Re: [Emacs-diffs] master 6cd5678: Clarify compiler-pacifier in frame.c

[Paul Eggert <eggert@cs.ucla.edu>]

[-- Attachment #1: Type: text/plain, Size: 2142 bytes --]

Eli Zaretskii wrote:

> Do you have a guess why GCC might get lost in that code?

Sure, lots of guesses. GCC may have a limit on how many variables it's willing 
to analyze before it gives up. Or maybe it has a limit on the complexity of the 
mask being used. Or maybe it's just using range and oddity analyses and that 
works only up to two-bit masks. (I should say that GCC is "lost" only in the 
sense that it's issuing bogus warnings; it's still generating correct code.)

To illustrate, please see the attached file maybe-uninitialized.c. When I 
compile it with 'gcc -O2 -Wmaybe-uninitialized -S maybe-uninitialized.c' on 
Fedora 30 x86-64, GCC warns that w, x, and y might be used uninitialized in 
bar4. It doesn't warn about h in bar4 even though h is initialized using the 
same strategy as the others. GCC also doesn't warn about bar2's locals even 
though bar 2 uses the same strategy as bar4. All the warnings GCC issues are 
false alarms, and these false alarms come from bar4 (with 4 variables) rather 
than from bar2 (with only 2).

>>> And how should GCC know that?
>>
>> GCC could use the same sort of reasoning I used.
> 
> Which reasoning is that?  You haven't presented your reasoning for
> XParseGeometry, AFAICT.  You presented reasoning for some other code,
> which you consider similar.

I presented the reasoning GCC's bogus warning about XParseGeometry's caller in 
this earlier email:

https://lists.gnu.org/r/emacs-devel/2019-08/msg00531.html

If that presentation wasn't clear enough I can go into it in more detail; just 
let me know which parts weren't clear. Perhaps maybe-uninitialized.c will help 
explain things too.

> UNINIT is for you to be able to use your tools of choice, and
> perhaps also to cater to your personal stylistic preferences.

It's not just me using those tools. And UNINIT is not my stylistic preference: I 
don't like using UNINIT and when variables need not be initialized I'd rather 
just not initialize them (this is the longstanding tradition in GNU and UNIX 
code). The only reason for UNINIT is that it's more useful for preventing and 
catching bugs than its alternatives are.

[-- Attachment #2: maybe-uninitialized.c --]
[-- Type: text/x-csrc, Size: 932 bytes --]

extern int v;
extern int bar2 (void);
extern int bar4 (void);

int v;

static int
foo2 (int *x, int *y)
{
  int mask = 0;
  if (v & 1)
    {
      mask = 1;
      *x = v;
    }
  if (v & 2)
    {
      mask |= 2;
      *y = v;
    }
  return mask;
}

int
bar2 (void)
{
  int sum = 0;
  int x, y;
  int mask = foo2 (&x, &y);
  if (mask & 1)
    sum += x;
  if (mask & 2)
    sum += y;
  return sum;
}

static int
foo4 (int *x, int *y, int *w, int *h)
{
  int mask = 0;
  if (v & 1)
    {
      mask = 1;
      *x = v;
    }
  if (v & 2)
    {
      mask |= 2;
      *y = v;
    }
  if (v & 4)
    {
      mask |= 4;
      *w = v;
    }
  if (v & 8)
    {
      mask |= 8;
      *h = v;
    }
  return mask;
}

int
bar4 (void)
{
  int sum = 0;
  int x, y, w, h;
  int mask = foo4 (&x, &y, &w, &h);
  if (mask & 1)
    sum += x;
  if (mask & 2)
    sum += y;
  if (mask & 4)
    sum += w;
  if (mask & 8)
    sum += h;
  return sum;
}