Re: [9fans] fun and scary evil C code

Re: [9fans] fun and scary evil C code

[Russ Cox]

> Does C99 or any other C mandate the actual memory layout
> of floats and doubles or the exact conversion of constant  
> representations?
> I'm fairly sure they somehow mandate IEEE 754 properties,
> but do they actually say that floats and doubles have to be stored  
> exactly that way in 4 or 8 bytes?
> 
> Even if we assume sizeof(double) == 8,
> what if my implementation is perverse and interleaves the exponent  
> bits amongst the mantissa bits?
> Where is this disallowed in the standard(s)?

It doesn't matter what the standard says; it matters what implementations do.
C implementations are going to provide what the underlying
hardware does, and almost all hardware does IEEE 754.

> as long as the conversion you're after has an exact ieee representation,
> i can't see how two compliant implementations could come up with
> differing representations.  (two different numbers can't have the same
> ieee representation, except -0 and +0.)  the conversion process doesn't
> need any floating point itself and the only interpolation comes when
> numbers don't have exact representations.

Historically, it has not always been true that decimal <-> binary
conversion of doubles has been precise enough to replicate a 
specific bit pattern.  While conversion doesn't require floating
point, it is often done in floating point anyway for convenience.

Russ

Re: [9fans] fun and scary evil C code

[erik quanstrom]

as long as the conversion you're after has an exact ieee representation,
i can't see how two compliant implementations could come up with
differing representations.  (two different numbers can't have the same
ieee representation, except -0 and +0.)  the conversion process doesn't
need any floating point itself and the only interpolation comes when
numbers don't have exact representations.

it didn't occur to me immediately after russ' post that what he said
was literally correct:

#include<u.h>
#include<libc.h>

#define Magici(x) (((uchar*)&magic)[7-(x)])
static double magic = 7.949928895127363e-275;

void
main(void)
{
	uint sign, exp;
	uvlong sig, *v;

	v = (uvlong*)&magic;
	sign = *v>>63;
	exp = *v>>52;
	sig = *v&~(4096-1LL<<52);
	print("%b*%b*%ullx\n", sign, exp, sig);
	print("%ullx\n", *v);
}

- erik

Re: [9fans] fun and scary evil C code

[roger peppe]

On Dec 20, 2007 12:51 AM, erik quanstrom <quanstro@quanstro.net> wrote:
> in any event, if ieee is used, i think the trick is safe.

surely it depends crucially on the consistency of the decimal to float
conversion?
is that standard?

Re: [9fans] fun and scary evil C code

[erik quanstrom]

> I lied.
> They're actually using that stuff to pick apart the (presumed IEEE 754)
> doubles into exponent, mantissa, ...
> 
> My questions still stand ...
> 
> Also, that file is such a tangled nest of ifdefs ...
> is anyone sure whether (and if so, when) that code is invoked?

interesting conflicts on the web.  http://en.wikipedia.org/wiki/IEEE_754
says c doesn't require ieee, http://home.datacomm.ch/t_wolf/tw/c/c9x_changes.html point #21.
says it does.

in any event, if ieee is used, i think the trick is safe.

- erik

Re: [9fans] fun and scary evil C code

[dave.l]

> As far as I can see,
> they're assuming that double is IEEE 754 and that you can
> rip apart such a representation on an random-endian machine
> and re-assemble it on an any other random-endian machine.

I lied.
They're actually using that stuff to pick apart the (presumed IEEE 754)
doubles into exponent, mantissa, ...

My questions still stand ...

Also, that file is such a tangled nest of ifdefs ...
is anyone sure whether (and if so, when) that code is invoked?

DaveL

Re: [9fans] fun and scary evil C code

[dave.l]

> there, of course, we have a real compiler and don't have to
> write uvlong constants as floating point numbers
> (wow that seems fragile).

Scarily: they're not:
if you read on, that macro is for picking apart a double into bytes  
and vice-versa.
(i.e. it's insanity is internally consistent).
It's still ludicrous and fragile.

As far as I can see,
they're assuming that double is IEEE 754 and that you can
rip apart such a representation on an random-endian machine
and re-assemble it on an any other random-endian machine.

Does C99 or any other C mandate the actual memory layout
of floats and doubles or the exact conversion of constant  
representations?
I'm fairly sure they somehow mandate IEEE 754 properties,
but do they actually say that floats and doubles have to be stored  
exactly that way in 4 or 8 bytes?

Even if we assume sizeof(double) == 8,
what if my implementation is perverse and interleaves the exponent  
bits amongst the mantissa bits?
Where is this disallowed in the standard(s)?

DaveL

Re: [9fans] fun and scary evil C code

[Bakul Shah]

> > #define TRIO_DOUBLE_INDEX(x) (((unsigned char *)&internalEndianMagic)[7-(x)
> ])
> 
> this is actually done in /sys/src/9/port/devcons.c too:
> 
> 	static uvlong uvorder = 0x0001020304050607ULL;
> 	
> 	static uchar*
> 	le2vlong(vlong *to, uchar *f)
> 	{
> 		uchar *t, *o;
> 		int i;
> 	
> 		t = (uchar*)to;
> 		o = (uchar*)&uvorder;
> 		for(i = 0; i < sizeof(vlong); i++)
> 			t[o[i]] = f[i];
> 		return f+sizeof(vlong);
> 	}
> 	
> 	static uchar*
> 	vlong2le(uchar *t, vlong from)
> 	{
> 		uchar *f, *o;
> 		int i;
> 	
> 		f = (uchar*)&from;
> 		o = (uchar*)&uvorder;
> 		for(i = 0; i < sizeof(vlong); i++)
> 			t[i] = f[o[i]];
> 		return t+sizeof(vlong);
> 	}
> 
> presotto wrote the code but said he learned the trick from ken.
> 
> there, of course, we have a real compiler and don't have to
> write uvlong constants as floating point numbers
> (wow that seems fragile).

Most likely the TRIO_DOUBLE_INDEX macro came from some code
from the pre-long-long days, when you had no other way to
write down an 8 byte value in host endian order (even in a
less than portable way).

Re: [9fans] fun and scary evil C code

[erik quanstrom]

> > presotto wrote the code but said he learned the trick from ken.
> > 
> > there, of course, we have a real compiler and don't have to
> > write uvlong constants as floating point numbers
> > (wow that seems fragile).
> > 
> > russ
> 
> Now THAT's something to be proud of.  Especially without comments.

which.  writing the code or learning tricks from ken?

☺

- erik

Re: [9fans] fun and scary evil C code

[Robert William Fuller]

Russ Cox wrote:
>> #define TRIO_DOUBLE_INDEX(x) (((unsigned char *)&internalEndianMagic)[7-(x)])
> 
> this is actually done in /sys/src/9/port/devcons.c too:
> 
> 	static uvlong uvorder = 0x0001020304050607ULL;
> 	
> 	static uchar*
> 	le2vlong(vlong *to, uchar *f)
> 	{
> 		uchar *t, *o;
> 		int i;
> 	
> 		t = (uchar*)to;
> 		o = (uchar*)&uvorder;
> 		for(i = 0; i < sizeof(vlong); i++)
> 			t[o[i]] = f[i];
> 		return f+sizeof(vlong);
> 	}
> 	
> 	static uchar*
> 	vlong2le(uchar *t, vlong from)
> 	{
> 		uchar *f, *o;
> 		int i;
> 	
> 		f = (uchar*)&from;
> 		o = (uchar*)&uvorder;
> 		for(i = 0; i < sizeof(vlong); i++)
> 			t[i] = f[o[i]];
> 		return t+sizeof(vlong);
> 	}
> 
> presotto wrote the code but said he learned the trick from ken.
> 
> there, of course, we have a real compiler and don't have to
> write uvlong constants as floating point numbers
> (wow that seems fragile).
> 
> russ

Now THAT's something to be proud of.  Especially without comments.

Re: [9fans] fun and scary evil C code

[Russ Cox]

> #define TRIO_DOUBLE_INDEX(x) (((unsigned char *)&internalEndianMagic)[7-(x)])

this is actually done in /sys/src/9/port/devcons.c too:

	static uvlong uvorder = 0x0001020304050607ULL;
	
	static uchar*
	le2vlong(vlong *to, uchar *f)
	{
		uchar *t, *o;
		int i;
	
		t = (uchar*)to;
		o = (uchar*)&uvorder;
		for(i = 0; i < sizeof(vlong); i++)
			t[o[i]] = f[i];
		return f+sizeof(vlong);
	}
	
	static uchar*
	vlong2le(uchar *t, vlong from)
	{
		uchar *f, *o;
		int i;
	
		f = (uchar*)&from;
		o = (uchar*)&uvorder;
		for(i = 0; i < sizeof(vlong); i++)
			t[i] = f[o[i]];
		return t+sizeof(vlong);
	}

presotto wrote the code but said he learned the trick from ken.

there, of course, we have a real compiler and don't have to
write uvlong constants as floating point numbers
(wow that seems fragile).

russ

[9fans] fun and scary evil C code

[David Leimbach]

[-- Attachment #1: Type: text/plain, Size: 1037 bytes --]

I was amused by this:
http://www.steike.com/code/useless/evil-c/

I particularly liked the "internalEndianMagic".

I see this in some XML libs, as well as GMP and other open sourced code.

http://unix.derkeiler.com/Newsgroups/comp.unix.programmer/2005-12/msg00198.html


From:
https://svn.r-project.org/R/trunk/src/extra/trio/trionan.c

/*
 * Endian-agnostic indexing macro.
 *
 * The value of internalEndianMagic, when converted into a 64-bit
 * integer, becomes 0x0706050403020100 (we could have used a 64-bit
 * integer value instead of a double, but not all platforms supports
 * that type). The value is automatically encoded with the correct
 * endianess by the compiler, which means that we can support any
 * kind of endianess. The individual bytes are then used as an index
 * for the IEEE 754 bit-patterns and masks.
 */
#define TRIO_DOUBLE_INDEX(x) (((unsigned char *)&internalEndianMagic)[7-(x)])
static TRIO_CONST double internalEndianMagic = 7.949928895127363e-275;
#endif

pretty weird stuff.

[-- Attachment #2: Type: text/html, Size: 2084 bytes --]