xref: /llvm-project/llvm/lib/Target/X86/X86SpeculativeLoadHardening.cpp (revision 19032bfe87fa0f4a3a7b3e68daafc93331b71e0d)
1 //====- X86SpeculativeLoadHardening.cpp - A Spectre v1 mitigation ---------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 /// \file
9 ///
10 /// Provide a pass which mitigates speculative execution attacks which operate
11 /// by speculating incorrectly past some predicate (a type check, bounds check,
12 /// or other condition) to reach a load with invalid inputs and leak the data
13 /// accessed by that load using a side channel out of the speculative domain.
14 ///
15 /// For details on the attacks, see the first variant in both the Project Zero
16 /// writeup and the Spectre paper:
17 /// https://googleprojectzero.blogspot.com/2018/01/reading-privileged-memory-with-side.html
18 /// https://spectreattack.com/spectre.pdf
19 ///
20 //===----------------------------------------------------------------------===//
21 
22 #include "X86.h"
23 #include "X86InstrInfo.h"
24 #include "X86Subtarget.h"
25 #include "llvm/ADT/ArrayRef.h"
26 #include "llvm/ADT/DenseMap.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/ADT/SmallPtrSet.h"
29 #include "llvm/ADT/SmallSet.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/SparseBitVector.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/CodeGen/MachineBasicBlock.h"
34 #include "llvm/CodeGen/MachineConstantPool.h"
35 #include "llvm/CodeGen/MachineFunction.h"
36 #include "llvm/CodeGen/MachineFunctionPass.h"
37 #include "llvm/CodeGen/MachineInstr.h"
38 #include "llvm/CodeGen/MachineInstrBuilder.h"
39 #include "llvm/CodeGen/MachineModuleInfo.h"
40 #include "llvm/CodeGen/MachineOperand.h"
41 #include "llvm/CodeGen/MachineRegisterInfo.h"
42 #include "llvm/CodeGen/MachineSSAUpdater.h"
43 #include "llvm/CodeGen/TargetInstrInfo.h"
44 #include "llvm/CodeGen/TargetRegisterInfo.h"
45 #include "llvm/CodeGen/TargetSchedule.h"
46 #include "llvm/CodeGen/TargetSubtargetInfo.h"
47 #include "llvm/IR/DebugLoc.h"
48 #include "llvm/MC/MCSchedule.h"
49 #include "llvm/Pass.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/raw_ostream.h"
53 #include "llvm/Target/TargetMachine.h"
54 #include <cassert>
55 #include <iterator>
56 #include <optional>
57 #include <utility>
58 
59 using namespace llvm;
60 
61 #define PASS_KEY "x86-slh"
62 #define DEBUG_TYPE PASS_KEY
63 
64 STATISTIC(NumCondBranchesTraced, "Number of conditional branches traced");
65 STATISTIC(NumBranchesUntraced, "Number of branches unable to trace");
66 STATISTIC(NumAddrRegsHardened,
67           "Number of address mode used registers hardaned");
68 STATISTIC(NumPostLoadRegsHardened,
69           "Number of post-load register values hardened");
70 STATISTIC(NumCallsOrJumpsHardened,
71           "Number of calls or jumps requiring extra hardening");
72 STATISTIC(NumInstsInserted, "Number of instructions inserted");
73 STATISTIC(NumLFENCEsInserted, "Number of lfence instructions inserted");
74 
75 static cl::opt<bool> EnableSpeculativeLoadHardening(
76     "x86-speculative-load-hardening",
77     cl::desc("Force enable speculative load hardening"), cl::init(false),
78     cl::Hidden);
79 
80 static cl::opt<bool> HardenEdgesWithLFENCE(
81     PASS_KEY "-lfence",
82     cl::desc(
83         "Use LFENCE along each conditional edge to harden against speculative "
84         "loads rather than conditional movs and poisoned pointers."),
85     cl::init(false), cl::Hidden);
86 
87 static cl::opt<bool> EnablePostLoadHardening(
88     PASS_KEY "-post-load",
89     cl::desc("Harden the value loaded *after* it is loaded by "
90              "flushing the loaded bits to 1. This is hard to do "
91              "in general but can be done easily for GPRs."),
92     cl::init(true), cl::Hidden);
93 
94 static cl::opt<bool> FenceCallAndRet(
95     PASS_KEY "-fence-call-and-ret",
96     cl::desc("Use a full speculation fence to harden both call and ret edges "
97              "rather than a lighter weight mitigation."),
98     cl::init(false), cl::Hidden);
99 
100 static cl::opt<bool> HardenInterprocedurally(
101     PASS_KEY "-ip",
102     cl::desc("Harden interprocedurally by passing our state in and out of "
103              "functions in the high bits of the stack pointer."),
104     cl::init(true), cl::Hidden);
105 
106 static cl::opt<bool>
107     HardenLoads(PASS_KEY "-loads",
108                 cl::desc("Sanitize loads from memory. When disable, no "
109                          "significant security is provided."),
110                 cl::init(true), cl::Hidden);
111 
112 static cl::opt<bool> HardenIndirectCallsAndJumps(
113     PASS_KEY "-indirect",
114     cl::desc("Harden indirect calls and jumps against using speculatively "
115              "stored attacker controlled addresses. This is designed to "
116              "mitigate Spectre v1.2 style attacks."),
117     cl::init(true), cl::Hidden);
118 
119 namespace {
120 
121 class X86SpeculativeLoadHardeningPass : public MachineFunctionPass {
122 public:
123   X86SpeculativeLoadHardeningPass() : MachineFunctionPass(ID) { }
124 
125   StringRef getPassName() const override {
126     return "X86 speculative load hardening";
127   }
128   bool runOnMachineFunction(MachineFunction &MF) override;
129   void getAnalysisUsage(AnalysisUsage &AU) const override;
130 
131   /// Pass identification, replacement for typeid.
132   static char ID;
133 
134 private:
135   /// The information about a block's conditional terminators needed to trace
136   /// our predicate state through the exiting edges.
137   struct BlockCondInfo {
138     MachineBasicBlock *MBB;
139 
140     // We mostly have one conditional branch, and in extremely rare cases have
141     // two. Three and more are so rare as to be unimportant for compile time.
142     SmallVector<MachineInstr *, 2> CondBrs;
143 
144     MachineInstr *UncondBr;
145   };
146 
147   /// Manages the predicate state traced through the program.
148   struct PredState {
149     unsigned InitialReg = 0;
150     unsigned PoisonReg = 0;
151 
152     const TargetRegisterClass *RC;
153     MachineSSAUpdater SSA;
154 
155     PredState(MachineFunction &MF, const TargetRegisterClass *RC)
156         : RC(RC), SSA(MF) {}
157   };
158 
159   const X86Subtarget *Subtarget = nullptr;
160   MachineRegisterInfo *MRI = nullptr;
161   const X86InstrInfo *TII = nullptr;
162   const TargetRegisterInfo *TRI = nullptr;
163 
164   std::optional<PredState> PS;
165 
166   void hardenEdgesWithLFENCE(MachineFunction &MF);
167 
168   SmallVector<BlockCondInfo, 16> collectBlockCondInfo(MachineFunction &MF);
169 
170   SmallVector<MachineInstr *, 16>
171   tracePredStateThroughCFG(MachineFunction &MF, ArrayRef<BlockCondInfo> Infos);
172 
173   void unfoldCallAndJumpLoads(MachineFunction &MF);
174 
175   SmallVector<MachineInstr *, 16>
176   tracePredStateThroughIndirectBranches(MachineFunction &MF);
177 
178   void tracePredStateThroughBlocksAndHarden(MachineFunction &MF);
179 
180   unsigned saveEFLAGS(MachineBasicBlock &MBB,
181                       MachineBasicBlock::iterator InsertPt,
182                       const DebugLoc &Loc);
183   void restoreEFLAGS(MachineBasicBlock &MBB,
184                      MachineBasicBlock::iterator InsertPt, const DebugLoc &Loc,
185                      Register Reg);
186 
187   void mergePredStateIntoSP(MachineBasicBlock &MBB,
188                             MachineBasicBlock::iterator InsertPt,
189                             const DebugLoc &Loc, unsigned PredStateReg);
190   unsigned extractPredStateFromSP(MachineBasicBlock &MBB,
191                                   MachineBasicBlock::iterator InsertPt,
192                                   const DebugLoc &Loc);
193 
194   void
195   hardenLoadAddr(MachineInstr &MI, MachineOperand &BaseMO,
196                  MachineOperand &IndexMO,
197                  SmallDenseMap<unsigned, unsigned, 32> &AddrRegToHardenedReg);
198   MachineInstr *
199   sinkPostLoadHardenedInst(MachineInstr &MI,
200                            SmallPtrSetImpl<MachineInstr *> &HardenedInstrs);
201   bool canHardenRegister(Register Reg);
202   unsigned hardenValueInRegister(Register Reg, MachineBasicBlock &MBB,
203                                  MachineBasicBlock::iterator InsertPt,
204                                  const DebugLoc &Loc);
205   unsigned hardenPostLoad(MachineInstr &MI);
206   void hardenReturnInstr(MachineInstr &MI);
207   void tracePredStateThroughCall(MachineInstr &MI);
208   void hardenIndirectCallOrJumpInstr(
209       MachineInstr &MI,
210       SmallDenseMap<unsigned, unsigned, 32> &AddrRegToHardenedReg);
211 };
212 
213 } // end anonymous namespace
214 
215 char X86SpeculativeLoadHardeningPass::ID = 0;
216 
217 void X86SpeculativeLoadHardeningPass::getAnalysisUsage(
218     AnalysisUsage &AU) const {
219   MachineFunctionPass::getAnalysisUsage(AU);
220 }
221 
222 static MachineBasicBlock &splitEdge(MachineBasicBlock &MBB,
223                                     MachineBasicBlock &Succ, int SuccCount,
224                                     MachineInstr *Br, MachineInstr *&UncondBr,
225                                     const X86InstrInfo &TII) {
226   assert(!Succ.isEHPad() && "Shouldn't get edges to EH pads!");
227 
228   MachineFunction &MF = *MBB.getParent();
229 
230   MachineBasicBlock &NewMBB = *MF.CreateMachineBasicBlock();
231 
232   // We have to insert the new block immediately after the current one as we
233   // don't know what layout-successor relationships the successor has and we
234   // may not be able to (and generally don't want to) try to fix those up.
235   MF.insert(std::next(MachineFunction::iterator(&MBB)), &NewMBB);
236 
237   // Update the branch instruction if necessary.
238   if (Br) {
239     assert(Br->getOperand(0).getMBB() == &Succ &&
240            "Didn't start with the right target!");
241     Br->getOperand(0).setMBB(&NewMBB);
242 
243     // If this successor was reached through a branch rather than fallthrough,
244     // we might have *broken* fallthrough and so need to inject a new
245     // unconditional branch.
246     if (!UncondBr) {
247       MachineBasicBlock &OldLayoutSucc =
248           *std::next(MachineFunction::iterator(&NewMBB));
249       assert(MBB.isSuccessor(&OldLayoutSucc) &&
250              "Without an unconditional branch, the old layout successor should "
251              "be an actual successor!");
252       auto BrBuilder =
253           BuildMI(&MBB, DebugLoc(), TII.get(X86::JMP_1)).addMBB(&OldLayoutSucc);
254       // Update the unconditional branch now that we've added one.
255       UncondBr = &*BrBuilder;
256     }
257 
258     // Insert unconditional "jump Succ" instruction in the new block if
259     // necessary.
260     if (!NewMBB.isLayoutSuccessor(&Succ)) {
261       SmallVector<MachineOperand, 4> Cond;
262       TII.insertBranch(NewMBB, &Succ, nullptr, Cond, Br->getDebugLoc());
263     }
264   } else {
265     assert(!UncondBr &&
266            "Cannot have a branchless successor and an unconditional branch!");
267     assert(NewMBB.isLayoutSuccessor(&Succ) &&
268            "A non-branch successor must have been a layout successor before "
269            "and now is a layout successor of the new block.");
270   }
271 
272   // If this is the only edge to the successor, we can just replace it in the
273   // CFG. Otherwise we need to add a new entry in the CFG for the new
274   // successor.
275   if (SuccCount == 1) {
276     MBB.replaceSuccessor(&Succ, &NewMBB);
277   } else {
278     MBB.splitSuccessor(&Succ, &NewMBB);
279   }
280 
281   // Hook up the edge from the new basic block to the old successor in the CFG.
282   NewMBB.addSuccessor(&Succ);
283 
284   // Fix PHI nodes in Succ so they refer to NewMBB instead of MBB.
285   for (MachineInstr &MI : Succ) {
286     if (!MI.isPHI())
287       break;
288     for (int OpIdx = 1, NumOps = MI.getNumOperands(); OpIdx < NumOps;
289          OpIdx += 2) {
290       MachineOperand &OpV = MI.getOperand(OpIdx);
291       MachineOperand &OpMBB = MI.getOperand(OpIdx + 1);
292       assert(OpMBB.isMBB() && "Block operand to a PHI is not a block!");
293       if (OpMBB.getMBB() != &MBB)
294         continue;
295 
296       // If this is the last edge to the succesor, just replace MBB in the PHI
297       if (SuccCount == 1) {
298         OpMBB.setMBB(&NewMBB);
299         break;
300       }
301 
302       // Otherwise, append a new pair of operands for the new incoming edge.
303       MI.addOperand(MF, OpV);
304       MI.addOperand(MF, MachineOperand::CreateMBB(&NewMBB));
305       break;
306     }
307   }
308 
309   // Inherit live-ins from the successor
310   for (auto &LI : Succ.liveins())
311     NewMBB.addLiveIn(LI);
312 
313   LLVM_DEBUG(dbgs() << "  Split edge from '" << MBB.getName() << "' to '"
314                     << Succ.getName() << "'.\n");
315   return NewMBB;
316 }
317 
318 /// Removing duplicate PHI operands to leave the PHI in a canonical and
319 /// predictable form.
320 ///
321 /// FIXME: It's really frustrating that we have to do this, but SSA-form in MIR
322 /// isn't what you might expect. We may have multiple entries in PHI nodes for
323 /// a single predecessor. This makes CFG-updating extremely complex, so here we
324 /// simplify all PHI nodes to a model even simpler than the IR's model: exactly
325 /// one entry per predecessor, regardless of how many edges there are.
326 static void canonicalizePHIOperands(MachineFunction &MF) {
327   SmallPtrSet<MachineBasicBlock *, 4> Preds;
328   SmallVector<int, 4> DupIndices;
329   for (auto &MBB : MF)
330     for (auto &MI : MBB) {
331       if (!MI.isPHI())
332         break;
333 
334       // First we scan the operands of the PHI looking for duplicate entries
335       // a particular predecessor. We retain the operand index of each duplicate
336       // entry found.
337       for (int OpIdx = 1, NumOps = MI.getNumOperands(); OpIdx < NumOps;
338            OpIdx += 2)
339         if (!Preds.insert(MI.getOperand(OpIdx + 1).getMBB()).second)
340           DupIndices.push_back(OpIdx);
341 
342       // Now walk the duplicate indices, removing both the block and value. Note
343       // that these are stored as a vector making this element-wise removal
344       // :w
345       // potentially quadratic.
346       //
347       // FIXME: It is really frustrating that we have to use a quadratic
348       // removal algorithm here. There should be a better way, but the use-def
349       // updates required make that impossible using the public API.
350       //
351       // Note that we have to process these backwards so that we don't
352       // invalidate other indices with each removal.
353       while (!DupIndices.empty()) {
354         int OpIdx = DupIndices.pop_back_val();
355         // Remove both the block and value operand, again in reverse order to
356         // preserve indices.
357         MI.removeOperand(OpIdx + 1);
358         MI.removeOperand(OpIdx);
359       }
360 
361       Preds.clear();
362     }
363 }
364 
365 /// Helper to scan a function for loads vulnerable to misspeculation that we
366 /// want to harden.
367 ///
368 /// We use this to avoid making changes to functions where there is nothing we
369 /// need to do to harden against misspeculation.
370 static bool hasVulnerableLoad(MachineFunction &MF) {
371   for (MachineBasicBlock &MBB : MF) {
372     for (MachineInstr &MI : MBB) {
373       // Loads within this basic block after an LFENCE are not at risk of
374       // speculatively executing with invalid predicates from prior control
375       // flow. So break out of this block but continue scanning the function.
376       if (MI.getOpcode() == X86::LFENCE)
377         break;
378 
379       // Looking for loads only.
380       if (!MI.mayLoad())
381         continue;
382 
383       // An MFENCE is modeled as a load but isn't vulnerable to misspeculation.
384       if (MI.getOpcode() == X86::MFENCE)
385         continue;
386 
387       // We found a load.
388       return true;
389     }
390   }
391 
392   // No loads found.
393   return false;
394 }
395 
396 bool X86SpeculativeLoadHardeningPass::runOnMachineFunction(
397     MachineFunction &MF) {
398   LLVM_DEBUG(dbgs() << "********** " << getPassName() << " : " << MF.getName()
399                     << " **********\n");
400 
401   // Only run if this pass is forced enabled or we detect the relevant function
402   // attribute requesting SLH.
403   if (!EnableSpeculativeLoadHardening &&
404       !MF.getFunction().hasFnAttribute(Attribute::SpeculativeLoadHardening))
405     return false;
406 
407   Subtarget = &MF.getSubtarget<X86Subtarget>();
408   MRI = &MF.getRegInfo();
409   TII = Subtarget->getInstrInfo();
410   TRI = Subtarget->getRegisterInfo();
411 
412   // FIXME: Support for 32-bit.
413   PS.emplace(MF, &X86::GR64_NOSPRegClass);
414 
415   if (MF.begin() == MF.end())
416     // Nothing to do for a degenerate empty function...
417     return false;
418 
419   // We support an alternative hardening technique based on a debug flag.
420   if (HardenEdgesWithLFENCE) {
421     hardenEdgesWithLFENCE(MF);
422     return true;
423   }
424 
425   // Create a dummy debug loc to use for all the generated code here.
426   DebugLoc Loc;
427 
428   MachineBasicBlock &Entry = *MF.begin();
429   auto EntryInsertPt = Entry.SkipPHIsLabelsAndDebug(Entry.begin());
430 
431   // Do a quick scan to see if we have any checkable loads.
432   bool HasVulnerableLoad = hasVulnerableLoad(MF);
433 
434   // See if we have any conditional branching blocks that we will need to trace
435   // predicate state through.
436   SmallVector<BlockCondInfo, 16> Infos = collectBlockCondInfo(MF);
437 
438   // If we have no interesting conditions or loads, nothing to do here.
439   if (!HasVulnerableLoad && Infos.empty())
440     return true;
441 
442   // The poison value is required to be an all-ones value for many aspects of
443   // this mitigation.
444   const int PoisonVal = -1;
445   PS->PoisonReg = MRI->createVirtualRegister(PS->RC);
446   BuildMI(Entry, EntryInsertPt, Loc, TII->get(X86::MOV64ri32), PS->PoisonReg)
447       .addImm(PoisonVal);
448   ++NumInstsInserted;
449 
450   // If we have loads being hardened and we've asked for call and ret edges to
451   // get a full fence-based mitigation, inject that fence.
452   if (HasVulnerableLoad && FenceCallAndRet) {
453     // We need to insert an LFENCE at the start of the function to suspend any
454     // incoming misspeculation from the caller. This helps two-fold: the caller
455     // may not have been protected as this code has been, and this code gets to
456     // not take any specific action to protect across calls.
457     // FIXME: We could skip this for functions which unconditionally return
458     // a constant.
459     BuildMI(Entry, EntryInsertPt, Loc, TII->get(X86::LFENCE));
460     ++NumInstsInserted;
461     ++NumLFENCEsInserted;
462   }
463 
464   // If we guarded the entry with an LFENCE and have no conditionals to protect
465   // in blocks, then we're done.
466   if (FenceCallAndRet && Infos.empty())
467     // We may have changed the function's code at this point to insert fences.
468     return true;
469 
470   // For every basic block in the function which can b
471   if (HardenInterprocedurally && !FenceCallAndRet) {
472     // Set up the predicate state by extracting it from the incoming stack
473     // pointer so we pick up any misspeculation in our caller.
474     PS->InitialReg = extractPredStateFromSP(Entry, EntryInsertPt, Loc);
475   } else {
476     // Otherwise, just build the predicate state itself by zeroing a register
477     // as we don't need any initial state.
478     PS->InitialReg = MRI->createVirtualRegister(PS->RC);
479     Register PredStateSubReg = MRI->createVirtualRegister(&X86::GR32RegClass);
480     auto ZeroI = BuildMI(Entry, EntryInsertPt, Loc, TII->get(X86::MOV32r0),
481                          PredStateSubReg);
482     ++NumInstsInserted;
483     MachineOperand *ZeroEFLAGSDefOp =
484         ZeroI->findRegisterDefOperand(X86::EFLAGS, /*TRI=*/nullptr);
485     assert(ZeroEFLAGSDefOp && ZeroEFLAGSDefOp->isImplicit() &&
486            "Must have an implicit def of EFLAGS!");
487     ZeroEFLAGSDefOp->setIsDead(true);
488     BuildMI(Entry, EntryInsertPt, Loc, TII->get(X86::SUBREG_TO_REG),
489             PS->InitialReg)
490         .addImm(0)
491         .addReg(PredStateSubReg)
492         .addImm(X86::sub_32bit);
493   }
494 
495   // We're going to need to trace predicate state throughout the function's
496   // CFG. Prepare for this by setting up our initial state of PHIs with unique
497   // predecessor entries and all the initial predicate state.
498   canonicalizePHIOperands(MF);
499 
500   // Track the updated values in an SSA updater to rewrite into SSA form at the
501   // end.
502   PS->SSA.Initialize(PS->InitialReg);
503   PS->SSA.AddAvailableValue(&Entry, PS->InitialReg);
504 
505   // Trace through the CFG.
506   auto CMovs = tracePredStateThroughCFG(MF, Infos);
507 
508   // We may also enter basic blocks in this function via exception handling
509   // control flow. Here, if we are hardening interprocedurally, we need to
510   // re-capture the predicate state from the throwing code. In the Itanium ABI,
511   // the throw will always look like a call to __cxa_throw and will have the
512   // predicate state in the stack pointer, so extract fresh predicate state from
513   // the stack pointer and make it available in SSA.
514   // FIXME: Handle non-itanium ABI EH models.
515   if (HardenInterprocedurally) {
516     for (MachineBasicBlock &MBB : MF) {
517       assert(!MBB.isEHScopeEntry() && "Only Itanium ABI EH supported!");
518       assert(!MBB.isEHFuncletEntry() && "Only Itanium ABI EH supported!");
519       assert(!MBB.isCleanupFuncletEntry() && "Only Itanium ABI EH supported!");
520       if (!MBB.isEHPad())
521         continue;
522       PS->SSA.AddAvailableValue(
523           &MBB,
524           extractPredStateFromSP(MBB, MBB.SkipPHIsAndLabels(MBB.begin()), Loc));
525     }
526   }
527 
528   if (HardenIndirectCallsAndJumps) {
529     // If we are going to harden calls and jumps we need to unfold their memory
530     // operands.
531     unfoldCallAndJumpLoads(MF);
532 
533     // Then we trace predicate state through the indirect branches.
534     auto IndirectBrCMovs = tracePredStateThroughIndirectBranches(MF);
535     CMovs.append(IndirectBrCMovs.begin(), IndirectBrCMovs.end());
536   }
537 
538   // Now that we have the predicate state available at the start of each block
539   // in the CFG, trace it through each block, hardening vulnerable instructions
540   // as we go.
541   tracePredStateThroughBlocksAndHarden(MF);
542 
543   // Now rewrite all the uses of the pred state using the SSA updater to insert
544   // PHIs connecting the state between blocks along the CFG edges.
545   for (MachineInstr *CMovI : CMovs)
546     for (MachineOperand &Op : CMovI->operands()) {
547       if (!Op.isReg() || Op.getReg() != PS->InitialReg)
548         continue;
549 
550       PS->SSA.RewriteUse(Op);
551     }
552 
553   LLVM_DEBUG(dbgs() << "Final speculative load hardened function:\n"; MF.dump();
554              dbgs() << "\n"; MF.verify(this));
555   return true;
556 }
557 
558 /// Implements the naive hardening approach of putting an LFENCE after every
559 /// potentially mis-predicted control flow construct.
560 ///
561 /// We include this as an alternative mostly for the purpose of comparison. The
562 /// performance impact of this is expected to be extremely severe and not
563 /// practical for any real-world users.
564 void X86SpeculativeLoadHardeningPass::hardenEdgesWithLFENCE(
565     MachineFunction &MF) {
566   // First, we scan the function looking for blocks that are reached along edges
567   // that we might want to harden.
568   SmallSetVector<MachineBasicBlock *, 8> Blocks;
569   for (MachineBasicBlock &MBB : MF) {
570     // If there are no or only one successor, nothing to do here.
571     if (MBB.succ_size() <= 1)
572       continue;
573 
574     // Skip blocks unless their terminators start with a branch. Other
575     // terminators don't seem interesting for guarding against misspeculation.
576     auto TermIt = MBB.getFirstTerminator();
577     if (TermIt == MBB.end() || !TermIt->isBranch())
578       continue;
579 
580     // Add all the non-EH-pad succossors to the blocks we want to harden. We
581     // skip EH pads because there isn't really a condition of interest on
582     // entering.
583     for (MachineBasicBlock *SuccMBB : MBB.successors())
584       if (!SuccMBB->isEHPad())
585         Blocks.insert(SuccMBB);
586   }
587 
588   for (MachineBasicBlock *MBB : Blocks) {
589     auto InsertPt = MBB->SkipPHIsAndLabels(MBB->begin());
590     BuildMI(*MBB, InsertPt, DebugLoc(), TII->get(X86::LFENCE));
591     ++NumInstsInserted;
592     ++NumLFENCEsInserted;
593   }
594 }
595 
596 SmallVector<X86SpeculativeLoadHardeningPass::BlockCondInfo, 16>
597 X86SpeculativeLoadHardeningPass::collectBlockCondInfo(MachineFunction &MF) {
598   SmallVector<BlockCondInfo, 16> Infos;
599 
600   // Walk the function and build up a summary for each block's conditions that
601   // we need to trace through.
602   for (MachineBasicBlock &MBB : MF) {
603     // If there are no or only one successor, nothing to do here.
604     if (MBB.succ_size() <= 1)
605       continue;
606 
607     // We want to reliably handle any conditional branch terminators in the
608     // MBB, so we manually analyze the branch. We can handle all of the
609     // permutations here, including ones that analyze branch cannot.
610     //
611     // The approach is to walk backwards across the terminators, resetting at
612     // any unconditional non-indirect branch, and track all conditional edges
613     // to basic blocks as well as the fallthrough or unconditional successor
614     // edge. For each conditional edge, we track the target and the opposite
615     // condition code in order to inject a "no-op" cmov into that successor
616     // that will harden the predicate. For the fallthrough/unconditional
617     // edge, we inject a separate cmov for each conditional branch with
618     // matching condition codes. This effectively implements an "and" of the
619     // condition flags, even if there isn't a single condition flag that would
620     // directly implement that. We don't bother trying to optimize either of
621     // these cases because if such an optimization is possible, LLVM should
622     // have optimized the conditional *branches* in that way already to reduce
623     // instruction count. This late, we simply assume the minimal number of
624     // branch instructions is being emitted and use that to guide our cmov
625     // insertion.
626 
627     BlockCondInfo Info = {&MBB, {}, nullptr};
628 
629     // Now walk backwards through the terminators and build up successors they
630     // reach and the conditions.
631     for (MachineInstr &MI : llvm::reverse(MBB)) {
632       // Once we've handled all the terminators, we're done.
633       if (!MI.isTerminator())
634         break;
635 
636       // If we see a non-branch terminator, we can't handle anything so bail.
637       if (!MI.isBranch()) {
638         Info.CondBrs.clear();
639         break;
640       }
641 
642       // If we see an unconditional branch, reset our state, clear any
643       // fallthrough, and set this is the "else" successor.
644       if (MI.getOpcode() == X86::JMP_1) {
645         Info.CondBrs.clear();
646         Info.UncondBr = &MI;
647         continue;
648       }
649 
650       // If we get an invalid condition, we have an indirect branch or some
651       // other unanalyzable "fallthrough" case. We model this as a nullptr for
652       // the destination so we can still guard any conditional successors.
653       // Consider code sequences like:
654       // ```
655       //   jCC L1
656       //   jmpq *%rax
657       // ```
658       // We still want to harden the edge to `L1`.
659       if (X86::getCondFromBranch(MI) == X86::COND_INVALID) {
660         Info.CondBrs.clear();
661         Info.UncondBr = &MI;
662         continue;
663       }
664 
665       // We have a vanilla conditional branch, add it to our list.
666       Info.CondBrs.push_back(&MI);
667     }
668     if (Info.CondBrs.empty()) {
669       ++NumBranchesUntraced;
670       LLVM_DEBUG(dbgs() << "WARNING: unable to secure successors of block:\n";
671                  MBB.dump());
672       continue;
673     }
674 
675     Infos.push_back(Info);
676   }
677 
678   return Infos;
679 }
680 
681 /// Trace the predicate state through the CFG, instrumenting each conditional
682 /// branch such that misspeculation through an edge will poison the predicate
683 /// state.
684 ///
685 /// Returns the list of inserted CMov instructions so that they can have their
686 /// uses of the predicate state rewritten into proper SSA form once it is
687 /// complete.
688 SmallVector<MachineInstr *, 16>
689 X86SpeculativeLoadHardeningPass::tracePredStateThroughCFG(
690     MachineFunction &MF, ArrayRef<BlockCondInfo> Infos) {
691   // Collect the inserted cmov instructions so we can rewrite their uses of the
692   // predicate state into SSA form.
693   SmallVector<MachineInstr *, 16> CMovs;
694 
695   // Now walk all of the basic blocks looking for ones that end in conditional
696   // jumps where we need to update this register along each edge.
697   for (const BlockCondInfo &Info : Infos) {
698     MachineBasicBlock &MBB = *Info.MBB;
699     const SmallVectorImpl<MachineInstr *> &CondBrs = Info.CondBrs;
700     MachineInstr *UncondBr = Info.UncondBr;
701 
702     LLVM_DEBUG(dbgs() << "Tracing predicate through block: " << MBB.getName()
703                       << "\n");
704     ++NumCondBranchesTraced;
705 
706     // Compute the non-conditional successor as either the target of any
707     // unconditional branch or the layout successor.
708     MachineBasicBlock *UncondSucc =
709         UncondBr ? (UncondBr->getOpcode() == X86::JMP_1
710                         ? UncondBr->getOperand(0).getMBB()
711                         : nullptr)
712                  : &*std::next(MachineFunction::iterator(&MBB));
713 
714     // Count how many edges there are to any given successor.
715     SmallDenseMap<MachineBasicBlock *, int> SuccCounts;
716     if (UncondSucc)
717       ++SuccCounts[UncondSucc];
718     for (auto *CondBr : CondBrs)
719       ++SuccCounts[CondBr->getOperand(0).getMBB()];
720 
721     // A lambda to insert cmov instructions into a block checking all of the
722     // condition codes in a sequence.
723     auto BuildCheckingBlockForSuccAndConds =
724         [&](MachineBasicBlock &MBB, MachineBasicBlock &Succ, int SuccCount,
725             MachineInstr *Br, MachineInstr *&UncondBr,
726             ArrayRef<X86::CondCode> Conds) {
727           // First, we split the edge to insert the checking block into a safe
728           // location.
729           auto &CheckingMBB =
730               (SuccCount == 1 && Succ.pred_size() == 1)
731                   ? Succ
732                   : splitEdge(MBB, Succ, SuccCount, Br, UncondBr, *TII);
733 
734           bool LiveEFLAGS = Succ.isLiveIn(X86::EFLAGS);
735           if (!LiveEFLAGS)
736             CheckingMBB.addLiveIn(X86::EFLAGS);
737 
738           // Now insert the cmovs to implement the checks.
739           auto InsertPt = CheckingMBB.begin();
740           assert((InsertPt == CheckingMBB.end() || !InsertPt->isPHI()) &&
741                  "Should never have a PHI in the initial checking block as it "
742                  "always has a single predecessor!");
743 
744           // We will wire each cmov to each other, but need to start with the
745           // incoming pred state.
746           unsigned CurStateReg = PS->InitialReg;
747 
748           for (X86::CondCode Cond : Conds) {
749             int PredStateSizeInBytes = TRI->getRegSizeInBits(*PS->RC) / 8;
750             auto CMovOp = X86::getCMovOpcode(PredStateSizeInBytes);
751 
752             Register UpdatedStateReg = MRI->createVirtualRegister(PS->RC);
753             // Note that we intentionally use an empty debug location so that
754             // this picks up the preceding location.
755             auto CMovI = BuildMI(CheckingMBB, InsertPt, DebugLoc(),
756                                  TII->get(CMovOp), UpdatedStateReg)
757                              .addReg(CurStateReg)
758                              .addReg(PS->PoisonReg)
759                              .addImm(Cond);
760             // If this is the last cmov and the EFLAGS weren't originally
761             // live-in, mark them as killed.
762             if (!LiveEFLAGS && Cond == Conds.back())
763               CMovI->findRegisterUseOperand(X86::EFLAGS, /*TRI=*/nullptr)
764                   ->setIsKill(true);
765 
766             ++NumInstsInserted;
767             LLVM_DEBUG(dbgs() << "  Inserting cmov: "; CMovI->dump();
768                        dbgs() << "\n");
769 
770             // The first one of the cmovs will be using the top level
771             // `PredStateReg` and need to get rewritten into SSA form.
772             if (CurStateReg == PS->InitialReg)
773               CMovs.push_back(&*CMovI);
774 
775             // The next cmov should start from this one's def.
776             CurStateReg = UpdatedStateReg;
777           }
778 
779           // And put the last one into the available values for SSA form of our
780           // predicate state.
781           PS->SSA.AddAvailableValue(&CheckingMBB, CurStateReg);
782         };
783 
784     std::vector<X86::CondCode> UncondCodeSeq;
785     for (auto *CondBr : CondBrs) {
786       MachineBasicBlock &Succ = *CondBr->getOperand(0).getMBB();
787       int &SuccCount = SuccCounts[&Succ];
788 
789       X86::CondCode Cond = X86::getCondFromBranch(*CondBr);
790       X86::CondCode InvCond = X86::GetOppositeBranchCondition(Cond);
791       UncondCodeSeq.push_back(Cond);
792 
793       BuildCheckingBlockForSuccAndConds(MBB, Succ, SuccCount, CondBr, UncondBr,
794                                         {InvCond});
795 
796       // Decrement the successor count now that we've split one of the edges.
797       // We need to keep the count of edges to the successor accurate in order
798       // to know above when to *replace* the successor in the CFG vs. just
799       // adding the new successor.
800       --SuccCount;
801     }
802 
803     // Since we may have split edges and changed the number of successors,
804     // normalize the probabilities. This avoids doing it each time we split an
805     // edge.
806     MBB.normalizeSuccProbs();
807 
808     // Finally, we need to insert cmovs into the "fallthrough" edge. Here, we
809     // need to intersect the other condition codes. We can do this by just
810     // doing a cmov for each one.
811     if (!UncondSucc)
812       // If we have no fallthrough to protect (perhaps it is an indirect jump?)
813       // just skip this and continue.
814       continue;
815 
816     assert(SuccCounts[UncondSucc] == 1 &&
817            "We should never have more than one edge to the unconditional "
818            "successor at this point because every other edge must have been "
819            "split above!");
820 
821     // Sort and unique the codes to minimize them.
822     llvm::sort(UncondCodeSeq);
823     UncondCodeSeq.erase(llvm::unique(UncondCodeSeq), UncondCodeSeq.end());
824 
825     // Build a checking version of the successor.
826     BuildCheckingBlockForSuccAndConds(MBB, *UncondSucc, /*SuccCount*/ 1,
827                                       UncondBr, UncondBr, UncondCodeSeq);
828   }
829 
830   return CMovs;
831 }
832 
833 /// Compute the register class for the unfolded load.
834 ///
835 /// FIXME: This should probably live in X86InstrInfo, potentially by adding
836 /// a way to unfold into a newly created vreg rather than requiring a register
837 /// input.
838 static const TargetRegisterClass *
839 getRegClassForUnfoldedLoad(MachineFunction &MF, const X86InstrInfo &TII,
840                            unsigned Opcode) {
841   unsigned Index;
842   unsigned UnfoldedOpc = TII.getOpcodeAfterMemoryUnfold(
843       Opcode, /*UnfoldLoad*/ true, /*UnfoldStore*/ false, &Index);
844   const MCInstrDesc &MCID = TII.get(UnfoldedOpc);
845   return TII.getRegClass(MCID, Index, &TII.getRegisterInfo(), MF);
846 }
847 
848 void X86SpeculativeLoadHardeningPass::unfoldCallAndJumpLoads(
849     MachineFunction &MF) {
850   for (MachineBasicBlock &MBB : MF)
851     // We use make_early_inc_range here so we can remove instructions if needed
852     // without disturbing the iteration.
853     for (MachineInstr &MI : llvm::make_early_inc_range(MBB.instrs())) {
854       // Must either be a call or a branch.
855       if (!MI.isCall() && !MI.isBranch())
856         continue;
857       // We only care about loading variants of these instructions.
858       if (!MI.mayLoad())
859         continue;
860 
861       switch (MI.getOpcode()) {
862       default: {
863         LLVM_DEBUG(
864             dbgs() << "ERROR: Found an unexpected loading branch or call "
865                       "instruction:\n";
866             MI.dump(); dbgs() << "\n");
867         report_fatal_error("Unexpected loading branch or call!");
868       }
869 
870       case X86::FARCALL16m:
871       case X86::FARCALL32m:
872       case X86::FARCALL64m:
873       case X86::FARJMP16m:
874       case X86::FARJMP32m:
875       case X86::FARJMP64m:
876         // We cannot mitigate far jumps or calls, but we also don't expect them
877         // to be vulnerable to Spectre v1.2 style attacks.
878         continue;
879 
880       case X86::CALL16m:
881       case X86::CALL16m_NT:
882       case X86::CALL32m:
883       case X86::CALL32m_NT:
884       case X86::CALL64m:
885       case X86::CALL64m_NT:
886       case X86::JMP16m:
887       case X86::JMP16m_NT:
888       case X86::JMP32m:
889       case X86::JMP32m_NT:
890       case X86::JMP64m:
891       case X86::JMP64m_NT:
892       case X86::TAILJMPm64:
893       case X86::TAILJMPm64_REX:
894       case X86::TAILJMPm:
895       case X86::TCRETURNmi64:
896       case X86::TCRETURNmi: {
897         // Use the generic unfold logic now that we know we're dealing with
898         // expected instructions.
899         // FIXME: We don't have test coverage for all of these!
900         auto *UnfoldedRC = getRegClassForUnfoldedLoad(MF, *TII, MI.getOpcode());
901         if (!UnfoldedRC) {
902           LLVM_DEBUG(dbgs()
903                          << "ERROR: Unable to unfold load from instruction:\n";
904                      MI.dump(); dbgs() << "\n");
905           report_fatal_error("Unable to unfold load!");
906         }
907         Register Reg = MRI->createVirtualRegister(UnfoldedRC);
908         SmallVector<MachineInstr *, 2> NewMIs;
909         // If we were able to compute an unfolded reg class, any failure here
910         // is just a programming error so just assert.
911         bool Unfolded =
912             TII->unfoldMemoryOperand(MF, MI, Reg, /*UnfoldLoad*/ true,
913                                      /*UnfoldStore*/ false, NewMIs);
914         (void)Unfolded;
915         assert(Unfolded &&
916                "Computed unfolded register class but failed to unfold");
917         // Now stitch the new instructions into place and erase the old one.
918         for (auto *NewMI : NewMIs)
919           MBB.insert(MI.getIterator(), NewMI);
920 
921         // Update the call info.
922         if (MI.isCandidateForAdditionalCallInfo())
923           MF.eraseAdditionalCallInfo(&MI);
924 
925         MI.eraseFromParent();
926         LLVM_DEBUG({
927           dbgs() << "Unfolded load successfully into:\n";
928           for (auto *NewMI : NewMIs) {
929             NewMI->dump();
930             dbgs() << "\n";
931           }
932         });
933         continue;
934       }
935       }
936       llvm_unreachable("Escaped switch with default!");
937     }
938 }
939 
940 /// Trace the predicate state through indirect branches, instrumenting them to
941 /// poison the state if a target is reached that does not match the expected
942 /// target.
943 ///
944 /// This is designed to mitigate Spectre variant 1 attacks where an indirect
945 /// branch is trained to predict a particular target and then mispredicts that
946 /// target in a way that can leak data. Despite using an indirect branch, this
947 /// is really a variant 1 style attack: it does not steer execution to an
948 /// arbitrary or attacker controlled address, and it does not require any
949 /// special code executing next to the victim. This attack can also be mitigated
950 /// through retpolines, but those require either replacing indirect branches
951 /// with conditional direct branches or lowering them through a device that
952 /// blocks speculation. This mitigation can replace these retpoline-style
953 /// mitigations for jump tables and other indirect branches within a function
954 /// when variant 2 isn't a risk while allowing limited speculation. Indirect
955 /// calls, however, cannot be mitigated through this technique without changing
956 /// the ABI in a fundamental way.
957 SmallVector<MachineInstr *, 16>
958 X86SpeculativeLoadHardeningPass::tracePredStateThroughIndirectBranches(
959     MachineFunction &MF) {
960   // We use the SSAUpdater to insert PHI nodes for the target addresses of
961   // indirect branches. We don't actually need the full power of the SSA updater
962   // in this particular case as we always have immediately available values, but
963   // this avoids us having to re-implement the PHI construction logic.
964   MachineSSAUpdater TargetAddrSSA(MF);
965   TargetAddrSSA.Initialize(MRI->createVirtualRegister(&X86::GR64RegClass));
966 
967   // Track which blocks were terminated with an indirect branch.
968   SmallPtrSet<MachineBasicBlock *, 4> IndirectTerminatedMBBs;
969 
970   // We need to know what blocks end up reached via indirect branches. We
971   // expect this to be a subset of those whose address is taken and so track it
972   // directly via the CFG.
973   SmallPtrSet<MachineBasicBlock *, 4> IndirectTargetMBBs;
974 
975   // Walk all the blocks which end in an indirect branch and make the
976   // target address available.
977   for (MachineBasicBlock &MBB : MF) {
978     // Find the last terminator.
979     auto MII = MBB.instr_rbegin();
980     while (MII != MBB.instr_rend() && MII->isDebugInstr())
981       ++MII;
982     if (MII == MBB.instr_rend())
983       continue;
984     MachineInstr &TI = *MII;
985     if (!TI.isTerminator() || !TI.isBranch())
986       // No terminator or non-branch terminator.
987       continue;
988 
989     unsigned TargetReg;
990 
991     switch (TI.getOpcode()) {
992     default:
993       // Direct branch or conditional branch (leading to fallthrough).
994       continue;
995 
996     case X86::FARJMP16m:
997     case X86::FARJMP32m:
998     case X86::FARJMP64m:
999       // We cannot mitigate far jumps or calls, but we also don't expect them
1000       // to be vulnerable to Spectre v1.2 or v2 (self trained) style attacks.
1001       continue;
1002 
1003     case X86::JMP16m:
1004     case X86::JMP16m_NT:
1005     case X86::JMP32m:
1006     case X86::JMP32m_NT:
1007     case X86::JMP64m:
1008     case X86::JMP64m_NT:
1009       // Mostly as documentation.
1010       report_fatal_error("Memory operand jumps should have been unfolded!");
1011 
1012     case X86::JMP16r:
1013       report_fatal_error(
1014           "Support for 16-bit indirect branches is not implemented.");
1015     case X86::JMP32r:
1016       report_fatal_error(
1017           "Support for 32-bit indirect branches is not implemented.");
1018 
1019     case X86::JMP64r:
1020       TargetReg = TI.getOperand(0).getReg();
1021     }
1022 
1023     // We have definitely found an indirect  branch. Verify that there are no
1024     // preceding conditional branches as we don't yet support that.
1025     if (llvm::any_of(MBB.terminators(), [&](MachineInstr &OtherTI) {
1026           return !OtherTI.isDebugInstr() && &OtherTI != &TI;
1027         })) {
1028       LLVM_DEBUG({
1029         dbgs() << "ERROR: Found other terminators in a block with an indirect "
1030                   "branch! This is not yet supported! Terminator sequence:\n";
1031         for (MachineInstr &MI : MBB.terminators()) {
1032           MI.dump();
1033           dbgs() << '\n';
1034         }
1035       });
1036       report_fatal_error("Unimplemented terminator sequence!");
1037     }
1038 
1039     // Make the target register an available value for this block.
1040     TargetAddrSSA.AddAvailableValue(&MBB, TargetReg);
1041     IndirectTerminatedMBBs.insert(&MBB);
1042 
1043     // Add all the successors to our target candidates.
1044     for (MachineBasicBlock *Succ : MBB.successors())
1045       IndirectTargetMBBs.insert(Succ);
1046   }
1047 
1048   // Keep track of the cmov instructions we insert so we can return them.
1049   SmallVector<MachineInstr *, 16> CMovs;
1050 
1051   // If we didn't find any indirect branches with targets, nothing to do here.
1052   if (IndirectTargetMBBs.empty())
1053     return CMovs;
1054 
1055   // We found indirect branches and targets that need to be instrumented to
1056   // harden loads within them. Walk the blocks of the function (to get a stable
1057   // ordering) and instrument each target of an indirect branch.
1058   for (MachineBasicBlock &MBB : MF) {
1059     // Skip the blocks that aren't candidate targets.
1060     if (!IndirectTargetMBBs.count(&MBB))
1061       continue;
1062 
1063     // We don't expect EH pads to ever be reached via an indirect branch. If
1064     // this is desired for some reason, we could simply skip them here rather
1065     // than asserting.
1066     assert(!MBB.isEHPad() &&
1067            "Unexpected EH pad as target of an indirect branch!");
1068 
1069     // We should never end up threading EFLAGS into a block to harden
1070     // conditional jumps as there would be an additional successor via the
1071     // indirect branch. As a consequence, all such edges would be split before
1072     // reaching here, and the inserted block will handle the EFLAGS-based
1073     // hardening.
1074     assert(!MBB.isLiveIn(X86::EFLAGS) &&
1075            "Cannot check within a block that already has live-in EFLAGS!");
1076 
1077     // We can't handle having non-indirect edges into this block unless this is
1078     // the only successor and we can synthesize the necessary target address.
1079     for (MachineBasicBlock *Pred : MBB.predecessors()) {
1080       // If we've already handled this by extracting the target directly,
1081       // nothing to do.
1082       if (IndirectTerminatedMBBs.count(Pred))
1083         continue;
1084 
1085       // Otherwise, we have to be the only successor. We generally expect this
1086       // to be true as conditional branches should have had a critical edge
1087       // split already. We don't however need to worry about EH pad successors
1088       // as they'll happily ignore the target and their hardening strategy is
1089       // resilient to all ways in which they could be reached speculatively.
1090       if (!llvm::all_of(Pred->successors(), [&](MachineBasicBlock *Succ) {
1091             return Succ->isEHPad() || Succ == &MBB;
1092           })) {
1093         LLVM_DEBUG({
1094           dbgs() << "ERROR: Found conditional entry to target of indirect "
1095                     "branch!\n";
1096           Pred->dump();
1097           MBB.dump();
1098         });
1099         report_fatal_error("Cannot harden a conditional entry to a target of "
1100                            "an indirect branch!");
1101       }
1102 
1103       // Now we need to compute the address of this block and install it as a
1104       // synthetic target in the predecessor. We do this at the bottom of the
1105       // predecessor.
1106       auto InsertPt = Pred->getFirstTerminator();
1107       Register TargetReg = MRI->createVirtualRegister(&X86::GR64RegClass);
1108       if (MF.getTarget().getCodeModel() == CodeModel::Small &&
1109           !Subtarget->isPositionIndependent()) {
1110         // Directly materialize it into an immediate.
1111         auto AddrI = BuildMI(*Pred, InsertPt, DebugLoc(),
1112                              TII->get(X86::MOV64ri32), TargetReg)
1113                          .addMBB(&MBB);
1114         ++NumInstsInserted;
1115         (void)AddrI;
1116         LLVM_DEBUG(dbgs() << "  Inserting mov: "; AddrI->dump();
1117                    dbgs() << "\n");
1118       } else {
1119         auto AddrI = BuildMI(*Pred, InsertPt, DebugLoc(), TII->get(X86::LEA64r),
1120                              TargetReg)
1121                          .addReg(/*Base*/ X86::RIP)
1122                          .addImm(/*Scale*/ 1)
1123                          .addReg(/*Index*/ 0)
1124                          .addMBB(&MBB)
1125                          .addReg(/*Segment*/ 0);
1126         ++NumInstsInserted;
1127         (void)AddrI;
1128         LLVM_DEBUG(dbgs() << "  Inserting lea: "; AddrI->dump();
1129                    dbgs() << "\n");
1130       }
1131       // And make this available.
1132       TargetAddrSSA.AddAvailableValue(Pred, TargetReg);
1133     }
1134 
1135     // Materialize the needed SSA value of the target. Note that we need the
1136     // middle of the block as this block might at the bottom have an indirect
1137     // branch back to itself. We can do this here because at this point, every
1138     // predecessor of this block has an available value. This is basically just
1139     // automating the construction of a PHI node for this target.
1140     Register TargetReg = TargetAddrSSA.GetValueInMiddleOfBlock(&MBB);
1141 
1142     // Insert a comparison of the incoming target register with this block's
1143     // address. This also requires us to mark the block as having its address
1144     // taken explicitly.
1145     MBB.setMachineBlockAddressTaken();
1146     auto InsertPt = MBB.SkipPHIsLabelsAndDebug(MBB.begin());
1147     if (MF.getTarget().getCodeModel() == CodeModel::Small &&
1148         !Subtarget->isPositionIndependent()) {
1149       // Check directly against a relocated immediate when we can.
1150       auto CheckI = BuildMI(MBB, InsertPt, DebugLoc(), TII->get(X86::CMP64ri32))
1151                         .addReg(TargetReg, RegState::Kill)
1152                         .addMBB(&MBB);
1153       ++NumInstsInserted;
1154       (void)CheckI;
1155       LLVM_DEBUG(dbgs() << "  Inserting cmp: "; CheckI->dump(); dbgs() << "\n");
1156     } else {
1157       // Otherwise compute the address into a register first.
1158       Register AddrReg = MRI->createVirtualRegister(&X86::GR64RegClass);
1159       auto AddrI =
1160           BuildMI(MBB, InsertPt, DebugLoc(), TII->get(X86::LEA64r), AddrReg)
1161               .addReg(/*Base*/ X86::RIP)
1162               .addImm(/*Scale*/ 1)
1163               .addReg(/*Index*/ 0)
1164               .addMBB(&MBB)
1165               .addReg(/*Segment*/ 0);
1166       ++NumInstsInserted;
1167       (void)AddrI;
1168       LLVM_DEBUG(dbgs() << "  Inserting lea: "; AddrI->dump(); dbgs() << "\n");
1169       auto CheckI = BuildMI(MBB, InsertPt, DebugLoc(), TII->get(X86::CMP64rr))
1170                         .addReg(TargetReg, RegState::Kill)
1171                         .addReg(AddrReg, RegState::Kill);
1172       ++NumInstsInserted;
1173       (void)CheckI;
1174       LLVM_DEBUG(dbgs() << "  Inserting cmp: "; CheckI->dump(); dbgs() << "\n");
1175     }
1176 
1177     // Now cmov over the predicate if the comparison wasn't equal.
1178     int PredStateSizeInBytes = TRI->getRegSizeInBits(*PS->RC) / 8;
1179     auto CMovOp = X86::getCMovOpcode(PredStateSizeInBytes);
1180     Register UpdatedStateReg = MRI->createVirtualRegister(PS->RC);
1181     auto CMovI =
1182         BuildMI(MBB, InsertPt, DebugLoc(), TII->get(CMovOp), UpdatedStateReg)
1183             .addReg(PS->InitialReg)
1184             .addReg(PS->PoisonReg)
1185             .addImm(X86::COND_NE);
1186     CMovI->findRegisterUseOperand(X86::EFLAGS, /*TRI=*/nullptr)
1187         ->setIsKill(true);
1188     ++NumInstsInserted;
1189     LLVM_DEBUG(dbgs() << "  Inserting cmov: "; CMovI->dump(); dbgs() << "\n");
1190     CMovs.push_back(&*CMovI);
1191 
1192     // And put the new value into the available values for SSA form of our
1193     // predicate state.
1194     PS->SSA.AddAvailableValue(&MBB, UpdatedStateReg);
1195   }
1196 
1197   // Return all the newly inserted cmov instructions of the predicate state.
1198   return CMovs;
1199 }
1200 
1201 // Returns true if the MI has EFLAGS as a register def operand and it's live,
1202 // otherwise it returns false
1203 static bool isEFLAGSDefLive(const MachineInstr &MI) {
1204   if (const MachineOperand *DefOp =
1205           MI.findRegisterDefOperand(X86::EFLAGS, /*TRI=*/nullptr)) {
1206     return !DefOp->isDead();
1207   }
1208   return false;
1209 }
1210 
1211 static bool isEFLAGSLive(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
1212                          const TargetRegisterInfo &TRI) {
1213   // Check if EFLAGS are alive by seeing if there is a def of them or they
1214   // live-in, and then seeing if that def is in turn used.
1215   for (MachineInstr &MI : llvm::reverse(llvm::make_range(MBB.begin(), I))) {
1216     if (MachineOperand *DefOp =
1217             MI.findRegisterDefOperand(X86::EFLAGS, /*TRI=*/nullptr)) {
1218       // If the def is dead, then EFLAGS is not live.
1219       if (DefOp->isDead())
1220         return false;
1221 
1222       // Otherwise we've def'ed it, and it is live.
1223       return true;
1224     }
1225     // While at this instruction, also check if we use and kill EFLAGS
1226     // which means it isn't live.
1227     if (MI.killsRegister(X86::EFLAGS, &TRI))
1228       return false;
1229   }
1230 
1231   // If we didn't find anything conclusive (neither definitely alive or
1232   // definitely dead) return whether it lives into the block.
1233   return MBB.isLiveIn(X86::EFLAGS);
1234 }
1235 
1236 /// Trace the predicate state through each of the blocks in the function,
1237 /// hardening everything necessary along the way.
1238 ///
1239 /// We call this routine once the initial predicate state has been established
1240 /// for each basic block in the function in the SSA updater. This routine traces
1241 /// it through the instructions within each basic block, and for non-returning
1242 /// blocks informs the SSA updater about the final state that lives out of the
1243 /// block. Along the way, it hardens any vulnerable instruction using the
1244 /// currently valid predicate state. We have to do these two things together
1245 /// because the SSA updater only works across blocks. Within a block, we track
1246 /// the current predicate state directly and update it as it changes.
1247 ///
1248 /// This operates in two passes over each block. First, we analyze the loads in
1249 /// the block to determine which strategy will be used to harden them: hardening
1250 /// the address or hardening the loaded value when loaded into a register
1251 /// amenable to hardening. We have to process these first because the two
1252 /// strategies may interact -- later hardening may change what strategy we wish
1253 /// to use. We also will analyze data dependencies between loads and avoid
1254 /// hardening those loads that are data dependent on a load with a hardened
1255 /// address. We also skip hardening loads already behind an LFENCE as that is
1256 /// sufficient to harden them against misspeculation.
1257 ///
1258 /// Second, we actively trace the predicate state through the block, applying
1259 /// the hardening steps we determined necessary in the first pass as we go.
1260 ///
1261 /// These two passes are applied to each basic block. We operate one block at a
1262 /// time to simplify reasoning about reachability and sequencing.
1263 void X86SpeculativeLoadHardeningPass::tracePredStateThroughBlocksAndHarden(
1264     MachineFunction &MF) {
1265   SmallPtrSet<MachineInstr *, 16> HardenPostLoad;
1266   SmallPtrSet<MachineInstr *, 16> HardenLoadAddr;
1267 
1268   SmallSet<unsigned, 16> HardenedAddrRegs;
1269 
1270   SmallDenseMap<unsigned, unsigned, 32> AddrRegToHardenedReg;
1271 
1272   // Track the set of load-dependent registers through the basic block. Because
1273   // the values of these registers have an existing data dependency on a loaded
1274   // value which we would have checked, we can omit any checks on them.
1275   SparseBitVector<> LoadDepRegs;
1276 
1277   for (MachineBasicBlock &MBB : MF) {
1278     // The first pass over the block: collect all the loads which can have their
1279     // loaded value hardened and all the loads that instead need their address
1280     // hardened. During this walk we propagate load dependence for address
1281     // hardened loads and also look for LFENCE to stop hardening wherever
1282     // possible. When deciding whether or not to harden the loaded value or not,
1283     // we check to see if any registers used in the address will have been
1284     // hardened at this point and if so, harden any remaining address registers
1285     // as that often successfully re-uses hardened addresses and minimizes
1286     // instructions.
1287     //
1288     // FIXME: We should consider an aggressive mode where we continue to keep as
1289     // many loads value hardened even when some address register hardening would
1290     // be free (due to reuse).
1291     //
1292     // Note that we only need this pass if we are actually hardening loads.
1293     if (HardenLoads)
1294       for (MachineInstr &MI : MBB) {
1295         // We naively assume that all def'ed registers of an instruction have
1296         // a data dependency on all of their operands.
1297         // FIXME: Do a more careful analysis of x86 to build a conservative
1298         // model here.
1299         if (llvm::any_of(MI.uses(), [&](MachineOperand &Op) {
1300               return Op.isReg() && LoadDepRegs.test(Op.getReg());
1301             }))
1302           for (MachineOperand &Def : MI.defs())
1303             if (Def.isReg())
1304               LoadDepRegs.set(Def.getReg());
1305 
1306         // Both Intel and AMD are guiding that they will change the semantics of
1307         // LFENCE to be a speculation barrier, so if we see an LFENCE, there is
1308         // no more need to guard things in this block.
1309         if (MI.getOpcode() == X86::LFENCE)
1310           break;
1311 
1312         // If this instruction cannot load, nothing to do.
1313         if (!MI.mayLoad())
1314           continue;
1315 
1316         // Some instructions which "load" are trivially safe or unimportant.
1317         if (MI.getOpcode() == X86::MFENCE)
1318           continue;
1319 
1320         // Extract the memory operand information about this instruction.
1321         const int MemRefBeginIdx = X86::getFirstAddrOperandIdx(MI);
1322         if (MemRefBeginIdx < 0) {
1323           LLVM_DEBUG(dbgs()
1324                          << "WARNING: unable to harden loading instruction: ";
1325                      MI.dump());
1326           continue;
1327         }
1328 
1329         MachineOperand &BaseMO =
1330             MI.getOperand(MemRefBeginIdx + X86::AddrBaseReg);
1331         MachineOperand &IndexMO =
1332             MI.getOperand(MemRefBeginIdx + X86::AddrIndexReg);
1333 
1334         // If we have at least one (non-frame-index, non-RIP) register operand,
1335         // and neither operand is load-dependent, we need to check the load.
1336         unsigned BaseReg = 0, IndexReg = 0;
1337         if (!BaseMO.isFI() && BaseMO.getReg() != X86::RIP &&
1338             BaseMO.getReg() != X86::NoRegister)
1339           BaseReg = BaseMO.getReg();
1340         if (IndexMO.getReg() != X86::NoRegister)
1341           IndexReg = IndexMO.getReg();
1342 
1343         if (!BaseReg && !IndexReg)
1344           // No register operands!
1345           continue;
1346 
1347         // If any register operand is dependent, this load is dependent and we
1348         // needn't check it.
1349         // FIXME: Is this true in the case where we are hardening loads after
1350         // they complete? Unclear, need to investigate.
1351         if ((BaseReg && LoadDepRegs.test(BaseReg)) ||
1352             (IndexReg && LoadDepRegs.test(IndexReg)))
1353           continue;
1354 
1355         // If post-load hardening is enabled, this load is compatible with
1356         // post-load hardening, and we aren't already going to harden one of the
1357         // address registers, queue it up to be hardened post-load. Notably,
1358         // even once hardened this won't introduce a useful dependency that
1359         // could prune out subsequent loads.
1360         if (EnablePostLoadHardening && X86InstrInfo::isDataInvariantLoad(MI) &&
1361             !isEFLAGSDefLive(MI) && MI.getDesc().getNumDefs() == 1 &&
1362             MI.getOperand(0).isReg() &&
1363             canHardenRegister(MI.getOperand(0).getReg()) &&
1364             !HardenedAddrRegs.count(BaseReg) &&
1365             !HardenedAddrRegs.count(IndexReg)) {
1366           HardenPostLoad.insert(&MI);
1367           HardenedAddrRegs.insert(MI.getOperand(0).getReg());
1368           continue;
1369         }
1370 
1371         // Record this instruction for address hardening and record its register
1372         // operands as being address-hardened.
1373         HardenLoadAddr.insert(&MI);
1374         if (BaseReg)
1375           HardenedAddrRegs.insert(BaseReg);
1376         if (IndexReg)
1377           HardenedAddrRegs.insert(IndexReg);
1378 
1379         for (MachineOperand &Def : MI.defs())
1380           if (Def.isReg())
1381             LoadDepRegs.set(Def.getReg());
1382       }
1383 
1384     // Now re-walk the instructions in the basic block, and apply whichever
1385     // hardening strategy we have elected. Note that we do this in a second
1386     // pass specifically so that we have the complete set of instructions for
1387     // which we will do post-load hardening and can defer it in certain
1388     // circumstances.
1389     for (MachineInstr &MI : MBB) {
1390       if (HardenLoads) {
1391         // We cannot both require hardening the def of a load and its address.
1392         assert(!(HardenLoadAddr.count(&MI) && HardenPostLoad.count(&MI)) &&
1393                "Requested to harden both the address and def of a load!");
1394 
1395         // Check if this is a load whose address needs to be hardened.
1396         if (HardenLoadAddr.erase(&MI)) {
1397           const int MemRefBeginIdx = X86::getFirstAddrOperandIdx(MI);
1398           assert(MemRefBeginIdx >= 0 && "Cannot have an invalid index here!");
1399 
1400           MachineOperand &BaseMO =
1401               MI.getOperand(MemRefBeginIdx + X86::AddrBaseReg);
1402           MachineOperand &IndexMO =
1403               MI.getOperand(MemRefBeginIdx + X86::AddrIndexReg);
1404           hardenLoadAddr(MI, BaseMO, IndexMO, AddrRegToHardenedReg);
1405           continue;
1406         }
1407 
1408         // Test if this instruction is one of our post load instructions (and
1409         // remove it from the set if so).
1410         if (HardenPostLoad.erase(&MI)) {
1411           assert(!MI.isCall() && "Must not try to post-load harden a call!");
1412 
1413           // If this is a data-invariant load and there is no EFLAGS
1414           // interference, we want to try and sink any hardening as far as
1415           // possible.
1416           if (X86InstrInfo::isDataInvariantLoad(MI) && !isEFLAGSDefLive(MI)) {
1417             // Sink the instruction we'll need to harden as far as we can down
1418             // the graph.
1419             MachineInstr *SunkMI = sinkPostLoadHardenedInst(MI, HardenPostLoad);
1420 
1421             // If we managed to sink this instruction, update everything so we
1422             // harden that instruction when we reach it in the instruction
1423             // sequence.
1424             if (SunkMI != &MI) {
1425               // If in sinking there was no instruction needing to be hardened,
1426               // we're done.
1427               if (!SunkMI)
1428                 continue;
1429 
1430               // Otherwise, add this to the set of defs we harden.
1431               HardenPostLoad.insert(SunkMI);
1432               continue;
1433             }
1434           }
1435 
1436           unsigned HardenedReg = hardenPostLoad(MI);
1437 
1438           // Mark the resulting hardened register as such so we don't re-harden.
1439           AddrRegToHardenedReg[HardenedReg] = HardenedReg;
1440 
1441           continue;
1442         }
1443 
1444         // Check for an indirect call or branch that may need its input hardened
1445         // even if we couldn't find the specific load used, or were able to
1446         // avoid hardening it for some reason. Note that here we cannot break
1447         // out afterward as we may still need to handle any call aspect of this
1448         // instruction.
1449         if ((MI.isCall() || MI.isBranch()) && HardenIndirectCallsAndJumps)
1450           hardenIndirectCallOrJumpInstr(MI, AddrRegToHardenedReg);
1451       }
1452 
1453       // After we finish hardening loads we handle interprocedural hardening if
1454       // enabled and relevant for this instruction.
1455       if (!HardenInterprocedurally)
1456         continue;
1457       if (!MI.isCall() && !MI.isReturn())
1458         continue;
1459 
1460       // If this is a direct return (IE, not a tail call) just directly harden
1461       // it.
1462       if (MI.isReturn() && !MI.isCall()) {
1463         hardenReturnInstr(MI);
1464         continue;
1465       }
1466 
1467       // Otherwise we have a call. We need to handle transferring the predicate
1468       // state into a call and recovering it after the call returns (unless this
1469       // is a tail call).
1470       assert(MI.isCall() && "Should only reach here for calls!");
1471       tracePredStateThroughCall(MI);
1472     }
1473 
1474     HardenPostLoad.clear();
1475     HardenLoadAddr.clear();
1476     HardenedAddrRegs.clear();
1477     AddrRegToHardenedReg.clear();
1478 
1479     // Currently, we only track data-dependent loads within a basic block.
1480     // FIXME: We should see if this is necessary or if we could be more
1481     // aggressive here without opening up attack avenues.
1482     LoadDepRegs.clear();
1483   }
1484 }
1485 
1486 /// Save EFLAGS into the returned GPR. This can in turn be restored with
1487 /// `restoreEFLAGS`.
1488 ///
1489 /// Note that LLVM can only lower very simple patterns of saved and restored
1490 /// EFLAGS registers. The restore should always be within the same basic block
1491 /// as the save so that no PHI nodes are inserted.
1492 unsigned X86SpeculativeLoadHardeningPass::saveEFLAGS(
1493     MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertPt,
1494     const DebugLoc &Loc) {
1495   // FIXME: Hard coding this to a 32-bit register class seems weird, but matches
1496   // what instruction selection does.
1497   Register Reg = MRI->createVirtualRegister(&X86::GR32RegClass);
1498   // We directly copy the FLAGS register and rely on later lowering to clean
1499   // this up into the appropriate setCC instructions.
1500   BuildMI(MBB, InsertPt, Loc, TII->get(X86::COPY), Reg).addReg(X86::EFLAGS);
1501   ++NumInstsInserted;
1502   return Reg;
1503 }
1504 
1505 /// Restore EFLAGS from the provided GPR. This should be produced by
1506 /// `saveEFLAGS`.
1507 ///
1508 /// This must be done within the same basic block as the save in order to
1509 /// reliably lower.
1510 void X86SpeculativeLoadHardeningPass::restoreEFLAGS(
1511     MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertPt,
1512     const DebugLoc &Loc, Register Reg) {
1513   BuildMI(MBB, InsertPt, Loc, TII->get(X86::COPY), X86::EFLAGS).addReg(Reg);
1514   ++NumInstsInserted;
1515 }
1516 
1517 /// Takes the current predicate state (in a register) and merges it into the
1518 /// stack pointer. The state is essentially a single bit, but we merge this in
1519 /// a way that won't form non-canonical pointers and also will be preserved
1520 /// across normal stack adjustments.
1521 void X86SpeculativeLoadHardeningPass::mergePredStateIntoSP(
1522     MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertPt,
1523     const DebugLoc &Loc, unsigned PredStateReg) {
1524   Register TmpReg = MRI->createVirtualRegister(PS->RC);
1525   // FIXME: This hard codes a shift distance based on the number of bits needed
1526   // to stay canonical on 64-bit. We should compute this somehow and support
1527   // 32-bit as part of that.
1528   auto ShiftI = BuildMI(MBB, InsertPt, Loc, TII->get(X86::SHL64ri), TmpReg)
1529                     .addReg(PredStateReg, RegState::Kill)
1530                     .addImm(47);
1531   ShiftI->addRegisterDead(X86::EFLAGS, TRI);
1532   ++NumInstsInserted;
1533   auto OrI = BuildMI(MBB, InsertPt, Loc, TII->get(X86::OR64rr), X86::RSP)
1534                  .addReg(X86::RSP)
1535                  .addReg(TmpReg, RegState::Kill);
1536   OrI->addRegisterDead(X86::EFLAGS, TRI);
1537   ++NumInstsInserted;
1538 }
1539 
1540 /// Extracts the predicate state stored in the high bits of the stack pointer.
1541 unsigned X86SpeculativeLoadHardeningPass::extractPredStateFromSP(
1542     MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertPt,
1543     const DebugLoc &Loc) {
1544   Register PredStateReg = MRI->createVirtualRegister(PS->RC);
1545   Register TmpReg = MRI->createVirtualRegister(PS->RC);
1546 
1547   // We know that the stack pointer will have any preserved predicate state in
1548   // its high bit. We just want to smear this across the other bits. Turns out,
1549   // this is exactly what an arithmetic right shift does.
1550   BuildMI(MBB, InsertPt, Loc, TII->get(TargetOpcode::COPY), TmpReg)
1551       .addReg(X86::RSP);
1552   auto ShiftI =
1553       BuildMI(MBB, InsertPt, Loc, TII->get(X86::SAR64ri), PredStateReg)
1554           .addReg(TmpReg, RegState::Kill)
1555           .addImm(TRI->getRegSizeInBits(*PS->RC) - 1);
1556   ShiftI->addRegisterDead(X86::EFLAGS, TRI);
1557   ++NumInstsInserted;
1558 
1559   return PredStateReg;
1560 }
1561 
1562 void X86SpeculativeLoadHardeningPass::hardenLoadAddr(
1563     MachineInstr &MI, MachineOperand &BaseMO, MachineOperand &IndexMO,
1564     SmallDenseMap<unsigned, unsigned, 32> &AddrRegToHardenedReg) {
1565   MachineBasicBlock &MBB = *MI.getParent();
1566   const DebugLoc &Loc = MI.getDebugLoc();
1567 
1568   // Check if EFLAGS are alive by seeing if there is a def of them or they
1569   // live-in, and then seeing if that def is in turn used.
1570   bool EFLAGSLive = isEFLAGSLive(MBB, MI.getIterator(), *TRI);
1571 
1572   SmallVector<MachineOperand *, 2> HardenOpRegs;
1573 
1574   if (BaseMO.isFI()) {
1575     // A frame index is never a dynamically controllable load, so only
1576     // harden it if we're covering fixed address loads as well.
1577     LLVM_DEBUG(
1578         dbgs() << "  Skipping hardening base of explicit stack frame load: ";
1579         MI.dump(); dbgs() << "\n");
1580   } else if (BaseMO.getReg() == X86::RSP) {
1581     // Some idempotent atomic operations are lowered directly to a locked
1582     // OR with 0 to the top of stack(or slightly offset from top) which uses an
1583     // explicit RSP register as the base.
1584     assert(IndexMO.getReg() == X86::NoRegister &&
1585            "Explicit RSP access with dynamic index!");
1586     LLVM_DEBUG(
1587         dbgs() << "  Cannot harden base of explicit RSP offset in a load!");
1588   } else if (BaseMO.getReg() == X86::RIP ||
1589              BaseMO.getReg() == X86::NoRegister) {
1590     // For both RIP-relative addressed loads or absolute loads, we cannot
1591     // meaningfully harden them because the address being loaded has no
1592     // dynamic component.
1593     //
1594     // FIXME: When using a segment base (like TLS does) we end up with the
1595     // dynamic address being the base plus -1 because we can't mutate the
1596     // segment register here. This allows the signed 32-bit offset to point at
1597     // valid segment-relative addresses and load them successfully.
1598     LLVM_DEBUG(
1599         dbgs() << "  Cannot harden base of "
1600                << (BaseMO.getReg() == X86::RIP ? "RIP-relative" : "no-base")
1601                << " address in a load!");
1602   } else {
1603     assert(BaseMO.isReg() &&
1604            "Only allowed to have a frame index or register base.");
1605     HardenOpRegs.push_back(&BaseMO);
1606   }
1607 
1608   if (IndexMO.getReg() != X86::NoRegister &&
1609       (HardenOpRegs.empty() ||
1610        HardenOpRegs.front()->getReg() != IndexMO.getReg()))
1611     HardenOpRegs.push_back(&IndexMO);
1612 
1613   assert((HardenOpRegs.size() == 1 || HardenOpRegs.size() == 2) &&
1614          "Should have exactly one or two registers to harden!");
1615   assert((HardenOpRegs.size() == 1 ||
1616           HardenOpRegs[0]->getReg() != HardenOpRegs[1]->getReg()) &&
1617          "Should not have two of the same registers!");
1618 
1619   // Remove any registers that have alreaded been checked.
1620   llvm::erase_if(HardenOpRegs, [&](MachineOperand *Op) {
1621     // See if this operand's register has already been checked.
1622     auto It = AddrRegToHardenedReg.find(Op->getReg());
1623     if (It == AddrRegToHardenedReg.end())
1624       // Not checked, so retain this one.
1625       return false;
1626 
1627     // Otherwise, we can directly update this operand and remove it.
1628     Op->setReg(It->second);
1629     return true;
1630   });
1631   // If there are none left, we're done.
1632   if (HardenOpRegs.empty())
1633     return;
1634 
1635   // Compute the current predicate state.
1636   Register StateReg = PS->SSA.GetValueAtEndOfBlock(&MBB);
1637 
1638   auto InsertPt = MI.getIterator();
1639 
1640   // If EFLAGS are live and we don't have access to instructions that avoid
1641   // clobbering EFLAGS we need to save and restore them. This in turn makes
1642   // the EFLAGS no longer live.
1643   unsigned FlagsReg = 0;
1644   if (EFLAGSLive && !Subtarget->hasBMI2()) {
1645     EFLAGSLive = false;
1646     FlagsReg = saveEFLAGS(MBB, InsertPt, Loc);
1647   }
1648 
1649   for (MachineOperand *Op : HardenOpRegs) {
1650     Register OpReg = Op->getReg();
1651     auto *OpRC = MRI->getRegClass(OpReg);
1652     Register TmpReg = MRI->createVirtualRegister(OpRC);
1653 
1654     // If this is a vector register, we'll need somewhat custom logic to handle
1655     // hardening it.
1656     if (!Subtarget->hasVLX() && (OpRC->hasSuperClassEq(&X86::VR128RegClass) ||
1657                                  OpRC->hasSuperClassEq(&X86::VR256RegClass))) {
1658       assert(Subtarget->hasAVX2() && "AVX2-specific register classes!");
1659       bool Is128Bit = OpRC->hasSuperClassEq(&X86::VR128RegClass);
1660 
1661       // Move our state into a vector register.
1662       // FIXME: We could skip this at the cost of longer encodings with AVX-512
1663       // but that doesn't seem likely worth it.
1664       Register VStateReg = MRI->createVirtualRegister(&X86::VR128RegClass);
1665       auto MovI =
1666           BuildMI(MBB, InsertPt, Loc, TII->get(X86::VMOV64toPQIrr), VStateReg)
1667               .addReg(StateReg);
1668       (void)MovI;
1669       ++NumInstsInserted;
1670       LLVM_DEBUG(dbgs() << "  Inserting mov: "; MovI->dump(); dbgs() << "\n");
1671 
1672       // Broadcast it across the vector register.
1673       Register VBStateReg = MRI->createVirtualRegister(OpRC);
1674       auto BroadcastI = BuildMI(MBB, InsertPt, Loc,
1675                                 TII->get(Is128Bit ? X86::VPBROADCASTQrr
1676                                                   : X86::VPBROADCASTQYrr),
1677                                 VBStateReg)
1678                             .addReg(VStateReg);
1679       (void)BroadcastI;
1680       ++NumInstsInserted;
1681       LLVM_DEBUG(dbgs() << "  Inserting broadcast: "; BroadcastI->dump();
1682                  dbgs() << "\n");
1683 
1684       // Merge our potential poison state into the value with a vector or.
1685       auto OrI =
1686           BuildMI(MBB, InsertPt, Loc,
1687                   TII->get(Is128Bit ? X86::VPORrr : X86::VPORYrr), TmpReg)
1688               .addReg(VBStateReg)
1689               .addReg(OpReg);
1690       (void)OrI;
1691       ++NumInstsInserted;
1692       LLVM_DEBUG(dbgs() << "  Inserting or: "; OrI->dump(); dbgs() << "\n");
1693     } else if (OpRC->hasSuperClassEq(&X86::VR128XRegClass) ||
1694                OpRC->hasSuperClassEq(&X86::VR256XRegClass) ||
1695                OpRC->hasSuperClassEq(&X86::VR512RegClass)) {
1696       assert(Subtarget->hasAVX512() && "AVX512-specific register classes!");
1697       bool Is128Bit = OpRC->hasSuperClassEq(&X86::VR128XRegClass);
1698       bool Is256Bit = OpRC->hasSuperClassEq(&X86::VR256XRegClass);
1699       if (Is128Bit || Is256Bit)
1700         assert(Subtarget->hasVLX() && "AVX512VL-specific register classes!");
1701 
1702       // Broadcast our state into a vector register.
1703       Register VStateReg = MRI->createVirtualRegister(OpRC);
1704       unsigned BroadcastOp = Is128Bit ? X86::VPBROADCASTQrZ128rr
1705                                       : Is256Bit ? X86::VPBROADCASTQrZ256rr
1706                                                  : X86::VPBROADCASTQrZrr;
1707       auto BroadcastI =
1708           BuildMI(MBB, InsertPt, Loc, TII->get(BroadcastOp), VStateReg)
1709               .addReg(StateReg);
1710       (void)BroadcastI;
1711       ++NumInstsInserted;
1712       LLVM_DEBUG(dbgs() << "  Inserting broadcast: "; BroadcastI->dump();
1713                  dbgs() << "\n");
1714 
1715       // Merge our potential poison state into the value with a vector or.
1716       unsigned OrOp = Is128Bit ? X86::VPORQZ128rr
1717                                : Is256Bit ? X86::VPORQZ256rr : X86::VPORQZrr;
1718       auto OrI = BuildMI(MBB, InsertPt, Loc, TII->get(OrOp), TmpReg)
1719                      .addReg(VStateReg)
1720                      .addReg(OpReg);
1721       (void)OrI;
1722       ++NumInstsInserted;
1723       LLVM_DEBUG(dbgs() << "  Inserting or: "; OrI->dump(); dbgs() << "\n");
1724     } else {
1725       // FIXME: Need to support GR32 here for 32-bit code.
1726       assert(OpRC->hasSuperClassEq(&X86::GR64RegClass) &&
1727              "Not a supported register class for address hardening!");
1728 
1729       if (!EFLAGSLive) {
1730         // Merge our potential poison state into the value with an or.
1731         auto OrI = BuildMI(MBB, InsertPt, Loc, TII->get(X86::OR64rr), TmpReg)
1732                        .addReg(StateReg)
1733                        .addReg(OpReg);
1734         OrI->addRegisterDead(X86::EFLAGS, TRI);
1735         ++NumInstsInserted;
1736         LLVM_DEBUG(dbgs() << "  Inserting or: "; OrI->dump(); dbgs() << "\n");
1737       } else {
1738         // We need to avoid touching EFLAGS so shift out all but the least
1739         // significant bit using the instruction that doesn't update flags.
1740         auto ShiftI =
1741             BuildMI(MBB, InsertPt, Loc, TII->get(X86::SHRX64rr), TmpReg)
1742                 .addReg(OpReg)
1743                 .addReg(StateReg);
1744         (void)ShiftI;
1745         ++NumInstsInserted;
1746         LLVM_DEBUG(dbgs() << "  Inserting shrx: "; ShiftI->dump();
1747                    dbgs() << "\n");
1748       }
1749     }
1750 
1751     // Record this register as checked and update the operand.
1752     assert(!AddrRegToHardenedReg.count(Op->getReg()) &&
1753            "Should not have checked this register yet!");
1754     AddrRegToHardenedReg[Op->getReg()] = TmpReg;
1755     Op->setReg(TmpReg);
1756     ++NumAddrRegsHardened;
1757   }
1758 
1759   // And restore the flags if needed.
1760   if (FlagsReg)
1761     restoreEFLAGS(MBB, InsertPt, Loc, FlagsReg);
1762 }
1763 
1764 MachineInstr *X86SpeculativeLoadHardeningPass::sinkPostLoadHardenedInst(
1765     MachineInstr &InitialMI, SmallPtrSetImpl<MachineInstr *> &HardenedInstrs) {
1766   assert(X86InstrInfo::isDataInvariantLoad(InitialMI) &&
1767          "Cannot get here with a non-invariant load!");
1768   assert(!isEFLAGSDefLive(InitialMI) &&
1769          "Cannot get here with a data invariant load "
1770          "that interferes with EFLAGS!");
1771 
1772   // See if we can sink hardening the loaded value.
1773   auto SinkCheckToSingleUse =
1774       [&](MachineInstr &MI) -> std::optional<MachineInstr *> {
1775     Register DefReg = MI.getOperand(0).getReg();
1776 
1777     // We need to find a single use which we can sink the check. We can
1778     // primarily do this because many uses may already end up checked on their
1779     // own.
1780     MachineInstr *SingleUseMI = nullptr;
1781     for (MachineInstr &UseMI : MRI->use_instructions(DefReg)) {
1782       // If we're already going to harden this use, it is data invariant, it
1783       // does not interfere with EFLAGS, and within our block.
1784       if (HardenedInstrs.count(&UseMI)) {
1785         if (!X86InstrInfo::isDataInvariantLoad(UseMI) || isEFLAGSDefLive(UseMI)) {
1786           // If we've already decided to harden a non-load, we must have sunk
1787           // some other post-load hardened instruction to it and it must itself
1788           // be data-invariant.
1789           assert(X86InstrInfo::isDataInvariant(UseMI) &&
1790                  "Data variant instruction being hardened!");
1791           continue;
1792         }
1793 
1794         // Otherwise, this is a load and the load component can't be data
1795         // invariant so check how this register is being used.
1796         const int MemRefBeginIdx = X86::getFirstAddrOperandIdx(UseMI);
1797         assert(MemRefBeginIdx >= 0 &&
1798                "Should always have mem references here!");
1799 
1800         MachineOperand &BaseMO =
1801             UseMI.getOperand(MemRefBeginIdx + X86::AddrBaseReg);
1802         MachineOperand &IndexMO =
1803             UseMI.getOperand(MemRefBeginIdx + X86::AddrIndexReg);
1804         if ((BaseMO.isReg() && BaseMO.getReg() == DefReg) ||
1805             (IndexMO.isReg() && IndexMO.getReg() == DefReg))
1806           // The load uses the register as part of its address making it not
1807           // invariant.
1808           return {};
1809 
1810         continue;
1811       }
1812 
1813       if (SingleUseMI)
1814         // We already have a single use, this would make two. Bail.
1815         return {};
1816 
1817       // If this single use isn't data invariant, isn't in this block, or has
1818       // interfering EFLAGS, we can't sink the hardening to it.
1819       if (!X86InstrInfo::isDataInvariant(UseMI) || UseMI.getParent() != MI.getParent() ||
1820           isEFLAGSDefLive(UseMI))
1821         return {};
1822 
1823       // If this instruction defines multiple registers bail as we won't harden
1824       // all of them.
1825       if (UseMI.getDesc().getNumDefs() > 1)
1826         return {};
1827 
1828       // If this register isn't a virtual register we can't walk uses of sanely,
1829       // just bail. Also check that its register class is one of the ones we
1830       // can harden.
1831       Register UseDefReg = UseMI.getOperand(0).getReg();
1832       if (!canHardenRegister(UseDefReg))
1833         return {};
1834 
1835       SingleUseMI = &UseMI;
1836     }
1837 
1838     // If SingleUseMI is still null, there is no use that needs its own
1839     // checking. Otherwise, it is the single use that needs checking.
1840     return {SingleUseMI};
1841   };
1842 
1843   MachineInstr *MI = &InitialMI;
1844   while (std::optional<MachineInstr *> SingleUse = SinkCheckToSingleUse(*MI)) {
1845     // Update which MI we're checking now.
1846     MI = *SingleUse;
1847     if (!MI)
1848       break;
1849   }
1850 
1851   return MI;
1852 }
1853 
1854 bool X86SpeculativeLoadHardeningPass::canHardenRegister(Register Reg) {
1855   // We only support hardening virtual registers.
1856   if (!Reg.isVirtual())
1857     return false;
1858 
1859   auto *RC = MRI->getRegClass(Reg);
1860   int RegBytes = TRI->getRegSizeInBits(*RC) / 8;
1861   if (RegBytes > 8)
1862     // We don't support post-load hardening of vectors.
1863     return false;
1864 
1865   unsigned RegIdx = Log2_32(RegBytes);
1866   assert(RegIdx < 4 && "Unsupported register size");
1867 
1868   // If this register class is explicitly constrained to a class that doesn't
1869   // require REX prefix, we may not be able to satisfy that constraint when
1870   // emitting the hardening instructions, so bail out here.
1871   // FIXME: This seems like a pretty lame hack. The way this comes up is when we
1872   // end up both with a NOREX and REX-only register as operands to the hardening
1873   // instructions. It would be better to fix that code to handle this situation
1874   // rather than hack around it in this way.
1875   const TargetRegisterClass *NOREXRegClasses[] = {
1876       &X86::GR8_NOREXRegClass, &X86::GR16_NOREXRegClass,
1877       &X86::GR32_NOREXRegClass, &X86::GR64_NOREXRegClass};
1878   if (RC == NOREXRegClasses[RegIdx])
1879     return false;
1880 
1881   const TargetRegisterClass *GPRRegClasses[] = {
1882       &X86::GR8RegClass, &X86::GR16RegClass, &X86::GR32RegClass,
1883       &X86::GR64RegClass};
1884   return RC->hasSuperClassEq(GPRRegClasses[RegIdx]);
1885 }
1886 
1887 /// Harden a value in a register.
1888 ///
1889 /// This is the low-level logic to fully harden a value sitting in a register
1890 /// against leaking during speculative execution.
1891 ///
1892 /// Unlike hardening an address that is used by a load, this routine is required
1893 /// to hide *all* incoming bits in the register.
1894 ///
1895 /// `Reg` must be a virtual register. Currently, it is required to be a GPR no
1896 /// larger than the predicate state register. FIXME: We should support vector
1897 /// registers here by broadcasting the predicate state.
1898 ///
1899 /// The new, hardened virtual register is returned. It will have the same
1900 /// register class as `Reg`.
1901 unsigned X86SpeculativeLoadHardeningPass::hardenValueInRegister(
1902     Register Reg, MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertPt,
1903     const DebugLoc &Loc) {
1904   assert(canHardenRegister(Reg) && "Cannot harden this register!");
1905 
1906   auto *RC = MRI->getRegClass(Reg);
1907   int Bytes = TRI->getRegSizeInBits(*RC) / 8;
1908   Register StateReg = PS->SSA.GetValueAtEndOfBlock(&MBB);
1909   assert((Bytes == 1 || Bytes == 2 || Bytes == 4 || Bytes == 8) &&
1910          "Unknown register size");
1911 
1912   // FIXME: Need to teach this about 32-bit mode.
1913   if (Bytes != 8) {
1914     unsigned SubRegImms[] = {X86::sub_8bit, X86::sub_16bit, X86::sub_32bit};
1915     unsigned SubRegImm = SubRegImms[Log2_32(Bytes)];
1916     Register NarrowStateReg = MRI->createVirtualRegister(RC);
1917     BuildMI(MBB, InsertPt, Loc, TII->get(TargetOpcode::COPY), NarrowStateReg)
1918         .addReg(StateReg, 0, SubRegImm);
1919     StateReg = NarrowStateReg;
1920   }
1921 
1922   unsigned FlagsReg = 0;
1923   if (isEFLAGSLive(MBB, InsertPt, *TRI))
1924     FlagsReg = saveEFLAGS(MBB, InsertPt, Loc);
1925 
1926   Register NewReg = MRI->createVirtualRegister(RC);
1927   unsigned OrOpCodes[] = {X86::OR8rr, X86::OR16rr, X86::OR32rr, X86::OR64rr};
1928   unsigned OrOpCode = OrOpCodes[Log2_32(Bytes)];
1929   auto OrI = BuildMI(MBB, InsertPt, Loc, TII->get(OrOpCode), NewReg)
1930                  .addReg(StateReg)
1931                  .addReg(Reg);
1932   OrI->addRegisterDead(X86::EFLAGS, TRI);
1933   ++NumInstsInserted;
1934   LLVM_DEBUG(dbgs() << "  Inserting or: "; OrI->dump(); dbgs() << "\n");
1935 
1936   if (FlagsReg)
1937     restoreEFLAGS(MBB, InsertPt, Loc, FlagsReg);
1938 
1939   return NewReg;
1940 }
1941 
1942 /// Harden a load by hardening the loaded value in the defined register.
1943 ///
1944 /// We can harden a non-leaking load into a register without touching the
1945 /// address by just hiding all of the loaded bits during misspeculation. We use
1946 /// an `or` instruction to do this because we set up our poison value as all
1947 /// ones. And the goal is just for the loaded bits to not be exposed to
1948 /// execution and coercing them to one is sufficient.
1949 ///
1950 /// Returns the newly hardened register.
1951 unsigned X86SpeculativeLoadHardeningPass::hardenPostLoad(MachineInstr &MI) {
1952   MachineBasicBlock &MBB = *MI.getParent();
1953   const DebugLoc &Loc = MI.getDebugLoc();
1954 
1955   auto &DefOp = MI.getOperand(0);
1956   Register OldDefReg = DefOp.getReg();
1957   auto *DefRC = MRI->getRegClass(OldDefReg);
1958 
1959   // Because we want to completely replace the uses of this def'ed value with
1960   // the hardened value, create a dedicated new register that will only be used
1961   // to communicate the unhardened value to the hardening.
1962   Register UnhardenedReg = MRI->createVirtualRegister(DefRC);
1963   DefOp.setReg(UnhardenedReg);
1964 
1965   // Now harden this register's value, getting a hardened reg that is safe to
1966   // use. Note that we insert the instructions to compute this *after* the
1967   // defining instruction, not before it.
1968   unsigned HardenedReg = hardenValueInRegister(
1969       UnhardenedReg, MBB, std::next(MI.getIterator()), Loc);
1970 
1971   // Finally, replace the old register (which now only has the uses of the
1972   // original def) with the hardened register.
1973   MRI->replaceRegWith(/*FromReg*/ OldDefReg, /*ToReg*/ HardenedReg);
1974 
1975   ++NumPostLoadRegsHardened;
1976   return HardenedReg;
1977 }
1978 
1979 /// Harden a return instruction.
1980 ///
1981 /// Returns implicitly perform a load which we need to harden. Without hardening
1982 /// this load, an attacker my speculatively write over the return address to
1983 /// steer speculation of the return to an attacker controlled address. This is
1984 /// called Spectre v1.1 or Bounds Check Bypass Store (BCBS) and is described in
1985 /// this paper:
1986 /// https://people.csail.mit.edu/vlk/spectre11.pdf
1987 ///
1988 /// We can harden this by introducing an LFENCE that will delay any load of the
1989 /// return address until prior instructions have retired (and thus are not being
1990 /// speculated), or we can harden the address used by the implicit load: the
1991 /// stack pointer.
1992 ///
1993 /// If we are not using an LFENCE, hardening the stack pointer has an additional
1994 /// benefit: it allows us to pass the predicate state accumulated in this
1995 /// function back to the caller. In the absence of a BCBS attack on the return,
1996 /// the caller will typically be resumed and speculatively executed due to the
1997 /// Return Stack Buffer (RSB) prediction which is very accurate and has a high
1998 /// priority. It is possible that some code from the caller will be executed
1999 /// speculatively even during a BCBS-attacked return until the steering takes
2000 /// effect. Whenever this happens, the caller can recover the (poisoned)
2001 /// predicate state from the stack pointer and continue to harden loads.
2002 void X86SpeculativeLoadHardeningPass::hardenReturnInstr(MachineInstr &MI) {
2003   MachineBasicBlock &MBB = *MI.getParent();
2004   const DebugLoc &Loc = MI.getDebugLoc();
2005   auto InsertPt = MI.getIterator();
2006 
2007   if (FenceCallAndRet)
2008     // No need to fence here as we'll fence at the return site itself. That
2009     // handles more cases than we can handle here.
2010     return;
2011 
2012   // Take our predicate state, shift it to the high 17 bits (so that we keep
2013   // pointers canonical) and merge it into RSP. This will allow the caller to
2014   // extract it when we return (speculatively).
2015   mergePredStateIntoSP(MBB, InsertPt, Loc, PS->SSA.GetValueAtEndOfBlock(&MBB));
2016 }
2017 
2018 /// Trace the predicate state through a call.
2019 ///
2020 /// There are several layers of this needed to handle the full complexity of
2021 /// calls.
2022 ///
2023 /// First, we need to send the predicate state into the called function. We do
2024 /// this by merging it into the high bits of the stack pointer.
2025 ///
2026 /// For tail calls, this is all we need to do.
2027 ///
2028 /// For calls where we might return and resume the control flow, we need to
2029 /// extract the predicate state from the high bits of the stack pointer after
2030 /// control returns from the called function.
2031 ///
2032 /// We also need to verify that we intended to return to this location in the
2033 /// code. An attacker might arrange for the processor to mispredict the return
2034 /// to this valid but incorrect return address in the program rather than the
2035 /// correct one. See the paper on this attack, called "ret2spec" by the
2036 /// researchers, here:
2037 /// https://christian-rossow.de/publications/ret2spec-ccs2018.pdf
2038 ///
2039 /// The way we verify that we returned to the correct location is by preserving
2040 /// the expected return address across the call. One technique involves taking
2041 /// advantage of the red-zone to load the return address from `8(%rsp)` where it
2042 /// was left by the RET instruction when it popped `%rsp`. Alternatively, we can
2043 /// directly save the address into a register that will be preserved across the
2044 /// call. We compare this intended return address against the address
2045 /// immediately following the call (the observed return address). If these
2046 /// mismatch, we have detected misspeculation and can poison our predicate
2047 /// state.
2048 void X86SpeculativeLoadHardeningPass::tracePredStateThroughCall(
2049     MachineInstr &MI) {
2050   MachineBasicBlock &MBB = *MI.getParent();
2051   MachineFunction &MF = *MBB.getParent();
2052   auto InsertPt = MI.getIterator();
2053   const DebugLoc &Loc = MI.getDebugLoc();
2054 
2055   if (FenceCallAndRet) {
2056     if (MI.isReturn())
2057       // Tail call, we don't return to this function.
2058       // FIXME: We should also handle noreturn calls.
2059       return;
2060 
2061     // We don't need to fence before the call because the function should fence
2062     // in its entry. However, we do need to fence after the call returns.
2063     // Fencing before the return doesn't correctly handle cases where the return
2064     // itself is mispredicted.
2065     BuildMI(MBB, std::next(InsertPt), Loc, TII->get(X86::LFENCE));
2066     ++NumInstsInserted;
2067     ++NumLFENCEsInserted;
2068     return;
2069   }
2070 
2071   // First, we transfer the predicate state into the called function by merging
2072   // it into the stack pointer. This will kill the current def of the state.
2073   Register StateReg = PS->SSA.GetValueAtEndOfBlock(&MBB);
2074   mergePredStateIntoSP(MBB, InsertPt, Loc, StateReg);
2075 
2076   // If this call is also a return, it is a tail call and we don't need anything
2077   // else to handle it so just return. Also, if there are no further
2078   // instructions and no successors, this call does not return so we can also
2079   // bail.
2080   if (MI.isReturn() || (std::next(InsertPt) == MBB.end() && MBB.succ_empty()))
2081     return;
2082 
2083   // Create a symbol to track the return address and attach it to the call
2084   // machine instruction. We will lower extra symbols attached to call
2085   // instructions as label immediately following the call.
2086   MCSymbol *RetSymbol =
2087       MF.getContext().createTempSymbol("slh_ret_addr",
2088                                        /*AlwaysAddSuffix*/ true);
2089   MI.setPostInstrSymbol(MF, RetSymbol);
2090 
2091   const TargetRegisterClass *AddrRC = &X86::GR64RegClass;
2092   unsigned ExpectedRetAddrReg = 0;
2093 
2094   // If we have no red zones or if the function returns twice (possibly without
2095   // using the `ret` instruction) like setjmp, we need to save the expected
2096   // return address prior to the call.
2097   if (!Subtarget->getFrameLowering()->has128ByteRedZone(MF) ||
2098       MF.exposesReturnsTwice()) {
2099     // If we don't have red zones, we need to compute the expected return
2100     // address prior to the call and store it in a register that lives across
2101     // the call.
2102     //
2103     // In some ways, this is doubly satisfying as a mitigation because it will
2104     // also successfully detect stack smashing bugs in some cases (typically,
2105     // when a callee-saved register is used and the callee doesn't push it onto
2106     // the stack). But that isn't our primary goal, so we only use it as
2107     // a fallback.
2108     //
2109     // FIXME: It isn't clear that this is reliable in the face of
2110     // rematerialization in the register allocator. We somehow need to force
2111     // that to not occur for this particular instruction, and instead to spill
2112     // or otherwise preserve the value computed *prior* to the call.
2113     //
2114     // FIXME: It is even less clear why MachineCSE can't just fold this when we
2115     // end up having to use identical instructions both before and after the
2116     // call to feed the comparison.
2117     ExpectedRetAddrReg = MRI->createVirtualRegister(AddrRC);
2118     if (MF.getTarget().getCodeModel() == CodeModel::Small &&
2119         !Subtarget->isPositionIndependent()) {
2120       BuildMI(MBB, InsertPt, Loc, TII->get(X86::MOV64ri32), ExpectedRetAddrReg)
2121           .addSym(RetSymbol);
2122     } else {
2123       BuildMI(MBB, InsertPt, Loc, TII->get(X86::LEA64r), ExpectedRetAddrReg)
2124           .addReg(/*Base*/ X86::RIP)
2125           .addImm(/*Scale*/ 1)
2126           .addReg(/*Index*/ 0)
2127           .addSym(RetSymbol)
2128           .addReg(/*Segment*/ 0);
2129     }
2130   }
2131 
2132   // Step past the call to handle when it returns.
2133   ++InsertPt;
2134 
2135   // If we didn't pre-compute the expected return address into a register, then
2136   // red zones are enabled and the return address is still available on the
2137   // stack immediately after the call. As the very first instruction, we load it
2138   // into a register.
2139   if (!ExpectedRetAddrReg) {
2140     ExpectedRetAddrReg = MRI->createVirtualRegister(AddrRC);
2141     BuildMI(MBB, InsertPt, Loc, TII->get(X86::MOV64rm), ExpectedRetAddrReg)
2142         .addReg(/*Base*/ X86::RSP)
2143         .addImm(/*Scale*/ 1)
2144         .addReg(/*Index*/ 0)
2145         .addImm(/*Displacement*/ -8) // The stack pointer has been popped, so
2146                                      // the return address is 8-bytes past it.
2147         .addReg(/*Segment*/ 0);
2148   }
2149 
2150   // Now we extract the callee's predicate state from the stack pointer.
2151   unsigned NewStateReg = extractPredStateFromSP(MBB, InsertPt, Loc);
2152 
2153   // Test the expected return address against our actual address. If we can
2154   // form this basic block's address as an immediate, this is easy. Otherwise
2155   // we compute it.
2156   if (MF.getTarget().getCodeModel() == CodeModel::Small &&
2157       !Subtarget->isPositionIndependent()) {
2158     // FIXME: Could we fold this with the load? It would require careful EFLAGS
2159     // management.
2160     BuildMI(MBB, InsertPt, Loc, TII->get(X86::CMP64ri32))
2161         .addReg(ExpectedRetAddrReg, RegState::Kill)
2162         .addSym(RetSymbol);
2163   } else {
2164     Register ActualRetAddrReg = MRI->createVirtualRegister(AddrRC);
2165     BuildMI(MBB, InsertPt, Loc, TII->get(X86::LEA64r), ActualRetAddrReg)
2166         .addReg(/*Base*/ X86::RIP)
2167         .addImm(/*Scale*/ 1)
2168         .addReg(/*Index*/ 0)
2169         .addSym(RetSymbol)
2170         .addReg(/*Segment*/ 0);
2171     BuildMI(MBB, InsertPt, Loc, TII->get(X86::CMP64rr))
2172         .addReg(ExpectedRetAddrReg, RegState::Kill)
2173         .addReg(ActualRetAddrReg, RegState::Kill);
2174   }
2175 
2176   // Now conditionally update the predicate state we just extracted if we ended
2177   // up at a different return address than expected.
2178   int PredStateSizeInBytes = TRI->getRegSizeInBits(*PS->RC) / 8;
2179   auto CMovOp = X86::getCMovOpcode(PredStateSizeInBytes);
2180 
2181   Register UpdatedStateReg = MRI->createVirtualRegister(PS->RC);
2182   auto CMovI = BuildMI(MBB, InsertPt, Loc, TII->get(CMovOp), UpdatedStateReg)
2183                    .addReg(NewStateReg, RegState::Kill)
2184                    .addReg(PS->PoisonReg)
2185                    .addImm(X86::COND_NE);
2186   CMovI->findRegisterUseOperand(X86::EFLAGS, /*TRI=*/nullptr)->setIsKill(true);
2187   ++NumInstsInserted;
2188   LLVM_DEBUG(dbgs() << "  Inserting cmov: "; CMovI->dump(); dbgs() << "\n");
2189 
2190   PS->SSA.AddAvailableValue(&MBB, UpdatedStateReg);
2191 }
2192 
2193 /// An attacker may speculatively store over a value that is then speculatively
2194 /// loaded and used as the target of an indirect call or jump instruction. This
2195 /// is called Spectre v1.2 or Bounds Check Bypass Store (BCBS) and is described
2196 /// in this paper:
2197 /// https://people.csail.mit.edu/vlk/spectre11.pdf
2198 ///
2199 /// When this happens, the speculative execution of the call or jump will end up
2200 /// being steered to this attacker controlled address. While most such loads
2201 /// will be adequately hardened already, we want to ensure that they are
2202 /// definitively treated as needing post-load hardening. While address hardening
2203 /// is sufficient to prevent secret data from leaking to the attacker, it may
2204 /// not be sufficient to prevent an attacker from steering speculative
2205 /// execution. We forcibly unfolded all relevant loads above and so will always
2206 /// have an opportunity to post-load harden here, we just need to scan for cases
2207 /// not already flagged and add them.
2208 void X86SpeculativeLoadHardeningPass::hardenIndirectCallOrJumpInstr(
2209     MachineInstr &MI,
2210     SmallDenseMap<unsigned, unsigned, 32> &AddrRegToHardenedReg) {
2211   switch (MI.getOpcode()) {
2212   case X86::FARCALL16m:
2213   case X86::FARCALL32m:
2214   case X86::FARCALL64m:
2215   case X86::FARJMP16m:
2216   case X86::FARJMP32m:
2217   case X86::FARJMP64m:
2218     // We don't need to harden either far calls or far jumps as they are
2219     // safe from Spectre.
2220     return;
2221 
2222   default:
2223     break;
2224   }
2225 
2226   // We should never see a loading instruction at this point, as those should
2227   // have been unfolded.
2228   assert(!MI.mayLoad() && "Found a lingering loading instruction!");
2229 
2230   // If the first operand isn't a register, this is a branch or call
2231   // instruction with an immediate operand which doesn't need to be hardened.
2232   if (!MI.getOperand(0).isReg())
2233     return;
2234 
2235   // For all of these, the target register is the first operand of the
2236   // instruction.
2237   auto &TargetOp = MI.getOperand(0);
2238   Register OldTargetReg = TargetOp.getReg();
2239 
2240   // Try to lookup a hardened version of this register. We retain a reference
2241   // here as we want to update the map to track any newly computed hardened
2242   // register.
2243   unsigned &HardenedTargetReg = AddrRegToHardenedReg[OldTargetReg];
2244 
2245   // If we don't have a hardened register yet, compute one. Otherwise, just use
2246   // the already hardened register.
2247   //
2248   // FIXME: It is a little suspect that we use partially hardened registers that
2249   // only feed addresses. The complexity of partial hardening with SHRX
2250   // continues to pile up. Should definitively measure its value and consider
2251   // eliminating it.
2252   if (!HardenedTargetReg)
2253     HardenedTargetReg = hardenValueInRegister(
2254         OldTargetReg, *MI.getParent(), MI.getIterator(), MI.getDebugLoc());
2255 
2256   // Set the target operand to the hardened register.
2257   TargetOp.setReg(HardenedTargetReg);
2258 
2259   ++NumCallsOrJumpsHardened;
2260 }
2261 
2262 INITIALIZE_PASS_BEGIN(X86SpeculativeLoadHardeningPass, PASS_KEY,
2263                       "X86 speculative load hardener", false, false)
2264 INITIALIZE_PASS_END(X86SpeculativeLoadHardeningPass, PASS_KEY,
2265                     "X86 speculative load hardener", false, false)
2266 
2267 FunctionPass *llvm::createX86SpeculativeLoadHardeningPass() {
2268   return new X86SpeculativeLoadHardeningPass();
2269 }
2270