[Bug tree-optimization/68835] [5/6 Regression] ICE in set_value_range, at tree-vrp.c:387, with __int128 bit field

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68835

--- Comment #10 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Seems get_int_cst_ext_nunits already has some code to handle the case of
"negative" extra large integers, but only if they have precision multiple of
HOST_BITS_PER_WIDE_INT.  So I believe we want something for the precisions that
aren't multiple of HOST_BITS_PER_WIDE_INT, but are > HOST_BITS_PER_WIDE_INT,
too.
But not sure what exactly, simply returning larger ext_len in that case does
not help.