Pengxuan Zheng <quic_pzh...@quicinc.com> writes:
> We can still use SVE's INDEX instruction to construct vectors even if not all
> elements are constants. For example, { 0, x, 2, 3 } can be constructed by
> first
> using "INDEX #0, #1" to generate { 0, 1, 2, 3 }, and then set the elements
> which
> are non-constants separately.
>
> PR target/113328
>
> gcc/ChangeLog:
>
> * config/aarch64/aarch64.cc (aarch64_expand_vector_init_fallback):
> Improve part-variable vector generation with SVE's INDEX if TARGET_SVE
> is available.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/aarch64/sve/acle/general/dupq_1.c: Update test to use
> check-function-bodies.
> * gcc.target/aarch64/sve/acle/general/dupq_2.c: Likewise.
> * gcc.target/aarch64/sve/acle/general/dupq_3.c: Likewise.
> * gcc.target/aarch64/sve/acle/general/dupq_4.c: Likewise.
> * gcc.target/aarch64/sve/vec_init_4.c: New test.
> * gcc.target/aarch64/sve/vec_init_5.c: New test.
>
> Signed-off-by: Pengxuan Zheng <quic_pzh...@quicinc.com>
> ---
> gcc/config/aarch64/aarch64.cc | 81 ++++++++++++++++++-
> .../aarch64/sve/acle/general/dupq_1.c | 18 ++++-
> .../aarch64/sve/acle/general/dupq_2.c | 18 ++++-
> .../aarch64/sve/acle/general/dupq_3.c | 18 ++++-
> .../aarch64/sve/acle/general/dupq_4.c | 18 ++++-
> .../gcc.target/aarch64/sve/vec_init_4.c | 47 +++++++++++
> .../gcc.target/aarch64/sve/vec_init_5.c | 12 +++
> 7 files changed, 199 insertions(+), 13 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/vec_init_4.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/vec_init_5.c
>
> diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
> index 6b3ca57d0eb..7305a5c6375 100644
> --- a/gcc/config/aarch64/aarch64.cc
> +++ b/gcc/config/aarch64/aarch64.cc
> @@ -23942,12 +23942,91 @@ aarch64_expand_vector_init_fallback (rtx target,
> rtx vals)
> if (n_var != n_elts)
> {
> rtx copy = copy_rtx (vals);
> + bool is_index_seq = false;
> +
> + /* If at least half of the elements of the vector are constants and all
> + these constant elements form a linear sequence of the form { B, B + S,
> + B + 2 * S, B + 3 * S, ... }, we can generate the vector with SVE's
> + INDEX instruction if SVE is available and then set the elements which
> + are not constant separately. More precisely, each constant element I
> + has to be B + I * S where B and S must be valid immediate operand for
> + an SVE INDEX instruction.
> +
> + For example, { X, 1, 2, 3} is a vector satisfying these conditions and
> + we can generate a vector of all constants (i.e., { 0, 1, 2, 3 }) first
> + and then set the first element of the vector to X. */
> +
> + if (TARGET_SVE && GET_MODE_CLASS (mode) == MODE_VECTOR_INT
> + && n_var <= n_elts / 2)
> + {
> + int const_idx = -1;
> + HOST_WIDE_INT const_val = 0;
> + int base = 16;
> + int step = 16;
> +
> + for (int i = 0; i < n_elts; ++i)
> + {
> + rtx x = XVECEXP (vals, 0, i);
> +
> + if (!CONST_INT_P (x))
> + continue;
> +
> + if (const_idx == -1)
> + {
> + const_idx = i;
> + const_val = INTVAL (x);
> + }
> + else
> + {
> + if ((INTVAL (x) - const_val) % (i - const_idx) == 0)
> + {
> + HOST_WIDE_INT s
> + = (INTVAL (x) - const_val) / (i - const_idx);
> + if (s >= -16 && s <= 15)
> + {
> + int b = const_val - s * const_idx;
> + if (b >= -16 && b <= 15)
> + {
> + base = b;
> + step = s;
> + }
> + }
> + }
> + break;
> + }
> + }
> +
> + if (base != 16
> + && (!CONST_INT_P (v0)
> + || (CONST_INT_P (v0) && INTVAL (v0) == base)))
> + {
> + if (!CONST_INT_P (v0))
> + XVECEXP (copy, 0, 0) = GEN_INT (base);
> +
> + is_index_seq = true;
> + for (int i = 1; i < n_elts; ++i)
> + {
> + rtx x = XVECEXP (copy, 0, i);
> +
> + if (CONST_INT_P (x))
> + {
> + if (INTVAL (x) != base + i * step)
> + {
> + is_index_seq = false;
> + break;
> + }
> + }
> + else
> + XVECEXP (copy, 0, i) = GEN_INT (base + i * step);
> + }
> + }
> + }
This seems a bit more complex than I was hoping for, although the
complexity is probably justified.
Seeing how awkard it is to do this using current interfaces, I think
I'd instead prefer to do something that I'd been vaguely hoping to do
for a while: extend vector-builder.h to accept wildcard/don't care values.
finalize () could then replace the wildcards with whatever gives the
"nicest" encoding.
That's also going to be relatively complex, but I think it'd be more
general, and might help with the existing vec_init code as well.
It would also be a step towards optimising -1 indices for
__builtin_shufflevector. It might be a few weeks before I can post
something though.
Pushing 1/2 without 2/2 has meant that the dupq tests will fail in the
meantime, but that's ok. In general, though, it's better not to push
individual patches from a series unless they've been tested in isolation
and are known to give clean test results.
Thanks,
Richard
