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 "X86InstrBuilder.h"
24 #include "X86InstrInfo.h"
25 #include "X86Subtarget.h"
26 #include "llvm/ADT/ArrayRef.h"
27 #include "llvm/ADT/DenseMap.h"
28 #include "llvm/ADT/Optional.h"
29 #include "llvm/ADT/STLExtras.h"
30 #include "llvm/ADT/ScopeExit.h"
31 #include "llvm/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/SmallSet.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/ADT/SparseBitVector.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/CodeGen/MachineBasicBlock.h"
37 #include "llvm/CodeGen/MachineConstantPool.h"
38 #include "llvm/CodeGen/MachineFunction.h"
39 #include "llvm/CodeGen/MachineFunctionPass.h"
40 #include "llvm/CodeGen/MachineInstr.h"
41 #include "llvm/CodeGen/MachineInstrBuilder.h"
42 #include "llvm/CodeGen/MachineModuleInfo.h"
43 #include "llvm/CodeGen/MachineOperand.h"
44 #include "llvm/CodeGen/MachineRegisterInfo.h"
45 #include "llvm/CodeGen/MachineSSAUpdater.h"
46 #include "llvm/CodeGen/TargetInstrInfo.h"
47 #include "llvm/CodeGen/TargetRegisterInfo.h"
48 #include "llvm/CodeGen/TargetSchedule.h"
49 #include "llvm/CodeGen/TargetSubtargetInfo.h"
50 #include "llvm/IR/DebugLoc.h"
51 #include "llvm/MC/MCSchedule.h"
52 #include "llvm/Pass.h"
53 #include "llvm/Support/CommandLine.h"
54 #include "llvm/Support/Debug.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Target/TargetMachine.h"
57 #include <algorithm>
58 #include <cassert>
59 #include <iterator>
60 #include <utility>
61
62 using namespace llvm;
63
64 #define PASS_KEY "x86-slh"
65 #define DEBUG_TYPE PASS_KEY
66
67 STATISTIC(NumCondBranchesTraced, "Number of conditional branches traced");
68 STATISTIC(NumBranchesUntraced, "Number of branches unable to trace");
69 STATISTIC(NumAddrRegsHardened,
70 "Number of address mode used registers hardaned");
71 STATISTIC(NumPostLoadRegsHardened,
72 "Number of post-load register values hardened");
73 STATISTIC(NumCallsOrJumpsHardened,
74 "Number of calls or jumps requiring extra hardening");
75 STATISTIC(NumInstsInserted, "Number of instructions inserted");
76 STATISTIC(NumLFENCEsInserted, "Number of lfence instructions inserted");
77
78 static cl::opt<bool> EnableSpeculativeLoadHardening(
79 "x86-speculative-load-hardening",
80 cl::desc("Force enable speculative load hardening"), cl::init(false),
81 cl::Hidden);
82
83 static cl::opt<bool> HardenEdgesWithLFENCE(
84 PASS_KEY "-lfence",
85 cl::desc(
86 "Use LFENCE along each conditional edge to harden against speculative "
87 "loads rather than conditional movs and poisoned pointers."),
88 cl::init(false), cl::Hidden);
89
90 static cl::opt<bool> EnablePostLoadHardening(
91 PASS_KEY "-post-load",
92 cl::desc("Harden the value loaded *after* it is loaded by "
93 "flushing the loaded bits to 1. This is hard to do "
94 "in general but can be done easily for GPRs."),
95 cl::init(true), cl::Hidden);
96
97 static cl::opt<bool> FenceCallAndRet(
98 PASS_KEY "-fence-call-and-ret",
99 cl::desc("Use a full speculation fence to harden both call and ret edges "
100 "rather than a lighter weight mitigation."),
101 cl::init(false), cl::Hidden);
102
103 static cl::opt<bool> HardenInterprocedurally(
104 PASS_KEY "-ip",
105 cl::desc("Harden interprocedurally by passing our state in and out of "
106 "functions in the high bits of the stack pointer."),
107 cl::init(true), cl::Hidden);
108
109 static cl::opt<bool>
110 HardenLoads(PASS_KEY "-loads",
111 cl::desc("Sanitize loads from memory. When disable, no "
112 "significant security is provided."),
113 cl::init(true), cl::Hidden);
114
115 static cl::opt<bool> HardenIndirectCallsAndJumps(
116 PASS_KEY "-indirect",
117 cl::desc("Harden indirect calls and jumps against using speculatively "
118 "stored attacker controlled addresses. This is designed to "
119 "mitigate Spectre v1.2 style attacks."),
120 cl::init(true), cl::Hidden);
121
122 namespace {
123
124 class X86SpeculativeLoadHardeningPass : public MachineFunctionPass {
125 public:
X86SpeculativeLoadHardeningPass()126 X86SpeculativeLoadHardeningPass() : MachineFunctionPass(ID) { }
127
getPassName() const128 StringRef getPassName() const override {
129 return "X86 speculative load hardening";
130 }
131 bool runOnMachineFunction(MachineFunction &MF) override;
132 void getAnalysisUsage(AnalysisUsage &AU) const override;
133
134 /// Pass identification, replacement for typeid.
135 static char ID;
136
137 private:
138 /// The information about a block's conditional terminators needed to trace
139 /// our predicate state through the exiting edges.
140 struct BlockCondInfo {
141 MachineBasicBlock *MBB;
142
143 // We mostly have one conditional branch, and in extremely rare cases have
144 // two. Three and more are so rare as to be unimportant for compile time.
145 SmallVector<MachineInstr *, 2> CondBrs;
146
147 MachineInstr *UncondBr;
148 };
149
150 /// Manages the predicate state traced through the program.
151 struct PredState {
152 unsigned InitialReg = 0;
153 unsigned PoisonReg = 0;
154
155 const TargetRegisterClass *RC;
156 MachineSSAUpdater SSA;
157
PredState__anon2dbf20440111::X86SpeculativeLoadHardeningPass::PredState158 PredState(MachineFunction &MF, const TargetRegisterClass *RC)
159 : RC(RC), SSA(MF) {}
160 };
161
162 const X86Subtarget *Subtarget = nullptr;
163 MachineRegisterInfo *MRI = nullptr;
164 const X86InstrInfo *TII = nullptr;
165 const TargetRegisterInfo *TRI = nullptr;
166
167 Optional<PredState> PS;
168
169 void hardenEdgesWithLFENCE(MachineFunction &MF);
170
171 SmallVector<BlockCondInfo, 16> collectBlockCondInfo(MachineFunction &MF);
172
173 SmallVector<MachineInstr *, 16>
174 tracePredStateThroughCFG(MachineFunction &MF, ArrayRef<BlockCondInfo> Infos);
175
176 void unfoldCallAndJumpLoads(MachineFunction &MF);
177
178 SmallVector<MachineInstr *, 16>
179 tracePredStateThroughIndirectBranches(MachineFunction &MF);
180
181 void tracePredStateThroughBlocksAndHarden(MachineFunction &MF);
182
183 unsigned saveEFLAGS(MachineBasicBlock &MBB,
184 MachineBasicBlock::iterator InsertPt, DebugLoc Loc);
185 void restoreEFLAGS(MachineBasicBlock &MBB,
186 MachineBasicBlock::iterator InsertPt, DebugLoc Loc,
187 unsigned OFReg);
188
189 void mergePredStateIntoSP(MachineBasicBlock &MBB,
190 MachineBasicBlock::iterator InsertPt, DebugLoc Loc,
191 unsigned PredStateReg);
192 unsigned extractPredStateFromSP(MachineBasicBlock &MBB,
193 MachineBasicBlock::iterator InsertPt,
194 DebugLoc Loc);
195
196 void
197 hardenLoadAddr(MachineInstr &MI, MachineOperand &BaseMO,
198 MachineOperand &IndexMO,
199 SmallDenseMap<unsigned, unsigned, 32> &AddrRegToHardenedReg);
200 MachineInstr *
201 sinkPostLoadHardenedInst(MachineInstr &MI,
202 SmallPtrSetImpl<MachineInstr *> &HardenedInstrs);
203 bool canHardenRegister(unsigned Reg);
204 unsigned hardenValueInRegister(unsigned Reg, MachineBasicBlock &MBB,
205 MachineBasicBlock::iterator InsertPt,
206 DebugLoc Loc);
207 unsigned hardenPostLoad(MachineInstr &MI);
208 void hardenReturnInstr(MachineInstr &MI);
209 void tracePredStateThroughCall(MachineInstr &MI);
210 void hardenIndirectCallOrJumpInstr(
211 MachineInstr &MI,
212 SmallDenseMap<unsigned, unsigned, 32> &AddrRegToHardenedReg);
213 };
214
215 } // end anonymous namespace
216
217 char X86SpeculativeLoadHardeningPass::ID = 0;
218
getAnalysisUsage(AnalysisUsage & AU) const219 void X86SpeculativeLoadHardeningPass::getAnalysisUsage(
220 AnalysisUsage &AU) const {
221 MachineFunctionPass::getAnalysisUsage(AU);
222 }
223
splitEdge(MachineBasicBlock & MBB,MachineBasicBlock & Succ,int SuccCount,MachineInstr * Br,MachineInstr * & UncondBr,const X86InstrInfo & TII)224 static MachineBasicBlock &splitEdge(MachineBasicBlock &MBB,
225 MachineBasicBlock &Succ, int SuccCount,
226 MachineInstr *Br, MachineInstr *&UncondBr,
227 const X86InstrInfo &TII) {
228 assert(!Succ.isEHPad() && "Shouldn't get edges to EH pads!");
229
230 MachineFunction &MF = *MBB.getParent();
231
232 MachineBasicBlock &NewMBB = *MF.CreateMachineBasicBlock();
233
234 // We have to insert the new block immediately after the current one as we
235 // don't know what layout-successor relationships the successor has and we
236 // may not be able to (and generally don't want to) try to fix those up.
237 MF.insert(std::next(MachineFunction::iterator(&MBB)), &NewMBB);
238
239 // Update the branch instruction if necessary.
240 if (Br) {
241 assert(Br->getOperand(0).getMBB() == &Succ &&
242 "Didn't start with the right target!");
243 Br->getOperand(0).setMBB(&NewMBB);
244
245 // If this successor was reached through a branch rather than fallthrough,
246 // we might have *broken* fallthrough and so need to inject a new
247 // unconditional branch.
248 if (!UncondBr) {
249 MachineBasicBlock &OldLayoutSucc =
250 *std::next(MachineFunction::iterator(&NewMBB));
251 assert(MBB.isSuccessor(&OldLayoutSucc) &&
252 "Without an unconditional branch, the old layout successor should "
253 "be an actual successor!");
254 auto BrBuilder =
255 BuildMI(&MBB, DebugLoc(), TII.get(X86::JMP_1)).addMBB(&OldLayoutSucc);
256 // Update the unconditional branch now that we've added one.
257 UncondBr = &*BrBuilder;
258 }
259
260 // Insert unconditional "jump Succ" instruction in the new block if
261 // necessary.
262 if (!NewMBB.isLayoutSuccessor(&Succ)) {
263 SmallVector<MachineOperand, 4> Cond;
264 TII.insertBranch(NewMBB, &Succ, nullptr, Cond, Br->getDebugLoc());
265 }
266 } else {
267 assert(!UncondBr &&
268 "Cannot have a branchless successor and an unconditional branch!");
269 assert(NewMBB.isLayoutSuccessor(&Succ) &&
270 "A non-branch successor must have been a layout successor before "
271 "and now is a layout successor of the new block.");
272 }
273
274 // If this is the only edge to the successor, we can just replace it in the
275 // CFG. Otherwise we need to add a new entry in the CFG for the new
276 // successor.
277 if (SuccCount == 1) {
278 MBB.replaceSuccessor(&Succ, &NewMBB);
279 } else {
280 MBB.splitSuccessor(&Succ, &NewMBB);
281 }
282
283 // Hook up the edge from the new basic block to the old successor in the CFG.
284 NewMBB.addSuccessor(&Succ);
285
286 // Fix PHI nodes in Succ so they refer to NewMBB instead of MBB.
287 for (MachineInstr &MI : Succ) {
288 if (!MI.isPHI())
289 break;
290 for (int OpIdx = 1, NumOps = MI.getNumOperands(); OpIdx < NumOps;
291 OpIdx += 2) {
292 MachineOperand &OpV = MI.getOperand(OpIdx);
293 MachineOperand &OpMBB = MI.getOperand(OpIdx + 1);
294 assert(OpMBB.isMBB() && "Block operand to a PHI is not a block!");
295 if (OpMBB.getMBB() != &MBB)
296 continue;
297
298 // If this is the last edge to the succesor, just replace MBB in the PHI
299 if (SuccCount == 1) {
300 OpMBB.setMBB(&NewMBB);
301 break;
302 }
303
304 // Otherwise, append a new pair of operands for the new incoming edge.
305 MI.addOperand(MF, OpV);
306 MI.addOperand(MF, MachineOperand::CreateMBB(&NewMBB));
307 break;
308 }
309 }
310
311 // Inherit live-ins from the successor
312 for (auto &LI : Succ.liveins())
313 NewMBB.addLiveIn(LI);
314
315 LLVM_DEBUG(dbgs() << " Split edge from '" << MBB.getName() << "' to '"
316 << Succ.getName() << "'.\n");
317 return NewMBB;
318 }
319
320 /// Removing duplicate PHI operands to leave the PHI in a canonical and
321 /// predictable form.
322 ///
323 /// FIXME: It's really frustrating that we have to do this, but SSA-form in MIR
324 /// isn't what you might expect. We may have multiple entries in PHI nodes for
325 /// a single predecessor. This makes CFG-updating extremely complex, so here we
326 /// simplify all PHI nodes to a model even simpler than the IR's model: exactly
327 /// one entry per predecessor, regardless of how many edges there are.
canonicalizePHIOperands(MachineFunction & MF)328 static void canonicalizePHIOperands(MachineFunction &MF) {
329 SmallPtrSet<MachineBasicBlock *, 4> Preds;
330 SmallVector<int, 4> DupIndices;
331 for (auto &MBB : MF)
332 for (auto &MI : MBB) {
333 if (!MI.isPHI())
334 break;
335
336 // First we scan the operands of the PHI looking for duplicate entries
337 // a particular predecessor. We retain the operand index of each duplicate
338 // entry found.
339 for (int OpIdx = 1, NumOps = MI.getNumOperands(); OpIdx < NumOps;
340 OpIdx += 2)
341 if (!Preds.insert(MI.getOperand(OpIdx + 1).getMBB()).second)
342 DupIndices.push_back(OpIdx);
343
344 // Now walk the duplicate indices, removing both the block and value. Note
345 // that these are stored as a vector making this element-wise removal
346 // :w
347 // potentially quadratic.
348 //
349 // FIXME: It is really frustrating that we have to use a quadratic
350 // removal algorithm here. There should be a better way, but the use-def
351 // updates required make that impossible using the public API.
352 //
353 // Note that we have to process these backwards so that we don't
354 // invalidate other indices with each removal.
355 while (!DupIndices.empty()) {
356 int OpIdx = DupIndices.pop_back_val();
357 // Remove both the block and value operand, again in reverse order to
358 // preserve indices.
359 MI.RemoveOperand(OpIdx + 1);
360 MI.RemoveOperand(OpIdx);
361 }
362
363 Preds.clear();
364 }
365 }
366
367 /// Helper to scan a function for loads vulnerable to misspeculation that we
368 /// want to harden.
369 ///
370 /// We use this to avoid making changes to functions where there is nothing we
371 /// need to do to harden against misspeculation.
hasVulnerableLoad(MachineFunction & MF)372 static bool hasVulnerableLoad(MachineFunction &MF) {
373 for (MachineBasicBlock &MBB : MF) {
374 for (MachineInstr &MI : MBB) {
375 // Loads within this basic block after an LFENCE are not at risk of
376 // speculatively executing with invalid predicates from prior control
377 // flow. So break out of this block but continue scanning the function.
378 if (MI.getOpcode() == X86::LFENCE)
379 break;
380
381 // Looking for loads only.
382 if (!MI.mayLoad())
383 continue;
384
385 // An MFENCE is modeled as a load but isn't vulnerable to misspeculation.
386 if (MI.getOpcode() == X86::MFENCE)
387 continue;
388
389 // We found a load.
390 return true;
391 }
392 }
393
394 // No loads found.
395 return false;
396 }
397
runOnMachineFunction(MachineFunction & MF)398 bool X86SpeculativeLoadHardeningPass::runOnMachineFunction(
399 MachineFunction &MF) {
400 LLVM_DEBUG(dbgs() << "********** " << getPassName() << " : " << MF.getName()
401 << " **********\n");
402
403 // Only run if this pass is forced enabled or we detect the relevant function
404 // attribute requesting SLH.
405 if (!EnableSpeculativeLoadHardening &&
406 !MF.getFunction().hasFnAttribute(Attribute::SpeculativeLoadHardening))
407 return false;
408
409 Subtarget = &MF.getSubtarget<X86Subtarget>();
410 MRI = &MF.getRegInfo();
411 TII = Subtarget->getInstrInfo();
412 TRI = Subtarget->getRegisterInfo();
413
414 // FIXME: Support for 32-bit.
415 PS.emplace(MF, &X86::GR64_NOSPRegClass);
416
417 if (MF.begin() == MF.end())
418 // Nothing to do for a degenerate empty function...
419 return false;
420
421 // We support an alternative hardening technique based on a debug flag.
422 if (HardenEdgesWithLFENCE) {
423 hardenEdgesWithLFENCE(MF);
424 return true;
425 }
426
427 // Create a dummy debug loc to use for all the generated code here.
428 DebugLoc Loc;
429
430 MachineBasicBlock &Entry = *MF.begin();
431 auto EntryInsertPt = Entry.SkipPHIsLabelsAndDebug(Entry.begin());
432
433 // Do a quick scan to see if we have any checkable loads.
434 bool HasVulnerableLoad = hasVulnerableLoad(MF);
435
436 // See if we have any conditional branching blocks that we will need to trace
437 // predicate state through.
438 SmallVector<BlockCondInfo, 16> Infos = collectBlockCondInfo(MF);
439
440 // If we have no interesting conditions or loads, nothing to do here.
441 if (!HasVulnerableLoad && Infos.empty())
442 return true;
443
444 // The poison value is required to be an all-ones value for many aspects of
445 // this mitigation.
446 const int PoisonVal = -1;
447 PS->PoisonReg = MRI->createVirtualRegister(PS->RC);
448 BuildMI(Entry, EntryInsertPt, Loc, TII->get(X86::MOV64ri32), PS->PoisonReg)
449 .addImm(PoisonVal);
450 ++NumInstsInserted;
451
452 // If we have loads being hardened and we've asked for call and ret edges to
453 // get a full fence-based mitigation, inject that fence.
454 if (HasVulnerableLoad && FenceCallAndRet) {
455 // We need to insert an LFENCE at the start of the function to suspend any
456 // incoming misspeculation from the caller. This helps two-fold: the caller
457 // may not have been protected as this code has been, and this code gets to
458 // not take any specific action to protect across calls.
459 // FIXME: We could skip this for functions which unconditionally return
460 // a constant.
461 BuildMI(Entry, EntryInsertPt, Loc, TII->get(X86::LFENCE));
462 ++NumInstsInserted;
463 ++NumLFENCEsInserted;
464 }
465
466 // If we guarded the entry with an LFENCE and have no conditionals to protect
467 // in blocks, then we're done.
468 if (FenceCallAndRet && Infos.empty())
469 // We may have changed the function's code at this point to insert fences.
470 return true;
471
472 // For every basic block in the function which can b
473 if (HardenInterprocedurally && !FenceCallAndRet) {
474 // Set up the predicate state by extracting it from the incoming stack
475 // pointer so we pick up any misspeculation in our caller.
476 PS->InitialReg = extractPredStateFromSP(Entry, EntryInsertPt, Loc);
477 } else {
478 // Otherwise, just build the predicate state itself by zeroing a register
479 // as we don't need any initial state.
480 PS->InitialReg = MRI->createVirtualRegister(PS->RC);
481 Register PredStateSubReg = MRI->createVirtualRegister(&X86::GR32RegClass);
482 auto ZeroI = BuildMI(Entry, EntryInsertPt, Loc, TII->get(X86::MOV32r0),
483 PredStateSubReg);
484 ++NumInstsInserted;
485 MachineOperand *ZeroEFLAGSDefOp =
486 ZeroI->findRegisterDefOperand(X86::EFLAGS);
487 assert(ZeroEFLAGSDefOp && ZeroEFLAGSDefOp->isImplicit() &&
488 "Must have an implicit def of EFLAGS!");
489 ZeroEFLAGSDefOp->setIsDead(true);
490 BuildMI(Entry, EntryInsertPt, Loc, TII->get(X86::SUBREG_TO_REG),
491 PS->InitialReg)
492 .addImm(0)
493 .addReg(PredStateSubReg)
494 .addImm(X86::sub_32bit);
495 }
496
497 // We're going to need to trace predicate state throughout the function's
498 // CFG. Prepare for this by setting up our initial state of PHIs with unique
499 // predecessor entries and all the initial predicate state.
500 canonicalizePHIOperands(MF);
501
502 // Track the updated values in an SSA updater to rewrite into SSA form at the
503 // end.
504 PS->SSA.Initialize(PS->InitialReg);
505 PS->SSA.AddAvailableValue(&Entry, PS->InitialReg);
506
507 // Trace through the CFG.
508 auto CMovs = tracePredStateThroughCFG(MF, Infos);
509
510 // We may also enter basic blocks in this function via exception handling
511 // control flow. Here, if we are hardening interprocedurally, we need to
512 // re-capture the predicate state from the throwing code. In the Itanium ABI,
513 // the throw will always look like a call to __cxa_throw and will have the
514 // predicate state in the stack pointer, so extract fresh predicate state from
515 // the stack pointer and make it available in SSA.
516 // FIXME: Handle non-itanium ABI EH models.
517 if (HardenInterprocedurally) {
518 for (MachineBasicBlock &MBB : MF) {
519 assert(!MBB.isEHScopeEntry() && "Only Itanium ABI EH supported!");
520 assert(!MBB.isEHFuncletEntry() && "Only Itanium ABI EH supported!");
521 assert(!MBB.isCleanupFuncletEntry() && "Only Itanium ABI EH supported!");
522 if (!MBB.isEHPad())
523 continue;
524 PS->SSA.AddAvailableValue(
525 &MBB,
526 extractPredStateFromSP(MBB, MBB.SkipPHIsAndLabels(MBB.begin()), Loc));
527 }
528 }
529
530 if (HardenIndirectCallsAndJumps) {
531 // If we are going to harden calls and jumps we need to unfold their memory
532 // operands.
533 unfoldCallAndJumpLoads(MF);
534
535 // Then we trace predicate state through the indirect branches.
536 auto IndirectBrCMovs = tracePredStateThroughIndirectBranches(MF);
537 CMovs.append(IndirectBrCMovs.begin(), IndirectBrCMovs.end());
538 }
539
540 // Now that we have the predicate state available at the start of each block
541 // in the CFG, trace it through each block, hardening vulnerable instructions
542 // as we go.
543 tracePredStateThroughBlocksAndHarden(MF);
544
545 // Now rewrite all the uses of the pred state using the SSA updater to insert
546 // PHIs connecting the state between blocks along the CFG edges.
547 for (MachineInstr *CMovI : CMovs)
548 for (MachineOperand &Op : CMovI->operands()) {
549 if (!Op.isReg() || Op.getReg() != PS->InitialReg)
550 continue;
551
552 PS->SSA.RewriteUse(Op);
553 }
554
555 LLVM_DEBUG(dbgs() << "Final speculative load hardened function:\n"; MF.dump();
556 dbgs() << "\n"; MF.verify(this));
557 return true;
558 }
559
560 /// Implements the naive hardening approach of putting an LFENCE after every
561 /// potentially mis-predicted control flow construct.
562 ///
563 /// We include this as an alternative mostly for the purpose of comparison. The
564 /// performance impact of this is expected to be extremely severe and not
565 /// practical for any real-world users.
hardenEdgesWithLFENCE(MachineFunction & MF)566 void X86SpeculativeLoadHardeningPass::hardenEdgesWithLFENCE(
567 MachineFunction &MF) {
568 // First, we scan the function looking for blocks that are reached along edges
569 // that we might want to harden.
570 SmallSetVector<MachineBasicBlock *, 8> Blocks;
571 for (MachineBasicBlock &MBB : MF) {
572 // If there are no or only one successor, nothing to do here.
573 if (MBB.succ_size() <= 1)
574 continue;
575
576 // Skip blocks unless their terminators start with a branch. Other
577 // terminators don't seem interesting for guarding against misspeculation.
578 auto TermIt = MBB.getFirstTerminator();
579 if (TermIt == MBB.end() || !TermIt->isBranch())
580 continue;
581
582 // Add all the non-EH-pad succossors to the blocks we want to harden. We
583 // skip EH pads because there isn't really a condition of interest on
584 // entering.
585 for (MachineBasicBlock *SuccMBB : MBB.successors())
586 if (!SuccMBB->isEHPad())
587 Blocks.insert(SuccMBB);
588 }
589
590 for (MachineBasicBlock *MBB : Blocks) {
591 auto InsertPt = MBB->SkipPHIsAndLabels(MBB->begin());
592 BuildMI(*MBB, InsertPt, DebugLoc(), TII->get(X86::LFENCE));
593 ++NumInstsInserted;
594 ++NumLFENCEsInserted;
595 }
596 }
597
598 SmallVector<X86SpeculativeLoadHardeningPass::BlockCondInfo, 16>
collectBlockCondInfo(MachineFunction & MF)599 X86SpeculativeLoadHardeningPass::collectBlockCondInfo(MachineFunction &MF) {
600 SmallVector<BlockCondInfo, 16> Infos;
601
602 // Walk the function and build up a summary for each block's conditions that
603 // we need to trace through.
604 for (MachineBasicBlock &MBB : MF) {
605 // If there are no or only one successor, nothing to do here.
606 if (MBB.succ_size() <= 1)
607 continue;
608
609 // We want to reliably handle any conditional branch terminators in the
610 // MBB, so we manually analyze the branch. We can handle all of the
611 // permutations here, including ones that analyze branch cannot.
612 //
613 // The approach is to walk backwards across the terminators, resetting at
614 // any unconditional non-indirect branch, and track all conditional edges
615 // to basic blocks as well as the fallthrough or unconditional successor
616 // edge. For each conditional edge, we track the target and the opposite
617 // condition code in order to inject a "no-op" cmov into that successor
618 // that will harden the predicate. For the fallthrough/unconditional
619 // edge, we inject a separate cmov for each conditional branch with
620 // matching condition codes. This effectively implements an "and" of the
621 // condition flags, even if there isn't a single condition flag that would
622 // directly implement that. We don't bother trying to optimize either of
623 // these cases because if such an optimization is possible, LLVM should
624 // have optimized the conditional *branches* in that way already to reduce
625 // instruction count. This late, we simply assume the minimal number of
626 // branch instructions is being emitted and use that to guide our cmov
627 // insertion.
628
629 BlockCondInfo Info = {&MBB, {}, nullptr};
630
631 // Now walk backwards through the terminators and build up successors they
632 // reach and the conditions.
633 for (MachineInstr &MI : llvm::reverse(MBB)) {
634 // Once we've handled all the terminators, we're done.
635 if (!MI.isTerminator())
636 break;
637
638 // If we see a non-branch terminator, we can't handle anything so bail.
639 if (!MI.isBranch()) {
640 Info.CondBrs.clear();
641 break;
642 }
643
644 // If we see an unconditional branch, reset our state, clear any
645 // fallthrough, and set this is the "else" successor.
646 if (MI.getOpcode() == X86::JMP_1) {
647 Info.CondBrs.clear();
648 Info.UncondBr = &MI;
649 continue;
650 }
651
652 // If we get an invalid condition, we have an indirect branch or some
653 // other unanalyzable "fallthrough" case. We model this as a nullptr for
654 // the destination so we can still guard any conditional successors.
655 // Consider code sequences like:
656 // ```
657 // jCC L1
658 // jmpq *%rax
659 // ```
660 // We still want to harden the edge to `L1`.
661 if (X86::getCondFromBranch(MI) == X86::COND_INVALID) {
662 Info.CondBrs.clear();
663 Info.UncondBr = &MI;
664 continue;
665 }
666
667 // We have a vanilla conditional branch, add it to our list.
668 Info.CondBrs.push_back(&MI);
669 }
670 if (Info.CondBrs.empty()) {
671 ++NumBranchesUntraced;
672 LLVM_DEBUG(dbgs() << "WARNING: unable to secure successors of block:\n";
673 MBB.dump());
674 continue;
675 }
676
677 Infos.push_back(Info);
678 }
679
680 return Infos;
681 }
682
683 /// Trace the predicate state through the CFG, instrumenting each conditional
684 /// branch such that misspeculation through an edge will poison the predicate
685 /// state.
686 ///
687 /// Returns the list of inserted CMov instructions so that they can have their
688 /// uses of the predicate state rewritten into proper SSA form once it is
689 /// complete.
690 SmallVector<MachineInstr *, 16>
tracePredStateThroughCFG(MachineFunction & MF,ArrayRef<BlockCondInfo> Infos)691 X86SpeculativeLoadHardeningPass::tracePredStateThroughCFG(
692 MachineFunction &MF, ArrayRef<BlockCondInfo> Infos) {
693 // Collect the inserted cmov instructions so we can rewrite their uses of the
694 // predicate state into SSA form.
695 SmallVector<MachineInstr *, 16> CMovs;
696
697 // Now walk all of the basic blocks looking for ones that end in conditional
698 // jumps where we need to update this register along each edge.
699 for (const BlockCondInfo &Info : Infos) {
700 MachineBasicBlock &MBB = *Info.MBB;
701 const SmallVectorImpl<MachineInstr *> &CondBrs = Info.CondBrs;
702 MachineInstr *UncondBr = Info.UncondBr;
703
704 LLVM_DEBUG(dbgs() << "Tracing predicate through block: " << MBB.getName()
705 << "\n");
706 ++NumCondBranchesTraced;
707
708 // Compute the non-conditional successor as either the target of any
709 // unconditional branch or the layout successor.
710 MachineBasicBlock *UncondSucc =
711 UncondBr ? (UncondBr->getOpcode() == X86::JMP_1
712 ? UncondBr->getOperand(0).getMBB()
713 : nullptr)
714 : &*std::next(MachineFunction::iterator(&MBB));
715
716 // Count how many edges there are to any given successor.
717 SmallDenseMap<MachineBasicBlock *, int> SuccCounts;
718 if (UncondSucc)
719 ++SuccCounts[UncondSucc];
720 for (auto *CondBr : CondBrs)
721 ++SuccCounts[CondBr->getOperand(0).getMBB()];
722
723 // A lambda to insert cmov instructions into a block checking all of the
724 // condition codes in a sequence.
725 auto BuildCheckingBlockForSuccAndConds =
726 [&](MachineBasicBlock &MBB, MachineBasicBlock &Succ, int SuccCount,
727 MachineInstr *Br, MachineInstr *&UncondBr,
728 ArrayRef<X86::CondCode> Conds) {
729 // First, we split the edge to insert the checking block into a safe
730 // location.
731 auto &CheckingMBB =
732 (SuccCount == 1 && Succ.pred_size() == 1)
733 ? Succ
734 : splitEdge(MBB, Succ, SuccCount, Br, UncondBr, *TII);
735
736 bool LiveEFLAGS = Succ.isLiveIn(X86::EFLAGS);
737 if (!LiveEFLAGS)
738 CheckingMBB.addLiveIn(X86::EFLAGS);
739
740 // Now insert the cmovs to implement the checks.
741 auto InsertPt = CheckingMBB.begin();
742 assert((InsertPt == CheckingMBB.end() || !InsertPt->isPHI()) &&
743 "Should never have a PHI in the initial checking block as it "
744 "always has a single predecessor!");
745
746 // We will wire each cmov to each other, but need to start with the
747 // incoming pred state.
748 unsigned CurStateReg = PS->InitialReg;
749
750 for (X86::CondCode Cond : Conds) {
751 int PredStateSizeInBytes = TRI->getRegSizeInBits(*PS->RC) / 8;
752 auto CMovOp = X86::getCMovOpcode(PredStateSizeInBytes);
753
754 Register UpdatedStateReg = MRI->createVirtualRegister(PS->RC);
755 // Note that we intentionally use an empty debug location so that
756 // this picks up the preceding location.
757 auto CMovI = BuildMI(CheckingMBB, InsertPt, DebugLoc(),
758 TII->get(CMovOp), UpdatedStateReg)
759 .addReg(CurStateReg)
760 .addReg(PS->PoisonReg)
761 .addImm(Cond);
762 // If this is the last cmov and the EFLAGS weren't originally
763 // live-in, mark them as killed.
764 if (!LiveEFLAGS && Cond == Conds.back())
765 CMovI->findRegisterUseOperand(X86::EFLAGS)->setIsKill(true);
766
767 ++NumInstsInserted;
768 LLVM_DEBUG(dbgs() << " Inserting cmov: "; CMovI->dump();
769 dbgs() << "\n");
770
771 // The first one of the cmovs will be using the top level
772 // `PredStateReg` and need to get rewritten into SSA form.
773 if (CurStateReg == PS->InitialReg)
774 CMovs.push_back(&*CMovI);
775
776 // The next cmov should start from this one's def.
777 CurStateReg = UpdatedStateReg;
778 }
779
780 // And put the last one into the available values for SSA form of our
781 // predicate state.
782 PS->SSA.AddAvailableValue(&CheckingMBB, CurStateReg);
783 };
784
785 std::vector<X86::CondCode> UncondCodeSeq;
786 for (auto *CondBr : CondBrs) {
787 MachineBasicBlock &Succ = *CondBr->getOperand(0).getMBB();
788 int &SuccCount = SuccCounts[&Succ];
789
790 X86::CondCode Cond = X86::getCondFromBranch(*CondBr);
791 X86::CondCode InvCond = X86::GetOppositeBranchCondition(Cond);
792 UncondCodeSeq.push_back(Cond);
793
794 BuildCheckingBlockForSuccAndConds(MBB, Succ, SuccCount, CondBr, UncondBr,
795 {InvCond});
796
797 // Decrement the successor count now that we've split one of the edges.
798 // We need to keep the count of edges to the successor accurate in order
799 // to know above when to *replace* the successor in the CFG vs. just
800 // adding the new successor.
801 --SuccCount;
802 }
803
804 // Since we may have split edges and changed the number of successors,
805 // normalize the probabilities. This avoids doing it each time we split an
806 // edge.
807 MBB.normalizeSuccProbs();
808
809 // Finally, we need to insert cmovs into the "fallthrough" edge. Here, we
810 // need to intersect the other condition codes. We can do this by just
811 // doing a cmov for each one.
812 if (!UncondSucc)
813 // If we have no fallthrough to protect (perhaps it is an indirect jump?)
814 // just skip this and continue.
815 continue;
816
817 assert(SuccCounts[UncondSucc] == 1 &&
818 "We should never have more than one edge to the unconditional "
819 "successor at this point because every other edge must have been "
820 "split above!");
821
822 // Sort and unique the codes to minimize them.
823 llvm::sort(UncondCodeSeq);
824 UncondCodeSeq.erase(std::unique(UncondCodeSeq.begin(), UncondCodeSeq.end()),
825 UncondCodeSeq.end());
826
827 // Build a checking version of the successor.
828 BuildCheckingBlockForSuccAndConds(MBB, *UncondSucc, /*SuccCount*/ 1,
829 UncondBr, UncondBr, UncondCodeSeq);
830 }
831
832 return CMovs;
833 }
834
835 /// Compute the register class for the unfolded load.
836 ///
837 /// FIXME: This should probably live in X86InstrInfo, potentially by adding
838 /// a way to unfold into a newly created vreg rather than requiring a register
839 /// input.
840 static const TargetRegisterClass *
getRegClassForUnfoldedLoad(MachineFunction & MF,const X86InstrInfo & TII,unsigned Opcode)841 getRegClassForUnfoldedLoad(MachineFunction &MF, const X86InstrInfo &TII,
842 unsigned Opcode) {
843 unsigned Index;
844 unsigned UnfoldedOpc = TII.getOpcodeAfterMemoryUnfold(
845 Opcode, /*UnfoldLoad*/ true, /*UnfoldStore*/ false, &Index);
846 const MCInstrDesc &MCID = TII.get(UnfoldedOpc);
847 return TII.getRegClass(MCID, Index, &TII.getRegisterInfo(), MF);
848 }
849
unfoldCallAndJumpLoads(MachineFunction & MF)850 void X86SpeculativeLoadHardeningPass::unfoldCallAndJumpLoads(
851 MachineFunction &MF) {
852 for (MachineBasicBlock &MBB : MF)
853 for (auto MII = MBB.instr_begin(), MIE = MBB.instr_end(); MII != MIE;) {
854 // Grab a reference and increment the iterator so we can remove this
855 // instruction if needed without disturbing the iteration.
856 MachineInstr &MI = *MII++;
857
858 // Must either be a call or a branch.
859 if (!MI.isCall() && !MI.isBranch())
860 continue;
861 // We only care about loading variants of these instructions.
862 if (!MI.mayLoad())
863 continue;
864
865 switch (MI.getOpcode()) {
866 default: {
867 LLVM_DEBUG(
868 dbgs() << "ERROR: Found an unexpected loading branch or call "
869 "instruction:\n";
870 MI.dump(); dbgs() << "\n");
871 report_fatal_error("Unexpected loading branch or call!");
872 }
873
874 case X86::FARCALL16m:
875 case X86::FARCALL32m:
876 case X86::FARCALL64m:
877 case X86::FARJMP16m:
878 case X86::FARJMP32m:
879 case X86::FARJMP64m:
880 // We cannot mitigate far jumps or calls, but we also don't expect them
881 // to be vulnerable to Spectre v1.2 style attacks.
882 continue;
883
884 case X86::CALL16m:
885 case X86::CALL16m_NT:
886 case X86::CALL32m:
887 case X86::CALL32m_NT:
888 case X86::CALL64m:
889 case X86::CALL64m_NT:
890 case X86::JMP16m:
891 case X86::JMP16m_NT:
892 case X86::JMP32m:
893 case X86::JMP32m_NT:
894 case X86::JMP64m:
895 case X86::JMP64m_NT:
896 case X86::TAILJMPm64:
897 case X86::TAILJMPm64_REX:
898 case X86::TAILJMPm:
899 case X86::TCRETURNmi64:
900 case X86::TCRETURNmi: {
901 // Use the generic unfold logic now that we know we're dealing with
902 // expected instructions.
903 // FIXME: We don't have test coverage for all of these!
904 auto *UnfoldedRC = getRegClassForUnfoldedLoad(MF, *TII, MI.getOpcode());
905 if (!UnfoldedRC) {
906 LLVM_DEBUG(dbgs()
907 << "ERROR: Unable to unfold load from instruction:\n";
908 MI.dump(); dbgs() << "\n");
909 report_fatal_error("Unable to unfold load!");
910 }
911 Register Reg = MRI->createVirtualRegister(UnfoldedRC);
912 SmallVector<MachineInstr *, 2> NewMIs;
913 // If we were able to compute an unfolded reg class, any failure here
914 // is just a programming error so just assert.
915 bool Unfolded =
916 TII->unfoldMemoryOperand(MF, MI, Reg, /*UnfoldLoad*/ true,
917 /*UnfoldStore*/ false, NewMIs);
918 (void)Unfolded;
919 assert(Unfolded &&
920 "Computed unfolded register class but failed to unfold");
921 // Now stitch the new instructions into place and erase the old one.
922 for (auto *NewMI : NewMIs)
923 MBB.insert(MI.getIterator(), NewMI);
924
925 // Update the call site info.
926 if (MI.isCandidateForCallSiteEntry())
927 MF.eraseCallSiteInfo(&MI);
928
929 MI.eraseFromParent();
930 LLVM_DEBUG({
931 dbgs() << "Unfolded load successfully into:\n";
932 for (auto *NewMI : NewMIs) {
933 NewMI->dump();
934 dbgs() << "\n";
935 }
936 });
937 continue;
938 }
939 }
940 llvm_unreachable("Escaped switch with default!");
941 }
942 }
943
944 /// Trace the predicate state through indirect branches, instrumenting them to
945 /// poison the state if a target is reached that does not match the expected
946 /// target.
947 ///
948 /// This is designed to mitigate Spectre variant 1 attacks where an indirect
949 /// branch is trained to predict a particular target and then mispredicts that
950 /// target in a way that can leak data. Despite using an indirect branch, this
951 /// is really a variant 1 style attack: it does not steer execution to an
952 /// arbitrary or attacker controlled address, and it does not require any
953 /// special code executing next to the victim. This attack can also be mitigated
954 /// through retpolines, but those require either replacing indirect branches
955 /// with conditional direct branches or lowering them through a device that
956 /// blocks speculation. This mitigation can replace these retpoline-style
957 /// mitigations for jump tables and other indirect branches within a function
958 /// when variant 2 isn't a risk while allowing limited speculation. Indirect
959 /// calls, however, cannot be mitigated through this technique without changing
960 /// the ABI in a fundamental way.
961 SmallVector<MachineInstr *, 16>
tracePredStateThroughIndirectBranches(MachineFunction & MF)962 X86SpeculativeLoadHardeningPass::tracePredStateThroughIndirectBranches(
963 MachineFunction &MF) {
964 // We use the SSAUpdater to insert PHI nodes for the target addresses of
965 // indirect branches. We don't actually need the full power of the SSA updater
966 // in this particular case as we always have immediately available values, but
967 // this avoids us having to re-implement the PHI construction logic.
968 MachineSSAUpdater TargetAddrSSA(MF);
969 TargetAddrSSA.Initialize(MRI->createVirtualRegister(&X86::GR64RegClass));
970
971 // Track which blocks were terminated with an indirect branch.
972 SmallPtrSet<MachineBasicBlock *, 4> IndirectTerminatedMBBs;
973
974 // We need to know what blocks end up reached via indirect branches. We
975 // expect this to be a subset of those whose address is taken and so track it
976 // directly via the CFG.
977 SmallPtrSet<MachineBasicBlock *, 4> IndirectTargetMBBs;
978
979 // Walk all the blocks which end in an indirect branch and make the
980 // target address available.
981 for (MachineBasicBlock &MBB : MF) {
982 // Find the last terminator.
983 auto MII = MBB.instr_rbegin();
984 while (MII != MBB.instr_rend() && MII->isDebugInstr())
985 ++MII;
986 if (MII == MBB.instr_rend())
987 continue;
988 MachineInstr &TI = *MII;
989 if (!TI.isTerminator() || !TI.isBranch())
990 // No terminator or non-branch terminator.
991 continue;
992
993 unsigned TargetReg;
994
995 switch (TI.getOpcode()) {
996 default:
997 // Direct branch or conditional branch (leading to fallthrough).
998 continue;
999
1000 case X86::FARJMP16m:
1001 case X86::FARJMP32m:
1002 case X86::FARJMP64m:
1003 // We cannot mitigate far jumps or calls, but we also don't expect them
1004 // to be vulnerable to Spectre v1.2 or v2 (self trained) style attacks.
1005 continue;
1006
1007 case X86::JMP16m:
1008 case X86::JMP16m_NT:
1009 case X86::JMP32m:
1010 case X86::JMP32m_NT:
1011 case X86::JMP64m:
1012 case X86::JMP64m_NT:
1013 // Mostly as documentation.
1014 report_fatal_error("Memory operand jumps should have been unfolded!");
1015
1016 case X86::JMP16r:
1017 report_fatal_error(
1018 "Support for 16-bit indirect branches is not implemented.");
1019 case X86::JMP32r:
1020 report_fatal_error(
1021 "Support for 32-bit indirect branches is not implemented.");
1022
1023 case X86::JMP64r:
1024 TargetReg = TI.getOperand(0).getReg();
1025 }
1026
1027 // We have definitely found an indirect branch. Verify that there are no
1028 // preceding conditional branches as we don't yet support that.
1029 if (llvm::any_of(MBB.terminators(), [&](MachineInstr &OtherTI) {
1030 return !OtherTI.isDebugInstr() && &OtherTI != &TI;
1031 })) {
1032 LLVM_DEBUG({
1033 dbgs() << "ERROR: Found other terminators in a block with an indirect "
1034 "branch! This is not yet supported! Terminator sequence:\n";
1035 for (MachineInstr &MI : MBB.terminators()) {
1036 MI.dump();
1037 dbgs() << '\n';
1038 }
1039 });
1040 report_fatal_error("Unimplemented terminator sequence!");
1041 }
1042
1043 // Make the target register an available value for this block.
1044 TargetAddrSSA.AddAvailableValue(&MBB, TargetReg);
1045 IndirectTerminatedMBBs.insert(&MBB);
1046
1047 // Add all the successors to our target candidates.
1048 for (MachineBasicBlock *Succ : MBB.successors())
1049 IndirectTargetMBBs.insert(Succ);
1050 }
1051
1052 // Keep track of the cmov instructions we insert so we can return them.
1053 SmallVector<MachineInstr *, 16> CMovs;
1054
1055 // If we didn't find any indirect branches with targets, nothing to do here.
1056 if (IndirectTargetMBBs.empty())
1057 return CMovs;
1058
1059 // We found indirect branches and targets that need to be instrumented to
1060 // harden loads within them. Walk the blocks of the function (to get a stable
1061 // ordering) and instrument each target of an indirect branch.
1062 for (MachineBasicBlock &MBB : MF) {
1063 // Skip the blocks that aren't candidate targets.
1064 if (!IndirectTargetMBBs.count(&MBB))
1065 continue;
1066
1067 // We don't expect EH pads to ever be reached via an indirect branch. If
1068 // this is desired for some reason, we could simply skip them here rather
1069 // than asserting.
1070 assert(!MBB.isEHPad() &&
1071 "Unexpected EH pad as target of an indirect branch!");
1072
1073 // We should never end up threading EFLAGS into a block to harden
1074 // conditional jumps as there would be an additional successor via the
1075 // indirect branch. As a consequence, all such edges would be split before
1076 // reaching here, and the inserted block will handle the EFLAGS-based
1077 // hardening.
1078 assert(!MBB.isLiveIn(X86::EFLAGS) &&
1079 "Cannot check within a block that already has live-in EFLAGS!");
1080
1081 // We can't handle having non-indirect edges into this block unless this is
1082 // the only successor and we can synthesize the necessary target address.
1083 for (MachineBasicBlock *Pred : MBB.predecessors()) {
1084 // If we've already handled this by extracting the target directly,
1085 // nothing to do.
1086 if (IndirectTerminatedMBBs.count(Pred))
1087 continue;
1088
1089 // Otherwise, we have to be the only successor. We generally expect this
1090 // to be true as conditional branches should have had a critical edge
1091 // split already. We don't however need to worry about EH pad successors
1092 // as they'll happily ignore the target and their hardening strategy is
1093 // resilient to all ways in which they could be reached speculatively.
1094 if (!llvm::all_of(Pred->successors(), [&](MachineBasicBlock *Succ) {
1095 return Succ->isEHPad() || Succ == &MBB;
1096 })) {
1097 LLVM_DEBUG({
1098 dbgs() << "ERROR: Found conditional entry to target of indirect "
1099 "branch!\n";
1100 Pred->dump();
1101 MBB.dump();
1102 });
1103 report_fatal_error("Cannot harden a conditional entry to a target of "
1104 "an indirect branch!");
1105 }
1106
1107 // Now we need to compute the address of this block and install it as a
1108 // synthetic target in the predecessor. We do this at the bottom of the
1109 // predecessor.
1110 auto InsertPt = Pred->getFirstTerminator();
1111 Register TargetReg = MRI->createVirtualRegister(&X86::GR64RegClass);
1112 if (MF.getTarget().getCodeModel() == CodeModel::Small &&
1113 !Subtarget->isPositionIndependent()) {
1114 // Directly materialize it into an immediate.
1115 auto AddrI = BuildMI(*Pred, InsertPt, DebugLoc(),
1116 TII->get(X86::MOV64ri32), TargetReg)
1117 .addMBB(&MBB);
1118 ++NumInstsInserted;
1119 (void)AddrI;
1120 LLVM_DEBUG(dbgs() << " Inserting mov: "; AddrI->dump();
1121 dbgs() << "\n");
1122 } else {
1123 auto AddrI = BuildMI(*Pred, InsertPt, DebugLoc(), TII->get(X86::LEA64r),
1124 TargetReg)
1125 .addReg(/*Base*/ X86::RIP)
1126 .addImm(/*Scale*/ 1)
1127 .addReg(/*Index*/ 0)
1128 .addMBB(&MBB)
1129 .addReg(/*Segment*/ 0);
1130 ++NumInstsInserted;
1131 (void)AddrI;
1132 LLVM_DEBUG(dbgs() << " Inserting lea: "; AddrI->dump();
1133 dbgs() << "\n");
1134 }
1135 // And make this available.
1136 TargetAddrSSA.AddAvailableValue(Pred, TargetReg);
1137 }
1138
1139 // Materialize the needed SSA value of the target. Note that we need the
1140 // middle of the block as this block might at the bottom have an indirect
1141 // branch back to itself. We can do this here because at this point, every
1142 // predecessor of this block has an available value. This is basically just
1143 // automating the construction of a PHI node for this target.
1144 unsigned TargetReg = TargetAddrSSA.GetValueInMiddleOfBlock(&MBB);
1145
1146 // Insert a comparison of the incoming target register with this block's
1147 // address. This also requires us to mark the block as having its address
1148 // taken explicitly.
1149 MBB.setHasAddressTaken();
1150 auto InsertPt = MBB.SkipPHIsLabelsAndDebug(MBB.begin());
1151 if (MF.getTarget().getCodeModel() == CodeModel::Small &&
1152 !Subtarget->isPositionIndependent()) {
1153 // Check directly against a relocated immediate when we can.
1154 auto CheckI = BuildMI(MBB, InsertPt, DebugLoc(), TII->get(X86::CMP64ri32))
1155 .addReg(TargetReg, RegState::Kill)
1156 .addMBB(&MBB);
1157 ++NumInstsInserted;
1158 (void)CheckI;
1159 LLVM_DEBUG(dbgs() << " Inserting cmp: "; CheckI->dump(); dbgs() << "\n");
1160 } else {
1161 // Otherwise compute the address into a register first.
1162 Register AddrReg = MRI->createVirtualRegister(&X86::GR64RegClass);
1163 auto AddrI =
1164 BuildMI(MBB, InsertPt, DebugLoc(), TII->get(X86::LEA64r), AddrReg)
1165 .addReg(/*Base*/ X86::RIP)
1166 .addImm(/*Scale*/ 1)
1167 .addReg(/*Index*/ 0)
1168 .addMBB(&MBB)
1169 .addReg(/*Segment*/ 0);
1170 ++NumInstsInserted;
1171 (void)AddrI;
1172 LLVM_DEBUG(dbgs() << " Inserting lea: "; AddrI->dump(); dbgs() << "\n");
1173 auto CheckI = BuildMI(MBB, InsertPt, DebugLoc(), TII->get(X86::CMP64rr))
1174 .addReg(TargetReg, RegState::Kill)
1175 .addReg(AddrReg, RegState::Kill);
1176 ++NumInstsInserted;
1177 (void)CheckI;
1178 LLVM_DEBUG(dbgs() << " Inserting cmp: "; CheckI->dump(); dbgs() << "\n");
1179 }
1180
1181 // Now cmov over the predicate if the comparison wasn't equal.
1182 int PredStateSizeInBytes = TRI->getRegSizeInBits(*PS->RC) / 8;
1183 auto CMovOp = X86::getCMovOpcode(PredStateSizeInBytes);
1184 Register UpdatedStateReg = MRI->createVirtualRegister(PS->RC);
1185 auto CMovI =
1186 BuildMI(MBB, InsertPt, DebugLoc(), TII->get(CMovOp), UpdatedStateReg)
1187 .addReg(PS->InitialReg)
1188 .addReg(PS->PoisonReg)
1189 .addImm(X86::COND_NE);
1190 CMovI->findRegisterUseOperand(X86::EFLAGS)->setIsKill(true);
1191 ++NumInstsInserted;
1192 LLVM_DEBUG(dbgs() << " Inserting cmov: "; CMovI->dump(); dbgs() << "\n");
1193 CMovs.push_back(&*CMovI);
1194
1195 // And put the new value into the available values for SSA form of our
1196 // predicate state.
1197 PS->SSA.AddAvailableValue(&MBB, UpdatedStateReg);
1198 }
1199
1200 // Return all the newly inserted cmov instructions of the predicate state.
1201 return CMovs;
1202 }
1203
1204 // Returns true if the MI has EFLAGS as a register def operand and it's live,
1205 // otherwise it returns false
isEFLAGSDefLive(const MachineInstr & MI)1206 static bool isEFLAGSDefLive(const MachineInstr &MI) {
1207 if (const MachineOperand *DefOp = MI.findRegisterDefOperand(X86::EFLAGS)) {
1208 return !DefOp->isDead();
1209 }
1210 return false;
1211 }
1212
isEFLAGSLive(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,const TargetRegisterInfo & TRI)1213 static bool isEFLAGSLive(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
1214 const TargetRegisterInfo &TRI) {
1215 // Check if EFLAGS are alive by seeing if there is a def of them or they
1216 // live-in, and then seeing if that def is in turn used.
1217 for (MachineInstr &MI : llvm::reverse(llvm::make_range(MBB.begin(), I))) {
1218 if (MachineOperand *DefOp = MI.findRegisterDefOperand(X86::EFLAGS)) {
1219 // If the def is dead, then EFLAGS is not live.
1220 if (DefOp->isDead())
1221 return false;
1222
1223 // Otherwise we've def'ed it, and it is live.
1224 return true;
1225 }
1226 // While at this instruction, also check if we use and kill EFLAGS
1227 // which means it isn't live.
1228 if (MI.killsRegister(X86::EFLAGS, &TRI))
1229 return false;
1230 }
1231
1232 // If we didn't find anything conclusive (neither definitely alive or
1233 // definitely dead) return whether it lives into the block.
1234 return MBB.isLiveIn(X86::EFLAGS);
1235 }
1236
1237 /// Trace the predicate state through each of the blocks in the function,
1238 /// hardening everything necessary along the way.
1239 ///
1240 /// We call this routine once the initial predicate state has been established
1241 /// for each basic block in the function in the SSA updater. This routine traces
1242 /// it through the instructions within each basic block, and for non-returning
1243 /// blocks informs the SSA updater about the final state that lives out of the
1244 /// block. Along the way, it hardens any vulnerable instruction using the
1245 /// currently valid predicate state. We have to do these two things together
1246 /// because the SSA updater only works across blocks. Within a block, we track
1247 /// the current predicate state directly and update it as it changes.
1248 ///
1249 /// This operates in two passes over each block. First, we analyze the loads in
1250 /// the block to determine which strategy will be used to harden them: hardening
1251 /// the address or hardening the loaded value when loaded into a register
1252 /// amenable to hardening. We have to process these first because the two
1253 /// strategies may interact -- later hardening may change what strategy we wish
1254 /// to use. We also will analyze data dependencies between loads and avoid
1255 /// hardening those loads that are data dependent on a load with a hardened
1256 /// address. We also skip hardening loads already behind an LFENCE as that is
1257 /// sufficient to harden them against misspeculation.
1258 ///
1259 /// Second, we actively trace the predicate state through the block, applying
1260 /// the hardening steps we determined necessary in the first pass as we go.
1261 ///
1262 /// These two passes are applied to each basic block. We operate one block at a
1263 /// time to simplify reasoning about reachability and sequencing.
tracePredStateThroughBlocksAndHarden(MachineFunction & MF)1264 void X86SpeculativeLoadHardeningPass::tracePredStateThroughBlocksAndHarden(
1265 MachineFunction &MF) {
1266 SmallPtrSet<MachineInstr *, 16> HardenPostLoad;
1267 SmallPtrSet<MachineInstr *, 16> HardenLoadAddr;
1268
1269 SmallSet<unsigned, 16> HardenedAddrRegs;
1270
1271 SmallDenseMap<unsigned, unsigned, 32> AddrRegToHardenedReg;
1272
1273 // Track the set of load-dependent registers through the basic block. Because
1274 // the values of these registers have an existing data dependency on a loaded
1275 // value which we would have checked, we can omit any checks on them.
1276 SparseBitVector<> LoadDepRegs;
1277
1278 for (MachineBasicBlock &MBB : MF) {
1279 // The first pass over the block: collect all the loads which can have their
1280 // loaded value hardened and all the loads that instead need their address
1281 // hardened. During this walk we propagate load dependence for address
1282 // hardened loads and also look for LFENCE to stop hardening wherever
1283 // possible. When deciding whether or not to harden the loaded value or not,
1284 // we check to see if any registers used in the address will have been
1285 // hardened at this point and if so, harden any remaining address registers
1286 // as that often successfully re-uses hardened addresses and minimizes
1287 // instructions.
1288 //
1289 // FIXME: We should consider an aggressive mode where we continue to keep as
1290 // many loads value hardened even when some address register hardening would
1291 // be free (due to reuse).
1292 //
1293 // Note that we only need this pass if we are actually hardening loads.
1294 if (HardenLoads)
1295 for (MachineInstr &MI : MBB) {
1296 // We naively assume that all def'ed registers of an instruction have
1297 // a data dependency on all of their operands.
1298 // FIXME: Do a more careful analysis of x86 to build a conservative
1299 // model here.
1300 if (llvm::any_of(MI.uses(), [&](MachineOperand &Op) {
1301 return Op.isReg() && LoadDepRegs.test(Op.getReg());
1302 }))
1303 for (MachineOperand &Def : MI.defs())
1304 if (Def.isReg())
1305 LoadDepRegs.set(Def.getReg());
1306
1307 // Both Intel and AMD are guiding that they will change the semantics of
1308 // LFENCE to be a speculation barrier, so if we see an LFENCE, there is
1309 // no more need to guard things in this block.
1310 if (MI.getOpcode() == X86::LFENCE)
1311 break;
1312
1313 // If this instruction cannot load, nothing to do.
1314 if (!MI.mayLoad())
1315 continue;
1316
1317 // Some instructions which "load" are trivially safe or unimportant.
1318 if (MI.getOpcode() == X86::MFENCE)
1319 continue;
1320
1321 // Extract the memory operand information about this instruction.
1322 // FIXME: This doesn't handle loading pseudo instructions which we often
1323 // could handle with similarly generic logic. We probably need to add an
1324 // MI-layer routine similar to the MC-layer one we use here which maps
1325 // pseudos much like this maps real instructions.
1326 const MCInstrDesc &Desc = MI.getDesc();
1327 int MemRefBeginIdx = X86II::getMemoryOperandNo(Desc.TSFlags);
1328 if (MemRefBeginIdx < 0) {
1329 LLVM_DEBUG(dbgs()
1330 << "WARNING: unable to harden loading instruction: ";
1331 MI.dump());
1332 continue;
1333 }
1334
1335 MemRefBeginIdx += X86II::getOperandBias(Desc);
1336
1337 MachineOperand &BaseMO =
1338 MI.getOperand(MemRefBeginIdx + X86::AddrBaseReg);
1339 MachineOperand &IndexMO =
1340 MI.getOperand(MemRefBeginIdx + X86::AddrIndexReg);
1341
1342 // If we have at least one (non-frame-index, non-RIP) register operand,
1343 // and neither operand is load-dependent, we need to check the load.
1344 unsigned BaseReg = 0, IndexReg = 0;
1345 if (!BaseMO.isFI() && BaseMO.getReg() != X86::RIP &&
1346 BaseMO.getReg() != X86::NoRegister)
1347 BaseReg = BaseMO.getReg();
1348 if (IndexMO.getReg() != X86::NoRegister)
1349 IndexReg = IndexMO.getReg();
1350
1351 if (!BaseReg && !IndexReg)
1352 // No register operands!
1353 continue;
1354
1355 // If any register operand is dependent, this load is dependent and we
1356 // needn't check it.
1357 // FIXME: Is this true in the case where we are hardening loads after
1358 // they complete? Unclear, need to investigate.
1359 if ((BaseReg && LoadDepRegs.test(BaseReg)) ||
1360 (IndexReg && LoadDepRegs.test(IndexReg)))
1361 continue;
1362
1363 // If post-load hardening is enabled, this load is compatible with
1364 // post-load hardening, and we aren't already going to harden one of the
1365 // address registers, queue it up to be hardened post-load. Notably,
1366 // even once hardened this won't introduce a useful dependency that
1367 // could prune out subsequent loads.
1368 if (EnablePostLoadHardening && X86InstrInfo::isDataInvariantLoad(MI) &&
1369 !isEFLAGSDefLive(MI) && MI.getDesc().getNumDefs() == 1 &&
1370 MI.getOperand(0).isReg() &&
1371 canHardenRegister(MI.getOperand(0).getReg()) &&
1372 !HardenedAddrRegs.count(BaseReg) &&
1373 !HardenedAddrRegs.count(IndexReg)) {
1374 HardenPostLoad.insert(&MI);
1375 HardenedAddrRegs.insert(MI.getOperand(0).getReg());
1376 continue;
1377 }
1378
1379 // Record this instruction for address hardening and record its register
1380 // operands as being address-hardened.
1381 HardenLoadAddr.insert(&MI);
1382 if (BaseReg)
1383 HardenedAddrRegs.insert(BaseReg);
1384 if (IndexReg)
1385 HardenedAddrRegs.insert(IndexReg);
1386
1387 for (MachineOperand &Def : MI.defs())
1388 if (Def.isReg())
1389 LoadDepRegs.set(Def.getReg());
1390 }
1391
1392 // Now re-walk the instructions in the basic block, and apply whichever
1393 // hardening strategy we have elected. Note that we do this in a second
1394 // pass specifically so that we have the complete set of instructions for
1395 // which we will do post-load hardening and can defer it in certain
1396 // circumstances.
1397 for (MachineInstr &MI : MBB) {
1398 if (HardenLoads) {
1399 // We cannot both require hardening the def of a load and its address.
1400 assert(!(HardenLoadAddr.count(&MI) && HardenPostLoad.count(&MI)) &&
1401 "Requested to harden both the address and def of a load!");
1402
1403 // Check if this is a load whose address needs to be hardened.
1404 if (HardenLoadAddr.erase(&MI)) {
1405 const MCInstrDesc &Desc = MI.getDesc();
1406 int MemRefBeginIdx = X86II::getMemoryOperandNo(Desc.TSFlags);
1407 assert(MemRefBeginIdx >= 0 && "Cannot have an invalid index here!");
1408
1409 MemRefBeginIdx += X86II::getOperandBias(Desc);
1410
1411 MachineOperand &BaseMO =
1412 MI.getOperand(MemRefBeginIdx + X86::AddrBaseReg);
1413 MachineOperand &IndexMO =
1414 MI.getOperand(MemRefBeginIdx + X86::AddrIndexReg);
1415 hardenLoadAddr(MI, BaseMO, IndexMO, AddrRegToHardenedReg);
1416 continue;
1417 }
1418
1419 // Test if this instruction is one of our post load instructions (and
1420 // remove it from the set if so).
1421 if (HardenPostLoad.erase(&MI)) {
1422 assert(!MI.isCall() && "Must not try to post-load harden a call!");
1423
1424 // If this is a data-invariant load and there is no EFLAGS
1425 // interference, we want to try and sink any hardening as far as
1426 // possible.
1427 if (X86InstrInfo::isDataInvariantLoad(MI) && !isEFLAGSDefLive(MI)) {
1428 // Sink the instruction we'll need to harden as far as we can down
1429 // the graph.
1430 MachineInstr *SunkMI = sinkPostLoadHardenedInst(MI, HardenPostLoad);
1431
1432 // If we managed to sink this instruction, update everything so we
1433 // harden that instruction when we reach it in the instruction
1434 // sequence.
1435 if (SunkMI != &MI) {
1436 // If in sinking there was no instruction needing to be hardened,
1437 // we're done.
1438 if (!SunkMI)
1439 continue;
1440
1441 // Otherwise, add this to the set of defs we harden.
1442 HardenPostLoad.insert(SunkMI);
1443 continue;
1444 }
1445 }
1446
1447 unsigned HardenedReg = hardenPostLoad(MI);
1448
1449 // Mark the resulting hardened register as such so we don't re-harden.
1450 AddrRegToHardenedReg[HardenedReg] = HardenedReg;
1451
1452 continue;
1453 }
1454
1455 // Check for an indirect call or branch that may need its input hardened
1456 // even if we couldn't find the specific load used, or were able to
1457 // avoid hardening it for some reason. Note that here we cannot break
1458 // out afterward as we may still need to handle any call aspect of this
1459 // instruction.
1460 if ((MI.isCall() || MI.isBranch()) && HardenIndirectCallsAndJumps)
1461 hardenIndirectCallOrJumpInstr(MI, AddrRegToHardenedReg);
1462 }
1463
1464 // After we finish hardening loads we handle interprocedural hardening if
1465 // enabled and relevant for this instruction.
1466 if (!HardenInterprocedurally)
1467 continue;
1468 if (!MI.isCall() && !MI.isReturn())
1469 continue;
1470
1471 // If this is a direct return (IE, not a tail call) just directly harden
1472 // it.
1473 if (MI.isReturn() && !MI.isCall()) {
1474 hardenReturnInstr(MI);
1475 continue;
1476 }
1477
1478 // Otherwise we have a call. We need to handle transferring the predicate
1479 // state into a call and recovering it after the call returns (unless this
1480 // is a tail call).
1481 assert(MI.isCall() && "Should only reach here for calls!");
1482 tracePredStateThroughCall(MI);
1483 }
1484
1485 HardenPostLoad.clear();
1486 HardenLoadAddr.clear();
1487 HardenedAddrRegs.clear();
1488 AddrRegToHardenedReg.clear();
1489
1490 // Currently, we only track data-dependent loads within a basic block.
1491 // FIXME: We should see if this is necessary or if we could be more
1492 // aggressive here without opening up attack avenues.
1493 LoadDepRegs.clear();
1494 }
1495 }
1496
1497 /// Save EFLAGS into the returned GPR. This can in turn be restored with
1498 /// `restoreEFLAGS`.
1499 ///
1500 /// Note that LLVM can only lower very simple patterns of saved and restored
1501 /// EFLAGS registers. The restore should always be within the same basic block
1502 /// as the save so that no PHI nodes are inserted.
saveEFLAGS(MachineBasicBlock & MBB,MachineBasicBlock::iterator InsertPt,DebugLoc Loc)1503 unsigned X86SpeculativeLoadHardeningPass::saveEFLAGS(
1504 MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertPt,
1505 DebugLoc Loc) {
1506 // FIXME: Hard coding this to a 32-bit register class seems weird, but matches
1507 // what instruction selection does.
1508 Register Reg = MRI->createVirtualRegister(&X86::GR32RegClass);
1509 // We directly copy the FLAGS register and rely on later lowering to clean
1510 // this up into the appropriate setCC instructions.
1511 BuildMI(MBB, InsertPt, Loc, TII->get(X86::COPY), Reg).addReg(X86::EFLAGS);
1512 ++NumInstsInserted;
1513 return Reg;
1514 }
1515
1516 /// Restore EFLAGS from the provided GPR. This should be produced by
1517 /// `saveEFLAGS`.
1518 ///
1519 /// This must be done within the same basic block as the save in order to
1520 /// reliably lower.
restoreEFLAGS(MachineBasicBlock & MBB,MachineBasicBlock::iterator InsertPt,DebugLoc Loc,unsigned Reg)1521 void X86SpeculativeLoadHardeningPass::restoreEFLAGS(
1522 MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertPt, DebugLoc Loc,
1523 unsigned Reg) {
1524 BuildMI(MBB, InsertPt, Loc, TII->get(X86::COPY), X86::EFLAGS).addReg(Reg);
1525 ++NumInstsInserted;
1526 }
1527
1528 /// Takes the current predicate state (in a register) and merges it into the
1529 /// stack pointer. The state is essentially a single bit, but we merge this in
1530 /// a way that won't form non-canonical pointers and also will be preserved
1531 /// across normal stack adjustments.
mergePredStateIntoSP(MachineBasicBlock & MBB,MachineBasicBlock::iterator InsertPt,DebugLoc Loc,unsigned PredStateReg)1532 void X86SpeculativeLoadHardeningPass::mergePredStateIntoSP(
1533 MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertPt, DebugLoc Loc,
1534 unsigned PredStateReg) {
1535 Register TmpReg = MRI->createVirtualRegister(PS->RC);
1536 // FIXME: This hard codes a shift distance based on the number of bits needed
1537 // to stay canonical on 64-bit. We should compute this somehow and support
1538 // 32-bit as part of that.
1539 auto ShiftI = BuildMI(MBB, InsertPt, Loc, TII->get(X86::SHL64ri), TmpReg)
1540 .addReg(PredStateReg, RegState::Kill)
1541 .addImm(47);
1542 ShiftI->addRegisterDead(X86::EFLAGS, TRI);
1543 ++NumInstsInserted;
1544 auto OrI = BuildMI(MBB, InsertPt, Loc, TII->get(X86::OR64rr), X86::RSP)
1545 .addReg(X86::RSP)
1546 .addReg(TmpReg, RegState::Kill);
1547 OrI->addRegisterDead(X86::EFLAGS, TRI);
1548 ++NumInstsInserted;
1549 }
1550
1551 /// Extracts the predicate state stored in the high bits of the stack pointer.
extractPredStateFromSP(MachineBasicBlock & MBB,MachineBasicBlock::iterator InsertPt,DebugLoc Loc)1552 unsigned X86SpeculativeLoadHardeningPass::extractPredStateFromSP(
1553 MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertPt,
1554 DebugLoc Loc) {
1555 Register PredStateReg = MRI->createVirtualRegister(PS->RC);
1556 Register TmpReg = MRI->createVirtualRegister(PS->RC);
1557
1558 // We know that the stack pointer will have any preserved predicate state in
1559 // its high bit. We just want to smear this across the other bits. Turns out,
1560 // this is exactly what an arithmetic right shift does.
1561 BuildMI(MBB, InsertPt, Loc, TII->get(TargetOpcode::COPY), TmpReg)
1562 .addReg(X86::RSP);
1563 auto ShiftI =
1564 BuildMI(MBB, InsertPt, Loc, TII->get(X86::SAR64ri), PredStateReg)
1565 .addReg(TmpReg, RegState::Kill)
1566 .addImm(TRI->getRegSizeInBits(*PS->RC) - 1);
1567 ShiftI->addRegisterDead(X86::EFLAGS, TRI);
1568 ++NumInstsInserted;
1569
1570 return PredStateReg;
1571 }
1572
hardenLoadAddr(MachineInstr & MI,MachineOperand & BaseMO,MachineOperand & IndexMO,SmallDenseMap<unsigned,unsigned,32> & AddrRegToHardenedReg)1573 void X86SpeculativeLoadHardeningPass::hardenLoadAddr(
1574 MachineInstr &MI, MachineOperand &BaseMO, MachineOperand &IndexMO,
1575 SmallDenseMap<unsigned, unsigned, 32> &AddrRegToHardenedReg) {
1576 MachineBasicBlock &MBB = *MI.getParent();
1577 DebugLoc Loc = MI.getDebugLoc();
1578
1579 // Check if EFLAGS are alive by seeing if there is a def of them or they
1580 // live-in, and then seeing if that def is in turn used.
1581 bool EFLAGSLive = isEFLAGSLive(MBB, MI.getIterator(), *TRI);
1582
1583 SmallVector<MachineOperand *, 2> HardenOpRegs;
1584
1585 if (BaseMO.isFI()) {
1586 // A frame index is never a dynamically controllable load, so only
1587 // harden it if we're covering fixed address loads as well.
1588 LLVM_DEBUG(
1589 dbgs() << " Skipping hardening base of explicit stack frame load: ";
1590 MI.dump(); dbgs() << "\n");
1591 } else if (BaseMO.getReg() == X86::RSP) {
1592 // Some idempotent atomic operations are lowered directly to a locked
1593 // OR with 0 to the top of stack(or slightly offset from top) which uses an
1594 // explicit RSP register as the base.
1595 assert(IndexMO.getReg() == X86::NoRegister &&
1596 "Explicit RSP access with dynamic index!");
1597 LLVM_DEBUG(
1598 dbgs() << " Cannot harden base of explicit RSP offset in a load!");
1599 } else if (BaseMO.getReg() == X86::RIP ||
1600 BaseMO.getReg() == X86::NoRegister) {
1601 // For both RIP-relative addressed loads or absolute loads, we cannot
1602 // meaningfully harden them because the address being loaded has no
1603 // dynamic component.
1604 //
1605 // FIXME: When using a segment base (like TLS does) we end up with the
1606 // dynamic address being the base plus -1 because we can't mutate the
1607 // segment register here. This allows the signed 32-bit offset to point at
1608 // valid segment-relative addresses and load them successfully.
1609 LLVM_DEBUG(
1610 dbgs() << " Cannot harden base of "
1611 << (BaseMO.getReg() == X86::RIP ? "RIP-relative" : "no-base")
1612 << " address in a load!");
1613 } else {
1614 assert(BaseMO.isReg() &&
1615 "Only allowed to have a frame index or register base.");
1616 HardenOpRegs.push_back(&BaseMO);
1617 }
1618
1619 if (IndexMO.getReg() != X86::NoRegister &&
1620 (HardenOpRegs.empty() ||
1621 HardenOpRegs.front()->getReg() != IndexMO.getReg()))
1622 HardenOpRegs.push_back(&IndexMO);
1623
1624 assert((HardenOpRegs.size() == 1 || HardenOpRegs.size() == 2) &&
1625 "Should have exactly one or two registers to harden!");
1626 assert((HardenOpRegs.size() == 1 ||
1627 HardenOpRegs[0]->getReg() != HardenOpRegs[1]->getReg()) &&
1628 "Should not have two of the same registers!");
1629
1630 // Remove any registers that have alreaded been checked.
1631 llvm::erase_if(HardenOpRegs, [&](MachineOperand *Op) {
1632 // See if this operand's register has already been checked.
1633 auto It = AddrRegToHardenedReg.find(Op->getReg());
1634 if (It == AddrRegToHardenedReg.end())
1635 // Not checked, so retain this one.
1636 return false;
1637
1638 // Otherwise, we can directly update this operand and remove it.
1639 Op->setReg(It->second);
1640 return true;
1641 });
1642 // If there are none left, we're done.
1643 if (HardenOpRegs.empty())
1644 return;
1645
1646 // Compute the current predicate state.
1647 unsigned StateReg = PS->SSA.GetValueAtEndOfBlock(&MBB);
1648
1649 auto InsertPt = MI.getIterator();
1650
1651 // If EFLAGS are live and we don't have access to instructions that avoid
1652 // clobbering EFLAGS we need to save and restore them. This in turn makes
1653 // the EFLAGS no longer live.
1654 unsigned FlagsReg = 0;
1655 if (EFLAGSLive && !Subtarget->hasBMI2()) {
1656 EFLAGSLive = false;
1657 FlagsReg = saveEFLAGS(MBB, InsertPt, Loc);
1658 }
1659
1660 for (MachineOperand *Op : HardenOpRegs) {
1661 Register OpReg = Op->getReg();
1662 auto *OpRC = MRI->getRegClass(OpReg);
1663 Register TmpReg = MRI->createVirtualRegister(OpRC);
1664
1665 // If this is a vector register, we'll need somewhat custom logic to handle
1666 // hardening it.
1667 if (!Subtarget->hasVLX() && (OpRC->hasSuperClassEq(&X86::VR128RegClass) ||
1668 OpRC->hasSuperClassEq(&X86::VR256RegClass))) {
1669 assert(Subtarget->hasAVX2() && "AVX2-specific register classes!");
1670 bool Is128Bit = OpRC->hasSuperClassEq(&X86::VR128RegClass);
1671
1672 // Move our state into a vector register.
1673 // FIXME: We could skip this at the cost of longer encodings with AVX-512
1674 // but that doesn't seem likely worth it.
1675 Register VStateReg = MRI->createVirtualRegister(&X86::VR128RegClass);
1676 auto MovI =
1677 BuildMI(MBB, InsertPt, Loc, TII->get(X86::VMOV64toPQIrr), VStateReg)
1678 .addReg(StateReg);
1679 (void)MovI;
1680 ++NumInstsInserted;
1681 LLVM_DEBUG(dbgs() << " Inserting mov: "; MovI->dump(); dbgs() << "\n");
1682
1683 // Broadcast it across the vector register.
1684 Register VBStateReg = MRI->createVirtualRegister(OpRC);
1685 auto BroadcastI = BuildMI(MBB, InsertPt, Loc,
1686 TII->get(Is128Bit ? X86::VPBROADCASTQrr
1687 : X86::VPBROADCASTQYrr),
1688 VBStateReg)
1689 .addReg(VStateReg);
1690 (void)BroadcastI;
1691 ++NumInstsInserted;
1692 LLVM_DEBUG(dbgs() << " Inserting broadcast: "; BroadcastI->dump();
1693 dbgs() << "\n");
1694
1695 // Merge our potential poison state into the value with a vector or.
1696 auto OrI =
1697 BuildMI(MBB, InsertPt, Loc,
1698 TII->get(Is128Bit ? X86::VPORrr : X86::VPORYrr), TmpReg)
1699 .addReg(VBStateReg)
1700 .addReg(OpReg);
1701 (void)OrI;
1702 ++NumInstsInserted;
1703 LLVM_DEBUG(dbgs() << " Inserting or: "; OrI->dump(); dbgs() << "\n");
1704 } else if (OpRC->hasSuperClassEq(&X86::VR128XRegClass) ||
1705 OpRC->hasSuperClassEq(&X86::VR256XRegClass) ||
1706 OpRC->hasSuperClassEq(&X86::VR512RegClass)) {
1707 assert(Subtarget->hasAVX512() && "AVX512-specific register classes!");
1708 bool Is128Bit = OpRC->hasSuperClassEq(&X86::VR128XRegClass);
1709 bool Is256Bit = OpRC->hasSuperClassEq(&X86::VR256XRegClass);
1710 if (Is128Bit || Is256Bit)
1711 assert(Subtarget->hasVLX() && "AVX512VL-specific register classes!");
1712
1713 // Broadcast our state into a vector register.
1714 Register VStateReg = MRI->createVirtualRegister(OpRC);
1715 unsigned BroadcastOp = Is128Bit ? X86::VPBROADCASTQrZ128rr
1716 : Is256Bit ? X86::VPBROADCASTQrZ256rr
1717 : X86::VPBROADCASTQrZrr;
1718 auto BroadcastI =
1719 BuildMI(MBB, InsertPt, Loc, TII->get(BroadcastOp), VStateReg)
1720 .addReg(StateReg);
1721 (void)BroadcastI;
1722 ++NumInstsInserted;
1723 LLVM_DEBUG(dbgs() << " Inserting broadcast: "; BroadcastI->dump();
1724 dbgs() << "\n");
1725
1726 // Merge our potential poison state into the value with a vector or.
1727 unsigned OrOp = Is128Bit ? X86::VPORQZ128rr
1728 : Is256Bit ? X86::VPORQZ256rr : X86::VPORQZrr;
1729 auto OrI = BuildMI(MBB, InsertPt, Loc, TII->get(OrOp), TmpReg)
1730 .addReg(VStateReg)
1731 .addReg(OpReg);
1732 (void)OrI;
1733 ++NumInstsInserted;
1734 LLVM_DEBUG(dbgs() << " Inserting or: "; OrI->dump(); dbgs() << "\n");
1735 } else {
1736 // FIXME: Need to support GR32 here for 32-bit code.
1737 assert(OpRC->hasSuperClassEq(&X86::GR64RegClass) &&
1738 "Not a supported register class for address hardening!");
1739
1740 if (!EFLAGSLive) {
1741 // Merge our potential poison state into the value with an or.
1742 auto OrI = BuildMI(MBB, InsertPt, Loc, TII->get(X86::OR64rr), TmpReg)
1743 .addReg(StateReg)
1744 .addReg(OpReg);
1745 OrI->addRegisterDead(X86::EFLAGS, TRI);
1746 ++NumInstsInserted;
1747 LLVM_DEBUG(dbgs() << " Inserting or: "; OrI->dump(); dbgs() << "\n");
1748 } else {
1749 // We need to avoid touching EFLAGS so shift out all but the least
1750 // significant bit using the instruction that doesn't update flags.
1751 auto ShiftI =
1752 BuildMI(MBB, InsertPt, Loc, TII->get(X86::SHRX64rr), TmpReg)
1753 .addReg(OpReg)
1754 .addReg(StateReg);
1755 (void)ShiftI;
1756 ++NumInstsInserted;
1757 LLVM_DEBUG(dbgs() << " Inserting shrx: "; ShiftI->dump();
1758 dbgs() << "\n");
1759 }
1760 }
1761
1762 // Record this register as checked and update the operand.
1763 assert(!AddrRegToHardenedReg.count(Op->getReg()) &&
1764 "Should not have checked this register yet!");
1765 AddrRegToHardenedReg[Op->getReg()] = TmpReg;
1766 Op->setReg(TmpReg);
1767 ++NumAddrRegsHardened;
1768 }
1769
1770 // And restore the flags if needed.
1771 if (FlagsReg)
1772 restoreEFLAGS(MBB, InsertPt, Loc, FlagsReg);
1773 }
1774
sinkPostLoadHardenedInst(MachineInstr & InitialMI,SmallPtrSetImpl<MachineInstr * > & HardenedInstrs)1775 MachineInstr *X86SpeculativeLoadHardeningPass::sinkPostLoadHardenedInst(
1776 MachineInstr &InitialMI, SmallPtrSetImpl<MachineInstr *> &HardenedInstrs) {
1777 assert(X86InstrInfo::isDataInvariantLoad(InitialMI) &&
1778 "Cannot get here with a non-invariant load!");
1779 assert(!isEFLAGSDefLive(InitialMI) &&
1780 "Cannot get here with a data invariant load "
1781 "that interferes with EFLAGS!");
1782
1783 // See if we can sink hardening the loaded value.
1784 auto SinkCheckToSingleUse =
1785 [&](MachineInstr &MI) -> Optional<MachineInstr *> {
1786 Register DefReg = MI.getOperand(0).getReg();
1787
1788 // We need to find a single use which we can sink the check. We can
1789 // primarily do this because many uses may already end up checked on their
1790 // own.
1791 MachineInstr *SingleUseMI = nullptr;
1792 for (MachineInstr &UseMI : MRI->use_instructions(DefReg)) {
1793 // If we're already going to harden this use, it is data invariant, it
1794 // does not interfere with EFLAGS, and within our block.
1795 if (HardenedInstrs.count(&UseMI)) {
1796 if (!X86InstrInfo::isDataInvariantLoad(UseMI) || isEFLAGSDefLive(UseMI)) {
1797 // If we've already decided to harden a non-load, we must have sunk
1798 // some other post-load hardened instruction to it and it must itself
1799 // be data-invariant.
1800 assert(X86InstrInfo::isDataInvariant(UseMI) &&
1801 "Data variant instruction being hardened!");
1802 continue;
1803 }
1804
1805 // Otherwise, this is a load and the load component can't be data
1806 // invariant so check how this register is being used.
1807 const MCInstrDesc &Desc = UseMI.getDesc();
1808 int MemRefBeginIdx = X86II::getMemoryOperandNo(Desc.TSFlags);
1809 assert(MemRefBeginIdx >= 0 &&
1810 "Should always have mem references here!");
1811 MemRefBeginIdx += X86II::getOperandBias(Desc);
1812
1813 MachineOperand &BaseMO =
1814 UseMI.getOperand(MemRefBeginIdx + X86::AddrBaseReg);
1815 MachineOperand &IndexMO =
1816 UseMI.getOperand(MemRefBeginIdx + X86::AddrIndexReg);
1817 if ((BaseMO.isReg() && BaseMO.getReg() == DefReg) ||
1818 (IndexMO.isReg() && IndexMO.getReg() == DefReg))
1819 // The load uses the register as part of its address making it not
1820 // invariant.
1821 return {};
1822
1823 continue;
1824 }
1825
1826 if (SingleUseMI)
1827 // We already have a single use, this would make two. Bail.
1828 return {};
1829
1830 // If this single use isn't data invariant, isn't in this block, or has
1831 // interfering EFLAGS, we can't sink the hardening to it.
1832 if (!X86InstrInfo::isDataInvariant(UseMI) || UseMI.getParent() != MI.getParent() ||
1833 isEFLAGSDefLive(UseMI))
1834 return {};
1835
1836 // If this instruction defines multiple registers bail as we won't harden
1837 // all of them.
1838 if (UseMI.getDesc().getNumDefs() > 1)
1839 return {};
1840
1841 // If this register isn't a virtual register we can't walk uses of sanely,
1842 // just bail. Also check that its register class is one of the ones we
1843 // can harden.
1844 Register UseDefReg = UseMI.getOperand(0).getReg();
1845 if (!Register::isVirtualRegister(UseDefReg) ||
1846 !canHardenRegister(UseDefReg))
1847 return {};
1848
1849 SingleUseMI = &UseMI;
1850 }
1851
1852 // If SingleUseMI is still null, there is no use that needs its own
1853 // checking. Otherwise, it is the single use that needs checking.
1854 return {SingleUseMI};
1855 };
1856
1857 MachineInstr *MI = &InitialMI;
1858 while (Optional<MachineInstr *> SingleUse = SinkCheckToSingleUse(*MI)) {
1859 // Update which MI we're checking now.
1860 MI = *SingleUse;
1861 if (!MI)
1862 break;
1863 }
1864
1865 return MI;
1866 }
1867
canHardenRegister(unsigned Reg)1868 bool X86SpeculativeLoadHardeningPass::canHardenRegister(unsigned Reg) {
1869 auto *RC = MRI->getRegClass(Reg);
1870 int RegBytes = TRI->getRegSizeInBits(*RC) / 8;
1871 if (RegBytes > 8)
1872 // We don't support post-load hardening of vectors.
1873 return false;
1874
1875 unsigned RegIdx = Log2_32(RegBytes);
1876 assert(RegIdx < 4 && "Unsupported register size");
1877
1878 // If this register class is explicitly constrained to a class that doesn't
1879 // require REX prefix, we may not be able to satisfy that constraint when
1880 // emitting the hardening instructions, so bail out here.
1881 // FIXME: This seems like a pretty lame hack. The way this comes up is when we
1882 // end up both with a NOREX and REX-only register as operands to the hardening
1883 // instructions. It would be better to fix that code to handle this situation
1884 // rather than hack around it in this way.
1885 const TargetRegisterClass *NOREXRegClasses[] = {
1886 &X86::GR8_NOREXRegClass, &X86::GR16_NOREXRegClass,
1887 &X86::GR32_NOREXRegClass, &X86::GR64_NOREXRegClass};
1888 if (RC == NOREXRegClasses[RegIdx])
1889 return false;
1890
1891 const TargetRegisterClass *GPRRegClasses[] = {
1892 &X86::GR8RegClass, &X86::GR16RegClass, &X86::GR32RegClass,
1893 &X86::GR64RegClass};
1894 return RC->hasSuperClassEq(GPRRegClasses[RegIdx]);
1895 }
1896
1897 /// Harden a value in a register.
1898 ///
1899 /// This is the low-level logic to fully harden a value sitting in a register
1900 /// against leaking during speculative execution.
1901 ///
1902 /// Unlike hardening an address that is used by a load, this routine is required
1903 /// to hide *all* incoming bits in the register.
1904 ///
1905 /// `Reg` must be a virtual register. Currently, it is required to be a GPR no
1906 /// larger than the predicate state register. FIXME: We should support vector
1907 /// registers here by broadcasting the predicate state.
1908 ///
1909 /// The new, hardened virtual register is returned. It will have the same
1910 /// register class as `Reg`.
hardenValueInRegister(unsigned Reg,MachineBasicBlock & MBB,MachineBasicBlock::iterator InsertPt,DebugLoc Loc)1911 unsigned X86SpeculativeLoadHardeningPass::hardenValueInRegister(
1912 unsigned Reg, MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertPt,
1913 DebugLoc Loc) {
1914 assert(canHardenRegister(Reg) && "Cannot harden this register!");
1915 assert(Register::isVirtualRegister(Reg) && "Cannot harden a physical register!");
1916
1917 auto *RC = MRI->getRegClass(Reg);
1918 int Bytes = TRI->getRegSizeInBits(*RC) / 8;
1919
1920 unsigned StateReg = PS->SSA.GetValueAtEndOfBlock(&MBB);
1921
1922 // FIXME: Need to teach this about 32-bit mode.
1923 if (Bytes != 8) {
1924 unsigned SubRegImms[] = {X86::sub_8bit, X86::sub_16bit, X86::sub_32bit};
1925 unsigned SubRegImm = SubRegImms[Log2_32(Bytes)];
1926 Register NarrowStateReg = MRI->createVirtualRegister(RC);
1927 BuildMI(MBB, InsertPt, Loc, TII->get(TargetOpcode::COPY), NarrowStateReg)
1928 .addReg(StateReg, 0, SubRegImm);
1929 StateReg = NarrowStateReg;
1930 }
1931
1932 unsigned FlagsReg = 0;
1933 if (isEFLAGSLive(MBB, InsertPt, *TRI))
1934 FlagsReg = saveEFLAGS(MBB, InsertPt, Loc);
1935
1936 Register NewReg = MRI->createVirtualRegister(RC);
1937 unsigned OrOpCodes[] = {X86::OR8rr, X86::OR16rr, X86::OR32rr, X86::OR64rr};
1938 unsigned OrOpCode = OrOpCodes[Log2_32(Bytes)];
1939 auto OrI = BuildMI(MBB, InsertPt, Loc, TII->get(OrOpCode), NewReg)
1940 .addReg(StateReg)
1941 .addReg(Reg);
1942 OrI->addRegisterDead(X86::EFLAGS, TRI);
1943 ++NumInstsInserted;
1944 LLVM_DEBUG(dbgs() << " Inserting or: "; OrI->dump(); dbgs() << "\n");
1945
1946 if (FlagsReg)
1947 restoreEFLAGS(MBB, InsertPt, Loc, FlagsReg);
1948
1949 return NewReg;
1950 }
1951
1952 /// Harden a load by hardening the loaded value in the defined register.
1953 ///
1954 /// We can harden a non-leaking load into a register without touching the
1955 /// address by just hiding all of the loaded bits during misspeculation. We use
1956 /// an `or` instruction to do this because we set up our poison value as all
1957 /// ones. And the goal is just for the loaded bits to not be exposed to
1958 /// execution and coercing them to one is sufficient.
1959 ///
1960 /// Returns the newly hardened register.
hardenPostLoad(MachineInstr & MI)1961 unsigned X86SpeculativeLoadHardeningPass::hardenPostLoad(MachineInstr &MI) {
1962 MachineBasicBlock &MBB = *MI.getParent();
1963 DebugLoc Loc = MI.getDebugLoc();
1964
1965 auto &DefOp = MI.getOperand(0);
1966 Register OldDefReg = DefOp.getReg();
1967 auto *DefRC = MRI->getRegClass(OldDefReg);
1968
1969 // Because we want to completely replace the uses of this def'ed value with
1970 // the hardened value, create a dedicated new register that will only be used
1971 // to communicate the unhardened value to the hardening.
1972 Register UnhardenedReg = MRI->createVirtualRegister(DefRC);
1973 DefOp.setReg(UnhardenedReg);
1974
1975 // Now harden this register's value, getting a hardened reg that is safe to
1976 // use. Note that we insert the instructions to compute this *after* the
1977 // defining instruction, not before it.
1978 unsigned HardenedReg = hardenValueInRegister(
1979 UnhardenedReg, MBB, std::next(MI.getIterator()), Loc);
1980
1981 // Finally, replace the old register (which now only has the uses of the
1982 // original def) with the hardened register.
1983 MRI->replaceRegWith(/*FromReg*/ OldDefReg, /*ToReg*/ HardenedReg);
1984
1985 ++NumPostLoadRegsHardened;
1986 return HardenedReg;
1987 }
1988
1989 /// Harden a return instruction.
1990 ///
1991 /// Returns implicitly perform a load which we need to harden. Without hardening
1992 /// this load, an attacker my speculatively write over the return address to
1993 /// steer speculation of the return to an attacker controlled address. This is
1994 /// called Spectre v1.1 or Bounds Check Bypass Store (BCBS) and is described in
1995 /// this paper:
1996 /// https://people.csail.mit.edu/vlk/spectre11.pdf
1997 ///
1998 /// We can harden this by introducing an LFENCE that will delay any load of the
1999 /// return address until prior instructions have retired (and thus are not being
2000 /// speculated), or we can harden the address used by the implicit load: the
2001 /// stack pointer.
2002 ///
2003 /// If we are not using an LFENCE, hardening the stack pointer has an additional
2004 /// benefit: it allows us to pass the predicate state accumulated in this
2005 /// function back to the caller. In the absence of a BCBS attack on the return,
2006 /// the caller will typically be resumed and speculatively executed due to the
2007 /// Return Stack Buffer (RSB) prediction which is very accurate and has a high
2008 /// priority. It is possible that some code from the caller will be executed
2009 /// speculatively even during a BCBS-attacked return until the steering takes
2010 /// effect. Whenever this happens, the caller can recover the (poisoned)
2011 /// predicate state from the stack pointer and continue to harden loads.
hardenReturnInstr(MachineInstr & MI)2012 void X86SpeculativeLoadHardeningPass::hardenReturnInstr(MachineInstr &MI) {
2013 MachineBasicBlock &MBB = *MI.getParent();
2014 DebugLoc Loc = MI.getDebugLoc();
2015 auto InsertPt = MI.getIterator();
2016
2017 if (FenceCallAndRet)
2018 // No need to fence here as we'll fence at the return site itself. That
2019 // handles more cases than we can handle here.
2020 return;
2021
2022 // Take our predicate state, shift it to the high 17 bits (so that we keep
2023 // pointers canonical) and merge it into RSP. This will allow the caller to
2024 // extract it when we return (speculatively).
2025 mergePredStateIntoSP(MBB, InsertPt, Loc, PS->SSA.GetValueAtEndOfBlock(&MBB));
2026 }
2027
2028 /// Trace the predicate state through a call.
2029 ///
2030 /// There are several layers of this needed to handle the full complexity of
2031 /// calls.
2032 ///
2033 /// First, we need to send the predicate state into the called function. We do
2034 /// this by merging it into the high bits of the stack pointer.
2035 ///
2036 /// For tail calls, this is all we need to do.
2037 ///
2038 /// For calls where we might return and resume the control flow, we need to
2039 /// extract the predicate state from the high bits of the stack pointer after
2040 /// control returns from the called function.
2041 ///
2042 /// We also need to verify that we intended to return to this location in the
2043 /// code. An attacker might arrange for the processor to mispredict the return
2044 /// to this valid but incorrect return address in the program rather than the
2045 /// correct one. See the paper on this attack, called "ret2spec" by the
2046 /// researchers, here:
2047 /// https://christian-rossow.de/publications/ret2spec-ccs2018.pdf
2048 ///
2049 /// The way we verify that we returned to the correct location is by preserving
2050 /// the expected return address across the call. One technique involves taking
2051 /// advantage of the red-zone to load the return address from `8(%rsp)` where it
2052 /// was left by the RET instruction when it popped `%rsp`. Alternatively, we can
2053 /// directly save the address into a register that will be preserved across the
2054 /// call. We compare this intended return address against the address
2055 /// immediately following the call (the observed return address). If these
2056 /// mismatch, we have detected misspeculation and can poison our predicate
2057 /// state.
tracePredStateThroughCall(MachineInstr & MI)2058 void X86SpeculativeLoadHardeningPass::tracePredStateThroughCall(
2059 MachineInstr &MI) {
2060 MachineBasicBlock &MBB = *MI.getParent();
2061 MachineFunction &MF = *MBB.getParent();
2062 auto InsertPt = MI.getIterator();
2063 DebugLoc Loc = MI.getDebugLoc();
2064
2065 if (FenceCallAndRet) {
2066 if (MI.isReturn())
2067 // Tail call, we don't return to this function.
2068 // FIXME: We should also handle noreturn calls.
2069 return;
2070
2071 // We don't need to fence before the call because the function should fence
2072 // in its entry. However, we do need to fence after the call returns.
2073 // Fencing before the return doesn't correctly handle cases where the return
2074 // itself is mispredicted.
2075 BuildMI(MBB, std::next(InsertPt), Loc, TII->get(X86::LFENCE));
2076 ++NumInstsInserted;
2077 ++NumLFENCEsInserted;
2078 return;
2079 }
2080
2081 // First, we transfer the predicate state into the called function by merging
2082 // it into the stack pointer. This will kill the current def of the state.
2083 unsigned StateReg = PS->SSA.GetValueAtEndOfBlock(&MBB);
2084 mergePredStateIntoSP(MBB, InsertPt, Loc, StateReg);
2085
2086 // If this call is also a return, it is a tail call and we don't need anything
2087 // else to handle it so just return. Also, if there are no further
2088 // instructions and no successors, this call does not return so we can also
2089 // bail.
2090 if (MI.isReturn() || (std::next(InsertPt) == MBB.end() && MBB.succ_empty()))
2091 return;
2092
2093 // Create a symbol to track the return address and attach it to the call
2094 // machine instruction. We will lower extra symbols attached to call
2095 // instructions as label immediately following the call.
2096 MCSymbol *RetSymbol =
2097 MF.getContext().createTempSymbol("slh_ret_addr",
2098 /*AlwaysAddSuffix*/ true);
2099 MI.setPostInstrSymbol(MF, RetSymbol);
2100
2101 const TargetRegisterClass *AddrRC = &X86::GR64RegClass;
2102 unsigned ExpectedRetAddrReg = 0;
2103
2104 // If we have no red zones or if the function returns twice (possibly without
2105 // using the `ret` instruction) like setjmp, we need to save the expected
2106 // return address prior to the call.
2107 if (!Subtarget->getFrameLowering()->has128ByteRedZone(MF) ||
2108 MF.exposesReturnsTwice()) {
2109 // If we don't have red zones, we need to compute the expected return
2110 // address prior to the call and store it in a register that lives across
2111 // the call.
2112 //
2113 // In some ways, this is doubly satisfying as a mitigation because it will
2114 // also successfully detect stack smashing bugs in some cases (typically,
2115 // when a callee-saved register is used and the callee doesn't push it onto
2116 // the stack). But that isn't our primary goal, so we only use it as
2117 // a fallback.
2118 //
2119 // FIXME: It isn't clear that this is reliable in the face of
2120 // rematerialization in the register allocator. We somehow need to force
2121 // that to not occur for this particular instruction, and instead to spill
2122 // or otherwise preserve the value computed *prior* to the call.
2123 //
2124 // FIXME: It is even less clear why MachineCSE can't just fold this when we
2125 // end up having to use identical instructions both before and after the
2126 // call to feed the comparison.
2127 ExpectedRetAddrReg = MRI->createVirtualRegister(AddrRC);
2128 if (MF.getTarget().getCodeModel() == CodeModel::Small &&
2129 !Subtarget->isPositionIndependent()) {
2130 BuildMI(MBB, InsertPt, Loc, TII->get(X86::MOV64ri32), ExpectedRetAddrReg)
2131 .addSym(RetSymbol);
2132 } else {
2133 BuildMI(MBB, InsertPt, Loc, TII->get(X86::LEA64r), ExpectedRetAddrReg)
2134 .addReg(/*Base*/ X86::RIP)
2135 .addImm(/*Scale*/ 1)
2136 .addReg(/*Index*/ 0)
2137 .addSym(RetSymbol)
2138 .addReg(/*Segment*/ 0);
2139 }
2140 }
2141
2142 // Step past the call to handle when it returns.
2143 ++InsertPt;
2144
2145 // If we didn't pre-compute the expected return address into a register, then
2146 // red zones are enabled and the return address is still available on the
2147 // stack immediately after the call. As the very first instruction, we load it
2148 // into a register.
2149 if (!ExpectedRetAddrReg) {
2150 ExpectedRetAddrReg = MRI->createVirtualRegister(AddrRC);
2151 BuildMI(MBB, InsertPt, Loc, TII->get(X86::MOV64rm), ExpectedRetAddrReg)
2152 .addReg(/*Base*/ X86::RSP)
2153 .addImm(/*Scale*/ 1)
2154 .addReg(/*Index*/ 0)
2155 .addImm(/*Displacement*/ -8) // The stack pointer has been popped, so
2156 // the return address is 8-bytes past it.
2157 .addReg(/*Segment*/ 0);
2158 }
2159
2160 // Now we extract the callee's predicate state from the stack pointer.
2161 unsigned NewStateReg = extractPredStateFromSP(MBB, InsertPt, Loc);
2162
2163 // Test the expected return address against our actual address. If we can
2164 // form this basic block's address as an immediate, this is easy. Otherwise
2165 // we compute it.
2166 if (MF.getTarget().getCodeModel() == CodeModel::Small &&
2167 !Subtarget->isPositionIndependent()) {
2168 // FIXME: Could we fold this with the load? It would require careful EFLAGS
2169 // management.
2170 BuildMI(MBB, InsertPt, Loc, TII->get(X86::CMP64ri32))
2171 .addReg(ExpectedRetAddrReg, RegState::Kill)
2172 .addSym(RetSymbol);
2173 } else {
2174 Register ActualRetAddrReg = MRI->createVirtualRegister(AddrRC);
2175 BuildMI(MBB, InsertPt, Loc, TII->get(X86::LEA64r), ActualRetAddrReg)
2176 .addReg(/*Base*/ X86::RIP)
2177 .addImm(/*Scale*/ 1)
2178 .addReg(/*Index*/ 0)
2179 .addSym(RetSymbol)
2180 .addReg(/*Segment*/ 0);
2181 BuildMI(MBB, InsertPt, Loc, TII->get(X86::CMP64rr))
2182 .addReg(ExpectedRetAddrReg, RegState::Kill)
2183 .addReg(ActualRetAddrReg, RegState::Kill);
2184 }
2185
2186 // Now conditionally update the predicate state we just extracted if we ended
2187 // up at a different return address than expected.
2188 int PredStateSizeInBytes = TRI->getRegSizeInBits(*PS->RC) / 8;
2189 auto CMovOp = X86::getCMovOpcode(PredStateSizeInBytes);
2190
2191 Register UpdatedStateReg = MRI->createVirtualRegister(PS->RC);
2192 auto CMovI = BuildMI(MBB, InsertPt, Loc, TII->get(CMovOp), UpdatedStateReg)
2193 .addReg(NewStateReg, RegState::Kill)
2194 .addReg(PS->PoisonReg)
2195 .addImm(X86::COND_NE);
2196 CMovI->findRegisterUseOperand(X86::EFLAGS)->setIsKill(true);
2197 ++NumInstsInserted;
2198 LLVM_DEBUG(dbgs() << " Inserting cmov: "; CMovI->dump(); dbgs() << "\n");
2199
2200 PS->SSA.AddAvailableValue(&MBB, UpdatedStateReg);
2201 }
2202
2203 /// An attacker may speculatively store over a value that is then speculatively
2204 /// loaded and used as the target of an indirect call or jump instruction. This
2205 /// is called Spectre v1.2 or Bounds Check Bypass Store (BCBS) and is described
2206 /// in this paper:
2207 /// https://people.csail.mit.edu/vlk/spectre11.pdf
2208 ///
2209 /// When this happens, the speculative execution of the call or jump will end up
2210 /// being steered to this attacker controlled address. While most such loads
2211 /// will be adequately hardened already, we want to ensure that they are
2212 /// definitively treated as needing post-load hardening. While address hardening
2213 /// is sufficient to prevent secret data from leaking to the attacker, it may
2214 /// not be sufficient to prevent an attacker from steering speculative
2215 /// execution. We forcibly unfolded all relevant loads above and so will always
2216 /// have an opportunity to post-load harden here, we just need to scan for cases
2217 /// not already flagged and add them.
hardenIndirectCallOrJumpInstr(MachineInstr & MI,SmallDenseMap<unsigned,unsigned,32> & AddrRegToHardenedReg)2218 void X86SpeculativeLoadHardeningPass::hardenIndirectCallOrJumpInstr(
2219 MachineInstr &MI,
2220 SmallDenseMap<unsigned, unsigned, 32> &AddrRegToHardenedReg) {
2221 switch (MI.getOpcode()) {
2222 case X86::FARCALL16m:
2223 case X86::FARCALL32m:
2224 case X86::FARCALL64m:
2225 case X86::FARJMP16m:
2226 case X86::FARJMP32m:
2227 case X86::FARJMP64m:
2228 // We don't need to harden either far calls or far jumps as they are
2229 // safe from Spectre.
2230 return;
2231
2232 default:
2233 break;
2234 }
2235
2236 // We should never see a loading instruction at this point, as those should
2237 // have been unfolded.
2238 assert(!MI.mayLoad() && "Found a lingering loading instruction!");
2239
2240 // If the first operand isn't a register, this is a branch or call
2241 // instruction with an immediate operand which doesn't need to be hardened.
2242 if (!MI.getOperand(0).isReg())
2243 return;
2244
2245 // For all of these, the target register is the first operand of the
2246 // instruction.
2247 auto &TargetOp = MI.getOperand(0);
2248 Register OldTargetReg = TargetOp.getReg();
2249
2250 // Try to lookup a hardened version of this register. We retain a reference
2251 // here as we want to update the map to track any newly computed hardened
2252 // register.
2253 unsigned &HardenedTargetReg = AddrRegToHardenedReg[OldTargetReg];
2254
2255 // If we don't have a hardened register yet, compute one. Otherwise, just use
2256 // the already hardened register.
2257 //
2258 // FIXME: It is a little suspect that we use partially hardened registers that
2259 // only feed addresses. The complexity of partial hardening with SHRX
2260 // continues to pile up. Should definitively measure its value and consider
2261 // eliminating it.
2262 if (!HardenedTargetReg)
2263 HardenedTargetReg = hardenValueInRegister(
2264 OldTargetReg, *MI.getParent(), MI.getIterator(), MI.getDebugLoc());
2265
2266 // Set the target operand to the hardened register.
2267 TargetOp.setReg(HardenedTargetReg);
2268
2269 ++NumCallsOrJumpsHardened;
2270 }
2271
2272 INITIALIZE_PASS_BEGIN(X86SpeculativeLoadHardeningPass, PASS_KEY,
2273 "X86 speculative load hardener", false, false)
2274 INITIALIZE_PASS_END(X86SpeculativeLoadHardeningPass, PASS_KEY,
2275 "X86 speculative load hardener", false, false)
2276
createX86SpeculativeLoadHardeningPass()2277 FunctionPass *llvm::createX86SpeculativeLoadHardeningPass() {
2278 return new X86SpeculativeLoadHardeningPass();
2279 }
2280