[Bug tree-optimization/98552] New: Make more use of __builtin_undefined for assuring that variables do not change

[Bug tree-optimization/98552] New: Make more use of __builtin_undefined for assuring that variables do not change

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98552

            Bug ID: 98552
           Summary: Make more use of __builtin_undefined for assuring that
                    variables do not change
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: tkoenig at gcc dot gnu.org
  Target Milestone: ---

Consider

void foo (int *);

void bar (int n)
{
  int i;
  for (i=0; i<n; i++)
    foo(&i);
}

void baz(int n)
{
  int i, j;
  for (i=0; i<n; i++)
    {
      j = i;
      foo (&i);
      if (j /= i)
        __builtin_unreachable();
    }
}

Assembly for bar and baz are identical, the loop is

.L9:
        leaq    12(%rsp), %rdi
        call    foo
        movl    12(%rsp), %eax
        addl    $1, %eax
        movl    %eax, 12(%rsp)
        cmpl    %ebx, %eax
        jl      .L9

In function bar, things are clear - the value of i has to be
reloaded from the stack, foo might have changed it.

However, this is not possible in baz.  j cannot be changed, and the
__builtin_unreachable ensures that i has the same value before and
after the call to foo. It need not be reloaded from the stack.

(The reason why I'm submitting this is another way to approach PR 31593 -
the Fortran front end could annotate code like this to inform
the middle end that DO loops variables are, in fact, invariant).

[Bug tree-optimization/98552] Make more use of __builtin_undefined for assuring that variables do not change

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98552

Thomas Koenig <tkoenig at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Version|unknown                     |11.0
           Keywords|                            |missed-optimization
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=31593
           Severity|normal                      |enhancement

[Bug tree-optimization/98552] Make more use of __builtin_undefined for assuring that variables do not change

[tobi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98552

Tobias Schlüter <tobi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |tobi at gcc dot gnu.org

--- Comment #1 from Tobias Schlüter <tobi at gcc dot gnu.org> ---
There's a typo in the example, /= instead of !=.  Fixed example below:

I sprinkled const's all over foo's prototype and 'i' still gets reloaded so it
stands to reason that the compiler doesn't see the optimization opportunity. 

======================

void foo (int *);

void bar (int n)
{
  int i;
  for (i=0; i<n; i++)
    foo(&i);
}

void baz(int n)
{
  int i, j;
  for (i=0; i<n; i++)
    {
      j = i;
      foo (&i);
      if (j != i)
        __builtin_unreachable();
    }
}

======================

[Bug tree-optimization/98552] Make more use of __builtin_undefined for assuring that variables do not change

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98552

--- Comment #2 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
(In reply to Tobias Schlüter from comment #1)
> There's a typo in the example, /= instead of !=.  Fixed example below:

The disease of a Fortran programmer writing C, I guess :-)

[Bug tree-optimization/98552] Make more use of __builtin_undefined for assuring that variables do not change

[tobi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98552

--- Comment #3 from Tobias Schlüter <tobi at gcc dot gnu.org> ---
Don't ask how long I'd been staring at the assembly in disbelief until I
figured out what had gone wrong :-)

In this particular case the problem can be addressed by passing &j into the
function instead of &i, so I wonder why the Fortran f.e. can't do that?  One
possible problem is that if j's address escapes (in the sense that the compiler
assumes that subsequent function calls can do whatever to it), you'd need a new
fake memory location.

[Bug tree-optimization/98552] Make more use of __builtin_undefined for assuring that variables do not change

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98552

--- Comment #4 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
Yes, I think that translating a DO loop into something like

int i;

for (i=0; i<n; i++)
{
  const int j = i;
  /* Use j from here on where i is referenced.  */
}

might solve a few of the issues in DO loop optimization.

[Bug tree-optimization/98552] Make more use of __builtin_undefined for assuring that variables do not change

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98552

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2021-12-23

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed.