operations are (almost by definition) easy to widen (`zext (and %iv, 42)` == `and zext(%iv), zext(42)`).  `-instcombine` is usually able to clean these (the `trunc` s) up, but if we later find interesting cases that `-instcombine` cannot clean up then we could make `cloneBitwiseIVUser` smarter, which would further diverge the implementations.


http://reviews.llvm.org/D13716