[clang] [llvm] [LV] Mask off possibly aliasing vector lanes (PR #100579)

[Sam Tebbs via llvm-commits]

================
@@ -428,6 +431,85 @@ Value *VPInstruction::generatePerPart(VPTransformState &State, unsigned Part) {
                                    {PredTy, ScalarTC->getType()},
                                    {VIVElem0, ScalarTC}, nullptr, Name);
   }
+  case VPInstruction::AliasLaneMask: {
+    // Given a pointer A that is being stored to, and pointer B that is being
+    // read from, both with unknown lengths, create a mask that disables
+    // elements which could overlap across a loop iteration. For example, if A
+    // is X and B is X + 2 with VF being 4, only the final two elements of the
+    // loaded vector can be stored since they don't overlap with the stored
+    // vector. %b.vec = load %b ; = [s, t, u, v]
+    // [...]
+    // store %a, %b.vec ; only u and v can be stored as their addresses don't
+    // overlap with %a + (VF - 1)
+    Value *ReadPtr = State.get(getOperand(0), VPIteration(Part, 0));
+    Value *StorePtr = State.get(getOperand(1), VPIteration(Part, 0));
+    unsigned ElementSize = 0;
+
+    // We expect the operands to the alias mask to be ptrtoint. Sometimes it's
+    // an add of a ptrtoint.
+    auto *ReadInsn = cast<Instruction>(ReadPtr);
+    auto *ReadCast = dyn_cast<CastInst>(ReadPtr);
+    if (ReadInsn->getOpcode() == Instruction::Add)
+      ReadCast = dyn_cast<CastInst>(ReadInsn->getOperand(0));
+
+    if (ReadCast && ReadCast->getOpcode() == Instruction::PtrToInt) {
+      Value *Ptr = ReadCast->getOperand(0);
+      for (auto *Use : Ptr->users()) {
+        if (auto *GEP = dyn_cast<GetElementPtrInst>(Use)) {
+          auto *EltVT = GEP->getSourceElementType();
+          if (EltVT->isArrayTy())
+            ElementSize = EltVT->getArrayElementType()->getScalarSizeInBits() *
+                          EltVT->getArrayNumElements();
+          else
+            ElementSize =
+                GEP->getSourceElementType()->getScalarSizeInBits() / 8;
----------------
SamTebbs33 wrote:

I've managed to figure it out now and am using the `AccessSize` provided by `PointerDIffInfo`. Thanks for pointing out the flakiness here!

https://github.com/llvm/llvm-project/pull/100579