1 | // Copyright 2013 The Go Authors. All rights reserved. |
---|---|
2 | // Use of this source code is governed by a BSD-style |
3 | // license that can be found in the LICENSE file. |
4 | |
5 | package pointer |
6 | |
7 | // This file implements Hash-Value Numbering (HVN), a pre-solver |
8 | // constraint optimization described in Hardekopf & Lin, SAS'07 (see |
9 | // doc.go) that analyses the graph topology to determine which sets of |
10 | // variables are "pointer equivalent" (PE), i.e. must have identical |
11 | // points-to sets in the solution. |
12 | // |
13 | // A separate ("offline") graph is constructed. Its nodes are those of |
14 | // the main-graph, plus an additional node *X for each pointer node X. |
15 | // With this graph we can reason about the unknown points-to set of |
16 | // dereferenced pointers. (We do not generalize this to represent |
17 | // unknown fields x->f, perhaps because such fields would be numerous, |
18 | // though it might be worth an experiment.) |
19 | // |
20 | // Nodes whose points-to relations are not entirely captured by the |
21 | // graph are marked as "indirect": the *X nodes, the parameters of |
22 | // address-taken functions (which includes all functions in method |
23 | // sets), or nodes updated by the solver rules for reflection, etc. |
24 | // |
25 | // All addr (y=&x) nodes are initially assigned a pointer-equivalence |
26 | // (PE) label equal to x's nodeid in the main graph. (These are the |
27 | // only PE labels that are less than len(a.nodes).) |
28 | // |
29 | // All offsetAddr (y=&x.f) constraints are initially assigned a PE |
30 | // label; such labels are memoized, keyed by (x, f), so that equivalent |
31 | // nodes y as assigned the same label. |
32 | // |
33 | // Then we process each strongly connected component (SCC) of the graph |
34 | // in topological order, assigning it a PE label based on the set P of |
35 | // PE labels that flow to it from its immediate dependencies. |
36 | // |
37 | // If any node in P is "indirect", the entire SCC is assigned a fresh PE |
38 | // label. Otherwise: |
39 | // |
40 | // |P|=0 if P is empty, all nodes in the SCC are non-pointers (e.g. |
41 | // uninitialized variables, or formal params of dead functions) |
42 | // and the SCC is assigned the PE label of zero. |
43 | // |
44 | // |P|=1 if P is a singleton, the SCC is assigned the same label as the |
45 | // sole element of P. |
46 | // |
47 | // |P|>1 if P contains multiple labels, a unique label representing P is |
48 | // invented and recorded in an hash table, so that other |
49 | // equivalent SCCs may also be assigned this label, akin to |
50 | // conventional hash-value numbering in a compiler. |
51 | // |
52 | // Finally, a renumbering is computed such that each node is replaced by |
53 | // the lowest-numbered node with the same PE label. All constraints are |
54 | // renumbered, and any resulting duplicates are eliminated. |
55 | // |
56 | // The only nodes that are not renumbered are the objects x in addr |
57 | // (y=&x) constraints, since the ids of these nodes (and fields derived |
58 | // from them via offsetAddr rules) are the elements of all points-to |
59 | // sets, so they must remain as they are if we want the same solution. |
60 | // |
61 | // The solverStates (node.solve) for nodes in the same equivalence class |
62 | // are linked together so that all nodes in the class have the same |
63 | // solution. This avoids the need to renumber nodeids buried in |
64 | // Queries, cgnodes, etc (like (*analysis).renumber() does) since only |
65 | // the solution is needed. |
66 | // |
67 | // The result of HVN is that the number of distinct nodes and |
68 | // constraints is reduced, but the solution is identical (almost---see |
69 | // CROSS-CHECK below). In particular, both linear and cyclic chains of |
70 | // copies are each replaced by a single node. |
71 | // |
72 | // Nodes and constraints created "online" (e.g. while solving reflection |
73 | // constraints) are not subject to this optimization. |
74 | // |
75 | // PERFORMANCE |
76 | // |
77 | // In two benchmarks (guru and godoc), HVN eliminates about two thirds |
78 | // of nodes, the majority accounted for by non-pointers: nodes of |
79 | // non-pointer type, pointers that remain nil, formal parameters of dead |
80 | // functions, nodes of untracked types, etc. It also reduces the number |
81 | // of constraints, also by about two thirds, and the solving time by |
82 | // 30--42%, although we must pay about 15% for the running time of HVN |
83 | // itself. The benefit is greater for larger applications. |
84 | // |
85 | // There are many possible optimizations to improve the performance: |
86 | // * Use fewer than 1:1 onodes to main graph nodes: many of the onodes |
87 | // we create are not needed. |
88 | // * HU (HVN with Union---see paper): coalesce "union" peLabels when |
89 | // their expanded-out sets are equal. |
90 | // * HR (HVN with deReference---see paper): this will require that we |
91 | // apply HVN until fixed point, which may need more bookkeeping of the |
92 | // correspondence of main nodes to onodes. |
93 | // * Location Equivalence (see paper): have points-to sets contain not |
94 | // locations but location-equivalence class labels, each representing |
95 | // a set of locations. |
96 | // * HVN with field-sensitive ref: model each of the fields of a |
97 | // pointer-to-struct. |
98 | // |
99 | // CROSS-CHECK |
100 | // |
101 | // To verify the soundness of the optimization, when the |
102 | // debugHVNCrossCheck option is enabled, we run the solver twice, once |
103 | // before and once after running HVN, dumping the solution to disk, and |
104 | // then we compare the results. If they are not identical, the analysis |
105 | // panics. |
106 | // |
107 | // The solution dumped to disk includes only the N*N submatrix of the |
108 | // complete solution where N is the number of nodes after generation. |
109 | // In other words, we ignore pointer variables and objects created by |
110 | // the solver itself, since their numbering depends on the solver order, |
111 | // which is affected by the optimization. In any case, that's the only |
112 | // part the client cares about. |
113 | // |
114 | // The cross-check is too strict and may fail spuriously. Although the |
115 | // H&L paper describing HVN states that the solutions obtained should be |
116 | // identical, this is not the case in practice because HVN can collapse |
117 | // cycles involving *p even when pts(p)={}. Consider this example |
118 | // distilled from testdata/hello.go: |
119 | // |
120 | // var x T |
121 | // func f(p **T) { |
122 | // t0 = *p |
123 | // ... |
124 | // t1 = φ(t0, &x) |
125 | // *p = t1 |
126 | // } |
127 | // |
128 | // If f is dead code, we get: |
129 | // unoptimized: pts(p)={} pts(t0)={} pts(t1)={&x} |
130 | // optimized: pts(p)={} pts(t0)=pts(t1)=pts(*p)={&x} |
131 | // |
132 | // It's hard to argue that this is a bug: the result is sound and the |
133 | // loss of precision is inconsequential---f is dead code, after all. |
134 | // But unfortunately it limits the usefulness of the cross-check since |
135 | // failures must be carefully analyzed. Ben Hardekopf suggests (in |
136 | // personal correspondence) some approaches to mitigating it: |
137 | // |
138 | // If there is a node with an HVN points-to set that is a superset |
139 | // of the NORM points-to set, then either it's a bug or it's a |
140 | // result of this issue. If it's a result of this issue, then in |
141 | // the offline constraint graph there should be a REF node inside |
142 | // some cycle that reaches this node, and in the NORM solution the |
143 | // pointer being dereferenced by that REF node should be the empty |
144 | // set. If that isn't true then this is a bug. If it is true, then |
145 | // you can further check that in the NORM solution the "extra" |
146 | // points-to info in the HVN solution does in fact come from that |
147 | // purported cycle (if it doesn't, then this is still a bug). If |
148 | // you're doing the further check then you'll need to do it for |
149 | // each "extra" points-to element in the HVN points-to set. |
150 | // |
151 | // There are probably ways to optimize these checks by taking |
152 | // advantage of graph properties. For example, extraneous points-to |
153 | // info will flow through the graph and end up in many |
154 | // nodes. Rather than checking every node with extra info, you |
155 | // could probably work out the "origin point" of the extra info and |
156 | // just check there. Note that the check in the first bullet is |
157 | // looking for soundness bugs, while the check in the second bullet |
158 | // is looking for precision bugs; depending on your needs, you may |
159 | // care more about one than the other. |
160 | // |
161 | // which we should evaluate. The cross-check is nonetheless invaluable |
162 | // for all but one of the programs in the pointer_test suite. |
163 | |
164 | import ( |
165 | "fmt" |
166 | "go/types" |
167 | "io" |
168 | "reflect" |
169 | |
170 | "golang.org/x/tools/container/intsets" |
171 | ) |
172 | |
173 | // A peLabel is a pointer-equivalence label: two nodes with the same |
174 | // peLabel have identical points-to solutions. |
175 | // |
176 | // The numbers are allocated consecutively like so: |
177 | // |
178 | // 0 not a pointer |
179 | // 1..N-1 addrConstraints (equals the constraint's .src field, hence sparse) |
180 | // ... offsetAddr constraints |
181 | // ... SCCs (with indirect nodes or multiple inputs) |
182 | // |
183 | // Each PE label denotes a set of pointers containing a single addr, a |
184 | // single offsetAddr, or some set of other PE labels. |
185 | type peLabel int |
186 | |
187 | type hvn struct { |
188 | a *analysis |
189 | N int // len(a.nodes) immediately after constraint generation |
190 | log io.Writer // (optional) log of HVN lemmas |
191 | onodes []*onode // nodes of the offline graph |
192 | label peLabel // the next available PE label |
193 | hvnLabel map[string]peLabel // hash-value numbering (PE label) for each set of onodeids |
194 | stack []onodeid // DFS stack |
195 | index int32 // next onode.index, from Tarjan's SCC algorithm |
196 | |
197 | // For each distinct offsetAddrConstraint (src, offset) pair, |
198 | // offsetAddrLabels records a unique PE label >= N. |
199 | offsetAddrLabels map[offsetAddr]peLabel |
200 | } |
201 | |
202 | // The index of an node in the offline graph. |
203 | // (Currently the first N align with the main nodes, |
204 | // but this may change with HRU.) |
205 | type onodeid uint32 |
206 | |
207 | // An onode is a node in the offline constraint graph. |
208 | // (Where ambiguous, members of analysis.nodes are referred to as |
209 | // "main graph" nodes.) |
210 | // |
211 | // Edges in the offline constraint graph (edges and implicit) point to |
212 | // the source, i.e. against the flow of values: they are dependencies. |
213 | // Implicit edges are used for SCC computation, but not for gathering |
214 | // incoming labels. |
215 | type onode struct { |
216 | rep onodeid // index of representative of SCC in offline constraint graph |
217 | |
218 | edges intsets.Sparse // constraint edges X-->Y (this onode is X) |
219 | implicit intsets.Sparse // implicit edges *X-->*Y (this onode is X) |
220 | peLabels intsets.Sparse // set of peLabels are pointer-equivalent to this one |
221 | indirect bool // node has points-to relations not represented in graph |
222 | |
223 | // Tarjan's SCC algorithm |
224 | index, lowlink int32 // Tarjan numbering |
225 | scc int32 // -ve => on stack; 0 => unvisited; +ve => node is root of a found SCC |
226 | } |
227 | |
228 | type offsetAddr struct { |
229 | ptr nodeid |
230 | offset uint32 |
231 | } |
232 | |
233 | // nextLabel issues the next unused pointer-equivalence label. |
234 | func (h *hvn) nextLabel() peLabel { |
235 | h.label++ |
236 | return h.label |
237 | } |
238 | |
239 | // ref(X) returns the index of the onode for *X. |
240 | func (h *hvn) ref(id onodeid) onodeid { |
241 | return id + onodeid(len(h.a.nodes)) |
242 | } |
243 | |
244 | // hvn computes pointer-equivalence labels (peLabels) using the Hash-based |
245 | // Value Numbering (HVN) algorithm described in Hardekopf & Lin, SAS'07. |
246 | func (a *analysis) hvn() { |
247 | start("HVN") |
248 | |
249 | if a.log != nil { |
250 | fmt.Fprintf(a.log, "\n\n==== Pointer equivalence optimization\n\n") |
251 | } |
252 | |
253 | h := hvn{ |
254 | a: a, |
255 | N: len(a.nodes), |
256 | log: a.log, |
257 | hvnLabel: make(map[string]peLabel), |
258 | offsetAddrLabels: make(map[offsetAddr]peLabel), |
259 | } |
260 | |
261 | if h.log != nil { |
262 | fmt.Fprintf(h.log, "\nCreating offline graph nodes...\n") |
263 | } |
264 | |
265 | // Create offline nodes. The first N nodes correspond to main |
266 | // graph nodes; the next N are their corresponding ref() nodes. |
267 | h.onodes = make([]*onode, 2*h.N) |
268 | for id := range a.nodes { |
269 | id := onodeid(id) |
270 | h.onodes[id] = &onode{} |
271 | h.onodes[h.ref(id)] = &onode{indirect: true} |
272 | } |
273 | |
274 | // Each node initially represents just itself. |
275 | for id, o := range h.onodes { |
276 | o.rep = onodeid(id) |
277 | } |
278 | |
279 | h.markIndirectNodes() |
280 | |
281 | // Reserve the first N PE labels for addrConstraints. |
282 | h.label = peLabel(h.N) |
283 | |
284 | // Add offline constraint edges. |
285 | if h.log != nil { |
286 | fmt.Fprintf(h.log, "\nAdding offline graph edges...\n") |
287 | } |
288 | for _, c := range a.constraints { |
289 | if debugHVNVerbose && h.log != nil { |
290 | fmt.Fprintf(h.log, "; %s\n", c) |
291 | } |
292 | c.presolve(&h) |
293 | } |
294 | |
295 | // Find and collapse SCCs. |
296 | if h.log != nil { |
297 | fmt.Fprintf(h.log, "\nFinding SCCs...\n") |
298 | } |
299 | h.index = 1 |
300 | for id, o := range h.onodes { |
301 | if id > 0 && o.index == 0 { |
302 | // Start depth-first search at each unvisited node. |
303 | h.visit(onodeid(id)) |
304 | } |
305 | } |
306 | |
307 | // Dump the solution |
308 | // (NB: somewhat redundant with logging from simplify().) |
309 | if debugHVNVerbose && h.log != nil { |
310 | fmt.Fprintf(h.log, "\nPointer equivalences:\n") |
311 | for id, o := range h.onodes { |
312 | if id == 0 { |
313 | continue |
314 | } |
315 | if id == int(h.N) { |
316 | fmt.Fprintf(h.log, "---\n") |
317 | } |
318 | fmt.Fprintf(h.log, "o%d\t", id) |
319 | if o.rep != onodeid(id) { |
320 | fmt.Fprintf(h.log, "rep=o%d", o.rep) |
321 | } else { |
322 | fmt.Fprintf(h.log, "p%d", o.peLabels.Min()) |
323 | if o.indirect { |
324 | fmt.Fprint(h.log, " indirect") |
325 | } |
326 | } |
327 | fmt.Fprintln(h.log) |
328 | } |
329 | } |
330 | |
331 | // Simplify the main constraint graph |
332 | h.simplify() |
333 | |
334 | a.showCounts() |
335 | |
336 | stop("HVN") |
337 | } |
338 | |
339 | // ---- constraint-specific rules ---- |
340 | |
341 | // dst := &src |
342 | func (c *addrConstraint) presolve(h *hvn) { |
343 | // Each object (src) is an initial PE label. |
344 | label := peLabel(c.src) // label < N |
345 | if debugHVNVerbose && h.log != nil { |
346 | // duplicate log messages are possible |
347 | fmt.Fprintf(h.log, "\tcreate p%d: {&n%d}\n", label, c.src) |
348 | } |
349 | odst := onodeid(c.dst) |
350 | osrc := onodeid(c.src) |
351 | |
352 | // Assign dst this label. |
353 | h.onodes[odst].peLabels.Insert(int(label)) |
354 | if debugHVNVerbose && h.log != nil { |
355 | fmt.Fprintf(h.log, "\to%d has p%d\n", odst, label) |
356 | } |
357 | |
358 | h.addImplicitEdge(h.ref(odst), osrc) // *dst ~~> src. |
359 | } |
360 | |
361 | // dst = src |
362 | func (c *copyConstraint) presolve(h *hvn) { |
363 | odst := onodeid(c.dst) |
364 | osrc := onodeid(c.src) |
365 | h.addEdge(odst, osrc) // dst --> src |
366 | h.addImplicitEdge(h.ref(odst), h.ref(osrc)) // *dst ~~> *src |
367 | } |
368 | |
369 | // dst = *src + offset |
370 | func (c *loadConstraint) presolve(h *hvn) { |
371 | odst := onodeid(c.dst) |
372 | osrc := onodeid(c.src) |
373 | if c.offset == 0 { |
374 | h.addEdge(odst, h.ref(osrc)) // dst --> *src |
375 | } else { |
376 | // We don't interpret load-with-offset, e.g. results |
377 | // of map value lookup, R-block of dynamic call, slice |
378 | // copy/append, reflection. |
379 | h.markIndirect(odst, "load with offset") |
380 | } |
381 | } |
382 | |
383 | // *dst + offset = src |
384 | func (c *storeConstraint) presolve(h *hvn) { |
385 | odst := onodeid(c.dst) |
386 | osrc := onodeid(c.src) |
387 | if c.offset == 0 { |
388 | h.onodes[h.ref(odst)].edges.Insert(int(osrc)) // *dst --> src |
389 | if debugHVNVerbose && h.log != nil { |
390 | fmt.Fprintf(h.log, "\to%d --> o%d\n", h.ref(odst), osrc) |
391 | } |
392 | } |
393 | // We don't interpret store-with-offset. |
394 | // See discussion of soundness at markIndirectNodes. |
395 | } |
396 | |
397 | // dst = &src.offset |
398 | func (c *offsetAddrConstraint) presolve(h *hvn) { |
399 | // Give each distinct (addr, offset) pair a fresh PE label. |
400 | // The cache performs CSE, effectively. |
401 | key := offsetAddr{c.src, c.offset} |
402 | label, ok := h.offsetAddrLabels[key] |
403 | if !ok { |
404 | label = h.nextLabel() |
405 | h.offsetAddrLabels[key] = label |
406 | if debugHVNVerbose && h.log != nil { |
407 | fmt.Fprintf(h.log, "\tcreate p%d: {&n%d.#%d}\n", |
408 | label, c.src, c.offset) |
409 | } |
410 | } |
411 | |
412 | // Assign dst this label. |
413 | h.onodes[c.dst].peLabels.Insert(int(label)) |
414 | if debugHVNVerbose && h.log != nil { |
415 | fmt.Fprintf(h.log, "\to%d has p%d\n", c.dst, label) |
416 | } |
417 | } |
418 | |
419 | // dst = src.(typ) where typ is an interface |
420 | func (c *typeFilterConstraint) presolve(h *hvn) { |
421 | h.markIndirect(onodeid(c.dst), "typeFilter result") |
422 | } |
423 | |
424 | // dst = src.(typ) where typ is concrete |
425 | func (c *untagConstraint) presolve(h *hvn) { |
426 | odst := onodeid(c.dst) |
427 | for end := odst + onodeid(h.a.sizeof(c.typ)); odst < end; odst++ { |
428 | h.markIndirect(odst, "untag result") |
429 | } |
430 | } |
431 | |
432 | // dst = src.method(c.params...) |
433 | func (c *invokeConstraint) presolve(h *hvn) { |
434 | // All methods are address-taken functions, so |
435 | // their formal P-blocks were already marked indirect. |
436 | |
437 | // Mark the caller's targets node as indirect. |
438 | sig := c.method.Type().(*types.Signature) |
439 | id := c.params |
440 | h.markIndirect(onodeid(c.params), "invoke targets node") |
441 | id++ |
442 | |
443 | id += nodeid(h.a.sizeof(sig.Params())) |
444 | |
445 | // Mark the caller's R-block as indirect. |
446 | end := id + nodeid(h.a.sizeof(sig.Results())) |
447 | for id < end { |
448 | h.markIndirect(onodeid(id), "invoke R-block") |
449 | id++ |
450 | } |
451 | } |
452 | |
453 | // markIndirectNodes marks as indirect nodes whose points-to relations |
454 | // are not entirely captured by the offline graph, including: |
455 | // |
456 | // (a) All address-taken nodes (including the following nodes within |
457 | // the same object). This is described in the paper. |
458 | // |
459 | // The most subtle cause of indirect nodes is the generation of |
460 | // store-with-offset constraints since the offline graph doesn't |
461 | // represent them. A global audit of constraint generation reveals the |
462 | // following uses of store-with-offset: |
463 | // |
464 | // (b) genDynamicCall, for P-blocks of dynamically called functions, |
465 | // to which dynamic copy edges will be added to them during |
466 | // solving: from storeConstraint for standalone functions, |
467 | // and from invokeConstraint for methods. |
468 | // All such P-blocks must be marked indirect. |
469 | // (c) MakeUpdate, to update the value part of a map object. |
470 | // All MakeMap objects's value parts must be marked indirect. |
471 | // (d) copyElems, to update the destination array. |
472 | // All array elements must be marked indirect. |
473 | // |
474 | // Not all indirect marking happens here. ref() nodes are marked |
475 | // indirect at construction, and each constraint's presolve() method may |
476 | // mark additional nodes. |
477 | func (h *hvn) markIndirectNodes() { |
478 | // (a) all address-taken nodes, plus all nodes following them |
479 | // within the same object, since these may be indirectly |
480 | // stored or address-taken. |
481 | for _, c := range h.a.constraints { |
482 | if c, ok := c.(*addrConstraint); ok { |
483 | start := h.a.enclosingObj(c.src) |
484 | end := start + nodeid(h.a.nodes[start].obj.size) |
485 | for id := c.src; id < end; id++ { |
486 | h.markIndirect(onodeid(id), "A-T object") |
487 | } |
488 | } |
489 | } |
490 | |
491 | // (b) P-blocks of all address-taken functions. |
492 | for id := 0; id < h.N; id++ { |
493 | obj := h.a.nodes[id].obj |
494 | |
495 | // TODO(adonovan): opt: if obj.cgn.fn is a method and |
496 | // obj.cgn is not its shared contour, this is an |
497 | // "inlined" static method call. We needn't consider it |
498 | // address-taken since no invokeConstraint will affect it. |
499 | |
500 | if obj != nil && obj.flags&otFunction != 0 && h.a.atFuncs[obj.cgn.fn] { |
501 | // address-taken function |
502 | if debugHVNVerbose && h.log != nil { |
503 | fmt.Fprintf(h.log, "n%d is address-taken: %s\n", id, obj.cgn.fn) |
504 | } |
505 | h.markIndirect(onodeid(id), "A-T func identity") |
506 | id++ |
507 | sig := obj.cgn.fn.Signature |
508 | psize := h.a.sizeof(sig.Params()) |
509 | if sig.Recv() != nil { |
510 | psize += h.a.sizeof(sig.Recv().Type()) |
511 | } |
512 | for end := id + int(psize); id < end; id++ { |
513 | h.markIndirect(onodeid(id), "A-T func P-block") |
514 | } |
515 | id-- |
516 | continue |
517 | } |
518 | } |
519 | |
520 | // (c) all map objects' value fields. |
521 | for _, id := range h.a.mapValues { |
522 | h.markIndirect(onodeid(id), "makemap.value") |
523 | } |
524 | |
525 | // (d) all array element objects. |
526 | // TODO(adonovan): opt: can we do better? |
527 | for id := 0; id < h.N; id++ { |
528 | // Identity node for an object of array type? |
529 | if tArray, ok := h.a.nodes[id].typ.(*types.Array); ok { |
530 | // Mark the array element nodes indirect. |
531 | // (Skip past the identity field.) |
532 | for range h.a.flatten(tArray.Elem()) { |
533 | id++ |
534 | h.markIndirect(onodeid(id), "array elem") |
535 | } |
536 | } |
537 | } |
538 | } |
539 | |
540 | func (h *hvn) markIndirect(oid onodeid, comment string) { |
541 | h.onodes[oid].indirect = true |
542 | if debugHVNVerbose && h.log != nil { |
543 | fmt.Fprintf(h.log, "\to%d is indirect: %s\n", oid, comment) |
544 | } |
545 | } |
546 | |
547 | // Adds an edge dst-->src. |
548 | // Note the unusual convention: edges are dependency (contraflow) edges. |
549 | func (h *hvn) addEdge(odst, osrc onodeid) { |
550 | h.onodes[odst].edges.Insert(int(osrc)) |
551 | if debugHVNVerbose && h.log != nil { |
552 | fmt.Fprintf(h.log, "\to%d --> o%d\n", odst, osrc) |
553 | } |
554 | } |
555 | |
556 | func (h *hvn) addImplicitEdge(odst, osrc onodeid) { |
557 | h.onodes[odst].implicit.Insert(int(osrc)) |
558 | if debugHVNVerbose && h.log != nil { |
559 | fmt.Fprintf(h.log, "\to%d ~~> o%d\n", odst, osrc) |
560 | } |
561 | } |
562 | |
563 | // visit implements the depth-first search of Tarjan's SCC algorithm. |
564 | // Precondition: x is canonical. |
565 | func (h *hvn) visit(x onodeid) { |
566 | h.checkCanonical(x) |
567 | xo := h.onodes[x] |
568 | xo.index = h.index |
569 | xo.lowlink = h.index |
570 | h.index++ |
571 | |
572 | h.stack = append(h.stack, x) // push |
573 | assert(xo.scc == 0, "node revisited") |
574 | xo.scc = -1 |
575 | |
576 | var deps []int |
577 | deps = xo.edges.AppendTo(deps) |
578 | deps = xo.implicit.AppendTo(deps) |
579 | |
580 | for _, y := range deps { |
581 | // Loop invariant: x is canonical. |
582 | |
583 | y := h.find(onodeid(y)) |
584 | |
585 | if x == y { |
586 | continue // nodes already coalesced |
587 | } |
588 | |
589 | xo := h.onodes[x] |
590 | yo := h.onodes[y] |
591 | |
592 | switch { |
593 | case yo.scc > 0: |
594 | // y is already a collapsed SCC |
595 | |
596 | case yo.scc < 0: |
597 | // y is on the stack, and thus in the current SCC. |
598 | if yo.index < xo.lowlink { |
599 | xo.lowlink = yo.index |
600 | } |
601 | |
602 | default: |
603 | // y is unvisited; visit it now. |
604 | h.visit(y) |
605 | // Note: x and y are now non-canonical. |
606 | |
607 | x = h.find(onodeid(x)) |
608 | |
609 | if yo.lowlink < xo.lowlink { |
610 | xo.lowlink = yo.lowlink |
611 | } |
612 | } |
613 | } |
614 | h.checkCanonical(x) |
615 | |
616 | // Is x the root of an SCC? |
617 | if xo.lowlink == xo.index { |
618 | // Coalesce all nodes in the SCC. |
619 | if debugHVNVerbose && h.log != nil { |
620 | fmt.Fprintf(h.log, "scc o%d\n", x) |
621 | } |
622 | for { |
623 | // Pop y from stack. |
624 | i := len(h.stack) - 1 |
625 | y := h.stack[i] |
626 | h.stack = h.stack[:i] |
627 | |
628 | h.checkCanonical(x) |
629 | xo := h.onodes[x] |
630 | h.checkCanonical(y) |
631 | yo := h.onodes[y] |
632 | |
633 | if xo == yo { |
634 | // SCC is complete. |
635 | xo.scc = 1 |
636 | h.labelSCC(x) |
637 | break |
638 | } |
639 | h.coalesce(x, y) |
640 | } |
641 | } |
642 | } |
643 | |
644 | // Precondition: x is canonical. |
645 | func (h *hvn) labelSCC(x onodeid) { |
646 | h.checkCanonical(x) |
647 | xo := h.onodes[x] |
648 | xpe := &xo.peLabels |
649 | |
650 | // All indirect nodes get new labels. |
651 | if xo.indirect { |
652 | label := h.nextLabel() |
653 | if debugHVNVerbose && h.log != nil { |
654 | fmt.Fprintf(h.log, "\tcreate p%d: indirect SCC\n", label) |
655 | fmt.Fprintf(h.log, "\to%d has p%d\n", x, label) |
656 | } |
657 | |
658 | // Remove pre-labeling, in case a direct pre-labeled node was |
659 | // merged with an indirect one. |
660 | xpe.Clear() |
661 | xpe.Insert(int(label)) |
662 | |
663 | return |
664 | } |
665 | |
666 | // Invariant: all peLabels sets are non-empty. |
667 | // Those that are logically empty contain zero as their sole element. |
668 | // No other sets contains zero. |
669 | |
670 | // Find all labels coming in to the coalesced SCC node. |
671 | for _, y := range xo.edges.AppendTo(nil) { |
672 | y := h.find(onodeid(y)) |
673 | if y == x { |
674 | continue // already coalesced |
675 | } |
676 | ype := &h.onodes[y].peLabels |
677 | if debugHVNVerbose && h.log != nil { |
678 | fmt.Fprintf(h.log, "\tedge from o%d = %s\n", y, ype) |
679 | } |
680 | |
681 | if ype.IsEmpty() { |
682 | if debugHVNVerbose && h.log != nil { |
683 | fmt.Fprintf(h.log, "\tnode has no PE label\n") |
684 | } |
685 | } |
686 | assert(!ype.IsEmpty(), "incoming node has no PE label") |
687 | |
688 | if ype.Has(0) { |
689 | // {0} represents a non-pointer. |
690 | assert(ype.Len() == 1, "PE set contains {0, ...}") |
691 | } else { |
692 | xpe.UnionWith(ype) |
693 | } |
694 | } |
695 | |
696 | switch xpe.Len() { |
697 | case 0: |
698 | // SCC has no incoming non-zero PE labels: it is a non-pointer. |
699 | xpe.Insert(0) |
700 | |
701 | case 1: |
702 | // already a singleton |
703 | |
704 | default: |
705 | // SCC has multiple incoming non-zero PE labels. |
706 | // Find the canonical label representing this set. |
707 | // We use String() as a fingerprint consistent with Equals(). |
708 | key := xpe.String() |
709 | label, ok := h.hvnLabel[key] |
710 | if !ok { |
711 | label = h.nextLabel() |
712 | if debugHVNVerbose && h.log != nil { |
713 | fmt.Fprintf(h.log, "\tcreate p%d: union %s\n", label, xpe.String()) |
714 | } |
715 | h.hvnLabel[key] = label |
716 | } |
717 | xpe.Clear() |
718 | xpe.Insert(int(label)) |
719 | } |
720 | |
721 | if debugHVNVerbose && h.log != nil { |
722 | fmt.Fprintf(h.log, "\to%d has p%d\n", x, xpe.Min()) |
723 | } |
724 | } |
725 | |
726 | // coalesce combines two nodes in the offline constraint graph. |
727 | // Precondition: x and y are canonical. |
728 | func (h *hvn) coalesce(x, y onodeid) { |
729 | xo := h.onodes[x] |
730 | yo := h.onodes[y] |
731 | |
732 | // x becomes y's canonical representative. |
733 | yo.rep = x |
734 | |
735 | if debugHVNVerbose && h.log != nil { |
736 | fmt.Fprintf(h.log, "\tcoalesce o%d into o%d\n", y, x) |
737 | } |
738 | |
739 | // x accumulates y's edges. |
740 | xo.edges.UnionWith(&yo.edges) |
741 | yo.edges.Clear() |
742 | |
743 | // x accumulates y's implicit edges. |
744 | xo.implicit.UnionWith(&yo.implicit) |
745 | yo.implicit.Clear() |
746 | |
747 | // x accumulates y's pointer-equivalence labels. |
748 | xo.peLabels.UnionWith(&yo.peLabels) |
749 | yo.peLabels.Clear() |
750 | |
751 | // x accumulates y's indirect flag. |
752 | if yo.indirect { |
753 | xo.indirect = true |
754 | } |
755 | } |
756 | |
757 | // simplify computes a degenerate renumbering of nodeids from the PE |
758 | // labels assigned by the hvn, and uses it to simplify the main |
759 | // constraint graph, eliminating non-pointer nodes and duplicate |
760 | // constraints. |
761 | func (h *hvn) simplify() { |
762 | // canon maps each peLabel to its canonical main node. |
763 | canon := make([]nodeid, h.label) |
764 | for i := range canon { |
765 | canon[i] = nodeid(h.N) // indicates "unset" |
766 | } |
767 | |
768 | // mapping maps each main node index to the index of the canonical node. |
769 | mapping := make([]nodeid, len(h.a.nodes)) |
770 | |
771 | for id := range h.a.nodes { |
772 | id := nodeid(id) |
773 | if id == 0 { |
774 | canon[0] = 0 |
775 | mapping[0] = 0 |
776 | continue |
777 | } |
778 | oid := h.find(onodeid(id)) |
779 | peLabels := &h.onodes[oid].peLabels |
780 | assert(peLabels.Len() == 1, "PE class is not a singleton") |
781 | label := peLabel(peLabels.Min()) |
782 | |
783 | canonID := canon[label] |
784 | if canonID == nodeid(h.N) { |
785 | // id becomes the representative of the PE label. |
786 | canonID = id |
787 | canon[label] = canonID |
788 | |
789 | if h.a.log != nil { |
790 | fmt.Fprintf(h.a.log, "\tpts(n%d) is canonical : \t(%s)\n", |
791 | id, h.a.nodes[id].typ) |
792 | } |
793 | |
794 | } else { |
795 | // Link the solver states for the two nodes. |
796 | assert(h.a.nodes[canonID].solve != nil, "missing solver state") |
797 | h.a.nodes[id].solve = h.a.nodes[canonID].solve |
798 | |
799 | if h.a.log != nil { |
800 | // TODO(adonovan): debug: reorganize the log so it prints |
801 | // one line: |
802 | // pe y = x1, ..., xn |
803 | // for each canonical y. Requires allocation. |
804 | fmt.Fprintf(h.a.log, "\tpts(n%d) = pts(n%d) : %s\n", |
805 | id, canonID, h.a.nodes[id].typ) |
806 | } |
807 | } |
808 | |
809 | mapping[id] = canonID |
810 | } |
811 | |
812 | // Renumber the constraints, eliminate duplicates, and eliminate |
813 | // any containing non-pointers (n0). |
814 | addrs := make(map[addrConstraint]bool) |
815 | copys := make(map[copyConstraint]bool) |
816 | loads := make(map[loadConstraint]bool) |
817 | stores := make(map[storeConstraint]bool) |
818 | offsetAddrs := make(map[offsetAddrConstraint]bool) |
819 | untags := make(map[untagConstraint]bool) |
820 | typeFilters := make(map[typeFilterConstraint]bool) |
821 | invokes := make(map[invokeConstraint]bool) |
822 | |
823 | nbefore := len(h.a.constraints) |
824 | cc := h.a.constraints[:0] // in-situ compaction |
825 | for _, c := range h.a.constraints { |
826 | // Renumber. |
827 | switch c := c.(type) { |
828 | case *addrConstraint: |
829 | // Don't renumber c.src since it is the label of |
830 | // an addressable object and will appear in PT sets. |
831 | c.dst = mapping[c.dst] |
832 | default: |
833 | c.renumber(mapping) |
834 | } |
835 | |
836 | if c.ptr() == 0 { |
837 | continue // skip: constraint attached to non-pointer |
838 | } |
839 | |
840 | var dup bool |
841 | switch c := c.(type) { |
842 | case *addrConstraint: |
843 | _, dup = addrs[*c] |
844 | addrs[*c] = true |
845 | |
846 | case *copyConstraint: |
847 | if c.src == c.dst { |
848 | continue // skip degenerate copies |
849 | } |
850 | if c.src == 0 { |
851 | continue // skip copy from non-pointer |
852 | } |
853 | _, dup = copys[*c] |
854 | copys[*c] = true |
855 | |
856 | case *loadConstraint: |
857 | if c.src == 0 { |
858 | continue // skip load from non-pointer |
859 | } |
860 | _, dup = loads[*c] |
861 | loads[*c] = true |
862 | |
863 | case *storeConstraint: |
864 | if c.src == 0 { |
865 | continue // skip store from non-pointer |
866 | } |
867 | _, dup = stores[*c] |
868 | stores[*c] = true |
869 | |
870 | case *offsetAddrConstraint: |
871 | if c.src == 0 { |
872 | continue // skip offset from non-pointer |
873 | } |
874 | _, dup = offsetAddrs[*c] |
875 | offsetAddrs[*c] = true |
876 | |
877 | case *untagConstraint: |
878 | if c.src == 0 { |
879 | continue // skip untag of non-pointer |
880 | } |
881 | _, dup = untags[*c] |
882 | untags[*c] = true |
883 | |
884 | case *typeFilterConstraint: |
885 | if c.src == 0 { |
886 | continue // skip filter of non-pointer |
887 | } |
888 | _, dup = typeFilters[*c] |
889 | typeFilters[*c] = true |
890 | |
891 | case *invokeConstraint: |
892 | if c.params == 0 { |
893 | panic("non-pointer invoke.params") |
894 | } |
895 | if c.iface == 0 { |
896 | continue // skip invoke on non-pointer |
897 | } |
898 | _, dup = invokes[*c] |
899 | invokes[*c] = true |
900 | |
901 | default: |
902 | // We don't bother de-duping advanced constraints |
903 | // (e.g. reflection) since they are uncommon. |
904 | |
905 | // Eliminate constraints containing non-pointer nodeids. |
906 | // |
907 | // We use reflection to find the fields to avoid |
908 | // adding yet another method to constraint. |
909 | // |
910 | // TODO(adonovan): experiment with a constraint |
911 | // method that returns a slice of pointers to |
912 | // nodeids fields to enable uniform iteration; |
913 | // the renumber() method could be removed and |
914 | // implemented using the new one. |
915 | // |
916 | // TODO(adonovan): opt: this is unsound since |
917 | // some constraints still have an effect if one |
918 | // of the operands is zero: rVCall, rVMapIndex, |
919 | // rvSetMapIndex. Handle them specially. |
920 | rtNodeid := reflect.TypeOf(nodeid(0)) |
921 | x := reflect.ValueOf(c).Elem() |
922 | for i, nf := 0, x.NumField(); i < nf; i++ { |
923 | f := x.Field(i) |
924 | if f.Type() == rtNodeid { |
925 | if f.Uint() == 0 { |
926 | dup = true // skip it |
927 | break |
928 | } |
929 | } |
930 | } |
931 | } |
932 | if dup { |
933 | continue // skip duplicates |
934 | } |
935 | |
936 | cc = append(cc, c) |
937 | } |
938 | h.a.constraints = cc |
939 | |
940 | if h.log != nil { |
941 | fmt.Fprintf(h.log, "#constraints: was %d, now %d\n", nbefore, len(h.a.constraints)) |
942 | } |
943 | } |
944 | |
945 | // find returns the canonical onodeid for x. |
946 | // (The onodes form a disjoint set forest.) |
947 | func (h *hvn) find(x onodeid) onodeid { |
948 | // TODO(adonovan): opt: this is a CPU hotspot. Try "union by rank". |
949 | xo := h.onodes[x] |
950 | rep := xo.rep |
951 | if rep != x { |
952 | rep = h.find(rep) // simple path compression |
953 | xo.rep = rep |
954 | } |
955 | return rep |
956 | } |
957 | |
958 | func (h *hvn) checkCanonical(x onodeid) { |
959 | if debugHVN { |
960 | assert(x == h.find(x), "not canonical") |
961 | } |
962 | } |
963 | |
964 | func assert(p bool, msg string) { |
965 | if debugHVN && !p { |
966 | panic("assertion failed: " + msg) |
967 | } |
968 | } |
969 |
Members