diff options
| author | Richard Sandiford <richard.sandiford@arm.com> | 2018-08-24 13:05:36 +0000 |
|---|---|---|
| committer | Richard Sandiford <rsandifo@gcc.gnu.org> | 2018-08-24 13:05:36 +0000 |
| commit | ab7e60cec1a6f4185b0428f3a2b3e71df0bae533 (patch) | |
| tree | c666f7a31ae111b41e904e26bb4df9009c6db203 /gcc/genmodes.c | |
| parent | 1ade64c9d8cd37c8db0a07383189f1719c7164da (diff) | |
| download | gcc-ab7e60cec1a6f4185b0428f3a2b3e71df0bae533.zip gcc-ab7e60cec1a6f4185b0428f3a2b3e71df0bae533.tar.gz gcc-ab7e60cec1a6f4185b0428f3a2b3e71df0bae533.tar.bz2 | |
Handle SLP permutations for variable-length vectors
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
Diffstat (limited to 'gcc/genmodes.c')
0 files changed, 0 insertions, 0 deletions