[PATCH V2] [RFC]Support vectorization for Complex type.

Richard Biener richard.guenther@gmail.com
Tue Jul 19 08:43:01 GMT 2022


On Mon, Jul 18, 2022 at 4:31 AM liuhongt <hongtao.liu@intel.com> wrote:
>
> V2 update:
>    Handle VMAT_ELEMENTWISE, VMAT_CONTIGUOUS_PERMUTE, VMAT_STRIDED_SLP,
>    VMAT_CONTIGUOUS_REVERSE, VMAT_CONTIGUOUS_DOWN for complex type.
>
> I've run SPECspeed@2017 627.cam4_s, there's some vectorization cases,
> but no big performance impact(since this patch only handle load/store).
>
> Any comments?

My original comments still stand (it feels like this should be more generic).
Can we go the way lowering complex loads/stores first?  A large part
of the testcases
added by the patch should pass after that.

Thanks,
Richard.

> gcc/ChangeLog:
>
>         PR tree-optimization/106010
>         * tree-vect-data-refs.cc (vect_get_data_access_cost):
>         Pass complex_p to vect_get_num_copies to avoid ICE.
>         (vect_analyze_data_refs): Support vectorization for Complex
>         type with vector scalar types.
>         (vect_permute_load_chain): Handle Complex type.
>         * tree-vect-loop.cc (vect_determine_vf_for_stmt_1): VF should
>         be half of TYPE_VECTOR_SUBPARTS when complex_p.
>         * tree-vect-slp.cc (vect_record_max_nunits): nunits should be
>         half of TYPE_VECTOR_SUBPARTS when complex_p.
>         (vect_optimize_slp): Support permutation for complex type.
>         (vect_slp_analyze_node_operations_1): Double nunits in
>         vect_get_num_vectors to get right SLP_TREE_NUMBER_OF_VEC_STMTS
>         when complex_p.
>         (vect_slp_analyze_node_operations): Ditto.
>         (vect_create_constant_vectors): Support CTOR for complex type.
>         (vect_transform_slp_perm_load): Support permutation for
>         complex type.
>         * tree-vect-stmts.cc (vect_init_vector): Support complex type.
>         (vect_get_vec_defs_for_operand): Get vector type for
>         complex type.
>         (vectorizable_store): Get right ncopies/nunits and
>         elem_type for complex type vector, also return false when
>         complex_p and !TYPE_VECTOR_SUBPARTS.is_constant ().
>         (vect_truncate_gather_scatter_offset): Return false for
>         complex type.
>         (vectorizable_load): Ditto.
>         (vect_get_vector_types_for_stmt): Get vector type for
>         complex type.
>         (get_group_load_store_type): Hanlde complex type for
>         nunits.
>         (perm_mask_for_reverse): New overload.
>         (get_negative_load_store_type): Handle complex type,
>         p_offset should be N - 2 beofre addres of DR.
>         (vect_check_scalar_mask): Return false for complex type.
>         * tree-vectorizer.h (STMT_VINFO_COMPLEX_P): New macro.
>         (vect_get_num_copies): New overload.
>
> gcc/testsuite/ChangeLog:
>
>         * gcc.target/i386/pr106010-1a.c: New test.
>         * gcc.target/i386/pr106010-1b.c: New test.
>         * gcc.target/i386/pr106010-1c.c: New test.
>         * gcc.target/i386/pr106010-2a.c: New test.
>         * gcc.target/i386/pr106010-2b.c: New test.
>         * gcc.target/i386/pr106010-2c.c: New test.
>         * gcc.target/i386/pr106010-3a.c: New test.
>         * gcc.target/i386/pr106010-3b.c: New test.
>         * gcc.target/i386/pr106010-3c.c: New test.
>         * gcc.target/i386/pr106010-4a.c: New test.
>         * gcc.target/i386/pr106010-4b.c: New test.
>         * gcc.target/i386/pr106010-4c.c: New test.
>         * gcc.target/i386/pr106010-5a.c: New test.
>         * gcc.target/i386/pr106010-5b.c: New test.
>         * gcc.target/i386/pr106010-5c.c: New test.
>         * gcc.target/i386/pr106010-6a.c: New test.
>         * gcc.target/i386/pr106010-6b.c: New test.
>         * gcc.target/i386/pr106010-6c.c: New test.
>         * gcc.target/i386/pr106010-7a.c: New test.
>         * gcc.target/i386/pr106010-7b.c: New test.
>         * gcc.target/i386/pr106010-7c.c: New test.
>         * gcc.target/i386/pr106010-8a.c: New test.
>         * gcc.target/i386/pr106010-8b.c: New test.
>         * gcc.target/i386/pr106010-8c.c: New test.
>         * gcc.target/i386/pr106010-9a.c: New test.
>         * gcc.target/i386/pr106010-9b.c: New test.
>         * gcc.target/i386/pr106010-9c.c: New test.
>         * gcc.target/i386/pr106010-9d.c: New test.
> ---
>  gcc/testsuite/gcc.target/i386/pr106010-1a.c |  58 +++++
>  gcc/testsuite/gcc.target/i386/pr106010-1b.c |  63 ++++++
>  gcc/testsuite/gcc.target/i386/pr106010-1c.c |  41 ++++
>  gcc/testsuite/gcc.target/i386/pr106010-2a.c |  82 +++++++
>  gcc/testsuite/gcc.target/i386/pr106010-2b.c |  62 ++++++
>  gcc/testsuite/gcc.target/i386/pr106010-2c.c |  47 ++++
>  gcc/testsuite/gcc.target/i386/pr106010-3a.c |  80 +++++++
>  gcc/testsuite/gcc.target/i386/pr106010-3b.c | 126 +++++++++++
>  gcc/testsuite/gcc.target/i386/pr106010-3c.c |  69 ++++++
>  gcc/testsuite/gcc.target/i386/pr106010-4a.c | 101 +++++++++
>  gcc/testsuite/gcc.target/i386/pr106010-4b.c |  67 ++++++
>  gcc/testsuite/gcc.target/i386/pr106010-4c.c |  54 +++++
>  gcc/testsuite/gcc.target/i386/pr106010-5a.c | 117 ++++++++++
>  gcc/testsuite/gcc.target/i386/pr106010-5b.c |  80 +++++++
>  gcc/testsuite/gcc.target/i386/pr106010-5c.c |  62 ++++++
>  gcc/testsuite/gcc.target/i386/pr106010-6a.c | 115 ++++++++++
>  gcc/testsuite/gcc.target/i386/pr106010-6b.c | 157 +++++++++++++
>  gcc/testsuite/gcc.target/i386/pr106010-6c.c |  80 +++++++
>  gcc/testsuite/gcc.target/i386/pr106010-7a.c |  58 +++++
>  gcc/testsuite/gcc.target/i386/pr106010-7b.c |  63 ++++++
>  gcc/testsuite/gcc.target/i386/pr106010-7c.c |  41 ++++
>  gcc/testsuite/gcc.target/i386/pr106010-8a.c |  58 +++++
>  gcc/testsuite/gcc.target/i386/pr106010-8b.c |  53 +++++
>  gcc/testsuite/gcc.target/i386/pr106010-8c.c |  38 ++++
>  gcc/testsuite/gcc.target/i386/pr106010-9a.c |  89 ++++++++
>  gcc/testsuite/gcc.target/i386/pr106010-9b.c |  90 ++++++++
>  gcc/testsuite/gcc.target/i386/pr106010-9c.c |  90 ++++++++
>  gcc/testsuite/gcc.target/i386/pr106010-9d.c |  92 ++++++++
>  gcc/tree-vect-data-refs.cc                  | 134 +++++++++---
>  gcc/tree-vect-loop.cc                       |   7 +-
>  gcc/tree-vect-slp.cc                        | 174 +++++++++++----
>  gcc/tree-vect-stmts.cc                      | 231 +++++++++++++++++---
>  gcc/tree-vectorizer.h                       |  13 ++
>  33 files changed, 2594 insertions(+), 98 deletions(-)
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-1a.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-1b.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-1c.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-2a.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-2b.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-2c.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-3a.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-3b.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-3c.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-4a.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-4b.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-4c.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-5a.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-5b.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-5c.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-6a.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-6b.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-6c.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-7a.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-7b.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-7c.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-8a.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-8b.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-8c.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-9a.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-9b.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-9c.c
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr106010-9d.c
>
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-1a.c b/gcc/testsuite/gcc.target/i386/pr106010-1a.c
> new file mode 100644
> index 00000000000..b608f484934
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-1a.c
> @@ -0,0 +1,58 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx -ftree-vectorize -fvect-cost-model=unlimited -fdump-tree-vect-details -mprefer-vector-width=256" } */
> +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 6 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(4\) double>} 2 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(8\) float>} 2 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(4\) long long int>} 2 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(8\) int>} 2 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(16\) short int>} 2 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(32\) char>} 2 "vect" } } */
> +
> +#define N 10000
> +void
> +__attribute__((noipa))
> +foo_pd (_Complex double* a, _Complex double* b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b[i];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_ps (_Complex float* a, _Complex float* b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b[i];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi64 (_Complex long long* a, _Complex long long* b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b[i];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi32 (_Complex int* a, _Complex int* b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b[i];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi16 (_Complex short* a, _Complex short* b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b[i];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi8 (_Complex char* a, _Complex char* b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b[i];
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-1b.c b/gcc/testsuite/gcc.target/i386/pr106010-1b.c
> new file mode 100644
> index 00000000000..0f377c3a548
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-1b.c
> @@ -0,0 +1,63 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256" } */
> +/* { dg-require-effective-target avx } */
> +
> +#include "avx-check.h"
> +#include <string.h>
> +#include "pr106010-1a.c"
> +
> +void
> +avx_test (void)
> +{
> +  _Complex double* pd_src = (_Complex double*) malloc (2 * N * sizeof (double));
> +  _Complex double* pd_dst = (_Complex double*) malloc (2 * N * sizeof (double));
> +  _Complex float* ps_src = (_Complex float*) malloc (2 * N * sizeof (float));
> +  _Complex float* ps_dst = (_Complex float*) malloc (2 * N * sizeof (float));
> +  _Complex long long* epi64_src = (_Complex long long*) malloc (2 * N * sizeof (long long));
> +  _Complex long long* epi64_dst = (_Complex long long*) malloc (2 * N * sizeof (long long));
> +  _Complex int* epi32_src = (_Complex int*) malloc (2 * N * sizeof (int));
> +  _Complex int* epi32_dst = (_Complex int*) malloc (2 * N * sizeof (int));
> +  _Complex short* epi16_src = (_Complex short*) malloc (2 * N * sizeof (short));
> +  _Complex short* epi16_dst = (_Complex short*) malloc (2 * N * sizeof (short));
> +  _Complex char* epi8_src = (_Complex char*) malloc (2 * N * sizeof (char));
> +  _Complex char* epi8_dst = (_Complex char*) malloc (2 * N * sizeof (char));
> +  char* p_init = (char*) malloc (2 * N * sizeof (double));
> +
> +  __builtin_memset (pd_dst, 0, 2 * N * sizeof (double));
> +  __builtin_memset (ps_dst, 0, 2 * N * sizeof (float));
> +  __builtin_memset (epi64_dst, 0, 2 * N * sizeof (long long));
> +  __builtin_memset (epi32_dst, 0, 2 * N * sizeof (int));
> +  __builtin_memset (epi16_dst, 0, 2 * N * sizeof (short));
> +  __builtin_memset (epi8_dst, 0, 2 * N * sizeof (char));
> +
> +  for (int i = 0; i != 2 * N * sizeof (double); i++)
> +    p_init[i] = i;
> +
> +  memcpy (pd_src, p_init, 2 * N * sizeof (double));
> +  memcpy (ps_src, p_init, 2 * N * sizeof (float));
> +  memcpy (epi64_src, p_init, 2 * N * sizeof (long long));
> +  memcpy (epi32_src, p_init, 2 * N * sizeof (int));
> +  memcpy (epi16_src, p_init, 2 * N * sizeof (short));
> +  memcpy (epi8_src, p_init, 2 * N * sizeof (char));
> +
> +  foo_pd (pd_dst, pd_src);
> +  foo_ps (ps_dst, ps_src);
> +  foo_epi64 (epi64_dst, epi64_src);
> +  foo_epi32 (epi32_dst, epi32_src);
> +  foo_epi16 (epi16_dst, epi16_src);
> +  foo_epi8 (epi8_dst, epi8_src);
> +  if (__builtin_memcmp (pd_dst, pd_src, N * 2 * sizeof (double)) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (ps_dst, ps_src, N * 2 * sizeof (float)) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi64_dst, epi64_src, N * 2 * sizeof (long long)) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi32_dst, epi32_src, N * 2 * sizeof (int)) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi16_dst, epi16_src, N * 2 * sizeof (short)) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi8_dst, epi8_src, N * 2 * sizeof (char)) != 0)
> +    __builtin_abort ();
> +
> +  return;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-1c.c b/gcc/testsuite/gcc.target/i386/pr106010-1c.c
> new file mode 100644
> index 00000000000..f07e9fb2d3d
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-1c.c
> @@ -0,0 +1,41 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx512fp16 -mavx512vl -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256 -fdump-tree-vect-details" } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(16\) _Float16>} 2 "vect" } } */
> +/* { dg-require-effective-target avx512fp16 } */
> +
> +#include <string.h>
> +
> +static void do_test (void);
> +
> +#define DO_TEST do_test
> +#define AVX512FP16
> +#include "avx512-check.h"
> +
> +#define N 10000
> +
> +void
> +__attribute__((noipa))
> +foo_ph (_Complex _Float16* a, _Complex _Float16* b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b[i];
> +}
> +
> +static void
> +do_test (void)
> +{
> +  _Complex _Float16* ph_src = (_Complex _Float16*) malloc (2 * N * sizeof (_Float16));
> +  _Complex _Float16* ph_dst = (_Complex _Float16*) malloc (2 * N * sizeof (_Float16));
> +  char* p_init = (char*) malloc (2 * N * sizeof (_Float16));
> +
> +  __builtin_memset (ph_dst, 0, 2 * N * sizeof (_Float16));
> +
> +  for (int i = 0; i != 2 * N * sizeof (_Float16); i++)
> +    p_init[i] = i;
> +
> +  memcpy (ph_src, p_init, 2 * N * sizeof (_Float16));
> +
> +  foo_ph (ph_dst, ph_src);
> +  if (__builtin_memcmp (ph_dst, ph_src, N * 2 * sizeof (_Float16)) != 0)
> +    __builtin_abort ();
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-2a.c b/gcc/testsuite/gcc.target/i386/pr106010-2a.c
> new file mode 100644
> index 00000000000..d2e2f8d4f43
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-2a.c
> @@ -0,0 +1,82 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx -ftree-vectorize -fvect-cost-model=unlimited -fdump-tree-slp-details -mprefer-vector-width=256" } */
> +/* { dg-final { scan-tree-dump-times "basic block part vectorized using (?:32|64) byte vectors" 6 "slp2" } }*/
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(4\) double>} 2 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(8\) float>} 2 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(4\) long long int>} 2 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(8\) int>} 2 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(16\) short int>} 2 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(32\) char>} 2 "slp2" } } */
> +
> +void
> +__attribute__((noipa))
> +foo_pd (_Complex double* a, _Complex double* __restrict b)
> +{
> +  a[0] = b[0];
> +  a[1] = b[1];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_ps (_Complex float* a, _Complex float* __restrict b)
> +{
> +  a[0] = b[0];
> +  a[1] = b[1];
> +  a[2] = b[2];
> +  a[3] = b[3];
> +
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi64 (_Complex long long* a, _Complex long long* __restrict b)
> +{
> +  a[0] = b[0];
> +  a[1] = b[1];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi32 (_Complex int* a, _Complex int* __restrict b)
> +{
> +  a[0] = b[0];
> +  a[1] = b[1];
> +  a[2] = b[2];
> +  a[3] = b[3];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi16 (_Complex short* a, _Complex short* __restrict b)
> +{
> +  a[0] = b[0];
> +  a[1] = b[1];
> +  a[2] = b[2];
> +  a[3] = b[3];
> +  a[4] = b[4];
> +  a[5] = b[5];
> +  a[6] = b[6];
> +  a[7] = b[7];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi8 (_Complex char* a, _Complex char* __restrict b)
> +{
> +  a[0] = b[0];
> +  a[1] = b[1];
> +  a[2] = b[2];
> +  a[3] = b[3];
> +  a[4] = b[4];
> +  a[5] = b[5];
> +  a[6] = b[6];
> +  a[7] = b[7];
> +  a[8] = b[8];
> +  a[9] = b[9];
> +  a[10] = b[10];
> +  a[11] = b[11];
> +  a[12] = b[12];
> +  a[13] = b[13];
> +  a[14] = b[14];
> +  a[15] = b[15];
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-2b.c b/gcc/testsuite/gcc.target/i386/pr106010-2b.c
> new file mode 100644
> index 00000000000..ac360752693
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-2b.c
> @@ -0,0 +1,62 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256" } */
> +/* { dg-require-effective-target avx } */
> +
> +#include "avx-check.h"
> +#include <string.h>
> +#include "pr106010-2a.c"
> +
> +void
> +avx_test (void)
> +{
> +  _Complex double* pd_src = (_Complex double*) malloc (32);
> +  _Complex double* pd_dst = (_Complex double*) malloc (32);
> +  _Complex float* ps_src = (_Complex float*) malloc (32);
> +  _Complex float* ps_dst = (_Complex float*) malloc (32);
> +  _Complex long long* epi64_src = (_Complex long long*) malloc (32);
> +  _Complex long long* epi64_dst = (_Complex long long*) malloc (32);
> +  _Complex int* epi32_src = (_Complex int*) malloc (32);
> +  _Complex int* epi32_dst = (_Complex int*) malloc (32);
> +  _Complex short* epi16_src = (_Complex short*) malloc (32);
> +  _Complex short* epi16_dst = (_Complex short*) malloc (32);
> +  _Complex char* epi8_src = (_Complex char*) malloc (32);
> +  _Complex char* epi8_dst = (_Complex char*) malloc (32);
> +  char* p = (char* ) malloc (32);
> +
> +  __builtin_memset (pd_dst, 0, 32);
> +  __builtin_memset (ps_dst, 0, 32);
> +  __builtin_memset (epi64_dst, 0, 32);
> +  __builtin_memset (epi32_dst, 0, 32);
> +  __builtin_memset (epi16_dst, 0, 32);
> +  __builtin_memset (epi8_dst, 0, 32);
> +
> +  for (int i = 0; i != 32; i++)
> +    p[i] = i;
> +  __builtin_memcpy (pd_src, p, 32);
> +  __builtin_memcpy (ps_src, p, 32);
> +  __builtin_memcpy (epi64_src, p, 32);
> +  __builtin_memcpy (epi32_src, p, 32);
> +  __builtin_memcpy (epi16_src, p, 32);
> +  __builtin_memcpy (epi8_src, p, 32);
> +
> +  foo_pd (pd_dst, pd_src);
> +  foo_ps (ps_dst, ps_src);
> +  foo_epi64 (epi64_dst, epi64_src);
> +  foo_epi32 (epi32_dst, epi32_src);
> +  foo_epi16 (epi16_dst, epi16_src);
> +  foo_epi8 (epi8_dst, epi8_src);
> +  if (__builtin_memcmp (pd_dst, pd_src, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (ps_dst, ps_src, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi64_dst, epi64_src, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi32_dst, epi32_src, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi16_dst, epi16_src, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi16_dst, epi16_src, 32) != 0)
> +    __builtin_abort ();
> +
> +  return;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-2c.c b/gcc/testsuite/gcc.target/i386/pr106010-2c.c
> new file mode 100644
> index 00000000000..a002f209ec9
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-2c.c
> @@ -0,0 +1,47 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx512fp16 -mavx512vl -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256 -fdump-tree-slp-details" } */
> +/* { dg-require-effective-target avx512fp16 } */
> +
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(16\) _Float16>} 2 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times "basic block part vectorized using (?:32|64) byte vectors" 1 "slp2" } }*/
> +
> +#include <string.h>
> +
> +static void do_test (void);
> +#define DO_TEST do_test
> +#define AVX512FP16
> +#include "avx512-check.h"
> +
> +void
> +__attribute__((noipa))
> +foo_ph (_Complex _Float16* a, _Complex _Float16* __restrict b)
> +{
> +  a[0] = b[0];
> +  a[1] = b[1];
> +  a[2] = b[2];
> +  a[3] = b[3];
> +  a[4] = b[4];
> +  a[5] = b[5];
> +  a[6] = b[6];
> +  a[7] = b[7];
> +}
> +
> +void
> +do_test (void)
> +{
> +  _Complex _Float16* ph_src = (_Complex _Float16*) malloc (32);
> +  _Complex _Float16* ph_dst = (_Complex _Float16*) malloc (32);
> +  char* p = (char* ) malloc (32);
> +
> +   __builtin_memset (ph_dst, 0, 32);
> +
> +  for (int i = 0; i != 32; i++)
> +    p[i] = i;
> +  __builtin_memcpy (ph_src, p, 32);
> +
> +  foo_ph (ph_dst, ph_src);
> +  if (__builtin_memcmp (ph_dst, ph_src, 32) != 0)
> +    __builtin_abort ();
> +
> +  return;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-3a.c b/gcc/testsuite/gcc.target/i386/pr106010-3a.c
> new file mode 100644
> index 00000000000..c1b64b56b1c
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-3a.c
> @@ -0,0 +1,80 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx2 -ftree-vectorize -fvect-cost-model=unlimited -fdump-tree-slp-details" } */
> +/* { dg-final { scan-tree-dump-times "basic block part vectorized using (?:32|64) byte vectors" 6 "slp2" } }*/
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*VEC_PERM_EXPR.*\{ 2, 3, 0, 1 \}} 2 "slp2" } }  */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*VEC_PERM_EXPR.*\{ 6, 7, 4, 5, 2, 3, 0, 1 \}} 1 "slp2" } }  */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*VEC_PERM_EXPR.*\{ 2, 3, 0, 1, 6, 7, 4, 5 \}} 1 "slp2" } }  */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*VEC_PERM_EXPR.*\{ 14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1 \}} 1 "slp2" } }  */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*VEC_PERM_EXPR.*\{ 14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1, 30, 31, 28, 29, 26, 27, 24, 25, 22, 23, 20, 21, 18, 19, 16, 17 \}} 1 "slp2" } }  */
> +
> +void
> +__attribute__((noipa))
> +foo_pd (_Complex double* a, _Complex double* __restrict b)
> +{
> +  a[0] = b[1];
> +  a[1] = b[0];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_ps (_Complex float* a, _Complex float* __restrict b)
> +{
> +  a[0] = b[1];
> +  a[1] = b[0];
> +  a[2] = b[3];
> +  a[3] = b[2];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi64 (_Complex long long* a, _Complex long long* __restrict b)
> +{
> +  a[0] = b[1];
> +  a[1] = b[0];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi32 (_Complex int* a, _Complex int* __restrict b)
> +{
> +  a[0] = b[3];
> +  a[1] = b[2];
> +  a[2] = b[1];
> +  a[3] = b[0];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi16 (_Complex short* a, _Complex short* __restrict b)
> +{
> +  a[0] = b[7];
> +  a[1] = b[6];
> +  a[2] = b[5];
> +  a[3] = b[4];
> +  a[4] = b[3];
> +  a[5] = b[2];
> +  a[6] = b[1];
> +  a[7] = b[0];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi8 (_Complex char* a, _Complex char* __restrict b)
> +{
> +  a[0] = b[7];
> +  a[1] = b[6];
> +  a[2] = b[5];
> +  a[3] = b[4];
> +  a[4] = b[3];
> +  a[5] = b[2];
> +  a[6] = b[1];
> +  a[7] = b[0];
> +  a[8] = b[15];
> +  a[9] = b[14];
> +  a[10] = b[13];
> +  a[11] = b[12];
> +  a[12] = b[11];
> +  a[13] = b[10];
> +  a[14] = b[9];
> +  a[15] = b[8];
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-3b.c b/gcc/testsuite/gcc.target/i386/pr106010-3b.c
> new file mode 100644
> index 00000000000..e4fa3f3a541
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-3b.c
> @@ -0,0 +1,126 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx2 -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256" } */
> +/* { dg-require-effective-target avx2 } */
> +
> +#include "avx2-check.h"
> +#include <string.h>
> +#include "pr106010-3a.c"
> +
> +void
> +avx2_test (void)
> +{
> +  _Complex double* pd_src = (_Complex double*) malloc (32);
> +  _Complex double* pd_dst = (_Complex double*) malloc (32);
> +  _Complex double* pd_exp = (_Complex double*) malloc (32);
> +  _Complex float* ps_src = (_Complex float*) malloc (32);
> +  _Complex float* ps_dst = (_Complex float*) malloc (32);
> +  _Complex float* ps_exp = (_Complex float*) malloc (32);
> +  _Complex long long* epi64_src = (_Complex long long*) malloc (32);
> +  _Complex long long* epi64_dst = (_Complex long long*) malloc (32);
> +  _Complex long long* epi64_exp = (_Complex long long*) malloc (32);
> +  _Complex int* epi32_src = (_Complex int*) malloc (32);
> +  _Complex int* epi32_dst = (_Complex int*) malloc (32);
> +  _Complex int* epi32_exp = (_Complex int*) malloc (32);
> +  _Complex short* epi16_src = (_Complex short*) malloc (32);
> +  _Complex short* epi16_dst = (_Complex short*) malloc (32);
> +  _Complex short* epi16_exp = (_Complex short*) malloc (32);
> +  _Complex char* epi8_src = (_Complex char*) malloc (32);
> +  _Complex char* epi8_dst = (_Complex char*) malloc (32);
> +  _Complex char* epi8_exp = (_Complex char*) malloc (32);
> +  char* p = (char* ) malloc (32);
> +  char* q = (char* ) malloc (32);
> +
> +  __builtin_memset (pd_dst, 0, 32);
> +  __builtin_memset (ps_dst, 0, 32);
> +  __builtin_memset (epi64_dst, 0, 32);
> +  __builtin_memset (epi32_dst, 0, 32);
> +  __builtin_memset (epi16_dst, 0, 32);
> +  __builtin_memset (epi8_dst, 0, 32);
> +
> +  for (int i = 0; i != 32; i++)
> +    p[i] = i;
> +  __builtin_memcpy (pd_src, p, 32);
> +  __builtin_memcpy (ps_src, p, 32);
> +  __builtin_memcpy (epi64_src, p, 32);
> +  __builtin_memcpy (epi32_src, p, 32);
> +  __builtin_memcpy (epi16_src, p, 32);
> +  __builtin_memcpy (epi8_src, p, 32);
> +
> +  for (int i = 0; i != 16; i++)
> +    {
> +      p[i] = i + 16;
> +      p[i + 16] = i;
> +    }
> +  __builtin_memcpy (pd_exp, p, 32);
> +  __builtin_memcpy (epi64_exp, p, 32);
> +
> +  for (int i = 0; i != 8; i++)
> +    {
> +      p[i] = i + 8;
> +      p[i + 8] = i;
> +      p[i + 16] = i + 24;
> +      p[i + 24] = i + 16;
> +      q[i] = i + 24;
> +      q[i + 8] = i + 16;
> +      q[i + 16] = i + 8;
> +      q[i + 24] = i;
> +    }
> +  __builtin_memcpy (ps_exp, p, 32);
> +  __builtin_memcpy (epi32_exp, q, 32);
> +
> +
> +  for (int i = 0; i != 4; i++)
> +    {
> +      q[i] = i + 28;
> +      q[i + 4] = i + 24;
> +      q[i + 8] = i + 20;
> +      q[i + 12] = i + 16;
> +      q[i + 16] = i + 12;
> +      q[i + 20] = i + 8;
> +      q[i + 24] = i + 4;
> +      q[i + 28] = i;
> +    }
> +  __builtin_memcpy (epi16_exp, q, 32);
> +
> +  for (int i = 0; i != 2; i++)
> +    {
> +      q[i] = i + 14;
> +      q[i + 2] = i + 12;
> +      q[i + 4] = i + 10;
> +      q[i + 6] = i + 8;
> +      q[i + 8] = i + 6;
> +      q[i + 10] = i + 4;
> +      q[i + 12] = i + 2;
> +      q[i + 14] = i;
> +      q[i + 16] = i + 30;
> +      q[i + 18] = i + 28;
> +      q[i + 20] = i + 26;
> +      q[i + 22] = i + 24;
> +      q[i + 24] = i + 22;
> +      q[i + 26] = i + 20;
> +      q[i + 28] = i + 18;
> +      q[i + 30] = i + 16;
> +    }
> +  __builtin_memcpy (epi8_exp, q, 32);
> +
> +  foo_pd (pd_dst, pd_src);
> +  foo_ps (ps_dst, ps_src);
> +  foo_epi64 (epi64_dst, epi64_src);
> +  foo_epi32 (epi32_dst, epi32_src);
> +  foo_epi16 (epi16_dst, epi16_src);
> +  foo_epi8 (epi8_dst, epi8_src);
> +  if (__builtin_memcmp (pd_dst, pd_exp, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (ps_dst, ps_exp, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi64_dst, epi64_exp, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi32_dst, epi32_exp, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi16_dst, epi16_exp, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi8_dst, epi8_exp, 32) != 0)
> +    __builtin_abort ();
> +
> +  return;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-3c.c b/gcc/testsuite/gcc.target/i386/pr106010-3c.c
> new file mode 100644
> index 00000000000..5a5a3d4b992
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-3c.c
> @@ -0,0 +1,69 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx512fp16 -mavx512vl -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256 -fdump-tree-slp-details" } */
> +/* { dg-require-effective-target avx512fp16 } */
> +/* { dg-final { scan-tree-dump-times "basic block part vectorized using (?:32|64) byte vectors" 1 "slp2" } }*/
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*VEC_PERM_EXPR.*\{ 2, 3, 0, 1, 8, 9, 6, 7, 14, 15, 12, 13, 4, 5, 10, 11 \}} 1 "slp2" } }  */
> +
> +#include <string.h>
> +
> +static void do_test (void);
> +#define DO_TEST do_test
> +#define AVX512FP16
> +#include "avx512-check.h"
> +
> +void
> +__attribute__((noipa))
> +foo_ph (_Complex _Float16* a, _Complex _Float16* __restrict b)
> +{
> +  a[0] = b[1];
> +  a[1] = b[0];
> +  a[2] = b[4];
> +  a[3] = b[3];
> +  a[4] = b[7];
> +  a[5] = b[6];
> +  a[6] = b[2];
> +  a[7] = b[5];
> +}
> +
> +void
> +do_test (void)
> +{
> +  _Complex _Float16* ph_src = (_Complex _Float16*) malloc (32);
> +  _Complex _Float16* ph_dst = (_Complex _Float16*) malloc (32);
> +  _Complex _Float16* ph_exp = (_Complex _Float16*) malloc (32);
> +  char* p = (char* ) malloc (32);
> +  char* q = (char* ) malloc (32);
> +
> +  __builtin_memset (ph_dst, 0, 32);
> +
> +  for (int i = 0; i != 32; i++)
> +    p[i] = i;
> +  __builtin_memcpy (ph_src, p, 32);
> +
> +  for (int i = 0; i != 4; i++)
> +    {
> +      p[i] = i + 4;
> +      p[i + 4] = i;
> +      p[i + 8] = i + 16;
> +      p[i + 12] = i + 12;
> +      p[i + 16] = i + 28;
> +      p[i + 20] = i + 24;
> +      p[i + 24] = i + 8;
> +      p[i + 28] = i + 20;
> +      q[i] = i + 28;
> +      q[i + 4] = i + 24;
> +      q[i + 8] = i + 20;
> +      q[i + 12] = i + 16;
> +      q[i + 16] = i + 12;
> +      q[i + 20] = i + 8;
> +      q[i + 24] = i + 4;
> +      q[i + 28] = i;
> +    }
> +  __builtin_memcpy (ph_exp, p, 32);
> +
> +  foo_ph (ph_dst, ph_src);
> +  if (__builtin_memcmp (ph_dst, ph_exp, 32) != 0)
> +    __builtin_abort ();
> +
> +  return;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-4a.c b/gcc/testsuite/gcc.target/i386/pr106010-4a.c
> new file mode 100644
> index 00000000000..b7b0b532bb1
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-4a.c
> @@ -0,0 +1,101 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx -ftree-vectorize -fvect-cost-model=unlimited -fdump-tree-slp-details" } */
> +/* { dg-final { scan-tree-dump-times "basic block part vectorized using (?:32|64) byte vectors" 6 "slp2" } }*/
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(4\) double>} 1 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(8\) float>} 1 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(4\) long long int>} 1 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(8\) int>} 1 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(16\) short int>} 1 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(32\) char>} 1 "slp2" } } */
> +
> +void
> +__attribute__((noipa))
> +foo_pd (_Complex double* a,
> +       _Complex double b1,
> +       _Complex double b2)
> +{
> +  a[0] = b1;
> +  a[1] = b2;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_ps (_Complex float* a,
> +       _Complex float b1, _Complex float b2,
> +       _Complex float b3, _Complex float b4)
> +{
> +  a[0] = b1;
> +  a[1] = b2;
> +  a[2] = b3;
> +  a[3] = b4;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi64 (_Complex long long* a,
> +          _Complex long long b1,
> +          _Complex long long b2)
> +{
> +  a[0] = b1;
> +  a[1] = b2;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi32 (_Complex int* a,
> +          _Complex int b1, _Complex int b2,
> +          _Complex int b3, _Complex int b4)
> +{
> +  a[0] = b1;
> +  a[1] = b2;
> +  a[2] = b3;
> +  a[3] = b4;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi16 (_Complex short* a,
> +          _Complex short b1, _Complex short b2,
> +          _Complex short b3, _Complex short b4,
> +          _Complex short b5, _Complex short b6,
> +          _Complex short b7,_Complex short b8)
> +{
> +  a[0] = b1;
> +  a[1] = b2;
> +  a[2] = b3;
> +  a[3] = b4;
> +  a[4] = b5;
> +  a[5] = b6;
> +  a[6] = b7;
> +  a[7] = b8;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi8 (_Complex char* a,
> +         _Complex char b1, _Complex char b2,
> +         _Complex char b3, _Complex char b4,
> +         _Complex char b5, _Complex char b6,
> +         _Complex char b7,_Complex char b8,
> +         _Complex char b9, _Complex char b10,
> +         _Complex char b11, _Complex char b12,
> +         _Complex char b13, _Complex char b14,
> +         _Complex char b15,_Complex char b16)
> +{
> +  a[0] = b1;
> +  a[1] = b2;
> +  a[2] = b3;
> +  a[3] = b4;
> +  a[4] = b5;
> +  a[5] = b6;
> +  a[6] = b7;
> +  a[7] = b8;
> +  a[8] = b9;
> +  a[9] = b10;
> +  a[10] = b11;
> +  a[11] = b12;
> +  a[12] = b13;
> +  a[13] = b14;
> +  a[14] = b15;
> +  a[15] = b16;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-4b.c b/gcc/testsuite/gcc.target/i386/pr106010-4b.c
> new file mode 100644
> index 00000000000..e2e79508c4b
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-4b.c
> @@ -0,0 +1,67 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256" } */
> +/* { dg-require-effective-target avx } */
> +
> +#include "avx-check.h"
> +#include <string.h>
> +#include "pr106010-4a.c"
> +
> +void
> +avx_test (void)
> +{
> +  _Complex double* pd_src = (_Complex double*) malloc (32);
> +  _Complex double* pd_dst = (_Complex double*) malloc (32);
> +  _Complex float* ps_src = (_Complex float*) malloc (32);
> +  _Complex float* ps_dst = (_Complex float*) malloc (32);
> +  _Complex long long* epi64_src = (_Complex long long*) malloc (32);
> +  _Complex long long* epi64_dst = (_Complex long long*) malloc (32);
> +  _Complex int* epi32_src = (_Complex int*) malloc (32);
> +  _Complex int* epi32_dst = (_Complex int*) malloc (32);
> +  _Complex short* epi16_src = (_Complex short*) malloc (32);
> +  _Complex short* epi16_dst = (_Complex short*) malloc (32);
> +  _Complex char* epi8_src = (_Complex char*) malloc (32);
> +  _Complex char* epi8_dst = (_Complex char*) malloc (32);
> +  char* p = (char* ) malloc (32);
> +
> +  __builtin_memset (pd_dst, 0, 32);
> +  __builtin_memset (ps_dst, 0, 32);
> +  __builtin_memset (epi64_dst, 0, 32);
> +  __builtin_memset (epi32_dst, 0, 32);
> +  __builtin_memset (epi16_dst, 0, 32);
> +  __builtin_memset (epi8_dst, 0, 32);
> +
> +  for (int i = 0; i != 32; i++)
> +    p[i] = i;
> +  __builtin_memcpy (pd_src, p, 32);
> +  __builtin_memcpy (ps_src, p, 32);
> +  __builtin_memcpy (epi64_src, p, 32);
> +  __builtin_memcpy (epi32_src, p, 32);
> +  __builtin_memcpy (epi16_src, p, 32);
> +  __builtin_memcpy (epi8_src, p, 32);
> +
> +  foo_pd (pd_dst, pd_src[0], pd_src[1]);
> +  foo_ps (ps_dst, ps_src[0], ps_src[1], ps_src[2], ps_src[3]);
> +  foo_epi64 (epi64_dst, epi64_src[0], epi64_src[1]);
> +  foo_epi32 (epi32_dst, epi32_src[0], epi32_src[1], epi32_src[2], epi32_src[3]);
> +  foo_epi16 (epi16_dst, epi16_src[0], epi16_src[1], epi16_src[2], epi16_src[3],
> +            epi16_src[4], epi16_src[5], epi16_src[6], epi16_src[7]);
> +  foo_epi8 (epi8_dst, epi8_src[0], epi8_src[1], epi8_src[2], epi8_src[3],
> +           epi8_src[4], epi8_src[5], epi8_src[6], epi8_src[7],
> +           epi8_src[8], epi8_src[9], epi8_src[10], epi8_src[11],
> +           epi8_src[12], epi8_src[13], epi8_src[14], epi8_src[15]);
> +
> +  if (__builtin_memcmp (pd_dst, pd_src, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (ps_dst, ps_src, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi64_dst, epi64_src, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi32_dst, epi32_src, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi16_dst, epi16_src, 32) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi8_dst, epi8_src, 32) != 0)
> +    __builtin_abort ();
> +
> +  return;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-4c.c b/gcc/testsuite/gcc.target/i386/pr106010-4c.c
> new file mode 100644
> index 00000000000..8e02aefe3b5
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-4c.c
> @@ -0,0 +1,54 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx512fp16 -mavx512vl -fdump-tree-slp-details -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256" } */
> +/* { dg-require-effective-target avx512fp16 } */
> +/* { dg-final { scan-tree-dump-times "basic block part vectorized using (?:32|64) byte vectors" 1 "slp2" } }*/
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(16\) _Float16>} 1 "slp2" } } */
> +
> +#include <string.h>
> +
> +static void do_test (void);
> +#define DO_TEST do_test
> +#define AVX512FP16
> +#include "avx512-check.h"
> +
> +void
> +__attribute__((noipa))
> +foo_ph (_Complex _Float16* a,
> +       _Complex _Float16 b1, _Complex _Float16 b2,
> +       _Complex _Float16 b3, _Complex _Float16 b4,
> +       _Complex _Float16 b5, _Complex _Float16 b6,
> +       _Complex _Float16 b7,_Complex _Float16 b8)
> +{
> +  a[0] = b1;
> +  a[1] = b2;
> +  a[2] = b3;
> +  a[3] = b4;
> +  a[4] = b5;
> +  a[5] = b6;
> +  a[6] = b7;
> +  a[7] = b8;
> +}
> +
> +void
> +do_test (void)
> +{
> +
> +  _Complex _Float16* ph_src = (_Complex _Float16*) malloc (32);
> +  _Complex _Float16* ph_dst = (_Complex _Float16*) malloc (32);
> +
> +  char* p = (char* ) malloc (32);
> +
> +  __builtin_memset (ph_dst, 0, 32);
> +
> +  for (int i = 0; i != 32; i++)
> +    p[i] = i;
> +
> +  __builtin_memcpy (ph_src, p, 32);
> +
> +  foo_ph (ph_dst, ph_src[0], ph_src[1], ph_src[2], ph_src[3],
> +         ph_src[4], ph_src[5], ph_src[6], ph_src[7]);
> +
> +  if (__builtin_memcmp (ph_dst, ph_src, 32) != 0)
> +    __builtin_abort ();
> +  return;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-5a.c b/gcc/testsuite/gcc.target/i386/pr106010-5a.c
> new file mode 100644
> index 00000000000..9d4a6f9846b
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-5a.c
> @@ -0,0 +1,117 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx -ftree-vectorize -fvect-cost-model=unlimited -fdump-tree-slp-details -mprefer-vector-width=256" } */
> +/* { dg-final { scan-tree-dump-times "basic block part vectorized using (?:32|64) byte vectors" 6 "slp2" } }*/
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(4\) double>} 4 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(8\) float>} 4 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(4\) long long int>} 4 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(8\) int>} 4 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(16\) short int>} 4 "slp2" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(32\) char>} 4 "slp2" } } */
> +
> +void
> +__attribute__((noipa))
> +foo_pd (_Complex double* a, _Complex double* __restrict b)
> +{
> +  a[0] = b[2];
> +  a[1] = b[3];
> +  a[2] = b[0];
> +  a[3] = b[1];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_ps (_Complex float* a, _Complex float* __restrict b)
> +{
> +  a[0] = b[4];
> +  a[1] = b[5];
> +  a[2] = b[6];
> +  a[3] = b[7];
> +  a[4] = b[0];
> +  a[5] = b[1];
> +  a[6] = b[2];
> +  a[7] = b[3];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi64 (_Complex long long* a, _Complex long long* __restrict b)
> +{
> +  a[0] = b[2];
> +  a[1] = b[3];
> +  a[2] = b[0];
> +  a[3] = b[1];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi32 (_Complex int* a, _Complex int* __restrict b)
> +{
> +  a[0] = b[4];
> +  a[1] = b[5];
> +  a[2] = b[6];
> +  a[3] = b[7];
> +  a[4] = b[0];
> +  a[5] = b[1];
> +  a[6] = b[2];
> +  a[7] = b[3];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi16 (_Complex short* a, _Complex short* __restrict b)
> +{
> +  a[0] = b[8];
> +  a[1] = b[9];
> +  a[2] = b[10];
> +  a[3] = b[11];
> +  a[4] = b[12];
> +  a[5] = b[13];
> +  a[6] = b[14];
> +  a[7] = b[15];
> +  a[8] = b[0];
> +  a[9] = b[1];
> +  a[10] = b[2];
> +  a[11] = b[3];
> +  a[12] = b[4];
> +  a[13] = b[5];
> +  a[14] = b[6];
> +  a[15] = b[7];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi8 (_Complex char* a, _Complex char* __restrict b)
> +{
> +  a[0] = b[16];
> +  a[1] = b[17];
> +  a[2] = b[18];
> +  a[3] = b[19];
> +  a[4] = b[20];
> +  a[5] = b[21];
> +  a[6] = b[22];
> +  a[7] = b[23];
> +  a[8] = b[24];
> +  a[9] = b[25];
> +  a[10] = b[26];
> +  a[11] = b[27];
> +  a[12] = b[28];
> +  a[13] = b[29];
> +  a[14] = b[30];
> +  a[15] = b[31];
> +  a[16] = b[0];
> +  a[17] = b[1];
> +  a[18] = b[2];
> +  a[19] = b[3];
> +  a[20] = b[4];
> +  a[21] = b[5];
> +  a[22] = b[6];
> +  a[23] = b[7];
> +  a[24] = b[8];
> +  a[25] = b[9];
> +  a[26] = b[10];
> +  a[27] = b[11];
> +  a[28] = b[12];
> +  a[29] = b[13];
> +  a[30] = b[14];
> +  a[31] = b[15];
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-5b.c b/gcc/testsuite/gcc.target/i386/pr106010-5b.c
> new file mode 100644
> index 00000000000..d5c6ebeb5cf
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-5b.c
> @@ -0,0 +1,80 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256" } */
> +/* { dg-require-effective-target avx } */
> +
> +#include "avx-check.h"
> +#include <string.h>
> +#include "pr106010-5a.c"
> +
> +void
> +avx_test (void)
> +{
> +  _Complex double* pd_src = (_Complex double*) malloc (64);
> +  _Complex double* pd_dst = (_Complex double*) malloc (64);
> +  _Complex double* pd_exp = (_Complex double*) malloc (64);
> +  _Complex float* ps_src = (_Complex float*) malloc (64);
> +  _Complex float* ps_dst = (_Complex float*) malloc (64);
> +  _Complex float* ps_exp = (_Complex float*) malloc (64);
> +  _Complex long long* epi64_src = (_Complex long long*) malloc (64);
> +  _Complex long long* epi64_dst = (_Complex long long*) malloc (64);
> +  _Complex long long* epi64_exp = (_Complex long long*) malloc (64);
> +  _Complex int* epi32_src = (_Complex int*) malloc (64);
> +  _Complex int* epi32_dst = (_Complex int*) malloc (64);
> +  _Complex int* epi32_exp = (_Complex int*) malloc (64);
> +  _Complex short* epi16_src = (_Complex short*) malloc (64);
> +  _Complex short* epi16_dst = (_Complex short*) malloc (64);
> +  _Complex short* epi16_exp = (_Complex short*) malloc (64);
> +  _Complex char* epi8_src = (_Complex char*) malloc (64);
> +  _Complex char* epi8_dst = (_Complex char*) malloc (64);
> +  _Complex char* epi8_exp = (_Complex char*) malloc (64);
> +  char* p = (char* ) malloc (64);
> +  char* q = (char* ) malloc (64);
> +
> +  __builtin_memset (pd_dst, 0, 64);
> +  __builtin_memset (ps_dst, 0, 64);
> +  __builtin_memset (epi64_dst, 0, 64);
> +  __builtin_memset (epi32_dst, 0, 64);
> +  __builtin_memset (epi16_dst, 0, 64);
> +  __builtin_memset (epi8_dst, 0, 64);
> +
> +  for (int i = 0; i != 64; i++)
> +    {
> +      p[i] = i;
> +      q[i] = (i + 32) % 64;
> +    }
> +  __builtin_memcpy (pd_src, p, 64);
> +  __builtin_memcpy (ps_src, p, 64);
> +  __builtin_memcpy (epi64_src, p, 64);
> +  __builtin_memcpy (epi32_src, p, 64);
> +  __builtin_memcpy (epi16_src, p, 64);
> +  __builtin_memcpy (epi8_src, p, 64);
> +
> +  __builtin_memcpy (pd_exp, q, 64);
> +  __builtin_memcpy (ps_exp, q, 64);
> +  __builtin_memcpy (epi64_exp, q, 64);
> +  __builtin_memcpy (epi32_exp, q, 64);
> +  __builtin_memcpy (epi16_exp, q, 64);
> +  __builtin_memcpy (epi8_exp, q, 64);
> +
> +  foo_pd (pd_dst, pd_src);
> +  foo_ps (ps_dst, ps_src);
> +  foo_epi64 (epi64_dst, epi64_src);
> +  foo_epi32 (epi32_dst, epi32_src);
> +  foo_epi16 (epi16_dst, epi16_src);
> +  foo_epi8 (epi8_dst, epi8_src);
> +
> +  if (__builtin_memcmp (pd_dst, pd_exp, 64) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (ps_dst, ps_exp, 64) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi64_dst, epi64_exp, 64) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi32_dst, epi32_exp, 64) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi16_dst, epi16_exp, 64) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi8_dst, epi8_exp, 64) != 0)
> +    __builtin_abort ();
> +
> +  return;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-5c.c b/gcc/testsuite/gcc.target/i386/pr106010-5c.c
> new file mode 100644
> index 00000000000..9ce4e6dd5c0
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-5c.c
> @@ -0,0 +1,62 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx512fp16 -mavx512vl -ftree-vectorize -fvect-cost-model=unlimited -fdump-tree-slp-details -mprefer-vector-width=256" } */
> +/* { dg-require-effective-target avx512fp16 } */
> +/* { dg-final { scan-tree-dump-times "basic block part vectorized using (?:32|64) byte vectors" 1 "slp2" } }*/
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(16\) _Float16>} 4 "slp2" } } */
> +
> +#include <string.h>
> +
> +static void do_test (void);
> +#define DO_TEST do_test
> +#define AVX512FP16
> +#include "avx512-check.h"
> +
> +void
> +__attribute__((noipa))
> +foo_ph (_Complex _Float16* a, _Complex _Float16* __restrict b)
> +{
> +  a[0] = b[8];
> +  a[1] = b[9];
> +  a[2] = b[10];
> +  a[3] = b[11];
> +  a[4] = b[12];
> +  a[5] = b[13];
> +  a[6] = b[14];
> +  a[7] = b[15];
> +  a[8] = b[0];
> +  a[9] = b[1];
> +  a[10] = b[2];
> +  a[11] = b[3];
> +  a[12] = b[4];
> +  a[13] = b[5];
> +  a[14] = b[6];
> +  a[15] = b[7];
> +}
> +
> +void
> +do_test (void)
> +{
> +  _Complex _Float16* ph_src = (_Complex _Float16*) malloc (64);
> +  _Complex _Float16* ph_dst = (_Complex _Float16*) malloc (64);
> +  _Complex _Float16* ph_exp = (_Complex _Float16*) malloc (64);
> +  char* p = (char* ) malloc (64);
> +  char* q = (char* ) malloc (64);
> +
> +  __builtin_memset (ph_dst, 0, 64);
> +
> +  for (int i = 0; i != 64; i++)
> +    {
> +      p[i] = i;
> +      q[i] = (i + 32) % 64;
> +    }
> +  __builtin_memcpy (ph_src, p, 64);
> +
> +  __builtin_memcpy (ph_exp, q, 64);
> +
> +  foo_ph (ph_dst, ph_src);
> +
> +  if (__builtin_memcmp (ph_dst, ph_exp, 64) != 0)
> +    __builtin_abort ();
> +
> +  return;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-6a.c b/gcc/testsuite/gcc.target/i386/pr106010-6a.c
> new file mode 100644
> index 00000000000..65a90d03684
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-6a.c
> @@ -0,0 +1,115 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx2 -ftree-vectorize -fvect-cost-model=unlimited -fdump-tree-slp-details -mprefer-vector-width=256" } */
> +/* { dg-final { scan-tree-dump-times "basic block part vectorized using (?:32|64) byte vectors" 6 "slp2" } }*/
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*VEC_PERM_EXPR.*\{ 2, 3, 0, 1 \}} 4 "slp2" } }  */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*VEC_PERM_EXPR.*\{ 6, 7, 4, 5, 2, 3, 0, 1 \}} 4 "slp2" } }  */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*VEC_PERM_EXPR.*\{ 14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1 \}} 2 "slp2" } }  */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*VEC_PERM_EXPR.*\{ 30, 31, 28, 29, 26, 27, 24, 25, 22, 23, 20, 21, 18, 19, 16, 17, 14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1 \}} 2 "slp2" } }  */
> +
> +void
> +__attribute__((noipa))
> +foo_pd (_Complex double* a, _Complex double* __restrict b)
> +{
> +  a[0] = b[3];
> +  a[1] = b[2];
> +  a[2] = b[1];
> +  a[3] = b[0];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_ps (_Complex float* a, _Complex float* __restrict b)
> +{
> +  a[0] = b[7];
> +  a[1] = b[6];
> +  a[2] = b[5];
> +  a[3] = b[4];
> +  a[4] = b[3];
> +  a[5] = b[2];
> +  a[6] = b[1];
> +  a[7] = b[0];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi64 (_Complex long long* a, _Complex long long* __restrict b)
> +{
> +  a[0] = b[3];
> +  a[1] = b[2];
> +  a[2] = b[1];
> +  a[3] = b[0];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi32 (_Complex int* a, _Complex int* __restrict b)
> +{
> +  a[0] = b[7];
> +  a[1] = b[6];
> +  a[2] = b[5];
> +  a[3] = b[4];
> +  a[4] = b[3];
> +  a[5] = b[2];
> +  a[6] = b[1];
> +  a[7] = b[0];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi16 (_Complex short* a, _Complex short* __restrict b)
> +{
> +  a[0] = b[15];
> +  a[1] = b[14];
> +  a[2] = b[13];
> +  a[3] = b[12];
> +  a[4] = b[11];
> +  a[5] = b[10];
> +  a[6] = b[9];
> +  a[7] = b[8];
> +  a[8] = b[7];
> +  a[9] = b[6];
> +  a[10] = b[5];
> +  a[11] = b[4];
> +  a[12] = b[3];
> +  a[13] = b[2];
> +  a[14] = b[1];
> +  a[15] = b[0];
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi8 (_Complex char* a, _Complex char* __restrict b)
> +{
> +  a[0] = b[31];
> +  a[1] = b[30];
> +  a[2] = b[29];
> +  a[3] = b[28];
> +  a[4] = b[27];
> +  a[5] = b[26];
> +  a[6] = b[25];
> +  a[7] = b[24];
> +  a[8] = b[23];
> +  a[9] = b[22];
> +  a[10] = b[21];
> +  a[11] = b[20];
> +  a[12] = b[19];
> +  a[13] = b[18];
> +  a[14] = b[17];
> +  a[15] = b[16];
> +  a[16] = b[15];
> +  a[17] = b[14];
> +  a[18] = b[13];
> +  a[19] = b[12];
> +  a[20] = b[11];
> +  a[21] = b[10];
> +  a[22] = b[9];
> +  a[23] = b[8];
> +  a[24] = b[7];
> +  a[25] = b[6];
> +  a[26] = b[5];
> +  a[27] = b[4];
> +  a[28] = b[3];
> +  a[29] = b[2];
> +  a[30] = b[1];
> +  a[31] = b[0];
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-6b.c b/gcc/testsuite/gcc.target/i386/pr106010-6b.c
> new file mode 100644
> index 00000000000..1c5bb020939
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-6b.c
> @@ -0,0 +1,157 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx2 -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256" } */
> +/* { dg-require-effective-target avx2 } */
> +
> +#include "avx2-check.h"
> +#include <string.h>
> +#include "pr106010-6a.c"
> +
> +void
> +avx2_test (void)
> +{
> +  _Complex double* pd_src = (_Complex double*) malloc (64);
> +  _Complex double* pd_dst = (_Complex double*) malloc (64);
> +  _Complex double* pd_exp = (_Complex double*) malloc (64);
> +  _Complex float* ps_src = (_Complex float*) malloc (64);
> +  _Complex float* ps_dst = (_Complex float*) malloc (64);
> +  _Complex float* ps_exp = (_Complex float*) malloc (64);
> +  _Complex long long* epi64_src = (_Complex long long*) malloc (64);
> +  _Complex long long* epi64_dst = (_Complex long long*) malloc (64);
> +  _Complex long long* epi64_exp = (_Complex long long*) malloc (64);
> +  _Complex int* epi32_src = (_Complex int*) malloc (64);
> +  _Complex int* epi32_dst = (_Complex int*) malloc (64);
> +  _Complex int* epi32_exp = (_Complex int*) malloc (64);
> +  _Complex short* epi16_src = (_Complex short*) malloc (64);
> +  _Complex short* epi16_dst = (_Complex short*) malloc (64);
> +  _Complex short* epi16_exp = (_Complex short*) malloc (64);
> +  _Complex char* epi8_src = (_Complex char*) malloc (64);
> +  _Complex char* epi8_dst = (_Complex char*) malloc (64);
> +  _Complex char* epi8_exp = (_Complex char*) malloc (64);
> +  char* p = (char* ) malloc (64);
> +  char* q = (char* ) malloc (64);
> +
> +  __builtin_memset (pd_dst, 0, 64);
> +  __builtin_memset (ps_dst, 0, 64);
> +  __builtin_memset (epi64_dst, 0, 64);
> +  __builtin_memset (epi32_dst, 0, 64);
> +  __builtin_memset (epi16_dst, 0, 64);
> +  __builtin_memset (epi8_dst, 0, 64);
> +
> +  for (int i = 0; i != 64; i++)
> +    p[i] = i;
> +
> +  __builtin_memcpy (pd_src, p, 64);
> +  __builtin_memcpy (ps_src, p, 64);
> +  __builtin_memcpy (epi64_src, p, 64);
> +  __builtin_memcpy (epi32_src, p, 64);
> +  __builtin_memcpy (epi16_src, p, 64);
> +  __builtin_memcpy (epi8_src, p, 64);
> +
> +
> +  for (int i = 0; i != 16; i++)
> +    {
> +      q[i] = i + 48;
> +      q[i + 16] = i + 32;
> +      q[i + 32] = i + 16;
> +      q[i + 48] = i;
> +    }
> +
> +  __builtin_memcpy (pd_exp, q, 64);
> +  __builtin_memcpy (epi64_exp, q, 64);
> +
> +   for (int i = 0; i != 8; i++)
> +    {
> +      q[i] = i + 56;
> +      q[i + 8] = i + 48;
> +      q[i + 16] = i + 40;
> +      q[i + 24] = i + 32;
> +      q[i + 32] = i + 24;
> +      q[i + 40] = i + 16;
> +      q[i + 48] = i + 8;
> +      q[i + 56] = i;
> +    }
> +
> +  __builtin_memcpy (ps_exp, q, 64);
> +  __builtin_memcpy (epi32_exp, q, 64);
> +
> +  for (int i = 0; i != 4; i++)
> +    {
> +      q[i] = i + 60;
> +      q[i + 4] = i + 56;
> +      q[i + 8] = i + 52;
> +      q[i + 12] = i + 48;
> +      q[i + 16] = i + 44;
> +      q[i + 20] = i + 40;
> +      q[i + 24] = i + 36;
> +      q[i + 28] = i + 32;
> +      q[i + 32] = i + 28;
> +      q[i + 36] = i + 24;
> +      q[i + 40] = i + 20;
> +      q[i + 44] = i + 16;
> +      q[i + 48] = i + 12;
> +      q[i + 52] = i + 8;
> +      q[i + 56] = i + 4;
> +      q[i + 60] = i;
> +    }
> +
> +  __builtin_memcpy (epi16_exp, q, 64);
> +
> +  for (int i = 0; i != 2; i++)
> +    {
> +      q[i] = i + 62;
> +      q[i + 2] = i + 60;
> +      q[i + 4] = i + 58;
> +      q[i + 6] = i + 56;
> +      q[i + 8] = i + 54;
> +      q[i + 10] = i + 52;
> +      q[i + 12] = i + 50;
> +      q[i + 14] = i + 48;
> +      q[i + 16] = i + 46;
> +      q[i + 18] = i + 44;
> +      q[i + 20] = i + 42;
> +      q[i + 22] = i + 40;
> +      q[i + 24] = i + 38;
> +      q[i + 26] = i + 36;
> +      q[i + 28] = i + 34;
> +      q[i + 30] = i + 32;
> +      q[i + 32] = i + 30;
> +      q[i + 34] = i + 28;
> +      q[i + 36] = i + 26;
> +      q[i + 38] = i + 24;
> +      q[i + 40] = i + 22;
> +      q[i + 42] = i + 20;
> +      q[i + 44] = i + 18;
> +      q[i + 46] = i + 16;
> +      q[i + 48] = i + 14;
> +      q[i + 50] = i + 12;
> +      q[i + 52] = i + 10;
> +      q[i + 54] = i + 8;
> +      q[i + 56] = i + 6;
> +      q[i + 58] = i + 4;
> +      q[i + 60] = i + 2;
> +      q[i + 62] = i;
> +    }
> +  __builtin_memcpy (epi8_exp, q, 64);
> +
> +  foo_pd (pd_dst, pd_src);
> +  foo_ps (ps_dst, ps_src);
> +  foo_epi64 (epi64_dst, epi64_src);
> +  foo_epi32 (epi32_dst, epi32_src);
> +  foo_epi16 (epi16_dst, epi16_src);
> +  foo_epi8 (epi8_dst, epi8_src);
> +
> +  if (__builtin_memcmp (pd_dst, pd_exp, 64) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (ps_dst, ps_exp, 64) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi64_dst, epi64_exp, 64) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi32_dst, epi32_exp, 64) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi16_dst, epi16_exp, 64) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi8_dst, epi8_exp, 64) != 0)
> +    __builtin_abort ();
> +
> +  return;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-6c.c b/gcc/testsuite/gcc.target/i386/pr106010-6c.c
> new file mode 100644
> index 00000000000..b859d884a7f
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-6c.c
> @@ -0,0 +1,80 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx512fp16 -mavx512vl -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256 -fdump-tree-slp-details" } */
> +/* { dg-require-effective-target avx512fp16 } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*VEC_PERM_EXPR.*\{ 14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1 \}} 2 "slp2" } }  */
> +/* { dg-final { scan-tree-dump-times "basic block part vectorized using (?:32|64) byte vectors" 1 "slp2" } } */
> +
> +#include <string.h>
> +
> +static void do_test (void);
> +#define DO_TEST do_test
> +#define AVX512FP16
> +#include "avx512-check.h"
> +
> +void
> +__attribute__((noipa))
> +foo_ph (_Complex _Float16* a, _Complex _Float16* __restrict b)
> +{
> +  a[0] = b[15];
> +  a[1] = b[14];
> +  a[2] = b[13];
> +  a[3] = b[12];
> +  a[4] = b[11];
> +  a[5] = b[10];
> +  a[6] = b[9];
> +  a[7] = b[8];
> +  a[8] = b[7];
> +  a[9] = b[6];
> +  a[10] = b[5];
> +  a[11] = b[4];
> +  a[12] = b[3];
> +  a[13] = b[2];
> +  a[14] = b[1];
> +  a[15] = b[0];
> +}
> +
> +void
> +do_test (void)
> +{
> +  _Complex _Float16* ph_src = (_Complex _Float16*) malloc (64);
> +  _Complex _Float16* ph_dst = (_Complex _Float16*) malloc (64);
> +  _Complex _Float16* ph_exp = (_Complex _Float16*) malloc (64);
> +  char* p = (char* ) malloc (64);
> +  char* q = (char* ) malloc (64);
> +
> +  __builtin_memset (ph_dst, 0, 64);
> +
> +  for (int i = 0; i != 64; i++)
> +    p[i] = i;
> +
> +  __builtin_memcpy (ph_src, p, 64);
> +
> +  for (int i = 0; i != 4; i++)
> +    {
> +      q[i] = i + 60;
> +      q[i + 4] = i + 56;
> +      q[i + 8] = i + 52;
> +      q[i + 12] = i + 48;
> +      q[i + 16] = i + 44;
> +      q[i + 20] = i + 40;
> +      q[i + 24] = i + 36;
> +      q[i + 28] = i + 32;
> +      q[i + 32] = i + 28;
> +      q[i + 36] = i + 24;
> +      q[i + 40] = i + 20;
> +      q[i + 44] = i + 16;
> +      q[i + 48] = i + 12;
> +      q[i + 52] = i + 8;
> +      q[i + 56] = i + 4;
> +      q[i + 60] = i;
> +    }
> +
> +  __builtin_memcpy (ph_exp, q, 64);
> +
> +  foo_ph (ph_dst, ph_src);
> +
> +  if (__builtin_memcmp (ph_dst, ph_exp, 64) != 0)
> +    __builtin_abort ();
> +
> +  return;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-7a.c b/gcc/testsuite/gcc.target/i386/pr106010-7a.c
> new file mode 100644
> index 00000000000..2ea01fac927
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-7a.c
> @@ -0,0 +1,58 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx -ftree-vectorize -fvect-cost-model=unlimited -fdump-tree-vect-details -mprefer-vector-width=256" } */
> +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 6 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(4\) double>} 1 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(8\) float>} 1 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(4\) long long int>} 1 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(8\) int>} 1 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(16\) short int>} 1 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(32\) char>} 1 "vect" } } */
> +
> +#define N 10000
> +void
> +__attribute__((noipa))
> +foo_pd (_Complex double* a, _Complex double b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_ps (_Complex float* a, _Complex float b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi64 (_Complex long long* a, _Complex long long b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi32 (_Complex int* a, _Complex int b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi16 (_Complex short* a, _Complex short b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi8 (_Complex char* a, _Complex char b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-7b.c b/gcc/testsuite/gcc.target/i386/pr106010-7b.c
> new file mode 100644
> index 00000000000..26482cc10f5
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-7b.c
> @@ -0,0 +1,63 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256" } */
> +/* { dg-require-effective-target avx } */
> +
> +#include "avx-check.h"
> +#include <string.h>
> +#include "pr106010-7a.c"
> +
> +void
> +avx_test (void)
> +{
> +  _Complex double* pd_src = (_Complex double*) malloc (2 * N * sizeof (double));
> +  _Complex double* pd_dst = (_Complex double*) malloc (2 * N * sizeof (double));
> +  _Complex float* ps_src = (_Complex float*) malloc (2 * N * sizeof (float));
> +  _Complex float* ps_dst = (_Complex float*) malloc (2 * N * sizeof (float));
> +  _Complex long long* epi64_src = (_Complex long long*) malloc (2 * N * sizeof (long long));
> +  _Complex long long* epi64_dst = (_Complex long long*) malloc (2 * N * sizeof (long long));
> +  _Complex int* epi32_src = (_Complex int*) malloc (2 * N * sizeof (int));
> +  _Complex int* epi32_dst = (_Complex int*) malloc (2 * N * sizeof (int));
> +  _Complex short* epi16_src = (_Complex short*) malloc (2 * N * sizeof (short));
> +  _Complex short* epi16_dst = (_Complex short*) malloc (2 * N * sizeof (short));
> +  _Complex char* epi8_src = (_Complex char*) malloc (2 * N * sizeof (char));
> +  _Complex char* epi8_dst = (_Complex char*) malloc (2 * N * sizeof (char));
> +  char* p_init = (char*) malloc (2 * N * sizeof (double));
> +
> +  __builtin_memset (pd_dst, 0, 2 * N * sizeof (double));
> +  __builtin_memset (ps_dst, 0, 2 * N * sizeof (float));
> +  __builtin_memset (epi64_dst, 0, 2 * N * sizeof (long long));
> +  __builtin_memset (epi32_dst, 0, 2 * N * sizeof (int));
> +  __builtin_memset (epi16_dst, 0, 2 * N * sizeof (short));
> +  __builtin_memset (epi8_dst, 0, 2 * N * sizeof (char));
> +
> +  for (int i = 0; i != 2 * N * sizeof (double); i++)
> +    p_init[i] = i % 2 + 3;
> +
> +  memcpy (pd_src, p_init, 2 * N * sizeof (double));
> +  memcpy (ps_dst, p_init, 2 * N * sizeof (float));
> +  memcpy (epi64_dst, p_init, 2 * N * sizeof (long long));
> +  memcpy (epi32_dst, p_init, 2 * N * sizeof (int));
> +  memcpy (epi16_dst, p_init, 2 * N * sizeof (short));
> +  memcpy (epi8_dst, p_init, 2 * N * sizeof (char));
> +
> +  foo_pd (pd_dst, pd_src[0]);
> +  foo_ps (ps_dst, ps_src[0]);
> +  foo_epi64 (epi64_dst, epi64_src[0]);
> +  foo_epi32 (epi32_dst, epi32_src[0]);
> +  foo_epi16 (epi16_dst, epi16_src[0]);
> +  foo_epi8 (epi8_dst, epi8_src[0]);
> +  if (__builtin_memcmp (pd_dst, pd_src, N * 2 * sizeof (double)) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (ps_dst, ps_src, N * 2 * sizeof (float)) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi64_dst, epi64_src, N * 2 * sizeof (long long)) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi32_dst, epi32_src, N * 2 * sizeof (int)) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi16_dst, epi16_src, N * 2 * sizeof (short)) != 0)
> +    __builtin_abort ();
> +  if (__builtin_memcmp (epi8_dst, epi8_src, N * 2 * sizeof (char)) != 0)
> +    __builtin_abort ();
> +
> +  return;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-7c.c b/gcc/testsuite/gcc.target/i386/pr106010-7c.c
> new file mode 100644
> index 00000000000..7f4056a5ecc
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-7c.c
> @@ -0,0 +1,41 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx512fp16 -mavx512vl -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256 -fdump-tree-vect-details" } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(16\) _Float16>} 1 "vect" } } */
> +/* { dg-require-effective-target avx512fp16 } */
> +
> +#include <string.h>
> +
> +static void do_test (void);
> +
> +#define DO_TEST do_test
> +#define AVX512FP16
> +#include "avx512-check.h"
> +
> +#define N 10000
> +
> +void
> +__attribute__((noipa))
> +foo_ph (_Complex _Float16* a, _Complex _Float16 b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b;
> +}
> +
> +static void
> +do_test (void)
> +{
> +  _Complex _Float16* ph_src = (_Complex _Float16*) malloc (2 * N * sizeof (_Float16));
> +  _Complex _Float16* ph_dst = (_Complex _Float16*) malloc (2 * N * sizeof (_Float16));
> +  char* p_init = (char*) malloc (2 * N * sizeof (_Float16));
> +
> +  __builtin_memset (ph_dst, 0, 2 * N * sizeof (_Float16));
> +
> +  for (int i = 0; i != 2 * N * sizeof (_Float16); i++)
> +    p_init[i] = i % 2 + 3;
> +
> +  memcpy (ph_src, p_init, 2 * N * sizeof (_Float16));
> +
> +  foo_ph (ph_dst, ph_src[0]);
> +  if (__builtin_memcmp (ph_dst, ph_src, N * 2 * sizeof (_Float16)) != 0)
> +    __builtin_abort ();
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-8a.c b/gcc/testsuite/gcc.target/i386/pr106010-8a.c
> new file mode 100644
> index 00000000000..11054b60d30
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-8a.c
> @@ -0,0 +1,58 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx -ftree-vectorize -fvect-cost-model=unlimited -fdump-tree-vect-details -mprefer-vector-width=256" } */
> +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 6 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(4\) double>} 1 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(8\) float>} 1 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(4\) long long int>} 1 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(8\) int>} 1 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(16\) short int>} 1 "vect" } } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(32\) char>} 1 "vect" } } */
> +
> +#define N 10000
> +void
> +__attribute__((noipa))
> +foo_pd (_Complex double* a)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = 1.0 + 2.0i;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_ps (_Complex float* a)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = 1.0f + 2.0fi;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi64 (_Complex long long* a)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = 1 + 2i;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi32 (_Complex int* a)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = 1 + 2i;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi16 (_Complex short* a)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = 1 + 2i;
> +}
> +
> +void
> +__attribute__((noipa))
> +foo_epi8 (_Complex char* a)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = 1 + 2i;
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-8b.c b/gcc/testsuite/gcc.target/i386/pr106010-8b.c
> new file mode 100644
> index 00000000000..6bb0073b691
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-8b.c
> @@ -0,0 +1,53 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256" } */
> +/* { dg-require-effective-target avx } */
> +
> +#include "avx-check.h"
> +#include <string.h>
> +#include "pr106010-8a.c"
> +
> +void
> +avx_test (void)
> +{
> +  _Complex double pd_src = 1.0 + 2.0i;
> +  _Complex double* pd_dst = (_Complex double*) malloc (2 * N * sizeof (double));
> +  _Complex float ps_src = 1.0 + 2.0i;
> +  _Complex float* ps_dst = (_Complex float*) malloc (2 * N * sizeof (float));
> +  _Complex long long epi64_src = 1 + 2i;;
> +  _Complex long long* epi64_dst = (_Complex long long*) malloc (2 * N * sizeof (long long));
> +  _Complex int epi32_src = 1 + 2i;
> +  _Complex int* epi32_dst = (_Complex int*) malloc (2 * N * sizeof (int));
> +  _Complex short epi16_src = 1 + 2i;
> +  _Complex short* epi16_dst = (_Complex short*) malloc (2 * N * sizeof (short));
> +  _Complex char epi8_src = 1 + 2i;
> +  _Complex char* epi8_dst = (_Complex char*) malloc (2 * N * sizeof (char));
> +
> +  __builtin_memset (pd_dst, 0, 2 * N * sizeof (double));
> +  __builtin_memset (ps_dst, 0, 2 * N * sizeof (float));
> +  __builtin_memset (epi64_dst, 0, 2 * N * sizeof (long long));
> +  __builtin_memset (epi32_dst, 0, 2 * N * sizeof (int));
> +  __builtin_memset (epi16_dst, 0, 2 * N * sizeof (short));
> +  __builtin_memset (epi8_dst, 0, 2 * N * sizeof (char));
> +
> +  foo_pd (pd_dst);
> +  foo_ps (ps_dst);
> +  foo_epi64 (epi64_dst);
> +  foo_epi32 (epi32_dst);
> +  foo_epi16 (epi16_dst);
> +  foo_epi8 (epi8_dst);
> +  for (int i = 0 ; i != N; i++)
> +    {
> +      if (pd_dst[i] != pd_src)
> +       __builtin_abort ();
> +      if (ps_dst[i] != ps_src)
> +       __builtin_abort ();
> +      if (epi64_dst[i] != epi64_src)
> +       __builtin_abort ();
> +      if (epi32_dst[i] != epi32_src)
> +       __builtin_abort ();
> +      if (epi16_dst[i] != epi16_src)
> +       __builtin_abort ();
> +      if (epi8_dst[i] != epi8_src)
> +       __builtin_abort ();
> +    }
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-8c.c b/gcc/testsuite/gcc.target/i386/pr106010-8c.c
> new file mode 100644
> index 00000000000..61ae131829d
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-8c.c
> @@ -0,0 +1,38 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx512fp16 -mavx512vl -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256 -fdump-tree-vect-details" } */
> +/* { dg-final { scan-tree-dump-times {(?n)add new stmt:.*MEM <vector\(16\) _Float16>} 1 "vect" } } */
> +/* { dg-require-effective-target avx512fp16 } */
> +
> +#include <string.h>
> +
> +static void do_test (void);
> +
> +#define DO_TEST do_test
> +#define AVX512FP16
> +#include "avx512-check.h"
> +
> +#define N 10000
> +
> +void
> +__attribute__((noipa))
> +foo_ph (_Complex _Float16* a)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = 1.0f16 + 2.0f16i;
> +}
> +
> +static void
> +do_test (void)
> +{
> +  _Complex _Float16 ph_src = 1.0f16 + 2.0f16i;
> +  _Complex _Float16* ph_dst = (_Complex _Float16*) malloc (2 * N * sizeof (_Float16));
> +
> +  __builtin_memset (ph_dst, 0, 2 * N * sizeof (_Float16));
> +
> +  foo_ph (ph_dst);
> +  for (int i = 0; i != N; i++)
> +    {
> +      if (ph_dst[i] != ph_src)
> +       __builtin_abort ();
> +    }
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-9a.c b/gcc/testsuite/gcc.target/i386/pr106010-9a.c
> new file mode 100644
> index 00000000000..e922f7b5400
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-9a.c
> @@ -0,0 +1,89 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O3 -mavx2 -fvect-cost-model=unlimited -fdump-tree-vect-details" } */
> +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 6 "vect" } } */
> +
> +typedef struct { _Complex double c; double a1; double a2;}
> +  cdf;
> +typedef struct { _Complex double c; double a1; double a2; double a3; double a4;}
> +  cdf2;
> +typedef struct { _Complex double c1; _Complex double c2; double a1; double a2; double a3; double a4;}
> +  cdf3;
> +typedef struct { _Complex double c1; _Complex double c2; double a1; double a2;}
> +  cdf4;
> +
> +#define N 100
> +/* VMAT_ELEMENTWISE.  */
> +void
> +__attribute__((noipa))
> +foo (cdf* a, cdf* __restrict b)
> +{
> +   for (int i = 0; i < N; ++i)
> +    {
> +      a[i].c = b[i].c;
> +      a[i].a1 = b[i].a1;
> +      a[i].a2 = b[i].a2;
> +    }
> +}
> +
> +/* VMAT_CONTIGUOUS_PERMUTE.  */
> +void
> +__attribute__((noipa))
> +foo1 (cdf2* a, cdf2* __restrict b)
> +{
> +   for (int i = 0; i < N; ++i)
> +    {
> +      a[i].c = b[i].c;
> +      a[i].a1 = b[i].a1;
> +      a[i].a2 = b[i].a2;
> +      a[i].a3 = b[i].a3;
> +      a[i].a4 = b[i].a4;
> +    }
> +}
> +
> +/* VMAT_CONTIGUOUS.  */
> +void
> +__attribute__((noipa))
> +foo2 (cdf3* a, cdf3* __restrict b)
> +{
> +   for (int i = 0; i < N; ++i)
> +    {
> +      a[i].c1 = b[i].c1;
> +      a[i].c2 = b[i].c2;
> +      a[i].a1 = b[i].a1;
> +      a[i].a2 = b[i].a2;
> +      a[i].a3 = b[i].a3;
> +      a[i].a4 = b[i].a4;
> +    }
> +}
> +
> +/* VMAT_STRIDED_SLP.  */
> +void
> +__attribute__((noipa))
> +foo3 (cdf4* a, cdf4* __restrict b)
> +{
> +   for (int i = 0; i < N; ++i)
> +    {
> +      a[i].c1 = b[i].c1;
> +      a[i].c2 = b[i].c2;
> +      a[i].a1 = b[i].a1;
> +      a[i].a2 = b[i].a2;
> +    }
> +}
> +
> +/* VMAT_CONTIGUOUS_REVERSE.  */
> +void
> +__attribute__((noipa))
> +foo4 (_Complex double* a, _Complex double* __restrict b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[i] = b[N-i-1];
> +}
> +
> +/* VMAT_CONTIGUOUS_DOWN.  */
> +void
> +__attribute__((noipa))
> +foo5 (_Complex double* a, _Complex double* __restrict b)
> +{
> +  for (int i = 0; i != N; i++)
> +    a[N-i-1] = b[0];
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-9b.c b/gcc/testsuite/gcc.target/i386/pr106010-9b.c
> new file mode 100644
> index 00000000000..e220445e6e3
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-9b.c
> @@ -0,0 +1,90 @@
> +/* { dg-do run } */
> +/* { dg-options "-O3 -msse2 -fvect-cost-model=unlimited" } */
> +/* { dg-require-effective-target sse2 } */
> +
> +#include <string.h>
> +#include "sse2-check.h"
> +#include "pr106010-9a.c"
> +
> +static void
> +sse2_test (void)
> +{
> +  _Complex double* pd_src = (_Complex double*) malloc (N * sizeof (_Complex double));
> +  _Complex double* pd_dst = (_Complex double*) malloc (N * sizeof (_Complex double));
> +  _Complex double* pd_src2 = (_Complex double*) malloc (N * sizeof (_Complex double));
> +  _Complex double* pd_dst2 = (_Complex double*) malloc (N * sizeof (_Complex double));
> +  cdf* cdf_src = (cdf*) malloc (N * sizeof (cdf));
> +  cdf* cdf_dst = (cdf*) malloc (N * sizeof (cdf));
> +  cdf2* cdf2_src = (cdf2*) malloc (N * sizeof (cdf2));
> +  cdf2* cdf2_dst = (cdf2*) malloc (N * sizeof (cdf2));
> +  cdf3* cdf3_src = (cdf3*) malloc (N * sizeof (cdf3));
> +  cdf3* cdf3_dst = (cdf3*) malloc (N * sizeof (cdf3));
> +  cdf4* cdf4_src = (cdf4*) malloc (N * sizeof (cdf4));
> +  cdf4* cdf4_dst = (cdf4*) malloc (N * sizeof (cdf4));
> +
> +  char* p_init = (char*) malloc (N * sizeof (cdf3));
> +
> +  __builtin_memset (cdf_dst, 0, N * sizeof (cdf));
> +  __builtin_memset (cdf2_dst, 0, N * sizeof (cdf2));
> +  __builtin_memset (cdf3_dst, 0, N * sizeof (cdf3));
> +  __builtin_memset (cdf4_dst, 0, N * sizeof (cdf4));
> +  __builtin_memset (pd_dst, 0, N * sizeof (_Complex double));
> +  __builtin_memset (pd_dst2, 0, N * sizeof (_Complex double));
> +
> +  for (int i = 0; i != N * sizeof (cdf3); i++)
> +    p_init[i] = i;
> +
> +  memcpy (cdf_src, p_init, N * sizeof (cdf));
> +  memcpy (cdf2_src, p_init, N * sizeof (cdf2));
> +  memcpy (cdf3_src, p_init, N * sizeof (cdf3));
> +  memcpy (cdf4_src, p_init, N * sizeof (cdf4));
> +  memcpy (pd_src, p_init, N * sizeof (_Complex double));
> +  for (int i = 0; i != 2 * N * sizeof (double); i++)
> +    p_init[i] = i % 16;
> +  memcpy (pd_src2, p_init, N * sizeof (_Complex double));
> +
> +  foo (cdf_dst, cdf_src);
> +  foo1 (cdf2_dst, cdf2_src);
> +  foo2 (cdf3_dst, cdf3_src);
> +  foo3 (cdf4_dst, cdf4_src);
> +  foo4 (pd_dst, pd_src);
> +  foo5 (pd_dst2, pd_src2);
> +  for (int i = 0; i != N; i++)
> +    {
> +      p_init[(N - i - 1) * 16] = i * 16;
> +      p_init[(N - i - 1) * 16 + 1] = i * 16 + 1;
> +      p_init[(N - i - 1) * 16 + 2] = i * 16 + 2;
> +      p_init[(N - i - 1) * 16 + 3] = i * 16 + 3;
> +      p_init[(N - i - 1) * 16 + 4] = i * 16 + 4;
> +      p_init[(N - i - 1) * 16 + 5] = i * 16 + 5;
> +      p_init[(N - i - 1) * 16 + 6] = i * 16 + 6;
> +      p_init[(N - i - 1) * 16 + 7] = i * 16 + 7;
> +      p_init[(N - i - 1) * 16 + 8] = i * 16 + 8;
> +      p_init[(N - i - 1) * 16 + 9] = i * 16 + 9;
> +      p_init[(N - i - 1) * 16 + 10] = i * 16 + 10;
> +      p_init[(N - i - 1) * 16 + 11] = i * 16 + 11;
> +      p_init[(N - i - 1) * 16 + 12] = i * 16 + 12;
> +      p_init[(N - i - 1) * 16 + 13] = i * 16 + 13;
> +      p_init[(N - i - 1) * 16 + 14] = i * 16 + 14;
> +      p_init[(N - i - 1) * 16 + 15] = i * 16 + 15;
> +    }
> +  memcpy (pd_src, p_init, N * 16);
> +
> +  if (__builtin_memcmp (pd_dst, pd_src, N * 2 * sizeof (double)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (pd_dst2, pd_src2, N * 2 * sizeof (double)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (cdf_dst, cdf_src, N * sizeof (cdf)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (cdf2_dst, cdf2_src, N * sizeof (cdf2)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (cdf3_dst, cdf3_src, N * sizeof (cdf3)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (cdf4_dst, cdf4_src, N * sizeof (cdf4)) != 0)
> +    __builtin_abort ();
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-9c.c b/gcc/testsuite/gcc.target/i386/pr106010-9c.c
> new file mode 100644
> index 00000000000..ff51f6195b7
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-9c.c
> @@ -0,0 +1,90 @@
> +/* { dg-do run } */
> +/* { dg-options "-O3 -mavx2 -fvect-cost-model=unlimited" } */
> +/* { dg-require-effective-target avx2 } */
> +
> +#include <string.h>
> +#include "avx2-check.h"
> +#include "pr106010-9a.c"
> +
> +static void
> +avx2_test (void)
> +{
> +  _Complex double* pd_src = (_Complex double*) malloc (N * sizeof (_Complex double));
> +  _Complex double* pd_dst = (_Complex double*) malloc (N * sizeof (_Complex double));
> +  _Complex double* pd_src2 = (_Complex double*) malloc (N * sizeof (_Complex double));
> +  _Complex double* pd_dst2 = (_Complex double*) malloc (N * sizeof (_Complex double));
> +  cdf* cdf_src = (cdf*) malloc (N * sizeof (cdf));
> +  cdf* cdf_dst = (cdf*) malloc (N * sizeof (cdf));
> +  cdf2* cdf2_src = (cdf2*) malloc (N * sizeof (cdf2));
> +  cdf2* cdf2_dst = (cdf2*) malloc (N * sizeof (cdf2));
> +  cdf3* cdf3_src = (cdf3*) malloc (N * sizeof (cdf3));
> +  cdf3* cdf3_dst = (cdf3*) malloc (N * sizeof (cdf3));
> +  cdf4* cdf4_src = (cdf4*) malloc (N * sizeof (cdf4));
> +  cdf4* cdf4_dst = (cdf4*) malloc (N * sizeof (cdf4));
> +
> +  char* p_init = (char*) malloc (N * sizeof (cdf3));
> +
> +  __builtin_memset (cdf_dst, 0, N * sizeof (cdf));
> +  __builtin_memset (cdf2_dst, 0, N * sizeof (cdf2));
> +  __builtin_memset (cdf3_dst, 0, N * sizeof (cdf3));
> +  __builtin_memset (cdf4_dst, 0, N * sizeof (cdf4));
> +  __builtin_memset (pd_dst, 0, N * sizeof (_Complex double));
> +  __builtin_memset (pd_dst2, 0, N * sizeof (_Complex double));
> +
> +  for (int i = 0; i != N * sizeof (cdf3); i++)
> +    p_init[i] = i;
> +
> +  memcpy (cdf_src, p_init, N * sizeof (cdf));
> +  memcpy (cdf2_src, p_init, N * sizeof (cdf2));
> +  memcpy (cdf3_src, p_init, N * sizeof (cdf3));
> +  memcpy (cdf4_src, p_init, N * sizeof (cdf4));
> +  memcpy (pd_src, p_init, N * sizeof (_Complex double));
> +  for (int i = 0; i != 2 * N * sizeof (double); i++)
> +    p_init[i] = i % 16;
> +  memcpy (pd_src2, p_init, N * sizeof (_Complex double));
> +
> +  foo (cdf_dst, cdf_src);
> +  foo1 (cdf2_dst, cdf2_src);
> +  foo2 (cdf3_dst, cdf3_src);
> +  foo3 (cdf4_dst, cdf4_src);
> +  foo4 (pd_dst, pd_src);
> +  foo5 (pd_dst2, pd_src2);
> +  for (int i = 0; i != N; i++)
> +    {
> +      p_init[(N - i - 1) * 16] = i * 16;
> +      p_init[(N - i - 1) * 16 + 1] = i * 16 + 1;
> +      p_init[(N - i - 1) * 16 + 2] = i * 16 + 2;
> +      p_init[(N - i - 1) * 16 + 3] = i * 16 + 3;
> +      p_init[(N - i - 1) * 16 + 4] = i * 16 + 4;
> +      p_init[(N - i - 1) * 16 + 5] = i * 16 + 5;
> +      p_init[(N - i - 1) * 16 + 6] = i * 16 + 6;
> +      p_init[(N - i - 1) * 16 + 7] = i * 16 + 7;
> +      p_init[(N - i - 1) * 16 + 8] = i * 16 + 8;
> +      p_init[(N - i - 1) * 16 + 9] = i * 16 + 9;
> +      p_init[(N - i - 1) * 16 + 10] = i * 16 + 10;
> +      p_init[(N - i - 1) * 16 + 11] = i * 16 + 11;
> +      p_init[(N - i - 1) * 16 + 12] = i * 16 + 12;
> +      p_init[(N - i - 1) * 16 + 13] = i * 16 + 13;
> +      p_init[(N - i - 1) * 16 + 14] = i * 16 + 14;
> +      p_init[(N - i - 1) * 16 + 15] = i * 16 + 15;
> +    }
> +  memcpy (pd_src, p_init, N * 16);
> +
> +  if (__builtin_memcmp (pd_dst, pd_src, N * 2 * sizeof (double)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (pd_dst2, pd_src2, N * 2 * sizeof (double)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (cdf_dst, cdf_src, N * sizeof (cdf)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (cdf2_dst, cdf2_src, N * sizeof (cdf2)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (cdf3_dst, cdf3_src, N * sizeof (cdf3)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (cdf4_dst, cdf4_src, N * sizeof (cdf4)) != 0)
> +    __builtin_abort ();
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr106010-9d.c b/gcc/testsuite/gcc.target/i386/pr106010-9d.c
> new file mode 100644
> index 00000000000..d4d8f1dd722
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr106010-9d.c
> @@ -0,0 +1,92 @@
> +/* { dg-do run } */
> +/* { dg-options "-O3 -mavx512f -mavx512vl -fvect-cost-model=unlimited -mprefer-vector-width=512" } */
> +/* { dg-require-effective-target avx512f } */
> +
> +#include <string.h>
> +#include <stdlib.h>
> +#define AVX512F
> +#include "avx512-check.h"
> +#include "pr106010-9a.c"
> +
> +static void
> +test_512 (void)
> +{
> +  _Complex double* pd_src = (_Complex double*) malloc (N * sizeof (_Complex double));
> +  _Complex double* pd_dst = (_Complex double*) malloc (N * sizeof (_Complex double));
> +  _Complex double* pd_src2 = (_Complex double*) malloc (N * sizeof (_Complex double));
> +  _Complex double* pd_dst2 = (_Complex double*) malloc (N * sizeof (_Complex double));
> +  cdf* cdf_src = (cdf*) malloc (N * sizeof (cdf));
> +  cdf* cdf_dst = (cdf*) malloc (N * sizeof (cdf));
> +  cdf2* cdf2_src = (cdf2*) malloc (N * sizeof (cdf2));
> +  cdf2* cdf2_dst = (cdf2*) malloc (N * sizeof (cdf2));
> +  cdf3* cdf3_src = (cdf3*) malloc (N * sizeof (cdf3));
> +  cdf3* cdf3_dst = (cdf3*) malloc (N * sizeof (cdf3));
> +  cdf4* cdf4_src = (cdf4*) malloc (N * sizeof (cdf4));
> +  cdf4* cdf4_dst = (cdf4*) malloc (N * sizeof (cdf4));
> +
> +  char* p_init = (char*) malloc (N * sizeof (cdf3));
> +
> +  __builtin_memset (cdf_dst, 0, N * sizeof (cdf));
> +  __builtin_memset (cdf2_dst, 0, N * sizeof (cdf2));
> +  __builtin_memset (cdf3_dst, 0, N * sizeof (cdf3));
> +  __builtin_memset (cdf4_dst, 0, N * sizeof (cdf4));
> +  __builtin_memset (pd_dst, 0, N * sizeof (_Complex double));
> +  __builtin_memset (pd_dst2, 0, N * sizeof (_Complex double));
> +
> +  for (int i = 0; i != N * sizeof (cdf3); i++)
> +    p_init[i] = i;
> +
> +  memcpy (cdf_src, p_init, N * sizeof (cdf));
> +  memcpy (cdf2_src, p_init, N * sizeof (cdf2));
> +  memcpy (cdf3_src, p_init, N * sizeof (cdf3));
> +  memcpy (cdf4_src, p_init, N * sizeof (cdf4));
> +  memcpy (pd_src, p_init, N * sizeof (_Complex double));
> +  for (int i = 0; i != 2 * N * sizeof (double); i++)
> +    p_init[i] = i % 16;
> +  memcpy (pd_src2, p_init, N * sizeof (_Complex double));
> +
> +  foo (cdf_dst, cdf_src);
> +  foo1 (cdf2_dst, cdf2_src);
> +  foo2 (cdf3_dst, cdf3_src);
> +  foo3 (cdf4_dst, cdf4_src);
> +  foo4 (pd_dst, pd_src);
> +  foo5 (pd_dst2, pd_src2);
> +  for (int i = 0; i != N; i++)
> +    {
> +      p_init[(N - i - 1) * 16] = i * 16;
> +      p_init[(N - i - 1) * 16 + 1] = i * 16 + 1;
> +      p_init[(N - i - 1) * 16 + 2] = i * 16 + 2;
> +      p_init[(N - i - 1) * 16 + 3] = i * 16 + 3;
> +      p_init[(N - i - 1) * 16 + 4] = i * 16 + 4;
> +      p_init[(N - i - 1) * 16 + 5] = i * 16 + 5;
> +      p_init[(N - i - 1) * 16 + 6] = i * 16 + 6;
> +      p_init[(N - i - 1) * 16 + 7] = i * 16 + 7;
> +      p_init[(N - i - 1) * 16 + 8] = i * 16 + 8;
> +      p_init[(N - i - 1) * 16 + 9] = i * 16 + 9;
> +      p_init[(N - i - 1) * 16 + 10] = i * 16 + 10;
> +      p_init[(N - i - 1) * 16 + 11] = i * 16 + 11;
> +      p_init[(N - i - 1) * 16 + 12] = i * 16 + 12;
> +      p_init[(N - i - 1) * 16 + 13] = i * 16 + 13;
> +      p_init[(N - i - 1) * 16 + 14] = i * 16 + 14;
> +      p_init[(N - i - 1) * 16 + 15] = i * 16 + 15;
> +    }
> +  memcpy (pd_src, p_init, N * 16);
> +
> +  if (__builtin_memcmp (pd_dst, pd_src, N * 2 * sizeof (double)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (pd_dst2, pd_src2, N * 2 * sizeof (double)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (cdf_dst, cdf_src, N * sizeof (cdf)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (cdf2_dst, cdf2_src, N * sizeof (cdf2)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (cdf3_dst, cdf3_src, N * sizeof (cdf3)) != 0)
> +    __builtin_abort ();
> +
> +  if (__builtin_memcmp (cdf4_dst, cdf4_src, N * sizeof (cdf4)) != 0)
> +    __builtin_abort ();
> +}
> diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
> index d20a10a1524..19567bb338a 100644
> --- a/gcc/tree-vect-data-refs.cc
> +++ b/gcc/tree-vect-data-refs.cc
> @@ -1403,7 +1403,8 @@ vect_get_data_access_cost (vec_info *vinfo, dr_vec_info *dr_info,
>    if (PURE_SLP_STMT (stmt_info))
>      ncopies = 1;
>    else
> -    ncopies = vect_get_num_copies (loop_vinfo, STMT_VINFO_VECTYPE (stmt_info));
> +    ncopies = vect_get_num_copies (loop_vinfo, STMT_VINFO_VECTYPE (stmt_info),
> +                                  STMT_VINFO_COMPLEX_P (stmt_info));
>
>    if (DR_IS_READ (dr_info->dr))
>      vect_get_load_cost (vinfo, stmt_info, ncopies, alignment_support_scheme,
> @@ -4597,8 +4598,22 @@ vect_analyze_data_refs (vec_info *vinfo, poly_uint64 *min_vf, bool *fatal)
>
>        /* Set vectype for STMT.  */
>        scalar_type = TREE_TYPE (DR_REF (dr));
> -      tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type);
> -      if (!vectype)
> +      tree adjust_scalar_type = scalar_type;
> +      /* Support Complex type access. Note that the complex type of load/store
> +        does not support gather/scatter.  */
> +      if (TREE_CODE (scalar_type) == COMPLEX_TYPE
> +         && gatherscatter == SG_NONE)
> +       {
> +         adjust_scalar_type = TREE_TYPE (scalar_type);
> +         STMT_VINFO_COMPLEX_P (stmt_info) = true;
> +       }
> +      tree vectype = get_vectype_for_scalar_type (vinfo, adjust_scalar_type);
> +      unsigned HOST_WIDE_INT constant_nunits;
> +      if (!vectype
> +         /* For complex type, V1DI doesn't make sense.  */
> +         || (STMT_VINFO_COMPLEX_P (stmt_info)
> +             && (!TYPE_VECTOR_SUBPARTS (vectype).is_constant (&constant_nunits)
> +                 || constant_nunits == 1)))
>          {
>            if (dump_enabled_p ())
>              {
> @@ -4635,8 +4650,11 @@ vect_analyze_data_refs (vec_info *vinfo, poly_uint64 *min_vf, bool *fatal)
>         }
>
>        /* Adjust the minimal vectorization factor according to the
> -        vector type.  */
> +        vector type. Note for complex type, VF is half of
> +        TYPE_VECTOR_SUBPARTS.  */
>        vf = TYPE_VECTOR_SUBPARTS (vectype);
> +      if (STMT_VINFO_COMPLEX_P (stmt_info))
> +       vf = exact_div (vf, 2);
>        *min_vf = upper_bound (*min_vf, vf);
>
>        /* Leave the BB vectorizer to pick the vector type later, based on
> @@ -6140,21 +6158,55 @@ vect_permute_load_chain (vec_info *vinfo, vec<tree> dr_chain,
>        vec_perm_indices indices;
>        for (k = 0; k < 3; k++)
>         {
> -         for (i = 0; i < nelt; i++)
> -           if (3 * i + k < 2 * nelt)
> -             sel[i] = 3 * i + k;
> -           else
> -             sel[i] = 0;
> -         indices.new_vector (sel, 2, nelt);
> -         perm3_mask_low = vect_gen_perm_mask_checked (vectype, indices);
> +         if (STMT_VINFO_COMPLEX_P (stmt_info))
> +           {
> +             for (i = 0; i < nelt / 2; i++)
> +               if (6 * i + 2 * k + 1 < 2 * nelt)
> +                 {
> +                   sel[2 * i] = 6 * i + 2 * k;
> +                   sel[2 * i + 1] = 6 * i + 2 * k + 1;
> +                 }
> +               else
> +                 {
> +                   sel[2 * i] = 0;
> +                   sel[2 * i + 1] = 0;
> +                 }
>
> -         for (i = 0, j = 0; i < nelt; i++)
> -           if (3 * i + k < 2 * nelt)
> -             sel[i] = i;
> -           else
> -             sel[i] = nelt + ((nelt + k) % 3) + 3 * (j++);
> -         indices.new_vector (sel, 2, nelt);
> -         perm3_mask_high = vect_gen_perm_mask_checked (vectype, indices);
> +             indices.new_vector (sel, 2, nelt);
> +             perm3_mask_low = vect_gen_perm_mask_checked (vectype, indices);
> +
> +             for (i = 0, j = 0; i < nelt / 2; i++)
> +               if (6 * i + 2 * k + 1 < 2 * nelt)
> +                 {
> +                   sel[2 * i] = 2 * i;
> +                   sel[2 * i + 1] = 2 * i + 1;
> +                 }
> +               else
> +                 {
> +                   sel[2 * i] = nelt + ((nelt + 2 * k) % 6) + 6 * j;
> +                   sel[2 * i + 1] = nelt + ((nelt + 2 * k) % 6) + 6 * (j++) + 1;
> +                 }
> +             indices.new_vector (sel, 2, nelt);
> +             perm3_mask_high = vect_gen_perm_mask_checked (vectype, indices);
> +           }
> +         else
> +           {
> +             for (i = 0; i < nelt; i++)
> +               if (3 * i + k < 2 * nelt)
> +                 sel[i] = 3 * i + k;
> +               else
> +                 sel[i] = 0;
> +             indices.new_vector (sel, 2, nelt);
> +             perm3_mask_low = vect_gen_perm_mask_checked (vectype, indices);
> +
> +             for (i = 0, j = 0; i < nelt; i++)
> +               if (3 * i + k < 2 * nelt)
> +                 sel[i] = i;
> +               else
> +                 sel[i] = nelt + ((nelt + k) % 3) + 3 * (j++);
> +             indices.new_vector (sel, 2, nelt);
> +             perm3_mask_high = vect_gen_perm_mask_checked (vectype, indices);
> +           }
>
>           first_vect = dr_chain[0];
>           second_vect = dr_chain[1];
> @@ -6186,17 +6238,43 @@ vect_permute_load_chain (vec_info *vinfo, vec<tree> dr_chain,
>
>        /* The encoding has a single stepped pattern.  */
>        poly_uint64 nelt = TYPE_VECTOR_SUBPARTS (vectype);
> -      vec_perm_builder sel (nelt, 1, 3);
> -      sel.quick_grow (3);
> -      for (i = 0; i < 3; ++i)
> -       sel[i] = i * 2;
> -      vec_perm_indices indices (sel, 2, nelt);
> -      perm_mask_even = vect_gen_perm_mask_checked (vectype, indices);
> +      if (STMT_VINFO_COMPLEX_P (stmt_info))
> +       {
> +         vec_perm_builder sel;
> +         unsigned neltc = nelt.to_constant ();
> +         sel.new_vector (neltc, neltc, 1);
> +         sel.quick_grow (neltc);
> +         for (unsigned i = 0; i != neltc / 2; i++)
> +           {
> +             sel[2 * i] = i * 4;
> +             sel[2 * i + 1] = i * 4 + 1;
> +           }
> +         vec_perm_indices indices (sel, 2, nelt);
> +         perm_mask_even = vect_gen_perm_mask_checked (vectype, indices);
>
> -      for (i = 0; i < 3; ++i)
> -       sel[i] = i * 2 + 1;
> -      indices.new_vector (sel, 2, nelt);
> -      perm_mask_odd = vect_gen_perm_mask_checked (vectype, indices);
> +         for (unsigned i = 0; i != nelt.to_constant() / 2; i++)
> +           {
> +             sel[2 * i] = i * 4 + 2;
> +             sel[2 * i + 1] = i * 4 + 3;
> +           }
> +         indices.new_vector (sel, 2, nelt);
> +         perm_mask_odd = vect_gen_perm_mask_checked (vectype, indices);
> +       }
> +      else
> +       {
> +         vec_perm_builder sel (nelt, 1, 3);
> +         sel.quick_grow (3);
> +         for (i = 0; i < 3; ++i)
> +           sel[i] = i * 2;
> +
> +         vec_perm_indices indices (sel, 2, nelt);
> +         perm_mask_even = vect_gen_perm_mask_checked (vectype, indices);
> +
> +         for (i = 0; i < 3; ++i)
> +           sel[i] = i * 2 + 1;
> +         indices.new_vector (sel, 2, nelt);
> +         perm_mask_odd = vect_gen_perm_mask_checked (vectype, indices);
> +       }
>
>        for (i = 0; i < log_length; i++)
>         {
> diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> index 3a70c15b593..365fa738022 100644
> --- a/gcc/tree-vect-loop.cc
> +++ b/gcc/tree-vect-loop.cc
> @@ -200,7 +200,12 @@ vect_determine_vf_for_stmt_1 (vec_info *vinfo, stmt_vec_info stmt_info,
>      }
>
>    if (nunits_vectype)
> -    vect_update_max_nunits (vf, nunits_vectype);
> +    {
> +      poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (nunits_vectype);
> +      if (STMT_VINFO_COMPLEX_P (stmt_info))
> +       nunits = exact_div (nunits, 2);
> +      vect_update_max_nunits (vf, nunits);
> +    }
>
>    return opt_result::success ();
>  }
> diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
> index dab5daddcc5..5d66ea2f286 100644
> --- a/gcc/tree-vect-slp.cc
> +++ b/gcc/tree-vect-slp.cc
> @@ -877,10 +877,14 @@ vect_record_max_nunits (vec_info *vinfo, stmt_vec_info stmt_info,
>        return false;
>      }
>
> +  poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
> +  if (STMT_VINFO_COMPLEX_P (stmt_info))
> +    nunits = exact_div (nunits, 2);
> +
>    /* If populating the vector type requires unrolling then fail
>       before adjusting *max_nunits for basic-block vectorization.  */
>    if (is_a <bb_vec_info> (vinfo)
> -      && !multiple_p (group_size, TYPE_VECTOR_SUBPARTS (vectype)))
> +      && !multiple_p (group_size , nunits))
>      {
>        if (dump_enabled_p ())
>         dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> @@ -891,7 +895,7 @@ vect_record_max_nunits (vec_info *vinfo, stmt_vec_info stmt_info,
>      }
>
>    /* In case of multiple types we need to detect the smallest type.  */
> -  vect_update_max_nunits (max_nunits, vectype);
> +  vect_update_max_nunits (max_nunits, nunits);
>    return true;
>  }
>
> @@ -3720,22 +3724,54 @@ vect_optimize_slp (vec_info *vinfo)
>          vect_attempt_slp_rearrange_stmts did.  This allows us to be lazy
>          when permuting constants and invariants keeping the permute
>          bijective.  */
> -      auto_sbitmap load_index (SLP_TREE_LANES (node));
> -      bitmap_clear (load_index);
> -      for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
> -       bitmap_set_bit (load_index, SLP_TREE_LOAD_PERMUTATION (node)[j] - imin);
> -      unsigned j;
> -      for (j = 0; j < SLP_TREE_LANES (node); ++j)
> -       if (!bitmap_bit_p (load_index, j))
> -         break;
> -      if (j != SLP_TREE_LANES (node))
> -       continue;
> +      /* Permutation of Complex type.  */
> +      if (STMT_VINFO_COMPLEX_P (dr_stmt))
> +       {
> +         auto_sbitmap load_index (SLP_TREE_LANES (node) * 2);
> +         bitmap_clear (load_index);
> +         for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
> +           {
> +             unsigned bit = SLP_TREE_LOAD_PERMUTATION (node)[j] - imin;
> +             bitmap_set_bit (load_index, 2 * bit);
> +             bitmap_set_bit (load_index, 2 * bit + 1);
> +           }
> +         unsigned j;
> +         for (j = 0; j < SLP_TREE_LANES (node) * 2; ++j)
> +           if (!bitmap_bit_p (load_index, j))
> +             break;
> +         if (j != SLP_TREE_LANES (node) * 2)
> +           continue;
>
> -      vec<unsigned> perm = vNULL;
> -      perm.safe_grow (SLP_TREE_LANES (node), true);
> -      for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
> -       perm[j] = SLP_TREE_LOAD_PERMUTATION (node)[j] - imin;
> -      perms.safe_push (perm);
> +         vec<unsigned> perm = vNULL;
> +         perm.safe_grow (SLP_TREE_LANES (node) * 2, true);
> +         for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
> +           {
> +             unsigned cidx = SLP_TREE_LOAD_PERMUTATION (node)[j] - imin;
> +             perm[2 * j] = 2 * cidx;
> +             perm[2 * j + 1] = 2 * cidx + 1;
> +           }
> +         perms.safe_push (perm);
> +       }
> +      else
> +       {
> +         auto_sbitmap load_index (SLP_TREE_LANES (node));
> +         bitmap_clear (load_index);
> +         for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
> +           bitmap_set_bit (load_index,
> +                           SLP_TREE_LOAD_PERMUTATION (node)[j] - imin);
> +         unsigned j;
> +         for (j = 0; j < SLP_TREE_LANES (node); ++j)
> +           if (!bitmap_bit_p (load_index, j))
> +             break;
> +         if (j != SLP_TREE_LANES (node))
> +           continue;
> +
> +         vec<unsigned> perm = vNULL;
> +         perm.safe_grow (SLP_TREE_LANES (node), true);
> +         for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
> +           perm[j] = SLP_TREE_LOAD_PERMUTATION (node)[j] - imin;
> +         perms.safe_push (perm);
> +       }
>        vertices[idx].perm_in = perms.length () - 1;
>        vertices[idx].perm_out = perms.length () - 1;
>      }
> @@ -4518,6 +4554,12 @@ vect_slp_analyze_node_operations_1 (vec_info *vinfo, slp_tree node,
>         vf = loop_vinfo->vectorization_factor;
>        else
>         vf = 1;
> +      /* For complex type and SLP, double vf to get right vectype.
> +        .i.e vector(4) double for complex double, group size is 2, double vf
> +        to map vf * group_size to TYPE_VECTOR_SUBPARTS.  */
> +     if (STMT_VINFO_COMPLEX_P (stmt_info))
> +       vf *= 2;
> +
>        unsigned int group_size = SLP_TREE_LANES (node);
>        tree vectype = SLP_TREE_VECTYPE (node);
>        SLP_TREE_NUMBER_OF_VEC_STMTS (node)
> @@ -4763,10 +4805,17 @@ vect_slp_analyze_node_operations (vec_info *vinfo, slp_tree node,
>             }
>           unsigned group_size = SLP_TREE_LANES (child);
>           poly_uint64 vf = 1;
> +
>           if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
>             vf = loop_vinfo->vectorization_factor;
> +
> +         /* V2SF is just 1 complex type, so mutiply by 2
> +            to get release vector numbers.  */
> +         unsigned cp
> +           = STMT_VINFO_COMPLEX_P (SLP_TREE_REPRESENTATIVE (node)) ? 2 : 1;
> +
>           SLP_TREE_NUMBER_OF_VEC_STMTS (child)
> -           = vect_get_num_vectors (vf * group_size, vector_type);
> +           = vect_get_num_vectors (vf * group_size * cp, vector_type);
>           /* And cost them.  */
>           vect_prologue_cost_for_slp (child, cost_vec);
>         }
> @@ -6402,6 +6451,11 @@ vect_create_constant_vectors (vec_info *vinfo, slp_tree op_node)
>
>    /* We always want SLP_TREE_VECTYPE (op_node) here correctly set.  */
>    vector_type = SLP_TREE_VECTYPE (op_node);
> +  unsigned int cp = 1;
> +  /* Handle Complex type vector init.
> +     SLP_TREE_REPRESENTATIVE (op_node) could be NULL.  */
> +  if (TREE_CODE (TREE_TYPE (op_node->ops[0])) == COMPLEX_TYPE)
> +    cp = 2;
>
>    unsigned int number_of_vectors = SLP_TREE_NUMBER_OF_VEC_STMTS (op_node);
>    SLP_TREE_VEC_DEFS (op_node).create (number_of_vectors);
> @@ -6426,9 +6480,9 @@ vect_create_constant_vectors (vec_info *vinfo, slp_tree op_node)
>    /* When using duplicate_and_interleave, we just need one element for
>       each scalar statement.  */
>    if (!TYPE_VECTOR_SUBPARTS (vector_type).is_constant (&nunits))
> -    nunits = group_size;
> +    nunits = group_size * cp;
>
> -  number_of_copies = nunits * number_of_vectors / group_size;
> +  number_of_copies = nunits * number_of_vectors / (group_size * cp);
>
>    number_of_places_left_in_vector = nunits;
>    constant_p = true;
> @@ -6460,8 +6514,23 @@ vect_create_constant_vectors (vec_info *vinfo, slp_tree op_node)
>                         gcc_unreachable ();
>                     }
>                   else
> -                   op = fold_unary (VIEW_CONVERT_EXPR,
> -                                    TREE_TYPE (vector_type), op);
> +                   {
> +                     tree scalar_type = TREE_TYPE (vector_type);
> +                     /* For complex type, insert real and imag part
> +                        separately.  */
> +                     if (cp == 2)
> +                       {
> +                         gcc_assert ((TREE_CODE (TREE_TYPE (op))
> +                                      == COMPLEX_TYPE)
> +                                     && (scalar_type
> +                                         == TREE_TYPE (TREE_TYPE (op))));
> +                         elts[number_of_places_left_in_vector--]
> +                           = fold_unary (IMAGPART_EXPR, scalar_type, op);
> +                         op = fold_unary (REALPART_EXPR, scalar_type, op);
> +                       }
> +                     else
> +                       op = fold_unary (VIEW_CONVERT_EXPR, scalar_type, op);
> +                   }
>                   gcc_assert (op && CONSTANT_CLASS_P (op));
>                 }
>               else
> @@ -6481,11 +6550,28 @@ vect_create_constant_vectors (vec_info *vinfo, slp_tree op_node)
>                     }
>                   else
>                     {
> -                     op = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (vector_type),
> -                                  op);
> -                     init_stmt
> -                       = gimple_build_assign (new_temp, VIEW_CONVERT_EXPR,
> -                                              op);
> +                     tree scalar_type = TREE_TYPE (vector_type);
> +                     if (cp == 2)
> +                       {
> +                         gcc_assert ((TREE_CODE (TREE_TYPE (op))
> +                                      == COMPLEX_TYPE)
> +                                     && (scalar_type
> +                                         == TREE_TYPE (TREE_TYPE (op))));
> +                         tree imag = build1 (IMAGPART_EXPR, scalar_type, op);
> +                         op = build1 (REALPART_EXPR, scalar_type, op);
> +                         tree imag_temp = make_ssa_name (scalar_type);
> +                         elts[number_of_places_left_in_vector--] = imag_temp;
> +                         init_stmt = gimple_build_assign (imag_temp, imag);
> +                         gimple_seq_add_stmt (&ctor_seq, init_stmt);
> +                         init_stmt = gimple_build_assign (new_temp, op);
> +                       }
> +                     else
> +                       {
> +                         op = build1 (VIEW_CONVERT_EXPR, scalar_type, op);
> +                         init_stmt
> +                           = gimple_build_assign (new_temp, VIEW_CONVERT_EXPR,
> +                                                  op);
> +                       }
>                     }
>                   gimple_seq_add_stmt (&ctor_seq, init_stmt);
>                   op = new_temp;
> @@ -6696,15 +6782,17 @@ vect_transform_slp_perm_load (vec_info *vinfo,
>    unsigned int nelts_to_build;
>    unsigned int nvectors_per_build;
>    unsigned int in_nlanes;
> +  unsigned int cp = STMT_VINFO_COMPLEX_P (stmt_info) ? 2 : 1;
>    bool repeating_p = (group_size == DR_GROUP_SIZE (stmt_info)
> -                     && multiple_p (nunits, group_size));
> +                     && multiple_p (nunits, group_size * cp));
>    if (repeating_p)
>      {
>        /* A single vector contains a whole number of copies of the node, so:
>          (a) all permutes can use the same mask; and
>          (b) the permutes only need a single vector input.  */
> -      mask.new_vector (nunits, group_size, 3);
> -      nelts_to_build = mask.encoded_nelts ();
> +      /* For complex type, mask size should be double of nelts_to_build.  */
> +      mask.new_vector (nunits, group_size * cp, 3);
> +      nelts_to_build = mask.encoded_nelts () / cp;
>        nvectors_per_build = SLP_TREE_VEC_STMTS (node).length ();
>        in_nlanes = DR_GROUP_SIZE (stmt_info) * 3;
>      }
> @@ -6744,8 +6832,8 @@ vect_transform_slp_perm_load (vec_info *vinfo,
>         {
>           /* Enforced before the loop when !repeating_p.  */
>           unsigned int const_nunits = nunits.to_constant ();
> -         vec_index = i / const_nunits;
> -         mask_element = i % const_nunits;
> +         vec_index = i / (const_nunits / cp);
> +         mask_element = i % (const_nunits / cp);
>           if (vec_index == first_vec_index
>               || first_vec_index == -1)
>             {
> @@ -6755,7 +6843,7 @@ vect_transform_slp_perm_load (vec_info *vinfo,
>                    || second_vec_index == -1)
>             {
>               second_vec_index = vec_index;
> -             mask_element += const_nunits;
> +             mask_element += (const_nunits / cp);
>             }
>           else
>             {
> @@ -6768,14 +6856,24 @@ vect_transform_slp_perm_load (vec_info *vinfo,
>               return false;
>             }
>
> -         gcc_assert (mask_element < 2 * const_nunits);
> +         gcc_assert (mask_element < 2 * const_nunits / cp);
>         }
>
>        if (mask_element != index)
>         noop_p = false;
> -      mask[index++] = mask_element;
> +      /* Set index for Complex _type.
> +        i.e. mask like [1,0] is actually [2, 3, 0, 1]
> +        for vector scalar type.  */
> +      if (cp == 2)
> +       {
> +         mask[2 * index] = 2 * mask_element;
> +         mask[2 * index + 1] = 2 * mask_element + 1;
> +       }
> +      else
> +       mask[index] = mask_element;
> +      index++;
>
> -      if (index == count && !noop_p)
> +      if (index * cp == count && !noop_p)
>         {
>           indices.new_vector (mask, second_vec_index == -1 ? 1 : 2, nunits);
>           if (!can_vec_perm_const_p (mode, mode, indices))
> @@ -6799,7 +6897,7 @@ vect_transform_slp_perm_load (vec_info *vinfo,
>           ++*n_perms;
>         }
>
> -      if (index == count)
> +      if (index * cp == count)
>         {
>           if (!analyze_only)
>             {
> @@ -6869,7 +6967,7 @@ vect_transform_slp_perm_load (vec_info *vinfo,
>           bool load_seen = false;
>           for (unsigned i = 0; i < in_nlanes; ++i)
>             {
> -             if (i % const_nunits == 0)
> +             if (i % (const_nunits * cp) == 0)
>                 {
>                   if (load_seen)
>                     *n_loads += 1;
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 72107afc883..d6223c28f1c 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -1397,25 +1397,70 @@ vect_init_vector (vec_info *vinfo, stmt_vec_info stmt_info, tree val, tree type,
>  {
>    gimple *init_stmt;
>    tree new_temp;
> +  tree scalar_type = TREE_TYPE (type);
> +  gimple_seq stmts = NULL;
> +
> +  if (TREE_CODE (TREE_TYPE (val)) == COMPLEX_TYPE)
> +    {
> +      unsigned HOST_WIDE_INT nunits;
> +      gcc_assert (TYPE_VECTOR_SUBPARTS (type).is_constant (&nunits));
> +
> +      tree_vector_builder elts (type, nunits, 1);
> +      tree imag, real;
> +      if (TREE_CODE (val) == COMPLEX_CST)
> +       {
> +         real = fold_unary (REALPART_EXPR, scalar_type, val);
> +         imag = fold_unary (IMAGPART_EXPR, scalar_type, val);
> +       }
> +      else
> +       {
> +         real = make_ssa_name (scalar_type);
> +         imag = make_ssa_name (scalar_type);
> +         init_stmt
> +           = gimple_build_assign (real,
> +                                  build1 (REALPART_EXPR, scalar_type, val));
> +         gimple_seq_add_stmt (&stmts, init_stmt);
> +         init_stmt
> +           = gimple_build_assign (imag,
> +                                  build1 (IMAGPART_EXPR, scalar_type, val));
> +         gimple_seq_add_stmt (&stmts, init_stmt);
> +       }
>
> +      /* Build vector as [real,imag,real,imag,...].  */
> +      for (unsigned i = 0; i != nunits; i++)
> +       {
> +         if (i % 2)
> +           elts.quick_push (imag);
> +         else
> +           elts.quick_push (real);
> +       }
> +      val = gimple_build_vector (&stmts, &elts);
> +      if (!gimple_seq_empty_p (stmts))
> +       {
> +         if (gsi)
> +           gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
> +         else
> +           vinfo->insert_seq_on_entry (stmt_info, stmts);
> +       }
> +    }
>    /* We abuse this function to push sth to a SSA name with initial 'val'.  */
> -  if (! useless_type_conversion_p (type, TREE_TYPE (val)))
> +  else if (! useless_type_conversion_p (type, TREE_TYPE (val)))
>      {
>        gcc_assert (TREE_CODE (type) == VECTOR_TYPE);
> -      if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val)))
> +      if (! types_compatible_p (scalar_type, TREE_TYPE (val)))
>         {
>           /* Scalar boolean value should be transformed into
>              all zeros or all ones value before building a vector.  */
>           if (VECTOR_BOOLEAN_TYPE_P (type))
>             {
> -             tree true_val = build_all_ones_cst (TREE_TYPE (type));
> -             tree false_val = build_zero_cst (TREE_TYPE (type));
> +             tree true_val = build_all_ones_cst (scalar_type);
> +             tree false_val = build_zero_cst (scalar_type);
>
>               if (CONSTANT_CLASS_P (val))
>                 val = integer_zerop (val) ? false_val : true_val;
>               else
>                 {
> -                 new_temp = make_ssa_name (TREE_TYPE (type));
> +                 new_temp = make_ssa_name (scalar_type);
>                   init_stmt = gimple_build_assign (new_temp, COND_EXPR,
>                                                    val, true_val, false_val);
>                   vect_init_vector_1 (vinfo, stmt_info, init_stmt, gsi);
> @@ -1424,14 +1469,13 @@ vect_init_vector (vec_info *vinfo, stmt_vec_info stmt_info, tree val, tree type,
>             }
>           else
>             {
> -             gimple_seq stmts = NULL;
>               if (! INTEGRAL_TYPE_P (TREE_TYPE (val)))
>                 val = gimple_build (&stmts, VIEW_CONVERT_EXPR,
> -                                   TREE_TYPE (type), val);
> +                                   scalar_type, val);
>               else
>                 /* ???  Condition vectorization expects us to do
>                    promotion of invariant/external defs.  */
> -               val = gimple_convert (&stmts, TREE_TYPE (type), val);
> +               val = gimple_convert (&stmts, scalar_type, val);
>               for (gimple_stmt_iterator gsi2 = gsi_start (stmts);
>                    !gsi_end_p (gsi2); )
>                 {
> @@ -1496,7 +1540,12 @@ vect_get_vec_defs_for_operand (vec_info *vinfo, stmt_vec_info stmt_vinfo,
>                && VECTOR_BOOLEAN_TYPE_P (stmt_vectype))
>         vector_type = truth_type_for (stmt_vectype);
>        else
> -       vector_type = get_vectype_for_scalar_type (loop_vinfo, TREE_TYPE (op));
> +       {
> +         tree scalar_type = TREE_TYPE (op);
> +         if (STMT_VINFO_COMPLEX_P (stmt_vinfo))
> +           scalar_type = TREE_TYPE (scalar_type);
> +         vector_type = get_vectype_for_scalar_type (loop_vinfo, scalar_type);
> +       }
>
>        gcc_assert (vector_type);
>        tree vop = vect_init_vector (vinfo, stmt_vinfo, op, vector_type, NULL);
> @@ -1892,6 +1941,13 @@ vect_truncate_gather_scatter_offset (stmt_vec_info stmt_info,
>        return false;
>      }
>
> +  if (STMT_VINFO_COMPLEX_P (stmt_info))
> +    {
> +      if (dump_enabled_p ())
> +       dump_printf_loc (MSG_NOTE, vect_location,
> +                        "Complex type doens't support gather_scatter.\n");
> +      return false;
> +    }
>    /* Get the number of bits in an element.  */
>    tree vectype = STMT_VINFO_VECTYPE (stmt_info);
>    scalar_mode element_mode = SCALAR_TYPE_MODE (TREE_TYPE (vectype));
> @@ -2022,6 +2078,30 @@ perm_mask_for_reverse (tree vectype)
>    return vect_gen_perm_mask_checked (vectype, indices);
>  }
>
> +static tree
> +perm_mask_for_reverse (tree vectype, bool complex_p)
> +{
> +  if (!complex_p)
> +    return perm_mask_for_reverse (vectype);
> +
> +  unsigned HOST_WIDE_INT nunits;
> +  gcc_assert (TYPE_VECTOR_SUBPARTS (vectype).is_constant (&nunits));
> +
> +  /* The encoding has a single stepped pattern.  */
> +  vec_perm_builder sel (nunits, nunits, 1);
> +  for (unsigned i = 0; i < nunits; i+=2)
> +    {
> +      sel.quick_push (nunits - 2 - i);
> +      sel.quick_push (nunits - 1 - i);
> +    }
> +
> +  vec_perm_indices indices (sel, 1, nunits);
> +  if (!can_vec_perm_const_p (TYPE_MODE (vectype), TYPE_MODE (vectype),
> +                            indices))
> +    return NULL_TREE;
> +  return vect_gen_perm_mask_checked (vectype, indices);
> +}
> +
>  /* A subroutine of get_load_store_type, with a subset of the same
>     arguments.  Handle the case where STMT_INFO is a load or store that
>     accesses consecutive elements with a negative step.  Sets *POFFSET
> @@ -2045,8 +2125,12 @@ get_negative_load_store_type (vec_info *vinfo,
>      }
>
>    /* For backward running DRs the first access in vectype actually is
> -     N-1 elements before the address of the DR.  */
> -  *poffset = ((-TYPE_VECTOR_SUBPARTS (vectype) + 1)
> +     N-1 elements before the address of the DR.
> +     for Complex type, it's N - 2.  */
> +  unsigned cp = 1;
> +  if (STMT_VINFO_COMPLEX_P (stmt_info))
> +    cp = 2;
> +  *poffset = ((-TYPE_VECTOR_SUBPARTS (vectype) + cp)
>               * TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (vectype))));
>
>    int misalignment = dr_misalignment (dr_info, vectype, *poffset);
> @@ -2071,7 +2155,7 @@ get_negative_load_store_type (vec_info *vinfo,
>        return VMAT_CONTIGUOUS_DOWN;
>      }
>
> -  if (!perm_mask_for_reverse (vectype))
> +  if (!perm_mask_for_reverse (vectype, STMT_VINFO_COMPLEX_P (stmt_info)))
>      {
>        if (dump_enabled_p ())
>         dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> @@ -2188,6 +2272,8 @@ get_group_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
>                            && !DR_GROUP_NEXT_ELEMENT (stmt_info));
>    unsigned HOST_WIDE_INT gap = DR_GROUP_GAP (first_stmt_info);
>    poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
> +  if (STMT_VINFO_COMPLEX_P (stmt_info))
> +    nunits = exact_div (nunits, 2);
>
>    /* True if the vectorized statements would access beyond the last
>       statement in the group.  */
> @@ -2352,7 +2438,11 @@ get_group_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
>         {
>           /* First cope with the degenerate case of a single-element
>              vector.  */
> -         if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U))
> +         poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
> +         if (STMT_VINFO_COMPLEX_P (stmt_info))
> +           nunits = exact_div (nunits, 2);
> +
> +         if (known_eq (nunits, 1U))
>             ;
>
>           /* Otherwise try using LOAD/STORE_LANES.  */
> @@ -2361,6 +2451,8 @@ get_group_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
>                    : vect_store_lanes_supported (vectype, group_size,
>                                                  masked_p))
>             {
> +             if (STMT_VINFO_COMPLEX_P (stmt_info))
> +               return false;
>               *memory_access_type = VMAT_LOAD_STORE_LANES;
>               overrun_p = would_overrun_p;
>             }
> @@ -2620,6 +2712,14 @@ vect_check_scalar_mask (vec_info *vinfo, stmt_vec_info stmt_info,
>        return false;
>      }
>
> +  if (STMT_VINFO_COMPLEX_P (stmt_info))
> +    {
> +      if (dump_enabled_p ())
> +       dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> +                        "Complex type doesn't support mask argument.\n");
> +      return false;
> +    }
> +
>    if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (*mask)))
>      {
>        if (dump_enabled_p ())
> @@ -7509,8 +7609,17 @@ vectorizable_store (vec_info *vinfo,
>       same location twice.  */
>    gcc_assert (slp == PURE_SLP_STMT (stmt_info));
>
> +  if (!STMT_VINFO_DATA_REF (stmt_info))
> +    return false;
> +
>    tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE;
>    poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
> +  if (STMT_VINFO_COMPLEX_P (stmt_info))
> +    {
> +      if (!nunits.is_constant ())
> +       return false;
> +      nunits = exact_div (nunits, 2);
> +    }
>
>    if (loop_vinfo)
>      {
> @@ -7526,7 +7635,8 @@ vectorizable_store (vec_info *vinfo,
>    if (slp)
>      ncopies = 1;
>    else
> -    ncopies = vect_get_num_copies (loop_vinfo, vectype);
> +    ncopies = vect_get_num_copies (loop_vinfo, vectype,
> +                                  STMT_VINFO_COMPLEX_P (stmt_info));
>
>    gcc_assert (ncopies >= 1);
>
> @@ -7544,11 +7654,10 @@ vectorizable_store (vec_info *vinfo,
>      return false;
>
>    elem_type = TREE_TYPE (vectype);
> +  if (STMT_VINFO_COMPLEX_P (stmt_info))
> +    elem_type = build_complex_type (elem_type);
>    vec_mode = TYPE_MODE (vectype);
>
> -  if (!STMT_VINFO_DATA_REF (stmt_info))
> -    return false;
> -
>    vect_memory_access_type memory_access_type;
>    enum dr_alignment_support alignment_support_scheme;
>    int misalignment;
> @@ -7951,21 +8060,31 @@ vectorizable_store (vec_info *vinfo,
>        tree lvectype = vectype;
>        if (slp)
>         {
> +         scalar_mode elmode;
>           if (group_size < const_nunits
>               && const_nunits % group_size == 0)
>             {
>               nstores = const_nunits / group_size;
> -             lnel = group_size;
> -             ltype = build_vector_type (elem_type, group_size);
> +             if (STMT_VINFO_COMPLEX_P (stmt_info))
> +               {
> +                 lnel = group_size * 2;
> +                 ltype = build_vector_type (TREE_TYPE (elem_type), group_size * 2);
> +                 elmode = SCALAR_TYPE_MODE (TREE_TYPE (elem_type));
> +               }
> +             else
> +               {
> +                 ltype = build_vector_type (elem_type, group_size);
> +                 lnel = group_size;
> +                 elmode = SCALAR_TYPE_MODE (elem_type);
> +               }
>               lvectype = vectype;
>
>               /* First check if vec_extract optab doesn't support extraction
>                  of vector elts directly.  */
> -             scalar_mode elmode = SCALAR_TYPE_MODE (elem_type);
>               machine_mode vmode;
>               if (!VECTOR_MODE_P (TYPE_MODE (vectype))
>                   || !related_vector_mode (TYPE_MODE (vectype), elmode,
> -                                          group_size).exists (&vmode)
> +                                          lnel).exists (&vmode)
>                   || (convert_optab_handler (vec_extract_optab,
>                                              TYPE_MODE (vectype), vmode)
>                       == CODE_FOR_nothing))
> @@ -8051,6 +8170,8 @@ vectorizable_store (vec_info *vinfo,
>           unsigned int group_el = 0;
>           unsigned HOST_WIDE_INT
>             elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype)));
> +         if (STMT_VINFO_COMPLEX_P (stmt_info))
> +           elsz *= 2;
>           for (j = 0; j < ncopies; j++)
>             {
>               vec_oprnd = vec_oprnds[j];
> @@ -8448,7 +8569,9 @@ vectorizable_store (vec_info *vinfo,
>
>               if (memory_access_type == VMAT_CONTIGUOUS_REVERSE)
>                 {
> -                 tree perm_mask = perm_mask_for_reverse (vectype);
> +                 tree perm_mask
> +                   = perm_mask_for_reverse (vectype,
> +                                            STMT_VINFO_COMPLEX_P (stmt_info));
>                   tree perm_dest = vect_create_destination_var
>                     (vect_get_store_rhs (stmt_info), vectype);
>                   tree new_temp = make_ssa_name (perm_dest);
> @@ -8778,6 +8901,12 @@ vectorizable_load (vec_info *vinfo,
>
>    tree vectype = STMT_VINFO_VECTYPE (stmt_info);
>    poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
> +  if (STMT_VINFO_COMPLEX_P (stmt_info))
> +    {
> +      if (!nunits.is_constant ())
> +       return false;
> +      nunits = exact_div (nunits, 2);
> +    }
>
>    if (loop_vinfo)
>      {
> @@ -8794,7 +8923,8 @@ vectorizable_load (vec_info *vinfo,
>    if (slp)
>      ncopies = 1;
>    else
> -    ncopies = vect_get_num_copies (loop_vinfo, vectype);
> +    ncopies = vect_get_num_copies (loop_vinfo, vectype,
> +                                  STMT_VINFO_COMPLEX_P (stmt_info));
>
>    gcc_assert (ncopies >= 1);
>
> @@ -8822,6 +8952,8 @@ vectorizable_load (vec_info *vinfo,
>      }
>
>    elem_type = TREE_TYPE (vectype);
> +  if (STMT_VINFO_COMPLEX_P (stmt_info))
> +    elem_type = build_complex_type (elem_type);
>    mode = TYPE_MODE (vectype);
>
>    /* FORNOW. In some cases can vectorize even if data-type not supported
> @@ -8870,8 +9002,11 @@ vectorizable_load (vec_info *vinfo,
>                 if (k > maxk)
>                   maxk = k;
>               tree vectype = SLP_TREE_VECTYPE (slp_node);
> +             /* For complex type, half the nunits.  */
>               if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant (&nunits)
> -                 || maxk >= (DR_GROUP_SIZE (group_info) & ~(nunits - 1)))
> +                 || maxk >= (DR_GROUP_SIZE (group_info)
> +                             & ~((STMT_VINFO_COMPLEX_P (group_info)
> +                                  ? nunits >> 1 : nunits) - 1)))
>                 {
>                   if (dump_enabled_p ())
>                     dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> @@ -9098,9 +9233,10 @@ vectorizable_load (vec_info *vinfo,
>         }
>        else
>         {
> +         unsigned cp = STMT_VINFO_COMPLEX_P (stmt_info) ? 2 : 1;
>           if (grouped_load)
>             cst_offset
> -             = (tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype)))
> +             = (tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))) * cp
>                  * vect_get_place_in_interleaving_chain (stmt_info,
>                                                          first_stmt_info));
>           group_size = 1;
> @@ -9150,6 +9286,8 @@ vectorizable_load (vec_info *vinfo,
>        int nloads = const_nunits;
>        int lnel = 1;
>        tree ltype = TREE_TYPE (vectype);
> +      if (STMT_VINFO_COMPLEX_P (stmt_info))
> +       ltype = build_complex_type (ltype);
>        tree lvectype = vectype;
>        auto_vec<tree> dr_chain;
>        if (memory_access_type == VMAT_STRIDED_SLP)
> @@ -10080,7 +10218,9 @@ vectorizable_load (vec_info *vinfo,
>
>               if (memory_access_type == VMAT_CONTIGUOUS_REVERSE)
>                 {
> -                 tree perm_mask = perm_mask_for_reverse (vectype);
> +                 tree perm_mask
> +                   = perm_mask_for_reverse (vectype,
> +                                            STMT_VINFO_COMPLEX_P (stmt_info));
>                   new_temp = permute_vec_elements (vinfo, new_temp, new_temp,
>                                                    perm_mask, stmt_info, gsi);
>                   new_stmt = SSA_NAME_DEF_STMT (new_temp);
> @@ -12499,12 +12639,27 @@ vect_get_vector_types_for_stmt (vec_info *vinfo, stmt_vec_info stmt_info,
>             dump_printf_loc (MSG_NOTE, vect_location,
>                              "get vectype for scalar type: %T\n", scalar_type);
>         }
> +
> +      tree orig_scalar_type = scalar_type;
> +      if (TREE_CODE (scalar_type) == COMPLEX_TYPE)
> +       {
> +         /* Set complex_p for BB vectorizer.  */
> +         STMT_VINFO_COMPLEX_P (stmt_info) = true;
> +         scalar_type = TREE_TYPE (scalar_type);
> +         /* Double group_size for BB vectorizer to make
> +            following 2 get_vectype_for_scalar_type return wanted vectype.
> +            Real group size is not changed, just make the "faked" input
> +            group_size.  */
> +         group_size *= 2;
> +       }
>        vectype = get_vectype_for_scalar_type (vinfo, scalar_type, group_size);
> -      if (!vectype)
> +      if (!vectype
> +         || (STMT_VINFO_COMPLEX_P (stmt_info)
> +             && !TYPE_VECTOR_SUBPARTS (vectype).is_constant ()))
>         return opt_result::failure_at (stmt,
>                                        "not vectorized:"
>                                        " unsupported data-type %T\n",
> -                                      scalar_type);
> +                                      orig_scalar_type);
>
>        if (dump_enabled_p ())
>         dump_printf_loc (MSG_NOTE, vect_location, "vectype: %T\n", vectype);
> @@ -12529,16 +12684,30 @@ vect_get_vector_types_for_stmt (vec_info *vinfo, stmt_vec_info stmt_info,
>                                                    TREE_TYPE (vectype));
>        if (scalar_type != TREE_TYPE (vectype))
>         {
> -         if (dump_enabled_p ())
> +         tree orig_scalar_type = scalar_type;
> +         if (TREE_CODE (scalar_type) == COMPLEX_TYPE)
> +           {
> +             /* Set complex_p for Loop vectorizer.  */
> +             STMT_VINFO_COMPLEX_P (stmt_info) = true;
> +             scalar_type = TREE_TYPE (scalar_type);
> +             if (dump_enabled_p ())
> +               dump_printf_loc (MSG_NOTE, vect_location,
> +                            "get complex for smallest scalar type: %T\n",
> +                            scalar_type);
> +
> +           }
> +         else if (dump_enabled_p ())
>             dump_printf_loc (MSG_NOTE, vect_location,
>                              "get vectype for smallest scalar type: %T\n",
>                              scalar_type);
>           nunits_vectype = get_vectype_for_scalar_type (vinfo, scalar_type,
>                                                         group_size);
> -         if (!nunits_vectype)
> +         if (!nunits_vectype
> +             || (STMT_VINFO_COMPLEX_P (stmt_info)
> +                 && !TYPE_VECTOR_SUBPARTS (nunits_vectype).is_constant ()))
>             return opt_result::failure_at
>               (stmt, "not vectorized: unsupported data-type %T\n",
> -              scalar_type);
> +              orig_scalar_type);
>           if (dump_enabled_p ())
>             dump_printf_loc (MSG_NOTE, vect_location, "nunits vectype: %T\n",
>                              nunits_vectype);
> diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
> index e5fdc9e0a14..4a809e492c4 100644
> --- a/gcc/tree-vectorizer.h
> +++ b/gcc/tree-vectorizer.h
> @@ -1161,6 +1161,9 @@ public:
>       vectorization.  */
>    bool vectorizable;
>
> +  /* The scalar type of the LHS of this statement is complex type.  */
> +  bool complex_p;
> +
>    /* The stmt to which this info struct refers to.  */
>    gimple *stmt;
>
> @@ -1395,6 +1398,7 @@ struct gather_scatter_info {
>  #define STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT(S) (S)->reduc_epilogue_adjustment
>  #define STMT_VINFO_REDUC_IDX(S)                   (S)->reduc_idx
>  #define STMT_VINFO_FORCE_SINGLE_CYCLE(S)   (S)->force_single_cycle
> +#define STMT_VINFO_COMPLEX_P(S)            (S)->complex_p
>
>  #define STMT_VINFO_DR_WRT_VEC_LOOP(S)      (S)->dr_wrt_vec_loop
>  #define STMT_VINFO_DR_BASE_ADDRESS(S)      (S)->dr_wrt_vec_loop.base_address
> @@ -1970,6 +1974,15 @@ vect_get_num_copies (loop_vec_info loop_vinfo, tree vectype)
>    return vect_get_num_vectors (LOOP_VINFO_VECT_FACTOR (loop_vinfo), vectype);
>  }
>
> +static inline unsigned int
> +vect_get_num_copies (loop_vec_info loop_vinfo, tree vectype, bool complex_p)
> +{
> +  poly_uint64 nunits = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
> +  if (complex_p)
> +    nunits *= 2;
> +  return vect_get_num_vectors (nunits, vectype);
> +}
> +
>  /* Update maximum unit count *MAX_NUNITS so that it accounts for
>     NUNITS.  *MAX_NUNITS can be 1 if we haven't yet recorded anything.  */
>
> --
> 2.18.1
>


More information about the Gcc-patches mailing list