I did what Nickolas suggested. It compiled fine, but failed horribly at runtime with error messages that don't seem to be related to the fix I made. I also tried another way using PERL_MAGIC_ext, but it resulted in the same issues.

So I guess Math::Pari is relying on more <=5.8 behaviour that has changed in 5.10. I wasn't able to figure out what exactly that is though.

I've tried it with the following patch:

1153c1153
<     ((CV*)cv)->sv_any->xiv_u.xivu_iv = numargs;    /* XXXX Nasty of 
+us... */
---
>     ((CV*)cv)->sv_any->xof_off = numargs;    /* XXXX Nasty of us... 
+*/
1219c1219
<     int numargs = ((CV*)cv)->sv_any->xiv_u.xivu_iv;    /* XXXX Nasty
+ of us... */
---
>     int numargs = ((CV*)cv)->sv_any->xof_off;    /* XXXX Nasty of us
+... */
[download]

It seems to work but increases the 'Nasty' level

That way it might compile, but it's still using SvPVX and SvIVX on an AV to store internal information. That breaks as those seem to be used by the AV itself in 5.10.