Well, if the user knows that the pointer is usually NULL, then if (p) free (p); is a significant optimization over free (p).  So as code size optimization, removing it is fine, but otherwise, it isn't at all clear.

I think a warning like this might be useful, although I probably wouldn't want to enable under -Wall or -Wextra.   (GCC could also use branch probabilities to decide whether or not to warn.)

That said, I would expect GCC to get rid of the tests with -fdelete-null-pointer-checks.  As it is, it eliminates the call to free when the pointer is known to be null but not the test itself (see below).  That seems quite suboptimal to me since in the absence of any other information a pointer being passed to free is far more likely to be non-null than not.  Clang and ICC both optimize the function below into an unconditional jump to free.  So while I think implementing the optimization is the better solution here, the warning could also help improve the efficiency of the emitted code in its absence.

$ cat b.c && gcc -O2 -S -Wall -Wextra -fdump-tree-optimized=/dev/stdout  b.c
void free (void*);

void g (void *p)
{
  if (p)        // test retained
    free (p);
  else
    free (p);   // eliminated
}


;; Function g (g, funcdef_no=0, decl_uid=1817, cgraph_uid=0, symbol_order=0)

Removing basic block 5
g (void * p)
{
  <bb 2> [100.00%] [count: INV]:
  if (p_2(D) != 0B)
    goto <bb 3>; [53.47%] [count: INV]
  else
    goto <bb 4>; [46.53%] [count: INV]

  <bb 3> [53.47%] [count: INV]:
  free (p_2(D)); [tail call]

  <bb 4> [100.00%] [count: INV]:
  return;

}

cc-ing Jim Meyering, the author of the original gnulib script, to see if he has any input

Thanks for considering the addition. IME, the vast majority (probably "all I've seen") of such "useless" if stmts have been attempts to avoid what used to be UB/segfault on old systems, or simply due to people not realizing that free(NULL) is now known to be ok on all reasonable portability targets.

One thing to consider now that GCC has a static analyzer is that putting the kind of useless "if" in question before a "free" can make the analyzer shut up about possible double-free warnings, so if this were to become its own diagnostic, there might be separate diagnostics that prompt users to do opposite things...

Recent patchset for fixing problems in GCC based on this same principle: https://gcc.gnu.org/pipermail/gcc-patches/2023-March/613116.html