[llvm] [SeparateConstOffsetFromGEP] Decompose constant xor operand if possible (PR #135788)

[Matt Arsenault via llvm-commits]

================
@@ -1162,6 +1167,179 @@ bool SeparateConstOffsetFromGEP::splitGEP(GetElementPtrInst *GEP) {
   return true;
 }
 
+bool SeparateConstOffsetFromGEP::decomposeXor(Function &F) {
+  bool FunctionChanged = false;
+  SmallVector<std::pair<Instruction *, Value *>, 16> ReplacementsToMake;
+
+  for (BasicBlock &BB : F) {
+    for (Instruction &I : BB) {
+      if (I.getOpcode() == Instruction::Xor) {
+        if (Value *Replacement = tryFoldXorToOrDisjoint(I)) {
+          ReplacementsToMake.push_back({&I, Replacement});
+          FunctionChanged = true;
+        }
+      }
+    }
+  }
+
+  if (!ReplacementsToMake.empty()) {
+    LLVM_DEBUG(dbgs() << "Applying " << ReplacementsToMake.size()
+                      << " XOR->OR Disjoint replacements in " << F.getName()
+                      << "\n");
+    for (auto &Pair : ReplacementsToMake)
+      Pair.first->replaceAllUsesWith(Pair.second);
+
+    for (auto &Pair : ReplacementsToMake)
+      Pair.first->eraseFromParent();
+  }
+
+  return FunctionChanged;
+}
+
+static llvm::Instruction *findClosestSequentialXor(Value *A, Instruction &I) {
+  llvm::Instruction *ClosestUser = nullptr;
+  for (llvm::User *User : A->users()) {
+    if (auto *UserInst = llvm::dyn_cast<llvm::Instruction>(User)) {
+      if (UserInst->getOpcode() != Instruction::Xor || UserInst == &I)
+        continue;
+      if (!ClosestUser)
+        ClosestUser = UserInst;
+      else {
+        // Compare instruction positions.
+        if (UserInst->comesBefore(ClosestUser)) {
+          ClosestUser = UserInst;
+        }
+      }
+    }
+  }
+  return ClosestUser;
+}
+
+/// Try to transform I = xor(A, C1) into or_disjoint(Y, C2)
+/// where Y = xor(A, C0) is another existing instruction dominating I,
+/// C2 = C1 - C0, and A is known to be disjoint with C2.
+///
+/// This transformation is beneficial particularly for GEPs because:
+/// 1. OR operations often map better to addressing modes than XOR
+/// 2. Disjoint OR operations preserve the semantics of the original XOR
+/// 3. This can enable further optimizations in the GEP offset folding pipeline
+///
+/// @param I  The XOR instruction being visited.
+/// @return The replacement Value* if successful, nullptr otherwise.
+Value *SeparateConstOffsetFromGEP::tryFoldXorToOrDisjoint(Instruction &I) {
+  assert(I.getOpcode() == Instruction::Xor && "Instruction must be XOR");
+
+  // Check if I has at least one GEP user.
+  bool HasGepUser = false;
+  for (User *U : I.users()) {
+    if (isa<GetElementPtrInst>(U)) {
+      HasGepUser = true;
+      break;
+    }
+  }
----------------
arsenm wrote:

Generally shouldn't be looking at users but I also don't see why this cares 

https://github.com/llvm/llvm-project/pull/135788