switch question in recog.c

switch question in recog.c

[Andrew MacLeod]

Is this really valid? Sure seems screwy to me.


in the function asm_operand_ok(), the code looks like:

<...>
        case 'X':
          result = 1;

        case 'g':
          if (general_operand (op, VOIDmode))
            result = 1;
          break;

        default:
          /* For all other letters, we first check for a register class,
             otherwise it is an EXTRA_CONSTRAINT.  */
          if (REG_CLASS_FROM_CONSTRAINT (c, constraint) != NO_REGS)
            {
            case 'r':
              if (GET_MODE (op) == BLKmode)
                break;
              if (register_operand (op, VOIDmode))
                result = 1;
            }
#ifdef EXTRA_CONSTRAINT_STR
          if (EXTRA_CONSTRAINT_STR (op, c, constraint))
            result = 1;
          if (EXTRA_MEMORY_CONSTRAINT (c, constraint))
<...>

Look at where the case 'r' is.. its inside the 'default:' case. Is that
really kosher?


Andrew

Re: switch question in recog.c

[Daniel Jacobowitz]

On Tue, Oct 07, 2003 at 02:54:40PM -0400, Andrew MacLeod wrote:
> Is this really valid? Sure seems screwy to me.
> 
> 
> in the function asm_operand_ok(), the code looks like:
> 
> <...>
>         case 'X':
>           result = 1;
> 
>         case 'g':
>           if (general_operand (op, VOIDmode))
>             result = 1;
>           break;
> 
>         default:
>           /* For all other letters, we first check for a register class,
>              otherwise it is an EXTRA_CONSTRAINT.  */
>           if (REG_CLASS_FROM_CONSTRAINT (c, constraint) != NO_REGS)
>             {
>             case 'r':
>               if (GET_MODE (op) == BLKmode)
>                 break;
>               if (register_operand (op, VOIDmode))
>                 result = 1;
>             }
> #ifdef EXTRA_CONSTRAINT_STR
>           if (EXTRA_CONSTRAINT_STR (op, c, constraint))
>             result = 1;
>           if (EXTRA_MEMORY_CONSTRAINT (c, constraint))
> <...>
> 
> Look at where the case 'r' is.. its inside the 'default:' case. Is that
> really kosher?

Yep.  It acts basically as a label, they're only scoped by additional
switch statements.  At least that's my reading of the standard.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer

Re: switch question in recog.c

[Andrew MacLeod]

On Tue, 2003-10-07 at 14:59, Daniel Jacobowitz wrote:
> On Tue, Oct 07, 2003 at 02:54:40PM -0400, Andrew MacLeod wrote:

> >         default:
> >           /* For all other letters, we first check for a register class,
> >              otherwise it is an EXTRA_CONSTRAINT.  */
> >           if (REG_CLASS_FROM_CONSTRAINT (c, constraint) != NO_REGS)
> >             {
> >             case 'r':
> >               if (GET_MODE (op) == BLKmode)
> >                 break;
> >               if (register_operand (op, VOIDmode))
> >                 result = 1;
> >             }
> > #ifdef EXTRA_CONSTRAINT_STR
> >           if (EXTRA_CONSTRAINT_STR (op, c, constraint))
> >             result = 1;
> >           if (EXTRA_MEMORY_CONSTRAINT (c, constraint))
> > <...>
> > 
> > Look at where the case 'r' is.. its inside the 'default:' case. Is that
> > really kosher?
> 
> Yep.  It acts basically as a label, they're only scoped by additional
> switch statements.  At least that's my reading of the standard.
> 
Huh. I would have failed that question miserably after using the
language for 20 years. It would never occur to me to put a case after
the default label, let alone inside a control structure of some sort.
How miserable :-) But you are right, the standard doesnt appear to
disallow it that I can see.

Thanks
Andrew

Re: switch question in recog.c

[Joe Buck]

On Tue, Oct 07, 2003 at 03:15:49PM -0400, Andrew MacLeod wrote:
> On Tue, 2003-10-07 at 14:59, Daniel Jacobowitz wrote:
> > > Look at where the case 'r' is.. its inside the 'default:' case. Is that
> > > really kosher?
> > 
> > Yep.  It acts basically as a label, they're only scoped by additional
> > switch statements.  At least that's my reading of the standard.
> > 
> Huh. I would have failed that question miserably after using the
> language for 20 years. 

Unfortunately, C doesn't have a high-level switch statement, just a
prettified computed goto.  There's an argument for warning about
code like this.

Re: switch question in recog.c

[Michael Matz]

Hi,

On 7 Oct 2003, Andrew MacLeod wrote:

> Huh. I would have failed that question miserably after using the
> language for 20 years.

Then you never used a variant of the holy, mighty device from Tom Duff ;-)

Constructed useless example (i.e. not the usual memcpy() unrolling):

------------------
extern void abort ();
int f(int i)
{
  switch (i & 1) {
    while (1) {
      if (i > 32) {
        abort ();
        case 0: return 1;
      } else {
        abort ();
        case 1: return 2;
      }
      default: break;
    }
  }
}

int main()
{
  if (f (42) != 1)
    abort ();
  return 0;
}
------------------

"case x" is a normal label (with the exception that it's scoped to the
enclosing case statement) introducing a labeled statement, hence it can be
placed everywhere as long as it's inside a switch.  Sometimes I love C ;-)


Ciao,
Michael.

Re: switch question in recog.c

[DJ Delorie]

> Constructed useless example (i.e. not the usual memcpy() unrolling):

Duff's device wasn't a memcpy() unrolling.

Re: switch question in recog.c

[Michael Matz]

Hi,

On Tue, 7 Oct 2003, DJ Delorie wrote:

> > Constructed useless example (i.e. not the usual memcpy() unrolling):
>
> Duff's device wasn't a memcpy() unrolling.

See again http://www.lysator.liu.se/c/duffs-device.html commenting from
Tom himself.  His initial posting was about the loop
	send(short *to, short *from, int count)
        { do *to = *from++; while(--count>0); }

That's basically a memcpy().  I'm not old enough to know if Tom really is
the inventor of this "use" of C features, but not even Tom remembers ;-)


Ciao,
Michael.

Re: switch question in recog.c

[Erik Trulsson]

On Wed, Oct 08, 2003 at 12:38:54AM +0200, Michael Matz wrote:
> Hi,
> 
> On Tue, 7 Oct 2003, DJ Delorie wrote:
> 
> > > Constructed useless example (i.e. not the usual memcpy() unrolling):
> >
> > Duff's device wasn't a memcpy() unrolling.
> 
> See again http://www.lysator.liu.se/c/duffs-device.html commenting from
> Tom himself.  His initial posting was about the loop
> 	send(short *to, short *from, int count)
>         { do *to = *from++; while(--count>0); }
> 
> That's basically a memcpy().  I'm not old enough to know if Tom really is
> the inventor of this "use" of C features, but not even Tom remembers ;-)

It is not a memcpy(), but similar in structure. If it had been memcpy()
it would have been '*to++ = from++' instead of '*to = *from++'. Note
that the address written to never changes.
What the code does is write a buffer to some memory-mapped
I/O-register (pointed to by 'to').


-- 
<Insert your favourite quote here.>
Erik Trulsson
ertr1013@student.uu.se

Re: switch question in recog.c

[Michael Matz]

Hi,

On Wed, 8 Oct 2003, Erik Trulsson wrote:

> >         { do *to = *from++; while(--count>0); }
> >
> It is not a memcpy(), but similar in structure.

Hmm ... yeah ... well, I'm drunken and can't read anymore, you're right
;-)


Ciao,
Michael.

Re: switch question in recog.c

[Zack Weinberg]

Michael Matz <matz@suse.de> writes:

> Hi,
>
> On 7 Oct 2003, Andrew MacLeod wrote:
>
>> Huh. I would have failed that question miserably after using the
>> language for 20 years.
>
> Then you never used a variant of the holy, mighty device from Tom Duff ;-)
>
> Constructed useless example (i.e. not the usual memcpy() unrolling):

"For maximum obfuscation, the outer braces can be removed":

 extern void abort ();
 int f(int i)
 {
   switch (i & 1)
     while (1)
       if (i > 32) {
         abort ();
         case 0: return 1;
       } else {
         abort ();
         case 1: return 2;
       }
 }

> "case x" is a normal label (with the exception that it's scoped to
> the enclosing case statement) introducing a labeled statement, hence
> it can be placed everywhere as long as it's inside a switch.

Except you can't goto a case label.  Occasionally I want to be able to
write "goto default;".

zw

Re: switch question in recog.c

[Nix]

On Tue, 7 Oct 2003, Zack Weinberg moaned:
> "For maximum obfuscation, the outer braces can be removed":

The most extreme example of this I've ever seen came from the
UNIX-HATERS archives, attributed to Harbison and Steele's "C: A
Reference Manual".. I'm sure that everyone looking at this will
agree that it is a nice clear and sensible micro-optimization:

  switch (x)
    default:
      if (prime(x))
        case 2: case 3: case 5: case 7:
	  process_prime(x);
      else
        case 4: case 6: case 8: case 9: case 10:
	  process_composite(x);

-- 
`If you want a vision of the future, it is a wireless broadband network
 feeding requests for foreign money-laundering assistance into a human
 temporal lobe, forever. With banner ads.' --- John M. Ford

Re: switch question in recog.c

[Dale Johannesen]

> in the function asm_operand_ok(), the code looks like:
>
> <...>
>         case 'X':
>           result = 1;
>
>         case 'g':
>           if (general_operand (op, VOIDmode))
>             result = 1;
>           break;
>
>         default:
>           /* For all other letters, we first check for a register 
> class,
>              otherwise it is an EXTRA_CONSTRAINT.  */
>           if (REG_CLASS_FROM_CONSTRAINT (c, constraint) != NO_REGS)
>             {
>             case 'r':
>               if (GET_MODE (op) == BLKmode)
>                 break;
>               if (register_operand (op, VOIDmode))
>                 result = 1;
>             }
> #ifdef EXTRA_CONSTRAINT_STR
>           if (EXTRA_CONSTRAINT_STR (op, c, constraint))
>             result = 1;
>           if (EXTRA_MEMORY_CONSTRAINT (c, constraint))

That is valid, but it looks like this code got munged since 3.3.  All 
those
"result=1"s used to be "return 1", so the control flow is now weird, for
example after 'X",   I expect most of them should be "result=1; break" .

Re: switch question in recog.c

[Richard Henderson]

On Tue, Oct 07, 2003 at 12:08:07PM -0700, Dale Johannesen wrote:
> That is valid, but it looks like this code got munged since 3.3.  All 
> those
> "result=1"s used to be "return 1", so the control flow is now weird, for
> example after 'X",   I expect most of them should be "result=1; break" .

Indeed.  Patches solicited... ;-)


r~

Re: switch question in recog.c

[Chris Lattner]

> Huh. I would have failed that question miserably after using the
> language for 20 years. It would never occur to me to put a case after
> the default label, let alone inside a control structure of some sort.
> How miserable :-) But you are right, the standard doesnt appear to
> disallow it that I can see.

If you like this, you might like "Duff's Device", which abuses switch
statements even worse:
http://www.lysator.liu.se/c/duffs-device.html

I'm happy LLVM doesn't have to worry about this stuff. :)

-Chris

-- 
http://llvm.cs.uiuc.edu/
http://www.nondot.org/~sabre/Projects/