dump reason for throwing away SLP instance

This adds dumping to vect_slp_analyze_node_alignment when it fails
an SLP instance due to shared vector type conflicts.

2020-10-28  Richard Biener  <rguenther@suse.de>

* tree-vect-data-refs.c (vect_slp_analyze_node_alignment):
Dump when vect_update_shared_vectype fails.

libstdc++: Fix name clash with _Cosh in QNX headers [PR 95592]

This replaces unqualified names like _Cosh with struct std::_Cosh to
ensure there is no ambiguity with other entities with the same name.

libstdc++-v3/ChangeLog:

PR libstdc++/95592
* include/bits/valarray_after.h (_DEFINE_EXPR_UNARY_OPERATOR)
(_DEFINE_EXPR_BINARY_OPERATOR, _DEFINE_EXPR_BINARY_FUNCTION):
Use elaborated-type-specifier and qualified-id to avoid
ambiguities with QNX system headers.
* testsuite/26_numerics/valarray/95592.cc: New test.

libstdc++: Make std::span layout-compatible with struct iovec [PR 95609]

This change reorders the data members of std::span so that span<byte> is
layout-compatible with common implementations of struct iovec. This will
allow span<byte> to be used directly in places that use a struct iovec
to do scatter-gather I/O.

It's important to note that POSIX doesn't specify the order of members
in iovec. Also the equivalent type on Windows has members in the other
order, and uses type ULONG (which is always 32-bit whereas size_t is
64-bit for Win64). So this change will only help for certain targets and
an indirection between std::span and I/O system calls will still be
needed for the general case.

libstdc++-v3/ChangeLog:

PR libstdc++/95609
* include/std/span (span): Reorder data members to match common
implementations of struct iovec.
* testsuite/23_containers/span/layout_compat.cc: New test.

aarch64: Add vstN_lane_bf16 + vstNq_lane_bf16 intrinsics

gcc/ChangeLog

2020-10-19  Andrea Corallo  <andrea.corallo@arm.com>

* config/aarch64/arm_neon.h (__ST2_LANE_FUNC, __ST3_LANE_FUNC)
(__ST4_LANE_FUNC): Rename the macro generating the 'q' variants
into __ST2Q_LANE_FUNC, __ST2Q_LANE_FUNC, __ST2Q_LANE_FUNC so they
all can be undefed at the and of the file.
(vst2_lane_bf16, vst2q_lane_bf16, vst3_lane_bf16, vst3q_lane_bf16)
(vst4_lane_bf16, vst4q_lane_bf16): Add new intrinsics.

gcc/testsuite/ChangeLog

2020-10-19  Andrea Corallo  <andrea.corallo@arm.com>

* gcc.target/aarch64/advsimd-intrinsics/arm-neon-ref.h
(hbfloat16_t): Define type.
(CHECK_FP): Make it working for bfloat types.
* gcc.target/aarch64/advsimd-intrinsics/bf16_vstN_lane_1.c: New file.
* gcc.target/aarch64/advsimd-intrinsics/bf16_vstN_lane_2.c: Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vst2_lane_bf16_indices_1.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vst2q_lane_bf16_indices_1.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vst3_lane_bf16_indices_1.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vst3q_lane_bf16_indices_1.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vst4_lane_bf16_indices_1.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vst4q_lane_bf16_indices_1.c:
Likewise.

aarch64: Add bfloat16 vldN_lane_bf16 + vldNq_lane_bf16 intrisics

gcc/ChangeLog

2020-10-15  Andrea Corallo  <andrea.corallo@arm.com>

* config/aarch64/arm_neon.h (__LD2_LANE_FUNC, __LD3_LANE_FUNC)
(__LD4_LANE_FUNC): Rename the macro generating the 'q' variants
into __LD2Q_LANE_FUNC, __LD2Q_LANE_FUNC, __LD2Q_LANE_FUNC so they
all can be undefed at the and of the file.
(vld2_lane_bf16, vld2q_lane_bf16, vld3_lane_bf16, vld3q_lane_bf16)
(vld4_lane_bf16, vld4q_lane_bf16): Add new intrinsics.

gcc/testsuite/ChangeLog

2020-10-15  Andrea Corallo  <andrea.corallo@arm.com>

* gcc.target/aarch64/advsimd-intrinsics/bf16_vldN_lane_1.c: New
testcase.
* gcc.target/aarch64/advsimd-intrinsics/bf16_vldN_lane_2.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vld2_lane_bf16_indices_1.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vld2q_lane_bf16_indices_1.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vld3_lane_bf16_indices_1.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vld3q_lane_bf16_indices_1.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vld4_lane_bf16_indices_1.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vld4q_lane_bf16_indices_1.c:
Likewise.

[PR97504] riscv needs wraplf for aux_long_long_float too

riscv is another platform on which GNAT maps Long_Long_Float to double
rather than long double, so we have to explicitly avoid the long
double intrinsics.

for  gcc/ada/ChangeLog

PR ada/97504
* Makefile.rtl (LIBGNAT_TARGET_PAIRS> <riscv*-*-*>: Use wraplf
version of Aux_Long_Long_Float.

openmp: Parsing and some semantic analysis of OpenMP allocate clause

This patch adds parsing of OpenMP allocate clause, but still ignores
it during OpenMP lowering where we should for privatized variables
with allocate clause use the corresponding allocators rather than
allocating them on the stack.

2020-10-28  Jakub Jelinek  <jakub@redhat.com>

gcc/
* tree-core.h (enum omp_clause_code): Add OMP_CLAUSE_ALLOCATE.
* tree.h (OMP_CLAUSE_ALLOCATE_ALLOCATOR,
OMP_CLAUSE_ALLOCATE_COMBINED): Define.
* tree.c (omp_clause_num_ops, omp_clause_code_name): Add allocate
clause.
(walk_tree_1): Handle OMP_CLAUSE_ALLOCATE.
* tree-pretty-print.c (dump_omp_clause): Likewise.
* gimplify.c (gimplify_scan_omp_clauses, gimplify_adjust_omp_clauses,
gimplify_omp_for): Likewise.
* tree-nested.c (convert_nonlocal_omp_clauses,
convert_local_omp_clauses): Likewise.
* omp-low.c (scan_sharing_clauses): Likewise.
gcc/c-family/
* c-pragma.h (enum pragma_omp_clause): Add PRAGMA_OMP_CLAUSE_ALLOCATE.
* c-omp.c: Include bitmap.h.
(c_omp_split_clauses): Handle OMP_CLAUSE_ALLOCATE.
gcc/c/
* c-parser.c (c_parser_omp_clause_name): Handle allocate.
(c_parser_omp_clause_allocate): New function.
(c_parser_omp_all_clauses): Handle PRAGMA_OMP_CLAUSE_ALLOCATE.
(OMP_FOR_CLAUSE_MASK, OMP_SECTIONS_CLAUSE_MASK,
OMP_PARALLEL_CLAUSE_MASK, OMP_SINGLE_CLAUSE_MASK,
OMP_TASK_CLAUSE_MASK, OMP_TASKGROUP_CLAUSE_MASK,
OMP_DISTRIBUTE_CLAUSE_MASK, OMP_TEAMS_CLAUSE_MASK,
OMP_TARGET_CLAUSE_MASK, OMP_TASKLOOP_CLAUSE_MASK): Add
PRAGMA_OMP_CLAUSE_ALLOCATE.
* c-typeck.c (c_finish_omp_clauses): Handle OMP_CLAUSE_ALLOCATE.
gcc/cp/
* parser.c (cp_parser_omp_clause_name): Handle allocate.
(cp_parser_omp_clause_allocate): New function.
(cp_parser_omp_all_clauses): Handle PRAGMA_OMP_CLAUSE_ALLOCATE.
(OMP_FOR_CLAUSE_MASK, OMP_SECTIONS_CLAUSE_MASK,
OMP_PARALLEL_CLAUSE_MASK, OMP_SINGLE_CLAUSE_MASK,
OMP_TASK_CLAUSE_MASK, OMP_TASKGROUP_CLAUSE_MASK,
OMP_DISTRIBUTE_CLAUSE_MASK, OMP_TEAMS_CLAUSE_MASK,
OMP_TARGET_CLAUSE_MASK, OMP_TASKLOOP_CLAUSE_MASK): Add
PRAGMA_OMP_CLAUSE_ALLOCATE.
* semantics.c (finish_omp_clauses): Handle OMP_CLAUSE_ALLOCATE.
* pt.c (tsubst_omp_clauses): Likewise.
gcc/testsuite/
* c-c++-common/gomp/allocate-1.c: New test.
* c-c++-common/gomp/allocate-2.c: New test.
* c-c++-common/gomp/clauses-1.c (omp_allocator_handle_t): New typedef.
(foo, bar, baz): Add allocate clauses where allowed.

openmp: Implicitly discover declare target for variants of declare variant calls

This marks all variants of declare variant also declare target if the base
functions are called directly in target regions or declare target functions.

2020-10-28  Jakub Jelinek  <jakub@redhat.com>

gcc/
* omp-offload.c (omp_declare_target_tgt_fn_r): Handle direct calls to
declare variant base functions.
libgomp/
* testsuite/libgomp.c/target-42.c: New test.

xfail and improve some failing libgomp tests [PR81690]

With the patch I've posted today to fix up declare variant LTO handling,
Tobias reported the patch still doesn't work, and there are two
reasons for that.
One is that when the base function is marked implicitly as declare target,
we don't mark also implicitly the variants.  I'll need to ask on omp-lang
about details for that, but generally the compiler should do it some way.
The other one is that the way base_delay is written, it will always
call the usleep function, which is undesirable for nvptx.  While the
compiler will replace all direct calls to base_delay to nvptx_delay,
the base_delay definition which calls usleep stays.

2020-10-28  Jakub Jelinek  <jakub@redhat.com>
    Tom de Vries  <tdevries@suse.de>

PR testsuite/81690
* testsuite/libgomp.c/usleep.h: New file.
* testsuite/libgomp.c/target-32.c: Include usleep.h.
(main): Use tgt_usleep instead of usleep.
* testsuite/libgomp.c/thread-limit-2.c: Include usleep.h.
(main): Use tgt_usleep instead of usleep.

lto: LTO cgraph support for late declare variant resolution [PR96680]

> I've tried to add the saving/restoring next to ipa refs saving/restoring, as
> the declare variant alt stuff is kind of extension of those, unfortunately
> following doesn't compile, because I need to also write or read a tree there
> (ctx is a portion of DECL_ATTRIBUTES of the base function), but the ipa refs
> write/read back functions don't have arguments that can be used for that.

This patch adds the streaming out and in of those omp_declare_variant_alt
hash table on the side data for the declare_variant_alt cgraph_nodes and
treats for LTO purposes the declare_variant_alt nodes (which have no body)
as if they contained a body that calls all the possible variants.
After IPA all the calls to these magic declare_variant_alt calls are
replaced with call to one of the variant depending on which one has the
highest score in the context.

2020-10-28  Jakub Jelinek  <jakub@redhat.com>

PR lto/96680
gcc/
* lto-streamer.h (omp_lto_output_declare_variant_alt,
omp_lto_input_declare_variant_alt): Declare variant.
* symtab.c (symtab_node::get_partitioning_class): Return
SYMBOL_DUPLICATE for declare_variant_alt nodes.
* passes.c (ipa_write_summaries): Add declare_variant_alt to
partition.
* lto-cgraph.c (output_refs): Call omp_lto_output_declare_variant_alt
on declare_variant_alt nodes.
(input_refs): Call omp_lto_input_declare_variant_alt on
declare_variant_alt nodes.
* lto-streamer-out.c (output_function): Don't call
collect_block_tree_leafs if DECL_INITIAL is error_mark_node.
(lto_output): Call output_function even for declare_variant_alt
nodes.
* omp-general.c (omp_lto_output_declare_variant_alt,
omp_lto_input_declare_variant_alt): New functions.
gcc/lto/
* lto-common.c (lto_fixup_prevailing_decls): Don't use
LTO_NO_PREVAIL on TREE_LIST's TREE_PURPOSE.
* lto-partition.c (lto_balanced_map): Treat declare_variant_alt
nodes like definitions.
libgomp/
* testsuite/libgomp.c/declare-variant-1.c: New test.

wide-int: Fix up set_bit_large

> >> wide_int new_lb = wi::set_bit (r.lower_bound (0), 127)
> >>
> >> and creates the value:
> >>
> >> p new_lb
> >> {<wide_int_storage> = {val = {-65535, -1, 0}, len = 2, precision = 128},
> >> static is_sign_extended = true}
> >
> > This is non-canonical and so invalid, if the low HWI has the MSB set
> > and the high HWI is -1, it should have been just
> > val = {-65535}, len = 1, precision = 128}
> >
> > I guess the bug is that wi::set_bit_large doesn't call canonize.
>
> Yeah, looks like a micro-optimisation gone wrong.

2020-10-28  Jakub Jelinek  <jakub@redhat.com>

* wide-int.cc (wi::set_bit_large): Call canonize unless setting
msb bit and clearing bits above it.

[RS6000] power10 scan-assembler tests

On power10 these are "dg-do run" tests, so need -save-temps for the
assembler scanning.

* gcc.target/powerpc/vsx-load-element-extend-char.c: Add -save-temps.
* gcc.target/powerpc/vsx-load-element-extend-int.c: Likewise.
* gcc.target/powerpc/vsx-load-element-extend-longlong.c: Likewise.
* gcc.target/powerpc/vsx-load-element-extend-short.c: Likewise.
* gcc.target/powerpc/vsx-store-element-truncate-char.c: Likewise.
* gcc.target/powerpc/vsx-store-element-truncate-int.c: Likewise.
* gcc.target/powerpc/vsx-store-element-truncate-longlong.c: Likewise.
* gcc.target/powerpc/vsx-store-element-truncate-short.c: Likewise.

[RS6000] dg-do !compile and scan-assembler

These tests never checked assembly, because .s files were not
produced.  One was looking for the wrong instructions.

A typical error log
PASS: gcc.target/powerpc/vec-permute-ext-runnable.c (test for excess errors)
gcc.target/powerpc/vec-permute-ext-runnable.c output file does not exist
UNRESOLVED: gcc.target/powerpc/vec-permute-ext-runnable.c scan-assembler-times \\mpermx\\M 10

* gcc.target/powerpc/vec-blend-runnable.c: Add save-temps.
* gcc.target/powerpc/vec-insert-word-runnable.c: Likewise.
* gcc.target/powerpc/vec-permute-ext-runnable.c: Likewise.
* gcc.target/powerpc/vec-replace-word-runnable.c: Likewise.
* gcc.target/powerpc/vec-splati-runnable.c: Likewise.
* gcc.target/powerpc/vec-ternarylogic-3.c: Likewise.
* gcc.target/powerpc/vec-ternarylogic-9.c: Likewise.
* gcc.target/powerpc/vsx_mask-count-runnable.c: Likewise.
* gcc.target/powerpc/vsx_mask-expand-runnable.c: Likewise.
* gcc.target/powerpc/vsx_mask-extract-runnable.c: Likewise.
* gcc.target/powerpc/vsx_mask-move-runnable.c: Likewise.
* gcc.target/powerpc/vec-shift-double-runnable.c: Likewise,
and correct assembly match.

Tweaks to ranger API routines.

Remove the gcc_assert wrappers that contain statements that need to be
executed.
Audit routines to ensure range is set to UNDEFINED when false is returned.

* gimple-range-gori.cc (gori_compute_cache::cache_stmt): Accumulate
return values and only set cache when everything returned true.
* gimple-range.cc (get_tree_range): Set the return range to UNDEFINED
when the range isn't supported.
(gimple_ranger::calc_stmt): Return varying if the type is supported,
even if the stmt processing failed.  False otherwise.
(range_of_builtin_ubsan_call): Don't use gcc_assert.
(range_of_builtin_call): Ditto.
(gimple_ranger::range_of_cond_expr): Ditto.
(gimple_ranger::range_of_expr): Ditto
(gimple_ranger::range_on_entry): Ditto.
(gimple_ranger::range_on_exit): Ditto.
(gimple_ranger::range_on_edge): DItto.
(gimple_ranger::range_of_stmt): Don't use gcc_assert, and initialize
return value to UNDEFINED.

Daily bump.

c: Allow duplicate C2x standard attributes

N2557, accepted into C2x at the October WG14 meeting, removes the
requirement that duplicates of standard attributes cannot appear
within an attribute list (so allowing e.g. [[deprecated, deprecated]],
where previously that was disallowed but [[deprecated]] [[deprecated]]
was OK).  Remove the code checking for this (standard attributes
aren't in any released version of the C standard) and update tests
accordingly.

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

gcc/c/
2020-10-27  Joseph Myers  <joseph@codesourcery.com>

* c-parser.c (c_parser_std_attribute_specifier): Allow duplicate
standard attributes.

gcc/testsuite/
2020-10-27  Joseph Myers  <joseph@codesourcery.com>

* gcc.dg/c2x-attr-deprecated-4.c, gcc.dg/c2x-attr-fallthrough-4.c,
gcc.dg/c2x-attr-maybe_unused-4.c: Allow duplicate attributes.

libgo: update to Go 1.15.3 release

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/265717

Fix PR97497

This works around a limitation of gcse with handling of partially
clobbered registers.  With this patch our GOT pointer register r12 is
not marked as partially clobbered anymore for the -m31 -mzarch -fpic
combination. This is correct since all the bits in r12 we actually
care about are in fact preserved.

gcc/ChangeLog:

PR rtl-optimization/97497
* config/s390/s390.c (s390_hard_regno_call_part_clobbered): Do not
return true for r12 when -fpic is used.

gcc/testsuite/ChangeLog:

* gcc.target/s390/pr97497.c: New test.

PR fortran/97491 - Wrong restriction for VALUE arguments of pure procedures

A dummy argument with the VALUE attribute may be redefined in a PURE or
ELEMENTAL procedure.  Adjust the associated purity check.

gcc/fortran/ChangeLog:

* resolve.c (gfc_impure_variable): A dummy argument with the VALUE
attribute may be redefined without making a procedure impure.

gcc/testsuite/ChangeLog:

* gfortran.dg/value_8.f90: New test.

PPC testsuite fixes

2020-10-27  Carl Love  <cel@us.ibm.com>

gcc/testsuite
* gcc.target/powerpc/vec-blend-runnable.c: Change #ifdef
DEBUG to #if DEBUG.
Fix printf line so it is less then 80 characters long.
* gcc.target/powerpc/vec-insert-word-runnable.c: Change
#ifdef DEBUG to #if DEBUG.
* gcc.target/powerpc/vec-permute-ext-runnable.c: Change
#ifdef DEBUG to #if DEBUG.
* gcc.target/powerpc/vec-replace-word-runnable.c: Change
#ifdef DEBUG to #if DEBUG.
Fix printf lines so they are less then 80 characters long.
* gcc.target/powerpc/vec-shift-double-runnable.c: Change
#ifdef DEBUG to #if DEBUG.

compiler, go/internal/gccgoimporter: export notinheap annotation

This is the gofrontend version of https://golang.org/cl/259297.
This is required now because that change is in the 1.15.3 release.

This requires changing the go/internal/gccgoimporter package, to skip
the new annotation. This change will need to be ported to the gc and
x/tools repos.

For golang/go#41761

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/265258

compiler: remove unused Type::in_heap_ member variable

This member variable was added in https://golang.org/cl/46490, but it
was never used. The code uses Named_type::in_heap_ instead.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/265257

c++: Kill nested_udts

During the implementation of modules I added myself a note to
implement nested_udt handling.  It wasn't obvious to me what they were
for and nothing seemed to be broken in ignoring them.  I figured
something would eventually pop up and I'd add support.  Nothing popped up.

Investigating on trunk discovered 3 places where we look at the
nested-udts.  I couldn't figure how the one in lookup_field_r was
needed -- surely the regular lookup would find the type.  It turned
out that code was unreachable.  So we can delete it.

Next in do_type_instantiation, we walk the nested-utd table
instantiating types.  But those types are also on the TYPE_FIELDS
list, which we've just iterated over.  So I can move the handling into
that loop.

The final use is in handling structs that have a typedef name for
linkage purposes.  Again, we can just iterate over TYPE_FIELDS.  (As
commented, we probably don't need to do even that, as a DR, whose
number I forget, requires such structs to only have C-like things in
them.  But I didn't go that far.

Having removed all the uses of nested-udts, I can remove their
creation from name-lookup, and as the only instance of a binding_table
object, we can remove all that code too.

gcc/cp/
* cp-tree.h (struct lang_type): Delete nested_udts field.
(CLASSTYPE_NESTED_UTDS): Delete.
* name-lookup.h (binding_table, binding_entry): Delete typedefs.
(bt_foreach_proc): Likewise.
(struct binding_entry_s): Delete.
(SCOPE_DEFAULT_HT_SIZE, CLASS_SCOPE_HT_SIZE)
(NAMESPACE_ORDINARY_HT_SIZE, NAMESPACE_STD_HT_SIZE)
(GLOBAL_SCOPE_HT_SIZE): Delete.
(binding_table_foreach, binding_table_find): Delete declarations.
* name-lookup.c (ENTRY_INDEX): Delete.
(free_binding_entry): Delete.
(binding_entry_make, binding_entry_free): Delete.
(struct binding_table_s): Delete.
(binding_table_construct, binding_table_free): Delete.
(binding_table_new, binding_table_expand): Delete.
(binding_table_insert, binding_table_find): Delete.
(binding_table_foreach): Delete.
(maybe_process_template_type_declaration): Delete
CLASSTYPE_NESTED_UTDS insertion.
(do_pushtag): Likewise.
* decl2.c (bt_reset_linkage_1): Fold into reset_type_linkage_1.
(reset_type_linkage_2, bt_reset_linkage_2): Fold into
reset_type_linkage.
* pt.c (instantiate_class_template_1): Delete NESTED_UTDs comment.
(bt_instantiate_type_proc): Delete.
(do_type_instantiation): Instantiate implicit typedef fields.
Delete NESTED_UTD walk.
* search.c (lookup_field_r): Delete unreachable NESTED_UTD
search.

c++: Small cleanup for do_type_instantiation

In working on a bigger cleanup I noticed some opportunities to make
do_type_instantiation's control flow simpler.

gcc/cp/
* parser.c (cp_parser_explicit_instantiation): Refactor some RAII.
* pt.c (bt_instantiate_type_proc): DATA is the tree, pass type to
do_type_instantiation.
(do_type_instantiation): Require T to be a type.  Refactor for
some RAII.

AArch64: Fix overflow in memcopy expansion on aarch64.

Currently the inline memcpy expansion code for AArch64 is using a signed int
to hold the number of elements to copy.  When you giver give it a value larger
than INT_MAX it will overflow.

The overflow causes the maximum number of instructions we want to expand to
check to fail since this assumes an unsigned number.

This patch changes the maximum isns arithmetic to be unsigned HOST_WIDE_INT.

note that the calculation *must* remained signed as the memcopy issues
overlapping unaligned copies.  This means the pointer must be moved back and
so you need signed arithmetic.

gcc/ChangeLog:

PR target/97535
* config/aarch64/aarch64.c (aarch64_expand_cpymem): Use unsigned
arithmetic in check.

gcc/testsuite/ChangeLog:

PR target/97535
* gcc.target/aarch64/pr97535.c: New test.

aarch64: Add vcopy(q)__lane(q)_bf16 intrinsics

gcc/ChangeLog

2020-10-20  Andrea Corallo  <andrea.corallo@arm.com>

* config/aarch64/arm_neon.h (vcopy_lane_bf16, vcopyq_lane_bf16)
(vcopyq_laneq_bf16, vcopy_laneq_bf16): New intrinsics.

gcc/testsuite/ChangeLog

2020-10-20  Andrea Corallo  <andrea.corallo@arm.com>

* gcc.target/aarch64/advsimd-intrinsics/bf16_vect_copy_lane_1.c:
New test.
* gcc.target/aarch64/advsimd-intrinsics/vcopy_lane_bf16_indices_1.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vcopy_lane_bf16_indices_2.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vcopy_laneq_bf16_indices_1.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vcopy_laneq_bf16_indices_2.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vcopyq_lane_bf16_indices_1.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vcopyq_lane_bf16_indices_2.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vcopyq_laneq_bf16_indices_1.c:
Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vcopyq_laneq_bf16_indices_2.c:
Likewise.

libstdc++: Fix ODR violations caused by <tr1/functional>

The placeholders for std::tr1::bind are defined in an anonymous
namespace, which means they have internal linkage. This will cause ODR
violations when used in function templates (such as std::tr1::bind) from
multiple translation units. Although probably harmless (every definition
will generate identical code, even if technically ill-formed) we can
avoid the ODR violations by reusing the std::placeholder objects as the
std::tr1::placeholder objects.

To make this work, the std::_Placeholder type needs to be defined for
C++98 mode, so that <tr1/functional> can use it. The members of the
std::placeholder namespace must not be defined by <functional> in C++98
mode, because "placeholders", "_1", "_2" etc. are not reserved names in
C++98. Instead they can be declared in <tr1/functional>, because those
names *are* reserved in that header. With the std::placeholders objects
declared, a simple using-directive suffices to redeclare them in
namespace std::tr1::placeholders. This means any use of the TR1
placeholders actually refers to the C++11 placeholders, which are
defined with external linkage and exported from the library, so don't
cause ODR violations.

libstdc++-v3/ChangeLog:

* include/std/functional (std::_Placeholder): Define for C++98
as well as later standards.
* include/tr1/functional (std::placeholders::_1 etc): Declare
for C++98.
(tr1::_Placeholder): Replace with using-declaration for
std::_Placeholder.
(tr1::placeholders::_1 etc.): Replace with using-directive for
std::placeholders.

libstdc++: Remove unused variables in special functions

libstdc++-v3/ChangeLog:

* include/tr1/ell_integral.tcc (__ellint_rf, __ellint_rd)
(__ellint_rc, __ellint_rj): Remove unused variables.
* include/tr1/modified_bessel_func.tcc (__airy): Likewise.

libstdc++: Fix -Wsign-compare warnings in headers

libstdc++-v3/ChangeLog:

* include/bits/locale_conv.h (__str_codecvt_out_all):
Add cast to compare operands of the same signedness.
* include/bits/locale_facets_nonio.tcc
(time_get::_M_extract_wday_or_month): Likewise.
* include/bits/sstream.tcc (basic_stringbuf::overflow):
Likewise.
* include/tr1/legendre_function.tcc (__sph_legendre): Use
unsigned for loop variable.

Extract VX_CPU_PREFIX up into config/vxworks.h

Move VX_CPU_PREFIX to a place where it can be
reused by multiple target ports.

2020-10-21  Olivier Hainque  <hainque@adacore.com>

gcc/
* config/vxworks.h (VX_CPU_PREFIX): #define here.
* config/rs6000/vxworks.h: Remove #definition.

Fix glitch on VX_CPU selection for E6500

Proper macro name is PPCE6500, not E6500.
Introduced accidentally during a pre-commit minor rearrangement.

2020-10-27  Olivier Hainque  <hainque@adacore.com>

gcc/
* config/rs6000/vxworks.h (CPP_SPEC): Fix macro definition
for -mcpu=e6500.

Fix BB store group splitting group size compute

This fixes a mistake in the previous change in this area to what
was desired - figure the largest power-of-two group size fitting
in the matching area.

2020-10-27  Richard Biener  <rguenther@suse.de>

* tree-vect-slp.c (vect_build_slp_instance): Use ceil_log2
to compute maximum group-size.

* gcc.dg/vect/bb-slp-67.c: New testcase.

Fix ipa-modref signature updates

PR ipa/97586
* ipa-modref-tree.h (modref_tree::remap_params): New member function.
* ipa-modref.c (modref_summaries_lto::duplicate): Check that
optimization summaries are not duplicated.
(remap_arguments): Remove.
(modref_transform): Rename to ...
(update_signature): ... this one; handle also lto summary.
(pass_ipa_modref::execute): Update signatures here rather
than in transform hook.

libstdc++: Add missing noexcept to std::from_chars declarations

libstdc++-v3/ChangeLog:

* include/std/charconv (from_chars): Add noexcept to match
definitions in src/c++17/floating_from_chars.cc

libstdc++: Fix directory_iterator exception specification

libstdc++-v3/ChangeLog:

* src/c++17/fs_dir.cc (fs::directory_iterator::operator*):
Add noexcept. Do not throw on precondition violation.

libstdc++: Add noexcept to declaration of path::_List members

libstdc++-v3/ChangeLog:

* include/bits/fs_path.h (path::_List::begin, path::_List::end):
Add noexcept to match definitions in src/c++17/fs_path.cc.

Add tests for PR92942 - missing -Wstringop-overflow for allocations with a negative lower bound size.

gcc/testsuite/ChangeLog:

PR middle-end/92942
* gcc.dg/Wstringop-overflow-56.c: New test.
* gcc.dg/Wstringop-overflow-57.c: Same.

Remove .s file.

gcc/testsuite/ChangeLog:
* gcc.dg/Wstringop-overflow-44.s: Remove.

Combine logical OR ranges properly. pr97567

update testcase to work on 32 bit targets

gcc/testsuite
* gcc.dg/pr97567.c: Update to work with 32 bit targets.

Adjust BB vectorization function splitting

This adjusts the condition when to split at control altering stmts,
only when there's a definition.  It also removes the only use
of --param slp-max-insns-in-bb which a previous change left doing
nothing (but repeatedly print a message for each successive
instruction...).

2020-10-27  Richard Biener  <rguenther@suse.de>

* tree-vect-slp.c (vect_slp_bbs): Remove no-op
slp-max-insns-in-bb check.
(vect_slp_function): Dump when splitting the function.
Adjust the split condition for control altering stmts.
* params.opt (-param=slp-max-insns-in-bb): Remove.
* doc/invoke.texi (-param=slp-max-insns-in-bb): Likewise.

analyzer: don't assume extern const vars are zero-initialized [PR97568]

gcc/analyzer/ChangeLog:
PR analyzer/97568
* region-model.cc (region_model::get_initial_value_for_global):
Move check that !DECL_EXTERNAL from here to...
* region.cc (decl_region::get_svalue_for_initializer): ...here,
using it to reject zero initialization.

gcc/testsuite/ChangeLog:
PR analyzer/97568
* gcc.dg/analyzer/pr97568.c: New test.

analyzer: Change cast from long to intptr_t [PR96608]

Casting to intptr_t states the intent of an integer to pointer cast
more clearly and ensures that the cast causes no loss of precision on
any platforms. LLP64 platforms eg. have a long value of 4 bytes and
pointer values of 8 bytes which may even cause compiler errors.

gcc/analyzer/ChangeLog:
PR analyzer/96608
* store.h (hash): Cast to intptr_t instead of long

analyzer: eliminate non-deterministic behavior

This patch is a followup to the previous one, eliminating
non-determinism in the behavior of the analyzer (rather than just in
the logs), by sorting whenever the result previously depended on
pointer values.  Tested as per the previous patch.

gcc/analyzer/ChangeLog:
* constraint-manager.cc (svalue_cmp_by_ptr): Delete.
(equiv_class::canonicalize): Use svalue::cmp_ptr_ptr instead.
(equiv_class_cmp): Eliminate pointer comparison.
* diagnostic-manager.cc (dedupe_key::comparator): If they are at
the same location, also compare epath ength and pending_diagnostic
kind.
* engine.cc (readability_comparator): If two path_vars have the
same readability, then impose an arbitrary ordering on them.
(worklist::key_t::cmp): If two points have the same plan ordering,
continue the comparison.  Call sm_state_map::cmp rather than
comparing hash values.
* program-state.cc (sm_state_map::entry_t::cmp): New.
(sm_state_map::cmp): New.
* program-state.h (sm_state_map::entry_t::cmp): New decl.
(sm_state_map::elements): New.
(sm_state_map::cmp): New.

analyzer: eliminate non-determinism in logs

This patch and the followup eliminate various forms of non-determinism
in the analyzer due to changing pointer values.

This patch fixes churn seen when diffing analyzer logs.  The patch
avoids embedding pointers in various places, and adds sorting when
dumping hash_set and hash_map for various analyzer types.  Doing so
requires implementing a way to sort svalue instances, and assigning UIDs
to gimple statements.

Tested both patches together via a script that runs a testcase 100 times,
and then using diff and md5sum to verify that the results are consistent
in the face of address space randomization:

FILENAME=$1
rm $FILENAME.*
for i in `seq 1 100`; do
    echo "iteration: $i"
    ./xgcc -B. -fanalyzer -c ../../src/gcc/testsuite/gcc.dg/analyzer/$FILENAME \
       --Wanalyzer-too-complex \
       -fdump-analyzer-supergraph \
       -fdump-analyzer-exploded-graph \
       -fdump-analyzer \
       -fdump-noaddr \
       -fdump-analyzer-exploded-nodes-2
    mv $FILENAME.supergraph.dot $FILENAME.$i.supergraph.dot
    mv $FILENAME.analyzer.txt $FILENAME.$i.analyzer.txt
    mv $FILENAME.supergraph-eg.dot $FILENAME.$i.supergraph-eg.dot
    mv $FILENAME.eg.txt $FILENAME.$i.eg.txt
    mv $FILENAME.eg.dot $FILENAME.$i.eg.dot
done

gcc/analyzer/ChangeLog:
* engine.cc (setjmp_record::cmp): New.
(supernode_cluster::dump_dot): Avoid embedding pointer in cluster
name.
(supernode_cluster::cmp_ptr_ptr): New.
(function_call_string_cluster::dump_dot): Avoid embedding pointer
in cluster name.  Sort m_map when dumping child clusters.
(function_call_string_cluster::cmp_ptr_ptr): New.
(root_cluster::dump_dot): Sort m_map when dumping child clusters.
* program-point.cc (function_point::cmp): New.
(function_point::cmp_ptr): New.
* program-point.h (function_point::cmp): New decl.
(function_point::cmp_ptr): New decl.
* program-state.cc (sm_state_map::print): Sort the values.  Guard
the printing of pointers with !flag_dump_noaddr.
(program_state::prune_for_point): Sort the regions.
(log_set_of_svalues): Sort the values.  Guard the printing of
pointers with !flag_dump_noaddr.
* region-model-manager.cc (log_uniq_map): Sort the values.
* region-model-reachability.cc (dump_set): New function template.
(reachable_regions::dump_to_pp): Use it.
* region-model.h (svalue::cmp_ptr): New decl.
(svalue::cmp_ptr_ptr): New decl.
(setjmp_record::cmp): New decl.
(placeholder_svalue::get_name): New accessor.
(widening_svalue::get_point): New accessor.
(compound_svalue::get_map): New accessor.
(conjured_svalue::get_stmt): New accessor.
(conjured_svalue::get_id_region): New accessor.
(region::cmp_ptrs): Rename to...
(region::cmp_ptr_ptr): ...this.
* region.cc (region::cmp_ptrs): Rename to...
(region::cmp_ptr_ptr): ...this.
* state-purge.cc
(state_purge_per_ssa_name::state_purge_per_ssa_name): Sort
m_points_needing_name when dumping.
* store.cc (concrete_binding::cmp_ptr_ptr): New.
(symbolic_binding::cmp_ptr_ptr): New.
(binding_map::cmp): New.
(get_sorted_parent_regions): Update for renaming of
region::cmp_ptrs to region::cmp_ptr_ptr.
(store::dump_to_pp): Likewise.
(store::to_json): Likewise.
(store::can_merge_p): Sort the base regions before considering
them.
* store.h (concrete_binding::cmp_ptr_ptr): New decl.
(symbolic_binding::cmp_ptr_ptr): New decl.
(binding_map::cmp): New decl.
* supergraph.cc (supergraph::supergraph): Assign UIDs to the
gimple stmts.
* svalue.cc (cmp_cst): New.
(svalue::cmp_ptr): New.
(svalue::cmp_ptr_ptr): New.

analyzer: fix param "analyzer-max-enodes-per-program-point"

This was effectively checking for one beyond the limit, rather than
the limit itself.

Seen when fixing PR analyzer/97514.

gcc/analyzer/ChangeLog:
* engine.cc (exploded_graph::get_or_create_node): Fix off-by-one
when imposing param_analyzer_max_enodes_per_program_point limit.

libstdc++: Include <cstdint> in tests that use std::uintptr_t

libstdc++-v3/ChangeLog:

* testsuite/experimental/memory_resource/new_delete_resource.cc:
Add missing <cstdint> header.
* testsuite/experimental/memory_resource/resource_adaptor.cc:
Likewise.

analyzer: implement region_model::get_representative_path_var for labels

This fixes an ICE seen e.g. with gcc.dg/analyzer/data-model-16.c when
enabling -fdump-analyzer.

gcc/analyzer/ChangeLog:
* region-model.cc (region_model::get_representative_path_var):
Implement case RK_LABEL.
* region-model.h (label_region::get_label): New accessor.

testsuite: restrict test to c++11 and later [PR97590]

2020-10-27  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/97560
PR testsuite/97590
* g++.dg/pr97560.C: Require c++11 effective target and add comment
with PR number.

Refactor array descriptor field access

This refactors the array descriptor component access tree building
to commonize code into new helpers to provide a single place to
fix correctness issues with respect to TBAA.

The only interesting part is the gfc_conv_descriptor_data_get change
to drop broken special-casing of REFERENCE_TYPE desc which, when hit,
would build invalid GENERIC trees, missing an INDIRECT_REF before
subsetting the descriptor with a COMPONENT_REF.

2020-10-16  Richard Biener  <rguenther@suse.de>

gcc/fortran/ChangeLog:
* trans-array.c (gfc_get_descriptor_field): New helper.
(gfc_conv_descriptor_data_get): Use it - drop strange
REFERENCE_TYPE handling and make sure we don't trigger it.
(gfc_conv_descriptor_data_addr): Use gfc_get_descriptor_field.
(gfc_conv_descriptor_data_set): Likewise.
(gfc_conv_descriptor_offset): Likewise.
(gfc_conv_descriptor_dtype): Likewise.
(gfc_conv_descriptor_span): Likewise.
(gfc_get_descriptor_dimension): Likewise.
(gfc_conv_descriptor_token): Likewise.
(gfc_conv_descriptor_subfield): New helper.
(gfc_conv_descriptor_stride): Use it.
(gfc_conv_descriptor_lbound): Likewise.
(gfc_conv_descriptor_ubound): Likewise.

SLP vectorize across PHI nodes

This makes SLP discovery detect backedges by seeding the bst_map with
the node to be analyzed so it can be picked up from recursive calls.
This removes the need to discover backedges in a separate walk.

This enables SLP build to handle PHI nodes in full, continuing
the SLP build to non-backedges.  For loop vectorization this
enables outer loop vectorization of nested SLP cycles and for
BB vectorization this enables vectorization of PHIs at CFG merges.

It also turns code generation into a SCC discovery walk to handle
irreducible regions and nodes only reachable via backedges where
we now also fill in vectorized backedge defs.

This requires sanitizing the SLP tree for SLP reduction chains even
more, manually filling the backedge SLP def.

This also exposes the fact that CFG copying (and edge splitting
until I fixed that) ends up with different edge order in the
copy which doesn't play well with the desired 1:1 mapping of
SLP PHI node children and edges for epilogue vectorization.
I've tried to fixup CFG copying here but this really looks
like a dead (or expensive) end there so I've done fixup in
slpeel_tree_duplicate_loop_to_edge_cfg instead for the cases
we can run into.

There's still NULLs in the SLP_TREE_CHILDREN vectors and I'm
not sure it's possible to eliminate them all this stage1 so the
patch has quite some checks for this case all over the place.

Bootstrapped and tested on x86_64-unknown-linux-gnu.  SPEC CPU 2017
and SPEC CPU 2006 successfully built and tested.

2020-10-27  Richard Biener  <rguenther@suse.de>

* gimple.h (gimple_expr_type): For PHIs return the type
of the result.
* tree-vect-loop-manip.c (slpeel_tree_duplicate_loop_to_edge_cfg):
Make sure edge order into copied loop headers line up with the
originals.
* tree-vect-loop.c (vect_transform_cycle_phi): Handle nested
loops with SLP.
(vectorizable_phi): New function.
(vectorizable_live_operation): For BB vectorization compute insert
location here.
* tree-vect-slp.c (vect_free_slp_tree): Deal with NULL
SLP_TREE_CHILDREN entries.
(vect_create_new_slp_node): Add overloads with pre-existing node
argument.
(vect_print_slp_graph): Likewise.
(vect_mark_slp_stmts): Likewise.
(vect_mark_slp_stmts_relevant): Likewise.
(vect_gather_slp_loads): Likewise.
(vect_optimize_slp): Likewise.
(vect_slp_analyze_node_operations): Likewise.
(vect_bb_slp_scalar_cost): Likewise.
(vect_remove_slp_scalar_calls): Likewise.
(vect_get_and_check_slp_defs): Handle PHIs.
(vect_build_slp_tree_1): Handle PHIs.
(vect_build_slp_tree_2): Continue SLP build, following PHI
arguments.  Fix memory leak.
(vect_build_slp_tree): Put stub node into the hash-map so
we can discover cycles directly.
(vect_build_slp_instance): Set the backedge SLP def for
reduction chains.
(vect_analyze_slp_backedges): Remove.
(vect_analyze_slp): Do not call it.
(vect_slp_convert_to_external): Release SLP_TREE_LOAD_PERMUTATION.
(vect_slp_analyze_node_operations): Handle stray failed
backedge defs by failing.
(vect_slp_build_vertices): Adjust leaf condition.
(vect_bb_slp_mark_live_stmts): Handle PHIs, use visited
hash-set to handle cycles.
(vect_slp_analyze_operations): Adjust.
(vect_bb_partition_graph_r): Likewise.
(vect_slp_function): Adjust split condition to allow CFG
merges.
(vect_schedule_slp_instance): Rename to ...
(vect_schedule_slp_node): ... this.  Move DFS walk to ...
(vect_schedule_scc): ... this new function.
(vect_schedule_slp): Call it.  Remove ad-hoc vectorized
backedge fill code.
* tree-vect-stmts.c (vect_analyze_stmt): Call
vectorizable_phi.
(vect_transform_stmt): Likewise.
(vect_is_simple_use): Handle vect_backedge_def.
* tree-vectorizer.c (vec_info::new_stmt_vec_info): Only
set loop header PHIs to vect_unknown_def_type for loop
vectorization.
* tree-vectorizer.h (enum vect_def_type): Add vect_backedge_def.
(enum stmt_vec_info_type): Add phi_info_type.
(vectorizable_phi): Declare.

* gcc.dg/vect/bb-slp-54.c: New test.
* gcc.dg/vect/bb-slp-55.c: Likewise.
* gcc.dg/vect/bb-slp-56.c: Likewise.
* gcc.dg/vect/bb-slp-57.c: Likewise.
* gcc.dg/vect/bb-slp-58.c: Likewise.
* gcc.dg/vect/bb-slp-59.c: Likewise.
* gcc.dg/vect/bb-slp-60.c: Likewise.
* gcc.dg/vect/bb-slp-61.c: Likewise.
* gcc.dg/vect/bb-slp-62.c: Likewise.
* gcc.dg/vect/bb-slp-63.c: Likewise.
* gcc.dg/vect/bb-slp-64.c: Likewise.
* gcc.dg/vect/bb-slp-65.c: Likewise.
* gcc.dg/vect/bb-slp-66.c: Likewise.
* gcc.dg/vect/vect-outer-slp-1.c: Likewise.
* gfortran.dg/vect/O3-bb-slp-1.f: Likewise.
* gfortran.dg/vect/O3-bb-slp-2.f: Likewise.
* g++.dg/vect/simd-11.cc: Likewise.

Avoid uniform lane BB vectorization

This makes sure to use splats early when facing uniform internal
operands in BB SLP discovery rather than relying on the late
heuristincs re-building nodes from scratch.

2020-10-27  Richard Biener  <rguenther@suse.de>

* tree-vect-slp.c (vect_build_slp_tree_2): When vectorizing
BBs splat uniform operands and stop SLP discovery.

* gcc.target/i386/pr95866-1.c: Adjust.

[RS6000] Unsupported test options for -m32

FAIL: gcc.target/powerpc/swaps-p8-22.c (test for excess errors)
Excess errors:
cc1: error: '-mcmodel' not supported in this configuration

* gcc.target/powerpc/swaps-p8-22.c: Enable only for aix and
-m64 linux.

d: Remove the d_critsec_size target hook.

The allocation of mutex objects for synchronized statements has been
moved to the library as of merging druntime 58560d51.  All support code
in the compiler for getting the OS critical section size has been
removed along with it.

Reviewed-on: https://github.com/dlang/dmd/pull/11902
     https://github.com/dlang/druntime/pull/3248

gcc/ChangeLog:

* config/aarch64/aarch64-linux.h (GNU_USER_TARGET_D_CRITSEC_SIZE):
Remove.
* config/glibc-d.c (glibc_d_critsec_size): Likewise.
(TARGET_D_CRITSEC_SIZE): Likewise.
* config/i386/linux-common.h (GNU_USER_TARGET_D_CRITSEC_SIZE):
Likewise.
* config/sol2-d.c (solaris_d_critsec_size): Likewise.
(TARGET_D_CRITSEC_SIZE):  Likewise.
* doc/tm.texi.in (TARGET_D_CRITSEC_SIZE): Likewise.
* doc/tm.texi: Regenerate.

gcc/d/ChangeLog:

* dmd/MERGE: Merge upstream dmd bec5973b0.
* d-target.cc (Target::critsecsize): Remove.
* d-target.def: Remove d_critsec_size.

libphobos/ChangeLog:

* libdruntime/MERGE: Merge upstream druntime 58560d51.

d: Merge upstream dmd 0fcdaab32

Fixes a bug where there was undefined template references when compiling
upstream dmd mainline.

In `TemplateInstance::semantic`, there exists special handling of
matching template instances for the same template declaration to ensure
that only at most one instance gets codegen'd.

If the primary instance `inst` originated from a non-root module, the
`minst` field will be updated so it is now coming from a root module,
however all Dsymbol `inst->members` of the instance still have their
`_scope->minst` pointing at the original non-root module. We must now
propagate `minst` to all members so that forward referenced dependencies
that get instantiated will also be appended to the root module,
otherwise there will be undefined references at link-time.

This doesn't affect compilations where all modules are compiled
together, as every module is a root module in that situation.  What this
primarily affects are cases where there is a mix of root and non-root
modules, and a template was first instantiated in a non-root context,
then later instantiated again in a root context.

Reviewed-on: https://github.com/dlang/dmd/pull/11867

gcc/d/ChangeLog:

* dmd/MERGE: Merge upstream dmd 0fcdaab32

gcov-profile: use static pool for TOPN first

gcc/ChangeLog:

PR gcov-profile/97461
* gcov-io.h (GCOV_PREALLOCATED_KVP): Pre-allocate 64
static counters.

libgcc/ChangeLog:

PR gcov-profile/97461
* libgcov.h (gcov_counter_add): Use first static counters
as it should help to have malloc wrappers set up.

gcc/testsuite/ChangeLog:

PR gcov-profile/97461
* gcc.dg/tree-prof/pr97461.c: New test.

Re-enable fnspec checking once fortran frontend is fixed.

* tree-ssa-alias.c (attr_fnspec::verify): Re-enabl checking.

[Ada] vx7r2cert: port ARM, PowerPC and x86 targets

gcc/ada/

* Makefile.rtl: Add vx7r2cert spec file to ARM, PowerPC and x86
targets.
* vxworks7-cert-rtp-link.spec: New spec file.

[Ada] Warnings on g-spogwa.adb

gcc/ada/

* Makefile.rtl (GNATRTL_NONTASKING_OBJS): Add g-spogwa object.
* libgnat/g-spogwa.adb: Fix style errors.

[Ada] Fix GNATprove support for iterated_component_associations

gcc/ada/

* exp_spark.adb (Expand_SPARK_Array_Aggregate): Dedicated
routine for array aggregates; mostly reuses existing code, but
calls itself recursively for multi-dimensional array aggregates.
(Expand_SPARK_N_Aggregate): Call Expand_SPARK_Array_Aggregate to
do the actual expansion, starting from the first index of the
array type.

[Ada] Multidimensional arrays with Iterated_Component_Associations

gcc/ada/

* sem_aggr.adb (Resolve_Iterated_Component_Association): new
internal subprogram Remove_References, to reset semantic
information on each reference to the index variable of the
association, so that Collect_Aggregate_Bounds can work properly
on multidimensional arrays with nested associations, and
subsequent expansion into loops can verify that dimensions of
each subaggregate are compatible.

[Ada] CUDA: handle parameter associations in expansion of CUDA_Execute

gcc/ada/

* exp_prag.adb (Append_Copies): Handle N_Parameter_Associations.

[Ada] Remove Digits_From_Size and Width_From_Size

gcc/ada/

* ada_get_targ.adb (Digits_From_Size): Delete.
(Width_From_Size): Likewise.
* get_targ.adb (Digits_From_Size): Likewise.
(Width_From_Size): Likewise.
* get_targ.ads (Digits_From_Size): Likewise.
(Width_From_Size): Likewise.
* ttypes.ads: Remove with clause for Get_Targ.
(Standard_Short_Short_Integer_Width): Delete.
(Standard_Short_Integer_Width): Likewise.
(Standard_Integer_Width): Likewise.
(Standard_Long_Integer_Width): Likewise.
(Standard_Long_Long_Integer_Width): Likewise.
(Standard_Long_Long_Long_Integer_Width): Likewise.
(Standard_Short_Float_Digits): Likewise.
(Standard_Float_Digits): Likewise.
(Standard_Long_Float_Digits): Likewise.
(Standard_Long_Long_Float_Digits): Likewise.
* gnat1drv.adb (Adjust_Global_Switches): Adjust.

[Ada] New flag Transform_Function_Array

gcc/ada/

* exp_ch6.adb, freeze.adb, gnat1drv.adb, opt.ads, sem_ch6.adb
(Transform_Function_Array): New flag, split from Modify_Tree_For_C.
* exp_unst.adb: Minor reformatting.

[Ada] Improve EINTR error handling on "poll" system call

gcc/ada/

* libgnat/g-socpol.adb (Wait): Do not exit from loop on EINTR
error and timeout is over.

Fix fnspec of math builtins

* builtin-attrs.def (STRERRNOC): New macro.
(STRERRNOP): New macro.
(ATTR_ERRNOCONST_NOTHROW_LEAF_LIST): New attr list.
(ATTR_ERRNOPURE_NOTHROW_LEAF_LIST): New attr list.
* builtins.def (ATTR_MATHFN_ERRNO): Use
ATTR_ERRNOCONST_NOTHROW_LEAF_LIST.
(ATTR_MATHFN_FPROUNDING_ERRNO): Use ATTR_ERRNOCONST_NOTHROW_LEAF_LIST
or ATTR_ERRNOPURE_NOTHROW_LEAF_LIST.

RISC-V: Refine riscv_parse_arch_string

 - Generalize logic for translating arch to internal flags, this patch
   is infrastructure for supporing sub-extension parsing.

gcc/ChangeLog

* common/config/riscv/riscv-common.c (opt_var_ref_t): New.
(riscv_ext_flag_table_t): New.
(riscv_ext_flag_table): New.
(riscv_parse_arch_string): Pass gcc_options* instead of
&opts->x_target_flags only, and using riscv_arch_option_table to
setup flags.
(riscv_handle_option): Update argument for riscv_parse_arch_string.
(riscv_expand_arch): Ditto.
(riscv_expand_arch_from_cpu): Ditto.

Use EAF_RETURN_ARG in tree-ssa-ccp.c

* tree-ssa-ccp.c (evaluate_stmt): Use EAF_RETURN_ARG; do not handle
string buitings specially.

Fix builtin decls generated in tree.c

* tree.c (set_call_expr_flags): Fix string for ECF_RET1.
(build_common_builtin_nodes): Do not set ECF_RET1 for memcpy, memmove,
and memset. They are handled by builtin_fnspec.

Add string builtins to builtin_fnspec

* builtins.c (builtin_fnspec): Add bzero, memcmp, memcmp_eq, bcmp,
strncmp, strncmp_eq, strncasecmp, rindex, strlen, strlnen, strcasecmp,
strcspn, strspn, strcmp, strcmp_eq.

Move SLP nodes to an alloc-pool

This introduces a global alloc-pool for SLP nodes to reduce overhead
on SLP allocation churn which will get worse and to eventually release
SLP cycles which will retain a refcount of one and thus are never
freed at the moment.

2020-10-26  Richard Biener  <rguenther@suse.de>

* tree-vectorizer.h (slp_tree_pool): Declare.
(_slp_tree::operator new): Likewise.
(_slp_tree::operator delete): Likewise.
* tree-vectorizer.c (vectorize_loops): Allocate and free the
slp_tree_pool.
(pass_slp_vectorize::execute): Likewise.
* tree-vect-slp.c (slp_tree_pool): Define.
(_slp_tree::operator new): Likewise.
(_slp_tree::operator delete): Likewise.

lto: no sub-make when --jobserver-auth= is missing

We newly correctly detect that a job server is not active for
a LTO linking:

lto-wrapper: warning: jobserver is not available: '--jobserver-auth=' is not present in 'MAKEFLAGS'

In that situation we should not call make -f abc.mk as it can leed
to N^2 LTRANS units.

gcc/ChangeLog:

* lto-wrapper.c (run_gcc): Do not use sub-make when jobserver is
not detected properly.

call_summary: move hooks to base.

gcc/ChangeLog:

* symbol-summary.h (call_summary_base): Pass symtab hooks to
base and register (or unregister) hooks directly.

call_summary: modify cgraph hooks.

gcc/ChangeLog:

* symbol-summary.h (function_summary_base::unregister_hooks):
Call disable_insertion_hook and disable_duplication_hook.
(function_summary_base::symtab_insertion): New field.
(function_summary_base::symtab_removal): Likewise.
(function_summary_base::symtab_duplication): Likewise.
Register hooks in function_summary_base and directly register
(or unregister) hooks.

Add test for pr97560.

gcc/testsuite/ChangeLog:

PR tree-optimization/97560
* g++.dg/pr97560.C: New test.

[RS6000] Separate dg-require-effective-target options

* gcc.target/powerpc/vsx_mask-count-runnable.c: Separate options
passed to dg-require-effective-target.
* gcc.target/powerpc/vsx_mask-expand-runnable.c: Likewise.
* gcc.target/powerpc/vsx_mask-extract-runnable.c: Likewise.
* gcc.target/powerpc/vsx_mask-move-runnable.c: Likewise.

Daily bump.

[RS6000] Remove -mpcrel from tests

When running with -m32
FAIL: gcc.target/powerpc/pr94740.c (test for excess errors)
Excess errors:
cc1: error: '-mpcrel' requires '-mcmodel=medium'

The others don't run for -m32, but remove the unnecessary -mpcrel
anyway.

* gcc.target/powerpc/localentry-1.c: Remove -mpcrel from options.
* gcc.target/powerpc/notoc-direct-1.c: Likewise.
* gcc.target/powerpc/pr94740.c: Likewise.

[RS6000] biarch test fail

* gcc.target/powerpc/bswap64-4.c: Comment.

[RS6000] Replace -mcpu with -mdejagnu-cpu

* gcc.target/powerpc/pr93122.c: Replace -mcpu with -mdejagnu-cpu.
* gcc.target/powerpc/vsx_mask-count-runnable.c: Likewise.
* gcc.target/powerpc/vsx_mask-expand-runnable.c: Likewise.
* gcc.target/powerpc/vsx_mask-extract-runnable.c: Likewise.
* gcc.target/powerpc/vsx_mask-move-runnable.c: Likewise.

[RS6000] Tests that use int128_t and -m32

All these tests fail with -m32 due to lack of int128 support, in some
cases with what I thought was not the best error message.  For example
vsx_mask-move-runnable.c:34:3: error: unknown type name 'vector'
is misleading.  The problem isn't "vector" but "vector __uint128_t".

* gcc.target/powerpc/vsx-load-element-extend-char.c: Require int128.
* gcc.target/powerpc/vsx-load-element-extend-int.c: Likewise.
* gcc.target/powerpc/vsx-load-element-extend-longlong.c: Likewise.
* gcc.target/powerpc/vsx-load-element-extend-short.c: Likewise.
* gcc.target/powerpc/vsx-store-element-truncate-char.c: Likewise.
* gcc.target/powerpc/vsx-store-element-truncate-int.c: Likewise.
* gcc.target/powerpc/vsx-store-element-truncate-longlong.c: Likewise.
* gcc.target/powerpc/vsx-store-element-truncate-short.c: Likewise.
* gcc.target/powerpc/vsx_mask-count-runnable.c: Likewise.
* gcc.target/powerpc/vsx_mask-expand-runnable.c: Likewise.
* gcc.target/powerpc/vsx_mask-extract-runnable.c: Likewise.
* gcc.target/powerpc/vsx_mask-move-runnable.c: Likewise.

[RS6000] Link power10 testcases

Running the assembler and linker catches more errors.

* gcc.target/powerpc/cfuged-1.c,
gcc.target/powerpc/cntlzdm-1.c,
gcc.target/powerpc/cnttzdm-1.c,
gcc.target/powerpc/dg-future-1.c,
gcc.target/powerpc/lsbb-runnable.c,
gcc.target/powerpc/mma-double-test.c,
gcc.target/powerpc/mma-single-test.c,
gcc.target/powerpc/p10-arch31.c,
gcc.target/powerpc/p10-identify.c,
gcc.target/powerpc/pdep-1.c,
gcc.target/powerpc/pextd-1.c,
gcc.target/powerpc/pr96787-2.c,
gcc.target/powerpc/vec-blend-runnable.c,
gcc.target/powerpc/vec-cfuged-1.c,
gcc.target/powerpc/vec-clrl-1.c,
gcc.target/powerpc/vec-clrl-3.c,
gcc.target/powerpc/vec-clrr-1.c,
gcc.target/powerpc/vec-clrr-3.c,
gcc.target/powerpc/vec-cntlzm-1.c,
gcc.target/powerpc/vec-cnttzm-1.c,
gcc.target/powerpc/vec-extracth-1.c,
gcc.target/powerpc/vec-extracth-3.c,
gcc.target/powerpc/vec-extracth-5.c,
gcc.target/powerpc/vec-extracth-7.c,
gcc.target/powerpc/vec-extractl-1.c,
gcc.target/powerpc/vec-extractl-3.c,
gcc.target/powerpc/vec-extractl-5.c,
gcc.target/powerpc/vec-extractl-7.c,
gcc.target/powerpc/vec-gnb-1.c,
gcc.target/powerpc/vec-insert-word-runnable.c,
gcc.target/powerpc/vec-pdep-1.c,
gcc.target/powerpc/vec-permute-ext-runnable.c,
gcc.target/powerpc/vec-pext-1.c,
gcc.target/powerpc/vec-replace-word-runnable.c,
gcc.target/powerpc/vec-shift-double-runnable.c,
gcc.target/powerpc/vec-splati-runnable.c,
gcc.target/powerpc/vec-stril-1.c,
gcc.target/powerpc/vec-stril-16.c,
gcc.target/powerpc/vec-stril-17.c,
gcc.target/powerpc/vec-stril-18.c,
gcc.target/powerpc/vec-stril-19.c,
gcc.target/powerpc/vec-stril-20.c,
gcc.target/powerpc/vec-stril-21.c,
gcc.target/powerpc/vec-stril-22.c,
gcc.target/powerpc/vec-stril-23.c,
gcc.target/powerpc/vec-stril-3.c,
gcc.target/powerpc/vec-stril-5.c,
gcc.target/powerpc/vec-stril-7.c,
gcc.target/powerpc/vec-stril_p-1.c,
gcc.target/powerpc/vec-stril_p-3.c,
gcc.target/powerpc/vec-stril_p-5.c,
gcc.target/powerpc/vec-stril_p-7.c,
gcc.target/powerpc/vec-strir-1.c,
gcc.target/powerpc/vec-strir-16.c,
gcc.target/powerpc/vec-strir-17.c,
gcc.target/powerpc/vec-strir-18.c,
gcc.target/powerpc/vec-strir-19.c,
gcc.target/powerpc/vec-strir-20.c,
gcc.target/powerpc/vec-strir-21.c,
gcc.target/powerpc/vec-strir-22.c,
gcc.target/powerpc/vec-strir-23.c,
gcc.target/powerpc/vec-strir-3.c,
gcc.target/powerpc/vec-strir-5.c,
gcc.target/powerpc/vec-strir-7.c,
gcc.target/powerpc/vec-strir_p-1.c,
gcc.target/powerpc/vec-strir_p-3.c,
gcc.target/powerpc/vec-strir_p-5.c,
gcc.target/powerpc/vec-strir_p-7.c,
gcc.target/powerpc/vec-ternarylogic-1.c,
gcc.target/powerpc/vec-ternarylogic-3.c,
gcc.target/powerpc/vec-ternarylogic-5.c,
gcc.target/powerpc/vec-ternarylogic-7.c,
gcc.target/powerpc/vec-ternarylogic-9.c,
gcc.target/powerpc/vsx_mask-count-runnable.c,
gcc.target/powerpc/vsx_mask-expand-runnable.c,
gcc.target/powerpc/vsx_mask-extract-runnable.c,
gcc.target/powerpc/vsx_mask-move-runnable.c,
gcc.target/powerpc/xxgenpc-runnable.c: Link testcase when it
can't be run.

[RS6000] dimode_off.c test

This tests behaviour near the limit of 16-bit signed offsets.  If
power10 prefix instructions are enabled, no such testing occurs.

* gcc.target/powerpc/dimode_off.c: Add -mno-prefixed to options.

[RS6000] Non-pcrel tests when power10

These tests require -mno-pcrel because they are testing features
of the non-pcrel ABI.

* gcc.target/powerpc/cprophard.c: Add -mno-pcrel to options.
* gcc.target/powerpc/float128-hw3.c: Likewise.
* gcc.target/powerpc/pr79439-1.c: Likewise.
* gcc.target/powerpc/pr79439-2.c: Likewise.
* gcc.target/powerpc/r2_shrink-wrap.c: Likewise.

syscall: import additional BSD-specific syscall wrappers

Import additional code from upstream for handing system
calls on BSD systems. This makes the syscall package on
NetBSD complete enough to compile the standard library.

Updates golang/go#38538.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/265123

Combine logical OR ranges properly.

When combining logical OR operands with a FALSE result, union the false
ranges for operand1 and operand2... not intersection.

gcc/
PR tree-optimization/97567
* gimple-range-gori.cc (gori_compute::logical_combine): Union the
ranges of operand1 and operand2, not intersect.
gcc/testsuite/
* gcc.dg/pr97567.c: New.

libstdc++: Fix error in experimental::strand

libstdc++-v3/ChangeLog:

* include/experimental/executor (strand::_State): Fix thinko.

Extend builtin fnspecs

* attr-fnspec.h: Update toplevel comment.
(attr_fnspec::attr_fnspec): New constructor.
(attr_fnspec::arg_read_p,
attr_fnspec::arg_written_p,
attr_fnspec::arg_access_size_given_by_arg_p,
attr_fnspec::arg_single_access_p
attr_fnspec::loads_known_p
attr_fnspec::stores_known_p,
attr_fnspec::clobbers_errno_p): New member functions.
(gimple_call_fnspec): Declare.
(builtin_fnspec): Declare.
* builtins.c: Include attr-fnspec.h
(builtin_fnspec): New function.
* builtins.def (BUILT_IN_MEMCPY): Do not specify RET1 fnspec.
(BUILT_IN_MEMMOVE): Do not specify RET1 fnspec.
(BUILT_IN_MEMSET): Do not specify RET1 fnspec.
(BUILT_IN_STRCAT): Do not specify RET1 fnspec.
(BUILT_IN_STRCPY): Do not specify RET1 fnspec.
(BUILT_IN_STRNCAT): Do not specify RET1 fnspec.
(BUILT_IN_STRNCPY): Do not specify RET1 fnspec.
(BUILT_IN_MEMCPY_CHK): Do not specify RET1 fnspec.
(BUILT_IN_MEMMOVE_CHK): Do not specify RET1 fnspec.
(BUILT_IN_MEMSET_CHK): Do not specify RET1 fnspec.
(BUILT_IN_STRCAT_CHK): Do not specify RET1 fnspec.
(BUILT_IN_STRCPY_CHK): Do not specify RET1 fnspec.
(BUILT_IN_STRNCAT_CHK): Do not specify RET1 fnspec.
(BUILT_IN_STRNCPY_CHK): Do not specify RET1 fnspec.
* gimple.c (gimple_call_fnspec): Return attr_fnspec.
(gimple_call_arg_flags): Update.
(gimple_call_return_flags): Update.
* tree-ssa-alias.c (check_fnspec): New function.
(ref_maybe_used_by_call_p_1): Use fnspec for builtin handling.
(call_may_clobber_ref_p_1): Likewise.
(attr_fnspec::verify): Update verifier.
* calls.c (decl_fnspec): New function.
(decl_return_flags): Use it.

Handle signed 1-bit ranges in irange::invert.

The problem here is we are trying to add 1 to a -1 in a signed 1-bit
field and coming up with UNDEFINED because of the overflow.

Signed 1-bits are annoying because you can't really add or subtract
one, because the one is unrepresentable.  For invert() we have a
special subtract_one() function that handles 1-bit signed fields.

This patch implements the analogous add_one() function so that invert
works.

gcc/ChangeLog:

PR tree-optimization/97555
* range-op.cc (range_tests): Test 1-bit signed invert.
* value-range.cc (subtract_one): Adjust comment.
(add_one): New.
(irange::invert): Call add_one.

gcc/testsuite/ChangeLog:

* gcc.dg/pr97555.c: New test.

Implement three-level optimize_for_size predicates

this patch implements thre two-state optimize_for_size predicates, so with -Os
and with profile feedback for never executed code it returns OPTIMIZE_SIZE_MAX
while in cases we decide to optimize for size based on branch prediction logic
it return OPTIMIZE_SIZE_BALLANCED.

The idea is that for places where we guess that code is unlikely we do not
want to do extreme optimizations for size that leads to many fold slowdowns
(using idiv rather than few shigts or using rep based inlined stringops).

I will update RTL handling code to also support this with BB granuality (which
we don't currently).  LLVM has -Os and -Oz levels where -Oz is our -Os and
LLVM's -Os would ocrrespond to OPTIMIZE_SIZE_BALLANCED.  I wonder if we want
to export this to command line somehow?  For me it would be definitly useful
to test things, I am not sure how "weaker" -Os is desired in practice.

gcc/ChangeLog:

* cgraph.h (cgraph_node::optimize_for_size_p): Return
optimize_size_level.
(cgraph_node::optimize_for_size_p): Update.
* coretypes.h (enum optimize_size_level): New enum.
* predict.c (unlikely_executed_edge_p): Microoptimize.
(optimize_function_for_size_p): Return optimize_size_level.
(optimize_bb_for_size_p): Likewise.
(optimize_edge_for_size_p): Likewise.
(optimize_insn_for_size_p): Likewise.
(optimize_loop_nest_for_size_p): Likewise.
* predict.h (optimize_function_for_size_p): Update declaration.
(optimize_bb_for_size_p): Update declaration.
(optimize_edge_for_size_p): Update declaration.
(optimize_insn_for_size_p): Update declaration.
(optimize_loop_for_size_p): Update declaration.
(optimize_loop_nest_for_size_p): Update declaration.

Refactor SLP instance analysis

This refactors the toplevel entry to analyze an SLP instance to
expose a worker analyzing from a vector of stmts and an SLP entry
kind.

2020-10-26  Richard Biener  <rguenther@suse.de>

* tree-vect-slp.c (enum slp_instance_kind): New.
(vect_build_slp_instance): Split out from...
(vect_analyze_slp_instance): ... this.

Re: error: ‘EVRP_MODE_DEBUG’ was not declared – was: [PUSHED] Ranger classes.

Initialize zerov to match vr-values.c.

* gimple-range.cc (range_of_builtin_call): Initialize zerov to 0.

c++: Implement __is_nothrow_constructible and __is_nothrow_assignable

gcc/c-family/ChangeLog:

* c-common.c (__is_nothrow_assignable): New.
(__is_nothrow_constructible): Likewise.
* c-common.h (RID_IS_NOTHROW_ASSIGNABLE): New.
(RID_IS_NOTHROW_CONSTRUCTIBLE): Likewise.

gcc/cp/ChangeLog:

* cp-tree.h (CPTK_IS_NOTHROW_ASSIGNABLE): New.
(CPTK_IS_NOTHROW_CONSTRUCTIBLE): Likewise.
(is_nothrow_xible): Likewise.
* method.c (is_nothrow_xible): New.
(is_trivially_xible): Tweak.
* parser.c (cp_parser_primary_expression): Handle the new RID_*.
(cp_parser_trait_expr): Likewise.
* semantics.c (trait_expr_value): Handle the new RID_*.
(finish_trait_expr): Likewise.

libstdc++-v3/ChangeLog:

* include/std/type_traits (__is_nt_constructible_impl): Remove.
(__is_nothrow_constructible_impl): Adjust.
(is_nothrow_default_constructible): Likewise.
(__is_nt_assignable_impl): Remove.
(__is_nothrow_assignable_impl): Adjust.

Fix simdclones

gcc/ChangeLog:

PR ipa/97576
* cgraphclones.c (cgraph_node::materialize_clone): Clear stmt
references.
* cgraphunit.c (mark_functions_to_output): Do not clear them here.
* ipa-inline-transform.c (inline_transform): Clear stmt references.
* symtab.c (symtab_node::clear_stmts_in_references): Make recursive
for clones.
* tree-ssa-structalias.c (ipa_pta_execute): Do not clear references.

gcc/testsuite/ChangeLog:

PR ipa/97576
* gcc.c-torture/compile/pr97576.c: New test.

AArch64: Add FLAG for store intrinsics [PR94442]

2020-10-26  Zhiheng Xie  <xiezhiheng@huawei.com>
    Nannan Zheng  <zhengnannan@huawei.com>

gcc/ChangeLog:

* config/aarch64/aarch64-builtins.c: Add FLAG STORE.
* config/aarch64/aarch64-simd-builtins.def: Add proper FLAG
for store intrinsics.

libstdc++: Fix declarations of memalign etc. for freestanding [PR 97570]

libstdc++-v3/ChangeLog:

PR libstdc++/97570
* libsupc++/new_opa.cc: Declare size_t in global namespace.
Remove unused header.

PR tree-optimization/97546 Bail out of find_bswap_or_nop on non-INTEGER_CST sizes

This patch fixes the ICE in the PR by bailing out of find_bswap_or_nop
on poly_int sizes.
I don't think it intends to handle them and from my reading of the code
it's the most appropriate place to reject them
here rather than in the callers.

Bootstrapped and tested on aarch64-none-linux-gnu.

gcc/
PR tree-optimization/97546
* gimple-ssa-store-merging.c (find_bswap_or_nop): Return NULL if
type is not INTEGER_CST.

gcc/testsuite/
PR tree-optimization/97546
* gcc.target/aarch64/sve/acle/general/pr97546.c: New test.

middle-end/97521 - always use single-bit bools in mask vector types

This makes us always use a single-bit boolean type component type
for integer mode mask VECTOR_BOOLEAN_TYPE_P to match the RTL and target
representation.  This aovids the need for magic translation and
the inconsistencies from the translation requirement now that
we expose temporaries of those types on the GIMPLE level.

2020-10-23  Richard Biener  <rguenther@suse.de>

PR middle-end/97521
* expr.c (const_scalar_mask_from_tree): Remove.
(expand_expr_real_1): Always VIEW_CONVERT integer mode
vector constants to an integer type.
* tree.c (build_truth_vector_type_for_mode): Use a single-bit
boolean component type for non-vector-mode mask_mode.

* gcc.target/i386/pr97521.c: New testcase.

x86: Inline strncmp only with -minline-all-stringops

Expand strncmp to "repz cmpsb" only with -minline-all-stringops since
"repz cmpsb" can be much slower than strncmp function implemented with
vector instructions, see

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=43052

gcc/

PR target/95458
* config/i386/i386-expand.c (ix86_expand_cmpstrn_or_cmpmem):
Return false for -mno-inline-all-stringops.

gcc/testsuite/

PR target/95458
* gcc.target/i386/pr95458-1.c: New test.
* gcc.target/i386/pr95458-2.c: Likewise.

x86: Add cmpmemsi for -minline-all-stringops

We used to expand memcmp to "repz cmpsb" via cmpstrnsi.  It was changed
by

commit 9b0f6f5e511ca512e4faeabc81d2fd3abad9b02f
Author: Nick Clifton <nickc@redhat.com>
Date:   Fri Aug 12 16:26:11 2011 +0000

    builtins.c (expand_builtin_memcmp): Do not use cmpstrnsi pattern.

            * builtins.c (expand_builtin_memcmp): Do not use cmpstrnsi
            pattern.
            * doc/md.texi (cmpstrn): Note that the comparison stops if both
            fetched bytes are zero.
            (cmpstr): Likewise.
            (cmpmem): Note that the comparison does not stop if both of the
            fetched bytes are zero.

Duplicate the cmpstrn pattern for cmpmem.  The only difference is that
the length argument of cmpmem is guaranteed to be less than or equal to
lengths of 2 memory areas.  Since "repz cmpsb" can be much slower than
memcmp function implemented with vector instruction, see

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=43052

expand cmpmem to "repz cmpsb" only for -minline-all-stringops.

gcc/

PR target/95151
* config/i386/i386-expand.c (ix86_expand_cmpstrn_or_cmpmem): New
function.
* config/i386/i386-protos.h (ix86_expand_cmpstrn_or_cmpmem): New
prototype.
* config/i386/i386.md (cmpmemsi): New pattern.

gcc/testsuite/

PR target/95151
* gcc.target/i386/pr95151-1.c: New test.
* gcc.target/i386/pr95151-2.c: Likewise.
* gcc.target/i386/pr95151-3.c: Likewise.
* gcc.target/i386/pr95151-4.c: Likewise.

IBM Z: Add vcond_mask expander

After adding vec_cmp expanders we have seen various performance
related regression in the testsuite.  These appear to be caused by a
missing vcond_mask definition in the backend.  Fixed with this patch.

The patch fixes the following testsuite fails:

FAIL: gcc.dg/vect/vect-21.c -flto -ffat-lto-objects  scan-tree-dump-times vect "vectorized 3 loops" 1
FAIL: gcc.dg/vect/vect-21.c scan-tree-dump-times vect "vectorized 3 loops" 1
FAIL: gcc.dg/vect/vect-23.c -flto -ffat-lto-objects  scan-tree-dump-times vect "vectorized 3 loops" 1
FAIL: gcc.dg/vect/vect-23.c scan-tree-dump-times vect "vectorized 3 loops" 1
FAIL: gcc.dg/vect/vect-24.c -flto -ffat-lto-objects  scan-tree-dump-times vect "vectorized 3 loops" 1
FAIL: gcc.dg/vect/vect-24.c scan-tree-dump-times vect "vectorized 3 loops" 1
FAIL: gcc.dg/vect/vect-live-6.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
FAIL: gcc.dg/vect/vect-live-6.c scan-tree-dump vect "vectorized 1 loops"
FAIL: gcc.target/s390/vector/vcond-shift.c scan-assembler-times vesrab\\t%v.?,%v.?,7 6
FAIL: gcc.target/s390/vector/vcond-shift.c scan-assembler-times vesraf\\t%v.?,%v.?,31 6
FAIL: gcc.target/s390/vector/vcond-shift.c scan-assembler-times vesrah\\t%v.?,%v.?,15 6
FAIL: gcc.target/s390/vector/vcond-shift.c scan-assembler-times vesrlb\\t%v.?,%v.?,7 4
FAIL: gcc.target/s390/vector/vcond-shift.c scan-assembler-times vesrlf\\t%v.?,%v.?,31 4
FAIL: gcc.target/s390/vector/vcond-shift.c scan-assembler-times vesrlh\\t%v.?,%v.?,15 4
FAIL: gcc.dg/vect/vect-21.c -flto -ffat-lto-objects  scan-tree-dump-times vect "vectorized 3 loops" 1
FAIL: gcc.dg/vect/vect-21.c scan-tree-dump-times vect "vectorized 3 loops" 1
FAIL: gcc.dg/vect/vect-23.c -flto -ffat-lto-objects  scan-tree-dump-times vect "vectorized 3 loops" 1
FAIL: gcc.dg/vect/vect-23.c scan-tree-dump-times vect "vectorized 3 loops" 1
FAIL: gcc.dg/vect/vect-24.c -flto -ffat-lto-objects  scan-tree-dump-times vect "vectorized 3 loops" 1
FAIL: gcc.dg/vect/vect-24.c scan-tree-dump-times vect "vectorized 3 loops" 1
FAIL: gcc.dg/vect/vect-live-6.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
FAIL: gcc.dg/vect/vect-live-6.c scan-tree-dump vect "vectorized 1 loops"
FAIL: gcc.target/s390/vector/vcond-shift.c scan-assembler-times vesrab\\t%v.?,%v.?,7 6
FAIL: gcc.target/s390/vector/vcond-shift.c scan-assembler-times vesraf\\t%v.?,%v.?,31 6
FAIL: gcc.target/s390/vector/vcond-shift.c scan-assembler-times vesrah\\t%v.?,%v.?,15 6
FAIL: gcc.target/s390/vector/vcond-shift.c scan-assembler-times vesrlb\\t%v.?,%v.?,7 4
FAIL: gcc.target/s390/vector/vcond-shift.c scan-assembler-times vesrlf\\t%v.?,%v.?,31 4
FAIL: gcc.target/s390/vector/vcond-shift.c scan-assembler-times vesrlh\\t%v.?,%v.?,15 4

gcc/ChangeLog:

* config/s390/vector.md ("vcond_mask_<mode><mode>"): New expander.