[llvm] [AArch64MIPeepholeOpt] Coalesce sibling base-address materializations (PR #223682)

[via llvm-commits]

================
@@ -957,6 +966,106 @@ bool AArch64MIPeepholeOptImpl::visitCopy(MachineInstr &MI) {
   return true;
 }
 
+bool AArch64MIPeepholeOptImpl::shareBaseAddresses(MachineBasicBlock &MBB) {
+  if (!EnableBaseAddressCSE)
+    return false;
+
+  bool Changed = false;
+
+  struct Cand {
+    MachineInstr *DefMI;
+    Register SrcReg;
+    unsigned Opc;
+    int64_t C;
+    Register BaseReg;
+    MachineInstr *UserMI;
+    int Scale;
+    int64_t OldOff;
+  };
+  SmallVector<Cand, 8> Cands;
+  using Key = std::pair<Register, unsigned>;
+  MapVector<Key, SmallVector<unsigned, 4>> Groups;
+
+  DenseMap<MachineInstr *, unsigned> Pos;
+  unsigned Idx = 0;
+  for (MachineInstr &MI : MBB.instrs()) {
+    Pos[&MI] = Idx++;
+    unsigned Opc = MI.getOpcode();
+    if (Opc != AArch64::ADDXri && Opc != AArch64::SUBXri)
+      continue;
+    if (!MI.getOperand(2).isImm() || !MI.getOperand(3).isImm() ||
+        MI.getOperand(3).getImm() != 0) // shift must be 0 (no LSL #12)
+      continue;
+    Register SrcReg = MI.getOperand(1).getReg();
+    Register BaseReg = MI.getOperand(0).getReg();
+    if (!SrcReg.isVirtual() || !BaseReg.isVirtual() ||
+        !MRI->hasOneUse(BaseReg)) // secondary def is erased below
+      continue;
+    MachineInstr &User = *MRI->use_instr_begin(BaseReg);
+    if (!User.mayLoadOrStore() || AArch64InstrInfo::isPairedLdSt(User) ||
+        AArch64InstrInfo::isPreLdSt(User) || // different operand layout
+        TII->hasUnscaledLdStOffset(User) ||  // rewrite below assumes scaled
+        !AArch64InstrInfo::getLdStOffsetOp(User).isImm())
+      continue;
----------------
whokeke wrote:

Added a comment above the exclusion list. After the four checks (mayLoadOrStore, !isPairedLdSt, !isPreLdSt, !hasUnscaledLdStOffset, offset isImm()), what's left is a scaled, non-paired, non-pre-indexed load/store with an immediate offset — the exact shape the base+uimm12-offset rewrite below assumes. I kept the exclusion style rather than switching to an inclusive opcode allowlist, since isPairedLdSt/isPreLdSt/hasUnscaledLdStOffset are existing shared helpers that already partition the opcode space correctly, and re-deriving an allowlist here risks missing an opcode they cover. Happy to switch if you'd still prefer an explicit list.

https://github.com/llvm/llvm-project/pull/223682