Created attachment 17393 [details]
the preprocessed source

Reducing.

It is most likely caused by revision 144497:

http://gcc.gnu.org/ml/gcc-cvs/2009-02/msg00665.html

In TODO_recompute_alias after PRE we have something (alc_capture_mixer1, found
through DECL_INITIAL of caps) in gimple_addressable_vars that is no longer available.

Honza, this is yours again.  I guess during inlining we no longer add
referenced vars to the inlined-to function if they are addresses of globals
referenced through initializers of local statics?

static struct snd_kcontrol_new alc_capture_mixer1[];

static void set_capture_mixer(struct alc_spec *spec)
{
 static struct snd_kcontrol_new *caps[3] = {
  alc_capture_mixer1,
  alc_capture_mixer2,
  alc_capture_mixer3,
 };
 if (spec->num_adc_nids > 0 && spec->num_adc_nids <= 3)
  spec->cap_mixer = caps[spec->num_adc_nids - 1];
}

inlined into some other function.  Proper testcase still reducing.

Reduced testcase:

typedef unsigned short u16;
struct snd_kcontrol_new { };
typedef u16 hda_nid_t;
struct alc_spec {
    struct snd_kcontrol_new *cap_mixer;
    unsigned int num_adc_nids;
};
static struct snd_kcontrol_new alc_capture_mixer1[] = { };
static struct snd_kcontrol_new alc_capture_mixer2[] = { };
static struct snd_kcontrol_new alc_capture_mixer3[] = { };
static void set_capture_mixer(struct alc_spec *spec)
{
  static struct snd_kcontrol_new *caps[3] = {
      alc_capture_mixer1,   alc_capture_mixer2,   alc_capture_mixer3
  };
  if (spec->num_adc_nids > 0 && spec->num_adc_nids <= 3)
    spec->cap_mixer = caps[spec->num_adc_nids - 1];
}
static hda_nid_t alc662_adc_nids[1] = { 0x09, };
int patch_alc662(struct hda_codec *codec)
{
  struct alc_spec *spec;
  spec->num_adc_nids = (sizeof(alc662_adc_nids) / sizeof((alc662_adc_nids)[0])
                        + (sizeof(char[1 - 2 * !!(__builtin_types_compatible_p(typeof(alc662_adc_nids), typeof(&alc662_adc_nids[0])))]) - 1));
  if (!spec->cap_mixer)
    set_capture_mixer(spec);
}

This is curious, since we should see the initializer when adding variable at first time.  I am looking into this.

Honza

Smaller testcase:
static int a[] = { 1 };

static inline void
bar (int **x)
{
  static int *c[2] = { 0, a };
  *x = c[1];
}

int
foo (int **x)
{
  bar (x);
}

I think it is the new:
3384      if (TREE_STATIC (var) && !TREE_ASM_WRITTEN (var))
3385        {
3386          if (var_ann (var) && referenced_var_check_and_insert (var))
3387            cfun->local_decls = tree_cons (NULL_TREE, var,
3388          cfun->local_decls);
3389        }
that causes this.  Compared to add_referenced_var this doesn't do the:
      /* Scan DECL_INITIAL for pointer variables as they may contain
         address arithmetic referencing the address of other
         variables.
         Even non-constant initializers need to be walked, because
         IPA passes might prove that their are invariant later on.  */
      if (DECL_INITIAL (var)
          /* Initializers of external variables are not useful to the
             optimizers.  */
          && !DECL_EXTERNAL (var))
        walk_tree (&DECL_INITIAL (var), find_vars_r, NULL, 0);
part.  So, either tree-inline.c needs to do the same, or it can't use referenced_vars bit as a test whether it has been queued already onto local_decls or not.  Honza?

Subject: Re:  [4.4 Regression] ICE in referenced_var_lookup, at tree-dfa.c:563

> part.  So, either tree-inline.c needs to do the same, or it can't use
> referenced_vars bit as a test whether it has been queued already onto
> local_decls or not.  Honza?

Ah I see.  Actually original version of patch added function to tree-dfa
to check if variable is referenced and used this predicate followed by
add_referenced_var if it was not, but then I noticed that
referenced_var_check_and_insert is available to I decided to use it.
OK, I will prepare this version of patch tomorrow.

To "do the same" the easiest would be probably change add_referenced_var to return bool, whatever referenced_var_check_and_insert returned, and use add_referenced_var instead of referenced_var_check_and_insert in expand_call_inline.

Created attachment 17412 [details]
gcc44-pr39360.patch

Patch I'm bootstrapping/regtesting now.

Subject: Bug 39360

Author: jakub
Date: Fri Mar  6 22:51:28 2009
New Revision: 144683

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=144683
Log:
	PR middle-end/39360
	* tree-flow.h (add_referenced_var): Return bool instead of void.
	* tree-dfa.c (add_referenced_var): Return result of
	referenced_var_check_and_insert call.
	* tree-inline.c (expand_call_inline): Call add_referenced_var instead
	of referenced_var_check_and_insert.

	* gcc.c-torture/compile/pr39360.c: New test.

Added:
    trunk/gcc/testsuite/gcc.c-torture/compile/pr39360.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree-dfa.c
    trunk/gcc/tree-flow.h
    trunk/gcc/tree-inline.c

Fixed.

thanks a lot.