forwprop: allow subvectors in simplify_vector_constructor ()

This is an attempt to fix
https://gcc.gnu.org/pipermail/gcc-patches/2025-October/697879.html in the
middle-end; the motivation in that patch was to teach gcc to compile:

int16x8_t foo (int16x8_t x)
{
  return vcombine_s16 (vget_high_s16 (x), vget_low_s16 (x));
}

into one instruction:

foo:
        ext     v0.16b, v0.16b, v0.16b, #8
        ret

rather than the two we are generating now:

foo:
        dup     d31, v0.d[1]
        uzp1    v0.2d, v31.2d, v0.2d
        ret

Instead of adding a define_insn in the backend, this patch relaxes the
precondition of tree-ssa-forwprop.cc:simplify_vector_constructor () to
accept subvectors as constructor elements.  During initial argument
processing (ll. 3817-3916), subvectors are decomposed into individual
elements before populating the ELTS array; this allows the rest of the
function to remain unchanged.  Special handling is also implemented for
constant and splat subvector elements of a constructor (the latter with
the use of ssa_uniform_vector_p () from tree-vect-generic.cc, which this
patch moves to tree.cc).

Add GIMPLE tests to gcc.dg/tree-ssa demonstrating the intended behavior
with various combinations of subvectors as constructor arguments,
including constant and splat subvectors; also add some aarch64-specific
tests to show that the change leads to us picking the "ext" instruction
for the resulting VEC_PERM_EXPR.

Bootstrapped and regtested on aarch64 and x86_64, regtested on aarch64_be.

gcc/ChangeLog:

	* tree-ssa-forwprop.cc (simplify_vector_constructor): Support
	vector constructor elements.
	* tree-vect-generic.cc (ssa_uniform_vector_p): Make non-static and
	move ...
	* tree.cc (ssa_uniform_vector_p): ... here.
	* tree.h (ssa_uniform_vector_p): Declare it.

gcc/testsuite/ChangeLog:

	* gcc.dg/tree-ssa/forwprop-43.c: New test.
	* gcc.target/aarch64/simd/combine_ext.c: New test.

0 files changed, 0 insertions, 0 deletions