Interesting, I can't reproduce:

$ ./cc1 -quiet -Wimplicit-fallthrough ft.c
$

Oops. That must have been the "worked-around" version.
Swap the #undef and fallthrough comment to repro:

int
foo (int x)
{
  switch (x)
    {
    case 1:
      x = 3;
      /* fallthrough */
#undef X
    case 2:
      x = 4;
    }
  return x;
}
--------------------
$ /p/p/gcc-2016-10-02.12h48/libexec/gcc/x86_64-pc-linux-gnu/7.0.0/cc1 -Wimplicit-fallthrough /t/ft.c
 foo
Analyzing compilation unit

/t/ft.c: In function ‘foo’:
/t/ft.c:20:9: warning: this statement may fall through [-Wimplicit-fallthroug ]
       x = 3;
       ~~^~~
/t/ft.c:23:5: note: here
     case 2:
     ^~~~

Confirmed.  See also pr77853

While the comment after /* FALLTHRU */ comment is easily solvable, this one is I'm afraid going to be too difficult, there are many tokens in between and it is hard to propagate the info across the preprocessing directives.  Just use attributes?

Yeah, I think this one's WONTFIX.

(In reply to Marek Polacek from comment #5)
> Yeah, I think this one's WONTFIX.

Well, in theory we could for the comments that satisfy fallthrough_comment_p create CPP_COMMENT token and insert it, even without -C or -CC (perhaps only if -Wimplicit-fallthrough).  But, when without those options, we'd have to cope with the behavior changes where it makes a difference.  E.g. I believe
/* FALLTHROUGH */ #ifdef SOMETHING
etc. would change behavior.
As a workaround, with the patches I've posted -Wimplicit-fallthrough -C (or -CC) would work in this case.

Thanks for investigating.
I already pushed a workaround for gnulib's vasnprintf problem (http://git.savannah.gnu.org/cgit/gnulib.git/commit/?id=e63f5eb55570a1ea3e51ce47df33689785e085c1).
I may just leave that there.

The alternative of using an empty "__attribute__((__fallthrough__));" statement would require a macro expanding to that or to nothing.

I see the same in conditional code. Example fix that I applied to Qt locally, but won't push, because the next compiler will warn that the *comment is bogus*: 

       case QVariant::Bool:
           return qulonglong(d->data.b);
   #ifndef QT_BOOTSTRAPPED
       case QMetaType::QJsonValue:
           if (!v_cast<QJsonValue>(d)->isDouble())
               break;
  -        // fall through
   #endif
  +        // fall through
       case QVariant::Double:

We have about a dozen of those in QtBase alone, and while Qt 5.8 has a macro for the attribute, the 5.6 LTS version does not, and still has >2 years of support ahead of it, so a comment solution would be strongly preferred, and it would *really* help if this was fixed before this version of the compiler went into production.

*** Bug 77955 has been marked as a duplicate of this bug. ***

The testcase from PR77955:

markus@x4 /tmp % cat fall.c
void bar(int);

void foo(int i) {
  switch (i) {
  case 1: {
    bar(1);
    // fall-through
  }
  case 2:
    bar(2);
  default:
    break;
  }
}
markus@x4 /tmp % gcc -Wimplicit-fallthrough=1 -c fall.c
fall.c: In function ‘foo’:
fall.c:6:5: warning: this statement may fall through [-Wimplicit-fallthrough=]
     bar(1);
     ^~~~~~
fall.c:9:3: note: here
   case 2:
   ^~~~

Two issues; the first is the bogus warning, the second that we don't
warn for the case 2 fall-through.

(In reply to Markus Trippelsdorf from comment #10)
> The testcase from PR77955:
> 
> markus@x4 /tmp % cat fall.c
> void bar(int);
> 
> void foo(int i) {
>   switch (i) {
>   case 1: {
>     bar(1);
>     // fall-through
>   }
>   case 2:

Well, this one is even much harder than the preprocessor issue, it isn't solvable by preserving CPP_COMMENT tokens and ignoring them, there are tokens in between the fallthru comment and case keyword even after preprocessing here.
I'm afraid this is a clear WONTFIX, move the comment or better use __attribute__((fallthrough))/[[fallthrough]] instead.

>     bar(2);
>   default:
>     break;

Not warning here is completely intentional, it is a fallthru into break;

Is replacing a matching comment with __attribute__(fallthrough)) so complicated as to make this a wontfix?

(In reply to Marc Mutz from comment #12)
> Is replacing a matching comment with __attribute__(fallthrough)) so
> complicated as to make this a wontfix?

It is not really possible.
__attribute__((fallthrough)) has precise rules on where it can appear, while /* FALLTHRU */ comments, being comments, can appear anywhere.  Especially with -Wimplicit-fallthrough=1 when all comments are considered fallthru comments...

*** Bug 79750 has been marked as a duplicate of this bug. ***

(In reply to Jakub Jelinek from comment #13)
> (In reply to Marc Mutz from comment #12)
> > Is replacing a matching comment with __attribute__(fallthrough)) so
> > complicated as to make this a wontfix?
> 
> It is not really possible.
> __attribute__((fallthrough)) has precise rules on where it can appear, while
> /* FALLTHRU */ comments, being comments, can appear anywhere.  Especially
> with -Wimplicit-fallthrough=1 when all comments are considered fallthru
> comments...

I don't buy this. You don't need to use the 'official' attribute. You can invent an internal one, say __attribute__((__replaced_fallthrough_comment)), which does not have the limitations of [[fallthrough]]. It's just a way to preserve a comment over the preprocessor stage.

Hi,

I have the same problem:


int main() {
    int s = 1;

    switch (s) {
        case 2:
            s = 2;
#if B
            break;
#else
            s = 4;
            /* falls through */
#endif
        case 3:
            s = 3;
    }
} 

It appears

prog.c: In function 'main':
prog.c:11:15: warning: this statement may fall through [-Wimplicit-fallthrough=]
             s = 4;
             ~~^~~
prog.c:14:9: note: here
         case 3:
         ^~~~

At the moment I use a workaround:

#if (__GNUC__ >= 7)
#    define FALLTHROUGH __attribute__ ((fallthrough));
#else
#    define FALLTHROUGH
#endif

int main() {
    int s = 1;
    
    switch (s) {
        case 2:
            s = 2;
#if B
            break;
#else            
            s = 4;
            FALLTHROUGH
#endif
        case 3:   
            s = 3;
    }
}

This code is based on the real code from libexpat

(In reply to Anton Maklakov from comment #16)
> Hi,
> 
> I have the same problem:
> 
> 
> int main() {
>     int s = 1;
> 
>     switch (s) {
>         case 2:
>             s = 2;
> #if B
>             break;
> #else
>             s = 4;
>             /* falls through */
> #endif
>         case 3:

That works as documented:
'The comment needs to be followed after optional whitespace and other comments by "case" or "default" keywords or by a user label that precedes some "case" or "default" label.'

So, either do what you've mentioned, or just move the comment right before the case, where it is recognized (and not anywhere else).

Thank you for the answer!

I've got. Using the hint after #endif is not pretty good, by the logic. But this is easier.

*** Bug 90457 has been marked as a duplicate of this bug. ***

Re-doing some old CCs that somehow went missing without being marked as removed in the bug history (presumably due to the server transfer a few years ago)

*** Bug 111598 has been marked as a duplicate of this bug. ***