[Bug middle-end/57649] [4.7 Regression] ICE in set_is_used

[Bug middle-end/57649] [4.7 Regression] ICE in set_is_used

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57649

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |4.7.4

[Bug middle-end/57649] New: [4.7 Regression] ICE in set_is_used

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57649

            Bug ID: 57649
           Summary: [4.7 Regression] ICE in set_is_used
           Product: gcc
           Version: 4.7.3
            Status: UNCONFIRMED
          Keywords: ice-on-valid-code
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jakub at gcc dot gnu.org
                CC: matz at gcc dot gnu.org

void bar (const char *);
void
foo ()
{
  static const char *s = __PRETTY_FUNCTION__;
  struct T { T () { bar (s); } } t;
}

ICEs at -O2 starting with r179618 on 4.7 branch, went away with r187719 for
4.8+.

Similarly C -O2:
void bar (const char *);
void
foo (void)
{
  static const char *s = __PRETTY_FUNCTION__;
  auto void baz (void) { bar (s); }
  baz ();
}

The problem is that the s variable isn't referenced in foo during
*referenced_vars pass, is referenced in another function (nested or local class
method), but at that point the DECL_INITIAL of the var isn't marked, and later
on when the nested/local class method is inlined nothing marks the
DECL_INITIAL.
Perhaps the inliner should somewhere go through all gimple_referenced_vars of
the callee and for each TREE_STATIC var with DECL_CONTEXT of the caller just
call add_referenced_var?

[Bug middle-end/57649] [4.7 Regression] ICE in set_is_used

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57649

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-06-19
                 CC|                            |ebotcazou at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #1 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
We have this patchlet in our 4.7 tree:

--- a/gcc/tree-dfa.c
+++ b/gcc/tree-dfa.c
@@ -590,10 +590,13 @@ add_referenced_var (tree var)
       /* Scan DECL_INITIAL for pointer variables as they may contain
         address arithmetic referencing the address of other
         variables.  As we are only interested in directly referenced
-        globals or referenced locals restrict this to initializers
-        than can refer to local variables.  */
+        globals or referenced locals, restrict this to initializers
+        than can refer to local variables.  But we need to do it for
+        initializers of global variables declared in parent functions
+        lest these variables aren't referenced anywhere else.  */
       if (DECL_INITIAL (var)
-          && DECL_CONTEXT (var) == current_function_decl)
+         && (DECL_CONTEXT (var) == current_function_decl
+             || (!DECL_FILE_SCOPE_P (var) && is_global_var (var))))
        walk_tree (&DECL_INITIAL (var), find_vars_r, NULL, 0);

       return true;


    * tree-dfa.c (add_referenced_var): Find variables referenced in the
    initializers of global variables declared in parent functions.

[Bug middle-end/57649] [4.7 Regression] ICE in set_is_used

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57649

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Actually, it ICEs even without inlining, so guess the issue is just that we
have a local static var that is referenced by some other function (thus used)
and we don't mark its DECL_INITIAL as referenced.

[Bug middle-end/57649] [4.7 Regression] ICE in set_is_used

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57649

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Other testcases without __PRETTY_FUNCTION__:

void bar (const char *);
void
foo (void)
{
  static const char r[] = "abc";
  static const char *s = t;
  auto void baz (void) { bar (s); }
  baz ();
}

void bar (const char *);
void
foo ()
{
  static const char r[] = "abc";
  static const char *s = r;
  struct T { T () { bar (s); } } t;
}

[Bug middle-end/57649] [4.7 Regression] ICE in set_is_used

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57649

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
How about backporting r187719?  I don't remember whether there are any
dependencies here, but ... it seems it's the first patch in the series
to cleanup all this stuff.

[Bug middle-end/57649] [4.7 Regression] ICE in set_is_used

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57649

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Isn't it going to cause problems if we fold some accesses to the static var
using the expressions in the initializer, we'd end up with not having the vars
referenced?

If backporting anything, my preference would be to actually backport just the
tree-ssa-live.c change and not the tree-dfa.c change.
Or Eric's patch.

[Bug middle-end/57649] [4.7 Regression] ICE in set_is_used

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57649

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #5)
> Isn't it going to cause problems if we fold some accesses to the static var
> using the expressions in the initializer, we'd end up with not having the
> vars referenced?

Well - we are carefully trying to add them at the point we fold ...

> If backporting anything, my preference would be to actually backport just
> the tree-ssa-live.c change and not the tree-dfa.c change.
> Or Eric's patch.

Backporting just the tree-ssa-live.c change also works for me.

[Bug middle-end/57649] [4.7 Regression] ICE in set_is_used

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57649

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |FIXED
   Target Milestone|4.7.4                       |4.8.0
      Known to fail|                            |4.7.4

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed in 4.8.0.