From fd634998f813340768c333cdad638498602856e5 Mon Sep 17 00:00:00 2001 From: Ulya Trofimovich Date: Tue, 21 Apr 2020 21:28:32 +0100 Subject: [PATCH] Rewrite recursion into iteration (Tarjan's SCC algorithm and YYFILL states). This is to avoid stack overflow on large RE (especially on instrumented builds that have larger stack frames, like AddressSanitizer). Stack overflow reported by Agostino Sarubbo. Related to #219 "overflow-1.re test fails on system with small stack". Upstram-Status: Backport: https://github.com/skvadrik/re2c/commit/fd634998f813340768c333cdad638498602856e5 CVE: CVE-2018-21232 Signed-off-by: Davide Gardenal --- diff --git a/src/dfa/fillpoints.cc b/src/dfa/fillpoints.cc --- a/src/dfa/fillpoints.cc (revision e58939b34bb4c37cd990f82dc286f21cb405743e) +++ b/src/dfa/fillpoints.cc (date 1646929180243) @@ -5,151 +5,186 @@ #include "src/dfa/dfa.h" -namespace re2c -{ + +/* + * note [finding strongly connected components of DFA] + * + * A slight modification of Tarjan's algorithm. + * + * The algorithm traverses the DFA in depth-first order. It maintains a stack + * of states that have already been visited but haven't been assigned to an SCC + * yet. For each state the algorithm calculates 'lowlink': index of the highest + * ancestor state reachable in one step from a descendant of this state. + * Lowlink is used to determine when a set of states should be popped off stack + * into a new SCC. + * + * We use lowlink to hold different kinds of information: + * - values in range [0 .. stack size] mean that the state is on stack (a + * link to a state with the smallest index reachable from this one) + * - SCC_UND means that this state has not been visited yet + * - SCC_INF means that this state has already been popped off stack + * + * We use stack size (rather than topological sort index) as a unique index of + * the state on stack. This is safe because the indices of states on stack are + * unique and less than the indices of states that have been popped off stack + * (SCC_INF). + */ + +namespace re2c { + namespace { -static const size_t SCC_INF = std::numeric_limits::max(); -static const size_t SCC_UND = SCC_INF - 1; + static const size_t SCC_INF = std::numeric_limits::max(); + static const size_t SCC_UND = SCC_INF - 1; -static bool loopback(size_t node, size_t narcs, const size_t *arcs) -{ - for (size_t i = 0; i < narcs; ++i) - { - if (arcs[i] == node) - { - return true; - } - } - return false; -} + static bool loopback(size_t state, size_t narcs, const size_t *arcs) + { + for (size_t i = 0; i < narcs; ++i) { + if (arcs[i] == state) return true; + } + return false; + } -/* - * node [finding strongly connected components of DFA] - * - * A slight modification of Tarjan's algorithm. - * - * The algorithm walks graph in deep-first order. It maintains a stack - * of nodes that have already been visited but haven't been assigned to - * SCC yet. For each node the algorithm calculates 'lowlink': index of - * the highest ancestor node reachable in one step from a descendant of - * the node. Lowlink is used to determine when a set of nodes should be - * popped off the stack into a new SCC. - * - * We use lowlink to hold different kinds of information: - * - values in range [0 .. stack size] mean that this node is on stack - * (link to a node with the smallest index reachable from this one) - * - SCC_UND means that this node has not been visited yet - * - SCC_INF means that this node has already been popped off stack - * - * We use stack size (rather than topological sort index) as unique index - * of a node on stack. This is safe because indices of nodes on stack are - * still unique and less than indices of nodes that have been popped off - * stack (SCC_INF). - * - */ -static void scc( - const dfa_t &dfa, - std::stack &stack, - std::vector &lowlink, - std::vector &trivial, - size_t i) -{ - const size_t link = stack.size(); - lowlink[i] = link; - stack.push(i); + struct StackItem { + size_t state; // current state + size_t symbol; // next arc to be visited in this state + size_t link; // Tarjan's "lowlink" + }; + +// Tarjan's algorithm + static void scc(const dfa_t &dfa, std::vector &trivial, + std::vector &stack_dfs) + { + std::vector lowlink(dfa.states.size(), SCC_UND); + std::stack stack; + + StackItem x0 = {0, 0, 0}; + stack_dfs.push_back(x0); + + while (!stack_dfs.empty()) { + const size_t i = stack_dfs.back().state; + size_t c = stack_dfs.back().symbol; + size_t link = stack_dfs.back().link; + stack_dfs.pop_back(); + + const size_t *arcs = dfa.states[i]->arcs; + + if (c == 0) { + // DFS recursive enter + //DASSERT(lowlink[i] == SCC_UND); + link = lowlink[i] = stack.size(); + stack.push(i); + } + else { + // DFS recursive return (from one of successor states) + const size_t j = arcs[c - 1]; + //DASSERT(lowlink[j] != SCC_UND); + lowlink[i] = std::min(lowlink[i], lowlink[j]); + } - const size_t *arcs = dfa.states[i]->arcs; - for (size_t c = 0; c < dfa.nchars; ++c) - { - const size_t j = arcs[c]; - if (j != dfa_t::NIL) - { - if (lowlink[j] == SCC_UND) - { - scc(dfa, stack, lowlink, trivial, j); - } - if (lowlink[j] < lowlink[i]) - { - lowlink[i] = lowlink[j]; - } - } - } + // find the next successor state that hasn't been visited yet + for (; c < dfa.nchars; ++c) { + const size_t j = arcs[c]; + if (j != dfa_t::NIL) { + if (lowlink[j] == SCC_UND) { + break; + } + lowlink[i] = std::min(lowlink[i], lowlink[j]); + } + } - if (lowlink[i] == link) - { - // SCC is non-trivial (has loops) iff it either: - // - consists of multiple nodes (they all must be interconnected) - // - consists of single node which loops back to itself - trivial[i] = i == stack.top() - && !loopback(i, dfa.nchars, arcs); + if (c < dfa.nchars) { + // recurse into the next successor state + StackItem x1 = {i, c + 1, link}; + stack_dfs.push_back(x1); + StackItem x2 = {arcs[c], 0, SCC_UND}; + stack_dfs.push_back(x2); + } + else if (lowlink[i] == link) { + // all successors have been visited + // SCC is non-trivial (has loops) if either: + // - it contains multiple interconnected states + // - it contains a single self-looping state + trivial[i] = i == stack.top() && !loopback(i, dfa.nchars, arcs); - size_t j; - do - { - j = stack.top(); - stack.pop(); - lowlink[j] = SCC_INF; - } - while (j != i); - } -} + for (;;) { + const size_t j = stack.top(); + stack.pop(); + lowlink[j] = SCC_INF; + if (i == j) break; + } + } + } + } -static void calc_fill( - const dfa_t &dfa, - const std::vector &trivial, - std::vector &fill, - size_t i) -{ - if (fill[i] == SCC_UND) - { - fill[i] = 0; - const size_t *arcs = dfa.states[i]->arcs; - for (size_t c = 0; c < dfa.nchars; ++c) - { - const size_t j = arcs[c]; - if (j != dfa_t::NIL) - { - calc_fill(dfa, trivial, fill, j); - size_t max = 1; - if (trivial[j]) - { - max += fill[j]; - } - if (max > fill[i]) - { - fill[i] = max; - } - } - } - } -} - -void fillpoints(const dfa_t &dfa, std::vector &fill) -{ - const size_t size = dfa.states.size(); - - // find DFA states that belong to non-trivial SCC - std::stack stack; - std::vector lowlink(size, SCC_UND); - std::vector trivial(size, false); - scc(dfa, stack, lowlink, trivial, 0); - - // for each DFA state, calculate YYFILL argument: - // maximal path length to the next YYFILL state - fill.resize(size, SCC_UND); - calc_fill(dfa, trivial, fill, 0); + static void calc_fill(const dfa_t &dfa, const std::vector &trivial, + std::vector &stack_dfs, std::vector &fill) + { + const size_t nstates = dfa.states.size(); + fill.resize(nstates, SCC_UND); + + StackItem x0 = {0, 0, SCC_INF}; + stack_dfs.push_back(x0); + + while (!stack_dfs.empty()) { + const size_t i = stack_dfs.back().state; + size_t c = stack_dfs.back().symbol; + stack_dfs.pop_back(); + + const size_t *arcs = dfa.states[i]->arcs; + + if (c == 0) { + // DFS recursive enter + if (fill[i] != SCC_UND) continue; + fill[i] = 0; + } + else { + // DFS recursive return (from one of successor states) + const size_t j = arcs[c - 1]; + //DASSERT(fill[i] != SCC_UND && fill[j] != SCC_UND); + fill[i] = std::max(fill[i], 1 + (trivial[j] ? fill[j] : 0)); + } + + // find the next successor state that hasn't been visited yet + for (; c < dfa.nchars; ++c) { + const size_t j = arcs[c]; + if (j != dfa_t::NIL) break; + } + + if (c < dfa.nchars) { + // recurse into the next successor state + StackItem x1 = {i, c + 1, SCC_INF}; + stack_dfs.push_back(x1); + StackItem x2 = {arcs[c], 0, SCC_INF}; + stack_dfs.push_back(x2); + } + } - // The following states must trigger YYFILL: - // - inital state - // - all states in non-trivial SCCs - // for other states, reset YYFILL argument to zero - for (size_t i = 1; i < size; ++i) - { - if (trivial[i]) - { - fill[i] = 0; - } - } -} + // The following states must trigger YYFILL: + // - inital state + // - all states in non-trivial SCCs + // for other states, reset YYFILL argument to zero + for (size_t i = 1; i < nstates; ++i) { + if (trivial[i]) { + fill[i] = 0; + } + } + } + } // anonymous namespace + + void fillpoints(const dfa_t &dfa, std::vector &fill) + { + const size_t nstates = dfa.states.size(); + std::vector trivial(nstates, false); + std::vector stack_dfs; + stack_dfs.reserve(nstates); + + // find DFA states that belong to non-trivial SCC + scc(dfa, trivial, stack_dfs); + + // for each DFA state, calculate YYFILL argument: + // maximal path length to the next YYFILL state + calc_fill(dfa, trivial, stack_dfs, fill); + } + } // namespace re2c