platform/upstream/gcc.git
3 years agoconfigure: Allow host fragments to react to --enable-host-shared.
Iain Sandoe [Sat, 7 Aug 2021 13:40:33 +0000 (14:40 +0100)]
configure: Allow host fragments to react to --enable-host-shared.

This makes the host_shared value available to host makefile
fragments.

It uses this to adjust Darwin's mdynamic-no-pic in the case that
shared host resources are required.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
ChangeLog:

* Makefile.in: Regenerate.
* Makefile.tpl: Make the state of the configured host
shared flag available to makefile fragements.

config/ChangeLog:

* mh-darwin: Require a non-shared host configuration to
enable  mdynamic-no-pic where that is supported.

3 years agoObjective-C: fix crash with -fobjc-nilcheck
Iain Sandoe [Sat, 14 Aug 2021 11:27:55 +0000 (12:27 +0100)]
Objective-C: fix crash with -fobjc-nilcheck

When -fobjc-nilcheck is enabled, messages that result in a struct type should
yield a zero-initialized struct when sent to nil.  Currently, the frontend
crashes when it encounters this situation.  This patch fixes the crash by
generating the tree for the `{}` initializer.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
Co-authored-by: Matt Jacobson <mhjacobson@me.com>
PR objc/101666

gcc/objc/ChangeLog:

* objc-act.c (objc_build_constructor): Handle empty constructor
lists.
* objc-next-runtime-abi-02.c (build_v2_objc_method_fixup_call):
Handle nil receivers.
(build_v2_build_objc_method_call): Likewise.

gcc/testsuite/ChangeLog:

* obj-c++.dg/pr101666-0.mm: New test.
* obj-c++.dg/pr101666-1.mm: New test.
* obj-c++.dg/pr101666.inc: New.
* objc.dg/pr101666-0.m: New test.
* objc.dg/pr101666-1.m: New test.
* objc.dg/pr101666.inc: New.

3 years agolibiberty, Darwin : Fix simple-object LTO table for cross-endian case.
Iain Sandoe [Fri, 9 Apr 2021 12:40:11 +0000 (13:40 +0100)]
libiberty, Darwin : Fix simple-object LTO table for cross-endian case.

We encapsulate streamed IR in three special sections with a table
that describes their entries.  The table is expected to be written
with native endianness for the target, but for cross-endian cross-
compilation the swapping was omitted.  Fixed thus.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
libiberty/ChangeLog:

* simple-object-mach-o.c (simple_object_mach_o_write_segment):
Arrange to swap the LTO index tables where needed.

3 years agoDarwin: Handle the -rpath command line option.
Iain Sandoe [Sun, 28 Mar 2021 13:48:17 +0000 (14:48 +0100)]
Darwin: Handle the -rpath command line option.

This handles the command line '-rpath' option by passing it through
to the static linker.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/ChangeLog:

* config.gcc: Include rpath.opt for Darwin.
* config/darwin.h (DRIVER_SELF_SPECS): Handle -rpath.

3 years agoanalyzer: detect and analyze virtual function calls
Ankur Saini [Sun, 15 Aug 2021 13:49:07 +0000 (19:19 +0530)]
analyzer: detect and analyze virtual function calls

2021-08-15  Ankur Saini  <arsenic@sourceware.org>

gcc/analyzer/ChangeLog:
PR analyzer/97114
* region-model.cc (region_model::get_rvalue_1): Add case for
OBJ_TYPE_REF.

gcc/testsuite/ChangeLog:
PR analyzer/97114
* g++.dg/analyzer/vfunc-2.C: New test.
* g++.dg/analyzer/vfunc-3.C: New test.
* g++.dg/analyzer/vfunc-4.C: New test.
* g++.dg/analyzer/vfunc-5.C: New test.

3 years agoanalyzer: detect and analyze calls via function pointer
Ankur Saini [Thu, 29 Jul 2021 10:18:07 +0000 (15:48 +0530)]
analyzer: detect and analyze calls via function pointer

2021-07-29  Ankur Saini  <arsenic@sourceware.org>

gcc/analyzer/ChangeLog:
PR analyzer/100546
* analysis-plan.cc (analysis_plan::use_summary_p): Don't use call
summaries if there is no callgraph edge
* checker-path.cc (call_event::call_event): Handle calls events that
are not represented by a supergraph call edge
(return_event::return_event): Likewise.
(call_event::get_desc): Work with new call_event structure.
(return_event::get_desc): Likeise.
* checker-path.h (call_event::m_src_snode): New field.
(call_event::m_dest_snode): New field.
(return_event::m_src_snode): New field.
(return_event::m_dest_snode): New field.
* diagnostic-manager.cc
(diagnostic_manager::prune_for_sm_diagnostic)<case EK_CALL_EDGE>:
Refactor to work with edges without callgraph edge.
(diagnostic_manager::prune_for_sm_diagnostic)<case EK_RETURN_EDGE>:
Likewise.
* engine.cc (dynamic_call_info_t::update_model): New function.
(dynamic_call_info_t::add_events_to_path): New function.
(exploded_graph::create_dynamic_call): New function.
(exploded_graph::process_node): Work with dynamically discovered calls.
* exploded-graph.h (class dynamic_call_info_t): New class.
(exploded_graph::create_dynamic_call): New decl.
* program-point.cc (program_point::push_to_call_stack): New function.
(program_point::pop_from_call_stack): New function.
* program-point.h (program_point::push_to_call_stack): New decl.
(program_point::pop_from_call_stack): New decl.
* program-state.cc (program_state::push_call): New function.
(program_state::returning_call): New function.
* program-state.h (program_state::push_call): New decl.
(program_state::returning_call): New decl.
* region-model.cc (region_model::update_for_gcall) New function.
(region_model::update_for_return_gcall): New function.
(egion_model::update_for_call_superedge): Get the underlying gcall and
update for gcall.
(region_model::update_for_return_superedge): Likewise.
* region-model.h (region_model::update_for_gcall): New decl.
(region_model::update_for_return_gcall): New decl.
* state-purge.cc (state_purge_per_ssa_name::process_point): Update to
work with calls without underlying cgraph edge.
* supergraph.cc (supergraph::supergraph) Split snodes at every callsite.
* supergraph.h (supernode::get_returning_call) New accessor.

gcc/testsuite/ChangeLog:
PR analyzer/100546
* gcc.dg/analyzer/function-ptr-4.c: New test.
* gcc.dg/analyzer/pr100546.c: New test.

3 years agoMake 'gcc/hash-map-tests.c:test_map_of_type_with_ctor_and_dtor_expand' work on 32...
Thomas Schwinge [Wed, 18 Aug 2021 15:20:40 +0000 (17:20 +0200)]
Make 'gcc/hash-map-tests.c:test_map_of_type_with_ctor_and_dtor_expand' work on 32-bit architectures [PR101959]

Bug fix for recent commit e4f16e9f357a38ec702fb69a0ffab9d292a6af9b
"Add more self-tests for 'hash_map' with Value type with non-trivial
constructor/destructor".

gcc/
PR bootstrap/101959
* hash-map-tests.c (test_map_of_type_with_ctor_and_dtor_expand):
Use an 'int_hash'.

3 years agoaarch64: Fix float <-> int errors in vld4[q]_lane intrinsics
Jonathan Wright [Wed, 18 Aug 2021 08:10:22 +0000 (09:10 +0100)]
aarch64: Fix float <-> int errors in vld4[q]_lane intrinsics

A previous commit "aarch64: Remove macros for vld4[q]_lane Neon
intrinsics" introduced some float <-> int type conversion errors.
This patch fixes those errors.

gcc/ChangeLog:

2021-08-18  Jonathan Wright  <jonathan.wright@arm.com>

* config/aarch64/arm_neon.h (vld3_lane_f64): Use float RTL
pattern and type cast.
(vld4_lane_f32): Use float RTL pattern.
(vld4q_lane_f64): Use float type cast.

3 years agolibstdc++: Improve doxygen documentation for std::unique_ptr
Jonathan Wakely [Wed, 18 Aug 2021 14:03:19 +0000 (15:03 +0100)]
libstdc++: Improve doxygen documentation for std::unique_ptr

Add more detailed documentation for unique_ptr and related components.

The new alias templates for the _MakeUniq SFINAE helper make the
generated docs look better too.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/unique_ptr.h (default_delete): Add @since tag.
(unique_ptr, unique_ptr<T[]>): Likewise. Improve @brief.
(make_unique, make_unique_for_overwrite): Likewise. Add @tparam,
@param, and @returns.
(_MakeUniq): Move to __detail namespace. Add alias template
helpers.

3 years agolibstdc++: Improve doxygen comments in <bits/stl_function.h>
Jonathan Wakely [Wed, 18 Aug 2021 11:18:35 +0000 (12:18 +0100)]
libstdc++: Improve doxygen comments in <bits/stl_function.h>

Add notes about deprecation and modern replacements. Fix bogus
"memory_adaptors" group name. Use markdown for formatting.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/stl_function.h: Improve doxygen comments.

3 years agolibstdc++: Enable doxygen processing for C++20 components
Jonathan Wakely [Tue, 17 Aug 2021 19:27:02 +0000 (20:27 +0100)]
libstdc++: Enable doxygen processing for C++20 components

Improve grouping, add @since and @deprecated information.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* doc/doxygen/user.cfg.in (PREDEFINED): Enable doxygen
processing for C++20 components and components that depend on
compiler features.
* include/bits/stl_algo.h (random_shuffle): Use @deprecated.
* include/std/type_traits: Improve doxygen comments for C++20
traits.

3 years agolibstdc++: Simplify n-ary arithmetic promotion traits
Jonathan Wakely [Tue, 17 Aug 2021 19:26:52 +0000 (20:26 +0100)]
libstdc++: Simplify n-ary arithmetic promotion traits

The std::complex partial specializations have been unnecessary since
774c3d8647cc7012937cfc9d2d6dacc85b6cf8e9

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/ext/type_traits.h (__promote_2, __promote_3)
(__promote_4): Redfine as alias templates using __promoted_t.
* include/std/complex (__promote_2): Remove partial
specializations for std::complex.

3 years agolibstdc++: Minor optimization for min/max/minmax
Jonathan Wakely [Wed, 18 Aug 2021 11:24:12 +0000 (12:24 +0100)]
libstdc++: Minor optimization for min/max/minmax

The debug mode checks for a valid range are redundant when we have an
initializer_list argument, because we know it's a valid range already.
By making std::min(initialier_list<T>) call the internal __min_element
function directly we avoid a function call and skip those checks. The
same can be done for the overload taking a comparison function, and also
for the std::max and std::minmax overloads for initializer_list
arguments.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/stl_algo.h (min(initializer_list<T>))
(min(initializer_list<T>, Compare)): Call __min_element directly to
avoid redundant debug checks for valid ranges.
(max(initializer_list<T>), max(initializer_list<T>, Compare)):
Likewise, for __max_element.
(minmax(initializer_list<T>), minmax(initializer_list<T>, Compare)):
Likewise, for __minmax_element.

3 years agolibstdc++: Fix CTAD for debug sequence containers
Jonathan Wakely [Tue, 17 Aug 2021 17:19:27 +0000 (18:19 +0100)]
libstdc++: Fix CTAD for debug sequence containers

This fixes some 23_containers/*/cons/deduction.cc failures seen with
-std=c++17/-D_GLIBCXX_DEBUG, caused by non-immediate errors when
substituting template arguments into an incorrect specialization of the
std::__cxx1998 base class. This happens because the size_type member of
the debug container is _Base_type::size_type, so is non-deducible, and
the deduced types get substituted into _Base_type, triggering the
static_assert that checks the allocator's value_type matches the
container's.

The solution is to make the C(size_type, const T&, const Alloc&)
constructors of the debug sequence containers non-deducible. In order to
make CTAD work again deduction guides that use std::size_t for the first
argument are added.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/debug/deque (deque(size_type, const T&, const A&)):
Prevent class template argument deduction and replace with a
deduction guide.
* include/debug/forward_list (forward_list(size_type, const T&, const A&)):
Likewise.
* include/debug/list (list(size_type, const T&, const A&)):
Likewise.
* include/debug/vector (vector(size_type, const T&, const A&)):
Likewise.

3 years agolibstdc++: Fix vector<bool> printer tests to work in debug mode
Jonathan Wakely [Tue, 17 Aug 2021 14:06:22 +0000 (15:06 +0100)]
libstdc++: Fix vector<bool> printer tests to work in debug mode

This fixes a compilation error in debug mode, due to std::_Bit_reference
not being defined, because it's in namespace std::__cxx1998 instead. We
can refer to it as vector<bool>::reference instead, which always works.

That fixes some compilation errors in debug mode, but the tests fail at
run-time instead because the printers for vector<bool> helpers are only
registered for the std namespace, not std::__cxx1998. That is fixed by
using add_container to register the printers instead of add_version, as
the former registers them in the std and std::__cxx1998 namespaces.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* python/libstdcxx/v6/printers.py (StdBitReferencePrinter): Use
'std::vector<bool>::reference' as type name, not _Bit_reference.
(build_libstdcxx_dictionary): Register printers for vector<bool>
types in debug mode too.
* testsuite/libstdc++-prettyprinters/simple.cc: Adjust expected
output for invalid _Bit_reference. Use vector<bool>::reference
instead of _Bit_reference.
* testsuite/libstdc++-prettyprinters/simple11.cc: Likewise.

3 years agoFortran: Add OpenMP's nothing directive support
Tobias Burnus [Wed, 18 Aug 2021 13:21:18 +0000 (15:21 +0200)]
Fortran: Add OpenMP's nothing directive support

Fortran version of commit 5079b7781a2c506dcdfb241347d74c7891268225

gcc/fortran/ChangeLog:

* match.h (gfc_match_omp_nothing): New.
* openmp.c (gfc_match_omp_nothing): New.
* parse.c (decode_omp_directive): Match 'nothing' directive.

gcc/testsuite/ChangeLog:

* gfortran.dg/nothing-1.f90: New test.
* gfortran.dg/nothing-2.f90: New test.

3 years agoc++: aggregate CTAD and brace elision [PR101344]
Patrick Palka [Wed, 18 Aug 2021 12:37:45 +0000 (08:37 -0400)]
c++: aggregate CTAD and brace elision [PR101344]

Here the problem is ultimately that collect_ctor_idx_types always
recurses into an eligible sub-CONSTRUCTOR regardless of whether the
corresponding pair of braces was elided in the original initializer.
This causes us to reject some completely-braced forms of aggregate
CTAD as in the first testcase below, because collect_ctor_idx_types
effectively assumes that the original initializer is always minimally
braced (and so the aggregate deduction candidate is given a function
type that's incompatible with the original completely-braced initializer).

In order to fix this, collect_ctor_idx_types needs to somehow know the
shape of the original initializer when iterating over the reshaped
initializer.  To that end this patch makes reshape_init flag sub-ctors
that were built to undo brace elision in the original ctor, so that
collect_ctor_idx_types that determine whether to recurse into a sub-ctor
by simply inspecting this flag.

This happens to also fix PR101820, which is about aggregate CTAD using
designated initializers, for much the same reasons.

A curious case is the "intermediately-braced" initialization of 'e3'
(which we reject) in the first testcase below.  It seems to me we're
behaving as specified here (according to [over.match.class.deduct]/1)
because the initializer element x_1={1, 2, 3, 4} corresponds to the
subobject e_1=E::t, hence the type T_1 of the first function parameter
of the aggregate deduction candidate is T(&&)[2][2], but T can't be
deduced from x_1 using this parameter type (as opposed to say T(&&)[4]).

PR c++/101344
PR c++/101820

gcc/cp/ChangeLog:

* cp-tree.h (CONSTRUCTOR_BRACES_ELIDED_P): Define.
* decl.c (reshape_init_r): Set it.
* pt.c (collect_ctor_idx_types): Recurse into a sub-CONSTRUCTOR
iff CONSTRUCTOR_BRACES_ELIDED_P.

gcc/testsuite/ChangeLog:

* g++.dg/cpp2a/class-deduction-aggr11.C: New test.
* g++.dg/cpp2a/class-deduction-aggr12.C: New test.

3 years agoc++: ignore explicit dguides during NTTP CTAD [PR101883]
Patrick Palka [Wed, 18 Aug 2021 12:37:42 +0000 (08:37 -0400)]
c++: ignore explicit dguides during NTTP CTAD [PR101883]

Since (template) argument passing is a copy-initialization context,
we mustn't consider explicit deduction guides when deducing a CTAD
placeholder type of an NTTP.

PR c++/101883

gcc/cp/ChangeLog:

* pt.c (convert_template_argument): Pass LOOKUP_IMPLICIT to
do_auto_deduction.

gcc/testsuite/ChangeLog:

* g++.dg/cpp2a/nontype-class49.C: New test.

3 years agoFix thinko in latest change for GNAT encodings
Eric Botcazou [Wed, 18 Aug 2021 12:22:07 +0000 (14:22 +0200)]
Fix thinko in latest change for GNAT encodings

 gcc/ada/
* gcc-interface/decl.c (gnat_to_gnu_entity) <discrete_type>: Fix
thinko in latest change.

3 years agoAdd EAF_NOREAD check to tree-ssa-uninit
Jan Hubicka [Wed, 18 Aug 2021 11:25:28 +0000 (13:25 +0200)]
Add EAF_NOREAD check to tree-ssa-uninit

gcc/ChangeLog:

2021-08-18  Jan Hubicka  <hubicka@ucw.cz>

* tree-ssa-uninit.c (maybe_warn_pass_by_reference): Check also
EAF_NOREAD.

3 years agoAdd more self-tests for 'hash_map' with Value type with non-trivial constructor/destr...
Thomas Schwinge [Fri, 13 Aug 2021 15:53:12 +0000 (17:53 +0200)]
Add more self-tests for 'hash_map' with Value type with non-trivial constructor/destructor

... to document the current behavior.

gcc/
* hash-map-tests.c (test_map_of_type_with_ctor_and_dtor): Extend.
(test_map_of_type_with_ctor_and_dtor_expand): Add function.
(hash_map_tests_c_tests): Call it.

3 years agoTurn 'bool force_collect' parameter to 'ggc_collect' into an 'enum ggc_collect mode'
Thomas Schwinge [Tue, 17 Aug 2021 19:15:46 +0000 (21:15 +0200)]
Turn 'bool force_collect' parameter to 'ggc_collect' into an 'enum ggc_collect mode'

... to make the meaning more explicit to the reader of the code.

Follow-up to recent commit 0edf2e81bb02cba43b649b3f6e7258b68a779ac0
"Turn global 'ggc_force_collect' variable into 'force_collect' parameter to
'ggc_collect'".

gcc/
* ggc.h (enum ggc_collect): New.
(ggc_collect): Use it.
* ggc-page.c: Adjust.
* ggc-common.c: Likewise.
* ggc-tests.c: Likewise.
* read-rtl-function.c: Likewise.
* selftest-run-tests.c: Likewise.
* doc/gty.texi (Invoking the garbage collector): Likewise.

Suggested-by: David Malcolm <dmalcolm@redhat.com>
3 years agoFortran/OpenMP: Add memory routines existing for C/C++
Tobias Burnus [Wed, 18 Aug 2021 09:14:05 +0000 (11:14 +0200)]
Fortran/OpenMP: Add memory routines existing for C/C++

This patch adds the Fortran interface for omp_alloc/omp_free
and the omp_target_* memory routines, which were added in
OpenMP 5.0 for C/C++ but only OpenMP 5.1 added them for Fortran.

Those functions use BIND(C), i.e. on the libgomp side, the same
interface as for C/C++ is used.

Note: By using BIND(C) in omp_lib.h, files including this file
no longer compiler with -std=f95 but require at least -std=f2003.

libgomp/ChangeLog:

* omp_lib.f90.in (omp_alloc, omp_free, omp_target_alloc,
omp_target_free. omp_target_is_present, omp_target_memcpy,
omp_target_memcpy_rect, omp_target_associate_ptr,
omp_target_disassociate_ptr): Add interface.
* omp_lib.h.in (omp_alloc, omp_free, omp_target_alloc,
omp_target_free. omp_target_is_present, omp_target_memcpy,
omp_target_memcpy_rect, omp_target_associate_ptr,
omp_target_disassociate_ptr): Add interface.
* testsuite/libgomp.fortran/alloc-1.F90: Remove local
interface block for omp_alloc + omp_free.
* testsuite/libgomp.fortran/alloc-4.f90: Likewise.
* testsuite/libgomp.fortran/refcount-1.f90: New test.
* testsuite/libgomp.fortran/target-12.f90: New test.

3 years agoopenmp: Add nothing directive support
Jakub Jelinek [Wed, 18 Aug 2021 09:10:43 +0000 (11:10 +0200)]
openmp: Add nothing directive support

As has been clarified, it is intentional that nothing directive is accepted
in substatements of selection and looping statements and after labels and
is handled as if the directive just isn't there, so that
void
foo (int x)
{
  if (x)
    #pragma omp metadirective when (...:nothing) when (...:parallel)
    bar ();
}
behaves consistently; declarative and stand-alone directives aren't allowed
at that point, but constructs are parsed with the following statement as
the construct body and nothing or missing default on metadirective therefore
should handle the following statement as part of the if substatement instead
of having nothing as the substatement and bar done unconditionally after the
if.

2021-08-18  Jakub Jelinek  <jakub@redhat.com>

gcc/c-family/
* c-pragma.h (enum pragma_kind): Add PRAGMA_OMP_NOTHING.
* c-pragma.c (omp_pragmas): Add nothing directive.
* c-omp.c (omp_directives): Uncomment nothing directive entry.
gcc/c/
* c-parser.c (c_parser_omp_nothing): New function.
(c_parser_pragma): Handle PRAGMA_OMP_NOTHING.
gcc/cp/
* parser.c (cp_parser_omp_nothing): New function.
(cp_parser_pragma): Handle PRAGMA_OMP_NOTHING.
gcc/testsuite/
* c-c++-common/gomp/nothing-1.c: New test.
* g++.dg/gomp/attrs-1.C (bar): Add nothing directive test.
* g++.dg/gomp/attrs-2.C (bar): Likewise.
* g++.dg/gomp/attrs-9.C: Likewise.
libgomp/
* testsuite/libgomp.c-c++-common/nothing-1.c: New test.

3 years agocommit-mklog: Add --co argument.
Martin Liska [Tue, 17 Aug 2021 12:57:40 +0000 (14:57 +0200)]
commit-mklog: Add --co argument.

The argument can be used for addition of Co-Authored-By lines
with --trailer='Co-Authored-By=Mona Lisa Octocat <mona@github.com>'.

contrib/ChangeLog:

* gcc-git-customization.sh: Wrap $@ in quotes.
* git-commit-mklog.py: Add new argument --co.
* mklog.py: Skip the Co-Authored-By lines.

3 years agoopenmp: Actually ignore pragma_stmt pragmas for which c_parser_pragma returns false
Jakub Jelinek [Wed, 18 Aug 2021 08:20:50 +0000 (10:20 +0200)]
openmp: Actually ignore pragma_stmt pragmas for which c_parser_pragma returns false

Unlike the C++ FE, the C FE ignored pragmas (as if they weren't there) in
pragma_stmt contexts if c*_parser_pragma returns false only when after labels,
not inside of substatements of selection or loop statements.
After making just that change, several gomp/goacc testcases started failing,
because extra diagnostics has been emitted (in C, in C++ it was emitted
already before).  Say
void
foo (int x)
{
  if (x)
    #pragma omp barrier
}
used to in C emit just an error that the pragma is not allowed in such
contexts, but in C++ emitted both that and a parsing error that
  if (x)
}
is invalid.  So, the rest of this patch is mostly about returning true
after we report that that certain pragma is not allowed in pragma_stmt
contexts, because for error-recovery it seems better to treat the
pragma in that case as something that is the substatement of such if etc.
c*_parser_pragma return value is only ever used for pragma_stmt context,
in which false means act as if the pragma isn't there (e.g. has been handled
already by preprocessor etc.), and true which means it was there.

2021-08-18  Jakub Jelinek  <jakub@redhat.com>

gcc/c/
* c-parser.c (c_parser_statement_after_labels): Add restart label
near the start of the function.  If c_parser_pragma returns false,
goto restart.
(c_parser_pragma): For PRAGMA_OMP_CANCELLATION_POINT return what
c_parser_omp_cancellation_point returned.  For PRAGMA_OMP_DECLARE
return what c_parser_omp_declare returned.  Return true instead of
false after emitting errors that the directive is not allowed in
pragma_stmt context.
(c_parser_omp_ordered): Return true instead of
false after emitting errors that the directive is not allowed in
pragma_stmt context.
(c_parser_omp_target_update): Likewise.
(c_parser_omp_target_enter_data, c_parser_omp_target_exit_data):
Change return type from tree to bool, return false if the
directive should be ignored in pragma_stmt contexts.
(c_parser_omp_target): Adjust callers of c_parser_omp_target_*_data,
return their result directly.
(c_parser_omp_cancellation_point): Change return type from void to
bool, return false if the directive should be ignored in pragma_stmt
contexts.
(c_parser_omp_declare): Likewise.
gcc/cp/
* parser.c (cp_parser_omp_ordered): Return true instead of
false after emitting errors that the directive is not allowed in
pragma_stmt context.
(cp_parser_omp_target_update): Likewise.
(cp_parser_omp_cancellation_point): Change return type from void to
bool, return false if the directive should be ignored in pragma_stmt
contexts.
(cp_parser_omp_target_enter_data, cp_parser_omp_target_exit_data):
Change return type from tree to bool, return false if the
directive should be ignored in pragma_stmt contexts.
(cp_parser_omp_target): Adjust callers of cp_parser_omp_target_*_data,
return their result directly.
(cp_parser_pragma): For PRAGMA_OMP_CANCELLATION_POINT return what
cp_parser_omp_cancellation_point returned.  Return true instead of
false after emitting errors that the directive is not allowed in
pragma_stmt context.
gcc/testsuite/
* c-c++-common/gomp/pr63326.c: Don't expect extra "before" errors
in C++.
* g++.dg/gomp/attrs-7.C: Don't expect one extra error.
* g++.dg/gomp/barrier-2.C: Likewise.
* gcc.dg/gomp/declare-simd-5.c: Likewise.
* gcc.dg/gomp/barrier-2.c: Likewise.
* gcc.dg/gomp/declare-variant-2.c: Likewise.

3 years agoAdd x86 tune to enable v2df vector reduction by paddpd.
liuhongt [Tue, 17 Aug 2021 05:11:26 +0000 (13:11 +0800)]
Add x86 tune to enable v2df vector reduction by paddpd.

The tune is disabled by default.

gcc/ChangeLog:

PR target/97147
* config/i386/i386.h (TARGET_V2DF_REDUCTION_PREFER_HADDPD):
New macro.
* config/i386/sse.md (*sse3_haddv2df3_low): Add
TARGET_V2DF_REDUCTION_PREFER_HADDPD.
(*sse3_hsubv2df3_low): Ditto.
* config/i386/x86-tune.def
(X86_TUNE_V2DF_REDUCTION_PREFER_HADDPD): New tune.

gcc/testsuite/ChangeLog:

PR target/97147
* gcc.target/i386/pr54400.c: Adjust testcase.
* gcc.target/i386/pr94147.c: New test.

3 years agoDaily bump.
GCC Administrator [Wed, 18 Aug 2021 00:16:48 +0000 (00:16 +0000)]
Daily bump.

3 years agoAdd GORI tracing faciltiies.
Andrew MacLeod [Thu, 12 Aug 2021 16:29:48 +0000 (12:29 -0400)]
Add GORI tracing faciltiies.

Debugging range-ops and gori unwinding needed some help.

* gimple-range-gori.cc (gori_compute::gori_compute): Enable tracing.
(gori_compute::compute_operand_range): Add tracing.
(gori_compute::logical_combine): Ditto.
(gori_compute::compute_logical_operands): Ditto.
(gori_compute::compute_operand1_range): Ditto.
(gori_compute::compute_operand2_range): Ditto.
(gori_compute::outgoing_edge_range_p): Ditto.
* gimple-range-gori.h (class gori_compute): Add range_tracer.

3 years agoChange evrp-mode options.
Andrew MacLeod [Thu, 12 Aug 2021 18:02:20 +0000 (14:02 -0400)]
Change evrp-mode options.

Remove tracing in hybrid mode. Add trace/gori/cache tracing options.
tracing options are now  'trace', 'gori', 'cache', or all combined in 'debug'

* flag-types.h (enum evrp_mode): Adjust evrp-mode values.
* gimple-range-cache.cc (DEBUG_RANGE_CACHE): Relocate from.
* gimple-range-trace.h (DEBUG_RANGE_CACHE): Here.
* params.opt (--param=evrp-mode): Adjust options.

3 years agoAbstract tracing routines into a class.
Andrew MacLeod [Fri, 30 Jul 2021 19:15:29 +0000 (15:15 -0400)]
Abstract tracing routines into a class.

Generalize range tracing into a class and integrae it with gimple_ranger.
Remove the old derived trace_ranger class.

* Makefile.in (OBJS): Add gimple-range-trace.o.
* gimple-range-cache.h (enable_new_values): Remove unused prototype.
* gimple-range-fold.cc: Adjust headers.
* gimple-range-trace.cc: New.
* gimple-range-trace.h: New.
* gimple-range.cc (gimple_ranger::gimple_ranger): Enable tracer.
(gimple_ranger::range_of_expr): Add tracing.
(gimple_ranger::range_on_entry): Ditto.
(gimple_ranger::range_on_exit): Ditto.
(gimple_ranger::range_on_edge): Ditto.
(gimple_ranger::fold_range_internal): Ditto.
(gimple_ranger::dump_bb): Do not calculate edge range twice.
(trace_ranger::*): Remove.
(enable_ranger): Never create a trace_ranger.
(debug_seed_ranger): Move to gimple-range-trace.cc.
(dump_ranger): Ditto.
(debug_ranger): Ditto.
* gimple-range.h: Include gimple-range-trace.h.
(range_on_entry, range_on_exit): No longer virtual.
(class trace_ranger): Remove.
(DEBUG_RANGE_CACHE): Move to gimple-range-trace.h.

3 years agoMove more warning code to gimple-ssa-warn-access etc.
Martin Sebor [Tue, 17 Aug 2021 20:49:05 +0000 (14:49 -0600)]
Move more warning code to gimple-ssa-warn-access etc.

Also resolves:
PR middle-end/101854 - Invalid warning -Wstringop-overflow wrong argument

gcc/ChangeLog:

PR middle-end/101854
* builtins.c (expand_builtin_alloca): Move warning code to check_alloca
in gimple-ssa-warn-access.cc.
* calls.c (alloc_max_size): Move code to check_alloca.
(get_size_range): Move to pointer-query.cc.
(maybe_warn_alloc_args_overflow): Move to gimple-ssa-warn-access.cc.
(get_attr_nonstring_decl): Move to tree.c.
(fntype_argno_type): Move to gimple-ssa-warn-access.cc.
(append_attrname): Same.
(maybe_warn_rdwr_sizes): Same.
(initialize_argument_information): Move code to
gimple-ssa-warn-access.cc.
* calls.h (maybe_warn_alloc_args_overflow): Move to
gimple-ssa-warn-access.h.
(get_attr_nonstring_decl): Move to tree.h.
(maybe_warn_nonstring_arg):  Move to gimple-ssa-warn-access.h.
(enum size_range_flags): Move to pointer-query.h.
(get_size_range): Same.
* gimple-ssa-warn-access.cc (has_location): Remove unused overload
to avoid Clang -Wunused-function.
(get_size_range): Declare static.
(maybe_emit_free_warning): Rename...
(maybe_check_dealloc_call): ...to this for consistency.
(class pass_waccess): Add members.
(pass_waccess::~pass_waccess): Defined.
(alloc_max_size): Move here from calls.c.
(maybe_warn_alloc_args_overflow): Same.
(check_alloca): New function.
(check_alloc_size_call): New function.
(check_strncat): Handle another warning flag.
(pass_waccess::check_builtin): Handle alloca.
(fntype_argno_type): Move here from calls.c.
(append_attrname): Same.
(maybe_warn_rdwr_sizes): Same.
(pass_waccess::check_call): Define.
(check_nonstring_args): New function.
(pass_waccess::check): Call new member functions.
(pass_waccess::execute): Enable ranger.
* gimple-ssa-warn-access.h (get_size_range): Move here from calls.h.
(maybe_warn_nonstring_arg): Same.
* gimple-ssa-warn-restrict.c: Remove #include.
* pointer-query.cc (get_size_range): Move here from calls.c.
* pointer-query.h (enum size_range_flags): Same.
(get_size_range): Same.
* tree.c (get_attr_nonstring_decl): Move here from calls.c.
* tree.h (get_attr_nonstring_decl): Move here from calls.h.

gcc/testsuite/ChangeLog:

* gcc.dg/attr-alloc_size-5.c: Adjust optimization to -O1.
* gcc.dg/attr-alloc_size-7.c: Use #pragmas to adjust optimization.
* gcc.dg/attr-alloc_size-8.c: Adjust optimization to -O1.

PR middle-end/101854
* gcc.dg/Wstringop-overflow-72.c: New test.

3 years agoc++: Implement P0466R5 __cpp_lib_is_layout_compatible compiler helpers [PR101539]
Jakub Jelinek [Tue, 17 Aug 2021 19:06:39 +0000 (21:06 +0200)]
c++: Implement P0466R5 __cpp_lib_is_layout_compatible compiler helpers [PR101539]

The following patch implements __is_layout_compatible trait and
__builtin_is_corresponding_member helper function for the
std::is_corresponding_member template function.

As the current definition of layout compatible type has various problems,
which result e.g. in corresponding members in layout compatible types having
different member offsets, the patch anticipates some changes to the C++
standard:
1) class or enumeral types aren't layout compatible if they have different
   alignment or size
2) if two members have different offsets, they can't be corresponding members
   ([[no_unique_address]] with empty types can change that, or alignas
   on the member decls)
3) in unions, bitfields can't correspond to non-unions, or bitfields can't
   correspond to bitfields with different widths, or members with
   [[no_unique_address]] can't correspond to members without that attribute

__builtin_is_corresponding_member for anonymous structs (GCC extension) will
recurse into the anonymous structs.  For anonymous unions it will emit
a sorry if it can't prove such member types can't appear in the
anonymous unions or anonymous aggregates in that union, because
corresponding member is defined only using common initial sequence which is
only defined for std-layout non-union class types and so I have no idea what
to do otherwise in that case.

2021-08-17  Jakub Jelinek  <jakub@redhat.com>

PR c++/101539
gcc/c-family/
* c-common.h (enum rid): Add RID_IS_LAYOUT_COMPATIBLE.
* c-common.c (c_common_reswords): Add __is_layout_compatible.
gcc/cp/
* cp-tree.h (enum cp_trait_kind): Add CPTK_IS_LAYOUT_COMPATIBLE.
(enum cp_built_in_function): Add CP_BUILT_IN_IS_CORRESPONDING_MEMBER.
(fold_builtin_is_corresponding_member, next_common_initial_seqence,
layout_compatible_type_p): Declare.
* parser.c (cp_parser_primary_expression): Handle
RID_IS_LAYOUT_COMPATIBLE.
(cp_parser_trait_expr): Likewise.
* cp-objcp-common.c (names_builtin_p): Likewise.
* constraint.cc (diagnose_trait_expr): Handle
CPTK_IS_LAYOUT_COMPATIBLE.
* decl.c (cxx_init_decl_processing): Register
__builtin_is_corresponding_member builtin.
* constexpr.c (cxx_eval_builtin_function_call): Handle
CP_BUILT_IN_IS_CORRESPONDING_MEMBER builtin.
* semantics.c (is_corresponding_member_union,
is_corresponding_member_aggr, fold_builtin_is_corresponding_member):
New functions.
(trait_expr_value): Handle CPTK_IS_LAYOUT_COMPATIBLE.
(finish_trait_expr): Likewise.
* typeck.c (next_common_initial_seqence, layout_compatible_type_p):
New functions.
* cp-gimplify.c (cp_gimplify_expr): Fold
CP_BUILT_IN_IS_CORRESPONDING_MEMBER.
(cp_fold): Likewise.
* tree.c (builtin_valid_in_constant_expr_p): Handle
CP_BUILT_IN_IS_CORRESPONDING_MEMBER.
* cxx-pretty-print.c (pp_cxx_trait_expression): Handle
CPTK_IS_LAYOUT_COMPATIBLE.
* class.c (remove_zero_width_bit_fields): Remove.
(layout_class_type): Don't call it.
gcc/testsuite/
* g++.dg/cpp2a/is-corresponding-member1.C: New test.
* g++.dg/cpp2a/is-corresponding-member2.C: New test.
* g++.dg/cpp2a/is-corresponding-member3.C: New test.
* g++.dg/cpp2a/is-corresponding-member4.C: New test.
* g++.dg/cpp2a/is-corresponding-member5.C: New test.
* g++.dg/cpp2a/is-corresponding-member6.C: New test.
* g++.dg/cpp2a/is-corresponding-member7.C: New test.
* g++.dg/cpp2a/is-corresponding-member8.C: New test.
* g++.dg/cpp2a/is-layout-compatible1.C: New test.
* g++.dg/cpp2a/is-layout-compatible2.C: New test.
* g++.dg/cpp2a/is-layout-compatible3.C: New test.

3 years agoObjective-C: Default flag_objc_sjlj_exceptions off for NeXT ABI >= 2.
Matt Jacobson [Thu, 29 Jul 2021 08:57:23 +0000 (09:57 +0100)]
Objective-C: Default flag_objc_sjlj_exceptions off for NeXT ABI >= 2.

Signed-off-by: Matt Jacobson <mhjacobson@me.com>
gcc/c-family/ChangeLog:

* c-opts.c (c_common_post_options): Default to
flag_objc_sjlj_exceptions = 1 only when flag_objc_abi < 2.

gcc/objc/ChangeLog:

* objc-next-runtime-abi-02.c
(objc_next_runtime_abi_02_init): Warn about and reset
flag_objc_sjlj_exceptions regardless of flag_objc_exceptions.
(next_runtime_02_initialize): Use a checking assert that
flag_objc_sjlj_exceptions is off.

3 years agolibstdc++: Avoid illegal argument to verbose in dg-test callback, continued
Thomas Schwinge [Tue, 17 Aug 2021 15:58:30 +0000 (17:58 +0200)]
libstdc++: Avoid illegal argument to verbose in dg-test callback, continued

This is a follow-up to commit 697b94cfaef4a958132faf0cf4b35b15dfb29acc
"libstdc++: Avoid illegal argument to verbose in dg-test callback".
I'm confirming the original problem, but on one system, it's not
resolved by this change, because instead we get:

    extra_tool_flags are:
    ERROR: tcl error sourcing [...]/libstdc++-v3/testsuite/libstdc++-dg/conformance.exp.
    ERROR: usage: send [args] string
        while executing
    "send_log "$message\n""
        (procedure "verbose" line 48)
        invoked from within
    "verbose -log -- $extra_tool_flags"
        (procedure "libstdc++-dg-test" line 45)
        invoked from within
    "${tool}-dg-test $prog [lindex ${dg-do-what} 0] "$tool_flags ${dg-extra-tool-flags}""
        (procedure "saved-dg-test" line 115)
        invoked from within
    [...]

That's Ubuntu's dejagnu 1.5-3ubuntu1 being so old that it doesn't include
DejaGnu commit 57c22601afe43d2c2b8819df4f2ecacb034516fd "Protect from leading
dash in message".  (I suppose that's what'd make this work, but have not
verified.)

libstdc++-v3/
* testsuite/lib/libstdc++.exp: Avoid illegal argument to verbose,
continued.

3 years agoDarwin, libcc1: Handle hosts with mdynamic-no-pic support.
Iain Sandoe [Sun, 4 Jul 2021 16:56:05 +0000 (17:56 +0100)]
Darwin, libcc1: Handle hosts with mdynamic-no-pic support.

The default for building host-side binaries for mdynamic-no-pic
hosts is to enable this.  However, it is not compatible with
dynamic libraries, so must be switched off for libcc1.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
libcc1/ChangeLog:

* Makefile.am: Switch mdynamic-no-pic to fPIC.
* Makefile.in: Regenerated.

3 years agoTurn global 'ggc_force_collect' variable into 'force_collect' parameter to 'ggc_collect'
Thomas Schwinge [Tue, 17 Aug 2021 08:47:02 +0000 (10:47 +0200)]
Turn global 'ggc_force_collect' variable into 'force_collect' parameter to 'ggc_collect'

This simplifies the interface and gets us rid of a global variable.
No change in behavior.

Clean-up for 2004-09-02 CVS commit (Subversion r86974,
Git commit 0772402279c0161fe41784911b52c77e12803c42)
"Better memory statistics, take 2".

gcc/
* ggc.h (ggc_collect): Add 'force_collect' parameter.
* ggc-page.c (ggc_collect): Use that one instead of global
'ggc_force_collect'.  Adjust all users.
* doc/gty.texi (Invoking the garbage collector): Update.
* ggc-internal.h (ggc_force_collect): Remove.
* ggc-common.c (ggc_force_collect): Likewise.
* selftest.h (forcibly_ggc_collect): Remove.
* ggc-tests.c (selftest::forcibly_ggc_collect): Likewise.
* read-rtl-function.c (test_loading_labels): Adjust.
* selftest-run-tests.c (run_tests): Likewise.

3 years agoRestore 'gcc.dg/pr78213.c' testing
Thomas Schwinge [Tue, 17 Aug 2021 06:45:18 +0000 (08:45 +0200)]
Restore 'gcc.dg/pr78213.c' testing

... after it had gotten disabled in r243681 (Git
commit ecfc21ff34ddc6f8aa517251fb51494c68ff741f)
"Introduce selftest::locate_file".

gcc/testsuite/
* gcc.dg/pr78213.c: Restore testing.

3 years agoDarwin: Reset section names table at the end of compile.
Iain Sandoe [Fri, 13 Aug 2021 19:20:04 +0000 (20:20 +0100)]
Darwin: Reset section names table at the end of compile.

For a single use (typical compile) this vector will be reclaimed
as GGC.  For JIT this is not sufficient since it does not reset
the pointer to NULL (and thus we think the the vector is already
allocated when a context is reused).

The clears the vector and sets the pointer to NULL at the end
of object output.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/ChangeLog:

* config/darwin.c (darwin_file_end): Reset and reclaim the
section names table at the end of compile.

3 years agoDarwin, X86, config: Adjust 'as' command lines [PR100340].
Iain Sandoe [Sat, 31 Jul 2021 15:29:03 +0000 (16:29 +0100)]
Darwin, X86, config: Adjust 'as' command lines [PR100340].

Versions of the assembler using clang from XCode 12.5/12.5.1
have a bug which produces different code layout between debug and
non-debug input, leading to a compare fail for default configure
parameters.

This is a workaround fix to disable the optimisation that is
responsible for the bug.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
PR target/100340 - Bootstrap fails with Clang 12.0.5 (XCode 12.5)

PR target/100340

gcc/ChangeLog:

* config.in: Regenerate.
* config/i386/darwin.h (EXTRA_ASM_OPTS): New
(ASM_SPEC): Pass options to disable branch shortening where
needed.
* configure: Regenerate.
* configure.ac: Detect versions of 'as' that support the
optimisation which has the bug.

3 years agoFall back to masked_gather_load/masked_scatter_store
Richard Biener [Tue, 17 Aug 2021 13:50:31 +0000 (15:50 +0200)]
Fall back to masked_gather_load/masked_scatter_store

This adds a fallback to the masked_ variants for gather_load
and scatter_store if the latter are not available.

2021-08-17  Richard Biener  <rguenther@suse.de>

* optabs-query.c (supports_vec_gather_load_p): Also check
for masked optabs.
(supports_vec_scatter_store_p): Likewise.
* tree-vect-data-refs.c (vect_gather_scatter_fn_p): Fall
back to masked variants if non-masked are not supported.
* tree-vect-patterns.c (vect_recog_gather_scatter_pattern):
When we need to use masked gather/scatter but do not have
a mask set up a constant true one.
* tree-vect-stmts.c (vect_check_scalar_mask): Also allow
non-SSA_NAME masks.

3 years agolibstdc++: Fix testsuite for skipping gdb tests on remote/non-native target
Luc Michel [Tue, 17 Aug 2021 15:34:37 +0000 (16:34 +0100)]
libstdc++: Fix testsuite for skipping gdb tests on remote/non-native target

This fixes an incorrect invocation of gdb on remote targets where
DejaGNU would try to run host's gdb in remote target simulator.
gdb-test skips the testing when target is remote or non native but the
gdb version check function does not.

Suggested-by: Jonathan Wakely <jwakely@redhat.com>
Signed-off-by: Luc Michel <lmichel@kalray.eu>
Co-authored-by: Marc Poulhies <mpoulhies@kalrayinc.com>
libstdc++-v3/ChangeLog:

* testsuite/lib/gdb-test.exp (gdb_version_check)
(gdb_version_check_xmethods): Only check the GDB version for
local native targets.

3 years agolibstdc++: Optimize std::seed_seq construction
Antony Polukhin [Tue, 17 Aug 2021 12:50:53 +0000 (13:50 +0100)]
libstdc++: Optimize std::seed_seq construction

When std::seed_seq is constructed from random access iterators we can
detect the internal vector size in O(1). Reserving memory for elements
in such cases may avoid multiple memory allocations.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/random.tcc (seed_seq::seed_seq): Reserve capacity
if distance is O(1).
* testsuite/26_numerics/random/pr60037-neg.cc: Adjust dg-error
line number.

Co-authored-by: Jonathan Wakely <jwakely@redhat.com>
3 years agoImproved handling of MINUS_EXPR in bit CCP.
Roger Sayle [Tue, 17 Aug 2021 13:59:14 +0000 (14:59 +0100)]
Improved handling of MINUS_EXPR in bit CCP.

This patch improves the bit bounds for MINUS_EXPR during tree-ssa's
conditional constant propagation (CCP) pass (and as an added bonus
adds support for POINTER_DIFF_EXPR).

The pessimistic assumptions made by the current algorithm are
demonstrated by considering 1 - (x&1).  Intuitively this should
have possible values 0 and 1, and therefore an unknown mask of 1.
Alas by treating subtraction as a negation followed by addition,
the second operand first becomes 0 or -1, with an unknown mask
of all ones, which results in the addition containing no known bits.

Improved bounds are achieved by using the same approach used for
PLUS_EXPR, determining the result with the minimum number of borrows,
the result from the maximum number of borrows, and examining the bits
they have in common.  One additional benefit of this approach
is that it is applicable to POINTER_DIFF_EXPR, where previously the
negation of a pointer didn't/doesn't make sense.

A more convincing example, where a transformation missed by .032t.cpp
isn't caught a few passes later by .038t.evrp, is the expression
(7 - (x&5)) & 2, which (in the new test case) currently survives the
tree-level optimizers but with this patch is now simplified to the
constant value 2.

2021-08-17  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* tree-ssa-ccp.c (bit_value_binop) [MINUS_EXPR]: Use same
algorithm as PLUS_EXPR to improve subtraction bit bounds.
[POINTER_DIFF_EXPR]: Treat as synonymous with MINUS_EXPR.

gcc/testsuite/ChangeLog
* gcc.dg/tree-ssa/ssa-ccp-40.c: New test case.

3 years agoImproved handling of MULT_EXPR in bit CCP.
Roger Sayle [Tue, 17 Aug 2021 13:50:54 +0000 (14:50 +0100)]
Improved handling of MULT_EXPR in bit CCP.

This patch allows GCC to constant fold (i | (i<<16)) | ((i<<24) | (i<<8)),
where i is an unsigned char, or the equivalent (i*65537) | (i*16777472), to
i*16843009.  The trick is to teach tree_nonzero_bits which bits may be
set in the result of a multiplication by a constant given which bits are
potentially set in the operands.  This allows the optimizations recently
added to match.pd to catch more cases.

The required mask/value pair from a multiplication may be calculated using
a classical shift-and-add algorithm, given we already have implementations
for both addition and shift by constant.  To keep this optimization "cheap",
this functionality is only used if the constant multiplier has a few bits
set (unless flag_expensive_optimizations), and we provide a special case
fast-path implementation for the common case where the (non-constant)
operand has no bits that are guaranteed to be set.  I have no evidence
that this functionality causes performance issues, it's just that sparse
multipliers provide the largest benefit to CCP.

2021-08-17  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* tree-ssa-ccp.c (bit_value_mult_const): New helper function to
calculate the mask-value pair result of a multiplication by an
unsigned constant.
(bit_value_binop) [MULT_EXPR]:  Call it from here for
multiplications by (sparse) non-negative constants.

gcc/testsuite/ChangeLog
* gcc.dg/fold-ior-5.c: New test case.

3 years agoFortran: Implement OpenMP 5.1 scope construct
Tobias Burnus [Tue, 17 Aug 2021 13:50:11 +0000 (15:50 +0200)]
Fortran: Implement OpenMP 5.1 scope construct

Fortran version to commit e45483c7c4badc4bf2d6ced22360ce1ab172967f,
which implemented OpenMP's scope construct for C and C++.
Most testcases are based on the C testcases; it also contains some
testcases which existed previously but had no Fortran equivalent.

gcc/fortran/ChangeLog:

* dump-parse-tree.c (show_omp_node, show_code_node): Handle
EXEC_OMP_SCOPE.
* gfortran.h (enum gfc_statement): Add ST_OMP_(END_)SCOPE.
(enum gfc_exec_op): Add EXEC_OMP_SCOPE.
* match.h (gfc_match_omp_scope): New.
* openmp.c (OMP_SCOPE_CLAUSES): Define
(gfc_match_omp_scope): New.
(gfc_match_omp_cancellation_point, gfc_match_omp_end_nowait):
Improve error diagnostic.
(omp_code_to_statement): Handle ST_OMP_SCOPE.
(gfc_resolve_omp_directive): Handle EXEC_OMP_SCOPE.
* parse.c (decode_omp_directive, next_statement,
gfc_ascii_statement, parse_omp_structured_block,
parse_executable): Handle OpenMP's scope construct.
* resolve.c (gfc_resolve_blocks): Likewise
* st.c (gfc_free_statement): Likewise
* trans-openmp.c (gfc_trans_omp_scope): New.
(gfc_trans_omp_directive): Call it.
* trans.c (trans_code): handle EXEC_OMP_SCOPE.

libgomp/ChangeLog:

* testsuite/libgomp.fortran/scope-1.f90: New test.
* testsuite/libgomp.fortran/task-reduction-16.f90: New test.

gcc/testsuite/ChangeLog:

* gfortran.dg/gomp/scan-1.f90:
* gfortran.dg/gomp/cancel-1.f90: New test.
* gfortran.dg/gomp/cancel-4.f90: New test.
* gfortran.dg/gomp/loop-4.f90: New test.
* gfortran.dg/gomp/nesting-1.f90: New test.
* gfortran.dg/gomp/nesting-2.f90: New test.
* gfortran.dg/gomp/nesting-3.f90: New test.
* gfortran.dg/gomp/nowait-1.f90: New test.
* gfortran.dg/gomp/reduction-task-1.f90: New test.
* gfortran.dg/gomp/reduction-task-2.f90: New test.
* gfortran.dg/gomp/reduction-task-2a.f90: New test.
* gfortran.dg/gomp/reduction-task-3.f90: New test.
* gfortran.dg/gomp/scope-1.f90: New test.
* gfortran.dg/gomp/scope-2.f90: New test.

3 years agolibstdc++: Test std::seed_seq construction from input iterators
Jonathan Wakely [Tue, 17 Aug 2021 13:18:58 +0000 (14:18 +0100)]
libstdc++: Test std::seed_seq construction from input iterators

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* testsuite/26_numerics/random/seed_seq/cons/range.cc: Check
construction from input iterators.

3 years agolibstdc++: Remove pretty printer committed by mistake
Jonathan Wakely [Tue, 17 Aug 2021 13:29:53 +0000 (14:29 +0100)]
libstdc++: Remove pretty printer committed by mistake

The std::error_category printer wasn't meant to be part of the commit
adding std::error_code and std::error_condition printers.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* python/libstdcxx/v6/printers.py (StdErrorCatPrinter): Remove.

3 years agolibstdc++: Optimize std::function move constructor [PR101923]
Jonathan Wakely [Tue, 17 Aug 2021 10:30:56 +0000 (11:30 +0100)]
libstdc++: Optimize std::function move constructor [PR101923]

PR 101923 points out that the unconditional swap in the std::function
move constructor makes it slower than copying an empty std::function.
The copy constructor has to check for the empty case before doing
anything, and that makes it very fast for the empty case.

Adding the same check to the move constructor avoids copying the
_Any_data POD when we don't need to. We can also inline the effects of
swap, by copying each member and then zeroing the pointer members.

This makes moving an empty object at least as fast as copying an empty
object.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

PR libstdc++/101923
* include/bits/std_function.h (function(function&&)): Check for
non-empty parameter before doing any work.

3 years agolibstdc++: Only define basic_string::contains for C++23
Jonathan Wakely [Mon, 16 Aug 2021 19:42:54 +0000 (20:42 +0100)]
libstdc++: Only define basic_string::contains for C++23

The new contains member of the COW string is defined for non-strict
gnu++20 mode as well as for C++23 modes. I think that was left in the
committed patch unintentionally. It is inconsistent with the SSO string,
and doesn't actually compile because it uses the
basic_string_view::contains member which only defined for C++23.

This makes it only defined for C++23.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/cow_string.h (basic_string::contains): Do not
define for -std=gnu++20.

3 years agolibstdc++: Rename __detail::__not_same_as helper
Jonathan Wakely [Mon, 16 Aug 2021 17:10:04 +0000 (18:10 +0100)]
libstdc++: Rename __detail::__not_same_as helper

This is done to match an editorial change in the working draft, to
rename the exposition-only not-same-as helper to different-from.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/ranges_util.h (__not_same_as): Rename to
__different_from.
* include/std/ranges (__not_same_as): Likewise.

3 years agolibstdc++: Add conditional noexcept to std::exchange
Jonathan Wakely [Mon, 16 Aug 2021 17:00:08 +0000 (18:00 +0100)]
libstdc++: Add conditional noexcept to std::exchange

This is not required by the standard, but seems useful.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/std/utility (exchange): Add noexcept-specifier.
* testsuite/20_util/exchange/noexcept.cc: New test.

3 years agolibstdc++: Add pretty printer for std::error_code and std::error_condition
Jonathan Wakely [Mon, 16 Aug 2021 16:41:50 +0000 (17:41 +0100)]
libstdc++: Add pretty printer for std::error_code and std::error_condition

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* python/libstdcxx/v6/printers.py (StdErrorCodePrinter): Define.
(build_libstdcxx_dictionary): Register printer for
std::error_code and std::error_condition.
* testsuite/libstdc++-prettyprinters/cxx11.cc: Test it.

3 years agoDo not enable DT_INIT_ARRAY/DT_FINI_ARRAY on uclinuxfdpiceabi
Christophe Lyon [Tue, 17 Aug 2021 12:59:26 +0000 (12:59 +0000)]
Do not enable DT_INIT_ARRAY/DT_FINI_ARRAY on uclinuxfdpiceabi

Commit r12-1328 enabled DT_INIT_ARRAY/DT_FINI_ARRAY for all Linux
targets, but this does not work for arm-none-uclinuxfdpiceabi: it
makes all the execution tests fail.

This patch restores the original behavior for uclinuxfdpiceabi.

2021-08-12  Christophe Lyon  <christophe.lyon@foss.st.com>

gcc/
PR target/100896
* config.gcc (gcc_cv_initfini_array): Leave undefined for
uclinuxfdpiceabi targets.

3 years agoretain debug stmt order when moving to successors
Alexandre Oliva [Sun, 25 Jul 2021 02:05:33 +0000 (23:05 -0300)]
retain debug stmt order when moving to successors

We iterate over debug stmts from the last one in new_bb, and we insert
them before the first post-label stmt in each dest block, without
moving the insertion iterator, so they end up reversed.  Moving the
insertion iterator fixes this.

for  gcc/ChangeLog

* tree-inline.c (maybe_move_debug_stmts_to_successors): Don't
reverse debug stmts.

3 years agodon't access cfun in dump_function_to_file
Alexandre Oliva [Sun, 25 Jul 2021 02:05:32 +0000 (23:05 -0300)]
don't access cfun in dump_function_to_file

dump_function_to_file takes the function to dump as a parameter, and
parts of it use the local fun variable where cfun would be used
elsewhere.  Others use cfun, presumably in error.  Fixed to use fun
uniformly.  Added a few more tests for non-NULL fun before
dereferencing it.

for  gcc/ChangeLog

* tree-cfg.c (dump_function_to_file): Use fun, not cfun.

3 years agoaarch64: Remove macros for vld4[q]_lane Neon intrinsics
Jonathan Wright [Mon, 16 Aug 2021 13:37:18 +0000 (14:37 +0100)]
aarch64: Remove macros for vld4[q]_lane Neon intrinsics

Remove macros for vld4[q]_lane Neon intrinsics. This is a preparatory
step before adding new modes for structures of Advanced SIMD vectors.

gcc/ChangeLog:

2021-08-16  Jonathan Wright  <jonathan.wright@arm.com>

* config/aarch64/arm_neon.h (__LD4_LANE_FUNC): Delete.
(__LD4Q_LANE_FUNC): Likewise.
(vld4_lane_u8): Define without macro.
(vld4_lane_u16): Likewise.
(vld4_lane_u32): Likewise.
(vld4_lane_u64): Likewise.
(vld4_lane_s8): Likewise.
(vld4_lane_s16): Likewise.
(vld4_lane_s32): Likewise.
(vld4_lane_s64): Likewise.
(vld4_lane_f16): Likewise.
(vld4_lane_f32): Likewise.
(vld4_lane_f64): Likewise.
(vld4_lane_p8): Likewise.
(vld4_lane_p16): Likewise.
(vld4_lane_p64): Likewise.
(vld4q_lane_u8): Likewise.
(vld4q_lane_u16): Likewise.
(vld4q_lane_u32): Likewise.
(vld4q_lane_u64): Likewise.
(vld4q_lane_s8): Likewise.
(vld4q_lane_s16): Likewise.
(vld4q_lane_s32): Likewise.
(vld4q_lane_s64): Likewise.
(vld4q_lane_f16): Likewise.
(vld4q_lane_f32): Likewise.
(vld4q_lane_f64): Likewise.
(vld4q_lane_p8): Likewise.
(vld4q_lane_p16): Likewise.
(vld4q_lane_p64): Likewise.
(vld4_lane_bf16): Likewise.
(vld4q_lane_bf16): Likewise.

3 years agoaarch64: Remove macros for vld3[q]_lane Neon intrinsics
Jonathan Wright [Mon, 16 Aug 2021 08:59:44 +0000 (09:59 +0100)]
aarch64: Remove macros for vld3[q]_lane Neon intrinsics

Remove macros for vld3[q]_lane Neon intrinsics. This is a preparatory
step before adding new modes for structures of Advanced SIMD vectors.

gcc/ChangeLog:

2021-08-16  Jonathan Wright  <jonathan.wright@arm.com>

* config/aarch64/arm_neon.h (__LD3_LANE_FUNC): Delete.
(__LD3Q_LANE_FUNC): Delete.
(vld3_lane_u8): Define without macro.
(vld3_lane_u16): Likewise.
(vld3_lane_u32): Likewise.
(vld3_lane_u64): Likewise.
(vld3_lane_s8): Likewise.
(vld3_lane_s16): Likewise.
(vld3_lane_s32): Likewise.
(vld3_lane_s64): Likewise.
(vld3_lane_f16): Likewise.
(vld3_lane_f32): Likewise.
(vld3_lane_f64): Likewise.
(vld3_lane_p8): Likewise.
(vld3_lane_p16): Likewise.
(vld3_lane_p64): Likewise.
(vld3q_lane_u8): Likewise.
(vld3q_lane_u16): Likewise.
(vld3q_lane_u32): Likewise.
(vld3q_lane_u64): Likewise.
(vld3q_lane_s8): Likewise.
(vld3q_lane_s16): Likewise.
(vld3q_lane_s32): Likewise.
(vld3q_lane_s64): Likewise.
(vld3q_lane_f16): Likewise.
(vld3q_lane_f32): Likewise.
(vld3q_lane_f64): Likewise.
(vld3q_lane_p8): Likewise.
(vld3q_lane_p16): Likewise.
(vld3q_lane_p64): Likewise.
(vld3_lane_bf16): Likewise.
(vld3q_lane_bf16): Likewise.

3 years agoaarch64: Remove macros for vld2[q]_lane Neon intrinsics
Jonathan Wright [Thu, 12 Aug 2021 11:27:15 +0000 (12:27 +0100)]
aarch64: Remove macros for vld2[q]_lane Neon intrinsics

Remove macros for vld2[q]_lane Neon intrinsics. This is a preparatory
step before adding new modes for structures of Advanced SIMD vectors.

gcc/ChangeLog:

2021-08-12  Jonathan Wright  <jonathan.wright@arm.com>

* config/aarch64/arm_neon.h (__LD2_LANE_FUNC): Delete.
(__LD2Q_LANE_FUNC): Likewise.
(vld2_lane_u8): Define without macro.
(vld2_lane_u16): Likewise.
(vld2_lane_u32): Likewise.
(vld2_lane_u64): Likewise.
(vld2_lane_s8): Likewise.
(vld2_lane_s16): Likewise.
(vld2_lane_s32): Likewise.
(vld2_lane_s64): Likewise.
(vld2_lane_f16): Likewise.
(vld2_lane_f32): Likewise.
(vld2_lane_f64): Likewise.
(vld2_lane_p8): Likewise.
(vld2_lane_p16): Likewise.
(vld2_lane_p64): Likewise.
(vld2q_lane_u8): Likewise.
(vld2q_lane_u16): Likewise.
(vld2q_lane_u32): Likewise.
(vld2q_lane_u64): Likewise.
(vld2q_lane_s8): Likewise.
(vld2q_lane_s16): Likewise.
(vld2q_lane_s32): Likewise.
(vld2q_lane_s64): Likewise.
(vld2q_lane_f16): Likewise.
(vld2q_lane_f32): Likewise.
(vld2q_lane_f64): Likewise.
(vld2q_lane_p8): Likewise.
(vld2q_lane_p16): Likewise.
(vld2q_lane_p64): Likewise.
(vld2_lane_bf16): Likewise.
(vld2q_lane_bf16): Likewise.

3 years agoImprove diff-ability of scheduler logs
Maxim Kuvyrkov [Thu, 29 Aug 2019 15:27:30 +0000 (15:27 +0000)]
Improve diff-ability of scheduler logs

* haifa-sched.c (advance_one_cycle): Output more context-synchronization
lines for diff.

3 years agoAdd missing entry for rank_for_schedule stats.
Maxim Kuvyrkov [Thu, 29 Aug 2019 15:24:37 +0000 (15:24 +0000)]
Add missing entry for rank_for_schedule stats.

* haifa-sched.c (enum rfs_decision, rfs_str): Add RFS_AUTOPREF.
(rank_for_schedule): Use it.

3 years agoImprove autoprefetcher heuristic (partly fix regression in PR91598)
Maxim Kuvyrkov [Thu, 29 Aug 2019 15:21:36 +0000 (15:21 +0000)]
Improve autoprefetcher heuristic (partly fix regression in PR91598)

PR rtl-optimization/91598
* haifa-sched.c (autopref_rank_for_schedule): Prioritize "irrelevant"
insns after memory reads and before memory writes.

3 years agoaarch64: Replace some uses of GET_CODE with RTL predicate macros
Alistair Lee [Tue, 17 Aug 2021 09:49:35 +0000 (10:49 +0100)]
aarch64: Replace some uses of GET_CODE with RTL predicate macros

gcc/
2021-08-17  Alistair_Lee  <alistair.lee@arm.com>

* rtl.h (CONST_VECTOR_P): New macro.
* config/aarch64/aarch64.c (aarch64_get_sve_pred_bits): Use RTL
code testing macros.
(aarch64_ptrue_all_mode): Likewise.
(aarch64_expand_mov_immediate): Likewise.
(aarch64_const_vec_all_in_range_p): Likewise.
(aarch64_rtx_costs): Likewise.
(aarch64_legitimate_constant_p): Likewise.
(aarch64_simd_valid_immediate): Likewise.
(aarch64_simd_make_constant): Likewise.
(aarch64_convert_mult_to_shift): Likewise.
(aarch64_expand_sve_vec_perm): Likewise.
(aarch64_vec_fpconst_pow_of_2): Likewise.

3 years agoSpecial case -TYPE_MIN_VALUE for flag_wrapv in operator_abs::op1_range.
Andrew MacLeod [Tue, 17 Aug 2021 08:50:56 +0000 (10:50 +0200)]
Special case -TYPE_MIN_VALUE for flag_wrapv in operator_abs::op1_range.

With flag_wrapv, -TYPE_MIN_VALUE = TYPE_MIN_VALUE which is
unrepresentable.  We currently special case this in the ABS folding
routine, but are missing similar treatment in operator_abs::op1_range.

Tested on x86-64 Linux.

PR tree-optimization/101938

gcc/ChangeLog:

* range-op.cc (operator_abs::op1_range): Special case
-TYPE_MIN_VALUE for flag_wrapv.

gcc/testsuite/ChangeLog:

* gcc.dg/pr101938.c: New test.

3 years agotree-optimization/101868 - avoid PRE of trapping mems across calls
Richard Biener [Tue, 17 Aug 2021 09:23:06 +0000 (11:23 +0200)]
tree-optimization/101868 - avoid PRE of trapping mems across calls

This adds the testcase from the fix for the PR.

2021-08-17  Richard Biener  <rguenther@suse.de>

PR tree-optimization/101868
* gcc.dg/lto/pr101868_0.c: New testcase.
* gcc.dg/lto/pr101868_1.c: Likewise.
* gcc.dg/lto/pr101868_2.c: Likewise.
* gcc.dg/lto/pr101868_3.c: Likewise.

3 years agovect: Add extraction cost for slp reduc
Kewen Lin [Tue, 17 Aug 2021 08:18:02 +0000 (03:18 -0500)]
vect: Add extraction cost for slp reduc

As Richi pointed out, currently for BB reductions we don't
actually build a SLP node with IFN_REDUC_* information,
ideally we may end up with that eventually.

For now, it's costed as shuffles and reduc operations, it
misses the cost of lane extraction.  This patch is to add
one time of vec_to_scalar cost for lane extraction, it's to
make the costings consistent and conservative for now.

gcc/ChangeLog:

* tree-vect-slp.c (vectorizable_bb_reduc_epilogue): Add the cost for
value extraction.

3 years agoopenmp: Implement OpenMP 5.1 scope construct
Jakub Jelinek [Tue, 17 Aug 2021 07:30:09 +0000 (09:30 +0200)]
openmp: Implement OpenMP 5.1 scope construct

This patch implements the OpenMP 5.1 scope construct, which is similar
to worksharing constructs in many regards, but isn't one of them.
The body of the construct is encountered by all threads though, it can
be nested in itself or intermixed with taskgroup and worksharing etc.
constructs can appear inside of it (but it can't be nested in
worksharing etc. constructs).  The main purpose of the construct
is to allow reductions (normal and task ones) without the need to
close the parallel and reopen another one.

If it doesn't have task reductions, it can be implemented without
any new library support, with nowait it just does the privatizations
at the start if any and reductions before the end of the body, with
without nowait emits a normal GOMP_barrier{,_cancel} at the end too.

For task reductions, we need to ensure only one thread initializes
the task reduction library data structures and other threads copy from that,
so a new GOMP_scope_start routine is added to the library for that.
It acts as if the start of the scope construct is a nowait worksharing
construct (that is ok, it can't be nested in other worksharing
constructs and all threads need to encounter the start in the same
order) which does the task reduction initialization, but as the body
can have other scope constructs and/or worksharing constructs, that is
all where we use this dummy worksharing construct.  With task reductions,
the construct must not have nowait and ends with a GOMP_barrier{,_cancel},
followed by task reductions followed by GOMP_workshare_task_reduction_unregister.

Only C/C++ FE support is done.

2021-08-17  Jakub Jelinek  <jakub@redhat.com>

gcc/
* tree.def (OMP_SCOPE): New tree code.
* tree.h (OMP_SCOPE_BODY, OMP_SCOPE_CLAUSES): Define.
* tree-nested.c (convert_nonlocal_reference_stmt,
convert_local_reference_stmt, convert_gimple_call): Handle
GIMPLE_OMP_SCOPE.
* tree-pretty-print.c (dump_generic_node): Handle OMP_SCOPE.
* gimple.def (GIMPLE_OMP_SCOPE): New gimple code.
* gimple.c (gimple_build_omp_scope): New function.
(gimple_copy): Handle GIMPLE_OMP_SCOPE.
* gimple.h (gimple_build_omp_scope): Declare.
(gimple_has_substatements): Handle GIMPLE_OMP_SCOPE.
(gimple_omp_scope_clauses, gimple_omp_scope_clauses_ptr,
gimple_omp_scope_set_clauses): New inline functions.
(CASE_GIMPLE_OMP): Add GIMPLE_OMP_SCOPE.
* gimple-pretty-print.c (dump_gimple_omp_scope): New function.
(pp_gimple_stmt_1): Handle GIMPLE_OMP_SCOPE.
* gimple-walk.c (walk_gimple_stmt): Likewise.
* gimple-low.c (lower_stmt): Likewise.
* gimplify.c (is_gimple_stmt): Handle OMP_MASTER.
(gimplify_scan_omp_clauses): For task reductions, handle OMP_SCOPE
like ORT_WORKSHARE constructs.  Adjust diagnostics for %<scope%>
allowing task reductions.  Reject inscan reductions on scope.
(omp_find_stores_stmt): Handle GIMPLE_OMP_SCOPE.
(gimplify_omp_workshare, gimplify_expr): Handle OMP_SCOPE.
* tree-inline.c (remap_gimple_stmt): Handle GIMPLE_OMP_SCOPE.
(estimate_num_insns): Likewise.
* omp-low.c (build_outer_var_ref): Look through GIMPLE_OMP_SCOPE
contexts if var isn't privatized there.
(check_omp_nesting_restrictions): Handle GIMPLE_OMP_SCOPE.
(scan_omp_1_stmt): Likewise.
(maybe_add_implicit_barrier_cancel): Look through outer
scope constructs.
(lower_omp_scope): New function.
(lower_omp_task_reductions): Handle OMP_SCOPE.
(lower_omp_1): Handle GIMPLE_OMP_SCOPE.
(diagnose_sb_1, diagnose_sb_2): Likewise.
* omp-expand.c (expand_omp_single): Support also GIMPLE_OMP_SCOPE.
(expand_omp): Handle GIMPLE_OMP_SCOPE.
(omp_make_gimple_edges): Likewise.
* omp-builtins.def (BUILT_IN_GOMP_SCOPE_START): New built-in.
gcc/c-family/
* c-pragma.h (enum pragma_kind): Add PRAGMA_OMP_SCOPE.
* c-pragma.c (omp_pragmas): Add scope construct.
* c-omp.c (omp_directives): Uncomment scope directive entry.
gcc/c/
* c-parser.c (OMP_SCOPE_CLAUSE_MASK): Define.
(c_parser_omp_scope): New function.
(c_parser_omp_construct): Handle PRAGMA_OMP_SCOPE.
gcc/cp/
* parser.c (OMP_SCOPE_CLAUSE_MASK): Define.
(cp_parser_omp_scope): New function.
(cp_parser_omp_construct, cp_parser_pragma): Handle PRAGMA_OMP_SCOPE.
* pt.c (tsubst_expr): Handle OMP_SCOPE.
gcc/testsuite/
* c-c++-common/gomp/nesting-2.c (foo): Add scope and masked
construct tests.
* c-c++-common/gomp/scan-1.c (f3): Add scope construct test..
* c-c++-common/gomp/cancel-1.c (f2): Add scope and masked
construct tests.
* c-c++-common/gomp/reduction-task-2.c (bar): Add scope construct
test.  Adjust diagnostics for the addition of scope.
* c-c++-common/gomp/loop-1.c (f5): Add master, masked and scope
construct tests.
* c-c++-common/gomp/clause-dups-1.c (f1): Add scope construct test.
* gcc.dg/gomp/nesting-1.c (f1, f2, f3): Add scope construct tests.
* c-c++-common/gomp/scope-1.c: New test.
* c-c++-common/gomp/scope-2.c: New test.
* g++.dg/gomp/attrs-1.C (bar): Add scope construct tests.
* g++.dg/gomp/attrs-2.C (bar): Likewise.
* gfortran.dg/gomp/reduction4.f90: Adjust expected diagnostics.
* gfortran.dg/gomp/reduction7.f90: Likewise.
libgomp/
* Makefile.am (libgomp_la_SOURCES): Add scope.c
* Makefile.in: Regenerated.
* libgomp_g.h (GOMP_scope_start): Declare.
* libgomp.map: Add GOMP_scope_start@@GOMP_5.1.
* scope.c: New file.
* testsuite/libgomp.c-c++-common/scope-1.c: New test.
* testsuite/libgomp.c-c++-common/task-reduction-16.c: New test.

3 years agoc++: Add C++20 #__VA_OPT__ support
Jakub Jelinek [Tue, 17 Aug 2021 07:25:56 +0000 (09:25 +0200)]
c++: Add C++20 #__VA_OPT__ support

The following patch implements C++20 # __VA_OPT__ (...) support.
Testcases cover what I came up with myself and what LLVM has for #__VA_OPT__
in its testsuite and the string literals are identical between the two
compilers on the va-opt-5.c testcase.

2021-08-17  Jakub Jelinek  <jakub@redhat.com>

libcpp/
* macro.c (vaopt_state): Add m_stringify member.
(vaopt_state::vaopt_state): Initialize it.
(vaopt_state::update): Overwrite it.
(vaopt_state::stringify): New method.
(stringify_arg): Replace arg argument with first, count arguments
and add va_opt argument.  Use first instead of arg->first and
count instead of arg->count, for va_opt add paste_tokens handling.
(paste_tokens): Fix up len calculation.  Don't spell rhs twice,
instead use %.*s to supply lhs and rhs spelling lengths.  Don't call
_cpp_backup_tokens here.
(paste_all_tokens): Call it here instead.
(replace_args): Adjust stringify_arg caller.  For vaopt_state::END
if stringify is true handle __VA_OPT__ stringification.
(create_iso_definition): Handle # __VA_OPT__ similarly to # macro_arg.
gcc/testsuite/
* c-c++-common/cpp/va-opt-5.c: New test.
* c-c++-common/cpp/va-opt-6.c: New test.

3 years agotree-optimization/101925 - fix VN with reverse storage order
Richard Biener [Mon, 16 Aug 2021 13:17:08 +0000 (15:17 +0200)]
tree-optimization/101925 - fix VN with reverse storage order

This fixes value-numbering breaking reverse storage order accesses
due to a missed check.  It adds a new overload for
reverse_storage_order_for_component_p and sets reversed on the
VN IL ops for component and array accesses accordingly.
It also compares the reversed reference ops flag on reference
lookup.

2021-08-16  Richard Biener  <rguenther@suse.de>

PR tree-optimization/101925
* tree-ssa-sccvn.c (copy_reference_ops_from_ref): Set
reverse on COMPONENT_REF and ARRAY_REF according to
what reverse_storage_order_for_component_p does.
(vn_reference_eq): Compare reversed on reference ops.
(reverse_storage_order_for_component_p): New overload.
(vn_reference_lookup_3): Check reverse_storage_order_for_component_p
on the reference looked up.

* gcc.dg/sso-16.c: New testcase.

3 years agoImprove SImode shifts for H8
Jeff Law [Tue, 17 Aug 2021 02:23:30 +0000 (22:23 -0400)]
Improve SImode shifts for H8

Similar to the H8/300H patch, this improves SImode shifts for the H8/S.
It's not as big a win on the H8/S since we can shift two positions at a
time.  But that also means that we can handle more residuals with minimal
ode growth after a special shift-by-16 or shift-by-24 sequence.

I think there's more to do here, but this seemed like as good a checkpoint
as any.  Tested without regressions.

gcc/
* config/h8300/h8300.c (shift_alg_si): Avoid loops for most SImode
shifts on the H8/S.
(h8300_option_override): Use loops on H8/S more often when optimizing
for size.
(get_shift_alg): Handle new "special" cases on H8/S.  Simplify
accordingly.  Handle various arithmetic right shifts with special
sequences that we couldn't handle before.

3 years agoAdjust testcase.
liuhongt [Tue, 17 Aug 2021 01:32:35 +0000 (09:32 +0800)]
Adjust testcase.

This testcase is used to detect reuse of perm mask in the main loop,
in epilog, vpermi2b can still be used, so add the option
--param=vect-epilogues-nomask=0.

gcc/testsuite/ChangeLog:

* gcc.target/i386/pr82460-2.c: Adjust testcase by adding
--param=vect-epilogues-nomask=0

3 years agoDaily bump.
GCC Administrator [Tue, 17 Aug 2021 00:16:32 +0000 (00:16 +0000)]
Daily bump.

3 years agoDrop embeded stabs from rl78 port
Jeff Law [Mon, 16 Aug 2021 23:03:56 +0000 (19:03 -0400)]
Drop embeded stabs from rl78 port

gcc/
* config.gcc (rl78-*-elf*): Do not include dbxelf.h.

3 years agoUpdate cpplib de.po
Joseph Myers [Mon, 16 Aug 2021 19:16:23 +0000 (19:16 +0000)]
Update cpplib de.po

* de.po: Update.

3 years agolibstdc++: Use qualified-id for class member constant [PR101937]
Jonathan Wakely [Mon, 16 Aug 2021 14:35:58 +0000 (15:35 +0100)]
libstdc++: Use qualified-id for class member constant [PR101937]

The expression ctx._M_indent is not a constant expression when ctx is a
reference parameter, even though _M_indent is an enumerator. Rename it
to _S_indent to be consistent with our conventions, and refer to it as
PrintContext::_S_indent to be valid C++ code (at least until P2280 is
accepted as a DR).

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

PR libstdc++/101937
* src/c++11/debug.cc (PrintContext::_M_indent): Replace with a
static data member.
(print_word): Use qualified-id to access it.

3 years agolibstdc++: Install GDB pretty printers for debug library
Jonathan Wakely [Thu, 12 Aug 2021 18:56:14 +0000 (19:56 +0100)]
libstdc++: Install GDB pretty printers for debug library

The additional libraries installed by --enable-libstdcxx-debug are built
without optimization to aid debugging, but the Python pretty printers
are not installed alongside them. This means that you can step through
the unoptimized library code, but at the expense of pretty printing the
library types.

This remedies the situation by installing another copy of the GDB hooks
alongside the debug version of libstdc++.so.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* python/Makefile.am [GLIBCXX_BUILD_DEBUG] (install-data-local):
Install another copy of the GDB hook.
* python/Makefile.in: Regenerate.

3 years agogcov: Add TARGET_GCOV_TYPE_SIZE target hook
Sebastian Huber [Mon, 9 Aug 2021 07:06:14 +0000 (09:06 +0200)]
gcov: Add TARGET_GCOV_TYPE_SIZE target hook

If -fprofile-update=atomic is used, then the target must provide atomic
operations for the counters of the type returned by get_gcov_type().
This is a 64-bit type for targets which have a 64-bit long long type.
On 32-bit targets this could be an issue since they may not provide
64-bit atomic operations.  Allow targets to override the default type
size with the new TARGET_GCOV_TYPE_SIZE target hook.

If a 32-bit gcov type size is used, then there is currently a warning in
libgcov-driver.c in a dead code block due to
sizeof (counter) == sizeof (gcov_unsigned_t):

libgcc/libgcov-driver.c: In function 'dump_counter':
libgcc/libgcov-driver.c:401:46: warning: right shift count >= width of type [-Wshift-count-overflow]
  401 |     dump_unsigned ((gcov_unsigned_t)(counter >> 32), dump_fn, arg);
      |                                              ^~

gcc/c-family/

* c-cppbuiltin.c (c_cpp_builtins): Define
__LIBGCC_GCOV_TYPE_SIZE if flag_building_libgcc is true.

gcc/

* config/sparc/rtemself.h (SPARC_GCOV_TYPE_SIZE): Define.
* config/sparc/sparc.c (sparc_gcov_type_size): New.
(TARGET_GCOV_TYPE_SIZE): Redefine if SPARC_GCOV_TYPE_SIZE is defined.
* coverage.c (get_gcov_type): Use targetm.gcov_type_size().
* doc/tm.texi (TARGET_GCOV_TYPE_SIZE): Add hook under "Misc".
* doc/tm.texi.in: Regenerate.
* target.def (gcov_type_size): New target hook.
* targhooks.c (default_gcov_type_size): New.
* targhooks.h (default_gcov_type_size): Declare.
* tree-profile.c (gimple_gen_edge_profiler): Use precision of
gcov_type_node.
(gimple_gen_time_profiler): Likewise.

libgcc/

* libgcov.h (gcov_type): Define using __LIBGCC_GCOV_TYPE_SIZE.
(gcov_type_unsigned): Likewise.

3 years agoFix regression in debug info for Ada with DWARF 5
Eric Botcazou [Mon, 16 Aug 2021 13:26:22 +0000 (15:26 +0200)]
Fix regression in debug info for Ada with DWARF 5

add_scalar_info can directly generate a reference to an existing DIE for a
scalar attribute, e.g the upper bound of a VLA, but it does so only if this
existing DIE has a location or is a constant:

              if (get_AT (decl_die, DW_AT_location)
                  || get_AT (decl_die, DW_AT_data_member_location)
                  || get_AT (decl_die, DW_AT_const_value))

Now, in DWARF 5, members of a structure that are bitfields no longer have a
DW_AT_data_member_location but a DW_AT_data_bit_offset attribute instead, so
the condition is bypassed.

gcc/
* dwarf2out.c (add_scalar_info): Deal with DW_AT_data_bit_offset.

3 years ago[OpenMP] Update omp-low.c's omp_runtime_api_call [PR101931]
Tobias Burnus [Mon, 16 Aug 2021 12:00:13 +0000 (14:00 +0200)]
[OpenMP] Update omp-low.c's omp_runtime_api_call [PR101931]

gcc/ChangeLog:

PR middle-end/101931
* omp-low.c (omp_runtime_api_call): Update for routines
added in the meanwhile.

3 years agoSpeed up jump table switch detection.
Martin Liska [Fri, 13 Aug 2021 15:22:35 +0000 (17:22 +0200)]
Speed up jump table switch detection.

PR tree-optimization/100393

gcc/ChangeLog:

* tree-switch-conversion.c (group_cluster::dump): Use
  get_comparison_count.
(jump_table_cluster::find_jump_tables): Pre-compute number of
comparisons and then decrement it. Cache also max_ratio.
(jump_table_cluster::can_be_handled): Change signature.
* tree-switch-conversion.h (get_comparison_count): New.

3 years agoDisable GNAT encodings by default
Eric Botcazou [Mon, 16 Aug 2021 10:14:12 +0000 (12:14 +0200)]
Disable GNAT encodings by default

Given the latest work in the compiler and debugger, we no longer need to use
most GNAT-specific encodings in the debug info generated for an Ada program,
so the attached patch disables them, except with -fgnat-encodings=all.

gcc/
* dwarf2out.c (add_data_member_location_attribute): Use GNAT
encodings only when -fgnat-encodings=all is specified.
(add_bound_info): Likewise.
(add_byte_size_attribute): Likewise.
(gen_member_die): Likewise.

3 years agoAddress '?:' issues in 'libgomp.oacc-c-c++-common/mode-transitions.c'
Thomas Schwinge [Wed, 11 Aug 2021 09:59:19 +0000 (11:59 +0200)]
Address '?:' issues in 'libgomp.oacc-c-c++-common/mode-transitions.c'

    [...]/libgomp.oacc-c-c++-common/mode-transitions.c: In function ‘t3’:
    [...]/libgomp.oacc-c-c++-common/mode-transitions.c:127:43: warning: ‘?:’ using integer constants in boolean context, the expression will always evaluate to ‘true’ [-Wint-in-bool-context]
      127 |     assert (arr[i] == ((i % 64) < 32) ? 1 : -1);
          |                                           ^

    [...]/libgomp.oacc-c-c++-common/mode-transitions.c: In function ‘t9’:
    [...]/libgomp.oacc-c-c++-common/mode-transitions.c:359:46: warning: ‘?:’ using integer constants in boolean context, the expression will always evaluate to ‘true’ [-Wint-in-bool-context]
      359 |         assert (arr[i] == ((i % 3) == 0) ? 1 : 2);
          |                                              ^

..., and PR101862 "[C, C++] Potential '?:' diagnostic for always-true
expressions in boolean context".

libgomp/
* testsuite/libgomp.oacc-c-c++-common/mode-transitions.c: Address
'?:' issues.

3 years agoPlug 'par' memory leak in 'gcc/omp-oacc-neuter-broadcast.cc:execute_omp_oacc_neuter_b...
Thomas Schwinge [Fri, 6 Aug 2021 13:34:25 +0000 (15:34 +0200)]
Plug 'par' memory leak in 'gcc/omp-oacc-neuter-broadcast.cc:execute_omp_oacc_neuter_broadcast'

Fix-up for recent commit e2a58ed6dc5293602d0d168475109caa81ad0f0d
"openacc: Middle-end worker-partitioning support".

gcc/
* omp-oacc-neuter-broadcast.cc
(execute_omp_oacc_neuter_broadcast): Plug 'par' memory leak.

3 years agoClarify memory management for 'prop_set' in 'gcc/omp-oacc-neuter-broadcast.cc'
Thomas Schwinge [Wed, 11 Aug 2021 20:31:55 +0000 (22:31 +0200)]
Clarify memory management for 'prop_set' in 'gcc/omp-oacc-neuter-broadcast.cc'

Clean-up for recent commit e2a58ed6dc5293602d0d168475109caa81ad0f0d
"openacc: Middle-end worker-partitioning support".

gcc/
* omp-oacc-neuter-broadcast.cc
(execute_omp_oacc_neuter_broadcast): Clarify memory management for
'prop_set'.

3 years agoAvoid 'GTY' use for 'gcc/omp-oacc-neuter-broadcast.cc:field_map'
Thomas Schwinge [Fri, 6 Aug 2021 10:09:12 +0000 (12:09 +0200)]
Avoid 'GTY' use for 'gcc/omp-oacc-neuter-broadcast.cc:field_map'

... and further simplify related things a bit.

Fix-up/clean-up for recent commit e2a58ed6dc5293602d0d168475109caa81ad0f0d
"openacc: Middle-end worker-partitioning support".

gcc/
* omp-oacc-neuter-broadcast.cc (field_map): Move variable into...
(execute_omp_oacc_neuter_broadcast): ... here.
(install_var_field, build_receiver_ref, build_sender_ref): Take
'field_map_t *' parameter.  Adjust all users.
(worker_single_copy, neuter_worker_single): Take a
'record_field_map_t *' parameter.  Adjust all users.

3 years agoFix ICE.
liuhongt [Mon, 16 Aug 2021 09:12:21 +0000 (17:12 +0800)]
Fix ICE.

gcc/ChangeLog:

PR target/101930
* config/i386/i386.md (ldexp<mode>3): Force operands[1] to
reg.

gcc/testsuite/ChangeLog:

PR target/101930
* gcc.target/i386/pr101930.c: New test.

3 years agoipa: make target_clone default decl local [PR101726]
Martin Liska [Thu, 12 Aug 2021 14:01:01 +0000 (16:01 +0200)]
ipa: make target_clone default decl local [PR101726]

PR ipa/101726

gcc/ChangeLog:

* multiple_target.c (create_dispatcher_calls): Make default
  function local only if it is a definition.

3 years agoipa: ICF should check SSA_NAME_IS_DEFAULT_DEF
Martin Liska [Fri, 13 Aug 2021 10:35:47 +0000 (12:35 +0200)]
ipa: ICF should check SSA_NAME_IS_DEFAULT_DEF

PR ipa/100600

gcc/ChangeLog:

* ipa-icf-gimple.c (func_checker::compare_ssa_name): Do not
  consider equal SSA_NAMEs when one is a param.

gcc/testsuite/ChangeLog:

* gcc.dg/ipa/pr100600.c: New test.

3 years agoOptimize __builtin_shuffle_vector.
liuhongt [Mon, 16 Aug 2021 03:16:52 +0000 (11:16 +0800)]
Optimize __builtin_shuffle_vector.

1. Support vpermw/vpermb in ix86_expand_vec_one_operand_perm_avx512.
2. Support 256/128-bits vpermi2b ix86_expand_vec_perm_vpermt2.
3. Add define_insn_and_split to optimize specific vector permutation to opmov{dw,wb,qd}.

gcc/ChangeLog:

PR target/101846
* config/i386/i386-expand.c (ix86_expand_vec_perm_vpermt2):
Support vpermi2b for V32QI/V16QImode.
(ix86_extract_perm_from_pool_constant): New function.
(ix86_expand_vec_one_operand_perm_avx512): Support
vpermw/vpermb under TARGET_AVX512BW/TARGET_AVX512VBMI.
(expand_vec_perm_1): Adjust comments for upper.
* config/i386/i386-protos.h (ix86_extract_perm_from_pool_constant):
New declare.
* config/i386/predicates.md (permvar_truncate_operand): New predicate.
(pshufb_truncv4siv4hi_operand): Ditto.
(pshufb_truncv8hiv8qi_operand): Ditto.
* config/i386/sse.md (*avx512bw_permvar_truncv16siv16hi_1):
New pre_reload define_insn_and_split.
(*avx512f_permvar_truncv8siv8hi_1): Ditto.
(*avx512f_vpermvar_truncv8div8si_1): Ditto.
(*avx512f_permvar_truncv32hiv32qi_1): Ditto.
(*avx512f_permvar_truncv16hiv16qi_1): Ditto.
(*avx512f_permvar_truncv4div4si_1): Ditto.
(*avx512f_pshufb_truncv8hiv8qi_1): Ditto.
(*avx512f_pshufb_truncv4siv4hi_1): Ditto.
(*avx512f_pshufd_truncv2div2si_1): Ditto.

gcc/testsuite/ChangeLog:

PR target/101846
* gcc.target/i386/pr101846-2.c: New test.
* gcc.target/i386/pr101846-3.c: New test.
* gcc.target/i386/pr101846-4.c: New test.

3 years agoFortran/OpenMP: Add support for OpenMP 5.1 masked construct
Tobias Burnus [Mon, 16 Aug 2021 07:26:26 +0000 (09:26 +0200)]
Fortran/OpenMP: Add support for OpenMP 5.1 masked construct

Commit r12-2891-gd0befed793b94f3f407be44e6f69f81a02f5f073 added C/C++
support for the masked construct. This patch extends it to
Fortran.

gcc/fortran/ChangeLog:

* dump-parse-tree.c (show_omp_clauses): Handle 'filter' clause.
(show_omp_node, show_code_node): Handle (combined) omp masked construct.
* frontend-passes.c (gfc_code_walker): Likewise.
* gfortran.h (enum gfc_statement): Add ST_OMP_*_MASKED*.
(enum gfc_exec_op): Add EXEC_OMP_*_MASKED*.
* match.h (gfc_match_omp_masked, gfc_match_omp_masked_taskloop,
gfc_match_omp_masked_taskloop_simd, gfc_match_omp_parallel_masked,
gfc_match_omp_parallel_masked_taskloop,
gfc_match_omp_parallel_masked_taskloop_simd): New prototypes.
* openmp.c (enum omp_mask1): Add OMP_CLAUSE_FILTER.
(gfc_match_omp_clauses): Match it.
(OMP_MASKED_CLAUSES, gfc_match_omp_parallel_masked,
gfc_match_omp_parallel_masked_taskloop,
gfc_match_omp_parallel_masked_taskloop_simd,
gfc_match_omp_masked, gfc_match_omp_masked_taskloop,
gfc_match_omp_masked_taskloop_simd): New.
(resolve_omp_clauses): Resolve filter clause.
(gfc_resolve_omp_parallel_blocks, resolve_omp_do,
omp_code_to_statement, gfc_resolve_omp_directive): Handle
omp masked constructs.
* parse.c (decode_omp_directive, case_exec_markers,
gfc_ascii_statement, parse_omp_do, parse_omp_structured_block,
parse_executable): Likewise.
* resolve.c (gfc_resolve_blocks, gfc_resolve_code): Likewise.
* st.c (gfc_free_statement): Likewise.
* trans-openmp.c (gfc_trans_omp_clauses): Handle filter clause.
(GFC_OMP_SPLIT_MASKED, GFC_OMP_MASK_MASKED): New enum values.
(gfc_trans_omp_masked): New.
(gfc_split_omp_clauses): Handle combined masked directives.
(gfc_trans_omp_master_taskloop): Rename to ...
(gfc_trans_omp_master_masked_taskloop): ... this; handle also
combined masked directives.
(gfc_trans_omp_parallel_master): Rename to ...
(gfc_trans_omp_parallel_master_masked): ... this; handle
combined masked directives.
(gfc_trans_omp_directive): Handle EXEC_OMP_*_MASKED*.
* trans.c (trans_code): Likewise.

libgomp/ChangeLog:

* testsuite/libgomp.fortran/masked-1.f90: New test.

gcc/testsuite/ChangeLog:

* gfortran.dg/gomp/masked-1.f90: New test.
* gfortran.dg/gomp/masked-2.f90: New test.
* gfortran.dg/gomp/masked-3.f90: New test.
* gfortran.dg/gomp/masked-combined-1.f90: New test.
* gfortran.dg/gomp/masked-combined-2.f90: New test.

3 years agoRISC-V: Allow multi-lib build with different code model
Kito Cheng [Tue, 20 Jul 2021 02:53:18 +0000 (10:53 +0800)]
RISC-V: Allow multi-lib build with different code model

--with-multilib-generator was only support for different ISA/ABI
combination, however code model is effect the code gen a lots it
should able to handled in multilib mechanism.

Adding `--cmodel=` option to `--with-multilib-generator` to generating
multilib combination with different code model.

E.g.
--with-multilib-generator="rv64ima-lp64--;--cmodel=medlow,medany"
will generate 3 multi-lib suppport:
1) rv64ima with lp64
2) rv64ima with lp64 and medlow code model
3) rv64ima with lp64 and medany code model

gcc/

* config/riscv/multilib-generator: Support code model option for
multi-lib.
* doc/install.texi: Add document of new option for
--with-multilib-generator.

3 years agoDaily bump.
GCC Administrator [Mon, 16 Aug 2021 00:16:32 +0000 (00:16 +0000)]
Daily bump.

3 years agoFortran: fix checks for STAT= and ERRMSG= arguments of SYNC ALL/SYNC IMAGES
Harald Anlauf [Sun, 15 Aug 2021 18:13:11 +0000 (20:13 +0200)]
Fortran: fix checks for STAT= and ERRMSG= arguments of SYNC ALL/SYNC IMAGES

gcc/fortran/ChangeLog:

PR fortran/99351
* match.c (sync_statement): Replace %v code by %e in gfc_match to
allow for function references as STAT and ERRMSG arguments.
* resolve.c (resolve_sync): Adjust checks of STAT= and ERRMSG= to
being definable arguments.  Function references with a data
pointer result are accepted.
* trans-stmt.c (gfc_trans_sync): Adjust assertion.

gcc/testsuite/ChangeLog:

PR fortran/99351
* gfortran.dg/coarray_sync.f90: New test.
* gfortran.dg/coarray_3.f90: Adjust error messages.

3 years agoaix: 64 bit AIX TLS libpthread dependency.
Clément Chigot [Thu, 12 Aug 2021 17:17:15 +0000 (13:17 -0400)]
aix: 64 bit AIX TLS libpthread dependency.

64bit XCOFF files will generated TLS access, with local-exec or
global-exec models, by an access to R13. Thus, there isn't
any reference to a TLS symbol.

The problem is that it allows programs with TLS to be compiled and
linked even without -pthread. Most of the time, it will result in
a segfault when trying to access a TLS variable. But sometimes, it
might create a memory corruption.

This patch forces a reference to __tls_get_addr() to ensure link
will fail without -pthread.

gcc/ChangeLog:
2021-08-11  Clément Chigot  <clement.chigot@atos.net>

* config/rs6000/rs6000.c (xcoff_tls_exec_model_detected): New.
(rs6000_legitimize_tls_address_aix): Use it.
(rs6000_xcoff_file_end): Add ".ref __tls_get_addr" when
xcoff_tls_exec_model_detected is true.

3 years agoImprove many SImode shifts on the H8/300H
Jeff Law [Sun, 15 Aug 2021 04:13:23 +0000 (00:13 -0400)]
Improve many SImode shifts on the H8/300H

As I've mentioned before, the H8/300H can only shift a single bit position at a time.  Naturally this means many shifts are implemented as loops.  There's a variety of special cases that we can do without loops by using rotates, sub-word moves, etc.  The general guidance for the port has been to only use inline or special sequences if they're shorter or just one instruction longer than the loop.

This was pretty reasonable guidance for QI/HI mode.  It was relaxed a bit about 10 years ago for HImode in particular where the kpit team realized they could save 50-100 cycles for some shifts by allowing 2 instructions of code growth over the loop implementation.

But they only re-tuned HImode shifts.  There's even bigger benefits for re-tuning SImode shifts.  There's cases where we can save close to 200 cycles by allowing 2 additional instructions.

This patch re-tunes SImode shifts on the H8/300H primarily by inlining more often or using a special sequence + inlining for residuals.  Both cases were already supported and this just uses those existing capabilities more often, so it was trivial to implement.  I think there's some cases were entirely new special sequences could be used, but I haven't tried those yet.

gcc/

* config/h8300/h8300.c (shift_alg_si): Retune H8/300H shifts
to allow a bit more code growth, saving many dozens of cycles.
(h8300_option_override): Adjus shift_alg_si if optimizing for
code size.
(get_shift_alg): Use special + inline shifts for residuals
in more cases.

3 years agolibgo: various fixes for Solaris support
Ian Lance Taylor [Sat, 14 Aug 2021 00:21:54 +0000 (17:21 -0700)]
libgo: various fixes for Solaris support

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/342189

3 years agoDaily bump.
GCC Administrator [Sun, 15 Aug 2021 00:16:27 +0000 (00:16 +0000)]
Daily bump.

3 years agoor1k: Use cmodel=large when building crtstuff
Stafford Horne [Tue, 20 Apr 2021 20:33:15 +0000 (05:33 +0900)]
or1k: Use cmodel=large when building crtstuff

When linking gcc runtime objects into large binaries the link may fail
with the below errors.  This will happen even if we are building with
-mcmodel=large.

    /home/shorne/work/openrisc/output/host/lib/gcc/or1k-buildroot-linux-uclibc/10.3.0/crtbeginS.o: in function `deregister_tm_clones':
    crtstuff.c:(.text+0x3c): relocation truncated to fit: R_OR1K_GOT16 against undefined symbol `_ITM_deregisterTMCloneTable'
    /home/shorne/work/openrisc/output/host/lib/gcc/or1k-buildroot-linux-uclibc/10.3.0/crtbeginS.o: in function `register_tm_clones':
    crtstuff.c:(.text+0xc0): relocation truncated to fit: R_OR1K_GOT16 against undefined symbol `_ITM_registerTMCloneTable'

This patch builds the gcc crtstuff binaries always with the
-mcmodel=large option to ensure they can be linked into large binaries.

libgcc/ChangeLog:

PR target/99783
* config.host (or1k-*, tmake_file): Add or1k/t-crtstuff.
* config/or1k/t-crtstuff: New file.

3 years agoor1k: Add mcmodel option to handle large GOTs
Stafford Horne [Mon, 5 Apr 2021 20:47:17 +0000 (05:47 +0900)]
or1k: Add mcmodel option to handle large GOTs

When building libgeos we get an error with:

    linux-uclibc/9.3.0/crtbeginS.o: in function `__do_global_dtors_aux':
    crtstuff.c:(.text+0x118): relocation truncated to fit: R_OR1K_GOT16 against symbol `__cxa_finalize' defined in .text section in
    /home/shorne/work/openrisc/3eb9f9d0f6d8274b2d19753c006bd83f7d536e3c/output/host/or1k-buildroot-linux-uclibc/sysroot/lib/libc.so.

This is caused by GOT code having a limit of 64k.  In OpenRISC this
looks to be the only relocation code pattern to be limited to 64k.

This patch allows specifying a new option -mcmodel=large which can be
used to generate 2 more instructions to construct 32-bit addresses for
up to 4G GOTs.

gcc/ChangeLog:

PR target/99783
* config/or1k/or1k-opts.h: New file.
* config/or1k/or1k.c (or1k_legitimize_address_1, print_reloc):
Support generating gotha relocations if -mcmodel=large is
specified.
* config/or1k/or1k.h (TARGET_CMODEL_SMALL, TARGET_CMODEL_LARGE):
New macros.
* config/or1k/or1k.opt (mcmodel=): New option.
* doc/invoke.texi (OpenRISC Options): Document mcmodel.

3 years agoDiagnose mismatches between array and scalar new and delete [PR101791].
Martin Sebor [Sat, 14 Aug 2021 19:25:41 +0000 (13:25 -0600)]
Diagnose mismatches between array and scalar new and delete [PR101791].

Resolves:
PR middle-end/101791 - missing warning on a mismatch between scalar and array forms of new and delete

gcc/ChangeLog:

PR middle-end/101791
* gimple-ssa-warn-access.cc (new_delete_mismatch_p): Use new argument
to valid_new_delete_pair_p.
* tree.c (valid_new_delete_pair_p): Add argument.
* tree.h (valid_new_delete_pair_p): Same.

gcc/testsuite/ChangeLog:

PR middle-end/101791
* g++.dg/warn/Wmismatched-new-delete-6.C: New test.
* g++.dg/warn/Wmismatched-new-delete-7.C: New test.