[Bug c/31878] New: Spurious warnings generated due to not optimizing first

[Bug c/31878] New: Spurious warnings generated due to not optimizing first

[lloyd at randombit dot net]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 1735 bytes --]

$ cat return.c
#include <assert.h>

int f(int x) {
    if(x)
        return x;
    assert(x);
}

$ gcc -O2 -c -Wall -W return.c 
return.c: In function ‘int f(int)’:
return.c:9: warning: control reaches end of non-void function

The call to assert expands to (glibc 2.4):

    (static_cast<void> ((x) ? 0 : (__assert_fail ("x", "return.c", 8,
__PRETTY_FUNCTION__), 0)));

So unless I'm missing something f() is equivalent to:

if(x)
  return x;
else
  __assert_fail(/* etc */);

However when explicitly written in this way, we don't get a warning:

$ cat return2.c
#include <assert.h>

int f(int x)
{
    if(x)
        return x;
    else
        __assert_fail ("x", __FILE__, __LINE__, __PRETTY_FUNCTION__);
}

$ gcc -O2 -Wall -c return2.c
$

Looking at the produced assembly on an x86-64 system, GCC is only performing
the comparison once (when optimizing), so it should see that there is no
codepath that doesn't either return or call a noreturn function. However I'm
guessing warnings are generated prior to the optimizer running and thus don't
see this?

Marked as Severity: trivial because one can always insert a dummy return after
the assert. Which is good because AFAICT fixing this would require rewriting
GCC. :)


-- 
           Summary: Spurious warnings generated due to not optimizing first
           Product: gcc
           Version: 4.1.0
            Status: UNCONFIRMED
          Severity: trivial
          Priority: P3
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: lloyd at randombit dot net
 GCC build triplet: x86_64-redhat-linux
  GCC host triplet: x86_64-redhat-linux
GCC target triplet: x86_64-redhat-linux


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31878

[Bug c/31878] Spurious warnings generated due to not optimizing first

[pinskia at gcc dot gnu dot org]

------- Comment #1 from pinskia at gcc dot gnu dot org  2007-05-09 22:07 -------
We should get the same warning while compiling with optimizations and while
compiling without so I don't think this is a bug.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31878

[Bug c/31878] Spurious warnings generated due to not optimizing first

[lloyd at randombit dot net]

------- Comment #2 from lloyd at randombit dot net  2007-05-09 22:16 -------
So are you saying that it is the case that the f() function below might return
without a value? Since that is what the warning suggests.

(My interpretation re the optimizer may be completely off, I don't know GCC
internals, but I have a hard time believing that GCC warning about something
that cannot possibly happen is not a bug)


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31878

[Bug c/31878] Spurious warnings generated due to not optimizing first

[manu at gcc dot gnu dot org]

------- Comment #3 from manu at gcc dot gnu dot org  2007-05-10 15:50 -------
(In reply to comment #2)
> So are you saying that it is the case that the f() function below might return
> without a value? Since that is what the warning suggests.

assert() does not return a value and it is the last . So, yes, as far as GCC
knows when it emits the warning, f() may return without a value. Another
example:

int f(int x)
{
    int y = 1;
    if(y)
        return x;
}

> (My interpretation re the optimizer may be completely off, I don't know GCC
> internals, but I have a hard time believing that GCC warning about something
> that cannot possibly happen is not a bug)

You may know that it cannot possibly happen but GCC may not know it. In the
general case, it is the halting problem. In practice, there should be a limit
to how much effort GCC should do. The current limit is that warnings (in
general) are given by front-ends and should be the same with and without
optimisation. But even if that weren't the case, the warning currently comes
before optimisation and changing this is not simple at all. Finally, you could
always build a testcase where GCC is not smart enough.

I agree with Andrew: this is not a bug. At most, it is an enhancement request
that conflicts with the current general rule that front-end warnings should be
the same with and without optimisation.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31878

[Bug c/31878] Spurious warnings generated due to not optimizing first

[lloyd at randombit dot net]

------- Comment #4 from lloyd at randombit dot net  2007-05-10 17:51 -------
Manuel,

For your example code, GCC _is_ aware that the function always returns, since
the code it generates for it (with optimization) is:

f:
        movl    %edi, %eax
        ret

So obviously it knows, at the level of the code generator, it's just a question
of propagating that information back to the frontend.

Speaking of the same warnings with-or-without optimizations - should I then
file a bug about:

int f(int x)
{
    int y;
    if(x)
        return y;
    return 0;
}

No warning about y being used uninitialized unless I compile with
optimizations...


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31878

[Bug c/31878] Spurious warnings generated due to not optimizing first

[manu at gcc dot gnu dot org]

------- Comment #5 from manu at gcc dot gnu dot org  2007-05-11 11:38 -------
(In reply to comment #4)
> So obviously it knows, at the level of the code generator, it's just a question
> of propagating that information back to the frontend.

I wrote: "yes, as far as GCC knows when it emits the warning, f() may return
without a value". So obviously, code generation happens way after the warning
is emitted. There is no "propagating back information to the front-end". Sorry.
This is not how GCC works. So, yes, fixing this may require rewriting GCC. But
even if it were trivial, that is, even if when using optimisation we could
detect that no warning is needed, there is the current policy of giving (in
general) the same warnings with and without optimisation. Thus, GCC may still
decide to give a warning.

So I still don't think this is worthy to be called a bug, at best it is
WONTFIX. But, hey!, if you want to keep it open in case some brave volunteer
developer appears from the void and decides to contribute some magical patch...
who knows, the hero might be you. ;-)

> Speaking of the same warnings with-or-without optimizations - should I then
> file a bug about:
[snip]
> No warning about y being used uninitialized unless I compile with
> optimizations...

Currently, -Wuninitialized needs optimisation to do any work. This may change
in the future, though. You chose one of the few exceptions to illustrate your
point. 


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31878

[Bug middle-end/31878] Spurious warnings generated due to not optimizing first

[pinskia at gcc dot gnu dot org]

------- Comment #6 from pinskia at gcc dot gnu dot org  2008-11-14 23:38 -------
This needs VRP really to be run at -O0 which I don't think is going to happen
any time soon.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|trivial                     |enhancement
          Component|c                           |middle-end
  GCC build triplet|x86_64-redhat-linux         |
   GCC host triplet|x86_64-redhat-linux         |
 GCC target triplet|x86_64-redhat-linux         |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31878

[Bug middle-end/31878] Spurious warnings with -Wreturn-type due to not performing CCP/VRP in the front-end

[manu at gcc dot gnu dot org]

------- Comment #7 from manu at gcc dot gnu dot org  2009-02-12 14:59 -------
Updating the description. Although anyone would agree that not warning is the
right thing to do, there doesn't seem to be any practical solution for this
problem, so I leave it as unconfirmed.

I wonder if clang/llvm gets this right.


-- 

manu at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |manu at gcc dot gnu dot org
            Summary|Spurious warnings generated |Spurious warnings with -
                   |due to not optimizing first |Wreturn-type due to not
                   |                            |performing CCP/VRP in the
                   |                            |front-end


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31878

[Bug middle-end/31878] Spurious warnings with -Wreturn-type due to not performing CCP/VRP in the front-end

[sabre at nondot dot org]

------- Comment #8 from sabre at nondot dot org  2009-12-29 06:07 -------
Clang does not produce diagnostics from the optimizer.  We intentionally do not
want to do this, because of the fragility of the diagnostics as the optimizer
evolves.  IMO it is much better to produce a consistent and predictable set of
diagnostics rather than trying to avoid all false positives by relying on the
optimizer to be tricky.


-- 

sabre at nondot dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |sabre at nondot dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31878

[Bug middle-end/31878] Spurious warnings with -Wreturn-type due to not performing CCP/VRP in the front-end

[manu at gcc dot gnu dot org]

------- Comment #9 from manu at gcc dot gnu dot org  2009-12-30 01:43 -------
(In reply to comment #8)
So the answer is that Clang also warns. I was wondering whether the minimal ccp
propagation that Clang does in the front-end would catch this. I guess the
answer is no. 

As for GCC, the general wisdom is to not move more warnings to the middle-end
unless really awful false positives/negatives, and to not perform any expensive
optimizations at -O0. So unless someone comes up with a more convincing
testcase, or an inexpensive way to detect this at -O0, this won't be fixed.

Closing then, we have far enough real bugs to worry about.


-- 

manu at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |WONTFIX


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31878

[Bug middle-end/31878] Spurious warnings with -Wreturn-type due to not performing CCP/VRP in the front-end

[sabre at nondot dot org]

------- Comment #10 from sabre at nondot dot org  2009-12-30 01:46 -------
Yes, clang warns:
$ clang t.c
t.c:7:1: warning: control may reach end of non-void function [-Wreturn-type]
}
^


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31878