Re: alias/flow patch for PR 3783

[Kazu Hirata <kazu at hxi dot com>]

Hi,

With the patch

http://gcc.gnu.org/ml/gcc-patches/2001-09/msg00252.html

the following code, a simplified version of gmtime.c from newlib,
generates a segfault when compiled with h8300-hms-gcc -mh -O2
(mainline).

struct dummy {
  int element;
};

char *foo (const int *);

char *
bar (const int *p)
{
  int tim = *p;

  return foo (&tim);
}

There is a backtrace with gdb.

Program received signal SIGSEGV, Segmentation fault.
remove_edge (e=0x4018ff40) at /home/kazu/gnu/gcc/gcc-head/gcc/gcc/flow.c:8907
8907      for (tmp = src->succ; tmp && tmp != e; tmp = tmp->succ_next)
(gdb) backtrace
#0  remove_edge (e=0x4018ff40) at /home/kazu/gnu/gcc/gcc-head/gcc/gcc/flow.c:8907
#1  0x80e444e in purge_dead_edges (bb=0x826d590) at /home/kazu/gnu/gcc/gcc-head/gcc/gcc/flow.c:10379
#2  0x80e4509 in purge_all_dead_edges () at /home/kazu/gnu/gcc/gcc-head/gcc/gcc/flow.c:10402
During symbol reading, type qualifier 'volatile' ignored.
#3  0x81642e2 in rest_of_compilation (decl=0x4017bc3c)
    at /home/kazu/gnu/gcc/gcc-head/gcc/gcc/toplev.c:3273
During symbol reading, type qualifier 'volatile' ignored.
#4  0x8060116 in c_expand_body (fndecl=0x4017bc3c, nested_p=0)
    at /home/kazu/gnu/gcc/gcc-head/gcc/gcc/c-decl.c:6724
#5  0x805fee2 in finish_function (nested=0) at /home/kazu/gnu/gcc/gcc-head/gcc/gcc/c-decl.c:6641
During symbol reading, type qualifier 'volatile' ignored.
#6  0x8049fd4 in yyparse_1 () at c-parse.y:358
#7  0x8162c33 in compile_file (name=0x8253be8 "kazu.c")
    at /home/kazu/gnu/gcc/gcc-head/gcc/gcc/toplev.c:2319
#8  0x81665bf in toplev_main (argc=4, argv=0xbffff9d4)
    at /home/kazu/gnu/gcc/gcc-head/gcc/gcc/toplev.c:4928
#9  0x4003d177 in __libc_start_main (main=0x8090464 <main>, argc=4, ubp_av=0xbffff9d4,
    init=0x804901c <_init>, fini=0x81d4d60 <_fini>, rtld_fini=0x4000e184 <_dl_fini>,
    stack_end=0xbffff9cc) at ../sysdeps/generic/libc-start.c:129
(gdb) print e
$4 = (struct edge_def *) 0x4018ff40
(gdb) print e->src
$5 = (struct basic_block_def *) 0x0
(gdb)

What's happening at the end is accessing src->succ even though src is
NULL.  I have not figured out why this happens.

One very weired thing is that removing the dummy struct definition
silences the segfault even though I don't see any connection between
the struct and the function bar.

Without reg_alias_info (q); around flow.c:7030, the segfault does not
happen, but this is pretty much the same as reversing the patch...

I would appreciate if somebody could take a look this.

Thanks,

Kazu Hirata