[X86][CET] Shadow stack fix for setjmp/longjmp

[mike.dvoretsky]

Summary

This patch adds a shadow stack fix when compiling setjmp/longjmp with the shadow stack enabled. This allows setjmp/longjmp to work correctly with CET.

[craig.topper]

llvm/lib/Target/X86/X86ISelLowering.cpp

27560 ↗

(On Diff #144313)

Variable names should be capitalized.

27797 ↗

(On Diff #144313)

Can we use MOV64ri32 instead of MOV64ri? That would only use 32-bits for the immediate instead of 64.

Or better yet, can you use XOR?

27807 ↗

(On Diff #144313)

Can the be "TEST reg, reg" instead of CMP? Should be shorter encoding that putting 0 in the immediate. Even if not CMP64ri8/CMP32ri8 should be shorter than the ri32/ri versions.

27863 ↗

(On Diff #144313)

Single shifts should use SHL64r1/SHL32r1

[mike.dvoretsky]

Updated per comments.

[hjl.tools]

llvm/test/CodeGen/X86/shadow-stack.ll

47 ↗

(On Diff #144543)

Why movabsq $128? incssp takes last 8 bits. "movl $255" works.

76 ↗

(On Diff #144543)

Why movl $128? incssp takes last 8 bits. "movl $255" works.

[mike.dvoretsky]

Style fix.

llvm/test/CodeGen/X86/shadow-stack.ll

76 ↗

(On Diff #144543)

This part ensures that we increment the SSP beyond the value in the low 8 bits of ecx on the last incsspd use. This means that we need to further increment the SSP by ecx * 256 at the start of this block. We cannot express that 256 in 8 bits, so instead we increment by (2 * ecx) * 128. So using 255 doesn't fit the logic of this part. Same goes for the 64-bit version.

[mike.dvoretsky]

Test fix.

[mike.dvoretsky]

@craig.topper, please review the latest changes.

[craig.topper]

LGTM

llvm/trunk/lib/Target/X86/X86ISelLowering.h

MachineBasicBlock *EmitLoweredTLSCall(MachineInstr &MI,
                                           MachineBasicBlock *BB) const;
 
     MachineBasicBlock *EmitLoweredRetpoline(MachineInstr &MI,
                                             MachineBasicBlock *BB) const;
 
     MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
                                         MachineBasicBlock *MBB) const;
 
+    void emitSetJmpShadowStackFix(MachineInstr &MI,
+                                  MachineBasicBlock *MBB) const;
+
     MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
                                          MachineBasicBlock *MBB) const;
 
+    MachineBasicBlock *emitLongJmpShadowStackFix(MachineInstr &MI,
+                                                 MachineBasicBlock *MBB) const;
+
     MachineBasicBlock *emitFMA3Instr(MachineInstr &MI,
                                      MachineBasicBlock *MBB) const;
 
     MachineBasicBlock *EmitSjLjDispatchBlock(MachineInstr &MI,
                                              MachineBasicBlock *MBB) const;
 
     /// Emit nodes that will be selected as "test Op0,Op0", or something
     /// equivalent, for use with the given x86 condition code.

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

BuildMI(*BB, MI, DL, TII->get(TargetOpcode::COPY), AvailableReg)
       .addReg(CalleeVReg);
   MI.getOperand(0).ChangeToES(Symbol);
   MI.setDesc(TII->get(Opc));
   MachineInstrBuilder(*BB->getParent(), &MI)
       .addReg(AvailableReg, RegState::Implicit | RegState::Kill);
   return BB;
 }
 
+/// SetJmp implies future control flow change upon calling the corresponding
+/// LongJmp.
+/// Instead of using the 'return' instruction, the long jump fixes the stack and
+/// performs an indirect branch. To do so it uses the registers that were stored
+/// in the jump buffer (when calling SetJmp).
+/// In case the shadow stack is enabled we need to fix it as well, because some
+/// return addresses will be skipped.
+/// The function will save the SSP for future fixing in the function
+/// emitLongJmpShadowStackFix.
+/// \param [in] MI The temporary Machine Instruction for the builtin.
+/// \param [in] MBB The Machine Basic Block that will be modified.
+void X86TargetLowering::emitSetJmpShadowStackFix(MachineInstr &MI,
+                                                 MachineBasicBlock *MBB) const {
+  DebugLoc DL = MI.getDebugLoc();
+  MachineFunction *MF = MBB->getParent();
+  const TargetInstrInfo *TII = Subtarget.getInstrInfo();
+  MachineRegisterInfo &MRI = MF->getRegInfo();
+  MachineInstrBuilder MIB;
+
+  // Memory Reference
+  MachineInstr::mmo_iterator MMOBegin = MI.memoperands_begin();
+  MachineInstr::mmo_iterator MMOEnd = MI.memoperands_end();
+
+  // Initialize a register with zero.
+  MVT PVT = getPointerTy(MF->getDataLayout());
+  const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
+  unsigned ZReg = MRI.createVirtualRegister(PtrRC);
+  unsigned XorRROpc = (PVT == MVT::i64) ? X86::XOR64rr : X86::XOR32rr;
+  BuildMI(*MBB, MI, DL, TII->get(XorRROpc))
+      .addDef(ZReg)
+      .addReg(ZReg, RegState::Undef)
+      .addReg(ZReg, RegState::Undef);
+
+  // Read the current SSP Register value to the zeroed register.
+  unsigned SSPCopyReg = MRI.createVirtualRegister(PtrRC);
+  unsigned RdsspOpc = (PVT == MVT::i64) ? X86::RDSSPQ : X86::RDSSPD;
+  BuildMI(*MBB, MI, DL, TII->get(RdsspOpc), SSPCopyReg).addReg(ZReg);
+
+  // Write the SSP register value to offset 3 in input memory buffer.
+  unsigned PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
+  MIB = BuildMI(*MBB, MI, DL, TII->get(PtrStoreOpc));
+  const int64_t SSPOffset = 3 * PVT.getStoreSize();
+  const unsigned MemOpndSlot = 1;
+  for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
+    if (i == X86::AddrDisp)
+      MIB.addDisp(MI.getOperand(MemOpndSlot + i), SSPOffset);
+    else
+      MIB.add(MI.getOperand(MemOpndSlot + i));
+  }
+  MIB.addReg(SSPCopyReg);
+  MIB.setMemRefs(MMOBegin, MMOEnd);
+}
+
 MachineBasicBlock *
 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr &MI,
                                     MachineBasicBlock *MBB) const {
   DebugLoc DL = MI.getDebugLoc();
   MachineFunction *MF = MBB->getParent();
   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
   const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
   MachineRegisterInfo &MRI = MF->getRegInfo();
for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
     else
       MIB.add(MI.getOperand(MemOpndSlot + i));
   }
   if (!UseImmLabel)
     MIB.addReg(LabelReg);
   else
     MIB.addMBB(restoreMBB);
   MIB.setMemRefs(MMOBegin, MMOEnd);
+
+  if (Subtarget.hasSHSTK()) {
+    emitSetJmpShadowStackFix(MI, thisMBB);
+  }
+
   // Setup
   MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
           .addMBB(restoreMBB);
 
   const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo();
   MIB.addRegMask(RegInfo->getNoPreservedMask());
   thisMBB->addSuccessor(mainMBB);
   thisMBB->addSuccessor(restoreMBB);
X86TargetLowering::emitEHSjLjSetJmp(MachineInstr &MI,
   BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
   BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
   restoreMBB->addSuccessor(sinkMBB);
 
   MI.eraseFromParent();
   return sinkMBB;
 }
 
+/// Fix the shadow stack using the previously saved SSP pointer.
+/// \sa emitSetJmpShadowStackFix
+/// \param [in] MI The temporary Machine Instruction for the builtin.
+/// \param [in] MBB The Machine Basic Block that will be modified.
+/// \return The sink MBB that will perform the future indirect branch.
+MachineBasicBlock *
+X86TargetLowering::emitLongJmpShadowStackFix(MachineInstr &MI,
+                                             MachineBasicBlock *MBB) const {
+  DebugLoc DL = MI.getDebugLoc();
+  MachineFunction *MF = MBB->getParent();
+  const TargetInstrInfo *TII = Subtarget.getInstrInfo();
+  MachineRegisterInfo &MRI = MF->getRegInfo();
+
+  // Memory Reference
+  MachineInstr::mmo_iterator MMOBegin = MI.memoperands_begin();
+  MachineInstr::mmo_iterator MMOEnd = MI.memoperands_end();
+
+  MVT PVT = getPointerTy(MF->getDataLayout());
+  const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
+
+  // checkSspMBB:
+  //         xor vreg1, vreg1
+  //         rdssp vreg1
+  //         test vreg1, vreg1
+  //         je sinkMBB   # Jump if Shadow Stack is not supported
+  // fallMBB:
+  //         mov buf+24/12(%rip), vreg2
+  //         sub vreg1, vreg2
+  //         jbe sinkMBB  # No need to fix the Shadow Stack
+  // fixShadowMBB:
+  //         shr 3/2, vreg2
+  //         incssp vreg2  # fix the SSP according to the lower 8 bits
+  //         shr 8, vreg2
+  //         je sinkMBB
+  // fixShadowLoopPrepareMBB:
+  //         shl vreg2
+  //         mov 255, vreg3
+  // fixShadowLoopMBB:
+  //         incssp vreg3
+  //         dec vreg2
+  //         jne fixShadowLoopMBB # Iterate until you finish fixing
+  //                              # the Shadow Stack
+  // sinkMBB:
+
+  MachineFunction::iterator I = ++MBB->getIterator();
+  const BasicBlock *BB = MBB->getBasicBlock();
+
+  MachineBasicBlock *checkSspMBB = MF->CreateMachineBasicBlock(BB);
+  MachineBasicBlock *fallMBB = MF->CreateMachineBasicBlock(BB);
+  MachineBasicBlock *fixShadowMBB = MF->CreateMachineBasicBlock(BB);
+  MachineBasicBlock *fixShadowLoopPrepareMBB = MF->CreateMachineBasicBlock(BB);
+  MachineBasicBlock *fixShadowLoopMBB = MF->CreateMachineBasicBlock(BB);
+  MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
+  MF->insert(I, checkSspMBB);
+  MF->insert(I, fallMBB);
+  MF->insert(I, fixShadowMBB);
+  MF->insert(I, fixShadowLoopPrepareMBB);
+  MF->insert(I, fixShadowLoopMBB);
+  MF->insert(I, sinkMBB);
+
+  MBB->addSuccessor(checkSspMBB);
+
+  // Initialize a register with zero.
+  unsigned ZReg = MRI.createVirtualRegister(PtrRC);
+  unsigned XorRROpc = (PVT == MVT::i64) ? X86::XOR64rr : X86::XOR32rr;
+  BuildMI(checkSspMBB, DL, TII->get(XorRROpc))
+      .addDef(ZReg)
+      .addReg(ZReg, RegState::Undef)
+      .addReg(ZReg, RegState::Undef);
+
+  // Read the current SSP Register value to the zeroed register.
+  unsigned SSPCopyReg = MRI.createVirtualRegister(PtrRC);
+  unsigned RdsspOpc = (PVT == MVT::i64) ? X86::RDSSPQ : X86::RDSSPD;
+  BuildMI(checkSspMBB, DL, TII->get(RdsspOpc), SSPCopyReg).addReg(ZReg);
+
+  // Check whether the result of the SSP register is zero and jump directly
+  // to the sink.
+  unsigned TestRROpc = (PVT == MVT::i64) ? X86::TEST64rr : X86::TEST32rr;
+  BuildMI(checkSspMBB, DL, TII->get(TestRROpc))
+      .addReg(SSPCopyReg)
+      .addReg(SSPCopyReg);
+  BuildMI(checkSspMBB, DL, TII->get(X86::JE_1)).addMBB(sinkMBB);
+  checkSspMBB->addSuccessor(sinkMBB);
+  checkSspMBB->addSuccessor(fallMBB);
+
+  // Reload the previously saved SSP register value.
+  unsigned PrevSSPReg = MRI.createVirtualRegister(PtrRC);
+  unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
+  const int64_t SPPOffset = 3 * PVT.getStoreSize();
+  MachineInstrBuilder MIB =
+      BuildMI(fallMBB, DL, TII->get(PtrLoadOpc), PrevSSPReg);
+  for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
+    if (i == X86::AddrDisp)
+      MIB.addDisp(MI.getOperand(i), SPPOffset);
+    else
+      MIB.add(MI.getOperand(i));
+  }
+  MIB.setMemRefs(MMOBegin, MMOEnd);
+
+  // Subtract the current SSP from the previous SSP.
+  unsigned SspSubReg = MRI.createVirtualRegister(PtrRC);
+  unsigned SubRROpc = (PVT == MVT::i64) ? X86::SUB64rr : X86::SUB32rr;
+  BuildMI(fallMBB, DL, TII->get(SubRROpc), SspSubReg)
+      .addReg(PrevSSPReg)
+      .addReg(SSPCopyReg);
+
+  // Jump to sink in case PrevSSPReg <= SSPCopyReg.
+  BuildMI(fallMBB, DL, TII->get(X86::JBE_1)).addMBB(sinkMBB);
+  fallMBB->addSuccessor(sinkMBB);
+  fallMBB->addSuccessor(fixShadowMBB);
+
+  // Shift right by 2/3 for 32/64 because incssp multiplies the argument by 4/8.
+  unsigned ShrRIOpc = (PVT == MVT::i64) ? X86::SHR64ri : X86::SHR32ri;
+  unsigned Offset = (PVT == MVT::i64) ? 3 : 2;
+  unsigned SspFirstShrReg = MRI.createVirtualRegister(PtrRC);
+  BuildMI(fixShadowMBB, DL, TII->get(ShrRIOpc), SspFirstShrReg)
+      .addReg(SspSubReg)
+      .addImm(Offset);
+
+  // Increase SSP when looking only on the lower 8 bits of the delta.
+  unsigned IncsspOpc = (PVT == MVT::i64) ? X86::INCSSPQ : X86::INCSSPD;
+  BuildMI(fixShadowMBB, DL, TII->get(IncsspOpc)).addReg(SspFirstShrReg);
+
+  // Reset the lower 8 bits.
+  unsigned SspSecondShrReg = MRI.createVirtualRegister(PtrRC);
+  BuildMI(fixShadowMBB, DL, TII->get(ShrRIOpc), SspSecondShrReg)
+      .addReg(SspFirstShrReg)
+      .addImm(8);
+
+  // Jump if the result of the shift is zero.
+  BuildMI(fixShadowMBB, DL, TII->get(X86::JE_1)).addMBB(sinkMBB);
+  fixShadowMBB->addSuccessor(sinkMBB);
+  fixShadowMBB->addSuccessor(fixShadowLoopPrepareMBB);
+
+  // Do a single shift left.
+  unsigned ShlR1Opc = (PVT == MVT::i64) ? X86::SHL64r1 : X86::SHL32r1;
+  unsigned SspAfterShlReg = MRI.createVirtualRegister(PtrRC);
+  BuildMI(fixShadowLoopPrepareMBB, DL, TII->get(ShlR1Opc), SspAfterShlReg)
+      .addReg(SspSecondShrReg);
+
+  // Save the value 128 to a register (will be used next with incssp).
+  unsigned Value128InReg = MRI.createVirtualRegister(PtrRC);
+  unsigned MovRIOpc = (PVT == MVT::i64) ? X86::MOV64ri32 : X86::MOV32ri;
+  BuildMI(fixShadowLoopPrepareMBB, DL, TII->get(MovRIOpc), Value128InReg)
+      .addImm(128);
+  fixShadowLoopPrepareMBB->addSuccessor(fixShadowLoopMBB);
+
+  // Since incssp only looks at the lower 8 bits, we might need to do several
+  // iterations of incssp until we finish fixing the shadow stack.
+  unsigned DecReg = MRI.createVirtualRegister(PtrRC);
+  unsigned CounterReg = MRI.createVirtualRegister(PtrRC);
+  BuildMI(fixShadowLoopMBB, DL, TII->get(X86::PHI), CounterReg)
+      .addReg(SspAfterShlReg)
+      .addMBB(fixShadowLoopPrepareMBB)
+      .addReg(DecReg)
+      .addMBB(fixShadowLoopMBB);
+
+  // Every iteration we increase the SSP by 128.
+  BuildMI(fixShadowLoopMBB, DL, TII->get(IncsspOpc)).addReg(Value128InReg);
+
+  // Every iteration we decrement the counter by 1.
+  unsigned DecROpc = (PVT == MVT::i64) ? X86::DEC64r : X86::DEC32r;
+  BuildMI(fixShadowLoopMBB, DL, TII->get(DecROpc), DecReg).addReg(CounterReg);
+
+  // Jump if the counter is not zero yet.
+  BuildMI(fixShadowLoopMBB, DL, TII->get(X86::JNE_1)).addMBB(fixShadowLoopMBB);
+  fixShadowLoopMBB->addSuccessor(sinkMBB);
+  fixShadowLoopMBB->addSuccessor(fixShadowLoopMBB);
+
+  return sinkMBB;
+}
+
 MachineBasicBlock *
 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr &MI,
                                      MachineBasicBlock *MBB) const {
   DebugLoc DL = MI.getDebugLoc();
   MachineFunction *MF = MBB->getParent();
   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
   MachineRegisterInfo &MRI = MF->getRegInfo();
 
X86TargetLowering::emitEHSjLjLongJmp(MachineInstr &MI,
   MachineInstrBuilder MIB;
 
   const int64_t LabelOffset = 1 * PVT.getStoreSize();
   const int64_t SPOffset = 2 * PVT.getStoreSize();
 
   unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
   unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
 
+  MachineBasicBlock *thisMBB = MBB;
+
+  // When CET and shadow stack is enabled, we need to fix the Shadow Stack.
+  if (Subtarget.hasSHSTK()) {
+    thisMBB = emitLongJmpShadowStackFix(MI, thisMBB);
+  }
+
   // Reload FP
-  MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
+  MIB = BuildMI(thisMBB, DL, TII->get(PtrLoadOpc), FP);
   for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
     MIB.add(MI.getOperand(i));
   MIB.setMemRefs(MMOBegin, MMOEnd);
+
   // Reload IP
-  MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
+  MIB = BuildMI(thisMBB, DL, TII->get(PtrLoadOpc), Tmp);
   for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
     if (i == X86::AddrDisp)
       MIB.addDisp(MI.getOperand(i), LabelOffset);
     else
       MIB.add(MI.getOperand(i));
   }
   MIB.setMemRefs(MMOBegin, MMOEnd);
+
   // Reload SP
-  MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
+  MIB = BuildMI(thisMBB, DL, TII->get(PtrLoadOpc), SP);
   for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
     if (i == X86::AddrDisp)
       MIB.addDisp(MI.getOperand(i), SPOffset);
     else
       MIB.add(MI.getOperand(i));
   }
   MIB.setMemRefs(MMOBegin, MMOEnd);
+
   // Jump
-  BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
+  BuildMI(thisMBB, DL, TII->get(IJmpOpc)).addReg(Tmp);
 
   MI.eraseFromParent();
-  return MBB;
+  return thisMBB;
 }
 
 void X86TargetLowering::SetupEntryBlockForSjLj(MachineInstr &MI,
                                                MachineBasicBlock *MBB,
                                                MachineBasicBlock *DispatchBB,
                                                int FI) const {
   DebugLoc DL = MI.getDebugLoc();
   MachineFunction *MF = MBB->getParent();

llvm/trunk/test/CodeGen/X86/shadow-stack.ll

+; RUN: llc -mtriple x86_64-unknown-unknown -mattr=+shstk < %s | FileCheck %s --check-prefix=X86_64
+; RUN: llc -mtriple i386-unknown-unknown -mattr=+shstk < %s | FileCheck %s --check-prefix=X86
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; The IR was created using the following C code:
+;; typedef void *jmp_buf;
+;; jmp_buf *buf;
+;;
+;; __attribute__((noinline)) int bar (int i) {
+;;   int j = i - 111;
+;;   __builtin_longjmp (buf, 1);
+;;   return j;
+;; }
+;;
+;; int foo (int i) {
+;;   int j = i * 11;
+;;   if (!__builtin_setjmp (buf)) {
+;;     j += 33 + bar (j);
+;;   }
+;;   return j + i;
+;; }
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+@buf = common local_unnamed_addr global i8* null, align 8
+
+; Functions that use LongJmp should fix the Shadow Stack using previosuly saved
+; ShadowStackPointer in the input buffer.
+; The fix requires unwinding the shadow stack to the last SSP.
+define i32 @bar(i32 %i) local_unnamed_addr {
+; X86_64-LABEL: bar:
+; X86_64:         movq {{.*}}(%rip), %rax
+; X86_64-NEXT:    xorq %rdx, %rdx
+; X86_64-NEXT:    rdsspq %rdx
+; X86_64-NEXT:    testq %rdx, %rdx
+; X86_64-NEXT:    je .LBB0_5
+; X86_64-NEXT:  # %bb.1: # %entry
+; X86_64-NEXT:    movq 24(%rax), %rcx
+; X86_64-NEXT:    subq %rdx, %rcx
+; X86_64-NEXT:    jbe .LBB0_5
+; X86_64-NEXT:  # %bb.2: # %entry
+; X86_64-NEXT:    shrq $3, %rcx
+; X86_64-NEXT:    incsspq %rcx
+; X86_64-NEXT:    shrq $8, %rcx
+; X86_64-NEXT:    je .LBB0_5
+; X86_64-NEXT:  # %bb.3: # %entry
+; X86_64-NEXT:    shlq %rcx
+; X86_64-NEXT:    movq $128, %rdx
+; X86_64-NEXT:  .LBB0_4: # %entry
+; X86_64-NEXT:    # =>This Inner Loop Header: Depth=1
+; X86_64-NEXT:    incsspq %rdx
+; X86_64-NEXT:    decq %rcx
+; X86_64-NEXT:    jne .LBB0_4
+; X86_64-NEXT:  .LBB0_5: # %entry
+; X86_64-NEXT:    movq (%rax), %rbp
+; X86_64-NEXT:    movq 8(%rax), %rcx
+; X86_64-NEXT:    movq 16(%rax), %rsp
+; X86_64-NEXT:    jmpq *%rcx
+;
+; X86-LABEL: bar:
+; X86:         movl buf, %eax
+; X86-NEXT:    xorl %edx, %edx
+; X86-NEXT:    rdsspd %edx
+; X86-NEXT:    testl %edx, %edx
+; X86-NEXT:    je .LBB0_5
+; X86-NEXT:  # %bb.1: # %entry
+; X86-NEXT:    movl 12(%eax), %ecx
+; X86-NEXT:    subl %edx, %ecx
+; X86-NEXT:    jbe .LBB0_5
+; X86-NEXT:  # %bb.2: # %entry
+; X86-NEXT:    shrl $2, %ecx
+; X86-NEXT:    incsspd %ecx
+; X86-NEXT:    shrl $8, %ecx
+; X86-NEXT:    je .LBB0_5
+; X86-NEXT:  # %bb.3: # %entry
+; X86-NEXT:    shll %ecx
+; X86-NEXT:    movl $128, %edx
+; X86-NEXT:  .LBB0_4: # %entry
+; X86-NEXT:    # =>This Inner Loop Header: Depth=1
+; X86-NEXT:    incsspd %edx
+; X86-NEXT:    decl %ecx
+; X86-NEXT:    jne .LBB0_4
+; X86-NEXT:  .LBB0_5: # %entry
+; X86-NEXT:    movl (%eax), %ebp
+; X86-NEXT:    movl 4(%eax), %ecx
+; X86-NEXT:    movl 8(%eax), %esp
+; X86-NEXT:    jmpl *%ecx
+entry:
+  %0 = load i8*, i8** @buf, align 8
+  tail call void @llvm.eh.sjlj.longjmp(i8* %0)
+  unreachable
+}
+
+declare void @llvm.eh.sjlj.longjmp(i8*)
+
+; Functions that call SetJmp should save the current ShadowStackPointer for
+; future fixing of the Shadow Stack.
+define i32 @foo(i32 %i) local_unnamed_addr {
+; X86_64-LABEL: foo:
+; X86_64:         xorq %rcx, %rcx
+; X86_64-NEXT:    rdsspq %rcx
+; X86_64-NEXT:    movq %rcx, 24(%rax)
+; X86_64:         callq bar
+;
+; X86-LABEL: foo:
+; X86:         xorl %ecx, %ecx
+; X86-NEXT:    rdsspd %ecx
+; X86-NEXT:    movl %ecx, 12(%eax)
+; X86:         calll bar
+entry:
+  %0 = load i8*, i8** @buf, align 8
+  %1 = bitcast i8* %0 to i8**
+  %2 = tail call i8* @llvm.frameaddress(i32 0)
+  store i8* %2, i8** %1, align 8
+  %3 = tail call i8* @llvm.stacksave()
+  %4 = getelementptr inbounds i8, i8* %0, i64 16
+  %5 = bitcast i8* %4 to i8**
+  store i8* %3, i8** %5, align 8
+  %6 = tail call i32 @llvm.eh.sjlj.setjmp(i8* %0)
+  %tobool = icmp eq i32 %6, 0
+  br i1 %tobool, label %if.then, label %if.end
+
+if.then:                                          ; preds = %entry
+  %call = tail call i32 @bar(i32 undef)
+  unreachable
+
+if.end:                                           ; preds = %entry
+  %add2 = mul nsw i32 %i, 12
+  ret i32 %add2
+}
+
+declare i8* @llvm.frameaddress(i32)
+declare i8* @llvm.stacksave()
+declare i32 @llvm.eh.sjlj.setjmp(i8*)