[llvm] [AArch64][GlobalISel] Add support for post-indexed loads/stores. (PR #69532)

[Jon Roelofs via llvm-commits]

================
@@ -945,42 +945,170 @@ void CombinerHelper::applySextInRegOfLoad(
   MI.eraseFromParent();
 }
 
+static Type *getTypeForLLT(LLT Ty, LLVMContext &C) {
+  if (Ty.isVector())
+    return FixedVectorType::get(IntegerType::get(C, Ty.getScalarSizeInBits()),
+                                Ty.getNumElements());
+  return IntegerType::get(C, Ty.getSizeInBits());
+}
+
+/// Return true if 'MI' is a load or a store that may be fold it's address
+/// operand into the load / store addressing mode.
+static bool canFoldInAddressingMode(GLoadStore *MI, const TargetLowering &TLI,
+                                    MachineRegisterInfo &MRI) {
+  TargetLowering::AddrMode AM;
+  auto *MF = MI->getMF();
+  auto *Addr = getOpcodeDef<GPtrAdd>(MI->getPointerReg(), MRI);
+  if (!Addr)
+    return false;
+
+  AM.HasBaseReg = true;
+  auto CstOff = getIConstantVRegVal(Addr->getOffsetReg(), MRI);
+  if (CstOff)
+    AM.BaseOffs = CstOff->getSExtValue(); // [reg +/- imm]
+  else
+    AM.Scale = 1; // [reg +/- reg]
+
+  return TLI.isLegalAddressingMode(
+      MF->getDataLayout(), AM,
+      getTypeForLLT(MI->getMMO().getMemoryType(),
+                    MF->getFunction().getContext()),
+      MI->getMMO().getAddrSpace());
+}
+
+namespace {
+unsigned getIndexedOpc(unsigned LdStOpc) {
+  switch (LdStOpc) {
+  case TargetOpcode::G_LOAD:
+    return TargetOpcode::G_INDEXED_LOAD;
+  case TargetOpcode::G_STORE:
+    return TargetOpcode::G_INDEXED_STORE;
+  case TargetOpcode::G_ZEXTLOAD:
+    return TargetOpcode::G_INDEXED_ZEXTLOAD;
+  case TargetOpcode::G_SEXTLOAD:
+    return TargetOpcode::G_INDEXED_SEXTLOAD;
+  default:
+    llvm_unreachable("Unexpected opcode");
+  }
+}
+} // namespace
+
+bool CombinerHelper::isIndexedLoadStoreLegal(GLoadStore &LdSt) const {
+  // Check for legality.
+  LLT PtrTy = MRI.getType(LdSt.getPointerReg());
+  LLT Ty = MRI.getType(LdSt.getReg(0));
+  LLT MemTy = LdSt.getMMO().getMemoryType();
+  SmallVector<LegalityQuery::MemDesc, 2> MemDescrs(
+      {{MemTy, MemTy.getSizeInBits(), AtomicOrdering::NotAtomic}});
+  unsigned IndexedOpc = getIndexedOpc(LdSt.getOpcode());
+  SmallVector<LLT> OpTys;
+  if (IndexedOpc == TargetOpcode::G_INDEXED_STORE)
+    OpTys = {PtrTy, Ty, Ty};
+  else
+    OpTys = {Ty, PtrTy}; // For G_INDEXED_LOAD, G_INDEXED_[SZ]EXTLOAD
+
+  LegalityQuery Q(IndexedOpc, OpTys, MemDescrs);
+  return isLegal(Q);
+}
+
+static cl::opt<unsigned> PostIndexUseThreshold(
+    "post-index-use-threshold", cl::Hidden, cl::init(32),
+    cl::desc("Number of uses of a base pointer to check before it is no longer "
+             "considered for post-indexing."));
+
 bool CombinerHelper::findPostIndexCandidate(GLoadStore &LdSt, Register &Addr,
-                                            Register &Base, Register &Offset) {
+                                            Register &Base, Register &Offset,
+                                            bool &RematOffset) {
+  // We're looking for the following pattern, for either load or store:
+  // %baseptr:_(p0) = ...
+  // G_STORE %val(s64), %baseptr(p0)
+  // %offset:_(s64) = G_CONSTANT i64 -256
+  // %new_addr:_(p0) = G_PTR_ADD %baseptr, %offset(s64)
   auto &MF = *LdSt.getParent()->getParent();
   const auto &TLI = *MF.getSubtarget().getTargetLowering();
 
-  Base = LdSt.getPointerReg();
+  Register Ptr = LdSt.getPointerReg();
+  // If the store is the only use, don't bother.
+  if (MRI.hasOneNonDBGUse(Ptr))
+    return false;
+
+  if (!isIndexedLoadStoreLegal(LdSt))
+    return false;
 
-  if (getOpcodeDef(TargetOpcode::G_FRAME_INDEX, Base, MRI))
+  if (getOpcodeDef(TargetOpcode::G_FRAME_INDEX, Ptr, MRI))
     return false;
 
-  // FIXME: The following use traversal needs a bail out for patholigical cases.
-  for (auto &Use : MRI.use_nodbg_instructions(Base)) {
+  MachineInstr *StoredValDef = getDefIgnoringCopies(LdSt.getReg(0), MRI);
+  auto *PtrDef = MRI.getVRegDef(Ptr);
+
+  unsigned NumUsesChecked = 0;
+  for (auto &Use : MRI.use_nodbg_instructions(Ptr)) {
+    if (++NumUsesChecked > PostIndexUseThreshold)
+      return false; // Try to avoid exploding compile time.
+
     auto *PtrAdd = dyn_cast<GPtrAdd>(&Use);
-    if (!PtrAdd)
+    // The use itself might be dead. This can happen during combines if DCE
+    // hasn't had a chance to run yet. Don't allow it to form an indexed op.
+    if (!PtrAdd || MRI.use_nodbg_empty(PtrAdd->getReg(0)))
+      continue;
+
+    // Check the user of this isn't the store, otherwise we'd be generate a
+    // indexed store defining its own use.
+    if (StoredValDef == &Use)
       continue;
 
     Offset = PtrAdd->getOffsetReg();
     if (!ForceLegalIndexing &&
-        !TLI.isIndexingLegal(LdSt, Base, Offset, /*IsPre*/ false, MRI))
+        !TLI.isIndexingLegal(LdSt, PtrAdd->getBaseReg(), Offset,
+                             /*IsPre*/ false, MRI))
       continue;
 
     // Make sure the offset calculation is before the potentially indexed op.
     MachineInstr *OffsetDef = MRI.getVRegDef(Offset);
-    if (!dominates(*OffsetDef, LdSt))
-      continue;
+    if (!dominates(*OffsetDef, LdSt)) {
+      // If the offset however is just a G_CONSTANT, we can always just
+      // rematerialize it where we need it.
+      if (OffsetDef->getOpcode() != TargetOpcode::G_CONSTANT)
+        continue;
+      RematOffset = true;
----------------
jroelofs wrote:

On return, it looks like `RematOffset` is either `true`, or uninitialized. Probably needs a `RematOffset = false` up near the beginning of the funciton.

https://github.com/llvm/llvm-project/pull/69532