Refactor BitcodeReader: move Metadata and ValueId handling in their own class/file

Summary:
I'm planning on changing the way we load metadata to enable laziness.
I'm getting lost in this gigantic files, and gigantic class that is the bitcode
reader. This is a first toward splitting it in a few coarse components that
are more easily understandable.

Reviewers: pcc, tejohnson

Subscribers: mgorny, llvm-commits, dexonsmith

Differential Revision: https://reviews.llvm.org/D27646

llvm-svn: 289461

llvm/lib/Bitcode/Reader/CMakeLists.txt

@@ -2,6 +2,8 @@ add_llvm_library(LLVMBitReader
   BitReader.cpp
   BitcodeReader.cpp
   BitstreamReader.cpp
+  MetadataLoader.cpp
+  ValueList.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${LLVM_MAIN_INCLUDE_DIR}/llvm/Bitcode

llvm/lib/Bitcode/Reader/MetadataLoader.h

@@ -0,0 +1,79 @@
+//===-- Bitcode/Reader/MetadataLoader.h - Load Metadatas -------*- C++ -*-====//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This class handles loading Metadatas.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_LIB_BITCODE_READER_METADATALOADER_H
+#define LLVM_LIB_BITCODE_READER_METADATALOADER_H
+
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/Support/Error.h"
+
+#include <functional>
+#include <memory>
+
+namespace llvm {
+class BitcodeReaderValueList;
+class BitstreamCursor;
+class DISubprogram;
+class Error;
+class Function;
+class Instruction;
+class Metadata;
+class MDNode;
+class Module;
+class Type;
+
+/// Helper class that handles loading Metadatas and keeping them available.
+class MetadataLoader {
+  class MetadataLoaderImpl;
+  std::unique_ptr<MetadataLoaderImpl> Pimpl;
+  Error parseMetadata(bool ModuleLevel);
+
+public:
+  ~MetadataLoader();
+  MetadataLoader(BitstreamCursor &Stream, Module &TheModule,
+                 BitcodeReaderValueList &ValueList,
+                 std::function<Type *(unsigned)> getTypeByID);
+  MetadataLoader &operator=(MetadataLoader &&);
+  MetadataLoader(MetadataLoader &&);
+
+  // Parse a module metadata block
+  Error parseModuleMetadata() { return parseMetadata(true); }
+
+  // Parse a function metadata block
+  Error parseFunctionMetadata() { return parseMetadata(false); }
+
+  // Return true there are remaining unresolved forward references.
+  bool hasFwdRefs() const;
+
+  /// Return the given metadata, creating a replaceable forward reference if
+  /// necessary.
+  Metadata *getMetadataFwdRef(unsigned Idx);
+
+  MDNode *getMDNodeFwdRefOrNull(unsigned Idx);
+
+  /// Return the DISubprogra metadata for a Function if any, null otherwise.
+  DISubprogram *lookupSubprogramForFunction(Function *F);
+
+  /// Parse a `METADATA_ATTACHMENT` block for a function.
+  Error parseMetadataAttachment(
+      Function &F, const SmallVectorImpl<Instruction *> &InstructionList);
+
+  /// Parse a `METADATA_KIND` block for the current module.
+  Error parseMetadataKinds();
+
+  unsigned size() const;
+  void shrinkTo(unsigned N);
+};
+}
+
+#endif // LLVM_LIB_BITCODE_READER_METADATALOADER_H

llvm/lib/Bitcode/Reader/ValueList.cpp

@@ -0,0 +1,199 @@
+//===----- ValueList.cpp - Internal BitcodeReader implementation ----------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#include "ValueList.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/Instructions.h"
+
+using namespace llvm;
+
+namespace llvm {
+namespace {
+
+/// \brief A class for maintaining the slot number definition
+/// as a placeholder for the actual definition for forward constants defs.
+class ConstantPlaceHolder : public ConstantExpr {
+  void operator=(const ConstantPlaceHolder &) = delete;
+
+public:
+  // allocate space for exactly one operand
+  void *operator new(size_t s) { return User::operator new(s, 1); }
+  explicit ConstantPlaceHolder(Type *Ty, LLVMContext &Context)
+      : ConstantExpr(Ty, Instruction::UserOp1, &Op<0>(), 1) {
+    Op<0>() = UndefValue::get(Type::getInt32Ty(Context));
+  }
+
+  /// \brief Methods to support type inquiry through isa, cast, and dyn_cast.
+  static bool classof(const Value *V) {
+    return isa<ConstantExpr>(V) &&
+           cast<ConstantExpr>(V)->getOpcode() == Instruction::UserOp1;
+  }
+
+  /// Provide fast operand accessors
+  DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value);
+};
+
+} // end anonymous namespace
+
+// FIXME: can we inherit this from ConstantExpr?
+template <>
+struct OperandTraits<ConstantPlaceHolder>
+    : public FixedNumOperandTraits<ConstantPlaceHolder, 1> {};
+DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ConstantPlaceHolder, Value)
+
+} // end namespace llvm
+
+void BitcodeReaderValueList::assignValue(Value *V, unsigned Idx) {
+  if (Idx == size()) {
+    push_back(V);
+    return;
+  }
+
+  if (Idx >= size())
+    resize(Idx + 1);
+
+  WeakVH &OldV = ValuePtrs[Idx];
+  if (!OldV) {
+    OldV = V;
+    return;
+  }
+
+  // Handle constants and non-constants (e.g. instrs) differently for
+  // efficiency.
+  if (Constant *PHC = dyn_cast<Constant>(&*OldV)) {
+    ResolveConstants.push_back(std::make_pair(PHC, Idx));
+    OldV = V;
+  } else {
+    // If there was a forward reference to this value, replace it.
+    Value *PrevVal = OldV;
+    OldV->replaceAllUsesWith(V);
+    delete PrevVal;
+  }
+}
+
+Constant *BitcodeReaderValueList::getConstantFwdRef(unsigned Idx, Type *Ty) {
+  if (Idx >= size())
+    resize(Idx + 1);
+
+  if (Value *V = ValuePtrs[Idx]) {
+    if (Ty != V->getType())
+      report_fatal_error("Type mismatch in constant table!");
+    return cast<Constant>(V);
+  }
+
+  // Create and return a placeholder, which will later be RAUW'd.
+  Constant *C = new ConstantPlaceHolder(Ty, Context);
+  ValuePtrs[Idx] = C;
+  return C;
+}
+
+Value *BitcodeReaderValueList::getValueFwdRef(unsigned Idx, Type *Ty) {
+  // Bail out for a clearly invalid value. This would make us call resize(0)
+  if (Idx == std::numeric_limits<unsigned>::max())
+    return nullptr;
+
+  if (Idx >= size())
+    resize(Idx + 1);
+
+  if (Value *V = ValuePtrs[Idx]) {
+    // If the types don't match, it's invalid.
+    if (Ty && Ty != V->getType())
+      return nullptr;
+    return V;
+  }
+
+  // No type specified, must be invalid reference.
+  if (!Ty)
+    return nullptr;
+
+  // Create and return a placeholder, which will later be RAUW'd.
+  Value *V = new Argument(Ty);
+  ValuePtrs[Idx] = V;
+  return V;
+}
+
+/// Once all constants are read, this method bulk resolves any forward
+/// references.  The idea behind this is that we sometimes get constants (such
+/// as large arrays) which reference *many* forward ref constants.  Replacing
+/// each of these causes a lot of thrashing when building/reuniquing the
+/// constant.  Instead of doing this, we look at all the uses and rewrite all
+/// the place holders at once for any constant that uses a placeholder.
+void BitcodeReaderValueList::resolveConstantForwardRefs() {
+  // Sort the values by-pointer so that they are efficient to look up with a
+  // binary search.
+  std::sort(ResolveConstants.begin(), ResolveConstants.end());
+
+  SmallVector<Constant *, 64> NewOps;
+
+  while (!ResolveConstants.empty()) {
+    Value *RealVal = operator[](ResolveConstants.back().second);
+    Constant *Placeholder = ResolveConstants.back().first;
+    ResolveConstants.pop_back();
+
+    // Loop over all users of the placeholder, updating them to reference the
+    // new value.  If they reference more than one placeholder, update them all
+    // at once.
+    while (!Placeholder->use_empty()) {
+      auto UI = Placeholder->user_begin();
+      User *U = *UI;
+
+      // If the using object isn't uniqued, just update the operands.  This
+      // handles instructions and initializers for global variables.
+      if (!isa<Constant>(U) || isa<GlobalValue>(U)) {
+        UI.getUse().set(RealVal);
+        continue;
+      }
+
+      // Otherwise, we have a constant that uses the placeholder.  Replace that
+      // constant with a new constant that has *all* placeholder uses updated.
+      Constant *UserC = cast<Constant>(U);
+      for (User::op_iterator I = UserC->op_begin(), E = UserC->op_end(); I != E;
+           ++I) {
+        Value *NewOp;
+        if (!isa<ConstantPlaceHolder>(*I)) {
+          // Not a placeholder reference.
+          NewOp = *I;
+        } else if (*I == Placeholder) {
+          // Common case is that it just references this one placeholder.
+          NewOp = RealVal;
+        } else {
+          // Otherwise, look up the placeholder in ResolveConstants.
+          ResolveConstantsTy::iterator It = std::lower_bound(
+              ResolveConstants.begin(), ResolveConstants.end(),
+              std::pair<Constant *, unsigned>(cast<Constant>(*I), 0));
+          assert(It != ResolveConstants.end() && It->first == *I);
+          NewOp = operator[](It->second);
+        }
+
+        NewOps.push_back(cast<Constant>(NewOp));
+      }
+
+      // Make the new constant.
+      Constant *NewC;
+      if (ConstantArray *UserCA = dyn_cast<ConstantArray>(UserC)) {
+        NewC = ConstantArray::get(UserCA->getType(), NewOps);
+      } else if (ConstantStruct *UserCS = dyn_cast<ConstantStruct>(UserC)) {
+        NewC = ConstantStruct::get(UserCS->getType(), NewOps);
+      } else if (isa<ConstantVector>(UserC)) {
+        NewC = ConstantVector::get(NewOps);
+      } else {
+        assert(isa<ConstantExpr>(UserC) && "Must be a ConstantExpr.");
+        NewC = cast<ConstantExpr>(UserC)->getWithOperands(NewOps);
+      }
+
+      UserC->replaceAllUsesWith(NewC);
+      UserC->destroyConstant();
+      NewOps.clear();
+    }
+
+    // Update all ValueHandles, they should be the only users at this point.
+    Placeholder->replaceAllUsesWith(RealVal);
+    delete Placeholder;
+  }
+}

llvm/lib/Bitcode/Reader/ValueList.h

@@ -0,0 +1,76 @@
+//===-- Bitcode/Reader/ValueEnumerator.h - Number values --------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This class gives values and types Unique ID's.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/IR/LLVMContext.h"
+#include "llvm/IR/ValueHandle.h"
+
+#include <vector>
+
+namespace llvm {
+class Constant;
+
+class BitcodeReaderValueList {
+  std::vector<WeakVH> ValuePtrs;
+
+  /// As we resolve forward-referenced constants, we add information about them
+  /// to this vector.  This allows us to resolve them in bulk instead of
+  /// resolving each reference at a time.  See the code in
+  /// ResolveConstantForwardRefs for more information about this.
+  ///
+  /// The key of this vector is the placeholder constant, the value is the slot
+  /// number that holds the resolved value.
+  typedef std::vector<std::pair<Constant *, unsigned>> ResolveConstantsTy;
+  ResolveConstantsTy ResolveConstants;
+  LLVMContext &Context;
+
+public:
+  BitcodeReaderValueList(LLVMContext &C) : Context(C) {}
+  ~BitcodeReaderValueList() {
+    assert(ResolveConstants.empty() && "Constants not resolved?");
+  }
+
+  // vector compatibility methods
+  unsigned size() const { return ValuePtrs.size(); }
+  void resize(unsigned N) { ValuePtrs.resize(N); }
+  void push_back(Value *V) { ValuePtrs.emplace_back(V); }
+
+  void clear() {
+    assert(ResolveConstants.empty() && "Constants not resolved?");
+    ValuePtrs.clear();
+  }
+
+  Value *operator[](unsigned i) const {
+    assert(i < ValuePtrs.size());
+    return ValuePtrs[i];
+  }
+
+  Value *back() const { return ValuePtrs.back(); }
+  void pop_back() { ValuePtrs.pop_back(); }
+  bool empty() const { return ValuePtrs.empty(); }
+
+  void shrinkTo(unsigned N) {
+    assert(N <= size() && "Invalid shrinkTo request!");
+    ValuePtrs.resize(N);
+  }
+
+  Constant *getConstantFwdRef(unsigned Idx, Type *Ty);
+  Value *getValueFwdRef(unsigned Idx, Type *Ty);
+
+  void assignValue(Value *V, unsigned Idx);
+
+  /// Once all constants are read, this method bulk resolves any forward
+  /// references.
+  void resolveConstantForwardRefs();
+};
+
+} // namespace llvm