[Bug tree-optimization/101031] New: wrong code at -O2 on x86_64-linux-gnu

[Bug tree-optimization/101031] New: wrong code at -O2 on x86_64-linux-gnu

[qrzhang at gatech dot edu]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101031

            Bug ID: 101031
           Summary: wrong code at -O2 on x86_64-linux-gnu
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: qrzhang at gatech dot edu
  Target Milestone: ---

It appears to be a recent regression.


$ gcc-trunk -v
Supported LTO compression algorithms: zlib
gcc version 12.0.0 20210611 (experimental) [master revision
336c41dbcb2:00da4bcb67d:36943c6bdd3d3b535b24872bbd802d91ef0c6299] (GCC)


$ gcc-trunk abc.c ; ./a.out
0

$ gcc-trunk -O2 abc.c ; ./a.out
1


$ cat abc.c
int a;
char b, e;
static char *c = &b;
static long d;
void f(void);
void h() {
  int g = 0;
  for (; g < 2; ++g) {
    d = *c;
    *c = 1;
    b = 0;
  }
  f();
}
void f() {
  if (d++)
    c = &e;
  for (; a;)
    ;
}
int main() {
  h();
  printf("%d\n", b);
}

[Bug tree-optimization/101031] wrong code at -O2 on x86_64-linux-gnu

[qrzhang at gatech dot edu]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101031

Qirun Zhang <qrzhang at gatech dot edu> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenther at suse dot de

--- Comment #1 from Qirun Zhang <qrzhang at gatech dot edu> ---
My bisection points to g:d1d01a66012a93cc8cb7dafbe1b5ec453e

[Bug tree-optimization/101031] [12 Regression] wrong code at -O2 on x86_64-linux-gnu

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101031

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
            Summary|wrong code at -O2 on        |[12 Regression] wrong code
                   |x86_64-linux-gnu            |at -O2 on x86_64-linux-gnu
            Version|unknown                     |12.0
           Priority|P3                          |P1
   Last reconfirmed|                            |2021-06-11
                 CC|                            |jakub at gcc dot gnu.org
   Target Milestone|---                         |12.0
             Status|UNCONFIRMED                 |NEW

[Bug tree-optimization/101031] [12 Regression] wrong code at -O2 on x86_64-linux-gnu

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101031

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed, mine.  Not a dup of PR101009.

[Bug tree-optimization/101031] [12 Regression] wrong code at -O2 on x86_64-linux-gnu

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101031

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Simpler IL:

int a;
char b, e;
static char *c = &b;
static long d;
void f(void);
void __attribute__((noipa)) h() {
  int g = 0;
  for (; g < 2; ++g) {
    d = *c;
    *c = 1;
    b = 0;
  }
  f();
}
void __attribute__((noipa)) f() {
  if (d++)
    c = &e;
  for (; a;)
    ;
}
int main() {
  h();
  if (b != 0)
    __builtin_abort ();
  return 0;
}

we end up with

void h ()
{
  char * c.0_14;
  char _24;
  long int _25;

  <bb 2> [local count: 357878152]:
  c.0_14 = c;
  *c.0_14 = 1;
  b = 0;
  _24 = *c.0_14;
  _25 = (long int) _24;
  d = _25;
  *c.0_14 = 1;
  f (); [tail call]
  return;

[Bug tree-optimization/101031] [12 Regression] wrong code at -O2 on x86_64-linux-gnu

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101031

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Interestingly the bogus store removal happens in the strlen pass.

> diff -u b/a-t3.c.190t.dom3 b/a-t3.c.191t.strlen1
--- b/a-t3.c.190t.dom3  2021-06-14 08:59:33.483709029 +0200
+++ b/a-t3.c.191t.strlen1       2021-06-14 08:59:33.487709082 +0200
@@ -78,7 +80,6 @@
   _25 = (long int) _24;
   d = _25;
   *c.0_14 = 1;
-  b = 0;
   f ();
   return;

in

#0  gsi_remove (i=0x7fffffffd830, remove_permanently=true)
    at /home/rguenther/src/gcc2/gcc/gimple-iterator.c:558
#1  0x0000000001619e63 in handle_store (gsi=0x7fffffffd830, 
    zero_write=0x7fffffffd7df, ptr_qry=...)
    at /home/rguenther/src/gcc2/gcc/tree-ssa-strlen.c:4852
#2  0x000000000161bb9a in check_and_optimize_stmt (gsi=0x7fffffffd830, 
    cleanup_eh=0x7fffffffd89f, ptr_qry=...)
    at /home/rguenther/src/gcc2/gcc/tree-ssa-strlen.c:5439
#3  0x000000000161c30c in strlen_dom_walker::before_dom_children (
    this=0x7fffffffd970, bb=<basic_block 0x7ffff6576270 (2)>)
    at /home/rguenther/src/gcc2/gcc/tree-ssa-strlen.c:5635
#4  0x000000000228e9b6 in dom_walker::walk (this=0x7fffffffd970, 
    bb=<basic_block 0x7ffff6576270 (2)>)
    at /home/rguenther/src/gcc2/gcc/domwalk.c:309

the *c stores have string info index 1 and the 'b' store have index 2.

Interestingly the first *c store is handled by handle_store but the
second is not and we return false from check_and_optimize_stmt.  In
particular we run into

      if (store_before_nul[1] > 0
          && storing_nonzero_p
          && lenrange[0] == lenrange[1]
          && lenrange[0] == lenrange[2]
          && TREE_CODE (TREE_TYPE (rhs)) == INTEGER_TYPE)
        {
          /* Handle a store of one or more non-nul characters that ends
             before the terminating nul of the destination and so does
             not affect its length
             If si->nonzero_chars > OFFSET, we aren't overwriting '\0',
             and if we aren't storing '\0', we know that the length of
             the string and any other zero terminated string in memory
             remains the same.  In that case we move to the next gimple
             statement and return to signal the caller that it shouldn't
             invalidate anything.
...
          gsi_next (gsi);
          return false;
        }

and skip the invalidation of 'b'.

Now, we're not invalidating the first *c store at the point we stroe to b
either because maybe_invalidate has

      ao_ref r;
      tree size = NULL_TREE;
      if (si->nonzero_chars)
        {
          /* Include the terminating nul in the size of the string
             to consider when determining possible clobber.  */
          tree type = TREE_TYPE (si->nonzero_chars);
          size = fold_build2 (PLUS_EXPR, type, si->nonzero_chars,
                              build_int_cst (type, 1));
        }
      ao_ref_init_from_ptr_and_size (&r, si->ptr, size);
      if (stmt_may_clobber_ref_p_1 (stmt, &r))

and thus we ask the oracle whether b = 0; may clobber a store/load of
*c of size 2 which it can't (an access of size two cannot refer to 'b').

Looking at compare_nonzero_chars the si->nonzero_chars check should likely
be si->nonzero_chars && !integer_zerop (si->nonzero_chars) as well.

Now, adding 1 is done for correctness AFAIU but then extending the access
beyond the size of the object is invalid.  So IMHO we have to adjust
max_size to be 1 bigger than size instead.

The following fixes the testcase:

diff --git a/gcc/tree-ssa-strlen.c b/gcc/tree-ssa-strlen.c
index 423075b2bd1..6add8c99032 100644
--- a/gcc/tree-ssa-strlen.c
+++ b/gcc/tree-ssa-strlen.c
@@ -1284,16 +1284,19 @@ maybe_invalidate (gimple *stmt, bool zero_write =
false)
        continue;

       ao_ref r;
-      tree size = NULL_TREE;
-      if (si->nonzero_chars)
+      tree size = si->nonzero_chars;
+      ao_ref_init_from_ptr_and_size (&r, si->ptr, size);
+      /* Include the terminating nul in the size of the string
+        to consider when determining possible clobber.  But do not
+        add it to 'size' since we don't know whether it would
+        actually fit the allocated area.  */
+      if (known_size_p (r.size))
        {
-         /* Include the terminating nul in the size of the string
-            to consider when determining possible clobber.  */
-         tree type = TREE_TYPE (si->nonzero_chars);
-         size = fold_build2 (PLUS_EXPR, type, si->nonzero_chars,
-                             build_int_cst (type, 1));
+         if (known_le (r.size, HOST_WIDE_INT_MAX - BITS_PER_UNIT))
+           r.max_size += BITS_PER_UNIT;
+         else
+           r.max_size = -1;
        }
-      ao_ref_init_from_ptr_and_size (&r, si->ptr, size);
       if (stmt_may_clobber_ref_p_1 (stmt, &r))
        {
          if (dump_file && (dump_flags & TDF_DETAILS))

[Bug tree-optimization/101031] [12 Regression] wrong code at -O2 on x86_64-linux-gnu

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101031

--- Comment #5 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:08ce1f4c5091b80b680d15c53a17237544a3cca8

commit r12-1419-g08ce1f4c5091b80b680d15c53a17237544a3cca8
Author: Richard Biener <rguenther@suse.de>
Date:   Mon Jun 14 09:37:24 2021 +0200

    tree-optimization/101031 - fix strlen opt invalidation logic

    strlen opt uses ao_ref_init_from_ptr_and_size to prepare alias
    queries to invalidate its knowledge about strings.  It constrains
    the size using the number of known-nonzero chars and adds one
    for a terminating nul - without knowing whether such nul exists
    or even fits the object.  The latter is now a problem since the
    oracle disambiguates an access of size two (as built so) against
    a store to a plain char variable (where a terminating nul does not
    fit).  The fix is to instead increment max_size but leave size to
    the number of chars we know are accessed.

    2021-06-14  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/101031
            * tree-ssa-strlen.c (maybe_invalidate): Increment max_size
            instead of size when accounting for a possibly string
            terminating nul.

            * gcc.dg/torture/pr101031.c: New testcase.

[Bug tree-optimization/101031] [12 Regression] wrong code at -O2 on x86_64-linux-gnu

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101031

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |FIXED

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed on trunk.