1 //===-- StackProtector.cpp - Stack Protector Insertion --------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass inserts stack protectors into functions which need them. A variable
11 // with a random value in it is stored onto the stack before the local variables
12 // are allocated. Upon exiting the block, the stored value is checked. If it's
13 // changed, then there was some sort of violation and the program aborts.
14 //
15 //===----------------------------------------------------------------------===//
16 
17 #include "llvm/CodeGen/StackProtector.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/BranchProbabilityInfo.h"
21 #include "llvm/Analysis/ValueTracking.h"
22 #include "llvm/CodeGen/Analysis.h"
23 #include "llvm/CodeGen/Passes.h"
24 #include "llvm/IR/Attributes.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/GlobalValue.h"
30 #include "llvm/IR/GlobalVariable.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/Intrinsics.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Module.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Target/TargetSubtargetInfo.h"
39 #include <cstdlib>
40 using namespace llvm;
41 
42 #define DEBUG_TYPE "stack-protector"
43 
44 STATISTIC(NumFunProtected, "Number of functions protected");
45 STATISTIC(NumAddrTaken, "Number of local variables that have their address"
46                         " taken.");
47 
48 static cl::opt<bool> EnableSelectionDAGSP("enable-selectiondag-sp",
49                                           cl::init(true), cl::Hidden);
50 
51 char StackProtector::ID = 0;
52 INITIALIZE_PASS(StackProtector, "stack-protector", "Insert stack protectors",
53                 false, true)
54 
createStackProtectorPass(const TargetMachine * TM)55 FunctionPass *llvm::createStackProtectorPass(const TargetMachine *TM) {
56   return new StackProtector(TM);
57 }
58 
59 StackProtector::SSPLayoutKind
getSSPLayout(const AllocaInst * AI) const60 StackProtector::getSSPLayout(const AllocaInst *AI) const {
61   return AI ? Layout.lookup(AI) : SSPLK_None;
62 }
63 
adjustForColoring(const AllocaInst * From,const AllocaInst * To)64 void StackProtector::adjustForColoring(const AllocaInst *From,
65                                        const AllocaInst *To) {
66   // When coloring replaces one alloca with another, transfer the SSPLayoutKind
67   // tag from the remapped to the target alloca. The remapped alloca should
68   // have a size smaller than or equal to the replacement alloca.
69   SSPLayoutMap::iterator I = Layout.find(From);
70   if (I != Layout.end()) {
71     SSPLayoutKind Kind = I->second;
72     Layout.erase(I);
73 
74     // Transfer the tag, but make sure that SSPLK_AddrOf does not overwrite
75     // SSPLK_SmallArray or SSPLK_LargeArray, and make sure that
76     // SSPLK_SmallArray does not overwrite SSPLK_LargeArray.
77     I = Layout.find(To);
78     if (I == Layout.end())
79       Layout.insert(std::make_pair(To, Kind));
80     else if (I->second != SSPLK_LargeArray && Kind != SSPLK_AddrOf)
81       I->second = Kind;
82   }
83 }
84 
runOnFunction(Function & Fn)85 bool StackProtector::runOnFunction(Function &Fn) {
86   F = &Fn;
87   M = F->getParent();
88   DominatorTreeWrapperPass *DTWP =
89       getAnalysisIfAvailable<DominatorTreeWrapperPass>();
90   DT = DTWP ? &DTWP->getDomTree() : nullptr;
91   TLI = TM->getSubtargetImpl(Fn)->getTargetLowering();
92 
93   Attribute Attr = Fn.getFnAttribute("stack-protector-buffer-size");
94   if (Attr.isStringAttribute() &&
95       Attr.getValueAsString().getAsInteger(10, SSPBufferSize))
96       return false; // Invalid integer string
97 
98   if (!RequiresStackProtector())
99     return false;
100 
101   ++NumFunProtected;
102   return InsertStackProtectors();
103 }
104 
105 /// \param [out] IsLarge is set to true if a protectable array is found and
106 /// it is "large" ( >= ssp-buffer-size).  In the case of a structure with
107 /// multiple arrays, this gets set if any of them is large.
ContainsProtectableArray(Type * Ty,bool & IsLarge,bool Strong,bool InStruct) const108 bool StackProtector::ContainsProtectableArray(Type *Ty, bool &IsLarge,
109                                               bool Strong,
110                                               bool InStruct) const {
111   if (!Ty)
112     return false;
113   if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
114     if (!AT->getElementType()->isIntegerTy(8)) {
115       // If we're on a non-Darwin platform or we're inside of a structure, don't
116       // add stack protectors unless the array is a character array.
117       // However, in strong mode any array, regardless of type and size,
118       // triggers a protector.
119       if (!Strong && (InStruct || !Trip.isOSDarwin()))
120         return false;
121     }
122 
123     // If an array has more than SSPBufferSize bytes of allocated space, then we
124     // emit stack protectors.
125     if (SSPBufferSize <= TLI->getDataLayout()->getTypeAllocSize(AT)) {
126       IsLarge = true;
127       return true;
128     }
129 
130     if (Strong)
131       // Require a protector for all arrays in strong mode
132       return true;
133   }
134 
135   const StructType *ST = dyn_cast<StructType>(Ty);
136   if (!ST)
137     return false;
138 
139   bool NeedsProtector = false;
140   for (StructType::element_iterator I = ST->element_begin(),
141                                     E = ST->element_end();
142        I != E; ++I)
143     if (ContainsProtectableArray(*I, IsLarge, Strong, true)) {
144       // If the element is a protectable array and is large (>= SSPBufferSize)
145       // then we are done.  If the protectable array is not large, then
146       // keep looking in case a subsequent element is a large array.
147       if (IsLarge)
148         return true;
149       NeedsProtector = true;
150     }
151 
152   return NeedsProtector;
153 }
154 
HasAddressTaken(const Instruction * AI)155 bool StackProtector::HasAddressTaken(const Instruction *AI) {
156   for (const User *U : AI->users()) {
157     if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
158       if (AI == SI->getValueOperand())
159         return true;
160     } else if (const PtrToIntInst *SI = dyn_cast<PtrToIntInst>(U)) {
161       if (AI == SI->getOperand(0))
162         return true;
163     } else if (isa<CallInst>(U)) {
164       return true;
165     } else if (isa<InvokeInst>(U)) {
166       return true;
167     } else if (const SelectInst *SI = dyn_cast<SelectInst>(U)) {
168       if (HasAddressTaken(SI))
169         return true;
170     } else if (const PHINode *PN = dyn_cast<PHINode>(U)) {
171       // Keep track of what PHI nodes we have already visited to ensure
172       // they are only visited once.
173       if (VisitedPHIs.insert(PN).second)
174         if (HasAddressTaken(PN))
175           return true;
176     } else if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
177       if (HasAddressTaken(GEP))
178         return true;
179     } else if (const BitCastInst *BI = dyn_cast<BitCastInst>(U)) {
180       if (HasAddressTaken(BI))
181         return true;
182     }
183   }
184   return false;
185 }
186 
187 /// \brief Check whether or not this function needs a stack protector based
188 /// upon the stack protector level.
189 ///
190 /// We use two heuristics: a standard (ssp) and strong (sspstrong).
191 /// The standard heuristic which will add a guard variable to functions that
192 /// call alloca with a either a variable size or a size >= SSPBufferSize,
193 /// functions with character buffers larger than SSPBufferSize, and functions
194 /// with aggregates containing character buffers larger than SSPBufferSize. The
195 /// strong heuristic will add a guard variables to functions that call alloca
196 /// regardless of size, functions with any buffer regardless of type and size,
197 /// functions with aggregates that contain any buffer regardless of type and
198 /// size, and functions that contain stack-based variables that have had their
199 /// address taken.
RequiresStackProtector()200 bool StackProtector::RequiresStackProtector() {
201   bool Strong = false;
202   bool NeedsProtector = false;
203   if (F->hasFnAttribute(Attribute::StackProtectReq)) {
204     NeedsProtector = true;
205     Strong = true; // Use the same heuristic as strong to determine SSPLayout
206   } else if (F->hasFnAttribute(Attribute::StackProtectStrong))
207     Strong = true;
208   else if (!F->hasFnAttribute(Attribute::StackProtect))
209     return false;
210 
211   for (const BasicBlock &BB : *F) {
212     for (const Instruction &I : BB) {
213       if (const AllocaInst *AI = dyn_cast<AllocaInst>(&I)) {
214         if (AI->isArrayAllocation()) {
215           // SSP-Strong: Enable protectors for any call to alloca, regardless
216           // of size.
217           if (Strong)
218             return true;
219 
220           if (const auto *CI = dyn_cast<ConstantInt>(AI->getArraySize())) {
221             if (CI->getLimitedValue(SSPBufferSize) >= SSPBufferSize) {
222               // A call to alloca with size >= SSPBufferSize requires
223               // stack protectors.
224               Layout.insert(std::make_pair(AI, SSPLK_LargeArray));
225               NeedsProtector = true;
226             } else if (Strong) {
227               // Require protectors for all alloca calls in strong mode.
228               Layout.insert(std::make_pair(AI, SSPLK_SmallArray));
229               NeedsProtector = true;
230             }
231           } else {
232             // A call to alloca with a variable size requires protectors.
233             Layout.insert(std::make_pair(AI, SSPLK_LargeArray));
234             NeedsProtector = true;
235           }
236           continue;
237         }
238 
239         bool IsLarge = false;
240         if (ContainsProtectableArray(AI->getAllocatedType(), IsLarge, Strong)) {
241           Layout.insert(std::make_pair(AI, IsLarge ? SSPLK_LargeArray
242                                                    : SSPLK_SmallArray));
243           NeedsProtector = true;
244           continue;
245         }
246 
247         if (Strong && HasAddressTaken(AI)) {
248           ++NumAddrTaken;
249           Layout.insert(std::make_pair(AI, SSPLK_AddrOf));
250           NeedsProtector = true;
251         }
252       }
253     }
254   }
255 
256   return NeedsProtector;
257 }
258 
InstructionWillNotHaveChain(const Instruction * I)259 static bool InstructionWillNotHaveChain(const Instruction *I) {
260   return !I->mayHaveSideEffects() && !I->mayReadFromMemory() &&
261          isSafeToSpeculativelyExecute(I);
262 }
263 
264 /// Identify if RI has a previous instruction in the "Tail Position" and return
265 /// it. Otherwise return 0.
266 ///
267 /// This is based off of the code in llvm::isInTailCallPosition. The difference
268 /// is that it inverts the first part of llvm::isInTailCallPosition since
269 /// isInTailCallPosition is checking if a call is in a tail call position, and
270 /// we are searching for an unknown tail call that might be in the tail call
271 /// position. Once we find the call though, the code uses the same refactored
272 /// code, returnTypeIsEligibleForTailCall.
FindPotentialTailCall(BasicBlock * BB,ReturnInst * RI,const TargetLoweringBase * TLI)273 static CallInst *FindPotentialTailCall(BasicBlock *BB, ReturnInst *RI,
274                                        const TargetLoweringBase *TLI) {
275   // Establish a reasonable upper bound on the maximum amount of instructions we
276   // will look through to find a tail call.
277   unsigned SearchCounter = 0;
278   const unsigned MaxSearch = 4;
279   bool NoInterposingChain = true;
280 
281   for (BasicBlock::reverse_iterator I = std::next(BB->rbegin()), E = BB->rend();
282        I != E && SearchCounter < MaxSearch; ++I) {
283     Instruction *Inst = &*I;
284 
285     // Skip over debug intrinsics and do not allow them to affect our MaxSearch
286     // counter.
287     if (isa<DbgInfoIntrinsic>(Inst))
288       continue;
289 
290     // If we find a call and the following conditions are satisifed, then we
291     // have found a tail call that satisfies at least the target independent
292     // requirements of a tail call:
293     //
294     // 1. The call site has the tail marker.
295     //
296     // 2. The call site either will not cause the creation of a chain or if a
297     // chain is necessary there are no instructions in between the callsite and
298     // the call which would create an interposing chain.
299     //
300     // 3. The return type of the function does not impede tail call
301     // optimization.
302     if (CallInst *CI = dyn_cast<CallInst>(Inst)) {
303       if (CI->isTailCall() &&
304           (InstructionWillNotHaveChain(CI) || NoInterposingChain) &&
305           returnTypeIsEligibleForTailCall(BB->getParent(), CI, RI, *TLI))
306         return CI;
307     }
308 
309     // If we did not find a call see if we have an instruction that may create
310     // an interposing chain.
311     NoInterposingChain =
312         NoInterposingChain && InstructionWillNotHaveChain(Inst);
313 
314     // Increment max search.
315     SearchCounter++;
316   }
317 
318   return nullptr;
319 }
320 
321 /// Insert code into the entry block that stores the __stack_chk_guard
322 /// variable onto the stack:
323 ///
324 ///   entry:
325 ///     StackGuardSlot = alloca i8*
326 ///     StackGuard = load __stack_chk_guard
327 ///     call void @llvm.stackprotect.create(StackGuard, StackGuardSlot)
328 ///
329 /// Returns true if the platform/triple supports the stackprotectorcreate pseudo
330 /// node.
CreatePrologue(Function * F,Module * M,ReturnInst * RI,const TargetLoweringBase * TLI,const Triple & TT,AllocaInst * & AI,Value * & StackGuardVar)331 static bool CreatePrologue(Function *F, Module *M, ReturnInst *RI,
332                            const TargetLoweringBase *TLI, const Triple &TT,
333                            AllocaInst *&AI, Value *&StackGuardVar) {
334   bool SupportsSelectionDAGSP = false;
335   PointerType *PtrTy = Type::getInt8PtrTy(RI->getContext());
336   unsigned AddressSpace, Offset;
337   if (TLI->getStackCookieLocation(AddressSpace, Offset)) {
338     Constant *OffsetVal =
339         ConstantInt::get(Type::getInt32Ty(RI->getContext()), Offset);
340 
341     StackGuardVar =
342         ConstantExpr::getIntToPtr(OffsetVal, PointerType::get(PtrTy,
343                                                               AddressSpace));
344   } else if (TT.isOSOpenBSD()) {
345     StackGuardVar = M->getOrInsertGlobal("__guard_local", PtrTy);
346     cast<GlobalValue>(StackGuardVar)
347         ->setVisibility(GlobalValue::HiddenVisibility);
348   } else {
349     SupportsSelectionDAGSP = true;
350     StackGuardVar = M->getOrInsertGlobal("__stack_chk_guard", PtrTy);
351   }
352 
353   IRBuilder<> B(&F->getEntryBlock().front());
354   AI = B.CreateAlloca(PtrTy, nullptr, "StackGuardSlot");
355   LoadInst *LI = B.CreateLoad(StackGuardVar, "StackGuard");
356   B.CreateCall2(Intrinsic::getDeclaration(M, Intrinsic::stackprotector), LI,
357                 AI);
358 
359   return SupportsSelectionDAGSP;
360 }
361 
362 /// InsertStackProtectors - Insert code into the prologue and epilogue of the
363 /// function.
364 ///
365 ///  - The prologue code loads and stores the stack guard onto the stack.
366 ///  - The epilogue checks the value stored in the prologue against the original
367 ///    value. It calls __stack_chk_fail if they differ.
InsertStackProtectors()368 bool StackProtector::InsertStackProtectors() {
369   bool HasPrologue = false;
370   bool SupportsSelectionDAGSP =
371       EnableSelectionDAGSP && !TM->Options.EnableFastISel;
372   AllocaInst *AI = nullptr;       // Place on stack that stores the stack guard.
373   Value *StackGuardVar = nullptr; // The stack guard variable.
374 
375   for (Function::iterator I = F->begin(), E = F->end(); I != E;) {
376     BasicBlock *BB = I++;
377     ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
378     if (!RI)
379       continue;
380 
381     if (!HasPrologue) {
382       HasPrologue = true;
383       SupportsSelectionDAGSP &=
384           CreatePrologue(F, M, RI, TLI, Trip, AI, StackGuardVar);
385     }
386 
387     if (SupportsSelectionDAGSP) {
388       // Since we have a potential tail call, insert the special stack check
389       // intrinsic.
390       Instruction *InsertionPt = nullptr;
391       if (CallInst *CI = FindPotentialTailCall(BB, RI, TLI)) {
392         InsertionPt = CI;
393       } else {
394         InsertionPt = RI;
395         // At this point we know that BB has a return statement so it *DOES*
396         // have a terminator.
397         assert(InsertionPt != nullptr &&
398                "BB must have a terminator instruction at this point.");
399       }
400 
401       Function *Intrinsic =
402           Intrinsic::getDeclaration(M, Intrinsic::stackprotectorcheck);
403       CallInst::Create(Intrinsic, StackGuardVar, "", InsertionPt);
404     } else {
405       // If we do not support SelectionDAG based tail calls, generate IR level
406       // tail calls.
407       //
408       // For each block with a return instruction, convert this:
409       //
410       //   return:
411       //     ...
412       //     ret ...
413       //
414       // into this:
415       //
416       //   return:
417       //     ...
418       //     %1 = load __stack_chk_guard
419       //     %2 = load StackGuardSlot
420       //     %3 = cmp i1 %1, %2
421       //     br i1 %3, label %SP_return, label %CallStackCheckFailBlk
422       //
423       //   SP_return:
424       //     ret ...
425       //
426       //   CallStackCheckFailBlk:
427       //     call void @__stack_chk_fail()
428       //     unreachable
429 
430       // Create the FailBB. We duplicate the BB every time since the MI tail
431       // merge pass will merge together all of the various BB into one including
432       // fail BB generated by the stack protector pseudo instruction.
433       BasicBlock *FailBB = CreateFailBB();
434 
435       // Split the basic block before the return instruction.
436       BasicBlock *NewBB = BB->splitBasicBlock(RI, "SP_return");
437 
438       // Update the dominator tree if we need to.
439       if (DT && DT->isReachableFromEntry(BB)) {
440         DT->addNewBlock(NewBB, BB);
441         DT->addNewBlock(FailBB, BB);
442       }
443 
444       // Remove default branch instruction to the new BB.
445       BB->getTerminator()->eraseFromParent();
446 
447       // Move the newly created basic block to the point right after the old
448       // basic block so that it's in the "fall through" position.
449       NewBB->moveAfter(BB);
450 
451       // Generate the stack protector instructions in the old basic block.
452       IRBuilder<> B(BB);
453       LoadInst *LI1 = B.CreateLoad(StackGuardVar);
454       LoadInst *LI2 = B.CreateLoad(AI);
455       Value *Cmp = B.CreateICmpEQ(LI1, LI2);
456       unsigned SuccessWeight =
457           BranchProbabilityInfo::getBranchWeightStackProtector(true);
458       unsigned FailureWeight =
459           BranchProbabilityInfo::getBranchWeightStackProtector(false);
460       MDNode *Weights = MDBuilder(F->getContext())
461                             .createBranchWeights(SuccessWeight, FailureWeight);
462       B.CreateCondBr(Cmp, NewBB, FailBB, Weights);
463     }
464   }
465 
466   // Return if we didn't modify any basic blocks. i.e., there are no return
467   // statements in the function.
468   if (!HasPrologue)
469     return false;
470 
471   return true;
472 }
473 
474 /// CreateFailBB - Create a basic block to jump to when the stack protector
475 /// check fails.
CreateFailBB()476 BasicBlock *StackProtector::CreateFailBB() {
477   LLVMContext &Context = F->getContext();
478   BasicBlock *FailBB = BasicBlock::Create(Context, "CallStackCheckFailBlk", F);
479   IRBuilder<> B(FailBB);
480   if (Trip.isOSOpenBSD()) {
481     Constant *StackChkFail =
482         M->getOrInsertFunction("__stack_smash_handler",
483                                Type::getVoidTy(Context),
484                                Type::getInt8PtrTy(Context), nullptr);
485 
486     B.CreateCall(StackChkFail, B.CreateGlobalStringPtr(F->getName(), "SSH"));
487   } else {
488     Constant *StackChkFail =
489         M->getOrInsertFunction("__stack_chk_fail", Type::getVoidTy(Context),
490                                nullptr);
491     B.CreateCall(StackChkFail);
492   }
493   B.CreateUnreachable();
494   return FailBB;
495 }
496