[llvm] [LLVM][Instrumentation] Add numerical sanitizer (PR #85916)

Fangrui Song via llvm-commits llvm-commits at lists.llvm.org
Thu Jun 20 13:43:39 PDT 2024


================
@@ -0,0 +1,2191 @@
+//===-- NumericalStabilitySanitizer.cpp -----------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of NumericalStabilitySanitizer.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Transforms/Instrumentation/NumericalStabilitySanitizer.h"
+
+#include <cstdint>
+
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallString.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/StringExtras.h"
+#include "llvm/Analysis/CaptureTracking.h"
+#include "llvm/Analysis/TargetLibraryInfo.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/Intrinsics.h"
+#include "llvm/IR/LLVMContext.h"
+#include "llvm/IR/MDBuilder.h"
+#include "llvm/IR/Metadata.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/Type.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/ProfileData/InstrProf.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/MathExtras.h"
+#include "llvm/Support/Regex.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Transforms/Instrumentation.h"
+#include "llvm/Transforms/Utils/BasicBlockUtils.h"
+#include "llvm/Transforms/Utils/EscapeEnumerator.h"
+#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/Transforms/Utils/ModuleUtils.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "nsan"
+
+STATISTIC(NumInstrumentedFTLoads,
+          "Number of instrumented floating-point loads");
+
+STATISTIC(NumInstrumentedFTCalls,
+          "Number of instrumented floating-point calls");
+STATISTIC(NumInstrumentedFTRets,
+          "Number of instrumented floating-point returns");
+STATISTIC(NumInstrumentedFTStores,
+          "Number of instrumented floating-point stores");
+STATISTIC(NumInstrumentedNonFTStores,
+          "Number of instrumented non floating-point stores");
+STATISTIC(
+    NumInstrumentedNonFTMemcpyStores,
+    "Number of instrumented non floating-point stores with memcpy semantics");
+STATISTIC(NumInstrumentedFCmp, "Number of instrumented fcmps");
+
+// Using smaller shadow types types can help improve speed. For example, `dlq`
+// is 3x slower to 5x faster in opt mode and 2-6x faster in dbg mode compared to
+// `dqq`.
+static cl::opt<std::string> ClShadowMapping(
+    "nsan-shadow-type-mapping", cl::init("dqq"),
+    cl::desc("One shadow type id for each of `float`, `double`, `long double`. "
+             "`d`,`l`,`q`,`e` mean double, x86_fp80, fp128 (quad) and "
+             "ppc_fp128 (extended double) respectively. The default is to "
+             "shadow `float` as `double`, and `double` and `x86_fp80` as "
+             "`fp128`"),
+    cl::Hidden);
+
+static cl::opt<bool>
+    ClInstrumentFCmp("nsan-instrument-fcmp", cl::init(true),
+                     cl::desc("Instrument floating-point comparisons"),
+                     cl::Hidden);
+
+static cl::opt<std::string> ClCheckFunctionsFilter(
+    "check-functions-filter",
+    cl::desc("Only emit checks for arguments of functions "
+             "whose names match the given regular expression"),
+    cl::value_desc("regex"));
+
+static cl::opt<bool> ClTruncateFCmpEq(
+    "nsan-truncate-fcmp-eq", cl::init(true),
+    cl::desc(
+        "This flag controls the behaviour of fcmp equality comparisons."
+        "For equality comparisons such as `x == 0.0f`, we can perform the "
+        "shadow check in the shadow (`x_shadow == 0.0) == (x == 0.0f)`) or app "
+        " domain (`(trunc(x_shadow) == 0.0f) == (x == 0.0f)`). This helps "
+        "catch the case when `x_shadow` is accurate enough (and therefore "
+        "close enough to zero) so that `trunc(x_shadow)` is zero even though "
+        "both `x` and `x_shadow` are not."),
+    cl::Hidden);
+
+// When there is external, uninstrumented code writing to memory, the shadow
+// memory can get out of sync with the application memory. Enabling this flag
+// emits consistency checks for loads to catch this situation.
+// When everything is instrumented, this is not strictly necessary because any
+// load should have a corresponding store, but can help debug cases when the
+// framework did a bad job at tracking shadow memory modifications by failing on
+// load rather than store.
+// FIXME: provide a way to resume computations from the FT value when the load
+// is inconsistent. This ensures that further computations are not polluted.
+static cl::opt<bool> ClCheckLoads("nsan-check-loads", cl::init(false),
+                                  cl::desc("Check floating-point load"),
+                                  cl::Hidden);
+
+static cl::opt<bool> ClCheckStores("nsan-check-stores", cl::init(true),
+                                   cl::desc("Check floating-point stores"),
+                                   cl::Hidden);
+
+static cl::opt<bool> ClCheckRet("nsan-check-ret", cl::init(true),
+                                cl::desc("Check floating-point return values"),
+                                cl::Hidden);
+
+// LLVM may store constant floats as bitcasted ints.
+// It's not really necessary to shadow such stores,
+// if the shadow value is unknown the framework will re-extend it on load
+// anyway. Moreover, because of size collisions (e.g. bf16 vs f16) it is
+// impossible to determine the floating-point type based on the size.
+// However, for debugging purposes it can be useful to model such stores.
+static cl::opt<bool> ClPropagateNonFTConstStoresAsFT(
+    "nsan-propagate-non-ft-const-stores-as-ft", cl::init(false),
+    cl::desc(
+        "Propagate non floating-point const stores as floating point values."
+        "For debugging purposes only"),
+    cl::Hidden);
+
+static constexpr StringLiteral kNsanModuleCtorName("nsan.module_ctor");
+static constexpr StringLiteral kNsanInitName("__nsan_init");
+
+// The following values must be kept in sync with the runtime.
+static constexpr int kShadowScale = 2;
+static constexpr int kMaxVectorWidth = 8;
+static constexpr int kMaxNumArgs = 128;
+static constexpr int kMaxShadowTypeSizeBytes = 16; // fp128
+
+namespace {
+
+// Defines the characteristics (type id, type, and floating-point semantics)
+// attached for all possible shadow types.
+class ShadowTypeConfig {
+public:
+  static std::unique_ptr<ShadowTypeConfig> fromNsanTypeId(char TypeId);
+
+  // The LLVM Type corresponding to the shadow type.
+  virtual Type *getType(LLVMContext &Context) const = 0;
+
+  // The nsan type id of the shadow type (`d`, `l`, `q`, ...).
+  virtual char getNsanTypeId() const = 0;
+
+  virtual ~ShadowTypeConfig() = default;
+};
+
+template <char NsanTypeId>
+class ShadowTypeConfigImpl : public ShadowTypeConfig {
+public:
+  char getNsanTypeId() const override { return NsanTypeId; }
+  static constexpr const char kNsanTypeId = NsanTypeId;
+};
+
+// `double` (`d`) shadow type.
+class F64ShadowConfig : public ShadowTypeConfigImpl<'d'> {
+  Type *getType(LLVMContext &Context) const override {
+    return Type::getDoubleTy(Context);
+  }
+};
+
+// `x86_fp80` (`l`) shadow type: X86 long double.
+class F80ShadowConfig : public ShadowTypeConfigImpl<'l'> {
+  Type *getType(LLVMContext &Context) const override {
+    return Type::getX86_FP80Ty(Context);
+  }
+};
+
+// `fp128` (`q`) shadow type.
+class F128ShadowConfig : public ShadowTypeConfigImpl<'q'> {
+  Type *getType(LLVMContext &Context) const override {
+    return Type::getFP128Ty(Context);
+  }
+};
+
+// `ppc_fp128` (`e`) shadow type: IBM extended double with 106 bits of mantissa.
+class PPC128ShadowConfig : public ShadowTypeConfigImpl<'e'> {
+  Type *getType(LLVMContext &Context) const override {
+    return Type::getPPC_FP128Ty(Context);
+  }
+};
+
+// Creates a ShadowTypeConfig given its type id.
+std::unique_ptr<ShadowTypeConfig>
+ShadowTypeConfig::fromNsanTypeId(const char TypeId) {
+  switch (TypeId) {
+  case F64ShadowConfig::kNsanTypeId:
+    return std::make_unique<F64ShadowConfig>();
+  case F80ShadowConfig::kNsanTypeId:
+    return std::make_unique<F80ShadowConfig>();
+  case F128ShadowConfig::kNsanTypeId:
+    return std::make_unique<F128ShadowConfig>();
+  case PPC128ShadowConfig::kNsanTypeId:
+    return std::make_unique<PPC128ShadowConfig>();
+  }
+  report_fatal_error("nsan: invalid shadow type id '" + Twine(TypeId) + "'");
+}
+
+// An enum corresponding to shadow value types. Used as indices in arrays, so
+// not an `enum class`.
+enum FTValueType { kFloat, kDouble, kLongDouble, kNumValueTypes };
+
+// If `FT` corresponds to a primitive FTValueType, return it.
+static std::optional<FTValueType> ftValueTypeFromType(Type *FT) {
+  if (FT->isFloatTy())
+    return kFloat;
+  if (FT->isDoubleTy())
+    return kDouble;
+  if (FT->isX86_FP80Ty())
+    return kLongDouble;
+  return {};
+}
+
+// Returns the LLVM type for an FTValueType.
+static Type *typeFromFTValueType(FTValueType VT, LLVMContext &Context) {
+  switch (VT) {
+  case kFloat:
+    return Type::getFloatTy(Context);
+  case kDouble:
+    return Type::getDoubleTy(Context);
+  case kLongDouble:
+    return Type::getX86_FP80Ty(Context);
+  case kNumValueTypes:
+    return nullptr;
+  }
+}
+
+// Returns the type name for an FTValueType.
+static const char *typeNameFromFTValueType(FTValueType VT) {
+  switch (VT) {
+  case kFloat:
+    return "float";
+  case kDouble:
+    return "double";
+  case kLongDouble:
+    return "longdouble";
+  case kNumValueTypes:
+    return nullptr;
+  }
+}
+
+// A specific mapping configuration of application type to shadow type for nsan
+// (see -nsan-shadow-mapping flag).
+class MappingConfig {
+public:
+  explicit MappingConfig(LLVMContext &C) : Context(C) {
+    if (ClShadowMapping.size() != 3)
+      report_fatal_error("Invalid nsan mapping: " + Twine(ClShadowMapping));
+    unsigned ShadowTypeSizeBits[kNumValueTypes];
+    for (int VT = 0; VT < kNumValueTypes; ++VT) {
+      auto Config = ShadowTypeConfig::fromNsanTypeId(ClShadowMapping[VT]);
+      if (!Config)
+        report_fatal_error("Failed to get ShadowTypeConfig for " +
+                           Twine(ClShadowMapping[VT]));
+      const unsigned AppTypeSize =
+          typeFromFTValueType(static_cast<FTValueType>(VT), Context)
+              ->getScalarSizeInBits();
+      const unsigned ShadowTypeSize =
+          Config->getType(Context)->getScalarSizeInBits();
+      // Check that the shadow type size is at most kShadowScale times the
+      // application type size, so that shadow memory compoutations are valid.
+      if (ShadowTypeSize > kShadowScale * AppTypeSize)
+        report_fatal_error("Invalid nsan mapping f" + Twine(AppTypeSize) +
+                           "->f" + Twine(ShadowTypeSize) +
+                           ": The shadow type size should be at most " +
+                           Twine(kShadowScale) +
+                           " times the application type size");
+      ShadowTypeSizeBits[VT] = ShadowTypeSize;
+      Configs[VT] = std::move(Config);
+    }
+
+    // Check that the mapping is monotonous. This is required because if one
+    // does an fpextend of `float->long double` in application code, nsan is
+    // going to do an fpextend of `shadow(float) -> shadow(long double)` in
+    // shadow code. This will fail in `qql` mode, since nsan would be
+    // fpextending `f128->long`, which is invalid.
+    // FIXME: Relax this.
+    if (ShadowTypeSizeBits[kFloat] > ShadowTypeSizeBits[kDouble] ||
+        ShadowTypeSizeBits[kDouble] > ShadowTypeSizeBits[kLongDouble])
+      report_fatal_error("Invalid nsan mapping: { float->f" +
+                         Twine(ShadowTypeSizeBits[kFloat]) + "; double->f" +
+                         Twine(ShadowTypeSizeBits[kDouble]) +
+                         "; long double->f" +
+                         Twine(ShadowTypeSizeBits[kLongDouble]) + " }");
+  }
+
+  const ShadowTypeConfig &byValueType(FTValueType VT) const {
+    assert(VT < FTValueType::kNumValueTypes && "invalid value type");
+    return *Configs[VT];
+  }
+
+  // Returns the extended shadow type for a given application type.
+  Type *getExtendedFPType(Type *FT) const {
+    if (const auto VT = ftValueTypeFromType(FT))
+      return Configs[*VT]->getType(Context);
+    if (FT->isVectorTy()) {
+      auto *VecTy = cast<VectorType>(FT);
+      // FIXME: add support for scalable vector types.
+      if (VecTy->isScalableTy())
+        return nullptr;
+      Type *ExtendedScalar = getExtendedFPType(VecTy->getElementType());
+      return ExtendedScalar
+                 ? VectorType::get(ExtendedScalar, VecTy->getElementCount())
+                 : nullptr;
+    }
+    return nullptr;
+  }
+
+private:
+  LLVMContext &Context;
+  std::unique_ptr<ShadowTypeConfig> Configs[FTValueType::kNumValueTypes];
+};
+
+// The memory extents of a type specifies how many elements of a given
+// FTValueType needs to be stored when storing this type.
+struct MemoryExtents {
+  FTValueType ValueType;
+  uint64_t NumElts;
+};
+static MemoryExtents getMemoryExtentsOrDie(Type *FT) {
+  if (const auto VT = ftValueTypeFromType(FT))
+    return {*VT, 1};
+  if (auto *VecTy = dyn_cast<VectorType>(FT)) {
+    const auto ScalarExtents = getMemoryExtentsOrDie(VecTy->getElementType());
+    return {ScalarExtents.ValueType,
+            ScalarExtents.NumElts * VecTy->getElementCount().getFixedValue()};
+  }
+  llvm_unreachable("invalid value type");
+}
+
+// The location of a check. Passed as parameters to runtime checking functions.
+class CheckLoc {
+public:
+  // Creates a location that references an application memory location.
+  static CheckLoc makeStore(Value *Address) {
+    CheckLoc Result(kStore);
+    Result.Address = Address;
+    return Result;
+  }
+  static CheckLoc makeLoad(Value *Address) {
+    CheckLoc Result(kLoad);
+    Result.Address = Address;
+    return Result;
+  }
+
+  // Creates a location that references an argument, given by id.
+  static CheckLoc makeArg(int ArgId) {
+    CheckLoc Result(kArg);
+    Result.ArgId = ArgId;
+    return Result;
+  }
+
+  // Creates a location that references the return value of a function.
+  static CheckLoc makeRet() { return CheckLoc(kRet); }
+
+  // Creates a location that references a vector insert.
+  static CheckLoc makeInsert() { return CheckLoc(kInsert); }
+
+  // Returns the CheckType of location this refers to, as an integer-typed LLVM
+  // IR value.
+  Value *getType(LLVMContext &C) const {
+    return ConstantInt::get(Type::getInt32Ty(C), static_cast<int>(CheckTy));
+  }
+
+  // Returns a CheckType-specific value representing details of the location
+  // (e.g. application address for loads or stores), as an `IntptrTy`-typed LLVM
+  // IR value.
+  Value *getValue(Type *IntptrTy, IRBuilder<> &Builder) const {
+    switch (CheckTy) {
+    case kUnknown:
+      llvm_unreachable("unknown type");
+    case kRet:
+    case kInsert:
+      return ConstantInt::get(IntptrTy, 0);
+    case kArg:
+      return ConstantInt::get(IntptrTy, ArgId);
+    case kLoad:
+    case kStore:
+      return Builder.CreatePtrToInt(Address, IntptrTy);
+    }
+  }
+
+private:
+  // Must be kept in sync with the runtime.
+  enum CheckType {
+    kUnknown = 0,
+    kRet,
+    kArg,
+    kLoad,
+    kStore,
+    kInsert,
+  };
+  explicit CheckLoc(CheckType CheckTy) : CheckTy(CheckTy) {}
+
+  const CheckType CheckTy;
+  Value *Address = nullptr;
+  int ArgId = -1;
+};
+
+// A map of LLVM IR values to shadow LLVM IR values.
+class ValueToShadowMap {
+public:
+  explicit ValueToShadowMap(const MappingConfig &Config) : Config(Config) {}
+
+  ValueToShadowMap(const ValueToShadowMap &) = delete;
+  ValueToShadowMap &operator=(const ValueToShadowMap &) = delete;
+
+  // Sets the shadow value for a value. Asserts that the value does not already
+  // have a value.
+  void setShadow(Value *V, Value *Shadow) {
+    assert(V);
+    assert(Shadow);
+    [[maybe_unused]] const bool Inserted = Map.try_emplace(V, Shadow).second;
+    LLVM_DEBUG({
+      if (!Inserted) {
+        if (auto *I = dyn_cast<Instruction>(V))
+          errs() << I->getFunction()->getName() << ": ";
+        errs() << "duplicate shadow (" << V << "): ";
+        V->dump();
+      }
+    });
+    assert(Inserted && "duplicate shadow");
+  }
+
+  // Returns true if the value already has a shadow (including if the value is a
+  // constant). If true, calling getShadow() is valid.
+  bool hasShadow(Value *V) const {
+    return isa<Constant>(V) || (Map.find(V) != Map.end());
+  }
+
+  // Returns the shadow value for a given value. Asserts that the value has
+  // a shadow value. Lazily creates shadows for constant values.
+  Value *getShadow(Value *V) const {
+    if (Constant *C = dyn_cast<Constant>(V))
+      return getShadowConstant(C);
+    const auto ShadowValIt = Map.find(V);
+    assert(ShadowValIt != Map.end() && "shadow val does not exist");
+    assert(ShadowValIt->second && "shadow val is null");
+    return ShadowValIt->second;
+  }
+
+  bool empty() const { return Map.empty(); }
+
+private:
+  // Extends a constant application value to its shadow counterpart.
+  APFloat extendConstantFP(APFloat CV, const fltSemantics &To) const {
+    bool LosesInfo = false;
+    CV.convert(To, APFloatBase::rmTowardZero, &LosesInfo);
+    return CV;
+  }
+
+  // Returns the shadow constant for the given application constant.
+  Constant *getShadowConstant(Constant *C) const {
+    if (UndefValue *U = dyn_cast<UndefValue>(C)) {
+      return UndefValue::get(Config.getExtendedFPType(U->getType()));
+    }
+    if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
+      // Floating-point constants.
+      Type *Ty = Config.getExtendedFPType(CFP->getType());
+      return ConstantFP::get(
+          Ty, extendConstantFP(CFP->getValueAPF(), Ty->getFltSemantics()));
+    }
+    // Vector, array, or aggregate constants.
+    if (C->getType()->isVectorTy()) {
+      SmallVector<Constant *, 8> Elements;
+      for (int I = 0, E = cast<VectorType>(C->getType())
+                              ->getElementCount()
+                              .getFixedValue();
+           I < E; ++I)
+        Elements.push_back(getShadowConstant(C->getAggregateElement(I)));
+      return ConstantVector::get(Elements);
+    }
+    llvm_unreachable("unimplemented");
+  }
+
+  const MappingConfig &Config;
+  DenseMap<Value *, Value *> Map;
+};
+
+/// Instantiating NumericalStabilitySanitizer inserts the nsan runtime library
+/// API function declarations into the module if they don't exist already.
+/// Instantiating ensures the __nsan_init function is in the list of global
+/// constructors for the module.
+class NumericalStabilitySanitizer {
+public:
+  NumericalStabilitySanitizer(Module &M);
+  bool sanitizeFunction(Function &F, const TargetLibraryInfo &TLI);
+
+private:
+  bool instrumentMemIntrinsic(MemIntrinsic *MI);
+  void maybeAddSuffixForNsanInterface(CallBase *CI);
+  bool addrPointsToConstantData(Value *Addr);
+  void maybeCreateShadowValue(Instruction &Root, const TargetLibraryInfo &TLI,
+                              ValueToShadowMap &Map);
+  Value *createShadowValueWithOperandsAvailable(Instruction &Inst,
+                                                const TargetLibraryInfo &TLI,
+                                                const ValueToShadowMap &Map);
+  PHINode *maybeCreateShadowPhi(PHINode &Phi, const TargetLibraryInfo &TLI);
+  void createShadowArguments(Function &F, const TargetLibraryInfo &TLI,
+                             ValueToShadowMap &Map);
+
+  void populateShadowStack(CallBase &CI, const TargetLibraryInfo &TLI,
+                           const ValueToShadowMap &Map);
+
+  void propagateShadowValues(Instruction &Inst, const TargetLibraryInfo &TLI,
+                             const ValueToShadowMap &Map);
+  Value *emitCheck(Value *V, Value *ShadowV, IRBuilder<> &Builder,
+                   CheckLoc Loc);
+  Value *emitCheckInternal(Value *V, Value *ShadowV, IRBuilder<> &Builder,
+                           CheckLoc Loc);
+  void emitFCmpCheck(FCmpInst &FCmp, const ValueToShadowMap &Map);
+
+  // Value creation handlers.
+  Value *handleLoad(LoadInst &Load, Type *VT, Type *ExtendedVT);
+  Value *handleCallBase(CallBase &Call, Type *VT, Type *ExtendedVT,
+                        const TargetLibraryInfo &TLI,
+                        const ValueToShadowMap &Map, IRBuilder<> &Builder);
+  Value *maybeHandleKnownCallBase(CallBase &Call, Type *VT, Type *ExtendedVT,
+                                  const TargetLibraryInfo &TLI,
+                                  const ValueToShadowMap &Map,
+                                  IRBuilder<> &Builder);
+  Value *handleTrunc(const FPTruncInst &Trunc, Type *VT, Type *ExtendedVT,
+                     const ValueToShadowMap &Map, IRBuilder<> &Builder);
+  Value *handleExt(const FPExtInst &Ext, Type *VT, Type *ExtendedVT,
+                   const ValueToShadowMap &Map, IRBuilder<> &Builder);
+
+  // Value propagation handlers.
+  void propagateFTStore(StoreInst &Store, Type *VT, Type *ExtendedVT,
+                        const ValueToShadowMap &Map);
+  void propagateNonFTStore(StoreInst &Store, Type *VT,
+                           const ValueToShadowMap &Map);
+
+  const DataLayout &DL;
+  LLVMContext &Context;
+  MappingConfig Config;
+  IntegerType *IntptrTy = nullptr;
+  FunctionCallee NsanGetShadowPtrForStore[FTValueType::kNumValueTypes] = {};
+  FunctionCallee NsanGetShadowPtrForLoad[FTValueType::kNumValueTypes] = {};
+  FunctionCallee NsanCheckValue[FTValueType::kNumValueTypes] = {};
+  FunctionCallee NsanFCmpFail[FTValueType::kNumValueTypes] = {};
+  FunctionCallee NsanCopyValues;
+  FunctionCallee NsanSetValueUnknown;
+  FunctionCallee NsanGetRawShadowTypePtr;
+  FunctionCallee NsanGetRawShadowPtr;
+  GlobalValue *NsanShadowRetTag = nullptr;
+
+  Type *NsanShadowRetType = nullptr;
+  GlobalValue *NsanShadowRetPtr = nullptr;
+
+  GlobalValue *NsanShadowArgsTag = nullptr;
+
+  Type *NsanShadowArgsType = nullptr;
+  GlobalValue *NsanShadowArgsPtr = nullptr;
+
+  std::optional<Regex> CheckFunctionsFilter;
+};
+
+void insertModuleCtor(Module &M) {
+  getOrCreateSanitizerCtorAndInitFunctions(
+      M, kNsanModuleCtorName, kNsanInitName, /*InitArgTypes=*/{},
+      /*InitArgs=*/{},
+      // This callback is invoked when the functions are created the first
+      // time. Hook them into the global ctors list in that case:
+      [&](Function *Ctor, FunctionCallee) { appendToGlobalCtors(M, Ctor, 0); });
+}
+
+} // end anonymous namespace
+
+PreservedAnalyses
+NumericalStabilitySanitizerPass::run(Module &M, ModuleAnalysisManager &MAM) {
+  insertModuleCtor(M);
+
+  NumericalStabilitySanitizer Nsan(M);
+  auto &FAM = MAM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
+  for (Function &F : M)
+    Nsan.sanitizeFunction(F, FAM.getResult<TargetLibraryAnalysis>(F));
+
+  return PreservedAnalyses::none();
+}
+
+static GlobalValue *createThreadLocalGV(const char *Name, Module &M, Type *Ty) {
+  return dyn_cast<GlobalValue>(M.getOrInsertGlobal(Name, Ty, [&M, Ty, Name] {
+    return new GlobalVariable(M, Ty, false, GlobalVariable::ExternalLinkage,
+                              nullptr, Name, nullptr,
+                              GlobalVariable::InitialExecTLSModel);
+  }));
+}
+
+NumericalStabilitySanitizer::NumericalStabilitySanitizer(Module &M)
+    : DL(M.getDataLayout()), Context(M.getContext()), Config(Context) {
+  IntptrTy = DL.getIntPtrType(Context);
+  Type *PtrTy = PointerType::getUnqual(Context);
+  Type *Int32Ty = Type::getInt32Ty(Context);
+  Type *Int1Ty = Type::getInt1Ty(Context);
+  Type *VoidTy = Type::getVoidTy(Context);
+
+  AttributeList Attr;
+  Attr = Attr.addFnAttribute(Context, Attribute::NoUnwind);
+  // Initialize the runtime values (functions and global variables).
+  for (int I = 0; I < kNumValueTypes; ++I) {
+    const FTValueType VT = static_cast<FTValueType>(I);
+    const char *VTName = typeNameFromFTValueType(VT);
+    Type *VTTy = typeFromFTValueType(VT, Context);
+
+    // Load/store.
+    const std::string GetterPrefix =
+        std::string("__nsan_get_shadow_ptr_for_") + VTName;
+    NsanGetShadowPtrForStore[VT] = M.getOrInsertFunction(
+        GetterPrefix + "_store", Attr, PtrTy, PtrTy, IntptrTy);
+    NsanGetShadowPtrForLoad[VT] = M.getOrInsertFunction(
+        GetterPrefix + "_load", Attr, PtrTy, PtrTy, IntptrTy);
+
+    // Check.
+    const auto &ShadowConfig = Config.byValueType(VT);
+    Type *ShadowTy = ShadowConfig.getType(Context);
+    NsanCheckValue[VT] =
+        M.getOrInsertFunction(std::string("__nsan_internal_check_") + VTName +
+                                  "_" + ShadowConfig.getNsanTypeId(),
+                              Attr, Int32Ty, VTTy, ShadowTy, Int32Ty, IntptrTy);
+    NsanFCmpFail[VT] = M.getOrInsertFunction(
+        std::string("__nsan_fcmp_fail_") + VTName + "_" +
+            ShadowConfig.getNsanTypeId(),
+        Attr, VoidTy, VTTy, VTTy, ShadowTy, ShadowTy, Int32Ty, Int1Ty, Int1Ty);
+  }
+
+  NsanCopyValues = M.getOrInsertFunction("__nsan_copy_values", Attr, VoidTy,
+                                         PtrTy, PtrTy, IntptrTy);
+  NsanSetValueUnknown = M.getOrInsertFunction("__nsan_set_value_unknown", Attr,
+                                              VoidTy, PtrTy, IntptrTy);
+
+  // FIXME: Add attributes nofree, nosync, readnone, readonly,
+  NsanGetRawShadowTypePtr = M.getOrInsertFunction(
+      "__nsan_internal_get_raw_shadow_type_ptr", Attr, PtrTy, PtrTy);
+  NsanGetRawShadowPtr = M.getOrInsertFunction(
+      "__nsan_internal_get_raw_shadow_ptr", Attr, PtrTy, PtrTy);
+
+  NsanShadowRetTag = createThreadLocalGV("__nsan_shadow_ret_tag", M, IntptrTy);
+
+  NsanShadowRetType = ArrayType::get(Type::getInt8Ty(Context),
+                                     kMaxVectorWidth * kMaxShadowTypeSizeBytes);
+  NsanShadowRetPtr =
+      createThreadLocalGV("__nsan_shadow_ret_ptr", M, NsanShadowRetType);
+
+  NsanShadowArgsTag =
+      createThreadLocalGV("__nsan_shadow_args_tag", M, IntptrTy);
+
+  NsanShadowArgsType =
+      ArrayType::get(Type::getInt8Ty(Context),
+                     kMaxVectorWidth * kMaxNumArgs * kMaxShadowTypeSizeBytes);
+
+  NsanShadowArgsPtr =
+      createThreadLocalGV("__nsan_shadow_args_ptr", M, NsanShadowArgsType);
+
+  if (!ClCheckFunctionsFilter.empty()) {
+    Regex R = Regex(ClCheckFunctionsFilter);
+    std::string RegexError;
+    assert(R.isValid(RegexError));
+    CheckFunctionsFilter = std::move(R);
+  }
+}
+
+// Returns true if the given LLVM Value points to constant data (typically, a
+// global variable reference).
+bool NumericalStabilitySanitizer::addrPointsToConstantData(Value *Addr) {
+  // If this is a GEP, just analyze its pointer operand.
+  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
+    Addr = GEP->getPointerOperand();
+
+  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr))
+    return GV->isConstant();
+  return false;
+}
+
+// This instruments the function entry to create shadow arguments.
+// Pseudocode:
+//   if (this_fn_ptr == __nsan_shadow_args_tag) {
+//     s(arg0) = LOAD<sizeof(arg0)>(__nsan_shadow_args);
+//     s(arg1) = LOAD<sizeof(arg1)>(__nsan_shadow_args + sizeof(arg0));
+//     ...
+//     __nsan_shadow_args_tag = 0;
+//   } else {
+//     s(arg0) = fext(arg0);
+//     s(arg1) = fext(arg1);
+//     ...
+//   }
+void NumericalStabilitySanitizer::createShadowArguments(
+    Function &F, const TargetLibraryInfo &TLI, ValueToShadowMap &Map) {
+  assert(!F.getIntrinsicID() && "found a definition of an intrinsic");
+
+  // Do not bother if there are no FP args.
+  if (all_of(F.args(), [this](const Argument &Arg) {
+        return Config.getExtendedFPType(Arg.getType()) == nullptr;
+      }))
+    return;
+
+  IRBuilder<> Builder(F.getEntryBlock().getFirstNonPHI());
+  // The function has shadow args if the shadow args tag matches the function
+  // address.
+  Value *HasShadowArgs = Builder.CreateICmpEQ(
+      Builder.CreateLoad(IntptrTy, NsanShadowArgsTag, /*isVolatile=*/false),
+      Builder.CreatePtrToInt(&F, IntptrTy));
+
+  unsigned ShadowArgsOffsetBytes = 0;
+  for (Argument &Arg : F.args()) {
+    Type *VT = Arg.getType();
+    Type *ExtendedVT = Config.getExtendedFPType(VT);
+    if (ExtendedVT == nullptr)
+      continue; // Not an FT value.
+    Value *L = Builder.CreateAlignedLoad(
+        ExtendedVT,
+        Builder.CreateConstGEP2_64(NsanShadowArgsType, NsanShadowArgsPtr, 0,
+                                   ShadowArgsOffsetBytes),
+        Align(1), /*isVolatile=*/false);
+    Value *Shadow = Builder.CreateSelect(HasShadowArgs, L,
+                                         Builder.CreateFPExt(&Arg, ExtendedVT));
+    Map.setShadow(&Arg, Shadow);
+    TypeSize SlotSize = DL.getTypeStoreSize(ExtendedVT);
+    assert(!SlotSize.isScalable() && "unsupported");
+    ShadowArgsOffsetBytes += SlotSize;
+  }
+  Builder.CreateStore(ConstantInt::get(IntptrTy, 0), NsanShadowArgsTag);
+}
+
+// Returns true if the instrumentation should emit code to check arguments
+// before a function call.
+static bool shouldCheckArgs(CallBase &CI, const TargetLibraryInfo &TLI,
+                            const std::optional<Regex> &CheckFunctionsFilter) {
+
+  Function *Fn = CI.getCalledFunction();
+
+  if (CheckFunctionsFilter) {
+    // Skip checking args of indirect calls.
+    if (Fn == nullptr)
+      return false;
+    if (CheckFunctionsFilter->match(Fn->getName()))
+      return true;
+    return false;
+  }
+
+  if (Fn == nullptr)
+    return true; // Always check args of indirect calls.
+
+  // Never check nsan functions, the user called them for a reason.
+  if (Fn->getName().starts_with("__nsan_"))
+    return false;
+
+  const auto ID = Fn->getIntrinsicID();
+  LibFunc LFunc = LibFunc::NumLibFuncs;
+  // Always check args of unknown functions.
+  if (ID == Intrinsic::ID() && !TLI.getLibFunc(*Fn, LFunc))
+    return true;
+
+  // Do not check args of an `fabs` call that is used for a comparison.
+  // This is typically used for `fabs(a-b) < tolerance`, where what matters is
+  // the result of the comparison, which is already caught be the fcmp checks.
+  if (ID == Intrinsic::fabs || LFunc == LibFunc_fabsf ||
+      LFunc == LibFunc_fabs || LFunc == LibFunc_fabsl)
+    for (const auto &U : CI.users())
+      if (isa<CmpInst>(U))
+        return false;
+
+  return true; // Default is check.
+}
+
+// Populates the shadow call stack (which contains shadow values for every
+// floating-point parameter to the function).
+void NumericalStabilitySanitizer::populateShadowStack(
+    CallBase &CI, const TargetLibraryInfo &TLI, const ValueToShadowMap &Map) {
+  // Do not create a shadow stack for inline asm.
+  if (CI.isInlineAsm())
+    return;
+
+  // Do not bother if there are no FP args.
+  if (all_of(CI.operands(), [this](const Value *Arg) {
+        return Config.getExtendedFPType(Arg->getType()) == nullptr;
+      }))
+    return;
+
+  IRBuilder<> Builder(&CI);
+  SmallVector<Value *, 8> ArgShadows;
+  const bool ShouldCheckArgs = shouldCheckArgs(CI, TLI, CheckFunctionsFilter);
+  int ArgId = -1;
+  for (Value *Arg : CI.operands()) {
+    ++ArgId;
+    if (Config.getExtendedFPType(Arg->getType()) == nullptr)
+      continue; // Not an FT value.
+    Value *ArgShadow = Map.getShadow(Arg);
+    ArgShadows.push_back(ShouldCheckArgs ? emitCheck(Arg, ArgShadow, Builder,
+                                                     CheckLoc::makeArg(ArgId))
+                                         : ArgShadow);
+  }
+
+  // Do not create shadow stacks for intrinsics/known lib funcs.
+  if (Function *Fn = CI.getCalledFunction()) {
+    LibFunc LFunc;
+    if (Fn->isIntrinsic() || TLI.getLibFunc(*Fn, LFunc))
+      return;
+  }
+
+  // Set the shadow stack tag.
+  Builder.CreateStore(CI.getCalledOperand(), NsanShadowArgsTag);
+  TypeSize ShadowArgsOffsetBytes = TypeSize::getFixed(0);
+
+  unsigned ShadowArgId = 0;
+  for (const Value *Arg : CI.operands()) {
+    Type *VT = Arg->getType();
+    Type *ExtendedVT = Config.getExtendedFPType(VT);
+    if (ExtendedVT == nullptr)
+      continue; // Not an FT value.
+    Builder.CreateAlignedStore(
+        ArgShadows[ShadowArgId++],
+        Builder.CreateConstGEP2_64(NsanShadowArgsType, NsanShadowArgsPtr, 0,
+                                   ShadowArgsOffsetBytes),
+        Align(1), /*isVolatile=*/false);
+    TypeSize SlotSize = DL.getTypeStoreSize(ExtendedVT);
+    assert(!SlotSize.isScalable() && "unsupported");
+    ShadowArgsOffsetBytes += SlotSize;
+  }
+}
+
+// Internal part of emitCheck(). Returns a value that indicates whether
+// computation should continue with the shadow or resume by re-fextending the
+// value.
+enum class ContinuationType { // Keep in sync with runtime.
+  ContinueWithShadow = 0,
+  ResumeFromValue = 1,
+};
+Value *NumericalStabilitySanitizer::emitCheckInternal(Value *V, Value *ShadowV,
+                                                      IRBuilder<> &Builder,
+                                                      CheckLoc Loc) {
+  // Do not emit checks for constant values, this is redundant.
+  if (isa<Constant>(V))
+    return ConstantInt::get(
+        Builder.getInt32Ty(),
+        static_cast<int>(ContinuationType::ContinueWithShadow));
+
+  Type *Ty = V->getType();
+  if (const auto VT = ftValueTypeFromType(Ty))
+    return Builder.CreateCall(
+        NsanCheckValue[*VT],
+        {V, ShadowV, Loc.getType(Context), Loc.getValue(IntptrTy, Builder)});
+
+  if (Ty->isVectorTy()) {
+    auto *VecTy = cast<VectorType>(Ty);
+    // We currently skip scalable vector types in MappingConfig,
+    // thus we should not encounter any such types here.
+    assert(!VecTy->isScalableTy() &&
+           "Scalable vector types are not supported yet");
+    Value *CheckResult = nullptr;
+    for (int I = 0, E = VecTy->getElementCount().getFixedValue(); I < E; ++I) {
+      // We resume if any element resumes. Another option would be to create a
+      // vector shuffle with the array of ContinueWithShadow, but that is too
+      // complex.
+      Value *ExtractV = Builder.CreateExtractElement(V, I);
+      Value *ExtractShadowV = Builder.CreateExtractElement(ShadowV, I);
+      Value *ComponentCheckResult =
+          emitCheckInternal(ExtractV, ExtractShadowV, Builder, Loc);
+      CheckResult = CheckResult
+                        ? Builder.CreateOr(CheckResult, ComponentCheckResult)
+                        : ComponentCheckResult;
+    }
+    return CheckResult;
+  }
+  if (Ty->isArrayTy()) {
+    Value *CheckResult = nullptr;
+    for (int I = 0, E = Ty->getArrayNumElements(); I < E; ++I) {
+      Value *ExtractV = Builder.CreateExtractElement(V, I);
+      Value *ExtractShadowV = Builder.CreateExtractElement(ShadowV, I);
+      Value *ComponentCheckResult =
+          emitCheckInternal(ExtractV, ExtractShadowV, Builder, Loc);
+      CheckResult = CheckResult
+                        ? Builder.CreateOr(CheckResult, ComponentCheckResult)
+                        : ComponentCheckResult;
+    }
+    return CheckResult;
+  }
+  if (Ty->isStructTy()) {
+    Value *CheckResult = nullptr;
+    for (int I = 0, E = Ty->getStructNumElements(); I < E; ++I) {
+      if (Config.getExtendedFPType(Ty->getStructElementType(I)) == nullptr)
+        continue; // Only check FT values.
+      Value *ExtractV = Builder.CreateExtractValue(V, I);
+      Value *ExtractShadowV = Builder.CreateExtractElement(ShadowV, I);
+      Value *ComponentCheckResult =
+          emitCheckInternal(ExtractV, ExtractShadowV, Builder, Loc);
+      CheckResult = CheckResult
+                        ? Builder.CreateOr(CheckResult, ComponentCheckResult)
+                        : ComponentCheckResult;
+    }
+    if (!CheckResult)
+      return ConstantInt::get(
+          Builder.getInt32Ty(),
+          static_cast<int>(ContinuationType::ContinueWithShadow));
+    return CheckResult;
+  }
+
+  llvm_unreachable("not implemented");
+}
+
+// Inserts a runtime check of V against its shadow value ShadowV.
+// We check values whenever they escape: on return, call, stores, and
+// insertvalue.
+// Returns the shadow value that should be used to continue the computations,
+// depending on the answer from the runtime.
+// FIXME: Should we check on select ? phi ?
+Value *NumericalStabilitySanitizer::emitCheck(Value *V, Value *ShadowV,
+                                              IRBuilder<> &Builder,
+                                              CheckLoc Loc) {
+  // Do not emit checks for constant values, this is redundant.
+  if (isa<Constant>(V))
+    return ShadowV;
+
+  if (Instruction *Inst = dyn_cast<Instruction>(V)) {
+    Function *F = Inst->getFunction();
+    if (CheckFunctionsFilter && !CheckFunctionsFilter->match(F->getName())) {
+      return ShadowV;
+    }
+  }
+
+  Value *CheckResult = emitCheckInternal(V, ShadowV, Builder, Loc);
+  Value *ICmpEQ = Builder.CreateICmpEQ(
+      CheckResult,
+      ConstantInt::get(Builder.getInt32Ty(),
+                       static_cast<int>(ContinuationType::ResumeFromValue)));
+  return Builder.CreateSelect(
+      ICmpEQ, Builder.CreateFPExt(V, Config.getExtendedFPType(V->getType())),
+      ShadowV);
+}
+
+static Instruction *getNextInstructionOrDie(Instruction &Inst) {
+  assert(Inst.getNextNode() && "instruction is a terminator");
+  return Inst.getNextNode();
+}
+
+// Inserts a check that fcmp on shadow values are consistent with that on base
+// values.
+void NumericalStabilitySanitizer::emitFCmpCheck(FCmpInst &FCmp,
+                                                const ValueToShadowMap &Map) {
+  if (!ClInstrumentFCmp)
+    return;
+
+  Function *F = FCmp.getFunction();
+  if (CheckFunctionsFilter && !CheckFunctionsFilter->match(F->getName())) {
+    return;
+  }
+
+  Value *LHS = FCmp.getOperand(0);
+  if (Config.getExtendedFPType(LHS->getType()) == nullptr)
+    return;
+  Value *RHS = FCmp.getOperand(1);
+
+  // Split the basic block. On mismatch, we'll jump to the new basic block with
+  // a call to the runtime for error reporting.
+  BasicBlock *FCmpBB = FCmp.getParent();
+  BasicBlock *NextBB = FCmpBB->splitBasicBlock(getNextInstructionOrDie(FCmp));
+  // Remove the newly created terminator unconditional branch.
+  FCmpBB->back().eraseFromParent();
+  BasicBlock *FailBB =
+      BasicBlock::Create(Context, "", FCmpBB->getParent(), NextBB);
+
+  // Create the shadow fcmp and comparison between the fcmps.
+  IRBuilder<> FCmpBuilder(FCmpBB);
+  FCmpBuilder.SetCurrentDebugLocation(FCmp.getDebugLoc());
+  Value *ShadowLHS = Map.getShadow(LHS);
+  Value *ShadowRHS = Map.getShadow(RHS);
+  // See comment on ClTruncateFCmpEq.
+  if (FCmp.isEquality() && ClTruncateFCmpEq) {
+    Type *Ty = ShadowLHS->getType();
+    ShadowLHS = FCmpBuilder.CreateFPExt(
+        FCmpBuilder.CreateFPTrunc(ShadowLHS, LHS->getType()), Ty);
+    ShadowRHS = FCmpBuilder.CreateFPExt(
+        FCmpBuilder.CreateFPTrunc(ShadowRHS, RHS->getType()), Ty);
+  }
+  Value *ShadowFCmp =
+      FCmpBuilder.CreateFCmp(FCmp.getPredicate(), ShadowLHS, ShadowRHS);
+  Value *OriginalAndShadowFcmpMatch =
+      FCmpBuilder.CreateICmpEQ(&FCmp, ShadowFCmp);
+
+  if (OriginalAndShadowFcmpMatch->getType()->isVectorTy()) {
+    // If we have a vector type, `OriginalAndShadowFcmpMatch` is a vector of i1,
+    // where an element is true if the corresponding elements in original and
+    // shadow are the same. We want all elements to be 1.
+    OriginalAndShadowFcmpMatch =
+        FCmpBuilder.CreateAndReduce(OriginalAndShadowFcmpMatch);
+  }
+
+  // Use MDBuilder(*C).createLikelyBranchWeights() because "match" is the common
+  // case.
+  FCmpBuilder.CreateCondBr(OriginalAndShadowFcmpMatch, NextBB, FailBB,
+                           MDBuilder(Context).createLikelyBranchWeights());
+
+  // Fill in FailBB.
+  IRBuilder<> FailBuilder(FailBB);
+  FailBuilder.SetCurrentDebugLocation(FCmp.getDebugLoc());
+
+  const auto EmitFailCall = [this, &FCmp, &FCmpBuilder,
+                             &FailBuilder](Value *L, Value *R, Value *ShadowL,
+                                           Value *ShadowR, Value *Result,
+                                           Value *ShadowResult) {
+    Type *FT = L->getType();
+    FunctionCallee *Callee = nullptr;
+    if (FT->isFloatTy()) {
+      Callee = &(NsanFCmpFail[kFloat]);
+    } else if (FT->isDoubleTy()) {
+      Callee = &(NsanFCmpFail[kDouble]);
+    } else if (FT->isX86_FP80Ty()) {
+      // FIXME: make NsanFCmpFailLongDouble work.
+      Callee = &(NsanFCmpFail[kDouble]);
+      L = FailBuilder.CreateFPTrunc(L, Type::getDoubleTy(Context));
+      R = FailBuilder.CreateFPTrunc(L, Type::getDoubleTy(Context));
+    } else {
+      llvm_unreachable("not implemented");
+    }
+    FailBuilder.CreateCall(*Callee, {L, R, ShadowL, ShadowR,
+                                     ConstantInt::get(FCmpBuilder.getInt32Ty(),
+                                                      FCmp.getPredicate()),
+                                     Result, ShadowResult});
+  };
+  if (LHS->getType()->isVectorTy()) {
+    for (int I = 0, E = cast<VectorType>(LHS->getType())
+                            ->getElementCount()
+                            .getFixedValue();
+         I < E; ++I) {
+      Value *ExtractLHS = FailBuilder.CreateExtractElement(LHS, I);
+      Value *ExtractRHS = FailBuilder.CreateExtractElement(RHS, I);
+      Value *ExtractShaodwLHS = FailBuilder.CreateExtractElement(ShadowLHS, I);
+      Value *ExtractShaodwRHS = FailBuilder.CreateExtractElement(ShadowRHS, I);
+      Value *ExtractFCmp = FailBuilder.CreateExtractElement(&FCmp, I);
+      Value *ExtractShadowFCmp =
+          FailBuilder.CreateExtractElement(ShadowFCmp, I);
+      EmitFailCall(ExtractLHS, ExtractRHS, ExtractShaodwLHS, ExtractShaodwRHS,
+                   ExtractFCmp, ExtractShadowFCmp);
+    }
+  } else {
+    EmitFailCall(LHS, RHS, ShadowLHS, ShadowRHS, &FCmp, ShadowFCmp);
+  }
+  FailBuilder.CreateBr(NextBB);
+
+  ++NumInstrumentedFCmp;
+}
+
+// Creates a shadow phi value for any phi that defines a value of FT type.
+PHINode *NumericalStabilitySanitizer::maybeCreateShadowPhi(
+    PHINode &Phi, const TargetLibraryInfo &TLI) {
+  Type *VT = Phi.getType();
+  Type *ExtendedVT = Config.getExtendedFPType(VT);
+  if (ExtendedVT == nullptr)
+    return nullptr; // Not an FT value.
+  // The phi operands are shadow values and are not available when the phi is
+  // created. They will be populated in a final phase, once all shadow values
+  // have been created.
+  PHINode *Shadow = PHINode::Create(ExtendedVT, Phi.getNumIncomingValues());
+  Shadow->insertAfter(&Phi);
+  return Shadow;
+}
+
+Value *NumericalStabilitySanitizer::handleLoad(LoadInst &Load, Type *VT,
+                                               Type *ExtendedVT) {
+  IRBuilder<> Builder(getNextInstructionOrDie(Load));
+  Builder.SetCurrentDebugLocation(Load.getDebugLoc());
+  if (addrPointsToConstantData(Load.getPointerOperand())) {
+    // No need to look into the shadow memory, the value is a constant. Just
+    // convert from FT to 2FT.
+    return Builder.CreateFPExt(&Load, ExtendedVT);
+  }
+
+  // if (%shadowptr == &)
+  //    %shadow = fpext %v
+  // else
+  //    %shadow = load (ptrcast %shadow_ptr))
+  // Considered options here:
+  //  - Have `NsanGetShadowPtrForLoad` return a fixed address
+  //    &__nsan_unknown_value_shadow_address that is valid to load from, and
+  //    use a select. This has the advantage that the generated IR is simpler.
+  //  - Have `NsanGetShadowPtrForLoad` return nullptr.  Because `select` does
+  //    not short-circuit, dereferencing the returned pointer is no longer an
+  //    option, have to split and create a separate basic block. This has the
+  //    advantage of being easier to debug because it crashes if we ever mess
+  //    up.
+
+  const auto Extents = getMemoryExtentsOrDie(VT);
+  Value *ShadowPtr = Builder.CreateCall(
+      NsanGetShadowPtrForLoad[Extents.ValueType],
+      {Load.getPointerOperand(), ConstantInt::get(IntptrTy, Extents.NumElts)});
+  ++NumInstrumentedFTLoads;
+
+  // Split the basic block.
+  BasicBlock *LoadBB = Load.getParent();
+  BasicBlock *NextBB = LoadBB->splitBasicBlock(Builder.GetInsertPoint());
+  // Create the two options for creating the shadow value.
+  BasicBlock *ShadowLoadBB =
+      BasicBlock::Create(Context, "", LoadBB->getParent(), NextBB);
+  BasicBlock *FExtBB =
+      BasicBlock::Create(Context, "", LoadBB->getParent(), NextBB);
+
+  // Replace the newly created terminator unconditional branch by a conditional
+  // branch to one of the options.
+  {
+    LoadBB->back().eraseFromParent();
+    IRBuilder<> LoadBBBuilder(LoadBB); // The old builder has been invalidated.
+    LoadBBBuilder.SetCurrentDebugLocation(Load.getDebugLoc());
+    LoadBBBuilder.CreateCondBr(LoadBBBuilder.CreateIsNull(ShadowPtr), FExtBB,
+                               ShadowLoadBB);
+  }
+
+  // Fill in ShadowLoadBB.
+  IRBuilder<> ShadowLoadBBBuilder(ShadowLoadBB);
+  ShadowLoadBBBuilder.SetCurrentDebugLocation(Load.getDebugLoc());
+  Value *ShadowLoad = ShadowLoadBBBuilder.CreateAlignedLoad(
+      ExtendedVT, ShadowPtr, Align(1), Load.isVolatile());
+  if (ClCheckLoads) {
+    ShadowLoad = emitCheck(&Load, ShadowLoad, ShadowLoadBBBuilder,
+                           CheckLoc::makeLoad(Load.getPointerOperand()));
+  }
+  ShadowLoadBBBuilder.CreateBr(NextBB);
+
+  // Fill in FExtBB.
+  IRBuilder<> FExtBBBuilder(FExtBB);
+  FExtBBBuilder.SetCurrentDebugLocation(Load.getDebugLoc());
+  Value *FExt = FExtBBBuilder.CreateFPExt(&Load, ExtendedVT);
+  FExtBBBuilder.CreateBr(NextBB);
+
+  // The shadow value come from any of the options.
+  IRBuilder<> NextBBBuilder(&*NextBB->begin());
+  NextBBBuilder.SetCurrentDebugLocation(Load.getDebugLoc());
+  PHINode *ShadowPhi = NextBBBuilder.CreatePHI(ExtendedVT, 2);
+  ShadowPhi->addIncoming(ShadowLoad, ShadowLoadBB);
+  ShadowPhi->addIncoming(FExt, FExtBB);
+  return ShadowPhi;
+}
+
+Value *NumericalStabilitySanitizer::handleTrunc(const FPTruncInst &Trunc,
+                                                Type *VT, Type *ExtendedVT,
+                                                const ValueToShadowMap &Map,
+                                                IRBuilder<> &Builder) {
+  Value *OrigSource = Trunc.getOperand(0);
+  Type *OrigSourceTy = OrigSource->getType();
+  Type *ExtendedSourceTy = Config.getExtendedFPType(OrigSourceTy);
+
+  // When truncating:
+  //  - (A) If the source has a shadow, we truncate from the shadow, else we
+  //    truncate from the original source.
+  //  - (B) If the shadow of the source is larger than the shadow of the dest,
+  //    we still need a truncate. Else, the shadow of the source is the same
+  //    type as the shadow of the dest (because mappings are non-decreasing), so
+  //   we don't need to emit a truncate.
+  // Examples,
+  //   with a mapping of {f32->f64;f64->f80;f80->f128}
+  //     fptrunc double   %1 to float     ->  fptrunc x86_fp80 s(%1) to double
+  //     fptrunc x86_fp80 %1 to float     ->  fptrunc fp128    s(%1) to double
+  //     fptrunc fp128    %1 to float     ->  fptrunc fp128    %1    to double
+  //     fptrunc x86_fp80 %1 to double    ->  x86_fp80 s(%1)
+  //     fptrunc fp128    %1 to double    ->  fptrunc fp128 %1 to x86_fp80
+  //     fptrunc fp128    %1 to x86_fp80  ->  fp128 %1
+  //   with a mapping of {f32->f64;f64->f128;f80->f128}
+  //     fptrunc double   %1 to float     ->  fptrunc fp128    s(%1) to double
+  //     fptrunc x86_fp80 %1 to float     ->  fptrunc fp128    s(%1) to double
+  //     fptrunc fp128    %1 to float     ->  fptrunc fp128    %1    to double
+  //     fptrunc x86_fp80 %1 to double    ->  fp128 %1
+  //     fptrunc fp128    %1 to double    ->  fp128 %1
+  //     fptrunc fp128    %1 to x86_fp80  ->  fp128 %1
+  //   with a mapping of {f32->f32;f64->f32;f80->f64}
+  //     fptrunc double   %1 to float     ->  float s(%1)
+  //     fptrunc x86_fp80 %1 to float     ->  fptrunc double    s(%1) to float
+  //     fptrunc fp128    %1 to float     ->  fptrunc fp128     %1    to float
+  //     fptrunc x86_fp80 %1 to double    ->  fptrunc double    s(%1) to float
+  //     fptrunc fp128    %1 to double    ->  fptrunc fp128     %1    to float
+  //     fptrunc fp128    %1 to x86_fp80  ->  fptrunc fp128     %1    to double
+
+  // See (A) above.
+  Value *Source = ExtendedSourceTy ? Map.getShadow(OrigSource) : OrigSource;
+  Type *SourceTy = ExtendedSourceTy ? ExtendedSourceTy : OrigSourceTy;
+  // See (B) above.
+  if (SourceTy == ExtendedVT)
+    return Source;
+
+  return Builder.CreateFPTrunc(Source, ExtendedVT);
+}
+
+Value *NumericalStabilitySanitizer::handleExt(const FPExtInst &Ext, Type *VT,
+                                              Type *ExtendedVT,
+                                              const ValueToShadowMap &Map,
+                                              IRBuilder<> &Builder) {
+  Value *OrigSource = Ext.getOperand(0);
+  Type *OrigSourceTy = OrigSource->getType();
+  Type *ExtendedSourceTy = Config.getExtendedFPType(OrigSourceTy);
+  // When extending:
+  //  - (A) If the source has a shadow, we extend from the shadow, else we
+  //    extend from the original source.
+  //  - (B) If the shadow of the dest is larger than the shadow of the source,
+  //    we still need an extend. Else, the shadow of the source is the same
+  //    type as the shadow of the dest (because mappings are non-decreasing), so
+  //    we don't need to emit an extend.
+  // Examples,
+  //   with a mapping of {f32->f64;f64->f80;f80->f128}
+  //     fpext half    %1 to float     ->  fpext half     %1    to double
+  //     fpext half    %1 to double    ->  fpext half     %1    to x86_fp80
+  //     fpext half    %1 to x86_fp80  ->  fpext half     %1    to fp128
+  //     fpext float   %1 to double    ->  double s(%1)
+  //     fpext float   %1 to x86_fp80  ->  fpext double   s(%1) to fp128
+  //     fpext double  %1 to x86_fp80  ->  fpext x86_fp80 s(%1) to fp128
+  //   with a mapping of {f32->f64;f64->f128;f80->f128}
+  //     fpext half    %1 to float     ->  fpext half     %1    to double
+  //     fpext half    %1 to double    ->  fpext half     %1    to fp128
+  //     fpext half    %1 to x86_fp80  ->  fpext half     %1    to fp128
+  //     fpext float   %1 to double    ->  fpext double   s(%1) to fp128
+  //     fpext float   %1 to x86_fp80  ->  fpext double   s(%1) to fp128
+  //     fpext double  %1 to x86_fp80  ->  fp128 s(%1)
+  //   with a mapping of {f32->f32;f64->f32;f80->f64}
+  //     fpext half    %1 to float     ->  fpext half     %1    to float
+  //     fpext half    %1 to double    ->  fpext half     %1    to float
+  //     fpext half    %1 to x86_fp80  ->  fpext half     %1    to double
+  //     fpext float   %1 to double    ->  s(%1)
+  //     fpext float   %1 to x86_fp80  ->  fpext float    s(%1) to double
+  //     fpext double  %1 to x86_fp80  ->  fpext float    s(%1) to double
+
+  // See (A) above.
+  Value *Source = ExtendedSourceTy ? Map.getShadow(OrigSource) : OrigSource;
+  Type *SourceTy = ExtendedSourceTy ? ExtendedSourceTy : OrigSourceTy;
+  // See (B) above.
+  if (SourceTy == ExtendedVT)
+    return Source;
+
+  return Builder.CreateFPExt(Source, ExtendedVT);
+}
+
+namespace {
+
+// FIXME: This should be tablegen-ed.
+
+struct KnownIntrinsic {
+  struct WidenedIntrinsic {
+    const char *NarrowName;
+    Intrinsic::ID ID; // wide id.
+    using FnTypeFactory = FunctionType *(*)(LLVMContext &);
+    FnTypeFactory MakeFnTy;
+  };
+
+  static const char *get(LibFunc LFunc);
+
+  // Given an intrinsic with an `FT` argument, try to find a wider intrinsic
+  // that applies the same operation on the shadow argument.
+  // Options are:
+  //  - pass in the ID and full function type,
+  //  - pass in the name, which includes the function type through mangling.
+  static const WidenedIntrinsic *widen(StringRef Name);
+
+private:
+  struct LFEntry {
+    LibFunc LFunc;
+    const char *IntrinsicName;
+  };
+  static const LFEntry kLibfuncIntrinsics[];
+
+  static const WidenedIntrinsic kWidenedIntrinsics[];
+};
+
+FunctionType *Make_Double_Double(LLVMContext &C) {
----------------
MaskRay wrote:

`functionName`

https://github.com/llvm/llvm-project/pull/85916


More information about the llvm-commits mailing list