[Bug rtl-optimization/66790] Invalid uninitialized register handling in REE

[bernds at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66790

--- Comment #28 from Bernd Schmidt <bernds at gcc dot gnu.org> ---
(In reply to Pierre-Marie de Rodat from comment #27)
> (In reply to Bernd Schmidt from comment #14)
> > It looks like your "visited" bit tries to avoid that, but I don't
> > think this works.
> (In reply to Paolo Bonzini from comment #19)
> > There's no code in GCC for must-initialized.  Pierre's patch gets it right
> > (except that I'm not sure about the comment at line 2033).
> The "visited" bit I introduced is used to fake an "all bits set"
> initialization. The confluence_n method is meant to do:
> 
>     IN(dest) = OUT(src_0) AND OUT(src_1) AND ...
> 
> ... by small increments of:
> 
>     IN(dest) := IN(dest) AND OUT(src_i)
> 
> When it is executed for the first time, IN(dest) is all-zeros, so we do
> a copy because "a AND all_ones = a". Then we can do regular AND later
> on. In order to get rid of this "visited" bit, I guess we could instead
> pretend that IN/OUT sets are set of never-or-may-initialized registers,
> but this complicates other parts of the code, so...

I know this is what it's trying to do, but when I experimented with it that
wasn't working, due to the problem I described.

> If I understand correctly, the BB_LAST_CHANGE_AGE mechanism in df-core.h
> makes sure computations comply with this dependency (c. before b., b.
> before a.).  So at the point the confluence_n function for the 3 -> 3
> edges is called for the first time, both IN(3) and OUT(3) can be not
> all-zeros.

It is sufficient for OUT(3) to be all-zeros. And I don't think the
LAST_CHANGE_AGE mechanism does anything to prevent it. Please try it out. I
think you have to initialize your bitmaps correctly rather than rely on
"visited".

I applied your patch and compiled the following with a breakpoint on
df_mir_confluence_n:
int foo (int n)
{
  int sum = 0;
  for (int i = 0; i < n; i++)
    sum += i;
  return sum;
}

and when it reaches the loop:

Breakpoint 5, df_mir_confluence_n (e=<edge 0x7ffff07352d8 (4 -> 4)>) at
../../trunk/gcc/df-problems.c:2040
(gdb) p debug_bitmap (op1)

first = 0x24bf418 current = 0x24bf418 indx = 0
        0x24bf418 next = (nil) prev = (nil) indx = 0
                bits = { 0 1 2 4 5 7 17 21 22 23 24 25
                         26 27 28 37 38 }
$2 = void
(gdb) p debug_bitmap (op2)

first = (nil) current = (nil) indx = 0

and

(gdb) p bb_info->visited 
$5 = true

because the edge 3->4 was seen first.

> (In reply to Bernd Schmidt from comment #14)
> > I do have to say that I am still uncomfortable with changing RRE to

I did not write this.