[CodeGen] Split out the notions of MI invariance and MI dereferenceability.

[jlebar]

Summary

An IR load can be invariant, dereferenceable, neither, or both. But
currently, MI's notion of invariance is IR-invariant &&
IR-dereferenceable.

This patch splits up the notions of invariance and dereferenceability at
the MI level. It's NFC, so adds some probably-unnecessary
"is-dereferenceable" checks, which we can remove later if desired.

[jlebar]

FYI I created this patch by going through all occurrences of "invariant" within the following files:

git grep -li invariant | grep -v '^test\|/Transforms\|/IR\|/Analysis\|^docs\|ADT'

[chandlerc]

This seems largely good (once preceding patches land). It'd be good to get someone to double check the PPC refactoring, as that's the only one that seemed even a little interesting.

[hfinkel]

In D23371#517565, @chandlerc wrote:

This seems largely good (once preceding patches land). It'd be good to get someone to double check the PPC refactoring, as that's the only one that seemed even a little interesting.

Quickly looked at the PPC bits; Looks fine to me.

[jlebar]

Preceding patches are in, not sure if you wanted to have another look before lg'ing this. I went back through and (I admit, halfheartedly) re-audited, and I didn't find anything that needed to change.

[jlebar]

Rebase to HEAD.

[chandlerc]

LGTM!

Thanks for all the prep work to get here and making this change!

llvm/include/llvm/CodeGen/MachineMemOperand.h

enum Flags : uint16_t {
     /// The memory access reads data.
     MOLoad = 1u << 0,
     /// The memory access writes data.
     MOStore = 1u << 1,
     /// The memory access is volatile.
     MOVolatile = 1u << 2,
     /// The memory access is non-temporal.
     MONonTemporal = 1u << 3,
-    /// The memory access is invariant.
-    MOInvariant = 1u << 4,
+    /// The memory access is dereferenceable (i.e., doesn't trap).
+    MODereferenceable = 1u << 4,
+    /// The memory access always returns the same value (or traps).
+    MOInvariant = 1u << 5,
 
     // Reserved for use by target-specific passes.
-    MOTargetFlag1 = 1u << 5,
-    MOTargetFlag2 = 1u << 6,
-    MOTargetFlag3 = 1u << 7,
+    MOTargetFlag1 = 1u << 6,
+    MOTargetFlag2 = 1u << 7,
+    MOTargetFlag3 = 1u << 8,
 
     LLVM_MARK_AS_BITMASK_ENUM(/* LargestFlag = */ MOTargetFlag3)
   };
 
 private:
   MachinePointerInfo PtrInfo;
   uint64_t Size;
   Flags FlagVals;
public:
 
   /// Return the range tag for the memory reference.
   const MDNode *getRanges() const { return Ranges; }
 
   bool isLoad() const { return FlagVals & MOLoad; }
   bool isStore() const { return FlagVals & MOStore; }
   bool isVolatile() const { return FlagVals & MOVolatile; }
   bool isNonTemporal() const { return FlagVals & MONonTemporal; }
+  bool isDereferenceable() const { return FlagVals & MODereferenceable; }
   bool isInvariant() const { return FlagVals & MOInvariant; }
 
   /// Returns true if this memory operation doesn't have any ordering
   /// constraints other than normal aliasing. Volatile and atomic memory
   /// operations can't be reordered.
   ///
   /// Currently, we don't model the difference between volatile and atomic
   /// operations. They should retain their ordering relative to all memory

llvm/include/llvm/CodeGen/SelectionDAGNodes.h

class MemSDNodeBitfields {
     friend class MemSDNode;
     friend class MemIntrinsicSDNode;
     friend class AtomicSDNode;
 
     uint16_t : NumSDNodeBits;
 
     uint16_t IsVolatile : 1;
     uint16_t IsNonTemporal : 1;
+    uint16_t IsDereferenceable : 1;
     uint16_t IsInvariant : 1;
     uint16_t SynchScope : 1; // enum SynchronizationScope
     uint16_t Ordering : 4;   // enum AtomicOrdering
   };
-  enum { NumMemSDNodeBits = NumSDNodeBits + 8 };
+  enum { NumMemSDNodeBits = NumSDNodeBits + 9 };
 
   class LSBaseSDNodeBitfields {
     friend class LSBaseSDNode;
     uint16_t : NumMemSDNodeBits;
 
     uint16_t AddressingMode : 3; // enum ISD::MemIndexedMode
   };
   enum { NumLSBaseSDNodeBits = NumMemSDNodeBits + 3 };
public:
   /// encoding of the volatile flag, as well as bits used by subclasses. This
   /// function should only be used to compute a FoldingSetNodeID value.
   unsigned getRawSubclassData() const {
     uint16_t Data;
     memcpy(&Data, &RawSDNodeBits, sizeof(RawSDNodeBits));
     return Data;
   }
 
-  // We access subclass data here so that we can check consistency
-  // with MachineMemOperand information.
   bool isVolatile() const { return MemSDNodeBits.IsVolatile; }
   bool isNonTemporal() const { return MemSDNodeBits.IsNonTemporal; }
+  bool isDereferenceable() const { return MemSDNodeBits.IsDereferenceable; }
   bool isInvariant() const { return MemSDNodeBits.IsInvariant; }
 
   AtomicOrdering getOrdering() const {
     return static_cast<AtomicOrdering>(MemSDNodeBits.Ordering);
   }
   SynchronizationScope getSynchScope() const {
     return static_cast<SynchronizationScope>(MemSDNodeBits.SynchScope);
   }

llvm/lib/CodeGen/MIRParser/MILexer.h

enum TokenKind {
     less,
     greater,
 
     // Keywords
     kw_implicit,
     kw_implicit_define,
     kw_def,
     kw_dead,
+    kw_dereferenceable,
     kw_killed,
     kw_undef,
     kw_internal,
     kw_early_clobber,
     kw_debug_use,
     kw_tied_def,
     kw_frame_setup,
     kw_debug_location,
bool isRegisterFlag() const {
     return Kind == kw_implicit || Kind == kw_implicit_define ||
            Kind == kw_def || Kind == kw_dead || Kind == kw_killed ||
            Kind == kw_undef || Kind == kw_internal ||
            Kind == kw_early_clobber || Kind == kw_debug_use;
   }
 
   bool isMemoryOperandFlag() const {
     return Kind == kw_volatile || Kind == kw_non_temporal ||
-           Kind == kw_invariant;
+           Kind == kw_dereferenceable || Kind == kw_invariant;
   }
 
   bool is(TokenKind K) const { return Kind == K; }
 
   bool isNot(TokenKind K) const { return Kind != K; }
 
   StringRef::iterator location() const { return Range.begin(); }
 

llvm/lib/CodeGen/MIRParser/MILexer.cpp

return StringSwitch<MIToken::TokenKind>(Identifier)
       .Case("float", MIToken::kw_float)
       .Case("double", MIToken::kw_double)
       .Case("x86_fp80", MIToken::kw_x86_fp80)
       .Case("fp128", MIToken::kw_fp128)
       .Case("ppc_fp128", MIToken::kw_ppc_fp128)
       .Case("target-flags", MIToken::kw_target_flags)
       .Case("volatile", MIToken::kw_volatile)
       .Case("non-temporal", MIToken::kw_non_temporal)
+      .Case("dereferenceable", MIToken::kw_dereferenceable)
       .Case("invariant", MIToken::kw_invariant)
       .Case("align", MIToken::kw_align)
       .Case("stack", MIToken::kw_stack)
       .Case("got", MIToken::kw_got)
       .Case("jump-table", MIToken::kw_jump_table)
       .Case("constant-pool", MIToken::kw_constant_pool)
       .Case("call-entry", MIToken::kw_call_entry)
       .Case("liveout", MIToken::kw_liveout)

llvm/lib/CodeGen/MIRParser/MIParser.cpp

bool MIParser::parseMemoryOperandFlag(MachineMemOperand::Flags &Flags) {
   const auto OldFlags = Flags;
   switch (Token.kind()) {
   case MIToken::kw_volatile:
     Flags |= MachineMemOperand::MOVolatile;
     break;
   case MIToken::kw_non_temporal:
     Flags |= MachineMemOperand::MONonTemporal;
     break;
+  case MIToken::kw_dereferenceable:
+    Flags |= MachineMemOperand::MODereferenceable;
+    break;
   case MIToken::kw_invariant:
     Flags |= MachineMemOperand::MOInvariant;
     break;
   // TODO: parse the target specific memory operand flags.
   default:
     llvm_unreachable("The current token should be a memory operand flag");
   }
   if (OldFlags == Flags)

llvm/lib/CodeGen/MIRPrinter.cpp

 
 void MIPrinter::print(const MachineMemOperand &Op) {
   OS << '(';
   // TODO: Print operand's target specific flags.
   if (Op.isVolatile())
     OS << "volatile ";
   if (Op.isNonTemporal())
     OS << "non-temporal ";
+  if (Op.isDereferenceable())
+    OS << "dereferenceable ";
   if (Op.isInvariant())
     OS << "invariant ";
   if (Op.isLoad())
     OS << "load ";
   else {
     assert(Op.isStore() && "Non load machine operand must be a store");
     OS << "store ";
   }

llvm/lib/CodeGen/MachineInstr.cpp

if (NoAliasInfo->getNumOperands() > 0)
         if (i != ie-1)
           OS << ",";
       }
     else
       OS << "<unknown>";
     OS << ")";
   }
 
-  // Print nontemporal info.
   if (isNonTemporal())
     OS << "(nontemporal)";
-
+  if (isDereferenceable())
+    OS << "(dereferenceable)";
   if (isInvariant())
     OS << "(invariant)";
 }
 
 //===----------------------------------------------------------------------===//
 // MachineInstr Implementation
 //===----------------------------------------------------------------------===//
 
bool MachineInstr::isDereferenceableInvariantLoad(AliasAnalysis *AA) const {
   if (memoperands_empty())
     return false;
 
   const MachineFrameInfo &MFI = getParent()->getParent()->getFrameInfo();
 
   for (MachineMemOperand *MMO : memoperands()) {
     if (MMO->isVolatile()) return false;
     if (MMO->isStore()) return false;
-    if (MMO->isInvariant()) continue;
+    if (MMO->isInvariant() && MMO->isDereferenceable())
+      continue;
 
     // A load from a constant PseudoSourceValue is invariant.
     if (const PseudoSourceValue *PSV = MMO->getPseudoValue())
       if (PSV->isConstant(&MFI))
         continue;
 
     if (const Value *V = MMO->getValue()) {
       // If we have an AliasAnalysis, ask it whether the memory is constant.

llvm/lib/CodeGen/MachinePipeliner.cpp

void SwingSchedulerDAG::updateMemOperands(MachineInstr &NewMI,
   // If the instruction has memory operands, then adjust the offset
   // when the instruction appears in different stages.
   unsigned NumRefs = NewMI.memoperands_end() - NewMI.memoperands_begin();
   if (NumRefs == 0)
     return;
   MachineInstr::mmo_iterator NewMemRefs = MF.allocateMemRefsArray(NumRefs);
   unsigned Refs = 0;
   for (MachineMemOperand *MMO : NewMI.memoperands()) {
-    if (MMO->isVolatile() || MMO->isInvariant() || (!MMO->getValue())) {
+    if (MMO->isVolatile() || (MMO->isInvariant() && MMO->isDereferenceable()) ||
+        (!MMO->getValue())) {
       NewMemRefs[Refs++] = MMO;
       continue;
     }
     unsigned Delta;
     if (computeDelta(OldMI, Delta)) {
       int64_t AdjOffset = Delta * Num;
       NewMemRefs[Refs++] =
           MF.getMachineMemOperand(MMO, AdjOffset, MMO->getSize());

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

if (LHS.getOpcode() == ISD::LOAD) {
     SDValue Load;
     // It is safe to replace the two loads if they have different alignments,
     // but the new load must be the minimum (most restrictive) alignment of the
     // inputs.
     unsigned Alignment = std::min(LLD->getAlignment(), RLD->getAlignment());
     MachineMemOperand::Flags MMOFlags = LLD->getMemOperand()->getFlags();
     if (!RLD->isInvariant())
       MMOFlags &= ~MachineMemOperand::MOInvariant;
+    if (!RLD->isDereferenceable())
+      MMOFlags &= ~MachineMemOperand::MODereferenceable;
     if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
       // FIXME: Discards pointer and AA info.
       Load = DAG.getLoad(TheSelect->getValueType(0), SDLoc(TheSelect),
                          LLD->getChain(), Addr, MachinePointerInfo(), Alignment,
                          MMOFlags);
     } else {
       // FIXME: Discards pointer and AA info.
       Load = DAG.getExtLoad(

llvm/lib/CodeGen/SelectionDAG/FastISel.cpp

if (const auto *LI = dyn_cast<LoadInst>(I)) {
     Flags = MachineMemOperand::MOStore;
     Ptr = SI->getPointerOperand();
     ValTy = SI->getValueOperand()->getType();
   } else
     return nullptr;
 
   bool IsNonTemporal = I->getMetadata(LLVMContext::MD_nontemporal) != nullptr;
   bool IsInvariant = I->getMetadata(LLVMContext::MD_invariant_load) != nullptr;
+  bool IsDereferenceable =
+      I->getMetadata(LLVMContext::MD_dereferenceable) != nullptr;
   const MDNode *Ranges = I->getMetadata(LLVMContext::MD_range);
 
   AAMDNodes AAInfo;
   I->getAAMetadata(AAInfo);
 
   if (Alignment == 0) // Ensure that codegen never sees alignment 0.
     Alignment = DL.getABITypeAlignment(ValTy);
 
   unsigned Size = DL.getTypeStoreSize(ValTy);
 
   if (IsVolatile)
     Flags |= MachineMemOperand::MOVolatile;
   if (IsNonTemporal)
     Flags |= MachineMemOperand::MONonTemporal;
+  if (IsDereferenceable)
+    Flags |= MachineMemOperand::MODereferenceable;
   if (IsInvariant)
     Flags |= MachineMemOperand::MOInvariant;
 
   return FuncInfo.MF->getMachineMemOperand(MachinePointerInfo(Ptr), Flags, Size,
                                            Alignment, AAInfo, Ranges);
 }
 
 CmpInst::Predicate FastISel::optimizeCmpPredicate(const CmpInst *CI) const {

llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp

 SDValue DAGTypeLegalizer::SoftenFloatRes_LOAD(SDNode *N, unsigned ResNo) {
   bool LegalInHWReg = isLegalInHWReg(N->getValueType(ResNo));
   LoadSDNode *L = cast<LoadSDNode>(N);
   EVT VT = N->getValueType(0);
   EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
   SDLoc dl(N);
 
   auto MMOFlags =
-      L->getMemOperand()->getFlags() & ~MachineMemOperand::MOInvariant;
+      L->getMemOperand()->getFlags() &
+      ~(MachineMemOperand::MOInvariant | MachineMemOperand::MODereferenceable);
   SDValue NewL;
   if (L->getExtensionType() == ISD::NON_EXTLOAD) {
     NewL = DAG.getLoad(L->getAddressingMode(), L->getExtensionType(), NVT, dl,
                        L->getChain(), L->getBasePtr(), L->getOffset(),
                        L->getPointerInfo(), NVT, L->getAlignment(), MMOFlags,
                        L->getAAInfo());
     // Legalized the chain result - switch anything that used the old chain to
     // use the new one.

 
 SDValue DAGTypeLegalizer::PromoteFloatRes_LOAD(SDNode *N) {
   LoadSDNode *L = cast<LoadSDNode>(N);
   EVT VT = N->getValueType(0);
 
   // Load the value as an integer value with the same number of bits.
   EVT IVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
   auto MMOFlags =
-      L->getMemOperand()->getFlags() & ~MachineMemOperand::MOInvariant;
+      L->getMemOperand()->getFlags() &
+      ~(MachineMemOperand::MOInvariant | MachineMemOperand::MODereferenceable);
   SDValue newL = DAG.getLoad(L->getAddressingMode(), L->getExtensionType(), IVT,
                              SDLoc(N), L->getChain(), L->getBasePtr(),
                              L->getOffset(), L->getPointerInfo(), IVT,
                              L->getAlignment(), MMOFlags, L->getAAInfo());
   // Legalize the chain result by replacing uses of the old value chain with the
   // new one
   ReplaceValueWith(SDValue(N, 1), newL.getValue(1));
 

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

return getLoad(ISD::UNINDEXED, ExtType, VT, dl, Chain, Ptr, Undef,
                  MemVT, MMO);
 }
 
 SDValue SelectionDAG::getIndexedLoad(SDValue OrigLoad, const SDLoc &dl,
                                      SDValue Base, SDValue Offset,
                                      ISD::MemIndexedMode AM) {
   LoadSDNode *LD = cast<LoadSDNode>(OrigLoad);
   assert(LD->getOffset().isUndef() && "Load is already a indexed load!");
-  // Don't propagate the invariant flag.
+  // Don't propagate the invariant or dereferenceable flags.
   auto MMOFlags =
-      LD->getMemOperand()->getFlags() & ~MachineMemOperand::MOInvariant;
+      LD->getMemOperand()->getFlags() &
+      ~(MachineMemOperand::MOInvariant | MachineMemOperand::MODereferenceable);
   return getLoad(AM, LD->getExtensionType(), OrigLoad.getValueType(), dl,
                  LD->getChain(), Base, Offset, LD->getPointerInfo(),
                  LD->getMemoryVT(), LD->getAlignment(), MMOFlags,
                  LD->getAAInfo());
 }
 
 SDValue SelectionDAG::getStore(SDValue Chain, const SDLoc &dl, SDValue Val,
                                SDValue Ptr, MachinePointerInfo PtrInfo,
AddrSpaceCastSDNode::AddrSpaceCastSDNode(unsigned Order, const DebugLoc &dl,
     : SDNode(ISD::ADDRSPACECAST, Order, dl, getSDVTList(VT)),
       SrcAddrSpace(SrcAS), DestAddrSpace(DestAS) {}
 
 MemSDNode::MemSDNode(unsigned Opc, unsigned Order, const DebugLoc &dl,
                      SDVTList VTs, EVT memvt, MachineMemOperand *mmo)
     : SDNode(Opc, Order, dl, VTs), MemoryVT(memvt), MMO(mmo) {
   MemSDNodeBits.IsVolatile = MMO->isVolatile();
   MemSDNodeBits.IsNonTemporal = MMO->isNonTemporal();
+  MemSDNodeBits.IsDereferenceable = MMO->isDereferenceable();
   MemSDNodeBits.IsInvariant = MMO->isInvariant();
 
   // We check here that the size of the memory operand fits within the size of
   // the MMO. This is because the MMO might indicate only a possible address
   // range instead of specifying the affected memory addresses precisely.
   assert(memvt.getStoreSize() <= MMO->getSize() && "Size mismatch!");
 }
 

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

static SDValue getLoadStackGuard(SelectionDAG &DAG, const SDLoc &DL,
   EVT PtrTy = TLI.getPointerTy(DAG.getDataLayout());
   MachineFunction &MF = DAG.getMachineFunction();
   Value *Global = TLI.getSDagStackGuard(*MF.getFunction()->getParent());
   MachineSDNode *Node =
       DAG.getMachineNode(TargetOpcode::LOAD_STACK_GUARD, DL, PtrTy, Chain);
   if (Global) {
     MachinePointerInfo MPInfo(Global);
     MachineInstr::mmo_iterator MemRefs = MF.allocateMemRefsArray(1);
-    auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant;
+    auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant |
+                 MachineMemOperand::MODereferenceable;
     *MemRefs = MF.getMachineMemOperand(MPInfo, Flags, PtrTy.getSizeInBits() / 8,
                                        DAG.getEVTAlignment(PtrTy));
     Node->setMemRefs(MemRefs, MemRefs + 1);
   }
   return SDValue(Node, 0);
 }
 
 /// Codegen a new tail for a stack protector check ParentMBB which has had its
void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
   }
 
   SDValue Ptr = getValue(SV);
 
   Type *Ty = I.getType();
 
   bool isVolatile = I.isVolatile();
   bool isNonTemporal = I.getMetadata(LLVMContext::MD_nontemporal) != nullptr;
-
-  // The IR notion of invariant_load only guarantees that all *non-faulting*
-  // invariant loads result in the same value.  The MI notion of invariant load
-  // guarantees that the load can be legally moved to any location within its
-  // containing function.  The MI notion of invariant_load is stronger than the
-  // IR notion of invariant_load -- an MI invariant_load is an IR invariant_load
-  // with a guarantee that the location being loaded from is dereferenceable
-  // throughout the function's lifetime.
-
-  bool isInvariant = I.getMetadata(LLVMContext::MD_invariant_load) != nullptr &&
-                     isDereferenceablePointer(SV, DAG.getDataLayout());
+  bool isInvariant = I.getMetadata(LLVMContext::MD_invariant_load) != nullptr;
+  bool isDereferenceable = isDereferenceablePointer(SV, DAG.getDataLayout());
   unsigned Alignment = I.getAlignment();
 
   AAMDNodes AAInfo;
   I.getAAMetadata(AAInfo);
   const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
 
   SmallVector<EVT, 4> ValueVTs;
   SmallVector<uint64_t, 4> Offsets;
SDValue A = DAG.getNode(ISD::ADD, dl,
                             &Flags);
     auto MMOFlags = MachineMemOperand::MONone;
     if (isVolatile)
       MMOFlags |= MachineMemOperand::MOVolatile;
     if (isNonTemporal)
       MMOFlags |= MachineMemOperand::MONonTemporal;
     if (isInvariant)
       MMOFlags |= MachineMemOperand::MOInvariant;
+    if (isDereferenceable)
+      MMOFlags |= MachineMemOperand::MODereferenceable;
 
     SDValue L = DAG.getLoad(ValueVTs[i], dl, Root, A,
                             MachinePointerInfo(SV, Offsets[i]), Alignment,
                             MMOFlags, AAInfo, Ranges);
 
     Values[i] = L;
     Chains[ChainI] = L.getValue(1);
   }

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
 
   SDValue TLVPAddr =
       DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
   SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr);
 
   // The first entry in the descriptor is a function pointer that we must call
   // to obtain the address of the variable.
   SDValue Chain = DAG.getEntryNode();
-  SDValue FuncTLVGet =
-      DAG.getLoad(MVT::i64, DL, Chain, DescAddr,
-                  MachinePointerInfo::getGOT(DAG.getMachineFunction()),
-                  /* Alignment = */ 8, MachineMemOperand::MONonTemporal |
-                                           MachineMemOperand::MOInvariant);
+  SDValue FuncTLVGet = DAG.getLoad(
+      MVT::i64, DL, Chain, DescAddr,
+      /* Alignment = */ 8,
+      MachineMemOperand::MONonTemporal | MachineMemOperand::MOInvariant |
+          MachineMemOperand::MODereferenceable);
   Chain = FuncTLVGet.getValue(1);
 
   MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
   MFI.setAdjustsStack(true);
 
   // TLS calls preserve all registers except those that absolutely must be
   // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
   // silly).

llvm/lib/Target/AMDGPU/R600ISelLowering.cpp

for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
     unsigned ValBase = ArgLocs[In.getOrigArgIndex()].getLocMemOffset();
     unsigned PartOffset = VA.getLocMemOffset();
     unsigned Offset = Subtarget->getExplicitKernelArgOffset() + VA.getLocMemOffset();
 
     MachinePointerInfo PtrInfo(UndefValue::get(PtrTy), PartOffset - ValBase);
     SDValue Arg = DAG.getLoad(
         ISD::UNINDEXED, Ext, VT, DL, Chain,
         DAG.getConstant(Offset, DL, MVT::i32), DAG.getUNDEF(MVT::i32), PtrInfo,
-        MemVT, /* Alignment = */ 4,
-        MachineMemOperand::MONonTemporal | MachineMemOperand::MOInvariant);
+        MemVT, /* Alignment = */ 4, MachineMemOperand::MONonTemporal |
+                                        MachineMemOperand::MODereferenceable |
+                                        MachineMemOperand::MOInvariant);
 
     // 4 is the preferred alignment for the CONSTANT memory space.
     InVals.push_back(Arg);
     MFI->setABIArgOffset(Offset + MemVT.getStoreSize());
   }
   return Chain;
 }
 

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
   unsigned Align = DL.getABITypeAlignment(Ty);
 
   ISD::LoadExtType ExtTy = Signed ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
   if (MemVT.isFloatingPoint())
     ExtTy = ISD::EXTLOAD;
 
   SDValue Ptr = LowerParameterPtr(DAG, SL, Chain, Offset);
   return DAG.getLoad(ISD::UNINDEXED, ExtTy, VT, SL, Chain, Ptr, PtrOffset,
-                     PtrInfo, MemVT, Align, MachineMemOperand::MONonTemporal |
-                                                MachineMemOperand::MOInvariant);
+                     PtrInfo, MemVT, Align,
+                     MachineMemOperand::MONonTemporal |
+                         MachineMemOperand::MODereferenceable |
 }
 
 SDValue SITargetLowering::LowerFormalArguments(
     SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
   const SIRegisterInfo *TRI = getSubtarget()->getRegisterInfo();
 
SDValue SITargetLowering::getSegmentAperture(unsigned AS,
   // TODO: We should use the value from the IR intrinsic call, but it might not
   // be available and how do we get it?
   Value *V = UndefValue::get(PointerType::get(Type::getInt8Ty(*DAG.getContext()),
                                               AMDGPUAS::CONSTANT_ADDRESS));
 
   MachinePointerInfo PtrInfo(V, StructOffset);
   return DAG.getLoad(MVT::i32, SL, QueuePtr.getValue(1), Ptr, PtrInfo,
                      MinAlign(64, StructOffset),
+                     MachineMemOperand::MODereferenceable |
                          MachineMemOperand::MOInvariant);
 }
 
 SDValue SITargetLowering::lowerADDRSPACECAST(SDValue Op,
                                              SelectionDAG &DAG) const {
   SDLoc SL(Op);
   const AddrSpaceCastSDNode *ASC = cast<AddrSpaceCastSDNode>(Op);
 
   SDValue Src = ASC->getOperand(0);
SDValue SITargetLowering::LowerGlobalAddress(AMDGPUMachineFunction *MFI,
   Type *Ty = PtrVT.getTypeForEVT(*DAG.getContext());
   PointerType *PtrTy = PointerType::get(Ty, AMDGPUAS::CONSTANT_ADDRESS);
   const DataLayout &DataLayout = DAG.getDataLayout();
   unsigned Align = DataLayout.getABITypeAlignment(PtrTy);
   // FIXME: Use a PseudoSourceValue once those can be assigned an address space.
   MachinePointerInfo PtrInfo(UndefValue::get(PtrTy));
 
   return DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), GOTAddr, PtrInfo, Align,
+                     MachineMemOperand::MODereferenceable |
                          MachineMemOperand::MOInvariant);
 }
 
 SDValue SITargetLowering::lowerTRAP(SDValue Op,
                                     SelectionDAG &DAG) const {
   const MachineFunction &MF = DAG.getMachineFunction();
   DiagnosticInfoUnsupported NoTrap(*MF.getFunction(),
                                    "trap handler not supported",
                                    Op.getDebugLoc(),
return CreateLiveInRegister(DAG, &AMDGPU::VGPR_32RegClass,
       TRI->getPreloadedValue(MF, SIRegisterInfo::WORKITEM_ID_Z), VT);
   case AMDGPUIntrinsic::SI_load_const: {
     SDValue Ops[] = {
       Op.getOperand(1),
       Op.getOperand(2)
     };
 
     MachineMemOperand *MMO = MF.getMachineMemOperand(
         MachinePointerInfo(),
-      MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant,
+        MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
+            MachineMemOperand::MOInvariant,
         VT.getStoreSize(), 4);
     return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
                                    Op->getVTList(), Ops, VT, MMO);
   }
   case AMDGPUIntrinsic::amdgcn_fdiv_fast: {
     return lowerFDIV_FAST(Op, DAG);
   }
   case AMDGPUIntrinsic::SI_vs_load_input:
     return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,

llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp

void ARMBaseInstrInfo::expandLoadStackGuardBase(MachineBasicBlock::iterator MI,
   MachineInstrBuilder MIB;
 
   BuildMI(MBB, MI, DL, get(LoadImmOpc), Reg)
       .addGlobalAddress(GV, 0, ARMII::MO_NONLAZY);
 
   if (Subtarget.isGVIndirectSymbol(GV)) {
     MIB = BuildMI(MBB, MI, DL, get(LoadOpc), Reg);
     MIB.addReg(Reg, RegState::Kill).addImm(0);
-    auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant;
+    auto Flags = MachineMemOperand::MOLoad |
+                 MachineMemOperand::MODereferenceable |
+                 MachineMemOperand::MOInvariant;
     MachineMemOperand *MMO = MBB.getParent()->getMachineMemOperand(
         MachinePointerInfo::getGOT(*MBB.getParent()), Flags, 4, 4);
     MIB.addMemOperand(MMO);
     AddDefaultPred(MIB);
   }
 
   MIB = BuildMI(MBB, MI, DL, get(LoadOpc), Reg);
   MIB.addReg(Reg, RegState::Kill).addImm(0);

llvm/lib/Target/ARM/ARMISelLowering.cpp

if (!PreferIndirect) {
       // ARM call to a local ARM function is predicable.
       isLocalARMFunc = !Subtarget->isThumb() && (isDef || !ARMInterworking);
       // tBX takes a register source operand.
       if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
         assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
         Callee = DAG.getNode(
             ARMISD::WrapperPIC, dl, PtrVt,
             DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, ARMII::MO_NONLAZY));
-        Callee =
-            DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), Callee,
-                        MachinePointerInfo::getGOT(DAG.getMachineFunction()),
-                        /* Alignment = */ 0, MachineMemOperand::MOInvariant);
+        Callee = DAG.getLoad(
+            PtrVt, dl, DAG.getEntryNode(), Callee,
+            /* Alignment = */ 0, MachineMemOperand::MODereferenceable |
+                                     MachineMemOperand::MOInvariant);
       } else if (Subtarget->isTargetCOFF()) {
         assert(Subtarget->isTargetWindows() &&
                "Windows is the only supported COFF target");
         unsigned TargetFlags = GV->hasDLLImportStorageClass()
                                    ? ARMII::MO_DLLIMPORT
                                    : ARMII::MO_NO_FLAG;
         Callee = DAG.getTargetGlobalAddress(GV, dl, PtrVt, /*Offset=*/0,
                                             TargetFlags);
ARMTargetLowering::LowerGlobalTLSAddressDarwin(SDValue Op,
 
   // First step is to get the address of the actua global symbol. This is where
   // the TLS descriptor lives.
   SDValue DescAddr = LowerGlobalAddressDarwin(Op, DAG);
 
   // The first entry in the descriptor is a function pointer that we must call
   // to obtain the address of the variable.
   SDValue Chain = DAG.getEntryNode();
-  SDValue FuncTLVGet =
-      DAG.getLoad(MVT::i32, DL, Chain, DescAddr,
-                  MachinePointerInfo::getGOT(DAG.getMachineFunction()),
-                  /* Alignment = */ 4, MachineMemOperand::MONonTemporal |
-                                           MachineMemOperand::MOInvariant);
+  SDValue FuncTLVGet = DAG.getLoad(
+      MVT::i32, DL, Chain, DescAddr,
+      /* Alignment = */ 4,
+      MachineMemOperand::MONonTemporal | MachineMemOperand::MODereferenceable |
   Chain = FuncTLVGet.getValue(1);
 
   MachineFunction &F = DAG.getMachineFunction();
   MachineFrameInfo &MFI = F.getFrameInfo();
   MFI.setAdjustsStack(true);
 
   // TLS calls preserve all registers except those that absolutely must be
   // trashed: R0 (it takes an argument), LR (it's a call) and CPSR (let's not be

llvm/lib/Target/ARM/ARMInstrInfo.cpp

void ARMInstrInfo::expandLoadStackGuard(MachineBasicBlock::iterator MI) const {
 
   MachineBasicBlock &MBB = *MI->getParent();
   DebugLoc DL = MI->getDebugLoc();
   unsigned Reg = MI->getOperand(0).getReg();
   MachineInstrBuilder MIB;
 
   MIB = BuildMI(MBB, MI, DL, get(ARM::MOV_ga_pcrel_ldr), Reg)
             .addGlobalAddress(GV, 0, ARMII::MO_NONLAZY);
-  auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant;
+  auto Flags = MachineMemOperand::MOLoad |
+               MachineMemOperand::MODereferenceable |
+               MachineMemOperand::MOInvariant;
   MachineMemOperand *MMO = MBB.getParent()->getMachineMemOperand(
       MachinePointerInfo::getGOT(*MBB.getParent()), Flags, 4, 4);
   MIB.addMemOperand(MMO);
   MIB = BuildMI(MBB, MI, DL, get(ARM::LDRi12), Reg);
   MIB.addReg(Reg, RegState::Kill).addImm(0);
   MIB.setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
   AddDefaultPred(MIB);
 }

llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp

if (!PAL.hasAttribute(i + 1, Attribute::ByVal)) {
         // f32 = load ...
         if (NumElts == 1) {
           // We only have one element, so just directly load it
           Value *SrcValue = Constant::getNullValue(PointerType::get(
               EltVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
           SDValue P = DAG.getLoad(
               EltVT, dl, Root, Arg, MachinePointerInfo(SrcValue),
               DL.getABITypeAlignment(EltVT.getTypeForEVT(F->getContext())),
+              MachineMemOperand::MODereferenceable |
                   MachineMemOperand::MOInvariant);
           if (P.getNode())
             P.getNode()->setIROrder(idx + 1);
 
           if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits())
             P = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, P);
           InVals.push_back(P);
           ++InsIdx;
         } else if (NumElts == 2) {
           // V2 load
           // f32,f32 = load ...
           EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, 2);
           Value *SrcValue = Constant::getNullValue(PointerType::get(
               VecVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
           SDValue P = DAG.getLoad(
               VecVT, dl, Root, Arg, MachinePointerInfo(SrcValue),
               DL.getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())),
+              MachineMemOperand::MODereferenceable |
                   MachineMemOperand::MOInvariant);
           if (P.getNode())
             P.getNode()->setIROrder(idx + 1);
 
           SDValue Elt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
                                      DAG.getIntPtrConstant(0, dl));
           SDValue Elt1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
                                      DAG.getIntPtrConstant(1, dl));
 
if (!PAL.hasAttribute(i + 1, Attribute::ByVal)) {
             Value *SrcValue = Constant::getNullValue(
                 PointerType::get(VecVT.getTypeForEVT(F->getContext()),
                                  llvm::ADDRESS_SPACE_PARAM));
             SDValue SrcAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Arg,
                                           DAG.getConstant(Ofst, dl, PtrVT));
             SDValue P = DAG.getLoad(
                 VecVT, dl, Root, SrcAddr, MachinePointerInfo(SrcValue),
                 DL.getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())),
+                MachineMemOperand::MODereferenceable |
                     MachineMemOperand::MOInvariant);
             if (P.getNode())
               P.getNode()->setIROrder(idx + 1);
 
             for (unsigned j = 0; j < VecSize; ++j) {
               if (i + j >= NumElts)
                 break;
               SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
                                         DAG.getIntPtrConstant(j, dl));

llvm/lib/Target/PowerPC/PPCISelLowering.h

public:
     void insertSSPDeclarations(Module &M) const override;
 
   private:
     struct ReuseLoadInfo {
       SDValue Ptr;
       SDValue Chain;
       SDValue ResChain;
       MachinePointerInfo MPI;
+      bool IsDereferenceable;
       bool IsInvariant;
       unsigned Alignment;
       AAMDNodes AAInfo;
       const MDNode *Ranges;
 
-      ReuseLoadInfo() : IsInvariant(false), Alignment(0), Ranges(nullptr) {}
+      ReuseLoadInfo()
+          : IsDereferenceable(false), IsInvariant(false), Alignment(0),
+            Ranges(nullptr) {}
+
+      MachineMemOperand::Flags MMOFlags() const {
+        MachineMemOperand::Flags F = MachineMemOperand::MONone;
+        if (IsDereferenceable)
+          F |= MachineMemOperand::MODereferenceable;
+        if (IsInvariant)
+          F |= MachineMemOperand::MOInvariant;
+        return F;
+      }
     };
 
     bool canReuseLoadAddress(SDValue Op, EVT MemVT, ReuseLoadInfo &RLI,
                              SelectionDAG &DAG,
                              ISD::LoadExtType ET = ISD::NON_EXTLOAD) const;
     void spliceIntoChain(SDValue ResChain, SDValue NewResChain,
                          SelectionDAG &DAG) const;
 

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

if (isSVR4ABI && isPPC64 && !isELFv2ABI) {
 
       // Load the address of the function entry point from the function
       // descriptor.
       SDValue LDChain = CallSeqStart.getValue(CallSeqStart->getNumValues()-1);
       if (LDChain.getValueType() == MVT::Glue)
         LDChain = CallSeqStart.getValue(CallSeqStart->getNumValues()-2);
 
       auto MMOFlags = Subtarget.hasInvariantFunctionDescriptors()
-                          ? MachineMemOperand::MOInvariant
+                          ? (MachineMemOperand::MODereferenceable |
+                             MachineMemOperand::MOInvariant)
                           : MachineMemOperand::MONone;
 
       MachinePointerInfo MPI(CS ? CS->getCalledValue() : nullptr);
       SDValue LoadFuncPtr = DAG.getLoad(MVT::i64, dl, LDChain, Callee, MPI,
                                         /* Alignment = */ 8, MMOFlags);
 
       // Load environment pointer into r11.
       SDValue PtrOff = DAG.getIntPtrConstant(16, dl);
SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
                                           const SDLoc &dl) const {
   if (Subtarget.hasDirectMove() && Subtarget.isPPC64())
     return LowerFP_TO_INTDirectMove(Op, DAG, dl);
 
   ReuseLoadInfo RLI;
   LowerFP_TO_INTForReuse(Op, RLI, DAG, dl);
 
   return DAG.getLoad(Op.getValueType(), dl, RLI.Chain, RLI.Ptr, RLI.MPI,
-                     RLI.Alignment,
-                     RLI.IsInvariant ? MachineMemOperand::MOInvariant
-                                     : MachineMemOperand::MONone,
-                     RLI.AAInfo, RLI.Ranges);
+                     RLI.Alignment, RLI.MMOFlags(), RLI.AAInfo, RLI.Ranges);
 }
 
 // We're trying to insert a regular store, S, and then a load, L. If the
 // incoming value, O, is a load, we might just be able to have our load use the
 // address used by O. However, we don't know if anything else will store to
 // that address before we can load from it. To prevent this situation, we need
 // to insert our load, L, into the chain as a peer of O. To do this, we give L
 // the same chain operand as O, we create a token factor from the chain results
if (LD->isIndexed() && !LD->getOffset().isUndef()) {
     assert(LD->getAddressingMode() == ISD::PRE_INC &&
            "Non-pre-inc AM on PPC?");
     RLI.Ptr = DAG.getNode(ISD::ADD, dl, RLI.Ptr.getValueType(), RLI.Ptr,
                           LD->getOffset());
   }
 
   RLI.Chain = LD->getChain();
   RLI.MPI = LD->getPointerInfo();
+  RLI.IsDereferenceable = LD->isDereferenceable();
   RLI.IsInvariant = LD->isInvariant();
   RLI.Alignment = LD->getAlignment();
   RLI.AAInfo = LD->getAAInfo();
   RLI.Ranges = LD->getRanges();
 
   RLI.ResChain = SDValue(LD, LD->isIndexed() ? 2 : 1);
   return true;
 }
if (Op.getValueType() == MVT::f32 &&
       SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT);
     }
 
     ReuseLoadInfo RLI;
     SDValue Bits;
 
     MachineFunction &MF = DAG.getMachineFunction();
     if (canReuseLoadAddress(SINT, MVT::i64, RLI, DAG)) {
-      Bits =
-          DAG.getLoad(MVT::f64, dl, RLI.Chain, RLI.Ptr, RLI.MPI, RLI.Alignment,
-                      RLI.IsInvariant ? MachineMemOperand::MOInvariant
-                                      : MachineMemOperand::MONone,
-                      RLI.AAInfo, RLI.Ranges);
+      Bits = DAG.getLoad(MVT::f64, dl, RLI.Chain, RLI.Ptr, RLI.MPI,
+                         RLI.Alignment, RLI.MMOFlags(), RLI.AAInfo, RLI.Ranges);
       spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG);
     } else if (Subtarget.hasLFIWAX() &&
                canReuseLoadAddress(SINT, MVT::i32, RLI, DAG, ISD::SEXTLOAD)) {
       MachineMemOperand *MMO =
         MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4,
                                 RLI.Alignment, RLI.AAInfo, RLI.Ranges);
       SDValue Ops[] = { RLI.Chain, RLI.Ptr };
       Bits = DAG.getMemIntrinsicNode(PPCISD::LFIWAX, dl,

llvm/lib/Target/SystemZ/SystemZISelDAGToDAG.cpp

bool SystemZDAGToDAGISel::canUseBlockOperation(StoreSDNode *Store,
   if (Load->getMemoryVT() != Store->getMemoryVT())
     return false;
 
   // Volatility stops an access from being decomposed.
   if (Load->isVolatile() || Store->isVolatile())
     return false;
 
   // There's no chance of overlap if the load is invariant.
-  if (Load->isInvariant())
+  if (Load->isInvariant() && Load->isDereferenceable())
     return true;
 
   // Otherwise we need to check whether there's an alias.
   const Value *V1 = Load->getMemOperand()->getValue();
   const Value *V2 = Store->getMemOperand()->getValue();
   if (!V1 || !V2)
     return false;
 

llvm/lib/Target/X86/X86InstrInfo.cpp

 // code sequence is needed for other targets.
 static void expandLoadStackGuard(MachineInstrBuilder &MIB,
                                  const TargetInstrInfo &TII) {
   MachineBasicBlock &MBB = *MIB->getParent();
   DebugLoc DL = MIB->getDebugLoc();
   unsigned Reg = MIB->getOperand(0).getReg();
   const GlobalValue *GV =
       cast<GlobalValue>((*MIB->memoperands_begin())->getValue());
-  auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant;
+  auto Flags = MachineMemOperand::MOLoad |
+               MachineMemOperand::MODereferenceable |
+               MachineMemOperand::MOInvariant;
   MachineMemOperand *MMO = MBB.getParent()->getMachineMemOperand(
       MachinePointerInfo::getGOT(*MBB.getParent()), Flags, 8, 8);
   MachineBasicBlock::iterator I = MIB.getInstr();
 
   BuildMI(MBB, I, DL, TII.get(X86::MOV64rm), Reg).addReg(X86::RIP).addImm(1)
       .addReg(0).addGlobalAddress(GV, 0, X86II::MO_GOTPCREL).addReg(0)
       .addMemOperand(MMO);
   MIB->setDebugLoc(DL);