【GCC编译器】计算支配树信息 Part2 - Lengauer - Tarjan 算法求直接必经节点

写在最前,原理参考: https://www.cnblogs.com/meowww/p/6475952.html#3706558

首先列出了GCC 实现Lengauer - Tarjan 算法使用到的数据结构dom_info:

/* This class holds various arrays reflecting the (sub)structure of the
   flowgraph.  Most of them are of type TBB and are also indexed by TBB.  */

class dom_info
{
public:
  dom_info (function *, cdi_direction);
  dom_info (vec <basic_block>, cdi_direction);
  ~dom_info ();
  void calc_dfs_tree ();
  void calc_idoms ();

  inline basic_block get_idom (basic_block);
private:
  void calc_dfs_tree_nonrec (basic_block);
  void compress (TBB);
  void dom_init (void);
  TBB eval (TBB);
  void link_roots (TBB, TBB);

  // m_dfs_parent[v]是结点v在深度为主生成树中的父节点
  TBB *m_dfs_parent;
  // m_key[v]是结点v的半必经结点的深度为主编号
  TBB *m_key;
  // m_path_min[v]是v的祖先链中其半必经结点的深度为主编号最小的一个结点
  TBB *m_path_min;
  // m_bucket[v]是其半必经结点是m_dfs_order[v]的集合中的第一个结点
  TBB *m_bucket;
  // m_next_bucket是m_bucket的辅助结构,用数组代替链表保存集合,加快查询速度
  TBB *m_next_bucket;
  // m_dom[v]是结点v的半必经结点的深度为主编号
  TBB *m_dom;

  /* The following few fields implement the structures needed for disjoint
     sets.  */
  // 如果v在树林中,m_set_chain[v] 是v的祖先的深度为主编号
  // 如果v是树林中一棵树的根,m_set_chain[v] 为 0
  TBB *m_set_chain;
  // m_set_child 和 m_set_size 是用来使树林中的树保持平衡,并使算法时间达到时间下界的两个数据结构
  unsigned int *m_set_size;
  TBB *m_set_child;
  // m_dfs_order[v]是结点v的深度为主编号
  TBB *m_dfs_order;
  TBB *m_dfs_last;
  // m_dfs_to_bb[i]是深度为主编号为i的结点
  basic_block *m_dfs_to_bb;

  /* This is the next free DFS number when creating the DFS tree.  */
  unsigned int m_dfsnum;
  /* The number of nodes in the DFS tree (==m_dfsnum-1).  */
  unsigned int m_nodes;

  /* Blocks with bits set here have a fake edge to EXIT.  These are used
     to turn a DFS forest into a proper tree.  */
  bitmap m_fake_exit_edge;

  /* Number of basic blocks in the function being compiled.  */
  size_t m_n_basic_blocks;

  /* True, if we are computing postdominators (rather than dominators).  */
  bool m_reverse;

  /* Start block (the entry block for forward problem, exit block for backward
     problem).  */
  basic_block m_start_block;
  /* Ending block.  */
  basic_block m_end_block;
};

Lengauer - Tarjan 算法流程:

① 计算CFG的深度为主生成树,获得每个结点的深度为主编号,详见 https://www.cnblogs.com/compiler-share/p/15100454.html

② 按深度为主编号由大到小的顺序,访问每一个节点v,对v的所有直接前驱k1做一次eval,求得sdom的最小值k。然后把节点v放到sdom的bucket里,连接v与par的边。

③ 现在par到v的这棵子树中已经处理完了,所以可以对par的bucket里的每个w求一次idom并且清空bucket。如果sdom(eval(w))=sdom(w),则idom(w)=sdom(w)=parent(v);否则可以记下idom(w)=idom(eval(w)),实现时我们可以写成idom(w)=eval(w),留到第4步处理。

④ 按深度为主编号由小到大的顺序,访问每一个v,如果idom(v)=sdom(v),已经是第3步求出的正确的idom了;否则就证明这是第3步留下的待处理点,令idom(v)=idom(idom(v))即可

/* This calculates the immediate dominators (or post-dominators). THIS is our
   working structure and should hold the DFS forest.
   On return the immediate dominator to node V is in m_dom[V].  */

void
dom_info::calc_idoms ()
{
  /* Go backwards in DFS order, to first look at the leafs.  */
  // Step2
  // 按深度为主编号由大到小的顺序,访问每一个节点v,对v的所有直接前驱k1做一次eval,求得sdom的最小值k。
  // 然后把节点v放到k的bucket里,连接v与par的边。
  for (TBB v = m_nodes; v > 1; v--)
    {
      basic_block bb = m_dfs_to_bb[v];
      edge e;

      TBB par = m_dfs_parent[v]; // 深度为主生成树中的父节点
      TBB k = v;

      edge_iterator ei = m_reverse ? ei_start (bb->succs)
				   : ei_start (bb->preds);
      edge_iterator einext;

      if (m_fake_exit_edge)
	{
	  /* If this block has a fake edge to exit, process that first.  */
	  if (bitmap_bit_p (m_fake_exit_edge, bb->index))
	    {
	      einext = ei;
	      einext.index = 0;
	      goto do_fake_exit_edge;
	    }
	}

      /* Search all direct predecessors for the smallest node with a path
         to them.  That way we have the smallest node with also a path to
         us only over nodes behind us.  In effect we search for our
         semidominator.  */
      while (!ei_end_p (ei))  // 遍历bb在控制流图中的所有前驱
	{
	  basic_block b;
	  TBB k1;

	  e = ei_edge (ei);
	  b = m_reverse ? e->dest : e->src;
	  einext = ei;
	  ei_next (&einext);

	  if (b == m_start_block)
	    {
	    do_fake_exit_edge:
	      k1 = *m_dfs_last;
	    }
	  else
	    k1 = m_dfs_order[b->index];

	  /* Call eval() only if really needed.  If k1 is above V in DFS tree,
	     then we know, that eval(k1) == k1 and key[k1] == k1.  */
	  if (k1 > v) // 如果e是cross edge或者back edge
	    k1 = m_key[eval (k1)]; // 找出已访问节点中,k1的祖先链中半必经结点的深度为主编号最小的那个节点
	  if (k1 < k) // sdom 取深度为主编号最小的那个
	    k = k1; 

	  ei = einext;
	}

      m_key[v] = k;
      link_roots (par, v); // 链接v和par结点
      m_next_bucket[v] = m_bucket[k]; // 结点v仅被处理一次,m_next_bucket[v]具有唯一性,故无需clean
      m_bucket[k] = v;

      /* Transform semidominators into dominators.  */
      // Step3
      // 现在par到v的这棵子树中已经处理完了,所以可以对par的bucket里的每个w求一次idom并且清空bucket。
      // 如果sdom(eval(w))=sdom(w),则idom(w)=sdom(w)=parent(v)
      // 否则可以记下idom(w)=idom(eval(w)),实现时我们可以写成idom(w)=eval(w),留到第4步处理。
      for (TBB w = m_bucket[par]; w; w = m_next_bucket[w])
	{
	  k = eval (w);
	  if (m_key[k] < m_key[w])
	    m_dom[w] = k;
	  else
	    m_dom[w] = par;
	}
      /* We don't need to cleanup next_bucket[].  */
      m_bucket[par] = 0;
    }

  /* Explicitly define the dominators.  */
  // Step4
  // 按深度为主编号由小到大的顺序,访问每一个v
  // 如果idom(v)=sdom(v)的话,就已经是第3步求出的正确的idom了
  // 否则就证明这是第3步留下的待处理点,令idom(v)=idom(idom(v))
  m_dom[1] = 0;
  for (TBB v = 2; v <= m_nodes; v++)
    if (m_dom[v] != m_key[v])
      m_dom[v] = m_dom[m_dom[v]];
}

link_roots 和 eval 管理一个辅助的数据结构,即由深度为主生成树构成的树林,用于追踪已处理过的结点。eval 使用compress对从数组m_set_chain导出的结点到深度为主生成树的树根的路径执行路径压缩。这里有4个数据结构:

① 如果v在树林中,m_set_chain[v] 是v的祖先的深度为主编号;如果v是树林中一棵树的根,m_set_chain[v] 为 0

② m_path_min[v] 是v的祖先链中其半必经结点的深度为主编号最小的一个结点

③ m_set_child 和 m_set_size 是用来使树林中的树保持平衡,并使算法时间达到时间下界的两个数据结构

link_roots 实现:

核心逻辑只有一行代码 m_set_chain[s] = v,设置v为s的祖先节点,把s和父节点v连接起来。其他算法是为了降低算法复杂度,这里不做讨论。

/* This essentially merges the two sets of V and W, giving a single set with
   the new root V.  The internal representation of these disjoint sets is a
   balanced tree.  Currently link(V,W) is only used with V being the parent
   of W.  */

void
dom_info::link_roots (TBB v, TBB w)
{
  TBB s = w;

  /* Rebalance the tree.  */
  while (m_key[m_path_min[w]] < m_key[m_path_min[m_set_child[s]]])
    {
      if (m_set_size[s] + m_set_size[m_set_child[m_set_child[s]]]
	  >= 2 * m_set_size[m_set_child[s]])
	{
	  m_set_chain[m_set_child[s]] = s;
	  m_set_child[s] = m_set_child[m_set_child[s]];
	}
      else
	{
	  m_set_size[m_set_child[s]] = m_set_size[s];
	  s = m_set_chain[s] = m_set_child[s];
	}
    }

  m_path_min[s] = m_path_min[w];
  m_set_size[v] += m_set_size[w];
  if (m_set_size[v] < 2 * m_set_size[w])
    std::swap (m_set_child[v], s);

  /* Merge all subtrees.  */
  while (s)
    {
      // 核心代码
      m_set_chain[s] = v;
      s = m_set_child[s];
    }
}

eval 和 compress 的实现:

递归查询已处理过的结点中v的祖先结点,把m_set_chain[v]设置为root结点,把m_path_min[v]设置为祖先链中半必经结点的深度为主编号最小的那个结点。

/* Compress the path from V to the root of its set and update path_min at the
   same time.  After compress(di, V) set_chain[V] is the root of the set V is
   in and path_min[V] is the node with the smallest key[] value on the path
   from V to that root.  */

void
dom_info::compress (TBB v)
{
  /* Btw. It's not worth to unrecurse compress() as the depth is usually not
     greater than 5 even for huge graphs (I've not seen call depth > 4).
     Also performance wise compress() ranges _far_ behind eval().  */
  TBB parent = m_set_chain[v];
  if (m_set_chain[parent])
    {
      compress (parent);
      if (m_key[m_path_min[parent]] < m_key[m_path_min[v]])
	m_path_min[v] = m_path_min[parent];
      m_set_chain[v] = m_set_chain[parent];
    }
}

/* Compress the path from V to the set root of V if needed (when the root has
   changed since the last call).  Returns the node with the smallest key[]
   value on the path from V to the root.  */

inline TBB
dom_info::eval (TBB v)
{
  /* The representative of the set V is in, also called root (as the set
     representation is a tree).  */
  TBB rep = m_set_chain[v];

  /* V itself is the root.  */
  if (!rep)
    return m_path_min[v];

  /* Compress only if necessary.  */
  if (m_set_chain[rep])
    {
      compress (v);
      rep = m_set_chain[v];
    }

  if (m_key[m_path_min[rep]] >= m_key[m_path_min[v]])
    return m_path_min[v];
  else
    return m_path_min[rep];
}

 

posted on 2021-08-21 14:09  Save-Reset  阅读(215)  评论(0编辑  收藏  举报

导航