This fixes a test failure with -std=gnu++98 -D_GLIBCXX_DEBUG:

FAIL: 21_strings/basic_string/modifiers/insert/char/1.cc (test for excess errors)

	* include/debug/string (__gnu_debug::basic_string::insert): Fix for
	C++98 where the member function of the base class returns void.

In implementing Jason's suggested direction for 93933, the compiler
exploded in a surprising way.  Turns out an assert had been passing
NULLS to comptypes, and therefore not checking what it intended.

Further comptypes, could silently accept such nulls under most
circumstances.

	* class.c (adjust_clone_args): Correct arg-checking assert.
	* typeck.c (comptypes): Assert not nulls.

Comparing value-initialized forward_iterator_wrapper<T> objects fails an
assertion, but should be valid in C++14 and later.

	* testsuite/util/testsuite_iterators.h (forward_iterator_wrapper): Add
	equality comparisons that support value-initialized iterators.

Value-numbering failed to handle __builtin_{memcpy,memset,...}_chk
variants when removing abstraction and also failed to use the
value-numbering lattice when requiring the length argument of the
call to be constant.

2020-02-27  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/93508
	* tree-ssa-sccvn.c (vn_reference_lookup_3): Handle _CHK like
	non-_CHK variants.  Valueize their length arguments.

	* gcc.dg/tree-ssa/ssa-fre-85.c: New testcase.

When possibly expanding a hash-map avoid keeping a reference to an
entry.

2020-02-27  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/93953
	* tree-vect-slp.c (slp_copy_subtree): Avoid keeping a reference
	to the hash-map entry.

	* gcc.dg/pr93953.c: New testcase.

BImode usually only requires one bit, but instructions that write to VCC also
clobber the reset of the DImode register pair, so gcn_class_max_nregs reports
that two registers are needed for BImode.  Paradoxically, accessing VCC via
SImode is therefore uses fewer registers than accessing via BImode.

The LRA checking code takes this into account, but the DF liveness data also
looks at the subreg, so it says (subreg:SI (reg:BI VCC) 0) only makes the low
part live.  Both are "correct", but they disagree, which causes an ICE.

This doesn't happen when writing conditions to VCC; it happens when accessing
VCC_LO via a regular move to a regular SImode register.

If we transform the subregs so that BImode is always the outer mode then it
basically means the same thing, except that now both LRA and DF calculate nregs
the same, and ICE goes away.

As soon as LRA is done the subregs all evaporate anyway.

2020-02-27  Andrew Stubbs  <ams@codesourcery.com>

	gcc/
	* config/gcn/gcn.md (mov<mode>): Add transformations for BI subregs.

Although most of the old-style "concept checks" are only really usable
with C++98 because they enforce the wrong things, this is a simple
change that makes them a bit more useful for C++14 and up.

	* include/bits/boost_concept_check.h (__function_requires): Add
	_GLIBCXX14_CONSTEXPR.
	* testsuite/25_algorithms/min/concept_checks.cc: New test.

This applies file mapping when emitting the directory table
directly instead of using the assemblers .file directive where
we already correctly apply the map.  Notably the non-assembler
path is used for the early debug emission for LTO.

2020-02-27  Mark Williams  <mwilliams@fb.com>

	* dwarf2out.c (file_name_acquire): Call remap_debug_filename.
	* lto-opts.c (lto_write_options): Drop -fdebug-prefix-map,
	-ffile-prefix-map and -fmacro-prefix-map.
	* lto-streamer-out.c: Include file-prefix-map.h.
	(lto_output_location): Remap the file part of locations.

The following testcase is rejected, while it was accepted in 3.4 and earlier
(before tree-ssa merge).
The problem is that we decide to promote the const variable to TREE_STATIC,
but TREE_STATIC DECL_REGISTER VAR_DECLs may only be the global register vars
and so assemble_variable/make_decl_rtl diagnoses it.

Either we do what the following patch does, where we could consider
register as a hint the user doesn't want such optimization, because if
something is forced static, it is not "register" anymore and register static
is not valid in C either, or we could clear DECL_REGISTER instead, but would
still need to punt at least on DECL_HARD_REGISTER cases.

2020-02-27  Jakub Jelinek  <jakub@redhat.com>

	PR c/93949
	* gimplify.c (gimplify_init_constructor): Don't promote readonly
	DECL_REGISTER variables to TREE_STATIC.

	* gcc.c-torture/compile/pr93949.c: New test.

The following is the last spot in vn_reference_lookup_3 that didn't allow
non-byte aligned offsets or sizes.  To be precise, it did allow size that
wasn't multiple of byte size and that caused a wrong-code issue on
big-endian, as the pr93945.c testcase shows, so for GCC 9 we should add
&& multiple_p (ref->size, BITS_PER_UNIT) check instead.
For the memset with SSA_NAME middle-argument, it still requires byte-aligned
offset, as we'd otherwise need to rotate the value at runtime.

2020-02-27  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/93582
	PR tree-optimization/93945
	* tree-ssa-sccvn.c (vn_reference_lookup_3): Handle memset with
	non-zero INTEGER_CST second argument and ref->offset or ref->size
	not a multiple of BITS_PER_UNIT.

	* gcc.dg/tree-ssa/pr93582-9.c: New test.
	* gcc.c-torture/execute/pr93945.c: New test.

	* doc/install.texi (Binaries): Update description of BullFreeware.

When doing the 8.4-rc1, I've noticed (probably also because of the dying
disk on sourceware) that git clone is extremely slow, and furthermore when
all of us have some local snapshots, it is a waste of resources to download
everything again.  Especially for the -f runs when we'll need to wait until
git tag -s asks us for a gpg password interactively.

The following patch adds an option through which one can point the script
at a local gcc .git directory from which it can --dissociate --reference ...
during cloning to speed it up.

2020-02-27  Jakub Jelinek  <jakub@redhat.com>

	* gcc_release: Add support for -b local-git-repo argument.

2020-02-26  Sandra Loosemore  <sandra@codesourcery.com>

	PR c++/90467

	gcc/
	* doc/invoke.texi (Option Summary): Re-alphabetize warnings in
	C++ Language Options, Warning Options, and Static Analyzer
	Options lists.  Document negative form of options enabled by
	default.  Move some things around to more accurately sort
	warnings by category.
	(C++ Dialect Options, Warning Options, Static Analyzer
	Options): Document negative form of options when enabled by
	default.  Move some things around to more accurately sort
	warnings by category.  Add some missing index entries.
	Light copy-editing.

r7-2111 introduced maybe_constant_value in cp_fully_fold.
maybe_constant_value uses cxx_eval_outermost_constant_expr, which
can clear TREE_CONSTANT:
6510   else if (non_constant_p && TREE_CONSTANT (r))
[...]
6529       TREE_CONSTANT (r) = false;

In this test the array size is '(long int) "h"'.  This used to be
TREE_CONSTANT but given the change above, the flag will be cleared
when we cp_fully_fold the array size in compute_array_index_type_loc.
That means we don't emit an error in the
10391   else if (TREE_CONSTANT (size)
block in the same function, and we go on.  Then we compute ITYPE
using cp_build_binary_op and use maybe_constant_value on it and
suddenly we have something that's TREE_CONSTANT again.  And then we
crash in reshape_init_array_1 in tree_to_uhwi, because what we have
doesn't fit in an unsigned HWI.

icc accepts this code, but since we used to reject it, I see no desire
to make this work, so don't use the folded result when we've lost
the TREE_CONSTANT flag while evaluating the size.

2020-02-26  Marek Polacek  <polacek@redhat.com>

	PR c++/93789 - ICE with invalid array bounds.
	* decl.c (compute_array_index_type_loc): Don't use the folded
	size when folding cleared TREE_CONSTANT.

	* g++.dg/ext/vla22.C: New test.

PR analyzer/93950 reports an ICE when pruning the path of a
-Wanalyzer-null-dereference diagnostic.

The root cause is a bug in the state-tracking code, in which the
variable of interest is tracked from the callee to a "nullptr" param
at the caller, whereupon we have an INTEGER_CST "variable", and
the attempt to look up its lvalue fails.

This code could use a rewrite; in the meantime this patch extends
the bulletproofing from g:8525d1f5
for PR analyzer/93544 to all of the various places where var can
be updated, fixing the ICE.

gcc/analyzer/ChangeLog:
	PR analyzer/93950
	* diagnostic-manager.cc
	(diagnostic_manager::prune_for_sm_diagnostic): Assert that var is
	either NULL or not a constant.  When updating var, bulletproof
	against constant values.

gcc/testsuite/ChangeLog:
	PR analyzer/93950
	* g++.dg/analyzer/pr93950.C: New test.

PR analyzer/93947 reports an ICE at -O1 when attempting to analyze a
call that has been optimized away as unreachable.

The root cause is a NULL dereference due to the fndecl having a NULL
cgraph_node: the cgraph_node was created by
pass_build_cgraph_edges::execute, but was later removed by
symbol_table::remove_unreachable_nodes before the analyzer pass.

This patch fixes it by checking for NULL before handling the
cgraph_node.

The reproducer demonstrates a weakness in the analyzer's constraint
handling, where region_model::apply_constraints_for_gswitch fails
to spot when the cases fully cover the data type, and thus make the
default impossible.  For now this is xfail-ed in the testcase.

gcc/analyzer/ChangeLog:
	PR analyzer/93947
	* region-model.cc (region_model::get_fndecl_for_call): Gracefully
	fail for fn_decls that don't have a cgraph_node.

gcc/testsuite/ChangeLog:
	PR analyzer/93947
	* gcc.dg/analyzer/torture/pr93947.c: New test.

PR target/91276 - Doc typos in __builtin_crypto_vpmsum*

gcc/ChangeLog:

2020-02-26  Carl Love  <cel@us.ibm.com>

	PR target/91276
	* doc/extend.texi (PowerPC AltiVec Built-in Functions available on
	ISA 3.0): The builtin-function name __builtin_crypto_vpmsumb is only
	for the vector unsigned short arguments.  It is also listed as the
	name of the built-in for arguments vector unsigned short,
	vector unsigned int and vector unsigned long long built-ins.  The
	name of the builtins for these arguments should be:
	__builtin_crypto_vpmsumh, __builtin_crypto_vpmsumw and
	__builtin_crypto_vpmsumd respectively.

Signed-off-by: Carl Love <carll@us.ibm.com>

In n4849 and preceding versions, [class.copy.elision] (1.3)
appears to confer additional permissions on coroutines to elide
parameter copies.

After considerable discussion on this topic by email and during
the February 2020 WG21 meeting, it has been determined that there
are no additional permissions applicable to coroutine parameter
copy elision.

The content of that clause in the standard is expected to be amended
eventually to clarify this.  Other than this, the handling of
parameter lifetimes is expected to be as per n4849:

 * A copy is made before the promise is constructed
 * If the promise CTOR uses the parms, then it should use the copy
   where appropriate.
 * The param copy lifetimes end after the promise is destroyed
   (during the coroutine frame destruction).
 * Otherwise, C++20 copy elision rules apply.

(as an aside) In practice, we expect that copy elision can only occur
when the coroutine body is fully inlined, possibly in conjunction with
heap allocation elision.

The patch:
 * Reorders the copying process to precede the promise CTOR and
    ensures the correct use.
 * Copies all params into the frame regardless of whether the coro
   body uses them (this is a bit unfortunate, and we should figure
   out an amendment for C++23).

gcc/cp/ChangeLog:

2020-02-26  Iain Sandoe  <iain@sandoe.co.uk>

	* class.c (classtype_has_non_deleted_copy_ctor): New.
	* coroutines.cc (struct param_info): Keep track of params
	that are references, and cache the original type and whether
	the DTOR is trivial.
	(build_actor_fn): Handle param copies always, and adjust the
	handling for references.
	(register_param_uses): Only handle uses here.
	(classtype_has_non_deleted_copy_ctor): New.
	(morph_fn_to_coro): Adjust param copy handling to match n4849
	by reordering ahead of the promise CTOR and always making a
	frame copy, even if the param is unused in the coroutine body.
	* cp-tree.h (classtype_has_non_deleted_copy_ctor): New.

gcc/testsuite/ChangeLog:

2020-02-26  Iain Sandoe  <iain@sandoe.co.uk>

	* g++.dg/coroutines/coro1-refs-and-ctors.h: New.
	* g++.dg/coroutines/torture/func-params-07.C: New test.
	* g++.dg/coroutines/torture/func-params-08.C: New test.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/221158

	PR target/93913
	* gcc.target/powerpc/fold-vec-st-char.c (scan-assembler-times): Allow
	stxv and stxvx instructions as well.
	* gcc.target/powerpc/fold-vec-st-float.c: Likewise.
	* gcc.target/powerpc/fold-vec-st-int.c: Likewise.
	* gcc.target/powerpc/fold-vec-st-short.c: Likewise.

This patch improves our concept diagnostics in two ways.  First, it sets a more
precise location for the constraint expressions built in
finish_constraint_binary_op.  As a result, when a disjunction is unsatisfied we
now print e.g.

.../include/bits/range_access.h:467:2: note: neither operand of the disjunction is satisfied
  466 |  requires is_bounded_array_v<remove_reference_t<_Tp>> || __member_end<_Tp>
      |           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  467 |  || __adl_end<_Tp>
      |  ^~~~~~~~~~~~~~~~~

instead of

.../include/bits/range_access.h:467:2: note: neither operand of the disjunction is satisfied
  467 |  || __adl_end<_Tp>
      |  ^~

Second, this patch changes diagnose_atomic_constraint to print unsatisfied
atomic constraint expressions with their template arguments.  So e.g. we now
print

cpp2a/concepts-pr67719.C:9:8: note: the expression ‘(... &&(C<Tx>)()) [with Tx = {int, long int, void}]’ evaluated to ‘false’

instead of

cpp2a/concepts-pr67719.C:9:8: note: the expression ‘(... &&(C<Tx>)())’ evaluated to ‘false’

Tested on x86_64-pc-linux-gnu, and verified that all the diagnostics emitted in
our concept tests are no worse with this patch.

gcc/cp/ChangeLog:

	* constraint.cc (finish_constraint_binary_op): Set expr's location range
	to the range of its operands.
	(satisfy_atom): Pass MAP instead of ARGS to diagnose_atomic_constraint.
	(diagnose_trait_expr): Take the instantiated parameter mapping MAP
	instead of the corresponding template arguments ARGS and adjust body
	accordingly.
	(diagnose_requires_expr): Likewise.
	(diagnose_atomic_constraint): Likewise.  When printing an atomic
	constraint expression, print the instantiated parameter mapping
	alongside it.
	* cxx-pretty-print.cc (cxx_pretty_printer::expression)
	[NONTYPE_ARGUMENT_PACK]: Print braces around a NONTYPE_ARGUMENT_PACK.
	(cxx_pretty_printer::type_id): Handle TYPE_ARGUMENT_PACK.

gcc/testsuite/ChangeLog:

	* g++.dg/concepts/diagnostic2.C: New test.
	* g++.dg/concepts/diagnostic3.C: New test.

Since <https://gcc.gnu.org/ml/gcc-patches/2015-02/msg00556.html> we
attempt to value-initialize in build_vec_init even when there's no
initializer but the type has a constexpr default constructor.  But
build_value_init doesn't work in templates, and build_vec_init
creates a lot of garbage that would not be used anyway, so don't
call it in a template.

	PR c++/93676 - value-init crash in template.
	* init.c (build_new_1): Don't call build_vec_init in a template.

	* g++.dg/cpp0x/nsdmi-template19.C: New test.

We are calling _OuterIter::__current from _InnerIter::operator==, but the former
is private within this non-member friend.  Fix this by calling
_OuterIter::operator== instead, which does the right thing here.

libstdc++-v3/ChangeLog:

	PR libstdc++/93936
	* include/std/ranges (split_view::_InnerIter::operator==): Compare
	the operands' _M_i rather than their _M_i.current().
	* testsuite/std/ranges/adaptors/split.cc: Augment test.

This adds constexpr to 11 algorithms defined in <numeric> as per P1645R1.

libstdc++-v3/ChangeLog:

	P1645R1 constexpr for <numeric> algorithms
	* include/bits/stl_numeric.h (iota, accumulate, inner_product,
	partial_sum, adjacent_difference): Make conditionally constexpr for
	C++20.
	* include/std/numeric (__cpp_lib_constexpr_numeric): Define this feature
	test macro.
	(reduce, transform_reduce, exclusive_scan, inclusive_scan,
	transform_exclusive_scan, transform_inclusive_scan): Make conditionally
	constexpr for C++20.
	* include/std/version (__cpp_lib_constexpr_numeric): Define.
	* testsuite/26_numerics/accumulate/constexpr.cc: New test.
	* testsuite/26_numerics/adjacent_difference/constexpr.cc: Likewise.
	* testsuite/26_numerics/exclusive_scan/constexpr.cc: Likewise.
	* testsuite/26_numerics/inclusive_scan/constexpr.cc: Likewise.
	* testsuite/26_numerics/inner_product/constexpr.cc: Likewise.
	* testsuite/26_numerics/iota/constexpr.cc: Likewise.
	* testsuite/26_numerics/partial_sum/constexpr.cc: Likewise.
	* testsuite/26_numerics/reduce/constexpr.cc: Likewise.
	* testsuite/26_numerics/transform_exclusive_scan/constexpr.cc: Likewise.
	* testsuite/26_numerics/transform_inclusive_scan/constexpr.cc: Likewise.
	* testsuite/26_numerics/transform_reduce/constexpr.cc: Likewise.

	* include/bits/ranges_algo.h (__lexicographical_compare_fn): Declare
	variables in smaller scope and avoid calling ranges::distance when we
	know they are pointers. Remove statically-unreachable use of
	__builtin_unreachable().
	* include/bits/stl_algobase.h (__lexicographical_compare::__lc):
	Define inline.

This introduces a couple of convenience alias templates to be used for
some repeated patterns using std::conditional_t.

	* include/std/ranges (__detail::__maybe_empty_t): Define new helper
	alias.
	(__detail::__maybe_const_t): Likewise.
	(__adaptor::_RangeAdaptor): Use __maybe_empty_t.
	(transform_view, take_view, take_while_view, elements_view): Use
	__maybe_const_t.
	(join_view, split_view): Use both.

This ICEs since my patch for P0388, which allowed conversions to arrays
of unknown bound, but not the reverse, so these two static_casts are
ill-formed.

[expr.static.cast]/3 says that "cv1 T1" and "cv2 T2" have to be
reference-compatible and the comment in build_static_cast_1 says it too
but then we actually use reference_related_p...  Fixed thus.

2020-02-26  Marek Polacek  <polacek@redhat.com>

	PR c++/93862 - ICE with static_cast when converting from int[].
	* call.c (reference_compatible_p): No longer static.
	* cp-tree.h (reference_compatible_p): Declare.
	* typeck.c (build_static_cast_1): Use reference_compatible_p instead
	of reference_related_p.

	* g++.dg/cpp0x/rv-cast7.C: New test.

DR 1423, which supersedes DR 654, says that you can't copy-init
a bool from a std::nullptr_t:

  bool b = nullptr;  // error

Conversely, it works with direct-initialization which is more
permissive than copy-initialization.

No code changes necessary since we handle it right.

2020-02-26  Marek Polacek  <polacek@redhat.com>

	DR 1423, Convertibility of nullptr to bool.
	* g++.dg/DRs/dr1423.C: New test.

Here we crash when constexpr-initializing a class member of empty class
type with [[no_unique_address]].  Without the attribute we would have
a ctor (that initializes bar) of the form

  { .D.2173 = { .x = {} } }

but with the attribute reduced_constant_expression_p gets

  { .x = {} }

That means that "idx != field" is true for the latter and we see that
foo, the base class of bar, is an empty class, so we want to look at
the next initializable field (since empty class fields may not have an
initializer).  But in this case there are no more, therefore accessing
DECL_CHAIN (field) crashes.  Long story short, we need to avoid a crash
on a null field when we're initializing a class that only contains an
empty base class.

While poking into this I discovered c++/93898, but that's a different
problem.

2020-02-26  Marek Polacek  <polacek@redhat.com>

	PR c++/93803 - ICE with constexpr init and [[no_unique_address]].
	* constexpr.c (reduced_constant_expression_p): Don't crash on a null
	field.

	* g++.dg/cpp2a/constexpr-init16.C: New test.
	* g++.dg/cpp2a/constexpr-init17.C: New test.

This adjusts dumping as proved useful in debugging.

2020-02-26  Richard Biener  <rguenther@suse.de>

	* tree-vect-slp.c (vect_print_slp_tree): Also dump ref count
	and load permutation.

In this PR we had a conversion between two integer vectors that
both had scalar integer modes.  We then tried to implement the
conversion using the scalar optab for those modes, instead of
doing the conversion elementwise.

I wondered about letting through scalar modes for single-element
vectors, but I don't have any evidence that that's useful/necessary,
so it seemed better to keep things simple.

2020-02-26  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
	PR middle-end/93843
	* optabs-tree.c (supportable_convert_operation): Reject types with
	scalar modes.

gcc/testsuite/
	PR middle-end/93843
	* gcc.dg/vect/pr93843-1.c: New test.
	* gcc.dg/vect/pr93843-2.c: Likewise.

This patch adds various information to -fdump-analyzer and
-fdump-analyzer-stderr to make it easier to track down
problems with state explosions in the exploded_graph.

It logs the number of unprocessed nodes in the worklist, for
the case where the upper limit on exploded nodes is reached.

It prints:
[a] a bar chart showing the number of exploded nodes by function, and

[b] bar charts for each function showing the number of exploded nodes
    per supernode/BB, and

[c] bar charts for each function showing the number of excess exploded
    nodes per supernode/BB beyond the limit
    (--param=analyzer-max-enodes-per-program-point), where that limit
    was reached

I've found these helpful in finding exactly where we fail to consolidate
state, leading to state explosions and false negatives due to the
thresholds being reached.

The patch also adds a "superedge::dump" member function I found myself
needing.

gcc/ChangeLog:
	* Makefile.in (ANALYZER_OBJS): Add analyzer/bar-chart.o.

gcc/analyzer/ChangeLog:
	* bar-chart.cc: New file.
	* bar-chart.h: New file.
	* engine.cc: Include "analyzer/bar-chart.h".
	(stats::log): Only log the m_num_nodes kinds that are non-zero.
	(stats::dump): Likewise when dumping.
	(stats::get_total_enodes): New.
	(exploded_graph::get_or_create_node): Increment the per-point-data
	m_excess_enodes when hitting the per-program-point limit on
	enodes.
	(exploded_graph::print_bar_charts): New.
	(exploded_graph::log_stats): Log the number of unprocessed enodes
	in the worklist.  Call print_bar_charts.
	(exploded_graph::dump_stats): Print the number of unprocessed
	enodes in the worklist.
	* exploded-graph.h (stats::get_total_enodes): New decl.
	(struct per_program_point_data): Add field m_excess_enodes.
	(exploded_graph::print_bar_charts): New decl.
	* supergraph.cc (superedge::dump): New.
	(superedge::dump): New.
	* supergraph.h (supernode::get_function): New.
	(superedge::dump): New decl.
	(superedge::dump): New decl.

2020-02-26  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/93820
	* gcc.dg/pr93820-2.c: New test.

The following testcase is miscompiled in 8+.
The problem is that check_no_overlap has a special case for INTEGER_CST
marked stores (i.e. stores of constants), if both all currenly merged stores
and the one under consideration for merging with them are marked that way,
it anticipates that other INTEGER_CST marked stores that overlap with those
and precede those (have smaller info->order) could be merged with those and
doesn't punt for them.
In PR86844 and PR87859 fixes I've then added quite large code that is
performed after check_no_overlap and tries to find out if we need and can
merge further INTEGER_CST marked stores, or need to punt.
Unfortunately, that code is there only in the overlapping case code and
the testcase below shows that we really need it even in the adjacent store
case.  After sort_by_bitpos we have:
bitpos	width	order	rhs_code
96	32	3	INTEGER_CST
128	32	1	INTEGER_CST
128	128	2	INTEGER_CST
192	32	0	MEM_REF
Because of the missing PR86844/PR87859-ish code in the adjacent store
case, we merge the adjacent (memory wise) stores 96/32/3 and 128/32/1,
and then we consider the 128-bit store which is in program-order in between
them, but in this case we punt, because the merging would extend the
merged store region from bitpos 96 and 64-bits to bitpos 96 and 160-bits
and that has an overlap with an incompatible store (the MEM_REF one).
The problem is that we can't really punt this way, because the 128-bit
store is in between those two we've merged already, so either we manage
to merge even that one together with the others, or would need to avoid
already merging the 96/32/3 and 128/32/1 stores together.
Now, rather than copying around the PR86844/PR87859 code to the other spot,
we can actually just use the overlapping code, merge_overlapping is really
a superset of merge_into, so that is what the patch does.  If doing
adjacent store merge for rhs_code other than INTEGER_CST, I believe the
current code is already fine, check_no_overlap in that case doesn't make
the exception and will punt if there is some earlier (smaller order)
non-mergeable overlapping store.  There is just one case that could be
problematic, if the merged_store has BIT_INSERT_EXPRs in them and the
new store is a constant store (INTEGER_CST rhs_code), then check_no_overlap
would do the exception and still would allow the special case.  But we
really shouldn't have the special case in that case, so this patch also
changes check_no_overlap to just have a bool whether we should have the
special case or not.

Note, as I said in the PR, for GCC11 we could consider performing some kind
of cheap DSE during the store merging (perhaps guarded with flag_tree_dse).
And another thing to consider is only consider as problematic non-mergeable
stores that not only have order smaller than last_order as currently, but
also have order larger than first_order, as in this testcase if we actually
ignored (not merged with anything at all) the 192/32/0 store, because it is
not in between the other stores we'd merge, it would be fine to merge the
other 3 stores, though of course the testcase can be easily adjusted by
putting the 192/32 store after the 128/32 store and then this patch would be
still needed.  Though, I think I'd need more time thinking this over.

2020-02-26  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/93820
	* gimple-ssa-store-merging.c (check_no_overlap): Change RHS_CODE
	argument to ALL_INTEGER_CST_P boolean.
	(imm_store_chain_info::try_coalesce_bswap): Adjust caller.
	(imm_store_chain_info::coalesce_immediate_stores): Likewise.  Handle
	adjacent INTEGER_CST store into merged_store->only_constants like
	overlapping one.

	* gcc.dg/pr93820.c: New test.

Add testcase for a bug that has been just on the 8 branch.

2020-02-26  Jakub Jelinek  <jakub@redhat.com>

	PR c++/93905
	* g++.dg/cpp0x/pr93905.C: New test.

2020-02-25  Jakub Jelinek  <jakub@redhat.com>

	PR other/93912
	* config/sh/sh.c (expand_cbranchdi4): Fix comment typo, probablity
	-> probability.
	* cfghooks.c (verify_flow_info): Likewise.
	* predict.c (combine_predictions_for_bb): Likewise.
	* bb-reorder.c (connect_better_edge_p): Likewise.  Fix comment typo,
	sucessor -> successor.
	(find_traces_1_round): Fix comment typo, destinarion -> destination.
	* omp-expand.c (expand_oacc_for): Fix comment typo, sucessors ->
	successors.
	* tree-ssa-loop-ch.c (should_duplicate_loop_header_p): Fix dump
	message typo, sucessors -> successors.
c/
	* gimple-parser.c (c_parser_gimple_parse_bb_spec_edge_probability):
	Rename last argument from probablity to probability.

gcc/ChangeLog:
	* doc/extend.texi (attribute access): Correct an example.

This patch adds the load/store bfloat16 intrinsics to the AArch64 back-end.
ACLE documents are at https://developer.arm.com/docs/101028/latest
ISA documents are at https://developer.arm.com/docs/ddi0596/latest

2020-02-25  Mihail Ionescu  <mihail.ionescu@arm.com>

gcc/
	* config/aarch64/aarch64-builtins.c (aarch64_scalar_builtin_types):
	Add simd_bf.
	(aarch64_init_simd_builtin_scalar_types): Register simd_bf.
	(VAR15, VAR16): New.
	* config/aarch64/iterators.md (VALLDIF): Enable for V4BF and V8BF.
	(VD): Enable for V4BF.
	(VDC): Likewise.
	(VQ): Enable for V8BF.
	(VQ2): Likewise.
	(VQ_NO2E): Likewise.
	(VDBL, Vdbl): Add V4BF.
	(V_INT_EQUIV, v_int_equiv): Add V4BF and V8BF.
	* config/aarch64/arm_neon.h (bfloat16x4x2_t): New typedef.
	(bfloat16x8x2_t): Likewise.
	(bfloat16x4x3_t): Likewise.
	(bfloat16x8x3_t): Likewise.
	(bfloat16x4x4_t): Likewise.
	(bfloat16x8x4_t): Likewise.
	(vcombine_bf16): New.
	(vld1_bf16, vld1_bf16_x2): New.
	(vld1_bf16_x3, vld1_bf16_x4): New.
	(vld1q_bf16, vld1q_bf16_x2): New.
	(vld1q_bf16_x3, vld1q_bf16_x4): New.
	(vld1_lane_bf16): New.
	(vld1q_lane_bf16): New.
	(vld1_dup_bf16): New.
	(vld1q_dup_bf16): New.
	(vld2_bf16): New.
	(vld2q_bf16): New.
	(vld2_dup_bf16): New.
	(vld2q_dup_bf16): New.
	(vld3_bf16): New.
	(vld3q_bf16): New.
	(vld3_dup_bf16): New.
	(vld3q_dup_bf16): New.
	(vld4_bf16): New.
	(vld4q_bf16): New.
	(vld4_dup_bf16): New.
	(vld4q_dup_bf16): New.
	(vst1_bf16, vst1_bf16_x2): New.
	(vst1_bf16_x3, vst1_bf16_x4): New.
	(vst1q_bf16, vst1q_bf16_x2): New.
	(vst1q_bf16_x3, vst1q_bf16_x4): New.
	(vst1_lane_bf16): New.
	(vst1q_lane_bf16): New.
	(vst2_bf16): New.
	(vst2q_bf16): New.
	(vst3_bf16): New.
	(vst3q_bf16): New.
	(vst4_bf16): New.
	(vst4q_bf16): New.

gcc/testsuite/
	* gcc.target/aarch64/advsimd-intrinsics/bf16_vstn.c: New test.
	* gcc.target/aarch64/advsimd-intrinsics/bf16_vldn.c: New test.