I managed to reproduce the problem in a standalone testcase.

It seems that the different coderef values from Perl and XS are a red herring. I guess dereferencing a coderef in Perl has some indirection magic that I don't understand, because I get different values in the following test script, but that doesn't stop it from working?

#! perl -slw
use strict;
use Inline 'NoClean', 'FORCE', 'INFO' ;
use Inline Config => WARNINGS => 4;
use Inline 'C' => 'DATA', NAME =>'test';


sub recorder{ print "record: @_"; }
sub player{ print "player: @_"; }

printf "r:%x p:%x\n", \&recorder, \&player;
setCallbacks( \&recorder, \&player );

record( 'test' );
#record( 'test' );
#play( 'test' );

__DATA__
__C__

SV *g_rec = 0, *g_play = 0;

int setCallbacks( SV *rec, SV *play ) {
    printf( "r:%x p:%x\n", rec, play );
    g_rec = rec;
    SvREFCNT_inc( g_rec );
    g_play = play;
    SvREFCNT_inc( g_play );
    return 0;
}

int record( SV *m ) {
    call_sv( g_rec, G_VOID );
    return 0;
}

int play( SV *m ) {
    call_sv( g_play, G_VOID );
    return 0;
}
[download]

As is, with the second call to record() and that to play() commented out, the program runs fine

P:\test>test

r:226004 p:2260b8
r:19499e4 p:2252a8
record: test
[download]

Different addresses(?), but it works. Remove either comment and it panics.

P:\test>test

r:226004 p:2260b8
r:19499e4 p:2252a8
record: test
panic: memory wrap at P:\test\test.pl line 16.
[download]

I assume that it is something to do with my (lack of) handling of the stack(?), but I've tried various combinations of the Inline_* macros without success.

Cluebats gratefully receieved.