On Tue, Apr 02, 2024 at 09:56:01AM +0200, Juergen Christ wrote:
> Loop vectorizer can generate vector permutes with constant indexes
> where all indexes are equal. Optimize this case to use vector
> replicate instead of vector permute.
>
> gcc/ChangeLog:
>
> * config/s390/s390.cc (expand_perm_as_replicate): Implement.
> (vectorize_vec_perm_const_1): Call new function.
> * config/s390/vx-builtins.md (vec_splat<mode>): Change to...
> (@vec_splat<mode>): ...this.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/s390/vector/vec-expand-replicate.c: New test.
>
> Bootstrapped and regtested on s390x. Ok for trunk?
>
> Signed-off-by: Juergen Christ <jchr...@linux.ibm.com>
> ---
> gcc/config/s390/s390.cc | 32 +++++++++++++++++++
> gcc/config/s390/vx-builtins.md | 2 +-
> .../s390/vector/vec-expand-replicate.c | 30 +++++++++++++++++
> 3 files changed, 63 insertions(+), 1 deletion(-)
> create mode 100644
> gcc/testsuite/gcc.target/s390/vector/vec-expand-replicate.c
>
> diff --git a/gcc/config/s390/s390.cc b/gcc/config/s390/s390.cc
> index 372a23244032..4b4014ebe444 100644
> --- a/gcc/config/s390/s390.cc
> +++ b/gcc/config/s390/s390.cc
> @@ -17923,6 +17923,35 @@ expand_perm_as_a_vlbr_vstbr_candidate (const struct
> expand_vec_perm_d &d)
> return false;
> }
>
> +static bool expand_perm_as_replicate (const struct expand_vec_perm_d &d)
^~~~~~~~~~~~~~~~~~~~~~~~
Function names start on a new line.
> +{
> + unsigned char i;
> + unsigned char elem;
> + rtx base = d.op0;
> + rtx insn;
> + /* Needed to silence maybe-uninitialized warning. */
> + gcc_assert(d.nelt > 0);
~~~~~~~~~~^~~~~~~~~~~~
Between function name and open bracket whitespace is missing.
Curiously enough, the error is about d which is a reference and cannot
be null. If you are eager you could reduce this and open a PR.
s390.cc:17935:8: warning: ‘d’ may be used uninitialized [-Wmaybe-uninitialized]
17935 | elem = d.perm[0];
| ~~~~~^~~~~~~~~~~
> + elem = d.perm[0];
> + for (i = 1; i < d.nelt; ++i)
> + if (d.perm[i] != elem)
> + return false;
> + if (!d.testing_p)
> + {
> + if (elem >= d.nelt)
> + {
> + base = d.op1;
> + elem -= d.nelt;
> + }
> + insn = maybe_gen_vec_splat (d.vmode, d.target, base, GEN_INT (elem));
> + if (insn == NULL_RTX)
> + return false;
> + emit_insn (insn);
> + return true;
> + }
> + else
> + return maybe_code_for_vec_splat (d.vmode) != CODE_FOR_nothing;
> +}
> +
> /* Try to find the best sequence for the vector permute operation
> described by D. Return true if the operation could be
> expanded. */
> @@ -17941,6 +17970,9 @@ vectorize_vec_perm_const_1 (const struct
> expand_vec_perm_d &d)
> if (expand_perm_as_a_vlbr_vstbr_candidate (d))
> return true;
>
> + if (expand_perm_as_replicate(d))
~~~~~~~~~~~~~~~~~~~~~~~~^~~
Between function name and open bracket whitespace is missing.
> + return true;
> +
> return false;
> }
>
> diff --git a/gcc/config/s390/vx-builtins.md b/gcc/config/s390/vx-builtins.md
> index 432d81a719fc..93c0d408a43e 100644
> --- a/gcc/config/s390/vx-builtins.md
> +++ b/gcc/config/s390/vx-builtins.md
> @@ -424,7 +424,7 @@
>
>
> ; Replicate from vector element
> -(define_expand "vec_splat<mode>"
> +(define_expand "@vec_splat<mode>"
> [(set (match_operand:V_HW 0 "register_operand" "")
> (vec_duplicate:V_HW (vec_select:<non_vec>
> (match_operand:V_HW 1 "register_operand" "")
> diff --git a/gcc/testsuite/gcc.target/s390/vector/vec-expand-replicate.c
> b/gcc/testsuite/gcc.target/s390/vector/vec-expand-replicate.c
> new file mode 100644
> index 000000000000..27563a00f22b
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/s390/vector/vec-expand-replicate.c
> @@ -0,0 +1,30 @@
> +/* Check that the vectorize_vec_perm_const expander correctly deals with
> + replication. Extracted from spec "nab". */
> +
> +/* { dg-do compile } */
> +/* { dg-options "-O3 -mzarch -march=z13 -fvect-cost-model=unlimited" } */
> +
> +
> +#define REAL_T double
> +typedef REAL_T MATRIX_T[ 4 ][ 4 ];
> +
> +int concat_mat_i, concat_mat_j;
> +static void concat_mat(MATRIX_T m1, MATRIX_T, MATRIX_T m3);
> +MATRIX_T *rot4p() {
> + MATRIX_T mat3, mat4;
> + static MATRIX_T mat5;
> + concat_mat(mat4, mat3, mat5);
> +}
> +void concat_mat(MATRIX_T m1, MATRIX_T, MATRIX_T m3) {
> + int k;
> + for (;; concat_mat_i++) {
> + concat_mat_j = 0;
> + for (; 4; concat_mat_j++) {
> + k = 0;
> + for (; k < 4; k++)
> + m3[concat_mat_i][concat_mat_j] += m1[concat_mat_i][k];
> + }
Just nitpicking, if we could come up with a test case which does not
involve integer overflows due to non-terminating loops, I would prefer
that.
Cheers,
Stefan
> + }
> +}
> +
> +/* { dg-final { scan-assembler-not "vperm" } } */
> --
> 2.39.3
>
