Strengthen a few OpenACC test cases

Rather than rubber-stamp whatever requested vs. actual device kernel launch
configuration happens, actually (again) verify the requested values (modulo
expected variations).

This better highlights that "AMD GCN has an upper limit of 'num_workers(16)'",
and the deficiency that "AMD GCN uses the autovectorizer for the vector
dimension: the use of a function call in vector-partitioned code [...] is not
currently supported".

And, this removes several instances of race conditions, where variables are
concurrently written to in OpenACC gang-redundant mode.

libgomp/
* testsuite/libgomp.oacc-c-c++-common/loop-gwv-1.c: Strengthen.
* testsuite/libgomp.oacc-c-c++-common/loop-gwv-2.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-red-gwv-1.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-red-v-1.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-red-v-2.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-red-w-1.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-red-w-2.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-red-wv-1.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-v-1.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-w-1.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/loop-wv-1.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/routine-gwv-1.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/routine-v-1.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/routine-w-1.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/routine-wv-1.c: Likewise.

c++: [[no_unique_address]] and virtual base [PR104139]

Fixing a thinko in my patch for 103681: when computing the size of a virtual
base, it would help to use its binfo instead of the one for the derived
class.

PR c++/104139
PR c++/103681

gcc/cp/ChangeLog:

* class.cc (end_of_class): Use base_binfo.

gcc/testsuite/ChangeLog:

* g++.dg/abi/no_unique_address2.C: Adjust to detect this on x86-64.

libstdc++: Fix typo in comment

libstdc++-v3/ChangeLog:

* testsuite/20_util/shared_ptr/cons/array.cc: Fix comment.

libstdc++: Ensure all feature test macros have type long [PR87193]

This defines all the __cpp_lib_xxx macros as type long, as required by
the standard. We had an inconsistent mix of int and long, sometimes even
for the same macro name.

The __cpp_lib_experimental_xxx macros are left as type int, because
that's what it says in the relevant TS specs.

libstdc++-v3/ChangeLog:

PR libstdc++/87193
PR libstdc++/104019
* include/bits/alloc_traits.h (__cpp_lib_allocator_traits_is_always_equal):
Define as type long.
* include/bits/allocator.h (__cpp_lib_incomplete_container_elements):
Likewise.
* include/bits/basic_string.h (__cpp_lib_string_udls): Likewise.
* include/bits/chrono.h (__cpp_lib_chrono): Likewise.
(__cpp_lib_chrono_udls): Likewise.
* include/bits/move.h (__cpp_lib_addressof_constexpr): Likewise.
* include/bits/node_handle.h (__cpp_lib_node_extract): Likewise.
* include/bits/range_access.h (__cpp_lib_nonmember_container_access):
Likewise.
* include/bits/shared_ptr.h (__cpp_lib_enable_shared_from_this):
Likewise.
* include/bits/stl_algo.h (__cpp_lib_clamp): Likewise.
(__cpp_lib_sample): Likewise.
* include/bits/stl_algobase.h (__cpp_lib_robust_nonmodifying_seq_ops):
Likewise.
* include/bits/stl_function.h (__cpp_lib_transparent_operators):
Likewise.
* include/bits/stl_iterator.h (__cpp_lib_make_reverse_iterator):
Likewise.
* include/bits/stl_map.h (__cpp_lib_map_try_emplace):
Likewise.
* include/bits/stl_tree.h (__cpp_lib_generic_associative_lookup):
Likewise.
* include/bits/unique_ptr.h (__cpp_lib_make_unique):
Likewise.
* include/bits/unordered_map.h (__cpp_lib_unordered_map_try_emplace):
Likewise.
* include/c_global/cmath (__cpp_lib_hypot): Likewise.
* include/c_global/cstddef (__cpp_lib_byte): Likewise.
* include/std/atomic (__cpp_lib_atomic_is_always_lock_free):
Likewise.
* include/std/complex (__cpp_lib_complex_udls): Likewise.
* include/std/filesystem (__cpp_lib_filesystem): Likewise.
* include/std/functional (__cpp_lib_not_fn): Likewise.
(__cpp_lib_boyer_moore_searcher): Likewise.
* include/std/iomanip (__cpp_lib_quoted_string_io): Likewise.
* include/std/mutex (__cpp_lib_scoped_lock): Likewise.
* include/std/numeric (__cpp_lib_gcd_lcm): Likewise.
(__cpp_lib_gcd, __cpp_lib_lcm): Likewise.
* include/std/tuple (__cpp_lib_apply): Likewise.
(__cpp_lib_make_from_tuple): Likewise.
* include/std/type_traits (__cpp_lib_integral_constant_callable)
(__cpp_lib_bool_constant, __cpp_lib_logical_traits)
(__cpp_lib_is_null_pointer, __cpp_lib_transformation_trait_aliases)
(__cpp_lib_result_of_sfinae, __cpp_lib_void_t)
(__cpp_lib_is_swappable, __cpp_lib_is_invocable)
(__cpp_lib_has_unique_object_representations)
(__cpp_lib_is_aggregate): Likewise.
* include/std/version: Likewise.
* libsupc++/new (__cpp_lib_launder): Likewise.

libstdc++: Fix condition for __cpp_lib_shared_ptr_arrays

I changed the preprocessor condition from <= to < in r12-6574 which
meant the macro was not defined by <version> for C++17.

libstdc++-v3/ChangeLog:

* include/std/version (__cpp_lib_shared_ptr_arrays): Fix
condition for C++17 definition.

Enable configure detection of ld.mold.

gcc/ChangeLog:

* configure.ac: Detect ld_is_mold and use it for
comdat_group=yes and gcc_cv_ld_hidden=yes.
* configure: Regenerate.

tree-optimization/100089 - BB vectorization of if-converted loop bodies

The PR complains that when we only partially BB vectorize an
if-converted loop body that this can leave unvectorized code
unconditionally executed and thus effectively slow down code.
For -O2 we already mitigated the issue by not doing BB vectorization
when not all if-converted stmts were covered but the issue is
present with -O3 as well.  Thus the following simply extends the
fix to cover all but the unlimited cost models.  It is after all
very likely that we vectorize some stmts, if only a single
paired store.

2022-01-21  Richard Biener  <rguenther@suse.de>

PR tree-optimization/100089
* tree-vect-slp.cc (vect_slp_region): Reject BB vectorization
of if-converted loops with unvectorized COND_EXPRs for
all but the unlimited cost models.

libstdc++: Fix constexpr constructor for atomic<shared_ptr<T>>

libstdc++-v3/ChangeLog:

* include/bits/shared_ptr_atomic.h (_Sp_atomic::_Atomic_count):
Add constexpr.
(_Sp_atomic::_M_ptr): Add default member-initializer.
* testsuite/20_util/shared_ptr/atomic/atomic_shared_ptr.cc:
Check constant initialization.
* testsuite/20_util/weak_ptr/atomic_weak_ptr.cc: Likewise.

sanitizer/99673 - bad -Wstringop-overread diagnostic with asan

The testcase got fixed by lowering of &MEM[ptr + CST] to ptr + CST.

2022-01-21  Richard Biener  <rguenther@suse.de>

PR sanitizer/99673
* gcc.dg/asan/pr99673.c: New testcase.

[ARM] Add support for TLS register based stack protector canary access

Add support for accessing the stack canary value via the TLS register,
so that multiple threads running in the same address space can use
distinct canary values. This is intended for the Linux kernel running in
SMP mode, where processes entering the kernel are essentially threads
running the same program concurrently: using a global variable for the
canary in that context is problematic because it can never be rotated,
and so the OS is forced to use the same value as long as it remains up.

Using the TLS register to index the stack canary helps with this, as it
allows each CPU to context switch the TLS register along with the rest
of the process, permitting each process to use its own value for the
stack canary.

gcc/ChangeLog:

* config/arm/arm-opts.h (enum stack_protector_guard): New.
* config/arm/arm-protos.h (arm_stack_protect_tls_canary_mem):
New.
* config/arm/arm.cc (TARGET_STACK_PROTECT_GUARD): Define.
(arm_option_override_internal): Handle and put in error checks.
for stack protector guard options.
(arm_option_reconfigure_globals): Likewise.
(arm_stack_protect_tls_canary_mem): New.
(arm_stack_protect_guard): New.
* config/arm/arm.md (stack_protect_set): New.
(stack_protect_set_tls): Likewise.
(stack_protect_test): Likewise.
(stack_protect_test_tls): Likewise.
(reload_tp_hard): Likewise.
* config/arm/arm.opt (-mstack-protector-guard): New
(-mstack-protector-guard-offset): New.
* doc/invoke.texi: Document new options.

gcc/testsuite/ChangeLog:

* gcc.target/arm/stack-protector-7.c: New test.
* gcc.target/arm/stack-protector-8.c: New test.

aarch64: Adjust spellcheck tests for recent quotation-mark changes

Adjust tests for quotation-mark and punctuation changes.

gcc/testsuite:

* gcc.target/aarch64/spellcheck_1.c: Adjust tests for new output.
* gcc.target/aarch64/spellcheck_2.c: Likewise.
* gcc.target/aarch64/spellcheck_3.c: Likewise.
* gcc.target/aarch64/spellcheck_7.c: Likewise.

tree-optimization/104156 - fix unswitching compare-debug issue

When hoisting guards the unswitching pass does not properly ignore
debug stmts when looking for uses outside of the loop of defs
produced in the skipped region.  The following rectifies this
by instead collecting them and resetting them after the transform.

2022-01-21  Richard Biener  <rguenther@suse.de>

PR tree-optimization/104156
* tree-ssa-loop-unswitch.cc (tree_unswitch_outer_loop):
Collect and reset debug stmts with out-of-loop uses when
hoisting guards.
(find_loop_guard): Adjust.
(empty_bb_without_guard_p): Likewise.  Ignore debug stmts.
(used_outside_loop_p): Push debug uses to a vector of
debug stmts to reset.
(hoist_guard): Adjust -fopt-info category.

* gcc.dg/loop-unswitch-6.c: New testcase.

tree-optimization/104152 - add missing check for SLP constant build

This adds a missing check to verify we can actually build an
invariant vector from components when SLP vectorizing an associatable
chain.

2022-01-21  Richard Biener  <rguenther@suse.de>

PR tree-optimization/104152
* tree-vect-slp.cc (vect_build_slp_tree_2): Add missing
can_duplicate_and_interleave_p check.

* gcc.dg/vect/pr104152.c: New testcase.

warn-access: Fix up warning_at arguments

A warning regression fix I'm about to post warns (and breaks bootstrap due
to that) on the following spot.  Seems it is a copy and paste from
earlier code that mentions the %qD variable instead of talking about
unnamed temporary.

2022-01-21  Jakub Jelinek  <jakub@redhat.com>

* gimple-ssa-warn-access.cc (pass_waccess::warn_invalid_pointer):
Avoid passing var to warning_at when the format string doesn't
refer to it.

Reset relations when crossing backedges.

As discussed in PR103721, the problem here is that we are crossing a
backedge and causing us to use relations from a previous iteration of a
loop.

This handles the testcases in both PR103721 and PR104067 which are variants
of the same thing.

Tested on x86-64 Linux with the usual regstrap as well as verifying the
thread count before and after the patch.  The number of threads is
reduced by a miniscule amount.

gcc/ChangeLog:

PR tree-optimization/103721
* gimple-range-path.cc
(path_range_query::relations_may_be_invalidated): New.
(path_range_query::compute_ranges_in_block): Reset relations if
they may be invalidated.
(path_range_query::maybe_register_phi_relation): Exit if relations
may be invalidated on incoming edge.
(path_range_query::compute_phi_relations): Pass incoming PHI edge
to maybe_register_phi_relation.
* gimple-range-path.h (relations_may_be_invalidated): New.
(maybe_register_phi_relation): Pass edge instead of tree.
* tree-ssa-threadbackward.cc (back_threader::back_threader):
Mark DFS edges.
* value-relation.cc (path_oracle::path_oracle): Call
mark_dfs_back_edges.
(path_oracle::register_relation): Add SSA names to m_registered
bitmap.
(path_oracle::reset_path): Clear m_registered bitmap.
* value-relation.h (path_oracle::set_root_oracle): New.

gcc/testsuite/ChangeLog:

* gcc.dg/pr103721-2.c: New test.
* gcc.dg/pr103721.c: New test.

optabs: Don't create pseudos in prepare_cmp_insn when not allowed [PR102478]

cond traps can be created during ce3 after reload (and e.g. PR103028
recently fixed some ce3 cond trap related bug, so I think often that
works fine and we shouldn't disable cond traps after RA altogether),
but it calls prepare_cmp_insn.  This function can fail, so I don't
see why we couldn't make it work after RA (in most cases it already
just works).  The first hunk is just an optimization which doesn't
make sense after RA, so I've guarded it with can_create_pseudo_p.
The second hunk is just a theoretical case, I don't have a testcase for it.
prepare_cmp_insn has some other spots that can create pseudos, like when
both operands have VOIDmode, or when it is BLKmode comparison, or
not OPTAB_DIRECT, but I think none of that applies to ce3, we punt on
BLKmode earlier, use OPTAB_DIRECT and shouldn't be comparing two
VOIDmode CONST_INTs.

2022-01-21  Jakub Jelinek  <jakub@redhat.com>

PR rtl-optimization/102478
* optabs.cc (prepare_cmp_insn): If !can_create_pseudo_p (), don't
force_reg constants and for -fnon-call-exceptions fail if copy_to_reg
would be needed.

* gcc.dg/pr102478.c: New test.

[committed] Fix expected output for various MIPS multiplication tests

The recent multiply-highpart work twiddled code generation on the MIPS targets
and is causing mips.exp failures.

The resultant code is actually better and matches a comment in the test files
which indicates that it would be better to generate a mult-highpart.  So I'm
pretty confident in removing the undesired mflo & changing the name of the
target pattern we expect to see.

This fixes the mips64 and mips64el failures in my tester.  I suspect it'll
also fix the failures on mipsisa32, but that target is bootstrapped with qemu --
which takes forever so it only runs once a week ;-)

gcc/testsuite
* gcc.target/mips/fix-r4000-2.c: Update expected output.
* gcc.target/mips/fix-r4000-3.c: Update expected output.  Add
-fexpensive-optimizations for consistency in output.
* gcc.target/mips/fix-r4000-7.c: Update expected output.
* gcc.target/mips/fix-r4000-8.c: Update expected output.

runtime: build panic32.go on amd64p32

Fixes https://gcc.gnu.org/PR104149

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/380054

Daily bump.

analyzer: reject ((i + 1 > 0) && (i < 0)) for integers [PR94362]

PR analyzer/94362 reports a false positive from
-Wanalyzer-null-dereference seen when analyzing OpenSSL.

The root cause is that the analyzer's path feasibility checker
erroneously considers this to be feasible:
  (R + 1 > 0) && (R < 0)
for int R (the return value from sk_EVP_PKEY_ASN1_METHOD_num),
whereas it's not satisfiable for any int R.

This patch makes the constraint manager try harder to reject
such combinations of conditions, fixing the false positive;
perhaps in the longer term we ought to use an SMT solver.

gcc/analyzer/ChangeLog:
PR analyzer/94362
* constraint-manager.cc (bound::ensure_closed): Convert param to
enum bound_kind.
(range::constrained_to_single_element): Likewise.
(range::add_bound): New.
(constraint_manager::add_constraint): Handle SVAL + OFFSET
compared to a constant.
(constraint_manager::get_ec_bounds): Rewrite in terms of
range::add_bound.
(constraint_manager::eval_condition): Reject if range::add_bound
fails.
(selftest::test_constant_comparisons): Add test coverage for
various impossible combinations of integer comparisons.
* constraint-manager.h (enum bound_kind): New.
(struct bound): Likewise.
(bound::ensure_closed): Convert to param to enum bound_kind.
(struct range): Convert to...
(class range): ...this, making fields private.
(range::add_bound): New decls.
* region-model.cc (region_model::add_constraint): Fail if
constraint_manager::add_constraint fails.

gcc/testsuite/ChangeLog:
PR analyzer/94362
* gcc.dg/analyzer/pr94362-1.c: New test.
* gcc.dg/analyzer/pr94362-2.c: New test.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

analyzer: add regression test [PR103685]

PR analyzer/103685 reports a false positive from -Wanalyzer-null-dereference
seen at -O2 with GCC 11.  I can reproduce it with GCC 11, but not with
trunk; this patch adds a reduced test case that reproduces it with
GCC 11 as a regression test for GCC 12 onwards.

gcc/testsuite/ChangeLog:
PR analyzer/103685
* gcc.dg/analyzer/torture/pr103685.c: New test.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

Fortran: Fix scope for OMP AFFINITY clause iterator variables [PR103695]

gfc_finish_var_decl was confused by the undocumented overloading of
the proc_name field in struct gfc_namespace to contain iterator
variables for the OpenMP AFFINITY clause, causing it to insert the
decls in the wrong scope.  This patch adds a new distinct field to
hold these variables.

2022-01-20  Sandra Loosemore  <sandra@codesourcery.com>

PR fortran/103695
PR fortran/102621

gcc/fortran
* gfortran.h (struct gfc_namespace) Add omp_affinity_iterator
field.
* dump-parse-tree.cc (show_iterator): Use it.
* openmp.cc (gfc_match_iterator): Likewise.
(resolve_omp_clauses): Likewise.
* trans-decl.cc (gfc_finish_var_decl): Likewise.
* trans-openmp.cc (handle_iterator): Likewise.

gcc/testsuite/
* gfortran.dg/gomp/affinity-clause-3.f90: Adjust pattern.
* gfortran.dg/gomp/pr102621.f90: New.
* gfortran.dg/gomp/pr103695.f90: New.

c++: designator for base class member [PR101405]

A C++20 designator must name a direct non-static member of the class; in
this case it names a member of a base class, and we should give an error
instead of crashing.

PR c++/101405

gcc/cp/ChangeLog:

* decl.cc (reshape_init_class): Reject designator for a member of
another class.

gcc/testsuite/ChangeLog:

* g++.dg/cpp2a/desig20.C: New test.

c++: Add test for fixed PR [PR102338]

This was fixed by r12-6025 and is sufficiently different from
noexcept71.C that I think we should add it.

PR c++/102338

gcc/testsuite/ChangeLog:

* g++.dg/cpp1y/noexcept2.C: New test.

c++: add testcase for recently fixed PR [PR103631]

We accept this testcase after r12-6773.

PR c++/103631

gcc/testsuite/ChangeLog:

* g++.dg/cpp2a/nontype-class51.C: New test.

middle-end/100786 - constant folding from incompatible alias

The following avoids us ICEing doing constant folding from variables
with aliases of different types.  The issue appears both in
folding and CCP and FRE can do more fancy stuff to still constant
fold cases where the load is smaller than the initializer so
defer it to there.

2022-01-20  Richard Biener  <rguenther@suse.de>

PR middle-end/100786
* gimple-fold.cc (get_symbol_constant_value): Only return
values of compatible type to the symbol.

* gcc.dg/torture/pr100786.c: New testcase.

 aarch64: allow ld1/stq in test output [PR102517]

Following the changes to the inline memcpy operations get expanded, we
now generate ld1/st1 using a 128-bit vector register rather than ldp
with Q registers.  The behaviour is equivalent, so relax the tests to
permit either variant.

gcc/testsuite/ChangeLog:

PR target/102517
* gcc.target/aarch64/cpymem-q-reg_1.c: Allow ld1 and st1 for the
memcpy expansion.

Only add equivalencies that are still valid.

When equivalencies sets are merged, each member of the set should be queried
to ensure its still valid rather than a bulk union.

* value-relation.cc (relation_oracle::valid_equivs): Query and add
if valid members of a set.
(equiv_oracle::register_equiv): Call valid_equivs rather than
bitmap direct operations.
(path_oracle::register_equiv): Ditto.
* value-relation.h (relation_oracle::valid_equivs): New prototype.

c++: CTAD inside alias template [PR91911, PR103672]

In the first testcase below, when processing the alias template
ConstSpanType, transparency of alias template specializations means we
replace SpanType<T> with its instantiated definition.  But this
instantiation lowers the level of the CTAD placeholder for span{T()} from
2 to 1, and so the later instantiation of ConstSpanType<int> erroneously
substitutes this CTAD placeholder with the template argument at level 1
index 0, i.e. with int, before we get a chance to perform the CTAD.

Although we represent CTAD placeholders as template parameters, we never
actually want to replace them via tsubst.  So this patch adjusts tsubst
to handle CTAD placeholders by simply substituting the template and
returning a new CTAD placeholder.  Moreover, this means that the level
of a CTAD placeholder doesn't matter, so we may as well give them all
the same level.  This patch gives them the special level 0.

The change in tsubst_decl removes a likely dead !CHECKING_P safeguard
added in 2017, which would otherwise now get triggered for variables
with CTAD placeholder types (since their level is 0).

PR c++/91911
PR c++/103672

gcc/cp/ChangeLog:

* pt.cc (keep_template_parm): Punt on a level 0 template parm.
(tsubst_decl) <case VAR_DECL>: Remove !CHECKING_P safeguard.
(tsubst) <case TEMPLATE_TYPE_PARM>: Handle CTAD placeholders
specially.
(make_auto_1): Add defaulted 'level' parameter.
(make_template_placeholder): Pass 0 as 'level' to make_auto_1.

gcc/testsuite/ChangeLog:

* g++.dg/cpp1z/class-deduction101.C: New test.
* g++.dg/cpp1z/class-deduction101a.C: New test.
* g++.dg/cpp1z/class-deduction101b.C: New test.
* g++.dg/cpp1z/class-deduction102.C: New test.
* g++.dg/cpp1z/class-deduction102a.C: New test.
* g++.dg/cpp1z/class-deduction102b.C: New test.
* g++.dg/cpp1z/class-deduction103.C: New test.

c++: consistently diagnose bare CTAD placeholder in fn return type

Relax slightly the existing check for diagnosing a bare CTAD placeholder
as the return type of a function declarator to also handle the abstract
declarator case.

gcc/cp/ChangeLog:

* decl.cc (grokdeclarator): Diagnose a CTAD placeholder as
function return type even when !funcdecl_p.

gcc/testsuite/ChangeLog:

* g++.dg/other/pr88187.C: Adjust expected C++17 diagnostic.

target/100784 - avoid ICE with folding __builtin_ia32_shufpd

This avoids ICEing when there is no LHS on the call by following
what foldings of other builtins do in , namely not folding.

2022-01-20  Richard Biener  <rguenther@suse.de>

PR target/100784
* config/i386/i386.cc (ix86_gimple_fold_builtin): Check for
LHS before folding __builtin_ia32_shufpd and friends.

MAINTAINERS: adjust requested email address

ChangeLog:

* MAINTAINERS: Adjust email address based on IRC request.

libstdc++: Use Clang attribute instead of __constinit

Clang doesn't support the __constinit extension that we use pre-C++20,
but it does have its own equivalent attribute that can be used instead.

This makes it a little easier to use Clang to build libstdc++ (which
isn't supported. but is sometimes attempted for esoteric targets).

libstdc++-v3/ChangeLog:

* src/c++11/cxx11-ios_failure.cc (__constinit): Define as
equivalent attribute for Clang.
* src/c++11/future.cc (__constinit): Likewise.
* src/c++11/system_error.cc (__constinit): Likewise.
* src/c++17/memory_resource.cc (__constinit): Likewise.

libstdc++: Only add valid -L paths to testsuite linker options

The MacOS linker warns about -L arguments that don't exist, which causes
all tests to fail for the defauly configuration (because libbacktrace
isn't built).

libstdc++-v3/ChangeLog:

* scripts/testsuite_flags.in: Only add src/filesystem/.libs and
src/libbacktrace/.libs to LDFLAGS if those directories exist.

arm: Add test for AES erratum mitigation

Add a testcase for the erratum mitigation.  To improve coverage
use -dp on the assembler output and match the pattern names (and where
needed the alternative number).

gcc/testsuite/ChangeLog:

* gcc.target/arm/crypto-vaese-erratum1.c: New test.

arm: elide some cases where the AES erratum workaround is not required.

Some common cases where the AES erratum workaround are not required
are when there are 64- or 128-bit loads from memory, moving a 128-bit
value from core registers, and where a 128-bit constant is being
loaded from a literal pool.  The loads may also be misaligned or
generated via a neon intrinsic function.

gcc/ChangeLog:

* config/arm/crypto.md (aes_op_protect): Allow moves from core
registers and from memory.
(aes_op_protect_misalign_load): New pattern.
(aes_op_protect_neon_vld1v16qi): New pattern.

arm: suppress aes erratum when forwarding from aes

AES operations are commonly chained and since the result of one AES
operation is never a 32-bit value, they do not need an additional
mitigation instruction for the forwarded result.  We handle this
common case by adding additional patterns that allow for this.

gcc/ChangeLog:

* config/arm/crypto.md (crypto_<CRYPTO_AESMC:crypto_pattern>_protected):
New pattern.
(aarch32_crypto_aese_fused_protected): Likewise.
(aarch32_crypto_aesd_fused_protected): Likewise.

arm: add basic mitigation for Cortex-A AES errata

This patch adds the basic patterns for mitigation of the erratum, but no
attempt is made at this point to optimize the results for the cases where
the erratum mitigation is not needed.

The mitigation is done by guaranteeing that the input operands are fed
from a full-width operation by using an identity operation on the input
values.

gcc/ChangeLog:

* config/arm/crypto.md (crypto_<CRYPTO_AES:crypto_pattern>): Convert
to define_expand.  Add mitigation for the Cortex-A AES erratum
when enabled.
(*crypto_<CRYPTO_AES:crypto_pattern>_insn): New pattern, based
on original crypto_<CRYPTO_AES:crypto_pattern> insn.
(aes_op_protect): New pattern.
* config/arm/unspecs.md (unspec): Add UNSPEC_AES_PROTECT.

arm: Add option for mitigating against Cortex-A CPU erratum for AES

Add a new option -mfix-cortex-a-aes for enabling the Cortex-A AES
erratum work-around and enable it automatically for the affected
products (Cortex-A57 and Cortex-A72).

gcc/ChangeLog:

* config/arm/arm-cpus.in (quirk_aes_1742098): New quirk feature
(ALL_QUIRKS): Add it.
(cortex-a57, cortex-a72): Enable it.
(cortex-a57.cortex-a53, cortex-a72.cortex-a53): Likewise.
* config/arm/arm.opt (mfix-cortex-a57-aes-1742098): New command-line
option.
(mfix-cortex-a72-aes-1655431): New option alias.
* config/arm/arm.cc (arm_option_override): Handle default settings
for AES erratum switch.
* doc/invoke.texi (Arm Options): Document new options.

arm: Consistently use crypto_mode attribute in crypto patterns

A couple of patterns in the crypto support code were hard-coding the
mode rather than using the iterators.  While not incorrect, it was
slightly confusing, so adapt those patterns to the style of the rest
of the file.

Also fix some white space issues.

gcc/ChangeLog:

* config/arm/crypto.md (crypto_<CYRPTO_AES:crypto_pattern>): Use
<crypto_mode> rather than hard-coding the mode.
(crypto_<CRYPTO_AESMC:crypto_pattern>): Fix white space.
(crypto_<CRYPTO_AES:crypto_pattern>): Likewise.
(*aarch32_crypto_aese_fused): Likewise.
(*aarch32_crypto_aesd_fused): Likewise.
(crypto_<CRYPTO_BINARY:crypto_pattern>): Likewise.
(crypto_<CRYPTO_TERNARY:crypto_pattern>): Likewise.
(crypto_sha1h_lb): Likewise.
(crypto_vmullp64): Likewise.
(crypto_<CRYPTO_SELECTING:crypto_pattern>): Likewise.
(crypto_<CRYPTO_SELECTING:crypto_pattern>_lb): Likewise.

arm: Disambiguate multiple crypto patterns with the same name.

No functional change, but arm/crypto.md has multiple pattenrs all
called crypto_<crypto_pattern>, which makes references to them
ambiguous, so add the iterator base to the pattern name so that it is
distinct in the commit logs.

gcc/ChangeLog:

* config/arm/crypto.md (crypto_<CRYPTO_AESMC:crypto_pattern>): Add
iterator to pattern name to disambiguate.
(crypto_<CRYPTO_AES:crypto_pattern>): Likewise.
(crypto_<CRYPTO_BINARY:crypto_pattern>): Likewise.
(crypto_<CRYPTO_TERNARY:crypto_pattern>): Likewise.
(crypto_<CRYPTO_SELECTING:crypto_pattern>): Likewise.
(crypto_<CRYPTO_SELECTING:crypto_pattern>_lb): Likewise.

Fix Werror=format-diag with --disable-nls.

PR c++/104134

gcc/cp/ChangeLog:

* error.cc (dump_aggr_type): Partially disable the warning.

Fix -Werror=format-diag with RTL checking

PR bootstrap/104135

gcc/ChangeLog:

* emit-rtl.cc (make_insn_raw): Fix -Wformat-diag warnings.
* rtl.cc: Partially disable -Wformat-diag for RTL checking
error messages.

dwarf2out: Fix -gsplit-dwarf on riscv [PR103874]

riscv*-*-* are the only modern targets that !HAVE_AS_LEB128 (apparently
due to some aggressive linker optimizations).
As the following testcase shows, we mishandle in index_rnglists the
!HAVE_AS_LEB128 && !have_multiple_function_sections case.

output_rnglists does roughly:
  FOR_EACH_VEC_SAFE_ELT (ranges_table, i, r)
    {
...
      if (block_num > 0)
        {
...
          if (HAVE_AS_LEB128)
            {
              if (!have_multiple_function_sections)
                {
                  // code not using r->*_entry
                  continue;
                }
              // code that sometimes doesn't use r->*_entry,
              // sometimes r->begin_entry
            }
          else if (dwarf_split_debug_info)
            {
              // code that uses both r->begin_entry and r->end_entry
            }
          else
            {
              // code not using r->*_entry
            }
        }
      else if (block_num < 0)
        {
          if (!have_multiple_function_sections)
            gcc_unreachable ();
...
        }
    }
and index_rnglists is what sets up those r->{begin,end}_entry members.
The code did an early if (!have_multiple_function_sections) continue;
which is fine for the HAVE_AS_LEB128 case, because r->*_entry is not
used in that case, but not for !HAVE_AS_LEB128 that uses it anyway.

2022-01-20  Jakub Jelinek  <jakub@redhat.com>

PR debug/103874
* dwarf2out.cc (index_rnglists): For !HAVE_AS_LEB128 and
block_num > 0, index entry even if !have_multiple_function_sections.

* gcc.dg/debug/dwarf2/pr103874.c: New test.

testsuite: Add -Wno-psabi to pr47639.C testcase

This patch fixes
gcc/testsuite/g++.dg/opt/pr47639.C:6:24: warning: MMX vector return without MMX enabled changes the ABI [-Wpsabi]
gcc/testsuite/g++.dg/opt/pr47639.C:6:5: warning: MMX vector argument without MMX enabled changes the ABI [-Wpsabi]
FAILs on i686-linux.

2022-01-20  Jakub Jelinek  <jakub@redhat.com>

* g++.dg/opt/pr47639.C: Add -Wno-psabi to dg-options.

Enhance vec_pack_trunc for integral mode mask.

For testcase in PR, the patch supports QI:4 -> HI:16 pack with
multi steps(first pack QI:4 -> QI:8 through vec_pack_sbool_trunc_qi,
then pack QI:8 -> HI:16 through vec_pack_trunc_hi).
Similar for QI:2 -> HI:16 which is test4 in mask-pack-prefer-128.c.

gcc/ChangeLog:

PR target/103771
* tree-vect-stmts.cc (supportable_narrowing_operation): Enhance
integral mode mask pack by multi steps which takes
vec_pack_sbool_trunc_optab as start when elements number is
less than BITS_PER_UNITS.

gcc/testsuite/ChangeLog:

* gcc.target/i386/mask-pack-prefer128.c: New test.
* gcc.target/i386/mask-pack-prefer256.c: New test.
* gcc.target/i386/pr103771.c: New test.

tree-optimization/104114 - avoid diagnosing V1mode lowering

Currently we diagnose vector lowering of V1mode operations that
are not natively supported into V_C_E, scalar op plus CTOR with
-Wvector-operation-performance but that's hardly useful behavior
even though the way we lower things can be improved.

The following disables the diagnostics for the cases the vect.exp
testsuite runs into, on x86 that are vect-cond-11.c and
vect-singleton_1.c.

2022-01-19  Richard Biener  <rguenther@suse.de>

PR tree-optimization/104114
* tree-vect-generic.cc (expand_vector_piecewise): Do not diagnose
single element vector decomposition.

c++: template-id with current inst qualifier [PR102300]

The patch for PR41723 properly changed one place to look into the current
instantiation; now we need to fix this place as well.

PR c++/102300

gcc/cp/ChangeLog:

* parser.cc (cp_parser_template_name): Use dependent_scope_p.

gcc/testsuite/ChangeLog:

* g++.dg/parse/no-typename1.C: Remove expected error.
* g++.dg/template/nested7.C: New test.

testsuite: sldoi requires VSX.

gcc/testsuite/ChangeLog:

* gcc.target/powerpc/sldoi_to_mov.c: Require VSX.

Daily bump.

libstdc++: Remove -gdwarf-4 from flags for debug library

The default is -gdwarf-5 now, so this is hurting rather than improving
things.

libstdc++-v3/ChangeLog:

* configure.ac (GLIBCXX_ENABLE_DEBUG_FLAGS): Remove -gdwarf-4
from default flags.
* configure: Regenerate.

libstdc++: Include <stddef.h> for size_t and ptrdiff_t [PR104123]

libstdc++-v3/ChangeLog:

PR libstdc++/104123
* testsuite/29_atomics/headers/stdatomic.h/c_compat.cc: Include
<stddef.h>.

ifcvt: Run second pass if it is possible to omit a temporary.

If one of the to-be-converted SETs requires the original comparison
(i.e. in order to generate a min/max insn) but no other insn after it
does, we can omit creating temporaries, thus facilitating costing.

gcc/ChangeLog:

* ifcvt.cc (noce_convert_multiple_sets_1): New function.
(noce_convert_multiple_sets): Call function a second time if we can
improve the first try.

testsuite/s390: Add tests for noce_convert_multiple.

Add new s390-specific tests that check if we convert two SETs into two
loads on condition.  Remove the s390-specific target-check in
gcc.dg/ifcvt-4.c.

gcc/testsuite/ChangeLog:

* gcc.dg/ifcvt-4.c: Remove s390-specific check.
* gcc.target/s390/ifcvt-two-insns-bool.c: New test.
* gcc.target/s390/ifcvt-two-insns-int.c: New test.
* gcc.target/s390/ifcvt-two-insns-long.c: New test.

ifcvt: Try re-using CC for conditional moves.

Following up on the previous patch, this patch makes
noce_convert_multiple emit two cmov sequences:  The same one as before
and a second one that tries to re-use the existing CC.  Then their costs
are compared and the cheaper one is selected.

gcc/ChangeLog:

* ifcvt.cc (cond_exec_get_condition): New parameter to allow getting the
reversed comparison.
(try_emit_cmove_seq): New function to facilitate creating a cmov
sequence.
(noce_convert_multiple_sets): Create two sequences and use the less
expensive one.

ifcvt/optabs: Allow using a CC comparison for emit_conditional_move.

Currently we only ever call emit_conditional_move with the comparison
(as well as its comparands) we got from the jump.  Thus, backends are
going to emit a CC comparison for every conditional move that is being
generated instead of re-using the existing CC.
This, combined with emitting temporaries for each conditional move,
causes sky-high costs for conditional moves.

This patch allows to re-use a CC so the costing situation is improved a
bit.

gcc/ChangeLog:

* rtl.h (struct rtx_comparison): New struct that holds an rtx
comparison.
* config/rs6000/rs6000.cc (rs6000_emit_minmax): Use struct instead of
single parameters.
(rs6000_emit_swsqrt): Likewise.
* expmed.cc (expand_sdiv_pow2): Likewise.
(emit_store_flag): Likewise.
* expr.cc (expand_cond_expr_using_cmove): Likewise.
(expand_expr_real_2): Likewise.
* ifcvt.cc (noce_emit_cmove): Add compare and reversed compare
parameters.
* optabs.cc (emit_conditional_move_1): New function.
(expand_doubleword_shift_condmove): Use struct.
(emit_conditional_move): Use struct and allow to call directly
without going through preparation steps.
* optabs.h (emit_conditional_move): Use struct.

ifcvt: Improve costs handling for noce_convert_multiple.

When noce_convert_multiple is called the original costs are not yet
initialized.  Therefore, up to now, costs were only ever unfairly
compared against COSTS_N_INSNS (2).  This would lead to
default_noce_conversion_profitable_p () rejecting all but the most
contrived of sequences.

This patch temporarily initializes the original costs by counting
adding costs for all sets inside the then_bb.

gcc/ChangeLog:

* ifcvt.cc (bb_ok_for_noce_convert_multiple_sets): Estimate insns costs.
(noce_process_if_block): Use potential costs.

ifcvt: Allow constants for noce_convert_multiple.

This lifts the restriction of not allowing constants for
noce_convert_multiple.  The code later checks if a valid sequence
is produced anyway.

gcc/ChangeLog:

* ifcvt.cc (noce_convert_multiple_sets): Allow constants.
(bb_ok_for_noce_convert_multiple_sets): Likewise.

ifcvt: Check if cmovs are needed.

When if-converting multiple SETs and we encounter a swap-style idiom

  if (a > b)
    {
      tmp = c;   // [1]
      c = d;
      d = tmp;
    }

ifcvt should not generate a conditional move for the instruction at
[1].

In order to achieve that, this patch goes through all relevant SETs
and marks the relevant instructions.  This helps to evaluate costs.

On top, only generate temporaries if the current cmov is going to
overwrite one of the comparands of the initial compare.

gcc/ChangeLog:

* ifcvt.cc (need_cmov_or_rewire): New function.
(noce_convert_multiple_sets): Call it.

Update per-file selftest and finalization hooks for .c to .cc renaming

This is mostly a mechanical change, apart from:
- fix the name of attribute_c_tests to match its filename (attribs.cc)
- fix the name of opt_proposer_c to match its filename (opt-suggestions.cc)
- delete a bogus "modref_c_tests" decl from ipa-modref-tree.h that's been
  present since the initial commit of that file
  (d119f34c952f8718fdbabc63e2f369a16e92fa07)

gcc/ChangeLog:
* attribs.cc (attribute_c_tests): Rename to...
(attribs_cc_tests): ...this.
* bitmap.cc (bitmap_c_tests): Rename to...
(bitmap_cc_tests): ...this.
* cgraph.cc (cgraph_c_finalize): Rename to...
(cgraph_cc_finalize): ...this.
(cgraph_c_tests): Rename to...
(cgraph_cc_tests): ...this.
* cgraph.h (cgraph_c_finalize): Rename to...
(cgraph_cc_finalize): ...this.
(cgraphunit_c_finalize): Rename to...
(cgraphunit_cc_finalize): ...this.
* cgraphunit.cc (cgraphunit_c_finalize): Rename to...
(cgraphunit_cc_finalize): ...this.
* convert.cc (convert_c_tests): Rename to...
(convert_cc_tests): ...this.
* dbgcnt.cc (dbgcnt_c_tests): Rename to...
(dbgcnt_cc_tests): ...this.
* diagnostic-show-locus.cc (diagnostic_show_locus_c_tests): Rename to...
(diagnostic_show_locus_cc_tests): ...this.
* diagnostic.cc (diagnostic_c_tests): Rename to...
(diagnostic_cc_tests): ...this.
* dumpfile.cc (dumpfile_c_tests): Rename to...
(dumpfile_cc_tests): ...this.
* dwarf2out.cc (dwarf2out_c_finalize): Rename to...
(dwarf2out_cc_finalize): ...this.
* dwarf2out.h (dwarf2out_c_finalize): Rename to...
(dwarf2out_cc_finalize): ...this.
* edit-context.cc (edit_context_c_tests): Rename to...
(edit_context_cc_tests): ...this.
* et-forest.cc (et_forest_c_tests): Rename to...
(et_forest_cc_tests): ...this.
* fibonacci_heap.cc (fibonacci_heap_c_tests): Rename to...
(fibonacci_heap_cc_tests): ...this.
* fold-const.cc (fold_const_c_tests): Rename to...
(fold_const_cc_tests): ...this.
* function-tests.cc (function_tests_c_tests): Rename to...
(function_tests_cc_tests): ...this.
* gcse.cc (gcse_c_finalize): Rename to...
(gcse_cc_finalize): ...this.
* gcse.h (gcse_c_finalize): Rename to...
(gcse_cc_finalize): ...this.
* ggc-tests.cc (ggc_tests_c_tests): Rename to...
(ggc_tests_cc_tests): ...this.
* gimple-ssa-store-merging.cc (store_merging_c_tests): Rename to...
(store_merging_cc_tests): ...this.
* gimple.cc (gimple_c_tests): Rename to...
(gimple_cc_tests): ...this.
* hash-map-tests.cc (hash_map_tests_c_tests): Rename to...
(hash_map_tests_cc_tests): ...this.
* hash-set-tests.cc (hash_set_tests_c_tests): Rename to...
(hash_set_tests_cc_tests): ...this.
* input.cc (input_c_tests): Rename to...
(input_cc_tests): ...this.
* ipa-cp.cc (ipa_cp_c_finalize): Rename to...
(ipa_cp_cc_finalize): ...this.
* ipa-fnsummary.cc (ipa_fnsummary_c_finalize): Rename to...
(ipa_fnsummary_cc_finalize): ...this.
* ipa-fnsummary.h (ipa_fnsummary_c_finalize): Rename to...
(ipa_fnsummary_cc_finalize): ...this.
* ipa-modref-tree.cc (ipa_modref_tree_c_tests): Rename to...
(ipa_modref_tree_cc_tests): ...this.
* ipa-modref-tree.h (modref_c_tests): Delete bogus decl.
* ipa-modref.cc (ipa_modref_c_finalize): Rename to...
(ipa_modref_cc_finalize): ...this.
* ipa-modref.h (ipa_modref_c_finalize): Rename to...
(ipa_modref_cc_finalize): ...this.
* ipa-prop.h (ipa_cp_c_finalize): Rename to...
(ipa_cp_cc_finalize): ...this.
* ipa-reference.cc (ipa_reference_c_finalize): Rename to...
(ipa_reference_cc_finalize): ...this.
* ipa-reference.h (ipa_reference_c_finalize): Rename to...
(ipa_reference_cc_finalize): ...this.
* ira-costs.cc (ira_costs_c_finalize): Rename to...
(ira_costs_cc_finalize): ...this.
* ira.h (ira_costs_c_finalize): Rename to...
(ira_costs_cc_finalize): ...this.
* opt-suggestions.cc (opt_proposer_c_tests): Rename to...
(opt_suggestions_cc_tests): ...this.
* opts.cc (opts_c_tests): Rename to...
(opts_cc_tests): ...this.
* predict.cc (predict_c_tests): Rename to...
(predict_cc_tests): ...this.
* pretty-print.cc (pretty_print_c_tests): Rename to...
(pretty_print_cc_tests): ...this.
* read-rtl-function.cc (read_rtl_function_c_tests): Rename to...
(read_rtl_function_cc_tests): ...this.
* rtl-tests.cc (rtl_tests_c_tests): Rename to...
(rtl_tests_cc_tests): ...this.
* sbitmap.cc (sbitmap_c_tests): Rename to...
(sbitmap_cc_tests): ...this.
* selftest-run-tests.cc (selftest::run_tests): Update calls for
_c_ to _cc_ function renamings; fix names of attribs and
opt-suggestions tests.
* selftest.cc (selftest_c_tests): Rename to...
(selftest_cc_tests): ...this.
* selftest.h (attribute_c_tests): Rename to...
(attribs_cc_tests): ...this.
(bitmap_c_tests): Rename to...
(bitmap_cc_tests): ...this.
(cgraph_c_tests): Rename to...
(cgraph_cc_tests): ...this.
(convert_c_tests): Rename to...
(convert_cc_tests): ...this.
(diagnostic_c_tests): Rename to...
(diagnostic_cc_tests): ...this.
(diagnostic_show_locus_c_tests): Rename to...
(diagnostic_show_locus_cc_tests): ...this.
(dumpfile_c_tests): Rename to...
(dumpfile_cc_tests): ...this.
(edit_context_c_tests): Rename to...
(edit_context_cc_tests): ...this.
(et_forest_c_tests): Rename to...
(et_forest_cc_tests): ...this.
(fibonacci_heap_c_tests): Rename to...
(fibonacci_heap_cc_tests): ...this.
(fold_const_c_tests): Rename to...
(fold_const_cc_tests): ...this.
(function_tests_c_tests): Rename to...
(function_tests_cc_tests): ...this.
(ggc_tests_c_tests): Rename to...
(ggc_tests_cc_tests): ...this.
(gimple_c_tests): Rename to...
(gimple_cc_tests): ...this.
(hash_map_tests_c_tests): Rename to...
(hash_map_tests_cc_tests): ...this.
(hash_set_tests_c_tests): Rename to...
(hash_set_tests_cc_tests): ...this.
(input_c_tests): Rename to...
(input_cc_tests): ...this.
(opts_c_tests): Rename to...
(opts_cc_tests): ...this.
(predict_c_tests): Rename to...
(predict_cc_tests): ...this.
(pretty_print_c_tests): Rename to...
(pretty_print_cc_tests): ...this.
(read_rtl_function_c_tests): Rename to...
(read_rtl_function_cc_tests): ...this.
(rtl_tests_c_tests): Rename to...
(rtl_tests_cc_tests): ...this.
(sbitmap_c_tests): Rename to...
(sbitmap_cc_tests): ...this.
(selftest_c_tests): Rename to...
(selftest_cc_tests): ...this.
(simplify_rtx_c_tests): Rename to...
(simplify_rtx_cc_tests): ...this.
(spellcheck_c_tests): Rename to...
(spellcheck_cc_tests): ...this.
(spellcheck_tree_c_tests): Rename to...
(spellcheck_tree_cc_tests): ...this.
(sreal_c_tests): Rename to...
(sreal_cc_tests): ...this.
(store_merging_c_tests): Rename to...
(store_merging_cc_tests): ...this.
(tree_c_tests): Rename to...
(tree_cc_tests): ...this.
(tree_cfg_c_tests): Rename to...
(tree_cfg_cc_tests): ...this.
(typed_splay_tree_c_tests): Rename to...
(typed_splay_tree_cc_tests): ...this.
(vec_c_tests): Rename to...
(vec_cc_tests): ...this.
(vec_perm_indices_c_tests): Rename to...
(vec_perm_indices_cc_tests): ..this.
(opt_proposer_c_tests): Rename to...
(opt_suggestions_cc_tests): ...this.
(dbgcnt_c_tests): Rename to...
(dbgcnt_cc_tests): ...this.
(ipa_modref_tree_c_tests): Rename to...
(ipa_modref_tree_cc_tests): ...this.
* simplify-rtx.cc (simplify_rtx_c_tests): Rename to...
(simplify_rtx_cc_tests): ...this.
* spellcheck-tree.cc (spellcheck_tree_c_tests): Rename to...
(spellcheck_tree_cc_tests): ...this.
* spellcheck.cc (spellcheck_c_tests): Rename to...
(spellcheck_cc_tests): ...this.
* sreal.cc (sreal_c_tests): Rename to...
(sreal_cc_tests): ...this.
* toplev.cc (toplev::finalize): Update calls for _c_ to _cc_
function renamings.
* tree-cfg.cc (tree_cfg_c_tests): Rename to...
(tree_cfg_cc_tests): ...this.
* tree.cc (tree_c_tests): Rename to...
(tree_cc_tests): ...this.
* typed-splay-tree.cc (typed_splay_tree_c_tests): Rename to...
(typed_splay_tree_cc_tests): ...this.
* vec-perm-indices.cc (vec_perm_indices_c_tests): Rename to...
(vec_perm_indices_cc_tests): ...this.
* vec.cc (vec_c_tests): Rename to...
(vec_cc_tests): ...this.

gcc/c-family/ChangeLog:
* c-common.cc (c_common_c_tests): Rename to...
(c_common_cc_tests): ...this.
(c_family_tests): Update calls for .c to .cc renaming.
* c-common.h (c_format_c_tests): Rename to...
(c_format_cc_tests): ...this.
(c_indentation_c_tests): Rename to...
(c_indentation_cc_tests): ...this.
(c_pretty_print_c_tests): Rename to...
(c_pretty_print_cc_tests): ...this.
* c-format.cc (c_format_c_tests): Rename to...
(c_format_cc_tests): ...this.
* c-indentation.cc (c_indentation_c_tests): Rename to...
(c_indentation_cc_tests): ...this.
* c-pretty-print.cc (c_pretty_print_c_tests): Rename to...
(c_pretty_print_cc_tests): ...this.

gcc/cp/ChangeLog:
* cp-lang.cc (selftest::run_cp_tests): Update calls for .c to .cc
renaming.
* cp-tree.h (cp_pt_c_tests): Rename to...
(cp_pt_cc_tests): ...this.
(cp_tree_c_tests): Rename to...
(cp_tree_cc_tests): ...this.
* pt.cc (cp_pt_c_tests): Rename to...
(cp_pt_cc_tests): ...this.
* tree.cc (cp_tree_c_tests): Rename to...
(cp_tree_cc_tests): ...this.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

libstdc++: Fix libbacktrace build files

This makes it possible to combine --enable-libstdcxx-debug with
--enable-libstdcxx-backtrace, by adding a rule to src/Makefile to copy
the backtrace-supported.h header into the src/debug/libbacktrace
directory.

Add libbacktrace path to testsuite flags so the tests can link without
having the library installed.

Also fix some warnings when running automake for the libbacktrace
makefile.

Use a per-library CPPFLAGS variable to fix:

src/libbacktrace/Makefile.am:38: warning: AM_CPPFLAGS multiply defined in condition TRUE ...
fragment.am:43: ... 'AM_CPPFLAGS' previously defined here
src/libbacktrace/Makefile.am:32:   'fragment.am' included from here

Create symlinks to the libbacktrace sources to fix:

src/libbacktrace/Makefile.am:55: warning: source file '../../../libbacktrace/atomic.c' is in a subdirectory,
src/libbacktrace/Makefile.am:55: but option 'subdir-objects' is disabled

libstdc++-v3/ChangeLog:

* scripts/testsuite_flags.in: Add src/libbacktrace/.libs to
linker search paths.
* src/Makefile.am: Fix src/debug/libbacktrace build.
* src/Makefile.in: Regenerate.
* src/libbacktrace/Makefile.am: Use per-library CPPFLAGS
variable. Use symlinks for the source files.
* src/libbacktrace/Makefile.in: Regenerate.

vect: Fix epilogue mode skipping

gcc/ChangeLog:

PR tree-optimization/103997
* tree-vect-loop.cc (vect_analyze_loop): Fix mode skipping for epilogue
vectorization.

match.pd, optabs: Avoid vectorization of {FLOOR,CEIL,ROUND}_{DIV,MOD}_EXPR [PR102860]

power10 has modv4si3 expander and so vectorizes the following testcase
where Fortran modulo is FLOOR_MOD_EXPR.
optabs_for_tree_code indicates that the optab for all the *_MOD_EXPR
variants is umod_optab or smod_optab, but that isn't true, that optab
actually expands just TRUNC_MOD_EXPR.  For the other tree codes expmed.cc
has code how to adjust the TRUNC_MOD_EXPR into those by emitting some
extra comparisons and conditional updates.  Similarly for *_DIV_EXPR,
except in that case it actually needs both division and modulo.

While it would be possible to handle it in expmed.cc for vectors as well,
we'd need to be sure all the vector operations we need for that are
available, and furthermore we wouldn't account for that in the costing.

So, IMHO it is better to stop pretending those non-truncating (and
non-exact) div/mod operations have an optab.  For GCC 13, we should
IMHO pattern match these in tree-vect-patterns.cc and transform them
to truncating div/mod with follow-up adjustments and let the vectorizer
vectorize that.  As written in the PR, for signed operands:
r = x %[fl] y;
is
r = x % y; if (r && (x ^ y) < 0) r += y;
and
d = x /[fl] y;
is
r = x % y; d = x / y; if (r && (x ^ y) < 0) --d;
and
r = x %[cl] y;
is
r = x % y; if (r && (x ^ y) >= 0) r -= y;
and
d = /[cl] y;
is
r = x % y; d = x / y; if (r && (x ^ y) >= 0) ++d;
(too lazy to figure out rounding div/mod now).  I'll create a PR
for that.
The patch also extends a match.pd optimization that floor_mod on
unsigned operands is actually trunc_mod.

2022-01-19  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/102860
* match.pd (x %[fl] y -> x % y): New simplification for
unsigned integral types.
* optabs-tree.cc (optab_for_tree_code): Return unknown_optab
for {CEIL,FLOOR,ROUND}_{DIV,MOD}_EXPR with VECTOR_TYPE.

* gfortran.dg/pr102860.f90: New test.

testsuite: Test evrp-trans.c also with unsigned types [PR104115]

The testcase from the PR got fixed with r12-3119-g675a3e40567e1d
and looks quite similar to the evrp-trans.c test, except evrp-trans.c
is tested on signed integer types.
I think it would be useful to test it for unsigned comparisons too.

2022-01-19  Jakub Jelinek  <jakub@redhat.com>

PR c/104115
* gcc.dg/tree-ssa/evrp-trans2.c: New test.

tree-optimization/104112 - add check for vect epilogue reduc reuse

This adds a missing check for the availability of intermediate vector
types required to re-use the accumulator of a vectorized reduction
in the vectorized epilogue.  For SVE and VNx2DF vs V2DF with
-msve-vector-bits=512 for example V4DF is not available.

In addition to that we have to verify the reduction operation is
supported, otherwise we for example on i?86 get vector code that's
later decomposed again by vector lowering when trying to use
a V2HI epilogue for a V8HI reduction with a target without
TARGET_MMX_WITH_SSE.

It might be we want -Wvector-operation-performance for all vect.exp
tests but that seems to have existing regressions.

2022-01-19  Richard Biener  <rguenther@suse.de>

PR tree-optimization/104112
* tree-vect-loop.cc (vect_find_reusable_accumulator): Check
for required intermediate vector types.

* gcc.dg/vect/pr104112-1.c: New testcase.
* gcc.dg/vect/pr104112-2.c: New testcase.

check_GNU_style_lib: skip .py files (one should use flake8).

contrib/ChangeLog:

* check_GNU_style_lib.py: Skip Python files.

libgomp, OpenMP: Fix issue for omp_get_device_num on gcn targets.

Currently omp_get_device_num does not work on gcn targets with more than one
offload device. The reason is that GOMP_DEVICE_NUM_VAR is static in
icv-device.c and thus "__gomp_device_num" is not visible in the offload image.

This patch removes "static" such that "__gomp_device_num" is now part of the
offload image and can now be found in GOMP_OFFLOAD_load_image in the plugin.

This is not an issue for nvptx. There, "__gomp_device_num" is in the offload
image even with "static".

libgomp/ChangeLog:

* config/gcn/icv-device.c: Make GOMP_DEVICE_NUM_VAR public (remove
"static") to make the device num available in the offload image.

libstdc++: Fix for non-constexpr math_errhandling

Use SFINAE magic to support: "It is unspecified whether math_errhandling
is a macro or an identifier with external linkage." [C Standard]

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>
libstdc++-v3/ChangeLog:

* include/experimental/bits/simd.h (__floating_point_flags): Do
not rely on math_errhandling to expand to a constant expression.

[Ada] Set target_cpu to x32 for x86_64-linux-gnux32

Since the x86_64-linux-gnux32 compiler is actually an x32 compiler, set
target_cpu to x32 for x86_64-linux-gnux32.

PR ada/103538
* gcc-interface/Makefile.in (target_cpu): Set to x32 for
x86_64-linux-gnux32.

[Ada] Compile s-mmap and 128bit on x86_64-linux-gnux32

PR ada/103538
* Makefile.rtl (LIBGNAT_TARGET_PAIRS): Add
$(TRASYM_DWARF_UNIX_PAIRS),
s-tsmona.adb<libgnat/s-tsmona__linux.adb,
$(GNATRTL_128BIT_PAIRS).
(EXTRA_GNATRTL_NONTASKING_OBJS): Add $(TRASYM_DWARF_UNIX_OBJS)
and $(GNATRTL_128BIT_OBJS).

testsuite: Rename a few C++ to .C or .cc.

The tests are C++ code, so use a proper file extension.

gcc/testsuite/ChangeLog:

* g++.dg/ext/boolcomplex-1.c: Moved to...
* g++.dg/ext/boolcomplex-1.C: ...here.
* g++.dg/opt/pr47639.c: Moved to...
* g++.dg/opt/pr47639.C: ...here.
* g++.dg/pr83979.c: Moved to...
* g++.dg/pr83979.C: ...here.
* g++.dg/tm/asm-1.c: Moved to...
* g++.dg/tm/asm-1.C: ...here.
* g++.dg/vect/pr71483.c: Moved to...
* g++.dg/vect/pr71483.cc: ...here.

rs6000: Fix bootstrap

> On 18/01/2022 22:42, Segher Boessenkool wrote:
> > > +    default:
> > > +      break;
> > Please don't do that.  You can do
> >
> >    default:
> >      break;
> >      break;
> >      /* And just to make sure:  */
> >      break;
> >      break;
> >
> > and it will do exactly the same as not having a default at all.  Not
> > having such useless code is by far the most readable, so please don't
> > include a default case at all.
>
> I removed the default case. I hope this is what you wanted.

Unfortunately the removal of default: break; breaks bootstrap:
../../gcc/config/rs6000/rs6000.cc: In function ‘const char* rs6000_machine_from_flags()’:
../../gcc/config/rs6000/rs6000.cc:5940:10: error: enumeration value ‘PROCESSOR_PPC601’ not handled in switch [-Werror=switch]
 5940 |   switch (rs6000_cpu)
      |          ^
../../gcc/config/rs6000/rs6000.cc:5940:10: error: enumeration value ‘PROCESSOR_PPC603’ not handled in switch [-Werror=switch]
...
default: break; is needed to tell the -Wswitch warning that it is intentional
that not all enumerators are handled in the switch.

2022-01-19  Jakub Jelinek  <jakub@redhat.com>

* config/rs6000/rs6000.cc (rs6000_machine_from_flags): Add default:.

Enable -Werror=format-diag during bootstrap.

gcc/ChangeLog:

* configure.ac: Remove -Wno-error=format-diag.
* configure: Regenerate.

Update on riscv -Wformat-diag string.

gcc/ChangeLog:

* config/riscv/riscv.cc (riscv_handle_type_attribute):
Update one -Wformat-diag string in warning message.

gimple-ssa-warn-access: Fix up asan_test.C -Wdangling-pointer regression [PR104103]

As reported in the PR or as I've seen since the weekend, asan_test.C fails
because of many warnings like:
gcc/testsuite/g++.dg/asan/asan_test.cc:1157:10: error: using a dangling pointer to an unnamed temporary [-Werror=dangling-pointer=]
gcc/testsuite/g++.dg/asan/asan_test.cc:1157:10: error: using a dangling pointer to an unnamed temporary [-Werror=dangling-pointer=]
gcc/testsuite/g++.dg/asan/asan_test.cc:1162:27: error: using a dangling pointer to an unnamed temporary [-Werror=dangling-pointer=]
...
(lots of them).
There are no dangling pointers though, the warning pass sees:
  some_automatic_var ={v} {CLOBBER};
  .ASAN_MARK (POISON, &some_automatic_var, 8);
and warns on that (both on user vars and on e.g. TARGET_EXPR temporaries).
There is nothing wrong on that, .ASAN_MARK is compiler instrumentation,
which doesn't even touch the variable in any way nor make it escaped.
What it instead does is change bytes in the shadow memory corresponding
to the variable to reflect that the variable is out of scope and make
sure that access to it would be diagnosed at runtime.
So, for all purposes of the -Wdangling-pointer and -Wuse-after-free
warnings, we should ignore this internal call.

2022-01-19  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/104103
* gimple-ssa-warn-access.cc (pass_waccess::check_call): Don't check
.ASAN_MARK calls.

fold-const: Optimize &"foo"[0] == "foo" [PR89074]

This is a non-C++ related part from the PR89074 address_compare changes.
For "foo" == "foo" we already optimize this from the (cmp @0 @0)
simplification, because we use operand_equal_p in that case
and operand_equal_p also compares the STRING_CSTs bytes rather than
just addresses.

2022-01-19  Jakub Jelinek  <jakub@redhat.com>

PR c++/89074
* fold-const.cc (address_compare): Consider different STRING_CSTs
with the same lengths that memcmp the same as equal, not different.

* gcc.dg/tree-ssa/pr89074.c: New test.

i386: Fix *aes<aeswideklvariant>u8

grep '{[^|}]*}"' *.md

found another spot, though dunno if we have sufficient effective targets
etc. to add an -masm=intel test for it (and my installed binutils doesn't
support it anyway).
Binutils trunk testsuite shows the argument isn't omitted even in the Intel
syntax:
grep aesencwide *.s
keylocker.s:    aesencwide128kl 126(%edx)
keylocker.s:    aesencwide256kl 126(%edx)
keylocker.s:    aesencwide128kl [edx+126]
keylocker.s:    aesencwide256kl [edx+126]
property-10.s:   aesencwide128kl (%eax)
x86-64-keylocker.s:     aesencwide128kl 126(%rdx)
x86-64-keylocker.s:     aesencwide256kl 126(%rdx)
x86-64-keylocker.s:     aesencwide128kl [rdx+126]
x86-64-keylocker.s:     aesencwide256kl [rdx+126]
and doesn't use any WHATEVER PTR.

2022-01-19  Jakub Jelinek  <jakub@redhat.com>

* config/i386/sse.md (*aes<aeswideklvariant>u*): Use %0 instead of
{%0}.

Come up with git-fix-changelog.py script.

contrib/ChangeLog:

* git-backport.py: Use it.
* git-fix-changelog.py: New file.
* gcc-git-customization.sh: Add new alias git gcc-fix-changelog.

nvptx: update fix for -Wformat-diag

gcc/ChangeLog:

* config/nvptx/nvptx.cc (nvptx_goacc_validate_dims_1): Update
warning messages.

libgomp/ChangeLog:

* testsuite/libgomp.oacc-c++/privatized-ref-2.C: Update scanning
patterns.
* testsuite/libgomp.oacc-c++/privatized-ref-3.C: Likewise.
* testsuite/libgomp.oacc-c-c++-common/acc_prof-kernels-1.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/kernels-loop-2.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/parallel-dims.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/pr85486.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/pr95270-1.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/routine-nohost-2.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/struct-copyout-1.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/struct-copyout-2.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/vector-length-64-1.c:
Likewise.
* testsuite/libgomp.oacc-fortran/attach-descriptor-1.f90:
Likewise.
* testsuite/libgomp.oacc-fortran/derivedtypes-arrays-1.f90:
Likewise.
* testsuite/libgomp.oacc-fortran/kernels-loop-2.f95: Likewise.
* testsuite/libgomp.oacc-fortran/parallel-dims.f90: Likewise.
* testsuite/libgomp.oacc-fortran/privatized-ref-1.f95: Likewise.

Co-authored-by: Thomas Schwinge <thomas@codesourcery.com>

testsuite/102833 - fix gcc.dg/vect/bb-slp-17.c dump scanning

The desired transform requires V2SI vector add support, the closest
we have is vect64 so check that which at least fixes the i?86 fail.

2022-01-19  Richard Biener  <rguenther@suse.de>

PR testsuite/102833
* gcc.dg/vect/bb-slp-17.c: Require vect64.

powerc: Fix asm machine directive for some CPUs

For some CPUs, the assembler machine directive cannot be determined by ISA
flags.

gcc/

PR target/104090
* config/rs6000/rs6000.cc (rs6000_machine_from_flags): Use also
rs6000_cpu.

Fix tsvc test build on DragonFly.

Currently all tsvc tests fail to build on DragonFly BSD because they
assume <malloc.h> and memalign() are available.

gcc/testsuite/ChangeLog:

PR testsuite/104021
* gcc.dg/vect/tsvc/tsvc.h: Do not include malloc.h on dragonfly
and use posix_memalign ().

Signed-off-by: Rimvydas Jasinskas <rimvydas.jas@gmail.com>

Add XFAIL parts of pr16855.C for dragonfly.

On DragonFly BSD profiling ends before these DTORs are invoked on dso cleanup.
The -static compilation works as expected.

gcc/testsuite/ChangeLog:

PR testsuite/104022
* g++.dg/gcov/pr16855.C: xfail the count lines for DTORs on dragonfly.
* g++.dg/gcov/pr16855-priority.C: Ditto.  Adjust source layout so that
dejagnu xfail expressions work.

Signed-off-by: Rimvydas Jasinskas <rimvydas.jas@gmail.com>

testsuite: fix pytest detection unsupported message.

PR testsuite/104109

gcc/testsuite/ChangeLog:

* lib/gcov.exp: Fix pytest detection unsupported message.

testsuite: Adjust possibly fragile slp-perm-9.c [PR104015]

As Richard pointed out in PR104015, the test case slp-perm-9.c
can be fragile when vectorizer tries to use different
vectorisation strategies.

As suggested, this patch tries to make the check not sensitive
on the re-trying times by removing the times checking.  To still
retain the test coverage on unnecessary re-trying, for example
it exposes this PR104015 on Power9, I added two test cases to
powerpc testsuite.

gcc/testsuite/ChangeLog:

PR tree-optimization/104015
* gcc.dg/vect/slp-perm-9.c: Adjust.
* gcc.target/powerpc/pr104015-1.c: New test.
* gcc.target/powerpc/pr104015-2.c: New test.

c++: fix PR104025 change

Somehow I pushed my earlier patch without it actually fixing the test; we
need input_location to be for the last consumed token, not the next one.

gcc/cp/ChangeLog:

* parser.cc (saved_token_sentinel::rollback): Use
cp_lexer_previous_token.

i386: Fix GLC tuning with -masm=intel [PR104104]

> > On Sat, Jan 15, 2022 at 5:39 PM Hongyu Wang <wwwhhhyyy333@gmail.com> wrote:
> > > Thanks for the suggestion, here is the updated patch that survived
> > > bootstrap/regtest.

Unfortunately the patch results in assembler failures with -masm=intel.

> > > > +  if (TARGET_DEST_FALSE_DEPENDENCY
> > > > +      && get_attr_dest_false_dep (insn) ==
> > > > +        DEST_FALSE_DEP_TRUE)
> > > > +    output_asm_insn ("vxorps\t{%x0, %x0, %x0}", operands);

All the vxorps insns were emitted like the above, which means for -masm=sysv
it looks like
        vxorps  %xmm3, %xmm3, %xmm3
but for -masm=intel like:
        vxorps
We want obviously
        vxorps  xmm3, xmm3, xmm3
so the following patch just drops the errorneous {}s.

2022-01-19  Jakub Jelinek  <jakub@redhat.com>

PR target/104104
* config/i386/sse.md
(<avx512>_<complexopname>_<mode><maskc_name><round_name>,
avx512fp16_<complexopname>sh_v8hf<mask_scalarc_name><round_scalarcz_name>,
avx512dq_mul<mode>3<mask_name>, <avx2_avx512>_permvar<mode><mask_name>,
avx2_perm<mode>_1<mask_name>, avx512f_perm<mode>_1<mask_name>,
avx512dq_rangep<mode><mask_name><round_saeonly_name>,
avx512dq_ranges<mode><mask_scalar_name><round_saeonly_scalar_name>,
<avx512>_getmant<mode><mask_name><round_saeonly_name>,
avx512f_vgetmant<mode><mask_scalar_name><round_saeonly_scalar_name>):
Use vxorps\t%x0, %x0, %x0 instead of vxorps\t{%x0, %x0, %x0}.

* gcc.target/i386/pr104104.c: New test.

Add test for bogus warning [PR104076].

Related:
PR middle-end/104076 - bogus -Wdangling-pointer on a conditional

gcc/testsuite/ChangeLog:
PR middle-end/104076
* g++.dg/warn/Wdangling-pointer-3.C: New test.

Suppress valid warning [PR104103].

gcc/testsuite/ChangeLog:
PR middle-end/104103
* gcc.dg/torture/pr57147-2.c: Prune out expected warning.

Handle failure to determine pointer provenance conservatively [PR104069].

Partly resolves:
PR middle-end/104069 - -Werror=use-after-free false positive on elfutils-0.186

gcc/ChangeLog:
PR middle-end/104069
* gimple-ssa-warn-access.cc (pointers_related_p): Return false for
an unknown result as documented.

gcc/testsuite/ChangeLog:
PR middle-end/104069
* gcc.dg/Wuse-after-free.c: New test.

libstdc++: Update documentation for C++17 deprecations

libstdc++-v3/ChangeLog:

* doc/xml/manual/evolution.xml: Document deprecations.
* doc/xml/manual/status_cxx2017.xml: Update status.
* doc/html/*: Regenerate.

libstdc++: Fix deprecated attribute for std::get_temporary_buffer

This was deprecated in C++17, not C++14.

libstdc++-v3/ChangeLog:

* include/bits/stl_tempbuf.h (get_temporary_buffer): Change
_GLIBCXX14_DEPRECATED to _GLIBCXX17_DEPRECATED.

libstdc++: Remove unused std::pair helper function

This function is no longer used since r12-6691 and can be removed.

libstdc++-v3/ChangeLog:

* include/bits/stl_pair.h (_PCC::_DeprConsPair): Remove unused
function.

libstdc++: Fix std::atomic<std::shared_ptr<T>> for AIX [PR104101]

This fixes an on AIX.

The lock function currently just spins, which should be changed to use
back-off, and maybe then _M_val.wait(__current) when supported.

libstdc++-v3/ChangeLog:

PR libstdc++/104101
* include/bits/shared_ptr_atomic.h (_Sp_atomic::_Atomic_count::lock):
Only use __thread_relax if __cpp_lib_atomic_wait is defined.

Daily bump.

[COMMITTED] Improve coment for the newly added code in ipa-split.

It was pointed out to me by Jakub, that the comment in front of
the new code which handles warning/error attribute was not really
understandable. This fixes the comment to be understandable; I
don't know why I wrote the original comment that way even.

Committed as obvious after a quick build.

gcc/ChangeLog:

* ipa-split.cc (visit_bb): Fix comment before the
warning/error attribute checking code.

c++: Fix handling of temporaries with consteval ctors and non-trivial dtors [PR104055]

The following testcase is miscompiled.  We see the constructor is immediate,
in build_over_call we trigger:
          if (obj_arg && is_dummy_object (obj_arg))
            {
              call = build_cplus_new (DECL_CONTEXT (fndecl), call, complain);
              obj_arg = NULL_TREE;
            }
which makes call a TARGET_EXPR with the dtor in TARGET_EXPR_CLEANUP,
but then call cxx_constant_value on it.  In cxx_eval_outermost_constant_expr
it triggers the:
      else if (TREE_CODE (t) != CONSTRUCTOR)
        {
          r = get_target_expr_sfinae (r, tf_warning_or_error | tf_no_cleanup);
          TREE_CONSTANT (r) = true;
        }
which wraps the CONSTRUCTOR r into a new TARGET_EXPR, but one without
dtors (I think we need e.g. the TREE_CONSTANT for the callers),
and finally build_over_call uses that.

The following patch fixes that by using get_target_expr instead
of get_target_expr_sfinae + TREE_CONSTANT (r) = true if t is
a TARGET_EXPR with non-NULL TARGET_EXPR_CLEANUP.

2022-01-19  Jakub Jelinek  <jakub@redhat.com>

PR c++/104055
* constexpr.cc (cxx_eval_outermost_constant_expr): If t is a
TARGET_EXPR with TARGET_EXPR_CLEANUP, use get_target_expr rather
than get_target_expr_sfinae with tf_no_cleanup, and don't set
TREE_CONSTANT.

* g++.dg/cpp2a/consteval27.C: New test.

c++: Use -std=c++20 in testsuite default std list

C++20 has been official for a while now.

gcc/testsuite/ChangeLog:

* lib/g++-dg.exp: Change 2a to 20.

c++: input_location and lookahead [PR104025]

Debug information was getting confused because input_location was different
depending on whether we had looked ahead to see if the next tokens look like
a template argument list.

I tried resetting input_location in cp_lexer_rollback_tokens itself, but
that caused regressions, so let's just do it here for now.

PR c++/104025

gcc/cp/ChangeLog:

* parser.cc (saved_token_sentinel::rollback): Call
cp_lexer_set_source_position.
(~saved_token_sentinel): Call rollback.

gcc/testsuite/ChangeLog:

* g++.dg/warn/pr104025.C: New test.

Co-authored-by: Jakub Jelinek <jakub@redhat.com>

bpf: ensure correct string offsets in BTF.ext

BPF CO-RE relocations contain offsets to strings buffered in the BTF
string table. These BTF-specific strings are stored in memory in the
CTF auxilliary strtab, which at output time is concatenated onto the end
of the standard strtab.

Previously, these string offsets were computed at the time the
relocations were created. But strings could be added to the standard
strtab after this point, causing the offsets to no longer be correct.

Compute the offsets just before output instead, when they are sure to no
longer change.

gcc/ChangeLog:

* config/bpf/coreout.cc (bpf_core_reloc_add): Do not account
for base strtab offset yet as it may change.
(output_asm_btfext_core_reloc): Do so here instead.
(output_btfext_core_sections): Likewise.