Last code gen change of LTGT didn't consider the situation of cbranch with LTGT,
branch only support few compare codes.

gcc/ChangeLog

	* config/riscv/riscv.c (riscv_emit_float_compare): Using NE to compare
	the result of IOR.

gcc/testsuite/ChangeLog

	* gcc.dg/pr93995.c: New.

My GCC 9 patch for C++20 P0846R0 (ADL and function templates) tweaked
cp_parser_template_name to only return an identifier if name lookup
didn't find anything.  In the deduce4.C case it means that we now
return an OVERLOAD.  That means that cp_parser_template_id will call
lookup_template_function whereby producing a TEMPLATE_ID_EXPR with
unknown_type_node.  Previously, we created a TEMPLATE_ID_EXPR with
no type, making it type-dependent.  What we have now is no longer
type-dependent.  And so, when we call finish_call_expr after we've
parsed "foo<int>(10)", even though we're in a template, we still do
the normal processing, thus perform overload resolution.  When adding
the template candidate foo we need to deduce the template arguments,
and that is where things go downhill.

When fn_type_unification sees that we have explicit template arguments,
but they aren't complete, it will use them to substitute the function
type.  So we substitute e.g. "void <T33d> (U)".  But the explicit
template argument was for a different parameter so we don't actually
substitute anything.  But the problem here was that we reduced the
template level of 'U' anyway.  So then when we're actually deducing
the template arguments via type_unification_real, we fail in unify:
22932       if (TEMPLATE_TYPE_LEVEL (parm)
22933           != template_decl_level (tparm))
22934         /* The PARM is not one we're trying to unify.  Just check
22935            to see if it matches ARG.  */
because 'parm' has been reduced but 'tparm' has not yet.

Therefore we shouldn't reduce the template level of template parameters
when tf_partial aka template argument deduction substitution.  But we
can only return after performing the cp_build_qualified_type etc.
business otherwise things break horribly.

2020-03-03  Jason Merrill  <jason@redhat.com>
	    Marek Polacek  <polacek@redhat.com>

	PR c++/90505 - mismatch in template argument deduction.
	* pt.c (tsubst): Don't reduce the template level of template
	parameters when tf_partial.

	* g++.dg/template/deduce4.C: New test.
	* g++.dg/template/deduce5.C: New test.
	* g++.dg/template/deduce6.C: New test.
	* g++.dg/template/deduce7.C: New test.

When deciding whether to perform the memset optimization in ranges::fill_n, we
were crucially neglecting to check that the output pointer's value type is a
byte type.  This patch adds such a check to the problematic condition in
ranges::fill_n.

At the same time, this patch relaxes the overly conservative
__is_byte<_Tp>::__value check that requires the fill type be a byte type.  It's
overly conservative because it means we won't enable the memset optimization in
the following example

  char c[100];
  ranges::fill(c, 37);

because the fill type is deduced to be int here.  Rather than requiring that the
fill type be a byte type, it seems safe to just require the fill type be an
integral type, which is what this patch does.

libstdc++-v3/ChangeLog:

	PR libstdc++/94017
	* include/bits/ranges_algobase.h (__fill_n_fn::operator()): Refine
	condition for when to use memset, making sure to additionally check that
	the output pointer's value type is a non-volatile byte type.  Instead of
	requiring that the fill type is a byte type, just require that it's an
	integral type.
	* testsuite/20_util/specialized_algorithms/uninitialized_fill/94017.cc:
	New test.
	* testsuite/20_util/specialized_algorithms/uninitialized_fill_n/94017.cc:
	New test.
	* testsuite/25_algorithms/fill/94013.cc: Uncomment part that was blocked
	by PR 94017.
	* testsuite/25_algorithms/fill/94017.cc: New test.
	* testsuite/25_algorithms/fill_n/94017.cc: New test.

This adds support for move-only input iterators in the ranges::unitialized_*
algorithms defined in <memory>, as per LWG 3355.  The only changes needed are to
add calls to std::move in the appropriate places and to use operator- instead of
ranges::distance because the latter cannot be used with a move-only iterator
that has a sized sentinel, as is the case here.  (This issue with
ranges::distance is LWG 3392.)

libstdc++-v3/ChangeLog:

	LWG 3355 The memory algorithms should support move-only input iterators
	introduced by P1207
	* include/bits/ranges_uninitialized.h
	(__uninitialized_copy_fn::operator()): Use std::move to avoid attempting
	to copy __ifirst, which could be a move-only input iterator.  Use
	operator- instead of ranges::distance to compute distance from a sized
	sentinel.
	(__uninitialized_copy_n_fn::operator()): Likewise.
	(__uninitialized_move_fn::operator()): Likewise.
	(__uninitialized_move_n_fn::operator()): Likewise.
	(__uninitialized_destroy_fn::operator()): Use std::move to avoid
	attempting to copy __first.
	(__uninitialized_destroy_n_fn::operator()): Likewise.
	* testsuite/20_util/specialized_algorithms/destroy/constrained.cc:
	Augment test.
	* .../specialized_algorithms/uninitialized_copy/constrained.cc:
	Likewise.
	* .../specialized_algorithms/uninitialized_move/constrained.cc:
	Likewise.

This adds a testsuite range type whose end() is a sized sentinel to
<testsuite_iterators.h>, which will be used in the tests that verify LWG 3355.

libstdc++-v3/ChangeLog:

	* testsuite/util/testsuite_iterators.h (test_range::get_iterator): Make
	protected instead of private.
	(test_sized_range_sized_sent): New.

This adds a move-only testsuite iterator wrapper to <testsuite_iterators.h>
which will be used in the tests for LWG 3355.  The tests for LWG 3389 and 3390
are adjusted to use this new iterator wrapper.

libstdc++-v3/ChangeLog:

	* testsuite/util/testsuite_iterators.h (input_iterator_wrapper_nocopy):
	New testsuite iterator.
	* testsuite/24_iterators/counted_iterator/lwg3389.cc: Use it.
	* testsuite/24_iterators/move_iterator/lwg3390.cc: Likewise.

We are passing a value type as the first argument to is_nothrow_assignable_v,
but the result of that is inevitably false.  Since this predicate is a part of
the condition that guards the corresponding optimizations for these algorithms,
this bug means these optimizations are never used.  We should be passing a
reference type to is_nothrow_assignable_v instead.

libstdc++-v3/ChangeLog:

	* include/bits/ranges_uninitialized.h
	(uninitialized_copy_fn::operator()): Pass a reference type as the first
	argument to is_nothrow_assignable_v.
	(uninitialized_copy_fn::operator()): Likewise.
	(uninitialized_move_fn::operator()): Likewise.  Return an in_out_result
	with the input iterator stripped of its move_iterator.
	(uninitialized_move_n_fn::operator()): Likewise.
	(uninitialized_fill_fn::operator()): Pass a reference type as the first
	argument to is_nothrow_assignable_v.
	(uninitialized_fill_n_fn::operator()): Likewise.

gcc/cp
        * coroutines.cc (captures_temporary): Strip component_ref
        to its base object.

gcc/testsuite
        * g++.dg/coroutines/torture/co-await-15-capture-comp-ref.C: New test.

Several algorithms check the is_trivially_copyable trait to decide
whether to dispatch to memmove or memcmp as an optimization. Since
r271435 (CWG DR 2094) the trait is true for volatile-qualified scalars,
but we can't use memmove or memcmp when the type is volatile. We need to
also check for volatile types.

This is complicated by the fact that in C++20 (but not earlier standards)
iterator_traits<volatile T*>::value_type is T, so we can't just check
whether the value_type is volatile.

The solution in this patch is to introduce new traits __memcpyable and
__memcmpable which combine into a single trait the checks for pointers,
the value types being the same, and the type being trivially copyable
but not volatile-qualified.

	PR libstdc++/94013
	* include/bits/cpp_type_traits.h (__memcpyable, __memcmpable): New
	traits to control when to use memmove and memcmp optimizations.
	(__is_nonvolatile_trivially_copyable): New helper trait.
	* include/bits/ranges_algo.h (__lexicographical_compare_fn): Do not
	use memcmp optimization with volatile data.
	* include/bits/ranges_algobase.h (__equal_fn): Use __memcmpable.
	(__copy_or_move, __copy_or_move_backward): Use __memcpyable.
	* include/bits/stl_algobase.h (__copy_move_a2): Use __memcpyable.
	(__copy_move_backward_a2): Likewise.
	(__equal_aux1): Use __memcmpable.
	(__lexicographical_compare_aux): Do not use memcmp optimization with
	volatile data.
	* testsuite/25_algorithms/copy/94013.cc: New test.
	* testsuite/25_algorithms/copy_backward/94013.cc: New test.
	* testsuite/25_algorithms/equal/94013.cc: New test.
	* testsuite/25_algorithms/fill/94013.cc: New test.
	* testsuite/25_algorithms/lexicographical_compare/94013.cc: New test.
	* testsuite/25_algorithms/move/94013.cc: New test.
	* testsuite/25_algorithms/move_backward/94013.cc: New test.

We ICE on the following testcase since I've added the SAVE_EXPR-like
constexpr handling where the TARGET_EXPR initializer (and cleanup) is
evaluated only once (because it might have side-effects like new or delete
expressions in it).
The problem is if the TARGET_EXPR (but I guess in theory SAVE_EXPR too)
initializer is *non_constant_p.  We still remember the result, but already
not that it is *non_constant_p.  Normally that wouldn't be a big problem,
if something is *non_constant_p, we only or into it and so the whole
expression will be non-constant too.  Except in the builtins handling,
we try to evaluate the arguments with non_constant_p pointing into a dummy1
bool which we ignore.  This is because some builtins might fold into a
constant even if they don't have a constexpr argument.  Unfortunately if
we evaluate the TARGET_EXPR first in the argument of such a builtin and then
once again, we don't set *non_constant_p.

So, either we don't remember the TARGET_EXPR/SAVE_EXPR result if it wasn't
constant, like the following patch does, or we could remember it, but in
some way that would make it clear that it is non-constant (e.g. by
pushing into the global->values SAVE_EXPR, SAVE_EXPR entry and perhaps
for TARGET_EXPR don't remember it on TARGET_EXPR_SLOT, but the TARGET_EXPR
itself and similarly push TARGET_EXPR, TARGET_EXPR and if we see those
after the lookup, diagnose + set *non_constant_p.  Or we could perhaps
during the builtin argument evaluation push expressions into a different
save_expr vec and undo them afterwards.

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

	PR c++/93998
	* constexpr.c (cxx_eval_constant_expression)
	<case TARGET_EXPR, case SAVE_EXPR>: Don't record anything if
	*non_constant_p is true.

	* g++.dg/ext/pr93998.C: New test.

Unified syntax has been the official syntax for thumb1 assembly for
over 10 years now.  It's time we made preparations for that becoming
the default in the assembler.  But before we can start doing that we
really need to clean up some laggards from the olden days.  Libgcc
support for thumb1 is one such example.

This patch converts all of the legacy (disjoint) syntax that I could
find over to unified code.  The identification was done by using a
trick version of gas that defaulted to unified mode which then faults
if legacy syntax is encountered.  The code produced was then compared
against the old code to check for differences.  One such difference
does exist, but that is because in unified syntax 'movs rd, rn' is
encoded as 'lsls rd, rn, #0', rather than 'adds rd, rn, #0'; but that
is a deliberate change that was introduced because the lsls encoding
more closely reflects the behaviour of 'movs' in arm state (where only
some of the condition flags are modified).

	* config/arm/bpabi-v6m.S (aeabi_lcmp): Convert thumb1 code to unified
	syntax.
	(aeabi_ulcmp, aeabi_ldivmod, aeabi_uldivmod): Likewise.
	(aeabi_frsub, aeabi_cfcmpeq, aeabi_fcmpeq): Likewise.
	(aeabi_fcmp, aeabi_drsub, aeabi_cdrcmple): Likewise.
	(aeabi_cdcmpeq, aeabi_dcmpeq, aeabi_dcmp): Likewise.
	* config/arm/lib1funcs.S (Lend_fde): Convert thumb1 code to unified
	syntax.
	(divsi3, modsi3): Likewise.
	(clzdi2, ctzsi2): Likewise.
	* config/arm/libunwind.S (restore_core_regs): Convert thumb1 code to
	unified syntax.
	(UNWIND_WRAPPER): Likewise.

This patch is part of a series adding support for Armv8.6-A features.
It implements intrinsics to convert between bfloat16 and float32
formats.

gcc/ChangeLog:

	* config/arm/arm_bf16.h (vcvtah_f32_bf16, vcvth_bf16_f32): New.
	* config/arm/arm_neon.h (vcvt_f32_bf16, vcvtq_low_f32_bf16): New.
	(vcvtq_high_f32_bf16, vcvt_bf16_f32): New.
	(vcvtq_low_bf16_f32, vcvtq_high_bf16_f32): New.
	* config/arm/arm_neon_builtins.def (vbfcvt, vbfcvt_high): New entries.
	(vbfcvtv4sf, vbfcvtv4sf_high): Likewise.
	* config/arm/iterators.md (VBFCVT, VBFCVTM): New mode iterators.
	(V_bf_low, V_bf_cvt_m): New mode attributes.
	* config/arm/neon.md (neon_vbfcvtv4sf<VBFCVT:mode>): New.
	(neon_vbfcvtv4sf_highv8bf, neon_vbfcvtsf): New.
	(neon_vbfcvt<VBFCVT:mode>, neon_vbfcvt_highv8bf): New.
	(neon_vbfcvtbf_cvtmode<mode>, neon_vbfcvtbf): New
	* config/arm/unspecs.md (UNSPEC_BFCVT, UNSPEC_BFCVT_HIG): New.

gcc/testsuite/ChangeLog:

	* gcc.target/arm/simd/bf16_cvt_1.c: New test.

As noted in LWG 3410 the specification in the C++20 draft performs more
iterator comparisons than necessary when the end of either range is
reached. Our implementation followed that specification. This removes
the redundant comparisons so that we do no unnecessary work as soon as
we find that we've reached the end of either range.

The odd-looking return statement is because it generates better code
than the original version that copied the global constants.

	* include/bits/stl_algobase.h (lexicographical_compare_three_way):
	Avoid redundant iterator comparisons (LWG 3410).

As mentioned in the PR and discussed on IRC, the following patch is the
patch that fixes the originally reported issue.
We have there because of the premature bitfield comparison -> BIT_FIELD_REF
optimization:
  s$s4_19 = 0;
  s.s4 = s$s4_19;
  _10 = BIT_FIELD_REF <s, 8, 0>;
  _13 = _10 & 8;
and no other s fields are initialized.  If they would be all initialized with
constants, then my earlier PR93582 bitfield handling patches would handle it
already, but if at least one bit we ignore after the BIT_AND_EXPR masking
is not initialized or is initialized earlier to non-constant, we aren't able
to look through it until combine, which is too late for the warnings on the
dead code.
This patch handles BIT_AND_EXPR where the first operand is a SSA_NAME
initialized with a memory load and second operand is INTEGER_CST, by trying
a partial def lookup after pushing the ranges of 0 bits in the mask as
artificial initializers.  In the above case on little-endian, we push
offset 0 size 3 {} partial def and offset 4 size 4 (the result is unsigned
char) and then perform normal partial def handling.
My initial version of the patch failed miserably during bootstrap, because
data->finish (...) called vn_reference_lookup_or_insert_for_pieces
which I believe tried to remember the masked value rather than real for the
reference, or for failed lookup visit_reference_op_load called
vn_reference_insert.  The following version makes sure we aren't calling
either of those functions in the masked case, as we don't know anything
better about the reference from whatever has been discovered when the load
stmt has been visited, the patch just calls vn_nary_op_insert_stmt on
failure with the lhs (apparently calling it with the INTEGER_CST doesn't
work).

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

	PR tree-optimization/93582
	* tree-ssa-sccvn.h (vn_reference_lookup): Add mask argument.
	* tree-ssa-sccvn.c (struct vn_walk_cb_data): Add mask and masked_result
	members, initialize them in the constructor and if mask is non-NULL,
	artificially push_partial_def {} for the portions of the mask that
	contain zeros.
	(vn_walk_cb_data::finish): If mask is non-NULL, set masked_result to
	val and return (void *)-1.  Formatting fix.
	(vn_reference_lookup_pieces): Adjust vn_walk_cb_data initialization.
	Formatting fix.
	(vn_reference_lookup): Add mask argument.  If non-NULL, don't call
	fully_constant_vn_reference_p nor vn_reference_lookup_1 and return
	data.mask_result.
	(visit_nary_op): Handle BIT_AND_EXPR of a memory load and INTEGER_CST
	mask.
	(visit_stmt): Formatting fix.

	* gcc.dg/tree-ssa/pr93582-10.c: New test.
	* gcc.dg/pr93582.c: New test.
	* gcc.c-torture/execute/pr93582.c: New test.

This fixes a common mistake in removing a store that looks redudnant but
is not because it changes the dynamic type of the memory and thus makes
a difference for following loads with TBAA.

2020-03-03  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/93946
	* alias.h (refs_same_for_tbaa_p): Declare.
	* alias.c (refs_same_for_tbaa_p): New function.
	* tree-ssa-alias.c (ao_ref_alias_set): For a NULL ref return
	zero.
	* tree-ssa-scopedtables.h
	(avail_exprs_stack::lookup_avail_expr): Add output argument
	giving access to the hashtable entry.
	* tree-ssa-scopedtables.c (avail_exprs_stack::lookup_avail_expr):
	Likewise.
	* tree-ssa-dom.c: Include alias.h.
	(dom_opt_dom_walker::optimize_stmt): Validate TBAA state before
	removing redundant store.
	* tree-ssa-sccvn.h (vn_reference_s::base_set): New member.
	(ao_ref_init_from_vn_reference): Adjust prototype.
	(vn_reference_lookup_pieces): Likewise.
	(vn_reference_insert_pieces): Likewise.
	* tree-ssa-sccvn.c: Track base alias set in addition to alias
	set everywhere.
	(eliminate_dom_walker::eliminate_stmt): Also check base alias
	set when removing redundant stores.
	(visit_reference_op_store): Likewise.
	* dse.c (record_store): Adjust valdity check for redundant
	store removal.

	* gcc.dg/torture/pr93946-1.c: New testcase.
	* gcc.dg/torture/pr93946-2.c: Likewise.

In Fedora we configure GCC with --with-arch=zEC12 --with-tune=z13 right now
and furthermore redhat-rpm-config adds to rpm packages -march=zEC12 -mtune=z13
options (among others).  While looking at the git compilation, I've been
surprised that -O2 actually behaves differently from -O2 -mtune=z13 in this
configuration, and indeed, seems --with-tune= is completely ignored on s390
if --with-arch= is specified.

i386 had the same problem, but got that fixed in 2006, see PR26877.
The thing is that for tune, we add -mtune=%(VALUE) only if neither -mtune=
nor -march= is present, but as arch is processed first, it adds
-march=%(VALUE) first and then -march= is always present and so -mtune= is
never added.
By reordering it in OPTION_DEFAULT_SPECS, we process tune first, add the
default -mtune=%(VALUE) if -mtune= or -march= isn't seen, and then
add -march=%(VALUE) if -march= isn't seen.  It is true that cc1 etc.
will be then invoked with -mtune=z13 -march=zEC12, but like if the user
specifies it in that order, it should still use z13 tuning and zEC12
ISA set.

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

	PR target/26877
	* config/s390/s390.h (OPTION_DEFAULT_SPECS): Reorder.

The following testcase ICEs in cross to riscv64-linux.  The problem is
that we have a DImode integral constant (that doesn't fit into SImode),
which is pushed into a constant pool and later access just the first half of
it using a MEM.  When plus_constant is called on such a MEM, if the constant
has mode, we verify the mode, but if it doesn't, we don't and ICE later on
when we think the CONST_INT is a valid SImode constant.

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

	PR rtl-optimization/94002
	* explow.c (plus_constant): Punt if cst has VOIDmode and
	get_pool_mode is different from mode.

	* gcc.dg/pr94002.c: New test.

All ARC's small data adressing is using address scaling feature of the
load/store instructions (i.e., the address is made of a general
pointer plus a shifted offset. The shift amount depends on the
addressing mode).  This patch is checking the offset of an address if
it fits the scalled constraint.  If so, a small data access is
generated.  This patch fixes execute' pr93249 failure.

gcc/
xxxx-xx-xx  Claudiu Zissulescu  <claziss@synopsys.com>

	* config/arc/arc.c (leigitimate_small_data_address_p): Check if an
	address has an offset which fits the scalling constraint for a
	load/store operation.
	(legitimate_scaled_address_p): Update use
	leigitimate_small_data_address_p.
	(arc_print_operand): Likewise.
	(arc_legitimate_address_p): Likewise.
	(legitimate_small_data_address_p): Likewise.

Signed-off-by: Claudiu Zissulescu <claziss@gmail.com>

With the refurbish of ARC600' accumulator support, the mlo_operand
doesn't reflect the proper low accumulator register for the newer
ARCv2 accumulator register used by the fma instructions.  Hence,
replace it with accl_operand predicate.

gcc/
xxxx-xx-xx  Claudiu Zissulescu  <claziss@synopsys.com>

	* config/arc/arc.md (fmasf4_fpu): Use accl_operand predicate.
	(fnmasf4_fpu): Likewise.

Early expand ADDDI3 and SUBDI3 for better code gen.

gcc/
xxxx-xx-xx  Claudiu Zissulescu  <claziss@synopsys.com>

	* config/arc/arc.md (adddi3): Early expand the 64bit operation into
	32bit ops.
	(subdi3): Likewise.
	(adddi3_i): Remove pattern.
	(subdi3_i): Likewise.

gcc/
xxxx-xx-xx  Claudiu Zissulescu  <claziss@synopsys.com>

	* config/arc/arc.md (eh_return): Add length info.

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

	PR tree-optimization/93927
	* gcc.c-torture/compile/pr93927-1.c: New test.
	* gcc.c-torture/compile/pr93927-2.c: New test.

gcc/cp
        * coroutines.cc (finish_co_await_expr): Build co_await_expr
        with unknown_type_node.
        (finish_co_yield_expr): Ditto.
        *pt.c (type_dependent_expression_p): Set co_await/yield_expr
        with unknown type as dependent.

gcc/testsuite
        * g++.dg/coroutines/torture/co-await-14-template-traits.C: New test.

The note about duplicates attached to analyzer diagnostics feels like an
implementation detail; it's likely just noise from the perspective of an
end-user.

This patch disables it by default, introducing a flag to re-enable it.

gcc/analyzer/ChangeLog:
	* analyzer.opt (fanalyzer-show-duplicate-count): New option.
	* diagnostic-manager.cc
	(diagnostic_manager::emit_saved_diagnostic): Use the above to
	guard the printing of the duplicate count.

gcc/ChangeLog:
	* doc/invoke.texi (-fanalyzer-show-duplicate-count): New.

gcc/testsuite/ChangeLog:
	* gcc.dg/analyzer/CVE-2005-1689-dedupe-issue.c: Add
	-fanalyzer-show-duplicate-count.

gcc/ChangeLog:
	* doc/invoke.texi (Static Analyzer Options): Add
	-Wanalyzer-stale-setjmp-buffer to the list of options enabled
	by -fanalyzer.

PR analyzer/93959 reported that g++.dg/analyzer/malloc.C was failing
with no output on Solaris.

The issue is that <stdlib.h> there has "using std::free;", converting
all the "free" calls to std::free, which fails the name-matching via
is_named_call_p.

This patch implements an is_std_named_call_p variant of is_named_call_p
to check for the name within "std", and uses it in sm-malloc.c to check
for std::malloc, std::calloc, and std::free.

gcc/analyzer/ChangeLog:
	PR analyzer/93959
	* analyzer.cc (is_std_function_p): New function.
	(is_std_named_call_p): New functions.
	* analyzer.h (is_std_named_call_p): New decl.
	* sm-malloc.cc (malloc_state_machine::on_stmt): Check for "std::"
	variants when checking for malloc, calloc and free.

gcc/testsuite/ChangeLog:
	PR analyzer/93959
	* g++.dg/analyzer/cstdlib-2.C: New test.
	* g++.dg/analyzer/cstdlib.C: New test.

In the absence of specific comment on the handling of closures I'd
implemented something more than was intended (extending the lifetime
of lambda capture-by-copy vars to the duration of the coro).

After discussion at WG21 in February and by email, the correct handling
is to treat the closure "this" pointer the same way as for a regular one,
and thus it is the user's responsibility to ensure that the lambda capture
object has suitable lifetime for the coroutine.  It is noted that users
frequently get this wrong, so it would be a good thing to revisit for C++23.

This patch removes the additional copying behaviour for lambda capture-by-
copy vars.

gcc/cp/ChangeLog:

2020-03-02  Iain Sandoe  <iain@sandoe.co.uk>

	* coroutines.cc (struct local_var_info): Adjust to remove the
	reference to the captured var, and just to note that this is a
	lambda capture proxy.
	(transform_local_var_uses): Handle lambda captures specially.
	(struct param_frame_data): Add a visited set.
	(register_param_uses): Also check for param uses in lambda
	capture proxies.
	(struct local_vars_frame_data): Remove captures list.
	(register_local_var_uses): Handle lambda capture proxies by
	noting and bypassing them.
	(morph_fn_to_coro): Update to remove lifetime extension of
	lambda capture-by-copy vars.

gcc/testsuite/ChangeLog:

2020-03-02  Iain Sandoe  <iain@sandoe.co.uk>
	    Jun Ma <JunMa@linux.alibaba.com>

	* g++.dg/coroutines/torture/class-05-lambda-capture-copy-local.C:
	* g++.dg/coroutines/torture/lambda-09-init-captures.C: New test.
	* g++.dg/coroutines/torture/lambda-10-mutable.C: New test.

*movstrict<mode>_1 insn pattern allows only general registers,
so we have to reject modes not suitable for general regs in
corresponding movstrict<mode> expander.

	PR target/93997
	* config/i386/i386.md (movstrict<mode>): Allow only
	registers with VALID_INT_MODE_P modes.

testsuite/ChangeLog:

	PR target/93997
	* gcc.target/i386/pr93997.c: New test.

gcc/testsuite/ChangeLog:
	PR tree-optimization/92982
	* gcc.dg/strlenopt-94.c: New test.

One missing bit from r10-6656.  The docs and target-supports.exp
already handle -std=gnu++20.

2020-03-02  Marek Polacek  <polacek@redhat.com>

	PR c++/93958 - add missing -std=gnu++20.
	* c.opt: Add -std=gnu++20.

        PR fortran/93486
        * module.c: Increase size of variables used to read module names
        when loading interfaces from module files to permit cases where
        the name is the concatenation of a module and submodule name.
        * gfortran.dg/pr93486.f90: New test.

The new 25_algorithms/lexicographical_compare/93972.cc test fails on
targets where char is unsigned, revealing an existing regression with
the std::__memcmp helper that had gone unnoticed in
std::lexicographical_compare. When comparing char and unsigned char, the
memcmp optimisation is enabled, but the new std::__memcmp function fails
to compile for mismatched types.

	PR libstdc++/93972
	* include/bits/stl_algobase.h (__memcmp): Allow pointer types to
	differ.
	* testsuite/25_algorithms/lexicographical_compare/uchar.cc: New test.

The key property of this alias is not that it may be an empty type, but
that the type argument may not be used. The fact it's replaced by an
empty type is just an implementation detail.  The name was also
backwards with respect to the bool argument.

This patch changes the name to better reflect its purpose.

	* include/std/ranges (__detail::__maybe_empty_t): Rename to
	__maybe_present_t.
	(__adaptor::_RangeAdaptor, join_view, split_view): Use new name.

In general, we need to manage the lifetime of compiler-
generated awaitable instances in the coroutine frame, since
these must persist across suspension points.

However, it is quite possible that the user might provide the
awaitable instances, either as function params or as a local
variable.  We will already generate a frame entry for these as
required.

At present, under this circumstance, we are duplicating these,
awaitable, initialising a second frame copy for them (which we
then subsequently destroy manually after the suspension point).
That's not efficient - so an undesirable thinko in the first place.
However, there is also an actual bug; if the compiler elects to
elide the copy (which is perfectly legal), it does not have visibility
of the manual management of the post-suspend destruction
- this subsequently leads to double-free errors.

The solution is not to make the second copy (as noted, params
and local vars already have frame copies with managed lifetimes).

gcc/cp/ChangeLog:

2020-03-02  Iain Sandoe  <iain@sandoe.co.uk>

	* coroutines.cc (build_co_await): Do not build frame
	proxy vars when the co_await expression is a function
	parameter or local var.
	(co_await_expander): Do not initialise a frame var with
	itself.
	(transform_await_expr): Only substitute the awaitable
	frame var if it's needed.
	(register_awaits): Do not make frame copies for param
	or local vars that are awaitables.

gcc/testsuite/ChangeLog:

2020-03-02  Iain Sandoe  <iain@sandoe.co.uk>

	* g++.dg/coroutines/torture/func-params-09-awaitable-parms.C: New test.
	* g++.dg/coroutines/torture/local-var-5-awaitable.C: New test.

Add support for V64DFmode addition, and V64DImode min, max.  There's no
direct hardware support for these, so we use regular vector instructions
and separate lane shift instructions.

Also add support for V64QI and V64HI reductions. Some of these require
additional extends and truncates, because AMD GCN has 32-bit vector lanes.

2020-03-02  Andrew Stubbs  <ams@codesourcery.com>

	gcc/
	* config/gcn/gcn-valu.md (dpp_move<mode>): New.
	(reduc_insn): Use 'U' and 'B' operand codes.
	(reduc_<reduc_op>_scal_<mode>): Allow all types.
	(reduc_<reduc_op>_scal_v64di): Delete.
	(*<reduc_op>_dpp_shr_<mode>): Allow all 1reg types.
	(*plus_carry_dpp_shr_v64si): Change to ...
	(*plus_carry_dpp_shr_<mode>): ... this and allow all 1reg int types.
	(mov_from_lane63_v64di): Change to ...
	(mov_from_lane63_<mode>): ... this, and allow all 64-bit modes.
	* config/gcn/gcn.c (gcn_expand_dpp_shr_insn): Increase buffer size.
	Support UNSPEC_MOV_DPP_SHR output formats.
	(gcn_expand_reduc_scalar): Add "use_moves" reductions.
	Add "use_extends" reductions.
	(print_operand_address): Add 'I' and 'U' codes.
	* config/gcn/gcn.md (unspec): Add UNSPEC_MOV_DPP_SHR.

	* gcc.target/arm/fuse-caller-save.c: Update expected output.

	* include/bits/ranges_algo.h (shift_right): Add 'typename' to
	dependent type.