[Bug tree-optimization/114908] fails to optimize avx2 in-register permute written with std::experimental::simd

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114908

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |mkretz at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org
             Target|x86-64-v3                   |x86_64-*-*

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
The memcpy() calls are definitely a hindrance.  I suppose that
update-address-taken could replace some of them with BIT_INSERT_EXPRs but then
it doesn't
handle any calls right now.  Replacing the memcpy on its own would be possible
(special-casing just the "sub-vector" case) like

  __builtin_memcpy (&__r, &data, 24);

to

  _1 = __r;
  _2 = data;
  _3 = VEC_PERM <_2, _1, {0, 1, 2, 7 }>;
  __r = _3;

or if copying a single element using BIT_INSERT_EXPR.  OTOH that's not
good if __r stays in memory (the whole vector store might be good to
avoid STLF fails, but the read will be bad for the same reason).

The update-address-taken pass would know __r and data become registers.
We already have a similar case involving ATOMIC_COMPARE_EXCHANGE that
has delayed processing requring register arguments.  It might or might
not be a good example how to deal with this.