[llvm] [LAA] Perform checks for no-wrap separately from getPtrStride. (PR #126971)

[Philip Reames via llvm-commits]

================
@@ -813,16 +813,102 @@ static bool hasComputableBounds(PredicatedScalarEvolution &PSE, Value *Ptr,
   return AR->isAffine();
 }
 
+/// Try to compute the stride for \p AR. Used by getPtrStride.
+static std::optional<int64_t>
+getStrideFromAddRec(const SCEVAddRecExpr *AR, const Loop *Lp, Type *AccessTy,
+                    Value *Ptr, PredicatedScalarEvolution &PSE) {
+  // The access function must stride over the innermost loop.
+  if (Lp != AR->getLoop()) {
+    LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop "
+                      << *Ptr << " SCEV: " << *AR << "\n");
+    return std::nullopt;
+  }
+
+  // Check the step is constant.
+  const SCEV *Step = AR->getStepRecurrence(*PSE.getSE());
+
+  // Calculate the pointer stride and check if it is constant.
+  const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
+  if (!C) {
+    LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr
+                      << " SCEV: " << *AR << "\n");
+    return std::nullopt;
+  }
+
+  const auto &DL = Lp->getHeader()->getDataLayout();
+  TypeSize AllocSize = DL.getTypeAllocSize(AccessTy);
+  int64_t Size = AllocSize.getFixedValue();
+  const APInt &APStepVal = C->getAPInt();
+
+  // Huge step value - give up.
+  if (APStepVal.getBitWidth() > 64)
+    return std::nullopt;
+
+  int64_t StepVal = APStepVal.getSExtValue();
+
+  // Strided access.
+  int64_t Stride = StepVal / Size;
+  int64_t Rem = StepVal % Size;
+  if (Rem)
+    return std::nullopt;
+
+  return Stride;
+}
+
+static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
+                           PredicatedScalarEvolution &PSE, const Loop *L);
+
 /// Check whether a pointer address cannot wrap.
 static bool isNoWrap(PredicatedScalarEvolution &PSE,
                      const DenseMap<Value *, const SCEV *> &Strides, Value *Ptr,
-                     Type *AccessTy, Loop *L, bool Assume) {
-  const SCEV *PtrScev = PSE.getSCEV(Ptr);
+                     Type *AccessTy, const Loop *L, bool Assume,
+                     std::optional<int64_t> Stride = std::nullopt) {
+  const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
   if (PSE.getSE()->isLoopInvariant(PtrScev, L))
     return true;
 
-  return getPtrStride(PSE, AccessTy, Ptr, L, Strides, Assume).has_value() ||
-         PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
+  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
+  if (Assume && !AR)
+    AR = PSE.getAsAddRec(Ptr);
+  if (!AR)
+    return false;
+
+  // The address calculation must not wrap. Otherwise, a dependence could be
+  // inverted.
+  if (isNoWrapAddRec(Ptr, AR, PSE, L))
+    return true;
+
+  // An nusw getelementptr that is an AddRec cannot wrap. If it would wrap,
+  // the distance between the previously accessed location and the wrapped
+  // location will be larger than half the pointer index type space. In that
+  // case, the GEP would be  poison and any memory access dependent on it would
+  // be immediate UB when executed.
+  if (auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
----------------
preames wrote:

isNoWrapAddRec has this whole dance for analyzing nusw.  It seems very odd to have the unchecked form right after  a call to that routine?

(Though, now I see that in the current code too?)

https://github.com/llvm/llvm-project/pull/126971