[llvm] [DA] do not handle array accesses of different offsets (PR #123436)

[Sjoerd Meijer via llvm-commits]

================
@@ -3569,6 +3569,123 @@ bool DependenceInfo::invalidate(Function &F, const PreservedAnalyses &PA,
          Inv.invalidate<LoopAnalysis>(F, PA);
 }
 
+// Check that memory access offsets in V are multiples of array element size
+// EltSize. Param records the first parametric expression. If the scalar
+// evolution V contains two or more parameters, we check that the subsequent
+// parametric expressions are multiples of the first parametric expression
+// Param.
+static bool checkOffsets(ScalarEvolution *SE, const SCEV *V, const SCEV *&Param,
+                         uint64_t EltSize) {
+  if (auto *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
+    if (!checkOffsets(SE, AddRec->getStart(), Param, EltSize))
+      return false;
+    return checkOffsets(SE, AddRec->getStepRecurrence(*SE), Param, EltSize);
+  }
+  if (auto *Cst = dyn_cast<SCEVConstant>(V)) {
+    APInt C = Cst->getAPInt();
+
+    // For example, alias_with_different_offsets in
+    // test/Analysis/DependenceAnalysis/DifferentOffsets.ll accesses "%A + 2":
+    //   %arrayidx = getelementptr inbounds i8, ptr %A, i64 2
+    //   store i32 42, ptr %arrayidx, align 1
+    // which is writing an i32, i.e., EltSize = 4 bytes, with an offset C = 2.
+    // checkOffsets returns false, as the offset C=2 is not a multiple of 4.
+    return C.srem(EltSize) == 0;
+  }
+
+  // Use a lambda helper function to check V for parametric expressions.
+  // Param records the first parametric expression. If the scalar evolution V
+  // contains two or more parameters, we check that the subsequent parametric
+  // expressions are multiples of the first parametric expression Param.
+  auto checkParamsMultipleOfSize = [&](const SCEV *V,
+                                       const SCEV *&Param) -> bool {
+    if (EltSize == 1)
+      return true;
+    if (!Param) {
+      Param = V;
+      return true;
----------------
sjoerdmeijer wrote:

> > DA is assuming that %n is part of all index expressions and therefore has the same offset everywhere. If some index expressions use %m instead, they may have different relative offsets.
> 
> You are right. If both %n and %m are used to index elements in a same array, the difference `%n - %m` is unknown at compilation time, and so the current dependence analysis result would be incorrect.

I am catching up on this, trying to let this sink in, but had this perhaps silly question in the meantime: if there are still correctness issues, can we not just return Confused for now when we detect that we have different SCEV expressions with different offsets? Or would this make DA so conservative that it will be useless?

https://github.com/llvm/llvm-project/pull/123436