Tamar Christina <tamar.christ...@arm.com> writes:
>> > + }
>> > + cgraph_simd_clone *sc = node->simdclone;
>> > +
>> > + for (unsigned i = 0; i < sc->nargs; ++i)
>> > + {
>> > + bool is_mask = false;
>> > + tree type;
>> > + switch (sc->args[i].arg_type)
>> > + {
>> > + case SIMD_CLONE_ARG_TYPE_MASK:
>> > + is_mask = true;
>> > + gcc_fallthrough ();
>> > + case SIMD_CLONE_ARG_TYPE_VECTOR:
>> > + case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP:
>> > + case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP:
>> > + type = sc->args[i].vector_type;
>> > + gcc_assert (VECTOR_TYPE_P (type));
>> > + if (node->simdclone->vecsize_mangle == 's')
>> > + type = simd_clone_adjust_sve_vector_type (type, is_mask,
>> > + sc->simdlen);
>> > + else if (is_mask)
>> > + type = truth_type_for (type);
>>
>> Sorry, I have a horrible feeling I knew this once and have forgotten,
>> but: why do we need to this for non-SVE, when we didn't before?
>>
>
> I don't think we do either. For Adv. SIMD the truth type is the same as
> the vector type anyway so this is a no-op. Removed.
>
>>
>> I should have noticed this last time, sorry, but we don't seem to have
>> any coverage for the linear cases above. Maybe that comes in a later
>> patch though.
>>
>
> No, Though I tried to make some examples of linear cases.
> On C the vectorizer just ignores the pragma.
Like we discussed off-list, the above cases only trigger for references
that have a val modifier (or no modifier), such as, to steal a variant
of your example:
#define TYPE int
#pragma omp declare simd linear(n)
extern TYPE __attribute__ ((const)) fn0 (TYPE x, TYPE &n);
void test_fn0 (TYPE *__restrict a, TYPE *__restrict b, TYPE n)
{
TYPE i;
#pragma omp simd linear(i)
for (i = 0; i < n; ++i)
a[i] += fn0 (b[i], i);
}
And as you pointed out, the vectoriser doesn't handle those yet:
case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP:
/* FORNOW */
i = -1;
break;
Since this series is only concerned with using externally-provided
simd implementations for vectorisation, the aarch64 cases above:
case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP:
will be nugatory until the FORNOW is fixed. But I went through some
#pragma omp declare simd examples locally, dumping the types that
aarch64_simd_clone_adjust produces, and I agree that they seem to be
correct. In particular, the elements always seem to be uintptr_ts, as
expected.
So while it seems unfortunate to be committing code that has no
test coverage, let's go with it.
> In C++ with a linear reference we fail to vectorize because we hit the
> Safe_len being an int and VF being a poly thing again and so we bail out.
>
> I did manage to create a testcase that generates an ICE, but that's due to
> an existing bug in the vectorizer with how it registers masks.
>
> Since that's an existing bug I'm hoping that's not a blocker for this series.
The two linear cases that the vectoriser can handle -- non-reference
parameters, and reference parameters with ref qualifiers -- wouldn't be
affected by the code above. The following ref case works for me locally:
#define TYPE long
#pragma omp declare simd linear(ref(n))
extern TYPE __attribute__ ((const)) fn0 (TYPE x, TYPE &n);
void test_fn0 (TYPE *__restrict a, TYPE *__restrict b, TYPE n)
{
TYPE i;
#pragma omp simd linear(i)
for (i = 0; i < n; ++i)
a[i] += fn0 (b[i], i);
}
as long as I disable:
/* If the function isn't const, only allow it in simd loops where user
has asserted that at least nunits consecutive iterations can be
performed using SIMD instructions. */
if ((loop == NULL || maybe_lt ((unsigned) loop->safelen, nunits))
&& gimple_vuse (stmt))
return false;
although the generated code is awful :) (Part of the problem seems
to be that .GOMP_SIMD_LANE as vectorised as ints rather than longs,
meaning that we need a doubled VF. Advanced SIMD behaves similarly,
although the code quality is better.)
If I change:
#define TYPE long
to:
#define TYPE int
then I get an ICE in vect_get_loop_mask, which might be what you hit,
or might be different. I think this case might be a backend bug.
lane_size includes:
/* For non map-to-vector types that are pointers we use the element type it
points to. */
if (POINTER_TYPE_P (type))
switch (clone_arg_type)
{
default:
break;
case SIMD_CLONE_ARG_TYPE_UNIFORM:
case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP:
type = TREE_TYPE (type);
break;
}
which doesn't cover this case. The code therefore continues to base the
lane size on the pointer size rather than the target of the reference
(i.e. 64 bits rather than 32 bits).
The VFABI says:
--------------------------------------------------------------------------
1. If ``<P>`` is an input parameter such that:
1. ``<P>`` is a ``uniform`` value, or
2. ``<P>`` is a ``linear`` value and not a reference marked with
``val`` or no linear modifiers,
then ``MTV(P)`` is ``false``.
--------------------------------------------------------------------------
where (2.) means that:
case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP:
should be treated as MTV(P) == false. It then says:
--------------------------------------------------------------------------
We then define the `Lane Size of P`, or ``LS(P)``, as follows.
1. If ``MTV(P)`` is ``false`` and ``P`` is a pointer or reference to
some type ``T`` for which ``PBV(T)`` is ``true``, ``LS(P) =
sizeof(T)``.
--------------------------------------------------------------------------
So I think the lane_size switch should be:
/* For non map-to-vector types that are pointers we use the element type it
points to. */
if (POINTER_TYPE_P (type))
switch (clone_arg_type)
{
default:
break;
case SIMD_CLONE_ARG_TYPE_UNIFORM:
case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP:
case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP:
type = TREE_TYPE (type);
break;
}
Making that change locally fixes the ICE for me.
> gcc/ChangeLog:
>
> PR target/96342
> * config/aarch64/aarch64-protos.h (add_sve_type_attribute): Declare.
> * config/aarch64/aarch64-sve-builtins.cc (add_sve_type_attribute): Make
> visibility global and support use for non_acle types.
> * config/aarch64/aarch64.cc
> (aarch64_simd_clone_compute_vecsize_and_simdlen): Create VLA simd clone
> when no simdlen is provided, according to ABI rules.
> (simd_clone_adjust_sve_vector_type): New helper function.
> (aarch64_simd_clone_adjust): Add '+sve' attribute to SVE simd clones
> and modify types to use SVE types.
> * omp-simd-clone.cc (simd_clone_mangle): Print 'x' for VLA simdlen.
> (simd_clone_adjust): Adapt safelen check to be compatible with VLA
> simdlen.
>
> gcc/testsuite/ChangeLog:
>
> PR target/96342
> * gcc.target/aarch64/declare-simd-2.c: Add SVE clone scan.
> * gcc.target/aarch64/vect-simd-clone-1.c: New test.
> * g++.target/aarch64/vect-simd-clone-1.c: New test.
OK, thanks.
Richard
> Co-authored-by: Victor Do Nascimento <victor.donascime...@arm.com>
> Co-authored-by: Tamar Christina <tamar.christ...@arm.com>
>
> Bootstrapped Regtested on aarch64-none-linux-gnu,
> arm-none-linux-gnueabihf, x86_64-pc-linux-gnu
> -m32, -m64 and no issues.
>
> Ok for master?
>
> Thanks,
> Tamar
>
> -- inline copy of patch --
>
> diff --git a/gcc/config/aarch64/aarch64-protos.h
> b/gcc/config/aarch64/aarch64-protos.h
> index
> bd17486e9128a21bd205ef1fb3ec3e323408ec59..7ab1316cf56850678d93b6fdb8d19eea18ad78f1
> 100644
> --- a/gcc/config/aarch64/aarch64-protos.h
> +++ b/gcc/config/aarch64/aarch64-protos.h
> @@ -1151,6 +1151,8 @@ namespace aarch64_sve {
> #ifdef GCC_TARGET_H
> bool verify_type_context (location_t, type_context_kind, const_tree, bool);
> #endif
> + void add_sve_type_attribute (tree, unsigned int, unsigned int,
> + const char *, const char *);
> }
>
> extern void aarch64_split_combinev16qi (rtx operands[3]);
> diff --git a/gcc/config/aarch64/aarch64-sve-builtins.cc
> b/gcc/config/aarch64/aarch64-sve-builtins.cc
> index
> 5acc56f99c65498cbf5593a9ee21540fa55098c2..e93c3a78e6d6c909f5de32ba8672503fc42b8d1c
> 100644
> --- a/gcc/config/aarch64/aarch64-sve-builtins.cc
> +++ b/gcc/config/aarch64/aarch64-sve-builtins.cc
> @@ -1032,15 +1032,18 @@ static GTY(()) hash_map<tree, registered_function *>
> *overload_names[2];
>
> /* Record that TYPE is an ABI-defined SVE type that contains NUM_ZR SVE
> vectors
> and NUM_PR SVE predicates. MANGLED_NAME, if nonnull, is the ABI-defined
> - mangling of the type. ACLE_NAME is the <arm_sve.h> name of the type. */
> -static void
> + mangling of the type. mangling of the type. ACLE_NAME is the <arm_sve.h>
> + name of the type, or null if <arm_sve.h> does not provide the type. */
> +void
> add_sve_type_attribute (tree type, unsigned int num_zr, unsigned int num_pr,
> const char *mangled_name, const char *acle_name)
> {
> tree mangled_name_tree
> = (mangled_name ? get_identifier (mangled_name) : NULL_TREE);
> + tree acle_name_tree
> + = (acle_name ? get_identifier (acle_name) : NULL_TREE);
>
> - tree value = tree_cons (NULL_TREE, get_identifier (acle_name), NULL_TREE);
> + tree value = tree_cons (NULL_TREE, acle_name_tree, NULL_TREE);
> value = tree_cons (NULL_TREE, mangled_name_tree, value);
> value = tree_cons (NULL_TREE, size_int (num_pr), value);
> value = tree_cons (NULL_TREE, size_int (num_zr), value);
> diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
> index
> 77a2a6bfa3a3a6cd678ceb820d310f44cacfe581..de4c0a0783912b54ac35d7c818c24574b27a4ca0
> 100644
> --- a/gcc/config/aarch64/aarch64.cc
> +++ b/gcc/config/aarch64/aarch64.cc
> @@ -29323,7 +29323,7 @@ aarch64_simd_clone_compute_vecsize_and_simdlen
> (struct cgraph_node *node,
> int num, bool explicit_p)
> {
> tree t, ret_type;
> - unsigned int nds_elt_bits;
> + unsigned int nds_elt_bits, wds_elt_bits;
> unsigned HOST_WIDE_INT const_simdlen;
>
> if (!TARGET_SIMD)
> @@ -29368,10 +29368,14 @@ aarch64_simd_clone_compute_vecsize_and_simdlen
> (struct cgraph_node *node,
> if (TREE_CODE (ret_type) != VOID_TYPE)
> {
> nds_elt_bits = lane_size (SIMD_CLONE_ARG_TYPE_VECTOR, ret_type);
> + wds_elt_bits = nds_elt_bits;
> vec_elts.safe_push (std::make_pair (ret_type, nds_elt_bits));
> }
> else
> - nds_elt_bits = POINTER_SIZE;
> + {
> + nds_elt_bits = POINTER_SIZE;
> + wds_elt_bits = 0;
> + }
>
> int i;
> tree type_arg_types = TYPE_ARG_TYPES (TREE_TYPE (node->decl));
> @@ -29379,44 +29383,65 @@ aarch64_simd_clone_compute_vecsize_and_simdlen
> (struct cgraph_node *node,
> for (t = (decl_arg_p ? DECL_ARGUMENTS (node->decl) : type_arg_types), i =
> 0;
> t && t != void_list_node; t = TREE_CHAIN (t), i++)
> {
> - tree arg_type = decl_arg_p ? TREE_TYPE (t) : TREE_VALUE (t);
> + tree type = decl_arg_p ? TREE_TYPE (t) : TREE_VALUE (t);
> if (clonei->args[i].arg_type != SIMD_CLONE_ARG_TYPE_UNIFORM
> - && !supported_simd_type (arg_type))
> + && !supported_simd_type (type))
> {
> if (!explicit_p)
> ;
> - else if (COMPLEX_FLOAT_TYPE_P (ret_type))
> + else if (COMPLEX_FLOAT_TYPE_P (type))
> warning_at (DECL_SOURCE_LOCATION (node->decl), 0,
> "GCC does not currently support argument type %qT "
> - "for simd", arg_type);
> + "for simd", type);
> else
> warning_at (DECL_SOURCE_LOCATION (node->decl), 0,
> "unsupported argument type %qT for simd",
> - arg_type);
> + type);
> return 0;
> }
> - unsigned lane_bits = lane_size (clonei->args[i].arg_type, arg_type);
> + unsigned lane_bits = lane_size (clonei->args[i].arg_type, type);
> if (clonei->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR)
> - vec_elts.safe_push (std::make_pair (arg_type, lane_bits));
> + vec_elts.safe_push (std::make_pair (type, lane_bits));
> if (nds_elt_bits > lane_bits)
> nds_elt_bits = lane_bits;
> + if (wds_elt_bits < lane_bits)
> + wds_elt_bits = lane_bits;
> }
>
> - clonei->vecsize_mangle = 'n';
> + /* If we could not determine the WDS type from available parameters/return,
> + then fallback to using uintptr_t. */
> + if (wds_elt_bits == 0)
> + wds_elt_bits = POINTER_SIZE;
> +
> clonei->mask_mode = VOIDmode;
> poly_uint64 simdlen;
> - auto_vec<poly_uint64> simdlens (2);
> + typedef struct
> + {
> + poly_uint64 len;
> + char mangle;
> + } aarch64_clone_info;
> + auto_vec<aarch64_clone_info, 3> clones;
> +
> /* Keep track of the possible simdlens the clones of this function can
> have,
> and check them later to see if we support them. */
> if (known_eq (clonei->simdlen, 0U))
> {
> simdlen = exact_div (poly_uint64 (64), nds_elt_bits);
> if (maybe_ne (simdlen, 1U))
> - simdlens.safe_push (simdlen);
> - simdlens.safe_push (simdlen * 2);
> + clones.safe_push ({simdlen, 'n'});
> + clones.safe_push ({simdlen * 2, 'n'});
> + /* Only create an SVE simd clone if we aren't dealing with an
> unprototyped
> + function.
> + We have also disabled support for creating SVE simdclones for functions
> + with function bodies and any simdclones when -msve-vector-bits is used.
> + TODO: add support for these. */
> + if (prototype_p (TREE_TYPE (node->decl))
> + && !node->definition
> + && !aarch64_sve_vg.is_constant ())
> + clones.safe_push ({exact_div (BITS_PER_SVE_VECTOR, wds_elt_bits), 's'});
> }
> else
> - simdlens.safe_push (clonei->simdlen);
> + clones.safe_push ({clonei->simdlen, 'n'});
>
> clonei->vecsize_int = 0;
> clonei->vecsize_float = 0;
> @@ -29430,11 +29455,12 @@ aarch64_simd_clone_compute_vecsize_and_simdlen
> (struct cgraph_node *node,
> simdclone would cause a vector type to be larger than 128-bits, and
> reject
> such a clone. */
> unsigned j = 0;
> - while (j < simdlens.length ())
> + while (j < clones.length ())
> {
> bool remove_simdlen = false;
> for (auto elt : vec_elts)
> - if (known_gt (simdlens[j] * elt.second, 128U))
> + if (clones[j].mangle == 'n'
> + && known_gt (clones[j].len * elt.second, 128U))
> {
> /* Don't issue a warning for every simdclone when there is no
> specific simdlen clause. */
> @@ -29442,18 +29468,17 @@ aarch64_simd_clone_compute_vecsize_and_simdlen
> (struct cgraph_node *node,
> warning_at (DECL_SOURCE_LOCATION (node->decl), 0,
> "GCC does not currently support simdlen %wd for "
> "type %qT",
> - constant_lower_bound (simdlens[j]), elt.first);
> + constant_lower_bound (clones[j].len), elt.first);
> remove_simdlen = true;
> break;
> }
> if (remove_simdlen)
> - simdlens.ordered_remove (j);
> + clones.ordered_remove (j);
> else
> j++;
> }
>
> -
> - int count = simdlens.length ();
> + int count = clones.length ();
> if (count == 0)
> {
> if (explicit_p && known_eq (clonei->simdlen, 0U))
> @@ -29470,21 +29495,112 @@ aarch64_simd_clone_compute_vecsize_and_simdlen
> (struct cgraph_node *node,
> }
>
> gcc_assert (num < count);
> - clonei->simdlen = simdlens[num];
> + clonei->simdlen = clones[num].len;
> + clonei->vecsize_mangle = clones[num].mangle;
> + /* SVE simdclones always have a Mask, so set inbranch to 1. */
> + if (clonei->vecsize_mangle == 's')
> + clonei->inbranch = 1;
> return count;
> }
>
> -/* Implement TARGET_SIMD_CLONE_ADJUST. */
> +/* Helper function to adjust an SVE vector type of an SVE simd clone.
> Returns
> + an SVE vector type based on the element type of the vector TYPE, with
> SIMDLEN
> + number of elements. If IS_MASK, returns an SVE mask type appropriate for
> use
> + with the SVE type it would otherwise return. */
> +
> +static tree
> +simd_clone_adjust_sve_vector_type (tree type, bool is_mask, poly_uint64
> simdlen)
> +{
> + unsigned int num_zr = 0;
> + unsigned int num_pr = 0;
> + machine_mode vector_mode;
> + type = TREE_TYPE (type);
> + scalar_mode scalar_m = SCALAR_TYPE_MODE (type);
> + vector_mode = aarch64_sve_data_mode (scalar_m, simdlen).require ();
> + type = build_vector_type_for_mode (type, vector_mode);
> + if (is_mask)
> + {
> + type = truth_type_for (type);
> + num_pr = 1;
> + }
> + else
> + num_zr = 1;
> +
> + /* We create new types here with the SVE type attribute instead of using
> ACLE
> + types as we need to support unpacked vectors which aren't available as
> + ACLE SVE types. */
> +
> + /* ??? This creates anonymous "SVE type" attributes for all types,
> + even those that correspond to <arm_sve.h> types. This affects type
> + compatibility in C/C++, but not in gimple. (Gimple type equivalence
> + is instead decided by TARGET_COMPATIBLE_VECTOR_TYPES_P.)
>
> + Thus a C/C++ definition of the implementation function will have a
> + different function type from the declaration that this code creates.
> + However, it doesn't seem worth trying to fix that until we have a
> + way of handling implementations that operate on unpacked types. */
> + type = build_distinct_type_copy (type);
> + aarch64_sve::add_sve_type_attribute (type, num_zr, num_pr, NULL, NULL);
> + return type;
> +}
> +
> +/* Implement TARGET_SIMD_CLONE_ADJUST. */
> static void
> aarch64_simd_clone_adjust (struct cgraph_node *node)
> {
> - /* Add aarch64_vector_pcs target attribute to SIMD clones so they
> - use the correct ABI. */
> -
> tree t = TREE_TYPE (node->decl);
> - TYPE_ATTRIBUTES (t) = make_attribute ("aarch64_vector_pcs", "default",
> - TYPE_ATTRIBUTES (t));
> +
> + if (node->simdclone->vecsize_mangle == 's')
> + {
> + /* This is additive and has no effect if SVE, or a superset thereof, is
> + already enabled. */
> + tree target = build_string (strlen ("+sve") + 1, "+sve");
> + if (!aarch64_option_valid_attribute_p (node->decl, NULL_TREE, target,
> 0))
> + gcc_unreachable ();
> + push_function_decl (node->decl);
> + }
> + else
> + {
> + /* Add aarch64_vector_pcs target attribute to SIMD clones so they
> + use the correct ABI. */
> + TYPE_ATTRIBUTES (t) = make_attribute ("aarch64_vector_pcs", "default",
> + TYPE_ATTRIBUTES (t));
> + }
> +
> + cgraph_simd_clone *sc = node->simdclone;
> +
> + for (unsigned i = 0; i < sc->nargs; ++i)
> + {
> + bool is_mask = false;
> + tree type;
> + switch (sc->args[i].arg_type)
> + {
> + case SIMD_CLONE_ARG_TYPE_MASK:
> + is_mask = true;
> + gcc_fallthrough ();
> + case SIMD_CLONE_ARG_TYPE_VECTOR:
> + case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP:
> + case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP:
> + type = sc->args[i].vector_type;
> + gcc_assert (VECTOR_TYPE_P (type));
> + if (node->simdclone->vecsize_mangle == 's')
> + type = simd_clone_adjust_sve_vector_type (type, is_mask,
> + sc->simdlen);
> + sc->args[i].vector_type = type;
> + break;
> + default:
> + continue;
> + }
> + }
> + if (node->simdclone->vecsize_mangle == 's')
> + {
> + tree ret_type = TREE_TYPE (t);
> + if (VECTOR_TYPE_P (ret_type))
> + TREE_TYPE (t)
> + = simd_clone_adjust_sve_vector_type (ret_type, false,
> + node->simdclone->simdlen);
> + pop_function_decl ();
> + }
> }
>
> /* Implement TARGET_SIMD_CLONE_USABLE. */
> @@ -29498,6 +29614,11 @@ aarch64_simd_clone_usable (struct cgraph_node *node,
> machine_mode vector_mode)
> if (!TARGET_SIMD || aarch64_sve_mode_p (vector_mode))
> return -1;
> return 0;
> + case 's':
> + if (!TARGET_SVE
> + || !aarch64_sve_mode_p (vector_mode))
> + return -1;
> + return 0;
> default:
> gcc_unreachable ();
> }
> diff --git a/gcc/omp-simd-clone.cc b/gcc/omp-simd-clone.cc
> index
> 864586207ee89269b5a2cf136487440212d59695..4be25539057251a318409e576e4bc43fc5fd4c40
> 100644
> --- a/gcc/omp-simd-clone.cc
> +++ b/gcc/omp-simd-clone.cc
> @@ -541,9 +541,12 @@ simd_clone_mangle (struct cgraph_node *node,
> pp_string (&pp, "_ZGV");
> pp_character (&pp, vecsize_mangle);
> pp_character (&pp, mask);
> - /* For now, simdlen is always constant, while variable simdlen pp 'n'. */
> - unsigned int len = simdlen.to_constant ();
> - pp_decimal_int (&pp, (len));
> +
> + unsigned HOST_WIDE_INT len;
> + if (simdlen.is_constant (&len))
> + pp_decimal_int (&pp, (int) (len));
> + else
> + pp_character (&pp, 'x');
>
> for (n = 0; n < clone_info->nargs; ++n)
> {
> @@ -1533,8 +1536,8 @@ simd_clone_adjust (struct cgraph_node *node)
> below). */
> loop = alloc_loop ();
> cfun->has_force_vectorize_loops = true;
> - /* For now, simlen is always constant. */
> - loop->safelen = node->simdclone->simdlen.to_constant ();
> + /* We can assert that safelen is the 'minimum' simdlen. */
> + loop->safelen = constant_lower_bound (node->simdclone->simdlen);
> loop->force_vectorize = true;
> loop->header = body_bb;
> }
> diff --git a/gcc/testsuite/g++.target/aarch64/vect-simd-clone-1.C
> b/gcc/testsuite/g++.target/aarch64/vect-simd-clone-1.C
> new file mode 100644
> index
> 0000000000000000000000000000000000000000..90febeca16e5126ed86f2f472b66c3bc3533c773
> --- /dev/null
> +++ b/gcc/testsuite/g++.target/aarch64/vect-simd-clone-1.C
> @@ -0,0 +1,88 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-O3 -march=armv8-a" } */
> +
> +/* Ensure correct creation of SVE Vector-length agnostic (VLA SVE) vector
> + function calls from scalar versions in accordance with the Vector
> Function
> + Application Binary Interface Specification for AArch64 (AAVPCS).
> +
> + We check for correctness in:
> + - Vector function name mangling, with the grammar:
> +
> + vector name := prefix "_" name
> + prefix := "_ZGV" isa mask <len> <parameters>
> +
> + Whereby:
> + - <isa> := "s" for SVE
> + - <mask> := "M" for Mask
> + - <len> := "x" for VLA SVE
> +
> + resulting in:
> + <prefix> := "_ZGVsMx" <parameters>
> +
> + with each vector parameter contributing a "v" to the prefix.
> +
> + - Parameter and return value mapping:
> + - Unless marked with uniform or linear OpenMP clauses, parameters and
> + return values are expected to map to vectors.
> + - Where the lane-size of a parameter is less than the widest data size
> + for a given function, the resulting vector should be unpacked and
> + populated via extending loads.
> +
> + - Finally, we also make sure we can correctly generate calls to the same
> + function, differing only in the target architecture (i.e. SVE vs SIMD),
> + ensuring that each call points to the correctly-mangled vector function
> + and employs the correct ABI. For example, for `fn' we may expect:
> +
> + for #pragma GCC target("+sve"): _ZGVsMxvv_fn
> + for #pragma GCC target("+simd): _ZGVnN4vv_fn */
> +
> +#pragma GCC target ("+sve")
> +/* { dg-final { scan-assembler {\s+_ZGVsMxv__Z3fn0i\n} } } */
> +extern int __attribute__ ((simd, const)) fn0 (int);
> +void test_fn0 (int *a, int *b, int n)
> +{
> + for (int i = 0; i < n; ++i)
> + a[i] += fn0 (b[i]);
> +}
> +
> +/* { dg-final { scan-assembler {\s+_ZGVsMxvv__Z3fn1si\n} } } */
> +extern int __attribute__ ((simd, const)) fn1 (short, int);
> +void test_fn1 (int *a, int *b, short *c, int n)
> +{
> + for (int i = 0; i < n; ++i)
> + a[i] = fn1 (c[i], b[i]);
> +}
> +
> +/* { dg-final { scan-assembler {\s+_ZGVsMxvv__Z3fn2si\n} } } */
> +extern short __attribute__ ((simd, const)) fn2 (short, int);
> +void test_fn2 (short *a, int *b, short *c, int n)
> +{
> + for (int i = 0; i < n; ++i)
> + a[i] = fn2 (c[i], b[i]);
> +}
> +
> +/* { dg-final { scan-assembler {\s+_ZGVsMxvv__Z3fn3ic\n} } } */
> +extern char __attribute__ ((simd, const)) fn3 (int, char);
> +void test_fn3 (int *a, int *b, char *c, int n)
> +{
> + for (int i = 0; i < n; ++i)
> + a[i] = (int) (fn3 (b[i], c[i]) + c[i]);
> +}
> +
> +/* { dg-final { scan-assembler {\s+_ZGVsMxvv__Z3fn4is\n} } } */
> +extern short __attribute__ ((simd, const)) fn4 (int, short);
> +void test_fn4 (int *a, int *b, short *c, int n)
> +{
> + for (int i = 0; i < n; ++i)
> + a[i] = (int) (fn4 (b[i], c[i]) + c[i]);
> +}
> +
> +#pragma GCC reset_options
> +#pragma GCC target ("+simd")
> +/* { dg-final { scan-assembler {\s+_ZGVsMxvv__Z3fn4is\n} } } */
> +extern short __attribute__ ((simd, const)) fn4 (int, short);
> +void test_fn5 (int *a, int *b, short *c, int n)
> +{
> + for (int i = 0; i < n; ++i)
> + a[i] = (int) (fn4 (b[i], c[i]) + c[i]);
> +}
> diff --git a/gcc/testsuite/gcc.target/aarch64/declare-simd-2.c
> b/gcc/testsuite/gcc.target/aarch64/declare-simd-2.c
> index
> e2e80f0c663dcc182b8cc48b0453558e794f4085..2f4d3a866e55018b8ac8b483b8c33db862a57071
> 100644
> --- a/gcc/testsuite/gcc.target/aarch64/declare-simd-2.c
> +++ b/gcc/testsuite/gcc.target/aarch64/declare-simd-2.c
> @@ -43,6 +43,7 @@ float f04 (double a)
> }
> /* { dg-final { scan-assembler {_ZGVnN2v_f04:} } } */
> /* { dg-final { scan-assembler {_ZGVnM2v_f04:} } } */
> +/* { dg-final { scan-assembler-not {_ZGVs[0-9a-z]*_f04:} } } */
>
> #pragma omp declare simd uniform(a) linear (b)
> void f05 (short a, short *b, short c)
> diff --git a/gcc/testsuite/gcc.target/aarch64/vect-simd-clone-1.c
> b/gcc/testsuite/gcc.target/aarch64/vect-simd-clone-1.c
> new file mode 100644
> index
> 0000000000000000000000000000000000000000..0d8f497644ca119529a5778b81ae8a78948306e4
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/vect-simd-clone-1.c
> @@ -0,0 +1,89 @@
> +/* { dg-do compile } */
> +/* { dg-options "-std=c99" } */
> +/* { dg-additional-options "-O3 -march=armv8-a" } */
> +
> +/* Ensure correct creation of SVE Vector-length agnostic (VLA SVE) vector
> + function calls from scalar versions in accordance with the Vector
> Function
> + Application Binary Interface Specification for AArch64 (AAVPCS).
> +
> + We check for correctness in:
> + - Vector function name mangling, with the grammar:
> +
> + vector name := prefix "_" name
> + prefix := "_ZGV" isa mask <len> <parameters>
> +
> + Whereby:
> + - <isa> := "s" for SVE
> + - <mask> := "M" for Mask
> + - <len> := "x" for VLA SVE
> +
> + resulting in:
> + <prefix> := "_ZGVsMx" <parameters>
> +
> + with each vector parameter contributing a "v" to the prefix.
> +
> + - Parameter and return value mapping:
> + - Unless marked with uniform or linear OpenMP clauses, parameters and
> + return values are expected to map to vectors.
> + - Where the lane-size of a parameter is less than the widest data size
> + for a given function, the resulting vector should be unpacked and
> + populated via extending loads.
> +
> + - Finally, we also make sure we can correctly generate calls to the same
> + function, differing only in the target architecture (i.e. SVE vs SIMD),
> + ensuring that each call points to the correctly-mangled vector function
> + and employs the correct ABI. For example, for `fn' we may expect:
> +
> + for #pragma GCC target("+sve"): _ZGVsMxvv_fn
> + for #pragma GCC target("+simd): _ZGVnN4vv_fn */
> +
> +#pragma GCC target ("+sve")
> +/* { dg-final { scan-assembler {\s+_ZGVsMxv_fn0\n} } } */
> +extern int __attribute__ ((simd, const)) fn0 (int);
> +void test_fn0 (int *a, int *b, int n)
> +{
> + for (int i = 0; i < n; ++i)
> + a[i] += fn0 (b[i]);
> +}
> +
> +/* { dg-final { scan-assembler {\s+_ZGVsMxvv_fn1\n} } } */
> +extern int __attribute__ ((simd, const)) fn1 (short, int);
> +void test_fn1 (int *a, int *b, short *c, int n)
> +{
> + for (int i = 0; i < n; ++i)
> + a[i] = fn1 (c[i], b[i]);
> +}
> +
> +/* { dg-final { scan-assembler {\s+_ZGVsMxvv_fn2\n} } } */
> +extern short __attribute__ ((simd, const)) fn2 (short, int);
> +void test_fn2 (short *a, int *b, short *c, int n)
> +{
> + for (int i = 0; i < n; ++i)
> + a[i] = fn2 (c[i], b[i]);
> +}
> +
> +/* { dg-final { scan-assembler {\s+_ZGVsMxvv_fn3\n} } } */
> +extern char __attribute__ ((simd, const)) fn3 (int, char);
> +void test_fn3 (int *a, int *b, char *c, int n)
> +{
> + for (int i = 0; i < n; ++i)
> + a[i] = (int) (fn3 (b[i], c[i]) + c[i]);
> +}
> +
> +/* { dg-final { scan-assembler {\s+_ZGVsMxvv_fn4\n} } } */
> +extern short __attribute__ ((simd, const)) fn4 (int, short);
> +void test_fn4 (int *a, int *b, short *c, int n)
> +{
> + for (int i = 0; i < n; ++i)
> + a[i] = (int) (fn4 (b[i], c[i]) + c[i]);
> +}
> +
> +#pragma GCC reset_options
> +#pragma GCC target ("+simd")
> +/* { dg-final { scan-assembler {\s+_ZGVnN4vv_fn4\n} } } */
> +extern short __attribute__ ((simd, const)) fn4 (int, short);
> +void test_fn5 (int *a, int *b, short *c, int n)
> +{
> + for (int i = 0; i < n; ++i)
> + a[i] = (int) (fn4 (b[i], c[i]) + c[i]);
> +}
