PR tree-optimization/86043 - strlen after memcpy partially overwriting a string not...
[official-gcc.git] / gcc / tree-ssa-strlen.c
blobeca88a56f72db73cc0e00f47fb83d585048d74f3
1 /* String length optimization
2 Copyright (C) 2011-2018 Free Software Foundation, Inc.
3 Contributed by Jakub Jelinek <jakub@redhat.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
10 any later version.
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License 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 "rtl.h"
26 #include "tree.h"
27 #include "gimple.h"
28 #include "alloc-pool.h"
29 #include "tree-pass.h"
30 #include "ssa.h"
31 #include "cgraph.h"
32 #include "gimple-pretty-print.h"
33 #include "gimple-ssa-warn-restrict.h"
34 #include "fold-const.h"
35 #include "stor-layout.h"
36 #include "gimple-fold.h"
37 #include "tree-eh.h"
38 #include "gimplify.h"
39 #include "gimple-iterator.h"
40 #include "gimplify-me.h"
41 #include "expr.h"
42 #include "tree-cfg.h"
43 #include "tree-dfa.h"
44 #include "domwalk.h"
45 #include "tree-ssa-alias.h"
46 #include "tree-ssa-propagate.h"
47 #include "tree-ssa-strlen.h"
48 #include "params.h"
49 #include "tree-hash-traits.h"
50 #include "tree-object-size.h"
51 #include "builtins.h"
52 #include "target.h"
53 #include "diagnostic-core.h"
54 #include "diagnostic.h"
55 #include "intl.h"
56 #include "attribs.h"
57 #include "calls.h"
59 /* A vector indexed by SSA_NAME_VERSION. 0 means unknown, positive value
60 is an index into strinfo vector, negative value stands for
61 string length of a string literal (~strlen). */
62 static vec<int> ssa_ver_to_stridx;
64 /* Number of currently active string indexes plus one. */
65 static int max_stridx;
67 /* String information record. */
68 struct strinfo
70 /* Number of leading characters that are known to be nonzero. This is
71 also the length of the string if FULL_STRING_P.
73 The values in a list of related string pointers must be consistent;
74 that is, if strinfo B comes X bytes after strinfo A, it must be
75 the case that A->nonzero_chars == X + B->nonzero_chars. */
76 tree nonzero_chars;
77 /* Any of the corresponding pointers for querying alias oracle. */
78 tree ptr;
79 /* This is used for two things:
81 - To record the statement that should be used for delayed length
82 computations. We maintain the invariant that all related strinfos
83 have delayed lengths or none do.
85 - To record the malloc or calloc call that produced this result. */
86 gimple *stmt;
87 /* Pointer to '\0' if known, if NULL, it can be computed as
88 ptr + length. */
89 tree endptr;
90 /* Reference count. Any changes to strinfo entry possibly shared
91 with dominating basic blocks need unshare_strinfo first, except
92 for dont_invalidate which affects only the immediately next
93 maybe_invalidate. */
94 int refcount;
95 /* Copy of index. get_strinfo (si->idx) should return si; */
96 int idx;
97 /* These 3 fields are for chaining related string pointers together.
98 E.g. for
99 bl = strlen (b); dl = strlen (d); strcpy (a, b); c = a + bl;
100 strcpy (c, d); e = c + dl;
101 strinfo(a) -> strinfo(c) -> strinfo(e)
102 All have ->first field equal to strinfo(a)->idx and are doubly
103 chained through prev/next fields. The later strinfos are required
104 to point into the same string with zero or more bytes after
105 the previous pointer and all bytes in between the two pointers
106 must be non-zero. Functions like strcpy or memcpy are supposed
107 to adjust all previous strinfo lengths, but not following strinfo
108 lengths (those are uncertain, usually invalidated during
109 maybe_invalidate, except when the alias oracle knows better).
110 Functions like strcat on the other side adjust the whole
111 related strinfo chain.
112 They are updated lazily, so to use the chain the same first fields
113 and si->prev->next == si->idx needs to be verified. */
114 int first;
115 int next;
116 int prev;
117 /* A flag whether the string is known to be written in the current
118 function. */
119 bool writable;
120 /* A flag for the next maybe_invalidate that this strinfo shouldn't
121 be invalidated. Always cleared by maybe_invalidate. */
122 bool dont_invalidate;
123 /* True if the string is known to be nul-terminated after NONZERO_CHARS
124 characters. False is useful when detecting strings that are built
125 up via successive memcpys. */
126 bool full_string_p;
129 /* Pool for allocating strinfo_struct entries. */
130 static object_allocator<strinfo> strinfo_pool ("strinfo pool");
132 /* Vector mapping positive string indexes to strinfo, for the
133 current basic block. The first pointer in the vector is special,
134 it is either NULL, meaning the vector isn't shared, or it is
135 a basic block pointer to the owner basic_block if shared.
136 If some other bb wants to modify the vector, the vector needs
137 to be unshared first, and only the owner bb is supposed to free it. */
138 static vec<strinfo *, va_heap, vl_embed> *stridx_to_strinfo;
140 /* One OFFSET->IDX mapping. */
141 struct stridxlist
143 struct stridxlist *next;
144 HOST_WIDE_INT offset;
145 int idx;
148 /* Hash table entry, mapping a DECL to a chain of OFFSET->IDX mappings. */
149 struct decl_stridxlist_map
151 struct tree_map_base base;
152 struct stridxlist list;
155 /* Hash table for mapping decls to a chained list of offset -> idx
156 mappings. */
157 static hash_map<tree_decl_hash, stridxlist> *decl_to_stridxlist_htab;
159 /* Hash table mapping strlen calls to stridx instances describing
160 the calls' arguments. Non-null only when warn_stringop_truncation
161 is non-zero. */
162 typedef std::pair<int, location_t> stridx_strlenloc;
163 static hash_map<tree, stridx_strlenloc> *strlen_to_stridx;
165 /* Obstack for struct stridxlist and struct decl_stridxlist_map. */
166 static struct obstack stridx_obstack;
168 /* Last memcpy statement if it could be adjusted if the trailing
169 '\0' written is immediately overwritten, or
170 *x = '\0' store that could be removed if it is immediately overwritten. */
171 struct laststmt_struct
173 gimple *stmt;
174 tree len;
175 int stridx;
176 } laststmt;
178 static int get_stridx_plus_constant (strinfo *, unsigned HOST_WIDE_INT, tree);
179 static void handle_builtin_stxncpy (built_in_function, gimple_stmt_iterator *);
181 /* Return:
183 - 1 if SI is known to start with more than OFF nonzero characters.
185 - 0 if SI is known to start with OFF nonzero characters,
186 but is not known to start with more.
188 - -1 if SI might not start with OFF nonzero characters. */
190 static inline int
191 compare_nonzero_chars (strinfo *si, unsigned HOST_WIDE_INT off)
193 if (si->nonzero_chars
194 && TREE_CODE (si->nonzero_chars) == INTEGER_CST)
195 return compare_tree_int (si->nonzero_chars, off);
196 else
197 return -1;
200 /* Return true if SI is known to be a zero-length string. */
202 static inline bool
203 zero_length_string_p (strinfo *si)
205 return si->full_string_p && integer_zerop (si->nonzero_chars);
208 /* Return strinfo vector entry IDX. */
210 static inline strinfo *
211 get_strinfo (int idx)
213 if (vec_safe_length (stridx_to_strinfo) <= (unsigned int) idx)
214 return NULL;
215 return (*stridx_to_strinfo)[idx];
218 /* Get the next strinfo in the chain after SI, or null if none. */
220 static inline strinfo *
221 get_next_strinfo (strinfo *si)
223 if (si->next == 0)
224 return NULL;
225 strinfo *nextsi = get_strinfo (si->next);
226 if (nextsi == NULL || nextsi->first != si->first || nextsi->prev != si->idx)
227 return NULL;
228 return nextsi;
231 /* Helper function for get_stridx. Return the strinfo index of the address
232 of EXP, which is available in PTR if nonnull. If OFFSET_OUT, it is
233 OK to return the index for some X <= &EXP and store &EXP - X in
234 *OFFSET_OUT. */
236 static int
237 get_addr_stridx (tree exp, tree ptr, unsigned HOST_WIDE_INT *offset_out)
239 HOST_WIDE_INT off;
240 struct stridxlist *list, *last = NULL;
241 tree base;
243 if (!decl_to_stridxlist_htab)
244 return 0;
246 poly_int64 poff;
247 base = get_addr_base_and_unit_offset (exp, &poff);
248 if (base == NULL || !DECL_P (base) || !poff.is_constant (&off))
249 return 0;
251 list = decl_to_stridxlist_htab->get (base);
252 if (list == NULL)
253 return 0;
257 if (list->offset == off)
259 if (offset_out)
260 *offset_out = 0;
261 return list->idx;
263 if (list->offset > off)
264 return 0;
265 last = list;
266 list = list->next;
268 while (list);
270 if ((offset_out || ptr) && last && last->idx > 0)
272 unsigned HOST_WIDE_INT rel_off
273 = (unsigned HOST_WIDE_INT) off - last->offset;
274 strinfo *si = get_strinfo (last->idx);
275 if (si && compare_nonzero_chars (si, rel_off) >= 0)
277 if (offset_out)
279 *offset_out = rel_off;
280 return last->idx;
282 else
283 return get_stridx_plus_constant (si, rel_off, ptr);
286 return 0;
289 /* Return string index for EXP. */
291 static int
292 get_stridx (tree exp)
294 tree s, o;
296 if (TREE_CODE (exp) == SSA_NAME)
298 if (ssa_ver_to_stridx[SSA_NAME_VERSION (exp)])
299 return ssa_ver_to_stridx[SSA_NAME_VERSION (exp)];
300 int i;
301 tree e = exp;
302 HOST_WIDE_INT off = 0;
303 for (i = 0; i < 5; i++)
305 gimple *def_stmt = SSA_NAME_DEF_STMT (e);
306 if (!is_gimple_assign (def_stmt)
307 || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR)
308 return 0;
309 tree rhs1 = gimple_assign_rhs1 (def_stmt);
310 tree rhs2 = gimple_assign_rhs2 (def_stmt);
311 if (TREE_CODE (rhs1) != SSA_NAME
312 || !tree_fits_shwi_p (rhs2))
313 return 0;
314 HOST_WIDE_INT this_off = tree_to_shwi (rhs2);
315 if (this_off < 0)
316 return 0;
317 off = (unsigned HOST_WIDE_INT) off + this_off;
318 if (off < 0)
319 return 0;
320 if (ssa_ver_to_stridx[SSA_NAME_VERSION (rhs1)])
322 strinfo *si
323 = get_strinfo (ssa_ver_to_stridx[SSA_NAME_VERSION (rhs1)]);
324 if (si && compare_nonzero_chars (si, off) >= 0)
325 return get_stridx_plus_constant (si, off, exp);
327 e = rhs1;
329 return 0;
332 if (TREE_CODE (exp) == ADDR_EXPR)
334 int idx = get_addr_stridx (TREE_OPERAND (exp, 0), exp, NULL);
335 if (idx != 0)
336 return idx;
339 s = string_constant (exp, &o);
340 if (s != NULL_TREE
341 && (o == NULL_TREE || tree_fits_shwi_p (o))
342 && TREE_STRING_LENGTH (s) > 0)
344 HOST_WIDE_INT offset = o ? tree_to_shwi (o) : 0;
345 const char *p = TREE_STRING_POINTER (s);
346 int max = TREE_STRING_LENGTH (s) - 1;
348 if (p[max] == '\0' && offset >= 0 && offset <= max)
349 return ~(int) strlen (p + offset);
351 return 0;
354 /* Return true if strinfo vector is shared with the immediate dominator. */
356 static inline bool
357 strinfo_shared (void)
359 return vec_safe_length (stridx_to_strinfo)
360 && (*stridx_to_strinfo)[0] != NULL;
363 /* Unshare strinfo vector that is shared with the immediate dominator. */
365 static void
366 unshare_strinfo_vec (void)
368 strinfo *si;
369 unsigned int i = 0;
371 gcc_assert (strinfo_shared ());
372 stridx_to_strinfo = vec_safe_copy (stridx_to_strinfo);
373 for (i = 1; vec_safe_iterate (stridx_to_strinfo, i, &si); ++i)
374 if (si != NULL)
375 si->refcount++;
376 (*stridx_to_strinfo)[0] = NULL;
379 /* Attempt to create a string index for exp, ADDR_EXPR's operand.
380 Return a pointer to the location where the string index can
381 be stored (if 0) or is stored, or NULL if this can't be tracked. */
383 static int *
384 addr_stridxptr (tree exp)
386 HOST_WIDE_INT off;
388 poly_int64 poff;
389 tree base = get_addr_base_and_unit_offset (exp, &poff);
390 if (base == NULL_TREE || !DECL_P (base) || !poff.is_constant (&off))
391 return NULL;
393 if (!decl_to_stridxlist_htab)
395 decl_to_stridxlist_htab
396 = new hash_map<tree_decl_hash, stridxlist> (64);
397 gcc_obstack_init (&stridx_obstack);
400 bool existed;
401 stridxlist *list = &decl_to_stridxlist_htab->get_or_insert (base, &existed);
402 if (existed)
404 int i;
405 stridxlist *before = NULL;
406 for (i = 0; i < 32; i++)
408 if (list->offset == off)
409 return &list->idx;
410 if (list->offset > off && before == NULL)
411 before = list;
412 if (list->next == NULL)
413 break;
414 list = list->next;
416 if (i == 32)
417 return NULL;
418 if (before)
420 list = before;
421 before = XOBNEW (&stridx_obstack, struct stridxlist);
422 *before = *list;
423 list->next = before;
424 list->offset = off;
425 list->idx = 0;
426 return &list->idx;
428 list->next = XOBNEW (&stridx_obstack, struct stridxlist);
429 list = list->next;
432 list->next = NULL;
433 list->offset = off;
434 list->idx = 0;
435 return &list->idx;
438 /* Create a new string index, or return 0 if reached limit. */
440 static int
441 new_stridx (tree exp)
443 int idx;
444 if (max_stridx >= PARAM_VALUE (PARAM_MAX_TRACKED_STRLENS))
445 return 0;
446 if (TREE_CODE (exp) == SSA_NAME)
448 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (exp))
449 return 0;
450 idx = max_stridx++;
451 ssa_ver_to_stridx[SSA_NAME_VERSION (exp)] = idx;
452 return idx;
454 if (TREE_CODE (exp) == ADDR_EXPR)
456 int *pidx = addr_stridxptr (TREE_OPERAND (exp, 0));
457 if (pidx != NULL)
459 gcc_assert (*pidx == 0);
460 *pidx = max_stridx++;
461 return *pidx;
464 return 0;
467 /* Like new_stridx, but for ADDR_EXPR's operand instead. */
469 static int
470 new_addr_stridx (tree exp)
472 int *pidx;
473 if (max_stridx >= PARAM_VALUE (PARAM_MAX_TRACKED_STRLENS))
474 return 0;
475 pidx = addr_stridxptr (exp);
476 if (pidx != NULL)
478 gcc_assert (*pidx == 0);
479 *pidx = max_stridx++;
480 return *pidx;
482 return 0;
485 /* Create a new strinfo. */
487 static strinfo *
488 new_strinfo (tree ptr, int idx, tree nonzero_chars, bool full_string_p)
490 strinfo *si = strinfo_pool.allocate ();
491 si->nonzero_chars = nonzero_chars;
492 si->ptr = ptr;
493 si->stmt = NULL;
494 si->endptr = NULL_TREE;
495 si->refcount = 1;
496 si->idx = idx;
497 si->first = 0;
498 si->prev = 0;
499 si->next = 0;
500 si->writable = false;
501 si->dont_invalidate = false;
502 si->full_string_p = full_string_p;
503 return si;
506 /* Decrease strinfo refcount and free it if not referenced anymore. */
508 static inline void
509 free_strinfo (strinfo *si)
511 if (si && --si->refcount == 0)
512 strinfo_pool.remove (si);
515 /* Set strinfo in the vector entry IDX to SI. */
517 static inline void
518 set_strinfo (int idx, strinfo *si)
520 if (vec_safe_length (stridx_to_strinfo) && (*stridx_to_strinfo)[0])
521 unshare_strinfo_vec ();
522 if (vec_safe_length (stridx_to_strinfo) <= (unsigned int) idx)
523 vec_safe_grow_cleared (stridx_to_strinfo, idx + 1);
524 (*stridx_to_strinfo)[idx] = si;
527 /* Return the first strinfo in the related strinfo chain
528 if all strinfos in between belong to the chain, otherwise NULL. */
530 static strinfo *
531 verify_related_strinfos (strinfo *origsi)
533 strinfo *si = origsi, *psi;
535 if (origsi->first == 0)
536 return NULL;
537 for (; si->prev; si = psi)
539 if (si->first != origsi->first)
540 return NULL;
541 psi = get_strinfo (si->prev);
542 if (psi == NULL)
543 return NULL;
544 if (psi->next != si->idx)
545 return NULL;
547 if (si->idx != si->first)
548 return NULL;
549 return si;
552 /* Set SI's endptr to ENDPTR and compute its length based on SI->ptr.
553 Use LOC for folding. */
555 static void
556 set_endptr_and_length (location_t loc, strinfo *si, tree endptr)
558 si->endptr = endptr;
559 si->stmt = NULL;
560 tree start_as_size = fold_convert_loc (loc, size_type_node, si->ptr);
561 tree end_as_size = fold_convert_loc (loc, size_type_node, endptr);
562 si->nonzero_chars = fold_build2_loc (loc, MINUS_EXPR, size_type_node,
563 end_as_size, start_as_size);
564 si->full_string_p = true;
567 /* Return string length, or NULL if it can't be computed. */
569 static tree
570 get_string_length (strinfo *si)
572 if (si->nonzero_chars)
573 return si->full_string_p ? si->nonzero_chars : NULL;
575 if (si->stmt)
577 gimple *stmt = si->stmt, *lenstmt;
578 tree callee, lhs, fn, tem;
579 location_t loc;
580 gimple_stmt_iterator gsi;
582 gcc_assert (is_gimple_call (stmt));
583 callee = gimple_call_fndecl (stmt);
584 gcc_assert (callee && DECL_BUILT_IN_CLASS (callee) == BUILT_IN_NORMAL);
585 lhs = gimple_call_lhs (stmt);
586 /* unshare_strinfo is intentionally not called here. The (delayed)
587 transformation of strcpy or strcat into stpcpy is done at the place
588 of the former strcpy/strcat call and so can affect all the strinfos
589 with the same stmt. If they were unshared before and transformation
590 has been already done, the handling of BUILT_IN_STPCPY{,_CHK} should
591 just compute the right length. */
592 switch (DECL_FUNCTION_CODE (callee))
594 case BUILT_IN_STRCAT:
595 case BUILT_IN_STRCAT_CHK:
596 gsi = gsi_for_stmt (stmt);
597 fn = builtin_decl_implicit (BUILT_IN_STRLEN);
598 gcc_assert (lhs == NULL_TREE);
599 tem = unshare_expr (gimple_call_arg (stmt, 0));
600 lenstmt = gimple_build_call (fn, 1, tem);
601 lhs = make_ssa_name (TREE_TYPE (TREE_TYPE (fn)), lenstmt);
602 gimple_call_set_lhs (lenstmt, lhs);
603 gimple_set_vuse (lenstmt, gimple_vuse (stmt));
604 gsi_insert_before (&gsi, lenstmt, GSI_SAME_STMT);
605 tem = gimple_call_arg (stmt, 0);
606 if (!ptrofftype_p (TREE_TYPE (lhs)))
608 lhs = convert_to_ptrofftype (lhs);
609 lhs = force_gimple_operand_gsi (&gsi, lhs, true, NULL_TREE,
610 true, GSI_SAME_STMT);
612 lenstmt = gimple_build_assign
613 (make_ssa_name (TREE_TYPE (gimple_call_arg (stmt, 0))),
614 POINTER_PLUS_EXPR,tem, lhs);
615 gsi_insert_before (&gsi, lenstmt, GSI_SAME_STMT);
616 gimple_call_set_arg (stmt, 0, gimple_assign_lhs (lenstmt));
617 lhs = NULL_TREE;
618 /* FALLTHRU */
619 case BUILT_IN_STRCPY:
620 case BUILT_IN_STRCPY_CHK:
621 gcc_assert (builtin_decl_implicit_p (BUILT_IN_STPCPY));
622 if (gimple_call_num_args (stmt) == 2)
623 fn = builtin_decl_implicit (BUILT_IN_STPCPY);
624 else
625 fn = builtin_decl_explicit (BUILT_IN_STPCPY_CHK);
626 gcc_assert (lhs == NULL_TREE);
627 if (dump_file && (dump_flags & TDF_DETAILS) != 0)
629 fprintf (dump_file, "Optimizing: ");
630 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
632 gimple_call_set_fndecl (stmt, fn);
633 lhs = make_ssa_name (TREE_TYPE (TREE_TYPE (fn)), stmt);
634 gimple_call_set_lhs (stmt, lhs);
635 update_stmt (stmt);
636 if (dump_file && (dump_flags & TDF_DETAILS) != 0)
638 fprintf (dump_file, "into: ");
639 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
641 /* FALLTHRU */
642 case BUILT_IN_STPCPY:
643 case BUILT_IN_STPCPY_CHK:
644 gcc_assert (lhs != NULL_TREE);
645 loc = gimple_location (stmt);
646 set_endptr_and_length (loc, si, lhs);
647 for (strinfo *chainsi = verify_related_strinfos (si);
648 chainsi != NULL;
649 chainsi = get_next_strinfo (chainsi))
650 if (chainsi->nonzero_chars == NULL)
651 set_endptr_and_length (loc, chainsi, lhs);
652 break;
653 case BUILT_IN_MALLOC:
654 break;
655 /* BUILT_IN_CALLOC always has si->nonzero_chars set. */
656 default:
657 gcc_unreachable ();
658 break;
662 return si->nonzero_chars;
665 /* Invalidate string length information for strings whose length
666 might change due to stores in stmt. */
668 static bool
669 maybe_invalidate (gimple *stmt)
671 strinfo *si;
672 unsigned int i;
673 bool nonempty = false;
675 for (i = 1; vec_safe_iterate (stridx_to_strinfo, i, &si); ++i)
676 if (si != NULL)
678 if (!si->dont_invalidate)
680 ao_ref r;
681 /* Do not use si->nonzero_chars. */
682 ao_ref_init_from_ptr_and_size (&r, si->ptr, NULL_TREE);
683 if (stmt_may_clobber_ref_p_1 (stmt, &r))
685 set_strinfo (i, NULL);
686 free_strinfo (si);
687 continue;
690 si->dont_invalidate = false;
691 nonempty = true;
693 return nonempty;
696 /* Unshare strinfo record SI, if it has refcount > 1 or
697 if stridx_to_strinfo vector is shared with some other
698 bbs. */
700 static strinfo *
701 unshare_strinfo (strinfo *si)
703 strinfo *nsi;
705 if (si->refcount == 1 && !strinfo_shared ())
706 return si;
708 nsi = new_strinfo (si->ptr, si->idx, si->nonzero_chars, si->full_string_p);
709 nsi->stmt = si->stmt;
710 nsi->endptr = si->endptr;
711 nsi->first = si->first;
712 nsi->prev = si->prev;
713 nsi->next = si->next;
714 nsi->writable = si->writable;
715 set_strinfo (si->idx, nsi);
716 free_strinfo (si);
717 return nsi;
720 /* Attempt to create a new strinfo for BASESI + OFF, or find existing
721 strinfo if there is any. Return it's idx, or 0 if no strinfo has
722 been created. */
724 static int
725 get_stridx_plus_constant (strinfo *basesi, unsigned HOST_WIDE_INT off,
726 tree ptr)
728 if (TREE_CODE (ptr) == SSA_NAME && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (ptr))
729 return 0;
731 if (compare_nonzero_chars (basesi, off) < 0
732 || !tree_fits_uhwi_p (basesi->nonzero_chars))
733 return 0;
735 unsigned HOST_WIDE_INT nonzero_chars
736 = tree_to_uhwi (basesi->nonzero_chars) - off;
737 strinfo *si = basesi, *chainsi;
738 if (si->first || si->prev || si->next)
739 si = verify_related_strinfos (basesi);
740 if (si == NULL
741 || si->nonzero_chars == NULL_TREE
742 || TREE_CODE (si->nonzero_chars) != INTEGER_CST)
743 return 0;
745 if (TREE_CODE (ptr) == SSA_NAME
746 && ssa_ver_to_stridx.length () <= SSA_NAME_VERSION (ptr))
747 ssa_ver_to_stridx.safe_grow_cleared (num_ssa_names);
749 gcc_checking_assert (compare_tree_int (si->nonzero_chars, off) != -1);
750 for (chainsi = si; chainsi->next; chainsi = si)
752 si = get_next_strinfo (chainsi);
753 if (si == NULL
754 || si->nonzero_chars == NULL_TREE
755 || TREE_CODE (si->nonzero_chars) != INTEGER_CST)
756 break;
757 int r = compare_tree_int (si->nonzero_chars, nonzero_chars);
758 if (r != 1)
760 if (r == 0)
762 if (TREE_CODE (ptr) == SSA_NAME)
763 ssa_ver_to_stridx[SSA_NAME_VERSION (ptr)] = si->idx;
764 else
766 int *pidx = addr_stridxptr (TREE_OPERAND (ptr, 0));
767 if (pidx != NULL && *pidx == 0)
768 *pidx = si->idx;
770 return si->idx;
772 break;
776 int idx = new_stridx (ptr);
777 if (idx == 0)
778 return 0;
779 si = new_strinfo (ptr, idx, build_int_cst (size_type_node, nonzero_chars),
780 basesi->full_string_p);
781 set_strinfo (idx, si);
782 if (strinfo *nextsi = get_strinfo (chainsi->next))
784 nextsi = unshare_strinfo (nextsi);
785 si->next = nextsi->idx;
786 nextsi->prev = idx;
788 chainsi = unshare_strinfo (chainsi);
789 if (chainsi->first == 0)
790 chainsi->first = chainsi->idx;
791 chainsi->next = idx;
792 if (chainsi->endptr == NULL_TREE && zero_length_string_p (si))
793 chainsi->endptr = ptr;
794 si->endptr = chainsi->endptr;
795 si->prev = chainsi->idx;
796 si->first = chainsi->first;
797 si->writable = chainsi->writable;
798 return si->idx;
801 /* Note that PTR, a pointer SSA_NAME initialized in the current stmt, points
802 to a zero-length string and if possible chain it to a related strinfo
803 chain whose part is or might be CHAINSI. */
805 static strinfo *
806 zero_length_string (tree ptr, strinfo *chainsi)
808 strinfo *si;
809 int idx;
810 if (ssa_ver_to_stridx.length () <= SSA_NAME_VERSION (ptr))
811 ssa_ver_to_stridx.safe_grow_cleared (num_ssa_names);
812 gcc_checking_assert (TREE_CODE (ptr) == SSA_NAME
813 && ssa_ver_to_stridx[SSA_NAME_VERSION (ptr)] == 0);
815 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (ptr))
816 return NULL;
817 if (chainsi != NULL)
819 si = verify_related_strinfos (chainsi);
820 if (si)
824 /* We shouldn't mix delayed and non-delayed lengths. */
825 gcc_assert (si->full_string_p);
826 if (si->endptr == NULL_TREE)
828 si = unshare_strinfo (si);
829 si->endptr = ptr;
831 chainsi = si;
832 si = get_next_strinfo (si);
834 while (si != NULL);
835 if (zero_length_string_p (chainsi))
837 if (chainsi->next)
839 chainsi = unshare_strinfo (chainsi);
840 chainsi->next = 0;
842 ssa_ver_to_stridx[SSA_NAME_VERSION (ptr)] = chainsi->idx;
843 return chainsi;
846 else
848 /* We shouldn't mix delayed and non-delayed lengths. */
849 gcc_assert (chainsi->full_string_p);
850 if (chainsi->first || chainsi->prev || chainsi->next)
852 chainsi = unshare_strinfo (chainsi);
853 chainsi->first = 0;
854 chainsi->prev = 0;
855 chainsi->next = 0;
859 idx = new_stridx (ptr);
860 if (idx == 0)
861 return NULL;
862 si = new_strinfo (ptr, idx, build_int_cst (size_type_node, 0), true);
863 set_strinfo (idx, si);
864 si->endptr = ptr;
865 if (chainsi != NULL)
867 chainsi = unshare_strinfo (chainsi);
868 if (chainsi->first == 0)
869 chainsi->first = chainsi->idx;
870 chainsi->next = idx;
871 if (chainsi->endptr == NULL_TREE)
872 chainsi->endptr = ptr;
873 si->prev = chainsi->idx;
874 si->first = chainsi->first;
875 si->writable = chainsi->writable;
877 return si;
880 /* For strinfo ORIGSI whose length has been just updated, adjust other
881 related strinfos so that they match the new ORIGSI. This involves:
883 - adding ADJ to the nonzero_chars fields
884 - copying full_string_p from the new ORIGSI. */
886 static void
887 adjust_related_strinfos (location_t loc, strinfo *origsi, tree adj)
889 strinfo *si = verify_related_strinfos (origsi);
891 if (si == NULL)
892 return;
894 while (1)
896 strinfo *nsi;
898 if (si != origsi)
900 tree tem;
902 si = unshare_strinfo (si);
903 /* We shouldn't see delayed lengths here; the caller must have
904 calculated the old length in order to calculate the
905 adjustment. */
906 gcc_assert (si->nonzero_chars);
907 tem = fold_convert_loc (loc, TREE_TYPE (si->nonzero_chars), adj);
908 si->nonzero_chars = fold_build2_loc (loc, PLUS_EXPR,
909 TREE_TYPE (si->nonzero_chars),
910 si->nonzero_chars, tem);
911 si->full_string_p = origsi->full_string_p;
913 si->endptr = NULL_TREE;
914 si->dont_invalidate = true;
916 nsi = get_next_strinfo (si);
917 if (nsi == NULL)
918 return;
919 si = nsi;
923 /* Find if there are other SSA_NAME pointers equal to PTR
924 for which we don't track their string lengths yet. If so, use
925 IDX for them. */
927 static void
928 find_equal_ptrs (tree ptr, int idx)
930 if (TREE_CODE (ptr) != SSA_NAME)
931 return;
932 while (1)
934 gimple *stmt = SSA_NAME_DEF_STMT (ptr);
935 if (!is_gimple_assign (stmt))
936 return;
937 ptr = gimple_assign_rhs1 (stmt);
938 switch (gimple_assign_rhs_code (stmt))
940 case SSA_NAME:
941 break;
942 CASE_CONVERT:
943 if (!POINTER_TYPE_P (TREE_TYPE (ptr)))
944 return;
945 if (TREE_CODE (ptr) == SSA_NAME)
946 break;
947 if (TREE_CODE (ptr) != ADDR_EXPR)
948 return;
949 /* FALLTHRU */
950 case ADDR_EXPR:
952 int *pidx = addr_stridxptr (TREE_OPERAND (ptr, 0));
953 if (pidx != NULL && *pidx == 0)
954 *pidx = idx;
955 return;
957 default:
958 return;
961 /* We might find an endptr created in this pass. Grow the
962 vector in that case. */
963 if (ssa_ver_to_stridx.length () <= SSA_NAME_VERSION (ptr))
964 ssa_ver_to_stridx.safe_grow_cleared (num_ssa_names);
966 if (ssa_ver_to_stridx[SSA_NAME_VERSION (ptr)] != 0)
967 return;
968 ssa_ver_to_stridx[SSA_NAME_VERSION (ptr)] = idx;
972 /* Return true if STMT is a call to a builtin function with the right
973 arguments and attributes that should be considered for optimization
974 by this pass. */
976 static bool
977 valid_builtin_call (gimple *stmt)
979 if (!gimple_call_builtin_p (stmt, BUILT_IN_NORMAL))
980 return false;
982 tree callee = gimple_call_fndecl (stmt);
983 switch (DECL_FUNCTION_CODE (callee))
985 case BUILT_IN_MEMCMP:
986 case BUILT_IN_MEMCMP_EQ:
987 case BUILT_IN_STRCHR:
988 case BUILT_IN_STRLEN:
989 /* The above functions should be pure. Punt if they aren't. */
990 if (gimple_vdef (stmt) || gimple_vuse (stmt) == NULL_TREE)
991 return false;
992 break;
994 case BUILT_IN_CALLOC:
995 case BUILT_IN_MALLOC:
996 case BUILT_IN_MEMCPY:
997 case BUILT_IN_MEMCPY_CHK:
998 case BUILT_IN_MEMPCPY:
999 case BUILT_IN_MEMPCPY_CHK:
1000 case BUILT_IN_MEMSET:
1001 case BUILT_IN_STPCPY:
1002 case BUILT_IN_STPCPY_CHK:
1003 case BUILT_IN_STRCAT:
1004 case BUILT_IN_STRCAT_CHK:
1005 case BUILT_IN_STRCPY:
1006 case BUILT_IN_STRCPY_CHK:
1007 /* The above functions should be neither const nor pure. Punt if they
1008 aren't. */
1009 if (gimple_vdef (stmt) == NULL_TREE || gimple_vuse (stmt) == NULL_TREE)
1010 return false;
1011 break;
1013 default:
1014 break;
1017 return true;
1020 /* If the last .MEM setter statement before STMT is
1021 memcpy (x, y, strlen (y) + 1), the only .MEM use of it is STMT
1022 and STMT is known to overwrite x[strlen (x)], adjust the last memcpy to
1023 just memcpy (x, y, strlen (y)). SI must be the zero length
1024 strinfo. */
1026 static void
1027 adjust_last_stmt (strinfo *si, gimple *stmt, bool is_strcat)
1029 tree vuse, callee, len;
1030 struct laststmt_struct last = laststmt;
1031 strinfo *lastsi, *firstsi;
1032 unsigned len_arg_no = 2;
1034 laststmt.stmt = NULL;
1035 laststmt.len = NULL_TREE;
1036 laststmt.stridx = 0;
1038 if (last.stmt == NULL)
1039 return;
1041 vuse = gimple_vuse (stmt);
1042 if (vuse == NULL_TREE
1043 || SSA_NAME_DEF_STMT (vuse) != last.stmt
1044 || !has_single_use (vuse))
1045 return;
1047 gcc_assert (last.stridx > 0);
1048 lastsi = get_strinfo (last.stridx);
1049 if (lastsi == NULL)
1050 return;
1052 if (lastsi != si)
1054 if (lastsi->first == 0 || lastsi->first != si->first)
1055 return;
1057 firstsi = verify_related_strinfos (si);
1058 if (firstsi == NULL)
1059 return;
1060 while (firstsi != lastsi)
1062 firstsi = get_next_strinfo (firstsi);
1063 if (firstsi == NULL)
1064 return;
1068 if (!is_strcat && !zero_length_string_p (si))
1069 return;
1071 if (is_gimple_assign (last.stmt))
1073 gimple_stmt_iterator gsi;
1075 if (!integer_zerop (gimple_assign_rhs1 (last.stmt)))
1076 return;
1077 if (stmt_could_throw_p (last.stmt))
1078 return;
1079 gsi = gsi_for_stmt (last.stmt);
1080 unlink_stmt_vdef (last.stmt);
1081 release_defs (last.stmt);
1082 gsi_remove (&gsi, true);
1083 return;
1086 if (!valid_builtin_call (last.stmt))
1087 return;
1089 callee = gimple_call_fndecl (last.stmt);
1090 switch (DECL_FUNCTION_CODE (callee))
1092 case BUILT_IN_MEMCPY:
1093 case BUILT_IN_MEMCPY_CHK:
1094 break;
1095 default:
1096 return;
1099 len = gimple_call_arg (last.stmt, len_arg_no);
1100 if (tree_fits_uhwi_p (len))
1102 if (!tree_fits_uhwi_p (last.len)
1103 || integer_zerop (len)
1104 || tree_to_uhwi (len) != tree_to_uhwi (last.len) + 1)
1105 return;
1106 /* Don't adjust the length if it is divisible by 4, it is more efficient
1107 to store the extra '\0' in that case. */
1108 if ((tree_to_uhwi (len) & 3) == 0)
1109 return;
1111 else if (TREE_CODE (len) == SSA_NAME)
1113 gimple *def_stmt = SSA_NAME_DEF_STMT (len);
1114 if (!is_gimple_assign (def_stmt)
1115 || gimple_assign_rhs_code (def_stmt) != PLUS_EXPR
1116 || gimple_assign_rhs1 (def_stmt) != last.len
1117 || !integer_onep (gimple_assign_rhs2 (def_stmt)))
1118 return;
1120 else
1121 return;
1123 gimple_call_set_arg (last.stmt, len_arg_no, last.len);
1124 update_stmt (last.stmt);
1127 /* For an LHS that is an SSA_NAME and for strlen() or strnlen() argument
1128 SRC, set LHS range info to [0, min (N, BOUND)] if SRC refers to
1129 a character array A[N] with unknown length bounded by N, and for
1130 strnlen(), by min (N, BOUND). */
1132 static tree
1133 maybe_set_strlen_range (tree lhs, tree src, tree bound)
1135 if (TREE_CODE (lhs) != SSA_NAME
1136 || !INTEGRAL_TYPE_P (TREE_TYPE (lhs)))
1137 return NULL_TREE;
1139 if (TREE_CODE (src) == SSA_NAME)
1141 gimple *def = SSA_NAME_DEF_STMT (src);
1142 if (is_gimple_assign (def)
1143 && gimple_assign_rhs_code (def) == ADDR_EXPR)
1144 src = gimple_assign_rhs1 (def);
1147 wide_int max = wi::to_wide (TYPE_MAX_VALUE (ptrdiff_type_node));
1148 wide_int min = wi::zero (max.get_precision ());
1150 if (TREE_CODE (src) == ADDR_EXPR)
1152 /* The last array member of a struct can be bigger than its size
1153 suggests if it's treated as a poor-man's flexible array member. */
1154 src = TREE_OPERAND (src, 0);
1155 bool src_is_array = TREE_CODE (TREE_TYPE (src)) == ARRAY_TYPE;
1156 if (src_is_array && !array_at_struct_end_p (src))
1158 tree type = TREE_TYPE (src);
1159 if (tree size = TYPE_SIZE_UNIT (type))
1160 if (size && TREE_CODE (size) == INTEGER_CST)
1161 max = wi::to_wide (size);
1163 /* For strlen() the upper bound above is equal to
1164 the longest string that can be stored in the array
1165 (i.e., it accounts for the terminating nul. For
1166 strnlen() bump up the maximum by one since the array
1167 need not be nul-terminated. */
1168 if (!bound && max != 0)
1169 --max;
1171 else
1173 if (TREE_CODE (src) == COMPONENT_REF && !src_is_array)
1174 src = TREE_OPERAND (src, 1);
1175 if (DECL_P (src))
1177 /* Handle the unlikely case of strlen (&c) where c is some
1178 variable. */
1179 if (tree size = DECL_SIZE_UNIT (src))
1180 if (TREE_CODE (size) == INTEGER_CST)
1181 max = wi::to_wide (size);
1186 if (bound)
1188 /* For strnlen, adjust MIN and MAX as necessary. If the bound
1189 is less than the size of the array set MAX to it. It it's
1190 greater than MAX and MAX is non-zero bump MAX down to account
1191 for the necessary terminating nul. Otherwise leave it alone. */
1192 if (TREE_CODE (bound) == INTEGER_CST)
1194 wide_int wibnd = wi::to_wide (bound);
1195 int cmp = wi::cmpu (wibnd, max);
1196 if (cmp < 0)
1197 max = wibnd;
1198 else if (cmp && wi::ne_p (max, min))
1199 --max;
1201 else if (TREE_CODE (bound) == SSA_NAME)
1203 wide_int minbound, maxbound;
1204 value_range_type rng = get_range_info (bound, &minbound, &maxbound);
1205 if (rng == VR_RANGE)
1207 /* For a bound in a known range, adjust the range determined
1208 above as necessary. For a bound in some anti-range or
1209 in an unknown range, use the range determined above. */
1210 if (wi::ltu_p (minbound, min))
1211 min = minbound;
1212 if (wi::ltu_p (maxbound, max))
1213 max = maxbound;
1218 if (min == max)
1219 return wide_int_to_tree (size_type_node, min);
1221 set_range_info (lhs, VR_RANGE, min, max);
1222 return lhs;
1225 /* Handle a strlen call. If strlen of the argument is known, replace
1226 the strlen call with the known value, otherwise remember that strlen
1227 of the argument is stored in the lhs SSA_NAME. */
1229 static void
1230 handle_builtin_strlen (gimple_stmt_iterator *gsi)
1232 gimple *stmt = gsi_stmt (*gsi);
1233 tree lhs = gimple_call_lhs (stmt);
1235 if (lhs == NULL_TREE)
1236 return;
1238 location_t loc = gimple_location (stmt);
1239 tree callee = gimple_call_fndecl (stmt);
1240 tree src = gimple_call_arg (stmt, 0);
1241 tree bound = (DECL_FUNCTION_CODE (callee) == BUILT_IN_STRNLEN
1242 ? gimple_call_arg (stmt, 1) : NULL_TREE);
1243 int idx = get_stridx (src);
1244 if (idx)
1246 strinfo *si = NULL;
1247 tree rhs;
1249 if (idx < 0)
1250 rhs = build_int_cst (TREE_TYPE (lhs), ~idx);
1251 else
1253 rhs = NULL_TREE;
1254 si = get_strinfo (idx);
1255 if (si != NULL)
1256 rhs = get_string_length (si);
1258 if (rhs != NULL_TREE)
1260 if (dump_file && (dump_flags & TDF_DETAILS) != 0)
1262 fprintf (dump_file, "Optimizing: ");
1263 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
1265 rhs = unshare_expr (rhs);
1266 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
1267 rhs = fold_convert_loc (loc, TREE_TYPE (lhs), rhs);
1269 /* Set for strnlen() calls with a non-constant bound. */
1270 bool noncst_bound = false;
1271 if (bound)
1273 tree new_rhs
1274 = fold_build2_loc (loc, MIN_EXPR, TREE_TYPE (rhs), rhs, bound);
1276 noncst_bound = (TREE_CODE (new_rhs) != INTEGER_CST
1277 || tree_int_cst_lt (new_rhs, rhs));
1279 rhs = new_rhs;
1282 if (!update_call_from_tree (gsi, rhs))
1283 gimplify_and_update_call_from_tree (gsi, rhs);
1284 stmt = gsi_stmt (*gsi);
1285 update_stmt (stmt);
1286 if (dump_file && (dump_flags & TDF_DETAILS) != 0)
1288 fprintf (dump_file, "into: ");
1289 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
1292 /* Avoid storing the length for calls to strnlen() with
1293 a non-constant bound. */
1294 if (noncst_bound)
1295 return;
1297 if (si != NULL
1298 && TREE_CODE (si->nonzero_chars) != SSA_NAME
1299 && TREE_CODE (si->nonzero_chars) != INTEGER_CST
1300 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
1302 si = unshare_strinfo (si);
1303 si->nonzero_chars = lhs;
1304 gcc_assert (si->full_string_p);
1307 if (strlen_to_stridx)
1308 strlen_to_stridx->put (lhs, stridx_strlenloc (idx, loc));
1310 return;
1313 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
1314 return;
1316 if (idx == 0)
1317 idx = new_stridx (src);
1318 else
1320 strinfo *si = get_strinfo (idx);
1321 if (si != NULL)
1323 if (!si->full_string_p && !si->stmt)
1325 /* Until now we only had a lower bound on the string length.
1326 Install LHS as the actual length. */
1327 si = unshare_strinfo (si);
1328 tree old = si->nonzero_chars;
1329 si->nonzero_chars = lhs;
1330 si->full_string_p = true;
1331 if (TREE_CODE (old) == INTEGER_CST)
1333 old = fold_convert_loc (loc, TREE_TYPE (lhs), old);
1334 tree adj = fold_build2_loc (loc, MINUS_EXPR,
1335 TREE_TYPE (lhs), lhs, old);
1336 adjust_related_strinfos (loc, si, adj);
1338 else
1340 si->first = 0;
1341 si->prev = 0;
1342 si->next = 0;
1345 return;
1348 if (idx)
1350 if (!bound)
1352 /* Only store the new length information for calls to strlen(),
1353 not for those to strnlen(). */
1354 strinfo *si = new_strinfo (src, idx, lhs, true);
1355 set_strinfo (idx, si);
1356 find_equal_ptrs (src, idx);
1359 /* For SRC that is an array of N elements, set LHS's range
1360 to [0, min (N, BOUND)]. A constant return value means
1361 the range would have consisted of a single value. In
1362 that case, fold the result into the returned constant. */
1363 if (tree ret = maybe_set_strlen_range (lhs, src, bound))
1364 if (TREE_CODE (ret) == INTEGER_CST)
1366 if (dump_file && (dump_flags & TDF_DETAILS) != 0)
1368 fprintf (dump_file, "Optimizing: ");
1369 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
1371 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (ret)))
1372 ret = fold_convert_loc (loc, TREE_TYPE (lhs), ret);
1373 if (!update_call_from_tree (gsi, ret))
1374 gimplify_and_update_call_from_tree (gsi, ret);
1375 stmt = gsi_stmt (*gsi);
1376 update_stmt (stmt);
1377 if (dump_file && (dump_flags & TDF_DETAILS) != 0)
1379 fprintf (dump_file, "into: ");
1380 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
1384 if (strlen_to_stridx && !bound)
1385 strlen_to_stridx->put (lhs, stridx_strlenloc (idx, loc));
1389 /* Handle a strchr call. If strlen of the first argument is known, replace
1390 the strchr (x, 0) call with the endptr or x + strlen, otherwise remember
1391 that lhs of the call is endptr and strlen of the argument is endptr - x. */
1393 static void
1394 handle_builtin_strchr (gimple_stmt_iterator *gsi)
1396 int idx;
1397 tree src;
1398 gimple *stmt = gsi_stmt (*gsi);
1399 tree lhs = gimple_call_lhs (stmt);
1401 if (lhs == NULL_TREE)
1402 return;
1404 if (!integer_zerop (gimple_call_arg (stmt, 1)))
1405 return;
1407 src = gimple_call_arg (stmt, 0);
1408 idx = get_stridx (src);
1409 if (idx)
1411 strinfo *si = NULL;
1412 tree rhs;
1414 if (idx < 0)
1415 rhs = build_int_cst (size_type_node, ~idx);
1416 else
1418 rhs = NULL_TREE;
1419 si = get_strinfo (idx);
1420 if (si != NULL)
1421 rhs = get_string_length (si);
1423 if (rhs != NULL_TREE)
1425 location_t loc = gimple_location (stmt);
1427 if (dump_file && (dump_flags & TDF_DETAILS) != 0)
1429 fprintf (dump_file, "Optimizing: ");
1430 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
1432 if (si != NULL && si->endptr != NULL_TREE)
1434 rhs = unshare_expr (si->endptr);
1435 if (!useless_type_conversion_p (TREE_TYPE (lhs),
1436 TREE_TYPE (rhs)))
1437 rhs = fold_convert_loc (loc, TREE_TYPE (lhs), rhs);
1439 else
1441 rhs = fold_convert_loc (loc, sizetype, unshare_expr (rhs));
1442 rhs = fold_build2_loc (loc, POINTER_PLUS_EXPR,
1443 TREE_TYPE (src), src, rhs);
1444 if (!useless_type_conversion_p (TREE_TYPE (lhs),
1445 TREE_TYPE (rhs)))
1446 rhs = fold_convert_loc (loc, TREE_TYPE (lhs), rhs);
1448 if (!update_call_from_tree (gsi, rhs))
1449 gimplify_and_update_call_from_tree (gsi, rhs);
1450 stmt = gsi_stmt (*gsi);
1451 update_stmt (stmt);
1452 if (dump_file && (dump_flags & TDF_DETAILS) != 0)
1454 fprintf (dump_file, "into: ");
1455 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
1457 if (si != NULL
1458 && si->endptr == NULL_TREE
1459 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
1461 si = unshare_strinfo (si);
1462 si->endptr = lhs;
1464 zero_length_string (lhs, si);
1465 return;
1468 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
1469 return;
1470 if (TREE_CODE (src) != SSA_NAME || !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (src))
1472 if (idx == 0)
1473 idx = new_stridx (src);
1474 else if (get_strinfo (idx) != NULL)
1476 zero_length_string (lhs, NULL);
1477 return;
1479 if (idx)
1481 location_t loc = gimple_location (stmt);
1482 tree lhsu = fold_convert_loc (loc, size_type_node, lhs);
1483 tree srcu = fold_convert_loc (loc, size_type_node, src);
1484 tree length = fold_build2_loc (loc, MINUS_EXPR,
1485 size_type_node, lhsu, srcu);
1486 strinfo *si = new_strinfo (src, idx, length, true);
1487 si->endptr = lhs;
1488 set_strinfo (idx, si);
1489 find_equal_ptrs (src, idx);
1490 zero_length_string (lhs, si);
1493 else
1494 zero_length_string (lhs, NULL);
1497 /* Handle a strcpy-like ({st{r,p}cpy,__st{r,p}cpy_chk}) call.
1498 If strlen of the second argument is known, strlen of the first argument
1499 is the same after this call. Furthermore, attempt to convert it to
1500 memcpy. */
1502 static void
1503 handle_builtin_strcpy (enum built_in_function bcode, gimple_stmt_iterator *gsi)
1505 int idx, didx;
1506 tree src, dst, srclen, len, lhs, type, fn, oldlen;
1507 bool success;
1508 gimple *stmt = gsi_stmt (*gsi);
1509 strinfo *si, *dsi, *olddsi, *zsi;
1510 location_t loc;
1512 src = gimple_call_arg (stmt, 1);
1513 dst = gimple_call_arg (stmt, 0);
1514 lhs = gimple_call_lhs (stmt);
1515 idx = get_stridx (src);
1516 si = NULL;
1517 if (idx > 0)
1518 si = get_strinfo (idx);
1520 didx = get_stridx (dst);
1521 olddsi = NULL;
1522 oldlen = NULL_TREE;
1523 if (didx > 0)
1524 olddsi = get_strinfo (didx);
1525 else if (didx < 0)
1526 return;
1528 if (olddsi != NULL)
1529 adjust_last_stmt (olddsi, stmt, false);
1531 srclen = NULL_TREE;
1532 if (si != NULL)
1533 srclen = get_string_length (si);
1534 else if (idx < 0)
1535 srclen = build_int_cst (size_type_node, ~idx);
1537 loc = gimple_location (stmt);
1538 if (srclen == NULL_TREE)
1539 switch (bcode)
1541 case BUILT_IN_STRCPY:
1542 case BUILT_IN_STRCPY_CHK:
1543 if (lhs != NULL_TREE || !builtin_decl_implicit_p (BUILT_IN_STPCPY))
1544 return;
1545 break;
1546 case BUILT_IN_STPCPY:
1547 case BUILT_IN_STPCPY_CHK:
1548 if (lhs == NULL_TREE)
1549 return;
1550 else
1552 tree lhsuint = fold_convert_loc (loc, size_type_node, lhs);
1553 srclen = fold_convert_loc (loc, size_type_node, dst);
1554 srclen = fold_build2_loc (loc, MINUS_EXPR, size_type_node,
1555 lhsuint, srclen);
1557 break;
1558 default:
1559 gcc_unreachable ();
1562 if (didx == 0)
1564 didx = new_stridx (dst);
1565 if (didx == 0)
1566 return;
1568 if (olddsi != NULL)
1570 oldlen = olddsi->nonzero_chars;
1571 dsi = unshare_strinfo (olddsi);
1572 dsi->nonzero_chars = srclen;
1573 dsi->full_string_p = (srclen != NULL_TREE);
1574 /* Break the chain, so adjust_related_strinfo on later pointers in
1575 the chain won't adjust this one anymore. */
1576 dsi->next = 0;
1577 dsi->stmt = NULL;
1578 dsi->endptr = NULL_TREE;
1580 else
1582 dsi = new_strinfo (dst, didx, srclen, srclen != NULL_TREE);
1583 set_strinfo (didx, dsi);
1584 find_equal_ptrs (dst, didx);
1586 dsi->writable = true;
1587 dsi->dont_invalidate = true;
1589 if (dsi->nonzero_chars == NULL_TREE)
1591 strinfo *chainsi;
1593 /* If string length of src is unknown, use delayed length
1594 computation. If string lenth of dst will be needed, it
1595 can be computed by transforming this strcpy call into
1596 stpcpy and subtracting dst from the return value. */
1598 /* Look for earlier strings whose length could be determined if
1599 this strcpy is turned into an stpcpy. */
1601 if (dsi->prev != 0 && (chainsi = verify_related_strinfos (dsi)) != NULL)
1603 for (; chainsi && chainsi != dsi; chainsi = get_strinfo (chainsi->next))
1605 /* When setting a stmt for delayed length computation
1606 prevent all strinfos through dsi from being
1607 invalidated. */
1608 chainsi = unshare_strinfo (chainsi);
1609 chainsi->stmt = stmt;
1610 chainsi->nonzero_chars = NULL_TREE;
1611 chainsi->full_string_p = false;
1612 chainsi->endptr = NULL_TREE;
1613 chainsi->dont_invalidate = true;
1616 dsi->stmt = stmt;
1618 /* Try to detect overlap before returning. This catches cases
1619 like strcpy (d, d + n) where n is non-constant whose range
1620 is such that (n <= strlen (d) holds).
1622 OLDDSI->NONZERO_chars may have been reset by this point with
1623 oldlen holding it original value. */
1624 if (olddsi && oldlen)
1626 /* Add 1 for the terminating NUL. */
1627 tree type = TREE_TYPE (oldlen);
1628 oldlen = fold_build2 (PLUS_EXPR, type, oldlen,
1629 build_int_cst (type, 1));
1630 check_bounds_or_overlap (as_a <gcall *>(stmt), olddsi->ptr, src,
1631 oldlen, NULL_TREE);
1634 return;
1637 if (olddsi != NULL)
1639 tree adj = NULL_TREE;
1640 if (oldlen == NULL_TREE)
1642 else if (integer_zerop (oldlen))
1643 adj = srclen;
1644 else if (TREE_CODE (oldlen) == INTEGER_CST
1645 || TREE_CODE (srclen) == INTEGER_CST)
1646 adj = fold_build2_loc (loc, MINUS_EXPR,
1647 TREE_TYPE (srclen), srclen,
1648 fold_convert_loc (loc, TREE_TYPE (srclen),
1649 oldlen));
1650 if (adj != NULL_TREE)
1651 adjust_related_strinfos (loc, dsi, adj);
1652 else
1653 dsi->prev = 0;
1655 /* strcpy src may not overlap dst, so src doesn't need to be
1656 invalidated either. */
1657 if (si != NULL)
1658 si->dont_invalidate = true;
1660 fn = NULL_TREE;
1661 zsi = NULL;
1662 switch (bcode)
1664 case BUILT_IN_STRCPY:
1665 fn = builtin_decl_implicit (BUILT_IN_MEMCPY);
1666 if (lhs)
1667 ssa_ver_to_stridx[SSA_NAME_VERSION (lhs)] = didx;
1668 break;
1669 case BUILT_IN_STRCPY_CHK:
1670 fn = builtin_decl_explicit (BUILT_IN_MEMCPY_CHK);
1671 if (lhs)
1672 ssa_ver_to_stridx[SSA_NAME_VERSION (lhs)] = didx;
1673 break;
1674 case BUILT_IN_STPCPY:
1675 /* This would need adjustment of the lhs (subtract one),
1676 or detection that the trailing '\0' doesn't need to be
1677 written, if it will be immediately overwritten.
1678 fn = builtin_decl_explicit (BUILT_IN_MEMPCPY); */
1679 if (lhs)
1681 dsi->endptr = lhs;
1682 zsi = zero_length_string (lhs, dsi);
1684 break;
1685 case BUILT_IN_STPCPY_CHK:
1686 /* This would need adjustment of the lhs (subtract one),
1687 or detection that the trailing '\0' doesn't need to be
1688 written, if it will be immediately overwritten.
1689 fn = builtin_decl_explicit (BUILT_IN_MEMPCPY_CHK); */
1690 if (lhs)
1692 dsi->endptr = lhs;
1693 zsi = zero_length_string (lhs, dsi);
1695 break;
1696 default:
1697 gcc_unreachable ();
1699 if (zsi != NULL)
1700 zsi->dont_invalidate = true;
1702 if (fn)
1704 tree args = TYPE_ARG_TYPES (TREE_TYPE (fn));
1705 type = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (args)));
1707 else
1708 type = size_type_node;
1710 len = fold_convert_loc (loc, type, unshare_expr (srclen));
1711 len = fold_build2_loc (loc, PLUS_EXPR, type, len, build_int_cst (type, 1));
1713 /* Set the no-warning bit on the transformed statement? */
1714 bool set_no_warning = false;
1716 if (const strinfo *chksi = olddsi ? olddsi : dsi)
1717 if (si
1718 && !check_bounds_or_overlap (as_a <gcall *>(stmt), chksi->ptr, si->ptr,
1719 NULL_TREE, len))
1721 gimple_set_no_warning (stmt, true);
1722 set_no_warning = true;
1725 if (fn == NULL_TREE)
1726 return;
1728 len = force_gimple_operand_gsi (gsi, len, true, NULL_TREE, true,
1729 GSI_SAME_STMT);
1730 if (dump_file && (dump_flags & TDF_DETAILS) != 0)
1732 fprintf (dump_file, "Optimizing: ");
1733 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
1735 if (gimple_call_num_args (stmt) == 2)
1736 success = update_gimple_call (gsi, fn, 3, dst, src, len);
1737 else
1738 success = update_gimple_call (gsi, fn, 4, dst, src, len,
1739 gimple_call_arg (stmt, 2));
1740 if (success)
1742 stmt = gsi_stmt (*gsi);
1743 update_stmt (stmt);
1744 if (dump_file && (dump_flags & TDF_DETAILS) != 0)
1746 fprintf (dump_file, "into: ");
1747 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
1749 /* Allow adjust_last_stmt to decrease this memcpy's size. */
1750 laststmt.stmt = stmt;
1751 laststmt.len = srclen;
1752 laststmt.stridx = dsi->idx;
1754 else if (dump_file && (dump_flags & TDF_DETAILS) != 0)
1755 fprintf (dump_file, "not possible.\n");
1757 if (set_no_warning)
1758 gimple_set_no_warning (stmt, true);
1761 /* Check the size argument to the built-in forms of stpncpy and strncpy
1762 for out-of-bounds offsets or overlapping access, and to see if the
1763 size argument is derived from a call to strlen() on the source argument,
1764 and if so, issue an appropriate warning. */
1766 static void
1767 handle_builtin_strncat (built_in_function bcode, gimple_stmt_iterator *gsi)
1769 /* Same as stxncpy(). */
1770 handle_builtin_stxncpy (bcode, gsi);
1773 /* Return true if LEN depends on a call to strlen(SRC) in an interesting
1774 way. LEN can either be an integer expression, or a pointer (to char).
1775 When it is the latter (such as in recursive calls to self) is is
1776 assumed to be the argument in some call to strlen() whose relationship
1777 to SRC is being ascertained. */
1779 bool
1780 is_strlen_related_p (tree src, tree len)
1782 if (TREE_CODE (TREE_TYPE (len)) == POINTER_TYPE
1783 && operand_equal_p (src, len, 0))
1784 return true;
1786 if (TREE_CODE (len) != SSA_NAME)
1787 return false;
1789 gimple *def_stmt = SSA_NAME_DEF_STMT (len);
1790 if (!def_stmt)
1791 return false;
1793 if (is_gimple_call (def_stmt))
1795 tree func = gimple_call_fndecl (def_stmt);
1796 if (!valid_builtin_call (def_stmt)
1797 || DECL_FUNCTION_CODE (func) != BUILT_IN_STRLEN)
1798 return false;
1800 tree arg = gimple_call_arg (def_stmt, 0);
1801 return is_strlen_related_p (src, arg);
1804 if (!is_gimple_assign (def_stmt))
1805 return false;
1807 tree_code code = gimple_assign_rhs_code (def_stmt);
1808 tree rhs1 = gimple_assign_rhs1 (def_stmt);
1809 tree rhstype = TREE_TYPE (rhs1);
1811 if ((TREE_CODE (rhstype) == POINTER_TYPE && code == POINTER_PLUS_EXPR)
1812 || (INTEGRAL_TYPE_P (rhstype)
1813 && (code == BIT_AND_EXPR
1814 || code == NOP_EXPR)))
1816 /* Pointer plus (an integer), and truncation are considered among
1817 the (potentially) related expressions to strlen. */
1818 return is_strlen_related_p (src, rhs1);
1821 if (tree rhs2 = gimple_assign_rhs2 (def_stmt))
1823 /* Integer subtraction is considered strlen-related when both
1824 arguments are integers and second one is strlen-related. */
1825 rhstype = TREE_TYPE (rhs2);
1826 if (INTEGRAL_TYPE_P (rhstype) && code == MINUS_EXPR)
1827 return is_strlen_related_p (src, rhs2);
1830 return false;
1833 /* Called by handle_builtin_stxncpy and by gimple_fold_builtin_strncpy
1834 in gimple-fold.c.
1835 Check to see if the specified bound is a) equal to the size of
1836 the destination DST and if so, b) if it's immediately followed by
1837 DST[CNT - 1] = '\0'. If a) holds and b) does not, warn. Otherwise,
1838 do nothing. Return true if diagnostic has been issued.
1840 The purpose is to diagnose calls to strncpy and stpncpy that do
1841 not nul-terminate the copy while allowing for the idiom where
1842 such a call is immediately followed by setting the last element
1843 to nul, as in:
1844 char a[32];
1845 strncpy (a, s, sizeof a);
1846 a[sizeof a - 1] = '\0';
1849 bool
1850 maybe_diag_stxncpy_trunc (gimple_stmt_iterator gsi, tree src, tree cnt)
1852 gimple *stmt = gsi_stmt (gsi);
1853 if (gimple_no_warning_p (stmt))
1854 return false;
1856 wide_int cntrange[2];
1858 if (TREE_CODE (cnt) == INTEGER_CST)
1859 cntrange[0] = cntrange[1] = wi::to_wide (cnt);
1860 else if (TREE_CODE (cnt) == SSA_NAME)
1862 enum value_range_type rng = get_range_info (cnt, cntrange, cntrange + 1);
1863 if (rng == VR_RANGE)
1865 else if (rng == VR_ANTI_RANGE)
1867 wide_int maxobjsize = wi::to_wide (TYPE_MAX_VALUE (ptrdiff_type_node));
1869 if (wi::ltu_p (cntrange[1], maxobjsize))
1871 cntrange[0] = cntrange[1] + 1;
1872 cntrange[1] = maxobjsize;
1874 else
1876 cntrange[1] = cntrange[0] - 1;
1877 cntrange[0] = wi::zero (TYPE_PRECISION (TREE_TYPE (cnt)));
1880 else
1881 return false;
1883 else
1884 return false;
1886 /* Negative value is the constant string length. If it's less than
1887 the lower bound there is no truncation. Avoid calling get_stridx()
1888 when ssa_ver_to_stridx is empty. That implies the caller isn't
1889 running under the control of this pass and ssa_ver_to_stridx hasn't
1890 been created yet. */
1891 int sidx = ssa_ver_to_stridx.length () ? get_stridx (src) : 0;
1892 if (sidx < 0 && wi::gtu_p (cntrange[0], ~sidx))
1893 return false;
1895 tree dst = gimple_call_arg (stmt, 0);
1896 tree dstdecl = dst;
1897 if (TREE_CODE (dstdecl) == ADDR_EXPR)
1898 dstdecl = TREE_OPERAND (dstdecl, 0);
1900 tree ref = NULL_TREE;
1902 if (!sidx)
1904 /* If the source is a non-string return early to avoid warning
1905 for possible truncation (if the truncation is certain SIDX
1906 is non-zero). */
1907 tree srcdecl = gimple_call_arg (stmt, 1);
1908 if (TREE_CODE (srcdecl) == ADDR_EXPR)
1909 srcdecl = TREE_OPERAND (srcdecl, 0);
1910 if (get_attr_nonstring_decl (srcdecl, &ref))
1911 return false;
1914 /* Likewise, if the destination refers to a an array/pointer declared
1915 nonstring return early. */
1916 if (get_attr_nonstring_decl (dstdecl, &ref))
1917 return false;
1919 /* Look for dst[i] = '\0'; after the stxncpy() call and if found
1920 avoid the truncation warning. */
1921 gsi_next_nondebug (&gsi);
1922 gimple *next_stmt = gsi_stmt (gsi);
1923 if (!next_stmt)
1925 /* When there is no statement in the same basic block check
1926 the immediate successor block. */
1927 if (basic_block bb = gimple_bb (stmt))
1929 if (single_succ_p (bb))
1931 /* For simplicity, ignore blocks with multiple outgoing
1932 edges for now and only consider successor blocks along
1933 normal edges. */
1934 edge e = EDGE_SUCC (bb, 0);
1935 if (!(e->flags & EDGE_ABNORMAL))
1937 gsi = gsi_start_bb (e->dest);
1938 next_stmt = gsi_stmt (gsi);
1939 if (next_stmt && is_gimple_debug (next_stmt))
1941 gsi_next_nondebug (&gsi);
1942 next_stmt = gsi_stmt (gsi);
1949 if (next_stmt && is_gimple_assign (next_stmt))
1951 tree lhs = gimple_assign_lhs (next_stmt);
1952 tree_code code = TREE_CODE (lhs);
1953 if (code == ARRAY_REF || code == MEM_REF)
1954 lhs = TREE_OPERAND (lhs, 0);
1956 tree func = gimple_call_fndecl (stmt);
1957 if (DECL_FUNCTION_CODE (func) == BUILT_IN_STPNCPY)
1959 tree ret = gimple_call_lhs (stmt);
1960 if (ret && operand_equal_p (ret, lhs, 0))
1961 return false;
1964 /* Determine the base address and offset of the reference,
1965 ignoring the innermost array index. */
1966 if (TREE_CODE (ref) == ARRAY_REF)
1967 ref = TREE_OPERAND (ref, 0);
1969 poly_int64 dstoff;
1970 tree dstbase = get_addr_base_and_unit_offset (ref, &dstoff);
1972 poly_int64 lhsoff;
1973 tree lhsbase = get_addr_base_and_unit_offset (lhs, &lhsoff);
1974 if (lhsbase
1975 && dstbase
1976 && known_eq (dstoff, lhsoff)
1977 && operand_equal_p (dstbase, lhsbase, 0))
1978 return false;
1981 int prec = TYPE_PRECISION (TREE_TYPE (cnt));
1982 wide_int lenrange[2];
1983 if (strinfo *sisrc = sidx > 0 ? get_strinfo (sidx) : NULL)
1985 lenrange[0] = (sisrc->nonzero_chars
1986 && TREE_CODE (sisrc->nonzero_chars) == INTEGER_CST
1987 ? wi::to_wide (sisrc->nonzero_chars)
1988 : wi::zero (prec));
1989 lenrange[1] = lenrange[0];
1991 else if (sidx < 0)
1992 lenrange[0] = lenrange[1] = wi::shwi (~sidx, prec);
1993 else
1995 tree range[2];
1996 get_range_strlen (src, range);
1997 if (range[0] != NULL_TREE
1998 && TREE_CODE (range[0]) == INTEGER_CST
1999 && range[1] != NULL_TREE
2000 && TREE_CODE (range[1]) == INTEGER_CST)
2002 lenrange[0] = wi::to_wide (range[0], prec);
2003 lenrange[1] = wi::to_wide (range[1], prec);
2005 else
2007 lenrange[0] = wi::shwi (0, prec);
2008 lenrange[1] = wi::shwi (-1, prec);
2012 location_t callloc = gimple_nonartificial_location (stmt);
2013 callloc = expansion_point_location_if_in_system_header (callloc);
2015 tree func = gimple_call_fndecl (stmt);
2017 if (lenrange[0] != 0 || !wi::neg_p (lenrange[1]))
2019 /* If the longest source string is shorter than the lower bound
2020 of the specified count the copy is definitely nul-terminated. */
2021 if (wi::ltu_p (lenrange[1], cntrange[0]))
2022 return false;
2024 if (wi::neg_p (lenrange[1]))
2026 /* The length of one of the strings is unknown but at least
2027 one has non-zero length and that length is stored in
2028 LENRANGE[1]. Swap the bounds to force a "may be truncated"
2029 warning below. */
2030 lenrange[1] = lenrange[0];
2031 lenrange[0] = wi::shwi (0, prec);
2034 gcall *call = as_a <gcall *> (stmt);
2036 /* Set to true for strncat whose bound is derived from the length
2037 of the destination (the expected usage pattern). */
2038 bool cat_dstlen_bounded = false;
2039 if (DECL_FUNCTION_CODE (func) == BUILT_IN_STRNCAT)
2040 cat_dstlen_bounded = is_strlen_related_p (dst, cnt);
2042 if (lenrange[0] == cntrange[1] && cntrange[0] == cntrange[1])
2043 return warning_n (callloc, OPT_Wstringop_truncation,
2044 cntrange[0].to_uhwi (),
2045 "%G%qD output truncated before terminating "
2046 "nul copying %E byte from a string of the "
2047 "same length",
2048 "%G%qD output truncated before terminating nul "
2049 "copying %E bytes from a string of the same "
2050 "length",
2051 call, func, cnt);
2052 else if (!cat_dstlen_bounded)
2054 if (wi::geu_p (lenrange[0], cntrange[1]))
2056 /* The shortest string is longer than the upper bound of
2057 the count so the truncation is certain. */
2058 if (cntrange[0] == cntrange[1])
2059 return warning_n (callloc, OPT_Wstringop_truncation,
2060 cntrange[0].to_uhwi (),
2061 "%G%qD output truncated copying %E byte "
2062 "from a string of length %wu",
2063 "%G%qD output truncated copying %E bytes "
2064 "from a string of length %wu",
2065 call, func, cnt, lenrange[0].to_uhwi ());
2067 return warning_at (callloc, OPT_Wstringop_truncation,
2068 "%G%qD output truncated copying between %wu "
2069 "and %wu bytes from a string of length %wu",
2070 call, func, cntrange[0].to_uhwi (),
2071 cntrange[1].to_uhwi (), lenrange[0].to_uhwi ());
2073 else if (wi::geu_p (lenrange[1], cntrange[1]))
2075 /* The longest string is longer than the upper bound of
2076 the count so the truncation is possible. */
2077 if (cntrange[0] == cntrange[1])
2078 return warning_n (callloc, OPT_Wstringop_truncation,
2079 cntrange[0].to_uhwi (),
2080 "%G%qD output may be truncated copying %E "
2081 "byte from a string of length %wu",
2082 "%G%qD output may be truncated copying %E "
2083 "bytes from a string of length %wu",
2084 call, func, cnt, lenrange[1].to_uhwi ());
2086 return warning_at (callloc, OPT_Wstringop_truncation,
2087 "%G%qD output may be truncated copying between "
2088 "%wu and %wu bytes from a string of length %wu",
2089 call, func, cntrange[0].to_uhwi (),
2090 cntrange[1].to_uhwi (), lenrange[1].to_uhwi ());
2094 if (!cat_dstlen_bounded
2095 && cntrange[0] != cntrange[1]
2096 && wi::leu_p (cntrange[0], lenrange[0])
2097 && wi::leu_p (cntrange[1], lenrange[0] + 1))
2099 /* If the source (including the terminating nul) is longer than
2100 the lower bound of the specified count but shorter than the
2101 upper bound the copy may (but need not) be truncated. */
2102 return warning_at (callloc, OPT_Wstringop_truncation,
2103 "%G%qD output may be truncated copying between "
2104 "%wu and %wu bytes from a string of length %wu",
2105 call, func, cntrange[0].to_uhwi (),
2106 cntrange[1].to_uhwi (), lenrange[0].to_uhwi ());
2110 if (tree dstsize = compute_objsize (dst, 1))
2112 /* The source length is uknown. Try to determine the destination
2113 size and see if it matches the specified bound. If not, bail.
2114 Otherwise go on to see if it should be diagnosed for possible
2115 truncation. */
2116 if (!dstsize)
2117 return false;
2119 if (wi::to_wide (dstsize) != cntrange[1])
2120 return false;
2122 if (cntrange[0] == cntrange[1])
2123 return warning_at (callloc, OPT_Wstringop_truncation,
2124 "%G%qD specified bound %E equals destination size",
2125 as_a <gcall *> (stmt), func, cnt);
2128 return false;
2131 /* Check the arguments to the built-in forms of stpncpy and strncpy for
2132 out-of-bounds offsets or overlapping access, and to see if the size
2133 is derived from calling strlen() on the source argument, and if so,
2134 issue the appropriate warning. */
2136 static void
2137 handle_builtin_stxncpy (built_in_function, gimple_stmt_iterator *gsi)
2139 if (!strlen_to_stridx)
2140 return;
2142 gimple *stmt = gsi_stmt (*gsi);
2144 tree dst = gimple_call_arg (stmt, 0);
2145 tree src = gimple_call_arg (stmt, 1);
2146 tree len = gimple_call_arg (stmt, 2);
2147 tree dstsize = NULL_TREE, srcsize = NULL_TREE;
2149 int didx = get_stridx (dst);
2150 if (strinfo *sidst = didx > 0 ? get_strinfo (didx) : NULL)
2152 /* Compute the size of the destination string including the NUL. */
2153 if (sidst->nonzero_chars)
2155 tree type = TREE_TYPE (sidst->nonzero_chars);
2156 dstsize = fold_build2 (PLUS_EXPR, type, sidst->nonzero_chars,
2157 build_int_cst (type, 1));
2159 dst = sidst->ptr;
2162 int sidx = get_stridx (src);
2163 strinfo *sisrc = sidx > 0 ? get_strinfo (sidx) : NULL;
2164 if (sisrc)
2166 /* strncat() and strncpy() can modify the source string by writing
2167 over the terminating nul so SISRC->DONT_INVALIDATE must be left
2168 clear. */
2170 /* Compute the size of the source string including the NUL. */
2171 if (sisrc->nonzero_chars)
2173 tree type = TREE_TYPE (sisrc->nonzero_chars);
2174 srcsize = fold_build2 (PLUS_EXPR, type, sisrc->nonzero_chars,
2175 build_int_cst (type, 1));
2178 src = sisrc->ptr;
2180 else
2181 srcsize = NULL_TREE;
2183 if (!check_bounds_or_overlap (as_a <gcall *>(stmt), dst, src,
2184 dstsize, srcsize))
2186 gimple_set_no_warning (stmt, true);
2187 return;
2190 /* If the length argument was computed from strlen(S) for some string
2191 S retrieve the strinfo index for the string (PSS->FIRST) alonng with
2192 the location of the strlen() call (PSS->SECOND). */
2193 stridx_strlenloc *pss = strlen_to_stridx->get (len);
2194 if (!pss || pss->first <= 0)
2196 if (maybe_diag_stxncpy_trunc (*gsi, src, len))
2197 gimple_set_no_warning (stmt, true);
2199 return;
2202 /* Retrieve the strinfo data for the string S that LEN was computed
2203 from as some function F of strlen (S) (i.e., LEN need not be equal
2204 to strlen(S)). */
2205 strinfo *silen = get_strinfo (pss->first);
2207 location_t callloc = gimple_nonartificial_location (stmt);
2208 callloc = expansion_point_location_if_in_system_header (callloc);
2210 tree func = gimple_call_fndecl (stmt);
2212 bool warned = false;
2214 /* When -Wstringop-truncation is set, try to determine truncation
2215 before diagnosing possible overflow. Truncation is implied by
2216 the LEN argument being equal to strlen(SRC), regardless of
2217 whether its value is known. Otherwise, issue the more generic
2218 -Wstringop-overflow which triggers for LEN arguments that in
2219 any meaningful way depend on strlen(SRC). */
2220 if (sisrc == silen
2221 && is_strlen_related_p (src, len)
2222 && warning_at (callloc, OPT_Wstringop_truncation,
2223 "%G%qD output truncated before terminating nul "
2224 "copying as many bytes from a string as its length",
2225 as_a <gcall *>(stmt), func))
2226 warned = true;
2227 else if (silen && is_strlen_related_p (src, silen->ptr))
2228 warned = warning_at (callloc, OPT_Wstringop_overflow_,
2229 "%G%qD specified bound depends on the length "
2230 "of the source argument",
2231 as_a <gcall *>(stmt), func);
2232 if (warned)
2234 location_t strlenloc = pss->second;
2235 if (strlenloc != UNKNOWN_LOCATION && strlenloc != callloc)
2236 inform (strlenloc, "length computed here");
2240 /* Handle a memcpy-like ({mem{,p}cpy,__mem{,p}cpy_chk}) call.
2241 If strlen of the second argument is known and length of the third argument
2242 is that plus one, strlen of the first argument is the same after this
2243 call. */
2245 static void
2246 handle_builtin_memcpy (enum built_in_function bcode, gimple_stmt_iterator *gsi)
2248 int idx, didx;
2249 tree src, dst, len, lhs, oldlen, newlen;
2250 gimple *stmt = gsi_stmt (*gsi);
2251 strinfo *si, *dsi, *olddsi;
2253 len = gimple_call_arg (stmt, 2);
2254 src = gimple_call_arg (stmt, 1);
2255 dst = gimple_call_arg (stmt, 0);
2256 idx = get_stridx (src);
2257 if (idx == 0)
2258 return;
2260 didx = get_stridx (dst);
2261 olddsi = NULL;
2262 if (didx > 0)
2263 olddsi = get_strinfo (didx);
2264 else if (didx < 0)
2265 return;
2267 if (olddsi != NULL
2268 && tree_fits_uhwi_p (len)
2269 && !integer_zerop (len))
2270 adjust_last_stmt (olddsi, stmt, false);
2272 bool full_string_p;
2273 if (idx > 0)
2275 gimple *def_stmt;
2277 /* Handle memcpy (x, y, l) where l's relationship with strlen (y)
2278 is known. */
2279 si = get_strinfo (idx);
2280 if (si == NULL || si->nonzero_chars == NULL_TREE)
2281 return;
2282 if (TREE_CODE (len) == INTEGER_CST
2283 && TREE_CODE (si->nonzero_chars) == INTEGER_CST)
2285 if (tree_int_cst_le (len, si->nonzero_chars))
2287 /* Copying LEN nonzero characters, where LEN is constant. */
2288 newlen = len;
2289 full_string_p = false;
2291 else
2293 /* Copying the whole of the analyzed part of SI. */
2294 newlen = si->nonzero_chars;
2295 full_string_p = si->full_string_p;
2298 else
2300 if (!si->full_string_p)
2301 return;
2302 if (TREE_CODE (len) != SSA_NAME)
2303 return;
2304 def_stmt = SSA_NAME_DEF_STMT (len);
2305 if (!is_gimple_assign (def_stmt)
2306 || gimple_assign_rhs_code (def_stmt) != PLUS_EXPR
2307 || gimple_assign_rhs1 (def_stmt) != si->nonzero_chars
2308 || !integer_onep (gimple_assign_rhs2 (def_stmt)))
2309 return;
2310 /* Copying variable-length string SI (and no more). */
2311 newlen = si->nonzero_chars;
2312 full_string_p = true;
2315 else
2317 si = NULL;
2318 /* Handle memcpy (x, "abcd", 5) or
2319 memcpy (x, "abc\0uvw", 7). */
2320 if (!tree_fits_uhwi_p (len))
2321 return;
2323 unsigned HOST_WIDE_INT clen = tree_to_uhwi (len);
2324 unsigned HOST_WIDE_INT nonzero_chars = ~idx;
2325 newlen = build_int_cst (size_type_node, MIN (nonzero_chars, clen));
2326 full_string_p = clen > nonzero_chars;
2329 if (!full_string_p
2330 && olddsi
2331 && olddsi->nonzero_chars
2332 && TREE_CODE (olddsi->nonzero_chars) == INTEGER_CST
2333 && tree_int_cst_le (newlen, olddsi->nonzero_chars))
2335 /* The SRC substring being written strictly overlaps
2336 a subsequence of the existing string OLDDSI. */
2337 newlen = olddsi->nonzero_chars;
2338 full_string_p = olddsi->full_string_p;
2341 if (olddsi != NULL && TREE_CODE (len) == SSA_NAME)
2342 adjust_last_stmt (olddsi, stmt, false);
2344 if (didx == 0)
2346 didx = new_stridx (dst);
2347 if (didx == 0)
2348 return;
2350 oldlen = NULL_TREE;
2351 if (olddsi != NULL)
2353 dsi = unshare_strinfo (olddsi);
2354 oldlen = olddsi->nonzero_chars;
2355 dsi->nonzero_chars = newlen;
2356 dsi->full_string_p = full_string_p;
2357 /* Break the chain, so adjust_related_strinfo on later pointers in
2358 the chain won't adjust this one anymore. */
2359 dsi->next = 0;
2360 dsi->stmt = NULL;
2361 dsi->endptr = NULL_TREE;
2363 else
2365 dsi = new_strinfo (dst, didx, newlen, full_string_p);
2366 set_strinfo (didx, dsi);
2367 find_equal_ptrs (dst, didx);
2369 dsi->writable = true;
2370 dsi->dont_invalidate = true;
2371 if (olddsi != NULL)
2373 tree adj = NULL_TREE;
2374 location_t loc = gimple_location (stmt);
2375 if (oldlen == NULL_TREE)
2377 else if (integer_zerop (oldlen))
2378 adj = newlen;
2379 else if (TREE_CODE (oldlen) == INTEGER_CST
2380 || TREE_CODE (newlen) == INTEGER_CST)
2381 adj = fold_build2_loc (loc, MINUS_EXPR, TREE_TYPE (newlen), newlen,
2382 fold_convert_loc (loc, TREE_TYPE (newlen),
2383 oldlen));
2384 if (adj != NULL_TREE)
2385 adjust_related_strinfos (loc, dsi, adj);
2386 else
2387 dsi->prev = 0;
2389 /* memcpy src may not overlap dst, so src doesn't need to be
2390 invalidated either. */
2391 if (si != NULL)
2392 si->dont_invalidate = true;
2394 if (full_string_p)
2396 lhs = gimple_call_lhs (stmt);
2397 switch (bcode)
2399 case BUILT_IN_MEMCPY:
2400 case BUILT_IN_MEMCPY_CHK:
2401 /* Allow adjust_last_stmt to decrease this memcpy's size. */
2402 laststmt.stmt = stmt;
2403 laststmt.len = dsi->nonzero_chars;
2404 laststmt.stridx = dsi->idx;
2405 if (lhs)
2406 ssa_ver_to_stridx[SSA_NAME_VERSION (lhs)] = didx;
2407 break;
2408 case BUILT_IN_MEMPCPY:
2409 case BUILT_IN_MEMPCPY_CHK:
2410 break;
2411 default:
2412 gcc_unreachable ();
2417 /* Handle a strcat-like ({strcat,__strcat_chk}) call.
2418 If strlen of the second argument is known, strlen of the first argument
2419 is increased by the length of the second argument. Furthermore, attempt
2420 to convert it to memcpy/strcpy if the length of the first argument
2421 is known. */
2423 static void
2424 handle_builtin_strcat (enum built_in_function bcode, gimple_stmt_iterator *gsi)
2426 int idx, didx;
2427 tree srclen, args, type, fn, objsz, endptr;
2428 bool success;
2429 gimple *stmt = gsi_stmt (*gsi);
2430 strinfo *si, *dsi;
2431 location_t loc = gimple_location (stmt);
2433 tree src = gimple_call_arg (stmt, 1);
2434 tree dst = gimple_call_arg (stmt, 0);
2436 /* Bail if the source is the same as destination. It will be diagnosed
2437 elsewhere. */
2438 if (operand_equal_p (src, dst, 0))
2439 return;
2441 tree lhs = gimple_call_lhs (stmt);
2443 didx = get_stridx (dst);
2444 if (didx < 0)
2445 return;
2447 dsi = NULL;
2448 if (didx > 0)
2449 dsi = get_strinfo (didx);
2451 srclen = NULL_TREE;
2452 si = NULL;
2453 idx = get_stridx (src);
2454 if (idx < 0)
2455 srclen = build_int_cst (size_type_node, ~idx);
2456 else if (idx > 0)
2458 si = get_strinfo (idx);
2459 if (si != NULL)
2460 srclen = get_string_length (si);
2463 /* Set the no-warning bit on the transformed statement? */
2464 bool set_no_warning = false;
2466 if (dsi == NULL || get_string_length (dsi) == NULL_TREE)
2469 /* The concatenation always involves copying at least one byte
2470 (the terminating nul), even if the source string is empty.
2471 If the source is unknown assume it's one character long and
2472 used that as both sizes. */
2473 tree slen = srclen;
2474 if (slen)
2476 tree type = TREE_TYPE (slen);
2477 slen = fold_build2 (PLUS_EXPR, type, slen, build_int_cst (type, 1));
2480 tree sptr = si && si->ptr ? si->ptr : src;
2482 if (!check_bounds_or_overlap (as_a <gcall *>(stmt), dst, sptr,
2483 NULL_TREE, slen))
2485 gimple_set_no_warning (stmt, true);
2486 set_no_warning = true;
2490 /* strcat (p, q) can be transformed into
2491 tmp = p + strlen (p); endptr = stpcpy (tmp, q);
2492 with length endptr - p if we need to compute the length
2493 later on. Don't do this transformation if we don't need
2494 it. */
2495 if (builtin_decl_implicit_p (BUILT_IN_STPCPY) && lhs == NULL_TREE)
2497 if (didx == 0)
2499 didx = new_stridx (dst);
2500 if (didx == 0)
2501 return;
2503 if (dsi == NULL)
2505 dsi = new_strinfo (dst, didx, NULL_TREE, false);
2506 set_strinfo (didx, dsi);
2507 find_equal_ptrs (dst, didx);
2509 else
2511 dsi = unshare_strinfo (dsi);
2512 dsi->nonzero_chars = NULL_TREE;
2513 dsi->full_string_p = false;
2514 dsi->next = 0;
2515 dsi->endptr = NULL_TREE;
2517 dsi->writable = true;
2518 dsi->stmt = stmt;
2519 dsi->dont_invalidate = true;
2521 return;
2524 tree dstlen = dsi->nonzero_chars;
2525 endptr = dsi->endptr;
2527 dsi = unshare_strinfo (dsi);
2528 dsi->endptr = NULL_TREE;
2529 dsi->stmt = NULL;
2530 dsi->writable = true;
2532 if (srclen != NULL_TREE)
2534 dsi->nonzero_chars = fold_build2_loc (loc, PLUS_EXPR,
2535 TREE_TYPE (dsi->nonzero_chars),
2536 dsi->nonzero_chars, srclen);
2537 gcc_assert (dsi->full_string_p);
2538 adjust_related_strinfos (loc, dsi, srclen);
2539 dsi->dont_invalidate = true;
2541 else
2543 dsi->nonzero_chars = NULL;
2544 dsi->full_string_p = false;
2545 if (lhs == NULL_TREE && builtin_decl_implicit_p (BUILT_IN_STPCPY))
2546 dsi->dont_invalidate = true;
2549 if (si != NULL)
2550 /* strcat src may not overlap dst, so src doesn't need to be
2551 invalidated either. */
2552 si->dont_invalidate = true;
2554 /* For now. Could remove the lhs from the call and add
2555 lhs = dst; afterwards. */
2556 if (lhs)
2557 return;
2559 fn = NULL_TREE;
2560 objsz = NULL_TREE;
2561 switch (bcode)
2563 case BUILT_IN_STRCAT:
2564 if (srclen != NULL_TREE)
2565 fn = builtin_decl_implicit (BUILT_IN_MEMCPY);
2566 else
2567 fn = builtin_decl_implicit (BUILT_IN_STRCPY);
2568 break;
2569 case BUILT_IN_STRCAT_CHK:
2570 if (srclen != NULL_TREE)
2571 fn = builtin_decl_explicit (BUILT_IN_MEMCPY_CHK);
2572 else
2573 fn = builtin_decl_explicit (BUILT_IN_STRCPY_CHK);
2574 objsz = gimple_call_arg (stmt, 2);
2575 break;
2576 default:
2577 gcc_unreachable ();
2580 if (fn == NULL_TREE)
2581 return;
2583 if (dsi && dstlen)
2585 tree type = TREE_TYPE (dstlen);
2587 /* Compute the size of the source sequence, including the nul. */
2588 tree srcsize = srclen ? srclen : size_zero_node;
2589 srcsize = fold_build2 (PLUS_EXPR, type, srcsize, build_int_cst (type, 1));
2591 tree sptr = si && si->ptr ? si->ptr : src;
2593 if (!check_bounds_or_overlap (as_a <gcall *>(stmt), dst, sptr,
2594 dstlen, srcsize))
2596 gimple_set_no_warning (stmt, true);
2597 set_no_warning = true;
2601 tree len = NULL_TREE;
2602 if (srclen != NULL_TREE)
2604 args = TYPE_ARG_TYPES (TREE_TYPE (fn));
2605 type = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (args)));
2607 len = fold_convert_loc (loc, type, unshare_expr (srclen));
2608 len = fold_build2_loc (loc, PLUS_EXPR, type, len,
2609 build_int_cst (type, 1));
2610 len = force_gimple_operand_gsi (gsi, len, true, NULL_TREE, true,
2611 GSI_SAME_STMT);
2613 if (endptr)
2614 dst = fold_convert_loc (loc, TREE_TYPE (dst), unshare_expr (endptr));
2615 else
2616 dst = fold_build2_loc (loc, POINTER_PLUS_EXPR,
2617 TREE_TYPE (dst), unshare_expr (dst),
2618 fold_convert_loc (loc, sizetype,
2619 unshare_expr (dstlen)));
2620 dst = force_gimple_operand_gsi (gsi, dst, true, NULL_TREE, true,
2621 GSI_SAME_STMT);
2622 if (dump_file && (dump_flags & TDF_DETAILS) != 0)
2624 fprintf (dump_file, "Optimizing: ");
2625 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
2627 if (srclen != NULL_TREE)
2628 success = update_gimple_call (gsi, fn, 3 + (objsz != NULL_TREE),
2629 dst, src, len, objsz);
2630 else
2631 success = update_gimple_call (gsi, fn, 2 + (objsz != NULL_TREE),
2632 dst, src, objsz);
2633 if (success)
2635 stmt = gsi_stmt (*gsi);
2636 update_stmt (stmt);
2637 if (dump_file && (dump_flags & TDF_DETAILS) != 0)
2639 fprintf (dump_file, "into: ");
2640 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
2642 /* If srclen == NULL, note that current string length can be
2643 computed by transforming this strcpy into stpcpy. */
2644 if (srclen == NULL_TREE && dsi->dont_invalidate)
2645 dsi->stmt = stmt;
2646 adjust_last_stmt (dsi, stmt, true);
2647 if (srclen != NULL_TREE)
2649 laststmt.stmt = stmt;
2650 laststmt.len = srclen;
2651 laststmt.stridx = dsi->idx;
2654 else if (dump_file && (dump_flags & TDF_DETAILS) != 0)
2655 fprintf (dump_file, "not possible.\n");
2657 if (set_no_warning)
2658 gimple_set_no_warning (stmt, true);
2661 /* Handle a call to malloc or calloc. */
2663 static void
2664 handle_builtin_malloc (enum built_in_function bcode, gimple_stmt_iterator *gsi)
2666 gimple *stmt = gsi_stmt (*gsi);
2667 tree lhs = gimple_call_lhs (stmt);
2668 if (lhs == NULL_TREE)
2669 return;
2671 gcc_assert (get_stridx (lhs) == 0);
2672 int idx = new_stridx (lhs);
2673 tree length = NULL_TREE;
2674 if (bcode == BUILT_IN_CALLOC)
2675 length = build_int_cst (size_type_node, 0);
2676 strinfo *si = new_strinfo (lhs, idx, length, length != NULL_TREE);
2677 if (bcode == BUILT_IN_CALLOC)
2678 si->endptr = lhs;
2679 set_strinfo (idx, si);
2680 si->writable = true;
2681 si->stmt = stmt;
2682 si->dont_invalidate = true;
2685 /* Handle a call to memset.
2686 After a call to calloc, memset(,0,) is unnecessary.
2687 memset(malloc(n),0,n) is calloc(n,1).
2688 return true when the call is transfomred, false otherwise. */
2690 static bool
2691 handle_builtin_memset (gimple_stmt_iterator *gsi)
2693 gimple *stmt2 = gsi_stmt (*gsi);
2694 if (!integer_zerop (gimple_call_arg (stmt2, 1)))
2695 return false;
2696 tree ptr = gimple_call_arg (stmt2, 0);
2697 int idx1 = get_stridx (ptr);
2698 if (idx1 <= 0)
2699 return false;
2700 strinfo *si1 = get_strinfo (idx1);
2701 if (!si1)
2702 return false;
2703 gimple *stmt1 = si1->stmt;
2704 if (!stmt1 || !is_gimple_call (stmt1))
2705 return false;
2706 tree callee1 = gimple_call_fndecl (stmt1);
2707 if (!valid_builtin_call (stmt1))
2708 return false;
2709 enum built_in_function code1 = DECL_FUNCTION_CODE (callee1);
2710 tree size = gimple_call_arg (stmt2, 2);
2711 if (code1 == BUILT_IN_CALLOC)
2712 /* Not touching stmt1 */ ;
2713 else if (code1 == BUILT_IN_MALLOC
2714 && operand_equal_p (gimple_call_arg (stmt1, 0), size, 0))
2716 gimple_stmt_iterator gsi1 = gsi_for_stmt (stmt1);
2717 update_gimple_call (&gsi1, builtin_decl_implicit (BUILT_IN_CALLOC), 2,
2718 size, build_one_cst (size_type_node));
2719 si1->nonzero_chars = build_int_cst (size_type_node, 0);
2720 si1->full_string_p = true;
2721 si1->stmt = gsi_stmt (gsi1);
2723 else
2724 return false;
2725 tree lhs = gimple_call_lhs (stmt2);
2726 unlink_stmt_vdef (stmt2);
2727 if (lhs)
2729 gimple *assign = gimple_build_assign (lhs, ptr);
2730 gsi_replace (gsi, assign, false);
2732 else
2734 gsi_remove (gsi, true);
2735 release_defs (stmt2);
2738 return true;
2741 /* Handle a call to memcmp. We try to handle small comparisons by
2742 converting them to load and compare, and replacing the call to memcmp
2743 with a __builtin_memcmp_eq call where possible.
2744 return true when call is transformed, return false otherwise. */
2746 static bool
2747 handle_builtin_memcmp (gimple_stmt_iterator *gsi)
2749 gcall *stmt2 = as_a <gcall *> (gsi_stmt (*gsi));
2750 tree res = gimple_call_lhs (stmt2);
2751 tree arg1 = gimple_call_arg (stmt2, 0);
2752 tree arg2 = gimple_call_arg (stmt2, 1);
2753 tree len = gimple_call_arg (stmt2, 2);
2754 unsigned HOST_WIDE_INT leni;
2755 use_operand_p use_p;
2756 imm_use_iterator iter;
2758 if (!res)
2759 return false;
2761 FOR_EACH_IMM_USE_FAST (use_p, iter, res)
2763 gimple *ustmt = USE_STMT (use_p);
2765 if (is_gimple_debug (ustmt))
2766 continue;
2767 if (gimple_code (ustmt) == GIMPLE_ASSIGN)
2769 gassign *asgn = as_a <gassign *> (ustmt);
2770 tree_code code = gimple_assign_rhs_code (asgn);
2771 if ((code != EQ_EXPR && code != NE_EXPR)
2772 || !integer_zerop (gimple_assign_rhs2 (asgn)))
2773 return false;
2775 else if (gimple_code (ustmt) == GIMPLE_COND)
2777 tree_code code = gimple_cond_code (ustmt);
2778 if ((code != EQ_EXPR && code != NE_EXPR)
2779 || !integer_zerop (gimple_cond_rhs (ustmt)))
2780 return false;
2782 else
2783 return false;
2786 if (tree_fits_uhwi_p (len)
2787 && (leni = tree_to_uhwi (len)) <= GET_MODE_SIZE (word_mode)
2788 && pow2p_hwi (leni))
2790 leni *= CHAR_TYPE_SIZE;
2791 unsigned align1 = get_pointer_alignment (arg1);
2792 unsigned align2 = get_pointer_alignment (arg2);
2793 unsigned align = MIN (align1, align2);
2794 scalar_int_mode mode;
2795 if (int_mode_for_size (leni, 1).exists (&mode)
2796 && (align >= leni || !targetm.slow_unaligned_access (mode, align)))
2798 location_t loc = gimple_location (stmt2);
2799 tree type, off;
2800 type = build_nonstandard_integer_type (leni, 1);
2801 gcc_assert (known_eq (GET_MODE_BITSIZE (TYPE_MODE (type)), leni));
2802 tree ptrtype = build_pointer_type_for_mode (char_type_node,
2803 ptr_mode, true);
2804 off = build_int_cst (ptrtype, 0);
2805 arg1 = build2_loc (loc, MEM_REF, type, arg1, off);
2806 arg2 = build2_loc (loc, MEM_REF, type, arg2, off);
2807 tree tem1 = fold_const_aggregate_ref (arg1);
2808 if (tem1)
2809 arg1 = tem1;
2810 tree tem2 = fold_const_aggregate_ref (arg2);
2811 if (tem2)
2812 arg2 = tem2;
2813 res = fold_convert_loc (loc, TREE_TYPE (res),
2814 fold_build2_loc (loc, NE_EXPR,
2815 boolean_type_node,
2816 arg1, arg2));
2817 gimplify_and_update_call_from_tree (gsi, res);
2818 return true;
2822 gimple_call_set_fndecl (stmt2, builtin_decl_explicit (BUILT_IN_MEMCMP_EQ));
2823 return true;
2826 /* Given an index to the strinfo vector, compute the string length for the
2827 corresponding string. Return -1 when unknown. */
2829 static HOST_WIDE_INT
2830 compute_string_length (int idx)
2832 HOST_WIDE_INT string_leni = -1;
2833 gcc_assert (idx != 0);
2835 if (idx < 0)
2836 return ~idx;
2838 strinfo *si = get_strinfo (idx);
2839 if (si)
2841 tree const_string_len = get_string_length (si);
2842 if (const_string_len && tree_fits_shwi_p (const_string_len))
2843 string_leni = tree_to_shwi (const_string_len);
2846 if (string_leni < 0)
2847 return -1;
2849 return string_leni;
2852 /* Determine the minimum size of the object referenced by DEST expression which
2853 must have a pointer type.
2854 Return the minimum size of the object if successful or NULL when the size
2855 cannot be determined. */
2856 static tree
2857 determine_min_objsize (tree dest)
2859 unsigned HOST_WIDE_INT size = 0;
2861 if (compute_builtin_object_size (dest, 2, &size))
2862 return build_int_cst (sizetype, size);
2864 /* Try to determine the size of the object through the RHS of the
2865 assign statement. */
2866 if (TREE_CODE (dest) == SSA_NAME)
2868 gimple *stmt = SSA_NAME_DEF_STMT (dest);
2869 if (!is_gimple_assign (stmt))
2870 return NULL_TREE;
2872 if (!gimple_assign_single_p (stmt)
2873 && !gimple_assign_unary_nop_p (stmt))
2874 return NULL_TREE;
2876 dest = gimple_assign_rhs1 (stmt);
2877 return determine_min_objsize (dest);
2880 /* Try to determine the size of the object from its type. */
2881 if (TREE_CODE (dest) != ADDR_EXPR)
2882 return NULL_TREE;
2884 tree type = TREE_TYPE (dest);
2885 if (TREE_CODE (type) == POINTER_TYPE)
2886 type = TREE_TYPE (type);
2888 type = TYPE_MAIN_VARIANT (type);
2890 /* We cannot determine the size of the array if it's a flexible array,
2891 which is declared at the end of a structure. */
2892 if (TREE_CODE (type) == ARRAY_TYPE
2893 && !array_at_struct_end_p (dest))
2895 tree size_t = TYPE_SIZE_UNIT (type);
2896 if (size_t && TREE_CODE (size_t) == INTEGER_CST
2897 && !integer_zerop (size_t))
2898 return size_t;
2901 return NULL_TREE;
2904 /* Handle a call to strcmp or strncmp. When the result is ONLY used to do
2905 equality test against zero:
2907 A. When the lengths of both arguments are constant and it's a strcmp:
2908 * if the lengths are NOT equal, we can safely fold the call
2909 to a non-zero value.
2910 * otherwise, do nothing now.
2912 B. When the length of one argument is constant, try to replace the call with
2913 a __builtin_str(n)cmp_eq call where possible, i.e:
2915 strncmp (s, STR, C) (!)= 0 in which, s is a pointer to a string, STR is a
2916 string with constant length , C is a constant.
2917 if (C <= strlen(STR) && sizeof_array(s) > C)
2919 replace this call with
2920 strncmp_eq (s, STR, C) (!)= 0
2922 if (C > strlen(STR)
2924 it can be safely treated as a call to strcmp (s, STR) (!)= 0
2925 can handled by the following strcmp.
2928 strcmp (s, STR) (!)= 0 in which, s is a pointer to a string, STR is a
2929 string with constant length.
2930 if (sizeof_array(s) > strlen(STR))
2932 replace this call with
2933 strcmp_eq (s, STR, strlen(STR)+1) (!)= 0
2936 Return true when the call is transformed, return false otherwise.
2939 static bool
2940 handle_builtin_string_cmp (gimple_stmt_iterator *gsi)
2942 gcall *stmt = as_a <gcall *> (gsi_stmt (*gsi));
2943 tree res = gimple_call_lhs (stmt);
2944 use_operand_p use_p;
2945 imm_use_iterator iter;
2946 tree arg1 = gimple_call_arg (stmt, 0);
2947 tree arg2 = gimple_call_arg (stmt, 1);
2948 int idx1 = get_stridx (arg1);
2949 int idx2 = get_stridx (arg2);
2950 HOST_WIDE_INT length = -1;
2951 bool is_ncmp = false;
2953 if (!res)
2954 return false;
2956 /* When both arguments are unknown, do nothing. */
2957 if (idx1 == 0 && idx2 == 0)
2958 return false;
2960 /* Handle strncmp function. */
2961 if (gimple_call_num_args (stmt) == 3)
2963 tree len = gimple_call_arg (stmt, 2);
2964 if (tree_fits_shwi_p (len))
2965 length = tree_to_shwi (len);
2967 is_ncmp = true;
2970 /* For strncmp, if the length argument is NOT known, do nothing. */
2971 if (is_ncmp && length < 0)
2972 return false;
2974 /* When the result is ONLY used to do equality test against zero. */
2975 FOR_EACH_IMM_USE_FAST (use_p, iter, res)
2977 gimple *use_stmt = USE_STMT (use_p);
2979 if (is_gimple_debug (use_stmt))
2980 continue;
2981 if (gimple_code (use_stmt) == GIMPLE_ASSIGN)
2983 tree_code code = gimple_assign_rhs_code (use_stmt);
2984 if (code == COND_EXPR)
2986 tree cond_expr = gimple_assign_rhs1 (use_stmt);
2987 if ((TREE_CODE (cond_expr) != EQ_EXPR
2988 && (TREE_CODE (cond_expr) != NE_EXPR))
2989 || !integer_zerop (TREE_OPERAND (cond_expr, 1)))
2990 return false;
2992 else if (code == EQ_EXPR || code == NE_EXPR)
2994 if (!integer_zerop (gimple_assign_rhs2 (use_stmt)))
2995 return false;
2997 else
2998 return false;
3000 else if (gimple_code (use_stmt) == GIMPLE_COND)
3002 tree_code code = gimple_cond_code (use_stmt);
3003 if ((code != EQ_EXPR && code != NE_EXPR)
3004 || !integer_zerop (gimple_cond_rhs (use_stmt)))
3005 return false;
3007 else
3008 return false;
3011 /* When the lengths of both arguments are known, and they are unequal, we can
3012 safely fold the call to a non-zero value for strcmp;
3013 othewise, do nothing now. */
3014 if (idx1 != 0 && idx2 != 0)
3016 HOST_WIDE_INT const_string_leni1 = compute_string_length (idx1);
3017 HOST_WIDE_INT const_string_leni2 = compute_string_length (idx2);
3019 if (!is_ncmp
3020 && const_string_leni1 != -1
3021 && const_string_leni2 != -1
3022 && const_string_leni1 != const_string_leni2)
3024 replace_call_with_value (gsi, integer_one_node);
3025 return true;
3027 return false;
3030 /* When the length of one argument is constant. */
3031 tree var_string = NULL_TREE;
3032 HOST_WIDE_INT const_string_leni = -1;
3034 if (idx1)
3036 const_string_leni = compute_string_length (idx1);
3037 var_string = arg2;
3039 else
3041 gcc_checking_assert (idx2);
3042 const_string_leni = compute_string_length (idx2);
3043 var_string = arg1;
3046 if (const_string_leni < 0)
3047 return false;
3049 unsigned HOST_WIDE_INT var_sizei = 0;
3050 /* try to determine the minimum size of the object pointed by var_string. */
3051 tree size = determine_min_objsize (var_string);
3053 if (!size)
3054 return false;
3056 if (tree_fits_uhwi_p (size))
3057 var_sizei = tree_to_uhwi (size);
3059 if (var_sizei == 0)
3060 return false;
3062 /* For strncmp, if length > const_string_leni , this call can be safely
3063 transformed to a strcmp. */
3064 if (is_ncmp && length > const_string_leni)
3065 is_ncmp = false;
3067 unsigned HOST_WIDE_INT final_length
3068 = is_ncmp ? length : const_string_leni + 1;
3070 /* Replace strcmp or strncmp with the corresponding str(n)cmp_eq. */
3071 if (var_sizei > final_length)
3073 tree fn
3074 = (is_ncmp
3075 ? builtin_decl_implicit (BUILT_IN_STRNCMP_EQ)
3076 : builtin_decl_implicit (BUILT_IN_STRCMP_EQ));
3077 if (!fn)
3078 return false;
3079 tree const_string_len = build_int_cst (size_type_node, final_length);
3080 update_gimple_call (gsi, fn, 3, arg1, arg2, const_string_len);
3082 else
3083 return false;
3085 return true;
3088 /* Handle a POINTER_PLUS_EXPR statement.
3089 For p = "abcd" + 2; compute associated length, or if
3090 p = q + off is pointing to a '\0' character of a string, call
3091 zero_length_string on it. */
3093 static void
3094 handle_pointer_plus (gimple_stmt_iterator *gsi)
3096 gimple *stmt = gsi_stmt (*gsi);
3097 tree lhs = gimple_assign_lhs (stmt), off;
3098 int idx = get_stridx (gimple_assign_rhs1 (stmt));
3099 strinfo *si, *zsi;
3101 if (idx == 0)
3102 return;
3104 if (idx < 0)
3106 tree off = gimple_assign_rhs2 (stmt);
3107 if (tree_fits_uhwi_p (off)
3108 && tree_to_uhwi (off) <= (unsigned HOST_WIDE_INT) ~idx)
3109 ssa_ver_to_stridx[SSA_NAME_VERSION (lhs)]
3110 = ~(~idx - (int) tree_to_uhwi (off));
3111 return;
3114 si = get_strinfo (idx);
3115 if (si == NULL || si->nonzero_chars == NULL_TREE)
3116 return;
3118 off = gimple_assign_rhs2 (stmt);
3119 zsi = NULL;
3120 if (si->full_string_p && operand_equal_p (si->nonzero_chars, off, 0))
3121 zsi = zero_length_string (lhs, si);
3122 else if (TREE_CODE (off) == SSA_NAME)
3124 gimple *def_stmt = SSA_NAME_DEF_STMT (off);
3125 if (gimple_assign_single_p (def_stmt)
3126 && si->full_string_p
3127 && operand_equal_p (si->nonzero_chars,
3128 gimple_assign_rhs1 (def_stmt), 0))
3129 zsi = zero_length_string (lhs, si);
3131 if (zsi != NULL
3132 && si->endptr != NULL_TREE
3133 && si->endptr != lhs
3134 && TREE_CODE (si->endptr) == SSA_NAME)
3136 enum tree_code rhs_code
3137 = useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (si->endptr))
3138 ? SSA_NAME : NOP_EXPR;
3139 gimple_assign_set_rhs_with_ops (gsi, rhs_code, si->endptr);
3140 gcc_assert (gsi_stmt (*gsi) == stmt);
3141 update_stmt (stmt);
3145 /* If RHS, either directly or indirectly, refers to a string of constant
3146 length, return the length. Otherwise, if it refers to a character
3147 constant, return 1 if the constant is non-zero and 0 if it is nul.
3148 Otherwise, return a negative value. */
3150 static HOST_WIDE_INT
3151 get_min_string_length (tree rhs, bool *full_string_p)
3153 if (TREE_CODE (TREE_TYPE (rhs)) == INTEGER_TYPE)
3155 if (tree_expr_nonzero_p (rhs))
3157 *full_string_p = false;
3158 return 1;
3161 *full_string_p = true;
3162 return 0;
3165 if (TREE_CODE (rhs) == MEM_REF
3166 && integer_zerop (TREE_OPERAND (rhs, 1)))
3168 rhs = TREE_OPERAND (rhs, 0);
3169 if (TREE_CODE (rhs) == ADDR_EXPR)
3171 tree rhs_addr = rhs;
3173 rhs = TREE_OPERAND (rhs, 0);
3174 if (TREE_CODE (rhs) != STRING_CST)
3176 int idx = get_stridx (rhs_addr);
3177 if (idx > 0)
3179 strinfo *si = get_strinfo (idx);
3180 if (si
3181 && tree_fits_shwi_p (si->nonzero_chars))
3183 *full_string_p = si->full_string_p;
3184 return tree_to_shwi (si->nonzero_chars);
3191 if (TREE_CODE (rhs) == VAR_DECL
3192 && TREE_READONLY (rhs))
3193 rhs = DECL_INITIAL (rhs);
3195 if (rhs && TREE_CODE (rhs) == STRING_CST)
3197 *full_string_p = true;
3198 return strlen (TREE_STRING_POINTER (rhs));
3201 return -1;
3204 /* Handle a single or multiple character store either by single
3205 character assignment or by multi-character assignment from
3206 MEM_REF. */
3208 static bool
3209 handle_char_store (gimple_stmt_iterator *gsi)
3211 int idx = -1;
3212 strinfo *si = NULL;
3213 gimple *stmt = gsi_stmt (*gsi);
3214 tree ssaname = NULL_TREE, lhs = gimple_assign_lhs (stmt);
3215 tree rhs = gimple_assign_rhs1 (stmt);
3216 unsigned HOST_WIDE_INT offset = 0;
3218 if (TREE_CODE (lhs) == MEM_REF
3219 && TREE_CODE (TREE_OPERAND (lhs, 0)) == SSA_NAME)
3221 tree mem_offset = TREE_OPERAND (lhs, 1);
3222 if (tree_fits_uhwi_p (mem_offset))
3224 /* Get the strinfo for the base, and use it if it starts with at
3225 least OFFSET nonzero characters. This is trivially true if
3226 OFFSET is zero. */
3227 offset = tree_to_uhwi (mem_offset);
3228 idx = get_stridx (TREE_OPERAND (lhs, 0));
3229 if (idx > 0)
3230 si = get_strinfo (idx);
3231 if (offset == 0)
3232 ssaname = TREE_OPERAND (lhs, 0);
3233 else if (si == NULL || compare_nonzero_chars (si, offset) < 0)
3234 return true;
3237 else
3239 idx = get_addr_stridx (lhs, NULL_TREE, &offset);
3240 if (idx > 0)
3241 si = get_strinfo (idx);
3244 /* STORING_NONZERO_P is true iff not all stored characters are zero.
3245 STORING_ALL_ZEROS_P is true iff all stored characters are zero.
3246 Both are false when it's impossible to determine which is true. */
3247 bool storing_nonzero_p;
3248 bool storing_all_zeros_p = initializer_zerop (rhs, &storing_nonzero_p);
3249 if (!storing_nonzero_p)
3250 storing_nonzero_p = tree_expr_nonzero_p (rhs);
3251 bool full_string_p = storing_all_zeros_p;
3253 /* Set to the length of the string being assigned if known. */
3254 HOST_WIDE_INT rhslen;
3256 if (si != NULL)
3258 int cmp = compare_nonzero_chars (si, offset);
3259 gcc_assert (offset == 0 || cmp >= 0);
3260 if (storing_all_zeros_p && cmp == 0 && si->full_string_p)
3262 /* When overwriting a '\0' with a '\0', the store can be removed
3263 if we know it has been stored in the current function. */
3264 if (!stmt_could_throw_p (stmt) && si->writable)
3266 unlink_stmt_vdef (stmt);
3267 release_defs (stmt);
3268 gsi_remove (gsi, true);
3269 return false;
3271 else
3273 si->writable = true;
3274 gsi_next (gsi);
3275 return false;
3278 /* If si->nonzero_chars > OFFSET, we aren't overwriting '\0',
3279 and if we aren't storing '\0', we know that the length of the
3280 string and any other zero terminated string in memory remains
3281 the same. In that case we move to the next gimple statement and
3282 return to signal the caller that it shouldn't invalidate anything.
3284 This is benefical for cases like:
3286 char p[20];
3287 void foo (char *q)
3289 strcpy (p, "foobar");
3290 size_t len = strlen (p); // This can be optimized into 6
3291 size_t len2 = strlen (q); // This has to be computed
3292 p[0] = 'X';
3293 size_t len3 = strlen (p); // This can be optimized into 6
3294 size_t len4 = strlen (q); // This can be optimized into len2
3295 bar (len, len2, len3, len4);
3298 else if (storing_nonzero_p && cmp > 0)
3300 gsi_next (gsi);
3301 return false;
3303 else if (storing_all_zeros_p || storing_nonzero_p || (offset != 0 && cmp > 0))
3305 /* When STORING_NONZERO_P, we know that the string will start
3306 with at least OFFSET + 1 nonzero characters. If storing
3307 a single character, set si->NONZERO_CHARS to the result.
3308 If storing multiple characters, try to determine the number
3309 of leading non-zero characters and set si->NONZERO_CHARS to
3310 the result instead.
3312 When STORING_ALL_ZEROS_P, we know that the string is now
3313 OFFSET characters long.
3315 Otherwise, we're storing an unknown value at offset OFFSET,
3316 so need to clip the nonzero_chars to OFFSET. */
3317 bool full_string_p = storing_all_zeros_p;
3318 HOST_WIDE_INT len = (storing_nonzero_p
3319 ? get_min_string_length (rhs, &full_string_p)
3320 : 1);
3322 location_t loc = gimple_location (stmt);
3323 tree oldlen = si->nonzero_chars;
3324 if (cmp == 0 && si->full_string_p)
3325 /* We're overwriting the nul terminator with a nonzero or
3326 unknown character. If the previous stmt was a memcpy,
3327 its length may be decreased. */
3328 adjust_last_stmt (si, stmt, false);
3329 si = unshare_strinfo (si);
3330 if (storing_nonzero_p)
3332 gcc_assert (len >= 0);
3333 si->nonzero_chars = build_int_cst (size_type_node, offset + len);
3335 else
3336 si->nonzero_chars = build_int_cst (size_type_node, offset);
3337 si->full_string_p = full_string_p;
3338 if (storing_all_zeros_p
3339 && ssaname
3340 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (ssaname))
3341 si->endptr = ssaname;
3342 else
3343 si->endptr = NULL;
3344 si->next = 0;
3345 si->stmt = NULL;
3346 si->writable = true;
3347 si->dont_invalidate = true;
3348 if (oldlen)
3350 tree adj = fold_build2_loc (loc, MINUS_EXPR, size_type_node,
3351 si->nonzero_chars, oldlen);
3352 adjust_related_strinfos (loc, si, adj);
3354 else
3355 si->prev = 0;
3358 else if (idx == 0 && (storing_all_zeros_p || storing_nonzero_p))
3360 if (ssaname)
3361 idx = new_stridx (ssaname);
3362 else
3363 idx = new_addr_stridx (lhs);
3364 if (idx != 0)
3366 tree ptr = (ssaname ? ssaname : build_fold_addr_expr (lhs));
3367 HOST_WIDE_INT slen = (storing_all_zeros_p
3369 : (storing_nonzero_p
3370 ? get_min_string_length (rhs, &full_string_p)
3371 : -1));
3372 tree len = (slen <= 0
3373 ? size_zero_node
3374 : build_int_cst (size_type_node, slen));
3375 si = new_strinfo (ptr, idx, len, slen >= 0 && full_string_p);
3376 set_strinfo (idx, si);
3377 if (storing_all_zeros_p
3378 && ssaname
3379 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (ssaname))
3380 si->endptr = ssaname;
3381 si->dont_invalidate = true;
3382 si->writable = true;
3385 else if (idx == 0
3386 && (rhslen = get_min_string_length (rhs, &full_string_p)) >= 0
3387 && ssaname == NULL_TREE
3388 && TREE_CODE (TREE_TYPE (lhs)) == ARRAY_TYPE)
3390 HOST_WIDE_INT a = int_size_in_bytes (TREE_TYPE (lhs));
3391 if (a > 0 && (unsigned HOST_WIDE_INT) a > (unsigned HOST_WIDE_INT) rhslen)
3393 int idx = new_addr_stridx (lhs);
3394 if (idx != 0)
3396 si = new_strinfo (build_fold_addr_expr (lhs), idx,
3397 build_int_cst (size_type_node, rhslen),
3398 full_string_p);
3399 set_strinfo (idx, si);
3400 si->dont_invalidate = true;
3405 if (si != NULL && offset == 0 && storing_all_zeros_p)
3407 /* Allow adjust_last_stmt to remove it if the stored '\0'
3408 is immediately overwritten. */
3409 laststmt.stmt = stmt;
3410 laststmt.len = build_int_cst (size_type_node, 1);
3411 laststmt.stridx = si->idx;
3413 return true;
3416 /* Try to fold strstr (s, t) eq/ne s to strncmp (s, t, strlen (t)) eq/ne 0. */
3418 static void
3419 fold_strstr_to_strncmp (tree rhs1, tree rhs2, gimple *stmt)
3421 if (TREE_CODE (rhs1) != SSA_NAME
3422 || TREE_CODE (rhs2) != SSA_NAME)
3423 return;
3425 gimple *call_stmt = NULL;
3426 for (int pass = 0; pass < 2; pass++)
3428 gimple *g = SSA_NAME_DEF_STMT (rhs1);
3429 if (gimple_call_builtin_p (g, BUILT_IN_STRSTR)
3430 && has_single_use (rhs1)
3431 && gimple_call_arg (g, 0) == rhs2)
3433 call_stmt = g;
3434 break;
3436 std::swap (rhs1, rhs2);
3439 if (call_stmt)
3441 tree arg0 = gimple_call_arg (call_stmt, 0);
3443 if (arg0 == rhs2)
3445 tree arg1 = gimple_call_arg (call_stmt, 1);
3446 tree arg1_len = NULL_TREE;
3447 int idx = get_stridx (arg1);
3449 if (idx)
3451 if (idx < 0)
3452 arg1_len = build_int_cst (size_type_node, ~idx);
3453 else
3455 strinfo *si = get_strinfo (idx);
3456 if (si)
3457 arg1_len = get_string_length (si);
3461 if (arg1_len != NULL_TREE)
3463 gimple_stmt_iterator gsi = gsi_for_stmt (call_stmt);
3464 tree strncmp_decl = builtin_decl_explicit (BUILT_IN_STRNCMP);
3466 if (!is_gimple_val (arg1_len))
3468 tree arg1_len_tmp = make_ssa_name (TREE_TYPE (arg1_len));
3469 gassign *arg1_stmt = gimple_build_assign (arg1_len_tmp,
3470 arg1_len);
3471 gsi_insert_before (&gsi, arg1_stmt, GSI_SAME_STMT);
3472 arg1_len = arg1_len_tmp;
3475 gcall *strncmp_call = gimple_build_call (strncmp_decl, 3,
3476 arg0, arg1, arg1_len);
3477 tree strncmp_lhs = make_ssa_name (integer_type_node);
3478 gimple_set_vuse (strncmp_call, gimple_vuse (call_stmt));
3479 gimple_call_set_lhs (strncmp_call, strncmp_lhs);
3480 gsi_remove (&gsi, true);
3481 gsi_insert_before (&gsi, strncmp_call, GSI_SAME_STMT);
3482 tree zero = build_zero_cst (TREE_TYPE (strncmp_lhs));
3484 if (is_gimple_assign (stmt))
3486 if (gimple_assign_rhs_code (stmt) == COND_EXPR)
3488 tree cond = gimple_assign_rhs1 (stmt);
3489 TREE_OPERAND (cond, 0) = strncmp_lhs;
3490 TREE_OPERAND (cond, 1) = zero;
3492 else
3494 gimple_assign_set_rhs1 (stmt, strncmp_lhs);
3495 gimple_assign_set_rhs2 (stmt, zero);
3498 else
3500 gcond *cond = as_a<gcond *> (stmt);
3501 gimple_cond_set_lhs (cond, strncmp_lhs);
3502 gimple_cond_set_rhs (cond, zero);
3504 update_stmt (stmt);
3510 /* Attempt to check for validity of the performed access a single statement
3511 at *GSI using string length knowledge, and to optimize it.
3512 If the given basic block needs clean-up of EH, CLEANUP_EH is set to
3513 true. */
3515 static bool
3516 strlen_check_and_optimize_stmt (gimple_stmt_iterator *gsi, bool *cleanup_eh)
3518 gimple *stmt = gsi_stmt (*gsi);
3520 if (is_gimple_call (stmt))
3522 tree callee = gimple_call_fndecl (stmt);
3523 if (valid_builtin_call (stmt))
3524 switch (DECL_FUNCTION_CODE (callee))
3526 case BUILT_IN_STRLEN:
3527 case BUILT_IN_STRNLEN:
3528 handle_builtin_strlen (gsi);
3529 break;
3530 case BUILT_IN_STRCHR:
3531 handle_builtin_strchr (gsi);
3532 break;
3533 case BUILT_IN_STRCPY:
3534 case BUILT_IN_STRCPY_CHK:
3535 case BUILT_IN_STPCPY:
3536 case BUILT_IN_STPCPY_CHK:
3537 handle_builtin_strcpy (DECL_FUNCTION_CODE (callee), gsi);
3538 break;
3540 case BUILT_IN_STRNCAT:
3541 case BUILT_IN_STRNCAT_CHK:
3542 handle_builtin_strncat (DECL_FUNCTION_CODE (callee), gsi);
3543 break;
3545 case BUILT_IN_STPNCPY:
3546 case BUILT_IN_STPNCPY_CHK:
3547 case BUILT_IN_STRNCPY:
3548 case BUILT_IN_STRNCPY_CHK:
3549 handle_builtin_stxncpy (DECL_FUNCTION_CODE (callee), gsi);
3550 break;
3552 case BUILT_IN_MEMCPY:
3553 case BUILT_IN_MEMCPY_CHK:
3554 case BUILT_IN_MEMPCPY:
3555 case BUILT_IN_MEMPCPY_CHK:
3556 handle_builtin_memcpy (DECL_FUNCTION_CODE (callee), gsi);
3557 break;
3558 case BUILT_IN_STRCAT:
3559 case BUILT_IN_STRCAT_CHK:
3560 handle_builtin_strcat (DECL_FUNCTION_CODE (callee), gsi);
3561 break;
3562 case BUILT_IN_MALLOC:
3563 case BUILT_IN_CALLOC:
3564 handle_builtin_malloc (DECL_FUNCTION_CODE (callee), gsi);
3565 break;
3566 case BUILT_IN_MEMSET:
3567 if (handle_builtin_memset (gsi))
3568 return false;
3569 break;
3570 case BUILT_IN_MEMCMP:
3571 if (handle_builtin_memcmp (gsi))
3572 return false;
3573 break;
3574 case BUILT_IN_STRCMP:
3575 case BUILT_IN_STRNCMP:
3576 if (handle_builtin_string_cmp (gsi))
3577 return false;
3578 break;
3579 default:
3580 break;
3583 else if (is_gimple_assign (stmt) && !gimple_clobber_p (stmt))
3585 tree lhs = gimple_assign_lhs (stmt);
3587 if (TREE_CODE (lhs) == SSA_NAME && POINTER_TYPE_P (TREE_TYPE (lhs)))
3589 if (gimple_assign_single_p (stmt)
3590 || (gimple_assign_cast_p (stmt)
3591 && POINTER_TYPE_P (TREE_TYPE (gimple_assign_rhs1 (stmt)))))
3593 int idx = get_stridx (gimple_assign_rhs1 (stmt));
3594 ssa_ver_to_stridx[SSA_NAME_VERSION (lhs)] = idx;
3596 else if (gimple_assign_rhs_code (stmt) == POINTER_PLUS_EXPR)
3597 handle_pointer_plus (gsi);
3599 else if (TREE_CODE (lhs) == SSA_NAME && INTEGRAL_TYPE_P (TREE_TYPE (lhs)))
3601 enum tree_code code = gimple_assign_rhs_code (stmt);
3602 if (code == COND_EXPR)
3604 tree cond = gimple_assign_rhs1 (stmt);
3605 enum tree_code cond_code = TREE_CODE (cond);
3607 if (cond_code == EQ_EXPR || cond_code == NE_EXPR)
3608 fold_strstr_to_strncmp (TREE_OPERAND (cond, 0),
3609 TREE_OPERAND (cond, 1), stmt);
3611 else if (code == EQ_EXPR || code == NE_EXPR)
3612 fold_strstr_to_strncmp (gimple_assign_rhs1 (stmt),
3613 gimple_assign_rhs2 (stmt), stmt);
3614 else if (gimple_assign_load_p (stmt)
3615 && TREE_CODE (TREE_TYPE (lhs)) == INTEGER_TYPE
3616 && TYPE_MODE (TREE_TYPE (lhs)) == TYPE_MODE (char_type_node)
3617 && (TYPE_PRECISION (TREE_TYPE (lhs))
3618 == TYPE_PRECISION (char_type_node))
3619 && !gimple_has_volatile_ops (stmt))
3621 tree off = integer_zero_node;
3622 unsigned HOST_WIDE_INT coff = 0;
3623 int idx = 0;
3624 tree rhs1 = gimple_assign_rhs1 (stmt);
3625 if (code == MEM_REF)
3627 idx = get_stridx (TREE_OPERAND (rhs1, 0));
3628 if (idx > 0)
3630 strinfo *si = get_strinfo (idx);
3631 if (si
3632 && si->nonzero_chars
3633 && TREE_CODE (si->nonzero_chars) == INTEGER_CST
3634 && (wi::to_widest (si->nonzero_chars)
3635 >= wi::to_widest (off)))
3636 off = TREE_OPERAND (rhs1, 1);
3637 else
3638 /* This case is not useful. See if get_addr_stridx
3639 returns something usable. */
3640 idx = 0;
3643 if (idx <= 0)
3644 idx = get_addr_stridx (rhs1, NULL_TREE, &coff);
3645 if (idx > 0)
3647 strinfo *si = get_strinfo (idx);
3648 if (si
3649 && si->nonzero_chars
3650 && TREE_CODE (si->nonzero_chars) == INTEGER_CST)
3652 widest_int w1 = wi::to_widest (si->nonzero_chars);
3653 widest_int w2 = wi::to_widest (off) + coff;
3654 if (w1 == w2
3655 && si->full_string_p)
3657 if (dump_file && (dump_flags & TDF_DETAILS) != 0)
3659 fprintf (dump_file, "Optimizing: ");
3660 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
3663 /* Reading the final '\0' character. */
3664 tree zero = build_int_cst (TREE_TYPE (lhs), 0);
3665 gimple_set_vuse (stmt, NULL_TREE);
3666 gimple_assign_set_rhs_from_tree (gsi, zero);
3667 *cleanup_eh
3668 |= maybe_clean_or_replace_eh_stmt (stmt,
3669 gsi_stmt (*gsi));
3670 stmt = gsi_stmt (*gsi);
3671 update_stmt (stmt);
3673 if (dump_file && (dump_flags & TDF_DETAILS) != 0)
3675 fprintf (dump_file, "into: ");
3676 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
3679 else if (w1 > w2)
3681 /* Reading a character before the final '\0'
3682 character. Just set the value range to ~[0, 0]
3683 if we don't have anything better. */
3684 wide_int min, max;
3685 tree type = TREE_TYPE (lhs);
3686 enum value_range_type vr
3687 = get_range_info (lhs, &min, &max);
3688 if (vr == VR_VARYING
3689 || (vr == VR_RANGE
3690 && min == wi::min_value (TYPE_PRECISION (type),
3691 TYPE_SIGN (type))
3692 && max == wi::max_value (TYPE_PRECISION (type),
3693 TYPE_SIGN (type))))
3694 set_range_info (lhs, VR_ANTI_RANGE,
3695 wi::zero (TYPE_PRECISION (type)),
3696 wi::zero (TYPE_PRECISION (type)));
3702 if (strlen_to_stridx)
3704 tree rhs1 = gimple_assign_rhs1 (stmt);
3705 if (stridx_strlenloc *ps = strlen_to_stridx->get (rhs1))
3706 strlen_to_stridx->put (lhs, stridx_strlenloc (*ps));
3709 else if (TREE_CODE (lhs) != SSA_NAME && !TREE_SIDE_EFFECTS (lhs))
3711 tree type = TREE_TYPE (lhs);
3712 if (TREE_CODE (type) == ARRAY_TYPE)
3713 type = TREE_TYPE (type);
3714 if (TREE_CODE (type) == INTEGER_TYPE
3715 && TYPE_MODE (type) == TYPE_MODE (char_type_node)
3716 && TYPE_PRECISION (type) == TYPE_PRECISION (char_type_node))
3718 if (! handle_char_store (gsi))
3719 return false;
3723 else if (gcond *cond = dyn_cast<gcond *> (stmt))
3725 enum tree_code code = gimple_cond_code (cond);
3726 if (code == EQ_EXPR || code == NE_EXPR)
3727 fold_strstr_to_strncmp (gimple_cond_lhs (stmt),
3728 gimple_cond_rhs (stmt), stmt);
3731 if (gimple_vdef (stmt))
3732 maybe_invalidate (stmt);
3733 return true;
3736 /* Recursively call maybe_invalidate on stmts that might be executed
3737 in between dombb and current bb and that contain a vdef. Stop when
3738 *count stmts are inspected, or if the whole strinfo vector has
3739 been invalidated. */
3741 static void
3742 do_invalidate (basic_block dombb, gimple *phi, bitmap visited, int *count)
3744 unsigned int i, n = gimple_phi_num_args (phi);
3746 for (i = 0; i < n; i++)
3748 tree vuse = gimple_phi_arg_def (phi, i);
3749 gimple *stmt = SSA_NAME_DEF_STMT (vuse);
3750 basic_block bb = gimple_bb (stmt);
3751 if (bb == NULL
3752 || bb == dombb
3753 || !bitmap_set_bit (visited, bb->index)
3754 || !dominated_by_p (CDI_DOMINATORS, bb, dombb))
3755 continue;
3756 while (1)
3758 if (gimple_code (stmt) == GIMPLE_PHI)
3760 do_invalidate (dombb, stmt, visited, count);
3761 if (*count == 0)
3762 return;
3763 break;
3765 if (--*count == 0)
3766 return;
3767 if (!maybe_invalidate (stmt))
3769 *count = 0;
3770 return;
3772 vuse = gimple_vuse (stmt);
3773 stmt = SSA_NAME_DEF_STMT (vuse);
3774 if (gimple_bb (stmt) != bb)
3776 bb = gimple_bb (stmt);
3777 if (bb == NULL
3778 || bb == dombb
3779 || !bitmap_set_bit (visited, bb->index)
3780 || !dominated_by_p (CDI_DOMINATORS, bb, dombb))
3781 break;
3787 class strlen_dom_walker : public dom_walker
3789 public:
3790 strlen_dom_walker (cdi_direction direction)
3791 : dom_walker (direction), m_cleanup_cfg (false)
3794 virtual edge before_dom_children (basic_block);
3795 virtual void after_dom_children (basic_block);
3797 /* Flag that will trigger TODO_cleanup_cfg to be returned in strlen
3798 execute function. */
3799 bool m_cleanup_cfg;
3802 /* Callback for walk_dominator_tree. Attempt to optimize various
3803 string ops by remembering string lengths pointed by pointer SSA_NAMEs. */
3805 edge
3806 strlen_dom_walker::before_dom_children (basic_block bb)
3808 basic_block dombb = get_immediate_dominator (CDI_DOMINATORS, bb);
3810 if (dombb == NULL)
3811 stridx_to_strinfo = NULL;
3812 else
3814 stridx_to_strinfo = ((vec<strinfo *, va_heap, vl_embed> *) dombb->aux);
3815 if (stridx_to_strinfo)
3817 for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi);
3818 gsi_next (&gsi))
3820 gphi *phi = gsi.phi ();
3821 if (virtual_operand_p (gimple_phi_result (phi)))
3823 bitmap visited = BITMAP_ALLOC (NULL);
3824 int count_vdef = 100;
3825 do_invalidate (dombb, phi, visited, &count_vdef);
3826 BITMAP_FREE (visited);
3827 if (count_vdef == 0)
3829 /* If there were too many vdefs in between immediate
3830 dominator and current bb, invalidate everything.
3831 If stridx_to_strinfo has been unshared, we need
3832 to free it, otherwise just set it to NULL. */
3833 if (!strinfo_shared ())
3835 unsigned int i;
3836 strinfo *si;
3838 for (i = 1;
3839 vec_safe_iterate (stridx_to_strinfo, i, &si);
3840 ++i)
3842 free_strinfo (si);
3843 (*stridx_to_strinfo)[i] = NULL;
3846 else
3847 stridx_to_strinfo = NULL;
3849 break;
3855 /* If all PHI arguments have the same string index, the PHI result
3856 has it as well. */
3857 for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi);
3858 gsi_next (&gsi))
3860 gphi *phi = gsi.phi ();
3861 tree result = gimple_phi_result (phi);
3862 if (!virtual_operand_p (result) && POINTER_TYPE_P (TREE_TYPE (result)))
3864 int idx = get_stridx (gimple_phi_arg_def (phi, 0));
3865 if (idx != 0)
3867 unsigned int i, n = gimple_phi_num_args (phi);
3868 for (i = 1; i < n; i++)
3869 if (idx != get_stridx (gimple_phi_arg_def (phi, i)))
3870 break;
3871 if (i == n)
3872 ssa_ver_to_stridx[SSA_NAME_VERSION (result)] = idx;
3877 bool cleanup_eh = false;
3879 /* Attempt to optimize individual statements. */
3880 for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi); )
3881 if (strlen_check_and_optimize_stmt (&gsi, &cleanup_eh))
3882 gsi_next (&gsi);
3884 if (cleanup_eh && gimple_purge_dead_eh_edges (bb))
3885 m_cleanup_cfg = true;
3887 bb->aux = stridx_to_strinfo;
3888 if (vec_safe_length (stridx_to_strinfo) && !strinfo_shared ())
3889 (*stridx_to_strinfo)[0] = (strinfo *) bb;
3890 return NULL;
3893 /* Callback for walk_dominator_tree. Free strinfo vector if it is
3894 owned by the current bb, clear bb->aux. */
3896 void
3897 strlen_dom_walker::after_dom_children (basic_block bb)
3899 if (bb->aux)
3901 stridx_to_strinfo = ((vec<strinfo *, va_heap, vl_embed> *) bb->aux);
3902 if (vec_safe_length (stridx_to_strinfo)
3903 && (*stridx_to_strinfo)[0] == (strinfo *) bb)
3905 unsigned int i;
3906 strinfo *si;
3908 for (i = 1; vec_safe_iterate (stridx_to_strinfo, i, &si); ++i)
3909 free_strinfo (si);
3910 vec_free (stridx_to_strinfo);
3912 bb->aux = NULL;
3916 /* Main entry point. */
3918 namespace {
3920 const pass_data pass_data_strlen =
3922 GIMPLE_PASS, /* type */
3923 "strlen", /* name */
3924 OPTGROUP_NONE, /* optinfo_flags */
3925 TV_TREE_STRLEN, /* tv_id */
3926 ( PROP_cfg | PROP_ssa ), /* properties_required */
3927 0, /* properties_provided */
3928 0, /* properties_destroyed */
3929 0, /* todo_flags_start */
3930 0, /* todo_flags_finish */
3933 class pass_strlen : public gimple_opt_pass
3935 public:
3936 pass_strlen (gcc::context *ctxt)
3937 : gimple_opt_pass (pass_data_strlen, ctxt)
3940 /* opt_pass methods: */
3941 virtual bool gate (function *) { return flag_optimize_strlen != 0; }
3942 virtual unsigned int execute (function *);
3944 }; // class pass_strlen
3946 unsigned int
3947 pass_strlen::execute (function *fun)
3949 gcc_assert (!strlen_to_stridx);
3950 if (warn_stringop_overflow || warn_stringop_truncation)
3951 strlen_to_stridx = new hash_map<tree, stridx_strlenloc> ();
3953 ssa_ver_to_stridx.safe_grow_cleared (num_ssa_names);
3954 max_stridx = 1;
3956 calculate_dominance_info (CDI_DOMINATORS);
3958 /* String length optimization is implemented as a walk of the dominator
3959 tree and a forward walk of statements within each block. */
3960 strlen_dom_walker walker (CDI_DOMINATORS);
3961 walker.walk (fun->cfg->x_entry_block_ptr);
3963 ssa_ver_to_stridx.release ();
3964 strinfo_pool.release ();
3965 if (decl_to_stridxlist_htab)
3967 obstack_free (&stridx_obstack, NULL);
3968 delete decl_to_stridxlist_htab;
3969 decl_to_stridxlist_htab = NULL;
3971 laststmt.stmt = NULL;
3972 laststmt.len = NULL_TREE;
3973 laststmt.stridx = 0;
3975 if (strlen_to_stridx)
3977 strlen_to_stridx->empty ();
3978 delete strlen_to_stridx;
3979 strlen_to_stridx = NULL;
3982 return walker.m_cleanup_cfg ? TODO_cleanup_cfg : 0;
3985 } // anon namespace
3987 gimple_opt_pass *
3988 make_pass_strlen (gcc::context *ctxt)
3990 return new pass_strlen (ctxt);