[Bug target/87532] bad results from vec_extract(unsigned char, foo) dependent upon function inline

[wschmidt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87532

--- Comment #11 from Bill Schmidt <wschmidt at gcc dot gnu.org> ---
Let me take back what I said earlier.  We've had full support for vec_extract
with a variable second argument for quite a long time.  So let me try again
responding to comment #4.

We have special-case code for ALTIVEC_BUILTIN_VEC_EXTRACT in rs6000-c.c when
handling overloaded built-ins.  There is short-circuit code there for cases
where the second argument is a constant and in-range, as well for cases where
the second argument is variable and we have a direct-move instruction
available.  In all other cases, we are supposed to expand this code into
address arithmetic:

/* Build *(((arg1_inner_type*)&(vector type){arg1})+arg2). */

But the fact that you are seeing the error message about the selector being out
of range indicates that we are expanding the vector built-in via rs6000.c:
altivec_expand_builtin, which in turn means that rs6000_overloaded_builtin_p
failed to expand the call.

We need to understand why the code in altivec_resolve_overloaded_builtin is not
being expanded as expected.  So please set a breakpoint there and look at the
fndecl to see what's going on.