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