[PATCH] [LAA] Try to prove non-wrapping of pointers if SCEV cannot

[Adam Nemet]

================
Comment at: lib/Analysis/LoopAccessAnalysis.cpp:509
@@ +508,3 @@
+/// i.e. monotonically increasing/decreasing.
+static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
+                           ScalarEvolution *SE, const Loop *L) {
----------------
sanjoy wrote:
> I'm not clear on how LAA uses this property, but I think this function should mention what kind of no-wrap (signed or unsigned) behavior it is trying to prove.  IOW, Ptr is supposed to be  monotonically increasing/decreasing in the signed or unsigned sense?
It's unsigned.  Pointers are unsigned in LLVM IR as mentioned in the GEP FAQ.  I'll add a comment.

================
Comment at: lib/Analysis/LoopAccessAnalysis.cpp:540
@@ +539,3 @@
+  // using a non-wrapping operation.
+  if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
+    if ((OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) &&
----------------
sanjoy wrote:
> What if `OBO` is `nuw` (and not `nsw`) and `OpAR` is `nsw` (and not `nuw`)?  Or vice-versa?
Yeah, this certainly required more thinking on my behalf, thanks for pressing it.

So to document why this was wrong, let's take this counterexample:

Consider the AddRec {0,+,100} <nuw> in i8.  The first three iterations of that yields: 0, 100, 200.

Putting this through a *signed* add of 3, the input is now interpreted as signed: 0, 100, -56.

No signed overflow on the result (3, 103, -53), yet the result is wrapped. 

Similarly, we can't take <nuw>-only for index.  Index is interpreted as signed.  With the above (i8) {0,+,100} example we'd get a wrapping range even though it may be inbounds for the array. 

Let me know if you or others disagree or have further comments.  Otherwise I'll update the patch accordingly.

http://reviews.llvm.org/D10472

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/