[PAC][AArch64] Lower ptrauth constants in code (#96879)

This re-applies #94241 after fixing buildbot failure, see
https://lab.llvm.org/buildbot/#/builders/51/builds/570

According to standard, `constexpr` variables and `const` variables
initialized with constant expressions can be used in lambdas w/o
capturing - see https://en.cppreference.com/w/cpp/language/lambda.
However, MSVC used on buildkite seems to ignore that rule and does not
allow using such uncaptured variables in lambdas: we have "error C3493:
'Mask16' cannot be implicitly captured because no default capture mode
has been specified" - see
https://buildkite.com/llvm-project/github-pull-requests/builds/73238

Explicitly capturing such a variable, however, makes buildbot fail with
"error: lambda capture 'Mask16' is not required to be captured for this
use [-Werror,-Wunused-lambda-capture]" - see
https://lab.llvm.org/buildbot/#/builders/51/builds/570.

Fix both cases by using `0xffff` value directly instead of giving a name
to it.

Original PR description below.

Depends on #94240.

Define the following pseudos for lowering ptrauth constants in code:

- non-`extern_weak`:
  - no GOT load needed: `MOVaddrPAC` - similar to `MOVaddr`, with added
PAC;
  - GOT load needed: `LOADgotPAC` - similar to `LOADgot`, with added PAC;
- `extern_weak`: `LOADauthptrstatic` - similar to `LOADgot`, but use a
special stub slot named `sym$auth_ptr$key$disc` filled by dynamic linker
during relocation resolving instead of a GOT slot.

---------

Co-authored-by: Ahmed Bougacha <ahmed@bougacha.org>

14 files changed, 666 insertions, 3 deletions

diff --git a/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp b/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
index ddf6d3c..c9986df 100644
--- a/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
@@ -3494,7 +3494,11 @@ bool IRTranslator::translate(const Constant &C, Register Reg) {
     EntryBuilder->buildConstant(Reg, 0);
   else if (auto GV = dyn_cast<GlobalValue>(&C))
     EntryBuilder->buildGlobalValue(Reg, GV);
-  else if (auto CAZ = dyn_cast<ConstantAggregateZero>(&C)) {
+  else if (auto CPA = dyn_cast<ConstantPtrAuth>(&C)) {
+    Register Addr = getOrCreateVReg(*CPA->getPointer());
+    Register AddrDisc = getOrCreateVReg(*CPA->getAddrDiscriminator());
+    EntryBuilder->buildConstantPtrAuth(Reg, CPA, Addr, AddrDisc);
+  } else if (auto CAZ = dyn_cast<ConstantAggregateZero>(&C)) {
     if (!isa<FixedVectorType>(CAZ->getType()))
       return false;
     // Return the scalar if it is a <1 x Ty> vector.
diff --git a/llvm/lib/CodeGen/GlobalISel/MachineIRBuilder.cpp b/llvm/lib/CodeGen/GlobalISel/MachineIRBuilder.cpp
index 37aa4e0..06a6c1f9 100644
--- a/llvm/lib/CodeGen/GlobalISel/MachineIRBuilder.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/MachineIRBuilder.cpp
@@ -397,6 +397,19 @@ MachineInstrBuilder MachineIRBuilder::buildFConstant(const DstOp &Res,
   return buildFConstant(Res, *CFP);
 }
 
+MachineInstrBuilder
+MachineIRBuilder::buildConstantPtrAuth(const DstOp &Res,
+                                       const ConstantPtrAuth *CPA,
+                                       Register Addr, Register AddrDisc) {
+  auto MIB = buildInstr(TargetOpcode::G_PTRAUTH_GLOBAL_VALUE);
+  Res.addDefToMIB(*getMRI(), MIB);
+  MIB.addUse(Addr);
+  MIB.addImm(CPA->getKey()->getZExtValue());
+  MIB.addUse(AddrDisc);
+  MIB.addImm(CPA->getDiscriminator()->getZExtValue());
+  return MIB;
+}
+
 MachineInstrBuilder MachineIRBuilder::buildBrCond(const SrcOp &Tst,
                                                   MachineBasicBlock &Dest) {
   assert(Tst.getLLTTy(*getMRI()).isScalar() && "invalid operand type");
diff --git a/llvm/lib/CodeGen/MachineModuleInfoImpls.cpp b/llvm/lib/CodeGen/MachineModuleInfoImpls.cpp
index 9c3b319..f114f1e 100644
--- a/llvm/lib/CodeGen/MachineModuleInfoImpls.cpp
+++ b/llvm/lib/CodeGen/MachineModuleInfoImpls.cpp
@@ -13,6 +13,7 @@
 
 #include "llvm/CodeGen/MachineModuleInfoImpls.h"
 #include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/STLExtras.h"
 #include "llvm/MC/MCSymbol.h"
 
 using namespace llvm;
@@ -41,3 +42,25 @@ MachineModuleInfoImpl::SymbolListTy MachineModuleInfoImpl::getSortedStubs(
   Map.clear();
   return List;
 }
+
+template <typename MachineModuleInfoTarget>
+static typename MachineModuleInfoTarget::AuthStubListTy getAuthGVStubListHelper(
+    DenseMap<MCSymbol *, typename MachineModuleInfoTarget::AuthStubInfo>
+        &AuthPtrStubs) {
+  typename MachineModuleInfoTarget::AuthStubListTy List(AuthPtrStubs.begin(),
+                                                        AuthPtrStubs.end());
+
+  if (!List.empty())
+    llvm::sort(List.begin(), List.end(),
+               [](const typename MachineModuleInfoTarget::AuthStubPairTy &LHS,
+                  const typename MachineModuleInfoTarget::AuthStubPairTy &RHS) {
+                 return LHS.first->getName() < RHS.first->getName();
+               });
+
+  AuthPtrStubs.clear();
+  return List;
+}
+
+MachineModuleInfoELF::AuthStubListTy MachineModuleInfoELF::getAuthGVStubList() {
+  return getAuthGVStubListHelper<MachineModuleInfoELF>(AuthPtrStubs);
+}
diff --git a/llvm/lib/CodeGen/MachineVerifier.cpp b/llvm/lib/CodeGen/MachineVerifier.cpp
index 9ea238c..0c8a0f2 100644
--- a/llvm/lib/CodeGen/MachineVerifier.cpp
+++ b/llvm/lib/CodeGen/MachineVerifier.cpp
@@ -2066,6 +2066,12 @@ void MachineVerifier::verifyPreISelGenericInstruction(const MachineInstr *MI) {
       report("Dst operand 0 must be a pointer", MI);
     break;
   }
+  case TargetOpcode::G_PTRAUTH_GLOBAL_VALUE: {
+    const MachineOperand &AddrOp = MI->getOperand(1);
+    if (!AddrOp.isReg() || !MRI->getType(AddrOp.getReg()).isPointer())
+      report("addr operand must be a pointer", &AddrOp, 1);
+    break;
+  }
   default:
     break;
   }
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index e45569b..465919d 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -1802,6 +1802,13 @@ SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
     if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
       return DAG.getGlobalAddress(GV, getCurSDLoc(), VT);
 
+    if (const ConstantPtrAuth *CPA = dyn_cast<ConstantPtrAuth>(C)) {
+      return DAG.getNode(ISD::PtrAuthGlobalAddress, getCurSDLoc(), VT,
+                         getValue(CPA->getPointer()), getValue(CPA->getKey()),
+                         getValue(CPA->getAddrDiscriminator()),
+                         getValue(CPA->getDiscriminator()));
+    }
+
     if (isa<ConstantPointerNull>(C)) {
       unsigned AS = V->getType()->getPointerAddressSpace();
       return DAG.getConstant(0, getCurSDLoc(),
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp
index a7555d6..c1d2c09 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp
@@ -75,6 +75,7 @@ std::string SDNode::getOperationName(const SelectionDAG *G) const {
     }
     return "<<Unknown Node #" + utostr(getOpcode()) + ">>";
 
+    // clang-format off
 #ifndef NDEBUG
   case ISD::DELETED_NODE:               return "<<Deleted Node!>>";
 #endif
@@ -126,6 +127,7 @@ std::string SDNode::getOperationName(const SelectionDAG *G) const {
   case ISD::ConstantFP:                 return "ConstantFP";
   case ISD::GlobalAddress:              return "GlobalAddress";
   case ISD::GlobalTLSAddress:           return "GlobalTLSAddress";
+  case ISD::PtrAuthGlobalAddress:       return "PtrAuthGlobalAddress";
   case ISD::FrameIndex:                 return "FrameIndex";
   case ISD::JumpTable:                  return "JumpTable";
   case ISD::JUMP_TABLE_DEBUG_INFO:
@@ -168,8 +170,6 @@ std::string SDNode::getOperationName(const SelectionDAG *G) const {
       return "OpaqueTargetConstant";
     return "TargetConstant";
 
-    // clang-format off
-
   case ISD::TargetConstantFP:           return "TargetConstantFP";
   case ISD::TargetGlobalAddress:        return "TargetGlobalAddress";
   case ISD::TargetGlobalTLSAddress:     return "TargetGlobalTLSAddress";
diff --git a/llvm/lib/Target/AArch64/AArch64AsmPrinter.cpp b/llvm/lib/Target/AArch64/AArch64AsmPrinter.cpp
index da11539..00b8d7b 100644
--- a/llvm/lib/Target/AArch64/AArch64AsmPrinter.cpp
+++ b/llvm/lib/Target/AArch64/AArch64AsmPrinter.cpp
@@ -133,6 +133,13 @@ public:
   unsigned emitPtrauthDiscriminator(uint16_t Disc, unsigned AddrDisc,
                                     unsigned &InstsEmitted);
 
+  // Emit the sequence for LOADauthptrstatic
+  void LowerLOADauthptrstatic(const MachineInstr &MI);
+
+  // Emit the sequence for LOADgotPAC/MOVaddrPAC (either GOT adrp-ldr or
+  // adrp-add followed by PAC sign)
+  void LowerMOVaddrPAC(const MachineInstr &MI);
+
   /// tblgen'erated driver function for lowering simple MI->MC
   /// pseudo instructions.
   bool emitPseudoExpansionLowering(MCStreamer &OutStreamer,
@@ -840,6 +847,15 @@ void AArch64AsmPrinter::emitHwasanMemaccessSymbols(Module &M) {
   }
 }
 
+template <typename MachineModuleInfoTarget>
+static void emitAuthenticatedPointer(
+    MCStreamer &OutStreamer, MCSymbol *StubLabel,
+    const typename MachineModuleInfoTarget::AuthStubInfo &StubInfo) {
+  // sym$auth_ptr$key$disc:
+  OutStreamer.emitLabel(StubLabel);
+  OutStreamer.emitValue(StubInfo.AuthPtrRef, /*size=*/8);
+}
+
 void AArch64AsmPrinter::emitEndOfAsmFile(Module &M) {
   emitHwasanMemaccessSymbols(M);
 
@@ -853,6 +869,25 @@ void AArch64AsmPrinter::emitEndOfAsmFile(Module &M) {
     OutStreamer->emitAssemblerFlag(MCAF_SubsectionsViaSymbols);
   }
 
+  if (TT.isOSBinFormatELF()) {
+    // Output authenticated pointers as indirect symbols, if we have any.
+    MachineModuleInfoELF &MMIELF = MMI->getObjFileInfo<MachineModuleInfoELF>();
+
+    auto Stubs = MMIELF.getAuthGVStubList();
+
+    if (!Stubs.empty()) {
+      const TargetLoweringObjectFile &TLOF = getObjFileLowering();
+      OutStreamer->switchSection(TLOF.getDataSection());
+      emitAlignment(Align(8));
+
+      for (const auto &Stub : Stubs)
+        emitAuthenticatedPointer<MachineModuleInfoELF>(*OutStreamer, Stub.first,
+                                                       Stub.second);
+
+      OutStreamer->addBlankLine();
+    }
+  }
+
   // Emit stack and fault map information.
   FM.serializeToFaultMapSection();
 
@@ -1623,6 +1658,213 @@ AArch64AsmPrinter::lowerConstantPtrAuth(const ConstantPtrAuth &CPA) {
                                    CPA.hasAddressDiscriminator(), Ctx);
 }
 
+void AArch64AsmPrinter::LowerLOADauthptrstatic(const MachineInstr &MI) {
+  unsigned DstReg = MI.getOperand(0).getReg();
+  const MachineOperand &GAOp = MI.getOperand(1);
+  const uint64_t KeyC = MI.getOperand(2).getImm();
+  assert(KeyC <= AArch64PACKey::LAST &&
+         "key is out of range [0, AArch64PACKey::LAST]");
+  const auto Key = (AArch64PACKey::ID)KeyC;
+  const uint64_t Disc = MI.getOperand(3).getImm();
+  assert(isUInt<16>(Disc) &&
+         "constant discriminator is out of range [0, 0xffff]");
+
+  // Emit instruction sequence like the following:
+  //   ADRP x16, symbol$auth_ptr$key$disc
+  //   LDR x16, [x16, :lo12:symbol$auth_ptr$key$disc]
+  //
+  // Where the $auth_ptr$ symbol is the stub slot containing the signed pointer
+  // to symbol.
+  assert(TM.getTargetTriple().isOSBinFormatELF() &&
+         "LOADauthptrstatic is implemented only for ELF");
+  const auto &TLOF =
+      static_cast<const AArch64_ELFTargetObjectFile &>(getObjFileLowering());
+
+  assert(GAOp.getOffset() == 0 &&
+         "non-zero offset for $auth_ptr$ stub slots is not supported");
+  const MCSymbol *GASym = TM.getSymbol(GAOp.getGlobal());
+  MCSymbol *AuthPtrStubSym =
+      TLOF.getAuthPtrSlotSymbol(TM, &MF->getMMI(), GASym, Key, Disc);
+
+  MachineOperand StubMOHi =
+      MachineOperand::CreateMCSymbol(AuthPtrStubSym, AArch64II::MO_PAGE);
+  MachineOperand StubMOLo = MachineOperand::CreateMCSymbol(
+      AuthPtrStubSym, AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
+  MCOperand StubMCHi, StubMCLo;
+
+  MCInstLowering.lowerOperand(StubMOHi, StubMCHi);
+  MCInstLowering.lowerOperand(StubMOLo, StubMCLo);
+
+  EmitToStreamer(
+      *OutStreamer,
+      MCInstBuilder(AArch64::ADRP).addReg(DstReg).addOperand(StubMCHi));
+
+  EmitToStreamer(*OutStreamer, MCInstBuilder(AArch64::LDRXui)
+                                   .addReg(DstReg)
+                                   .addReg(DstReg)
+                                   .addOperand(StubMCLo));
+}
+
+void AArch64AsmPrinter::LowerMOVaddrPAC(const MachineInstr &MI) {
+  unsigned InstsEmitted = 0;
+  auto EmitAndIncrement = [this, &InstsEmitted](const MCInst &Inst) {
+    EmitToStreamer(*OutStreamer, Inst);
+    ++InstsEmitted;
+  };
+
+  const bool IsGOTLoad = MI.getOpcode() == AArch64::LOADgotPAC;
+  MachineOperand GAOp = MI.getOperand(0);
+  const uint64_t KeyC = MI.getOperand(1).getImm();
+  assert(KeyC <= AArch64PACKey::LAST &&
+         "key is out of range [0, AArch64PACKey::LAST]");
+  const auto Key = (AArch64PACKey::ID)KeyC;
+  const unsigned AddrDisc = MI.getOperand(2).getReg();
+  const uint64_t Disc = MI.getOperand(3).getImm();
+  assert(isUInt<16>(Disc) &&
+         "constant discriminator is out of range [0, 0xffff]");
+
+  const int64_t Offset = GAOp.getOffset();
+  GAOp.setOffset(0);
+
+  // Emit:
+  // target materialization:
+  // - via GOT:
+  //     adrp x16, :got:target
+  //     ldr x16, [x16, :got_lo12:target]
+  //     add offset to x16 if offset != 0
+  //
+  // - direct:
+  //     adrp x16, target
+  //     add x16, x16, :lo12:target
+  //     add offset to x16 if offset != 0
+  //
+  // add offset to x16:
+  // - abs(offset) fits 24 bits:
+  //     add/sub x16, x16, #<offset>[, #lsl 12] (up to 2 instructions)
+  // - abs(offset) does not fit 24 bits:
+  //   - offset < 0:
+  //       movn+movk sequence filling x17 register with the offset (up to 4
+  //       instructions)
+  //       add x16, x16, x17
+  //   - offset > 0:
+  //       movz+movk sequence filling x17 register with the offset (up to 4
+  //       instructions)
+  //       add x16, x16, x17
+  //
+  // signing:
+  // - 0 discriminator:
+  //     paciza x16
+  // - Non-0 discriminator, no address discriminator:
+  //     mov x17, #Disc
+  //     pacia x16, x17
+  // - address discriminator (with potentially folded immediate discriminator):
+  //     pacia x16, xAddrDisc
+
+  MachineOperand GAMOHi(GAOp), GAMOLo(GAOp);
+  MCOperand GAMCHi, GAMCLo;
+
+  GAMOHi.setTargetFlags(AArch64II::MO_PAGE);
+  GAMOLo.setTargetFlags(AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
+  if (IsGOTLoad) {
+    GAMOHi.addTargetFlag(AArch64II::MO_GOT);
+    GAMOLo.addTargetFlag(AArch64II::MO_GOT);
+  }
+
+  MCInstLowering.lowerOperand(GAMOHi, GAMCHi);
+  MCInstLowering.lowerOperand(GAMOLo, GAMCLo);
+
+  EmitAndIncrement(
+      MCInstBuilder(AArch64::ADRP).addReg(AArch64::X16).addOperand(GAMCHi));
+
+  if (IsGOTLoad) {
+    EmitAndIncrement(MCInstBuilder(AArch64::LDRXui)
+                         .addReg(AArch64::X16)
+                         .addReg(AArch64::X16)
+                         .addOperand(GAMCLo));
+  } else {
+    EmitAndIncrement(MCInstBuilder(AArch64::ADDXri)
+                         .addReg(AArch64::X16)
+                         .addReg(AArch64::X16)
+                         .addOperand(GAMCLo)
+                         .addImm(0));
+  }
+
+  if (Offset != 0) {
+    const uint64_t AbsOffset = (Offset > 0 ? Offset : -((uint64_t)Offset));
+    const bool IsNeg = Offset < 0;
+    if (isUInt<24>(AbsOffset)) {
+      for (int BitPos = 0; BitPos != 24 && (AbsOffset >> BitPos);
+           BitPos += 12) {
+        EmitAndIncrement(
+            MCInstBuilder(IsNeg ? AArch64::SUBXri : AArch64::ADDXri)
+                .addReg(AArch64::X16)
+                .addReg(AArch64::X16)
+                .addImm((AbsOffset >> BitPos) & 0xfff)
+                .addImm(AArch64_AM::getShifterImm(AArch64_AM::LSL, BitPos)));
+      }
+    } else {
+      const uint64_t UOffset = Offset;
+      EmitAndIncrement(MCInstBuilder(IsNeg ? AArch64::MOVNXi : AArch64::MOVZXi)
+                           .addReg(AArch64::X17)
+                           .addImm((IsNeg ? ~UOffset : UOffset) & 0xffff)
+                           .addImm(/*shift=*/0));
+      auto NeedMovk = [IsNeg, UOffset](int BitPos) -> bool {
+        assert(BitPos == 16 || BitPos == 32 || BitPos == 48);
+        uint64_t Shifted = UOffset >> BitPos;
+        if (!IsNeg)
+          return Shifted != 0;
+        for (int I = 0; I != 64 - BitPos; I += 16)
+          if (((Shifted >> I) & 0xffff) != 0xffff)
+            return true;
+        return false;
+      };
+      for (int BitPos = 16; BitPos != 64 && NeedMovk(BitPos); BitPos += 16) {
+        EmitAndIncrement(MCInstBuilder(AArch64::MOVKXi)
+                             .addReg(AArch64::X17)
+                             .addReg(AArch64::X17)
+                             .addImm((UOffset >> BitPos) & 0xffff)
+                             .addImm(/*shift=*/BitPos));
+      }
+      EmitAndIncrement(MCInstBuilder(AArch64::ADDXrs)
+                           .addReg(AArch64::X16)
+                           .addReg(AArch64::X16)
+                           .addReg(AArch64::X17)
+                           .addImm(/*shift=*/0));
+    }
+  }
+
+  unsigned DiscReg = AddrDisc;
+  if (Disc != 0) {
+    if (AddrDisc != AArch64::XZR) {
+      EmitAndIncrement(MCInstBuilder(AArch64::ORRXrs)
+                           .addReg(AArch64::X17)
+                           .addReg(AArch64::XZR)
+                           .addReg(AddrDisc)
+                           .addImm(0));
+      EmitAndIncrement(MCInstBuilder(AArch64::MOVKXi)
+                           .addReg(AArch64::X17)
+                           .addReg(AArch64::X17)
+                           .addImm(Disc)
+                           .addImm(/*shift=*/48));
+    } else {
+      EmitAndIncrement(MCInstBuilder(AArch64::MOVZXi)
+                           .addReg(AArch64::X17)
+                           .addImm(Disc)
+                           .addImm(/*shift=*/0));
+    }
+    DiscReg = AArch64::X17;
+  }
+
+  auto MIB = MCInstBuilder(getPACOpcodeForKey(Key, DiscReg == AArch64::XZR))
+                 .addReg(AArch64::X16)
+                 .addReg(AArch64::X16);
+  if (DiscReg != AArch64::XZR)
+    MIB.addReg(DiscReg);
+  EmitAndIncrement(MIB);
+
+  assert(STI->getInstrInfo()->getInstSizeInBytes(MI) >= InstsEmitted * 4);
+}
+
 // Simple pseudo-instructions have their lowering (with expansion to real
 // instructions) auto-generated.
 #include "AArch64GenMCPseudoLowering.inc"
@@ -1758,6 +2000,15 @@ void AArch64AsmPrinter::emitInstruction(const MachineInstr *MI) {
     return;
   }
 
+  case AArch64::LOADauthptrstatic:
+    LowerLOADauthptrstatic(*MI);
+    return;
+
+  case AArch64::LOADgotPAC:
+  case AArch64::MOVaddrPAC:
+    LowerMOVaddrPAC(*MI);
+    return;
+
   case AArch64::BLRA:
     emitPtrauthBranch(MI);
     return;
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 61de30c..fba6b62 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -511,6 +511,8 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
   setOperationAction(ISD::JumpTable, MVT::i64, Custom);
   setOperationAction(ISD::SETCCCARRY, MVT::i64, Custom);
 
+  setOperationAction(ISD::PtrAuthGlobalAddress, MVT::i64, Custom);
+
   setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
   setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
   setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
@@ -6667,6 +6669,8 @@ SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
     return LowerGlobalAddress(Op, DAG);
   case ISD::GlobalTLSAddress:
     return LowerGlobalTLSAddress(Op, DAG);
+  case ISD::PtrAuthGlobalAddress:
+    return LowerPtrAuthGlobalAddress(Op, DAG);
   case ISD::SETCC:
   case ISD::STRICT_FSETCC:
   case ISD::STRICT_FSETCCS:
@@ -9517,6 +9521,133 @@ SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
   llvm_unreachable("Unexpected platform trying to use TLS");
 }
 
+//===----------------------------------------------------------------------===//
+//                      PtrAuthGlobalAddress lowering
+//
+// We have 3 lowering alternatives to choose from:
+// - MOVaddrPAC: similar to MOVaddr, with added PAC.
+//   If the GV doesn't need a GOT load (i.e., is locally defined)
+//   materialize the pointer using adrp+add+pac. See LowerMOVaddrPAC.
+//
+// - LOADgotPAC: similar to LOADgot, with added PAC.
+//   If the GV needs a GOT load, materialize the pointer using the usual
+//   GOT adrp+ldr, +pac. Pointers in GOT are assumed to be not signed, the GOT
+//   section is assumed to be read-only (for example, via relro mechanism). See
+//   LowerMOVaddrPAC.
+//
+// - LOADauthptrstatic: similar to LOADgot, but use a
+//   special stub slot instead of a GOT slot.
+//   Load a signed pointer for symbol 'sym' from a stub slot named
+//   'sym$auth_ptr$key$disc' filled by dynamic linker during relocation
+//   resolving. This usually lowers to adrp+ldr, but also emits an entry into
+//   .data with an
+//   @AUTH relocation. See LowerLOADauthptrstatic.
+//
+// All 3 are pseudos that are expand late to longer sequences: this lets us
+// provide integrity guarantees on the to-be-signed intermediate values.
+//
+// LOADauthptrstatic is undesirable because it requires a large section filled
+// with often similarly-signed pointers, making it a good harvesting target.
+// Thus, it's only used for ptrauth references to extern_weak to avoid null
+// checks.
+
+SDValue AArch64TargetLowering::LowerPtrAuthGlobalAddressStatically(
+    SDValue TGA, SDLoc DL, EVT VT, AArch64PACKey::ID KeyC,
+    SDValue Discriminator, SDValue AddrDiscriminator, SelectionDAG &DAG) const {
+  const auto *TGN = cast<GlobalAddressSDNode>(TGA.getNode());
+  assert(TGN->getGlobal()->hasExternalWeakLinkage());
+
+  // Offsets and extern_weak don't mix well: ptrauth aside, you'd get the
+  // offset alone as a pointer if the symbol wasn't available, which would
+  // probably break null checks in users. Ptrauth complicates things further:
+  // error out.
+  if (TGN->getOffset() != 0)
+    report_fatal_error(
+        "unsupported non-zero offset in weak ptrauth global reference");
+
+  if (!isNullConstant(AddrDiscriminator))
+    report_fatal_error("unsupported weak addr-div ptrauth global");
+
+  SDValue Key = DAG.getTargetConstant(KeyC, DL, MVT::i32);
+  return SDValue(DAG.getMachineNode(AArch64::LOADauthptrstatic, DL, MVT::i64,
+                                    {TGA, Key, Discriminator}),
+                 0);
+}
+
+SDValue
+AArch64TargetLowering::LowerPtrAuthGlobalAddress(SDValue Op,
+                                                 SelectionDAG &DAG) const {
+  SDValue Ptr = Op.getOperand(0);
+  uint64_t KeyC = Op.getConstantOperandVal(1);
+  SDValue AddrDiscriminator = Op.getOperand(2);
+  uint64_t DiscriminatorC = Op.getConstantOperandVal(3);
+  EVT VT = Op.getValueType();
+  SDLoc DL(Op);
+
+  if (KeyC > AArch64PACKey::LAST)
+    report_fatal_error("key in ptrauth global out of range [0, " +
+                       Twine((int)AArch64PACKey::LAST) + "]");
+
+  // Blend only works if the integer discriminator is 16-bit wide.
+  if (!isUInt<16>(DiscriminatorC))
+    report_fatal_error(
+        "constant discriminator in ptrauth global out of range [0, 0xffff]");
+
+  // Choosing between 3 lowering alternatives is target-specific.
+  if (!Subtarget->isTargetELF())
+    report_fatal_error("ptrauth global lowering is only implemented for ELF");
+
+  int64_t PtrOffsetC = 0;
+  if (Ptr.getOpcode() == ISD::ADD) {
+    PtrOffsetC = Ptr.getConstantOperandVal(1);
+    Ptr = Ptr.getOperand(0);
+  }
+  const auto *PtrN = cast<GlobalAddressSDNode>(Ptr.getNode());
+  const GlobalValue *PtrGV = PtrN->getGlobal();
+
+  // Classify the reference to determine whether it needs a GOT load.
+  const unsigned OpFlags =
+      Subtarget->ClassifyGlobalReference(PtrGV, getTargetMachine());
+  const bool NeedsGOTLoad = ((OpFlags & AArch64II::MO_GOT) != 0);
+  assert(((OpFlags & (~AArch64II::MO_GOT)) == 0) &&
+         "unsupported non-GOT op flags on ptrauth global reference");
+
+  // Fold any offset into the GV; our pseudos expect it there.
+  PtrOffsetC += PtrN->getOffset();
+  SDValue TPtr = DAG.getTargetGlobalAddress(PtrGV, DL, VT, PtrOffsetC,
+                                            /*TargetFlags=*/0);
+  assert(PtrN->getTargetFlags() == 0 &&
+         "unsupported target flags on ptrauth global");
+
+  SDValue Key = DAG.getTargetConstant(KeyC, DL, MVT::i32);
+  SDValue Discriminator = DAG.getTargetConstant(DiscriminatorC, DL, MVT::i64);
+  SDValue TAddrDiscriminator = !isNullConstant(AddrDiscriminator)
+                                   ? AddrDiscriminator
+                                   : DAG.getRegister(AArch64::XZR, MVT::i64);
+
+  // No GOT load needed -> MOVaddrPAC
+  if (!NeedsGOTLoad) {
+    assert(!PtrGV->hasExternalWeakLinkage() && "extern_weak should use GOT");
+    return SDValue(
+        DAG.getMachineNode(AArch64::MOVaddrPAC, DL, MVT::i64,
+                           {TPtr, Key, TAddrDiscriminator, Discriminator}),
+        0);
+  }
+
+  // GOT load -> LOADgotPAC
+  // Note that we disallow extern_weak refs to avoid null checks later.
+  if (!PtrGV->hasExternalWeakLinkage())
+    return SDValue(
+        DAG.getMachineNode(AArch64::LOADgotPAC, DL, MVT::i64,
+                           {TPtr, Key, TAddrDiscriminator, Discriminator}),
+        0);
+
+  // extern_weak ref -> LOADauthptrstatic
+  return LowerPtrAuthGlobalAddressStatically(
+      TPtr, DL, VT, (AArch64PACKey::ID)KeyC, Discriminator, AddrDiscriminator,
+      DAG);
+}
+
 // Looks through \param Val to determine the bit that can be used to
 // check the sign of the value. It returns the unextended value and
 // the sign bit position.
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.h b/llvm/lib/Target/AArch64/AArch64ISelLowering.h
index 5200b24..047c852bb 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.h
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.h
@@ -1127,6 +1127,12 @@ private:
   SDValue LowerELFTLSDescCallSeq(SDValue SymAddr, const SDLoc &DL,
                                  SelectionDAG &DAG) const;
   SDValue LowerWindowsGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerPtrAuthGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerPtrAuthGlobalAddressStatically(SDValue TGA, SDLoc DL, EVT VT,
+                                              AArch64PACKey::ID Key,
+                                              SDValue Discriminator,
+                                              SDValue AddrDiscriminator,
+                                              SelectionDAG &DAG) const;
   SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerSETCCCARRY(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
diff --git a/llvm/lib/Target/AArch64/AArch64InstrInfo.td b/llvm/lib/Target/AArch64/AArch64InstrInfo.td
index f3aac3b..1e06d5f 100644
--- a/llvm/lib/Target/AArch64/AArch64InstrInfo.td
+++ b/llvm/lib/Target/AArch64/AArch64InstrInfo.td
@@ -1765,6 +1765,38 @@ let Predicates = [HasPAuth] in {
   defm LDRAA  : AuthLoad<0, "ldraa", simm10Scaled>;
   defm LDRAB  : AuthLoad<1, "ldrab", simm10Scaled>;
 
+  // Materialize a signed global address, with adrp+add and PAC.
+  def MOVaddrPAC : Pseudo<(outs),
+                          (ins i64imm:$Addr, i32imm:$Key,
+                               GPR64noip:$AddrDisc, i64imm:$Disc), []>,
+               Sched<[WriteI, ReadI]> {
+    let isReMaterializable = 1;
+    let isCodeGenOnly = 1;
+    let Size = 40; // 12 fixed + 28 variable, for pointer offset, and discriminator
+    let Defs = [X16,X17];
+  }
+
+  // Materialize a signed global address, using a GOT load and PAC.
+  def LOADgotPAC : Pseudo<(outs),
+                          (ins i64imm:$Addr, i32imm:$Key,
+                               GPR64noip:$AddrDisc, i64imm:$Disc), []>,
+               Sched<[WriteI, ReadI]> {
+    let isReMaterializable = 1;
+    let isCodeGenOnly = 1;
+    let Size = 40; // 12 fixed + 28 variable, for pointer offset, and discriminator
+    let Defs = [X16,X17];
+  }
+
+  // Load a signed global address from a special $auth_ptr$ stub slot.
+  def LOADauthptrstatic : Pseudo<(outs GPR64:$dst),
+                              (ins i64imm:$Addr, i32imm:$Key,
+                                   i64imm:$Disc), []>,
+               Sched<[WriteI, ReadI]> {
+    let isReMaterializable = 1;
+    let isCodeGenOnly = 1;
+    let Size = 8;
+  }
+
   // Size 16: 4 fixed + 8 variable, to compute discriminator.
   let isCall = 1, isTerminator = 1, isReturn = 1, isBarrier = 1, Size = 16,
       Uses = [SP] in {
diff --git a/llvm/lib/Target/AArch64/AArch64TargetObjectFile.cpp b/llvm/lib/Target/AArch64/AArch64TargetObjectFile.cpp
index e5c6b6d..1c1e87d 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetObjectFile.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetObjectFile.cpp
@@ -8,7 +8,10 @@
 
 #include "AArch64TargetObjectFile.h"
 #include "AArch64TargetMachine.h"
+#include "MCTargetDesc/AArch64MCExpr.h"
+#include "llvm/ADT/StringExtras.h"
 #include "llvm/BinaryFormat/Dwarf.h"
+#include "llvm/CodeGen/MachineModuleInfoImpls.h"
 #include "llvm/IR/Mangler.h"
 #include "llvm/MC/MCContext.h"
 #include "llvm/MC/MCExpr.h"
@@ -88,3 +91,37 @@ void AArch64_MachoTargetObjectFile::getNameWithPrefix(
   // be accessed via at least a linker-private symbol.
   getMangler().getNameWithPrefix(OutName, GV, /* CannotUsePrivateLabel */ true);
 }
+
+template <typename MachineModuleInfoTarget>
+static MCSymbol *getAuthPtrSlotSymbolHelper(
+    MCContext &Ctx, const TargetMachine &TM, MachineModuleInfo *MMI,
+    MachineModuleInfoTarget &TargetMMI, const MCSymbol *RawSym,
+    AArch64PACKey::ID Key, uint16_t Discriminator) {
+  const DataLayout &DL = MMI->getModule()->getDataLayout();
+
+  MCSymbol *StubSym = Ctx.getOrCreateSymbol(
+      DL.getLinkerPrivateGlobalPrefix() + RawSym->getName() +
+      Twine("$auth_ptr$") + AArch64PACKeyIDToString(Key) + Twine('$') +
+      Twine(Discriminator));
+
+  typename MachineModuleInfoTarget::AuthStubInfo &StubInfo =
+      TargetMMI.getAuthPtrStubEntry(StubSym);
+
+  if (StubInfo.AuthPtrRef)
+    return StubSym;
+
+  const MCExpr *Sym = MCSymbolRefExpr::create(RawSym, Ctx);
+
+  StubInfo.AuthPtrRef =
+      AArch64AuthMCExpr::create(Sym, Discriminator, Key,
+                                /*HasAddressDiversity=*/false, Ctx);
+  return StubSym;
+}
+
+MCSymbol *AArch64_ELFTargetObjectFile::getAuthPtrSlotSymbol(
+    const TargetMachine &TM, MachineModuleInfo *MMI, const MCSymbol *RawSym,
+    AArch64PACKey::ID Key, uint16_t Discriminator) const {
+  auto &ELFMMI = MMI->getObjFileInfo<MachineModuleInfoELF>();
+  return getAuthPtrSlotSymbolHelper(getContext(), TM, MMI, ELFMMI, RawSym, Key,
+                                    Discriminator);
+}
diff --git a/llvm/lib/Target/AArch64/AArch64TargetObjectFile.h b/llvm/lib/Target/AArch64/AArch64TargetObjectFile.h
index 7cd56fc..c5ebf03 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetObjectFile.h
+++ b/llvm/lib/Target/AArch64/AArch64TargetObjectFile.h
@@ -9,6 +9,7 @@
 #ifndef LLVM_LIB_TARGET_AARCH64_AARCH64TARGETOBJECTFILE_H
 #define LLVM_LIB_TARGET_AARCH64_AARCH64TARGETOBJECTFILE_H
 
+#include "Utils/AArch64BaseInfo.h"
 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
 #include "llvm/Target/TargetLoweringObjectFile.h"
 
@@ -29,6 +30,11 @@ public:
                                           const MCValue &MV, int64_t Offset,
                                           MachineModuleInfo *MMI,
                                           MCStreamer &Streamer) const override;
+
+  MCSymbol *getAuthPtrSlotSymbol(const TargetMachine &TM,
+                                 MachineModuleInfo *MMI, const MCSymbol *RawSym,
+                                 AArch64PACKey::ID Key,
+                                 uint16_t Discriminator) const;
 };
 
 /// AArch64_MachoTargetObjectFile - This TLOF implementation is used for Darwin.
diff --git a/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp b/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp
index 0357a72..9e08609 100644
--- a/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp
@@ -224,6 +224,8 @@ private:
   bool selectJumpTable(MachineInstr &I, MachineRegisterInfo &MRI);
   bool selectBrJT(MachineInstr &I, MachineRegisterInfo &MRI);
   bool selectTLSGlobalValue(MachineInstr &I, MachineRegisterInfo &MRI);
+  bool selectPtrAuthGlobalValue(MachineInstr &I,
+                                MachineRegisterInfo &MRI) const;
   bool selectReduction(MachineInstr &I, MachineRegisterInfo &MRI);
   bool selectMOPS(MachineInstr &I, MachineRegisterInfo &MRI);
   bool selectUSMovFromExtend(MachineInstr &I, MachineRegisterInfo &MRI);
@@ -2848,6 +2850,9 @@ bool AArch64InstructionSelector::select(MachineInstr &I) {
     return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
   }
 
+  case TargetOpcode::G_PTRAUTH_GLOBAL_VALUE:
+    return selectPtrAuthGlobalValue(I, MRI);
+
   case TargetOpcode::G_ZEXTLOAD:
   case TargetOpcode::G_LOAD:
   case TargetOpcode::G_STORE: {
@@ -6583,6 +6588,145 @@ bool AArch64InstructionSelector::selectIntrinsic(MachineInstr &I,
   return false;
 }
 
+// G_PTRAUTH_GLOBAL_VALUE lowering
+//
+// We have 3 lowering alternatives to choose from:
+// - MOVaddrPAC: similar to MOVaddr, with added PAC.
+//   If the GV doesn't need a GOT load (i.e., is locally defined)
+//   materialize the pointer using adrp+add+pac. See LowerMOVaddrPAC.
+//
+// - LOADgotPAC: similar to LOADgot, with added PAC.
+//   If the GV needs a GOT load, materialize the pointer using the usual
+//   GOT adrp+ldr, +pac. Pointers in GOT are assumed to be not signed, the GOT
+//   section is assumed to be read-only (for example, via relro mechanism). See
+//   LowerMOVaddrPAC.
+//
+// - LOADauthptrstatic: similar to LOADgot, but use a
+//   special stub slot instead of a GOT slot.
+//   Load a signed pointer for symbol 'sym' from a stub slot named
+//   'sym$auth_ptr$key$disc' filled by dynamic linker during relocation
+//   resolving. This usually lowers to adrp+ldr, but also emits an entry into
+//   .data with an
+//   @AUTH relocation. See LowerLOADauthptrstatic.
+//
+// All 3 are pseudos that are expand late to longer sequences: this lets us
+// provide integrity guarantees on the to-be-signed intermediate values.
+//
+// LOADauthptrstatic is undesirable because it requires a large section filled
+// with often similarly-signed pointers, making it a good harvesting target.
+// Thus, it's only used for ptrauth references to extern_weak to avoid null
+// checks.
+
+bool AArch64InstructionSelector::selectPtrAuthGlobalValue(
+    MachineInstr &I, MachineRegisterInfo &MRI) const {
+  Register DefReg = I.getOperand(0).getReg();
+  Register Addr = I.getOperand(1).getReg();
+  uint64_t Key = I.getOperand(2).getImm();
+  Register AddrDisc = I.getOperand(3).getReg();
+  uint64_t Disc = I.getOperand(4).getImm();
+  int64_t Offset = 0;
+
+  if (Key > AArch64PACKey::LAST)
+    report_fatal_error("key in ptrauth global out of range [0, " +
+                       Twine((int)AArch64PACKey::LAST) + "]");
+
+  // Blend only works if the integer discriminator is 16-bit wide.
+  if (!isUInt<16>(Disc))
+    report_fatal_error(
+        "constant discriminator in ptrauth global out of range [0, 0xffff]");
+
+  // Choosing between 3 lowering alternatives is target-specific.
+  if (!STI.isTargetELF())
+    report_fatal_error("ptrauth global lowering is only implemented for ELF");
+
+  if (!MRI.hasOneDef(Addr))
+    return false;
+
+  // First match any offset we take from the real global.
+  const MachineInstr *DefMI = &*MRI.def_instr_begin(Addr);
+  if (DefMI->getOpcode() == TargetOpcode::G_PTR_ADD) {
+    Register OffsetReg = DefMI->getOperand(2).getReg();
+    if (!MRI.hasOneDef(OffsetReg))
+      return false;
+    const MachineInstr &OffsetMI = *MRI.def_instr_begin(OffsetReg);
+    if (OffsetMI.getOpcode() != TargetOpcode::G_CONSTANT)
+      return false;
+
+    Addr = DefMI->getOperand(1).getReg();
+    if (!MRI.hasOneDef(Addr))
+      return false;
+
+    DefMI = &*MRI.def_instr_begin(Addr);
+    Offset = OffsetMI.getOperand(1).getCImm()->getSExtValue();
+  }
+
+  // We should be left with a genuine unauthenticated GlobalValue.
+  const GlobalValue *GV;
+  if (DefMI->getOpcode() == TargetOpcode::G_GLOBAL_VALUE) {
+    GV = DefMI->getOperand(1).getGlobal();
+    Offset += DefMI->getOperand(1).getOffset();
+  } else if (DefMI->getOpcode() == AArch64::G_ADD_LOW) {
+    GV = DefMI->getOperand(2).getGlobal();
+    Offset += DefMI->getOperand(2).getOffset();
+  } else {
+    return false;
+  }
+
+  MachineIRBuilder MIB(I);
+
+  // Classify the reference to determine whether it needs a GOT load.
+  unsigned OpFlags = STI.ClassifyGlobalReference(GV, TM);
+  const bool NeedsGOTLoad = ((OpFlags & AArch64II::MO_GOT) != 0);
+  assert(((OpFlags & (~AArch64II::MO_GOT)) == 0) &&
+         "unsupported non-GOT op flags on ptrauth global reference");
+  assert((!GV->hasExternalWeakLinkage() || NeedsGOTLoad) &&
+         "unsupported non-GOT reference to weak ptrauth global");
+
+  std::optional<APInt> AddrDiscVal = getIConstantVRegVal(AddrDisc, MRI);
+  bool HasAddrDisc = !AddrDiscVal || *AddrDiscVal != 0;
+
+  // Non-extern_weak:
+  // - No GOT load needed -> MOVaddrPAC
+  // - GOT load for non-extern_weak -> LOADgotPAC
+  //   Note that we disallow extern_weak refs to avoid null checks later.
+  if (!GV->hasExternalWeakLinkage()) {
+    MIB.buildInstr(TargetOpcode::IMPLICIT_DEF, {AArch64::X16}, {});
+    MIB.buildInstr(TargetOpcode::IMPLICIT_DEF, {AArch64::X17}, {});
+    MIB.buildInstr(NeedsGOTLoad ? AArch64::LOADgotPAC : AArch64::MOVaddrPAC)
+        .addGlobalAddress(GV, Offset)
+        .addImm(Key)
+        .addReg(HasAddrDisc ? AddrDisc : AArch64::XZR)
+        .addImm(Disc)
+        .constrainAllUses(TII, TRI, RBI);
+    MIB.buildCopy(DefReg, Register(AArch64::X16));
+    RBI.constrainGenericRegister(DefReg, AArch64::GPR64RegClass, MRI);
+    I.eraseFromParent();
+    return true;
+  }
+
+  // extern_weak -> LOADauthptrstatic
+
+  // Offsets and extern_weak don't mix well: ptrauth aside, you'd get the
+  // offset alone as a pointer if the symbol wasn't available, which would
+  // probably break null checks in users. Ptrauth complicates things further:
+  // error out.
+  if (Offset != 0)
+    report_fatal_error(
+        "unsupported non-zero offset in weak ptrauth global reference");
+
+  if (HasAddrDisc)
+    report_fatal_error("unsupported weak addr-div ptrauth global");
+
+  MIB.buildInstr(AArch64::LOADauthptrstatic, {DefReg}, {})
+      .addGlobalAddress(GV, Offset)
+      .addImm(Key)
+      .addImm(Disc);
+  RBI.constrainGenericRegister(DefReg, AArch64::GPR64RegClass, MRI);
+
+  I.eraseFromParent();
+  return true;
+}
+
 void AArch64InstructionSelector::SelectTable(MachineInstr &I,
                                              MachineRegisterInfo &MRI,
                                              unsigned NumVec, unsigned Opc1,
diff --git a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
index fef0b72..50ba305 100644
--- a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
@@ -760,6 +760,9 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
   else
     getActionDefinitionsBuilder(G_GLOBAL_VALUE).legalFor({p0});
 
+  getActionDefinitionsBuilder(G_PTRAUTH_GLOBAL_VALUE)
+      .legalIf(all(typeIs(0, p0), typeIs(1, p0)));
+
   getActionDefinitionsBuilder(G_PTRTOINT)
       .legalFor({{s64, p0}, {v2s64, v2p0}})
       .widenScalarToNextPow2(0, 64)