Re^6: Windows / Linux puzzler

[ikegami]

If you compare the actual values by doing "%a" formatting on both the Linux and windows perls, you'll see that 0.1 is exactly the same double.

Your are correct. pack 'd' shows this more definitely.

>perl -E"say sprintf '%vX', pack 'd>', $ARGV[0]" 0.1
3F.B9.99.99.99.99.99.9A

$ perl -E'say sprintf "%vX", pack "d>", $ARGV[0]', 0.1
3F.B9.99.99.99.99.99.9A

I believe the OP's activestate perl was built with (mingw port of) gcc

It definitely allows mingw to be used to compile modules, but I don't know if that's what used to build Perl itself. (Update: perl -V hints that it uses mingw)

I beleive Strawberry Perl uses mingw, and I used Strawberry Perl in the test in the earlier post. If so, there are pertinent differences in the libraries between mingw and gcc.

[syphilis]

there are pertinent differences in the libraries between mingw and gcc

I think that with mingw, the MS runtime is still used by default for formatting of values.
However, that is avoided (by building with -D__USE_MINGW_ANSI_STDIO) in recent versions of Strawberry Perl, beginning perl-5.26.0.
(Look for "-D__USE_MINGW_ANSI_STDIO" in perl -V:ccflags)

Of course, ActiveState don't build with -D__USE_MINGW_ANSI_STDIO.

Cheers,
Rob

[ikegami]

Nope, that doesn't make differences go away, seeing as I was using SP 5.30.2 downloaded and installed less than a week ago. That's newer than 5.26, yet the output was different that the perl built gcc on Unbuntu on WSL on the same machine. The SP build does have -D__USE_MINGW_ANSI_STDIO

Strawberry Perl 5.30.2 on Windows:

>perl -e"CORE::say $^X"
C:\progs\sp5302x64\perl\bin\perl.exe

>perl -v | find "This is"
This is perl 5, version 30, subversion 2 (v5.30.2) built for MSWin32-x
+64-multi-thread

>perl -V:nvsize
nvsize='8';

>perl -V:ccflags
ccflags=' -s -O2 -DWIN32 -DWIN64 -DCONSERVATIVE -D__USE_MINGW_ANSI_STD
+IO -DPERL_TEXTMODE_SCRIPTS -DPERL_IMPLICIT_CONTEXT -DPERL_IMPLICIT_SY
+S -DUSE_PERLIO -fwrapv -fno-strict-aliasing -mms-bitfields';

>perl -e"CORE::say sprintf '%.20g', 0.1"
0.10000000000000000555

Self-built Perl 5.30.2 on Ubuntu 18.04 LTS on WSL on Windows:

$ uname -a
Linux tribble 4.4.0-18362-Microsoft #476-Microsoft Fri Nov 01 16:53:00
+ PST 2019 x86_64 x86_64 x86_64 GNU/Linux

$ perl -v | grep 'This is'
This is perl 5, version 30, subversion 2 (v5.30.2) built for x86_64-li
+nux-thread-multi

$ perl -V:nvsize
nvsize='8';

$ perl -e'CORE::say sprintf "%.20g", 0.1'
0.10000000000000001

[syphilis]

Interesting - thanks for persisting.

When I revert to an older Strawberry Perl that wasn't built with -D__USE_MINGW_ANSI_STDIO (5.24.1 / MSWin32-x64-multi-thread) I get the same as on Ubuntu:

C:\_64\strawberry5.24.1>perl -e"CORE::say sprintf '%.20g', 0.1"
0.10000000000000001

C:\_64\strawberry5.24.1>

I don't think that there's a lot of significance in the output of '%.20g' formatting of doubles.
As you're probably aware, both 0.10000000000000001 and 0.10000000000000000555 will assign to the same double anyway.
Both a 20-decimal-digit value, and that same value rounded to 17 decimal digits, will assign to the same double (unless something is buggy).

If the intention of "%.20g" formatting of 0.1 is to get the closest 20-decimal digit approximation to the double 0.1, then the latter is correct, and the former (ie the value consistently provided by Ubuntu) is incorrect:

C:\>perl -MMath::MPFR=":mpfr" -le "$x=Math::MPFR->new('0.1'); $base=10
+; $digits=20; Rmpfr_out_str($x, $base, $digits, MPFR_RNDN);"
1.0000000000000000555e-1
C:\>

Perhaps Linux has some catching up to do.
Even VC++ 2017 and later are "getting it right". On perl-5.30.0, MSWin32-x64-multi-thread, built with VC++ 2019:

C:\>perl -e"CORE::say sprintf '%.20g', 0.1"
0.10000000000000000555

C:\>

I'm not exactly sure when VC++ started "getting it right". I know it was some time after VC++ 2010.

Cheers,
Rob