[dfsan] Propagate origin tracking at load

[stephan.yichao.zhao]

Summary

This is a part of https://reviews.llvm.org/D95835.

One issue is about origin load optimization: see the
comments of useCallbackLoadLabelAndOrigin

@gbalats This change may have some conflicts with your 8bit change. PTAL the change at visitLoad.

[gbalats]

llvm/lib/Transforms/Instrumentation/DataFlowSanitizer.cpp

580–582

I think it would be cleaner if we retained the older function (with its older return value) and introduced a new one:
loadShadowAndOrigin

1969–1973

Add assertion before returning to check that !ShouldTrackOrigins => Shadows.empty() && Origins.empty()

1970–1971

Shouldn't this be guarded by ShouldTrackOrigins flag? Even if it turns out to be a noop (when Shadows, Origins is empty), it's a little confusing for the reader.

2060–2062

Could instead use the flag, based on the assertion above.

2103

Is the initialization required here given the next line?

[morehouse]

llvm/lib/Transforms/Instrumentation/DataFlowSanitizer.cpp

580–582

+1

1915

2052

What does "non-scaped" mean?

llvm/test/Instrumentation/DataFlowSanitizer/origin_ldst.ll

244

Nit: s/LABLE/LABEL

[stephan.yichao.zhao]

update

llvm/lib/Transforms/Instrumentation/DataFlowSanitizer.cpp

580–582

That would be great. I had also considered this.
At some points, loading origins depends on loading shadows.
One is loadFast16ShadowFast, using both shadows and origins to create some select instructions.
The other is by DFSanLoadLabelAndOriginFn.
So it is not straightforward to call loadShadow from loadShadowAndOrigin.

1969–1973

combineOrigins has a similar assert. Added a "if (ShouldTrackOrigins) {" here to improve readability.

2052

replaced by non-escaped, for alloca pointers are not used outside the function.

2103

Thank you.

[morehouse]

LGTM

[gbalats]

llvm/lib/Transforms/Instrumentation/DataFlowSanitizer.cpp

2182–2183

Initialization seems redundant given the std::tie below.

llvm/test/Instrumentation/DataFlowSanitizer/origin_ldst.ll

22

Can these i16 expressions be generalized as in https://reviews.llvm.org/D97723 so that it requires fewer changes when we also have fast8 support?

[stephan.yichao.zhao]

updated test cases to use meta data

[stephan.yichao.zhao]

updated test cases

[gbalats]

LGTM

[stephan.yichao.zhao]

updated test cases

llvm/lib/Transforms/Instrumentation/DataFlowSanitizer.cpp

 #include "llvm/IR/MDBuilder.h"
 #include "llvm/IR/Module.h"
 #include "llvm/IR/PassManager.h"
 #include "llvm/IR/Type.h"
 #include "llvm/IR/User.h"
 #include "llvm/IR/Value.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/Pass.h"
+#include "llvm/Support/Alignment.h"
 #include "llvm/Support/Casting.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/SpecialCaseList.h"
 #include "llvm/Support/VirtualFileSystem.h"
 #include "llvm/Transforms/Instrumentation.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
 #include "llvm/Transforms/Utils/Local.h"
class DataFlowSanitizer {
   MDNode *ColdCallWeights;
   DFSanABIList ABIList;
   DenseMap<Value *, Function *> UnwrappedFnMap;
   AttrBuilder ReadOnlyNoneAttrs;
   bool DFSanRuntimeShadowMask = false;
 
   Value *getShadowOffset(Value *Addr, IRBuilder<> &IRB);
   Value *getShadowAddress(Value *Addr, Instruction *Pos);
-  // std::pair<Value *, Value *>
-  // getShadowOriginAddress(Value *Addr, Align InstAlignment, Instruction *Pos);
+  std::pair<Value *, Value *>
+  getShadowOriginAddress(Value *Addr, Align InstAlignment, Instruction *Pos);
   bool isInstrumented(const Function *F);
   bool isInstrumented(const GlobalAlias *GA);
   FunctionType *getArgsFunctionType(FunctionType *T);
   FunctionType *getTrampolineFunctionType(FunctionType *T);
   TransformedFunction getCustomFunctionType(FunctionType *T);
   InstrumentedABI getInstrumentedABI();
   WrapperKind getWrapperKind(Function *F);
   void addGlobalNamePrefix(GlobalValue *GV);
struct DFSanFunction {
   DominatorTree DT;
   DataFlowSanitizer::InstrumentedABI IA;
   bool IsNativeABI;
   AllocaInst *LabelReturnAlloca = nullptr;
   AllocaInst *OriginReturnAlloca = nullptr;
   DenseMap<Value *, Value *> ValShadowMap;
   DenseMap<Value *, Value *> ValOriginMap;
   DenseMap<AllocaInst *, AllocaInst *> AllocaShadowMap;
+  DenseMap<AllocaInst *, AllocaInst *> AllocaOriginMap;
+
   std::vector<std::pair<PHINode *, PHINode *>> PHIFixups;
   DenseSet<Instruction *> SkipInsts;
   std::vector<Value *> NonZeroChecks;
   bool AvoidNewBlocks;
 
   struct CachedShadow {
     BasicBlock *Block; // The block where Shadow is defined.
     Value *Shadow;
struct DFSanFunction {
   /// Generates IR to compute the union of the two given shadows, inserting it
   /// before Pos. The combined value is with primitive type.
   Value *combineShadows(Value *V1, Value *V2, Instruction *Pos);
   /// Combines the shadow values of V1 and V2, then converts the combined value
   /// with primitive type into a shadow value with the original type T.
   Value *combineShadowsThenConvert(Type *T, Value *V1, Value *V2,
                                    Instruction *Pos);
   Value *combineOperandShadows(Instruction *Inst);
-  Value *loadShadow(Value *ShadowAddr, uint64_t Size, uint64_t Align,
+  std::pair<Value *, Value *> loadShadowOrigin(Value *ShadowAddr, uint64_t Size,
+                                               Align InstAlignment,
                                                Instruction *Pos);
   void storePrimitiveShadow(Value *Addr, uint64_t Size, Align Alignment,
                             Value *PrimitiveShadow, Instruction *Pos);
   /// Applies PrimitiveShadow to all primitive subtypes of T, returning
   /// the expanded shadow value.
   ///
   /// EFP({T1,T2, ...}, PS) = {EFP(T1,PS),EFP(T2,PS),...}
   /// EFP([n x T], PS) = [n x EFP(T,PS)]
   /// EFP(other types, PS) = PS
private:
   /// Returns the shadow value of an argument A.
   Value *getShadowForTLSArgument(Argument *A);
 
   /// The fast path of loading shadow in legacy mode.
   Value *loadLegacyShadowFast(Value *ShadowAddr, uint64_t Size,
                               Align ShadowAlign, Instruction *Pos);
 
   /// The fast path of loading shadow in fast-16-label mode.
-  Value *loadFast16ShadowFast(Value *ShadowAddr, uint64_t Size,
-                              Align ShadowAlign, Instruction *Pos);
+  std::pair<Value *, Value *>
+  loadFast16ShadowFast(Value *ShadowAddr, Value *OriginAddr, uint64_t Size,
+                       Align ShadowAlign, Align OriginAlign, Value *FirstOrigin,
+                       Instruction *Pos);
+
+  Align getOriginAlign(Align InstAlignment);
+
+  /// Because 4 contiguous bytes share one 4-byte origin, the most accurate load
+  /// is __dfsan_load_label_and_origin. This function returns the union of all
+  /// labels and the origin of the first taint label. However this is an
+  /// additional call with many instructions. To ensure common cases are fast,
+  /// checks if it is possible to load labels and origins without using the
+  /// callback function.
+  bool useCallbackLoadLabelAndOrigin(uint64_t Size, Align InstAlignment);
 };
 
 class DFSanVisitor : public InstVisitor<DFSanVisitor> {
 public:
   DFSanFunction &DFSF;
 
   DFSanVisitor(DFSanFunction &DFSF) : DFSF(DFSF) {}
 
Value *DataFlowSanitizer::getShadowOffset(Value *Addr, IRBuilder<> &IRB) {
   Value *ShadowPtrMaskValue;
   if (DFSanRuntimeShadowMask)
     ShadowPtrMaskValue = IRB.CreateLoad(IntptrTy, ExternalShadowMask);
   else
     ShadowPtrMaskValue = ShadowPtrMask;
   return IRB.CreateAnd(IRB.CreatePtrToInt(Addr, IntptrTy),
                        IRB.CreatePtrToInt(ShadowPtrMaskValue, IntptrTy));
 }
-/*
+
 std::pair<Value *, Value *>
 DataFlowSanitizer::getShadowOriginAddress(Value *Addr, Align InstAlignment,
                                           Instruction *Pos) {
   // Returns ((Addr & shadow_mask) + origin_base) & ~4UL
   IRBuilder<> IRB(Pos);
   Value *ShadowOffset = getShadowOffset(Addr, IRB);
   Value *ShadowPtr = IRB.CreateIntToPtr(
       IRB.CreateMul(ShadowOffset, ShadowPtrMul), PrimitiveShadowPtrTy);
   Value *OriginPtr = nullptr;
   if (shouldTrackOrigins()) {
     Value *OriginLong = IRB.CreateAdd(ShadowOffset, OriginBase);
     const Align Alignment = llvm::assumeAligned(InstAlignment.value());
     // When alignment is >= 4, Addr must be aligned to 4, otherwise it is UB.
     // So Mask is unnecessary.
     if (Alignment < kMinOriginAlignment) {
       uint64_t Mask = kMinOriginAlignment.value() - 1;
       OriginLong = IRB.CreateAnd(OriginLong, ConstantInt::get(IntptrTy, ~Mask));
     }
     OriginPtr = IRB.CreateIntToPtr(OriginLong, OriginPtrTy);
   }
   return {ShadowPtr, OriginPtr};
 }
-*/
+
 Value *DataFlowSanitizer::getShadowAddress(Value *Addr, Instruction *Pos) {
   // Returns (Addr & shadow_mask) x 2
   IRBuilder<> IRB(Pos);
   Value *ShadowOffset = getShadowOffset(Addr, IRB);
   return IRB.CreateIntToPtr(IRB.CreateMul(ShadowOffset, ShadowPtrMul),
                             PrimitiveShadowPtrTy);
 }
 
void DFSanVisitor::visitInstOperandOrigins(Instruction &I) {
   DFSF.setOrigin(&I, CombinedOrigin);
 }
 
 Align DFSanFunction::getShadowAlign(Align InstAlignment) {
   const Align Alignment = ClPreserveAlignment ? InstAlignment : Align(1);
   return Align(Alignment.value() * DFS.ShadowWidthBytes);
 }
 
-Value *DFSanFunction::loadFast16ShadowFast(Value *ShadowAddr, uint64_t Size,
-                                           Align ShadowAlign,
-                                           Instruction *Pos) {
+Align DFSanFunction::getOriginAlign(Align InstAlignment) {
+  const Align Alignment = llvm::assumeAligned(InstAlignment.value());
+  return Align(std::max(kMinOriginAlignment, Alignment));
+}
+
+bool DFSanFunction::useCallbackLoadLabelAndOrigin(uint64_t Size,
+                                                  Align InstAlignment) {
+  assert(Size != 0);
+  // * if Size == 1, it is sufficient to load its origin aligned at 4.
+  // * if Size == 2, we assume most cases Addr % 2 == 0, so it is sufficient to
+  //   load its origin aligned at 4. If not, although origins may be lost, it
+  //   should not happen very often.
+  // * if align >= 4, Addr must be aligned to 4, otherwise it is UB. When
+  //   Size % 4 == 0, it is more efficient to load origins without callbacks.
+  // * Otherwise we use __dfsan_load_label_and_origin.
+  // This should ensure that common cases run efficiently.
+  if (Size <= 2)
+    return false;
+
+  const Align Alignment = llvm::assumeAligned(InstAlignment.value());
+  if (Alignment >= kMinOriginAlignment &&
+      Size % (64 / DFS.ShadowWidthBits) == 0)
+    return false;
+
+  return true;
+}
+
+std::pair<Value *, Value *> DFSanFunction::loadFast16ShadowFast(
+    Value *ShadowAddr, Value *OriginAddr, uint64_t Size, Align ShadowAlign,
+    Align OriginAlign, Value *FirstOrigin, Instruction *Pos) {
   // First OR all the WideShadows, then OR individual shadows within the
   // combined WideShadow. This is fewer instructions than ORing shadows
   // individually.
+  const bool ShouldTrackOrigins = DFS.shouldTrackOrigins();
+  std::vector<Value *> Shadows;
+  std::vector<Value *> Origins;
   IRBuilder<> IRB(Pos);
   Value *WideAddr =
       IRB.CreateBitCast(ShadowAddr, Type::getInt64PtrTy(*DFS.Ctx));
   Value *CombinedWideShadow =
       IRB.CreateAlignedLoad(IRB.getInt64Ty(), WideAddr, ShadowAlign);
+  if (ShouldTrackOrigins) {
+    Shadows.push_back(CombinedWideShadow);
+    Origins.push_back(FirstOrigin);
+  }
   for (uint64_t Ofs = 64 / DFS.ShadowWidthBits; Ofs != Size;
        Ofs += 64 / DFS.ShadowWidthBits) {
     WideAddr = IRB.CreateGEP(Type::getInt64Ty(*DFS.Ctx), WideAddr,
                              ConstantInt::get(DFS.IntptrTy, 1));
     Value *NextWideShadow =
         IRB.CreateAlignedLoad(IRB.getInt64Ty(), WideAddr, ShadowAlign);
     CombinedWideShadow = IRB.CreateOr(CombinedWideShadow, NextWideShadow);
+    if (ShouldTrackOrigins) {
+      Shadows.push_back(NextWideShadow);
+      OriginAddr = IRB.CreateGEP(DFS.OriginTy, OriginAddr,
+                                 ConstantInt::get(DFS.IntptrTy, 1));
+      Origins.push_back(
+          IRB.CreateAlignedLoad(DFS.OriginTy, OriginAddr, OriginAlign));
+    }
   }
   for (unsigned Width = 32; Width >= DFS.ShadowWidthBits; Width >>= 1) {
     Value *ShrShadow = IRB.CreateLShr(CombinedWideShadow, Width);
     CombinedWideShadow = IRB.CreateOr(CombinedWideShadow, ShrShadow);
   }
-  return IRB.CreateTrunc(CombinedWideShadow, DFS.PrimitiveShadowTy);
+  return {IRB.CreateTrunc(CombinedWideShadow, DFS.PrimitiveShadowTy),
+          ShouldTrackOrigins
+              ? combineOrigins(Shadows, Origins, Pos,
+                               ConstantInt::getSigned(IRB.getInt64Ty(), 0))
+              : DFS.ZeroOrigin};
 }
 
 Value *DFSanFunction::loadLegacyShadowFast(Value *ShadowAddr, uint64_t Size,
                                            Align ShadowAlign,
                                            Instruction *Pos) {
   // Fast path for the common case where each byte has identical shadow: load
   // shadow 64 bits at a time, fall out to a __dfsan_union_load call if any
   // shadow is non-equal.
Value *DFSanFunction::loadLegacyShadowFast(Value *ShadowAddr, uint64_t Size,
   Shadow->addIncoming(FallbackCall, FallbackBB);
   Shadow->addIncoming(TruncShadow, LastBr->getParent());
   return Shadow;
 }
 
 // Generates IR to load shadow corresponding to bytes [Addr, Addr+Size), where
 // Addr has alignment Align, and take the union of each of those shadows. The
 // returned shadow always has primitive type.
-Value *DFSanFunction::loadShadow(Value *Addr, uint64_t Size, uint64_t Align,
+std::pair<Value *, Value *> DFSanFunction::loadShadowOrigin(Value *Addr,
+                                                            uint64_t Size,
+                                                            Align InstAlignment,
                                                             Instruction *Pos) {
+  const bool ShouldTrackOrigins = DFS.shouldTrackOrigins();
+
+  // Non-escaped loads.
   if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) {
-    const auto i = AllocaShadowMap.find(AI);
-    if (i != AllocaShadowMap.end()) {
+    const auto SI = AllocaShadowMap.find(AI);
+    if (SI != AllocaShadowMap.end()) {
       IRBuilder<> IRB(Pos);
-      return IRB.CreateLoad(DFS.PrimitiveShadowTy, i->second);
+      Value *ShadowLI = IRB.CreateLoad(DFS.PrimitiveShadowTy, SI->second);
+      const auto OI = AllocaOriginMap.find(AI);
+      assert(!ShouldTrackOrigins || OI != AllocaOriginMap.end());
+      return {ShadowLI, ShouldTrackOrigins
+                            ? IRB.CreateLoad(DFS.OriginTy, OI->second)
+                            : nullptr};
     }
   }
 
-  const llvm::Align ShadowAlign(Align * DFS.ShadowWidthBytes);
+  // Load from constant addresses.
   SmallVector<const Value *, 2> Objs;
   getUnderlyingObjects(Addr, Objs);
   bool AllConstants = true;
   for (const Value *Obj : Objs) {
     if (isa<Function>(Obj) || isa<BlockAddress>(Obj))
       continue;
     if (isa<GlobalVariable>(Obj) && cast<GlobalVariable>(Obj)->isConstant())
       continue;
 
     AllConstants = false;
     break;
   }
   if (AllConstants)
-    return DFS.ZeroPrimitiveShadow;
+    return {DFS.ZeroPrimitiveShadow,
+            ShouldTrackOrigins ? DFS.ZeroOrigin : nullptr};
+
+  if (Size == 0)
+    return {DFS.ZeroPrimitiveShadow,
+            ShouldTrackOrigins ? DFS.ZeroOrigin : nullptr};
+
+  // Use callback to load if this is not an optimizable case for origin
+  // tracking.
+  if (ShouldTrackOrigins &&
+      useCallbackLoadLabelAndOrigin(Size, InstAlignment)) {
+    IRBuilder<> IRB(Pos);
+    CallInst *Call =
+        IRB.CreateCall(DFS.DFSanLoadLabelAndOriginFn,
+                       {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
+                        ConstantInt::get(DFS.IntptrTy, Size)});
+    Call->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
+    return {IRB.CreateTrunc(IRB.CreateLShr(Call, DFS.OriginWidthBits),
+                            DFS.PrimitiveShadowTy),
+            IRB.CreateTrunc(Call, DFS.OriginTy)};
+  }
+
+  // Other cases that support loading shadows or origins in a fast way.
+  Value *ShadowAddr, *OriginAddr;
+  std::tie(ShadowAddr, OriginAddr) =
+      DFS.getShadowOriginAddress(Addr, InstAlignment, Pos);
+
+  const Align ShadowAlign = getShadowAlign(InstAlignment);
+  const Align OriginAlign = getOriginAlign(InstAlignment);
+  Value *Origin = nullptr;
+  if (ShouldTrackOrigins) {
+    IRBuilder<> IRB(Pos);
+    Origin = IRB.CreateAlignedLoad(DFS.OriginTy, OriginAddr, OriginAlign);
+  }
 
-  Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos);
   switch (Size) {
-  case 0:
-    return DFS.ZeroPrimitiveShadow;
   case 1: {
     LoadInst *LI = new LoadInst(DFS.PrimitiveShadowTy, ShadowAddr, "", Pos);
     LI->setAlignment(ShadowAlign);
-    return LI;
+    return {LI, Origin};
   }
   case 2: {
     IRBuilder<> IRB(Pos);
     Value *ShadowAddr1 = IRB.CreateGEP(DFS.PrimitiveShadowTy, ShadowAddr,
                                        ConstantInt::get(DFS.IntptrTy, 1));
-    return combineShadows(
-        IRB.CreateAlignedLoad(DFS.PrimitiveShadowTy, ShadowAddr, ShadowAlign),
-        IRB.CreateAlignedLoad(DFS.PrimitiveShadowTy, ShadowAddr1, ShadowAlign),
-        Pos);
+    Value *Load =
+        IRB.CreateAlignedLoad(DFS.PrimitiveShadowTy, ShadowAddr, ShadowAlign);
+    Value *Load1 =
+        IRB.CreateAlignedLoad(DFS.PrimitiveShadowTy, ShadowAddr1, ShadowAlign);
+    return {combineShadows(Load, Load1, Pos), Origin};
   }
   }
 
   if (ClFast16Labels && Size % (64 / DFS.ShadowWidthBits) == 0)
-    return loadFast16ShadowFast(ShadowAddr, Size, ShadowAlign, Pos);
+    return loadFast16ShadowFast(ShadowAddr, OriginAddr, Size, ShadowAlign,
+                                OriginAlign, Origin, Pos);
 
   if (!AvoidNewBlocks && Size % (64 / DFS.ShadowWidthBits) == 0)
-    return loadLegacyShadowFast(ShadowAddr, Size, ShadowAlign, Pos);
+    return {loadLegacyShadowFast(ShadowAddr, Size, ShadowAlign, Pos), Origin};
 
   IRBuilder<> IRB(Pos);
   FunctionCallee &UnionLoadFn =
       ClFast16Labels ? DFS.DFSanUnionLoadFast16LabelsFn : DFS.DFSanUnionLoadFn;
   CallInst *FallbackCall = IRB.CreateCall(
       UnionLoadFn, {ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size)});
   FallbackCall->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
-  return FallbackCall;
+  return {FallbackCall, Origin};
 }
 
 static AtomicOrdering addAcquireOrdering(AtomicOrdering AO) {
   switch (AO) {
   case AtomicOrdering::NotAtomic:
     return AtomicOrdering::NotAtomic;
   case AtomicOrdering::Unordered:
   case AtomicOrdering::Monotonic:
   case AtomicOrdering::Acquire:
     return AtomicOrdering::Acquire;
   case AtomicOrdering::Release:
   case AtomicOrdering::AcquireRelease:
     return AtomicOrdering::AcquireRelease;
   case AtomicOrdering::SequentiallyConsistent:
     return AtomicOrdering::SequentiallyConsistent;
   }
   llvm_unreachable("Unknown ordering");
 }
 
 void DFSanVisitor::visitLoadInst(LoadInst &LI) {
   auto &DL = LI.getModule()->getDataLayout();
   uint64_t Size = DL.getTypeStoreSize(LI.getType());
   if (Size == 0) {
     DFSF.setShadow(&LI, DFSF.DFS.getZeroShadow(&LI));
+    DFSF.setOrigin(&LI, DFSF.DFS.ZeroOrigin);
     return;
   }
 
   // When an application load is atomic, increase atomic ordering between
   // atomic application loads and stores to ensure happen-before order; load
   // shadow data after application data; store zero shadow data before
   // application data. This ensure shadow loads return either labels of the
   // initial application data or zeros.
   if (LI.isAtomic())
     LI.setOrdering(addAcquireOrdering(LI.getOrdering()));
 
-  Align Alignment = ClPreserveAlignment ? LI.getAlign() : Align(1);
   Instruction *Pos = LI.isAtomic() ? LI.getNextNode() : &LI;
-  Value *PrimitiveShadow =
-      DFSF.loadShadow(LI.getPointerOperand(), Size, Alignment.value(), Pos);
+  std::vector<Value *> Shadows;
+  std::vector<Value *> Origins;
+  Value *PrimitiveShadow, *Origin;
+  std::tie(PrimitiveShadow, Origin) =
+      DFSF.loadShadowOrigin(LI.getPointerOperand(), Size, LI.getAlign(), Pos);
+  const bool ShouldTrackOrigins = DFSF.DFS.shouldTrackOrigins();
+  if (ShouldTrackOrigins) {
+    Shadows.push_back(PrimitiveShadow);
+    Origins.push_back(Origin);
+  }
   if (ClCombinePointerLabelsOnLoad) {
     Value *PtrShadow = DFSF.getShadow(LI.getPointerOperand());
     PrimitiveShadow = DFSF.combineShadows(PrimitiveShadow, PtrShadow, Pos);
+    if (ShouldTrackOrigins) {
+      Shadows.push_back(PtrShadow);
+      Origins.push_back(DFSF.getOrigin(LI.getPointerOperand()));
+    }
   }
   if (!DFSF.DFS.isZeroShadow(PrimitiveShadow))
     DFSF.NonZeroChecks.push_back(PrimitiveShadow);
 
   Value *Shadow =
       DFSF.expandFromPrimitiveShadow(LI.getType(), PrimitiveShadow, Pos);
   DFSF.setShadow(&LI, Shadow);
+
+  if (ShouldTrackOrigins) {
+    DFSF.setOrigin(&LI, DFSF.combineOrigins(Shadows, Origins, Pos));
+  }
+
   if (ClEventCallbacks) {
     IRBuilder<> IRB(Pos);
     Value *Addr8 = IRB.CreateBitCast(LI.getPointerOperand(), DFSF.DFS.Int8Ptr);
     IRB.CreateCall(DFSF.DFS.DFSanLoadCallbackFn, {PrimitiveShadow, Addr8});
   }
 }
 
 void DFSanFunction::storeZeroPrimitiveShadow(Value *Addr, uint64_t Size,
for (User *U : I.users()) {
     }
 
     AllLoadsStores = false;
     break;
   }
   if (AllLoadsStores) {
     IRBuilder<> IRB(&I);
     DFSF.AllocaShadowMap[&I] = IRB.CreateAlloca(DFSF.DFS.PrimitiveShadowTy);
+    if (DFSF.DFS.shouldTrackOrigins()) {
+      DFSF.AllocaOriginMap[&I] =
+          IRB.CreateAlloca(DFSF.DFS.OriginTy, nullptr, "_dfsa");
+    }
   }
   DFSF.setShadow(&I, DFSF.DFS.ZeroPrimitiveShadow);
+  DFSF.setOrigin(&I, DFSF.DFS.ZeroOrigin);
 }
 
 void DFSanVisitor::visitSelectInst(SelectInst &I) {
   Value *CondShadow = DFSF.getShadow(I.getCondition());
   Value *TrueShadow = DFSF.getShadow(I.getTrueValue());
   Value *FalseShadow = DFSF.getShadow(I.getFalseValue());
   Value *ShadowSel = nullptr;
   const bool ShouldTrackOrigins = DFSF.DFS.shouldTrackOrigins();

llvm/test/Instrumentation/DataFlowSanitizer/origin_ldst.ll

+; RUN: opt < %s -dfsan -dfsan-track-origins=1 -dfsan-fast-16-labels=true -S | FileCheck %s --check-prefixes=CHECK_META,CHECK
+; RUN: opt < %s -dfsan -dfsan-track-origins=1 -dfsan-fast-16-labels=true -dfsan-combine-pointer-labels-on-load=false -S | FileCheck %s --check-prefixes=CHECK_META,NO_COMBINE_LOAD_PTR
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; CHECK_META: @__dfsan_shadow_width_bits = weak_odr constant i32 [[#SBITS:]]
+; CHECK_META: @__dfsan_shadow_width_bytes = weak_odr constant i32 [[#SBYTES:]]
+
+define {} @load0({}* %p) {
+  ; CHECK: @"dfs$load0"
+  ; CHECK-NEXT: %a = load {}, {}* %p, align 1
+  ; CHECK-NEXT: store {} zeroinitializer, {}* bitcast ([100 x i64]* @__dfsan_retval_tls to {}*), align [[#SBYTES]]
+  ; CHECK-NEXT: store i32 0, i32* @__dfsan_retval_origin_tls, align 4
+  ; CHECK-NEXT: ret {} %a
+
+  %a = load {}, {}* %p
+  ret {} %a
+}
+
+define i16 @load_non_escaped_alloca() {
+  ; CHECK: @"dfs$load_non_escaped_alloca"
+  ; CHECK: [[S_ALLOCA:%.*]] = alloca i[[#SBITS]], align [[#SBYTES]]
+  ; CHECK: [[O_ALLOCA:%.*]] = alloca i32, align 4
+  ; CHECK: [[SHADOW:%.*]] = load i[[#SBITS]], i[[#SBITS]]* [[S_ALLOCA]], align [[#SBYTES]]
+  ; CHECK: [[ORIGIN:%.*]] = load i32, i32* [[O_ALLOCA]], align 4
+  ; CHECK: %a = load i16, i16* %p, align 2
+  ; CHECK: store i[[#SBITS]] [[SHADOW]], i[[#SBITS]]* bitcast ([100 x i64]* @__dfsan_retval_tls to i[[#SBITS]]*), align [[#SBYTES]]
+  ; CHECK: store i32 [[ORIGIN]], i32* @__dfsan_retval_origin_tls, align 4
+  
+  %p = alloca i16
+  %a = load i16, i16* %p
+  ret i16 %a
+}
+
+define i16* @load_escaped_alloca() {
+  ; CHECK: @"dfs$load_escaped_alloca"
+  ; CHECK: [[INTP:%.*]] = ptrtoint i[[#SBITS]]* %p to i64
+  ; CHECK: [[OFFSET:%.*]] = and i64 [[INTP]], -123145302310913
+  ; CHECK: [[SHADOW_ADDR:%.*]] = mul i64 [[OFFSET]], 2
+  ; CHECK: [[SHADOW_PTR0:%.*]] = inttoptr i64 [[SHADOW_ADDR]] to i[[#SBITS]]*
+  ; CHECK: [[ORIGIN_OFFSET:%.*]] = add i64 [[OFFSET]], 35184372088832
+  ; CHECK: [[ORIGIN_ADDR:%.*]] = and i64 [[ORIGIN_OFFSET]], -4
+  ; CHECK: [[ORIGIN_PTR:%.*]] = inttoptr i64 [[ORIGIN_ADDR]] to i32*
+  ; CHECK: {{%.*}} = load i32, i32* [[ORIGIN_PTR]], align 4
+  ; CHECK: [[SHADOW_PTR1:%.*]] = getelementptr i[[#SBITS]], i[[#SBITS]]* [[SHADOW_PTR0]], i64 1
+  ; CHECK: [[SHADOW0:%.*]] = load i[[#SBITS]], i[[#SBITS]]* [[SHADOW_PTR0]], align [[#SBYTES]]
+  ; CHECK: [[SHADOW1:%.*]] = load i[[#SBITS]], i[[#SBITS]]* [[SHADOW_PTR1]], align [[#SBYTES]]
+  ; CHECK: {{%.*}} = or i[[#SBITS]] [[SHADOW0]], [[SHADOW1]]
+  ; CHECK: %a = load i16, i16* %p, align 2
+  ; CHECK: store i[[#SBITS]] 0, i[[#SBITS]]* bitcast ([100 x i64]* @__dfsan_retval_tls to i[[#SBITS]]*), align [[#SBYTES]]
+  ; CHECK: store i32 0, i32* @__dfsan_retval_origin_tls, align 4
+  
+  %p = alloca i16
+  %a = load i16, i16* %p
+  ret i16* %p
+}
+
+@X = constant i1 1
+define i1 @load_global() {
+  ; CHECK: @"dfs$load_global"
+  ; CHECK: %a = load i1, i1* @X, align 1
+  ; CHECK: store i[[#SBITS]] 0, i[[#SBITS]]* bitcast ([100 x i64]* @__dfsan_retval_tls to i[[#SBITS]]*), align [[#SBYTES]]
+  ; CHECK: store i32 0, i32* @__dfsan_retval_origin_tls, align 4
+
+  %a = load i1, i1* @X
+  ret i1 %a
+}
+
+define i1 @load1(i1* %p) {
+  ; CHECK: @"dfs$load1"
+  ; CHECK: [[PO:%.*]] = load i32, i32* getelementptr inbounds ([200 x i32], [200 x i32]* @__dfsan_arg_origin_tls, i64 0, i64 0), align 4
+  ; CHECK: [[PS:%.*]] = load i[[#SBITS]], i[[#SBITS]]* bitcast ([100 x i64]* @__dfsan_arg_tls to i[[#SBITS]]*), align [[#SBYTES]]
+  ; CHECK: [[INTP:%.*]] = ptrtoint {{.*}} %p to i64
+  ; CHECK: [[OFFSET:%.*]] = and i64 [[INTP]], -123145302310913
+  ; CHECK: [[SHADOW_ADDR:%.*]] = mul i64 [[OFFSET]], 2
+  ; CHECK: [[SHADOW_PTR:%.*]] = inttoptr i64 [[SHADOW_ADDR]] to i[[#SBITS]]*
+  ; CHECK: [[ORIGIN_OFFSET:%.*]] = add i64 [[OFFSET]], 35184372088832
+  ; CHECK: [[ORIGIN_ADDR:%.*]] = and i64 [[ORIGIN_OFFSET]], -4
+  ; CHECK: [[ORIGIN_PTR:%.*]] = inttoptr i64 [[ORIGIN_ADDR]] to i32*
+  ; CHECK: [[AO:%.*]] = load i32, i32* [[ORIGIN_PTR]], align 4
+  ; CHECK: [[AS:%.*]] = load i[[#SBITS]], i[[#SBITS]]* [[SHADOW_PTR]], align [[#SBYTES]]
+  ; CHECK: [[RS:%.*]] = or i[[#SBITS]] [[AS]], [[PS]]
+  ; CHECK: [[PS_NZ:%.*]] = icmp ne i[[#SBITS]] [[PS]], 0
+  ; CHECK: [[RO:%.*]] = select i1 [[PS_NZ]], i32 [[PO]], i32 [[AO]]
+  ; CHECK: %a = load i1, i1* %p, align 1
+  ; CHECK: store i[[#SBITS]] [[RS]], i[[#SBITS]]* bitcast ([100 x i64]* @__dfsan_retval_tls to i[[#SBITS]]*), align [[#SBYTES]]
+  ; CHECK: store i32 [[RO]], i32* @__dfsan_retval_origin_tls, align 4
+
+  %a = load i1, i1* %p
+  ret i1 %a
+}
+
+define i16 @load16(i1 %i, i16* %p) {
+  ; CHECK: @"dfs$load16"
+  ; CHECK: [[PO:%.*]] = load i32, i32* getelementptr inbounds ([200 x i32], [200 x i32]* @__dfsan_arg_origin_tls, i64 0, i64 1), align 4
+  ; CHECK: [[PS:%.*]] = load i[[#SBITS]], i[[#SBITS]]* inttoptr (i64 add (i64 ptrtoint ([100 x i64]* @__dfsan_arg_tls to i64), i64 2) to i[[#SBITS]]*), align [[#SBYTES]]
+  ; CHECK: [[INTP:%.*]] = ptrtoint {{.*}} %p to i64
+  ; CHECK: [[OFFSET:%.*]] = and i64 [[INTP]], -123145302310913
+  ; CHECK: [[SHADOW_ADDR:%.*]] = mul i64 [[OFFSET]], 2
+  ; CHECK: [[SHADOW_PTR0:%.*]] = inttoptr i64 [[SHADOW_ADDR]] to i[[#SBITS]]*
+  ; CHECK: [[ORIGIN_OFFSET:%.*]] = add i64 [[OFFSET]], 35184372088832
+  ; CHECK: [[ORIGIN_ADDR:%.*]] = and i64 [[ORIGIN_OFFSET]], -4
+  ; CHECK: [[ORIGIN_PTR:%.*]] = inttoptr i64 [[ORIGIN_ADDR]] to i32*
+  ; CHECK: [[AO:%.*]] = load i32, i32* [[ORIGIN_PTR]], align 4
+  ; CHECK: [[SHADOW_PTR1:%.*]] = getelementptr i[[#SBITS]], i[[#SBITS]]* [[SHADOW_PTR0]], i64 1
+  ; CHECK: [[SHADOW0:%.*]] = load i[[#SBITS]], i[[#SBITS]]* [[SHADOW_PTR0]], align [[#SBYTES]]
+  ; CHECK: [[SHADOW1:%.*]] = load i[[#SBITS]], i[[#SBITS]]* [[SHADOW_PTR1]], align [[#SBYTES]]
+  ; CHECK: [[AS:%.*]] = or i[[#SBITS]] [[SHADOW0]], [[SHADOW1]]
+  ; CHECK: [[RS:%.*]] = or i[[#SBITS]] [[AS]], [[PS]]
+  ; CHECK: [[PS_NZ:%.*]] = icmp ne i[[#SBITS]] [[PS]], 0
+  ; CHECK: [[RO:%.*]] = select i1 [[PS_NZ]], i32 [[PO]], i32 [[AO]]
+  ; CHECK: %a = load i16, i16* %p, align 2
+  ; CHECK: store i[[#SBITS]] [[RS]], i[[#SBITS]]* bitcast ([100 x i64]* @__dfsan_retval_tls to i[[#SBITS]]*), align [[#SBYTES]]
+  ; CHECK: store i32 [[RO]], i32* @__dfsan_retval_origin_tls, align 4
+
+  %a = load i16, i16* %p
+  ret i16 %a
+}
+
+define i32 @load32(i32* %p) {
+  ; CHECK: @"dfs$load32"
+
+  ; NO_COMBINE_LOAD_PTR: @"dfs$load32"
+  ; NO_COMBINE_LOAD_PTR: [[INTP:%.*]] = ptrtoint i32* %p to i64
+  ; NO_COMBINE_LOAD_PTR: [[OFFSET:%.*]] = and i64 [[INTP]], -123145302310913
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_ADDR:%.*]] = mul i64 [[OFFSET]], 2
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_PTR:%.*]] = inttoptr i64 [[SHADOW_ADDR]] to i[[#SBITS]]*
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_ADDR:%.*]] = add i64 [[OFFSET]], 35184372088832
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_PTR:%.*]] = inttoptr i64 [[ORIGIN_ADDR]] to i32*
+  ; NO_COMBINE_LOAD_PTR: [[AO:%.*]] = load i32, i32* [[ORIGIN_PTR]], align 4
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_PTR64:%.*]] = bitcast i[[#SBITS]]* [[SHADOW_PTR]] to i64*
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64:%.*]] = load i64, i64* [[SHADOW_PTR64]], align [[#SBYTES]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64_H32:%.*]] = lshr i64 [[SHADOW64]], 32
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64_HL32:%.*]] = or i64 [[SHADOW64]], [[SHADOW64_H32]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64_HL32_H16:%.*]] = lshr i64 [[SHADOW64_HL32]], 16
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64_HL32_HL16:%.*]] = or i64 [[SHADOW64_HL32]], [[SHADOW64_HL32_H16]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW:%.*]] = trunc i64 [[SHADOW64_HL32_HL16]] to i[[#SBITS]]
+  ; NO_COMBINE_LOAD_PTR: %a = load i32, i32* %p, align 4
+  ; NO_COMBINE_LOAD_PTR: store i[[#SBITS]] [[SHADOW]], i[[#SBITS]]* bitcast ([100 x i64]* @__dfsan_retval_tls to i[[#SBITS]]*), align [[#SBYTES]]
+  ; NO_COMBINE_LOAD_PTR: store i32 [[AO]], i32* @__dfsan_retval_origin_tls, align 4
+
+  %a = load i32, i32* %p
+  ret i32 %a
+}
+
+define i64 @load64(i64* %p) {
+  ; CHECK: @"dfs$load64"
+  
+  ; NO_COMBINE_LOAD_PTR: @"dfs$load64"
+  ; NO_COMBINE_LOAD_PTR: [[INTP:%.*]] = ptrtoint i64* %p to i64
+  ; NO_COMBINE_LOAD_PTR: [[OFFSET:%.*]] = and i64 [[INTP]], -123145302310913
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_ADDR:%.*]] = mul i64 [[OFFSET]], 2
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_PTR:%.*]] = inttoptr i64 [[SHADOW_ADDR]] to i[[#SBITS]]*
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_ADDR:%.*]] = add i64 [[OFFSET]], 35184372088832
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_PTR_0:%.*]] = inttoptr i64 [[ORIGIN_ADDR]] to i32*
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_0:%.*]] = load i32, i32* [[ORIGIN_PTR_0]], align 8
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_PTR_0:%.*]] = bitcast i[[#SBITS]]* [[SHADOW_PTR]] to i64*
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_0:%.*]] = load i64, i64* [[SHADOW_PTR_0]], align [[#SBYTES]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_PTR_1:%.*]] = getelementptr i64, i64* [[SHADOW_PTR_0]], i64 1
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_1:%.*]] = load i64, i64* [[SHADOW_PTR_1]], align [[#SBYTES]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64:%.*]] = or i64 [[SHADOW_0]], [[SHADOW_1]]
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_PTR_1:%.*]] = getelementptr i32, i32* [[ORIGIN_PTR_0]], i64 1
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_1:%.*]] = load i32, i32* [[ORIGIN_PTR_1]], align 8
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64_H32:%.*]] = lshr i64 [[SHADOW64]], 32
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64_HL32:%.*]] = or i64 [[SHADOW64]], [[SHADOW64_H32]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64_HL32_H16:%.*]] = lshr i64 [[SHADOW64_HL32]], 16
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64_HL32_HL16:%.*]] = or i64 [[SHADOW64_HL32]], [[SHADOW64_HL32_H16]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW:%.*]] = trunc i64 [[SHADOW64_HL32_HL16]] to i[[#SBITS]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_1_NZ:%.*]] = icmp ne i64 [[SHADOW_1]], 0
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN:%.*]] = select i1 [[SHADOW_1_NZ]], i32 [[ORIGIN_1]], i32 [[ORIGIN_0]]
+  ; NO_COMBINE_LOAD_PTR: %a = load i64, i64* %p, align 8
+  ; NO_COMBINE_LOAD_PTR: store i[[#SBITS]] [[SHADOW]], i[[#SBITS]]* bitcast ([100 x i64]* @__dfsan_retval_tls to i[[#SBITS]]*), align [[#SBYTES]]
+  ; NO_COMBINE_LOAD_PTR: store i32 [[ORIGIN]], i32* @__dfsan_retval_origin_tls, align 4
+
+  %a = load i64, i64* %p
+  ret i64 %a
+}
+
+define i64 @load64_align2(i64* %p) {
+  ; CHECK: @"dfs$load64_align2"
+
+  ; NO_COMBINE_LOAD_PTR: @"dfs$load64_align2"
+  ; NO_COMBINE_LOAD_PTR-NEXT: [[INTP:%.*]] = bitcast i64* %p to i8*
+  ; NO_COMBINE_LOAD_PTR-NEXT: [[LABEL_ORIGIN:%.*]] = call zeroext i64 @__dfsan_load_label_and_origin(i8* [[INTP]], i64 8)
+  ; NO_COMBINE_LOAD_PTR-NEXT: [[LABEL_ORIGIN_H32:%.*]] = lshr i64 [[LABEL_ORIGIN]], 32
+  ; NO_COMBINE_LOAD_PTR-NEXT: [[LABEL:%.*]] = trunc i64 [[LABEL_ORIGIN_H32]] to i[[#SBITS]]
+  ; NO_COMBINE_LOAD_PTR-NEXT: [[ORIGIN:%.*]] = trunc i64 [[LABEL_ORIGIN]] to i32
+  ; NO_COMBINE_LOAD_PTR-NEXT: %a = load i64, i64* %p, align [[#SBYTES]]
+  ; NO_COMBINE_LOAD_PTR-NEXT: store i[[#SBITS]] [[LABEL]], i[[#SBITS]]* bitcast ([100 x i64]* @__dfsan_retval_tls to i[[#SBITS]]*), align [[#SBYTES]]
+  ; NO_COMBINE_LOAD_PTR-NEXT: store i32 [[ORIGIN]], i32* @__dfsan_retval_origin_tls, align 4
+  
+  %a = load i64, i64* %p, align 2
+  ret i64 %a
+}
+
+define i92 @load92(i92* %p) {
+  ; CHECK: @"dfs$load92"
+
+  ; NO_COMBINE_LOAD_PTR: @"dfs$load92"
+  ; NO_COMBINE_LOAD_PTR: [[INTP:%.*]] = ptrtoint i92* %p to i64
+  ; NO_COMBINE_LOAD_PTR: [[OFFSET:%.*]] = and i64 [[INTP]], -123145302310913
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_ADDR:%.*]] = mul i64 [[OFFSET]], 2
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_PTR:%.*]] = inttoptr i64 [[SHADOW_ADDR]] to i[[#SBITS]]*
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_ADDR:%.*]] = add i64 [[OFFSET]], 35184372088832
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_PTR_0:%.*]] = inttoptr i64 [[ORIGIN_ADDR]] to i32*
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_0:%.*]] = load i32, i32* [[ORIGIN_PTR_0]], align 8
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_PTR_0:%.*]] = bitcast i[[#SBITS]]* [[SHADOW_PTR]] to i64*
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_0:%.*]] = load i64, i64* [[SHADOW_PTR_0]], align [[#SBYTES]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_PTR_1:%.*]] = getelementptr i64, i64* [[SHADOW_PTR_0]], i64 1
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_1:%.*]] = load i64, i64* [[SHADOW_PTR_1]], align [[#SBYTES]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_01:%.*]] = or i64 [[SHADOW_0]], [[SHADOW_1]]
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_PTR_1:%.*]] = getelementptr i32, i32* [[ORIGIN_PTR_0]], i64 1
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_1:%.*]] = load i32, i32* [[ORIGIN_PTR_1]], align 8
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_PTR_2:%.*]] = getelementptr i64, i64* [[SHADOW_PTR_1]], i64 1
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_2:%.*]] = load i64, i64* [[SHADOW_PTR_2]], align [[#SBYTES]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64:%.*]] = or i64 [[SHADOW_01]], [[SHADOW_2]]
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_PTR_2:%.*]] = getelementptr i32, i32* [[ORIGIN_PTR_1]], i64 1
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_2:%.*]] = load i32, i32* [[ORIGIN_PTR_2]], align 8
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64_H32:%.*]] = lshr i64 [[SHADOW64]], 32
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64_HL32:%.*]] = or i64 [[SHADOW64]], [[SHADOW64_H32]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64_HL32_H16:%.*]] = lshr i64 [[SHADOW64_HL32]], 16
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW64_HL32_HL16:%.*]] = or i64 [[SHADOW64_HL32]], [[SHADOW64_HL32_H16]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW:%.*]] = trunc i64 [[SHADOW64_HL32_HL16]] to i[[#SBITS]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_1_NZ:%.*]] = icmp ne i64 [[SHADOW_1]], 0
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN_10:%.*]] = select i1 [[SHADOW_1_NZ]], i32 [[ORIGIN_1]], i32 [[ORIGIN_0]]
+  ; NO_COMBINE_LOAD_PTR: [[SHADOW_2_NZ:%.*]] = icmp ne i64 [[SHADOW_2]], 0
+  ; NO_COMBINE_LOAD_PTR: [[ORIGIN:%.*]] = select i1 [[SHADOW_2_NZ]], i32 [[ORIGIN_2]], i32 [[ORIGIN_10]]
+  ; NO_COMBINE_LOAD_PTR: %a = load i92, i92* %p, align 8
+  ; NO_COMBINE_LOAD_PTR: store i[[#SBITS]] [[SHADOW]], i[[#SBITS]]* bitcast ([100 x i64]* @__dfsan_retval_tls to i[[#SBITS]]*), align [[#SBYTES]]
+  ; NO_COMBINE_LOAD_PTR: store i32 [[ORIGIN]], i32* @__dfsan_retval_origin_tls, align 4
+  
+  %a = load i92, i92* %p
+  ret i92 %a
+}
+
+define i17 @load17(i17* %p) {
+  ; CHECK: @"dfs$load17"
+
+  ; NO_COMBINE_LOAD_PTR: @"dfs$load17"
+  ; NO_COMBINE_LOAD_PTR-NEXT: [[INTP:%.*]] = bitcast i17* %p to i8*
+  ; NO_COMBINE_LOAD_PTR-NEXT: [[LABEL_ORIGIN:%.*]] = call zeroext i64 @__dfsan_load_label_and_origin(i8* [[INTP]], i64 3)
+  ; NO_COMBINE_LOAD_PTR-NEXT: [[LABEL_ORIGIN_H32:%.*]] = lshr i64 [[LABEL_ORIGIN]], 32
+  ; NO_COMBINE_LOAD_PTR-NEXT: [[LABEL:%.*]] = trunc i64 [[LABEL_ORIGIN_H32]] to i[[#SBITS]]
+  ; NO_COMBINE_LOAD_PTR-NEXT: [[ORIGIN:%.*]] = trunc i64 [[LABEL_ORIGIN]] to i32
+  ; NO_COMBINE_LOAD_PTR-NEXT: %a = load i17, i17* %p, align 4
+  ; NO_COMBINE_LOAD_PTR-NEXT: store i[[#SBITS]] [[LABEL]], i[[#SBITS]]* bitcast ([100 x i64]* @__dfsan_retval_tls to i[[#SBITS]]*), align [[#SBYTES]]
+  ; NO_COMBINE_LOAD_PTR-NEXT: store i32 [[ORIGIN]], i32* @__dfsan_retval_origin_tls, align 4
+  
+  %a = load i17, i17* %p, align 4
+  ret i17 %a
+}