bug#5114: 23.1.50; (string-to-number (number-to-string most-positive-fixnum))

bug#5114: 23.1.50; (string-to-number (number-to-string most-positive-fixnum))

[Helmut Eller]

(setq x (string-to-number (number-to-string most-positive-fixnum)))
(= most-positive-fixnum x) => nil
   
x is 2305843009213693440 but it should be most-positive-fixnum
which is 2305843009213693951.

The test
(= most-positive-fixnum 
   (string-to-number (number-to-string most-positive-fixnum)))
seems to work as expected on 32-bit machines but not so on 64 bit.

Helmut.

In GNU Emacs 23.1.50.1 (x86_64-unknown-linux-gnu)
 of 2009-12-03
configured using `configure  '--without-x''

Important settings:
  value of $LC_ALL: nil
  value of $LC_COLLATE: nil
  value of $LC_CTYPE: nil
  value of $LC_MESSAGES: nil
  value of $LC_MONETARY: nil
  value of $LC_NUMERIC: nil
  value of $LC_TIME: nil
  value of $LANG: en_US.UTF-8
  value of $XMODIFIERS: nil
  locale-coding-system: utf-8-unix
  default enable-multibyte-characters: t

bug#5114: 23.1.50; (string-to-number (number-to-string most-positive-fixnum))

[Stefan Monnier]

> (setq x (string-to-number (number-to-string most-positive-fixnum)))
> (= most-positive-fixnum x) => nil
   
> x is 2305843009213693440 but it should be most-positive-fixnum
> which is 2305843009213693951.

> The test
> (= most-positive-fixnum 
>    (string-to-number (number-to-string most-positive-fixnum)))
> seems to work as expected on 32-bit machines but not so on 64 bit.

Indeed, it passes through a floating point conversion, so there's only
abour 52 bit of precesion.


        Stefan

bug#5114: 23.1.50; (string-to-number (number-to-string most-positive-fixnum))

[Helmut Eller]

[-- Attachment #1: Type: text/plain, Size: 614 bytes --]

* Stefan Monnier [2009-12-03 21:52+0100] writes:

>> (setq x (string-to-number (number-to-string most-positive-fixnum)))
>> (= most-positive-fixnum x) => nil
>    
>> x is 2305843009213693440 but it should be most-positive-fixnum
>> which is 2305843009213693951.
>
>> The test
>> (= most-positive-fixnum 
>>    (string-to-number (number-to-string most-positive-fixnum)))
>> seems to work as expected on 32-bit machines but not so on 64 bit.
>
> Indeed, it passes through a floating point conversion, so there's only
> abour 52 bit of precesion.

Here is a patch for string-to-number to use the full fixnum range:


[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: string-to-number.patch --]
[-- Type: text/x-diff, Size: 962 bytes --]

--- data.c.~1.308.~	2009-12-05 08:07:48.000000000 +0100
+++ data.c	2009-12-05 13:27:33.000000000 +0100
@@ -2393,23 +2393,26 @@
     p++;
 
   if (isfloat_string (p, 1) && b == 10)
-    val = make_float (sign * atof (p));
-  else
-    {
-      double v = 0;
-
-      while (1)
-	{
-	  int digit = digit_to_number (*p++, b);
-	  if (digit < 0)
-	    break;
-	  v = v * b + digit;
-	}
-
-      val = make_fixnum_or_float (sign * v);
-    }
-
-  return val;
+    return make_float (sign * atof (p));
+  else {
+    unsigned long u = 0;
+    while (1)
+      {
+	int digit = digit_to_number (*p++, b);
+	if (digit < 0)
+	  return make_number (sign * u);
+	else if (u <= (MOST_POSITIVE_FIXNUM - digit) / b)
+	  u = u * b + digit;
+	else
+	  {
+	    /* overflow to flonums */
+	    double f = ((double)u) * b + digit;
+	    while (digit = digit_to_number (*p++, b), digit >= 0)
+	      f = f * b + digit;
+	    return make_float (sign * f);
+	  }
+      }
+  }
 }
 
 \f

[-- Attachment #3: Type: text/plain, Size: 8 bytes --]


Helmut

bug#5114: 23.1.50; (string-to-number (number-to-string most-positive-fixnum))

[Eli Zaretskii]

> From: Helmut Eller <eller.helmut@gmail.com>
> Date: Sat, 05 Dec 2009 13:36:41 +0100
> Cc: 5114@emacsbugs.donarmstrong.com
> 
> +  else {
> +    unsigned long u = 0;

This assumes that `unsigned long' is the same width as EMACS_INT.
This could be false, e.g., with 64-bit MS-Windows.  Isn't it better to
use EMACS_INT instead?

bug#5114: 23.1.50; (string-to-number (number-to-string most-positive-fixnum))

[Helmut Eller]

* Eli Zaretskii [2009-12-05 15:25+0100] writes:

>> From: Helmut Eller <eller.helmut@gmail.com>
>> Date: Sat, 05 Dec 2009 13:36:41 +0100
>> Cc: 5114@emacsbugs.donarmstrong.com
>> 
>> +  else {
>> +    unsigned long u = 0;
>
> This assumes that `unsigned long' is the same width as EMACS_INT.
> This could be false, e.g., with 64-bit MS-Windows.  Isn't it better to
> use EMACS_INT instead?

Using EMACS_UINT wouldn't hurt.  Does MOST_POSITIVE_FIXNUM not fit in a
unsigned long on Windows?  I assumed that longs are supposed to be as
wide as pointers.

Helmut

bug#5114: 23.1.50; (string-to-number (number-to-string most-positive-fixnum))

[Eli Zaretskii]

> From: Helmut Eller <eller.helmut@gmail.com>
> Date: Sat, 05 Dec 2009 16:18:13 +0100
> Cc: 
> 
> * Eli Zaretskii [2009-12-05 15:25+0100] writes:
> 
> >> From: Helmut Eller <eller.helmut@gmail.com>
> >> Date: Sat, 05 Dec 2009 13:36:41 +0100
> >> Cc: 5114@emacsbugs.donarmstrong.com
> >> 
> >> +  else {
> >> +    unsigned long u = 0;
> >
> > This assumes that `unsigned long' is the same width as EMACS_INT.
> > This could be false, e.g., with 64-bit MS-Windows.  Isn't it better to
> > use EMACS_INT instead?
> 
> Using EMACS_UINT wouldn't hurt.  Does MOST_POSITIVE_FIXNUM not fit in a
> unsigned long on Windows?

In the 32-bit Windows build, it does.  In the 64-bit Windows build
(which does not yet exist, since we don't yet support it), it will
not, because MS in their infinite wisdom (probably because backward
compatibility considerations wrt existing source code and headers)
decided to use the LLP64 programming model.

> I assumed that longs are supposed to be as wide as pointers.

Not on 64-bit Windows, they don't.

bug#5114: 23.1.50; (string-to-number (number-to-string most-positive-fixnum))

[Lars Magne Ingebrigtsen]

Helmut Eller <eller.helmut@gmail.com> writes:

> (setq x (string-to-number (number-to-string most-positive-fixnum)))
> (= most-positive-fixnum x) => nil
>
> x is 2305843009213693440 but it should be most-positive-fixnum
> which is 2305843009213693951.
>
> The test
> (= most-positive-fixnum 
>    (string-to-number (number-to-string most-positive-fixnum)))
> seems to work as expected on 32-bit machines but not so on 64 bit.

I'm unable to reproduce this on Emacs 24, so I think it's likely that
this has been fixed by all the numerical fixes that have been applied
the last year.

-- 
(domestic pets only, the antidote for overdose, milk.)
  bloggy blog http://lars.ingebrigtsen.no/

bug#5114: 23.1.50; (string-to-number (number-to-string most-positive-fixnum))

[Leo]

On 2011-09-18 17:47 +0800, Lars Magne Ingebrigtsen wrote:
> I'm unable to reproduce this on Emacs 24, so I think it's likely that
> this has been fixed by all the numerical fixes that have been applied
> the last year.

Indeed fixed in 24.x but present in emacs-23.

Leo