Re^3: Perl for Windows ?

[BrowserUk]

Thirded. I use AS builds, MSC v6 and nmake v7. I install most modules manually. Downloading the .gz; use gunzip and tar from UnxUtils; then the four part mantra.

Some modules that have particularly complex dependancies on third party modules, GD for example, I have built myself, but it is simply easier to use a PPM if it is available.

The main problem I encounter when building XS modules is unresolvable externals, usually from Perl%x.dll. The source of this problem is that win32 allows the programmer to specify which entrypoints in .dll are exported. This is used on win32 builds to ensure that only those entrypoints that are part of the official PerlApi are exported.

Under *nix using gcc et al. either a) has no way to limit what is exported from an .so; b) there is a mechanism but it isn't used. Either way, the result is that programmers writing XS modules under *nix, frequently use functions from perl58x.so that aren't officially a part of the PerlApi and don't notice because they get resolved and work. But when someone comes to try and build those modules on win32, those unofficial apis won't resolve. The originating authors without access to/interest in win32 simply put the problem down to "windoze" and that's where it ends.

If one #define was changed to allow all the functions in perl5x.dll to be exported, another huge swath of failures would disappear.

[syphilis]

If one #define was changed to allow all the functions in perl5x.dll to be exported, another huge swath of failures would disappear

This, I expect, alludes to (eg) your attempts to build Data::Alias as discussed in From whence comes perl-5.8.6-smartmatch.tar?.

Is it really just a matter of changing one #define ? If so, which one ? - and what's the change ? It could be interesting to build perl with that one define changed appropriately - and then see what happens as a result of that change. (Did we ever determine the reason that not all functions are exported ?)

Cheers,
Rob

[BrowserUk]

(Did we ever determine the reason that not all functions are exported ?)

I think you pretty much answered that question when you quoted Jan Dubois. Which you might summarise as "because they can" on some platforms.

Try editing perl.h line ~3943 (5.8.8 build)

#  define PERL_CALLCONV __declspec(dllexport)

Without the above edit, perl58.dll from the AS817/5.8.8 distribution:

C:\test>pushd \AS817\perl\bin

C:\AS817\perl\bin>dumpbin /exports perl58.dll | wc -l
   1457

C:\AS817\perl\bin>dumpbin /exports perl58.dll | find "av_reify"

C:\AS817\perl\bin>

With the above change, perl58.dll(5.8.8):

c:\Perl\src\perl-5.8.8>dumpbin /exports perl58.dll | wc -l
   1995

c:\Perl\src\perl-5.8.8>dumpbin /exports perl58.dll | find "av_reify"
        627  272 00001000 Perl_av_reify

How many problems those extra 500 exports will fix I'm not sure, but I'm betting quite a few.

Whether it will work with 5.10 I'll tell you in about 1/2 hour.

Update: Seems to work similarly for 5.10 also:

[10:50:31.10]C:\Perl\src\perl-5.10.0>dumpbin /exports perl510.dll | wc
+ -l
   1993

[10:50:47.42]C:\Perl\src\perl-5.10.0>cd \Perl510\bin

[10:50:59.53]C:\Perl510\bin>dumpbin /exports perl510.dll | wc -l
   1387

And the (summary) output from nmake test looks 'normal':

Failed Test          Stat Wstat Total Fail  List of Failed
----------------------------------------------------------------------
+---------
../ext/IO/t/io_dup.t                6    4  2-5
../lib/lib.t            5  1280    13    0  ??
comp/multiline.t                    6    2  5-6
io/dup.t                           29    6  2-7
 (1 subtest UNEXPECTEDLY SUCCEEDED), 56 tests and 1083 subtests skippe
+d.
Failed 4/1470 test scripts. 12/189314 subtests failed.
Files=1470, Tests=189314, 1264 wallclock secs ( 0.00 cusr +  0.00 csys
+ =  0.00 CPU)
Failed 4/1470 test programs. 12/189314 subtests failed.
NMAKE : fatal error U1077: '..\perl.exe' : return code '0xff'
Stop.

Whether it will break anything anywhere, or is backwards compatible with 5.0 again I do not know.

And where it leaves all the shenanigins played out by embed.pl & regen_lib.pl and all the work that goes into maintaining embed.fnc, pp.sym, intrpvar.h, perlvars.h and thrdvar.h is another open question.