[Bug middle-end/112824] Stack spills and vector splitting with vector builtins

["liuhongt at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112824

--- Comment #7 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
(In reply to Chris Elrod from comment #6)
> Hongtao Liu, I do think that one should ideally be able to get optimal
> codegen when using 512-bit builtin vectors or vector intrinsics, without
> needing to set `-mprefer-vector-width=512` (and, currently, also setting
> `-mtune-ctrl=avx512_move_by_pieces`).
> 
> 
> GCC respects the vector builtins and uses 512 bit ops, but then does splits
> and spills across function boundaries.
> So, what I'm arguing is, while it would be great to respect
> `-mprefer-vector-width=512`, it should ideally also be able to respect
> vector builtins/intrinsics, so that one can use full width vectors without
> also having to set `-mprefer-vector-width=512
> -mtune-control=avx512_move_by_pieces`.
If it's designed the way you want it to be， another issue would be like, should
we lower 512-bit vector builtins/intrinsic to ymm/xmm when
-mprefer-vector-width=256, the answer is we'd rather not. The intrinsic should
be closer to a one-to-one correspondence of instructions.(Though there're
several instrinics which are lowered to a sequence of instructions)
There're also others users using 512-bit intriniscs for specific kernel loops,
but still want compiler to generate xmm/ymm for other codes,
-mprefer-vector-width=256.

Originally -mprefer-vector-width=XXX is designed for auto-vectorization, and
-mtune-ctrl=avx512_move_by_pieces is for memory movement. Both of which are
orthogonal to codegen for builtin, intrinsic or explicit vector types. If user
explicitly use 512-bit vector type, builtins or intrinsics, gcc will generate
zmm no matter -mprefer-vector-width=.

And yes, there could be some mismatches between 512-bit intrinsic and
architecture tuning when you're using 512-bit intrinsic, and also rely on
compiler autogen to handle struct
(struct Dual<T,N> {
  Vector<T, N+1> data;
};). 
For such case, an explicit -mprefer-vector-width=512 is needed.