path: root/include/llvm/Analysis/RegionInfo.h

//===- RegionInfo.h - SESE region analysis ----------------------*- C++ -*-===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// Calculate a program structure tree built out of single entry single exit
// regions.
// The basic ideas are taken from "The Program Structure Tree - Richard Johnson,
// David Pearson, Keshav Pingali - 1994", however enriched with ideas from "The
// Refined Process Structure Tree - Jussi Vanhatalo, Hagen Voelyer, Jana
// Koehler - 2009".
// The algorithm to calculate these data structures however is completely
// different, as it takes advantage of existing information already available
// in (Post)dominace tree and dominance frontier passes. This leads to a simpler
// and in practice hopefully better performing algorithm. The runtime of the
// algorithms described in the papers above are both linear in graph size,
// O(V+E), whereas this algorithm is not, as the dominance frontier information
// itself is not, but in practice runtime seems to be in the order of magnitude
// of dominance tree calculation.
//
//===----------------------------------------------------------------------===//

#ifndef LLVM_ANALYSIS_REGIONINFO_H
#define LLVM_ANALYSIS_REGIONINFO_H

#include "llvm/ADT/PointerIntPair.h"
#include "llvm/Analysis/DominanceFrontier.h"
#include "llvm/Analysis/PostDominators.h"
#include "llvm/Support/Allocator.h"
#include <map>

namespace llvm {

class Region;
class RegionInfo;
class raw_ostream;
class Loop;
class LoopInfo;

/// @brief Marker class to iterate over the elements of a Region in flat mode.
///
/// The class is used to either iterate in Flat mode or by not using it to not
/// iterate in Flat mode.  During a Flat mode iteration all Regions are entered
/// and the iteration returns every BasicBlock.  If the Flat mode is not
/// selected for SubRegions just one RegionNode containing the subregion is
/// returned.
template <class GraphType>
class FlatIt {};

/// @brief A RegionNode represents a subregion or a BasicBlock that is part of a
/// Region.
class RegionNode {
  RegionNode(const RegionNode &) LLVM_DELETED_FUNCTION;
  const RegionNode &operator=(const RegionNode &) LLVM_DELETED_FUNCTION;

protected:
  /// This is the entry basic block that starts this region node.  If this is a
  /// BasicBlock RegionNode, then entry is just the basic block, that this
  /// RegionNode represents.  Otherwise it is the entry of this (Sub)RegionNode.
  ///
  /// In the BBtoRegionNode map of the parent of this node, BB will always map
  /// to this node no matter which kind of node this one is.
  ///
  /// The node can hold either a Region or a BasicBlock.
  /// Use one bit to save, if this RegionNode is a subregion or BasicBlock
  /// RegionNode.
  PointerIntPair<BasicBlock*, 1, bool> entry;

  /// @brief The parent Region of this RegionNode.
  /// @see getParent()
  Region* parent;

public:
  /// @brief Create a RegionNode.
  ///
  /// @param Parent      The parent of this RegionNode.
  /// @param Entry       The entry BasicBlock of the RegionNode.  If this
  ///                    RegionNode represents a BasicBlock, this is the
  ///                    BasicBlock itself.  If it represents a subregion, this
  ///                    is the entry BasicBlock of the subregion.
  /// @param isSubRegion If this RegionNode represents a SubRegion.
  inline RegionNode(Region* Parent, BasicBlock* Entry, bool isSubRegion = 0)
    : entry(Entry, isSubRegion), parent(Parent) {}

  /// @brief Get the parent Region of this RegionNode.
  ///
  /// The parent Region is the Region this RegionNode belongs to. If for
  /// example a BasicBlock is element of two Regions, there exist two
  /// RegionNodes for this BasicBlock. Each with the getParent() function
  /// pointing to the Region this RegionNode belongs to.
  ///
  /// @return Get the parent Region of this RegionNode.
  inline Region* getParent() const { return parent; }

  /// @brief Get the entry BasicBlock of this RegionNode.
  ///
  /// If this RegionNode represents a BasicBlock this is just the BasicBlock
  /// itself, otherwise we return the entry BasicBlock of the Subregion
  ///
  /// @return The entry BasicBlock of this RegionNode.
  inline BasicBlock* getEntry() const { return entry.getPointer(); }

  /// @brief Get the content of this RegionNode.
  ///
  /// This can be either a BasicBlock or a subregion. Before calling getNodeAs()
  /// check the type of the content with the isSubRegion() function call.
  ///
  /// @return The content of this RegionNode.
  template<class T>
  inline T* getNodeAs() const;

  /// @brief Is this RegionNode a subregion?
  ///
  /// @return True if it contains a subregion. False if it contains a
  ///         BasicBlock.
  inline bool isSubRegion() const {
    return entry.getInt();
  }
};

/// Print a RegionNode.
inline raw_ostream &operator<<(raw_ostream &OS, const RegionNode &Node);

template<>
inline BasicBlock* RegionNode::getNodeAs<BasicBlock>() const {
  assert(!isSubRegion() && "This is not a BasicBlock RegionNode!");
  return getEntry();
}

template<>
inline Region* RegionNode::getNodeAs<Region>() const {
  assert(isSubRegion() && "This is not a subregion RegionNode!");
  return reinterpret_cast<Region*>(const_cast<RegionNode*>(this));
}

//===----------------------------------------------------------------------===//
/// @brief A single entry single exit Region.
///
/// A Region is a connected subgraph of a control flow graph that has exactly
/// two connections to the remaining graph. It can be used to analyze or
/// optimize parts of the control flow graph.
///
/// A <em> simple Region </em> is connected to the remaining graph by just two
/// edges. One edge entering the Region and another one leaving the Region.
///
/// An <em> extended Region </em> (or just Region) is a subgraph that can be
/// transform into a simple Region. The transformation is done by adding
/// BasicBlocks that merge several entry or exit edges so that after the merge
/// just one entry and one exit edge exists.
///
/// The \e Entry of a Region is the first BasicBlock that is passed after
/// entering the Region. It is an element of the Region. The entry BasicBlock
/// dominates all BasicBlocks in the Region.
///
/// The \e Exit of a Region is the first BasicBlock that is passed after
/// leaving the Region. It is not an element of the Region. The exit BasicBlock,
/// postdominates all BasicBlocks in the Region.
///
/// A <em> canonical Region </em> cannot be constructed by combining smaller
/// Regions.
///
/// Region A is the \e parent of Region B, if B is completely contained in A.
///
/// Two canonical Regions either do not intersect at all or one is
/// the parent of the other.
///
/// The <em> Program Structure Tree</em> is a graph (V, E) where V is the set of
/// Regions in the control flow graph and E is the \e parent relation of these
/// Regions.
///
/// Example:
///
/// \verbatim
/// A simple control flow graph, that contains two regions.
///
///        1
///       / |
///      2   |
///     / \   3
///    4   5  |
///    |   |  |
///    6   7  8
///     \  | /
///      \ |/       Region A: 1 -> 9 {1,2,3,4,5,6,7,8}
///        9        Region B: 2 -> 9 {2,4,5,6,7}
/// \endverbatim
///
/// You can obtain more examples by either calling
///
/// <tt> "opt -regions -analyze anyprogram.ll" </tt>
/// or
/// <tt> "opt -view-regions-only anyprogram.ll" </tt>
///
/// on any LLVM file you are interested in.
///
/// The first call returns a textual representation of the program structure
/// tree, the second one creates a graphical representation using graphviz.
class Region : public RegionNode {
  friend class RegionInfo;
  Region(const Region &) LLVM_DELETED_FUNCTION;
  const Region &operator=(const Region &) LLVM_DELETED_FUNCTION;

  // Information necessary to manage this Region.
  RegionInfo* RI;
  DominatorTree *DT;

  // The exit BasicBlock of this region.
  // (The entry BasicBlock is part of RegionNode)
  BasicBlock *exit;

  typedef std::vector<Region*> RegionSet;

  // The subregions of this region.
  RegionSet children;

  typedef std::map<BasicBlock*, RegionNode*> BBNodeMapT;

  // Save the BasicBlock RegionNodes that are element of this Region.
  mutable BBNodeMapT BBNodeMap;

  /// verifyBBInRegion - Check if a BB is in this Region. This check also works
  /// if the region is incorrectly built. (EXPENSIVE!)
  void verifyBBInRegion(BasicBlock* BB) const;

  /// verifyWalk - Walk over all the BBs of the region starting from BB and
  /// verify that all reachable basic blocks are elements of the region.
  /// (EXPENSIVE!)
  void verifyWalk(BasicBlock* BB, std::set<BasicBlock*>* visitedBB) const;

  /// verifyRegionNest - Verify if the region and its children are valid
  /// regions (EXPENSIVE!)
  void verifyRegionNest() const;

public:
  /// @brief Create a new region.
  ///
  /// @param Entry  The entry basic block of the region.
  /// @param Exit   The exit basic block of the region.
  /// @param RI     The region info object that is managing this region.
  /// @param DT     The dominator tree of the current function.
  /// @param Parent The surrounding region or NULL if this is a top level
  ///               region.
  Region(BasicBlock *Entry, BasicBlock *Exit, RegionInfo* RI,
         DominatorTree *DT, Region *Parent = 0);

  /// Delete the Region and all its subregions.
  ~Region();

  /// @brief Get the entry BasicBlock of the Region.
  /// @return The entry BasicBlock of the region.
  BasicBlock *getEntry() const { return RegionNode::getEntry(); }

  /// @brief Replace the entry basic block of the region with the new basic
  ///        block.
  ///
  /// @param BB  The new entry basic block of the region.
  void replaceEntry(BasicBlock *BB);

  /// @brief Replace the exit basic block of the region with the new basic
  ///        block.
  ///
  /// @param BB  The new exit basic block of the region.
  void replaceExit(BasicBlock *BB);

  /// @brief Recursively replace the entry basic block of the region.
  ///
  /// This function replaces the entry basic block with a new basic block. It
  /// also updates all child regions that have the same entry basic block as
  /// this region.
  ///
  /// @param NewEntry The new entry basic block.
  void replaceEntryRecursive(BasicBlock *NewEntry);

  /// @brief Recursively replace the exit basic block of the region.
  ///
  /// This function replaces the exit basic block with a new basic block. It
  /// also updates all child regions that have the same exit basic block as
  /// this region.
  ///
  /// @param NewExit The new exit basic block.
  void replaceExitRecursive(BasicBlock *NewExit);

  /// @brief Get the exit BasicBlock of the Region.
  /// @return The exit BasicBlock of the Region, NULL if this is the TopLevel
  ///         Region.
  BasicBlock *getExit() const { return exit; }

  /// @brief Get the parent of the Region.
  /// @return The parent of the Region or NULL if this is a top level
  ///         Region.
  Region *getParent()