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

[Sam Tebbs via cfe-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()) {
----------------
SamTebbs33 wrote:

@fhahn  I'm not sure if figuring out the element size before execution time will work since at that point the source and destination pointers are SCEVS and the users of the pointers themselves aren't available. The SCEVs have to be expanded before that information is available. Do you have any pointers to help with this?

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