1//===- BoundsChecking.cpp - Instrumentation for run-time bounds checking --===//
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
9#include "llvm/Transforms/Instrumentation/BoundsChecking.h"
10#include "llvm/ADT/Statistic.h"
11#include "llvm/ADT/StringRef.h"
12#include "llvm/ADT/Twine.h"
13#include "llvm/Analysis/MemoryBuiltins.h"
14#include "llvm/Analysis/ScalarEvolution.h"
15#include "llvm/Analysis/TargetFolder.h"
16#include "llvm/Analysis/TargetLibraryInfo.h"
17#include "llvm/IR/BasicBlock.h"
18#include "llvm/IR/Constants.h"
19#include "llvm/IR/DataLayout.h"
20#include "llvm/IR/Function.h"
21#include "llvm/IR/IRBuilder.h"
22#include "llvm/IR/InstIterator.h"
23#include "llvm/IR/Instruction.h"
24#include "llvm/IR/Instructions.h"
25#include "llvm/IR/Intrinsics.h"
26#include "llvm/IR/Value.h"
27#include "llvm/Support/Casting.h"
28#include "llvm/Support/CommandLine.h"
29#include "llvm/Support/Debug.h"
30#include "llvm/Support/raw_ostream.h"
31#include <utility>
32
33using namespace llvm;
34
35#define DEBUG_TYPE "bounds-checking"
36
37static cl::opt<bool> SingleTrapBB("bounds-checking-single-trap",
38 cl::desc("Use one trap block per function"));
39
40STATISTIC(ChecksAdded, "Bounds checks added");
41STATISTIC(ChecksSkipped, "Bounds checks skipped");
42STATISTIC(ChecksUnable, "Bounds checks unable to add");
43
44class BuilderTy : public IRBuilder<TargetFolder> {
45public:
46 BuilderTy(BasicBlock *TheBB, BasicBlock::iterator IP, TargetFolder Folder)
47 : IRBuilder<TargetFolder>(TheBB, IP, Folder) {
48 SetNoSanitizeMetadata();
49 }
50};
51
52/// Gets the conditions under which memory accessing instructions will overflow.
53///
54/// \p Ptr is the pointer that will be read/written, and \p InstVal is either
55/// the result from the load or the value being stored. It is used to determine
56/// the size of memory block that is touched.
57///
58/// Returns the condition under which the access will overflow.
59static Value *getBoundsCheckCond(Value *Ptr, Value *InstVal,
60 const DataLayout &DL, TargetLibraryInfo &TLI,
61 ObjectSizeOffsetEvaluator &ObjSizeEval,
62 BuilderTy &IRB, ScalarEvolution &SE) {
63 TypeSize NeededSize = DL.getTypeStoreSize(Ty: InstVal->getType());
64 LLVM_DEBUG(dbgs() << "Instrument " << *Ptr << " for " << Twine(NeededSize)
65 << " bytes\n");
66
67 SizeOffsetValue SizeOffset = ObjSizeEval.compute(V: Ptr);
68
69 if (!SizeOffset.bothKnown()) {
70 ++ChecksUnable;
71 return nullptr;
72 }
73
74 Value *Size = SizeOffset.Size;
75 Value *Offset = SizeOffset.Offset;
76 ConstantInt *SizeCI = dyn_cast<ConstantInt>(Val: Size);
77
78 Type *IndexTy = DL.getIndexType(PtrTy: Ptr->getType());
79 Value *NeededSizeVal = IRB.CreateTypeSize(Ty: IndexTy, Size: NeededSize);
80
81 auto SizeRange = SE.getUnsignedRange(S: SE.getSCEV(V: Size));
82 auto OffsetRange = SE.getUnsignedRange(S: SE.getSCEV(V: Offset));
83 auto NeededSizeRange = SE.getUnsignedRange(S: SE.getSCEV(V: NeededSizeVal));
84
85 // three checks are required to ensure safety:
86 // . Offset >= 0 (since the offset is given from the base ptr)
87 // . Size >= Offset (unsigned)
88 // . Size - Offset >= NeededSize (unsigned)
89 //
90 // optimization: if Size >= 0 (signed), skip 1st check
91 // FIXME: add NSW/NUW here? -- we dont care if the subtraction overflows
92 Value *ObjSize = IRB.CreateSub(LHS: Size, RHS: Offset);
93 Value *Cmp2 = SizeRange.getUnsignedMin().uge(RHS: OffsetRange.getUnsignedMax())
94 ? ConstantInt::getFalse(Context&: Ptr->getContext())
95 : IRB.CreateICmpULT(LHS: Size, RHS: Offset);
96 Value *Cmp3 = SizeRange.sub(Other: OffsetRange)
97 .getUnsignedMin()
98 .uge(RHS: NeededSizeRange.getUnsignedMax())
99 ? ConstantInt::getFalse(Context&: Ptr->getContext())
100 : IRB.CreateICmpULT(LHS: ObjSize, RHS: NeededSizeVal);
101 Value *Or = IRB.CreateOr(LHS: Cmp2, RHS: Cmp3);
102 if ((!SizeCI || SizeCI->getValue().slt(RHS: 0)) &&
103 !SizeRange.getSignedMin().isNonNegative()) {
104 Value *Cmp1 = IRB.CreateICmpSLT(LHS: Offset, RHS: ConstantInt::get(Ty: IndexTy, V: 0));
105 Or = IRB.CreateOr(LHS: Cmp1, RHS: Or);
106 }
107
108 return Or;
109}
110
111static CallInst *InsertTrap(BuilderTy &IRB, bool DebugTrapBB,
112 std::optional<int8_t> GuardKind) {
113 if (!DebugTrapBB)
114 return IRB.CreateIntrinsic(ID: Intrinsic::trap, Args: {});
115
116 uint64_t ImmArg = GuardKind.has_value()
117 ? GuardKind.value()
118 : IRB.GetInsertBlock()->getParent()->size();
119 // Ensure we constrain ImmArg to fitting within a 8-but unsigned integer to
120 // prevent overflow.
121 if (ImmArg > 255)
122 ImmArg = 255;
123
124 return IRB.CreateIntrinsic(ID: Intrinsic::ubsantrap,
125 Args: ConstantInt::get(Ty: IRB.getInt8Ty(), V: ImmArg));
126}
127
128static CallInst *InsertCall(BuilderTy &IRB, bool MayReturn, StringRef Name) {
129 Function *Fn = IRB.GetInsertBlock()->getParent();
130 LLVMContext &Ctx = Fn->getContext();
131 llvm::AttrBuilder B(Ctx);
132 B.addAttribute(Val: llvm::Attribute::NoUnwind);
133 if (!MayReturn)
134 B.addAttribute(Val: llvm::Attribute::NoReturn);
135 FunctionCallee Callee = Fn->getParent()->getOrInsertFunction(
136 Name,
137 AttributeList: llvm::AttributeList::get(C&: Ctx, Index: llvm::AttributeList::FunctionIndex, B),
138 RetTy: Type::getVoidTy(C&: Ctx));
139 return IRB.CreateCall(Callee);
140}
141
142/// Adds run-time bounds checks to memory accessing instructions.
143///
144/// \p Or is the condition that should guard the trap.
145///
146/// \p GetTrapBB is a callable that returns the trap BB to use on failure.
147template <typename GetTrapBBT>
148static void insertBoundsCheck(Value *Or, BuilderTy &IRB, GetTrapBBT GetTrapBB) {
149 // check if the comparison is always false
150 ConstantInt *C = dyn_cast_or_null<ConstantInt>(Val: Or);
151 if (C) {
152 ++ChecksSkipped;
153 // If non-zero, nothing to do.
154 if (!C->getZExtValue())
155 return;
156 }
157 ++ChecksAdded;
158
159 BasicBlock::iterator SplitI = IRB.GetInsertPoint();
160 BasicBlock *OldBB = SplitI->getParent();
161 BasicBlock *Cont = OldBB->splitBasicBlock(I: SplitI);
162 OldBB->getTerminator()->eraseFromParent();
163
164 BasicBlock *TrapBB = GetTrapBB(IRB, Cont);
165
166 if (C) {
167 // If we have a constant zero, unconditionally branch.
168 // FIXME: We should really handle this differently to bypass the splitting
169 // the block.
170 BranchInst::Create(IfTrue: TrapBB, InsertBefore: OldBB);
171 return;
172 }
173
174 // Create the conditional branch.
175 BranchInst::Create(IfTrue: TrapBB, IfFalse: Cont, Cond: Or, InsertBefore: OldBB);
176}
177
178static std::string
179getRuntimeCallName(const BoundsCheckingPass::Options::Runtime &Opts) {
180 std::string Name = "__ubsan_handle_local_out_of_bounds";
181 if (Opts.MinRuntime)
182 Name += "_minimal";
183 if (!Opts.MayReturn)
184 Name += "_abort";
185 else if (Opts.HandlerPreserveAllRegs)
186 Name += "_preserve";
187 return Name;
188}
189
190static bool addBoundsChecking(Function &F, TargetLibraryInfo &TLI,
191 ScalarEvolution &SE,
192 const BoundsCheckingPass::Options &Opts) {
193 if (F.hasFnAttribute(Kind: Attribute::NoSanitizeBounds))
194 return false;
195
196 const DataLayout &DL = F.getDataLayout();
197 ObjectSizeOpts EvalOpts;
198 EvalOpts.RoundToAlign = true;
199 EvalOpts.EvalMode = ObjectSizeOpts::Mode::ExactUnderlyingSizeAndOffset;
200 ObjectSizeOffsetEvaluator ObjSizeEval(DL, &TLI, F.getContext(), EvalOpts);
201
202 // check HANDLE_MEMORY_INST in include/llvm/Instruction.def for memory
203 // touching instructions
204 SmallVector<std::pair<Instruction *, Value *>, 4> TrapInfo;
205 for (Instruction &I : instructions(F)) {
206 Value *Or = nullptr;
207 BuilderTy IRB(I.getParent(), BasicBlock::iterator(&I), TargetFolder(DL));
208 if (LoadInst *LI = dyn_cast<LoadInst>(Val: &I)) {
209 if (!LI->isVolatile())
210 Or = getBoundsCheckCond(Ptr: LI->getPointerOperand(), InstVal: LI, DL, TLI,
211 ObjSizeEval, IRB, SE);
212 } else if (StoreInst *SI = dyn_cast<StoreInst>(Val: &I)) {
213 if (!SI->isVolatile())
214 Or = getBoundsCheckCond(Ptr: SI->getPointerOperand(), InstVal: SI->getValueOperand(),
215 DL, TLI, ObjSizeEval, IRB, SE);
216 } else if (AtomicCmpXchgInst *AI = dyn_cast<AtomicCmpXchgInst>(Val: &I)) {
217 if (!AI->isVolatile())
218 Or =
219 getBoundsCheckCond(Ptr: AI->getPointerOperand(), InstVal: AI->getCompareOperand(),
220 DL, TLI, ObjSizeEval, IRB, SE);
221 } else if (AtomicRMWInst *AI = dyn_cast<AtomicRMWInst>(Val: &I)) {
222 if (!AI->isVolatile())
223 Or = getBoundsCheckCond(Ptr: AI->getPointerOperand(), InstVal: AI->getValOperand(),
224 DL, TLI, ObjSizeEval, IRB, SE);
225 }
226 if (Or) {
227 if (Opts.GuardKind) {
228 llvm::Value *Allow = IRB.CreateIntrinsic(
229 RetTy: IRB.getInt1Ty(), ID: Intrinsic::allow_ubsan_check,
230 Args: {llvm::ConstantInt::getSigned(Ty: IRB.getInt8Ty(), V: *Opts.GuardKind)});
231 Or = IRB.CreateAnd(LHS: Or, RHS: Allow);
232 }
233 TrapInfo.push_back(Elt: std::make_pair(x: &I, y&: Or));
234 }
235 }
236
237 std::string Name;
238 if (Opts.Rt)
239 Name = getRuntimeCallName(Opts: *Opts.Rt);
240
241 // Create a trapping basic block on demand using a callback. Depending on
242 // flags, this will either create a single block for the entire function or
243 // will create a fresh block every time it is called.
244 BasicBlock *ReuseTrapBB = nullptr;
245 auto GetTrapBB = [&ReuseTrapBB, &Opts, &Name](BuilderTy &IRB,
246 BasicBlock *Cont) {
247 Function *Fn = IRB.GetInsertBlock()->getParent();
248 auto DebugLoc = IRB.getCurrentDebugLocation();
249 IRBuilder<>::InsertPointGuard Guard(IRB);
250
251 // Create a trapping basic block on demand using a callback. Depending on
252 // flags, this will either create a single block for the entire function or
253 // will create a fresh block every time it is called.
254 if (ReuseTrapBB)
255 return ReuseTrapBB;
256
257 BasicBlock *TrapBB = BasicBlock::Create(Context&: Fn->getContext(), Name: "trap", Parent: Fn);
258 IRB.SetInsertPoint(TrapBB);
259
260 bool DebugTrapBB = !Opts.Merge;
261 CallInst *TrapCall = Opts.Rt ? InsertCall(IRB, MayReturn: Opts.Rt->MayReturn, Name)
262 : InsertTrap(IRB, DebugTrapBB, GuardKind: Opts.GuardKind);
263 if (DebugTrapBB)
264 TrapCall->addFnAttr(Kind: llvm::Attribute::NoMerge);
265
266 TrapCall->setDoesNotThrow();
267 TrapCall->setDebugLoc(DebugLoc);
268
269 bool MayReturn = Opts.Rt && Opts.Rt->MayReturn;
270 if (MayReturn) {
271 IRB.CreateBr(Dest: Cont);
272 } else {
273 TrapCall->setDoesNotReturn();
274 IRB.CreateUnreachable();
275 }
276 // The preserve-all logic is somewhat duplicated in CGExpr.cpp for
277 // local-bounds. Make sure to change that too.
278 if (Opts.Rt && Opts.Rt->HandlerPreserveAllRegs && MayReturn)
279 TrapCall->setCallingConv(CallingConv::PreserveAll);
280 if (!MayReturn && SingleTrapBB && !DebugTrapBB)
281 ReuseTrapBB = TrapBB;
282
283 return TrapBB;
284 };
285
286 for (const auto &Entry : TrapInfo) {
287 Instruction *Inst = Entry.first;
288 BuilderTy IRB(Inst->getParent(), BasicBlock::iterator(Inst), TargetFolder(DL));
289 insertBoundsCheck(Or: Entry.second, IRB, GetTrapBB);
290 }
291
292 return !TrapInfo.empty();
293}
294
295PreservedAnalyses BoundsCheckingPass::run(Function &F, FunctionAnalysisManager &AM) {
296 auto &TLI = AM.getResult<TargetLibraryAnalysis>(IR&: F);
297 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(IR&: F);
298
299 if (!addBoundsChecking(F, TLI, SE, Opts))
300 return PreservedAnalyses::all();
301
302 return PreservedAnalyses::none();
303}
304
305void BoundsCheckingPass::printPipeline(
306 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
307 static_cast<PassInfoMixin<BoundsCheckingPass> *>(this)->printPipeline(
308 OS, MapClassName2PassName);
309 OS << "<";
310 if (Opts.Rt) {
311 if (Opts.Rt->MinRuntime)
312 OS << "min-";
313 OS << "rt";
314 if (!Opts.Rt->MayReturn)
315 OS << "-abort";
316 } else {
317 OS << "trap";
318 }
319 if (Opts.Merge)
320 OS << ";merge";
321 if (Opts.GuardKind)
322 OS << ";guard=" << static_cast<int>(*Opts.GuardKind);
323 OS << ">";
324}
325