On Mon, Jul 12, 2021 at 7:55 PM Richard Sandiford
<[email protected]> wrote:
>
> Richard Biener via Gcc-patches <[email protected]> writes:
> > On Fri, Jul 9, 2021 at 3:12 PM Richard Sandiford
> > <[email protected]> wrote:
> >>
> >> Thanks for the review.
> >>
> >> Richard Biener <[email protected]> writes:
> >> >> @@ -588,6 +600,23 @@ public:
> >> >> /* Unrolling factor */
> >> >> poly_uint64 vectorization_factor;
> >> >>
> >> >> + /* If this loop is an epilogue loop whose main loop can be skipped,
> >> >> + MAIN_LOOP_EDGE is the edge from the main loop to this loop's
> >> >> + preheader. SKIP_MAIN_LOOP_EDGE is then the edge that skips the
> >> >> + main loop and goes straight to this loop's preheader.
> >> >> +
> >> >> + Both fields are null otherwise. */
> >> >> + edge main_loop_edge;
> >> >> + edge skip_main_loop_edge;
> >> >> +
> >> >> + /* If this loop is an epilogue loop that might be skipped after
> >> >> executing
> >> >> + the main loop, this edge is the one that skips the epilogue. */
> >> >> + edge skip_this_loop_edge;
> >> >> +
> >> >> + /* After vectorization, maps live-out SSA names to information about
> >> >> + the reductions that generated them. */
> >> >> + hash_map<tree, vect_reusable_accumulator> reusable_accumulators;
> >> >
> >> > Is that the LC PHI node defs or the definition inside of the loop?
> >> > If the latter we could attach the info directly to its stmt-info?
> >>
> >> Ah, yeah, I should improve the comment there. It's the vectoriser's
> >> replacement for the original LC PHI node, i.e. the final scalar result
> >> after the reduction has taken place.
> >
> > OK
> >
> >> >> @@ -1186,6 +1215,21 @@ public:
> >> >> /* The vector type for performing the actual reduction. */
> >> >> tree reduc_vectype;
> >> >>
> >> >> + /* If IS_REDUC_INFO is true and if the reduction is operating on N
> >> >> + elements in parallel, this vector gives the initial values of
> >> >> these
> >> >> + N elements. */
> >> >
> >> > That's N scalar elements or N vector elements? I suppose it's for
> >> > SLP reductions (rather than SLP reduction chains) and never non-SLP
> >> > reductions?
> >>
> >> Yeah, poor wording again, sorry. I meant something closer to:
> >>
> >> /* If IS_REDUC_INFO is true and if the vector code is performing
> >> N scalar reductions in parallel, this vector gives the initial
> >> scalar values of those N reductions. */
> >>
> >> >> + vec<tree> reduc_initial_values;
> >> >> +
> >> >> + /* If IS_REDUC_INFO is true and if the reduction is operating on N
> >> >> + elements in parallel, this vector gives the scalar result of each
> >> >> + reduction. */
> >> >> + vec<tree> reduc_scalar_results;
> >>
> >> Same change here.
> >>
> >> >> […]
> >> >> diff --git a/gcc/tree-vect-loop-manip.c b/gcc/tree-vect-loop-manip.c
> >> >> index 2909e8a0fc3..b7b0523e3c8 100644
> >> >> --- a/gcc/tree-vect-loop-manip.c
> >> >> +++ b/gcc/tree-vect-loop-manip.c
> >> >> @@ -2457,6 +2457,31 @@ vect_update_epilogue_niters (loop_vec_info
> >> >> epilogue_vinfo,
> >> >> return vect_determine_partial_vectors_and_peeling (epilogue_vinfo,
> >> >> true);
> >> >> }
> >> >>
> >> >> +/* LOOP_VINFO is an epilogue loop and MAIN_LOOP_VALUE is available on
> >> >> exit
> >> >> + from the corresponding main loop. Return a value that is available
> >> >> in
> >> >> + LOOP_VINFO's preheader, using SKIP_VALUE if the main loop is
> >> >> skipped.
> >> >> + Passing a null SKIP_VALUE is equivalent to passing zero. */
> >> >> +
> >> >> +tree
> >> >> +vect_get_main_loop_result (loop_vec_info loop_vinfo, tree
> >> >> main_loop_value,
> >> >> + tree skip_value)
> >> >> +{
> >> >> + if (!loop_vinfo->main_loop_edge)
> >> >> + return main_loop_value;
> >> >> +
> >> >> + if (!skip_value)
> >> >> + skip_value = build_zero_cst (TREE_TYPE (main_loop_value));
> >> >
> >> > shouldn't that be the initial value?
> >>
> >> For the current use case, the above two conditions are never true.
> >> I wrote it like this because I had a follow-on patch (which might
> >> not go anywhere) that needed this function for 0-based IVs.
> >>
> >> Maybe that's a bad risk/reward trade-off though. Not having to pass
> >> zero makes things only slightly simpler for the follow-on patch,
> >> and I guess could be dangerous in other cases.
> >>
> >> Perhaps in that case though I should change loop_vinfo->main_loop_edge
> >> into a gcc_assert as well.
> >
> > Yeah, I think asserts (and comments in case it's because we don't handle
> > some specific cases yet) are better than possibly wrong behavior.
>
> OK.
>
> >> >> + tree phi_result = make_ssa_name (TREE_TYPE (main_loop_value));
> >> >> + basic_block bb = loop_vinfo->main_loop_edge->dest;
> >> >> + gphi *new_phi = create_phi_node (phi_result, bb);
> >> >> + add_phi_arg (new_phi, main_loop_value, loop_vinfo->main_loop_edge,
> >> >> + UNKNOWN_LOCATION);
> >> >> + add_phi_arg (new_phi, skip_value,
> >> >> + loop_vinfo->skip_main_loop_edge, UNKNOWN_LOCATION);
> >> >> + return phi_result;
> >> >> +}
> >> >> +
> >> >> /* Function vect_do_peeling.
> >> >>
> >> >> Input:
> >> >> […]
> >> >> @@ -4823,6 +4842,100 @@ info_for_reduction (vec_info *vinfo,
> >> >> stmt_vec_info stmt_info)
> >> >> return stmt_info;
> >> >> }
> >> >>
> >> >> +/* PHI is a reduction in LOOP_VINFO that we are going to vectorize
> >> >> using vector
> >> >> + type VECTYPE. See if LOOP_VINFO is an epilogue loop whose main
> >> >> loop had a
> >> >> + matching reduction that we can build on. Adjust REDUC_INFO and
> >> >> return true
> >> >> + if so, otherwise return false. */
> >> >> +
> >> >> +static bool
> >> >> +vect_find_reusable_accumulator (loop_vec_info loop_vinfo,
> >> >> + stmt_vec_info reduc_info)
> >> >> +{
> >> >> + loop_vec_info main_loop_vinfo = LOOP_VINFO_ORIG_LOOP_INFO
> >> >> (loop_vinfo);
> >> >> + if (!main_loop_vinfo)
> >> >> + return false;
> >> >> +
> >> >> + if (STMT_VINFO_REDUC_TYPE (reduc_info) != TREE_CODE_REDUCTION)
> >> >> + return false;
> >> >> +
> >> >> + unsigned int num_phis = reduc_info->reduc_initial_values.length ();
> >> >> + auto_vec<tree, 16> main_loop_results (num_phis);
> >> >> + auto_vec<tree, 16> initial_values (num_phis);
> >> >> + if (edge main_loop_edge = loop_vinfo->main_loop_edge)
> >> >> + {
> >> >> + /* The epilogue loop can be entered either from the main loop or
> >> >> + from an earlier guard block. */
> >> >> + edge skip_edge = loop_vinfo->skip_main_loop_edge;
> >> >> + for (tree incoming_value : reduc_info->reduc_initial_values)
> >> >> + {
> >> >> + /* Look for:
> >> >> +
> >> >> + INCOMING_VALUE = phi<MAIN_LOOP_RESULT(main loop),
> >> >> + INITIAL_VALUE(guard block)>. */
> >> >> + gcc_assert (TREE_CODE (incoming_value) == SSA_NAME);
> >> >> +
> >> >> + gphi *phi = as_a <gphi *> (SSA_NAME_DEF_STMT
> >> >> (incoming_value));
> >> >> + gcc_assert (gimple_bb (phi) == main_loop_edge->dest);
> >> >> +
> >> >> + tree from_main_loop = PHI_ARG_DEF_FROM_EDGE (phi,
> >> >> main_loop_edge);
> >> >> + tree from_skip = PHI_ARG_DEF_FROM_EDGE (phi, skip_edge);
> >> >> +
> >> >> + main_loop_results.quick_push (from_main_loop);
> >> >> + initial_values.quick_push (from_skip);
> >> >> + }
> >> >> + }
> >> >> + else
> >> >> + /* The main loop dominates the epilogue loop. */
> >> >> + main_loop_results.splice (reduc_info->reduc_initial_values);
> >> >> +
> >> >> + /* See if the main loop has the kind of accumulator we need. */
> >> >> + vect_reusable_accumulator *accumulator
> >> >> + = main_loop_vinfo->reusable_accumulators.get
> >> >> (main_loop_results[0]);
> >> >> + if (!accumulator
> >> >> + || num_phis !=
> >> >> accumulator->reduc_info->reduc_scalar_results.length ()
> >> >> + || !std::equal (main_loop_results.begin (),
> >> >> main_loop_results.end (),
> >> >> +
> >> >> accumulator->reduc_info->reduc_scalar_results.begin ()))
> >> >> + return false;
> >> >> +
> >> >> + /* For now, only handle the case in which both loops are operating
> >> >> on the
> >> >> + same vector types. In future we could reduce wider vectors to
> >> >> narrower
> >> >> + ones as well. */
> >> >> + tree vectype = STMT_VINFO_VECTYPE (reduc_info);
> >> >> + tree old_vectype = TREE_TYPE (accumulator->reduc_input);
> >> >> + if (!useless_type_conversion_p (old_vectype, vectype))
> >> >
> >> > It should be indeed quite trivial to handle, likewise the case where we
> >> > have multiple PHIs - just reduce to a single input vector and have the
> >> > possibly multiple input vectors in the epilogue filled with neutral
> >> > elements. I'll see if I can cook up stuff for this next week.
> >>
> >> Yeah, agreed. The multi-vector epilogue case should be especially easy
> >> to handle, but it's not interesting for SVE as things stand, since:
> >>
> >> (a) non-SLP reductions use a single cycle for ncopies>1 (a misfeature
> >> IMO -- on targets with wide pipelines we want exactly the opposite)
> >>
> >> (b) SLP reductions are limited to single vectors for variable-length
> >> targets.
> >>
> >> So it wasn't possible to trigger multiple epilogue vectors for the
> >> motivating SVE use case.
> >
> > OK, I see. If the series is in I'll see to create testcases for x86_64.
> >
> >> >> […]
> >> >> @@ -5196,6 +5305,37 @@ vect_create_epilog_for_reduction (loop_vec_info
> >> >> loop_vinfo,
> >> >> reduc_inputs.safe_push (single_input);
> >> >> }
> >> >>
> >> >> + tree orig_reduc_input = reduc_inputs[0];
> >> >> +
> >> >> + /* If this loop is an epilogue loop that can be skipped after the
> >> >> + main loop, we can only share a reduction operation between the
> >> >> + main loop and the epilogue if we put it at the target of the
> >> >> + skip edge.
> >> >
> >> > Do you have a testcase where we cannot do this?
> >>
> >> No, it's being defensive. I wasn't sure how the epilogue code would
> >> evolve in future.
> >>
> >> >> + We can still reuse accumulators if this check fails. Doing so has
> >> >> + the minor(?) benefit of making the epilogue loop's scalar result
> >> >> + independent of the main loop's scalar result. */
> >> >> + bool unify_with_main_loop_p = false;
> >> >> + if (reduc_info->reused_accumulator
> >> >> + && loop_vinfo->skip_this_loop_edge
> >> >> + && single_succ_p (exit_bb)
> >> >> + && single_succ (exit_bb) ==
> >> >> loop_vinfo->skip_this_loop_edge->dest)
> >> >> + {
> >> >> + unify_with_main_loop_p = true;
> >> >> +
> >> >> + basic_block reduc_block = loop_vinfo->skip_this_loop_edge->dest;
> >> >> + reduc_inputs[0] = make_ssa_name (vectype);
> >> >> + gphi *new_phi = create_phi_node (reduc_inputs[0], reduc_block);
> >> >> + add_phi_arg (new_phi, orig_reduc_input, single_succ_edge
> >> >> (exit_bb),
> >> >> + UNKNOWN_LOCATION);
> >> >> + add_phi_arg (new_phi,
> >> >> reduc_info->reused_accumulator->reduc_input,
> >> >> + loop_vinfo->skip_this_loop_edge, UNKNOWN_LOCATION);
> >> >> + exit_gsi = gsi_after_labels (reduc_block);
> >> >> + }
> >> >> +
> >> >> + /* Shouldn't be used beyond this point. */
> >> >> + exit_bb = nullptr;
> >> >> +
> >> >> if (STMT_VINFO_REDUC_TYPE (reduc_info) == COND_REDUCTION
> >> >> && reduc_fn != IFN_LAST)
> >> >> {
> >> >> @@ -5819,6 +5958,12 @@ vect_create_epilog_for_reduction (loop_vec_info
> >> >> loop_vinfo,
> >> >> scalar_results[0] = new_temp;
> >> >> }
> >> >>
> >> >> + /* Record this operation if it could be reused by the epilogue loop.
> >> >> */
> >> >> + if (STMT_VINFO_REDUC_TYPE (reduc_info) == TREE_CODE_REDUCTION
> >> >> + && !double_reduc)
> >> >
> >> > what's the issue with double_reduc?
> >>
> >> Probably nothing TBH. I haven't been able to construct a case that
> >> uses predicated double reductions with vect-partial-vector-usage=1,
> >> but that's probably a missed optimisation.
> >>
> >> There again, double reductions themselves seem to be hard to trigger
> >> now that we have loop interchange. Is there a good way of testing
> >> them without -fno-loop-interchange?
> >
> > there are a bunch of testcases in gcc.dg/vect/vect-double-reduc-?.c,
> > I don't see how interchange avoids the double reduction, in fact when
> > doing interchange we no longer can apply outer loop vectorization (but
> > it's still a double reduction, just only inner loop vectorized).
> > But eventually we don't do epilogue vectorization for outer loop
> > vectorizations with reductions.
>
> Well, to take the vect-double-reduc-1.c loop:
>
> for (k = 0; k < K; k++)
> {
> sum = 0;
> for (j = 0; j < K; j++)
> for (i = 0; i < K; i++)
> sum += in[i+k][j] * coeff[i][j];
>
> out[k] = sum;
> }
>
> loop interchange converts this to:
>
> for (k = 0; k < K; k++)
> {
> sum = 0;
> for (i = 0; i < K; i++)
> for (j = 0; j < K; j++)
> sum += in[i+k][j] * coeff[i][j];
>
> out[k] = sum;
> }
>
> and then we vectorise the inner loop.
>
> I think in principle that's the right thing to do, since the double
> reduction is more like:
>
> for (j1 = 0; j1 < K; j1 += vsize)
> for (i = 0; i < K; i++)
> for (j = j1; j < MIN (K, j1 + vsize); j++)
> sum += in[i+k][j] * coeff[i][j];
>
> which isn't as nice an access pattern.
True, OTOH which version is faster probably depends on the inner
loop kernel and the number of memory streams involved.
> It would be good if we could sink the vector->scalar reduction
> in the interchanged form, but that feels like a separate optimisation,
> and could potentially happen regardless of whether we can vectorise
> any other code in the outer loop. E.g. it could happen for:
>
> for (i = 0; i < K; i++)
> {
> printf ("Hello, world!\n");
> for (j = 0; j < K; j++)
> sum += in[i+k][j] * coeff[i][j];
> }
Indeed.
> > Oh, and of course vect.exp runs with -O2 -ftree-vectorize, avoiding
> > any of the high-level loop opts ...
>
> Ah, yeah, of course. I was trying to use the vect-double-reduc*.c
> tests with -O3…
>
> How does this version look? Changes from v1:
>
> - Fix comments in new tree-vectorizer.h fields.
> - Fix an out-of-date comment above vect_find_reusable_accumulator
> - Remove !double_reduc condition.
> - Make vect_get_main_loop_result specific to the case in which a phi
> node is needed.
>
> Tested as above.
Looks good to me now, thus OK.
Thanks,
Richard.
> Thanks,
> Richard
>
> gcc/
> * tree-vectorizer.h (vect_reusable_accumulator): New structure.
> (_loop_vec_info::main_loop_edge): New field.
> (_loop_vec_info::skip_main_loop_edge): Likewise.
> (_loop_vec_info::skip_this_loop_edge): Likewise.
> (_loop_vec_info::reusable_accumulators): Likewise.
> (_stmt_vec_info::reduc_scalar_results): Likewise.
> (_stmt_vec_info::reused_accumulator): Likewise.
> (vect_get_main_loop_result): Declare.
> * tree-vectorizer.c (vec_info::new_stmt_vec_info): Initialize
> reduc_scalar_inputs.
> (vec_info::free_stmt_vec_info): Free reduc_scalar_inputs.
> * tree-vect-loop-manip.c (vect_get_main_loop_result): New function.
> (vect_do_peeling): Fill an epilogue loop's main_loop_edge,
> skip_main_loop_edge and skip_this_loop_edge fields.
> * tree-vect-loop.c (INCLUDE_ALGORITHM): Define.
> (vect_emit_reduction_init_stmts): New function.
> (get_initial_def_for_reduction): Use it.
> (get_initial_defs_for_reduction): Likewise. Change the vinfo
> parameter to a loop_vec_info.
> (vect_create_epilog_for_reduction): Store the scalar results
> in the reduc_info. If an epilogue loop is reusing an accumulator
> from the main loop, and if the epilogue loop can also be skipped,
> try to place the reduction code in the join block. Record
> accumulators that could potentially be reused by epilogue loops.
> (vect_transform_cycle_phi): When vectorizing epilogue loops,
> try to reuse accumulators from the main loop. Record the initial
> value in reduc_info for non-SLP reductions too.
>
> gcc/testsuite/
> * gcc.target/aarch64/sve/reduc_9.c: New test.
> * gcc.target/aarch64/sve/reduc_9_run.c: Likewise.
> * gcc.target/aarch64/sve/reduc_10.c: Likewise.
> * gcc.target/aarch64/sve/reduc_10_run.c: Likewise.
> * gcc.target/aarch64/sve/reduc_11.c: Likewise.
> * gcc.target/aarch64/sve/reduc_11_run.c: Likewise.
> * gcc.target/aarch64/sve/reduc_12.c: Likewise.
> * gcc.target/aarch64/sve/reduc_12_run.c: Likewise.
> * gcc.target/aarch64/sve/reduc_13.c: Likewise.
> * gcc.target/aarch64/sve/reduc_13_run.c: Likewise.
> * gcc.target/aarch64/sve/reduc_14.c: Likewise.
> * gcc.target/aarch64/sve/reduc_14_run.c: Likewise.
> * gcc.target/aarch64/sve/reduc_15.c: Likewise.
> * gcc.target/aarch64/sve/reduc_15_run.c: Likewise.
> ---
> .../gcc.target/aarch64/sve/reduc_10.c | 77 +++++
> .../gcc.target/aarch64/sve/reduc_10_run.c | 49 +++
> .../gcc.target/aarch64/sve/reduc_11.c | 71 ++++
> .../gcc.target/aarch64/sve/reduc_11_run.c | 34 ++
> .../gcc.target/aarch64/sve/reduc_12.c | 71 ++++
> .../gcc.target/aarch64/sve/reduc_12_run.c | 66 ++++
> .../gcc.target/aarch64/sve/reduc_13.c | 101 ++++++
> .../gcc.target/aarch64/sve/reduc_13_run.c | 61 ++++
> .../gcc.target/aarch64/sve/reduc_14.c | 107 ++++++
> .../gcc.target/aarch64/sve/reduc_14_run.c | 187 +++++++++++
> .../gcc.target/aarch64/sve/reduc_15.c | 16 +
> .../gcc.target/aarch64/sve/reduc_15_run.c | 22 ++
> .../gcc.target/aarch64/sve/reduc_9.c | 77 +++++
> .../gcc.target/aarch64/sve/reduc_9_run.c | 29 ++
> gcc/tree-vect-loop-manip.c | 26 ++
> gcc/tree-vect-loop.c | 307 ++++++++++++++----
> gcc/tree-vectorizer.c | 4 +
> gcc/tree-vectorizer.h | 56 +++-
> 18 files changed, 1297 insertions(+), 64 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/reduc_10.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/reduc_10_run.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/reduc_11.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/reduc_11_run.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/reduc_12.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/reduc_12_run.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/reduc_13.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/reduc_13_run.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/reduc_14.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/reduc_14_run.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/reduc_15.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/reduc_15_run.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/reduc_9.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/reduc_9_run.c
>
> diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
> index e2fd3609fee..d825b0c3723 100644
> --- a/gcc/tree-vectorizer.h
> +++ b/gcc/tree-vectorizer.h
> @@ -551,6 +551,18 @@ typedef auto_vec<rgroup_controls> vec_loop_lens;
>
> typedef auto_vec<std::pair<data_reference*, tree> > drs_init_vec;
>
> +/* Information about a reduction accumulator from the main loop that could
> + conceivably be reused as the input to a reduction in an epilogue loop. */
> +struct vect_reusable_accumulator {
> + /* The final value of the accumulator, which forms the input to the
> + reduction operation. */
> + tree reduc_input;
> +
> + /* The stmt_vec_info that describes the reduction (i.e. the one for
> + which is_reduc_info is true). */
> + stmt_vec_info reduc_info;
> +};
> +
> /*-----------------------------------------------------------------*/
> /* Info on vectorized loops. */
> /*-----------------------------------------------------------------*/
> @@ -588,6 +600,26 @@ public:
> /* Unrolling factor */
> poly_uint64 vectorization_factor;
>
> + /* If this loop is an epilogue loop whose main loop can be skipped,
> + MAIN_LOOP_EDGE is the edge from the main loop to this loop's
> + preheader. SKIP_MAIN_LOOP_EDGE is then the edge that skips the
> + main loop and goes straight to this loop's preheader.
> +
> + Both fields are null otherwise. */
> + edge main_loop_edge;
> + edge skip_main_loop_edge;
> +
> + /* If this loop is an epilogue loop that might be skipped after executing
> + the main loop, this edge is the one that skips the epilogue. */
> + edge skip_this_loop_edge;
> +
> + /* The vectorized form of a standard reduction replaces the original
> + scalar code's final result (a loop-closed SSA PHI) with the result
> + of a vector-to-scalar reduction operation. After vectorization,
> + this variable maps these vector-to-scalar results to information
> + about the reductions that generated them. */
> + hash_map<tree, vect_reusable_accumulator> reusable_accumulators;
> +
> /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR
> if there is no particular limit. */
> unsigned HOST_WIDE_INT max_vectorization_factor;
> @@ -1186,6 +1218,23 @@ public:
> /* The vector type for performing the actual reduction. */
> tree reduc_vectype;
>
> + /* If IS_REDUC_INFO is true and if the vector code is performing
> + N scalar reductions in parallel, this variable gives the initial
> + scalar values of those N reductions. */
> + vec<tree> reduc_initial_values;
> +
> + /* If IS_REDUC_INFO is true and if the vector code is performing
> + N scalar reductions in parallel, this variable gives the vectorized
> code's
> + final (scalar) result for each of those N reductions. In other words,
> + REDUC_SCALAR_RESULTS[I] replaces the original scalar code's loop-closed
> + SSA PHI for reduction number I. */
> + vec<tree> reduc_scalar_results;
> +
> + /* Only meaningful if IS_REDUC_INFO. If non-null, the reduction is
> + being performed by an epilogue loop and we have decided to reuse
> + this accumulator from the main loop. */
> + vect_reusable_accumulator *reused_accumulator;
> +
> /* Whether we force a single cycle PHI during reduction vectorization. */
> bool force_single_cycle;
>
> @@ -1382,12 +1431,6 @@ vect_phi_initial_value (gphi *phi)
> return PHI_ARG_DEF_FROM_EDGE (phi, pe);
> }
>
> -static inline tree
> -vect_phi_initial_value (stmt_vec_info stmt_info)
> -{
> - return vect_phi_initial_value (as_a <gphi *> (stmt_info->stmt));
> -}
> -
> /* Return true if STMT_INFO should produce a vector mask type rather than
> a normal nonmask type. */
>
> @@ -1818,6 +1861,7 @@ class loop *vect_loop_versioning (loop_vec_info, gimple
> *);
> extern class loop *vect_do_peeling (loop_vec_info, tree, tree,
> tree *, tree *, tree *, int, bool, bool,
> tree *);
> +extern tree vect_get_main_loop_result (loop_vec_info, tree, tree);
> extern void vect_prepare_for_masked_peels (loop_vec_info);
> extern dump_user_location_t find_loop_location (class loop *);
> extern bool vect_can_advance_ivs_p (loop_vec_info);
> diff --git a/gcc/tree-vectorizer.c b/gcc/tree-vectorizer.c
> index 9748043f3ee..f1035a83826 100644
> --- a/gcc/tree-vectorizer.c
> +++ b/gcc/tree-vectorizer.c
> @@ -694,6 +694,8 @@ vec_info::new_stmt_vec_info (gimple *stmt)
> STMT_VINFO_SLP_VECT_ONLY (res) = false;
> STMT_VINFO_SLP_VECT_ONLY_PATTERN (res) = false;
> STMT_VINFO_VEC_STMTS (res) = vNULL;
> + res->reduc_initial_values = vNULL;
> + res->reduc_scalar_results = vNULL;
>
> if (is_a <loop_vec_info> (this)
> && gimple_code (stmt) == GIMPLE_PHI
> @@ -755,6 +757,8 @@ vec_info::free_stmt_vec_info (stmt_vec_info stmt_info)
> release_ssa_name (lhs);
> }
>
> + stmt_info->reduc_initial_values.release ();
> + stmt_info->reduc_scalar_results.release ();
> STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release ();
> STMT_VINFO_VEC_STMTS (stmt_info).release ();
> free (stmt_info);
> diff --git a/gcc/tree-vect-loop-manip.c b/gcc/tree-vect-loop-manip.c
> index 2909e8a0fc3..c29ffb3356c 100644
> --- a/gcc/tree-vect-loop-manip.c
> +++ b/gcc/tree-vect-loop-manip.c
> @@ -2457,6 +2457,28 @@ vect_update_epilogue_niters (loop_vec_info
> epilogue_vinfo,
> return vect_determine_partial_vectors_and_peeling (epilogue_vinfo, true);
> }
>
> +/* LOOP_VINFO is an epilogue loop whose corresponding main loop can be
> skipped.
> + Return a value that equals:
> +
> + - MAIN_LOOP_VALUE when LOOP_VINFO is entered from the main loop and
> + - SKIP_VALUE when the main loop is skipped. */
> +
> +tree
> +vect_get_main_loop_result (loop_vec_info loop_vinfo, tree main_loop_value,
> + tree skip_value)
> +{
> + gcc_assert (loop_vinfo->main_loop_edge);
> +
> + tree phi_result = make_ssa_name (TREE_TYPE (main_loop_value));
> + basic_block bb = loop_vinfo->main_loop_edge->dest;
> + gphi *new_phi = create_phi_node (phi_result, bb);
> + add_phi_arg (new_phi, main_loop_value, loop_vinfo->main_loop_edge,
> + UNKNOWN_LOCATION);
> + add_phi_arg (new_phi, skip_value,
> + loop_vinfo->skip_main_loop_edge, UNKNOWN_LOCATION);
> + return phi_result;
> +}
> +
> /* Function vect_do_peeling.
>
> Input:
> @@ -2986,6 +3008,8 @@ vect_do_peeling (loop_vec_info loop_vinfo, tree niters,
> tree nitersm1,
> skip_vector ? anchor : guard_bb,
> prob_epilog.invert (),
> irred_flag);
> + if (vect_epilogues)
> + epilogue_vinfo->skip_this_loop_edge = guard_e;
> slpeel_update_phi_nodes_for_guard2 (loop, epilog, guard_e,
> single_exit (epilog));
> /* Only need to handle basic block before epilog loop if it's not
> @@ -3057,6 +3081,8 @@ vect_do_peeling (loop_vec_info loop_vinfo, tree niters,
> tree nitersm1,
> add_phi_arg (new_phi, build_zero_cst (TREE_TYPE (niters)), skip_e,
> UNKNOWN_LOCATION);
> niters = PHI_RESULT (new_phi);
> + epilogue_vinfo->main_loop_edge = update_e;
> + epilogue_vinfo->skip_main_loop_edge = skip_e;
> }
>
> /* Set ADVANCE to the number of iterations performed by the previous
> diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c
> index fe7e73f655f..8c27d75f889 100644
> --- a/gcc/tree-vect-loop.c
> +++ b/gcc/tree-vect-loop.c
> @@ -19,6 +19,7 @@ You should have received a copy of the GNU General Public
> License
> along with GCC; see the file COPYING3. If not see
> <http://www.gnu.org/licenses/>. */
>
> +#define INCLUDE_ALGORITHM
> #include "config.h"
> #include "system.h"
> #include "coretypes.h"
> @@ -823,6 +824,10 @@ _loop_vec_info::_loop_vec_info (class loop *loop_in,
> vec_info_shared *shared)
> th (0),
> versioning_threshold (0),
> vectorization_factor (0),
> + main_loop_edge (nullptr),
> + skip_main_loop_edge (nullptr),
> + skip_this_loop_edge (nullptr),
> + reusable_accumulators (),
> max_vectorization_factor (0),
> mask_skip_niters (NULL_TREE),
> rgroup_compare_type (NULL_TREE),
> @@ -4607,7 +4612,32 @@ vect_model_reduction_cost (loop_vec_info loop_vinfo,
> prologue_cost, epilogue_cost);
> }
>
> +/* SEQ is a sequence of instructions that initialize the reduction
> + described by REDUC_INFO. Emit them in the appropriate place. */
>
> +static void
> +vect_emit_reduction_init_stmts (loop_vec_info loop_vinfo,
> + stmt_vec_info reduc_info, gimple *seq)
> +{
> + if (reduc_info->reused_accumulator)
> + {
> + /* When reusing an accumulator from the main loop, we only need
> + initialization instructions if the main loop can be skipped.
> + In that case, emit the initialization instructions at the end
> + of the guard block that does the skip. */
> + edge skip_edge = loop_vinfo->skip_main_loop_edge;
> + gcc_assert (skip_edge);
> + gimple_stmt_iterator gsi = gsi_last_bb (skip_edge->src);
> + gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT);
> + }
> + else
> + {
> + /* The normal case: emit the initialization instructions on the
> + preheader edge. */
> + class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
> + gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), seq);
> + }
> +}
>
> /* Function get_initial_def_for_reduction
>
> @@ -4675,36 +4705,30 @@ get_initial_def_for_reduction (loop_vec_info
> loop_vinfo,
> }
>
> if (stmts)
> - gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
> + vect_emit_reduction_init_stmts (loop_vinfo, reduc_info, stmts);
> return init_def;
> }
>
> -/* Get at the initial defs for the reduction PHIs for REDUC_INFO, whose
> - associated SLP node is SLP_NODE. NUMBER_OF_VECTORS is the number of
> vector
> - defs to create. If NEUTRAL_OP is nonnull, introducing extra elements of
> - that value will not change the result. */
> +/* Get at the initial defs for the reduction PHIs for REDUC_INFO,
> + which performs a reduction involving GROUP_SIZE scalar statements.
> + NUMBER_OF_VECTORS is the number of vector defs to create. If NEUTRAL_OP
> + is nonnull, introducing extra elements of that value will not change the
> + result. */
>
> static void
> -get_initial_defs_for_reduction (vec_info *vinfo,
> +get_initial_defs_for_reduction (loop_vec_info loop_vinfo,
> stmt_vec_info reduc_info,
> - slp_tree slp_node,
> vec<tree> *vec_oprnds,
> unsigned int number_of_vectors,
> - bool reduc_chain, tree neutral_op)
> + unsigned int group_size, tree neutral_op)
> {
> - vec<stmt_vec_info> stmts = SLP_TREE_SCALAR_STMTS (slp_node);
> + vec<tree> &initial_values = reduc_info->reduc_initial_values;
> unsigned HOST_WIDE_INT nunits;
> unsigned j, number_of_places_left_in_vector;
> tree vector_type = STMT_VINFO_VECTYPE (reduc_info);
> - unsigned int group_size = stmts.length ();
> unsigned int i;
> - class loop *loop;
> -
> - loop = (gimple_bb (reduc_info->stmt))->loop_father;
> - gcc_assert (loop);
> - edge pe = loop_preheader_edge (loop);
>
> - gcc_assert (!reduc_chain || neutral_op);
> + gcc_assert (group_size == initial_values.length () || neutral_op);
>
> /* NUMBER_OF_COPIES is the number of times we need to use the same values
> in
> created vectors. It is greater than 1 if unrolling is performed.
> @@ -4734,18 +4758,13 @@ get_initial_defs_for_reduction (vec_info *vinfo,
> {
> tree op;
> i = j % group_size;
> - stmt_vec_info stmt_vinfo = stmts[i];
>
> /* Get the def before the loop. In reduction chain we have only
> one initial value. Else we have as many as PHIs in the group. */
> - if (reduc_chain)
> - op = j != 0 ? neutral_op : vect_phi_initial_value (stmt_vinfo);
> - else if (((vec_oprnds->length () + 1) * nunits
> - - number_of_places_left_in_vector >= group_size)
> - && neutral_op)
> + if (i >= initial_values.length () || (j > i && neutral_op))
> op = neutral_op;
> else
> - op = vect_phi_initial_value (stmt_vinfo);
> + op = initial_values[i];
>
> /* Create 'vect_ = {op0,op1,...,opn}'. */
> number_of_places_left_in_vector--;
> @@ -4781,8 +4800,8 @@ get_initial_defs_for_reduction (vec_info *vinfo,
> {
> /* First time round, duplicate ELTS to fill the
> required number of vectors. */
> - duplicate_and_interleave (vinfo, &ctor_seq, vector_type, elts,
> - number_of_vectors, *vec_oprnds);
> + duplicate_and_interleave (loop_vinfo, &ctor_seq, vector_type,
> + elts, number_of_vectors, *vec_oprnds);
> break;
> }
> vec_oprnds->quick_push (init);
> @@ -4794,7 +4813,7 @@ get_initial_defs_for_reduction (vec_info *vinfo,
> }
> }
> if (ctor_seq != NULL)
> - gsi_insert_seq_on_edge_immediate (pe, ctor_seq);
> + vect_emit_reduction_init_stmts (loop_vinfo, reduc_info, ctor_seq);
> }
>
> /* For a statement STMT_INFO taking part in a reduction operation return
> @@ -4823,6 +4842,99 @@ info_for_reduction (vec_info *vinfo, stmt_vec_info
> stmt_info)
> return stmt_info;
> }
>
> +/* See if LOOP_VINFO is an epilogue loop whose main loop had a reduction that
> + REDUC_INFO can build on. Adjust REDUC_INFO and return true if so,
> otherwise
> + return false. */
> +
> +static bool
> +vect_find_reusable_accumulator (loop_vec_info loop_vinfo,
> + stmt_vec_info reduc_info)
> +{
> + loop_vec_info main_loop_vinfo = LOOP_VINFO_ORIG_LOOP_INFO (loop_vinfo);
> + if (!main_loop_vinfo)
> + return false;
> +
> + if (STMT_VINFO_REDUC_TYPE (reduc_info) != TREE_CODE_REDUCTION)
> + return false;
> +
> + unsigned int num_phis = reduc_info->reduc_initial_values.length ();
> + auto_vec<tree, 16> main_loop_results (num_phis);
> + auto_vec<tree, 16> initial_values (num_phis);
> + if (edge main_loop_edge = loop_vinfo->main_loop_edge)
> + {
> + /* The epilogue loop can be entered either from the main loop or
> + from an earlier guard block. */
> + edge skip_edge = loop_vinfo->skip_main_loop_edge;
> + for (tree incoming_value : reduc_info->reduc_initial_values)
> + {
> + /* Look for:
> +
> + INCOMING_VALUE = phi<MAIN_LOOP_RESULT(main loop),
> + INITIAL_VALUE(guard block)>. */
> + gcc_assert (TREE_CODE (incoming_value) == SSA_NAME);
> +
> + gphi *phi = as_a <gphi *> (SSA_NAME_DEF_STMT (incoming_value));
> + gcc_assert (gimple_bb (phi) == main_loop_edge->dest);
> +
> + tree from_main_loop = PHI_ARG_DEF_FROM_EDGE (phi, main_loop_edge);
> + tree from_skip = PHI_ARG_DEF_FROM_EDGE (phi, skip_edge);
> +
> + main_loop_results.quick_push (from_main_loop);
> + initial_values.quick_push (from_skip);
> + }
> + }
> + else
> + /* The main loop dominates the epilogue loop. */
> + main_loop_results.splice (reduc_info->reduc_initial_values);
> +
> + /* See if the main loop has the kind of accumulator we need. */
> + vect_reusable_accumulator *accumulator
> + = main_loop_vinfo->reusable_accumulators.get (main_loop_results[0]);
> + if (!accumulator
> + || num_phis != accumulator->reduc_info->reduc_scalar_results.length ()
> + || !std::equal (main_loop_results.begin (), main_loop_results.end (),
> + accumulator->reduc_info->reduc_scalar_results.begin ()))
> + return false;
> +
> + /* For now, only handle the case in which both loops are operating on the
> + same vector types. In future we could reduce wider vectors to narrower
> + ones as well. */
> + tree vectype = STMT_VINFO_VECTYPE (reduc_info);
> + tree old_vectype = TREE_TYPE (accumulator->reduc_input);
> + if (!useless_type_conversion_p (old_vectype, vectype))
> + return false;
> +
> + /* Non-SLP reductions might apply an adjustment after the reduction
> + operation, in order to simplify the initialization of the accumulator.
> + If the epilogue loop carries on from where the main loop left off,
> + it should apply the same adjustment to the final reduction result.
> +
> + If the epilogue loop can also be entered directly (rather than via
> + the main loop), we need to be able to handle that case in the same way,
> + with the same adjustment. (In principle we could add a PHI node
> + to select the correct adjustment, but in practice that shouldn't be
> + necessary.) */
> + tree main_adjustment
> + = STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT (accumulator->reduc_info);
> + if (loop_vinfo->main_loop_edge && main_adjustment)
> + {
> + gcc_assert (num_phis == 1);
> + tree initial_value = initial_values[0];
> + /* Check that we can use INITIAL_VALUE as the adjustment and
> + initialize the accumulator with a neutral value instead. */
> + if (!operand_equal_p (initial_value, main_adjustment))
> + return false;
> + tree_code code = STMT_VINFO_REDUC_CODE (reduc_info);
> + initial_values[0] = neutral_op_for_reduction (TREE_TYPE
> (initial_value),
> + code, initial_value);
> + }
> + STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT (reduc_info) = main_adjustment;
> + reduc_info->reduc_initial_values.truncate (0);
> + reduc_info->reduc_initial_values.splice (initial_values);
> + reduc_info->reused_accumulator = accumulator;
> + return true;
> +}
> +
> /* Function vect_create_epilog_for_reduction
>
> Create code at the loop-epilog to finalize the result of a reduction
> @@ -4915,7 +5027,7 @@ vect_create_epilog_for_reduction (loop_vec_info
> loop_vinfo,
> gimple *use_stmt;
> auto_vec<tree> reduc_inputs;
> int j, i;
> - auto_vec<tree> scalar_results;
> + vec<tree> &scalar_results = reduc_info->reduc_scalar_results;
> unsigned int group_size = 1, k;
> auto_vec<gimple *> phis;
> /* SLP reduction without reduction chain, e.g.,
> @@ -4941,16 +5053,12 @@ vect_create_epilog_for_reduction (loop_vec_info
> loop_vinfo,
> gcc_assert (vectype);
> mode = TYPE_MODE (vectype);
>
> - tree initial_def = NULL;
> tree induc_val = NULL_TREE;
> tree adjustment_def = NULL;
> if (slp_node)
> ;
> else
> {
> - /* Get at the scalar def before the loop, that defines the initial
> value
> - of the reduction variable. */
> - initial_def = vect_phi_initial_value (reduc_def_stmt);
> /* Optimize: for induction condition reduction, if we can't use zero
> for induc_val, use initial_def. */
> if (STMT_VINFO_REDUC_TYPE (reduc_info) == INTEGER_INDUC_COND_REDUCTION)
> @@ -5196,6 +5304,37 @@ vect_create_epilog_for_reduction (loop_vec_info
> loop_vinfo,
> reduc_inputs.safe_push (single_input);
> }
>
> + tree orig_reduc_input = reduc_inputs[0];
> +
> + /* If this loop is an epilogue loop that can be skipped after the
> + main loop, we can only share a reduction operation between the
> + main loop and the epilogue if we put it at the target of the
> + skip edge.
> +
> + We can still reuse accumulators if this check fails. Doing so has
> + the minor(?) benefit of making the epilogue loop's scalar result
> + independent of the main loop's scalar result. */
> + bool unify_with_main_loop_p = false;
> + if (reduc_info->reused_accumulator
> + && loop_vinfo->skip_this_loop_edge
> + && single_succ_p (exit_bb)
> + && single_succ (exit_bb) == loop_vinfo->skip_this_loop_edge->dest)
> + {
> + unify_with_main_loop_p = true;
> +
> + basic_block reduc_block = loop_vinfo->skip_this_loop_edge->dest;
> + reduc_inputs[0] = make_ssa_name (vectype);
> + gphi *new_phi = create_phi_node (reduc_inputs[0], reduc_block);
> + add_phi_arg (new_phi, orig_reduc_input, single_succ_edge (exit_bb),
> + UNKNOWN_LOCATION);
> + add_phi_arg (new_phi, reduc_info->reused_accumulator->reduc_input,
> + loop_vinfo->skip_this_loop_edge, UNKNOWN_LOCATION);
> + exit_gsi = gsi_after_labels (reduc_block);
> + }
> +
> + /* Shouldn't be used beyond this point. */
> + exit_bb = nullptr;
> +
> if (STMT_VINFO_REDUC_TYPE (reduc_info) == COND_REDUCTION
> && reduc_fn != IFN_LAST)
> {
> @@ -5405,6 +5544,7 @@ vect_create_epilog_for_reduction (loop_vec_info
> loop_vinfo,
> the same as initial_def already. */
> tree zcompare = build2 (EQ_EXPR, boolean_type_node, new_temp,
> induc_val);
> + tree initial_def = reduc_info->reduc_initial_values[0];
>
> tmp = make_ssa_name (new_scalar_dest);
> epilog_stmt = gimple_build_assign (tmp, COND_EXPR, zcompare,
> @@ -5425,9 +5565,6 @@ vect_create_epilog_for_reduction (loop_vec_info
> loop_vinfo,
> gcc_assert (reduc_inputs.length () == 1);
> gcc_assert (pow2p_hwi (group_size));
>
> - slp_tree orig_phis_slp_node = slp_node_instance->reduc_phis;
> - vec<stmt_vec_info> orig_phis
> - = SLP_TREE_SCALAR_STMTS (orig_phis_slp_node);
> gimple_seq seq = NULL;
>
> /* Build a vector {0, 1, 2, ...}, with the same number of elements
> @@ -5452,7 +5589,7 @@ vect_create_epilog_for_reduction (loop_vec_info
> loop_vinfo,
> {
> tree initial_value = NULL_TREE;
> if (REDUC_GROUP_FIRST_ELEMENT (stmt_info))
> - initial_value = vect_phi_initial_value (orig_phis[0]);
> + initial_value = reduc_info->reduc_initial_values[0];
> neutral_op = neutral_op_for_reduction (TREE_TYPE (vectype), code,
> initial_value);
> }
> @@ -5466,7 +5603,7 @@ vect_create_epilog_for_reduction (loop_vec_info
> loop_vinfo,
> for MIN and MAX reduction, for example. */
> if (!neutral_op)
> {
> - tree scalar_value = vect_phi_initial_value (orig_phis[i]);
> + tree scalar_value = reduc_info->reduc_initial_values[i];
> scalar_value = gimple_convert (&seq, TREE_TYPE (vectype),
> scalar_value);
> vector_identity = gimple_build_vector_from_val (&seq, vectype,
> @@ -5780,6 +5917,7 @@ vect_create_epilog_for_reduction (loop_vec_info
> loop_vinfo,
> the same as initial_def already. */
> tree zcompare = build2 (EQ_EXPR, boolean_type_node, new_temp,
> induc_val);
> + tree initial_def = reduc_info->reduc_initial_values[0];
>
> tree tmp = make_ssa_name (new_scalar_dest);
> epilog_stmt = gimple_build_assign (tmp, COND_EXPR, zcompare,
> @@ -5819,6 +5957,11 @@ vect_create_epilog_for_reduction (loop_vec_info
> loop_vinfo,
> scalar_results[0] = new_temp;
> }
>
> + /* Record this operation if it could be reused by the epilogue loop. */
> + if (STMT_VINFO_REDUC_TYPE (reduc_info) == TREE_CODE_REDUCTION)
> + loop_vinfo->reusable_accumulators.put (scalar_results[0],
> + { orig_reduc_input, reduc_info });
> +
> if (double_reduc)
> loop = outer_loop;
>
> @@ -5886,6 +6029,17 @@ vect_create_epilog_for_reduction (loop_vec_info
> loop_vinfo,
> {
> /* Replace the uses: */
> orig_name = PHI_RESULT (exit_phi);
> +
> + /* Look for a single use at the target of the skip edge. */
> + if (unify_with_main_loop_p)
> + {
> + use_operand_p use_p;
> + gimple *user;
> + if (!single_imm_use (orig_name, &use_p, &user))
> + gcc_unreachable ();
> + orig_name = gimple_get_lhs (user);
> + }
> +
> scalar_result = scalar_results[k];
> FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, orig_name)
> {
> @@ -7421,16 +7575,32 @@ vect_transform_cycle_phi (loop_vec_info loop_vinfo,
> else
> {
> gcc_assert (slp_node == slp_node_instance->reduc_phis);
> - tree initial_value = NULL_TREE;
> + vec<tree> &initial_values = reduc_info->reduc_initial_values;
> + vec<stmt_vec_info> &stmts = SLP_TREE_SCALAR_STMTS (slp_node);
> +
> + unsigned int num_phis = stmts.length ();
> if (REDUC_GROUP_FIRST_ELEMENT (reduc_stmt_info))
> - initial_value = vect_phi_initial_value (phi);
> - tree_code code = STMT_VINFO_REDUC_CODE (reduc_info);
> - tree neutral_op = neutral_op_for_reduction (TREE_TYPE (vectype_out),
> - code, initial_value);
> - get_initial_defs_for_reduction (loop_vinfo, reduc_info,
> - slp_node_instance->reduc_phis,
> - &vec_initial_defs, vec_num,
> - initial_value != NULL, neutral_op);
> + num_phis = 1;
> + initial_values.reserve (num_phis);
> + for (unsigned int i = 0; i < num_phis; ++i)
> + {
> + gphi *this_phi = as_a<gphi *> (stmts[i]->stmt);
> + initial_values.quick_push (vect_phi_initial_value (this_phi));
> + }
> + if (vec_num == 1)
> + vect_find_reusable_accumulator (loop_vinfo, reduc_info);
> + if (!initial_values.is_empty ())
> + {
> + tree initial_value
> + = (num_phis == 1 ? initial_values[0] : NULL_TREE);
> + tree_code code = STMT_VINFO_REDUC_CODE (reduc_info);
> + tree neutral_op
> + = neutral_op_for_reduction (TREE_TYPE (vectype_out),
> + code, initial_value);
> + get_initial_defs_for_reduction (loop_vinfo, reduc_info,
> + &vec_initial_defs, vec_num,
> + stmts.length (), neutral_op);
> + }
> }
> }
> else
> @@ -7438,6 +7608,7 @@ vect_transform_cycle_phi (loop_vec_info loop_vinfo,
> /* Get at the scalar def before the loop, that defines the initial
> value of the reduction variable. */
> tree initial_def = vect_phi_initial_value (phi);
> + reduc_info->reduc_initial_values.safe_push (initial_def);
> /* Optimize: if initial_def is for REDUC_MAX smaller than the base
> and we can't use zero for induc_val, use initial_def. Similarly
> for REDUC_MIN and initial_def larger than the base. */
> @@ -7474,21 +7645,30 @@ vect_transform_cycle_phi (loop_vec_info loop_vinfo,
> initial_def, initial_def);
> else
> {
> - enum tree_code code = STMT_VINFO_REDUC_CODE (reduc_info);
> - tree neutral_op = neutral_op_for_reduction (TREE_TYPE (initial_def),
> - code, initial_def);
> - gcc_assert (neutral_op);
> - /* Try to simplify the vector initialization by applying an
> - adjustment after the reduction has been performed. */
> - if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def
> - && !operand_equal_p (neutral_op, initial_def))
> + if (ncopies == 1)
> + vect_find_reusable_accumulator (loop_vinfo, reduc_info);
> + if (!reduc_info->reduc_initial_values.is_empty ())
> {
> - STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT (reduc_info) = initial_def;
> - initial_def = neutral_op;
> + initial_def = reduc_info->reduc_initial_values[0];
> + enum tree_code code = STMT_VINFO_REDUC_CODE (reduc_info);
> + tree neutral_op
> + = neutral_op_for_reduction (TREE_TYPE (initial_def),
> + code, initial_def);
> + gcc_assert (neutral_op);
> + /* Try to simplify the vector initialization by applying an
> + adjustment after the reduction has been performed. */
> + if (!reduc_info->reused_accumulator
> + && STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def
> + && !operand_equal_p (neutral_op, initial_def))
> + {
> + STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT (reduc_info)
> + = initial_def;
> + initial_def = neutral_op;
> + }
> + vec_initial_def
> + = get_initial_def_for_reduction (loop_vinfo, reduc_info,
> + initial_def, neutral_op);
> }
> - vec_initial_def
> - = get_initial_def_for_reduction (loop_vinfo, reduc_info,
> - initial_def, neutral_op);
> }
> }
>
> @@ -7499,6 +7679,17 @@ vect_transform_cycle_phi (loop_vec_info loop_vinfo,
> vec_initial_defs.quick_push (vec_initial_def);
> }
>
> + if (auto *accumulator = reduc_info->reused_accumulator)
> + {
> + if (loop_vinfo->main_loop_edge)
> + vec_initial_defs[0]
> + = vect_get_main_loop_result (loop_vinfo, accumulator->reduc_input,
> + vec_initial_defs[0]);
> + else
> + vec_initial_defs.safe_push (accumulator->reduc_input);
> + gcc_assert (vec_initial_defs.length () == 1);
> + }
> +
> /* Generate the reduction PHIs upfront. */
> for (i = 0; i < vec_num; i++)
> {
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/reduc_10.c
> b/gcc/testsuite/gcc.target/aarch64/sve/reduc_10.c
> new file mode 100644
> index 00000000000..fb817b73d77
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/reduc_10.c
> @@ -0,0 +1,77 @@
> +/* { dg-options "-O3 --param vect-partial-vector-usage=1" } */
> +
> +unsigned short __attribute__((noipa))
> +add_loop (unsigned short *x, int n)
> +{
> + unsigned short res = 0;
> + for (int i = 0; i < n; ++i)
> + res += x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +min_loop (unsigned short *x, int n)
> +{
> + unsigned short res = ~0;
> + for (int i = 0; i < n; ++i)
> + res = res < x[i] ? res : x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +max_loop (unsigned short *x, int n)
> +{
> + unsigned short res = 0;
> + for (int i = 0; i < n; ++i)
> + res = res > x[i] ? res : x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +and_loop (unsigned short *x, int n)
> +{
> + unsigned short res = ~0;
> + for (int i = 0; i < n; ++i)
> + res &= x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +or_loop (unsigned short *x, int n)
> +{
> + unsigned short res = 0;
> + for (int i = 0; i < n; ++i)
> + res |= x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +eor_loop (unsigned short *x, int n)
> +{
> + unsigned short res = 0;
> + for (int i = 0; i < n; ++i)
> + res ^= x[i];
> + return res;
> +}
> +
> +/* { dg-final { scan-assembler-times {\tadd\tz[0-9]+\.h, z[0-9]+\.h,
> z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tadd\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tuaddv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\tumin\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 2 } } */
> +/* { dg-final { scan-assembler-times {\tuminv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\tumax\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 2 } } */
> +/* { dg-final { scan-assembler-times {\tumaxv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\tand\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tand\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tandv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\torr\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\torr\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\torv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\teor\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\teor\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\teorv\t} 1 } } */
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/reduc_10_run.c
> b/gcc/testsuite/gcc.target/aarch64/sve/reduc_10_run.c
> new file mode 100644
> index 00000000000..1dd579be701
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/reduc_10_run.c
> @@ -0,0 +1,49 @@
> +/* { dg-do run { target aarch64_sve_hw } } */
> +/* { dg-options "-O3 --param vect-partial-vector-usage=1" } */
> +
> +#define N 0x1100
> +
> +#include "reduc_10.c"
> +
> +int
> +main (void)
> +{
> + unsigned short x[N];
> + for (int i = 0; i < N; ++i)
> + x[i] = (i + 1) * (i + 2);
> +
> + if (add_loop (x, 0) != 0
> + || add_loop (x, 11) != 572
> + || add_loop (x, 0x100) != 22016
> + || add_loop (x, 0xfff) != 20480
> + || max_loop (x, 0) != 0
> + || max_loop (x, 11) != 132
> + || max_loop (x, 0x100) != 65280
> + || max_loop (x, 0xfff) != 65504
> + || or_loop (x, 0) != 0
> + || or_loop (x, 11) != 0xfe
> + || or_loop (x, 0x80) != 0x7ffe
> + || or_loop (x, 0xb4) != 0x7ffe
> + || or_loop (x, 0xb5) != 0xfffe
> + || eor_loop (x, 0) != 0
> + || eor_loop (x, 11) != 0xe8
> + || eor_loop (x, 0x100) != 0xcf00
> + || eor_loop (x, 0xfff) != 0xa000)
> + __builtin_abort ();
> +
> + for (int i = 0; i < N; ++i)
> + x[i] = ~x[i];
> +
> + if (min_loop (x, 0) != 65535
> + || min_loop (x, 11) != 65403
> + || min_loop (x, 0x100) != 255
> + || min_loop (x, 0xfff) != 31
> + || and_loop (x, 0) != 0xffff
> + || and_loop (x, 11) != 0xff01
> + || and_loop (x, 0x80) != 0x8001
> + || and_loop (x, 0xb4) != 0x8001
> + || and_loop (x, 0xb5) != 1)
> + __builtin_abort ();
> +
> + return 0;
> +}
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/reduc_11.c
> b/gcc/testsuite/gcc.target/aarch64/sve/reduc_11.c
> new file mode 100644
> index 00000000000..f99ef4aa865
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/reduc_11.c
> @@ -0,0 +1,71 @@
> +/* { dg-options "-O3 -msve-vector-bits=256 --param
> vect-partial-vector-usage=1" } */
> +
> +unsigned short __attribute__((noipa))
> +add_loop (unsigned short *x, unsigned short res)
> +{
> + for (int i = 0; i < 0xfff; ++i)
> + res += x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +min_loop (unsigned short *x, unsigned short res)
> +{
> + for (int i = 0; i < 0xfff; ++i)
> + res = res < x[i] ? res : x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +max_loop (unsigned short *x, unsigned short res)
> +{
> + for (int i = 0; i < 0xfff; ++i)
> + res = res > x[i] ? res : x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +and_loop (unsigned short *x, unsigned short res)
> +{
> + for (int i = 0; i < 0xfff; ++i)
> + res &= x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +or_loop (unsigned short *x, unsigned short res)
> +{
> + for (int i = 0; i < 0xfff; ++i)
> + res |= x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +eor_loop (unsigned short *x, unsigned short res)
> +{
> + for (int i = 0; i < 0xfff; ++i)
> + res ^= x[i];
> + return res;
> +}
> +
> +/* { dg-final { scan-assembler-times {\tadd\tz[0-9]+\.h, z[0-9]+\.h,
> z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tadd\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tuaddv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\tumin\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 2 } } */
> +/* { dg-final { scan-assembler-times {\tuminv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\tumax\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 2 } } */
> +/* { dg-final { scan-assembler-times {\tumaxv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\tand\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tand\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tandv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\torr\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\torr\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\torv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\teor\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\teor\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\teorv\t} 1 } } */
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/reduc_11_run.c
> b/gcc/testsuite/gcc.target/aarch64/sve/reduc_11_run.c
> new file mode 100644
> index 00000000000..5b41560d2ef
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/reduc_11_run.c
> @@ -0,0 +1,34 @@
> +/* { dg-do run { target aarch64_sve256_hw } } */
> +/* { dg-options "-O3 -msve-vector-bits=256 --param
> vect-partial-vector-usage=1" } */
> +
> +#define N 0x1100
> +
> +#include "reduc_11.c"
> +
> +int
> +main (void)
> +{
> + unsigned short x[N];
> + for (int i = 0; i < N; ++i)
> + x[i] = (i + 1) * (i + 2);
> +
> + if (add_loop (x, 42) != 20522
> + || max_loop (x, 65503) != 65504
> + || max_loop (x, 65505) != 65505
> + || or_loop (x, 0) != 0xfffe
> + || or_loop (x, 1) != 0xffff
> + || eor_loop (x, 0) != 0xa000
> + || eor_loop (x, 0xbfff) != 0x1fff)
> + __builtin_abort ();
> +
> + for (int i = 0; i < N; ++i)
> + x[i] = ~x[i];
> +
> + if (min_loop (x, 32) != 31
> + || min_loop (x, 30) != 30
> + || and_loop (x, 0xff) != 1
> + || and_loop (x, 0) != 0)
> + __builtin_abort ();
> +
> + return 0;
> +}
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/reduc_12.c
> b/gcc/testsuite/gcc.target/aarch64/sve/reduc_12.c
> new file mode 100644
> index 00000000000..d32b81a61bc
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/reduc_12.c
> @@ -0,0 +1,71 @@
> +/* { dg-options "-O3 --param vect-partial-vector-usage=1" } */
> +
> +unsigned short __attribute__((noipa))
> +add_loop (unsigned short *x, int n, unsigned short res)
> +{
> + for (int i = 0; i < n; ++i)
> + res += x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +min_loop (unsigned short *x, int n, unsigned short res)
> +{
> + for (int i = 0; i < n; ++i)
> + res = res < x[i] ? res : x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +max_loop (unsigned short *x, int n, unsigned short res)
> +{
> + for (int i = 0; i < n; ++i)
> + res = res > x[i] ? res : x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +and_loop (unsigned short *x, int n, unsigned short res)
> +{
> + for (int i = 0; i < n; ++i)
> + res &= x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +or_loop (unsigned short *x, int n, unsigned short res)
> +{
> + for (int i = 0; i < n; ++i)
> + res |= x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +eor_loop (unsigned short *x, int n, unsigned short res)
> +{
> + for (int i = 0; i < n; ++i)
> + res ^= x[i];
> + return res;
> +}
> +
> +/* { dg-final { scan-assembler-times {\tadd\tz[0-9]+\.h, z[0-9]+\.h,
> z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tadd\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tuaddv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\tumin\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 2 } } */
> +/* { dg-final { scan-assembler-times {\tuminv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\tumax\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 2 } } */
> +/* { dg-final { scan-assembler-times {\tumaxv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\tand\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tand\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tandv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\torr\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\torr\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\torv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\teor\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\teor\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\teorv\t} 1 } } */
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/reduc_12_run.c
> b/gcc/testsuite/gcc.target/aarch64/sve/reduc_12_run.c
> new file mode 100644
> index 00000000000..929b81a9705
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/reduc_12_run.c
> @@ -0,0 +1,66 @@
> +/* { dg-do run { target aarch64_sve_hw } } */
> +/* { dg-options "-O3 --param vect-partial-vector-usage=1" } */
> +
> +#define N 0x1100
> +
> +#include "reduc_12.c"
> +
> +int
> +main (void)
> +{
> + unsigned short x[N];
> + for (int i = 0; i < N; ++i)
> + x[i] = (i + 1) * (i + 2);
> +
> + if (add_loop (x, 0, 10) != 10
> + || add_loop (x, 11, 42) != 614
> + || add_loop (x, 0x100, 84) != 22100
> + || add_loop (x, 0xfff, 20) != 20500
> + || max_loop (x, 0, 10) != 10
> + || max_loop (x, 11, 131) != 132
> + || max_loop (x, 11, 133) != 133
> + || max_loop (x, 0x100, 65279) != 65280
> + || max_loop (x, 0x100, 65281) != 65281
> + || max_loop (x, 0xfff, 65503) != 65504
> + || max_loop (x, 0xfff, 65505) != 65505
> + || or_loop (x, 0, 0x71) != 0x71
> + || or_loop (x, 11, 0) != 0xfe
> + || or_loop (x, 11, 0xb3c) != 0xbfe
> + || or_loop (x, 0x80, 0) != 0x7ffe
> + || or_loop (x, 0x80, 1) != 0x7fff
> + || or_loop (x, 0xb4, 0) != 0x7ffe
> + || or_loop (x, 0xb4, 1) != 0x7fff
> + || or_loop (x, 0xb5, 0) != 0xfffe
> + || or_loop (x, 0xb5, 1) != 0xffff
> + || eor_loop (x, 0, 0x3e) != 0x3e
> + || eor_loop (x, 11, 0) != 0xe8
> + || eor_loop (x, 11, 0x1ff) != 0x117
> + || eor_loop (x, 0x100, 0) != 0xcf00
> + || eor_loop (x, 0x100, 0xeee) != 0xc1ee
> + || eor_loop (x, 0xfff, 0) != 0xa000
> + || eor_loop (x, 0xfff, 0x8888) != 0x2888)
> + __builtin_abort ();
> +
> + for (int i = 0; i < N; ++i)
> + x[i] = ~x[i];
> +
> + if (min_loop (x, 0, 10000) != 10000
> + || min_loop (x, 11, 65404) != 65403
> + || min_loop (x, 11, 65402) != 65402
> + || min_loop (x, 0x100, 256) != 255
> + || min_loop (x, 0x100, 254) != 254
> + || min_loop (x, 0xfff, 32) != 31
> + || min_loop (x, 0xfff, 30) != 30
> + || and_loop (x, 0, 0x1234) != 0x1234
> + || and_loop (x, 11, 0xffff) != 0xff01
> + || and_loop (x, 11, 0xcdef) != 0xcd01
> + || and_loop (x, 0x80, 0xffff) != 0x8001
> + || and_loop (x, 0x80, 0xfffe) != 0x8000
> + || and_loop (x, 0xb4, 0xffff) != 0x8001
> + || and_loop (x, 0xb4, 0xfffe) != 0x8000
> + || and_loop (x, 0xb5, 0xffff) != 1
> + || and_loop (x, 0xb5, 0xfffe) != 0)
> + __builtin_abort ();
> +
> + return 0;
> +}
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/reduc_13.c
> b/gcc/testsuite/gcc.target/aarch64/sve/reduc_13.c
> new file mode 100644
> index 00000000000..ce2b8f2fcdc
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/reduc_13.c
> @@ -0,0 +1,101 @@
> +/* { dg-options "-O3 -msve-vector-bits=256 --param
> vect-partial-vector-usage=1" } */
> +
> +void __attribute__((noipa))
> +add_loop (unsigned int *x, unsigned int *res)
> +{
> + unsigned int res0 = res[0];
> + unsigned int res1 = res[1];
> + for (int i = 0; i < 0x7ff; ++i)
> + {
> + res0 += x[i * 2];
> + res1 += x[i * 2 + 1];
> + }
> + res[0] = res0;
> + res[1] = res1;
> +}
> +
> +void __attribute__((noipa))
> +min_loop (unsigned int *x, unsigned int *res)
> +{
> + unsigned int res0 = res[0];
> + unsigned int res1 = res[1];
> + for (int i = 0; i < 0x7ff; ++i)
> + {
> + res0 = res0 < x[i * 2] ? res0 : x[i * 2];
> + res1 = res1 < x[i * 2 + 1] ? res1 : x[i * 2 + 1];
> + }
> + res[0] = res0;
> + res[1] = res1;
> +}
> +
> +void __attribute__((noipa))
> +max_loop (unsigned int *x, unsigned int *res)
> +{
> + unsigned int res0 = res[0];
> + unsigned int res1 = res[1];
> + for (int i = 0; i < 0x7ff; ++i)
> + {
> + res0 = res0 > x[i * 2] ? res0 : x[i * 2];
> + res1 = res1 > x[i * 2 + 1] ? res1 : x[i * 2 + 1];
> + }
> + res[0] = res0;
> + res[1] = res1;
> +}
> +
> +void __attribute__((noipa))
> +and_loop (unsigned int *x, unsigned int *res)
> +{
> + unsigned int res0 = res[0];
> + unsigned int res1 = res[1];
> + for (int i = 0; i < 0x7ff; ++i)
> + {
> + res0 &= x[i * 2];
> + res1 &= x[i * 2 + 1];
> + }
> + res[0] = res0;
> + res[1] = res1;
> +}
> +
> +void __attribute__((noipa))
> +or_loop (unsigned int *x, unsigned int *res)
> +{
> + unsigned int res0 = res[0];
> + unsigned int res1 = res[1];
> + for (int i = 0; i < 0x7ff; ++i)
> + {
> + res0 |= x[i * 2];
> + res1 |= x[i * 2 + 1];
> + }
> + res[0] = res0;
> + res[1] = res1;
> +}
> +
> +void __attribute__((noipa))
> +eor_loop (unsigned int *x, unsigned int *res)
> +{
> + unsigned int res0 = res[0];
> + unsigned int res1 = res[1];
> + for (int i = 0; i < 0x7ff; ++i)
> + {
> + res0 ^= x[i * 2];
> + res1 ^= x[i * 2 + 1];
> + }
> + res[0] = res0;
> + res[1] = res1;
> +}
> +
> +/* { dg-final { scan-assembler-times {\tadd\tz[0-9]+\.s, z[0-9]+\.s,
> z[0-9]+\.s\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tadd\tz[0-9]+\.s, p[0-7]/m,
> z[0-9]+\.s, z[0-9]+\.s\n} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\tumin\tz[0-9]+\.s, p[0-7]/m,
> z[0-9]+\.s, z[0-9]+\.s\n} 2 } } */
> +
> +/* { dg-final { scan-assembler-times {\tumax\tz[0-9]+\.s, p[0-7]/m,
> z[0-9]+\.s, z[0-9]+\.s\n} 2 } } */
> +
> +/* { dg-final { scan-assembler-times {\tand\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tand\tz[0-9]+\.s, p[0-7]/m,
> z[0-9]+\.s, z[0-9]+\.s\n} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\torr\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\torr\tz[0-9]+\.s, p[0-7]/m,
> z[0-9]+\.s, z[0-9]+\.s\n} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\teor\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\teor\tz[0-9]+\.s, p[0-7]/m,
> z[0-9]+\.s, z[0-9]+\.s\n} 1 } } */
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/reduc_13_run.c
> b/gcc/testsuite/gcc.target/aarch64/sve/reduc_13_run.c
> new file mode 100644
> index 00000000000..5514d8d6b3b
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/reduc_13_run.c
> @@ -0,0 +1,61 @@
> +/* { dg-do run { target aarch64_sve256_hw } } */
> +/* { dg-options "-O3 -msve-vector-bits=256 --param
> vect-partial-vector-usage=1" } */
> +
> +#define N 0x1100
> +
> +#include "reduc_13.c"
> +
> +int
> +main (void)
> +{
> + unsigned int x[N];
> + for (int i = 0; i < N; ++i)
> + x[i] = ((i + 1) * (i + 2)) & 0xfffff;
> +
> + unsigned int add_res[2] = { 42, 1111 };
> + add_loop (x, add_res);
> + if (add_res[0] != 968538154
> + || add_res[1] != 964340823)
> + __builtin_abort ();
> +
> + unsigned int max_res1[2] = { 0, 0 };
> + max_loop (x, max_res1);
> + if (max_res1[0] != 1048150
> + || max_res1[1] != 1045506)
> + __builtin_abort ();
> +
> + unsigned int max_res2[2] = { 1048151, 1045507 };
> + max_loop (x, max_res2);
> + if (max_res2[0] != 1048151
> + || max_res2[1] != 1045507)
> + __builtin_abort ();
> +
> + unsigned int or_res[2] = { 0x1000000, 0x2000000 };
> + or_loop (x, or_res);
> + if (or_res[0] != 0x10ffffe
> + || or_res[1] != 0x20ffffe)
> + __builtin_abort ();
> +
> + unsigned int eor_res[2] = { 0x1000000, 0x2000000 };
> + eor_loop (x, eor_res);
> + if (eor_res[0] != 0x1010000
> + || eor_res[1] != 0x20b5000)
> + __builtin_abort ();
> +
> + for (int i = 0; i < N; ++i)
> + x[i] = ~x[i] & 0xfffff;
> +
> + unsigned int min_res1[2] = { 500, 4000 };
> + min_loop (x, min_res1);
> + if (min_res1[0] != 425
> + || min_res1[1] != 3069)
> + __builtin_abort ();
> +
> + unsigned int min_res2[2] = { 424, 3068 };
> + min_loop (x, min_res2);
> + if (min_res2[0] != 424
> + || min_res2[1] != 3068)
> + __builtin_abort ();
> +
> + return 0;
> +}
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/reduc_14.c
> b/gcc/testsuite/gcc.target/aarch64/sve/reduc_14.c
> new file mode 100644
> index 00000000000..3be611e4b37
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/reduc_14.c
> @@ -0,0 +1,107 @@
> +/* { dg-options "-O3 --param vect-partial-vector-usage=1" } */
> +
> +void __attribute__((noipa))
> +add_loop (unsigned int *x, int n, unsigned int *res)
> +{
> + unsigned int res0 = res[0];
> + unsigned int res1 = res[1];
> + for (int i = 0; i < n; ++i)
> + {
> + res0 += x[i * 2];
> + res1 += x[i * 2 + 1];
> + }
> + res[0] = res0;
> + res[1] = res1;
> +}
> +
> +void __attribute__((noipa))
> +min_loop (unsigned int *x, int n, unsigned int *res)
> +{
> + unsigned int res0 = res[0];
> + unsigned int res1 = res[1];
> + for (int i = 0; i < n; ++i)
> + {
> + res0 = res0 < x[i * 2] ? res0 : x[i * 2];
> + res1 = res1 < x[i * 2 + 1] ? res1 : x[i * 2 + 1];
> + }
> + res[0] = res0;
> + res[1] = res1;
> +}
> +
> +void __attribute__((noipa))
> +max_loop (unsigned int *x, int n, unsigned int *res)
> +{
> + unsigned int res0 = res[0];
> + unsigned int res1 = res[1];
> + for (int i = 0; i < n; ++i)
> + {
> + res0 = res0 > x[i * 2] ? res0 : x[i * 2];
> + res1 = res1 > x[i * 2 + 1] ? res1 : x[i * 2 + 1];
> + }
> + res[0] = res0;
> + res[1] = res1;
> +}
> +
> +void __attribute__((noipa))
> +and_loop (unsigned int *x, int n, unsigned int *res)
> +{
> + unsigned int res0 = res[0];
> + unsigned int res1 = res[1];
> + for (int i = 0; i < n; ++i)
> + {
> + res0 &= x[i * 2];
> + res1 &= x[i * 2 + 1];
> + }
> + res[0] = res0;
> + res[1] = res1;
> +}
> +
> +void __attribute__((noipa))
> +or_loop (unsigned int *x, int n, unsigned int *res)
> +{
> + unsigned int res0 = res[0];
> + unsigned int res1 = res[1];
> + for (int i = 0; i < n; ++i)
> + {
> + res0 |= x[i * 2];
> + res1 |= x[i * 2 + 1];
> + }
> + res[0] = res0;
> + res[1] = res1;
> +}
> +
> +void __attribute__((noipa))
> +eor_loop (unsigned int *x, int n, unsigned int *res)
> +{
> + unsigned int res0 = res[0];
> + unsigned int res1 = res[1];
> + for (int i = 0; i < n; ++i)
> + {
> + res0 ^= x[i * 2];
> + res1 ^= x[i * 2 + 1];
> + }
> + res[0] = res0;
> + res[1] = res1;
> +}
> +
> +/* { dg-final { scan-assembler-times {\tadd\tz[0-9]+\.s, z[0-9]+\.s,
> z[0-9]+\.s\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tadd\tz[0-9]+\.s, p[0-7]/m,
> z[0-9]+\.s, z[0-9]+\.s\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tuaddv\t} 2 } } */
> +
> +/* { dg-final { scan-assembler-times {\tumin\tz[0-9]+\.s, p[0-7]/m,
> z[0-9]+\.s, z[0-9]+\.s\n} 2 } } */
> +/* { dg-final { scan-assembler-times {\tuminv\t} 2 } } */
> +
> +/* { dg-final { scan-assembler-times {\tumax\tz[0-9]+\.s, p[0-7]/m,
> z[0-9]+\.s, z[0-9]+\.s\n} 2 } } */
> +/* { dg-final { scan-assembler-times {\tumaxv\t} 2 } } */
> +
> +/* { dg-final { scan-assembler-times {\tand\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tand\tz[0-9]+\.s, p[0-7]/m,
> z[0-9]+\.s, z[0-9]+\.s\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tandv\t} 2 } } */
> +
> +/* { dg-final { scan-assembler-times {\torr\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\torr\tz[0-9]+\.s, p[0-7]/m,
> z[0-9]+\.s, z[0-9]+\.s\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\torv\t} 2 } } */
> +
> +/* { dg-final { scan-assembler-times {\teor\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\teor\tz[0-9]+\.s, p[0-7]/m,
> z[0-9]+\.s, z[0-9]+\.s\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\teorv\t} 2 } } */
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/reduc_14_run.c
> b/gcc/testsuite/gcc.target/aarch64/sve/reduc_14_run.c
> new file mode 100644
> index 00000000000..ccaa770e9b2
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/reduc_14_run.c
> @@ -0,0 +1,187 @@
> +/* { dg-do run { target aarch64_sve256_hw } } */
> +/* { dg-options "-O3 -msve-vector-bits=256 --param
> vect-partial-vector-usage=1" } */
> +
> +#define N 0x1100
> +
> +#include "reduc_14.c"
> +
> +int
> +main (void)
> +{
> + unsigned int x[N];
> + for (int i = 0; i < N; ++i)
> + x[i] = ((i + 1) * (i + 2)) & 0xfffff;
> +
> + unsigned int add_res1[2] = { 11, 22 };
> + add_loop (x, 0, add_res1);
> + if (add_res1[0] != 11
> + || add_res1[1] != 22)
> + __builtin_abort ();
> +
> + unsigned int add_res2[2] = { 10, 20 };
> + add_loop (x, 11, add_res2);
> + if (add_res2[0] != 1902
> + || add_res2[1] != 2176)
> + __builtin_abort ();
> +
> + unsigned int add_res3[2] = { 15, 30 };
> + add_loop (x, 0x100, add_res3);
> + if (add_res3[0] != 22435087
> + || add_res3[1] != 22566686)
> + __builtin_abort ();
> +
> + unsigned int add_res4[2] = { 100, 200 };
> + add_loop (x, 0x11f, add_res4);
> + if (add_res4[0] != 31602244
> + || add_res4[1] != 31767656)
> + __builtin_abort ();
> +
> + unsigned int max_res1[2] = { 461, 500 };
> + max_loop (x, 11, max_res1);
> + if (max_res1[0] != 462
> + || max_res1[1] != 506)
> + __builtin_abort ();
> +
> + unsigned int max_res2[2] = { 463, 507 };
> + max_loop (x, 11, max_res2);
> + if (max_res2[0] != 463
> + || max_res2[1] != 507)
> + __builtin_abort ();
> +
> + unsigned int max_res3[2] = { 1000000, 1000000 };
> + max_loop (x, 0x200, max_res3);
> + if (max_res3[0] != 1047552
> + || max_res3[1] != 1045506)
> + __builtin_abort ();
> +
> + unsigned int max_res4[2] = { 1047553, 1045507 };
> + max_loop (x, 0x200, max_res4);
> + if (max_res4[0] != 1047553
> + || max_res4[1] != 1045507)
> + __builtin_abort ();
> +
> + unsigned int max_res5[2] = { 300000, 30000 };
> + max_loop (x, 0x11f, max_res5);
> + if (max_res5[0] != 328902
> + || max_res5[1] != 330050)
> + __builtin_abort ();
> +
> + unsigned int max_res6[2] = { 328903, 330051 };
> + max_loop (x, 0x11f, max_res6);
> + if (max_res6[0] != 328903
> + || max_res6[1] != 330051)
> + __builtin_abort ();
> +
> + unsigned int or_res1[2] = { 11, 22 };
> + or_loop (x, 0, or_res1);
> + if (or_res1[0] != 11
> + || or_res1[1] != 22)
> + __builtin_abort ();
> +
> + unsigned int or_res2[2] = { 0x200000, 0xe00000 };
> + or_loop (x, 11, or_res2);
> + if (or_res2[0] != 0x2001fe
> + || or_res2[1] != 0xe001fe)
> + __builtin_abort ();
> +
> + unsigned int or_res3[2] = { 0x800000, 0x700000 };
> + or_loop (x, 0x40, or_res3);
> + if (or_res3[0] != 0x803ffe
> + || or_res3[1] != 0x707ffe)
> + __builtin_abort ();
> +
> + unsigned int or_res4[2] = { 0x100001, 0x300000 };
> + or_loop (x, 0x4f, or_res4);
> + if (or_res4[0] != 0x107fff
> + || or_res4[1] != 0x307ffe)
> + __builtin_abort ();
> +
> + unsigned int eor_res1[2] = { 11, 22 };
> + eor_loop (x, 0, eor_res1);
> + if (eor_res1[0] != 11
> + || eor_res1[1] != 22)
> + __builtin_abort ();
> +
> + unsigned int eor_res2[2] = { 0x2000ff, 0xe000ff };
> + eor_loop (x, 11, eor_res2);
> + if (eor_res2[0] != 0x2001cf
> + || eor_res2[1] != 0xe000b7)
> + __builtin_abort ();
> +
> + unsigned int eor_res3[2] = { 0x805000, 0x70f000 };
> + eor_loop (x, 0x100, eor_res3);
> + if (eor_res3[0] != 0x824200
> + || eor_res3[1] != 0x77dc00)
> + __builtin_abort ();
> +
> + unsigned int eor_res4[2] = { 0x101201, 0x300f00 };
> + eor_loop (x, 0x11f, eor_res4);
> + if (eor_res4[0] != 0x178801
> + || eor_res4[1] != 0x337240)
> + __builtin_abort ();
> +
> + for (int i = 0; i < N; ++i)
> + x[i] = ~x[i] & 0xfffff;
> +
> + unsigned int min_res1[2] = { 1048200, 1048100 };
> + min_loop (x, 11, min_res1);
> + if (min_res1[0] != 1048113
> + || min_res1[1] != 1048069)
> + __builtin_abort ();
> +
> + unsigned int min_res2[2] = { 1048112, 1048068 };
> + min_loop (x, 11, min_res2);
> + if (min_res2[0] != 1048112
> + || min_res2[1] != 1048068)
> + __builtin_abort ();
> +
> + unsigned int min_res3[2] = { 10000, 10000 };
> + min_loop (x, 0x200, min_res3);
> + if (min_res3[0] != 1023
> + || min_res3[1] != 3069)
> + __builtin_abort ();
> +
> + unsigned int min_res4[2] = { 1022, 3068 };
> + min_loop (x, 0x200, min_res4);
> + if (min_res4[0] != 1022
> + || min_res4[1] != 3068)
> + __builtin_abort ();
> +
> + unsigned int min_res5[2] = { 719680, 718530 };
> + min_loop (x, 0x11f, min_res5);
> + if (min_res5[0] != 719673
> + || min_res5[1] != 718525)
> + __builtin_abort ();
> +
> + unsigned int min_res6[2] = { 719672, 718524 };
> + min_loop (x, 0x11f, min_res6);
> + if (min_res6[0] != 719672
> + || min_res6[1] != 718524)
> + __builtin_abort ();
> +
> + unsigned int and_res1[2] = { 11, 22 };
> + and_loop (x, 0, and_res1);
> + if (and_res1[0] != 11
> + || and_res1[1] != 22)
> + __builtin_abort ();
> +
> + unsigned int and_res2[2] = { 0xf5cff, 0xf78ff };
> + and_loop (x, 11, and_res2);
> + if (and_res2[0] != 0xf5c01
> + || and_res2[1] != 0xf7801)
> + __builtin_abort ();
> +
> + unsigned int and_res3[2] = { 0x7efff, 0xecfff };
> + and_loop (x, 0x40, and_res3);
> + if (and_res3[0] != 0x7c001
> + || and_res3[1] != 0xe8001)
> + __builtin_abort ();
> +
> + unsigned int and_res4[2] = { 0xffffff, 0xffffff };
> + and_loop (x, 0x4f, and_res4);
> + if (and_res4[0] != 0xf8001
> + || and_res4[1] != 0xf8001)
> + __builtin_abort ();
> +
> + return 0;
> +}
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/reduc_15.c
> b/gcc/testsuite/gcc.target/aarch64/sve/reduc_15.c
> new file mode 100644
> index 00000000000..15b1ade30e2
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/reduc_15.c
> @@ -0,0 +1,16 @@
> +/* { dg-options "-O3 --param vect-partial-vector-usage=1" } */
> +
> +int __attribute__((noipa))
> +add_loop (int *x, int n, int res)
> +{
> + for (int i = 0; i < n; ++i)
> + {
> + res += x[i * 2];
> + res += x[i * 2 + 1];
> + }
> + return res;
> +}
> +
> +/* { dg-final { scan-assembler-times {\tadd\tz[0-9]+\.s, z[0-9]+\.s,
> z[0-9]+\.s\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tadd\tz[0-9]+\.s, p[0-7]/m,
> z[0-9]+\.s, z[0-9]+\.s\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tuaddv\t} 1 } } */
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/reduc_15_run.c
> b/gcc/testsuite/gcc.target/aarch64/sve/reduc_15_run.c
> new file mode 100644
> index 00000000000..3207fce5be3
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/reduc_15_run.c
> @@ -0,0 +1,22 @@
> +/* { dg-do run { target aarch64_sve256_hw } } */
> +/* { dg-options "-O3 -msve-vector-bits=256 --param
> vect-partial-vector-usage=1" } */
> +
> +#define N 0x1100
> +
> +#include "reduc_15.c"
> +
> +int
> +main (void)
> +{
> + int x[N];
> + for (int i = 0; i < N; ++i)
> + x[i] = ((i + 1) * (i + 2)) & 0xfffff;
> +
> + if (add_loop (x, 0, 33) != 33
> + || add_loop (x, 11, 30) != 4078
> + || add_loop (x, 0x100, 45) != 45001773
> + || add_loop (x, 0x11f, 300) != 63369900)
> + __builtin_abort ();
> +
> + return 0;
> +}
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/reduc_9.c
> b/gcc/testsuite/gcc.target/aarch64/sve/reduc_9.c
> new file mode 100644
> index 00000000000..b839821d6bb
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/reduc_9.c
> @@ -0,0 +1,77 @@
> +/* { dg-options "-O3 -msve-vector-bits=256 --param
> vect-partial-vector-usage=1" } */
> +
> +unsigned short __attribute__((noipa))
> +add_loop (unsigned short *x)
> +{
> + unsigned short res = 0;
> + for (int i = 0; i < 0xfff; ++i)
> + res += x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +min_loop (unsigned short *x)
> +{
> + unsigned short res = ~0;
> + for (int i = 0; i < 0xfff; ++i)
> + res = res < x[i] ? res : x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +max_loop (unsigned short *x)
> +{
> + unsigned short res = 0;
> + for (int i = 0; i < 0xfff; ++i)
> + res = res > x[i] ? res : x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +and_loop (unsigned short *x)
> +{
> + unsigned short res = ~0;
> + for (int i = 0; i < 0xfff; ++i)
> + res &= x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +or_loop (unsigned short *x)
> +{
> + unsigned short res = 0;
> + for (int i = 0; i < 0xfff; ++i)
> + res |= x[i];
> + return res;
> +}
> +
> +unsigned short __attribute__((noipa))
> +eor_loop (unsigned short *x)
> +{
> + unsigned short res = 0;
> + for (int i = 0; i < 0xfff; ++i)
> + res ^= x[i];
> + return res;
> +}
> +
> +/* { dg-final { scan-assembler-times {\tadd\tz[0-9]+\.h, z[0-9]+\.h,
> z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tadd\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tuaddv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\tumin\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 2 } } */
> +/* { dg-final { scan-assembler-times {\tuminv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\tumax\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 2 } } */
> +/* { dg-final { scan-assembler-times {\tumaxv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\tand\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tand\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\tandv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\torr\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\torr\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\torv\t} 1 } } */
> +
> +/* { dg-final { scan-assembler-times {\teor\tz[0-9]+\.d, z[0-9]+\.d,
> z[0-9]+\.d\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\teor\tz[0-9]+\.h, p[0-7]/m,
> z[0-9]+\.h, z[0-9]+\.h\n} 1 } } */
> +/* { dg-final { scan-assembler-times {\teorv\t} 1 } } */
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/reduc_9_run.c
> b/gcc/testsuite/gcc.target/aarch64/sve/reduc_9_run.c
> new file mode 100644
> index 00000000000..aa248f53eaa
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/reduc_9_run.c
> @@ -0,0 +1,29 @@
> +/* { dg-do run { target aarch64_sve256_hw } } */
> +/* { dg-options "-O3 -msve-vector-bits=256 --param
> vect-partial-vector-usage=1" } */
> +
> +#define N 0x1100
> +
> +#include "reduc_9.c"
> +
> +int
> +main (void)
> +{
> + unsigned short x[N];
> + for (int i = 0; i < N; ++i)
> + x[i] = (i + 1) * (i + 2);
> +
> + if (add_loop (x) != 20480
> + || max_loop (x) != 65504
> + || or_loop (x) != 0xfffe
> + || eor_loop (x) != 0xa000)
> + __builtin_abort ();
> +
> + for (int i = 0; i < N; ++i)
> + x[i] = ~x[i];
> +
> + if (min_loop (x) != 31
> + || and_loop (x) != 1)
> + __builtin_abort ();
> +
> + return 0;
> +}
