web # perl -e '$a="nan"; print sprintf("%.2f",$a)'
nan

web # perl -v
This is perl, v5.8.8 built for x86_64-linux
-----
For Win32:
This is perl, v5.8.4 built for MSWin32-x86-multi-thread

perl -e "$a='nan'; print sprintf('%.2f',$a)"
0.00
[download]

Yep - it's broken on Win32, and the nature of the breakage varies, depending upon which compiler (ie Visual Studio or MinGW) was used to build perl. I submitted a perlbug report about this (assigned a ticket number of 38779) a while back.

But the example that you've provided fails in the same way, irrespective of whether perl was compiled with Visual Studio 6.0 or MinGW.

I thought this had been fixed with Visual Studio 7.0 but I still I get (on my perl built with Visual Studio 7.0):

F:\>perl -we "$a='nan'; print sprintf('%.2f',$a)"
0.00
[download]


Let's face it ... handling of NaN's on Win32 is unreliable. I don't know how you can disable it. You could try re-compiling perl 5.8.8 with the following patch supplied by Jan Dubois in response to my perlbug report.

--- pp.c.orig    Wed Mar 22 19:35:45 2006
+++ pp.c    Wed Mar 22 21:43:04 2006
@@ -1732,8 +1732,15 @@
     }
 #endif
     {
+#if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
+      dPOPTOPnnrl;
+      if (Perl_isnan(left) || Perl_isnan(right))
+      RETSETNO;
+      SETs(boolSV(left < right));
+#else
       dPOPnv;
       SETs(boolSV(TOPn < value));
+#endif
       RETURN;
     }
 }
@@ -1808,8 +1815,15 @@
     }
 #endif
     {
+#if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
+      dPOPTOPnnrl;
+      if (Perl_isnan(left) || Perl_isnan(right))
+      RETSETNO;
+      SETs(boolSV(left > right));
+#else
       dPOPnv;
       SETs(boolSV(TOPn > value));
+#endif
       RETURN;
     }
 }
@@ -1884,8 +1898,15 @@
     }
 #endif
     {
+#if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
+      dPOPTOPnnrl;
+      if (Perl_isnan(left) || Perl_isnan(right))
+      RETSETNO;
+      SETs(boolSV(left <= right));
+#else
       dPOPnv;
       SETs(boolSV(TOPn <= value));
+#endif
       RETURN;
     }
 }
@@ -1960,8 +1981,15 @@
     }
 #endif
     {
+#if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
+      dPOPTOPnnrl;
+      if (Perl_isnan(left) || Perl_isnan(right))
+      RETSETNO;
+      SETs(boolSV(left >= right));
+#else
       dPOPnv;
       SETs(boolSV(TOPn >= value));
+#endif
       RETURN;
     }
 }
@@ -2029,8 +2057,15 @@
     }
 #endif
     {
+#if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
+      dPOPTOPnnrl;
+      if (Perl_isnan(left) || Perl_isnan(right))
+      RETSETYES;
+      SETs(boolSV(left != right));
+#else
       dPOPnv;
       SETs(boolSV(TOPn != value));
+#endif
       RETURN;
     }
 }
--- pp_hot.c.orig    Wed Mar 22 19:35:45 2006
+++ pp_hot.c    Wed Mar 22 21:42:38 2006
@@ -306,8 +306,15 @@
     }
 #endif
     {
+#if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
+      dPOPTOPnnrl;
+      if (Perl_isnan(left) || Perl_isnan(right))
+      RETSETNO;
+      SETs(boolSV(left == right));
+#else
       dPOPnv;
       SETs(boolSV(TOPn == value));
+#endif
       RETURN;
     }
 }
--- sv.c.orig    Wed Mar 22 21:47:13 2006
+++ sv.c    Wed Mar 22 21:31:27 2006
@@ -2134,6 +2134,13 @@
        certainly cast into the IV range at IV_MAX, whereas the correc
+t
        answer is the UV IV_MAX +1. Hence < ensures that dodgy boundar
+y
        cases go to UV */
+#if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
+    if (Perl_isnan(SvNVX(sv))) {
+        SvUV_set(sv, 0);
+        SvIsUV_on(sv);
+    }
+    else
+#endif
     if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
         SvIV_set(sv, I_V(SvNVX(sv)));
         if (SvNVX(sv) == (NV) SvIVX(sv)
--- win32/win32.h.orig    Wed Mar 22 19:35:49 2006
+++ win32/win32.h    Wed Mar 22 21:35:54 2006
@@ -210,6 +210,11 @@
 
 #define isnan        _isnan
 
+#if _MSC_VER < 1300
+/* VC6 has broken NaN semantics: NaN == NaN returns true instead of f
+alse */
+#define NAN_COMPARE_BROKEN 1
+#endif
+
 #endif /* _MSC_VER */
 
 #ifdef __MINGW32__        /* Minimal Gnu-Win32 */
[download]


I don't know if it will help. Good luck.

Cheers,
Rob