1 // Copyright 2012-2014 The Rust Project Developers. See the COPYRIGHT
2 // file at the top-level directory of this distribution and at
3 // http://rust-lang.org/COPYRIGHT.
5 // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
6 // http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
7 // <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
8 // option. This file may not be copied, modified, or distributed
9 // except according to those terms.
12 //! A module for propagating forward dataflow information. The analysis
13 //! assumes that the items to be propagated can be represented as bits
14 //! and thus uses bitvectors. Your job is simply to specify the so-called
15 //! GEN and KILL bits for each expression.
18 use middle::cfg::CFGIndex;
24 use syntax::ast_util::IdRange;
25 use syntax::print::pp;
26 use syntax::print::pprust::PrintState;
27 use util::nodemap::NodeMap;
29 use rustc_front::intravisit;
30 use rustc_front::print::pprust;
33 #[derive(Copy, Clone, Debug)]
34 pub enum EntryOrExit {
40 pub struct DataFlowContext<'a, 'tcx: 'a, O> {
41 tcx: &'a ty::ctxt<'tcx>,
43 /// a name for the analysis using this dataflow instance
44 analysis_name: &'static str,
46 /// the data flow operator
49 /// number of bits to propagate per id
52 /// number of words we will use to store bits_per_id.
53 /// equal to bits_per_id/usize::BITS rounded up.
56 // mapping from node to cfg node index
57 // FIXME (#6298): Shouldn't this go with CFG?
58 nodeid_to_index: NodeMap<Vec<CFGIndex>>,
60 // Bit sets per cfg node. The following three fields (`gens`, `kills`,
61 // and `on_entry`) all have the same structure. For each id in
62 // `id_range`, there is a range of words equal to `words_per_id`.
63 // So, to access the bits for any given id, you take a slice of
64 // the full vector (see the method `compute_id_range()`).
66 /// bits generated as we exit the cfg node. Updated by `add_gen()`.
69 /// bits killed as we exit the cfg node, or non-locally jump over
70 /// it. Updated by `add_kill(KillFrom::ScopeEnd)`.
71 scope_kills: Vec<usize>,
73 /// bits killed as we exit the cfg node directly; if it is jumped
74 /// over, e.g. via `break`, the kills are not reflected in the
75 /// jump's effects. Updated by `add_kill(KillFrom::Execution)`.
76 action_kills: Vec<usize>,
78 /// bits that are valid on entry to the cfg node. Updated by
83 pub trait BitwiseOperator {
84 /// Joins two predecessor bits together, typically either `|` or `&`
85 fn join(&self, succ: usize, pred: usize) -> usize;
88 /// Parameterization for the precise form of data flow that is used.
89 pub trait DataFlowOperator : BitwiseOperator {
90 /// Specifies the initial value for each bit in the `on_entry` set
91 fn initial_value(&self) -> bool;
94 struct PropagationContext<'a, 'b: 'a, 'tcx: 'b, O: 'a> {
95 dfcx: &'a mut DataFlowContext<'b, 'tcx, O>,
99 fn get_cfg_indices<'a>(id: ast::NodeId, index: &'a NodeMap<Vec<CFGIndex>>) -> &'a [CFGIndex] {
100 let opt_indices = index.get(&id);
101 opt_indices.map(|v| &v[..]).unwrap_or(&[])
104 impl<'a, 'tcx, O:DataFlowOperator> DataFlowContext<'a, 'tcx, O> {
105 fn has_bitset_for_nodeid(&self, n: ast::NodeId) -> bool {
106 assert!(n != ast::DUMMY_NODE_ID);
107 self.nodeid_to_index.contains_key(&n)
111 impl<'a, 'tcx, O:DataFlowOperator> pprust::PpAnn for DataFlowContext<'a, 'tcx, O> {
113 ps: &mut pprust::State,
114 node: pprust::AnnNode) -> io::Result<()> {
115 let id = match node {
116 pprust::NodeName(_) => 0,
117 pprust::NodeExpr(expr) => expr.id,
118 pprust::NodeBlock(blk) => blk.id,
119 pprust::NodeItem(_) | pprust::NodeSubItem(_) => 0,
120 pprust::NodePat(pat) => pat.id
123 if !self.has_bitset_for_nodeid(id) {
127 assert!(self.bits_per_id > 0);
128 let indices = get_cfg_indices(id, &self.nodeid_to_index);
129 for &cfgidx in indices {
130 let (start, end) = self.compute_id_range(cfgidx);
131 let on_entry = &self.on_entry[start.. end];
132 let entry_str = bits_to_string(on_entry);
134 let gens = &self.gens[start.. end];
135 let gens_str = if gens.iter().any(|&u| u != 0) {
136 format!(" gen: {}", bits_to_string(gens))
141 let action_kills = &self.action_kills[start .. end];
142 let action_kills_str = if action_kills.iter().any(|&u| u != 0) {
143 format!(" action_kill: {}", bits_to_string(action_kills))
148 let scope_kills = &self.scope_kills[start .. end];
149 let scope_kills_str = if scope_kills.iter().any(|&u| u != 0) {
150 format!(" scope_kill: {}", bits_to_string(scope_kills))
155 try!(ps.synth_comment(
156 format!("id {}: {}{}{}{}", id, entry_str,
157 gens_str, action_kills_str, scope_kills_str)));
158 try!(pp::space(&mut ps.s));
164 fn build_nodeid_to_index(decl: Option<&hir::FnDecl>,
165 cfg: &cfg::CFG) -> NodeMap<Vec<CFGIndex>> {
166 let mut index = NodeMap();
168 // FIXME (#6298): Would it be better to fold formals from decl
169 // into cfg itself? i.e. introduce a fn-based flow-graph in
170 // addition to the current block-based flow-graph, rather than
171 // have to put traversals like this here?
174 Some(decl) => add_entries_from_fn_decl(&mut index, decl, cfg.entry)
177 cfg.graph.each_node(|node_idx, node| {
178 if let cfg::CFGNodeData::AST(id) = node.data {
179 index.entry(id).or_insert(vec![]).push(node_idx);
186 fn add_entries_from_fn_decl(index: &mut NodeMap<Vec<CFGIndex>>,
189 //! add mappings from the ast nodes for the formal bindings to
190 //! the entry-node in the graph.
193 index: &'a mut NodeMap<Vec<CFGIndex>>,
195 let mut formals = Formals { entry: entry, index: index };
196 intravisit::walk_fn_decl(&mut formals, decl);
197 impl<'a, 'v> intravisit::Visitor<'v> for Formals<'a> {
198 fn visit_pat(&mut self, p: &hir::Pat) {
199 self.index.entry(p.id).or_insert(vec![]).push(self.entry);
200 intravisit::walk_pat(self, p)
206 /// Flag used by `add_kill` to indicate whether the provided kill
207 /// takes effect only when control flows directly through the node in
208 /// question, or if the kill's effect is associated with any
209 /// control-flow directly through or indirectly over the node.
210 #[derive(Copy, Clone, PartialEq, Debug)]
212 /// A `ScopeEnd` kill is one that takes effect when any control
213 /// flow goes over the node. A kill associated with the end of the
214 /// scope of a variable declaration `let x;` is an example of a
218 /// An `Execution` kill is one that takes effect only when control
219 /// flow goes through the node to completion. A kill associated
220 /// with an assignment statement `x = expr;` is an example of an
221 /// `Execution` kill.
225 impl<'a, 'tcx, O:DataFlowOperator> DataFlowContext<'a, 'tcx, O> {
226 pub fn new(tcx: &'a ty::ctxt<'tcx>,
227 analysis_name: &'static str,
228 decl: Option<&hir::FnDecl>,
232 bits_per_id: usize) -> DataFlowContext<'a, 'tcx, O> {
233 let usize_bits = mem::size_of::<usize>() * 8;
234 let words_per_id = (bits_per_id + usize_bits - 1) / usize_bits;
235 let num_nodes = cfg.graph.all_nodes().len();
237 debug!("DataFlowContext::new(analysis_name: {}, id_range={:?}, \
238 bits_per_id={}, words_per_id={}) \
240 analysis_name, id_range, bits_per_id, words_per_id,
243 let entry = if oper.initial_value() { usize::MAX } else {0};
245 let zeroes = vec![0; num_nodes * words_per_id];
246 let gens = zeroes.clone();
247 let kills1 = zeroes.clone();
249 let on_entry = vec![entry; num_nodes * words_per_id];
251 let nodeid_to_index = build_nodeid_to_index(decl, cfg);
255 analysis_name: analysis_name,
256 words_per_id: words_per_id,
257 nodeid_to_index: nodeid_to_index,
258 bits_per_id: bits_per_id,
261 action_kills: kills1,
267 pub fn add_gen(&mut self, id: ast::NodeId, bit: usize) {
268 //! Indicates that `id` generates `bit`
269 debug!("{} add_gen(id={}, bit={})",
270 self.analysis_name, id, bit);
271 assert!(self.nodeid_to_index.contains_key(&id));
272 assert!(self.bits_per_id > 0);
274 let indices = get_cfg_indices(id, &self.nodeid_to_index);
275 for &cfgidx in indices {
276 let (start, end) = self.compute_id_range(cfgidx);
277 let gens = &mut self.gens[start.. end];
282 pub fn add_kill(&mut self, kind: KillFrom, id: ast::NodeId, bit: usize) {
283 //! Indicates that `id` kills `bit`
284 debug!("{} add_kill(id={}, bit={})",
285 self.analysis_name, id, bit);
286 assert!(self.nodeid_to_index.contains_key(&id));
287 assert!(self.bits_per_id > 0);
289 let indices = get_cfg_indices(id, &self.nodeid_to_index);
290 for &cfgidx in indices {
291 let (start, end) = self.compute_id_range(cfgidx);
292 let kills = match kind {
293 KillFrom::Execution => &mut self.action_kills[start.. end],
294 KillFrom::ScopeEnd => &mut self.scope_kills[start.. end],
300 fn apply_gen_kill(&self, cfgidx: CFGIndex, bits: &mut [usize]) {
301 //! Applies the gen and kill sets for `cfgidx` to `bits`
302 debug!("{} apply_gen_kill(cfgidx={:?}, bits={}) [before]",
303 self.analysis_name, cfgidx, mut_bits_to_string(bits));
304 assert!(self.bits_per_id > 0);
306 let (start, end) = self.compute_id_range(cfgidx);
307 let gens = &self.gens[start.. end];
308 bitwise(bits, gens, &Union);
309 let kills = &self.action_kills[start.. end];
310 bitwise(bits, kills, &Subtract);
311 let kills = &self.scope_kills[start.. end];
312 bitwise(bits, kills, &Subtract);
314 debug!("{} apply_gen_kill(cfgidx={:?}, bits={}) [after]",
315 self.analysis_name, cfgidx, mut_bits_to_string(bits));
318 fn compute_id_range(&self, cfgidx: CFGIndex) -> (usize, usize) {
319 let n = cfgidx.node_id();
320 let start = n * self.words_per_id;
321 let end = start + self.words_per_id;
323 assert!(start < self.gens.len());
324 assert!(end <= self.gens.len());
325 assert!(self.gens.len() == self.action_kills.len());
326 assert!(self.gens.len() == self.scope_kills.len());
327 assert!(self.gens.len() == self.on_entry.len());
333 pub fn each_bit_on_entry<F>(&self, id: ast::NodeId, mut f: F) -> bool where
334 F: FnMut(usize) -> bool,
336 //! Iterates through each bit that is set on entry to `id`.
337 //! Only useful after `propagate()` has been called.
338 if !self.has_bitset_for_nodeid(id) {
341 let indices = get_cfg_indices(id, &self.nodeid_to_index);
342 for &cfgidx in indices {
343 if !self.each_bit_for_node(EntryOrExit::Entry, cfgidx, |i| f(i)) {
350 pub fn each_bit_for_node<F>(&self, e: EntryOrExit, cfgidx: CFGIndex, f: F) -> bool where
351 F: FnMut(usize) -> bool,
353 //! Iterates through each bit that is set on entry/exit to `cfgidx`.
354 //! Only useful after `propagate()` has been called.
356 if self.bits_per_id == 0 {
357 // Skip the surprisingly common degenerate case. (Note
358 // compute_id_range requires self.words_per_id > 0.)
362 let (start, end) = self.compute_id_range(cfgidx);
363 let on_entry = &self.on_entry[start.. end];
365 let slice = match e {
366 EntryOrExit::Entry => on_entry,
367 EntryOrExit::Exit => {
368 let mut t = on_entry.to_vec();
369 self.apply_gen_kill(cfgidx, &mut t);
374 debug!("{} each_bit_for_node({:?}, cfgidx={:?}) bits={}",
375 self.analysis_name, e, cfgidx, bits_to_string(slice));
376 self.each_bit(slice, f)
379 pub fn each_gen_bit<F>(&self, id: ast::NodeId, mut f: F) -> bool where
380 F: FnMut(usize) -> bool,
382 //! Iterates through each bit in the gen set for `id`.
383 if !self.has_bitset_for_nodeid(id) {
387 if self.bits_per_id == 0 {
388 // Skip the surprisingly common degenerate case. (Note
389 // compute_id_range requires self.words_per_id > 0.)
393 let indices = get_cfg_indices(id, &self.nodeid_to_index);
394 for &cfgidx in indices {
395 let (start, end) = self.compute_id_range(cfgidx);
396 let gens = &self.gens[start.. end];
397 debug!("{} each_gen_bit(id={}, gens={})",
398 self.analysis_name, id, bits_to_string(gens));
399 if !self.each_bit(gens, |i| f(i)) {
406 fn each_bit<F>(&self, words: &[usize], mut f: F) -> bool where
407 F: FnMut(usize) -> bool,
409 //! Helper for iterating over the bits in a bit set.
410 //! Returns false on the first call to `f` that returns false;
411 //! if all calls to `f` return true, then returns true.
413 let usize_bits = mem::size_of::<usize>() * 8;
414 for (word_index, &word) in words.iter().enumerate() {
416 let base_index = word_index * usize_bits;
417 for offset in 0..usize_bits {
418 let bit = 1 << offset;
419 if (word & bit) != 0 {
420 // NB: we round up the total number of bits
421 // that we store in any given bit set so that
422 // it is an even multiple of usize::BITS. This
423 // means that there may be some stray bits at
424 // the end that do not correspond to any
425 // actual value. So before we callback, check
426 // whether the bit_index is greater than the
427 // actual value the user specified and stop
429 let bit_index = base_index + offset as usize;
430 if bit_index >= self.bits_per_id {
432 } else if !f(bit_index) {
442 pub fn add_kills_from_flow_exits(&mut self, cfg: &cfg::CFG) {
443 //! Whenever you have a `break` or `continue` statement, flow
444 //! exits through any number of enclosing scopes on its way to
445 //! the new destination. This function infers the kill bits of
446 //! those control operators based on the kill bits associated
447 //! with those scopes.
449 //! This is usually called (if it is called at all), after
450 //! all add_gen and add_kill calls, but before propagate.
452 debug!("{} add_kills_from_flow_exits", self.analysis_name);
453 if self.bits_per_id == 0 {
454 // Skip the surprisingly common degenerate case. (Note
455 // compute_id_range requires self.words_per_id > 0.)
458 cfg.graph.each_edge(|_edge_index, edge| {
459 let flow_exit = edge.source();
460 let (start, end) = self.compute_id_range(flow_exit);
461 let mut orig_kills = self.scope_kills[start.. end].to_vec();
463 let mut changed = false;
464 for &node_id in &edge.data.exiting_scopes {
465 let opt_cfg_idx = self.nodeid_to_index.get(&node_id);
468 for &cfg_idx in indices {
469 let (start, end) = self.compute_id_range(cfg_idx);
470 let kills = &self.scope_kills[start.. end];
471 if bitwise(&mut orig_kills, kills, &Union) {
472 debug!("scope exits: scope id={} \
473 (node={:?} of {:?}) added killset: {}",
474 node_id, cfg_idx, indices,
475 bits_to_string(kills));
481 debug!("{} add_kills_from_flow_exits flow_exit={:?} \
482 no cfg_idx for exiting_scope={}",
483 self.analysis_name, flow_exit, node_id);
489 let bits = &mut self.scope_kills[start.. end];
490 debug!("{} add_kills_from_flow_exits flow_exit={:?} bits={} [before]",
491 self.analysis_name, flow_exit, mut_bits_to_string(bits));
492 bits.clone_from_slice(&orig_kills[..]);
493 debug!("{} add_kills_from_flow_exits flow_exit={:?} bits={} [after]",
494 self.analysis_name, flow_exit, mut_bits_to_string(bits));
501 impl<'a, 'tcx, O:DataFlowOperator+Clone+'static> DataFlowContext<'a, 'tcx, O> {
502 // ^^^^^^^^^^^^^ only needed for pretty printing
503 pub fn propagate(&mut self, cfg: &cfg::CFG, blk: &hir::Block) {
504 //! Performs the data flow analysis.
506 if self.bits_per_id == 0 {
507 // Optimize the surprisingly common degenerate case.
512 let words_per_id = self.words_per_id;
513 let mut propcx = PropagationContext {
518 let mut temp = vec![0; words_per_id];
519 while propcx.changed {
520 propcx.changed = false;
521 propcx.reset(&mut temp);
522 propcx.walk_cfg(cfg, &mut temp);
526 debug!("Dataflow result for {}:", self.analysis_name);
528 let mut v = Vec::new();
529 self.pretty_print_to(box &mut v, blk).unwrap();
530 String::from_utf8(v).unwrap()
534 fn pretty_print_to<'b>(&self, wr: Box<io::Write + 'b>,
535 blk: &hir::Block) -> io::Result<()> {
536 let mut ps = pprust::rust_printer_annotated(wr, self, None);
537 try!(ps.cbox(pprust::indent_unit));
539 try!(ps.print_block(blk));
544 impl<'a, 'b, 'tcx, O:DataFlowOperator> PropagationContext<'a, 'b, 'tcx, O> {
545 fn walk_cfg(&mut self,
547 in_out: &mut [usize]) {
548 debug!("DataFlowContext::walk_cfg(in_out={}) {}",
549 bits_to_string(in_out), self.dfcx.analysis_name);
550 assert!(self.dfcx.bits_per_id > 0);
552 cfg.graph.each_node(|node_index, node| {
553 debug!("DataFlowContext::walk_cfg idx={:?} id={} begin in_out={}",
554 node_index, node.data.id(), bits_to_string(in_out));
556 let (start, end) = self.dfcx.compute_id_range(node_index);
558 // Initialize local bitvector with state on-entry.
559 in_out.clone_from_slice(&self.dfcx.on_entry[start.. end]);
561 // Compute state on-exit by applying transfer function to
563 self.dfcx.apply_gen_kill(node_index, in_out);
565 // Propagate state on-exit from node into its successors.
566 self.propagate_bits_into_graph_successors_of(in_out, cfg, node_index);
567 true // continue to next node
571 fn reset(&mut self, bits: &mut [usize]) {
572 let e = if self.dfcx.oper.initial_value() {usize::MAX} else {0};
578 fn propagate_bits_into_graph_successors_of(&mut self,
582 for (_, edge) in cfg.graph.outgoing_edges(cfgidx) {
583 self.propagate_bits_into_entry_set_for(pred_bits, edge);
587 fn propagate_bits_into_entry_set_for(&mut self,
589 edge: &cfg::CFGEdge) {
590 let source = edge.source();
591 let cfgidx = edge.target();
592 debug!("{} propagate_bits_into_entry_set_for(pred_bits={}, {:?} to {:?})",
593 self.dfcx.analysis_name, bits_to_string(pred_bits), source, cfgidx);
594 assert!(self.dfcx.bits_per_id > 0);
596 let (start, end) = self.dfcx.compute_id_range(cfgidx);
598 // (scoping mutable borrow of self.dfcx.on_entry)
599 let on_entry = &mut self.dfcx.on_entry[start.. end];
600 bitwise(on_entry, pred_bits, &self.dfcx.oper)
603 debug!("{} changed entry set for {:?} to {}",
604 self.dfcx.analysis_name, cfgidx,
605 bits_to_string(&self.dfcx.on_entry[start.. end]));
611 fn mut_bits_to_string(words: &mut [usize]) -> String {
612 bits_to_string(words)
615 fn bits_to_string(words: &[usize]) -> String {
616 let mut result = String::new();
619 // Note: this is a little endian printout of bytes.
623 for _ in 0..mem::size_of::<usize>() {
625 result.push_str(&format!("{:02x}", v & 0xFF));
635 fn bitwise<Op:BitwiseOperator>(out_vec: &mut [usize],
638 assert_eq!(out_vec.len(), in_vec.len());
639 let mut changed = false;
640 for (out_elt, in_elt) in out_vec.iter_mut().zip(in_vec) {
641 let old_val = *out_elt;
642 let new_val = op.join(old_val, *in_elt);
644 changed |= old_val != new_val;
649 fn set_bit(words: &mut [usize], bit: usize) -> bool {
650 debug!("set_bit: words={} bit={}",
651 mut_bits_to_string(words), bit_str(bit));
652 let usize_bits = mem::size_of::<usize>() * 8;
653 let word = bit / usize_bits;
654 let bit_in_word = bit % usize_bits;
655 let bit_mask = 1 << bit_in_word;
656 debug!("word={} bit_in_word={} bit_mask={}", word, bit_in_word, word);
657 let oldv = words[word];
658 let newv = oldv | bit_mask;
663 fn bit_str(bit: usize) -> String {
665 let lobits = 1 << (bit & 0xFF);
666 format!("[{}:{}-{:02x}]", bit, byte, lobits)
670 impl BitwiseOperator for Union {
671 fn join(&self, a: usize, b: usize) -> usize { a | b }
674 impl BitwiseOperator for Subtract {
675 fn join(&self, a: usize, b: usize) -> usize { a & !b }