"H.J. Lu via Gcc-patches" <[email protected]> writes:
> On Sun, Jun 27, 2021 at 1:43 AM Richard Sandiford
> <[email protected]> wrote:
>>
>> "H.J. Lu" <[email protected]> writes:
>> > 1. Update vec_duplicate to allow to fail so that backend can only allow
>> > broadcasting an integer constant to a vector when broadcast instruction
>> > is available. This can be used by memset expander to avoid vec_duplicate
>> > when loading from constant pool is more efficient.
>>
>> I don't see any changes in target-independent code though, other than
>> the doc update. It's still the case that (existing) uses of
>> vec_duplicate_optab do not allow it to fail.
>
> I have a followup patch set on
>
> https://gitlab.com/x86-gcc/gcc/-/commits/users/hjl/pieces/broadcast
>
> to use it to expand memset with vector broadcast:
>
> https://gitlab.com/x86-gcc/gcc/-/commit/991c87f8a83ca736ae9ed92baa3ebadca289f6e3
>
> For SSE2 which doesn't have vector broadcast, the constant vector broadcast
> expander returns FAIL and load from constant pool will be used.
Hmm, but as Jeff and I mentioned in the earlier replies,
vec_duplicate_optab shouldn't be used for constants. Constants
should go via the move expanders instead.
In a previous message I suggested:
… would it work to change:
/* Try using vec_duplicate_optab for uniform vectors. */
if (!TREE_SIDE_EFFECTS (exp)
&& VECTOR_MODE_P (mode)
&& eltmode == GET_MODE_INNER (mode)
&& ((icode = optab_handler (vec_duplicate_optab, mode))
!= CODE_FOR_nothing)
&& (elt = uniform_vector_p (exp)))
to something like:
/* Try using vec_duplicate_optab for uniform vectors. */
if (!TREE_SIDE_EFFECTS (exp)
&& VECTOR_MODE_P (mode)
&& eltmode == GET_MODE_INNER (mode)
&& (elt = uniform_vector_p (exp)))
{
if (TREE_CODE (elt) == INTEGER_CST
|| TREE_CODE (elt) == POLY_INT_CST
|| TREE_CODE (elt) == REAL_CST
|| TREE_CODE (elt) == FIXED_CST)
{
rtx src = gen_const_vec_duplicate (mode, expand_normal (node));
emit_move_insn (target, src);
break;
}
…
}
if that code was the source of the constant operand. If we're adding a
new use of vec_duplicate_optab then that should be similarly protected
against constant operands.
Thanks,
Richard
