On Wed, Jun 3, 2026 at 5:21 PM Christopher Bazley <[email protected]> wrote:
>
> When given a constructor of variable-length vector type, the
> store_constructor function now builds a vector with the lower bound
> of the number of subparts in the vector type and uses it to emit the
> body of the kind of insn chosen by the convert_optab_handler function.
> Previously, this function used a fallback path of calling
> store_constructor_field upon discovering that the number of subparts
> in the vector type was not a constant multiple of the number of
> subparts in the element type.
>
> For example, this allows GCC to generate the following AArch64 assembly
> language output for the tail of a reduction in the slp_6 test:
>
> uaddv d31, p6, z31.b
> uaddv d27, p6, z27.b
> uaddv d26, p6, z26.b
> movi d30, #0
> insr z30.b, b26
> insr z30.b, b27
> insr z30.b, b31
> add z25.b, z25.b, z30.b
>
> instead of the following output (with predicated tails for basic block
> SLP vectorization but without this change):
>
> addvl x0, sp, #2
> movi d0, #0
> st1b z0.b, p6, [sp, #2, mul vl]
> uaddv d27, p6, z27.b
> uaddv d26, p6, z26.b
> uaddv d25, p6, z25.b
> str b27, [x0]
> addvl x0, sp, #1
> add x0, x0, 1
> ptrue p7.b, vl3
> ld1b z0.b, p6/z, [sp, #2, mul vl]
> st1b z0.b, p6, [sp, #1, mul vl]
> str b26, [x0]
> ld1b z0.b, p6/z, [sp, #1, mul vl]
> st1b z0.b, p6, [sp]
> str b25, [sp, 2]
> ld1b z0.b, p6/z, [sp]
> add z28.b, z28.b, z0.b
> st1b z28.b, p7, [x1]
> addvl sp, sp, #3
>
> or the original assembly language output (with neither predicated tails
> for basic block SLP vectorization nor this change):
>
> uaddv d31, p6, z31.b
> fmov x0, d31
> uaddv d31, p6, z26.b
> add w6, w6, w0
> fmov x0, d31
> uaddv d31, p6, z27.b
> add w5, w5, w0
> fmov x0, d31
> add w4, w4, w0
>
> gcc/ChangeLog:
>
> * expr.cc (store_constructor): Add an else block to handle
> cases of TREE_CODE (TREE_TYPE (exp)) == VECTOR_TYPE in which
> exact_div (n_elts, GET_MODE_NUNITS (eltmode)).is_constant
> (&const_n_elts) is false similar to the existing "element type
> is not a vector type" case except that const_n_elts is taken
> from the lower bound of the subparts of the vector type.
> ---
> gcc/expr.cc | 12 ++++++++++--
> 1 file changed, 10 insertions(+), 2 deletions(-)
>
> diff --git a/gcc/expr.cc b/gcc/expr.cc
> index de73215ccc6..f163ea16fe2 100644
> --- a/gcc/expr.cc
> +++ b/gcc/expr.cc
> @@ -7498,11 +7498,14 @@ fields_length (const_tree type)
> return count;
> }
>
> -
> /* Store the value of constructor EXP into the rtx TARGET.
> TARGET is either a REG or a MEM; we know it cannot conflict, since
> safe_from_p has been called.
> CLEARED is true if TARGET is known to have been zero'd.
> + If the constructor EXP has a vector type then elements of TARGET for which
> + there is no corresponding element in EXP are zero'd. For a
> variable-length
> + vector type, elements beyond the minimum number of subparts of the type
> are
> + not zero'd.
Huh? We agreed to zeroing even for VLA vectors.
> SIZE is the number of bytes of TARGET we are allowed to modify: this
> may not be the same as the size of EXP if we are assigning to a field
> which has been packed to exclude padding bits.
> @@ -8075,7 +8078,12 @@ store_constructor (tree exp, rtx target, int cleared,
> poly_int64 size,
> similarly non-const type vectors. */
> icode = convert_optab_handler (vec_init_optab, mode, eltmode);
> }
> -
> + else
> + {
> + /* Handle variable-length vector types. */
> + icode = convert_optab_handler (vec_init_optab, mode, eltmode);
> + const_n_elts = constant_lower_bound (n_elts);
> + }
And here I'd like to see
gcc_assert (icode != CODE_FOR_nothing);
for doc/md.texi the vec_init optab documentation should be amended that for
variable-size modes operand 1 contains elements for the minimum number of
elements and the remainder of elements should be zeroed.
> if (const_n_elts && icode != CODE_FOR_nothing)
> {
> vector = rtvec_alloc (const_n_elts);
> --
> 2.43.0
>
