On Mon, Jul 31, 2017 at 11:19:26AM -0500, Bill Schmidt wrote:
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81622 reports an ICE-on-invalid
> for the vec_ld and vec_st built-in functions. This fires when the last
> argument of the built-in is not a pointer or array type, as is required.
> We break on this during early expansion of the built-ins into tree code
> during parsing. The solution, as with other ill-formed uses of these
> built-ins, is to avoid the early expansion when the argument has an invalid
> type, so that normal error handling can kick in later.
>
> (The long-term solution is to move the vec_ld and vec_st built-ins to the
> gimple folding work that Will Schmidt has been doing, but that hasn't
> happened yet.)
>
> Bootstrapped and tested on powerpc64le-linux-gnu with no regressions.
> Is this ok for trunk and GCC 7? I'd like to get it into 7.2 since it
> is a 7 regression.
See the patch I've attached in the PR, this isn't sufficient
(and for the ARRAY_TYPE I wonder if you can ever see there ARRAY_TYPE),
the function has various other issues, including e.g. ICE on
vec_cmpne (1, 2) with -mpower9.
Jakub
