https://gcc.gnu.org/bugzilla/show_bug.cgi?id=43147
--- Comment #8 from Marc Glisse <glisse at gcc dot gnu.org> --- Created attachment 45306 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=45306&action=edit ix86_gimple_fold_builtin patch Like this then? I realized (because of the testsuite) that we do not currently validate that the 3rd argument is a 2-bit immediate (the error message on non-constants is misleading), and Intel's documentation indeed seems to indicate that any integer is valid, we only look at the 2 lowest bits. If possible, I think it would be nice to reduce the number of builtins, that's why I started with a patch to emmintrin.h (I now know that it should use (mask&2)?3:2 for the second number). For validation, something like: if(!__builtin_constant_p(__mask)) complain(); with extern void complain(void) __attribute__((error("argument must be a constant"))); may be good enough. But we still need to handle -O0... Yeah, maybe ix86_gimple_fold_builtin is easier :-(
