Created attachment 6764 [details]
preprocessed source

produced with Linux 3.3.4 compiler.

It is fixed in 3.4.0 but worked in 3.0.4 so this is a regression.
I have not reduced the testcase yet.

Confirmed. In fact, with the following snippet I can also crash 3.4, so 
it doesn't seem to be fixed there but only papered over: 
---------------- 
struct Ptrs { 
    int (*ptr1)(int offset); 
    int (*ptr2)(); 
} * ptrs; 
static int ** state; 
 
int foo(int stackIndex) { 
  int *p; 
  int sz; 
  int objOop; 
  int i; 
  int **pp; 
  ptrs->ptr1(stackIndex); 
  if (ptrs->ptr2()) { 
    return 0; 
  } 
  if (!((ptrs->ptr2()) == 0)) { 
    return 0; 
  } 
  sz = ptrs->ptr2(); 
  if (sz > 1)  
    return 0; 
        
  pp = state; 
  for (i = 0; i <= sz-1; ++i) { 
    objOop = ptrs->ptr1(0); 
    if (!ptrs->ptr2()) 
      return 0;	 
                
    *p = objOop; 
    pp[i] = p; 
  } 
} 
-------------------------- 
 
g/x> /home/bangerth/bin/gcc-3.3.4-pre/bin/gcc -O3 -c -fno-gcse x.i  
x.i: In function `foo': 
x.i:33: internal compiler error: in verify_local_live_at_start, at flow.c:575 
Please submit a full bug report, 
with preprocessed source if appropriate. 
See <URL:http://gcc.gnu.org/bugs.html> for instructions. 
g/x>  
g/x> /home/bangerth/bin/gcc-3.4-pre/bin/gcc -O3 -c -fno-gcse x.i  
x.i: In function `foo': 
x.i:33: internal compiler error: in verify_local_live_at_start, at flow.c:546 
Please submit a full bug report, 
with preprocessed source if appropriate. 
See <URL:http://gcc.gnu.org/bugs.html> for instructions. 
 
W. 

: Test is the same (pass) at both endpoints (2004-02-01-3.4,
: 2004-07-27-3.4), so there's nothing for me to do.

Broken on the mainline:
: Search converges between 2001-06-03-trunk (#22) and 2001-06-10-trunk (#23).

Fixed on the mainline:
: Search converges between 2003-06-04-trunk (#254) and 2003-06-05-trunk (#255).
So in my testing (on Phil's regression hunter) this is only a 3.3 regression but this was without -
march=pentiumpro which is what fixed it without using the driver (or something like that)

With using -march=pentiumpro:
: Search converges between 2004-04-20-trunk (#449) and 2004-05-01-trunk (#450).
This might still a bug on the mainline also but harder to get to as we get rid of some code before 
going into the RTL.

Moving the target to 3.4.2 per what Gaby said about regressions that are on both 3.3.x and 3.4.x.

Here's a slightly shorter testcase. It crashes everything from 3.1 until
3.4 branch.

======================================
int (**foo)(int);

int bar()
{
    static int **s;
    int i, j, k=i, **p;

    (*foo)(**s);
    if ((*foo)(0)) return 0;
    if (k > 0) return 0;

    p = s;
    for (i = 0; i <= k-1; ++i)
    {
        j = (*foo)(0);
        if ((*foo)(0)) return 0;
        **p = j;
    }
}
======================================

Postponed until GCC 3.4.3.

Postponed until GCC 3.4.3.

The testcases in comment #3 and #5 got fixed on the 3.4 branch by
Richard's patch for PR 17186:
http://gcc.gnu.org/ml/gcc-cvs/2004-09/msg00357.html

Richard, do you think a backport to the 3.3 branch is feasible?

It wouldn't be hard to backport to 3.3, no.

Since Richard reverted his patch for PR 17186 on the 3.4 branch,
this PR fails on the 3.4 branch again, too.
In addition, comment #9 is moot now.

Postponed until GCC 3.4.4.

This one is really tricky to fix: this is the classical case in which the return
value register's liveness propagates backwards until an unrelated use/clobber of
the register is encountered.  This works more or less until the regrename pass,
which locally renames the register in the latter def/use/clobber chain; so the
return value register's liveness is not masked anymore and propagates back to
the start of the basic block.  But, after reload, register liveness information
is not allowed to change at basic block boundaries so the compiler aborts on a
sanity check.

The easy workaround is to make sure that every code path in a function returning
a value really returns a value.

Downgrading to 'normal', this is an ICE whose workaround is obvious.

Created attachment 8448 [details]
simplified c++ testcase that appears to trigger the same ICE.

Attached is a simplified c++ testcase from the xorg-x11 6.8.1 distribution that
_appears_ to trigger the same ICE.  It needs gcc 3.3 or 3.4, -O3, and any ARM
target.  (It doesn't trigger when building for x86, when using -O2, or when
using gcc 2.95.3 or 4.0-20050305.  Didn't test other gcc versions.)

won't fix for 3.4.6.