debug: Make sure to output .file 0 when generating DWARF5.

When gas outputs DWARF5 .debug_line[_str] then we have to tell it the
comp_dir and main file name for the zero entry line table. Otherwise
gas has to guess at the CU compilation directory and file.

Before a gcc -gdwarf-5 ../src/hello.c line table looked like:

Directory table:
 0     ../src (24)
 1     ../src (24)
 2     /usr/include (31)

File name table:
 0     hello.c (16),  0
 1     hello.c (16),  1
 2     stdio.h (44),  2

With this patch it looks like:

Directory table:
 0     /tmp/obj (0)
 1     ../src (24)
 2     /usr/include (31)

File name table:
 0     ../src/hello.c (9),  0
 1     hello.c (16),  1
 2     stdio.h (44),  2

gcc/ChangeLog:

* dwarf2out.c (dwarf2out_finish): Emit .file 0 entry when
generating DWARF5 .debug_line table through gas.

Improve documentation of -fallow-store-data-races

2020-10-08  John Henning  <john.henning@oracle.com>

gcc/

PR other/97309
* doc/invoke.texi: Improve documentation of
-fallow-store-data-races.

libstdc++: Add assertions for preconditions in sampling distributions [PR 82584]

These three distributions all require 0 < S where S is the sum of the
weights. When the sum is zero there's an undefined FP division by zero.
Add assertions to help users diagnose the problem.

libstdc++-v3/ChangeLog:

PR libstdc++/82584
* include/bits/random.tcc
(discrete_distribution::param_type::_M_initialize)
(piecewise_constant_distribution::param_type::_M_initialize)
(piecewise_linear_distribution::param_type::_M_initialize):
Add assertions for positive sums..
* testsuite/26_numerics/random/pr60037-neg.cc: Adjust dg-error
line.

arm: [MVE] Add missing __arm_vcvtnq_u32_f32 intrinsic (PR 96914)

__arm_vcvtnq_u32_f32 was missing from arm_mve.h, although the s32_f32 and
[su]16_f16 versions were present.

This patch adds the missing version and testcase, which are
cut-and-paste from the other versions.

2020-10-08  Christophe Lyon  <christophe.lyon@linaro.org>

gcc/
PR target/96914
* config/arm/arm_mve.h (__arm_vcvtnq_u32_f32): New.

gcc/testsuite/
PR target/96914
* gcc.target/arm/mve/intrinsics/vcvtnq_u32_f32.c: New test.

SLP vectorize multiple BBs at once

This work from Martin Liska was motivated by gcc.dg/vect/bb-slp-22.c
which shows how poorly we currently BB vectorize code like

  a0 = in[0] + 23;
  a1 = in[1] + 142;
  a2 = in[2] + 2;
  a3 = in[3] + 31;

  if (x > y)
    {
      b[0] = a0;
      b[1] = a1;
      b[2] = a2;
      b[3] = a3;
    }
  else
    {
      out[0] = a0 * (x + 1);
      out[1] = a1 * (y + 1);
      out[2] = a2 * (x + 1);
      out[3] = a3 * (y + 1);
    }

namely by vectorizing the stores but not the common load (and add)
they are feeded with.

Thus with the following patch we change the BB vectorizer from
operating on a single basic-block at a time to consider somewhat
larger regions (but not the whole function yet because of issues
with vector size iteration).

I took the opportunity to remove the fancy region iterations again
now that we operate on BB granularity and in the end need to visit
PHI nodes as well.

2020-10-08  Martin Liska  <mliska@suse.cz>
    Richard Biener  <rguenther@suse.de>

* tree-vectorizer.h (_bb_vec_info::const_iterator): Remove.
(_bb_vec_info::const_reverse_iterator): Likewise.
(_bb_vec_info::region_stmts): Likewise.
(_bb_vec_info::reverse_region_stmts): Likewise.
(_bb_vec_info::_bb_vec_info): Adjust.
(_bb_vec_info::bb): Remove.
(_bb_vec_info::region_begin): Remove.
(_bb_vec_info::region_end): Remove.
(_bb_vec_info::bbs): New vector of BBs.
(vect_slp_function): Declare.
* tree-vect-patterns.c (vect_determine_precisions): Use
regular stmt iteration.
(vect_pattern_recog): Likewise.
* tree-vect-slp.c: Include cfganal.h, tree-eh.h and tree-cfg.h.
(vect_build_slp_tree_1): Properly refuse to vectorize
volatile and throwing stmts.
(vect_build_slp_tree_2): Pass group-size down to
get_vectype_for_scalar_type.
(_bb_vec_info::_bb_vec_info): Use regular stmt iteration,
adjust for changed region specification.
(_bb_vec_info::~_bb_vec_info): Likewise.
(vect_slp_check_for_constructors): Likewise.
(vect_slp_region): Likewise.
(vect_slp_bbs): New worker operating on a vector of BBs.
(vect_slp_bb): Wrap it.
(vect_slp_function): New function splitting the function
into multi-BB regions.
(vect_create_constant_vectors): Handle the case of inserting
after a throwing def.
(vect_schedule_slp_instance): Adjust.
* tree-vectorizer.c (vec_info::remove_stmt): Simplify again.
(vec_info::insert_seq_on_entry): Adjust.
(pass_slp_vectorize::execute): Also init PHIs.  Call
vect_slp_function.

* gcc.dg/vect/bb-slp-22.c: Adjust.
* gfortran.dg/pr68627.f: Likewise.

libstdc++: Add C++11 member functions for ios::failure in old ABI

The new constructors that C++11 added to std::ios_base::failure were
missing for the old ABI. This adds them, but just ignores the
std::error_code argument (because there's nowhere to store it).

This also adds a code() member, which should be provided by the
std::system_error base class, but that base class isn't present in the
old ABI.

This allows the old ios::failure to be used in code that expects the new
API, although with reduced functionality.

libstdc++-v3/ChangeLog:

* include/bits/ios_base.h (ios_base::failure): Add constructors
takeing error_code argument. Add code() member function.
* testsuite/27_io/ios_base/failure/cxx11.cc: Allow test to
run for the old ABI but do not check for derivation from
std::system_error.
* testsuite/27_io/ios_base/failure/error_code.cc: New test.

libstdc++: Avoid divide by zero in default template arguments

My previous attempt to fix this only worked when m is a power of two.
There is still a bug when a=00 and !has_single_bit(m).

Instead of trying to make _Mod work for a==0 this change ensures that we
never instantiate it with a==0. For C++17 we can use if-constexpr, but
otherwise we need to use a different multipler. It doesn't matter what
we use, as it won't actually be called, only instantiated.

libstdc++-v3/ChangeLog:

* include/bits/random.h (__detail::_Mod): Revert last change.
(__detail::__mod): Do not use _Mod for a==0 case.
* testsuite/26_numerics/random/linear_congruential_engine/operators/call.cc:
Check other cases with a==0. Also check runtime results.
* testsuite/26_numerics/random/pr60037-neg.cc: Adjust dg-error
line.

tree-optimization/97330 - fix bad load sinking

This fixes bad placement of sunk loads.

2020-10-08  Richard Biener  <rguenther@suse.de>

PR tree-optimization/97330
* tree-ssa-sink.c (statement_sink_location): Avoid skipping
PHIs when they dominate the insert location.

* gcc.dg/torture/pr97330-1.c: New testcase.
* gcc.dg/torture/pr97330-2.c: Likewise.

Fix handling of parm_offset in ipa-modref on 32bit targets.

* ipa-modref.c (get_access): Fix handling of offsets.
* tree-ssa-alias.c (modref_may_conflict): Watch for overflows.

IPA MOD REF: add debug counter.

gcc/ChangeLog:

* dbgcnt.def (DEBUG_COUNTER): Add ipa_mod_ref debug counter.
* tree-ssa-alias.c (modref_may_conflict): Handle the counter.

adjust BB vectorization dump scanning

This adjusts BB vectorization testcases to look for the number of
SLP subgraphs vectorized rather than for the number of basic blocks
we've found opportunities in because followup patches will play
with the granularity we work on, vectorizing multiple basic blocks
at a time.

Together with this, because I noticed when looking at non-obvious
mismatches, I avoid analyzing group-size 1 SLP instances which
result in pointless V1mode vectorizations.

It might be interesting to work on adding sth like
dg-warning to look for -fopt-info-{optimized,missing} so
we could directly annotate (not) vectorized loops instead of
relying on fragile counts.

2020-10-08  Richard Biener  <rguenther@suse.de>

* tree-vectorizer.c (try_vectorize_loop_1): Do not dump
"basic block vectorized".
(pass_slp_vectorize::execute): Likewise.
* tree-vect-slp.c (vect_analyze_slp_instance): Avoid
re-analyzing split single stmts.

* g++.dg/vect/slp-pr50819.cc: Adjust.
* gcc.dg/vect/bb-slp-1.c: Adjust.
* gcc.dg/vect/bb-slp-10.c: Adjust.
* gcc.dg/vect/bb-slp-11.c: Adjust.
* gcc.dg/vect/bb-slp-13.c: Adjust.
* gcc.dg/vect/bb-slp-14.c: Adjust.
* gcc.dg/vect/bb-slp-15.c: Adjust.
* gcc.dg/vect/bb-slp-16.c: Adjust.
* gcc.dg/vect/bb-slp-17.c: Adjust.
* gcc.dg/vect/bb-slp-18.c: Adjust.
* gcc.dg/vect/bb-slp-19.c: Adjust.
* gcc.dg/vect/bb-slp-2.c: Adjust.
* gcc.dg/vect/bb-slp-20.c: Adjust.
* gcc.dg/vect/bb-slp-21.c: Adjust.
* gcc.dg/vect/bb-slp-22.c: Adjust.
* gcc.dg/vect/bb-slp-23.c: Adjust.
* gcc.dg/vect/bb-slp-24.c: Adjust.
* gcc.dg/vect/bb-slp-25.c: Adjust.
* gcc.dg/vect/bb-slp-26.c: Adjust.
* gcc.dg/vect/bb-slp-27.c: Adjust.
* gcc.dg/vect/bb-slp-28.c: Adjust.
* gcc.dg/vect/bb-slp-29.c: Adjust.
* gcc.dg/vect/bb-slp-3.c: Adjust.
* gcc.dg/vect/bb-slp-30.c: Adjust.
* gcc.dg/vect/bb-slp-31.c: Adjust.
* gcc.dg/vect/bb-slp-34.c: Adjust.
* gcc.dg/vect/bb-slp-35.c: Adjust.
* gcc.dg/vect/bb-slp-36.c: Adjust.
* gcc.dg/vect/bb-slp-38.c: Adjust.
* gcc.dg/vect/bb-slp-4.c: Adjust.
* gcc.dg/vect/bb-slp-45.c: Adjust.
* gcc.dg/vect/bb-slp-46.c: Adjust.
* gcc.dg/vect/bb-slp-48.c: Adjust.
* gcc.dg/vect/bb-slp-5.c: Adjust.
* gcc.dg/vect/bb-slp-6.c: Adjust.
* gcc.dg/vect/bb-slp-7.c: Adjust.
* gcc.dg/vect/bb-slp-8.c: Adjust.
* gcc.dg/vect/bb-slp-8a.c: Adjust.
* gcc.dg/vect/bb-slp-8b.c: Adjust.
* gcc.dg/vect/bb-slp-9.c: Adjust.
* gcc.dg/vect/bb-slp-div-2.c: Adjust.
* gcc.dg/vect/bb-slp-over-widen-1.c: Adjust.
* gcc.dg/vect/bb-slp-over-widen-2.c: Adjust.
* gcc.dg/vect/bb-slp-pattern-2.c: Adjust.
* gcc.dg/vect/bb-slp-pow-1.c: Adjust.
* gcc.dg/vect/bb-slp-pr58135.c: Adjust.
* gcc.dg/vect/bb-slp-pr65935.c: Adjust.
* gcc.dg/vect/bb-slp-pr78205.c: Adjust.
* gcc.dg/vect/bb-slp-pr81635-1.c: Adjust.
* gcc.dg/vect/bb-slp-pr81635-3.c: Adjust.
* gcc.dg/vect/bb-slp-pr95839-2.c: Adjust.
* gcc.dg/vect/bb-slp-pr95839.c: Adjust.
* gcc.dg/vect/bb-slp-pr95866.c: Adjust.
* gcc.dg/vect/bb-slp-subgroups-1.c: Adjust.
* gcc.dg/vect/bb-slp-subgroups-2.c: Adjust.
* gcc.dg/vect/bb-slp-subgroups-3.c: Adjust.
* gcc.dg/vect/fast-math-bb-slp-call-1.c: Adjust.
* gcc.dg/vect/no-tree-reassoc-bb-slp-12.c: Adjust.
* gcc.dg/vect/no-tree-sra-bb-slp-pr50730.c: Adjust.
* gfortran.dg/vect/pr62283-2.f: Adjust.
* gcc.target/i386/pr68961.c: Adjust.
* gcc.target/i386/pr84101.c: Adjust.
* gcc.dg/vect/bb-slp-pr81635-2.c: Adjust.
* gcc.dg/vect/bb-slp-pr81635-4.c: Adjust.
* gcc.dg/vect/fast-math-bb-slp-call-2.c: Adjust.
* gcc.dg/vect/costmodel/ppc/costmodel-bb-slp-9a.c: Adjust.
* gcc.dg/vect/costmodel/x86_64/costmodel-vect-slp.c: Adjust.
* gcc.dg/vect/bb-slp-div-1.c: Adjust.
* gcc.dg/vect/bb-slp-pr90006.c: Adjust.
* g++.dg/vect/slp-pr50413.cc: Adjust.

arm: [MVE] Remove illegal intrinsics (PR target/96914)

A few MVE intrinsics had an unsigned variant implement while they are
supported by the hardware.  This patch removes them:
__arm_vqrdmlashq_n_u8
__arm_vqrdmlahq_n_u8
__arm_vqdmlahq_n_u8
__arm_vqrdmlashq_n_u16
__arm_vqrdmlahq_n_u16
__arm_vqdmlahq_n_u16
__arm_vqrdmlashq_n_u32
__arm_vqrdmlahq_n_u32
__arm_vqdmlahq_n_u32
__arm_vmlaldavaxq_p_u32
__arm_vmlaldavaxq_p_u16

2020-10-08  Christophe Lyon  <christophe.lyon@linaro.org>

gcc/
PR target/96914
* config/arm/arm_mve.h (vqrdmlashq_n_u8, vqrdmlashq_n_u16)
(vqrdmlashq_n_u32, vqrdmlahq_n_u8, vqrdmlahq_n_u16)
(vqrdmlahq_n_u32, vqdmlahq_n_u8, vqdmlahq_n_u16, vqdmlahq_n_u32)
(vmlaldavaxq_p_u16, vmlaldavaxq_p_u32): Remove.
* config/arm/arm_mve_builtins.def (vqrdmlashq_n_u, vqrdmlahq_n_u)
(vqdmlahq_n_u, vmlaldavaxq_p_u): Remove.
* config/arm/unspecs.md (VQDMLAHQ_N_U, VQRDMLAHQ_N_U)
(VQRDMLASHQ_N_U)
(VMLALDAVAXQ_P_U): Remove unspecs.
* config/arm/iterators.md (VQDMLAHQ_N_U, VQRDMLAHQ_N_U)
(VQRDMLASHQ_N_U, VMLALDAVAXQ_P_U): Remove attributes.
(VQDMLAHQ_N, VQRDMLAHQ_N, VQRDMLASHQ_N, VMLALDAVAXQ_P): Remove
unsigned variants from iterators.
* config/arm/mve.md (mve_vqdmlahq_n_<supf><mode>)
(mve_vqrdmlahq_n_<supf><mode>)
(mve_vqrdmlashq_n_<supf><mode>, mve_vmlaldavaxq_p_<supf><mode>):
Update comment.

gcc/testsuite/
PR target/96914
* gcc.target/arm/mve/intrinsics/vmlaldavaxq_p_u16.c: Remove.
* gcc.target/arm/mve/intrinsics/vmlaldavaxq_p_u32.c: Remove.
* gcc.target/arm/mve/intrinsics/vqdmlahq_n_u16.c: Remove.
* gcc.target/arm/mve/intrinsics/vqdmlahq_n_u32.c: Remove.
* gcc.target/arm/mve/intrinsics/vqdmlahq_n_u8.c: Remove.
* gcc.target/arm/mve/intrinsics/vqrdmlahq_n_u16.c: Remove.
* gcc.target/arm/mve/intrinsics/vqrdmlahq_n_u32.c: Remove.
* gcc.target/arm/mve/intrinsics/vqrdmlahq_n_u8.c: Remove.
* gcc.target/arm/mve/intrinsics/vqrdmlashq_n_u16.c: Remove.
* gcc.target/arm/mve/intrinsics/vqrdmlashq_n_u32.c: Remove.
* gcc.target/arm/mve/intrinsics/vqrdmlashq_n_u8.c: Remove.

arm: [MVE[ Add vqdmlashq intrinsics (PR target/96914)

This patch adds:
vqdmlashq_m_n_s16
vqdmlashq_m_n_s32
vqdmlashq_m_n_s8
vqdmlashq_n_s16
vqdmlashq_n_s32
vqdmlashq_n_s8

2020-10-08  Christophe Lyon  <christophe.lyon@linaro.org>

gcc/
PR target/96914
* config/arm/arm_mve.h (vqdmlashq, vqdmlashq_m): Define.
* config/arm/arm_mve_builtins.def (vqdmlashq_n_s)
(vqdmlashq_m_n_s,): New.
* config/arm/unspecs.md (VQDMLASHQ_N_S, VQDMLASHQ_M_N_S): New
unspecs.
* config/arm/iterators.md (VQDMLASHQ_N_S, VQDMLASHQ_M_N_S): New
attributes.
(VQDMLASHQ_N): New iterator.
* config/arm/mve.md (mve_vqdmlashq_n_, mve_vqdmlashq_m_n_s): New
patterns.

gcc/testsuite/
PR target/96914
* gcc.target/arm/mve/intrinsics/vqdmlashq_m_n_s16.c: New test.
* gcc.target/arm/mve/intrinsics/vqdmlashq_m_n_s32.c: New test.
* gcc.target/arm/mve/intrinsics/vqdmlashq_m_n_s8.c: New test.
* gcc.target/arm/mve/intrinsics/vqdmlashq_n_s16.c: New test.
* gcc.target/arm/mve/intrinsics/vqdmlashq_n_s32.c: New test.
* gcc.target/arm/mve/intrinsics/vqdmlashq_n_s8.c: New test.

arm: Fix ICE on glibc compilation after my DIVMOD optimization [PR97322]

The arm target hook for divmod wasn't prepared to handle constants passed to
the function.

2020-10-08  Jakub Jelinek  <jakub@redhat.com>

PR target/97322
* config/arm/arm.c (arm_expand_divmod_libfunc): Pass mode instead of
GET_MODE (op0) or GET_MODE (op1) to emit_library_call_value.

* gcc.dg/pr97322.c: New test.

Fix PR97325.

gcc/ChangeLog:

PR tree-optimization/97325
* gimple-range.cc (gimple_ranger::range_of_builtin_call): Handle
negative numbers in __builtin_ffs and __builtin_popcount.

Fix PR97315 (part 2 of 2)

gcc/ChangeLog:

PR tree-optimization/97315
* range-op.cc (value_range_with_overflow): Change any
non-overflow calculation in which both bounds are
overflow/underflow to be undefined.

gcc/testsuite/ChangeLog:

* gcc.dg/pr97315-2.c: New test.

Fix PR97315 (part 1 of 2)

gcc/ChangeLog:

PR tree-optimization/97315
* gimple-ssa-evrp.c (hybrid_folder::choose_value): Removes the
trap and instead annotates the listing.

gcc/testsuite/ChangeLog:

* gcc.dg/pr97315-1.c: New test.

openmp: Set cfun->calls_alloca when needed in OpenMP outlined regions [PR97294]

The following testcase FAILs, because we don't mark the child OpenMP function
as cfun->calls_alloca when it does call alloca.  When optimizing, during DCE we
reset those flags and recompute them again, but with -O0 DCE is not performed.

Fixed by calling notice_special_calls when moving insns to the child function.

cfun->calls_alloca is normally set during gimplification and most of the
alloca calls omp-low.c does go through the gimplifier, but one spot didn't
and built the gcall directly, so that one needs to set calls_alloca too.

2020-10-08  Jakub Jelinek  <jakub@redhat.com>

PR sanitizer/97294
* tree-cfg.c (move_block_to_fn): Call notice_special_calls on
call stmts being moved into dest_cfun.
* omp-low.c (lower_rec_input_clauses): Set cfun->calls_alloca when
adding __builtin_alloca_with_align call without gimplification.

* gcc.dg/asan/pr97294.c: New test.

[libgomp, nvptx] Report launch dimensions in GOMP_OFFLOAD_run

Using this patch, when using GOMP_DEBUG=1 and launching a kernel in
GOMP_OFFLOAD_run (used by the omp implementation), we see the kernel launch
dimensions:
...
  GOMP_OFFLOAD_run: kernel main$_omp_fn$0: \
    launch [(teams: 1), 1, 1] [(lanes: 32), (threads: 1), 1]
...

Build on x86_64-linux with nvptx accelerator, tested libgomp.

libgomp/ChangeLog:

2020-10-08  Tom de Vries  <tdevries@suse.de>

PR libgomp/81802
* plugin/plugin-nvptx.c (GOMP_OFFLOAD_run): Report launch
dimensions.

c++: ICE in dependent_type_p with constrained auto [PR97052]

This patch fixes an "unguarded" call to coerce_template_parms in
build_standard_check: processing_template_decl could be zero if we
get here during processing of the first 'auto' parameter of an
abbreviated function template, or if we're processing the type
constraint of a non-templated variable.  In the testcase below, this
leads to an ICE when coerce_template_parms instantiates C's dependent
default template argument.

gcc/cp/ChangeLog:

PR c++/97052
* constraint.cc (build_type_constraint): Temporarily increment
processing_template_decl before calling build_concept_check.
* pt.c (make_constrained_placeholder_type): Likewise.

gcc/testsuite/ChangeLog:

PR c++/97052
* g++.dg/cpp2a/concepts-defarg2.C: New test.

c++: Set the constraints of a class type sooner [PR96229]

In the testcase below, during processing (at parse time) of Y's base
class X<Y>, convert_template_argument calls is_compatible_template_arg
to check if the template argument Y is no more constrained than the
parameter P.  But at this point we haven't yet set Y's constraints, so
get_normalized_constraints_from_decl yields NULL_TREE as the normal form
and caches this result into the normalized_map.

We set Y's constraints later in cp_parser_class_specifier_1 but the
stale normal form in the normalized_map remains.  This ultimately causes
us to miss the constraint failure for Y<Z> because according to the
cached normal form, Y is not constrained.

This patch fixes this issue by moving up the call to
associate_classtype_constraints so that we set constraints before we
start processing a class's bases.

gcc/cp/ChangeLog:

PR c++/96229
* parser.c (cp_parser_class_specifier_1): Move call to
associate_classtype_constraints from here to ...
(cp_parser_class_head): ... here.
* pt.c (is_compatible_template_arg): Correct documentation to
say "argument is _no_ more constrained than the parameter".

gcc/testsuite/ChangeLog:

PR c++/96229
* g++.dg/cpp2a/concepts-class2.C: New test.

Daily bump.

libstdc++: Fix divide by zero in default template argument

libstdc++-v3/ChangeLog:

* include/bits/random.h (__detail::_Mod): Avoid divide by zero.
* testsuite/26_numerics/random/linear_congruential_engine/operators/call.cc:
New test.

libstdc++: Fix non-reserved names in headers

My recent changes to std::exception_ptr moved some members to be inline
in the header but didn't replace the variable names with reserved names.

The "tmp" variable must be fixed. The "other" parameter is actually a
reserved name because of std::allocator<T>::rebind<U>::other but should
be fixed anyway.

There are also some bad uses of "ForwardIterator" in <ranges>.

There's also a "il" parameter in a std::seed_seq constructor in <random>
which is only reserved since C++14.

libstdc++-v3/ChangeLog:

* include/bits/random.h (seed_seq(initializer_list<T>)): Rename
parameter to use reserved name.
* include/bits/ranges_algo.h (shift_left, shift_right): Rename
template parameters to use reserved name.
* libsupc++/exception_ptr.h (exception_ptr): Likewise for
parameters and local variables.
* testsuite/17_intro/names.cc: Check "il". Do not check "d" and
"y" in C++20 mode.

c++: Fix P0846 (ADL and function templates) in template [PR97010]

To quickly recap, P0846 says that a name is also considered to refer to
a template if it is an unqualified-id followed by a < and name lookup
finds either one or more functions or finds nothing.

In a template, when parsing a function call that has type-dependent
arguments, we can't perform ADL right away so we set KOENIG_LOOKUP_P in
the call to remember to do it when instantiating the call
(tsubst_copy_and_build/CALL_EXPR).  When the called function is a
function template, we represent the call with a TEMPLATE_ID_EXPR;
usually the operand is an OVERLOAD.

In the P0846 case though, the operand can be an IDENTIFIER_NODE, when
name lookup found nothing when parsing the template name.  But we
weren't handling this correctly in tsubst_copy_and_build.  First
we need to pass the FUNCTION_P argument from <case TEMPLATE_ID_EXPR> to
<case IDENTIFIER_NODE>, otherwise we give a bogus error.  And then in
<case CALL_EXPR> we need to perform ADL.  The rest of the changes is to
give better errors when ADL didn't find anything.

gcc/cp/ChangeLog:

PR c++/97010
* pt.c (tsubst_copy_and_build) <case TEMPLATE_ID_EXPR>: Call
tsubst_copy_and_build explicitly instead of using the RECUR macro.
Handle a TEMPLATE_ID_EXPR with an IDENTIFIER_NODE as its operand.
<case CALL_EXPR>: Perform ADL for a TEMPLATE_ID_EXPR with an
IDENTIFIER_NODE as its operand.

gcc/testsuite/ChangeLog:

PR c++/97010
* g++.dg/cpp2a/fn-template21.C: New test.
* g++.dg/cpp2a/fn-template22.C: New test.

libgo/configure: remove -fno-section-anchors for AIX

This option is no longer needed. There is no crash without it since
at least gcc-9.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/260157

libgo: handle go1.10+ correctly in match.sh

match.sh was not correctly handling build constraints for Go versions
that have a two-digit suffix, like "go1.10".

The same issue will arise with Go 1.100, but that is a long ways off.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/260077

Rename -fevrp-mode= to --param=evrp-mode=.

* common.opt (-fevrp-mode): Rename and move...
* params.opt (--param=evrp-mode): ...here.
* gimple-range.h (DEBUG_RANGE_CACHE): Use param_evrp_mode instead
of flag_evrp_mode.
* gimple-ssa-evrp.c (rvrp_folder): Same.
(hybrid_folder): Same.
(execute_early_vrp): Same.

tree-optimization/97307 - improve sinking of loads

This improves the heuristics finding a sink location for loads that does
not cross any store.

2020-10-07  Richard Biener  <rguenther@suse.de>

PR tree-optimization/97307
* tree-ssa-sink.c (statement_sink_location): Change heuristic
for not skipping stores to look for virtual definitions
rather than uses.

* gcc.dg/tree-ssa/ssa-sink-17.c: New testcase.
* gcc.dg/vect/pr65947-3.c: XFAIL.

c++: Distinguish alignof and __alignof__ in cp_tree_equal [PR97273]

cp_tree_equal currently considers alignof the same as __alignof__, but
these operators are semantically different ever since r8-7957.  In the
testcase below, this causes the second static_assert to fail on targets
where alignof(double) != __alignof__(double) because the specialization
table (which uses cp_tree_equal as its equality predicate) conflates the
two dependent specializations integral_constant<__alignof__(T)> and
integral_constant<alignof(T)>.

This patch makes cp_tree_equal distinguish between these two operators
by inspecting the ALIGNOF_EXPR_STD_P flag.

gcc/cp/ChangeLog:

PR c++/88115
PR libstdc++/97273
* tree.c (cp_tree_equal) <case ALIGNOF_EXPR>: Return false if
ALIGNOF_EXPR_STD_P differ.

gcc/testsuite/ChangeLog:

PR c++/88115
PR libstdc++/97273
* g++.dg/template/alignof3.C: New test.

Off by one final fix.

Allocate the memory in an approved portable way.

gcc/ChangeLog:

2020-10-06  Andrew MacLeod  <amacleod@redhat.com>

* value-range.h (irange_allocator::allocate): Allocate in two hunks
instead of using the variably-sized trailing array approach.

This patch fixes PR47469 - a trivial bit of tidying up.

2020-07-10  Paul Thomas  <pault@gcc.gnu.org>

gcc/fortran
PR fortran/47469
* trans-expr.c (arrayfunc_assign_needs_temporary): Tidy detection
of pointer and allocatable functions.

analyzer: handle C++ argument numbers and "this" [PR97116]

gcc/analyzer/ChangeLog:
PR analyzer/97116
* sm-malloc.cc (method_p): New.
(describe_argument_index): New.
(inform_nonnull_attribute): Use describe_argument_index.
(possible_null_arg::describe_final_event): Likewise.
(null_arg::describe_final_event): Likewise.

gcc/testsuite/ChangeLog:
PR analyzer/97116
* g++.dg/analyzer/pr97116.C: New test.

Add -fdiagnostics-path-format=separate-events to -fdiagnostics-plain-output

The path-printing default of -fdiagnostics-path-format=inline-events
interacted poorly with -fdiagnostics-plain-output, so it makes most
sense to add -fdiagnostics-path-format=separate-events to
-fdiagnostics-plain-output.

Seen when adding an experimental analyzer plugin to gcc.dg/plugin.exp.

gcc/ChangeLog:
* doc/invoke.texi (-fdiagnostics-plain-output): Add
-fdiagnostics-path-format=separate-events to list of
options injected by -fdiagnostics-plain-output.
* opts-common.c (decode_cmdline_options_to_array): Likewise.

gcc/testsuite/ChangeLog:
* g++.dg/analyzer/analyzer.exp (DEFAULT_CXXFLAGS): Remove
-fdiagnostics-path-format=separate-events.
* gcc.dg/analyzer/analyzer.exp (DEFAULT_CFLAGS): Likewise.
* gcc.dg/plugin/diagnostic-path-format-default.c: Rename to...
* gcc.dg/plugin/diagnostic-path-format-plain.c: ...this.  Remove
dg-options directive.  Copy remainder of test from
diagnostic-path-format-separate-events.c.
* gcc.dg/plugin/diagnostic-test-paths-2.c: Add
-fdiagnostics-path-format=inline-events to options.
Fix expected output for location of conditional within "for" loop.
* gcc.dg/plugin/plugin.exp (plugin_test_list): Update for
renaming.
* gfortran.dg/analyzer/analyzer.exp (DEFAULT_FFLAGS): Remove
-fdiagnostics-path-format=separate-events.

c++: block-scope externs get an alias [PR95677,PR31775,PR95677]

This patch improves block-scope extern handling by always injecting a
hidden copy into the enclosing namespace (or using a match already
there).  This hidden copy will be revealed if the user explicitly
declares it later.  We can get from the DECL_LOCAL_DECL_P local extern
to the alias via DECL_LOCAL_DECL_ALIAS.  This fixes several bugs and
removes the kludgy per-function extern_decl_map.  We only do this
pushing for non-dependent local externs -- dependent ones will be
pushed during instantiation.

User code that expected to be able to handle incompatible local
externs in different block-scopes will no longer work.  That code is
ill-formed.  (always was, despite what 31775 claimed).  I had to
adjust a number of testcases that fell into this.

I tried using DECL_VALUE_EXPR, but that didn't work out.  Due to
constexpr requirements we have to do the replacement very late (it
happens in the gimplifier).   Consider:

extern int l[]; // #1
constexpr bool foo ()
{
   extern int l[3]; // this does not complete the type of decl #1
   constexpr int *p = &l[2]; // ok
   return !p;
}

This requirement, coupled with our use of the common folding machinery
makes pr97306 hard to fix, as we end up with an expression containing
the two different decls for 'l', and only the c++ FE knows how to
reconcile those.  I punted on this.

gcc/cp/
* cp-tree.h (struct language_function): Delete extern_decl_map.
(DECL_LOCAL_DECL_ALIAS): New.
* name-lookup.h (is_local_extern): Delete.
* name-lookup.c (set_local_extern_decl_linkage): Replace with ...
(push_local_extern_decl): ... this new function.
(do_pushdecl): Call new function after pushing new decl.  Unhide
hidden non-functions.
(is_local_extern): Delete.
* decl.c (layout_var_decl): Do not allow VLA local externs.
* decl2.c (mark_used): Also mark DECL_LOCAL_DECL_ALIAS. Drop old
local-extern treatment.
* parser.c (cp_parser_oacc_declare): Deal with local extern aliases.
* pt.c (tsubst_expr): Adjust local extern instantiation.
* cp-gimplify.c (cp_genericize_r): Remap DECL_LOCAL_DECLs.
gcc/testsuite/
* g++.dg/cpp0x/lambda/lambda-sfinae1.C: Avoid ill-formed local extern
* g++.dg/init/pr42844.C: Add expected error.
* g++.dg/lookup/extern-redecl1.C: Likewise.
* g++.dg/lookup/koenig15.C: Avoid ill-formed.
* g++.dg/lto/pr95677.C: New.
* g++.dg/other/nested-extern-1.C: Correct expected behabviour.
* g++.dg/other/nested-extern-2.C: Likewise.
* g++.dg/other/nested-extern.cc: Split ...
* g++.dg/other/nested-extern-1.cc: ... here ...
* g++.dg/other/nested-extern-2.cc: ... here.
* g++.dg/template/scope5.C: Avoid ill-formed
* g++.old-deja/g++.law/missed-error2.C: Allow extension.
* g++.old-deja/g++.pt/crash3.C: Add expected error.

ipa-prop: Fix multiple-target speculation resolution

As the FIXME which this patch removes states, the current code does
not work when a call with multiple speculative targets gets resolved
through parameter tracking during inlining - it feeds the inliner an
edge it has already dealt with.  The patch makes the code which should
prevent it aware of the possibility that that speculation can have
more than one target now.

gcc/ChangeLog:

2020-09-30  Martin Jambor  <mjambor@suse.cz>

PR ipa/96394
* ipa-prop.c (update_indirect_edges_after_inlining): Do not add
resolved speculation edges to vector of new direct edges even in
presence of multiple speculative direct edges for a single call.

gcc/testsuite/ChangeLog:

2020-09-30  Martin Jambor  <mjambor@suse.cz>

PR ipa/96394
* gcc.dg/tree-prof/pr96394.c: New test.

c++: Rename DECL_BUILTIN_P to DECL_UNDECLARED_BUILTIN_P

I realized I'd misnamed DECL_BUILTIN_P, it's only true of compiler
builtins unless and until the user declares them -- at that point
they're real decls, and will have a location in the user's source.
(BUILT_IN_FN and friends still work though).  This renames them so
future-me is not confused as to why the predicate becomes false.

gcc/cp/
* cp-tree.h (DECL_BUILTIN_P): Rename to ...
(DECL_UNDECLARED_BUILTIN_P): ... here.
* decl.c (duplicate_decls): Adjust.
* name-lookup.c (anticipated_builtin_p): Adjust.
(do_nonmember_using_decl): Likewise.
libcc1/
* libcp1plugin.cc (supplement_binding): Rename
DECL_BUILTIN_P.

c++:  Adding exception specs can changed dependentness

Making an exception variant can cause a non-dependent function type to
become dependent (since c++17 eh-specs are part of the type).  The
same is (possibly?) true for adding a late return type.  Fixed thusly.
My upcoming local extern-decl changes have a test case that covers
this (which was how I found it).

gcc/cp/
* tree.c (build_cp_fntype_variant): Clear
TYPE_DEPENDENT_P_VALID if necessary.

amdgcn: Use scalar instructions for addptrdi3

Allow addptr to use SPGRs as well as VGPRs for pointers.  This ought to
prevent some unnecessary copying back and forth.

gcc/ChangeLog:

* config/gcn/gcn.md (unspec): Add UNSPEC_ADDPTR.
(addptrdi3): Add SGPR alternative.

Output filepath strings in .debug_line_str for DWARF5

DWARF5 has a new string table specially for file paths. .debug_line
file and dir tables reference strings in .debug_line_str.  If a
.debug_line_str section is emitted then also place CU DIE file
names and comp dirs there.

gcc/ChangeLog:

* dwarf2out.c (add_filepath_AT_string): New function.
(asm_outputs_debug_line_str): Likewise.
(add_filename_attribute): Likewise.
(add_comp_dir_attribute): Call add_filepath_AT_string.
(gen_compile_unit_die): Call add_filename_attribute for name.
(init_sections_and_labels): Init debug_line_str_section when
asm_outputs_debug_line_str return true.
(dwarf2out_early_finish): Remove DW_AT_name and DW_AT_comp_dir
hack and call add_filename_attribute for the remap_debug_filename.

debug: Pass --gdwarf-N to assembler if fixed gas is detected during configure

> > As for the test assembly, I'd say we should take
> > #define F void foo (void) {}
> > F
> > compile it with
> > gcc -S -O2 -g1 -dA -gno-as-loc-support -fno-merge-debug-strings
> > remove .cfi_* directives, remove the ret instruction, change @function
> > and @progbits to %function and %progbits, change .uleb128 to just .byte,
> > I think all the values should be small enough, maybe change .value to
> > .2byte and .long to .4byte (whatever is most portable across different
> > arches and gas versions), simplify (shorten) strings and adjust
> > sizes, and do something with the .quad directives, that is dependent on
> > the address size, perhaps just take those attributes out and adjust
> > .debug_abbrev?  Finally, remove all comments (emit them in the first case
> > just to better understand the debug info).
>
> I'm afraid it is hard to avoid the .quad or .8byte.
> Here is a 64-bit address version that assembles fine by both x86_64 and
> aarch64 as.
> Unfortunately doesn't fail with broken gas versions with -gdwarf-2 without
> the nop, so we'll need at least a nop in there.
> Fortunately gcc/configure.ac already determines the right nop insn for the
> target, in $insn.
> So I guess what we want next is have the 32-bit version of this with .4byte
> instead of .8byte and just let's try to assemble both versions, first
> without -gdwarf-2 and the one that succeeds assemble again with -gdwarf-2
> and check for the duplicate .debug_line sections error.

Ok, here it is in patch form.
I've briefly tested it, with the older binutils I have around (no --gdwarf-N
support), with latest gas (--gdwarf-N that can be passed to as even when
compiling C/C++ etc. code and emitting .debug_line) and latest gas with Mark's fix
reverted (--gdwarf-N support, but can only pass it to as when assembling
user .s/.S files, not when compiling C/C++ etc.).

2020-10-07  Jakub Jelinek  <jakub@redhat.com>

* configure.ac (HAVE_AS_GDWARF_5_DEBUG_FLAG,
HAVE_AS_WORKING_DWARF_4_FLAG): New tests.
* gcc.c (ASM_DEBUG_DWARF_OPTION): Define.
(ASM_DEBUG_SPEC): Use ASM_DEBUG_DWARF_OPTION instead of
"--gdwarf2".  Use %{cond:opt1;:opt2} style.
(ASM_DEBUG_OPTION_DWARF_OPT): Define.
(ASM_DEBUG_OPTION_SPEC): Define.
(asm_debug_option): New variable.
(asm_options): Add "%(asm_debug_option)".
(static_specs): Add asm_debug_option entry.
(static_spec_functions): Add dwarf-version-gt.
(debug_level_greater_than_spec_func): New function.
* config/darwin.h (ASM_DEBUG_OPTION_SPEC): Define.
* config/darwin9.h (ASM_DEBUG_OPTION_SPEC): Redefine.
* config.in: Regenerated.
* configure: Regenerated.

options: Avoid unused variable mask warning [PR97305]

> options-save.c: In function 'void cl_target_option_save(cl_target_option*, gcc_options*, gcc_options*)':
> options-save.c:8526:26: error: unused variable 'mask' [-Werror=unused-variable]
>  8526 |   unsigned HOST_WIDE_INT mask = 0;
>       |                          ^~~~
> options-save.c: In function 'void cl_target_option_restore(gcc_options*, gcc_options*, cl_target_option*)':
> options-save.c:8537:26: error: unused variable 'mask' [-Werror=unused-variable]
>  8537 |   unsigned HOST_WIDE_INT mask;
>       |                          ^~~~

Oops, missed that, sorry.

The following patch should fix that, tested on x86_64-linux make
options-save.c (same file as before) and -> ia64-linux cross make
options-save.o (no warning anymore, just the unwanted declarations gone).

2020-10-07  Jakub Jelinek  <jakub@redhat.com>

PR bootstrap/97305
* optc-save-gen.awk: Don't declare mask variable if explicit_mask
array is not present.

openmp: Improve composite simd vectorization

> > I was really hoping bbs 4 and 5 would be one loop (the one I set safelen
> > and force_vectorize etc. for) and that basic blocks 6 and 7 would be
> > together with that inner loop another loop, but apparently loop discovery
> > thinks it is just one loop.
> > Any ideas what I'm doing wrong or is there any way how to make it two loops
> > (that would also survive all the cfg cleanups until vectorization)?
>
> The early CFG looks like we have a common header with two latches
> so it boils down to how we disambiguate those in the end (we seem
> to unify the latches via a forwarder).  IIRC OMP lowering builds
> loops itself, could it not do the appropriate disambiguation itself?

I realized I emit the same stmts on both paths (before goto doit; and before
falling through it), at least the MIN_EXPR and PLUS_EXPR, so by forcing
there an extra bb which does those two and having the "doit" label before
that the innermost loop doesn't have multiple latches anymore and so is
vectorized fine.

2020-10-07  Jakub Jelinek  <jakub@redhat.com>

* omp-expand.c (expand_omp_simd): Don't emit MIN_EXPR and PLUS_EXPR
at the end of entry_bb and innermost init_bb, instead force arguments
for MIN_EXPR into temporaries in both cases and jump to a new bb that
performs MIN_EXPR and PLUS_EXPR.

* gcc.dg/gomp/simd-2.c: New test.
* gcc.dg/gomp/simd-3.c: New test.

[tree-ssa-loop-ch] Add missing NULL test for dump_file

If we change gimple_can_duplicate_bb_p to return false instead of true, we run
into a segfault in ch_base::copy_headers due to using dump_file while it's
NULL:
...
      if (!gimple_duplicate_sese_region (entry, exit, bbs, n_bbs, copied_bbs,
                                        true))
       {
         fprintf (dump_file, "Duplication failed.\n");
         continue;
       }
...

Fix this by adding the missing dump_file != NULL test.

Tested by rebuilding lto1 and rerunning the failing test-case.

gcc/ChangeLog:

2020-10-07  Tom de Vries  <tdevries@suse.de>

* tree-ssa-loop-ch.c (ch_base::copy_headers): Add missing NULL test
for dump_file.

Daily bump.

c++: typename in out-of-class member function definitions [PR97297]

I was notified that our P0634R3 (Down with typename) implementation has
a flaw: when we have an out-of-class member function definition, we
still required 'typename' for its parameters.  For example here:

  template <typename T> struct S {
    int simple(T::type);
  };
  template <typename T>
  int S<T>::simple(/* typename */T::type) { return 0; }

the 'typename' isn't necessary per [temp.res]/5.2.4.  We have a qualified
name here ("S<T>::simple") so we know it's already been declared so we
can look it up to see if it's a function template or a variable
template.

In this case, the P0634R3 code in cp_parser_direct_declarator wasn't
looking into uninstantiated templates and didn't find the member
function 'simple' -- cp_parser_lookup_name returned a SCOPE_REF which
means that the qualifying scope was dependent.  With this fix, we find
the BASELINK for 'simple', don't clear CP_PARSER_FLAGS_TYPENAME_OPTIONAL
from the flags, and the typename is implicitly assumed.

gcc/cp/ChangeLog:

PR c++/97297
* parser.c (cp_parser_direct_declarator): When checking if a
name is a function template declaration for the P0634R3 case,
look in uninstantiated templates too.

gcc/testsuite/ChangeLog:

PR c++/97297
* g++.dg/cpp2a/typename18.C: New test.

c-c++-common/goacc/declare-pr90861.c: Remove xfail

gcc/testsuite/ChangeLog
PR middle-end/90861
* c-c++-common/goacc/declare-pr90861.c: Remove xfail.

compiler: avoid undefined behavior in Import::read

For some implementations of Stream, advancing the stream will invalidate
the previously-returned peek buffer. Copy the peek buffer before
advancing in Import::read to avoid this undefined behavior.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/259438

Hybrid EVRP and testcases

Provide a hybrid EVRP pass which uses legacy EVRP and adds additonal
enhancements from the new ranger infrastructure.
A New option is also provided, -fevrp-mode=
And adjust testcases

gcc/ChangeLog:

2020-10-06  Andrew MacLeod  <amacleod@redhat.com>

* flag-types.h (enum evrp_mode): New enumerated type EVRP_MODE_*.
* common.opt (fevrp-mode): New undocumented flag.
* gimple-ssa-evrp.c: Include gimple-range.h
(class rvrp_folder): EVRP folding using ranger exclusively.
(rvrp_folder::rvrp_folder): New.
(rvrp_folder::~rvrp_folder): New.
(rvrp_folder::value_of_expr): New.  Use rangers value_of_expr.
(rvrp_folder::value_on_edge): New.  Use rangers value_on_edge.
(rvrp_folder::value_of_Stmt): New.  Use rangers value_of_stmt.
(rvrp_folder::fold_stmt): New.  Call the simplifier.
(class hybrid_folder): EVRP folding using both engines.
(hybrid_folder::hybrid_folder): New.
(hybrid_folder::~hybrid_folder): New.
(hybrid_folder::fold_stmt): New.  Simplify with one engne, then the
other.
(hybrid_folder::value_of_expr): New.  Use both value routines.
(hybrid_folder::value_on_edge): New.  Use both value routines.
(hybrid_folder::value_of_stmt): New.  Use both value routines.
(hybrid_folder::choose_value): New.  Choose between range_analzyer and
rangers values.
(execute_early_vrp): Choose a folder based on flag_evrp_mode.
* vr-values.c (simplify_using_ranges::fold_cond): Try range_of_stmt
first to see if it returns a value.
(simplify_using_ranges::simplify_switch_using_ranges): Return true if
any changes were made to the switch.

gcc/testsuite/ChangeLog:

2020-10-06  Andrew MacLeod  <amacleod@redhat.com>

* gcc.dg/pr81192.c: Disable EVRP pass.
* gcc.dg/tree-ssa/pr77445-2.c: Ditto.
* gcc.dg/tree-ssa/ssa-dom-thread-6.c: Adjust.
* gcc.dg/tree-ssa/ssa-dom-thread-7.c: Ditto.

Ranger classes.

Add the 8 ranger files and the Makefile changes to build it.

2020-10-06  Andrew MacLeod  <amacleod@redhat.com>

* Makefile.in (OBJS): Add gimple-range*.o.
* gimple-range.h: New file.
* gimple-range.cc: New file.
* gimple-range-cache.h: New file.
* gimple-range-cache.cc: New file.
* gimple-range-edge.h: New file.
* gimple-range-edge.cc: New file.
* gimple-range-gori.h: New file.
* gimple-range-gori.cc: New file.

[openacc, libgomp, testsuite] Xfail declare-5.f90

We're currently running into:
...
FAIL: libgomp.oacc-fortran/declare-5.f90 -DACC_DEVICE_TYPE_nvidia=1 \
  -DACC_MEM_SHARED=0 -foffload=nvptx-none  -O0  execution test
FAIL: libgomp.oacc-fortran/declare-5.f90 -DACC_DEVICE_TYPE_nvidia=1 \
  -DACC_MEM_SHARED=0 -foffload=nvptx-none  -O1  execution test
FAIL: libgomp.oacc-fortran/declare-5.f90 -DACC_DEVICE_TYPE_nvidia=1 \
  -DACC_MEM_SHARED=0 -foffload=nvptx-none  -O2  execution test
FAIL: libgomp.oacc-fortran/declare-5.f90 -DACC_DEVICE_TYPE_nvidia=1 \
  -DACC_MEM_SHARED=0 -foffload=nvptx-none  -O3 -fomit-frame-pointer \
  -funroll-loops -fpeel-loops -ftracer -finline-functions  execution test
FAIL: libgomp.oacc-fortran/declare-5.f90 -DACC_DEVICE_TYPE_nvidia=1 \
  -DACC_MEM_SHARED=0 -foffload=nvptx-none  -O3 -g  execution test
FAIL: libgomp.oacc-fortran/declare-5.f90 -DACC_DEVICE_TYPE_nvidia=1 \
  -DACC_MEM_SHARED=0 -foffload=nvptx-none  -Os  execution test
...

A PR was filed for this: PR92790 - "[OpenACC] declare device_resident -
Fortran common blocks not handled / libgomp.oacc-fortran/declare-5.f90 fails"

Xfail the fails.

Tested on x86_64-linux with nvptx accelerator.

libgomp/ChangeLog:

2020-10-06  Tom de Vries  <tdevries@suse.de>

* testsuite/libgomp.oacc-fortran/declare-5.f90: Add xfail for PR92790.

libstdc++: Inline std::exception_ptr members [PR 90295]

This inlines most members of std::exception_ptr so that all operations
on a null exception_ptr can be optimized away. This benefits code like
std::future and coroutines where an exception_ptr object is present to
cope with exceptional cases, but is usually not used and remains null.

Since those functions were previously non-inline we have to continue to
export them from the library, for objects that were compiled against the
old headers and expect to find definitions in the library.

In order to inline the copy constructor and destructor we need to export
the _M_addref() and _M_release() members that increment/decrement the
reference count when copying/destroying a non-null exception_ptr. The
copy ctor and dtor check for null and don't call _M_addref and
_M_release unless they need to. The checks for null pointers in
_M_addref and _M_release are still needed because old code might call
them without checking for null first. But we can use __builtin_expect to
predict that they are usually called for the non-null case.

libstdc++-v3/ChangeLog:

PR libstdc++/90295
* config/abi/pre/gnu.ver (CXXABI_1.3.13): New symbol version.
(exception_ptr::_M_addref(), exception_ptr::_M_release()):
Export symbols.
* libsupc++/eh_ptr.cc (exception_ptr::exception_ptr()):
Remove out-of-line definition.
(exception_ptr::exception_ptr(const exception_ptr&)):
Likewise.
(exception_ptr::~exception_ptr()): Likewise.
(exception_ptr::operator=(const exception_ptr&)):
Likewise.
(exception_ptr::swap(exception_ptr&)): Likewise.
(exception_ptr::_M_addref()): Add branch prediction.
* libsupc++/exception_ptr.h (exception_ptr::operator bool):
Add noexcept.
[!_GLIBCXX_EH_PTR_COMPAT] (operator==, operator!=): Define
inline as hidden friends. Remove declarations at namespace
scope.
(exception_ptr::exception_ptr()): Define inline.
(exception_ptr::exception_ptr(const exception_ptr&)):
Likewise.
(exception_ptr::~exception_ptr()): Likewise.
(exception_ptr::operator=(const exception_ptr&)):
Likewise.
(exception_ptr::swap(exception_ptr&)): Likewise.
* testsuite/util/testsuite_abi.cc: Add CXXABI_1.3.13.
* testsuite/18_support/exception_ptr/90295.cc: New test.

arm: Enable MVE SIMD modes for vectorization

This patch enables SIMD modes for MVE auto-vectorization.
In this patch, the integer and float MVE SIMD modes are returned by
arm_preferred_simd_mode (TARGET_VECTORIZE_PREFERRED_SIMD_MODE hook) when
MVE or MVE_FLOAT is enabled. Then the expanders for auto-vectorization
can be used for generating MVE SIMD code.

This patch also fixes bugs in MVE vreiterpretq_*.c tests which are
revealed by the enabled MVE SIMD modes.
The tests are for checking the MVE reinterpret intrinsics.
There are two functions in each of the tests. The two functions contain
the pattern of identical code so that they are folded in icf pass.
Because of icf, the instruction count only checks one function which is
8. However when the SIMD modes are enabled, the estimation of the code
size becomes smaller so that inlining is applied after icf, then the
instruction count becomes 16 which causes failure of the tests.
Because the icf is not the expected pattern to be tested but causes
above issues, -fno-ipa-icf is applied to the tests to avoid unstable
instruction count.

gcc/ChangeLog:

2020-10-05  Dennis Zhang  <dennis.zhang@arm.com>

* config/arm/arm.c (arm_preferred_simd_mode): Enable MVE SIMD modes.

gcc/testsuite/ChangeLog:

2020-10-05  Dennis Zhang  <dennis.zhang@arm.com>

* gcc.target/arm/mve/intrinsics/vreinterpretq_f16.c: Use additional
option -fno-ipa-icf and change the instruction count from 8 to 16.
* gcc.target/arm/mve/intrinsics/vreinterpretq_f32.c: Likewise.
* gcc.target/arm/mve/intrinsics/vreinterpretq_s16.c: Likewise.
* gcc.target/arm/mve/intrinsics/vreinterpretq_s32.c: Likewise.
* gcc.target/arm/mve/intrinsics/vreinterpretq_s64.c: Likewise.
* gcc.target/arm/mve/intrinsics/vreinterpretq_s8.c: Likewise.
* gcc.target/arm/mve/intrinsics/vreinterpretq_u16.c: Likewise.
* gcc.target/arm/mve/intrinsics/vreinterpretq_u32.c: Likewise.
* gcc.target/arm/mve/intrinsics/vreinterpretq_u64.c: Likewise.
* gcc.target/arm/mve/intrinsics/vreinterpretq_u8.c: Likewise.

[openacc] Fix acc declare for VLAs

Consider test-case test.c, with VLA A:
...
int main (void) {
  int N = 1000;
  int A[N];
  #pragma acc declare copy(A)
  return 0;
}
...
compiled using:
...
$ gcc test.c -fopenacc -S -fdump-tree-all
...

At original, we have:
...
  #pragma acc declare map(tofrom:A);
...
but at gimple, we have a map (to:A.1), but not a map (from:A.1):
...
  int[0:D.2074] * A.1;

  {
    int A[0:D.2074] [value-expr: *A.1];

    saved_stack.2 = __builtin_stack_save ();
    try
      {
        A.1 = __builtin_alloca_with_align (D.2078, 32);
        #pragma omp target oacc_declare map(to:(*A.1) [len: D.2076])
      }
    finally
      {
        __builtin_stack_restore (saved_stack.2);
      }
  }
...

This is caused by the following incompatibility.  When storing the desired
from clause in oacc_declare_returns, we use 'A.1' as the key:
...
10898                 oacc_declare_returns->put (decl, c);
(gdb) call debug_generic_expr (decl)
A.1
(gdb) call debug_generic_expr (c)
map(from:(*A.1))
...
but when looking it up, we use 'A' as the key:
...
(gdb)
1471                  tree *c = oacc_declare_returns->get (t);
(gdb) call debug_generic_expr (t)
A
...

Fix this by extracing the 'A.1' lookup key from 'A' using the decl-expr.

In addition, unshare the looked up value, to fix avoid running into
an "incorrect sharing of tree nodes" error.

Using these two fixes, we get our desired:
...
     finally
       {
+        #pragma omp target oacc_declare map(from:(*A.1))
         __builtin_stack_restore (saved_stack.2);
       }
...

Build on x86_64-linux with nvptx accelerator, tested libgomp.

gcc/ChangeLog:

2020-10-06  Tom de Vries  <tdevries@suse.de>

PR middle-end/90861
* gimplify.c (gimplify_bind_expr): Handle lookup in
oacc_declare_returns using key with decl-expr.

libgomp/ChangeLog:

2020-10-06  Tom de Vries  <tdevries@suse.de>

PR middle-end/90861
* testsuite/libgomp.oacc-c-c++-common/declare-vla.c: Remove xfail.

lto: fix LTO debug sections copying.

readelf -S prints:

There are 81999 section headers, starting at offset 0x1f488060:

Section Headers:
  [Nr] Name              Type            Address          Off    Size   ES Flg Lk Inf Al
  [ 0]                   NULL            0000000000000000 000000 01404f 00     81998   0  0
  [ 1] .group            GROUP           0000000000000000 000040 000008 04     81995 105027  4
...
  [81995] .symtab           SYMTAB          0000000000000000 d5d9298 2db310 18     81997 105026  8
  [81996] .symtab_shndx     SYMTAB SECTION INDICES 0000000000000000 d8b45a8 079dd8 04     81995   0  4
  [81997] .strtab           STRTAB          0000000000000000 d92e380 80460c 00      0   0  1
...

Looking at the documentation:
Table 7–15 ELF sh_link and sh_info Interpretation

sh_type - sh_link
SHT_SYMTAB - The section header index of the associated string table.
SHT_SYMTAB_SHNDX - The section header index of the associated symbol table.

As seen, sh_link of a SHT_SYMTAB always points to a .strtab and readelf
confirms that.

So we need to use reverse mapping taken from
  [81996] .symtab_shndx     SYMTAB SECTION INDICES 0000000000000000 d8b45a8 079dd8 04     81995   0  4

where sh_link points to 81995.

libiberty/ChangeLog:

PR lto/97290
* simple-object-elf.c (simple_object_elf_copy_lto_debug_sections):
Use sh_link of a .symtab_shndx section.

[PATCH][GCC] arm: Move iterators from mve.md to iterators.md to maintain consistency.

To maintain consistency with other Arm Architectures backend, iterators and iterator attributes are moved
from mve.md file to iterators.md. Also move enumerators for MVE unspecs from mve.md file to unspecs.md file.

gcc/ChangeLog:

2020-10-06  Srinath Parvathaneni  <srinath.parvathaneni@arm.com>

* config/arm/iterators.md (MVE_types): Move mode iterator from mve.md to
iterators.md.
(MVE_VLD_ST): Likewise.
(MVE_0): Likewise.
(MVE_1): Likewise.
(MVE_3): Likewise.
(MVE_2): Likewise.
(MVE_5): Likewise.
(MVE_6): Likewise.
(MVE_CNVT): Move mode attribute iterator from mve.md to iterators.md.
(MVE_LANES): Likewise.
(MVE_constraint): Likewise.
(MVE_constraint1): Likewise.
(MVE_constraint2): Likewise.
(MVE_constraint3): Likewise.
(MVE_pred): Likewise.
(MVE_pred1): Likewise.
(MVE_pred2): Likewise.
(MVE_pred3): Likewise.
(MVE_B_ELEM): Likewise.
(MVE_H_ELEM): Likewise.
(V_sz_elem1): Likewise.
(V_extr_elem): Likewise.
(earlyclobber_32): Likewise.
(supf): Move int attribute from mve.md to iterators.md.
(mode1): Likewise.
(VCVTQ_TO_F): Move int iterator from mve.md to iterators.md.
(VMVNQ_N): Likewise.
(VREV64Q): Likewise.
(VCVTQ_FROM_F): Likewise.
(VREV16Q): Likewise.
(VCVTAQ): Likewise.
(VMVNQ): Likewise.
(VDUPQ_N): Likewise.
(VCLZQ): Likewise.
(VADDVQ): Likewise.
(VREV32Q): Likewise.
(VMOVLBQ): Likewise.
(VMOVLTQ): Likewise.
(VCVTPQ): Likewise.
(VCVTNQ): Likewise.
(VCVTMQ): Likewise.
(VADDLVQ): Likewise.
(VCTPQ): Likewise.
(VCTPQ_M): Likewise.
(VCVTQ_N_TO_F): Likewise.
(VCREATEQ): Likewise.
(VSHRQ_N): Likewise.
(VCVTQ_N_FROM_F): Likewise.
(VADDLVQ_P): Likewise.
(VCMPNEQ): Likewise.
(VSHLQ): Likewise.
(VABDQ): Likewise.
(VADDQ_N): Likewise.
(VADDVAQ): Likewise.
(VADDVQ_P): Likewise.
(VANDQ): Likewise.
(VBICQ): Likewise.
(VBRSRQ_N): Likewise.
(VCADDQ_ROT270): Likewise.
(VCADDQ_ROT90): Likewise.
(VCMPEQQ): Likewise.
(VCMPEQQ_N): Likewise.
(VCMPNEQ_N): Likewise.
(VEORQ): Likewise.
(VHADDQ): Likewise.
(VHADDQ_N): Likewise.
(VHSUBQ): Likewise.
(VHSUBQ_N): Likewise.
(VMAXQ): Likewise.
(VMAXVQ): Likewise.
(VMINQ): Likewise.
(VMINVQ): Likewise.
(VMLADAVQ): Likewise.
(VMULHQ): Likewise.
(VMULLBQ_INT): Likewise.
(VMULLTQ_INT): Likewise.
(VMULQ): Likewise.
(VMULQ_N): Likewise.
(VORNQ): Likewise.
(VORRQ): Likewise.
(VQADDQ): Likewise.
(VQADDQ_N): Likewise.
(VQRSHLQ): Likewise.
(VQRSHLQ_N): Likewise.
(VQSHLQ): Likewise.
(VQSHLQ_N): Likewise.
(VQSHLQ_R): Likewise.
(VQSUBQ): Likewise.
(VQSUBQ_N): Likewise.
(VRHADDQ): Likewise.
(VRMULHQ): Likewise.
(VRSHLQ): Likewise.
(VRSHLQ_N): Likewise.
(VRSHRQ_N): Likewise.
(VSHLQ_N): Likewise.
(VSHLQ_R): Likewise.
(VSUBQ): Likewise.
(VSUBQ_N): Likewise.
(VADDLVAQ): Likewise.
(VBICQ_N): Likewise.
(VMLALDAVQ): Likewise.
(VMLALDAVXQ): Likewise.
(VMOVNBQ): Likewise.
(VMOVNTQ): Likewise.
(VORRQ_N): Likewise.
(VQMOVNBQ): Likewise.
(VQMOVNTQ): Likewise.
(VSHLLBQ_N): Likewise.
(VSHLLTQ_N): Likewise.
(VRMLALDAVHQ): Likewise.
(VBICQ_M_N): Likewise.
(VCVTAQ_M): Likewise.
(VCVTQ_M_TO_F): Likewise.
(VQRSHRNBQ_N): Likewise.
(VABAVQ): Likewise.
(VSHLCQ): Likewise.
(VRMLALDAVHAQ): Likewise.
(VADDVAQ_P): Likewise.
(VCLZQ_M): Likewise.
(VCMPEQQ_M_N): Likewise.
(VCMPEQQ_M): Likewise.
(VCMPNEQ_M_N): Likewise.
(VCMPNEQ_M): Likewise.
(VDUPQ_M_N): Likewise.
(VMAXVQ_P): Likewise.
(VMINVQ_P): Likewise.
(VMLADAVAQ): Likewise.
(VMLADAVQ_P): Likewise.
(VMLAQ_N): Likewise.
(VMLASQ_N): Likewise.
(VMVNQ_M): Likewise.
(VPSELQ): Likewise.
(VQDMLAHQ_N): Likewise.
(VQRDMLAHQ_N): Likewise.
(VQRDMLASHQ_N): Likewise.
(VQRSHLQ_M_N): Likewise.
(VQSHLQ_M_R): Likewise.
(VREV64Q_M): Likewise.
(VRSHLQ_M_N): Likewise.
(VSHLQ_M_R): Likewise.
(VSLIQ_N): Likewise.
(VSRIQ_N): Likewise.
(VMLALDAVQ_P): Likewise.
(VQMOVNBQ_M): Likewise.
(VMOVLTQ_M): Likewise.
(VMOVNBQ_M): Likewise.
(VRSHRNTQ_N): Likewise.
(VORRQ_M_N): Likewise.
(VREV32Q_M): Likewise.
(VREV16Q_M): Likewise.
(VQRSHRNTQ_N): Likewise.
(VMOVNTQ_M): Likewise.
(VMOVLBQ_M): Likewise.
(VMLALDAVAQ): Likewise.
(VQSHRNBQ_N): Likewise.
(VSHRNBQ_N): Likewise.
(VRSHRNBQ_N): Likewise.
(VMLALDAVXQ_P): Likewise.
(VQMOVNTQ_M): Likewise.
(VMVNQ_M_N): Likewise.
(VQSHRNTQ_N): Likewise.
(VMLALDAVAXQ): Likewise.
(VSHRNTQ_N): Likewise.
(VCVTMQ_M): Likewise.
(VCVTNQ_M): Likewise.
(VCVTPQ_M): Likewise.
(VCVTQ_M_N_FROM_F): Likewise.
(VCVTQ_M_FROM_F): Likewise.
(VRMLALDAVHQ_P): Likewise.
(VADDLVAQ_P): Likewise.
(VABAVQ_P): Likewise.
(VSHLQ_M): Likewise.
(VSRIQ_M_N): Likewise.
(VSUBQ_M): Likewise.
(VCVTQ_M_N_TO_F): Likewise.
(VHSUBQ_M): Likewise.
(VSLIQ_M_N): Likewise.
(VRSHLQ_M): Likewise.
(VMINQ_M): Likewise.
(VMULLBQ_INT_M): Likewise.
(VMULHQ_M): Likewise.
(VMULQ_M): Likewise.
(VHSUBQ_M_N): Likewise.
(VHADDQ_M_N): Likewise.
(VORRQ_M): Likewise.
(VRMULHQ_M): Likewise.
(VQADDQ_M): Likewise.
(VRSHRQ_M_N): Likewise.
(VQSUBQ_M_N): Likewise.
(VADDQ_M): Likewise.
(VORNQ_M): Likewise.
(VRHADDQ_M): Likewise.
(VQSHLQ_M): Likewise.
(VANDQ_M): Likewise.
(VBICQ_M): Likewise.
(VSHLQ_M_N): Likewise.
(VCADDQ_ROT270_M): Likewise.
(VQRSHLQ_M): Likewise.
(VQADDQ_M_N): Likewise.
(VADDQ_M_N): Likewise.
(VMAXQ_M): Likewise.
(VQSUBQ_M): Likewise.
(VMLASQ_M_N): Likewise.
(VMLADAVAQ_P): Likewise.
(VBRSRQ_M_N): Likewise.
(VMULQ_M_N): Likewise.
(VCADDQ_ROT90_M): Likewise.
(VMULLTQ_INT_M): Likewise.
(VEORQ_M): Likewise.
(VSHRQ_M_N): Likewise.
(VSUBQ_M_N): Likewise.
(VHADDQ_M): Likewise.
(VABDQ_M): Likewise.
(VMLAQ_M_N): Likewise.
(VQSHLQ_M_N): Likewise.
(VMLALDAVAQ_P): Likewise.
(VMLALDAVAXQ_P): Likewise.
(VQRSHRNBQ_M_N): Likewise.
(VQRSHRNTQ_M_N): Likewise.
(VQSHRNBQ_M_N): Likewise.
(VQSHRNTQ_M_N): Likewise.
(VRSHRNBQ_M_N): Likewise.
(VRSHRNTQ_M_N): Likewise.
(VSHLLBQ_M_N): Likewise.
(VSHLLTQ_M_N): Likewise.
(VSHRNBQ_M_N): Likewise.
(VSHRNTQ_M_N): Likewise.
(VSTRWSBQ): Likewise.
(VSTRBSOQ): Likewise.
(VSTRBQ): Likewise.
(VLDRBGOQ): Likewise.
(VLDRBQ): Likewise.
(VLDRWGBQ): Likewise.
(VLD1Q): Likewise.
(VLDRHGOQ): Likewise.
(VLDRHGSOQ): Likewise.
(VLDRHQ): Likewise.
(VLDRWQ): Likewise.
(VLDRDGBQ): Likewise.
(VLDRDGOQ): Likewise.
(VLDRDGSOQ): Likewise.
(VLDRWGOQ): Likewise.
(VLDRWGSOQ): Likewise.
(VST1Q): Likewise.
(VSTRHSOQ): Likewise.
(VSTRHSSOQ): Likewise.
(VSTRHQ): Likewise.
(VSTRWQ): Likewise.
(VSTRDSBQ): Likewise.
(VSTRDSOQ): Likewise.
(VSTRDSSOQ): Likewise.
(VSTRWSOQ): Likewise.
(VSTRWSSOQ): Likewise.
(VSTRWSBWBQ): Likewise.
(VLDRWGBWBQ): Likewise.
(VSTRDSBWBQ): Likewise.
(VLDRDGBWBQ): Likewise.
(VADCIQ): Likewise.
(VADCIQ_M): Likewise.
(VSBCQ): Likewise.
(VSBCQ_M): Likewise.
(VSBCIQ): Likewise.
(VSBCIQ_M): Likewise.
(VADCQ): Likewise.
(VADCQ_M): Likewise.
(UQRSHLLQ): Likewise.
(SQRSHRLQ): Likewise.
(VSHLCQ_M): Likewise.
* config/arm/mve.md (MVE_types): Move mode iterator to iterators.md from mve.md.
(MVE_VLD_ST): Likewise.
(MVE_0): Likewise.
(MVE_1): Likewise.
(MVE_3): Likewise.
(MVE_2): Likewise.
(MVE_5): Likewise.
(MVE_6): Likewise.
(MVE_CNVT): Move mode attribute iterator to iterators.md from mve.md.
(MVE_LANES): Likewise.
(MVE_constraint): Likewise.
(MVE_constraint1): Likewise.
(MVE_constraint2): Likewise.
(MVE_constraint3): Likewise.
(MVE_pred): Likewise.
(MVE_pred1): Likewise.
(MVE_pred2): Likewise.
(MVE_pred3): Likewise.
(MVE_B_ELEM): Likewise.
(MVE_H_ELEM): Likewise.
(V_sz_elem1): Likewise.
(V_extr_elem): Likewise.
(earlyclobber_32): Likewise.
(supf): Move int attribute to iterators.md from mve.md.
(mode1): Likewise.
(VCVTQ_TO_F): Move int iterator to iterators.md from mve.md.
(VMVNQ_N): Likewise.
(VREV64Q): Likewise.
(VCVTQ_FROM_F): Likewise.
(VREV16Q): Likewise.
(VCVTAQ): Likewise.
(VMVNQ): Likewise.
(VDUPQ_N): Likewise.
(VCLZQ): Likewise.
(VADDVQ): Likewise.
(VREV32Q): Likewise.
(VMOVLBQ): Likewise.
(VMOVLTQ): Likewise.
(VCVTPQ): Likewise.
(VCVTNQ): Likewise.
(VCVTMQ): Likewise.
(VADDLVQ): Likewise.
(VCTPQ): Likewise.
(VCTPQ_M): Likewise.
(VCVTQ_N_TO_F): Likewise.
(VCREATEQ): Likewise.
(VSHRQ_N): Likewise.
(VCVTQ_N_FROM_F): Likewise.
(VADDLVQ_P): Likewise.
(VCMPNEQ): Likewise.
(VSHLQ): Likewise.
(VABDQ): Likewise.
(VADDQ_N): Likewise.
(VADDVAQ): Likewise.
(VADDVQ_P): Likewise.
(VANDQ): Likewise.
(VBICQ): Likewise.
(VBRSRQ_N): Likewise.
(VCADDQ_ROT270): Likewise.
(VCADDQ_ROT90): Likewise.
(VCMPEQQ): Likewise.
(VCMPEQQ_N): Likewise.
(VCMPNEQ_N): Likewise.
(VEORQ): Likewise.
(VHADDQ): Likewise.
(VHADDQ_N): Likewise.
(VHSUBQ): Likewise.
(VHSUBQ_N): Likewise.
(VMAXQ): Likewise.
(VMAXVQ): Likewise.
(VMINQ): Likewise.
(VMINVQ): Likewise.
(VMLADAVQ): Likewise.
(VMULHQ): Likewise.
(VMULLBQ_INT): Likewise.
(VMULLTQ_INT): Likewise.
(VMULQ): Likewise.
(VMULQ_N): Likewise.
(VORNQ): Likewise.
(VORRQ): Likewise.
(VQADDQ): Likewise.
(VQADDQ_N): Likewise.
(VQRSHLQ): Likewise.
(VQRSHLQ_N): Likewise.
(VQSHLQ): Likewise.
(VQSHLQ_N): Likewise.
(VQSHLQ_R): Likewise.
(VQSUBQ): Likewise.
(VQSUBQ_N): Likewise.
(VRHADDQ): Likewise.
(VRMULHQ): Likewise.
(VRSHLQ): Likewise.
(VRSHLQ_N): Likewise.
(VRSHRQ_N): Likewise.
(VSHLQ_N): Likewise.
(VSHLQ_R): Likewise.
(VSUBQ): Likewise.
(VSUBQ_N): Likewise.
(VADDLVAQ): Likewise.
(VBICQ_N): Likewise.
(VMLALDAVQ): Likewise.
(VMLALDAVXQ): Likewise.
(VMOVNBQ): Likewise.
(VMOVNTQ): Likewise.
(VORRQ_N): Likewise.
(VQMOVNBQ): Likewise.
(VQMOVNTQ): Likewise.
(VSHLLBQ_N): Likewise.
(VSHLLTQ_N): Likewise.
(VRMLALDAVHQ): Likewise.
(VBICQ_M_N): Likewise.
(VCVTAQ_M): Likewise.
(VCVTQ_M_TO_F): Likewise.
(VQRSHRNBQ_N): Likewise.
(VABAVQ): Likewise.
(VSHLCQ): Likewise.
(VRMLALDAVHAQ): Likewise.
(VADDVAQ_P): Likewise.
(VCLZQ_M): Likewise.
(VCMPEQQ_M_N): Likewise.
(VCMPEQQ_M): Likewise.
(VCMPNEQ_M_N): Likewise.
(VCMPNEQ_M): Likewise.
(VDUPQ_M_N): Likewise.
(VMAXVQ_P): Likewise.
(VMINVQ_P): Likewise.
(VMLADAVAQ): Likewise.
(VMLADAVQ_P): Likewise.
(VMLAQ_N): Likewise.
(VMLASQ_N): Likewise.
(VMVNQ_M): Likewise.
(VPSELQ): Likewise.
(VQDMLAHQ_N): Likewise.
(VQRDMLAHQ_N): Likewise.
(VQRDMLASHQ_N): Likewise.
(VQRSHLQ_M_N): Likewise.
(VQSHLQ_M_R): Likewise.
(VREV64Q_M): Likewise.
(VRSHLQ_M_N): Likewise.
(VSHLQ_M_R): Likewise.
(VSLIQ_N): Likewise.
(VSRIQ_N): Likewise.
(VMLALDAVQ_P): Likewise.
(VQMOVNBQ_M): Likewise.
(VMOVLTQ_M): Likewise.
(VMOVNBQ_M): Likewise.
(VRSHRNTQ_N): Likewise.
(VORRQ_M_N): Likewise.
(VREV32Q_M): Likewise.
(VREV16Q_M): Likewise.
(VQRSHRNTQ_N): Likewise.
(VMOVNTQ_M): Likewise.
(VMOVLBQ_M): Likewise.
(VMLALDAVAQ): Likewise.
(VQSHRNBQ_N): Likewise.
(VSHRNBQ_N): Likewise.
(VRSHRNBQ_N): Likewise.
(VMLALDAVXQ_P): Likewise.
(VQMOVNTQ_M): Likewise.
(VMVNQ_M_N): Likewise.
(VQSHRNTQ_N): Likewise.
(VMLALDAVAXQ): Likewise.
(VSHRNTQ_N): Likewise.
(VCVTMQ_M): Likewise.
(VCVTNQ_M): Likewise.
(VCVTPQ_M): Likewise.
(VCVTQ_M_N_FROM_F): Likewise.
(VCVTQ_M_FROM_F): Likewise.
(VRMLALDAVHQ_P): Likewise.
(VADDLVAQ_P): Likewise.
(VABAVQ_P): Likewise.
(VSHLQ_M): Likewise.
(VSRIQ_M_N): Likewise.
(VSUBQ_M): Likewise.
(VCVTQ_M_N_TO_F): Likewise.
(VHSUBQ_M): Likewise.
(VSLIQ_M_N): Likewise.
(VRSHLQ_M): Likewise.
(VMINQ_M): Likewise.
(VMULLBQ_INT_M): Likewise.
(VMULHQ_M): Likewise.
(VMULQ_M): Likewise.
(VHSUBQ_M_N): Likewise.
(VHADDQ_M_N): Likewise.
(VORRQ_M): Likewise.
(VRMULHQ_M): Likewise.
(VQADDQ_M): Likewise.
(VRSHRQ_M_N): Likewise.
(VQSUBQ_M_N): Likewise.
(VADDQ_M): Likewise.
(VORNQ_M): Likewise.
(VRHADDQ_M): Likewise.
(VQSHLQ_M): Likewise.
(VANDQ_M): Likewise.
(VBICQ_M): Likewise.
(VSHLQ_M_N): Likewise.
(VCADDQ_ROT270_M): Likewise.
(VQRSHLQ_M): Likewise.
(VQADDQ_M_N): Likewise.
(VADDQ_M_N): Likewise.
(VMAXQ_M): Likewise.
(VQSUBQ_M): Likewise.
(VMLASQ_M_N): Likewise.
(VMLADAVAQ_P): Likewise.
(VBRSRQ_M_N): Likewise.
(VMULQ_M_N): Likewise.
(VCADDQ_ROT90_M): Likewise.
(VMULLTQ_INT_M): Likewise.
(VEORQ_M): Likewise.
(VSHRQ_M_N): Likewise.
(VSUBQ_M_N): Likewise.
(VHADDQ_M): Likewise.
(VABDQ_M): Likewise.
(VMLAQ_M_N): Likewise.
(VQSHLQ_M_N): Likewise.
(VMLALDAVAQ_P): Likewise.
(VMLALDAVAXQ_P): Likewise.
(VQRSHRNBQ_M_N): Likewise.
(VQRSHRNTQ_M_N): Likewise.
(VQSHRNBQ_M_N): Likewise.
(VQSHRNTQ_M_N): Likewise.
(VRSHRNBQ_M_N): Likewise.
(VRSHRNTQ_M_N): Likewise.
(VSHLLBQ_M_N): Likewise.
(VSHLLTQ_M_N): Likewise.
(VSHRNBQ_M_N): Likewise.
(VSHRNTQ_M_N): Likewise.
(VSTRWSBQ): Likewise.
(VSTRBSOQ): Likewise.
(VSTRBQ): Likewise.
(VLDRBGOQ): Likewise.
(VLDRBQ): Likewise.
(VLDRWGBQ): Likewise.
(VLD1Q): Likewise.
(VLDRHGOQ): Likewise.
(VLDRHGSOQ): Likewise.
(VLDRHQ): Likewise.
(VLDRWQ): Likewise.
(VLDRDGBQ): Likewise.
(VLDRDGOQ): Likewise.
(VLDRDGSOQ): Likewise.
(VLDRWGOQ): Likewise.
(VLDRWGSOQ): Likewise.
(VST1Q): Likewise.
(VSTRHSOQ): Likewise.
(VSTRHSSOQ): Likewise.
(VSTRHQ): Likewise.
(VSTRWQ): Likewise.
(VSTRDSBQ): Likewise.
(VSTRDSOQ): Likewise.
(VSTRDSSOQ): Likewise.
(VSTRWSOQ): Likewise.
(VSTRWSSOQ): Likewise.
(VSTRWSBWBQ): Likewise.
(VLDRWGBWBQ): Likewise.
(VSTRDSBWBQ): Likewise.
(VLDRDGBWBQ): Likewise.
(VADCIQ): Likewise.
(VADCIQ_M): Likewise.
(VSBCQ): Likewise.
(VSBCQ_M): Likewise.
(VSBCIQ): Likewise.
(VSBCIQ_M): Likewise.
(VADCQ): Likewise.
(VADCQ_M): Likewise.
(UQRSHLLQ): Likewise.
(SQRSHRLQ): Likewise.
(VSHLCQ_M): Likewise.
(define_c_enum "unspec"): Move MVE enumerator to unspecs.md from mve.md.
* config/arm/unspecs.md (define_c_enum "unspec"): Move MVE enumerator from
mve.md to unspecs.md.

dbgcnt: print list after compilation

gcc/ChangeLog:

* common.opt: Remove -fdbg-cnt-list from deferred options.
* dbgcnt.c (dbg_cnt_set_limit_by_index): Make a copy
to original_limits.
(dbg_cnt_list_all_counters): Print also current counter value
and print to stderr.
* opts-global.c (handle_common_deferred_options): Do not handle
-fdbg-cnt-list.
* opts.c (common_handle_option): Likewise.
* toplev.c (finalize): Handle it after compilation here.

dbgcnt: report upper limit when lower == upper

gcc/ChangeLog:

* dbgcnt.c (dbg_cnt): Report also upper limit.

configure: Fix in-tree building of GMP on BSD [PR97302]

ChangeLog:
PR target/97302
* configure.ac: Only set with_gmp to /usr/local
if not building in tree.
* configure: Regenerate.

[ftracer] Add caching of can_duplicate_bb_p

The fix "[omp, ftracer] Don't duplicate blocks in SIMT region" adds iteration
over insns in ignore_bb_p, which makes it more expensive.

Counteract this by piggybacking the computation of can_duplicate_bb_p onto
count_insns, which is called at the start of ftracer.

Bootstrapped and reg-tested on x86_64-linux.

gcc/ChangeLog:

2020-10-05  Tom de Vries  <tdevries@suse.de>

* tracer.c (count_insns): Rename to ...
(analyze_bb): ... this.
(cache_can_duplicate_bb_p, cached_can_duplicate_bb_p): New function.
(ignore_bb_p): Use cached_can_duplicate_bb_p.
(tail_duplicate): Call cache_can_duplicate_bb_p.

[ftracer] Factor out can_duplicate_bb_p

Factor out can_duplicate_bb_p out of ignore_bb_p.

Also factor out can_duplicate_insn_p and can_duplicate_bb_no_insn_iter_p to
expose the parts of can_duplicate_bb_p that are per-bb and per-insn.

Bootstrapped and reg-tested on x86_64-linux.

gcc/ChangeLog:

2020-10-05  Tom de Vries  <tdevries@suse.de>

* tracer.c (can_duplicate_insn_p, can_duplicate_bb_no_insn_iter_p)
(can_duplicate_bb_p): New function, factored out of ...
(ignore_bb_p): ... here.

libstdc++: Avoid CTAD for std::ranges::join_view [LWG 3474]

In commit ef275d1f2083f8a1fa1b59a3cd07fd3e8431023e I implemented the
wrong resolution of LWG 3474. This removes the deduction guide and
alters the views::join factory to create the right type explicitly.

libstdc++-v3/ChangeLog:

* include/std/ranges (join_view): Remove deduction guide.
(views::join): Add explicit template argument list to prevent
deducing the wrong type.
* testsuite/std/ranges/adaptors/join.cc: Move test for LWG 3474
here, from ...
* testsuite/std/ranges/adaptors/join_lwg3474.cc: Removed.

divmod: Match and expand DIVMOD even in some cases of constant divisor [PR97282]

As written in the comment, tree-ssa-math-opts.c wouldn't create a DIVMOD
ifn call for division + modulo by constant for the fear that during
expansion we could generate better code for those cases.
If the divisoris a power of two, that is certainly the case always,
but otherwise expand_divmod can punt in many cases, e.g. if the division
type's precision is above HOST_BITS_PER_WIDE_INT, we don't even call
choose_multiplier, because it works on HOST_WIDE_INTs (true, something
we should fix eventually now that we have wide_ints), or if pre/post shift
is larger than BITS_PER_WORD.

So, the following patch recognizes DIVMOD with constant last argument even
when it is unclear if expand_divmod will be able to optimize it, and then
during DIVMOD expansion if the divisor is constant attempts to expand it as
division + modulo and if they actually don't contain any libcalls or
division/modulo, they are kept as is, otherwise that sequence is thrown away
and divmod optab or libcall is used.

2020-10-06  Jakub Jelinek  <jakub@redhat.com>

PR rtl-optimization/97282
* tree-ssa-math-opts.c (divmod_candidate_p): Don't return false for
constant op2 if it is not a power of two and the type has precision
larger than HOST_BITS_PER_WIDE_INT or BITS_PER_WORD.
* internal-fn.c (contains_call_div_mod): New function.
(expand_DIVMOD): If last argument is a constant, try to expand it as
TRUNC_DIV_EXPR followed by TRUNC_MOD_EXPR, but if the sequence
contains any calls or {,U}{DIV,MOD} rtxes, throw it away and use
divmod optab or divmod libfunc.

* gcc.target/i386/pr97282.c: New test.

Fix off-by-one storage problem in irange_allocator.

gcc/ChangeLog:

* value-range.h (irange_allocator::allocate): Increase
newir storage by one.

openmp: Fix ICE in omp_discover_declare_target_tgt_fn_r

This ICEs because node->alias_target is (not yet) a FUNCTION_DECL, but
IDENTIFIER_NODE.

I guess we should retry the discovery before LTO streaming out, the reason
to do it this early is that it can affect the gimplification and omp lowering.

2020-10-06  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/97289
* omp-offload.c (omp_discover_declare_target_tgt_fn_r): Only follow
node->alias_target if it is a FUNCTION_DECL.

* c-c++-common/gomp/pr97289.c: New test.

arm: Add +nomve and +nomve.fp options to -mcpu=cortex-m55

This patch rearranges feature bits for MVE and FP to implement the
following flags for -mcpu=cortex-m55.

  - +nomve:    equivalent to armv8.1-m.main+fp.dp+dsp.
  - +nomve.fp: equivalent to armv8.1-m.main+mve+fp.dp (+dsp is implied by +mve).
  - +nofp:     equivalent to armv8.1-m.main+mve (+dsp is implied by +mve).
  - +nodsp:    equivalent to armv8.1-m.main+fp.dp.

Combinations of the above:

  - +nomve+nofp: equivalent to armv8.1-m.main+dsp.
  - +nodsp+nofp: equivalent to armv8.1-m.main.

Due to MVE and FP sharing vfp_base, some new syntax was required in the CPU
description to implement the concept of 'implied bits'. These are non-named
features added to the ISA late, depending on whether one or more features which
depend on them are present. This means vfp_base can be present when only one of
MVE and FP is removed, but absent when both are removed.

gcc/ChangeLog:

2020-07-31  Joe Ramsay  <joe.ramsay@arm.com>

* config/arm/arm-cpus.in:
(ALL_FPU_INTERNAL): Remove vfp_base.
(VFPv2): Remove vfp_base.
(MVE): Remove vfp_base.
(vfp_base): Redefine as implied bit dependent on MVE or FP
(cortex-m55): Add flags to disable MVE, MVE FP, FP and DSP extensions.
* config/arm/arm.c (arm_configure_build_target): Add implied bits to ISA.
* config/arm/parsecpu.awk:
(gen_isa): Print implied bits and their dependencies to ISA header.
(gen_data): Add parsing for implied feature bits.

gcc/testsuite/ChangeLog:

* gcc.target/arm/cortex-m55-nodsp-flag-hard.c: New test.
* gcc.target/arm/cortex-m55-nodsp-flag-softfp.c: New test.
* gcc.target/arm/cortex-m55-nodsp-nofp-flag-softfp.c: New test.
* gcc.target/arm/cortex-m55-nofp-flag-hard.c: New test.
* gcc.target/arm/cortex-m55-nofp-flag-softfp.c: New test.
* gcc.target/arm/cortex-m55-nofp-nomve-flag-softfp.c: New test.
* gcc.target/arm/cortex-m55-nomve-flag-hard.c: New test.
* gcc.target/arm/cortex-m55-nomve-flag-softfp.c: New test.
* gcc.target/arm/cortex-m55-nomve.fp-flag-hard.c: New test.
* gcc.target/arm/cortex-m55-nomve.fp-flag-softfp.c: New test.
* gcc.target/arm/multilib.exp: Add tests for -mcpu=cortex-m55.

IBM Z: Doc: Add z15/arch13 to the list of -march/-mtune options

gcc/ChangeLog:

* doc/invoke.texi: Add z15/arch13 to the list of documented
-march/-mtune options.

gofrontend: correct file reading logic in Stream_from_file

The implementation of Stream_from_file mishandled several cases:

  * It reversed the check for whether bytes were already available in
    the peek buffer.

  * It considered positive return values from lseek to be an error, when
    only a -1 return value indicates an error.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/259437

Daily bump.

libstdc++: Reduce uses of std::numeric_limits

This avoids unnecessary instantiations of std::numeric_limits or
inclusion of <limits> when a more lightweight alternative would work.
Some uses can be replaced with __gnu_cxx::__int_traits and some can just
use size_t(-1) directly where SIZE_MAX is needed.

libstdc++-v3/ChangeLog:

* include/bits/regex.h: Use __int_traits<int> instead of
std::numeric_limits<int>.
* include/bits/uniform_int_dist.h: Use __int_traits<T>::__max
instead of std::numeric_limits<T>::max().
* include/bits/hashtable_policy.h: Use size_t(-1) instead of
std::numeric_limits<size_t>::max().
* include/std/regex: Include <ext/numeric_traits.h>.
* include/std/string_view: Use typedef for __int_traits<int>.
* src/c++11/hashtable_c++0x.cc: Use size_t(-1) instead of
std::numeric_limits<size_t>::max().
* testsuite/std/ranges/iota/96042.cc: Include <limits>.
* testsuite/std/ranges/iota/difference_type.cc: Likewise.
* testsuite/std/ranges/subrange/96042.cc: Likewise.

c++: Fix typo in NON_UNION_CLASS_TYPE_P.

gcc/cp/ChangeLog:

* cp-tree.h (NON_UNION_CLASS_TYPE_P): Fix typo in a comment.

libstdc++: Minor header cleanup in <numeric>

When adding new features to <numeric> I included the required headers
adjacent to the new code. This cleans it up by moving all the includes
to the start of the file.

libstdc++-v3/ChangeLog:

* include/std/numeric: Move all #include directives to the top
of the header.
* testsuite/26_numerics/gcd/gcd_neg.cc: Adjust dg-error line
numbers.
* testsuite/26_numerics/lcm/lcm_neg.cc: Likewise.

Cleanup legacy_union and legacy intersect in value_range.

These are cleanups so that multi-range union/intersect doesn't
have to deal with legacy code.  Instead, these should be done in
legacy mode.

gcc/ChangeLog:

* value-range.cc (irange::legacy_intersect): Only handle
legacy ranges.
(irange::legacy_union): Same.
(irange::union_): When unioning legacy with non-legacy,
first convert to legacy and do everything in legacy mode.
(irange::intersect): Same, but for intersect.
* range-op.cc (range_tests): Adjust for above changes.

Import various range-op fixes from ranger branch.

This patch imports three fixes from the ranger branch:

1. Fold division by zero into varying instead of undefined.
This provides compatibility with existing stuff on trunk.

2. Solver changes for lshift.
This should not affect anything on trunk, as it only involves
the GORI solver which is yet to be contributed.

3. Preserve existing behavior for ABS([-MIN,-MIN]).
This is actually unrepresentable, but trunk has traditionally
treated this as [-MIN,-MIN] so this patch just syncs range-ops
with the rest of trunk.

gcc/ChangeLog:

* range-op.cc (operator_div::wi_fold): Return varying for
division by zero.
(class operator_rshift): Move class up.
(operator_abs::wi_fold): Return [-MIN,-MIN] for ABS([-MIN,-MIN]).
(operator_tests): Adjust tests.

support TARGET_MEM_REF in C/C++ error pretty-printing [PR97197]

> See my comment above for Martins attempts to improve things.  I don't
> really want to try decide what to do with those late diagnostic IL
> printing but my commit was blamed for showing target-mem-ref unsupported.
>
> I don't have much time to spend to think what to best print and what not,
> but yes, printing only the MEM_REF part is certainly imprecise.

Here is an updated version of the patch that prints TARGET_MEM_REF the way
it should be printed - as C representation of what it actually means.
Of course it would be better to have the original expressions, but with the
late diagnostics we no longer have them.

2020-10-05  Richard Biener  <rguenther@suse.de>
    Jakub Jelinek  <jakub@redhat.com>

PR c++/97197
gcc/cp/
* error.c (dump_expr): Handle TARGET_MEM_REF.
gcc/c-family/
* c-pretty-print.c: Include langhooks.h.
(c_pretty_printer::postfix_expression): Handle TARGET_MEM_REF as
expression.
(c_pretty_printer::expression): Handle TARGET_MEM_REF as
unary_expression.
(c_pretty_printer::unary_expression): Handle TARGET_MEM_REF.

libstdc++: Make allocators throw bad_array_new_length on overflow [LWG 3190]

std::allocator and std::pmr::polymorphic_allocator should throw
std::bad_array_new_length from their allocate member functions if the
number of bytes required cannot be represented in std::size_t.

libstdc++-v3/ChangeLog:

* config/abi/pre/gnu.ver: Add new symbol.
* include/bits/functexcept.h (__throw_bad_array_new_length):
Declare new function.
* include/ext/malloc_allocator.h (malloc_allocator::allocate):
Throw bad_array_new_length for impossible sizes (LWG 3190).
* include/ext/new_allocator.h (new_allocator::allocate):
Likewise.
* include/std/memory_resource (polymorphic_allocator::allocate)
(polymorphic_allocator::allocate_object): Use new function,
__throw_bad_array_new_length.
* src/c++11/functexcept.cc (__throw_bad_array_new_length):
Define.
* testsuite/20_util/allocator/lwg3190.cc: New test.

[omp, ftracer] Ignore IFN_GOMP_SIMT_XCHG_* in ignore_bb_p

As IFN_GOMP_SIMT_XCHG_* are part of the group marked by
IFN_GOMP_SIMT_ENTER_ALLOC/IFN_GOMP_SIMT_EXIT, handle them conservatively
in ignore_bb_p.

Build on x86_64-linux with nvptx accelerator, tested with libgomp.

gcc/ChangeLog:

2020-10-05  Tom de Vries  <tdevries@suse.de>

* tracer.c (ignore_bb_p): Ignore GOMP_SIMT_XCHG_*.

c++: Make spell corrections consistent

My change to namespace-scope spell corrections ignored the issue that
different targets might have different builtins, and therefore perturb
iteration order.  This fixes it by using an intermediate array of
identifier, which we sort before considering.

gcc/cp/
* name-lookup.c (maybe_add_fuzzy_decl): New.
(maybe_add_fuzzy_binding): New.
(consider_binding_level): Use intermediate sortable vector for
namespace bindings.
gcc/testsuite/
* c-c++-common/spellcheck-reserved.c: Restore diagnostic.

arm: Add missing part number for Neoverse V1

This patch adds vendor and part numbers which were missing from the
initial entry for Neoverse V1 in AArch32 GCC.

gcc/ChangeLog:

* config/arm/arm-cpus.in (neoverse-v1): Add missing vendor and
part numbers.

[omp, ftracer] Remove incorrect suggestion in ignore_bb_p

In commit ab3f4b27abe "[omp, ftracer] Don't duplicate blocks in SIMT region" I
added a comment in ignore_bb_p suggesting a reordering of SIMT_VOTE_ANY and
SIMT_EXIT, which is not possible since VOTE_ANY may have data dependencies to
storage that is deallocated by SIMT_EXIT.

I've now opened a PR (PR97291) to describe the problem the reordering was
intended to fix.

Remove the incorrect suggestion.

gcc/ChangeLog:

2020-10-05  Tom de Vries  <tdevries@suse.de>

* tracer.c (ignore_bb_p): Remove incorrect suggestion.

libstdc++: Use correct duration for atomic_futex wait on custom clock [PR 91486]

As Jonathan Wakely pointed out[1], my change in commit
f9ddb696a289cc48d24d3d23c0b324cb88de9573 should have been rounding to
the target clock duration type rather than the input clock duration type
in __atomic_futex_unsigned::_M_load_when_equal_until just as (e.g.)
condition_variable does.

As well as fixing this, let's create a rather contrived test that fails
with the previous code, but unfortunately only when run on a machine
with an uptime of over 208.5 days, and even then not always.

[1] https://gcc.gnu.org/pipermail/libstdc++/2020-September/051004.html

libstdc++-v3/ChangeLog:

PR libstdc++/91486
* include/bits/atomic_futex.h:
(__atomic_futex_unsigned::_M_load_when_equal_until): Use target
clock duration type when rounding.
* testsuite/30_threads/async/async.cc (test_pr91486_wait_for):
Rename from test_pr91486.
(float_steady_clock): New class for test.
(test_pr91486_wait_until): New test.

libstdc++: Test C++11 implementation of std::chrono::__detail::ceil

Commit 53ad6b1979f4bd7121e977c4a44151b14d8a0147 split the implementation
of std::chrono::__detail::ceil so that when compiling for C++17 and
later std::chrono::ceil is used but when compiling for earlier versions
a separate implementation is used to comply with C++11's limited
constexpr rules. Let's run the equivalent of the existing
std::chrono::ceil test cases on std::chrono::__detail::ceil too to make
sure that it doesn't get broken.

libstdc++-v3/ChangeLog:

* testsuite/20_util/duration_cast/rounding_c++11.cc: Copy
rounding.cc and alter to support compilation for C++11 and to
test std::chrono::__detail::ceil.

libstdc++: Add missing bugzilla PR numbers to ChangeLog

We missed these out of the git commit messages.

options: Save and restore opts_set for Optimization and Target options fallout

> This breaks ia64:
>
> In file included from ./tm.h:23,
>                  from ../../gcc/gencheck.c:23:
> ./options.h:7816:40: error: ISO C++ forbids zero-size array 'explicit_mask' [-Werror=pedantic]
>  7816 |   unsigned HOST_WIDE_INT explicit_mask[0];
>       |                                        ^
> ./options.h:7816:26: error: zero-size array member 'cl_target_option::explicit_mask' not at end of 'struct cl_target_option' [-Werror=pedantic]
>  7816 |   unsigned HOST_WIDE_INT explicit_mask[0];
>       |                          ^~~~~~~~~~~~~
> ./options.h:7812:16: note: in the definition of 'struct cl_target_option'
>  7812 | struct GTY(()) cl_target_option
>       |                ^~~~~~~~~~~~~~~~

Oops, sorry.

The following patch should fix that and should also fix streaming of the
new explicit_mask_* members.

2020-10-05  Jakub Jelinek  <jakub@redhat.com>

* opth-gen.awk: Don't emit explicit_mask array if n_target_explicit
is equal to n_target_explicit_mask.
* optc-save-gen.awk: Compute has_target_explicit_mask and if false,
don't emit code iterating over explicit_mask array elements.  Stream
also explicit_mask_* target members.

store-merging: Fix up -Wnarrowing warning

I've noticed a -Wnarrowing warning on gimple-ssa-store-merging.c, this
change fixes that up.

2020-10-05  Jakub Jelinek  <jakub@redhat.com>

* gimple-ssa-store-merging.c
(imm_store_chain_info::output_merged_store): Use ~0U instead of ~0 in
unsigned int array initializer.

[omp, ftracer] Don't duplicate blocks in SIMT region

When running the libgomp testsuite on x86_64-linux with nvptx accelerator on
the test-case included in this patch, we run into:
...
FAIL: libgomp.fortran/pr95654.f90 -O3 -fomit-frame-pointer -funroll-loops \
  -fpeel-loops -ftracer -finline-functions  execution test
...

The test-case is a minimal version of this FAIL:
...
FAIL: libgomp.fortran/pr66199-5.f90 -O3 -fomit-frame-pointer -funroll-loops \
  -fpeel-loops -ftracer -finline-functions  execution test
...
but that one has stopped failing at commit c2ebf4f10de "openmp: Add support
for non-rect simd and improve collapsed simd support".

The problem is that ftracer duplicates a block containing GOMP_SIMT_VOTE_ANY.

That is, before ftracer we have (dropping the GOMP_SIMT_ prefix):
...
bb4(ENTER_ALLOC)
*----------+
|           \
|            \
|             v
|             *
v             bb8
*<------------*
bb5(VOTE_ANY)
*-------------+
|             |
|             |
|             |
|             |
|             v
|             *
v             bb7(XCHG_IDX)
*<------------*
bb6(EXIT)
...

The XCHG_IDX internal-fn does inter-SIMT-lane communication, which for nvptx
maps onto shfl, an operator which has the requirement that the warp executing
the operator is convergent.  The warp diverges at bb4, and
reconverges at bb5, and does not diverge by going to bb7, so the shfl is
indeed executed by a convergent warp.

After ftracer, we have:
...
bb4(ENTER_ALLOC)
*----------+
|           \
|            \
|             \
|              \
v               v
*               *
bb5(VOTE_ANY)   bb8(VOTE_ANY)
*               *
|\             /|
| \  +--------+ |
|  \/           |
|  /\           |
| /  +----------v
|/              *
v               bb7(XCHG_IDX)
*<--------------*
bb6(EXIT)
...

The warp diverges again at bb5, but does not reconverge again before bb6, so
the shfl is executed by a divergent warp, which causes the FAIL.

Fix this by making ftracer ignore blocks containing ENTER_ALLOC, VOTE_ANY and
EXIT, effectively treating the SIMT region conservatively.

An argument can be made that the test needs to be added in a more
generic place, like gimple_can_duplicate_bb_p or some such, and that ftracer
then needs to use the generic test.  But that's a discussion with a much
broader scope, so I'm leaving that for another patch.

Bootstrapped and reg-tested on x86_64-linux.

Build on x86_64-linux with nvptx accelerator, tested with libgomp.

gcc/ChangeLog:

PR fortran/95654
* tracer.c (ignore_bb_p): Ignore GOMP_SIMT_ENTER_ALLOC,
GOMP_SIMT_VOTE_ANY and GOMP_SIMT_EXIT.

libgomp/ChangeLog:

2020-10-05  Tom de Vries  <tdevries@suse.de>

PR fortran/95654
* testsuite/libgomp.fortran/pr95654.f90: New test.

Daily bump.

PR fortran/97272 - Wrong answer from MAXLOC with character arg

The optional KIND argument to the MINLOC/MAXLOC intrinsic must not be
passed to the library function, as the kind conversion of the result
is treated explicitly elsewhere.

gcc/fortran/ChangeLog:

PR fortran/97272
* trans-intrinsic.c (strip_kind_from_actual): Helper function for
removal of KIND argument.
(gfc_conv_intrinsic_minmaxloc): Ignore KIND argument here, as it
is treated elsewhere.

gcc/testsuite/ChangeLog:

PR fortran/97272
* gfortran.dg/pr97272.f90: New test.

Daily bump.

aix: apply aix_malloc more narrowly.

In recent Technology Levels of AIX 7.2, new "#ifdef __cplusplus" have been
added. Thus, the aix_malloc fix was applied in wrong locations. This patch
increases the context to avoid this.

fixincludes/ChangeLog:

2020-10-03  Clément Chigot  <clement.chigot@atos.net>

* inclhack.def (aix_malloc): Add more context to select.
* fixincl.x: Regenerate.
* tests/base/malloc.h: Update expected results.

options: Fix up opts_set saving/restoring for underlying vars of Mask/InverseMask options

Seems I've missed that set_option has special treatment for
CLVC_BIT_CLEAR/CLVC_BIT_SET.
Which means I'll need to change the generic handling, so that for
global_options_set elements mentioned in CLVC_BIT_* options are treated
differently, instead of using the accumulated bitmasks they'll need to use
their specific bitmask variables during the option saving/restoring.

Here is a patch that implements that.

2020-10-03  Jakub Jelinek  <jakub@redhat.com>

* opth-gen.awk: For variables referenced in Mask and InverseMask,
don't use the explicit_mask bitmask array, but add separate
explicit_mask_* members with the same types as the variables.
* optc-save-gen.awk: Save, restore, compare and hash the separate
explicit_mask_* members.

Add gcc.dg/tree-ssa/modref-3.c testcase

* gcc.dg/tree-ssa/modref-3.c: New test.

Track access ranges in ipa-modref

this patch implements tracking of access ranges.  This is only applied when
base pointer is an arugment. Incrementally i will extend it to also track
TBAA basetype so we can disambiguate ranges for accesses to same basetype
(which makes is quite bit more effective). For this reason i track the access
offset separately from parameter offset (the second track combined adjustments
to the parameter). This is I think last feature I would like to add to the
memory access summary this stage1.

Further work will be needed to opitmize the summary and merge adjacent
range/make collapsing more intelingent (so we do not lose track that often),
but I wanted to keep basic patch simple.

According to the cc1plus stats:

Alias oracle query stats:
  refs_may_alias_p: 64108082 disambiguations, 74386675 queries
  ref_maybe_used_by_call_p: 142319 disambiguations, 65004781 queries
  call_may_clobber_ref_p: 23587 disambiguations, 29420 queries
  nonoverlapping_component_refs_p: 0 disambiguations, 38117 queries
  nonoverlapping_refs_since_match_p: 19489 disambiguations, 55748 must overlaps, 76044 queries
  aliasing_component_refs_p: 54763 disambiguations, 755876 queries
  TBAA oracle: 24184658 disambiguations 56823187 queries
               16260329 are in alias set 0
               10617146 queries asked about the same object
               125 queries asked about the same alias set
               0 access volatile
               3960555 are dependent in the DAG
               1800374 are aritificially in conflict with void *

Modref stats:
  modref use: 10656 disambiguations, 47037 queries
  modref clobber: 1473322 disambiguations, 1961464 queries
  5027242 tbaa queries (2.563005 per modref query)
  649087 base compares (0.330920 per modref query)

PTA query stats:
  pt_solution_includes: 977385 disambiguations, 13609749 queries
  pt_solutions_intersect: 1032703 disambiguations, 13187507 queries

Which should still compare with
https://gcc.gnu.org/pipermail/gcc-patches/2020-September/554930.html
there is about 2% more load disambiguations and 3.6% more store that is not
great, but the TBAA part helps noticeably more and also this should help
with -fno-strict-aliasing.

I plan to work on improving param tracking too.

Bootstrapped/regtested x86_64-linux with the other changes, OK?

2020-10-02  Jan Hubicka  <hubicka@ucw.cz>

* ipa-modref-tree.c (test_insert_search_collapse): Update andling
of accesses.
(test_merge): Likewise.
* ipa-modref-tree.h (struct modref_access_node): Add offset, size,
max_size, parm_offset and parm_offset_known.
(modref_access_node::useful_p): Constify.
(modref_access_node::range_info_useful_p): New predicate.
(modref_access_node::operator==): New.
(struct modref_parm_map): New structure.
(modref_tree::merge): Update for racking parameters)
* ipa-modref.c (dump_access): Dump new fields.
(get_access): Fill in new fields.
(merge_call_side_effects): Update handling of parm map.
(write_modref_records): Stream new fields.
(read_modref_records): Stream new fields.
(compute_parm_map): Update for new parm map.
(ipa_merge_modref_summary_after_inlining): Update.
(modref_propagate_in_scc): Update.
* tree-ssa-alias.c (modref_may_conflict): Handle known ranges.

doc: Replace roudnevenl with roundevenl

PR other/97280
* doc/extend.texi: Replace roudnevenl with roundevenl

Daily bump.

rs6000: clean up headers in rs6000.c and rs6000-call.c

When Andrew Macleod investigated the recent rs6000 bootstrap failure,
he suggested a clean up of the headers in rs6000.c and rs6000-call.c.
It now is recommended to include ssa.h instead of the individual headers.
This also ensures that value-range.h is included and in the correct order
so that the tree-ssa-propagate.h inclusion of value-query.h and its
dependencies are satisfied.

Bootstrapped on powerpc-ibm-aix7.2.0.0 and powerpc64le-linux.

gcc/ChangeLog:

2020-10-02  David Edelsohn  <dje.gcc@gmail.com>
    Andrew MacLeod  <amacleod@redhat.com>

* config/rs6000/rs6000.c: Include ssa.h. Reorder some headers.
* config/rs6000/rs6000-call.c: Same.

c++: Fix printing of C++20 template parameter object [PR97014]

No one is interested in the mangled name of the C++20 template parameter
object for a class NTTP.  So instead of printing

  required for the satisfaction of ‘positive<T::ratio>’ [with T = X<::_ZTAXtl5ratioLin1ELi2EEE>]

let's print

  required for the satisfaction of ‘positive<T::ratio>’ [with T = X<{-1, 2}>]

I don't think adding a test is necessary for this.

gcc/cp/ChangeLog:

PR c++/97014
* cxx-pretty-print.c (pp_cxx_template_argument_list): If the
argument is template_parm_object_p, print its DECL_INITIAL.

libstdc++: Change test to work without 64-bit atomics

This fixes a linker error for older ARM cores without 64-bit atomics.

I think the { dg-add-options libatomic } is no longer needed, but it's
harmless to keep it there.

libstdc++-v3/ChangeLog:

* testsuite/29_atomics/atomic_float/value_init.cc: Use float
instead of double so that __atomic_load_8 isn't needed.

libstdc++: Fix testcase by using terminate handler

This test was supposed to verify that when __libc_single_threaded is
available we successfully detect recursive static initialization even
when linked to libpthread. But I forgot to that when recursive init is
detected, we terminate, and so the test fails.

This adds a terminate handler that exits cleanly, so the test passes
when recursive init is detected.

libstdc++-v3/ChangeLog:

* testsuite/18_support/96817.cc: Use terminate handler that
calls _Exit(0).

c++: Kill DECL_ANTICIPATED

Here's the patch to remove DECL_ANTICIPATED, and with it hiddenness is
managed entirely in the symbol table.  Sadly I couldn't get rid of the
actual field without more investigation -- it's repurposed for
OMP_PRIVATIZED_MEMBER.  It looks like a the VAR-related flags in
lang_decl_base are not completely orthogonal, so perhaps some can be
turned into an enumeration or something.  But that's more than I want
to do right now.

DECL_FRIEND_P Is still slightly suspect as it appears to mean more
than just in-class definition.  However, I'm leaving that for now.

gcc/cp/
* cp-tree.h (lang_decl_base): anticipated_p is not used for
anticipatedness.
(DECL_ANTICIPATED): Delete.
* decl.c (duplicate_decls): Delete DECL_ANTICIPATED_management,
use was_hidden.
(cxx_builtin_function): Drop DECL_ANTICIPATED setting.
(xref_tag_1): Drop DECL_ANTICIPATED assert.
* name-lookup.c (name_lookup::adl_class_only): Drop
DECL_ANTICIPATED check.
(name_lookup::search_adl): Always dedup.
(anticipated_builtin_p): Reimplement.
(do_pushdecl): Drop DECL_ANTICIPATED asserts & update.
(lookup_elaborated_type_1): Drop DECL_ANTICIPATED update.
(do_pushtag): Drop DECL_ANTICIPATED setting.
* pt.c (push_template_decl): Likewise.
(tsubst_friend_class): Likewise.
libcc1/
* libcp1plugin.cc (libcp1plugin.cc): Drop DECL_ANTICIPATED test.