2016-07-14 Steven G. Kargl <kargl@gcc.gnu.org>
[official-gcc.git] / gcc / sanopt.c
blob26604539ca5110d206671d6eebcff1b131243aba
1 /* Optimize and expand sanitizer functions.
2 Copyright (C) 2014-2016 Free Software Foundation, Inc.
3 Contributed by Marek Polacek <polacek@redhat.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "backend.h"
25 #include "tree.h"
26 #include "gimple.h"
27 #include "tree-pass.h"
28 #include "tree-ssa-operands.h"
29 #include "gimple-pretty-print.h"
30 #include "fold-const.h"
31 #include "gimple-iterator.h"
32 #include "asan.h"
33 #include "ubsan.h"
34 #include "params.h"
35 #include "tree-hash-traits.h"
38 /* This is used to carry information about basic blocks. It is
39 attached to the AUX field of the standard CFG block. */
41 struct sanopt_info
43 /* True if this BB might call (directly or indirectly) free/munmap
44 or similar operation. */
45 bool has_freeing_call_p;
47 /* True if HAS_FREEING_CALL_P flag has been computed. */
48 bool has_freeing_call_computed_p;
50 /* True if there is a block with HAS_FREEING_CALL_P flag set
51 on any path between an immediate dominator of BB, denoted
52 imm(BB), and BB. */
53 bool imm_dom_path_with_freeing_call_p;
55 /* True if IMM_DOM_PATH_WITH_FREEING_CALL_P has been computed. */
56 bool imm_dom_path_with_freeing_call_computed_p;
58 /* Number of possibly freeing calls encountered in this bb
59 (so far). */
60 uint64_t freeing_call_events;
62 /* True if BB is currently being visited during computation
63 of IMM_DOM_PATH_WITH_FREEING_CALL_P flag. */
64 bool being_visited_p;
66 /* True if this BB has been visited in the dominator walk. */
67 bool visited_p;
70 /* If T has a single definition of form T = T2, return T2. */
72 static tree
73 maybe_get_single_definition (tree t)
75 if (TREE_CODE (t) == SSA_NAME)
77 gimple *g = SSA_NAME_DEF_STMT (t);
78 if (gimple_assign_single_p (g))
79 return gimple_assign_rhs1 (g);
81 return NULL_TREE;
84 /* Tree triplet for vptr_check_map. */
85 struct sanopt_tree_triplet
87 tree t1, t2, t3;
90 /* Traits class for tree triplet hash maps below. */
92 struct sanopt_tree_triplet_hash : typed_noop_remove <sanopt_tree_triplet>
94 typedef sanopt_tree_triplet value_type;
95 typedef sanopt_tree_triplet compare_type;
97 static inline hashval_t
98 hash (const sanopt_tree_triplet &ref)
100 inchash::hash hstate (0);
101 inchash::add_expr (ref.t1, hstate);
102 inchash::add_expr (ref.t2, hstate);
103 inchash::add_expr (ref.t3, hstate);
104 return hstate.end ();
107 static inline bool
108 equal (const sanopt_tree_triplet &ref1, const sanopt_tree_triplet &ref2)
110 return operand_equal_p (ref1.t1, ref2.t1, 0)
111 && operand_equal_p (ref1.t2, ref2.t2, 0)
112 && operand_equal_p (ref1.t3, ref2.t3, 0);
115 static inline void
116 mark_deleted (sanopt_tree_triplet &ref)
118 ref.t1 = reinterpret_cast<tree> (1);
121 static inline void
122 mark_empty (sanopt_tree_triplet &ref)
124 ref.t1 = NULL;
127 static inline bool
128 is_deleted (const sanopt_tree_triplet &ref)
130 return ref.t1 == (void *) 1;
133 static inline bool
134 is_empty (const sanopt_tree_triplet &ref)
136 return ref.t1 == NULL;
140 /* This is used to carry various hash maps and variables used
141 in sanopt_optimize_walker. */
143 struct sanopt_ctx
145 /* This map maps a pointer (the first argument of UBSAN_NULL) to
146 a vector of UBSAN_NULL call statements that check this pointer. */
147 hash_map<tree, auto_vec<gimple *> > null_check_map;
149 /* This map maps a pointer (the second argument of ASAN_CHECK) to
150 a vector of ASAN_CHECK call statements that check the access. */
151 hash_map<tree_operand_hash, auto_vec<gimple *> > asan_check_map;
153 /* This map maps a tree triplet (the first, second and fourth argument
154 of UBSAN_VPTR) to a vector of UBSAN_VPTR call statements that check
155 that virtual table pointer. */
156 hash_map<sanopt_tree_triplet_hash, auto_vec<gimple *> > vptr_check_map;
158 /* Number of IFN_ASAN_CHECK statements. */
159 int asan_num_accesses;
163 /* Return true if there might be any call to free/munmap operation
164 on any path in between DOM (which should be imm(BB)) and BB. */
166 static bool
167 imm_dom_path_with_freeing_call (basic_block bb, basic_block dom)
169 sanopt_info *info = (sanopt_info *) bb->aux;
170 edge e;
171 edge_iterator ei;
173 if (info->imm_dom_path_with_freeing_call_computed_p)
174 return info->imm_dom_path_with_freeing_call_p;
176 info->being_visited_p = true;
178 FOR_EACH_EDGE (e, ei, bb->preds)
180 sanopt_info *pred_info = (sanopt_info *) e->src->aux;
182 if (e->src == dom)
183 continue;
185 if ((pred_info->imm_dom_path_with_freeing_call_computed_p
186 && pred_info->imm_dom_path_with_freeing_call_p)
187 || (pred_info->has_freeing_call_computed_p
188 && pred_info->has_freeing_call_p))
190 info->imm_dom_path_with_freeing_call_computed_p = true;
191 info->imm_dom_path_with_freeing_call_p = true;
192 info->being_visited_p = false;
193 return true;
197 FOR_EACH_EDGE (e, ei, bb->preds)
199 sanopt_info *pred_info = (sanopt_info *) e->src->aux;
201 if (e->src == dom)
202 continue;
204 if (pred_info->has_freeing_call_computed_p)
205 continue;
207 gimple_stmt_iterator gsi;
208 for (gsi = gsi_start_bb (e->src); !gsi_end_p (gsi); gsi_next (&gsi))
210 gimple *stmt = gsi_stmt (gsi);
212 if (is_gimple_call (stmt) && !nonfreeing_call_p (stmt))
214 pred_info->has_freeing_call_p = true;
215 break;
219 pred_info->has_freeing_call_computed_p = true;
220 if (pred_info->has_freeing_call_p)
222 info->imm_dom_path_with_freeing_call_computed_p = true;
223 info->imm_dom_path_with_freeing_call_p = true;
224 info->being_visited_p = false;
225 return true;
229 FOR_EACH_EDGE (e, ei, bb->preds)
231 if (e->src == dom)
232 continue;
234 basic_block src;
235 for (src = e->src; src != dom; )
237 sanopt_info *pred_info = (sanopt_info *) src->aux;
238 if (pred_info->being_visited_p)
239 break;
240 basic_block imm = get_immediate_dominator (CDI_DOMINATORS, src);
241 if (imm_dom_path_with_freeing_call (src, imm))
243 info->imm_dom_path_with_freeing_call_computed_p = true;
244 info->imm_dom_path_with_freeing_call_p = true;
245 info->being_visited_p = false;
246 return true;
248 src = imm;
252 info->imm_dom_path_with_freeing_call_computed_p = true;
253 info->imm_dom_path_with_freeing_call_p = false;
254 info->being_visited_p = false;
255 return false;
258 /* Get the first dominating check from the list of stored checks.
259 Non-dominating checks are silently dropped. */
261 static gimple *
262 maybe_get_dominating_check (auto_vec<gimple *> &v)
264 for (; !v.is_empty (); v.pop ())
266 gimple *g = v.last ();
267 sanopt_info *si = (sanopt_info *) gimple_bb (g)->aux;
268 if (!si->visited_p)
269 /* At this point we shouldn't have any statements
270 that aren't dominating the current BB. */
271 return g;
273 return NULL;
276 /* Optimize away redundant UBSAN_NULL calls. */
278 static bool
279 maybe_optimize_ubsan_null_ifn (struct sanopt_ctx *ctx, gimple *stmt)
281 gcc_assert (gimple_call_num_args (stmt) == 3);
282 tree ptr = gimple_call_arg (stmt, 0);
283 tree cur_align = gimple_call_arg (stmt, 2);
284 gcc_assert (TREE_CODE (cur_align) == INTEGER_CST);
285 bool remove = false;
287 auto_vec<gimple *> &v = ctx->null_check_map.get_or_insert (ptr);
288 gimple *g = maybe_get_dominating_check (v);
289 if (!g)
291 /* For this PTR we don't have any UBSAN_NULL stmts recorded, so there's
292 nothing to optimize yet. */
293 v.safe_push (stmt);
294 return false;
297 /* We already have recorded a UBSAN_NULL check for this pointer. Perhaps we
298 can drop this one. But only if this check doesn't specify stricter
299 alignment. */
301 tree align = gimple_call_arg (g, 2);
302 int kind = tree_to_shwi (gimple_call_arg (g, 1));
303 /* If this is a NULL pointer check where we had segv anyway, we can
304 remove it. */
305 if (integer_zerop (align)
306 && (kind == UBSAN_LOAD_OF
307 || kind == UBSAN_STORE_OF
308 || kind == UBSAN_MEMBER_ACCESS))
309 remove = true;
310 /* Otherwise remove the check in non-recovering mode, or if the
311 stmts have same location. */
312 else if (integer_zerop (align))
313 remove = (flag_sanitize_recover & SANITIZE_NULL) == 0
314 || flag_sanitize_undefined_trap_on_error
315 || gimple_location (g) == gimple_location (stmt);
316 else if (tree_int_cst_le (cur_align, align))
317 remove = (flag_sanitize_recover & SANITIZE_ALIGNMENT) == 0
318 || flag_sanitize_undefined_trap_on_error
319 || gimple_location (g) == gimple_location (stmt);
321 if (!remove && gimple_bb (g) == gimple_bb (stmt)
322 && tree_int_cst_compare (cur_align, align) == 0)
323 v.pop ();
325 if (!remove)
326 v.safe_push (stmt);
327 return remove;
330 /* Optimize away redundant UBSAN_VPTR calls. The second argument
331 is the value loaded from the virtual table, so rely on FRE to find out
332 when we can actually optimize. */
334 static bool
335 maybe_optimize_ubsan_vptr_ifn (struct sanopt_ctx *ctx, gimple *stmt)
337 gcc_assert (gimple_call_num_args (stmt) == 5);
338 sanopt_tree_triplet triplet;
339 triplet.t1 = gimple_call_arg (stmt, 0);
340 triplet.t2 = gimple_call_arg (stmt, 1);
341 triplet.t3 = gimple_call_arg (stmt, 3);
343 auto_vec<gimple *> &v = ctx->vptr_check_map.get_or_insert (triplet);
344 gimple *g = maybe_get_dominating_check (v);
345 if (!g)
347 /* For this PTR we don't have any UBSAN_VPTR stmts recorded, so there's
348 nothing to optimize yet. */
349 v.safe_push (stmt);
350 return false;
353 return true;
356 /* Returns TRUE if ASan check of length LEN in block BB can be removed
357 if preceded by checks in V. */
359 static bool
360 can_remove_asan_check (auto_vec<gimple *> &v, tree len, basic_block bb)
362 unsigned int i;
363 gimple *g;
364 gimple *to_pop = NULL;
365 bool remove = false;
366 basic_block last_bb = bb;
367 bool cleanup = false;
369 FOR_EACH_VEC_ELT_REVERSE (v, i, g)
371 basic_block gbb = gimple_bb (g);
372 sanopt_info *si = (sanopt_info *) gbb->aux;
373 if (gimple_uid (g) < si->freeing_call_events)
375 /* If there is a potentially freeing call after g in gbb, we should
376 remove it from the vector, can't use in optimization. */
377 cleanup = true;
378 continue;
381 tree glen = gimple_call_arg (g, 2);
382 gcc_assert (TREE_CODE (glen) == INTEGER_CST);
384 /* If we've checked only smaller length than we want to check now,
385 we can't remove the current stmt. If g is in the same basic block,
386 we want to remove it though, as the current stmt is better. */
387 if (tree_int_cst_lt (glen, len))
389 if (gbb == bb)
391 to_pop = g;
392 cleanup = true;
394 continue;
397 while (last_bb != gbb)
399 /* Paths from last_bb to bb have been checked before.
400 gbb is necessarily a dominator of last_bb, but not necessarily
401 immediate dominator. */
402 if (((sanopt_info *) last_bb->aux)->freeing_call_events)
403 break;
405 basic_block imm = get_immediate_dominator (CDI_DOMINATORS, last_bb);
406 gcc_assert (imm);
407 if (imm_dom_path_with_freeing_call (last_bb, imm))
408 break;
410 last_bb = imm;
412 if (last_bb == gbb)
413 remove = true;
414 break;
417 if (cleanup)
419 unsigned int j = 0, l = v.length ();
420 for (i = 0; i < l; i++)
421 if (v[i] != to_pop
422 && (gimple_uid (v[i])
423 == ((sanopt_info *)
424 gimple_bb (v[i])->aux)->freeing_call_events))
426 if (i != j)
427 v[j] = v[i];
428 j++;
430 v.truncate (j);
433 return remove;
436 /* Optimize away redundant ASAN_CHECK calls. */
438 static bool
439 maybe_optimize_asan_check_ifn (struct sanopt_ctx *ctx, gimple *stmt)
441 gcc_assert (gimple_call_num_args (stmt) == 4);
442 tree ptr = gimple_call_arg (stmt, 1);
443 tree len = gimple_call_arg (stmt, 2);
444 basic_block bb = gimple_bb (stmt);
445 sanopt_info *info = (sanopt_info *) bb->aux;
447 if (TREE_CODE (len) != INTEGER_CST)
448 return false;
449 if (integer_zerop (len))
450 return false;
452 gimple_set_uid (stmt, info->freeing_call_events);
454 auto_vec<gimple *> *ptr_checks = &ctx->asan_check_map.get_or_insert (ptr);
456 tree base_addr = maybe_get_single_definition (ptr);
457 auto_vec<gimple *> *base_checks = NULL;
458 if (base_addr)
460 base_checks = &ctx->asan_check_map.get_or_insert (base_addr);
461 /* Original pointer might have been invalidated. */
462 ptr_checks = ctx->asan_check_map.get (ptr);
465 gimple *g = maybe_get_dominating_check (*ptr_checks);
466 gimple *g2 = NULL;
468 if (base_checks)
469 /* Try with base address as well. */
470 g2 = maybe_get_dominating_check (*base_checks);
472 if (g == NULL && g2 == NULL)
474 /* For this PTR we don't have any ASAN_CHECK stmts recorded, so there's
475 nothing to optimize yet. */
476 ptr_checks->safe_push (stmt);
477 if (base_checks)
478 base_checks->safe_push (stmt);
479 return false;
482 bool remove = false;
484 if (ptr_checks)
485 remove = can_remove_asan_check (*ptr_checks, len, bb);
487 if (!remove && base_checks)
488 /* Try with base address as well. */
489 remove = can_remove_asan_check (*base_checks, len, bb);
491 if (!remove)
493 ptr_checks->safe_push (stmt);
494 if (base_checks)
495 base_checks->safe_push (stmt);
498 return remove;
501 /* Try to optimize away redundant UBSAN_NULL and ASAN_CHECK calls.
503 We walk blocks in the CFG via a depth first search of the dominator
504 tree; we push unique UBSAN_NULL or ASAN_CHECK statements into a vector
505 in the NULL_CHECK_MAP or ASAN_CHECK_MAP hash maps as we enter the
506 blocks. When leaving a block, we mark the block as visited; then
507 when checking the statements in the vector, we ignore statements that
508 are coming from already visited blocks, because these cannot dominate
509 anything anymore. CTX is a sanopt context. */
511 static void
512 sanopt_optimize_walker (basic_block bb, struct sanopt_ctx *ctx)
514 basic_block son;
515 gimple_stmt_iterator gsi;
516 sanopt_info *info = (sanopt_info *) bb->aux;
517 bool asan_check_optimize = (flag_sanitize & SANITIZE_ADDRESS) != 0;
519 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi);)
521 gimple *stmt = gsi_stmt (gsi);
522 bool remove = false;
524 if (!is_gimple_call (stmt))
526 /* Handle asm volatile or asm with "memory" clobber
527 the same as potentionally freeing call. */
528 gasm *asm_stmt = dyn_cast <gasm *> (stmt);
529 if (asm_stmt
530 && asan_check_optimize
531 && (gimple_asm_clobbers_memory_p (asm_stmt)
532 || gimple_asm_volatile_p (asm_stmt)))
533 info->freeing_call_events++;
534 gsi_next (&gsi);
535 continue;
538 if (asan_check_optimize && !nonfreeing_call_p (stmt))
539 info->freeing_call_events++;
541 if (gimple_call_internal_p (stmt))
542 switch (gimple_call_internal_fn (stmt))
544 case IFN_UBSAN_NULL:
545 remove = maybe_optimize_ubsan_null_ifn (ctx, stmt);
546 break;
547 case IFN_UBSAN_VPTR:
548 remove = maybe_optimize_ubsan_vptr_ifn (ctx, stmt);
549 break;
550 case IFN_ASAN_CHECK:
551 if (asan_check_optimize)
552 remove = maybe_optimize_asan_check_ifn (ctx, stmt);
553 if (!remove)
554 ctx->asan_num_accesses++;
555 break;
556 default:
557 break;
560 if (remove)
562 /* Drop this check. */
563 if (dump_file && (dump_flags & TDF_DETAILS))
565 fprintf (dump_file, "Optimizing out\n ");
566 print_gimple_stmt (dump_file, stmt, 0, dump_flags);
567 fprintf (dump_file, "\n");
569 unlink_stmt_vdef (stmt);
570 gsi_remove (&gsi, true);
572 else
573 gsi_next (&gsi);
576 if (asan_check_optimize)
578 info->has_freeing_call_p = info->freeing_call_events != 0;
579 info->has_freeing_call_computed_p = true;
582 for (son = first_dom_son (CDI_DOMINATORS, bb);
583 son;
584 son = next_dom_son (CDI_DOMINATORS, son))
585 sanopt_optimize_walker (son, ctx);
587 /* We're leaving this BB, so mark it to that effect. */
588 info->visited_p = true;
591 /* Try to remove redundant sanitizer checks in function FUN. */
593 static int
594 sanopt_optimize (function *fun)
596 struct sanopt_ctx ctx;
597 ctx.asan_num_accesses = 0;
599 /* Set up block info for each basic block. */
600 alloc_aux_for_blocks (sizeof (sanopt_info));
602 /* We're going to do a dominator walk, so ensure that we have
603 dominance information. */
604 calculate_dominance_info (CDI_DOMINATORS);
606 /* Recursively walk the dominator tree optimizing away
607 redundant checks. */
608 sanopt_optimize_walker (ENTRY_BLOCK_PTR_FOR_FN (fun), &ctx);
610 free_aux_for_blocks ();
612 return ctx.asan_num_accesses;
615 /* Perform optimization of sanitize functions. */
617 namespace {
619 const pass_data pass_data_sanopt =
621 GIMPLE_PASS, /* type */
622 "sanopt", /* name */
623 OPTGROUP_NONE, /* optinfo_flags */
624 TV_NONE, /* tv_id */
625 ( PROP_ssa | PROP_cfg | PROP_gimple_leh ), /* properties_required */
626 0, /* properties_provided */
627 0, /* properties_destroyed */
628 0, /* todo_flags_start */
629 TODO_update_ssa, /* todo_flags_finish */
632 class pass_sanopt : public gimple_opt_pass
634 public:
635 pass_sanopt (gcc::context *ctxt)
636 : gimple_opt_pass (pass_data_sanopt, ctxt)
639 /* opt_pass methods: */
640 virtual bool gate (function *) { return flag_sanitize; }
641 virtual unsigned int execute (function *);
643 }; // class pass_sanopt
645 unsigned int
646 pass_sanopt::execute (function *fun)
648 basic_block bb;
649 int asan_num_accesses = 0;
651 /* Try to remove redundant checks. */
652 if (optimize
653 && (flag_sanitize
654 & (SANITIZE_NULL | SANITIZE_ALIGNMENT
655 | SANITIZE_ADDRESS | SANITIZE_VPTR)))
656 asan_num_accesses = sanopt_optimize (fun);
657 else if (flag_sanitize & SANITIZE_ADDRESS)
659 gimple_stmt_iterator gsi;
660 FOR_EACH_BB_FN (bb, fun)
661 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
663 gimple *stmt = gsi_stmt (gsi);
664 if (is_gimple_call (stmt) && gimple_call_internal_p (stmt)
665 && gimple_call_internal_fn (stmt) == IFN_ASAN_CHECK)
666 ++asan_num_accesses;
670 bool use_calls = ASAN_INSTRUMENTATION_WITH_CALL_THRESHOLD < INT_MAX
671 && asan_num_accesses >= ASAN_INSTRUMENTATION_WITH_CALL_THRESHOLD;
673 FOR_EACH_BB_FN (bb, fun)
675 gimple_stmt_iterator gsi;
676 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); )
678 gimple *stmt = gsi_stmt (gsi);
679 bool no_next = false;
681 if (!is_gimple_call (stmt))
683 gsi_next (&gsi);
684 continue;
687 if (gimple_call_internal_p (stmt))
689 enum internal_fn ifn = gimple_call_internal_fn (stmt);
690 switch (ifn)
692 case IFN_UBSAN_NULL:
693 no_next = ubsan_expand_null_ifn (&gsi);
694 break;
695 case IFN_UBSAN_BOUNDS:
696 no_next = ubsan_expand_bounds_ifn (&gsi);
697 break;
698 case IFN_UBSAN_OBJECT_SIZE:
699 no_next = ubsan_expand_objsize_ifn (&gsi);
700 break;
701 case IFN_UBSAN_VPTR:
702 no_next = ubsan_expand_vptr_ifn (&gsi);
703 break;
704 case IFN_ASAN_CHECK:
705 no_next = asan_expand_check_ifn (&gsi, use_calls);
706 break;
707 default:
708 break;
711 else if (gimple_call_builtin_p (stmt, BUILT_IN_NORMAL))
713 tree callee = gimple_call_fndecl (stmt);
714 switch (DECL_FUNCTION_CODE (callee))
716 case BUILT_IN_UNREACHABLE:
717 if (flag_sanitize & SANITIZE_UNREACHABLE
718 && !lookup_attribute ("no_sanitize_undefined",
719 DECL_ATTRIBUTES (fun->decl)))
720 no_next = ubsan_instrument_unreachable (&gsi);
721 break;
722 default:
723 break;
727 if (dump_file && (dump_flags & TDF_DETAILS))
729 fprintf (dump_file, "Expanded\n ");
730 print_gimple_stmt (dump_file, stmt, 0, dump_flags);
731 fprintf (dump_file, "\n");
734 if (!no_next)
735 gsi_next (&gsi);
738 return 0;
741 } // anon namespace
743 gimple_opt_pass *
744 make_pass_sanopt (gcc::context *ctxt)
746 return new pass_sanopt (ctxt);