]> git.saurik.com Git - apple/javascriptcore.git/blobdiff - dfg/DFGDominators.cpp
JavaScriptCore-7601.1.46.3.tar.gz
[apple/javascriptcore.git] / dfg / DFGDominators.cpp
index bdd6a6a5baeb8cf1eec7573ce33148bfaa476fcd..4c67e8b9e70a901337cb3bd67589f1c2310519da 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2011 Apple Inc. All rights reserved.
+ * Copyright (C) 2011, 2014 Apple Inc. All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions
 
 #if ENABLE(DFG_JIT)
 
+#include "DFGBlockMapInlines.h"
+#include "DFGBlockWorklist.h"
 #include "DFGGraph.h"
+#include "DFGNaiveDominators.h"
 #include "JSCInlines.h"
 
 namespace JSC { namespace DFG {
@@ -41,91 +44,429 @@ Dominators::~Dominators()
 {
 }
 
-void Dominators::compute(Graph& graph)
-{
-    // This implements a naive dominator solver.
+namespace {
+
+// This implements Lengauer and Tarjan's "A Fast Algorithm for Finding Dominators in a Flowgraph"
+// (TOPLAS 1979). It uses the "simple" implementation of LINK and EVAL, which yields an O(n log n)
+// solution. The full paper is linked below; this code attempts to closely follow the algorithm as
+// it is presented in the paper; in particular sections 3 and 4 as well as appendix B.
+// https://www.cs.princeton.edu/courses/archive/fall03/cs528/handouts/a%20fast%20algorithm%20for%20finding.pdf
+//
+// This code is very subtle. The Lengauer-Tarjan algorithm is incredibly deep to begin with. The
+// goal of this code is to follow the code in the paper, however our implementation must deviate
+// from the paper when it comes to recursion. The authors had used recursion to implement DFS, and
+// also to implement the "simple" EVAL. We convert both of those into worklist-based solutions.
+// Finally, once the algorithm gives us immediate dominators, we implement dominance tests by
+// walking the dominator tree and computing pre and post numbers. We then use the range inclusion
+// check trick that was first discovered by Paul F. Dietz in 1982 in "Maintaining order in a linked
+// list" (see http://dl.acm.org/citation.cfm?id=802184).
+
+class LengauerTarjan {
+public:
+    LengauerTarjan(Graph& graph)
+        : m_graph(graph)
+        , m_data(graph)
+    {
+        for (BlockIndex blockIndex = m_graph.numBlocks(); blockIndex--;) {
+            BasicBlock* block = m_graph.block(blockIndex);
+            if (!block)
+                continue;
+            m_data[block].label = block;
+        }
+    }
     
-    ASSERT(graph.block(0)->predecessors.isEmpty());
+    void compute()
+    {
+        computeDepthFirstPreNumbering(); // Step 1.
+        computeSemiDominatorsAndImplicitImmediateDominators(); // Steps 2 and 3.
+        computeExplicitImmediateDominators(); // Step 4.
+    }
     
-    unsigned numBlocks = graph.numBlocks();
+    BasicBlock* immediateDominator(BasicBlock* block)
+    {
+        return m_data[block].dom;
+    }
+
+private:
+    void computeDepthFirstPreNumbering()
+    {
+        // Use a block worklist that also tracks the index inside the successor list. This is
+        // necessary for ensuring that we don't attempt to visit a successor until the previous
+        // successors that we had visited are fully processed. This ends up being revealed in the
+        // output of this method because the first time we see an edge to a block, we set the
+        // block's parent. So, if we have:
+        //
+        // A -> B
+        // A -> C
+        // B -> C
+        //
+        // And we're processing A, then we want to ensure that if we see A->B first (and hence set
+        // B's prenumber before we set C's) then we also end up setting C's parent to B by virtue
+        // of not noticing A->C until we're done processing B.
+        
+        ExtendedBlockWorklist<unsigned> worklist;
+        worklist.push(m_graph.block(0), 0);
+        
+        while (BlockWith<unsigned> item = worklist.pop()) {
+            BasicBlock* block = item.block;
+            unsigned successorIndex = item.data;
+            
+            // We initially push with successorIndex = 0 regardless of whether or not we have any
+            // successors. This is so that we can assign our prenumber. Subsequently we get pushed
+            // with higher successorIndex values, but only if they are in range.
+            ASSERT(!successorIndex || successorIndex < block->numSuccessors());
+
+            if (!successorIndex) {
+                m_data[block].semiNumber = m_blockByPreNumber.size();
+                m_blockByPreNumber.append(block);
+            }
+            
+            if (successorIndex < block->numSuccessors()) {
+                unsigned nextSuccessorIndex = successorIndex + 1;
+                if (nextSuccessorIndex < block->numSuccessors())
+                    worklist.forcePush(block, nextSuccessorIndex);
+
+                BasicBlock* successorBlock = block->successor(successorIndex);
+                if (worklist.push(successorBlock, 0))
+                    m_data[successorBlock].parent = block;
+            }
+        }
+    }
     
-    // Allocate storage for the dense dominance matrix. 
-    if (numBlocks > m_results.size()) {
-        m_results.grow(numBlocks);
-        for (unsigned i = numBlocks; i--;)
-            m_results[i].resize(numBlocks);
-        m_scratch.resize(numBlocks);
+    void computeSemiDominatorsAndImplicitImmediateDominators()
+    {
+        for (unsigned currentPreNumber = m_blockByPreNumber.size(); currentPreNumber-- > 1;) {
+            BasicBlock* block = m_blockByPreNumber[currentPreNumber];
+            BlockData& blockData = m_data[block];
+            
+            // Step 2:
+            for (BasicBlock* predecessorBlock : block->predecessors) {
+                BasicBlock* intermediateBlock = eval(predecessorBlock);
+                blockData.semiNumber = std::min(
+                    m_data[intermediateBlock].semiNumber, blockData.semiNumber);
+            }
+            unsigned bucketPreNumber = blockData.semiNumber;
+            ASSERT(bucketPreNumber <= currentPreNumber);
+            m_data[m_blockByPreNumber[bucketPreNumber]].bucket.append(block);
+            link(blockData.parent, block);
+            
+            // Step 3:
+            for (BasicBlock* semiDominee : m_data[blockData.parent].bucket) {
+                BasicBlock* possibleDominator = eval(semiDominee);
+                BlockData& semiDomineeData = m_data[semiDominee];
+                ASSERT(m_blockByPreNumber[semiDomineeData.semiNumber] == blockData.parent);
+                BlockData& possibleDominatorData = m_data[possibleDominator];
+                if (possibleDominatorData.semiNumber < semiDomineeData.semiNumber)
+                    semiDomineeData.dom = possibleDominator;
+                else
+                    semiDomineeData.dom = blockData.parent;
+            }
+            m_data[blockData.parent].bucket.clear();
+        }
+    }
+    
+    void computeExplicitImmediateDominators()
+    {
+        for (unsigned currentPreNumber = 1; currentPreNumber < m_blockByPreNumber.size(); ++currentPreNumber) {
+            BasicBlock* block = m_blockByPreNumber[currentPreNumber];
+            BlockData& blockData = m_data[block];
+            
+            if (blockData.dom != m_blockByPreNumber[blockData.semiNumber])
+                blockData.dom = m_data[blockData.dom].dom;
+        }
+    }
+    
+    void link(BasicBlock* from, BasicBlock* to)
+    {
+        m_data[to].ancestor = from;
+    }
+    
+    BasicBlock* eval(BasicBlock* block)
+    {
+        if (!m_data[block].ancestor)
+            return block;
+        
+        compress(block);
+        return m_data[block].label;
+    }
+    
+    void compress(BasicBlock* initialBlock)
+    {
+        // This was meant to be a recursive function, but we don't like recursion because we don't
+        // want to blow the stack. The original function will call compress() recursively on the
+        // ancestor of anything that has an ancestor. So, we populate our worklist with the
+        // recursive ancestors of initialBlock. Then we process the list starting from the block
+        // that is furthest up the ancestor chain.
+        
+        BasicBlock* ancestor = m_data[initialBlock].ancestor;
+        ASSERT(ancestor);
+        if (!m_data[ancestor].ancestor)
+            return;
+        
+        Vector<BasicBlock*, 16> stack;
+        for (BasicBlock* block = initialBlock; block; block = m_data[block].ancestor)
+            stack.append(block);
+        
+        // We only care about blocks that have an ancestor that has an ancestor. The last two
+        // elements in the stack won't satisfy this property.
+        ASSERT(stack.size() >= 2);
+        ASSERT(!m_data[stack[stack.size() - 1]].ancestor);
+        ASSERT(!m_data[m_data[stack[stack.size() - 2]].ancestor].ancestor);
+        
+        for (unsigned i = stack.size() - 2; i--;) {
+            BasicBlock* block = stack[i];
+            BasicBlock*& labelOfBlock = m_data[block].label;
+            BasicBlock*& ancestorOfBlock = m_data[block].ancestor;
+            ASSERT(ancestorOfBlock);
+            ASSERT(m_data[ancestorOfBlock].ancestor);
+            
+            BasicBlock* labelOfAncestorOfBlock = m_data[ancestorOfBlock].label;
+            
+            if (m_data[labelOfAncestorOfBlock].semiNumber < m_data[labelOfBlock].semiNumber)
+                labelOfBlock = labelOfAncestorOfBlock;
+            ancestorOfBlock = m_data[ancestorOfBlock].ancestor;
+        }
     }
 
-    // We know that the entry block is only dominated by itself.
-    m_results[0].clearAll();
-    m_results[0].set(0);
+    struct BlockData {
+        BlockData()
+            : parent(nullptr)
+            , preNumber(UINT_MAX)
+            , semiNumber(UINT_MAX)
+            , ancestor(nullptr)
+            , label(nullptr)
+            , dom(nullptr)
+        {
+        }
+        
+        BasicBlock* parent;
+        unsigned preNumber;
+        unsigned semiNumber;
+        BasicBlock* ancestor;
+        BasicBlock* label;
+        Vector<BasicBlock*> bucket;
+        BasicBlock* dom;
+    };
+    
+    Graph& m_graph;
+    BlockMap<BlockData> m_data;
+    Vector<BasicBlock*> m_blockByPreNumber;
+};
 
-    // Find all of the valid blocks.
-    m_scratch.clearAll();
-    for (unsigned i = numBlocks; i--;) {
-        if (!graph.block(i))
-            continue;
-        m_scratch.set(i);
+struct ValidationContext {
+    ValidationContext(Graph& graph, Dominators& dominators)
+        : graph(graph)
+        , dominators(dominators)
+    {
     }
     
-    // Mark all nodes as dominated by everything.
-    for (unsigned i = numBlocks; i-- > 1;) {
-        if (!graph.block(i) || graph.block(i)->predecessors.isEmpty())
-            m_results[i].clearAll();
-        else
-            m_results[i].set(m_scratch);
+    void reportError(BasicBlock* from, BasicBlock* to, const char* message)
+    {
+        Error error;
+        error.from = from;
+        error.to = to;
+        error.message = message;
+        errors.append(error);
     }
+    
+    void handleErrors()
+    {
+        if (errors.isEmpty())
+            return;
+        
+        startCrashing();
+        dataLog("DFG DOMINATOR VALIDATION FAILED:\n");
+        dataLog("\n");
+        dataLog("For block domination relationships:\n");
+        for (unsigned i = 0; i < errors.size(); ++i) {
+            dataLog(
+                "    ", pointerDump(errors[i].from), " -> ", pointerDump(errors[i].to),
+                " (", errors[i].message, ")\n");
+        }
+        dataLog("\n");
+        dataLog("Control flow graph:\n");
+        for (BlockIndex blockIndex = 0; blockIndex < graph.numBlocks(); ++blockIndex) {
+            BasicBlock* block = graph.block(blockIndex);
+            if (!block)
+                continue;
+            dataLog("    Block #", blockIndex, ": successors = [");
+            CommaPrinter comma;
+            for (unsigned i = 0; i < block->numSuccessors(); ++i)
+                dataLog(comma, *block->successor(i));
+            dataLog("], predecessors = [");
+            comma = CommaPrinter();
+            for (unsigned i = 0; i < block->predecessors.size(); ++i)
+                dataLog(comma, *block->predecessors[i]);
+            dataLog("]\n");
+        }
+        dataLog("\n");
+        dataLog("Lengauer-Tarjan Dominators:\n");
+        dataLog(dominators);
+        dataLog("\n");
+        dataLog("Naive Dominators:\n");
+        naiveDominators.dump(graph, WTF::dataFile());
+        dataLog("\n");
+        dataLog("Graph at time of failure:\n");
+        graph.dump();
+        dataLog("\n");
+        dataLog("DFG DOMINATOR VALIDATION FAILIED!\n");
+        CRASH();
+    }
+    
+    Graph& graph;
+    Dominators& dominators;
+    NaiveDominators naiveDominators;
+    
+    struct Error {
+        BasicBlock* from;
+        BasicBlock* to;
+        const char* message;
+    };
+    
+    Vector<Error> errors;
+};
+
+} // anonymous namespace
 
-    // Iteratively eliminate nodes that are not dominator.
-    bool changed;
-    do {
-        changed = false;
-        // Prune dominators in all non entry blocks: forward scan.
-        for (unsigned i = 1; i < numBlocks; ++i)
-            changed |= pruneDominators(graph, i);
+void Dominators::compute(Graph& graph)
+{
+    LengauerTarjan lengauerTarjan(graph);
+    lengauerTarjan.compute();
 
-        if (!changed)
+    m_data = BlockMap<BlockData>(graph);
+    
+    // From here we want to build a spanning tree with both upward and downward links and we want
+    // to do a search over this tree to compute pre and post numbers that can be used for dominance
+    // tests.
+    
+    for (BlockIndex blockIndex = graph.numBlocks(); blockIndex--;) {
+        BasicBlock* block = graph.block(blockIndex);
+        if (!block)
+            continue;
+        
+        BasicBlock* idomBlock = lengauerTarjan.immediateDominator(block);
+        m_data[block].idomParent = idomBlock;
+        if (idomBlock)
+            m_data[idomBlock].idomKids.append(block);
+    }
+    
+    unsigned nextPreNumber = 0;
+    unsigned nextPostNumber = 0;
+    
+    // Plain stack-based worklist because we are guaranteed to see each block exactly once anyway.
+    Vector<BlockWithOrder> worklist;
+    worklist.append(BlockWithOrder(graph.block(0), PreOrder));
+    while (!worklist.isEmpty()) {
+        BlockWithOrder item = worklist.takeLast();
+        switch (item.order) {
+        case PreOrder:
+            m_data[item.block].preNumber = nextPreNumber++;
+            worklist.append(BlockWithOrder(item.block, PostOrder));
+            for (BasicBlock* kid : m_data[item.block].idomKids)
+                worklist.append(BlockWithOrder(kid, PreOrder));
             break;
+        case PostOrder:
+            m_data[item.block].postNumber = nextPostNumber++;
+            break;
+        }
+    }
+    
+    if (validationEnabled()) {
+        // Check our dominator calculation:
+        // 1) Check that our range-based ancestry test is the same as a naive ancestry test.
+        // 2) Check that our notion of who dominates whom is identical to a naive (not
+        //    Lengauer-Tarjan) dominator calculation.
+        
+        ValidationContext context(graph, *this);
+        context.naiveDominators.compute(graph);
+        
+        for (BlockIndex fromBlockIndex = graph.numBlocks(); fromBlockIndex--;) {
+            BasicBlock* fromBlock = graph.block(fromBlockIndex);
+            if (!fromBlock || m_data[fromBlock].preNumber == UINT_MAX)
+                continue;
+            for (BlockIndex toBlockIndex = graph.numBlocks(); toBlockIndex--;) {
+                BasicBlock* toBlock = graph.block(toBlockIndex);
+                if (!toBlock || m_data[toBlock].preNumber == UINT_MAX)
+                    continue;
+                
+                if (dominates(fromBlock, toBlock) != naiveDominates(fromBlock, toBlock))
+                    context.reportError(fromBlock, toBlock, "Range-based domination check is broken");
+                if (dominates(fromBlock, toBlock) != context.naiveDominators.dominates(fromBlock, toBlock))
+                    context.reportError(fromBlock, toBlock, "Lengauer-Tarjan domination is broken");
+            }
+        }
+        
+        context.handleErrors();
+    }
+}
 
-        // Prune dominators in all non entry blocks: backward scan.
-        changed = false;
-        for (unsigned i = numBlocks; i-- > 1;)
-            changed |= pruneDominators(graph, i);
-    } while (changed);
+BlockSet Dominators::strictDominatorsOf(BasicBlock* to) const
+{
+    BlockSet result;
+    forAllStrictDominatorsOf(to, BlockAdder(result));
+    return result;
 }
 
-bool Dominators::pruneDominators(Graph& graph, BlockIndex idx)
+BlockSet Dominators::dominatorsOf(BasicBlock* to) const
 {
-    BasicBlock* block = graph.block(idx);
+    BlockSet result;
+    forAllDominatorsOf(to, BlockAdder(result));
+    return result;
+}
 
-    if (!block || block->predecessors.isEmpty())
-        return false;
+BlockSet Dominators::blocksStrictlyDominatedBy(BasicBlock* from) const
+{
+    BlockSet result;
+    forAllBlocksStrictlyDominatedBy(from, BlockAdder(result));
+    return result;
+}
 
-    // Find the intersection of dom(preds).
-    m_scratch.set(m_results[block->predecessors[0]->index]);
-    for (unsigned j = block->predecessors.size(); j-- > 1;)
-        m_scratch.filter(m_results[block->predecessors[j]->index]);
+BlockSet Dominators::blocksDominatedBy(BasicBlock* from) const
+{
+    BlockSet result;
+    forAllBlocksDominatedBy(from, BlockAdder(result));
+    return result;
+}
 
-    // The block is also dominated by itself.
-    m_scratch.set(idx);
+BlockSet Dominators::dominanceFrontierOf(BasicBlock* from) const
+{
+    BlockSet result;
+    forAllBlocksInDominanceFrontierOfImpl(from, BlockAdder(result));
+    return result;
+}
 
-    return m_results[idx].setAndCheck(m_scratch);
+BlockSet Dominators::iteratedDominanceFrontierOf(const BlockList& from) const
+{
+    BlockSet result;
+    forAllBlocksInIteratedDominanceFrontierOfImpl(from, BlockAdder(result));
+    return result;
 }
 
-void Dominators::dump(Graph& graph, PrintStream& out) const
+bool Dominators::naiveDominates(BasicBlock* from, BasicBlock* to) const
 {
-    for (BlockIndex blockIndex = 0; blockIndex < graph.numBlocks(); ++blockIndex) {
-        BasicBlock* block = graph.block(blockIndex);
-        if (!block)
+    for (BasicBlock* block = to; block; block = m_data[block].idomParent) {
+        if (block == from)
+            return true;
+    }
+    return false;
+}
+
+void Dominators::dump(PrintStream& out) const
+{
+    if (!isValid()) {
+        out.print("    Not Valid.\n");
+        return;
+    }
+    
+    for (BlockIndex blockIndex = 0; blockIndex < m_data.size(); ++blockIndex) {
+        if (m_data[blockIndex].preNumber == UINT_MAX)
             continue;
-        out.print("    Block ", *block, ":");
-        for (BlockIndex otherIndex = 0; otherIndex < graph.numBlocks(); ++otherIndex) {
-            if (!dominates(block->index, otherIndex))
-                continue;
-            out.print(" #", otherIndex);
-        }
-        out.print("\n");
+        
+        out.print("    Block #", blockIndex, ": idom = ", pointerDump(m_data[blockIndex].idomParent), ", idomKids = [");
+        CommaPrinter comma;
+        for (unsigned i = 0; i < m_data[blockIndex].idomKids.size(); ++i)
+            out.print(comma, *m_data[blockIndex].idomKids[i]);
+        out.print("], pre/post = ", m_data[blockIndex].preNumber, "/", m_data[blockIndex].postNumber, "\n");
     }
 }