The SLP code currently punts for all variable-length permutes.
This patch makes it handle the easy case of N->N permutes in which
the number of vector lanes is a multiple of N.  Every permute then
uses the same mask, and that mask repeats (with a stride) every
N elements.

The patch uses the same path for constant-length vectors,
since it should be slightly cheaper in terms of compile time.

2018-08-24  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
	* tree-vect-slp.c (vect_transform_slp_perm_load): Separate out
	the case in which the permute needs only a single element and
	repeats for every vector of the result.  Extend that case to
	handle variable-length vectors.
	* tree-vect-stmts.c (vectorizable_load): Update accordingly.

gcc/testsuite/
	* gcc.target/aarch64/sve/slp_perm_1.c: New test.
	* gcc.target/aarch64/sve/slp_perm_2.c: Likewise.
	* gcc.target/aarch64/sve/slp_perm_3.c: Likewise.
	* gcc.target/aarch64/sve/slp_perm_4.c: Likewise.
	* gcc.target/aarch64/sve/slp_perm_5.c: Likewise.
	* gcc.target/aarch64/sve/slp_perm_6.c: Likewise.
	* gcc.target/aarch64/sve/slp_perm_7.c: Likewise.

From-SVN: r263832