platform/upstream/gcc.git
5 years agoAllow mixed vector sizes within a single vectorised stmt
Richard Sandiford [Thu, 14 Nov 2019 15:16:40 +0000 (15:16 +0000)]
Allow mixed vector sizes within a single vectorised stmt

Although a previous patch allowed mixed vector sizes within a vector
region, we generally still required equal vector sizes within a vector
stmt.  Specifically, vect_get_vector_types_for_stmt computes two vector
types: the vector type corresponding to STMT_VINFO_VECTYPE and the
vector type that determines the minimum vectorisation factor for the
stmt ("nunits_vectype").  It then required these two types to be
the same size.

There doesn't seem to be any need for that restriction though.  AFAICT,
all vectorizable_* functions either do their own compatibility checks
or don't need to do them (because gimple guarantees that the scalar
types are compatible).

It should always be the case that nunits_vectype has at least as many
elements as the other vectype, but that's something we can assert for.

I couldn't resist a couple of other tweaks while there:

- there's no need to compute nunits_vectype if its element type is
  the same as STMT_VINFO_VECTYPE's.

- it's useful to distinguish the nunits_vectype from the main vectype
  in dump messages

- when reusing the existing STMT_VINFO_VECTYPE, it's useful to say so
  in the dump, and say what the type is

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vect-stmts.c (vect_get_vector_types_for_stmt): Don't
require vectype and nunits_vectype to have the same size;
instead assert that nunits_vectype has at least as many
elements as vectype.  Don't compute a separate nunits_vectype
if the scalar type is obviously the same as vectype's.
Tweak dump messages.

From-SVN: r278244

5 years ago[AArch64] Support vectorising with multiple vector sizes
Richard Sandiford [Thu, 14 Nov 2019 15:15:34 +0000 (15:15 +0000)]
[AArch64] Support vectorising with multiple vector sizes

This patch makes the vectoriser try mixtures of 64-bit and 128-bit
vector modes on AArch64.  It fixes some existing XFAILs and allows
kernel 24 from the Livermore Loops test to be vectorised (by using
a mixture of V2DF and V2SI).

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64.c (aarch64_vectorize_related_mode): New
function.
(aarch64_autovectorize_vector_modes): Also add V4HImode and V2SImode.
(TARGET_VECTORIZE_RELATED_MODE): Define.

gcc/testsuite/
* gcc.dg/vect/vect-outer-4f.c: Expect the test to pass on aarch64
targets.
* gcc.dg/vect/vect-outer-4g.c: Likewise.
* gcc.dg/vect/vect-outer-4k.c: Likewise.
* gcc.dg/vect/vect-outer-4l.c: Likewise.
* gfortran.dg/vect/vect-8.f90: Expect kernel 24 to be vectorized
for aarch64.
* gcc.target/aarch64/vect_mixed_sizes_1.c: New test.
* gcc.target/aarch64/vect_mixed_sizes_2.c: Likewise.
* gcc.target/aarch64/vect_mixed_sizes_3.c: Likewise.
* gcc.target/aarch64/vect_mixed_sizes_4.c: Likewise.

From-SVN: r278243

5 years agoAvoid retrying with the same vector modes
Richard Sandiford [Thu, 14 Nov 2019 15:14:33 +0000 (15:14 +0000)]
Avoid retrying with the same vector modes

A later patch makes the AArch64 port add four entries to
autovectorize_vector_modes.  Each entry describes a different
vector mode assignment for vector code that mixes 8-bit, 16-bit,
32-bit and 64-bit elements.  But if (as usual) the vector code has
fewer element sizes than that, we could end up trying the same
combination of vector modes multiple times.  This patch adds a
check to prevent that.

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vectorizer.h (vec_info::mode_set): New typedef.
(vec_info::used_vector_mode): New member variable.
(vect_chooses_same_modes_p): Declare.
* tree-vect-stmts.c (get_vectype_for_scalar_type): Record each
chosen vector mode in vec_info::used_vector_mode.
(vect_chooses_same_modes_p): New function.
* tree-vect-loop.c (vect_analyze_loop): Use it to avoid trying
the same vector statements multiple times.
* tree-vect-slp.c (vect_slp_bb_region): Likewise.

From-SVN: r278242

5 years agoSupport vectorisation with mixed vector sizes
Richard Sandiford [Thu, 14 Nov 2019 15:12:58 +0000 (15:12 +0000)]
Support vectorisation with mixed vector sizes

After previous patches, it's now possible to make the vectoriser
support multiple vector sizes in the same vector region, using
related_vector_mode to pick the right vector mode for a given
element mode.  No port yet takes advantage of this, but I have
a follow-on patch for AArch64.

This patch also seemed like a good opportunity to add some more dump
messages: one to make it clear which vector size/mode was being used
when analysis passed or failed, and another to say when we've decided
to skip a redundant vector size/mode.

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* machmode.h (opt_machine_mode::operator==): New function.
(opt_machine_mode::operator!=): Likewise.
* tree-vectorizer.h (vec_info::vector_mode): Update comment.
(get_related_vectype_for_scalar_type): Delete.
(get_vectype_for_scalar_type_and_size): Declare.
* tree-vect-slp.c (vect_slp_bb_region): Print dump messages to say
whether analysis passed or failed, and with what vector modes.
Use related_vector_mode to check whether trying a particular
vector mode would be redundant with the autodetected mode,
and print a dump message if we decide to skip it.
* tree-vect-loop.c (vect_analyze_loop): Likewise.
(vect_create_epilog_for_reduction): Use
get_related_vectype_for_scalar_type instead of
get_vectype_for_scalar_type_and_size.
* tree-vect-stmts.c (get_vectype_for_scalar_type_and_size): Replace
with...
(get_related_vectype_for_scalar_type): ...this new function.
Take a starting/"prevailing" vector mode rather than a vector size.
Take an optional nunits argument, with the same meaning as for
related_vector_mode.  Use related_vector_mode when not
auto-detecting a mode, falling back to mode_for_vector if no
target mode exists.
(get_vectype_for_scalar_type): Update accordingly.
(get_same_sized_vectype): Likewise.
* tree-vectorizer.c (get_vec_alignment_for_array_type): Likewise.

From-SVN: r278240

5 years agoRequire equal type sizes for vectorised calls
Richard Sandiford [Thu, 14 Nov 2019 15:09:24 +0000 (15:09 +0000)]
Require equal type sizes for vectorised calls

As explained in the comment, vectorizable_call needs more work to
support mixtures of sizes.  This avoids testsuite fallout for
later SVE patches.

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vect-stmts.c (vectorizable_call): Require the types
to have the same size.

From-SVN: r278239

5 years agoMake less use of get_same_sized_vectype
Richard Sandiford [Thu, 14 Nov 2019 15:06:34 +0000 (15:06 +0000)]
Make less use of get_same_sized_vectype

Some callers of get_same_sized_vectype were dealing with operands that
are constant or defined externally, and so have no STMT_VINFO_VECTYPE
available.  Under the current model, using get_same_sized_vectype for
that case is equivalent to using get_vectype_for_scalar_type, since
get_vectype_for_scalar_type always returns vectors of the same size,
once a size is fixed.

Using get_vectype_for_scalar_type is arguably more obvious though:
if we're using the same scalar type as we would for internal
definitions, we should use the same vector type too.  (Constant and
external definitions sometimes let us change the original scalar type
to a "nicer" scalar type, but that isn't what's happening here.)

This is a prerequisite to supporting multiple vector sizes in the same
vec_info.

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vect-stmts.c (vectorizable_call): If an operand is
constant or external, use get_vectype_for_scalar_type
rather than get_same_sized_vectype to get its vector type.
(vectorizable_conversion, vectorizable_shift): Likewise.
(vectorizable_operation): Likewise.

From-SVN: r278238

5 years agoReplace vec_info::vector_size with vec_info::vector_mode
Richard Sandiford [Thu, 14 Nov 2019 15:05:37 +0000 (15:05 +0000)]
Replace vec_info::vector_size with vec_info::vector_mode

This patch replaces vec_info::vector_size with vec_info::vector_mode,
but for now continues to use it as a way of specifying a single
vector size.  This makes it easier for later patches to use
related_vector_mode instead.

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vectorizer.h (vec_info::vector_size): Replace with...
(vec_info::vector_mode): ...this new field.
* tree-vect-loop.c (vect_update_vf_for_slp): Update accordingly.
(vect_analyze_loop, vect_transform_loop): Likewise.
* tree-vect-loop-manip.c (vect_do_peeling): Likewise.
* tree-vect-slp.c (can_duplicate_and_interleave_p): Likewise.
(vect_make_slp_decision, vect_slp_bb_region): Likewise.
* tree-vect-stmts.c (get_vectype_for_scalar_type): Likewise.
* tree-vectorizer.c (try_vectorize_loop_1): Likewise.

gcc/testsuite/
* gcc.dg/vect/vect-tail-nomask-1.c: Update expected epilogue
vectorization message.

From-SVN: r278237

5 years agoReplace autovectorize_vector_sizes with autovectorize_vector_modes
Richard Sandiford [Thu, 14 Nov 2019 15:03:17 +0000 (15:03 +0000)]
Replace autovectorize_vector_sizes with autovectorize_vector_modes

This is another patch in the series to remove the assumption that
all modes involved in vectorisation have to be the same size.
Rather than have the target provide a list of vector sizes,
it makes the target provide a list of vector "approaches",
with each approach represented by a mode.

A later patch will pass this mode to targetm.vectorize.related_mode
to get the vector mode for a given element mode.  Until then, the modes
simply act as an alternative way of specifying the vector size.

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* target.h (vector_sizes, auto_vector_sizes): Delete.
(vector_modes, auto_vector_modes): New typedefs.
* target.def (autovectorize_vector_sizes): Replace with...
(autovectorize_vector_modes): ...this new hook.
* doc/tm.texi.in (TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES):
Replace with...
(TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_MODES): ...this new hook.
* doc/tm.texi: Regenerate.
* targhooks.h (default_autovectorize_vector_sizes): Delete.
(default_autovectorize_vector_modes): New function.
* targhooks.c (default_autovectorize_vector_sizes): Delete.
(default_autovectorize_vector_modes): New function.
* omp-general.c (omp_max_vf): Use autovectorize_vector_modes instead
of autovectorize_vector_sizes.  Use the number of units in the mode
to calculate the maximum VF.
* omp-low.c (omp_clause_aligned_alignment): Use
autovectorize_vector_modes instead of autovectorize_vector_sizes.
Use a loop based on related_mode to iterate through all supported
vector modes for a given scalar mode.
* optabs-query.c (can_vec_mask_load_store_p): Use
autovectorize_vector_modes instead of autovectorize_vector_sizes.
* tree-vect-loop.c (vect_analyze_loop, vect_transform_loop): Likewise.
* tree-vect-slp.c (vect_slp_bb_region): Likewise.
* config/aarch64/aarch64.c (aarch64_autovectorize_vector_sizes):
Replace with...
(aarch64_autovectorize_vector_modes): ...this new function.
(TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES): Delete.
(TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_MODES): Define.
* config/arc/arc.c (arc_autovectorize_vector_sizes): Replace with...
(arc_autovectorize_vector_modes): ...this new function.
(TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES): Delete.
(TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_MODES): Define.
* config/arm/arm.c (arm_autovectorize_vector_sizes): Replace with...
(arm_autovectorize_vector_modes): ...this new function.
(TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES): Delete.
(TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_MODES): Define.
* config/i386/i386.c (ix86_autovectorize_vector_sizes): Replace with...
(ix86_autovectorize_vector_modes): ...this new function.
(TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES): Delete.
(TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_MODES): Define.
* config/mips/mips.c (mips_autovectorize_vector_sizes): Replace with...
(mips_autovectorize_vector_modes): ...this new function.
(TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES): Delete.
(TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_MODES): Define.

From-SVN: r278236

5 years agoUse consistent compatibility checks in vectorizable_shift
Richard Sandiford [Thu, 14 Nov 2019 14:58:21 +0000 (14:58 +0000)]
Use consistent compatibility checks in vectorizable_shift

The validation phase of vectorizable_shift used TYPE_MODE to check
whether the shift amount vector was compatible with the shifted vector:

      if ((op1_vectype == NULL_TREE
   || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype))
    && (!slp_node
        || SLP_TREE_DEF_TYPE
     (SLP_TREE_CHILDREN (slp_node)[1]) != vect_constant_def))

But the generation phase was stricter and required the element types to
be equivalent:

   && !useless_type_conversion_p (TREE_TYPE (vectype),
  TREE_TYPE (op1)))

This difference led to an ICE with a later patch.

The first condition seems a bit too lax given that the function
supports vect_worthwhile_without_simd_p, where two different vector
types could have the same integer mode.  But it seems too strict
to reject signed shifts by unsigned amounts or unsigned shifts by
signed amounts; verify_gimple_assign_binary is happy with those.

This patch therefore goes for a middle ground of checking both TYPE_MODE
and TYPE_VECTOR_SUBPARTS, using the same condition in both places.

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vect-stmts.c (vectorizable_shift): Check the number
of vector elements as well as the type mode when deciding
whether an op1_vectype is compatible.  Reuse the result of
this check when generating vector statements.

From-SVN: r278235

5 years agoUse build_vector_type_for_mode in get_vectype_for_scalar_type_and_size
Richard Sandiford [Thu, 14 Nov 2019 14:57:26 +0000 (14:57 +0000)]
Use build_vector_type_for_mode in get_vectype_for_scalar_type_and_size

Except for one case, get_vectype_for_scalar_type_and_size calculates
what the vector mode should be and then calls build_vector_type,
which recomputes the mode from scratch.  This patch makes it use
build_vector_type_for_mode instead.

The exception mentioned above is when preferred_simd_mode returns
an integer mode, which it does if no appropriate vector mode exists.
The integer mode in question is usually word_mode, although epiphany
can return a doubleword mode in some cases.

There's no guarantee that this integer mode is appropriate, since for
example the scalar type could be a float.  The traditional behaviour is
therefore to use the integer mode to determine a size only, and leave
mode_for_vector to pick the TYPE_MODE.  (Note that it can actually end
up picking a vector mode if the target defines a disabled vector mode.
We therefore still need to check TYPE_MODE after building the type.)

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vect-stmts.c (get_vectype_for_scalar_type_and_size): If
targetm.vectorize.preferred_simd_mode returns an integer mode,
use mode_for_vector to decide what the vector type's mode
should actually be.  Use build_vector_type_for_mode instead
of build_vector_type.

From-SVN: r278234

5 years agoPass the data vector mode to get_mask_mode
Richard Sandiford [Thu, 14 Nov 2019 14:55:12 +0000 (14:55 +0000)]
Pass the data vector mode to get_mask_mode

This patch passes the data vector mode to get_mask_mode, rather than its
size and nunits.  This is a bit simpler and allows targets to distinguish
between modes that happen to have the same size and number of elements.

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* target.def (get_mask_mode): Take a vector mode itself as argument,
instead of properties about the vector mode.
* doc/tm.texi: Regenerate.
* targhooks.h (default_get_mask_mode): Update to reflect new
get_mode_mask interface.
* targhooks.c (default_get_mask_mode): Likewise.  Use
related_int_vector_mode.
* optabs-query.c (can_vec_mask_load_store_p): Update call
to get_mask_mode.
* tree-vect-stmts.c (check_load_store_masking): Likewise, checking
first that the original mode really is a vector.
* tree.c (build_truth_vector_type_for): Likewise.
* config/aarch64/aarch64.c (aarch64_get_mask_mode): Update for new
get_mode_mask interface.
(aarch64_expand_sve_vcond): Update call accordingly.
* config/gcn/gcn.c (gcn_vectorize_get_mask_mode): Update for new
get_mode_mask interface.
* config/i386/i386.c (ix86_get_mask_mode): Likewise.

From-SVN: r278233

5 years agoRemove build_{same_sized_,}truth_vector_type
Richard Sandiford [Thu, 14 Nov 2019 14:49:36 +0000 (14:49 +0000)]
Remove build_{same_sized_,}truth_vector_type

build_same_sized_truth_vector_type was confusingly named, since for
SVE and AVX512 the returned vector isn't the same byte size (although
it does have the same number of elements).  What it really returns
is the "truth" vector type for a given data vector type.

The more general truth_type_for provides the same thing when passed
a vector and IMO has a more descriptive name, so this patch replaces
all uses of build_same_sized_truth_vector_type with that.  It does
the same for a call to build_truth_vector_type, leaving truth_type_for
itself as the only remaining caller.

It's then more natural to pass build_truth_vector_type the original
vector type rather than its size and nunits, especially since the
given size isn't the size of the returned vector.  This in turn allows
a future patch to simplify the interface of get_mask_mode.  Doing this
also fixes a bug in which truth_type_for would pass a size of zero for
BLKmode vector types.

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree.h (build_truth_vector_type): Delete.
(build_same_sized_truth_vector_type): Likewise.
* tree.c (build_truth_vector_type): Rename to...
(build_truth_vector_type_for): ...this.  Make static and take
a vector type as argument.
(truth_type_for): Update accordingly.
(build_same_sized_truth_vector_type): Delete.
* tree-vect-generic.c (expand_vector_divmod): Use truth_type_for
instead of build_same_sized_truth_vector_type.
* tree-vect-loop.c (vect_create_epilog_for_reduction): Likewise.
(vect_record_loop_mask, vect_get_loop_mask): Likewise.
* tree-vect-patterns.c (build_mask_conversion): Likeise.
* tree-vect-slp.c (vect_get_constant_vectors): Likewise.
* tree-vect-stmts.c (vect_get_vec_def_for_operand): Likewise.
(vect_build_gather_load_calls, vectorizable_call): Likewise.
(scan_store_can_perm_p, vectorizable_scan_store): Likewise.
(vectorizable_store, vectorizable_condition): Likewise.
(get_mask_type_for_scalar_type, get_same_sized_vectype): Likewise.
(vect_get_mask_type_for_stmt): Use truth_type_for instead of
build_truth_vector_type.
* config/aarch64/aarch64-sve-builtins.cc (gimple_folder::convert_pred):
Use truth_type_for instead of build_same_sized_truth_vector_type.
* config/rs6000/rs6000-call.c (fold_build_vec_cmp): Likewise.

gcc/c/
* c-typeck.c (build_conditional_expr): Use truth_type_for instead
of build_same_sized_truth_vector_type.
(build_vec_cmp): Likewise.

gcc/cp/
* call.c (build_conditional_expr_1): Use truth_type_for instead
of build_same_sized_truth_vector_type.
* typeck.c (build_vec_cmp): Likewise.

gcc/d/
* d-codegen.cc (build_boolop): Use truth_type_for instead of
build_same_sized_truth_vector_type.

From-SVN: r278232

5 years agoAdd build_truth_vector_type_for_mode
Richard Sandiford [Thu, 14 Nov 2019 14:45:49 +0000 (14:45 +0000)]
Add build_truth_vector_type_for_mode

Callers of vect_halve_mask_nunits and vect_double_mask_nunits
already know what mode the resulting vector type should have,
so we might as well create the vector type directly with that mode,
just like build_vector_type_for_mode lets us build normal vectors
with a known mode.  This avoids the current awkwardness of having
to recompute the mode starting from vec_info::vector_size, which
hard-codes the assumption that all vectors have to be the same size.

A later patch gets rid of build_truth_vector_type and
build_same_sized_truth_vector_type, so the net effect of the
series is to reduce the number of type functions by one.

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree.h (build_truth_vector_type_for_mode): Declare.
* tree.c (build_truth_vector_type_for_mode): New function,
split out from...
(build_truth_vector_type): ...here.
(build_opaque_vector_type): Fix head comment.
* tree-vectorizer.h (supportable_narrowing_operation): Remove
vec_info parameter.
(vect_halve_mask_nunits): Replace vec_info parameter with the
mode of the new vector.
(vect_double_mask_nunits): Likewise.
* tree-vect-loop.c (vect_halve_mask_nunits): Likewise.
(vect_double_mask_nunits): Likewise.
* tree-vect-loop-manip.c: Include insn-config.h, rtl.h and recog.h.
(vect_maybe_permute_loop_masks): Remove vinfo parameter.  Update call
to vect_halve_mask_nunits, getting the required mode from the unpack
patterns.
(vect_set_loop_condition_masked): Update call accordingly.
* tree-vect-stmts.c (supportable_narrowing_operation): Remove vec_info
parameter and update call to vect_double_mask_nunits.
(vectorizable_conversion): Update call accordingly.
(simple_integer_narrowing): Likewise.  Remove vec_info parameter.
(vectorizable_call): Update call accordingly.
(supportable_widening_operation): Update call to
vect_halve_mask_nunits.
* config/aarch64/aarch64-sve-builtins.cc (register_builtin_types):
Use build_truth_vector_type_mode instead of build_truth_vector_type.

From-SVN: r278231

5 years agoReplace mode_for_int_vector with related_int_vector_mode
Richard Sandiford [Thu, 14 Nov 2019 14:39:57 +0000 (14:39 +0000)]
Replace mode_for_int_vector with related_int_vector_mode

mode_for_int_vector, like mode_for_vector, can sometimes return
an integer mode or an unsupported vector mode.  But no callers
are interested in that case, and only want supported vector modes.
This patch therefore replaces mode_for_int_vector with
related_int_vector_mode, which gives the target a chance to pick
its preferred vector mode for the given element mode and size.

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* machmode.h (mode_for_int_vector): Delete.
(related_int_vector_mode): Declare.
* stor-layout.c (mode_for_int_vector): Delete.
(related_int_vector_mode): New function.
* optabs.c (expand_vec_perm_1): Use related_int_vector_mode
instead of mode_for_int_vector.
(expand_vec_perm_const): Likewise.
* config/aarch64/aarch64.c (aarch64_emit_approx_sqrt): Likewise.
(aarch64_evpc_sve_tbl): Likewise.
* config/s390/s390.c (s390_expand_vec_compare_cc): Likewise.
(s390_expand_vcond): Likewise.

From-SVN: r278230

5 years agoAdd a targetm.vectorize.related_mode hook
Richard Sandiford [Thu, 14 Nov 2019 14:36:26 +0000 (14:36 +0000)]
Add a targetm.vectorize.related_mode hook

This patch is the first of a series that tries to remove two
assumptions:

(1) that all vectors involved in vectorisation must be the same size

(2) that there is only one vector mode for a given element mode and
    number of elements

Relaxing (1) helps with targets that support multiple vector sizes or
that require the number of elements to stay the same.  E.g. if we're
vectorising code that operates on narrow and wide elements, and the
narrow elements use 64-bit vectors, then on AArch64 it would normally
be better to use 128-bit vectors rather than pairs of 64-bit vectors
for the wide elements.

Relaxing (2) makes it possible for -msve-vector-bits=128 to produce
fixed-length code for SVE.  It also allows unpacked/half-size SVE
vectors to work with -msve-vector-bits=256.

The patch adds a new hook that targets can use to control how we
move from one vector mode to another.  The hook takes a starting vector
mode, a new element mode, and (optionally) a new number of elements.
The flexibility needed for (1) comes in when the number of elements
isn't specified.

All callers in this patch specify the number of elements, but a later
vectoriser patch doesn't.

2019-11-14  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* target.def (related_mode): New hook.
* doc/tm.texi.in (TARGET_VECTORIZE_RELATED_MODE): New hook.
* doc/tm.texi: Regenerate.
* targhooks.h (default_vectorize_related_mode): Declare.
* targhooks.c (default_vectorize_related_mode): New function.
* machmode.h (related_vector_mode): Declare.
* stor-layout.c (related_vector_mode): New function.
* expmed.c (extract_bit_field_1): Use it instead of mode_for_vector.
* optabs-query.c (qimode_for_vec_perm): Likewise.
* tree-vect-stmts.c (get_group_load_store_type): Likewise.
(vectorizable_store, vectorizable_load): Likewise

From-SVN: r278229

5 years agoaarch64: Add testsuite checks for asm-flag
Richard Henderson [Thu, 14 Nov 2019 13:45:01 +0000 (13:45 +0000)]
aarch64: Add testsuite checks for asm-flag

Inspired by the tests in gcc.target/i386.  Testing code generation,
diagnostics, and execution.

* gcc.target/aarch64/asm-flag-1.c: New test.
* gcc.target/aarch64/asm-flag-3.c: New test.
* gcc.target/aarch64/asm-flag-5.c: New test.
* gcc.target/aarch64/asm-flag-6.c: New test.

From-SVN: r278228

5 years agoarm: Add testsuite checks for asm-flag
Richard Henderson [Thu, 14 Nov 2019 13:44:48 +0000 (13:44 +0000)]
arm: Add testsuite checks for asm-flag

Inspired by the tests in gcc.target/i386.  Testing code generation,
diagnostics, and execution.

* gcc.target/arm/asm-flag-1.c: New test.
* gcc.target/arm/asm-flag-3.c: New test.
* gcc.target/arm/asm-flag-5.c: New test.
* gcc.target/arm/asm-flag-6.c: New test.

From-SVN: r278227

5 years agoarm, aarch64: Add support for __GCC_ASM_FLAG_OUTPUTS__
Richard Henderson [Thu, 14 Nov 2019 13:44:34 +0000 (13:44 +0000)]
arm, aarch64: Add support for __GCC_ASM_FLAG_OUTPUTS__

Since all but a couple of lines is shared between the two targets,
enable them both at once.

* config/arm/aarch-common-protos.h (arm_md_asm_adjust): Declare.
* config/arm/aarch-common.c (arm_md_asm_adjust): New.
* config/arm/arm-c.c (arm_cpu_builtins): Define
__GCC_ASM_FLAG_OUTPUTS__.
* config/arm/arm.c (TARGET_MD_ASM_ADJUST): New.
* config/aarch64/aarch64-c.c (aarch64_define_unconditional_macros):
Define __GCC_ASM_FLAG_OUTPUTS__.
* config/aarch64/aarch64.c (TARGET_MD_ASM_ADJUST): New.
* doc/extend.texi (FlagOutputOperands): Add documentation
for ARM and AArch64.

From-SVN: r278226

5 years agoarm: Rename CC_NOOVmode to CC_NZmode
Richard Henderson [Thu, 14 Nov 2019 13:44:18 +0000 (13:44 +0000)]
arm: Rename CC_NOOVmode to CC_NZmode

CC_NZmode is a more accurate description of what we require
from the mode, and matches up with the definition in aarch64.

Rename noov_comparison_operator to nz_comparison_operator
in order to match.

* config/arm/arm-modes.def (CC_NZ): Rename from CC_NOOV.
* config/arm/predicates.md (nz_comparison_operator): Rename
from noov_comparison_operator.
* config/arm/arm.c (arm_select_cc_mode): Use CC_NZmode name.
(arm_gen_dicompare_reg): Likewise.
(maybe_get_arm_condition_code): Likewise.
(thumb1_final_prescan_insn): Likewise.
(arm_emit_coreregs_64bit_shift): Likewise.
* config/arm/arm.md (addsi3_compare0): Likewise.
(*addsi3_compare0_scratch, subsi3_compare0): Likewise.
(*mulsi3_compare0, *mulsi3_compare0_v6): Likewise.
(*mulsi3_compare0_scratch, *mulsi3_compare0_scratch_v6): Likewise.
(*mulsi3addsi_compare0, *mulsi3addsi_compare0_v6): Likewise.
(*mulsi3addsi_compare0_scratch): Likewise.
(*mulsi3addsi_compare0_scratch_v6): Likewise.
(*andsi3_compare0, *andsi3_compare0_scratch): Likewise.
(*zeroextractsi_compare0_scratch): Likewise.
(*ne_zeroextractsi, *ne_zeroextractsi_shifted): Likewise.
(*ite_ne_zeroextractsi, *ite_ne_zeroextractsi_shifted): Likewise.
(andsi_not_shiftsi_si_scc_no_reuse): Likewise.
(andsi_not_shiftsi_si_scc): Likewise.
(*andsi_notsi_si_compare0, *andsi_notsi_si_compare0_scratch): Likewise.
(*iorsi3_compare0, *iorsi3_compare0_scratch): Likewise.
(*xorsi3_compare0, *xorsi3_compare0_scratch): Likewise.
(*shiftsi3_compare0, *shiftsi3_compare0_scratch): Likewise.
(*not_shiftsi_compare0, *not_shiftsi_compare0_scratch): Likewise.
(*notsi_compare0, *notsi_compare0_scratch): Likewise.
(return_addr_mask, *check_arch2): Likewise.
(*arith_shiftsi_compare0, *arith_shiftsi_compare0_scratch): Likewise.
(*sub_shiftsi_compare0, *sub_shiftsi_compare0_scratch): Likewise.
(compare_scc splitters): Likewise.
(movcond_addsi): Likewise.
* config/arm/thumb2.md (thumb2_addsi3_compare0): Likewise.
(*thumb2_addsi3_compare0_scratch): Likewise.
(*thumb2_mulsi_short_compare0): Likewise.
(*thumb2_mulsi_short_compare0_scratch): Likewise.
(compare peephole2s): Likewise.
* config/arm/thumb1.md (thumb1_cbz): Use CC_NZmode and
nz_comparison_operator names.
(cbranchsi4_insn): Likewise.

From-SVN: r278225

5 years agoarm: Fix the "c" constraint
Richard Henderson [Thu, 14 Nov 2019 13:44:05 +0000 (13:44 +0000)]
arm: Fix the "c" constraint

The existing definition using register class CC_REG does not
work because CC_REGNUM does not support normal modes, and so
fails to match register_operand.  Use a non-register constraint
and the cc_register predicate instead.

        * config/arm/constraints.md (c): Use cc_register predicate.

From-SVN: r278224

5 years agoaarch64: Add "c" constraint
Richard Henderson [Thu, 14 Nov 2019 13:43:50 +0000 (13:43 +0000)]
aarch64: Add "c" constraint

Mirror arm in letting "c" match the condition code register.

* config/aarch64/constraints.md (c): New constraint.

From-SVN: r278223

5 years agoipa-fnsummary.c (ipa_call_context::estimate_size_and_time, [...]): Micro optimize.
Jan Hubicka [Thu, 14 Nov 2019 13:38:13 +0000 (14:38 +0100)]
ipa-fnsummary.c (ipa_call_context::estimate_size_and_time, [...]): Micro optimize.

* ipa-fnsummary.c (ipa_call_context::estimate_size_and_time,
ipa_merge_fn_summary_after_inlining): Micro optimize.

From-SVN: r278222

5 years ago* params.opt (max-inline-insns-single-O2): Set to 70 (instead of 30).
Jan Hubicka [Thu, 14 Nov 2019 13:30:46 +0000 (14:30 +0100)]
* params.opt (max-inline-insns-single-O2): Set to 70 (instead of 30).

From-SVN: r278221

5 years agoSupport for value ranges in IPA predicates
Jan Hubicka [Thu, 14 Nov 2019 12:41:55 +0000 (12:41 +0000)]
Support for value ranges in IPA predicates

        * ipa-cp.c (ipa_vr_operation_and_type_effects): Move up in file.
(ipa_value_range_from_jfunc): New function.
* ipa-fnsummary.c (evaluate_conditions_for_known_args): Add
known_value_ranges parameter; use it to evalulate conditions.
(evaluate_properties_for_edge): Compute known value ranges.
(ipa_fn_summary_t::duplicate): Update use of
evaluate_conditions_for_known_args.
(estimate_ipcp_clone_size_and_time): Likewise.
(ipa_merge_fn_summary_after_inlining): Likewise.
* ipa-prop.h (ipa_value_range_from_jfunc): Declare.

        * gcc.dg/ipa/inline-9.c: New testcase.

From-SVN: r278220

5 years agoRemove usage of CIF_MAX_INLINE_INSNS_SINGLE_O2_LIMIT.
Martin Liska [Thu, 14 Nov 2019 12:12:26 +0000 (13:12 +0100)]
Remove usage of CIF_MAX_INLINE_INSNS_SINGLE_O2_LIMIT.

2019-11-14  Martin Liska  <mliska@suse.cz>

* ipa-inline.c (want_inline_small_function_p): Use
CIF_MAX_INLINE_INSNS_AUTO_LIMIT instead
of CIF_MAX_INLINE_INSNS_SINGLE_O2_LIMIT.

From-SVN: r278219

5 years agoAdd Optimization keyword for param_max_inline_insns_auto param.
Martin Liska [Thu, 14 Nov 2019 12:08:57 +0000 (13:08 +0100)]
Add Optimization keyword for param_max_inline_insns_auto param.

2019-11-14  Martin Liska  <mliska@suse.cz>

* ipa-cp.c (devirtualization_time_bonus): Use opt_for_fn
of a callee to get value of the param.
* ipa-inline.c (inline_insns_auto): Use proper
opt_for_fn.
* opts.c (maybe_default_option): Do not overwrite param
value if optimization level does not match.  Note that
params usually have default value set via Init() keyword.
* params.opt: Remove -param=max-inline-insns-auto-O2.
* cif-code.def (MAX_INLINE_INSNS_AUTO_O2_LIMIT): Remove.
* doc/invoke.texi: Remove documentation of
max-inline-insns-auto-O2.
2019-11-14  Martin Liska  <mliska@suse.cz>

* c-c++-common/asan/memcmp-1.c: Update expected backtrace.

From-SVN: r278218

5 years agoRemove dead code in switch conv pass.
Martin Liska [Thu, 14 Nov 2019 12:04:45 +0000 (13:04 +0100)]
Remove dead code in switch conv pass.

2019-11-14  Martin Liska  <mliska@suse.cz>

* tree-switch-conversion.c (switch_conversion::switch_conversion):
Do not initialize m_other_count.
(switch_conversion::collect): Do not count m_default_count and
m_other_count as we use frequencies for edges.
* tree-switch-conversion.h: Remove m_default_count and m_other_count.

From-SVN: r278217

5 years agoRemove wrong lto-dump: lto1 makefile dependency.
Martin Liska [Thu, 14 Nov 2019 08:57:57 +0000 (09:57 +0100)]
Remove wrong lto-dump: lto1 makefile dependency.

2019-11-14  Martin Liska  <mliska@suse.cz>

* Make-lang.in: Remove wrong dependency
of LTO_DUMP_EXE on LTO_EXE.

From-SVN: r278212

5 years agoDocument -fallocation-dce.
Martin Liska [Thu, 14 Nov 2019 08:43:29 +0000 (09:43 +0100)]
Document -fallocation-dce.

2019-11-14  Martin Liska  <mliska@suse.cz>

PR other/92329
* doc/invoke.texi: Document -fallocation-dce.

From-SVN: r278211

5 years agoEnable VPOPCNTDQ for icelake-{client,server} and tigerlake.
Martin Liska [Thu, 14 Nov 2019 08:36:34 +0000 (09:36 +0100)]
Enable VPOPCNTDQ for icelake-{client,server} and tigerlake.

2019-11-14  Martin Liska  <mliska@suse.cz>

PR target/92389
* config/i386/i386.h: Add PTA_AVX512VPOPCNTDQ to
PTA_ICELAKE_CLIENT which is later interited by
PTA_ICELAKE_SERVER and PTA_TIGERLAKE.

From-SVN: r278210

5 years agoUpdate statistics about needed symbols in IPA ICF.
Martin Liska [Thu, 14 Nov 2019 08:35:48 +0000 (09:35 +0100)]
Update statistics about needed symbols in IPA ICF.

2019-11-14  Martin Liska  <mliska@suse.cz>

* ipa-icf.c (sem_item_optimizer::execute): Save
loaded_symbols.
(sem_item_optimizer::parse_nonsingleton_classes):
Return number of loaded symbols.
(sem_item_optimizer::merge_classes): Print
statistics about totally needed symbols.
* ipa-icf.h (parse_nonsingleton_classes): Change return
type.
(merge_classes): Add one argument.

From-SVN: r278209

5 years agoHandle FIELD_DECL in IPA ICF.
Martin Liska [Thu, 14 Nov 2019 08:35:35 +0000 (09:35 +0100)]
Handle FIELD_DECL in IPA ICF.

2019-11-14  Martin Liska  <mliska@suse.cz>

* ipa-icf-gimple.c (func_checker::hash_operand): Improve
func_checker::hash_operand by handling of FIELD_DECLs.

From-SVN: r278208

5 years agoUse func_checker::hash_operand for hashing of GIMPLE operands.
Martin Liska [Thu, 14 Nov 2019 08:35:25 +0000 (09:35 +0100)]
Use func_checker::hash_operand for hashing of GIMPLE operands.

2019-11-14  Martin Liska  <mliska@suse.cz>

* ipa-icf-gimple.h (func_checker::func_checker): Add
default constructor.
* ipa-icf.c (sem_function::init): Make operand_equal_p
and hash_operand public.
(sem_item::add_expr): Remove.
(sem_item::add_type): Remove.
(sem_function::hash_stmt): Use m_checker for hashing
of GIMPLE statements.
(sem_function::parse): Init with checker.
(sem_variable::parse): Pass NULL as checker.
(sem_item_optimizer::parse_funcs_and_vars):
Pass checker to ::parse function.
(sem_item_optimizer::parse_nonsingleton_classes): Likewise.
(sem_variable::parse): New function.
(sem_variable::get_hash): Only return computed hash value.
(sem_variable::init): Initialize hash of a variable.
* ipa-icf.h: Remove add_expr, add_type and add func_checker
to couple of functions as a new argument.

From-SVN: r278207

5 years agoUpdate dump message in IPA ICF.
Martin Liska [Thu, 14 Nov 2019 08:35:05 +0000 (09:35 +0100)]
Update dump message in IPA ICF.

2019-11-14  Martin Liska  <mliska@suse.cz>

* ipa-icf-gimple.c (func_checker::compare_gimple_call): Update
bail out reason.
(func_checker::compare_gimple_assign): Likewise.

From-SVN: r278206

5 years agoi386-options.c (ix86_omp_device_kind_arch_isa): Don't change sse4.2 to sse4_2 and...
Jakub Jelinek [Thu, 14 Nov 2019 08:15:31 +0000 (09:15 +0100)]
i386-options.c (ix86_omp_device_kind_arch_isa): Don't change sse4.2 to sse4_2 and sse4.1 to sse4.1.

* config/i386/i386-options.c (ix86_omp_device_kind_arch_isa): Don't
change sse4.2 to sse4_2 and sse4.1 to sse4.1.
* config/i386/t-omp-device (omp-device-properties-i386): Likewise.

* c-c++-common/gomp/declare-variant-11.c: Add "sse4.2" and "sse4.1"
test.

From-SVN: r278205

5 years agoc-parser.c (c_parser_omp_context_selector): Don't require score argument to fit into...
Jakub Jelinek [Thu, 14 Nov 2019 08:14:16 +0000 (09:14 +0100)]
c-parser.c (c_parser_omp_context_selector): Don't require score argument to fit into shwi, just to be INTEGER_CST.

* c-parser.c (c_parser_omp_context_selector): Don't require score
argument to fit into shwi, just to be INTEGER_CST.  Diagnose
negative score.

* parser.c (cp_parser_omp_context_selector): Don't require score
argument to fit into shwi, just to be INTEGER_CST.  Diagnose
negative score.
* pt.c (tsubst_attribute): Likewise.

* c-c++-common/gomp/declare-variant-2.c: Add test for non-integral
score and for negative score.
* c-c++-common/gomp/declare-variant-3.c: Add test for zero score.
* g++.dg/gomp/declare-variant-8.C: Add test for negative and zero
scores.

From-SVN: r278204

5 years agoc-omp.c (c_omp_check_context_selector): Add nvidia to the list of valid vendors.
Jakub Jelinek [Thu, 14 Nov 2019 08:13:01 +0000 (09:13 +0100)]
c-omp.c (c_omp_check_context_selector): Add nvidia to the list of valid vendors.

* c-omp.c (c_omp_check_context_selector): Add nvidia to the list of
valid vendors.

* c-c++-common/gomp/declare-variant-3.c: Add testcase for vendor nvidia.

From-SVN: r278203

5 years agoomp-general.c (omp_context_name_list_prop): New function.
Jakub Jelinek [Thu, 14 Nov 2019 08:12:10 +0000 (09:12 +0100)]
omp-general.c (omp_context_name_list_prop): New function.

* omp-general.c (omp_context_name_list_prop): New function.
(omp_context_selector_matches): Use it.  Return 0 if it returns
NULL.
(omp_context_selector_props_compare): Allow equivalency of an
identifier and a string literal containing no embedded zeros.
c-family/
* c-omp.c (c_omp_check_context_selector): Handle name lists
containing string literals.  Don't diagnose atomic_default_mem_order
with multiple props.
c/
* c-parser.c (c_parser_omp_context_selector): Rename
CTX_PROPERTY_IDLIST to CTX_PROPERTY_NAME_LIST, add CTX_PROPERTY_ID.
Use CTX_PROPERTY_ID for atomic_default_mem_order, only allow a single
identifier in that.  For CTX_PROPERTY_NAME_LIST, allow identifiers
and string literals.
cp/
* parser.c (cp_parser_omp_context_selector): Rename
CTX_PROPERTY_IDLIST to CTX_PROPERTY_NAME_LIST, add CTX_PROPERTY_ID.
Use CTX_PROPERTY_ID for atomic_default_mem_order, only allow a single
identifier in that.  For CTX_PROPERTY_NAME_LIST, allow identifiers
and string literals.
* pt.c (tsubst_attribute): Fix up STRING_CST handling if allow_string.
testsuite/
* c-c++-common/gomp/declare-variant-2.c: Adjust expected diagnostics,
add a test for atomic_default_mem_order with a string literal.
* c-c++-common/gomp/declare-variant-3.c: Use string literal props
in a few random places, add a few string literal prop related tests.
* c-c++-common/gomp/declare-variant-8.c: Likewise.
* c-c++-common/gomp/declare-variant-9.c: Use string literal props
in a few random places.
* c-c++-common/gomp/declare-variant-10.c: Likewise.
* c-c++-common/gomp/declare-variant-11.c: Likewise.
* c-c++-common/gomp/declare-variant-12.c: Likewise.
* g++.dg/gomp/declare-variant-7.C: Likewise.

From-SVN: r278202

5 years agoRemove higher precision range tests because they are unused.
Aldy Hernandez [Thu, 14 Nov 2019 07:55:04 +0000 (07:55 +0000)]
Remove higher precision range tests because they are unused.

This removes the dependency on m_max_pairs from the selftests, which has
the ultimate effect of allowing us to put the tests in the selftest
namespace as was the original plan.

From-SVN: r278200

5 years agoupdate_version_svn (IGNORE_BRANCHES): Add 7.
Richard Biener [Thu, 14 Nov 2019 07:41:56 +0000 (07:41 +0000)]
update_version_svn (IGNORE_BRANCHES): Add 7.

2019-11-14  Richard Biener  <rguenther@suse.de>

* update_version_svn (IGNORE_BRANCHES): Add 7.

From-SVN: r278199

5 years agoMake normalize_addresses and normalize_symbolics work on THIS, instead
Aldy Hernandez [Thu, 14 Nov 2019 07:03:30 +0000 (07:03 +0000)]
Make normalize_addresses and normalize_symbolics work on THIS, instead
of returning by value.

From-SVN: r278196

5 years ago[testsuite] Fix PR92464 by adjust test case loop bound
Kewen Lin [Thu, 14 Nov 2019 05:57:12 +0000 (05:57 +0000)]
[testsuite] Fix PR92464 by adjust test case loop bound

  The recent vectorization cost adjustment on load leads
  the profitable min iteration count to change from 19 to 12.
  The case happens to hit the threshold.  This patch is to
  adjust the loop bound from 16 to 14.

  gcc/testsuite/ChangeLog

  2019-11-14  Kewen Lin  <linkw@gcc.gnu.org>

    PR target/92464
    * gcc.dg/vect/costmodel/ppc/costmodel-vect-76b.c: Adjust
    loop bound due to load cost adjustment.

From-SVN: r278195

5 years agoSupport C2x [[]] attributes for C.
Joseph Myers [Thu, 14 Nov 2019 03:49:43 +0000 (03:49 +0000)]
Support C2x [[]] attributes for C.

This patch adds support for the C2x [[]] attribute syntax to the C
front end.  Support is only added for C at this point, not for
Objective-C; I intend to add the unbounded lookahead required to
support it for Objective-C in a followup patch, but maybe not in
development stage 1.

The syntax is supported in all relevant places where the standard says
it is supported, but support is not added for the individual
attributes specified in C2x (all of which are optional to support).  I
expect to add support for some of them in followup patches; all except
nodiscard can be mapped directly to the semantics of an existing GNU
attribute (subject to extra checks for invalid usages such as the same
attribute being used more than once inside a single [[]]), and the
fallthrough attribute already works after this patch because of
existing special-case code handling it (but without some of the checks
for invalid usage being present).

Note that the four functions c_token_starts_declspecs,
c_token_starts_declaration, c_parser_next_token_starts_declspecs and
c_parser_next_tokens_start_declaration do *not* accept "[[".  This is
analogous with the handling of __extension__: both cases have the
property that they can start either a declaration or some other
statements and so need an unbounded number of tokens to be parsed in
the caller before it can find out what kind of syntactic construct
follows.  Note also that, while I updated all places calling those
functions for standard C syntax to handle attributes if applicable, I
did not do anything regarding calls to such functions for OpenMP or
OpenACC constructs.  Thus, if there are such constructs using such
functions where "[[" *should* be accepted as a possible start to a
declaration, the code for parsing those constructs should be updated
accordingly.

Although all cases of the syntax are handled, and attributes applied
to the constructs the standard says they should be (with less laxity
than there is for GNU attributes to allow an attribute applied to one
construct to be moved automatically to another one), there is a major
limitation in the existing language-independent code in attribs.c
preventing most cases of type attributes from working.  The following
code has been present with minor changes since the first support for
[[]] attributes for C++ was added:

      if (TYPE_P (*node)
          && cxx11_attr_p
          && !(flags & ATTR_FLAG_TYPE_IN_PLACE))
        {
          /* This is a c++11 attribute that appertains to a
             type-specifier, outside of the definition of, a class
             type.  Ignore it.  */
          auto_diagnostic_group d;
          if (warning (OPT_Wattributes, "attribute ignored"))
            inform (input_location,
                    "an attribute that appertains to a type-specifier "
                    "is ignored");
          continue;
        }

I see no justification for this in general for either C or C++ and so
propose to remove or restrict it in a followup bug-fix patch.  Both C
and C++ are clear about attributes in certain places (at the end of
declaration specifiers, or after function or array declarators)
appertaining to a specific type (and explicitly say, in the case of
attributes at the end of declaration specifiers, that they only apply
for that particular use of that type, not for subsequent uses of the
same type without the attributes).  Thus it seems clear to me that,
for example,

int [[gnu::mode(DI)]] x;

ought to be accepted as an analogue in [[]] syntax for

int __attribute__((mode(DI))) x;

(or strictly as an analogue for a version of that with extra
parentheses to make the GNU attribute bind properly to the type rather
than being automatically moved from the declaration to the type).
There are certain cases where an attribute *does* only make sense for
the definition of a type (e.g. "packed" on structure types), but those
should already be handled in the individual attribute handlers (such
as handle_packed_attribute, which already has code to deal with that
issue).  So my inclination is that the above-quoted check in attribs.c
should simply be removed, but failing that it should be restricted to
structure and union types (and such a change would be a bug-fix).
That would then allow various cases of [[]] attributes on types to
work properly.

Bootstrapped with no regressions on x86_64-pc-linux-gnu.

gcc/c:
* c-tree.h (enum c_typespec_kind): Add ctsk_tagref_attrs and
ctsk_tagfirstref_attrs.
(struct c_declspecs): Update description of attrs.  Add
postfix_attrs and non_std_attrs_seen_p.  Increase size of
typespec_kind bit-field.
(c_warn_unused_attributes): New declaration.
(parser_xref_tag): Update prototype.
* c-decl.c (c_warn_unused_attributes): New function.
(shadow_tag_warned): Handle ctsk_tagfirstref_attrs and
ctsk_tagref_attrs.  Handle attribute declarations.
(check_compound_literal_type): Handle ctsk_tagfirstref_attrs.
(grokdeclarator): Handle standard attributes.
(parser_xref_tag): Add arguments have_std_attrs and attrs.  Apply
attributes to incomplete type reference.
(xref_tag): Update call to parser_xref_tag.
(declspecs_add_addrspace, declspecs_add_type)
(declspecs_add_scspec, declspecs_add_attrs): Set
non_std_attrs_seen_p.
(finish_declspecs): Apply postfix standard attributes to type.
* c-parser.c (c_token_starts_declspecs)
(c_token_starts_declaration, c_parser_next_token_starts_declspecs)
(c_parser_next_tokens_start_declaration): Update comments.
(c_parser_consume_token, c_parser_consume_pragma): Handle moving
parser->tokens[2] to parser->tokens[1].
(c_parser_nth_token_starts_std_attributes)
(c_parser_std_attribute_specifier_sequence): New functions.
(c_parser_declaration_or_fndef): Add arguments have_attrs and
attrs.  All callers changed.  Handle standard attributes.
(c_parser_parms_declarator, c_parser_parms_list_declarator)
(c_parser_parameter_declaration): Add argument have_gnu_attrs.
All callers changed.
(c_parser_declspecs): Add arguments start_std_attr_ok and
end_std_attr_ok.  All callers changed.  Handle standard
attributes.
(c_parser_enum_specifier, c_parser_struct_or_union_specifier)
(c_parser_direct_declarator, c_parser_direct_declarator_inner)
(c_parser_compound_statement_nostart, c_parser_all_labels)
(c_parser_label, c_parser_statement, c_parser_for_statement):
Handle standard attributes.
* c-parser.h (c_parser_declspecs): Update prototype.
* gimple-parser.c (c_parser_gimple_declaration): Update call to
c_parser_declspecs.

gcc/testsuite:
* gcc.dg/c2x-attr-fallthrough-1.c, gcc.dg/c2x-attr-syntax-1.c,
gcc.dg/c2x-attr-syntax-2.c, gcc.dg/c2x-attr-syntax-3.c,
gcc.dg/gnu2x-attr-syntax-1.c, gcc.dg/gnu2x-attr-syntax-2.c,
gcc.dg/gnu2x-attrs-1.c: New tests.

From-SVN: r278194

5 years agoSupport extended aggregate jump function in ipa-cp
Feng Xue [Thu, 14 Nov 2019 03:19:15 +0000 (03:19 +0000)]
Support extended aggregate jump function in ipa-cp

2019-11-14  Feng Xue  <fxue@os.amperecomputing.com>

PR ipa/91682
* ipa-prop.h (jump_func_type): New value IPA_JF_LOAD_AGG.
(ipa_load_agg_data, ipa_agg_value, ipa_agg_value_set): New structs.
(ipa_agg_jf_item): Add new field jftype and type, redefine field value.
(ipa_agg_jump_function): Remove member function equal_to.
(ipa_agg_jump_function_p): Remove typedef.
(ipa_copy_agg_values, ipa_release_agg_values): New functions.
* ipa-prop.c (ipa_print_node_jump_functions_for_edge): Dump
information for aggregate jump function.
(get_ssa_def_if_simple_copy): Add new parameter rhs_stmt to
record last definition statement.
(load_from_unmodified_param_or_agg): New function.
(ipa_known_agg_contents_list): Add new field type and value, remove
field constant.
(build_agg_jump_func_from_list): Rename parameter const_count to
value_count, build aggregate jump function from ipa_load_agg_data.
(analyze_agg_content_value): New function.
(extract_mem_content): Analyze memory store assignment to prepare
information for aggregate jump function generation.
(determine_known_aggregate_parts): Add new parameter fbi, remove
parameter aa_walk_budeget_p.
(update_jump_functions_after_inlining): Update aggregate jump function.
(ipa_find_agg_cst_for_param): Change type of parameter agg.
(try_make_edge_direct_simple_call): Add new parameter new_root.
(try_make_edge_direct_virtual_call): Add new parameter new_root and
new_root_info.
(update_indirect_edges_after_inlining): Pass new argument to
try_make_edge_direct_simple_call and try_make_edge_direct_virtual_call.
(ipa_write_jump_function): Write aggregate jump function to file.
(ipa_read_jump_function): Read aggregate jump function from file.
(ipa_agg_value::equal_to): Migrate from ipa_agg_jf_item::equal_to.
* ipa-cp.c (ipa_get_jf_arith_result): New function.
(ipa_agg_value_from_node): Likewise.
(ipa_agg_value_set_from_jfunc): Likewise.
(propagate_vals_across_arith_jfunc): Likewise.
(propagate_aggregate_lattice): Likewise.
(ipa_get_jf_pass_through_result): Call ipa_get_jf_arith_result.
(propagate_vals_across_pass_through): Call
propagate_vals_across_arith_jfunc.
(get_clone_agg_value): Move forward.
(propagate_aggs_across_jump_function): Handle value propagation for
aggregate jump function.
(agg_jmp_p_vec_for_t_vec): Remove.
(context_independent_aggregate_values): Replace vec<ipa_agg_jf_item>
with vec<ipa_agg_value>.
(copy_plats_to_inter, intersect_with_plats): Likewise.
(agg_replacements_to_vector, intersect_with_agg_replacements): Likewise.
(intersect_aggregate_with_edge): Likewise.
(find_aggregate_values_for_callers_subset): Likewise.
(cgraph_edge_brings_all_agg_vals_for_node): Likewise.
(estimate_local_effects): Replace vec<ipa_agg_jump_function> and
vec<ipa_agg_jump_function_p> with vec<ipa_agg_value_set>.
(gather_context_independent_values): Likewise.
(perform_estimation_of_a_value, decide_whether_version_node): Likewise.
* ipa-fnsummary.c (evaluate_conditions_for_known_args): Replace
vec<ipa_agg_jump_function_p> with vec<ipa_agg_value_set>.
(evaluate_properties_for_edge): Likewise.
(estimate_edge_devirt_benefit): Likewise.
(estimate_edge_size_and_time):  Likewise.
(estimate_calls_size_and_time): Likewise.
(ipa_call_context::ipa_call_context): Likewise.
(estimate_ipcp_clone_size_and_time):  Likewise.
* ipa-fnsummary.h (ipa_call_context): Replace
vec<ipa_agg_jump_function_p> with vec<ipa_agg_value_set>.
* ipa-inline-analysis.c (do_estimate_edge_time): Replace
vec<ipa_agg_jump_function_p> with vec<ipa_agg_value_set>.
(do_estimate_edge_size): Likewise.
(do_estimate_edge_hints): Likewise.

2019-11-14  Feng Xue  <fxue@os.amperecomputing.com>

        PR ipa/91682
        * gcc.dg/ipa/ipcp-agg-10.c: Change dg-scan string.
        * gcc.dg/ipa/ipcp-agg-11.c: New test.

From-SVN: r278193

5 years agore PR ipa/92421 (ICE in inline_small_functions, at ipa-inline.c:2001 since r277759)
Jakub Jelinek [Thu, 14 Nov 2019 01:02:16 +0000 (02:02 +0100)]
re PR ipa/92421 (ICE in inline_small_functions, at ipa-inline.c:2001 since r277759)

PR ipa/92421
* g++.dg/torture/pr92421.C: Add -Wno-return-type to
dg-additional-options.  Avoid -Wwrite-string warnings, most of
-Wreturn-type warnings, define bf ctor.  Use struct instead of class
with public: at the start.

From-SVN: r278192

5 years agoalias-decl-pr92206-3.C: Require effective target c++14 rather than c++11.
Jakub Jelinek [Thu, 14 Nov 2019 00:47:05 +0000 (01:47 +0100)]
alias-decl-pr92206-3.C: Require effective target c++14 rather than c++11.

* g++.dg/cpp0x/alias-decl-pr92206-3.C: Require effective target c++14
rather than c++11.

From-SVN: r278191

5 years agoISO_Fortran_binding_17.c: Include ../../../libgfortran/ISO_Fortran_binding.h rather...
Jakub Jelinek [Thu, 14 Nov 2019 00:22:34 +0000 (01:22 +0100)]
ISO_Fortran_binding_17.c: Include ../../../libgfortran/ISO_Fortran_binding.h rather than...

* gfortran.dg/ISO_Fortran_binding_17.c: Include
../../../libgfortran/ISO_Fortran_binding.h rather than
ISO_Fortran_binding.h.

From-SVN: r278190

5 years agoDaily bump.
GCC Administrator [Thu, 14 Nov 2019 00:16:34 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r278189

5 years agoipa-cp.c (propagate_vr_across_jump_function): Propagate also across binary operations.
Jan Hubicka [Wed, 13 Nov 2019 23:40:49 +0000 (00:40 +0100)]
ipa-cp.c (propagate_vr_across_jump_function): Propagate also across binary operations.

* ipa-cp.c (propagate_vr_across_jump_function): Propagate also across
binary operations.

From-SVN: r278185

5 years agoipa-profile.c (check_argument_count): Check properly that e_info is non-NULL; do...
Jan Hubicka [Wed, 13 Nov 2019 23:39:44 +0000 (00:39 +0100)]
ipa-profile.c (check_argument_count): Check properly that e_info is non-NULL; do not check descriptors.

* ipa-profile.c (check_argument_count): Check properly that e_info
is non-NULL; do not check descriptors.

From-SVN: r278184

5 years agoipa-inline-analysis.c (do_estimate_edge_time): Relax check for ipa profiles.
Jan Hubicka [Wed, 13 Nov 2019 22:14:24 +0000 (23:14 +0100)]
ipa-inline-analysis.c (do_estimate_edge_time): Relax check for ipa profiles.

* ipa-inline-analysis.c (do_estimate_edge_time): Relax
check for ipa profiles.

From-SVN: r278179

5 years agore PR ipa/92421 (ICE in inline_small_functions, at ipa-inline.c:2001 since r277759)
Jan Hubicka [Wed, 13 Nov 2019 21:02:11 +0000 (22:02 +0100)]
re PR ipa/92421 (ICE in inline_small_functions, at ipa-inline.c:2001 since r277759)

PR c++/92421
* ipa-prop.c (update_indirect_edges_after_inlining):
Mark parameter as used.
* ipa-inline.c (recursive_inlining): Reset node cache
after inlining.
(inline_small_functions): Remove checking ifdef.
* ipa-inline-analysis.c (do_estimate_edge_time): Verify
cache consistency.
* g++.dg/torture/pr92421.C: New testcase.

From-SVN: r278159

5 years agore PR ipa/92498 (gcc.dg/tree-prof/crossmodule-indircall-1.c fails starting with r278100)
Jan Hubicka [Wed, 13 Nov 2019 19:44:35 +0000 (20:44 +0100)]
re PR ipa/92498 (gcc.dg/tree-prof/crossmodule-indircall-1.c fails starting with r278100)

PR ipa/92498
* ipa-profile.c (check_argument_count): Do not ICE when descriptors
is NULL.
(ipa_profile): Fix reversed test.

From-SVN: r278157

5 years agoipa-cp.c (ignore_edge_p): Do not check caller flags.
Jan Hubicka [Wed, 13 Nov 2019 19:42:02 +0000 (20:42 +0100)]
ipa-cp.c (ignore_edge_p): Do not check caller flags.

* ipa-cp.c (ignore_edge_p): Do not check caller flags.
(propagate_constants_topo): Fix typo.

From-SVN: r278156

5 years agoMove plain value_range things to value-range.[hc]*.
Aldy Hernandez [Wed, 13 Nov 2019 19:13:40 +0000 (19:13 +0000)]
Move plain value_range things to value-range.[hc]*.

From-SVN: r278153

5 years agoSanitize the constant argument for rotr<mode>3
Dragan Mladjenovic [Wed, 13 Nov 2019 18:50:15 +0000 (18:50 +0000)]
Sanitize the constant argument for rotr<mode>3

This was dormant for quite some time, but it started happening for me
on gcc.c-torture/compile/pr65153.c sometime after r276645 for -mabi=32 linux runs.

The pattern accepts any SMALL_OPERAND constant value while it asserts during the final
that the value is in the mode size range. I this case it happens that combine_and_move_insns
during ira makes a pattern with negative "shift count" which fails at final stage.

This simple fix just truncates the constant operand to mode size the same as shift patterns.

gcc/ChangeLog:

2019-11-13  Dragan Mladjenovic  <dmladjenovic@wavecomp.com>

* config/mips/mips.md (rotr<mode>3): Sanitize the constant argument
instead of asserting its value.

From-SVN: r278152

5 years agodownload_prerequisites: Use http instead of ftp for downloading.
Janne Blomqvist [Wed, 13 Nov 2019 17:48:52 +0000 (19:48 +0200)]
download_prerequisites: Use http instead of ftp for downloading.

2019-11-13  Janne Blomqvist  <jb@gcc.gnu.org>

        * download_prerequisites: Use http instead of ftp for downloading.

From-SVN: r278151

5 years agopr92090.c: Limit -mbig to powerpc64le-*-*.
David Edelsohn [Wed, 13 Nov 2019 17:13:03 +0000 (17:13 +0000)]
pr92090.c: Limit -mbig to powerpc64le-*-*.

        * gcc.target/powerpc/pr92090.c: Limit -mbig to powerpc64le-*-*.
        * gcc.target/powerpc/pr92090-2.c: Likewise.

From-SVN: r278150

5 years agolibsupc++: Implement comparison algorithms for C++20
Jonathan Wakely [Wed, 13 Nov 2019 16:26:18 +0000 (16:26 +0000)]
libsupc++: Implement comparison algorithms for C++20

This is incomplete because std::strong_order doesn't support
floating-point types.

The partial_order and weak_order tests use VERIFY instead of
static_assert because of PR 92431.

* libsupc++/compare (strong_order, weak_order, partial_order)
(compare_strong_order_fallback, compare_weak_order_fallback)
(compare_partial_order_fallback): Define customization point objects
for C++20.
* testsuite/18_support/comparisons/algorithms/partial_order.cc: New
test.
* testsuite/18_support/comparisons/algorithms/strong_order.cc: New
test.
* testsuite/18_support/comparisons/algorithms/weak_order.cc: New test.

From-SVN: r278149

5 years agoRewrite value_range constructors to the value_range_kind is at the end, and defaults...
Aldy Hernandez [Wed, 13 Nov 2019 16:03:27 +0000 (16:03 +0000)]
Rewrite value_range constructors to the value_range_kind is at the end, and defaults to VR_RANGE.

Rewrite value_range constructors to the value_range_kind is at the
end, and defaults to VR_RANGE.  Similarly for set() methods.

From-SVN: r278148

5 years agoPR c++/89070 - bogus [[nodiscard]] warning in SFINAE.
Marek Polacek [Wed, 13 Nov 2019 15:59:53 +0000 (15:59 +0000)]
PR c++/89070 - bogus [[nodiscard]] warning in SFINAE.

This is a complaint that we issue a [[nodiscard]] warning even in SFINAE
contexts.  Here 'complain' is tf_decltype, but not tf_warning so I guess
we can fix it as below.

* cvt.c (convert_to_void): Guard maybe_warn_nodiscard calls with
tf_warning.

* g++.dg/cpp1z/nodiscard7.C: New test.

From-SVN: r278147

5 years agotree-dump.c (dequeue_and_dump): Print first tree operand for VIEW_CONVERT_EXPR.
Ulrich Drepper [Wed, 13 Nov 2019 15:51:30 +0000 (15:51 +0000)]
tree-dump.c (dequeue_and_dump): Print first tree operand for VIEW_CONVERT_EXPR.

2019-11-13  Ulrich Drepper  <drepper@redhat.com>

* tree-dump.c (dequeue_and_dump): Print first tree operand
for VIEW_CONVERT_EXPR.

From-SVN: r278146

5 years agoAdd C2x *_NORM_MAX constants to <float.h>.
Joseph Myers [Wed, 13 Nov 2019 15:25:15 +0000 (15:25 +0000)]
Add C2x *_NORM_MAX constants to <float.h>.

C2x adds <float.h> constants FLT_NORM_MAX, DBL_NORM_MAX and
LDBL_NORM_MAX.  These are for the maximum "normalized" finite
floating-point number, where the given definition of normalized is
that all possible values with MANT_DIG significand digits (leading one
not zero) can be represented with that exponent.  The effect of that
definition is that these macros are the same as the corresponding MAX
macros for all formats except IBM long double, where the NORM_MAX
value has exponent 1 smaller than the MAX one so that all 106 digits
can be 1.

This patch adds those macros to GCC.  They are only defined for float,
double and long double; C2x does not include such macros for DFP
types, and while the integration of TS 18661-3 into C2x has not yet
occurred, the draft proposed text does not add them for the _FloatN /
_FloatNx types (where they would always be the same as the MAX
macros).

Bootstrapped with no regressions on x86_64-pc-linux-gnu.  Also tested
compilation of the new test for powerpc-linux-gnu to confirm the check
of LDBL_NORM_MAX in the IBM long double case does get properly
optimized out.

gcc:
* ginclude/float.c [__STDC_VERSION__ > 201710L] (FLT_NORM_MAX,
DBL_NORM_MAX, LDBL_NORM_MAX): Define.
* real.c (get_max_float): Add norm_max argument.
* real.h (get_max_float): Update prototype.
* builtins.c (fold_builtin_interclass_mathfn): Update calls to
get_max_float.

gcc/c-family:
* c-cppbuiltin.c (builtin_define_float_constants): Also define
NORM_MAX constants.  Update call to get_max_float.
(LAZY_HEX_FP_VALUES_CNT): Update value to include NORM_MAX
constants.

gcc/d:
* d-target.cc (define_float_constants): Update call to
get_max_float.

gcc/testsuite:
* gcc.dg/c11-float-3.c, gcc.dg/c2x-float-1.c: New tests.

From-SVN: r278145

5 years agoCome up with selftests for dbgcnt.
Martin Liska [Wed, 13 Nov 2019 15:13:16 +0000 (16:13 +0100)]
Come up with selftests for dbgcnt.

2019-11-13  Martin Liska  <mliska@suse.cz>

* dbgcnt.c (test_sorted_dbg_counters): New.
(dbgcnt_c_tests): Likewise.
* selftest-run-tests.c (selftest::run_tests): Likewise.
* selftest.h (dbgcnt_c_tests): Likewise.

From-SVN: r278144

5 years agoAdd a few missing checks that IPA_NODE_REF is not NULL (PR 92454)
Martin Jambor [Wed, 13 Nov 2019 14:12:58 +0000 (15:12 +0100)]
Add a few missing checks that IPA_NODE_REF is not NULL (PR 92454)

2019-11-13  Jan Hubicka  <hubicka@ucw.cz>
    Martin Jambor  <mjambor@suse.cz>

PR ipa/92454
* ipa-cp.c (spread_undeadness): Check that IPA_NODE_REF exists.
(identify_dead_nodes): Likewise.

testsuite/
* g++.dg/ipa/pr92454.C: New test.

From-SVN: r278142

5 years agoDo not overuse push/pop_cfun in IPA ICF.
Martin Liska [Wed, 13 Nov 2019 13:48:25 +0000 (14:48 +0100)]
Do not overuse push/pop_cfun in IPA ICF.

2019-11-13  Martin Liska  <mliska@suse.cz>

* ipa-icf.c (sem_function::equals_private): Do not overuse
push/pop_cfun functions.

From-SVN: r278141

5 years agoEnhance syntax of -fdbg-cnt.
Martin Liska [Wed, 13 Nov 2019 13:47:29 +0000 (14:47 +0100)]
Enhance syntax of -fdbg-cnt.

2019-11-13  Martin Liska  <mliska@suse.cz>

* common.opt: Document change of -fdbg-cnt option.
* dbgcnt.c (DEBUG_COUNTER): Remove.
(dbg_cnt_is_enabled): Remove.
(dbg_cnt): Work with new intervals.
(dbg_cnt_set_limit_by_index): Set to new
list of intervals.
(dbg_cnt_set_limit_by_name): Likewise.
(dbg_cnt_process_single_pair): Process new format.
(dbg_cnt_process_opt): Likewise.
(dbg_cnt_list_all_counters): Likewise.
* doc/invoke.texi: Document change of -fdbg-cnt option.
(cmp_tuples): New.
2019-11-13  Martin Liska  <mliska@suse.cz>

* gcc.dg/ipa/ipa-icf-39.c: Update -fdbg-cnt to the new format.
* gcc.dg/pr68766.c: Likewise.

From-SVN: r278140

5 years agoipa-inline.c (ipa_inline): Check that function is defined before flattening.
Jan Hubicka [Wed, 13 Nov 2019 13:29:27 +0000 (14:29 +0100)]
ipa-inline.c (ipa_inline): Check that function is defined before flattening.

* ipa-inline.c (ipa_inline): Check that function is defined before
flattening.
* gcc.c-torture/compile/flatten.c: New testcase.

From-SVN: r278139

5 years agoGCN Libgomp Plugin
Andrew Stubbs [Wed, 13 Nov 2019 12:38:18 +0000 (12:38 +0000)]
GCN Libgomp Plugin

2019-11-13  Andrew Stubbs  <ams@codesourcery.com>
    Kwok Cheung Yeung  <kcy@codesourcery.com>
    Julian Brown  <julian@codesourcery.com>
    Tom de Vries  <tom@codesourcery.com>

libgomp/
* plugin/Makefrag.am: Add amdgcn plugin support.
* plugin/configfrag.ac: Likewise.
* plugin/plugin-gcn.c: New file.
* configure: Regenerate.
* Makefile.in: Regenerate.
* testsuite/Makefile.in: Regenerate.

Co-Authored-By: Julian Brown <julian@codesourcery.com>
Co-Authored-By: Kwok Cheung Yeung <kcy@codesourcery.com>
Co-Authored-By: Tom de Vries <tom@codesourcery.com>
From-SVN: r278138

5 years agoUse a single worker for OpenACC on AMD GCN
Andrew Stubbs [Wed, 13 Nov 2019 12:38:13 +0000 (12:38 +0000)]
Use a single worker for OpenACC on AMD GCN

2019-11-13  Andrew Stubbs  <ams@codesourcery.com>
    Julian Brown  <julian@codesourcery.com>

gcc/
* config/gcn/gcn.c (gcn_goacc_validate_dims): Ensure
flag_worker_partitioning is not set.
(TARGET_GOACC_WORKER_PARTITIONING): Remove target hook definition.
* config/gcn/gcn.opt (macc-experimental-workers): Default to off.

Co-Authored-By: Julian Brown <julian@codesourcery.com>
From-SVN: r278137

5 years agoOptimize GCN OpenMP malloc performance
Andrew Stubbs [Wed, 13 Nov 2019 12:38:09 +0000 (12:38 +0000)]
Optimize GCN OpenMP malloc performance

2019-11-13  Andrew Stubbs  <ams@codesourcery.com>

libgomp/
* config/gcn/team.c (gomp_gcn_enter_kernel): Set up the team arena
and use team_malloc variants.
(gomp_gcn_exit_kernel): Use team_free.
* libgomp.h (TEAM_ARENA_SIZE): Define.
(TEAM_ARENA_START): Define.
(TEAM_ARENA_FREE): Define.
(TEAM_ARENA_END): Define.
(team_malloc): New function.
(team_malloc_cleared): New function.
(team_free): New function.
* team.c (gomp_new_team): Initialize and use team_malloc.
(free_team): Use team_free.
(gomp_free_thread): Use team_free.
(gomp_pause_host): Use team_free.
* work.c (gomp_init_work_share): Use team_malloc.
(gomp_fini_work_share): Use team_free.

From-SVN: r278136

5 years agoGCN libgomp port
Andrew Stubbs [Wed, 13 Nov 2019 12:38:04 +0000 (12:38 +0000)]
GCN libgomp port

2019-11-13  Andrew Stubbs  <ams@codesourcery.com>
    Kwok Cheung Yeung  <kcy@codesourcery.com>
    Julian Brown  <julian@codesourcery.com>
    Tom de Vries  <tom@codesourcery.com>

include/
* gomp-constants.h (GOMP_DEVICE_GCN): Define.
(GOMP_VERSION_GCN): Define.

libgomp/
* Makefile.am (libgomp_la_SOURCES): Add oacc-target.c.
* Makefile.in: Regenerate.
* config.h.in (PLUGIN_GCN): Add new undef.
* config/accel/openacc.f90 (acc_device_gcn): New parameter.
* config/gcn/affinity-fmt.c: New file.
* config/gcn/bar.c: New file.
* config/gcn/bar.h: New file.
* config/gcn/doacross.h: New file.
* config/gcn/icv-device.c: New file.
* config/gcn/oacc-target.c: New file.
* config/gcn/simple-bar.h: New file.
* config/gcn/target.c: New file.
* config/gcn/task.c: New file.
* config/gcn/team.c: New file.
* config/gcn/time.c: New file.
* configure.ac: Add amdgcn*-*-*.
* configure: Regenerate.
* configure.tgt: Add amdgcn*-*-*.
* libgomp-plugin.h (offload_target_type): Add OFFLOAD_TARGET_TYPE_GCN.
* libgomp.h (gcn_thrs): Add amdgcn variant.
(set_gcn_thrs): Likewise.
(gomp_thread): Likewise.
* oacc-int.h (goacc_thread): Likewise.
* oacc-target.c: New file.
* openacc.f90 (acc_device_gcn): New parameter.
* openacc.h (acc_device_t): Add acc_device_gcn.
* team.c (gomp_free_pool_helper): Add amdgcn support.

Co-Authored-By: Julian Brown <julian@codesourcery.com>
Co-Authored-By: Kwok Cheung Yeung <kcy@codesourcery.com>
Co-Authored-By: Tom de Vries <tom@codesourcery.com>
From-SVN: r278135

5 years agoAdd device number to GOMP_OFFLOAD_openacc_async_construct
Andrew Stubbs [Wed, 13 Nov 2019 12:37:59 +0000 (12:37 +0000)]
Add device number to GOMP_OFFLOAD_openacc_async_construct

2019-11-13  Andrew Stubbs  <ams@codesourcery.com>
    Julian Brown  <julian@codesourcery.com>

libgomp/
* libgomp-plugin.h (GOMP_OFFLOAD_openacc_async_construct): Add int
parameter.
* oacc-async.c (lookup_goacc_asyncqueue): Pass device number to the
queue constructor.
* oacc-host.c (host_openacc_async_construct): Add device parameter.
* plugin/plugin-nvptx.c (GOMP_OFFLOAD_openacc_async_construct): Add
device parameter.

Co-Authored-By: Julian Brown <julian@codesourcery.com>
From-SVN: r278134

5 years agoGCN mkoffload
Andrew Stubbs [Wed, 13 Nov 2019 12:37:54 +0000 (12:37 +0000)]
GCN mkoffload

2019-11-13  Andrew Stubbs  <ams@codesourcery.com>
    Kwok Cheung Yeung  <kcy@codesourcery.com>
    Julian Brown  <julian@codesourcery.com>
    Tom de Vries  <tom@codesourcery.com>

gcc/
* config/gcn/mkoffload.c: New file.
* config/gcn/offload.h: New file.

From-SVN: r278133

5 years agoMove generic libgomp files from nvptx to accel
Andrew Stubbs [Wed, 13 Nov 2019 12:37:50 +0000 (12:37 +0000)]
Move generic libgomp files from nvptx to accel

2019-11-13  Andrew Stubbs  <ams@codesourcery.com>

libgomp/
* configure.tgt (nvptx*-*-*): Add "accel" directory.
* config/nvptx/libgomp-plugin.c: Move ...
* config/accel/libgomp-plugin.c: ... to here.
* config/nvptx/lock.c: Move ...
* config/accel/lock.c: ... to here.
* config/nvptx/mutex.c: Move ...
* config/accel/mutex.c: ... to here.
* config/nvptx/mutex.h: Move ...
* config/accel/mutex.h: ... to here.
* config/nvptx/oacc-async.c: Move ...
* config/accel/oacc-async.c: ... to here.
* config/nvptx/oacc-cuda.c: Move ...
* config/accel/oacc-cuda.c: ... to here.
* config/nvptx/oacc-host.c: Move ...
* config/accel/oacc-host.c: ... to here.
* config/nvptx/oacc-init.c: Move ...
* config/accel/oacc-init.c: ... to here.
* config/nvptx/oacc-mem.c: Move ...
* config/accel/oacc-mem.c: ... to here.
* config/nvptx/oacc-plugin.c: Move ...
* config/accel/oacc-plugin.c: ... to here.
* config/nvptx/omp-lock.h: Move ...
* config/accel/omp-lock.h: ... to here.
* config/nvptx/openacc.f90: Move ...
* config/accel/openacc.f90: ... to here.
* config/nvptx/pool.h: Move ...
* config/accel/pool.h: ... to here.
* config/nvptx/proc.c: Move ...
* config/accel/proc.c: ... to here.
* config/nvptx/ptrlock.c: Move ...
* config/accel/ptrlock.c: ... to here.
* config/nvptx/ptrlock.h: Move ...
* config/accel/ptrlock.h: ... to here.
* config/nvptx/sem.c: Move ...
* config/accel/sem.c: ... to here.
* config/nvptx/sem.h: Move ...
* config/accel/sem.h: ... to here.
* config/nvptx/thread-stacksize.h: Move ...
* config/accel/thread-stacksize.h: ... to here.

From-SVN: r278132

5 years agoMove gcn-run heap into GPU memory.
Andrew Stubbs [Wed, 13 Nov 2019 12:37:44 +0000 (12:37 +0000)]
Move gcn-run heap into GPU memory.

2019-11-13  Andrew Stubbs  <ams@codesourcery.com>

gcc/
* config/gcn/gcn-run.c (heap_region): New global variable.
(struct hsa_runtime_fn_info): Add hsa_memory_assign_agent_fn.
(init_hsa_runtime_functions): Initialize hsa_memory_assign_agent.
(get_kernarg_region): Move contents to ....
(get_memory_region): .... here.
(get_heap_region): New function.
(init_device): Initialize the heap_region.
(device_malloc): Add region parameter.
(struct kernargs): Move heap ....
(heap): ... to global scope.
(main): Allocate heap separate to kernargs.

From-SVN: r278131

5 years agoc-ada-spec.c (get_underlying_decl): Do not look through typedefs.
Eric Botcazou [Wed, 13 Nov 2019 11:58:46 +0000 (11:58 +0000)]
c-ada-spec.c (get_underlying_decl): Do not look through typedefs.

* c-ada-spec.c (get_underlying_decl): Do not look through typedefs.
(dump_forward_type): Do not generate a declaration for function types.
(dump_nested_type) <ARRAY_TYPE>: Do not generate a nested declaration
of the component type if it is declared in another file.

From-SVN: r278129

5 years agoPR fortran/92470 Fixes for CFI_address
Tobias Burnus [Wed, 13 Nov 2019 11:13:57 +0000 (11:13 +0000)]
PR fortran/92470 Fixes for CFI_address

        libgfortran/
        PR fortran/92470
        * runtime/ISO_Fortran_binding.c (CFI_establish): Set lower_bound to 0
        also for CFI_attribute_other.

        gcc/testsuite/
        PR fortran/92470
        * gfortran.dg/ISO_Fortran_binding_1.c (establish_c): Add assert for
        lower_bound == 0.

From-SVN: r278128

5 years agoipa-prop.c (ipa_print_node_jump_functions, [...]): Print info about missing summaries.
Jan Hubicka [Wed, 13 Nov 2019 10:11:59 +0000 (11:11 +0100)]
ipa-prop.c (ipa_print_node_jump_functions, [...]): Print info about missing summaries.

* ipa-prop.c (ipa_print_node_jump_functions,
ipa_print_node_params): Print info about missing summaries.

From-SVN: r278127

5 years agoEnable libsanitizer build on riscv64
Andreas Schwab [Wed, 13 Nov 2019 09:45:15 +0000 (09:45 +0000)]
Enable libsanitizer build on riscv64

* configure.tgt (riscv64-*-linux*): Enable build.

From-SVN: r278126

5 years agoAccount for the cost of generating loop masks
Richard Sandiford [Wed, 13 Nov 2019 09:12:17 +0000 (09:12 +0000)]
Account for the cost of generating loop masks

We didn't take the cost of generating loop masks into account, and so
tended to underestimate the cost of loops that need multiple masks.

2019-11-13  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vect-loop.c (vect_estimate_min_profitable_iters): Include
the cost of generating loop masks.

gcc/testsuite/
* gcc.target/aarch64/sve/mask_struct_store_3.c: Add
-fno-vect-cost-model.
* gcc.target/aarch64/sve/mask_struct_store_3_run.c: Likewise.
* gcc.target/aarch64/sve/peel_ind_2.c: Likewise.
* gcc.target/aarch64/sve/peel_ind_2_run.c: Likewise.
* gcc.target/aarch64/sve/peel_ind_3.c: Likewise.
* gcc.target/aarch64/sve/peel_ind_3_run.c: Likewise.

From-SVN: r278125

5 years agoAvoid accounting for non-existent vector loop versioning
Richard Sandiford [Wed, 13 Nov 2019 09:05:59 +0000 (09:05 +0000)]
Avoid accounting for non-existent vector loop versioning

vect_analyze_loop_costing uses two profitability thresholds: a runtime
one and a static compile-time one.  The runtime one is simply the point
at which the vector loop is cheaper than the scalar loop, while the
static one also takes into account the cost of choosing between the
scalar and vector loops at runtime.  We compare this static cost against
the expected execution frequency to decide whether it's worth generating
any vector code at all.

However, we never reclaimed the cost of applying the runtime threshold
if it turned out that the vector code can always be used.  And we only
know whether that's true once we've calculated what the runtime
threshold would be.

2019-11-13  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vectorizer.h (vect_apply_runtime_profitability_check_p):
New function.
* tree-vect-loop-manip.c (vect_loop_versioning): Use it.
* tree-vect-loop.c (vect_analyze_loop_2): Likewise.
(vect_transform_loop): Likewise.
(vect_analyze_loop_costing): Don't take the cost of versioning
into account for the static profitability threshold if it turns
out that no versioning is needed.

From-SVN: r278124

5 years agoipa.c (cgraph_build_static_cdtor): Pass optimization_default_node and target_option_d...
Jan Hubicka [Wed, 13 Nov 2019 09:05:14 +0000 (10:05 +0100)]
ipa.c (cgraph_build_static_cdtor): Pass optimization_default_node and target_option_default_node to get...

* ipa.c (cgraph_build_static_cdtor): Pass optimization_default_node
and target_option_default_node to get -fprofile-generate ctors working
right with LTO.

From-SVN: r278123

5 years agoDon't assign a cost to vectorizable_assignment
Richard Sandiford [Wed, 13 Nov 2019 09:03:07 +0000 (09:03 +0000)]
Don't assign a cost to vectorizable_assignment

vectorizable_assignment handles true SSA-to-SSA copies (which hopefully
we don't see in practice) and no-op conversions that are required
to maintain correct gimple, such as changes between signed and
unsigned types.  These cases shouldn't generate any code and so
shouldn't count against either the scalar or vector costs.

Later patches test this, but it seemed worth splitting out.

2019-11-13  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vectorizer.h (vect_nop_conversion_p): Declare.
* tree-vect-stmts.c (vect_nop_conversion_p): New function.
(vectorizable_assignment): Don't add a cost for nop conversions.
* tree-vect-loop.c (vect_compute_single_scalar_iteration_cost):
Likewise.
* tree-vect-slp.c (vect_bb_slp_scalar_cost): Likewise.

From-SVN: r278122

5 years agoFix vectorizable_conversion costs
Richard Sandiford [Wed, 13 Nov 2019 09:00:53 +0000 (09:00 +0000)]
Fix vectorizable_conversion costs

This patch makes two tweaks to vectorizable_conversion.  The first
is to use "modifier" to distinguish between promotion, demotion,
and neither promotion nor demotion, rather than using a code for
some cases and "modifier" for others.  The second is to take ncopies
into account for the promotion and demotion costs; previously we gave
multiple copies the same cost as a single copy.

Later patches test this, but it seemed worth splitting out.

2019-11-13  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* tree-vect-stmts.c (vect_model_promotion_demotion_cost): Take the
number of ncopies as an additional argument.
(vectorizable_conversion): Update call accordingly.  Use "modifier"
to check whether a conversion is between vectors with the same
numbers of units.

From-SVN: r278121

5 years ago[AArch64] Use aarch64_sve_int_mode in SVE ACLE code
Richard Sandiford [Wed, 13 Nov 2019 08:56:54 +0000 (08:56 +0000)]
[AArch64] Use aarch64_sve_int_mode in SVE ACLE code

This is a like-for-like change at the moment, but is a prerequisite
for removing mode_for_int_vector.

2019-11-13  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/aarch64/aarch64-sve-builtins-functions.h
(unary_count::expand): Use aarch64_sve_int_mode instead of
mode_for_int_vector.

From-SVN: r278120

5 years ago[C++] Fix interaction between aka changes and DR1558 (PR92206)
Richard Sandiford [Wed, 13 Nov 2019 08:42:56 +0000 (08:42 +0000)]
[C++] Fix interaction between aka changes and DR1558 (PR92206)

One of the changes in r277281 was to make the typedef variant
handling in strip_typedefs pass the raw DECL_ORIGINAL_TYPE to the
recursive call, instead of applying TYPE_MAIN_VARIANT first.
This PR shows that that interacts badly with the implementation
of DR1558, because we then refuse to strip aliases with dependent
template parameters and trip:

  gcc_assert (!typedef_variant_p (result)
      || ((flags & STF_USER_VISIBLE)
  && !user_facing_original_type_p (result)));

Keeping the current behaviour but suppressing the ICE leads to a
duplicate error (the dg-bogus in the first test), so that didn't
seem like a good fix.

I assume keeping the alias should never actually be necessary for
DECL_ORIGINAL_TYPEs, because it will already have been checked
somewhere, even for implicit TYPE_DECLs.  This patch therefore
passes a flag to say that we can safely strip aliases with
dependent template parameters.

2019-11-13  Richard Sandiford  <richard.sandiford@arm.com>

gcc/cp/
PR c++/92206
* cp-tree.h (STF_STRIP_DEPENDENT): New constant.
* tree.c (strip_typedefs): Add STF_STRIP_DEPENDENT to the flags
when calling strip_typedefs recursively on a DECL_ORIGINAL_TYPE.
Don't apply the fix for DR1558 in that case; allow aliases with
dependent template parameters to be stripped instead.

gcc/testsuite/
PR c++/92206
* g++.dg/cpp0x/alias-decl-pr92206-1.C: New test.
* g++.dg/cpp0x/alias-decl-pr92206-2.C: Likewise.
* g++.dg/cpp0x/alias-decl-pr92206-3.C: Likewise.

From-SVN: r278119

5 years agoUpdate comment in opts.c.
Martin Liska [Wed, 13 Nov 2019 08:40:46 +0000 (09:40 +0100)]
Update comment in opts.c.

2019-11-13  Martin Liska  <mliska@suse.cz>

* opts.c: Update comment about OPT_LEVELS_2_PLUS_SPEED_ONLY.

From-SVN: r278118

5 years agoFix params.exp by parsing output of --help=params -Q.
Martin Liska [Wed, 13 Nov 2019 08:40:18 +0000 (09:40 +0100)]
Fix params.exp by parsing output of --help=params -Q.

2019-11-13  Martin Liska  <mliska@suse.cz>

* gcc.dg/params/params.exp: Restore test by parsing output
of --help=params -Q.

From-SVN: r278117

5 years agoRemove leftover call to finalize_options_struct.
Martin Liska [Wed, 13 Nov 2019 08:39:53 +0000 (09:39 +0100)]
Remove leftover call to finalize_options_struct.

2019-11-13  Martin Liska  <mliska@suse.cz>

* tree-streamer-in.c (lto_input_ts_function_decl_tree_pointers):
Remove call to finalize_options_struct.

From-SVN: r278116

5 years agore PR target/92055 ([avr] Support 64-bit double)
Georg-Johann Lay [Wed, 13 Nov 2019 08:18:35 +0000 (08:18 +0000)]
re PR target/92055 ([avr] Support 64-bit double)

PR target/92055
* config/avr/t-avr (avr-mcus): Do not depend on
$(srcdir)/config/avr/t-multilib.

From-SVN: r278115

5 years agoFortran] Use proper type for hidden is-present argument
Tobias Burnus [Wed, 13 Nov 2019 08:09:42 +0000 (09:09 +0100)]
Fortran] Use proper type for hidden is-present argument

        gcc/fortran/
        * trans-expr.c (gfc_conv_procedure_call): Fold hidden
        is-present argument to the right type.

From-SVN: r278114

5 years agore PR target/92473 (test pr92324-2.c fails on arm and aarch64)
Richard Biener [Wed, 13 Nov 2019 08:04:31 +0000 (08:04 +0000)]
re PR target/92473 (test pr92324-2.c fails on arm and aarch64)

2019-11-13  Richard Biener  <rguenther@suse.de>

PR tree-optimization/92473
* tree-vect-loop.c (vect_create_epilog_for_reduction): Perform
direct optab reduction in the correct type.

From-SVN: r278113

5 years agoAdd option -fweb for pr47763.c
Jiufu Guo [Wed, 13 Nov 2019 05:04:22 +0000 (05:04 +0000)]
Add option -fweb for pr47763.c

This case is testing 'web' on ignore naked clobber.
-funroll-loops no longer implies -fweb for powerpc.
So, add -fweb to enable 'web' for this case.

gcc.testsuite/
2019-11-13  Jiufu Guo  <guojiufu@linux.ibm.com>

PR target/92465
* gcc.dg/pr47763.c: Add option -fweb.

From-SVN: r278112

5 years agoDaily bump.
GCC Administrator [Wed, 13 Nov 2019 00:16:25 +0000 (00:16 +0000)]
Daily bump.

From-SVN: r278111

5 years agors6000: Use ULL on big hexadecimal literal
Segher Boessenkool [Tue, 12 Nov 2019 21:56:13 +0000 (22:56 +0100)]
rs6000: Use ULL on big hexadecimal literal

C++98 does not have long long int, and does not use (unsigned) long
long int for hexadecimal literals.  So let's use an ULL suffix here,
which is still not strict C++98, but which works with more compilers.

* config/rs6000/rs6000.md (rs6000_set_fpscr_drn): Use ULL on big
hexadecimal literal.

From-SVN: r278107

5 years ago[C++ PATCH] Merge some using-decl handling
Nathan Sidwell [Tue, 12 Nov 2019 21:21:13 +0000 (21:21 +0000)]
[C++ PATCH] Merge some using-decl handling

https://gcc.gnu.org/ml/gcc-patches/2019-11/msg00971.html
gcc/cp/
* name-lookup.c (lookup_using_decl): New function, merged from ...
(do_class_using_decl): ... here.  Call it.  And ...
(finish_nonmember_using_decl): ... here.  Call it.

gcc/testsuite/
* g++.dg/cpp0x/using-enum-2.C: Adjust expected error text.
* g++.dg/cpp0x/using-enum-3.C: Likewise.
* g++.dg/lookup/using4.C: Likewise.
* g++.dg/lookup/using7.C: Likewise.
* g++.dg/template/using12.C: Likewise.
* g++.dg/template/using18.C: Likewise.
* g++.dg/template/using22.C: Likewise.

From-SVN: r278106

5 years agore PR fortran/81651 (Enhancement request: have f951 print out fully qualified module...
Harald Anlauf [Tue, 12 Nov 2019 21:14:19 +0000 (21:14 +0000)]
re PR fortran/81651 (Enhancement request: have f951 print out fully qualified module file name)

2019-11-12  Harald Anlauf  <anlauf@gmx.de>

PR fortran/81651
* module.c (gzopen_included_file, gzopen_included_file_1)
(gzopen_intrinsic_module, bad_module, gfc_use_module): Use fully
qualified module path for error reporting.

From-SVN: r278105

5 years agotestsuite: Add testcases for PR92449
Segher Boessenkool [Tue, 12 Nov 2019 21:05:24 +0000 (22:05 +0100)]
testsuite: Add testcases for PR92449

PR target/92449
* gcc.c-torture/compile/pr92449.c: New test.
* gcc.target/powerpc/pr92449-1.c: New test.

From-SVN: r278104

5 years agors6000: Handle unordered for xscmpexp[dq]p without NaNs (PR92449)
Segher Boessenkool [Tue, 12 Nov 2019 21:02:03 +0000 (22:02 +0100)]
rs6000: Handle unordered for xscmpexp[dq]p without NaNs (PR92449)

* config/rs6000/vsx.md (xscmpexpdp_<code> for CMP_TEST): Handle
UNORDERED if !HONOR_NANS (DFmode).
(xscmpexpqp_<code>_<mode> for CMP_TEST and IEEE128): Handle UNORDERED
if !HONOR_NANS (<MODE>mode).

From-SVN: r278103